content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Recently, field theories with higher-order space derivatives have attracted attention, because of the improvement
of graph convergence
\cite{lifshitz}, at the price of violating Lorentz symmetry at high energies.
Ghosts are not introduced by this procedure, since the time derivative order remains minimal, such that
no new pole appears in the propagator of particles.
Also, theories based on anisotropic scaling of space and time (Lifshitz type theories) allow new
renormalizable interactions, and for
example a renormalizable exponential potential, in 3+1 dimensions, has been studied in \cite{LL}. Finally,
a renormalizable Lifshitz type theory of
Gravity has been proposed, which could lead to Quantum Gravity \cite{Horava}.\\
A Lorentz-violating extension of massless Quantum Electrodynamics ($QED$) is discussed here,
where higher order space derivatives are introduced for the photon field.
This model has isotropic scaling in space time though, such that the higher order space derivatives involve a mass scale $M$, whose
role will be discussed a bit further. Higher order space derivatives were already considered in a modified
Dirac equation \cite{Pavlopoulos}, where the resulting phenomenology concerning gamma ray bursts is also discussed.
Higher order space derivatives could arise, for example, from quantum gravitational space time foam \cite{foam}.
The mass $M$ naively suppresses the effect of higher order derivatives in the IR,
but as we will see, an IR signature of these higher orders remains: a fermion mass is generated dynamically.
This mass, although proportional to $M$, is orders of magnitude smaller than $M$
since it is suppressed by an exponentially small function of the fine structure constant.
We will study here this mass generation, using the non-perturbative Schwinger-Dyson approach, and we will discuss
the properties of the result.
We note here that another alternative to the Higgs model has been proposed in the context of anisotropic theories \cite{nohiggs},
and dynamical mass generation has been studied in \cite{fourfermion} for a Lifshitz-type
four-fermion model, and in \cite{yukawa} for a Lifshitz type Yukawa model.\\
We mention here the first works on dynamical mass generation in $QED$
\cite{johnson}, where a relation between the bare and dressed masses is derived, which involves a cut off $\Lambda$. In this
context, it was shown that, in order to have a finite theory, the limit $\Lambda\to\infty$ implies that the bare mass
should vanish, such that the dressed mass must be of dynamical origin. The Schwinger Dyson equation
with a finite cut off was studied later \cite{lambda}, where a critical value for the fine structure constant was found
in order to have dynamical mass generation in the limit $\Lambda\to\infty$. A summary of these results can be found in
\cite{miransky}.\\
Once the fermion dynamical mass generation is demonstrated, we study the perturbative quantization of the model, taking
into account the fermion mass.
We show that the model leads to the same renormalization flows as $QED$, when one considers the running scale $M$. The only difference is
the effective light cone seen by fermions, since space a time derivatives get different
quantum corrections.
We find that the maximum speed for fermions is smaller than the speed of light, by terms of one-loop order,
showing that the model is consistent with causality.
This is not in contradiction with the definition speed of light,
since the latter is obtained in the massless limit, which cannot be taken for fermions, due to dynamical mass generation.
The speed of light is given by the dispersion relation for the gauge field, which is not altered by the Lorentz-violating
model proposed here. \\
We note finally that more general higher order derivative extensions of $QED$ have been
presented in \cite{kostelecky} and references therein.
In these works, the authors consider Lorentz-violating vacuum expectation values for tensor fields,
which allow the introduction of higher order derivatives for the photon field. They explain that the Lorentz-violating Lagrangians
can be written as the Lagrangian for $QED$ in a medium, and they study for example the corresponding birefringence effects of the vacuum.
Our present study corresponds to a specific case of the latter models, where we develop quantum properties of a given Lorentz-violating
extension of $QED$.\\
In Section 2 we define the model and discuss some of its classical features. The fermion dynamical mass is then derived in
Section 3, and the properties of the quantum theory are developed in Section 4, where we first focus on one-loop, and
then develop general arguments for higher-order loops. The Conclusion presents future extensions involving non-Abelian
gauge fields, and an alternative to the Higgs mechanism. Finally, the Appendix contains the details of one-loop
calculations.
\section{Model}
The Lorentz-violating Lagrangian considered here is
\begin{equation}\label{bare}
{\cal L}=-\frac{1}{4}F^{\mu\nu}\left(1-\frac{\Delta}{M^2}\right)F_{\mu\nu}
-\frac{\xi}{2}\partial_\mu A^\mu\left(1-\frac{\Delta}{M^2}\right)\partial_\nu A^\nu
+i\overline\psi\hskip .25cm/\hskip -.25cm D\psi,
\end{equation}
where $D_\mu=\partial_\mu+ieA_\mu$, and $\Delta=-\partial_i\partial^i=\vec\partial\cdot\vec\partial$
(the metric used is (1,-1,-1,-1)),
which recovers $QED$ in a covariant gauge if $M\to\infty$.
The Lorentz-violating terms have two roles: introduce a mass scale, necessary to generate a fermion mass,
and lead to finite gap equation, as will be seen further.
We stress here that $M$ regularizes only loops with an internal photon line, and that
another regularization is necessary to deal with fermion loops.
Also, the Lorentz violating modifications proposed in the Lagrangian (\ref{bare}) do not
alter the photon dispersion relation, which remains relativistic.\\
No higher order space derivatives are introduced for the fermions, for the following reason:
in order to respect gauge invariance, such terms would need to be of the form
\begin{equation}
\frac{1}{M^{n-1}}\overline\psi (i\vec D\cdot\vec\gamma)^n\psi~~~~~~n\ge 2,
\end{equation}
such that new and non-renormalizable couplings would be introduced.\\
\vspace{0.5cm}
\noindent{\it Equations of motion}\\
The classical equation of motion for the free gauge field is
\begin{equation}
\left( 1-\frac{\Delta}{M^2}\right) \left( \Box\eta_{\mu\nu}-\partial_\mu\partial_\nu\right) A^\mu=0~,
\end{equation}
and has two solutions. One is the usual plane wave, with the usual dispersion relation, and the other is solution of
\begin{equation}\label{eqbis}
\left( 1-\frac{\Delta}{M^2}\right) A_\mu=0~.
\end{equation}
If the solution depends on time and one spatial direction $x$ only, it reads
\begin{equation}
A_\mu=K_\mu(t)e^{-Mx}~,
\end{equation}
where $K_\mu(t)$ is an homogeneous vector. If we consider
a spherically symmetric solution, depending on time and the radial spatial coordinate $r$ only, we find
\begin{equation}
A_\mu=K_\mu(t)\frac{e^{-Mr}}{r}~.
\end{equation}
In both cases, the only regular solution occurs when $K_\mu=0$. As a consequence, the only physical solution of the classical equation
of motion for the free gauge field is the usual plane wave.\\
\vspace{0.5cm}
\noindent{\it Lorentz transformation}\\
Finally, one can ask what the effect of a Lorentz transformation is on higher order space derivatives, as in the model (\ref{bare}).
For this, we consider two coordinate systems
$(t,\vec r)$ and $(t',\vec r~')$, with the following transformation law
\begin{eqnarray}\label{Lorentztransfo}
t&=&\gamma(t'+\vec v\cdot\vec r~')\\
\vec r&=&\vec r_\bot~'+\gamma(\vec r_\|~'+\vec v t')\nonumber~,
\end{eqnarray}
where the usual notations are used. In the Lorentz transformation (\ref{Lorentztransfo}), we have
\begin{equation}
-\partial_t^2+\Delta=-\partial_{t'}^2+\Delta'~~~~\mbox{with}~~~~
\partial_{t'}=\gamma\partial_t+\gamma\vec v\cdot\vec\nabla~,
\end{equation}
such that
\begin{equation}\label{Delta'}
\Delta'=\Delta+(\gamma^2-1)\partial_t^2+2\gamma^2\vec v\cdot\vec\nabla\partial_t+\gamma^2(\vec v\cdot\vec\nabla)^2~.
\end{equation}
Using eq.(\ref{Delta'}), the equation of motion $(M^2-\Delta')\phi(t',\vec r~')=0$ for a scalar field
\begin{equation}
\phi(t',\vec r~')=\phi(t,\vec r)=\phi_0\exp(i\omega t-i\vec k\cdot\vec r)~,
\end{equation}
leads then to the dispersion relation
\begin{equation}
\omega^2+2\omega\frac{\vec v\cdot\vec k}{v^2}+\frac{(\vec v\cdot\vec k)^2}{v^2}+\frac{M^2+k^2}{(v\gamma)^2}=0~,
\end{equation}
such that
\begin{equation}
\omega=-\frac{\vec v\cdot\vec k}{v^2}\pm\frac{i}{v\gamma}\sqrt{M^2+k^2-(\vec v\cdot\vec k)^2/v^2}~.
\end{equation}
As a consequence, the corresponding solution is not observable, since it decays in a time of the order of the Plank time
(we do not take into account the non-physical solution increasing exponentially in time).
Therefore the higher order time derivatives introduced by a Lorentz transformation do not affect the model presented here.
\section{Gap equation}
We review here the results of \cite{mdyn}. The Schwinger-Dyson equation for the fermion propagator is \cite{miransky}:
\begin{equation}\label{SD}
G^{-1}-G_{bare}^{-1}=\int D_{\mu\nu}(e\gamma^\mu) G\Gamma^\nu,
\end{equation}
where $\Gamma^\nu$, $G$ and $D_{\mu\nu}$ are respectively the dressed vertex, the dressed fermion propagator
and the dressed photon propagator. This equation gives an exact self consistent relation between dressed
$n$-point functions, and thus is not perturbative.
As a consequence, no redefinition of bare parameters can be done in order to absorb would-be divergences,
and for this reason one needs this equation to be regularized by $M$. \\
In the studies of dynamical mass generation
in $QED$ in the presence of an external magnetic field $B$ (``magnetic catalysis'' \cite{mag}),
the gap equation, in the Lowest Landau Level approximation, is finite because of the mass scale $\sqrt{|eB|}$
- where $e$ is the electric charge - which plays the role of a gauge invariant cut off.
Another example of dynamical mass generation, however of a different nature, is the Debye screening for the photon, at
finite temperature, which was found to be enhanced by a strong magnetic field \cite{Debye}.\\
In order to solve the Schwinger-Dyson equation (\ref{SD}), we consider the so-called ladder (or rainbow) approximation,
consisting in taking the bare vertex instead of $\Gamma^\nu$. It is known that this approximation
is not gauge invariant \cite{miransky}, but, as we will see,
the gauge coupling dependence of the dynamical mass is not affected by the choice of the gauge parameter $\xi$.
Then we will neglect corrections to the photon propagator, which is given by
\begin{equation}\label{D}
D_{\mu\nu}^{bare}(\omega,\vec p)=\frac{i}{1+p^2/M^2}\left( \frac{\eta_{\mu\nu}}{-\omega^2+p^2}+
\zeta\frac{p_\mu p_\nu}{(\omega^2-p^2)^2}\right) ,
\end{equation}
where $\zeta=1/\xi-1$, $p_0=\omega$ and $p^2=\vec p\cdot\vec p$.
Also, we neglect the fermion wave function renormalization:
only the fermion dynamical mass will be taken into account as a correction, such that the dressed fermion propagator
will be taken as
\begin{equation}\label{G}
G(\omega,\vec p)=i\frac{\omega\gamma^0-\vec p\cdot\vec\gamma+m_{dyn}}{\omega^2-p^2-m_{dyn}^2},
\end{equation}
where $m_{dyn}$ is the fermion dynamical mass.
With these approximations, the Schwinger-Dyson equation (\ref{SD}) - involving a convergent integral - leads to
\begin{equation}\label{mdyn=}
m_{dyn}=\frac{\alpha}{\pi^2}\int_{-\infty}^\infty d\omega\int_0^\infty
\frac{p^2dp}{1+p^2/M^2}\frac{m_{dyn}(4+\zeta)}{(\omega^2+p^2)(\omega^2+p^2+m^2_{dyn})}~,
\end{equation}
where $\alpha=e^2/4\pi$ is the fine structure constant.
This equation has the obvious solution $m_{dyn}=0$, and potentially a second solution, which must
satisfy the following gap equation, obtained after integration over the frequency $\omega$,
\begin{equation}\label{intgap}
\frac{\pi}{(4+\zeta)\alpha}=\int_0^\infty\frac{xdx}{1+\mu^2x^2}\left( 1-\frac{x}{\sqrt{1+x^2}}\right),
\end{equation}
where $\mu=m_{dyn}/M$ is the dimensionless dynamical mass, expected to be very small $\mu<<1$, and $x=p/m_{dyn}$.
Both terms in the last equation, if taken separately, lead to diverging integrals, with divergences which cancel each other.
An integration by parts for the second term leads then to
\begin{eqnarray}\label{8}
\frac{2\pi}{(4+\zeta)\alpha}&=&\frac{1}{\mu^2}\int_0^\infty dx\frac{\ln(1+\mu^2x^2)}{(1+x^2)^{3/2}}\nonumber\\
&=&\frac{2}{\mu^2}\left( \ln\left(\frac{\mu}{2}\right)+\frac{\cosh^{-1}(1/\mu)}{\sqrt{1-\mu^2}}\right)\nonumber\\
&=&\ln\left( \frac{2}{\mu}\right) -\frac{1}{2}+{\cal O}(\mu^2\ln\mu)
\end{eqnarray}
and the fermion dynamical mass is finally given by
\begin{equation}\label{mdyn}
m_{dyn}\simeq M\exp\left(-\frac{2\pi}{(4+\zeta)\alpha}\right).
\end{equation}
Note that the expression (\ref{mdyn}) for $m_{dyn}$ is not analytic in $\alpha$,
such that a perturbative expansion cannot lead to such a result, which
justifies the use of a non-perturbative approach. A perturbative expansion would lead to the solution $m_{dyn}=0$ only.\\
Among the two solutions $m_{dyn}=0$ and $m_{dyn}\ne0$, the physical system chooses the non-vanishing dynamical mass,
in order to avoid IR instabilities, not favorable energetically, which would otherwise occur in the theory.
\vspace{0.5cm}
\noindent{\it Gauge dependence} \\
There is an obvious dependence on the gauge parameter $\zeta$, which has a consequence on the value of $m_{dyn}$, but
the important point is the non-analytic $\alpha$-dependence of the dynamical mass, which is not affected by the choice of gauge:
the resulting dynamical mass is of the form $M\exp(-c/\alpha)$, where $c$ is a constant of order 1.
This feature is known in the studies of dynamical mass generation in $QED$ in the presence of an external magnetic field \cite{mag}.\\
It has been argued though, using the pinch technique \cite{pinch}, that for the calculation of physical quantities, which
must be gauge invariant, one should use the Feynman gauge $\zeta=0$ to calculate the fermion self energy. The reason for this is
the cancellation of terms arising from longitudinal contributions in the photon propagator.
Nevertheless, if $M$ is the Plank mass, the Feynman gauge for the
dynamical mass (\ref{mdyn}) leads to a too small result for the electron mass. But a multibrane scenario has been proposed in
\cite{branes}, where a Randal-Sundrum warp factor enhances the dynamical mass, which then acquires the right order of magnitude
for the electron, when $M$ is the Plank mass.
\section{Perturbative analysis of the model}
Taking into account the fermion mass $m_{dyn}$,
we discuss here the perturbative quantization of the model, first at a one-loop, and we give then the general arguments
at higher order loops.
\subsection{One-loop properties}
{\it Fermion kinetic term}\\
We calculate in the Appendix the one-loop quantum corrections for the fermion kinetic term, which are different for time and
space derivatives. We find the corrections
\begin{equation}
i\overline\psi \left((1+Z_0)\partial_0\gamma^0-(1+Z_1)\vec\partial\cdot\vec\gamma\right) \psi~,
\end{equation}
with
\begin{eqnarray}
Z_0&=&-\frac{\alpha}{2\pi}\left( \ln\left( \frac{1}{\mu}\right) +4\ln2-2\right) +{\cal O}(\mu^2\ln(1/\mu)) \nonumber\\
Z_1&=&-\frac{\alpha}{2\pi}\left( \ln\left( \frac{1}{\mu}\right) +\frac{50}{9}-\frac{20}{3}\ln2\right)+{\cal O}(\mu^2\ln(1/\mu))~.
\end{eqnarray}
Note that the dominant term, proportional to $\ln(1/\mu)$, is the same for $Z_0$ and $Z_1$, since in the
Lorentz symmetric situation ($M\to\infty$ for fixed fermion mass), we have $Z_0=Z_1$.
Also, the coefficient $-\alpha/(2\pi)$ in front
of the dominant term $\ln(1/\mu)$ is the coefficient found in $QED$ in $4-\epsilon$ dimensions
\begin{equation}
Z_0^{QED}=Z_1^{QED}=-\frac{\alpha}{2\pi\epsilon}+\mbox{finite}~.
\end{equation}
An important remark should be made here: because of the result (\ref{mdyn}), the ratio $\mu$ is actually finite in the limit where
$M\to\infty$, and one could think that no counter term is necessary to absorb terms proportional to $\ln(1/\mu)$.
But it is in fact necessary, in order
to respect the loop structure. Indeed, if one keeps $\ln(1/\mu)$-terms in the renormalized theory,
the would-be one-loop correction would become a tree-level one:
\begin{equation}
\alpha\ln\left( \frac{1}{\mu}\right) ={\cal O}(1)~.
\end{equation}
Therefore, provided one treats $\ln(1/\mu)$ as a $1/\epsilon$ term in dimensional regularization, one gets the usual $QED$ one-loop
structure.
\vspace{0.5cm}
\noindent{\it Maximum speed for fermions}\\
After redefinition of the bare parameters in the minimal substraction scheme, where
only the term proportional to $\ln(1/\mu)$ is absorbed, the fermion dispersion relation is
\begin{equation}
\left(1-\frac{\alpha}{\pi}[2\ln2-1]\right)^2\omega^2=\left(1-\frac{\alpha}{\pi}[25/9-(10/3)\ln2]\right)^2p^2+m_{dyn}^2~,
\end{equation}
and the product of the fermion phase velocity $v_\phi$ and group velocity $v_g$ is then
\begin{eqnarray}\label{phase}
v^2&\equiv&v_\phi v_g=\frac{\omega}{p}\frac{d\omega}{dp}\\
&=&\left( \frac{1-(\alpha/\pi)[25/9-(10/3)\ln2]}{1-(\alpha/\pi)[2\ln2-1]}\right)^2\nonumber\\
&=&1-\frac{2\alpha}{\pi}\left(\frac{34}{9}-\frac{16}{3}\ln2\right) +{\cal O}(\alpha^2)~<1~,
\end{eqnarray}
which shows that the effective light cone seen by fermions is consistent with causality. If we take $\alpha\simeq1/137$, we obtain
for the fermion maximum speed
\begin{equation}\label{vpvg}
v\simeq1-1.9\times10^{-4}~.
\end{equation}
\vspace{0.5cm}
\noindent{\it Gauge invariance and speed of light}\\
We calculate in the Appendix the one-loop vertex for vanishing incoming momentum, and we find the corrections
\begin{eqnarray}
\Gamma_{(1)}^0&=&Z_0e\gamma^0\nonumber\\
\Gamma_{(1)}^i&=&Z_1e\gamma^i~,
\end{eqnarray}
such that gauge invariance is respected, since quantum corrections to the vertex are identical with corrections to the fermion
kinetic term, for time and space components independently.
As a consequence, as in usual $QED$, quantum corrections to the coupling constant are given by the polarization tensor, and arise
only from the wave function renormalization of the gauge field. Since the one-loop polarization tensor
doesn't include any internal photon line, the one-loop running of the coupling constant is therefore also the
same as in $QED$.\\
Because $Z_0\ne Z_1$ and therefore the effective light cone seen by fermions involves the speed $v<1$, one
might see a problem with the definition of speed of light. We argue here that it is not the case, because of dynamical
mass generation for the fermion. Indeed, the speed of light $c$ is defined by
\begin{equation}\label{limit}
c=\lim_{m\to0}~\frac{\omega}{|\vec p|}~,
\end{equation}
for finite momentum $\vec p$ and frequency $\omega$.
But because the fermion is always massive, $m=m_{dyn}\ne 0$, the limit (\ref{limit}) cannot be taken, and the result (\ref{vpvg}) is
not in contradiction with the speed of light.
Such a conclusion was already obtained in \cite{yukawa} for a Lifshitz-type Yukawa interaction.\\
The speed of light is given by the gauge field dispersion relation, which is not modified at one-loop.
We show in the next subsection that this conclusion still holds at higher order loops.
\subsection{Higher order loops}
\noindent{\it Perturbative expansion}\\
We have seen at one-loop that, in order to recover the $QED$ renormalization structure, one needs to absorb terms proportional
to $\ln(1/\mu)$ in the redefinition of bare parameters. If we assume that, after including the appropriate counter terms,
we have a finite theory at $n-1$ loops, a graph calculated at $n$ loops will be of the form
\begin{equation}\label{finite}
G^{(n)}=P^{(n)}\frac{M^\epsilon}{\epsilon}+Q^{(n)}\ln\left( \frac{1}{\mu}\right)+\mbox{finite}~,
\end{equation}
where, compared to $(n-1)-$loop graphs:\\
$P^{(n)}$ contains graphs with an additional fermion loop, which needs to be regularized
dimensionally;\\
$Q^{(n)}$ contains graphs with an additional photon line, which are regularized by $M$. \\
The $n$-loop counter term then needs to absorb terms proportional to both $1/\epsilon$ and $\ln(1/\mu)$. The
contribution to the renormalization group flow is then
\begin{equation}
\lim_{\epsilon\to 0}\left\lbrace M\frac{\partial G^{(n)}}{\partial M}\right\rbrace =P^{(n)}+Q^{(n)}+{\cal O}(\mu\ln\mu)~,
\end{equation}
and corresponds to the usual result obtained in $QED$, since we have
\begin{equation}
G^{(n)}_{QED}=\left( P^{(n)}+Q^{(n)}\right) \frac{M^\epsilon}{\epsilon}+\mbox{finite}~,
\end{equation}
where the finite terms differ from those in eq.(\ref{finite}).
We stress here that the partial derivative with respect to $M$
should be taken at {\it fixed} fermion mass.
\vspace{0.5cm}
\noindent{\it Speed of light}\\
If one considers two-loop properties of the model or higher orders, the polarization tensor is affected by Lorentz violation,
unlike in the
one-loop case, and it is necessary to make sure that the model remains consistent, especially as far as gauge invariance
and speed of light are concerned.\\
From two loops and above, the field strength gets different corrections for time and space derivatives, and we obtain
\begin{equation}
2(1+Y_0)F_{0i}F^{0i}+~(1+Y_1)F_{ij}F^{ij}~,
\end{equation}
where $Y_0$ and $Y_1$ represent the finite quantum corrections to the operators $F_{0i}^2$ and $F_{ij}^2$ respectively,
after absorbing the regularization terms proportional to $1/\epsilon$ or $\ln(1/\mu)$.
In order to obtain corrections proportional to the Lorentz scalar $F_{\mu\nu}F^{\mu\nu}$, which is necessary to recover gauge
invariance and the speed of light $c=1$, we rescale the time coordinate and the component $A_0$ as:
\begin{equation}
t~\to~\frac{t}{\kappa}~~~~~\mbox{and}~~~~~~A_0~\to~\kappa A_0~~~~~\mbox{where}~~~~~~\kappa=\sqrt{\frac{1+Y_1}{1+Y_0}}~.
\end{equation}
One can easily see then, that this rescaling is consistent with gauge invariance of the fermion sector:
\begin{equation}
\overline\psi\left( i\partial_0-eA_0\right) \gamma^0\psi~\to~\kappa~\overline\psi\left( i\partial_0-eA_0\right) \gamma^0\psi
\end{equation}
The factor $\kappa$, which does not appear in the space components of the covariant derivatives, will then
contribute to the maximum speed for fermions, together with the corrections to the fermion kinetic terms,
as explained for the one-loop case. A final identical rescaling for all the gauge field components, by the factor $\sqrt{1+Y_1}$,
will lead to the redefinition of the coupling constant.
\section{Conclusion}
We have shown that Lorentz-violating higher-order derivatives for an Abelian gauge field lead to a fermion dynamical mass,
independently of the strength of the coupling constant, and that it also leads
to a consistent quantum theory. The order of magnitude of the fermion dynamical mass is too small for the electron, though, if
one identifies $M$ with the Plank mass, and this is a problem which could be solved with the introduction of a non-Abelian
gauge field, as explained below.\\
The main difference with $QED$ is the maximum speed for fermions, which is smaller than
the speed of light. This implies that the fermions see a different effective light cone, but it is not in contradiction with
the definition of speed of light, since the fermion is always massive.\\
The next step is to extend this study to non-Abelian gauge theories. It is known that vector fields coupled to the axial current
acquire a mass when a fermion condensate forms \cite{vectormass}, as the result of the massless bound state excitation, which
implies a pole in the polarization tensor for vanishing external momentum. The residue of this pole is the vector mass squared.
In the framework of our present model, the fermion condensate can occur, due to Lorentz-violation
in the Abelian sector, and then generate dynamically the non-Abelian vector masses.
We note that, in order not to break the non-Abelian
gauge symmetry at the classical level, a Lorentz-violating non-Abelian dynamics
would involve higher order {\it covariant}-space derivatives, and would therefore introduce new interactions,
which are not renormalizable. For this reason, we are planning to leave the non-Abelian sector unchanged, and
introduce Lorentz-violation in the Abelian sector only.\\
Another consequence of the coupling of fermions with several gauge fields will be to change the order of
magnitude of the fermion dynamical mass. Indeed, as was already seen in \cite{hosek}, where a similar study was done,
the contributions of different gauge fields lead to a partial
cancellation in the exponent in the exponential
of eq.(\ref{mdyn}), giving a phenomenologically more realistic dynamical mass.
\vspace{1cm}
\noindent{\bf Acknowledgments} This work is partly supported by the Royal Society, UK.
\section*{Appendix: One-loop Feynman graphs}
We calculate here one-loop corrections to the model (\ref{bare}), assuming that the fermion has mass $m$, which is dynamically generated.
The metric we use throughout is $(+,-,-,-)$.
\vspace{0.5cm}
\noindent{\it Fermion kinetic terms}\\
The fermion wave function renormalization is obtained by differentiating the fermion self energy with respect to the external
momentum $\nu,\vec k$, and then set $\nu=0, \vec k=0$. The fermion self energy is
\begin{eqnarray}
&&\Sigma^{(1)}(\nu,\vec k)\\
&=&i(ie)^2\int\frac{d\omega}{2\pi}\int\frac{d^3\vec p}{(2\pi)^3}\frac{\gamma^\mu(\hskip .25cm/\hskip -.25cm p+\hskip .25cm/\hskip -.25cm k+m)\gamma_\mu}{(1+p^2/M^2)(\omega^2-p^2)
[(\omega+\nu)^2-(\vec p+\vec k)^2-m^2]}\nonumber\\
&=&-i\frac{e^2}{(2\pi^4)}\int d\omega\int\frac{d^3\vec p}{1+p^2/M^2}\frac{4m-2(\hskip .25cm/\hskip -.25cm p+\hskip .25cm/\hskip -.25cm k)}{(\omega^2-p^2)[(\omega+\nu)^2-(\vec p+\vec k)^2-m^2]}
\nonumber~.
\end{eqnarray}
We write
\begin{equation}
\Sigma^{(1)}(\nu,\vec k)=Z_0\nu\gamma^0-Z_1\vec k\cdot\vec\gamma+{\cal O}(k^2)~,
\end{equation}
and a derivative with respect to $k^\rho$ gives, after setting $\nu=0, \vec k=0$,
\begin{eqnarray}
&&\left.\frac{\partial\Sigma^{(1)}}{\partial k_\rho}\right|_{k=0}\\
&=&\frac{2ie^2}{(2\pi)^4}\int\frac{d^3\vec p}{1+p^2/M^2}\int\frac{d\omega}{\omega^2-p^2}\left(\frac{\gamma^\rho}{\omega^2-p^2-m^2}
-\frac{2p^\rho\hskip .25cm/\hskip -.25cm p}{(\omega^2-p^2-m^2)^2}\right) \nonumber
\end{eqnarray}
For $\rho=0$, we obtain, after a Wick rotation,
\begin{equation}
Z_0=-\frac{e^2}{2\pi^3}\int\frac{p^2dp}{1+p^2/M^2}\int \frac{d\omega}{\omega^2+p^2}
\frac{-\omega^2+p^2+m^2}{(\omega^2+p^2+m^2)^2}
\end{equation}
and integration over the frequency $\omega$ gives
\begin{eqnarray}
Z_0&=&-\frac{e^2}{2\pi^2}\int_0^\infty\frac{x^2dx}{1+\mu^2x^2}\left(2x+\frac{1}{x}-2\sqrt{1+x^2}\right) \nonumber\\
&=&-\frac{2\alpha}{\pi}\left(\frac{1}{4}\ln(1/\mu) +\ln2 -\frac{1}{2}\right)+{\cal O}(\mu^2\ln(1/\mu)) ,
\end{eqnarray}
where $\mu=m/M<<1$ and $x=p/m$.\\
For $\rho=i$, we obtain, after a Wick rotation,
\begin{equation}
Z_1=-\frac{e^2}{2\pi^3}\int\frac{p^2dp}{1+p^2/M^2}\int\frac{d\omega}{\omega^2+p^2}\frac{\omega^2+p^2/3+m^2}{(\omega^2+p^2+m^2)^2}~,
\end{equation}
where the constant 1/3 in factor of $p^2$ in the numerator comes from the Galilean symmetry in the 3-dimensional space.
The integration over $\omega$ gives
\begin{eqnarray}
Z_1&=&-\frac{e^2}{2\pi^2}\int_0^\infty\frac{x^2dx}{1+\mu^2x^2}\left(-\frac{2x}{3}+\frac{1}{x}
+\frac{2x^2/3-1}{\sqrt{1+x^2}}+\frac{x^2/3}{(1+x^2)^{3/2}}\right) \nonumber\\
&=&-\frac{2\alpha}{\pi}\left(\frac{1}{4}\ln(1/\mu)+\frac{25}{18}-\frac{5}{3}\ln2\right)+{\cal O}(\mu^2\ln(1/\mu)) ~.
\end{eqnarray}
\vspace{0.5cm}
\noindent{\it Vertex}\\
The one-loop correction to the vertex is
\begin{eqnarray}
\Gamma_{(1)}^\rho
&=&(ie)^3\int\frac{d\omega}{2\pi}\int\frac{d^3\vec p}{(2\pi)^3}\frac{1}{1+p^2/M^2}
\frac{\gamma^\mu(\hskip .25cm/\hskip -.25cm p+m)\gamma^\rho(\hskip .25cm/\hskip -.25cm p+m)\gamma_\mu}{(\omega^2-p^2)(\omega^2-p^2-m^2)^2}\nonumber\\
&=&-\frac{e^3}{4\pi^3}\int \frac{p^2dp}{1+p^2/M^2}\int d\omega\frac{2\gamma^\rho(\omega^2-p^2)-4p^\rho\hskip .25cm/\hskip -.25cm p-2m^2\gamma^\rho}
{(\omega^2-p^2-m^2)^2(\omega^2-p^2)}~.
\end{eqnarray}
For $\rho=0$, we obtain after a Wick rotation
\begin{eqnarray}
\Gamma_{(1)}^0&=&-\gamma^0\frac{e^3}{2\pi^3}\int \frac{p^2dp}{1+p^2/M^2}\int d\omega
\frac{-\omega^2+p^2+m^2}{(\omega^2+p^2)(\omega^2+p^2+m^2)^2}\nonumber\\
&=&Z_0e\gamma^0~.
\end{eqnarray}
For $\rho=i$, we obtain after a Wick rotation
\begin{eqnarray}
\Gamma_{(1)}^i&=&-\gamma^0\frac{e^3}{2\pi^3}\int \frac{p^2dp}{1+p^2/M^2}\int d\omega
\frac{\omega^2+p^2/3+m^2}{(\omega^2+p^2)(\omega^2+p^2+m^2)^2}\nonumber\\
&=&Z_1e\gamma^i~.
\end{eqnarray}
|
\section{Introduction: The $\mu$-$b_\mu$ problem in gauge mediation of supersymmetry breaking}
\label{section_Intro}
Despite the many virtues that make theories with supersymmetry (SUSY) one of the most appealing candidates to extend the Standard Model (SM) at high energies, these are not absent of their own problems. The most obvious one is to explain why nature is not supersymmetric at low energies, i.e., to provide a satisfactory mechanism of SUSY breaking. This problem is intimately related with the predictive power of SUSY. Indeed, even for the minimal supersymmetric extension of the SM, the MSSM, with softly broken SUSY there are far too many free parameters to be determined experimentally before we can start making general predictions. One expects that once we know the details of the sector responsible of SUSY breaking all those parameters can be expressed in term of (hopefully) only a few (more fundamental) quantities. Models with Gauge Mediation of Supersymmetry Breaking (GMSB) \cite{Dine:1981gu,Nappi:1982hm,AlvarezGaume:1981wy,Dine:1993yw,Dine:1994vc,Dine:1995ag,Giudice:1998bp}
offer a complete pattern of soft interactions computable within a renormalizable framework in terms of very few parameters. In GMSB one assumes that the MSSM fields feel supersymmetry breaking through the SM gauge interactions of a set of messenger fields, $\Phi$, directly coupled to a hidden sector where SUSY breaking occurs. The latter is parametrized by a gauge-singlet chiral auxiliary superfield, $X$, whose vacuum expectation value (vev) $\left<X\right>=M+\theta^2 F$ is assumed to be the only source of SUSY breaking. Through superpotential interactions $W\supset X\Phi^2$, $\left<X\right>$ generates a splitting $\sim \sqrt{F}$ between the masses (of order $M$) of the scalar and fermionic components of the messenger fields. Those effects are translated into the visible sector through radiative corrections. In this way, gaugino masses are generated at the one-loop order while scalar masses squared receive contributions only through two-loop graphs:
\be
M_{\lambda_a} \sim \frac{g_a^2}{16\pi^2}\frac{F}{M},~~~~~~~~m_{\phi_i}^2\sim\sum_a\frac{g_a^4}{256\pi^4}\frac{F^2}{M^2}\sim M_\lambda^2.
\ee
Notice that, since gauge interactions are flavor blind, the gauge-mediation contributions to the scalar masses for the sfermions are family universal. Thus, GMSB provides a scenario where one can easily avoid introducing new sources of flavor violation in the low-energy theory. Together with its predictive power, this is another of the reasons why gauge mediation has become one of the most popular mechanisms of SUSY breaking.
Another problem common to all supersymmetric extensions of the SM is the origin of $\mu$, the Higgs bilinear term in the superpotential, $W\supset \mu \left(H_u\cdot H_d\right)$ (with ``$\cdot$'' being the $SU\left(2\right)_L$ product). For phenomenological reasons $\mu$ must be of the order of the weak scale and therefore it should not be present in the limit of exact supersymmetry, for it would be naturally of the order of the Planck or another large fundamental scale. Thus, $\mu$ must be generated upon SUSY breaking. The problem in GMSB is that the same interactions generating $\mu$ also generate $b_\mu$, the corresponding Higgs bilinear soft term in the scalar potential, $V\supset b_\mu \left(H_u\cdot H_d\right)+\hc$, in a way such that~\cite{Dvali:1996cu}
\be
\frac{b_\mu}{\mu}\sim \frac{F}{M}.
\label{mubmu}
\ee
Since $F/M\gtrsim 10$ - $100\units{TeV}$, in order to generate adequate masses for the supersymmetric particles, Eq. (\ref{mubmu}) requires an unnatural fine tuning to reproduce electroweak symmetry breaking (EWSB).
A simple solution to this naturalness problem can be achieved within the Next-to-Minimal Supersymmetric Standard Model (NMSSM). A bare $\mu$ term is forbidden in the NMSSM by imposing a discreet $\mathbb{Z}_3$ symmetry, broken only at the electroweak scale.\footnote{The introduction of such a symmetry has been criticized because of the potential cosmological problems due to the formation of stable domain walls. This kind of discussion, however, goes beyond the scope of this paper. One possibility for avoiding this problem can be found in \cite{Panagiotakopoulos:1998yw}.} An effective $\mu$ parameter can then be generated at low energies, by means of the vev of the new gauge-singlet chiral field, $S$, with superpotential couplings
\be
W\supset\lambda S \left(H_u\cdot H_d\right)-\frac{\kappa}{3}S^3.
\label{NMSSM ext}
\ee
Thus, the low energy dynamics generate $\mu^\mathrm{Eff}=\lambda \left<S\right>$. On the other hand, we have $b_\mu^\mathrm{Eff}=a_\lambda\left<S\right>+\lambda\kappa\left<S\right>^2$, with $a_\lambda$ the Higgs-singlet trilinear soft coupling in the scalar potential: $V \supset a_\lambda S \left(H_u\cdot H_d\right)+\hc$. Since the leading gauge-mediation contributions to $a$ terms arise at the two-loop level, and then are highly suppressed, $b_\mu^\mathrm{Eff}\sim \lambda\kappa\left<S\right>^2$ up to renormalization group (RG) running effects. Therefore, $b_\mu^\mt{Eff}/\mu^\mt{Eff}\sim \kappa \left<S\right>$, avoiding the hierarchy in (\ref{mubmu}). Note that a sizable value of the cubic coupling $\kappa$ is required to avoid the presence of too light pseudoscalars in the theory, for it explicitly breaks the global Peccei-Quinn symmetry of the NMSSM.
This minimal scenario with GMSB, however, is known to have problems to attain the correct EWSB and, at the same time, generate a phenomenologically viable spectrum~\cite{deGouvea:1997cx}. In particular, this requires to generate a large enough negative mass squared for the singlet, $m_S^2$, or large $a$ terms for the superpotential interactions in Eq.~(\ref{NMSSM ext}), $a_\lambda$ and $a_\kappa$. Neither of these can be obtained within pure GMSB since $S$ carries no gauge quantum numbers and, as explained above, $a$ terms are generically small in gauge mediation.
In Ref.~\cite{Giudice:1997ni} it was proposed that if, apart from gauge interactions, SUSY breaking is communicated also by means of direct superpotential interactions of the messengers with the NMSSM singlet $S$, negative values for $m_S^2$ and non-negligible contributions to $a_\lambda$ and $a_\kappa$ can be obtained. The phenomenological implications of such scenario were explored in \cite{Delgado:2007rz}, finding that EWSB is then compatible with a realistic spectrum. This scenario thus provides a viable solution to the $\mu$-$b_\mu$ problem of gauge mediation.\footnote{In this case $a_\lambda\sim \frac{\lambda}{16 \pi^2} \frac{F}{M}$ so, because of the loop factor, we can still avoid (\ref{mubmu}).} An alternative scenario where the singlet, the messengers and the Higgs doublets are coupled was discussed in \cite{Chacko:2001km}. Again, this provides sizable contributions to the soft $a$ terms and $m_S^2$, but the latter require $\left|\kappa\right|\sim 1$ at the messenger scale to obtain a negative value. Indeed, the characteristic feature in both scenarios is the use of singlet interactions to generate the required extra contributions to the soft terms for $S$, that then deviate from the standard gauge mediation scenario. Hence the name of singlet deflection of gauge mediation.
As explained in \cite{Delgado:2007rz}, one of the main phenomenological difficulties of the scenario with messengers coupled only to $S$ is to obtain regions in the parameter space where the lightest CP-even Higgs boson, $H^1$, passes the direct LEP 2 bound of 114 GeV~\cite{Barate:2003sz}. This has been one of the major problems of supersymmetric extensions of the SM, where the Higgs is naturally very light at tree level and one relies in large contributions from radiative corrections to lift the leading order value above the LEP 2 bound. In general, attaining a Higgs with masses above 114 GeV within this model required to assume both a large SUSY-breaking scale $F/M$ and a large messenger scale $M$ in order for top and stop loops to give such large radiative corrections to $m_{H^1}$. This lead to a heavy spectrum where, for instance, stops were required to have masses around 2 TeV. As we will illustrate, one can lower the mass scale of that model and still find regions in the parameter space where all experimental restrictions are passed, but such regions are quite small thus requiring a significant fine tuning.
Motivated by this tension between obtaining a relatively light spectrum and still having large allowed regions in the parameter space, in this paper we combine the ideas of \cite{Delgado:2007rz} and \cite{Chacko:2001km} to explore the scenario where deflection of gauge mediation is due to general interactions involving the singlet and the messenger fields. We show that, despite this more general scenario introduces several parameters with respect to \cite{Delgado:2007rz}, only one of them suffices to greatly increase the size of the allowed regions in the parameter space. Thus, the fine tuning problem for low SUSY-breaking scales is removed. In particular, this allows a TeV scale spectrum which may be eventually tested by the LHC.
The paper is organized as follows. In the next section we present the model and provide the soft SUSY-breaking terms in the effective theory below the messengers mass threshold. Section 3 explains the method employed in our phenomenological analysis of the model. We scan the parameter space searching for regions where EWSB can be reproduced. The resulting spectrum for those regions is detailed in section 4. Finally we present our conclusions.
\section{Description of the model}
\label{sec:Model}
We use the following conventions in writing the superpotential couplings and soft $a$ terms of the $\mathbb{Z}_3$-symmetric NMSSM:
\be
W=-y_t~u_3^c \left(H_u \cdot q_3\right) +y_b~d_3^c \left(H_d \cdot q_3\right)+y_\tau~e_3^c \left(H_d \cdot l_3\right) +\lambda~S \left(H_u\cdot H_d\right)-\frac{\kappa}{3}~S^3,
\label{NMSSMW}
\ee
\be
{\cal L}_\mathrm{soft}\supset a_t~u_3^c\left(H_u \cdot q_3\right) -a_b~d_3^c\left(H_d \cdot q_3\right)-a_\tau~e_3^c\left(H_d \cdot l_3\right) -a_\lambda~S \left(H_u\cdot H_d\right)+\frac{a_\kappa}{3}~S^3+\hc,
\ee
where, as usual, we have neglected Yukawa couplings other than those for the third SM family.
Following \cite{Delgado:2007rz}, above the messenger scale, $M$, we will assume $n=2$ pairs of messenger fields $\Phi_i$ and $\bar{\Phi}_i$, $i=1,2$, transforming as a 5 and a $\bar{5}$ of $SU\left(5\right)$, respectively. These fields decompose into a triplet and a doublet under $SU\left(3\right)_c$ and $SU\left(2\right)_L$, respectively: $\Phi=\left(\Phi^T,\Phi^D\right)$ and couple to the spurion field, $X=M+\theta^2 F$, responsible of SUSY breaking. This is considered only as a background nondynamical field. Apart from gauge interactions the messenger fields can also have direct superpotential couplings with the NMSSM fields, thus altering the gauge-mediation pattern of soft SUSY-breaking contributions. We consider the following superpotential couplings for the messenger fields at $M$
\be
\begin{split}
W_\mt{mess}&=X\sum_{i=1}^{n=2}\left(\kappa_i^D \bar{\Phi}^D_i\Phi^D_i+\kappa_i^T \bar{\Phi}^T_i\Phi^T_i\right)+\\
&+S\left(\xi_D \bar{\Phi}^D_1\Phi^D_2+\xi_T \bar{\Phi}^T_1\Phi^T_2+\xi_{H_u} \bar{\Phi}_1^D H_u+\xi_{H_d} \left(H_d\cdot \Phi_2^D\right)\right).
\end{split}
\label{MessW}
\ee
The structure of the above superpotential can be explained under the following asumptions. As stressed in \cite{Delgado:2007rz}, the structure of the singlet-messenger couplings can be explained by extending the discrete $\mathbb{Z}_3$ symmetry of the NMSSM with $\mathbb{Z}_3[\Phi_1]=\mathbb{Z}_3[\bar{\Phi}_2]=-1/3$, $\mathbb{Z}_3[\Phi_2]=\mathbb{Z}_3[\bar{\Phi}_1]=1/3~(=\mathbb{Z}_3[S]=\mathbb{Z}_3[H_u]=\mathbb{Z}_3[H_d])$, $\mathbb{Z}_3[X]=0$. Let us note that the couplings in (\ref{MessW}) generate a kinetic $\bar{\Phi}_1^D$-$H_d$ and $\Phi_2^D$-$H_u$ mixing at the one-loop level. Indeed, we have the following contributions to the off-diagonal elements of the matrix of anomalous dimensions $\gamma$~\footnote{~$\gamma_i^j\equiv d \log{Z_i^j}/d\log{Q}$, with $Z_i^j$ the corresponding wavefunction renormalization.} above the messenger scale:
\be
\gamma_{\bar{\Phi}_1^D}^{H_d}=-\frac{1}{8\pi^2}\left(\lambda \xi_{H_u}+\xi_D\xi_{H_d}\right),
\label{WFR1}
\ee
\be
\gamma_{\Phi_2^D}^{H_u}=-\frac{1}{8\pi^2}\left(\lambda \xi_{H_d}+\xi_D\xi_{H_u}\right).
\label{WFR2}
\ee
Thus, even if one of the couplings $\xi_D$, $\xi_{H_u}$ or $\xi_{H_d}$ vanishes at a given scale $Q\geq M$, a nonzero value is generated through renormalization above the messenger scale. This is not in contradiction with the nonrenormalization theorem, since the effect comes from the wavefunction renormalization. The above $\mathbb{Z}_3$ symmetry still allows for the possibility of including terms like $X \bar{\Phi}_2^D H_u $ and $X \left(H_d \cdot \Phi_1^D\right)$. Upon SUSY breaking, these introduce a mixing between the NMSSM Higgs fields and the doublet components of the messenger fields. Such mixing should be diagonalized in order to obtain the physical Higgses and messenger doublets, giving rise to additional interactions in the superpotential between the physical messengers and the NMSSM fields \cite{Chacko:2001km}. Note that, because of the mixing induced by (\ref{WFR1}) and (\ref{WFR2}), RG effects above $M$ would also generate the above-mentioned terms, even if we assume they vanish at a given scale. In this regard, for simplicity and to avoid the proliferation of new parameters controlling the contributions to soft terms we will assume there is no such mixing at the messenger scale. Thus, in the above superpotential $H_{d,u}$ describe the physical Higgses and these correspond to the NMSSM ones. Genuine trilinear couplings between the doublet components of the messengers and the quark/squark or lepton/slepton chiral fields could be also present:
\be
\Delta W_\mt{mess}=- \xi_u^{ij}~u_i^c \left(\Phi_2^D \cdot q_j\right) +\xi_d^{ij}~d_i^c~\bar{\Phi}_1^D q_j+\xi_e^{ij}~e_i^c~\bar{\Phi}_1^D l_j.
\label{MessW2}
\ee
Should they be present, a rather unnatural alignment of such couplings with the Yukawa ones would be required in order to avoid introducing extra sources of flavor violation. On the other hand, this kind of interactions can be forbidden within extra dimensional models by locating in different branes the MSSM quarks and leptons, and the hidden and messenger fields. Here, again for simplicity, we will just assume that the couplings in (\ref{MessW2}) vanish at the messenger scale. In this case, for energy scales $Q>M$ the wavefunction renormalizations (\ref{WFR1}) and (\ref{WFR2}) would generate nonzero values only for the couplings with the third family~\footnote{In turn, these feed (\ref{WFR1}) and (\ref{WFR2}) with extra contributions.} ($\xi_t\equiv \xi_u^{33}$, $\xi_b\equiv\xi_d^{33}$ and $\xi_\tau\equiv \xi_e^{33}$). Note also that in (\ref{MessW}) one can write similar couplings replacing $H_d$ by the left-handed lepton chiral fields $l_i$. These can be avoided by simply requiring $R$-parity conservation. Finally, just to mention that, as also stressed in \cite{Delgado:2007rz}, the two messengers are required in order to avoid kinetic mixing between $X$ and $S$, which could destabilize the weak scale. The structure of the minimal superpotential in Eq. (\ref{MessW}) is preserved if we include extra messenger fields, as long as they come in pairs and we enlarge consequently the assignments of the $\mathbb{Z}_3$ symmetry.
Upon integrating the messenger fields out, the couplings in Eq. (\ref{MessW}) generate the following contributions to the soft parameters at $Q=M$. These can be derived using the general method described in \cite{Giudice:1997ni,ArkaniHamed:1998kj}. For the scalar soft masses squared the contributions at the leading order in $F^2/M^2$ arise at two loops. These deviate from the gauge-mediation contributions by terms controlled by both the new interactions in (\ref{MessW}) and the standard superpotential couplings for each particle:
\be
\begin{split}
m_{q_3}^2=&\frac{1}{256 \pi^4}\left(\frac{g_1^4}{15}+3g_2^4+\frac{16}{3}g_3^4-y_b^2\xi_{H_d}^2-y_t^2\xi_{H_u}^2\right)\frac{F^2}{M^2},\\
m_{u^c_3}^2=&\frac{1}{256 \pi^4}\left(\frac{16}{15}g_1^4+\frac{16}{3}g_3^4-2y_t^2\xi_{H_u}^2\right)\frac{F^2}{M^2},\\
m_{d^c_3}^2=&\frac{1}{256 \pi^4}\left(\frac{4}{15}g_1^4+\frac{16}{3}g_3^4-2y_b^2\xi_{H_d}^2\right)\frac{F^2}{M^2},\\
m_{l_3}^2=&\frac{1}{256 \pi^4}\left(\frac{3}{5}g_1^4+3 g_2^4-y_\tau^2\xi_{H_d}^2\right)\frac{F^2}{M^2},\\
m_{e^c_3}^2=&\frac{1}{256 \pi^4}\left(\frac{12}{5}g_1^4-2y_\tau^2\xi_{H_d}^2\right)\frac{F^2}{M^2},
\end{split}
\label{sf3mass}
\ee
\be
\begin{split}
m_{H_d}^2=&\frac{1}{256 \pi^4}\left[\frac{3}{5}g_1^4+3g_2^4-\xi_{H_d}^2\left(\frac{3}{5}g_1^2+3g_2^2-2\kappa^2-4\xi_{H_d}^2-4\xi_D^2-3\xi_T^2-2\xi_{H_u}^2\right)-\right.\\
-&\left.2\lambda^2\left(\xi_D^2+\frac{3}{2}\xi_T^2+2\xi_{H_u}^2\right)\right]\frac{F^2}{M^2},\\
m_{H_u}^2=&\frac{1}{256 \pi^4}\left[\frac{3}{5}g_1^4+3g_2^4-\xi_{H_u}^2\left(\frac{3}{5}g_1^2+3g_2^2-2\kappa^2-4\xi_{H_u}^2-4\xi_D^2-3\xi_T^2-2\xi_{H_d}^2\right)-\right.\\
-&\left.2\lambda^2\left(\xi_D^2+\frac{3}{2}\xi_T^2+2\xi_{H_d}^2\right)\right]\frac{F^2}{M^2},\\
m_{S}^2=&\frac{1}{256 \pi^4}\left[-\frac{6}{5}g_1^2\left(\xi_{H_d}^2+\xi_{H_u}^2+\xi_D^2+\frac{2}{3}\xi_T^2\right)-6g_2^2\left(\xi_{H_d}^2+\xi_{H_u}^2+\xi_D^2\right)-16g_3^2\xi_T^2+\right.\\
+&8\xi_{H_u}^2\left(\frac{3}{4}y_t^2-\kappa^2+\xi_{H_u}^2+2\xi_D^2+\frac{3}{2}\xi_T^2\right)+8\xi_{H_d}^2\left(\frac{3}{4}y_b^2+\frac{1}{4}y_\tau^2-\kappa^2+\xi_{H_d}^2+2\xi_D^2+\frac{3}{2}\xi_T^2\right)+\\
+&\left.8\xi_{H_u}^2\xi_{H_d}^2+8\xi_D^2\left(\xi_D^2-\kappa^2\right)+3\xi_T^2\left(5\xi_T^2-4\kappa^2\right)+12\xi_D^2\xi_T^2+8\lambda\xi_{H_d}\xi_{H_u}\xi_D\right]\frac{F^2}{M^2},
\end{split}
\label{HSmass}
\ee
where $g_{1,2,3}$ are the SM gauge couplings with $g_1\equiv \sqrt{5/3}g^\prime$ and the hypercharge defined by $Q=T_3+Y$. These are the leading contributions provided $F/M^2\ll 1/4\pi$. Otherwise, one-loop corrections generated at order $F^4/M^6$ from nontrivial terms in the effective K{\"a}hler potential can be comparable. Here we will concentrate in the previous regime, where such one-loop contributions can be safely neglected. For the first and second sfermion families only the gauge-mediation contributions are considered since we neglect the corresponding Yukawa couplings.
The soft $a$ terms, which are negligible in GMSB, now receive the leading contributions at one loop:
\be
\begin{split}
a_t=&-\frac{y_t \xi_{H_u}^2}{16\pi^2}\frac{F}{M},\\
a_b=&-\frac{y_b \xi_{H_d}^2}{16\pi^2}\frac{F}{M},\\
a_\tau=&-\frac{y_\tau \xi_{H_d}^2}{16\pi^2}\frac{F}{M},\\
a_\lambda=&-\frac{1}{16\pi^2}\left[\lambda\left(4\xi_{H_d}^2+4\xi_{H_u}^2+2\xi^2_D+3\xi^2_T\right)+2\xi_{H_d}\xi_{H_u}\xi_D\right]\frac{F}{M},\\
a_\kappa=&-3\frac{\kappa}{16\pi^2}\left(2\xi_{H_d}^2+2\xi_{H_u}^2+2\xi_D^2+3\xi_T^2\right)\frac{F}{M}.
\end{split}
\label{aterms}
\ee
Thus, this model has enough freedom to provide for large enough negative contributions to $m_S^2$, $a_\lambda$ and $a_\kappa$, as required to make the correct EWSB compatible with a realistic spectrum. In particular, compared to the scenario with only $S \bar{\Phi}^D_1 H_u$ couplings in \cite{Chacko:2001km}, the singlet-messenger interactions can turn $m_S^2$ negative without requiring large values of $|\kappa|\sim 1$. On the other hand, it is also possible to generate a large $a_t$ and therefore a large stop mixing. This enhances the size of the radiative top/stop one-loop corrections to $m_{H^1}$. Thus, it should be possible to lift the Higgs mass prediction above the direct LEP 2 bound with relaxed (not too large) values of $F/M$ and $M$, compared to the case of the model without $H_u$ couplings to the messenger fields.
Finally, we have the gaugino masses. These are given by the pure GMSB contribution. At one loop,
\be
M_{\lambda_a}=\frac{g_a^2}{8\pi^2}\frac{F}{M},~~a=1,2,3,
\ee
for $n=2$ messenger fields.
\section{Electroweak symmetry breaking and the spectrum calculation}
Using the model described in the previous section we have performed a scan in the parameter space looking for those regions where a satisfactory EWSB is possible together with a phenomenologically acceptable spectrum. In what follows, and prior to show the results of this analysis we sketch the method employed in the scan. This largely follows that of Ref. \cite{Delgado:2007rz} so we refer to that reference for more details.
The model has eight unknown input parameters: the superpotential couplings $\lambda$, $\kappa$, $\xi_D$, $\xi_T$, $\xi_{H_d}$ and $\xi_{H_u}$, and the SUSY breaking and messenger scales $F/M$ and $M$. The values for the SM parameters $g_{1,2,3}$, $y_{t,b,\tau}$ and the electroweak scale, $v^2\equiv \left<H_d\right>^2+\left<H_u\right>^2\approx \left(174\units{GeV}\right)^2$, can be determined from the experimental values of $G_F$, $M_Z$, $\alpha_\mt{em}$, $\alpha_\mt{S}\left(M_Z^2\right)$, and the known fermion masses \cite{Nakamura:2010zzi} (for the top mass we use $m_t=173.3 \units{GeV}$ \cite{:1900yx} and also take into account one-loop QCD corrections in extracting the corresponding Yukawa coupling). Using the RG equations to evolve all these parameters from the scales where they are defined up to the messenger scale we can compute the values of the SUSY-breaking soft terms using Eqs. (\ref{sf3mass}), (\ref{HSmass}) and (\ref{aterms}). Then we must use again the RG to run all the couplings down to the scale where we minimize the scalar potential
\be
\begin{split}
V&=m_{H_u}^2 \left|H_u^0\right|^2+m_{H_d}^2 \left|H_d^0\right|^2+m_{S}^2 \left|S\right|^2+\left(-a_\lambda S H_u^0 H_d^0-\frac{a_\kappa}{3}S^3+\hc\right)+\\
&+\left|\lambda\right|^2\left|S\right|^2\left(\left|H_u^0\right|^2+\left|H_d^0\right|^2\right)+\left|\lambda H_u^0 H_d^0+\kappa S^2\right|^2+\frac{\frac{3}{5}g^2_1+g^2_2}{8}\left(\left|H_d^0\right|^2-\left|H_u^0\right|^2\right)^2.
\end{split}
\label{Vscal}
\ee
The dominant ${\cal O}\left(y_t^4,y_b^4\right)$ one-loop radiative corrections in the effective scalar potential are taken into account by replacing $V\rightarrow V+\Delta V$, where $\Delta V$ reads in the $\overline{\mt{DR}^\prime}$ scheme \cite{Barger:2006dh,Martin:2001vx}:
\be
\Delta V=\frac{3}{32\pi^2}\sum_{f=t,b}\left[\sum_{i=1}^2 \overline{m}_{\tilde{f}_i}^4\left(\log{\frac{\overline{m}_{\tilde{f}_i}^2}{Q^2}}-\frac 32 \right)-2\overline{m}_{f}^4\left(\log{\frac{\overline{m}_{f}^2}{Q^2}}-\frac 32\right)\right],
\ee
with $\overline{m}_f$, $\overline{m}_{\tilde{f}_i}$ the field-dependent fermion/sfermion masses.
In order to minimize the (leading) ${\cal O}\left(y_t^4\right)$ corrections, as in \cite{Delgado:2007rz} we choose to perform this last step at a matching scale given by the geometric average of the stop masses, $M_\mathrm{match}=\sqrt{m_{\tilde{t}_1}m_{\tilde{t}_2}}$. This is also the scale where we specify the value for $\lambda$. On the other hand, $\kappa$ is determined in the minimization from the extremal conditions of the potential at the electroweak vacuum, together with $\tan{\beta}$ and $\left<S\right>$. Thus, $\kappa$ is actually an output, which reduces the number of unknown parameters by one.
We will require in our analysis that the gauge and superpotential couplings remain perturbative up to the grand unification scale, $M_\mt{GUT}$, defined as the scale such that $g_1(M_\mt{GUT})=g_2(M_\mt{GUT})$. For simplicity we will assume that the singlet-messenger interactions unify at that scale, at a common value $\xi_U\equiv \xi_D,\xi_T(M_\mt{GUT})$. This removes another parameter from the list, leaving finally with a total of six unknown inputs. The $\xi_{H_u}$ and $\xi_{H_d}$ couplings are also specified at $M_\mt{GUT}$. For completeness, the corresponding RG equations for the region $M\leq Q < M_\mt{GUT}$ are given in the appendix. Recall that, despite we assume that the messenger couplings to quark and lepton chiral fields vanish at $M$, nonzero values for $\xi_t$, $\xi_b$ and $\xi_\tau$ will be generated above the messenger scale. We take their effect on the running into account. Finally note that since the boundary conditions for the different parameters are specified at different scales and also depend on the results of the minimization, after solving the RG equations and minimizing the Higgs potential we need to adjust $\tan{\beta}$, $\left<S\right>$ and $\kappa$ and iterate until the procedure converges.
Once we obtain a set of parameters that allows to reproduce the correct EWSB we make sure that the resulting spectrum passes all existing experimental bounds~\cite{Nakamura:2010zzi}. Although such bounds depend on a variety of assumptions and may not directly apply in some cases, we prefer to be conservative in the cuts we apply prior to accept a given point in the scan. In particular, we demand that the lightest neutralino is above $50\units{GeV}$ or $120 \units{GeV}$ depending on whether it behaves as an stable or unstable particle, respectively. For the lightest chargino we also impose the cut in $120\units{GeV}$. Current limits for sfermions are different for sleptons and squarks and in the latter case also distinguish between generic squarks or the third family ones. For the sleptons and the third family of squarks we impose a common bound $m_{\tilde{l_i},{\tilde t_i},{\tilde b_i}}>120\units{GeV}$. Limits for generic squarks are significantly stronger than those of stops and sbottoms. Here we use the latest experimental results on SUSY searches at the LHC \cite{Khachatryan:2011tk,daCosta:2011hh,Collaboration:2011qk}. These correspond to the minimal supergravity/constrained MSSM parameter space and impose bounds on generic squarks and gluino masses around $700$-$800\units{GeV}$, depending on the difference between $m_{\tilde{q}}$ and $M_{\tilde g}$. For the first two families of squarks and the gluino mass we thus require $m_{\tilde q},M_{\tilde g}>700\units{GeV}$. For the Higgs masses we make use of the existing bounds from LEP 2 searches though, as we will see, in practice only the limit on a SM-like Higgs mass of $114\units{GeV}$ applies.
The mass matrices for the sfermions, neutralinos, charginos and CP even and odd Higgses can be found in the literature (see for instance \cite{King:1995vk,Ellwanger:2009dp}) and here we only sketch the precision in our determination of the lightest CP-even Higgs mass in order to compare with the LEP 2 bound. These include the dominant ${\cal O}\left(y_{t,b}^4\right)$ one-loop corrections \cite{King:1995vk} as well as the two-loop leading logarithmic corrections of ${\cal O}\left(y_{t,b}^4~\!g_3^2,y_{t,b}^6\right)$ \cite{Ellwanger:2009dp}. We also include the one-loop leading logarithmic corrections of ${\cal O}\left(y_{t,b}^2~\!g_{1,2}^2,y_{t,b}^2~\!\lambda^2\right)$ that enter in the mass matrices through the wavefunction renormalization of the scalar fields \cite{Ellwanger:2009dp}.
Finally, we also include the one-loop leading logarithmic corrections of ${\cal O}\left(\lambda^4\right)$ to the mass of the lightest CP-even Higgs boson, computed in the limit $\left<S\right>\gg v$:
\be
\Delta m_{H^1}^2=-\frac{3\lambda_\mt{Eff}^2 v^2}{4\pi^2} \log{\frac{Q^2}{m_t^2}},
\ee
with $\lambda_\mt{Eff}$ the effective quartic coupling for the lightest CP-even Higgs boson.
\section{Phenomenological results}
Here we present the results of the scan and discuss the phenomenological features of the model. We present the results in two parts. In the first one we discuss the general aspects of the spectrum obtained from a scan over all the six free parameters. In the second, in order to compare the model with the one presented in \cite{Delgado:2007rz}, we consider a submanifold of the allowed region of the parameter space where some of the parameters are fixed to specific values. Along this section we will often refer to the model in that reference, with only singlet-messenger interactions, as N-GMSB (as denoted there) while for the general scenario of singlet deflection of GMSB presented here we use S-GMSB. We discuss the improvements with respect to \cite{Delgado:2007rz}, emphasizing whether or not these are worth increasing the number of free parameters.
We have explored values for the SUSY-breaking scale $F/M$ up to $175\units{TeV}$. For $F/M \gtrsim 175 \units{TeV}$ it was shown in \cite{Delgado:2007rz} that significantly large regions in the parameter space were allowed by all experimental bounds, provided we allow for a long enough running, i.e., a large $M$ ($F/M=172\units{TeV}$ and $M=10^{10}\units{TeV}$ in the example presented there). Here we focus on whether this can be achieved with a lower scale of SUSY breaking. Regarding the messenger mass scale we take $10^{4}\units{TeV}\leq M\leq 10^{11}\units{TeV}$. We restrict to values above $10^4\units{TeV}$ to ensure $F/M^2\ll1/4\pi$, so the leading corrections to soft masses are those in Eqs. (\ref{sf3mass}) and (\ref{HSmass}). For the remaining unknown parameters, the superpotential couplings $\lambda$, $\xi_U$, $\xi_{H_u}$ and $\xi_{H_d}$, we scan both positive and negative values. This is required since they contribute to terms with no definite sign in the boundary conditions (\ref{HSmass}) and (\ref{aterms}), and in the RG equations above the messenger scale.
As stressed in \cite{Delgado:2007rz}, one of the major difficulties in the model with only singlet-messenger interactions was finding a set of parameters that gives a large enough Higgs mass so it passes the direct LEP 2 bound. The problem in that model is the absence of direct contributions to $a_t$ at the messenger scale. Thus, a nonzero $a_t$ can be generated only through the RG evolution down to $M_\mt{match}$. This results in small values of the stop mixing, which then require large stop masses, and therefore large values of $F/M$, in order to lift $m_{H^1}$ above 114 GeV through top/stop radiative corrections. As we explained in section \ref{sec:Model}, achieving a significantly large stop mixing should be no longer an issue once we consider the new interactions between the singlet, the messengers and $H_u$. Therefore, since a larger stop mixing than in \cite{Delgado:2007rz} is in general predicted, lower values of $F/M$ should be naturally allowed, yielding also to a lighter spectrum. Still, overcoming the LEP 2 bound is the major constraint in $F/M$ and $M$, but the results of the scan show that values of the SUSY-breaking scale $F/M\gtrsim 50 \units{TeV}$ are within the allowed region. (For those values of $F/M$ the extra Higgs boson masses are still effectively decoupled. Therefore, the lightest CP-even Higgs is SM-like and the $114\units{GeV}$ limit applies, instead of the less stringent bounds from direct searches of the supersymmetric Higgses.) On the other hand, we find $m_{H^1}\lesssim 123 \units{GeV}$ for the entire set of scan points. Regarding also the Higgs mass and the radiative corrections we find
\be
1.5\lesssim\tan{\beta}\lesssim35,
\ee
with a general preference for low values. Therefore, contributions from bottom/sbottom loops to the Higgs mass can be safely neglected in general. Nevertheless, the small regions with large $\tan{\beta}$ require to include bottom/sbottom radiative corrections in the EWSB sector as explained in the previous section, for they might start to be noticeable at that point. Note also that in that case these corrections actually lower the prediction for the Higgs mass.
On general grounds the spectrum of the model is gauge mediation like, with sparticle masses ordered according to their gauge interactions. The lightest supersymmetric particle (LSP) is the gravitino, which may become a good dark matter candidate for relatively low values of $F/M$ and $M$. As explained above such values are now easily accessible. The characteristic SUSY mass scale (roughly $\sim F/(16\pi^2 M)$) is $\gtrsim 320 \units{GeV}$. In Fig. \ref{fig_Spectrum} left we show the masses for the third family of charged sfermions.
\begin{figure}[t]
\input{Spectrum}
\caption{(Left) Third family sfermion masses. (Right) Gluino and lightest neutralino and chargino masses versus the lightest CP-even Higgs mass. The sharp upper bounds correspond to the upper bound of $F/M$ in the scan. The lightest neutralino is in general the NLSP.}
\label{fig_Spectrum}
\end{figure}
These are typically lighter than the first and second generations because of the effect of the Yukawa couplings. Such effects enter not only in the running but also in the boundary conditions (\ref{sf3mass}). Hence, a characteristic feature of this scenario is a larger splitting between the third and the first two families of sfermions, compared to the one with pure gauge mediation. This is more pronounced in the case of stops, since not too large values of $\tan{\beta}$ are preferred. Still, the effect is only noticeable for sizable values of the $\xi_{H_u}$ coupling. In particular, it is noteworthy that stops (and to a less extent sbottoms) below the TeV are still compatible with all the constraints. Note also the presence of some points in the region of stop masses below $500\units{GeV}$. These correspond to configurations where an approximate cancellation in the boundary condition for the third family of squark masses is taking place (see Eq. (\ref{sf3mass})). They are, however, rather difficult to find in the scan. This may indicates that such configurations require a significant fine tuning and do not correspond to natural solutions. As in gauge mediation the lightest sfermions are the staus. Stau masses can be compared with the lightest neutralino mass, which is shown in Fig. \ref{fig_Spectrum} right, together with the gluino and the lightest chargino masses. In general, the lightest neutralino is lighter than staus, thus being the next-to-lightest supersymmetric particle (NLSP). This is the case for low $\tan \beta \lesssim 3$. For larger values there are regions where the stau can be nearly degenerate or slightly lighter than the lightest neutralino. In this case both may behave as co-NLSPs. Only for some regions with large $\tan{\beta}\gtrsim 20$ the effect of the $\tau$ Yukawa couplings can lower $m_{\tilde{\tau}_1}$ significantly below $M_{\chi^0_1}$. Regarding the lightest neutralino composition, this is in general bino like. However, we also find some regions where this can be singlino or Higgsino like. The main phenomenology of this model depends on the NLSP nature and lifetime. At particle colliders the most interesting scenarios occurs for those cases where the NLSP is stau, or it is neutralino but decays promptly within the detector. The latter occurs for $\sqrt{F}\lesssim 10^3\units{TeV}$, which is still accessible but only in a small region at the lower end of the range of allowed values of $F/M$ and $M$. For larger values the NLSP behaves like an stable particle. On the other hand, one must be aware that long-lived NLSP might pose cosmological problems for ordinary nucleosynthesis~\cite{Giudice:1998bp,Gherghetta:1998tq}. These can put additional constraints on the upper bounds of $F/M$ and $M$. For instance, for photon decays we are safe as long as $\tau_\mt{NLSP}< 10^7$ sec. For the range of values of $F/M$ explored this translates into an upper bound for the messenger scale $M\lesssim 10^{12}$-$10^{14}\units{TeV}$. Hadronic decays might impose even stronger constraints $M\lesssim 10^{9}$-$10^{11}\units{TeV}$.
Despite it is clear that the S-GMSB scenario favors heavier Higgs masses through radiative corrections than the one without the interactions involving Higgs doublets, we would like to determine quantitatively the actual improvement respect to N-GMSB, for we are introducing extra free parameters which reduce the predictive power of the model. In what follows we compare our results with those of \cite{Delgado:2007rz}. For that purpose we have considered two specific benchmark points with fixed values of $F/M$ and $M$. For the first one we have chosen $F/M=172~\units{TeV}$ and $M=10^{10}\units{TeV}$ as in the example presented in \cite{Delgado:2007rz}. As we have stressed previously, one of the major advantages of the model presented here is that it does not require of large values of $F/M$ and $M$ to pass all the constraints. Thus, we have also considered another point where both parameters take significantly lower values: $F/M=100\units{TeV}$ and $M=10^4\units{TeV}$. Let us note here that the $\xi_{H_d}$ coupling has a negligible effect over the Higgs mass prediction. Indeed, the impact of the parameter $\xi_{H_d}$ in generating a significant sbottom mixing is always irrelevant. Thus, it does not have any significant effect on the (already tiny) bottom/sbottom radiative corrections. The only relevant effect that this parameter may have is to help in the process of generating a negative $m_S^2$ and controlling the size of $a_\lambda$~\footnote{Actually, the only contribution to $a_\lambda$ that has no definite sign vanishes for $\xi_{H_d}\rightarrow0$.}. However, this r{\^o}le can be easily played by some of the other parameters. Therefore, for our purposes this coupling looks completely irrelevant. Hence, in order to emphasize the important effects in the comparison, we have gotten rid of this parameter by assuming that it vanishes at the scale $M_\mt{GUT}$, i.e., we only consider one extra parameter~\footnote{At any rate, since $\xi_{H_u}$ and $\xi_D$ are nonvanishing, small corrections from $\xi_{H_d}$ are still expected from the running down to the messenger scale (see discussion below Eq. (\ref{MessW}), and Eq. (\ref{MessCoupRGEs}) in the appendix).} compared to the N-GMSB scenario.
In Fig. \ref{fig_Comparison} left we compare the allowed regions for both models in the $\lambda(M_\mt{match})$-$\xi_U$ plane for $F/M=172~\units{TeV}$ and $M=10^{10}\units{TeV}$. As it is apparent the $\xi_{H_u}$ coupling significantly increases the area where Higgs masses above 114 GeV can be attained, beyond the three characteristic regions explained in \cite{Delgado:2007rz}. This enhancement of the allowed regions is even more apparent when we move to lower SUSY-breaking scales, where the model in \cite{Delgado:2007rz} has problems in obtaining large enough Higgs masses. Indeed, in Fig. \ref{fig_Comparison} right we can observe how for $F/M=100\units{TeV}$, $M=10^4\units{TeV}$ the N-GMSB scenario is almost ruled out by the experimental constraints, while still a large region in the $\lambda(M_\mt{match})$-$\xi_U$ plane is allowed once we introduce the extra coupling. Let us also note that this does not require too large values for $\xi_{H_u}$. For the regions in Fig. \ref{fig_Comparison} we find $\xi_{H_u}\lesssim 4$, with $\xi_{H_u}\sim 4$ only at the end of the tail in Fig. \ref{fig_Comparison} right ($\xi_{H_u}\lesssim 1.5$ in Fig. \ref{fig_Comparison} left). Characteristic values for the sparticle masses in this latter example, and thus not easily accessible to the N-GMSB scenario, are the following: the lightest stop (sbottom) mass is $\sim 1.1$-$1.3\units{TeV}$ ($1.4$-$1.5\units{TeV}$); for the lightest stau $m_{\tilde{\tau_1}}\sim 220$-$270\units{GeV}$; for the gaugino masses $M_{\lambda_a}\approx 280, 510, 1400\units{GeV}$ for bino, winos and gluinos, respectively. Notice that the stau is lighter than the bino. The lightest neutralino is in general an admixture of Higgsinos and the bino, though regions where it is purely singlino are still present. In the latter case as well as in those regions where the effective $\mu$ term is low enough the lightest neutralino can still be the NLSP. Notice also that this is a short-lived NLSP ($\sqrt{F}=10^3\units{TeV}$), which might still decay within the detector at collider experiments giving a hard photon or two jets plus missing $E_T$, depending on whether the NLSP is mostly a bino or has a singlino/Higgsino component. In the case where the NSLP is a stau the decay inside the detector will provide a hard jet plus missing energy.
\begin{figure}[t]
\input{Comparison}
\caption{Comparison of the allowed regions in the $\lambda(M_\mt{match})$-$\xi_U$ plane for the models with and without $S \bar{\Phi}_1^D H_u$ superpotential interactions, S-GMSB and N-GMSB, respectively. Only the regions with $m_{H^1}\geq 114 \units{GeV}$ are shown. (Left) $F/M=172 \units{TeV}$ and $M=10^{10}\units{TeV}$ as in \cite{Delgado:2007rz}. (Right) The same for $F/M=100 \units{TeV}$ and $M=10^5\units{TeV}$. }
\label{fig_Comparison}
\end{figure}
In summary, despite the model in \cite{Delgado:2007rz} is more predictive and it is a phenomenologically viable solution to the $\mu$-$b_\mu$ problem, at the price of only one more parameter the model presented here provides not only a significant enhancement of the allowed regions for those values of $F/M$ and $M$ where N-GMSB works but also extends the applicability of the gauge mediated NMSSM to scenarios where SUSY breaking occurs at lower energy scales.
\section{Conclusions}
In this paper we have extended the work presented in \cite{Delgado:2007rz}, where a model including singlet-messenger superpotential interactions was proposed in order to ameliorate the difficulties in generating the correct EWSB and a phenomenologically valid spectrum in gauge-mediation scenarios of the NMSSM. Such model thus also provides a satisfactory explanation to the $\mu$-$b_\mu$ problem of GMSB. However, it requires a relatively large scale of SUSY breaking in order to overcome all existing bounds and, in particular, the direct LEP 2 bound on the lightest CP-even Higgs boson mass. Considering an extension where we also include interactions between the NMSSM singlet, the messengers and the Higgs doublets we have proved that it is possible to lower significantly the scale of SUSY breaking without drastically constraining the size of the allowed regions in the parameter space. We have performed a scan over all the parameter space and discussed the general aspects of the spectrum of the model. Most of the basic features of a GMSB-like spectrum prevail in this scenario: strongly interacting particles are heavier, gravitino LSP, bino-like neutralino or stau NLSP, and so on. These are ``deflected'' by the new interactions, whose effects are multiplied by the NMSSM superpotential couplings of the corresponding particles. Thus, for instance, there is a significant splitting between the stops and first two families of up squarks that goes beyond the RG effects in GMSB. The overall scale of the spectrum can be relatively low. In particular, there are regions in the parameter space that allow stop masses below the TeV. Such values are not easily accessible without the new interactions. This low spectrum could be discovered at the early stages of the LHC.
We also emphasize that despite the general scenario introduces several extra free parameters, only one of them, the coupling $\xi_{H_u}$, is actually required in order to open the new regions in the parameter space. Thus, in practice, only one extra parameter is relevant and the model still remains quite predictive.
\section*{Acknowledgements}
It is a pleasure to thank G. F. Giudice and E. Ponton for reading the manuscript. We would like to especially thank P. Slavich for a lot of useful comments and clarifications. This work has been supported in part by the U.S. National Science Foundation under Grant PHY-0905283-ARRA.
\newpage
|
\section{Introduction}\label{sec.intro}
Let \(A\) be a Cohen-Macaulay ring of finite Krull dimension with a canonical module \(\omega_{A}\). Let \(\cat{MCM}_{A}\) and \(\cat{FID}_{A}\) denote the categories of maximal Cohen-Macaulay modules and of finite modules with finite injective dimension, respectively. M.\ Auslander and R.-O.\ Buchweitz proved in \cite{aus/buc:89} that for any finite \(A\)-module \(N\) there exists short exact sequences
\begin{equation}\label{eq.MCMseq}
0\rightarrow L\longrightarrow M\longrightarrow N\rightarrow 0\qquad\text{and}\qquad 0\rightarrow N\longrightarrow L'\longrightarrow M'\rightarrow 0
\end{equation}
with \(M\) and \(M'\) in \(\cat{MCM}_{A}\) and \(L\) and \(L'\) in \(\cat{FID}_{A}\). The maps \(M\rightarrow N\) and \(N\rightarrow L'\) in \eqref{eq.MCMseq} are called a maximal Cohen-Macaulay approximation and a hull of finite injective dimension, respectively, of the module \(N\).
The association \(N\mapsto X\) for \(X\) equal to \(M,M',L\) and \(L'\) define maps (functors) of corresponding stable categories.
In this article we study the continuous properties of these maps.
Linear representations provided by (sheaves) of modules and the associated homological algebra plays an important role in algebra and algebraic geometry, e.g.\ as a means for classification by providing invariants.
Finite complexes have particular properties as seen in the Buchsbaum-Eisenbud acyclicity criterion and the intersection theorems of Peskine, Szpiro and Roberts. However, for a non-regular local ring \(A\), the standard homological invariants are given by the (generally) infinite minimal \(A\)-free resolutions, of which very little is known. To stay within finite complexes one can enlarge or change the category of resolving objects and Cohen-Macaulay approximation is a structured way of doing this.
Let \(\cat{D}_{A}\) denote the subcategory \(\Add\{\omega_{A}\}\) of modules \(D\) isomorphic to direct summands of the \(\omega_{A}^{\oplus r}\), \(r>0\). A part of the approximation result says that all the modules in \(\cat{FID}_{A}\) have finite resolutions by objects in \(\cat{D}_{A}\). In particular the MCM approximation in \eqref{eq.MCMseq} can be extended to a finite resolution
\begin{equation}\label{eq.Dres}
0\rightarrow D^{-n}\longrightarrow D^{-n+1}\longrightarrow\dots \longrightarrow D^{-1}\longrightarrow M\longrightarrow N\rightarrow 0
\end{equation}
with the \(D^{i}\) in \(\cat{D}_{A}\). In the case \(A\) is Gorenstein, \(\cat{D}_{A}\) equals the category of finite projective modules \(\cat{P}_{\!A}\). This generalises: By a result of R.\ Y.\ Sharp \cite{sha:75b} the functor \(\hm{}{A}{\omega_{A}}{-}\) gives an exact equivalence \(\cat{D}_{A}\simeq\cat{P}_{\!A}\), hence a finite projective resolution is associated to \(N\). In the case \(A\) is local, the approximations and the complex can be chosen to be minimal and unique (with \(D^{i}\cong \omega_{A}^{\oplus d^{i}}\)) and in particular the \(d^{i}\) are invariants of \(N\).
The developments since Auslander and Buchweitz' fundamental work \cite{aus/buc:89} has included studies of invariants defined by Cohen-Macaulay approximation; \cite{din:92, aus/din/sol:93, has/shi:97} among several, `injectivity' and `surjectivity' properties of the approximation maps; \cite{kat:99, yos/iso:00, kat:07}, and characterisations of quasi-homogeneous isolated singularities; cf.\ \cite{her/mar:93, mar:00b}, all exclusively in the Gorenstein case. Noteworthy is \cite{sim/str:02} where A.-M.\ Simon and J.\ R.\ Strooker related some of these invariants with Hochster's Canonical Element Conjecture and the Monomial Conjecture. In particular these conjectures are equivalent to the vanishing of the \(\delta\)-invariant of certain cyclic modules over all Gorenstein rings.
S.\ P.\ Dutta applied the existence of a FID hull to prove a relationship between two of the Serre conjectures on intersection numbers: Failure of vanishing implies failure of higher non-negativity in the Gorenstein case under certain conditions, see \cite{dut:04}.
Buchweitz' unpublished manuscript \cite{buc:86}, an important precursor to \cite{aus/buc:89}, contains homological ideas which have become important in subsequent developments (e.g.\ \cite{kra:05}). Auslander and I.\ Reiten elaborated in \cite{aus/rei:91} on \cite{aus/buc:89}, mainly with a view towards artin algebras, instigating several generalisations and analogies to Cohen-Macaulay approximation.
However, the `relative' and continuous aspects have received surprisingly little attention.
In \cite{has:00} M.\ Hashimoto gave several new examples of Cohen-Macaulay approximation.
Perhaps closest to our results is \cite[IV 1.4.12]{has:00} where an affine algebraic group \(G\) acts on a positively graded Cohen-Macaulay ring \(T\) which is flat over a regular base ring \(R\). Hashimoto considers graded maximal Cohen-Macaulay \(T\)-modules (which automatically are \(R\)-flat) and graded modules locally of finite injective dimension (not \(R\)-flat in general), all with \(G\)-action. His result (with trivial group) is hence different from our Theorem \ref{thm.flatCMapprox}. We also note some explicit \(1\)-parameter families of indecomposable finite length modules \(N_{t}\) (for many Gorenstein rings) such that the minimal MCM approximation \(M_{t}\) is without free summands, see \cite{yos:99}.
A central part of the classification problem is to prove the existence of objects with certain properties and to estimate `how many' such objects there are.
A natural question is thus whether there is Cohen-Macaulay approximation for flat families of modules. In Theorem \ref{thm.flatCMapprox} we give a positive answer to this question. For a Cohen-Macaulay (CM) map \(h:S\rightarrow T\) (flat!) and an \(S\)-flat and finite \(T\)-module \(\mc{N}\) there are short exact sequences of \(S\)-flat and finite \(T\)-modules
\begin{equation}\label{eq.flatMCMseq}
0\rightarrow \mc{L}\longrightarrow \mc{M}\longrightarrow \mc{N}\rightarrow 0\qquad\text{and}\qquad 0\rightarrow \mc{N}\longrightarrow \mc{L}'\longrightarrow \mc{M}'\rightarrow 0
\end{equation}
such that the fibres of these sequences give `absolute' approximations and hulls as in the two sequences \eqref{eq.MCMseq}. Note that \(T\) in general is \emph{not} a Cohen-Macaulay ring although the fibres of \(h\) are.
We consider a category \(\cat{mod}{}^{\textnormal{fl}}\) of pairs \(\xi=(h:S\rightarrow T,\mc{N})\) and subcategories \(\cat{MCM}\), \(\cat{FID}\) and \(\cat{D}\).
They are fibred over the category \(\cat{CM}\) of CM maps (\(\xi\mapsto h\)) and also fibred over the base category of noetherian rings (\(\xi\mapsto S\)). The approximation and the hull \eqref{eq.flatMCMseq} induces functors of certain quotient categories fibred in additive categories over \(\cat{CM}\)
\begin{equation}
\cat{mod}{}^{\textnormal{fl}}/\cat{D}\rightarrow \cat{MCM}/\cat{D}\quad \text{and}\quad\cat{mod}{}^{\textnormal{fl}}/\cat{D}\rightarrow \cat{FID}/\cat{D}
\end{equation}
with analogous properties to the absolute case. If \(h:S\rightarrow T\) is a \emph{local} CM map, there is an analogous approximation result with minimal choices of the two sequences in \eqref{eq.flatMCMseq}, see Corollary \ref{cor.minapprox}.
A major consequence of these results is that any (numerical and additive) upper semi-continuous invariant of MCM or FID modules by the minimal approximations and hulls induces upper semi-continuous invariants for all finite modules, see Theorem \ref{thm.semicont}. Examples of such invariants are given by the \(\omega_{A}\)-ranks in the minimal \emph{representing complex} \(D^{*}(N)\) which is an (infinite) extension to the right of the \(\cat{D}_{A}\)-complex in \eqref{eq.Dres}.
Auslander's fundamental module \(E_{A}\) for a normal \(2\)-dimensional singularity \(\Spec A\) is given by the CM approximation of the maximal ideal;
\begin{equation}
0\rightarrow \omega_{A}\longrightarrow E_{A}\longrightarrow \fr{m}_{A}\rightarrow 0
\end{equation}
which in a certain sense generates all almost split sequences for \(A\), see \cite{aus:86}.
As a general example of flat Cohen-Macaulay approximation we define the fundamental module for any finite type CM map of pure relative dimension \(\geq 2\), see Corollary \ref{cor.Fmod}, and more generally a `fundamental' functor of projective modules in Proposition \ref{prop.Fmod}.
An attractive feature of Auslander and Buchweitz' theory is its axiomatic formulation with several applications besides the classical case described in the first paragraph, e.g.\ coherent rings with a cotilting module, the graded case, approximation with modules of Gorenstein dimension \(0\), and coherent sheaves on a projectively embedded Cohen-Macaulay scheme. See \cite{aus/buc:89} and \cite{has:00} for more examples. We formulate a relative Cohen-Macaulay approximation theory axiomatically in terms of categories \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\) fibred in abelian and additive subcategories over a base category \(\cat{C}\). In addition to the Auslander-Buchweitz axioms (AB1-4) for the fibre categories we formulate two axioms (BC1-2) regarding base change properties of the fibred categories. AB1-2 and BC1-2 imply the existence of an approximation and a hull which are preserved by any base change, see Theorem \ref{thm.cofapprox}. If AB3 holds too, we get functoriality and adjointness properties in suitable stable categories fibred in additive categories, see Theorem \ref{thm.cofmain}. In the case described above \(\cat{C}=\cat{CM}\), \(\cat{A}\) is the category \(\cat{mod}\) of pairs \((h:S\rightarrow T,\mc{N})\) where \(\mc{N}\) is a finite \(T\)-module (no \(S\)-flatness!) and \(\cat{X}=\cat{MCM}\). Presently we consider only this example (and local variants of it), but there are several other natural applications of the relative theory.
In the second half of the article we proceed to study properties of continuous families of MCM approximations and FID hulls by homological methods.
As a consequence of the existence of minimal approximations and hulls of local flat families there are induced natural maps of deformation functors of pairs of algebra and module:
\begin{equation}
\df{}{(A,N)}\longrightarrow\df{}{(A,X)}\quad\text{for}\quad X=M, M', L\,\,\text{and}\,\, L'\,,
\end{equation}
and there are corresponding maps \(\df{A}{N}\rightarrow\df{A}{X}\) of deformation functors of the modules where \(A\) only deforms trivially. Rather weak grade conditions on \(N\) imply the injectivity of these maps for \(X=L'\) (or \(L\)). If there in addition exists a versal family in \(\df{}{(A,N)}\) (or \(\df{A}{N}\)) then the maps are isomorphisms for the appropriate category of henselian rings, see Theorem \ref{thm.defgrade} and Corollary \ref{cor.defgrade}. As a consequence each CM algebraic \(k\)-algebra \(A\) with \(A/\fr{m}_{A}\cong k\) and \(\dim A\geq 2\) has a finite \(A\)-module \(Q'\) of finite \emph{projective} dimension with a \emph{universal} deformation in \(\df{A}{Q'}(A)\), see Corollary \ref{cor.defapprox}. Proposition \ref{prop.Gor} says that if \(A\) is Gorenstein and \(\dim A-\depth N=1\) then \(\df{}{(A,N)}\rightarrow \df{}{(A,M)}\) is smooth.
Consider a quotient ring \(B=A/I\) defined by a regular sequence \(I=(f_{1},\dots,f_{n})\) and an MCM \(B\)-module \(N\). Then \(N\) is also an \(A\)-module with an MCM approximation \(M\rightarrow N\). If \(N\) has a \emph{lifting} to \(A/I^{2}\), then the composition of natural maps \(\df{B}{N}\rightarrow \df{A}{N}\rightarrow\df{A}{M}\) is injective, see Theorem \ref{thm.defMCM}. It turns out that the lifting condition is equivalent to the splitting of \(B{\otimes}_{A}M\rightarrow N\). This generalises \cite[4.5]{aus/din/sol:93} and might have a certain independent interest.
The second part of the article also contains some general deformation theory of a pair \((h:S\rightarrow T,\mc{N})\) of an algebra and a \(T\)-module. We define the graded algebra \(\varGamma:=T\oplus \mc{N}\) and consider the graded Andr{\'e}-Quillen cohomology \(\gH{0}^{*}(S,\varGamma, J)\) which govern the obstruction theory of the pair. In the case the graded \(\varGamma\)-module \(J\) is concentrated in degree \(0\) and \(1\) there is a natural long-exact sequence which in the case \(J=\varGamma\) (with \(\gH{0}^{*}(S,\varGamma)=\gH{0}^{*}(S,\varGamma,\varGamma)\)) gives the suggestive
\begin{equation*}
\begin{split}
0 &{} \rightarrow\nd{}{T}{\mc{N}}\rightarrow\gDer{0}_{S}(\varGamma)\rightarrow\Der_{S}(T)\rightarrow\xt{1}{T}{\mc{N}}{\mc{N}}\rightarrow\gH{0}^{1}(S,\varGamma)\rightarrow\cH^{1}(S,T) \\
&{} \rightarrow\xt{2}{T}{\mc{N}}{\mc{N}}\rightarrow\gH{0}^{2}(S,\varGamma)\rightarrow\cH^{2}(S,T)\rightarrow\dots
\end{split}
\end{equation*}
It relates the cohomology of the pair with the cohomology groups governing the obstruction theory of the algebra \(T\) and of the module \(\mc{N}\).
The sequence is used in the proof of the existence of a versal element in \(\df{}{(A,N)}\) where \(\Spec A\) is an isolated equidimensional singularity and \(N\) is locally free on the smooth locus, see Theorem \ref{thm.ExVers}. It is also used to define and study the Kodaira-Spencer class \(\kappa(\varGamma/S/\mathcal{O})\) in \(\gH{0}^{1}(S,\varGamma,\Omega_{S/\mathcal{O}}{\otimes}\varGamma)\) (where \(\mathcal{O}\rightarrow S\) is a another ring homomorphism) which maps to the ungraded Kodaira-Spencer class \(\kappa(T/S/\mathcal{O})\). In the case the latter is zero we define a `secondary' Kodaira-Spencer class \(\kappa(\sigma,\mc{N})\) in \(\xt{1}{T}{\mc{N}}{\Omega_{S/\mathcal{O}}{\otimes} \mc{N}}\) which depends on a choice of an \(S\)-algebra splitting \(\sigma\). This enables us to define `global' Kodaira-Spencer maps
\begin{equation}
g^{\varGamma}:\Der_{\mathcal{O}}(S)\rightarrow\gH{0}^{1}(S,\varGamma,\varGamma)\quad \text{and}\quad g^{(\sigma,\,\mc{N})}:\Der_{\mathcal{O}}(S)\rightarrow\xt{1}{T}{\mc{N}}{\mc{N}}\,.
\end{equation}
We also describe how classes and maps are related to the Atiyah class \(\at_{T/\mathcal{O}}(\mc{N})\) in \(\xt{1}{T}{\mc{N}}{\Omega_{T/\mathcal{O}}{\otimes}\mc{N}}\). These results might have a certain independent interest. The arguments are general and can be formulated in the setting of L.\ Illusie's \cite{ill:71}.
Injectivity of the corresponding local Kodaira-Spencer maps gives a criterion for a global flat family to be non-trivial. The Kodaira-Spencer maps commute with Cohen-Macaulay approximation and this is applied to show that injectivity of the Kodaira-Spencer map is preserved by Cohen-Macaulay approximation under certain `global' conditions akin to the local ones in Theorem \ref{thm.defgrade} and \ref{thm.defMCM}.
To make the text more reader friendly we have included some background material, e.g.\ on Cohen-Macaualy approximation and Kodaira-Spencer maps, and some of the central technical tools such as a general `cohomology and base change' result and some language of fibred categories. Many results have analogous parts with similar arguments and the policy has been to give a fairly detailed proof of one case and leave the other cases to the reader.
\section{Preliminaries}\label{sec.CMapprox}
All rings are commutative. If \(A\) is a ring, \(\cat{Mod}_{A}\) denotes the category of \(A\)-modules and \(\cat{mod}_{A}\) denotes the full subcategory of finite \(A\)-modules. If \(A\) is local then \(\fr{m}_{A}\) denotes the maximal ideal. Subcategories are usually full and essential.
\subsection{Categorical Cohen-Macaulay approximation}\label{subs.CCM}
We briefly recall some of the main features of Cohen-Macaulay approximation as introduced by Auslander and Buchweitz in \cite{aus/buc:89}.
In this section let \(\cat{A}\) be an abelian category and \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\) additive subcategories.
Let \(\hat{\cat{X}}\) denote the subcategory of \(\cat{A}\) of objects \(N\) which have finite resolutions \(0\rightarrow M_{n}\rightarrow\dots\rightarrow M_{0}\rightarrow N\rightarrow 0\) with the \(M_{i}\) in \(\cat{X}\). If \(n\) is the smallest such number, then \(\resdim{\cat{X}}{N} = n\). Let \(\injdim{\cat{X}}{N}\) be the minimal \(n\) (possibly \(\infty\)) such that \(\xt{i}{\cat{A}}{M}{N}= 0\) for all \(i>n\) and all \(M\) in \(\cat{X}\). Let \(\cat{X}^{\perp}\) denote the subcategory of objects \(L\) in \(\cat{A}\) with \(\injdim{\cat{X}}{L}=0\); the right complement of \(\cat{X}\). The left complement \({}^{\perp}\cat{X}\) is defined analogously.
Let \(N\) be an object in \(\cat{A}\). An \emph{\(\cat{X}\)-approximation} and a \emph{\(\hat{\cat{D}}\)-hull}
of \(N\) are exact sequences as in \eqref{eq.MCMseq} with \(L\), \(L'\) in \(\hat{\cat{D}}\) and \(M\), \(M'\) in \(\cat{X}\).
In general any \(f:M\rightarrow N\) in \(\cat{A}\) is called a \emph{right \(\cat{X}\)-approximation of \(N\)} if \(M\) is in \(\cat{X}\) and any \(f':M'\rightarrow N\) with \(M'\) in \(\cat{X}\) factorises through \(f\). Dually, \(g:N\rightarrow L\) is called a \emph{left \(\cat{X}\)-approximation of \(N\)} if \(L\) is in \(\cat{X}\) and any \(g':N\rightarrow L'\) with \(L'\) in \(\cat{X}\) factorises through \(g\).
Consider the following conditions on the triple of categories \((\cat{A},\cat{X},\cat{D})\).
\begin{enumerate}
\item[(AB1)] \(\cat{X}\) is exact in \(\cat{A}\) (\(\cat{X}\) is closed under direct summands and extensions).
\item[(AB2)] \(\cat{D}\) is a \emph{cogenerator} for \(\cat{X}\), i.e.\ for each object \(M\) in \(\cat{X}\) there is an object
\(D\) in \(\cat{D}\) and a short exact sequence \(M\rightarrow D\rightarrow M'\) with \(M'\) in \(\cat{X}\).
\item[(AB3)] \(\cat{D}\) is \(\cat{X}\)-\emph{injective}, i.e.\ \(\cat{D}\subseteq\cat{X}^{\perp}\).
\item[(AB4)] \(\cat{A}\)-epimorphisms in \(\cat{X}\) are admissible (i.e.\ their kernels are contained in \(\cat{X}\)).
\end{enumerate}
If AB1 and AB2, there exist \(\cat{X}\)-approximations and \(\hat{\cat{D}}\)-hulls for all objects in \(\hat{\cat{X}}\) \cite[1.1]{aus/buc:89}. Assume AB1-3. Then any \(\cat{X}\)-approximation is a right \(\cat{X}\)-approximation and any \(\hat{\cat{D}}\)-hull is a left \(\hat{\cat{D}}\)-approximation. An \(\cat{X}\)-approximation determines a \(\hat{\cat{D}}\)-hull and vice versa through the following diagram of short exact sequences; the upper horizontal and right vertical being an \(\cat{X}\)-approximation and a \(\hat{\cat{D}}\)-hull of \(N\), \(D\) is in \(\cat{D}\). The boxed square is (co)cartesian (see \cite[1.4]{aus/buc:89}):
\begin{equation}\label{eq.pullpush}
\xymatrix@C-0pt@R-12pt@H-30pt{
L \ar[r]\ar@{=}[d] & M \ar[r]\ar[d]\ar@{}[dr]|{\Box} & N \ar[d] \\
L \ar[r] & D \ar[r]\ar[d] & L' \ar[d] \\
& M' \ar@{=}[r] & M'
}
\end{equation}
Moreover, the category \(\cat{D}\) is determined by \(\cat{X}\subset\cat{A}\). Indeed \(\cat{D}=\cat{X}\cap\cat{X}^{\perp}\). By \cite[3.9]{aus/buc:89} monomorphisms in \(\hat{\cat{D}}\) are admissible and \(\hat{\cat{D}}=\hat{\cat{X}}\cap\cat{X}^{\perp}\). Also \(\cat{X}={}^{\perp}\hat{\cat{D}}\cap\hat{\cat{X}}={}^{\perp}\cat{D}\cap\hat{\cat{X}}\). If \(\cat{X}/\cat{D}\) denotes the quotient category, the \(\cat{X}\)-approximation induces a right adjoint to the inclusion functor \(\cat{X}/\cat{D}\subseteq \hat{\cat{X}}/\cat{D}\) and the \(\hat{\cat{D}}\)-hull induces a left adjoint to the inclusion functor \(\hat{\cat{D}}/\cat{D}\subseteq\hat{\cat{X}}/\cat{D}\), see \cite[2.8]{aus/buc:89}.
A morphism \(f: M\rightarrow N\) in \(\cat{A}\) is called \emph{right minimal} if for any \(g:M\rightarrow M\) with \(fg=f\) it follows that \(g\) is an automorphism. Dually, \(f\) is called \emph{left minimal} if for any \(h:N\rightarrow N\) with \(hf=f\) it follows that \(h\) is an automorphism. Note that if \(f:M\rightarrow N\) and \(f: M'\rightarrow N\) both are right minimal then there exists an isomorphism \(g:M\rightarrow M'\) with \(f=f'g\), and similarly for left minimal morphisms.
We will simply call an \(\cat{X}\)-approximation (a \(\hat{\cat{D}}\)-hull) for minimal if it is right (left) minimal.
\begin{ex}\label{ex.MCMapprox}
Suppose \(A\) is a Cohen-Macaulay ring which posses a canonical module \(\omega_{A}\) in the sense that any localisation in a maximal ideal gives a maximal Cohen-Macaulay module of finite injective dimension and Cohen-Macaulay type \(1\), cf.\ \cite[3.3.16]{bru/her:98}. Let \(\cat{MCM}_{A}\) denote the category of maximal Cohen-Macaulay (MCM) \(A\)-modules and put \(\cat{D}_{A}:=\Add\{\omega_{A}\}\). Then the triple \((\cat{A},\cat{X},\cat{D})=(\cat{mod}_{A},\cat{MCM}_{A},\cat{D}_{A})\) satisfies properties AB1-4, cf.\ \cite[I 4.10.11]{has:00} and \(\hat{\cat{X}}=\cat{mod}_{A}\). If \(A\) in addition is a local ring, then the \(\cat{MCM}_{A}\)-approximation and the \(\hat{\cat{D}}_{A}\)-hull can be chosen to be minimal, cf.\ \cite[Sec.\ 3]{sim/str:02}.
Let \(\cat{FID}_{A}^{l}\) denote the subcategory of finite \(A\)-modules \(E\) which have locally finite injective dimension, i.e.\ \(\Injdim_{A_{\fr{p}}} E_{\fr{p}}<\infty\) for all \(\fr{p}\in\Spec A\). The approximation result implies that \(\cat{FID}_{A}^{l}=\hat{\cat{D}}_{A}\): Let \(L\) be in \(\hat{\cat{D}}_{A}\). By induction on \(\resdim{\cat{D}_{A}}{L}\) \(L\) is in \(\cat{FID}_{A}^{l}\). Conversely let \(E\) be in \(\cat{FID}_{A}^{l}\). If \(L\rightarrow M\rightarrow E\) is an \(\cat{MCM}_{A}\)-approximation of \(E\) then \(M\) also has locally finite injective dimension. Let \(M^{\vee}\) denote \(\hm{}{A}{M}{\omega_{A}}\) and choose a surjection \(A^{{\oplus}n}\rightarrow M^{\vee}\). Both \(M^{\vee}\) and the kernel \(M_{1}\) are MCM. Applying \(\hm{}{A}{-}{\omega_{A}}\) gives (by duality theory) the short exact sequence \(M\xra{i} \omega_{A}^{{\oplus}n}\rightarrow M_{1}^{\vee}\). But \(i\) splits since \(\xt{1}{A}{M_{1}^{\vee}}{M}=0\) by \cite[3.3.3]{bru/her:98} and so \(M\) is in \(\cat{D}_{A}\) and \(E\) is in \(\hat{\cat{D}}_{A}\).
\end{ex}
\subsection{The representing complex}\label{subsec.cplx}
Consider an abelian category \(\cat{A}\) and additive subcategories \(\cat{D}\subseteq\cat{X}\subseteq \cat{A}\). A \emph{\(\cat{DX}\)-resolution} of an object \(N\) in \(\cat{A}\) is a finite resolution \({}^{-}C^{*}\twoheadrightarrow N\) with \({}^{-}C^{i}\in\cat{D}\) for \(i<0\) and \({}^{-}C^{0}\in \cat{X}\). If \(L:=\coker (d^{-2}:{}^{-}C^{-2}\rightarrow {}^{-}C^{-1})\), then the short exact sequence \(L\rightarrow {}^{-}C^{0}\rightarrow N\) is an \(\cat{X}\)-approximation. A \emph{\(\hat{\cat{D}}\cat{D}\)-coresolution} of \(N\) is a coresolution \(N\rightarrowtail {}^{+}C^{*}(N)\) such that \({}^{+}C^{0}\in\hat{\cat{D}}\), \({}^{+}C^{i}\in\cat{D}\) and \(\ker d^{i}\in \cat{X}\) for \(i>0\). If \(M':=\ker d^{1}\) then the short exact sequence \(N\rightarrow {}^{+}C^{0}\rightarrow M'\) is a \(\hat{\cat{D}}\)-hull.
Given AB1 and AB2, each \(N\) in \(\hat{\cat{X}}\) has a \(\cat{DX}\)-resolution and a \(\hat{\cat{D}}\cat{D}\)-coresolution.
Finally, a bounded below \(\cat{D}\)-complex \(D^{*}(N):\dots\rightarrow D^{-1}\rightarrow D^{0}\rightarrow D^{1}\rightarrow\dots\) with \(\ker d^{i}\in\cat{X}\) for all \(i\geq 0\) and its only non-trivial cohomology in degree zero with \(\cH^{0}(D^{*})\cong N\) is called a \emph{\(\cat{D}\)-complex representing \(N\)}. A representing complex splits into (and is reconstructed from) a \(\cat{DX}\)-resolution given by \(\dots\rightarrow D^{-1}\rightarrow \ker d^{0}\twoheadrightarrow \cH^{0}(D^{*})=N\) and a \(\hat{\cat{D}}\cat{D}\)-coresolution \(N\rightarrowtail \coker d^{-1}\rightarrow D^{1}\rightarrow\dots\) where \(N\rightarrowtail \coker d^{-1}\) is induced by \(\ker d^{0}\rightarrowtail D^{0}\).
\begin{lem}\label{lem.res}
Assume \(\xt{1}{\cat{A}}{\cat{X}}{\hat{\cat{D}}}=0\).
Suppose \(f:N_{1}\rightarrow N_{2}\) is in \(\hat{\cat{X}}\)\textup{.} Assume \(F^{*}(N_{i})\) exists for \(i=1,2\) where \(F^{*}(N_{i})\) denotes one of the complexes \({}^{-}C^{*}(N_{i})\), \({}^{+}C^{*}(N_{i})\) or \(D^{*}(N_{i})\)\textup{.} Then \(f\) can be extended to an arrow of chain complexes \(f^{*}:F^{*}(N_{1})\rightarrow F^{*}(N_{2})\) which is uniquely defined up to homotopy\textup{.}
Assume \textup{AB1-3} for the triple of categories \((\cat{A},\cat{X},\cat{D})\)\textup{.} Then \(N\mapsto {}^{-}C^{*}(N)\)\textup{,} \(N\mapsto {}^{+}C^{*}(N)\) and \(N\mapsto D^{*}(N)\) induce functors to the homotopy categories of chain complexes as follows\textup{:}
\begin{equation*}
{}^{-}C^{*}:\hat{\cat{X}}\rightarrow \cat{K}^{\textnormal{b}}(\cat{X})\quad\quad
{}^{+}C^{*}:\hat{\cat{X}}\rightarrow \cat{K}^{+}(\hat{\cat{D}})\quad\quad
D^{*}:\hat{\cat{X}}/\cat{D}\rightarrow \cat{K}^{+}(\cat{D})
\end{equation*}
\end{lem}
\begin{proof}
The proof for \({}^{-}C^{*}(N)\) and \({}^{+}C^{*}(N)\) follows standard lines for constructing chain maps and homotopies. The assumption \(\xt{1}{\cat{A}}{\cat{X}}{\hat{\cat{D}}}=0\) is used every time a lifting or extension of an arrow is required.
Let \((D^{*}_{i}, d_{i}^{*})=D^{*}(N_{i})\) and let \(M_{i}=\ker d_{i}^{0}\) and \(L_{i}=\im d_{i}^{-1}\). Then there are short exact sequences \(L_{i}\rightarrow M_{i}\rightarrow N_{i}\) which by assumption are \(\cat{X}\)-approximations. Since \(\xt{1}{\cat{A}}{M_{1}}{L_{2}}=0\), the arrow \(N_{1}\rightarrow N_{2}\) extends to the \(\cat{X}\)-approximation and further on to the negative part of the complexes. If \(M'_{i}=\ker d^{1}_{i}\) then the \(M'_{i}\) are in \(\cat{X}\) by assumption and there are short exact sequences \(M_{i}\rightarrow D^{0}_{i}\rightarrow M'_{i}\). There is an extension of \(M_{1}\rightarrow D^{0}_{2}\) to \(D^{0}_{1}\rightarrow D^{0}_{2}\) and an induced arrow \(M'_{1}\rightarrow M'_{2}\) which again extends and so on to a chain map \(f^{*}:D^{*}_{1}\rightarrow D^{*}_{2}\).
Let \(g^{*}:D^{*}_{1}\rightarrow D^{*}_{2}\) be a chain map, put \(g=\cH^{0}(g^{*})\), \(s=f{-}g\) and \(s^{*}=f^{*}{-}g^{*}\). Suppose \(s\) factors through \(D\) in \(\cat{D}\); \(s=ab\) with \(a:D\rightarrow N_{2}\). Since \(\xt{1}{\cat{A}}{D}{L_{2}}=0\) there exist a lifting \(\tilde{a}:D\rightarrow M_{2}\) of \(a\). Put \(h_{N}=\tilde{a}b\) and continue similarly to construct a homotopy \(h\) for the extended negative part:
\begin{equation*}
\xymatrix@C+6pt@R-0pt@H+6pt{
& \dots \ar[r] & D_{1}^{-1} \ar[dl]_{h^{-1}}\ar[r]\ar[d]|-{s^{-1}} & M_{1} \ar[dl]|-{h_{M}}\ar[r]\ar[d]|-{\mr{Z}^{0}\!(s^{*})} & N_{1} \ar[dl]|-{h_{N}}\ar[d]^{s}\ar[r] & 0 \\
\dots \ar[r] & D_{2}^{-2}\ar[r] & D_{2}^{-1} \ar[r] & M_{2} \ar[r] & N_{2} \ar[r] & 0
}
\end{equation*}
In particular \(h_{M}:M_{1}\rightarrow D^{-1}_{2}\) can be extended to an \(h^{0}:D^{0}_{1}\rightarrow D^{-1}_{2}\) with \(s^{-1}=h^{0}d^{-1}_{1}{+}d^{-2}_{2}h^{-1}\). The construction of the \(h^{i}\) for \(i>0\) is standard.
\end{proof}
\begin{lem}\label{lem.xres}
Assume \textup{AB1-3} for the triple of categories \((\cat{A},\cat{X},\cat{D})\)\textup{.} Given an exact sequence \(\varepsilon:0\rightarrow N_{1}\rightarrow N_{2}\rightarrow N_{3}\rightarrow 0\) with objects in \(\hat{\cat{X}}\)\textup{.} Then there are exact sequences of complexes where \(\varepsilon\) equals the cohomology\textup{:}
\begin{enumerate}
\item[(i)] \(0\rightarrow {}^{-}C^{*}(N_{1})\longrightarrow {}^{-}C^{*}(N_{2})\longrightarrow {}^{-}C^{*}(N_{3})\rightarrow 0\)
\item[(ii)] \(0\rightarrow {}^{+}C^{*}(N_{1})\longrightarrow {}^{+}C^{*}(N_{2})\longrightarrow {}^{+}C^{*}(N_{3})\rightarrow 0\)
\item[(iii)] \(0\rightarrow D^{*}(N_{1})\longrightarrow D^{*}(N_{2})\longrightarrow D^{*}(N_{3})\rightarrow 0\) \textup{(}termwise split exact\textup{)}
\end{enumerate}
\end{lem}
\begin{proof}
Choose \(\cat{X}\)-approximations \(L_{i}\rightarrow M_{i}\rightarrow N_{i}\) for \(i=1,3\). There is an \(3{\times}3\) commutative diagram of \(6\) short exact sequences which extends the ``horseshoe'' diagram, cf.\ \cite[1.12.11]{has:00}. One obtains an \(\cat{X}\)-approximation of \(N_{2}\) and short exact sequences \(m:M_{1}\rightarrow M_{2}\rightarrow M_{3}\) and \(L_{1}\rightarrow L_{2}\rightarrow L_{3}\) in \(\cat{X}\) and \(\hat{\cat{D}}\) respectively since both categories are closed by extensions (by AB1 and \cite[3.8]{aus/buc:89}). If \(D^{*}_{i}[1]\xra{\eta_{i}} L_{i}\) are finite \(\cat{D}\)-resolutions then since \(\xt{1}{\cat{A}}{D^{-1}_{3}}{L_{1}}=0\) there is a lifting \(\tilde{\eta}_{3}:D^{-1}_{3}\rightarrow L_{2}\) of \(\eta_{3}\) which combined with \(\eta_{1}\) gives \(\eta_{2}:D^{-1}_{1}\coprod D^{-1}_{3}\rightarrow L_{2}\). The kernels of the resulting arrows between short exact sequences give a short exact sequence of objects in \(\hat{\cat{D}}(S)\). The argument is repeated. Splicing with \(m\) in degree zero the short exact sequence of \({}^{-}C^{*}\)-resolutions in (i) is obtained.
Choose short exact sequences \(M_{i}\rightarrow D^{0}_{i}\rightarrow M_{i}'\) for \(i=1,3\) as in AB2. Since \(\xt{1}{\cat{A}}{M_{3}}{D^{0}_{1}}=0\) there is an extension to an arrow of short exact sequences from \(m\) to \(D^{0}_{1}\rightarrow D^{0}_{2}\rightarrow D^{0}_{3}\) with \(D^{0}_{2}=D^{0}_{1}\coprod D^{0}_{3}\) and \(M_{2}':=\coker(M_{2}\rightarrow D^{0}_{2})\in \cat{X}\) by AB1. Repeated application of this argument gives a short exact sequence of \(\cat{D}\)-coresolutions and splicing with the sequences in (i) gives (iii). Pushout of \(M_{i}\rightarrow D^{0}_{i}\rightarrow M_{i}'\) along \(M_{i}\rightarrow N_{i}\) gives a short exact sequence of \(\hat{\cat{D}}\)-hulls and splicing with \(D^{1}_{i}\rightarrow D^{2}_{i}\rightarrow\dots\) gives (ii).
\end{proof}
\subsection{Base change}
The main tool for reducing properties to the fibres in a flat family will be the base change theorem.
We follow the quite elementary and general approach of A.\ Ogus and G.\ Bergman \cite{ogu/ber:72}.
\begin{defn}
Let \(h:S\rightarrow T\) be a ring homomorphism, \(I\) an \(S\)-module, \(N\) a \(T\)-module. For each \(u\in I\) there is a natural \(T\)-linear map \(\phi(u): N\rightarrow N{\otimes}_{S}I\); \(n\mapsto n{\otimes} u\). Let \(F\) be an \(S\)-linear functor of some additive subcategory of \(\cat{Mod}_{S}\) to \(\cat{Mod}_{T}\). Then the \emph{exchange map \(e_{I}\) for \(F\)} is defined as the \(T\)-linear map \(e_{I}:F(S){\otimes}_{S}I\rightarrow F(I)\) given by \(\xi{\otimes} u\mapsto \phi(u)_{*}\xi\). Let \(\mSpec T\) denote the set of closed points in \(\Spec T\).
\end{defn}
\begin{prop}\label{prop.nakayama}
Let \(h:S\rightarrow T\) be a ring homomorphism with \(S\) noetherian\textup{.} Suppose \(\{F^{q}:\cat{mod}_{S}\rightarrow \cat{mod}_{T}\}_{q\geq 0}\) is an \(h\)-linear cohomological \(\delta\)-functor\textup{.}
\begin{enumerate}
\item[(i)] If the exchange map \(e_{S/\fr{n}}^{q}:F^{q}(S){\otimes}_{S}S/\fr{n}\rightarrow F^{q}(S/\fr{n})\) is surjective for all \(\fr{n}\) in \(Z=\im\{\mSpec T\rightarrow\Spec S\}\)\textup{,} then \(e_{I}^{q}:F^{q}(S){\otimes}_{S}I\rightarrow F^{q}(I)\) is an isomorphism for all \(I\) in \(\cat{mod}_{S}\)\textup{.}
\item[(ii)] If \(e_{S/\fr{n}}^{q+1}\) is surjective for all \(\fr{n}\) in \(Z\)\textup{,} then \(e_{I}^{q}\) is an isomorphism for all \(I\) in \(\cat{mod}_{S}\) if and only if \(F^{q+1}(S)\) is \(S\)-flat\textup{.}
\end{enumerate}
\end{prop}
Note that if the \(F^{q}\) in addition extend to functors of all \(S\)-modules \(F^{q}:\cat{Mod}_{S}\rightarrow\cat{Mod}_{T}\) which commute with direct limits, then the conclusions are valied for all \(I\) in \(\cat{Mod}_{S}\).
\begin{ex}\label{ex.nakcplx}
Suppose \(S\) and \(T\) are noetherian. Let \(K^{*}:\, K^{0}\rightarrow K^{1}\rightarrow\dots\) be a complex of \(S\)-flat and finite \(T\)-modules. Define \(F^{q}:\cat{mod}_{S}\rightarrow\cat{mod}_{T}\) by \(F^{q}(I)=\cH^{q}(K^{*}{\otimes}_{S}I)\). Then \(\{F^{q}\}_{q\geq 0}\) is an \(h\)-linear cohomological \(\delta\)-functor which extends to all \(S\)-modules and commutes with direct limits.
\end{ex}
\begin{ex}\label{ex.nakayama}
Suppose \(S\) and \(T\) are noetherian. Let \(M\) and \(N\) be finite \(T\)-modules with \(N\) \(S\)-flat. Then the functors \(F^{q}:\cat{mod}_{S}\rightarrow\cat{mod}_{T}\) defined by \(F^{q}(I)=\xt{q}{T}{M}{N{\otimes}_{S}I}\) for \(q\geq 0\) give an \(h\)-linear cohomological \(\delta\)-functor which extends to all \(S\)-modules and commutes with direct limits.
\end{ex}
Let \(S\rightarrow T\) and \(S\rightarrow S'\) be ring homomorphisms, \(M\) a \(T\)-module, \(T'=T{\otimes}_{S}S'\) and \(N'\) a \(T'\)-module. Then there is a change of rings spectral sequence
\begin{equation}\label{eq.ss}
\cE_{2}^{p,q}=\xt{q}{T'}{\tor{S}{p}{M}{S'}}{N'}\Rightarrow\xt{p+q}{T}{M}{N'}
\end{equation}
which, in addition to the isomorphism \(\hm{}{T'}{M{\otimes}_{S}S'}{N'}\cong\hm{}{T}{M}{N'}\), gives edge maps \(\xt{q}{T'}{M{\otimes}_{S}S'}{N'}\rightarrow\xt{q}{T}{M}{N'}\) for \(q>0\) which are isomorphisms too if \(M\) (or \(S'\)) is \(S\)-flat. If \(I'\) is an \(S'\)-module we can compose the exchange map \(e^{q}_{I'}\) (regarding \(I'\) as \(S\)-module) with the inverse of this edge map for \(N'=N{\otimes}_{S}I'\) and obtain a map \(c^{q}_{I'}\) of \(T'\)-modules
\begin{equation}\label{eq.basechange}
c^{q}_{I'}:\xt{q}{T}{M}{N}{\otimes}_{S}I'\rightarrow\xt{q}{T'}{M{\otimes}_{S}S'}{N{\otimes}_{S}I'}\,.
\end{equation}
\begin{rem}
This is the base change map (in the affine case) considered by A.\ Altman and S.\ Kleiman, their conditions are slightly different, see \cite[1.9]{alt/kle:80}.
\end{rem}
We will use the following geometric notation. Suppose \(h:S\rightarrow T\) is a ring homomorphism, \(M\) is a \(T\)-module and \(s\) is a point in \(\Spec S\) with residue field \(k(s)\). Then \(M_{s}\) denotes the fibre \(M{\otimes}_{S}k(s)\) of \(M\) at \(s\) with its natural \(T_{s}=T{\otimes}_{S}k(s)\)-module structure. Now Proposition \ref{prop.nakayama} implies the following:
\begin{cor}\label{cor.xtdef}
Suppose \(S\rightarrow T\) and \(S\rightarrow S'\) are homomorphisms of noetherian rings\textup{,} \(M\) and \(N\) are finite \(T\)-modules\textup{,} \(Z=\im(\mSpec T\rightarrow\Spec S)\) and \(q\) is an integer\textup{.} Assume that \(M\) \textup{(}if \(q>0\)\textup{)} and \(N\) are \(S\)-flat\textup{.}
\begin{enumerate}
\item[(i)] If\, \(\xt{q+1}{T_{s}}{M_{s}}{N_{s}}=0\) for all \(s\) in \(Z\)\textup{,} then \(c^{q}_{I'}\) in \eqref{eq.basechange} is an isomorphism for all \(S'\)-modules \(I'\)\textup{.}
\item[(ii)] If in addition \(\xt{q-1}{T_{s}}{M_{s}}{N_{s}}=0\) for all \(s\in Z\)\textup{,} then \(\xt{q}{T}{M}{N}\) is \(S\)-flat\textup{.}
\end{enumerate}
\end{cor}
\section{Categories fibred in additive categories}\label{sec.cof}
We will phrase our results in the language of fibred categories.\footnote{We have chosen to work with rings instead of (affine) schemes. Our definition of a fibred category \(p:\cat{F}\rightarrow\cat{C}\) reflects this choice and is equivalent to the functor of opposite categories \(p^{\text{op}}:\cat{F}^{\text{op}}\rightarrow\cat{C}^{\text{op}}\) being a fibred category as defined in \cite{FAG}. M.\ Artin called \(p\) a cofibred category in \cite{art:74}, but this seems not to be common usage.} We therefore briefly recall some of the basic notions, taken mainly from A.\ Vistoli's article in \cite{FAG}. Then we define quotients of categories fibred in additive categories.
Consider a category \(\cat{C}\). Given a category over \(\cat{C}\), i.e.\ a functor \(p:\cat{F}\rightarrow\cat{C}\).
To an object \(T\) in \(\cat{C}\), let \(\cat{F}(T)\); the \emph{fiber of \(\cat{F}\) over \(T\)}, denote the subcategory of arrows \(\phi\) in \(\cat{F}\) such that \(p(\phi)=\id_{T}\).
An arrow \(\phi_{1}:\xi\rightarrow\xi_{1}\) in \(\cat{F}\) is \emph{cocartesian} if for any arrow \(\phi_{2}:\xi\rightarrow\xi_{2}\) in \(\cat{F}\) and any arrow \(f_{21}:p(\xi_{1})\rightarrow p(\xi_{2})\) in \(\cat{C}\) with \(f_{21} p(\phi_{1})=p(\phi_{2})\) there exists a unique arrow \(\phi_{21}:\xi_{1}\rightarrow\xi_{2}\) with \(p(\phi_{21})=f_{21}\) and \(\phi_{21}\phi_{1}=\phi_{2}\). If for any arrow \(f:T\rightarrow T'\) in \(\cat{C}\) and any object \(\xi\) in \(\cat{F}\) with \(p(\xi)=T\) there exists a cocartesian arrow \(\phi:\xi\rightarrow \xi'\) for some \(\xi'\) with \(p(\phi)=f\), then \(\cat{F}\) (or rather \(p:\cat{F}\rightarrow\cat{C}\)) is a \emph{fibred category}. Moreover, \(\xi'\) will be called a \emph{base change} of \(\xi\) by \(f\). If \(\xi''\) is another base change of \(\xi\) by \(f\) then \(\xi'\) and \(\xi''\) are isomorphic over \(T'\) by a unique isomorphism. We shall also say that a property \(P\) of objects in the fibres of \(\cat{F}\) is \emph{preserved by base change} if \(P(\xi)\) implies \(P(\xi')\) for any base change \(\xi'\) of \(\xi\).
A morphism of fibred categories is a functor \(F:\cat{F}_{1}\rightarrow\cat{F}_{2}\) with \(p_{2}F=p_{1}\) such that \(\phi\) cocartesian implies \(F(\phi)\) cocartesian. If \(F\) in addition is an inclusion of categories, \(\cat{F}_{1}\) is a fibred subcategory of \(\cat{F}_{2}\).
A category with all arrows being isomorphisms is a groupoid. A fibred category \(\cat{F}\) over \(\cat{C}\) is called a \emph{category fibred in groupoids} (often abbreviated to groupoid) if all fibres \(\cat{F}(T)\) are groupoids. Then all arrows in \(\cat{F}\) are cocartesian. If all fibres \(\cat{F}(T)\) only contain identities, then \(\cat{F}\) is called a \emph{category fibred in sets}.
\begin{lem}\label{lem.gpoid}
Given functors \(F:\cat{F}\rightarrow\cat{G}\) and \(q:\cat{G}\rightarrow\cat{C}\) and suppose \(q\) is fibred in sets\textup{.} Then \(F\) is fibred \textup{(}in groupoids\textup{/}sets\textup{)} if and only if \(qF\) is fibred \textup{(}in groupoids\textup{/}sets\textup{).}
\end{lem}
If \(T\) is an object in a category \(\cat{C}\) let \(\cat{C}/T\) denote the comma category of arrows to \(T\). Then the forgetful functor \(\cat{C}/T\rightarrow\cat{C}\) is fibred in sets. If \(p:\cat{F}\rightarrow\cat{C}\) is fibred (in groupoids/sets), \(\xi\) is an object in \(\cat{F}\) and \(T=p(\xi)\), then there is a natural functor \(p_{\xi}:\cat{F}/\xi\rightarrow\cat{C}/T\). The composition \(\cat{F}/\xi\rightarrow\cat{F}\rightarrow\cat{C}\) is clearly fibred (in groupoids/sets) and hence \(\cat{F}/\xi\rightarrow \cat{C}/T\) is fibred (in groupoids/sets) by Lemma \ref{lem.gpoid}.
If \(p:\cat{F}\rightarrow\cat{C}\) is a functor and \(\cat{C}'\) is a subcategory of \(\cat{C}\) we can define the \emph{restriction} \(p':\cat{F}_{\vert \cat{C}'}\rightarrow\cat{C}'\) of \(\cat{F}\) to \(\cat{C}'\) by picking for \(\cat{F}_{\vert \cat{C}'}\) the objects and morphisms in \(\cat{F}\) that \(p\) takes into \(\cat{C}'\). It follows that \(\cat{F}_{\vert \cat{C}'}\) is fibred (in groupoids/sets) if \(\cat{F}\) is.
The composition of two cocartesian arrows is cocartesian and isomorphisms are cocartesian. Hence the subcategory \(\cat{F}_{\text{coca}}\) of cocartesian arrows in a fibred category \(\cat{F}\) over \(\cat{C}\) is fibred in groupoids. If \(\cat{F}\) is fibred in groupoids there is an associated category fibred in sets \(\bar{\cat{F}}\rightarrow\cat{C}\) defined by identifying all isomorphic objects in all fibres \(\cat{F}(T)\) and identifying arrows accordingly. If \(\cat{F}\) is fibred in sets one defines a functor \(F:\cat{C}\rightarrow\Sets\) by \(F(T):=\cat{F}(T)\) and \(F(f):F(T)\rightarrow F(T')\) is defined by \(F(f)(\xi):=\eta_{\xi,f}\) where \(\phi_{\xi,f}:\xi\rightarrow \eta_{\xi,f}\) is the (in this case) unique cocartesian lifting of \(f\). From a functor \(G:\cat{C}\rightarrow\Sets\) one defines a category fibred in sets, and these two operations are inverse up to natural equivalences.
\begin{defn}\label{defn.addkof}
An \emph{additive \textup{(}abelian\textup{)} category \(\cat{F}\) over \(\cat{C}\)} is a functor \(p:\cat{F}\rightarrow\cat{C}\) such that:
\begin{enumerate}
\item[(i)] The fibre \(\cat{F}(T)\) is an additive (abelian) category for all objects \(T\) in \(\cat{C}\).
\item[(ii)] For all objects \(\xi_{1}\) and \(\xi_{2}\) in \(\cat{F}\) and arrows \(f:p(\xi_{1})\rightarrow p(\xi_{2})\) in \(\cat{C}\),
\begin{equation*}
\hm{}{f}{\xi_{1}}{\xi_{2}}:=\{\phi\in\hm{}{\cat{F}}{\xi_{1}}{\xi_{2}}\,\vert\,p(\phi)=f\}
\end{equation*}
is an abelian group, and composition of arrows
\begin{equation*}
\hm{}{f_{2}}{\xi_{2}}{\xi_{3}}\times\hm{}{f_{1}}{\xi_{1}}{\xi_{2}}\rightarrow\hm{}{f_{2}f_{1}}{\xi_{1}}{\xi_{3}}
\end{equation*}
is bilinear.
\end{enumerate}
A morphism \(F:\cat{F}_{1}\rightarrow\cat{F}_{2}\) of additive (abelian) categories over \(\cat{C}\) is a \emph{linear} functor \(F\) over \(\cat{C}\), i.e.\ which gives linear maps of \(\Hom\)-groups. If in addition \(F\) is an inclusion of categories then \(\cat{F}_{1}\) is an additive (abelian) subcategory of \(\cat{F}_{2}\) over \(\cat{C}\). A category \(\cat{F}\) over \(\cat{C}\) is \emph{fibred in additive \textup{(}abelian\textup{)} categories}, abbreviated by FAd (FAb), if \(\cat{F}\) is both fibred and additive (abelian) over \(\cat{C}\). Morphisms should be linear and preserve cocartesian arrows. A FAd subcategory is a morphism of FAds which is an inclusion of categories. For \(i=1,2\) let \(\cat{A}_{i}\) be a FAb over \(\cat{C}\) and \(\cat{X}_{i}\subseteq\cat{A}_{i}\) a FAd subcategory such that the fibre categories \(\cat{X}_{i}(T)\) are exact. Then a morphism of FAds \(F:\cat{X}_{1}\rightarrow\cat{X}_{2}\) is \emph{exact} if \(F\) preserves short exact sequences for all the fibre categories.
\end{defn}
Note that in a FAd finite (co)products in the fibres are preserved by base change.
Given a FAd subcategory \(\cat{D}\subseteq\cat{F}\). Two arrows \(\phi_{1}\) and \(\phi_{2}\) in \(\cat{F}\) are \(\cat{D}\)-equivalent if \(p(\phi_{1})=p(\phi_{2})\) and \(\phi_{1}-\phi_{2}\) factors through an object in \(\cat{D}\). Write \(\phi_{1}\sim\phi_{2}\).
Define the quotient category \(\cat{F}/\cat{D}\) over \(\cat{C}\) to have the same objects as \(\cat{F}\) and \(\hm{}{\cat{F}/\cat{D}}{\xi_{1}}{\xi_{2}}:=\hm{}{\cat{F}}{\xi_{1}}{\xi_{2}}/\sim\). The natural map to \(\cat{C}\) makes \(\cat{F}/\cat{D}\) an additive category over \(\cat{C}\) and the natural functor \(\cat{F}\rightarrow\cat{F}/\cat{D}\) is linear over \(\cat{C}\).
\begin{lem}\label{lem.keystable}
If \(\phi_{1}:\xi\rightarrow\xi_{1}\) is cocartesian in \(\cat{F}\) and \(\phi:\xi_{1}\rightarrow\xi_{2}\) is any arrow such that \(\phi\phi_{1}\sim 0\) then \(\phi\sim 0\)\textup{.}
\end{lem}
\begin{proof}
Suppose \(\phi\phi_{1}=\beta\alpha\) with \(\alpha:\xi\rightarrow\delta\) and with \(\delta\) in \(\cat{D}\). If \(p(\beta):T'\rightarrow T_{2}\) then since \(\cat{D}\) is a fibred subcategory there exists an arrow \(\delta\rightarrow\delta_{2}\) which is cocartesian in \(\cat{F}\) and with \(p(\delta_{2})=p(\xi_{2})=T_{2}\). Replacing \(\delta\) with \(\delta_{2}\) we assume \(p(\delta)=T_{2}\). Since \(\phi_{1}\) is cocartesian there exists a unique arrow \(\tau:\xi_{1}\rightarrow\delta\) with \(\tau\phi_{1}=\alpha\). Since \(\phi_{1}\) is cocartesian uniqueness implies that \(\beta\tau=\phi\).
\end{proof}
\begin{lem}\label{lem.stable}
Given a FAd subcategory \(\cat{D}\subseteq\cat{F}\) over \(\cat{C}\)\textup{,} then the quotient category \(\cat{F}/\cat{D}\) is FAd over \(\cat{C}\) and the quotient morphism \(\cat{F}\rightarrow\cat{F}/\cat{D}\) is a morphism of FAds\textup{.}
\end{lem}
\begin{proof}
We first show that if \(\phi_{1}:\xi\rightarrow\xi_{1}\) is cocartesian in \(\cat{F}\) then its image \([\phi_{1}]\) in \(\cat{F}/\cat{D}\) is cocartesian. Given \(\phi_{2}:\xi\rightarrow\xi_{2}\) and \(\theta:\xi_{1}\rightarrow\xi_{2}\) with \(\theta\phi_{1}=\phi_{2}\). Suppose \(\theta':\xi_{1}\rightarrow\xi_{2}\) with \(p(\theta')=p(\theta)\) satisfies \(\theta'\phi_{1}\sim\phi_{2}\). If \(\phi=\theta'-\theta\) then \(\phi\phi_{1}\sim 0\) so by Lemma \ref{lem.keystable} \(\phi\sim 0\). Now we show that \([\theta]\) is independent of the representations of the other maps. Let \(\phi_{i}':\xi\rightarrow\xi_{i}\) with \(\phi_{i}'\sim\phi_{i}\) and suppose (as we may) that \(\theta'\) satisfies \(\theta'\phi_{1}'=\phi_{2}'\) with \(p(\theta')=p(\theta)\). Again let \(\phi=\theta'-\theta\). Then \(0\sim\phi'_{2}-\phi_{2}=\theta'\phi'_{1}-\theta\phi_{1}=\theta'(\phi_{1}'-\phi_{1})+\phi\phi_{1}\sim\phi\phi_{1}\). By Lemma \ref{lem.keystable} \(\phi\sim 0\). Given \(f:T\rightarrow T_{1}\) and \(\xi\) in \(\cat{F}/\cat{D}\) with \(p(\xi)=T\) there exists a cocartesian \(\phi_{1}:\xi\rightarrow \xi_{1}\) in \(\cat{F}\) with \(p(\phi_{1})=f\) and by what we have done \([\phi_{1}]\) is cocartesian in \(\cat{F}/\cat{D}\).
\end{proof}
Note that there are in general more cocartesian arrows in \(\cat{F}/\cat{D}\) than those in the image of cocartesian arrows in \(\cat{F}\). The following lemma characterises the cocartesian arrows in the quotient category:
\begin{lem}\label{lem.stablecc}
If \(\rho\) and \(\theta\) are composeable arrows in \(\cat{F}\) with \(\rho\) cocartesian and \(\theta\) inducing an isomorphism in \(\cat{F}/\cat{D}\)\textup{,} then \([\theta\rho]\) is cocartesian in \(\cat{F}/\cat{D}\)\textup{.}
Conversely\textup{,} suppose \([\phi]:\xi_{1}\rightarrow\xi_{2}\) is cocartesian in \(\cat{F}/\cat{D}\) over \(f:T_{1}\rightarrow T_{2}\)\textup{.} Then for any base change
\(\rho:\xi_{1}\rightarrow \xi_{1}^{\#}\) of \(\xi_{1}\) over \(f\) in \(\cat{F}\)\textup{,} the induced arrow \(\phi^{\#}:\xi_{1}^{\#}\rightarrow\xi_{2}\) gives an isomorphism in \(\cat{F}/\cat{D}(T_{2})\)\textup{.}
\end{lem}
\begin{proof}
If \(\rho:\xi_{1}\rightarrow\xi_{2}\) is cocartesian and \([\theta]:\xi_{2}\rightarrow \xi_{3}\) is an isomorphism, let \(\phi=\theta\rho\). If \(\tau:\xi_{1}\rightarrow \xi_{4}\), \(p(\xi_{i})=T_{i}\) and there is a map \(f:T_{3}\rightarrow T_{4}\) with \(p(\tau)=f p(\theta) p(\rho)\), then there is a unique arrow \(\mu:\xi_{2}\rightarrow\xi_{4}\) above \(f p(\theta)\) with \(\mu\rho=\tau\). This gives the arrow \([\mu][\theta]^{-1}:\xi_{3}\rightarrow\xi_{4}\). If \(\mu_{i}:\xi_{3}\rightarrow\xi_{4}\) for \(i=1,2\) are two arrows with \([\mu_{i}][\phi]=[\tau]\), then \([\mu_{1}][\theta]=[\mu_{2}][\theta]\) since \([\rho]\) is cocartesian in \(\cat{F}/\cat{D}\) by Lemma \ref{lem.stable}. Since \([\theta]\) is an isomorphism, \([\mu_{1}]=[\mu_{2}]\).
Conversely, since \([\phi]\) is cocartesian there is a unique arrow \([\psi]:\xi_{2}\rightarrow\xi_{1}^{\#}\) in \(\cat{F}/\cat{D}(T_{2})\) with \([\psi\phi]=[\rho]\). By Lemma \ref{lem.stable} \([\rho]\) is cocartesian. It follows that \([\phi^{\#}]=[\psi]^{-1}\).
\end{proof}
\section{Cohen-Macaulay approximation in fibred categories}
Given a category \(\cat{C}\) and a category \(\cat{A}\) fibred in abelian categories over \(\cat{C}\). Base change by an \(f:T\rightarrow T'\) in \(\cat{C}\) applied to the objects in a complex \(\dots\rightarrow N_{d}\rightarrow N_{d-1}\rightarrow\dots\) in \(\cat{A}(T)\) can by Lemma \ref{lem.keystable} be uniquely extended to a complex and yield a commutative diagram where the vertical arrows are the cocartesian base change arrows:
\begin{equation*}
\xymatrix@C-0pt@R-12pt@H0pt{
\dots \ar[r] & N_{d+1} \ar[r]\ar[d] & N_{d} \ar[r]\ar[d] & N_{d-1} \ar[d]\ar[r] & \dots \\
\dots \ar[r] & N_{d+1}^{\#} \ar[r] & N_{d}^{\#} \ar[r] & N_{d-1}^{\#} \ar[r] & \dots
}
\end{equation*}
Similarly base change of a commutative diagram \(\Delta\) in \(\cat{A}(T)\) gives a commutative diagram \(\Delta^{\#}\) and the base change arrows give an arrow of diagrams \(\Delta\rightarrow\Delta^{\#}\).
Let \(\cat{X}\subseteq \cat{A}\) be a FAd subcategory.
Consider the following two conditions on the pair \((\cat{A},\cat{X})\) and an object \(T\) in \(\cat{C}\).
\begin{enumerate}
\item[(BC1)] If \(\alpha:A_{1}\rightarrow A_{2}\) is an epimorphism in \(\cat{A}(T)\) and \(f:T\rightarrow T'\) is an arrow in \(\cat{C}\) then any base change of \(\alpha\) by \(f\) is an epimorphism in \(\cat{A}(T')\).
\item[(BC2)] Let \(\xi:0\rightarrow A\rightarrow B\rightarrow M\rightarrow 0\) be an exact sequence in \(\cat{A}(T)\) with \(M\) in \(\cat{X}(T)\) and \(f:T\rightarrow T'\) is an arrow in \(\cat{C}\). Then any base change of \(\xi\) by \(f\) is an exact sequence in \(\cat{A}(T')\).
\end{enumerate}
The first condition would be satisfied if base change had a right adjoint. The second condition mimics flatness for all objects in \(\cat{X}(T)\).
The following is an elementary, but essential technical consequence of BC1.
\begin{lem}\label{lem.bs}
Let \(\cat{A}\) be a category fibred in abelian categories over \(\cat{C}\) which satisfies \textup{BC1} for \(T\) in \(\cat{C}\)\textup{.} Let \(c:\dots\rightarrow L_{n}\rightarrow L_{n-1}\rightarrow\dots\) be an acyclic complex in \(\cat{A}(T)\) which remains exact after a base change \(\dots\rightarrow L_{n}^{\#}\rightarrow L_{n-1}^{\#}\rightarrow\dots\) of \(c\) by \(f:T\rightarrow T'\)\textup{.} Then base change of \(K_{n}:=\ker\{d_{n-1}:L_{n-1}\rightarrow L_{n-2}\}\) by \(f\) is isomorphic to \(\ker d_{n-1}^{\#}\) for all \(n\)\textup{.}
\end{lem}
\begin{proof}
Let \(Q_{n}=\ker d_{n-1}^{\#}\). Since the composition \(K_{n}^{\#}\rightarrow L_{n-1}^{\#}\rightarrow L_{n-2}^{\#}\) by Lemma \ref{lem.keystable} is zero (as \(K_{n}\rightarrow K_{n}^{\#}\) is cocartesian), there is a factorisation \(\rho: K_{n}^{\#}\rightarrow Q_{n}\) of \(K_{n}^{\#}\rightarrow L_{n-1}^{\#}\). On the other hand the composition \(L_{n+1}^{\#}\rightarrow L_{n}^{\#}\rightarrow K_{n}^{\#}\) is zero too, hence there is an arrow from \(\coker d_{n+1}^{\#}\cong Q_{n}\) to \(K_{n}^{\#}\) which is a section of \(\rho\). By assumption \(L_{n}^{\#}\rightarrow K_{n}^{\#}\) is an epimorphism. It follows that \(Q_{n}\cong K_{n}^{\#}\).
\end{proof}
\begin{defn}\label{defn.f}
Given FAd subcategories \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\).
Let \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) denote the additive subcategory of \(\cat{A}(T)\) with objects \(N\) which have a finite \(\cat{X}\)-resolution \(M_{*}\rightarrow N\) which is preserved as resolution by any base change. Let \(\hat{\cat{X}}{}^{\textnormal{fl}}\subseteq \cat{A}\) denote the resulting FAd subcategory.
Let \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) denote the additive subcategory of \(\cat{A}(T)\) with objects \(L\) which have a \(\cat{D}(T)\)-resolution \(D^{*}\rightarrow L\) which is preserved as resolution by any base change. Let \(\hat{\cat{D}}{}^{\textnormal{fl}}\subseteq\cat{A}\) denote the resulting FAd subcategory.
\end{defn}
The reasoning in the beginning of this section combined with Lemma \ref{lem.res} gives the following.
\begin{lem}\label{lem.forall}
Let \(\eta:\dots \rightarrow E_{n}\rightarrow E_{n-1}\rightarrow\dots\) and \(\lambda:\dots\rightarrow F_{n}\rightarrow F_{n-1}\rightarrow\dots\) be complexes in \(\cat{A}(T)\) and \(\eta^{\#}\) and \(\lambda^{\#}\) the complexes resulting from base change over \(f:T\rightarrow T'\)\textup{.} If \(\eta\) is homotopic to \(\lambda\) then \(\eta^{\#}\) is homotopic to \(\lambda^{\#}\)\textup{.}
In particular\textup{;} if \(N\) in \(\cat{A}(T)\) has one \(\cat{DX}\)-resolution \textup{(}\(\cat{\hat{D}D}\)-coresolution\textup{)} which is preserved by base change then all \(\cat{DX}\)-resolutions \textup{(}\(\cat{\hat{D}D}\)-coresolutions\textup{)} are preserved by base change\textup{.}
\end{lem}
\begin{thm}\label{thm.cofapprox}
Let \(\cat{A}\) be a category fibred in abelian categories over \(\cat{C}\) and let \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\) be inclusion morphisms of categories fibred in additive categories\textup{.} Fix an object \(T\) in \(\cat{C}\)\textup{.} Assume \textup{BC1-2} for \((\cat{A},\cat{X})\) and \(T\)\textup{,} and \textup{AB1-2} for the triple of categories \((\cat{A}(T),\cat{X}(T),\cat{D}(T))\)\textup{.}
Then any object \(N\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) admits an \(\cat{X}(T)\)-approximation and a \(\hat{\cat{D}}(T)\)-hull\textup{;}
\begin{equation*}\label{eq.cofapprox}
0\rightarrow L\longrightarrow M\longrightarrow N\rightarrow 0 \quad \text{and} \quad 0\rightarrow N\longrightarrow L'\longrightarrow M'\rightarrow 0
\end{equation*}
with \(M\) and \(M'\) in \(\cat{X}(T)\) and \(L\) and \(L'\) in \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\)\textup{,} which are preserved by any base change\textup{.}
\end{thm}
\begin{proof}
The proof is a variation of the original proof of \cite[1.1]{aus/buc:89}.
For every \(N\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) let \(r(N)\) denote the minimal length of an \(\cat{X}(T)\)-resolution \(M_{*}\twoheadrightarrow N\) which is preserved by base change. The proof is by induction on \(r(N)\). If \(r(N)=0\) then \(N\) is in \(\cat{X}\) and so is its own \(\cat{X}\)-approximation, while AB2 provides a short exact sequence \(N\rightarrow D\rightarrow M'\) which is a \(\hat{\cat{D}}(T)\)-hull with \(D\) in \(\cat{D}(T)\subseteq\hat{\cat{D}}{}^{\textnormal{fl}}(T)\). The approximation is trivially preserved by base change, the hull because of BC2. Assume \(r=r(N)>0\) and let \(0\rightarrow M_{r}\rightarrow\dots\rightarrow M_{0}\twoheadrightarrow N\) be an \(\cat{X}(T)\)-resolution of minimal length preserved by base change. Then \(N_{1}=\ker(M_{0}\twoheadrightarrow N)\) is in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) by Lemma \ref{lem.bs} and \(r(N_{1})=r-1\). By induction there is a \(\hat{\cat{D}}(T)\)-hull \(N_{1}\rightarrow L\rightarrow M'_{1}\) with \(L\) in \(\hat{\cat{D}}{}^{\textnormal{fl}}\) which is preserved by base change. Pushout of \(e:N_{1}\rightarrow M_{0}\rightarrow N\) along \(N_{1}\rightarrow L\) gives an \(\cat{X}(T)\)-approximation \(L\rightarrow M\rightarrow N\) by AB1. In the commutative diagram obtained by a base change;
\begin{equation}
\xymatrix@C-0pt@R-12pt@H0pt{
N_{1}^{\#} \ar[r]\ar[d] & M_{0}^{\#} \ar[r]\ar[d] & N^{\#} \ar@{=}[d] \\
L^{\#} \ar[r]\ar[d] & M^{\#} \ar[r]\ar[d] & N^{\#} \\
(M'_{1})^{\#} \ar@{=}[r] & (M'_{1})^{\#}
}
\end{equation}
the upper row (by Lemma \ref{lem.bs}) and the columns (by BC2) are short exact sequences.
It follows that the middle row is a short exact sequence.
By AB2 there is a short exact sequence \(M\rightarrow D\rightarrow M'\) with \(D\) in \(\cat{D}(T)\) and \(M'\) in \(\cat{X}(T)\). Pushout of \(M\rightarrow D\rightarrow M'\) along \(M\rightarrow N\) gives a short exact sequence \(h:N\rightarrow L'\rightarrow M'\). Since the induced sequence \(L\rightarrow D\rightarrow L'\) is short exact, \(L'\) is contained in \(\hat{\cat{D}}(T)\). Applying a base change we obtain the following commutative diagram:
\begin{equation}
\xymatrix@C-0pt@R-12pt@H0pt{
L^{\#} \ar[r]\ar@{=}[d] & M^{\#} \ar[r]\ar[d] & N^{\#} \ar[d] \\
L^{\#} \ar[r] & D^{\#} \ar[r]\ar[d] & (L')^{\#} \ar[d] \\
& (M')^{\#} \ar@{=}[r] & (M')^{\#}
}
\end{equation}
The upper row and (by BC2) the two columns are short exact sequences. It follows that the middle row is a short exact sequence and hence that \(L'\) is contained in \(\hat{\cat{D}}{}^{\textnormal{fl}}\).
\end{proof}
Sequences as in Theorem \ref{thm.cofapprox} preserved by any base change will be called an \emph{\(\cat{X}\)-approximation} and a \emph{\(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull} of \(N\) respectively.
Lemma \ref{lem.stable} makes the following definition reasonable. Three categories fibred in additive categories (FAds) \(\cat{A}_{i}\), \(i=1,2,3\), an inclusion of FAds \(\cat{A}_{1}\subseteq \cat{A}_{2}\), and a morphism of FAds \(F:\cat{A}_{2}\rightarrow \cat{A}_{3}\) equivalent to the quotient morphism \(\cat{A}_{2}\rightarrow\cat{A}_{2}/\cat{A}_{1}\) is called a \emph{short exact sequence of categories fibred in additive categories} and is denoted by \(0\rightarrow \cat{A}_{1}\rightarrow\cat{A}_{2}\rightarrow\cat{A}_{3}\rightarrow 0\).
\begin{thm}\label{thm.cofmain}
Let \(\cat{A}\) be a category fibred in abelian categories over \(\cat{C}\) and let \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\) be inclusion morphisms of categories fibred in additive categories\textup{.} Assume \textup{BC1-2} for the pair \((\cat{A},\cat{X})\) and \textup{AB1-3} for the triple of categories \((\cat{A}(T),\cat{X}(T),\cat{D}(T))\)\textup{,} for all objects \(T\) in \(\cat{C}\)\textup{.}
Then\textup{:}
\begin{enumerate}
\item[(i)] The \(\cat{X}\)-approximation induces a morphism of categories fibred in additive categories \(j^{!}:\hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\rightarrow\cat{X}/\cat{D}\) which is a right adjoint to the full and faithful inclusion morphism \(j_{!}:\cat{X}/\cat{D}\rightarrow\hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\)\textup{.}
\item[(ii)] The \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull induces a morphism of categories fibred in additive categories \(i^{*}:\hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\rightarrow\hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\) which is a left adjoint to the full and faithful inclusion morphism \(i_{*}:\hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\rightarrow \hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\)\textup{.}
\item[(iii)] Together these maps give the following commutative diagram of short exact sequences of categories fibred in additive categories\textup{:}
\begin{equation*}
\xymatrix@C-0pt@R-8pt@H-30pt{
0\ar[r] & \, \hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D} \ar[r]^{i_{*}}\ar[d]_{\id} & \hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\ar[r]^{j^{!}}\ar@{=}[d] & \cat{X}/\cat{D} \ar[r] & 0 \\
0 & \hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\ar[l] & \hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\ar[l]_{i^{*}} & \, \cat{X}/\cat{D} \ar[l]_{j_{!}}\ar[u]_{\id} & 0 \ar[l]
}
\end{equation*}
\end{enumerate}
\end{thm}
\begin{proof}
In each fibre most of these statements are true by the arguments in the proof of \cite[2.8]{aus/buc:89} since we have Theorem \ref{thm.cofapprox}. The general cases are reduced to fibre cases by applying base change. First we have to establish the functors. Note that the quotient categories involved are FAds over \(\cat{C}\) by Lemma \ref{lem.stable}. Let \(p:\cat{A}\rightarrow\cat{C}\) denote the fibration. For each \(N_{i}\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}\) put \(T_{i}=p(N_{i})\) and choose a \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull \(\iota_{i}:N_{i}\rightarrow L_{i}\rightarrow M_{i}\) which exists by Theorem \ref{thm.cofapprox} and such that \(\iota_{i}=\id\) if \(N_{i}\) is in \(\hat{\cat{D}}{}^{\textnormal{fl}}\). For each arrow \(\psi:N_{1}\rightarrow N_{2}\) choose an arrow \(\lambda_{21}:L_{1}\rightarrow L_{2}\) commuting with \(\psi\). This arrow is obtained as a composition of a base change \(L_{1}\rightarrow L_{1}^{\#}\) over \(p(\psi):T_{1}\rightarrow T_{2}\) with an extension \(L_{1}^{\#}\rightarrow L_{2}\) of \(N_{1}^{\#}\rightarrow L_{2}\) obtained since \(\xt{1}{\cat{A}(T_{2})}{M_{1}^{\#}}{L_{2}}=0\) by \cite[2.5]{aus/buc:89}. If composeable it follows from \cite[2.8]{aus/buc:89} that \(\lambda_{32}\lambda_{21}\sim\lambda_{31}\).
There is a unique arrow \(\phi:N_{1}^{\#}\rightarrow N_{2}\) induced by \(\psi\).
If \(\lambda_{21}':L_{1}\rightarrow L_{2}\) is an extension of \(\psi':N_{1}\rightarrow N_{2}\) with \(p(\psi')=p(\psi)\) such that \(\delta_{1}:=\psi-\psi'\) is equivalent to \(0\), we have by Lemma \ref{lem.keystable} that \(\delta:N_{1}^{\#}\rightarrow N_{2}\) induced from \(\delta_{1}\) by base change factors through an object \(D\) in \(\cat{D}(T_{2})\). It follows that \(\delta\) factors through \(N_{1}^{\#}\rightarrow L_{1}^{\#}\). Let \(\tau\) denote the composition \(L_{1}^{\#}\rightarrow N_{2}\rightarrow L_{2}\) (so \(\tau\sim 0\)).
Let \(\eta\) be a base change over \(p(\psi)\) of the difference of the two extensions; \(\eta=(\lambda_{21}-\lambda_{21}')^{\#}\). One calculates that \((\eta-\tau)\iota_{1}^{\#}=0\), hence \(\eta-\tau\) is induced by an arrow \(M_{1}^{\#}\rightarrow L_{2}\) which lifts to an arrow \(M_{1}^{\#}\rightarrow D^{0}\) where \(D^{*}\twoheadrightarrow L_{2}\) is a finite \(\cat{D}\)-resolution of \(L_{2}\) (since \(\xt{1}{\cat{A}(T_{2})}{M_{1}^{\#}}{\hat{\cat{D}}(T_{2})}=0\)).
Hence \(\eta-\tau\sim 0\), so \(\eta\sim 0\) and \(\lambda_{21}\sim\lambda_{21}'\). We have shown that \(i^{*}:\hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\rightarrow \hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\) is a well defined functor. To show that \(i^{*}\) preserves cocartesian arrows we apply Lemma \ref{lem.stablecc}: If \([\psi]:N_{1}\rightarrow N_{2}\) is cocartesian in \(\hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\) then the induced map \([\phi]:N_{1}^{\#}\rightarrow N_{2}\) is an isomorphism and by \cite[2.8]{aus/buc:89} so is any extension \(L_{1}^{\#}\rightarrow L_{2}\) of \([\phi]\). Composed with the base change \(L_{1}\rightarrow L_{1}^{\#}\) we get a cocartesian arrow in \(\hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\) by Lemma \ref{lem.stablecc}.
A similar argument gives that the morphism \(j^{!}: \hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\rightarrow \cat{X}/\cat{D}\) induced by (choices of) \(\cat{X}\)-approximation also is well defined as a map of fibred categories.
To prove adjointness for the pair \((j_{!},j^{!})\) consider the chosen \(\cat{X}\)-approximation \(L\rightarrow M\xra{\pi} N\) of \(N\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\). Given \(\phi_{1}:M_{1}\rightarrow N\) with \(M_{1}\) in \(\cat{X}(T_{1})\) and \(f=p(\phi_{1})\). Let \(\phi: M_{1}^{\#}\rightarrow N\) be induced by a base change of \(M_{1}\) by \(f\). Since \(\xt{1}{\cat{A}(T)}{M_{1}^{\#}}{L}=0\), \(\phi\) can be lifted to an arrow \(\psi:M_{1}^{\#}\rightarrow M\). Composing \(\psi\) with the base change \(M_{1}\rightarrow M_{1}^{\#}\) gives a lifting of \(\phi_{1}\) which shows surjectivity of the adjointness map \([\pi\circ -]\). To prove injectivity consider for \(i=2,3\) arrows \(\psi_{i}:M_{1}\rightarrow M\) in \(\cat{X}\) with \(\pi\psi_{2}=\pi\psi_{3}\). Since \(p(\pi)=\id\) we have \(p(\psi_{2})=p(\psi_{3})=f\) and we can define \(\psi_{1}=\psi_{2}-\psi_{3}\) with \(\pi\psi_{1}\sim 0\). Base change by \(f\) induces a \(\psi:M_{1}^{\#}\rightarrow M\) from \(\psi_{1}\). Lemma \ref{lem.keystable} gives \(\pi\psi\sim 0\). The argument in \cite[2.8]{aus/buc:89} implies that \(\psi\) and hence \(\psi_{1}\) factors through an object in \(\cat{D}(T)\). Analogous arguing gives the adjointness of the pair \((i^{*},i_{*})\).
The commutativity of the diagram in \textup{(iii)} follows by definition. For \(i^{*}j_{!}=0=j^{!}i_{*}\) see \cite[2.8]{aus/buc:89}. We prove exactness in the upper row. Given \(\phi:N_{1}\rightarrow N_{2}\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}\) with \(f=p(\phi):T_{1}\rightarrow T_{2}\) such that \(j^{!}[\phi]=0\). If \(\pi_{i}:M_{i}\rightarrow N_{i}\) are the chosen \(\cat{X}\)-approximations, \(j^{!}[\phi]\) is represented by a lifting \(\psi:M_{1}\rightarrow M_{2}\) and the assumption is that \(\psi\) factors through an object \(D\) of \(\cat{D}\). We claim that \(\phi\) factors through an object in \(\hat{\cat{D}}{}^{\textnormal{fl}}\). By base change it's sufficient to prove the special case \(f=\id_{T_{2}}\). If \(M\) is any object in \(\cat{X}(T)\) we have that the composition \(\xt{1}{\cat{A}(T)}{M}{N_{1}}\cong \xt{1}{\cat{A}(T)}{M}{M_{1}} \xra{\psi_{*}}\xt{1}{\cat{A}(T)}{M}{M_{2}}\cong \xt{1}{\cat{A}(T)}{M}{N_{2}}\) is \(\phi_{*}\) which hence equals \(0\). If \(e:N_{1}\rightarrow L_{1}'\rightarrow M_{1}'\) is a \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull of \(N_{1}\), the connecting takes \(\phi\) to \(\phi_{*}e\in \xt{1}{\cat{A}(T)}{M_{1}'}{N_{2}}\), i.e.\ to \(0\), and so there exists a \(\delta:L_{1}'\rightarrow N_{2}\) which induces \(\phi\). Exactness in the lower row is analogous.
\end{proof}
\begin{prop}\label{prop.cof}
Let \(\cat{A}\) be a category fibred in abelian categories over \(\cat{C}\) and let \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\) be inclusion morphisms of categories fibred in additive categories\textup{.} Fix an object \(T\) in \(\cat{C}\)\textup{.} Assume \textup{BC1-2} for \((\cat{A},\cat{X})\) and \(T\), and \textup{AB1-3} for the triple of categories \((\cat{A}(T),\cat{X}(T),\cat{D}(T))\)\textup{.} Then\textup{:}
\begin{enumerate}
\item[(i)] \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)=\hat{\cat{X}}{}^{\textnormal{fl}}(T)\cap \cat{X}(T)^{\perp}\) and \(\cat{D}(T)=\cat{X}(T)\cap \hat{\cat{D}}{}^{\textnormal{fl}}(T)\)\textup{.}
\item[(ii)] \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) and \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) are closed under extensions\textup{.}
\item[(iii)] Exact sequences \(\dots\rightarrow N_{n}\xra{d_{n}} N_{n-1}\rightarrow \dots\) with objects \(N_{i}\) and kernels \(\ker d_{i}\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) remain exact after base change\textup{.}
\end{enumerate}
If in addition \textup{AB4}\textup{,} then\textup{:}
\begin{enumerate}
\item[(iv)] Epimorphisms in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) are admissible\textup{.}
\item[(v)] \(\Add \hat{\cat{X}}{}^{\textnormal{fl}}(T) =\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) and \(\Add \hat{\cat{D}}{}^{\textnormal{fl}}(T) = \hat{\cat{D}}{}^{\textnormal{fl}}(T)\)\textup{.}
\end{enumerate}
\end{prop}
\begin{proof}
For (i) the proofs of \cite[3.6-7]{aus/buc:89} works with the fl-s too by Theorem \ref{thm.cofapprox}. By (i) it's sufficient to prove (ii) for \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\). Let \(e: N_{1}\rightarrow N_{2}\rightarrow N_{3}\) be a short exact sequence in \(\hat{\cat{X}}(T)\).
If \(N_{1}\) and \(N_{3}\) are in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\), Theorem \ref{thm.cofapprox} and Lemma \ref{lem.forall} implies that \({}^{-}C^{*}(N_{1})\) and \({}^{-}C^{*}(N_{3})\) in Lemma \ref{lem.forall} are preserved as resolutions by base change. Together with BC2 this implies that base change of the short exact sequence of resolutions in Lemma \ref{lem.xres} (i) gives a short exact sequence of \(\cat{DX}\)-resolutions.
For (iii) the long exact sequence is broken into short exact sequences \(\xi_{i}\) with objects in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\). By Lemma \ref{lem.bs} it is sufficient to prove the claim for short exact sequences. By Lemma \ref{lem.xres} the short exact sequence of \(\cat{DX}\)-resolutions in Lemma \ref{lem.xres} (i) is preserved by base change. It follows that the short exact sequences \(\xi_{i}\) remain exact after base change.
For (iv); if \(N_{2}\) and \(N_{3}\) in \(e\) are in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) then \(N_{1}\) is in \(\hat{\cat{X}}(T)\) by AB4, see \cite[3.5]{aus/buc:89}. The argument proceeds as for extensions.
By (i) it's sufficient to prove (v) for \(\hat{\cat{X}}{}^{\textnormal{fl}}\). If \(N_{1}\coprod N_{2}\) is an object in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\), then \(N_{i}\) is in \(\hat{\cat{X}}(T)\) for \(i=1,2\) by \cite[3.4]{aus/buc:89}. By Lemma \ref{lem.forall} \({}^{-}C^{*}(N_{1})\coprod {}^{-}C^{*}(N_{2})\) is preserved by base change as resolution of \(N_{1}\coprod N_{2}\). It follows that the resolution \({}^{-}C^{*}(N_{i})\) is preserved by base change for \(i=1,2\).
\end{proof}
\begin{cor}\label{cor.Cbs}
Assume \textup{BC1-2} and \textup{AB1-3} as in \textup{Proposition \ref{prop.cof}.}
If \(N\) is in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) then any \(\cat{DX}\)-resolution \({}^{-}C^{*}(N)\twoheadrightarrow N\) and any \(\hat{\cat{D}}\cat{D}\)-coresolution \(N\rightarrowtail {}^{+}C^{*}(N)\) as in \textup{Section \ref{subsec.cplx}} is preserved by base change\textup{.}
\end{cor}
\begin{proof}
By Theorem \ref{thm.cofapprox} there exist a \(\cat{DX}\)-resolution and a \(\cat{\hat{D}D}\)-coresolution in \(\hat{\cat{X}}{}^{\textnormal{fl}}\).
The result follows from Proposition \ref{prop.cof} (iii) and Lemma \ref{lem.forall}.
\end{proof}
\begin{defn}\label{defn.cof}
Let \(\cat{A}\) be a category fibred in abelian categories over \(\cat{C}\) and let \(\cat{D}\subseteq\cat{A}\) be an inclusion morphism of a category fibred in additive categories. For \(T\) in \(\cat{C}\) let \(\check{\cat{D}}{}^{\textnormal{fl}}(T)\) denote the full subcategory of \(\cat{A}(T)\) of objects \(K\) with a finite coresolution \(K\rightarrow L^{*}\) with objects \(L^{i}\) in \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) for \(i\geq 0\).
\end{defn}
\begin{lem}\label{lem.cof2}
With these notions we have\textup{:}
\begin{enumerate}
\item[(i)] Epimorphisms in \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) are admissible if and only if\, \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)=\check{\cat{D}}{}^{\textnormal{fl}}(T)\)\textup{.}
\end{enumerate}
Assume \textup{BC1-2} and \textup{AB1-4} for \((\cat{A}(T),\cat{X}(T),\cat{D}(T))\)\textup{.} Then\textup{:}
\begin{enumerate}
\item[(ii)] \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)=\hat{\cat{D}}(T)\cap\check{\cat{D}}{}^{\textnormal{fl}}(T)\)\textup{.}
\item[(iii)] Epimorphisms in \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) are admissible if epimorphisms in \(\hat{\cat{D}}(T)\) are admissible\textup{.}
\end{enumerate}
\end{lem}
\begin{proof}
(i) is trivially true. In (ii) \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\subseteq\hat{\cat{D}}(T)\cap\check{\cat{D}}{}^{\textnormal{fl}}(T)\) is obvious. For the other inclusion, suppose \(K\) is an object in \(\hat{\cat{D}}(T)\cap\check{\cat{D}}{}^{\textnormal{fl}}(T)\setminus \hat{\cat{D}}{}^{\textnormal{fl}}(T)\) with a \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\)-coresolution \(K\rightarrow L^{*}\) of length \(n>0\). Since monomorphisms are admissible in \(\hat{\cat{D}}(T)=\hat{\cat{X}}(T)\cap \cat{X}(T)^{\perp}\), all \(K^{i}=\ker(L^{i}\rightarrow L^{i+1})\) are contained in \(\hat{\cat{D}}(T)\), and we can assume \(n=1\). But then \(K\) has to be in \(\hat{\cat{X}}{}^{\textnormal{fl}}\cap\hat{\cat{D}}=\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) by Proposition \ref{prop.cof}. Since (i) is true ``without the fl'' by \cite[4.1]{aus/buc:89}, (iii) follows immediately from (i) and (ii).
\end{proof}
\section{Cohen-Macaulay approximation of flat families}\label{sec.ft}
We define fibred categories of Cohen-Macaulay maps with flat modules and show that they allow Cohen-Macaulay approximation in the finite type case and the local, algebraic case.
\subsection{The finite type case}\label{subsec.ft}
Let \(h:S\rightarrow T\) be a ring homomorphism of noetherian rings. We say that \(h\) is a \emph{Cohen-Macaulay \textup{(}CM\textup{)} map} if it is of finite type, faithfully flat and all fibres are Cohen-Macaulay (cf.\ \cite[6.8.1]{EGAIV2}). In particular \(h\) is equidimensional (\cite[15.4.1]{EGAIV3}). B.\ Conrad has defined the \emph{dualising module} \(\omega_{h}\) for any CM \(h\), see \cite[Sec.\ 3.5]{con:00}.
Suppose \(h\) has pure relative dimension \(n\). For some \(N\geq n\) there is a surjective \(S\)-algebra map \(P\rightarrow T\) where \(P=S[t_{1},\dots,t_{N}]\). Let \(\omega_{P/S}:=\wedge^{N}\Omega_{P/S}\). Then there is an isomorphism \(\omega_{h}\cong\xt{N-n}{P}{T}{\omega_{P/S}}\) which is natural in the factorisation \(S\rightarrow P\rightarrow T\), see \cite[3.5.3-6]{con:00}. By (local) duality theory and Corollary \ref{cor.xtdef} \(\omega_{h}\) is \(S\)-flat
(or see \cite[Cor.\ 3.5.2]{con:00}).
If \(S\) is a field, we have that \(\omega_{h}\) is a canonical module of \(T\) as in Example \ref{ex.MCMapprox}, cf.\ \cite[3.3.7 and 16]{bru/her:98}.
Let \(\cat{CM}\) be the category with objects the CM maps and morphisms \((g,f):h_{1}\rightarrow h_{2}\) pairs of ring homomorphisms \(g:S_{1}\rightarrow S_{2}\) and \(f:T_{1}\rightarrow T_{2}\) such that \(h_{2}g=fh_{1}\) and such that the induced map \(f{\otimes} 1:T_{1}{\otimes} S_{2}\rightarrow T_{2}\) is an isomorphism:
\begin{equation*}
\xymatrix@C-0pt@R-8pt@H-30pt{
T_{1}\ar[r]^{f} & T_{2} & T_{1}{\otimes}_{S_{1}}S_{2} \ar[l]_(0.6){\simeq} \\
S_{1}\ar[u]^{h_{1}}\ar[r]^{g} & S_{2}\ar[u]_{h_{2}}
}
\end{equation*}
Let \(\cat{R}\) denote the category of noetherian rings. The forgetful functor \(p:\cat{CM}\rightarrow\cat{R}\); \((g,f)\mapsto g\), makes \(\cat{CM}\) fibred in groupoids over \(\cat{R}\). The essential part is that \(\cat{CM}\) should allow base change, i.e.\ given \(g:S_{1}\rightarrow S_{2}\) and \(h_{1}:S_{1}\rightarrow T_{1}\) as above there should exist a \(T_{2}\), an \(h_{2}:S_{2}\rightarrow T_{2}\) and an \(f\) such that \((g,f)\) is a morphism \(h_{1}\rightarrow h_{2}\) in \(\cat{CM}\). This follows from \cite[15.4.3]{EGAIV3}.
Let \(\cat{mod}\) be the category of pairs \((h:S\rightarrow T,N)\) with \(h\) in \(\cat{CM}\) and \(N\) a finite \(T\)-module. A morphism \((h_{1},N_{1})\rightarrow (h_{2},N_{2})\) is a morphism \((g,f):h_{1}\rightarrow h_{2}\) in \(\cat{CM}\) and a \(f\)-linear map \(\alpha:N_{1}\rightarrow N_{2}\). Then \(\alpha\) is cocartesian with respect to the forgetful functor \(F:\cat{mod}\rightarrow \cat{CM}\) if \(1{\otimes} \alpha: T_{2}{\otimes} N_{1}\rightarrow N_{2}\) is an isomorphism. It follows that \(\cat{mod}\) is fibred in abelian categories over \(\cat{CM}\).
Adding the property that \(N\) is \(S\)-flat gives the full subcategory \(\cat{mod}{}^{\textnormal{fl}}\). Moreover, let \(\cat{MCM}\) be the full subcategory of \(\cat{mod}{}^{\textnormal{fl}}\) where the fibre \(N_{s}=N{\otimes}_{S}k(s)\) is a maximal Cohen-Macaulay \(T_{s}\)-module for all \(s\in \Spec S\). The inclusions \(\cat{MCM}\subseteq\cat{mod}{}^{\textnormal{fl}}\subseteq\cat{mod}\) are inclusion morphisms of categories fibred in additive categories (FAds) over \(\cat{CM}\). For \(\cat{MCM}\) this follows from \cite[15.4.3]{EGAIV3}.
If \(h\) is a CM map let \(\cat{mod}_{h}\), \(\cat{MCM}_{h}\), \dots denote the fibre categories of \(\cat{mod}\), \(\cat{MCM}\), \dots over \(h\). An object in \(\cat{MCM}_{h}\) is called an (\(h\)-) family of maximal Cohen-Macaulay modules (or shorter; an MCM \(h\)-module).
Given a morphism \(h_{1}\rightarrow h_{2}\) in \(\cat{CM}\). By \cite[Thm.\ 3.6.1]{con:00} there is a natural isomorphism with base change \(T_{2}{\otimes}\omega_{h_{1}}\cong\omega_{h_{2}}\) which is compatible with localisation of \(T_{1}\) and is functorial with respect to composition \(h_{1}\rightarrow h_{2}\rightarrow h_{3}\). It follows that \(h\mapsto (h,\omega_{h})\) defines a morphism \(\omega:\cat{CM}\rightarrow\cat{MCM}\) of fibred categories over \(\cat{R}\) which is a section of the forgetful \(F:\cat{MCM}\rightarrow\cat{CM}\). Let \(\cat{D}\) be the full subcategory of \(\cat{MCM}\) over \(\cat{CM}\) with the objects \((h,D)\) where \(D\) is an object in \(\Add\{\omega_{h}\}\). The inclusion \(\cat{D}\subseteq\cat{MCM}\) is an inclusion of FAds over \(\cat{CM}\).
If \(\cat{U}\) denotes any of these FAds over \(\cat{CM}\), let \(\ul{\cat{U}}\) denote the quotient (`stable') category \(\cat{U}/\cat{D}\). With this notation we have the following.
\begin{thm}\label{thm.flatCMapprox}
The pair \((\cat{mod},\cat{MCM})\) over \(\cat{CM}\) satisfies \textup{BC1-2} and the triple of fibre categories \((\cat{mod}_{h},\cat{MCM}_{h},\cat{D}_{h})\) satisfies \textup{AB1-4} for all objects \(h\) in \(\cat{CM}\)\textup{.}
Moreover\textup{:}
\begin{enumerate}
\item[(i)] The fibred categories \(\widehat{\cat{MCM}}{}^{\textnormal{fl}}\) and \(\hat{\cat{D}}{}^{\textnormal{fl}}\) equals \(\cat{mod}{}^{\textnormal{fl}}\) and \(\hat{\cat{D}}\cap\cat{mod}{}^{\textnormal{fl}}\) respectively\textup{.}
\item[(ii)] For any object \((h,N)\) in \(\cat{mod}{}^{\textnormal{fl}}\), \(N\) admits an \(\cat{MCM}\)-approximation and a \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull which in particular are preserved by any base change\textup{.}
\item[(iii)] The \(\cat{MCM}\)-approximation induces a morphism of categories fibred in additive categories \(j^{!}:\underline{\cat{mod}}{}^{\textnormal{fl}}\rightarrow\ul{\cat{MCM}}\) which is a right adjoint to the full and faithful inclusion morphism \(j_{!}:\ul{\cat{MCM}}\rightarrow\underline{\cat{mod}}{}^{\textnormal{fl}}\)\textup{.}
\item[(iv)] The \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull induces a morphism of categories fibred in additive categories \(i^{*}:\underline{\cat{mod}}{}^{\textnormal{fl}}\rightarrow{\underline{\hat{\cat{D}}}{}^{\textnormal{fl}}}\) which is a left adjoint to the full and faithful inclusion morphism \(i_{*}:\underline{\hat{\cat{D}}}{}^{\textnormal{fl}}\rightarrow \underline{\cat{mod}}{}^{\textnormal{fl}}\)\textup{.}
\item[(v)] Together these maps give the following commutative diagram of short exact sequences of categories fibred in additive categories\textup{:}
\begin{equation*}
\xymatrix@C-0pt@R-8pt@H-30pt{
0\ar[r] & \,\underline{\hat{\cat{D}}}{}^{\textnormal{fl}} \ar[r]^(0.45){i_{*}}\ar[d]_(0.45){\id} & \underline{\cat{mod}}{}^{\textnormal{fl}}\ar[r]^{j^{!}}\ar@{=}[d] & \ul{\cat{MCM}}\ar[r] & 0 \\
0 & \,\underline{\hat{\cat{D}}}{}^{\textnormal{fl}}\ar[l] & \underline{\cat{mod}}{}^{\textnormal{fl}}\ar[l]_{i^{*}} & \ul{\cat{MCM}}\ar[l]_(0.45){j_{!}}\ar[u]_{\id} & 0 \ar[l]
}
\end{equation*}
\end{enumerate}
\end{thm}
\begin{proof}
Since base change is given by the tensor product BC1 and BC2 follows. In particular, a short exact sequence \(e:N_{1}\rightarrow N_{2}\rightarrow N_{3}\) in \(\cat{mod}_{h}\) with \(N_{3}\) and either \(N_{1}\) or \(N_{2}\) in \(\cat{MCM}_{h}\) gives short exact sequences of MCMs after base change to each fibre \(T_{s}\), i.e.\ AB1 and AB4.
For AB2, suppose \(M\) is in \(\cat{MCM}_{h}\). Then \(M^{\vee}=\hm{}{T}{M}{\omega}\) is in \(\cat{MCM}_{h}\) too by Corollary \ref{cor.xtdef}. Since \(M^{\vee}\) is finite there is a short exact sequence
\(M^{\vee}\leftarrow T^{r}\leftarrow M_{1}\). By AB1 \(M_{1}\) is in \(\cat{MCM}_{h}\). Applying \(\hm{}{T}{-}{\omega_{h}}\) gives the desired short exact sequence since Corollary \ref{cor.xtdef} implies that \(\xt{1}{T}{M^{\vee}}{\omega_{h}}=0\) and that the natural map \(M\rightarrow M^{\vee\vee}\) is an isomorphism. AB3 also follows from Corollary \ref{cor.xtdef}.
Any \(N\) in \(\widehat{\cat{MCM}}{}^{\textnormal{fl}}_{h}\) has by definition a finite \(\cat{MCM}\)-resolution (say of length \(n\)) preserved by base change. Since objects in \(\cat{MCM}\) are \(S\)-flat it follows by induction on \(n\) that \(\tor{S}{1}{N}{k(s)}=0\) for all \(s\in\Spec S\). Hence \(N\) is \(S\)-flat. Conversely, if \(N\) is in \(\cat{mod}{}^{\textnormal{fl}}_{h}\) it follows that a sufficiently high syzygy of \(N\) is in \(\cat{MCM}_{h}\), i.e.\ \(N\) is in \(\widehat{\cat{MCM}}{}^{\textnormal{fl}}_{h}\).
With Proposition \ref{prop.cof} this gives \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}=\cat{mod}{}^{\textnormal{fl}}_{h}\cap\cat{MCM}_{h}^{\perp}\). By induction on the length of the resolution \(\hat{\cat{D}}_{h}\subseteq \cat{MCM}_{h}^{\perp}\) and so \(\cat{mod}{}^{\textnormal{fl}}_{h}\cap\hat{\cat{D}}_{h}\subseteq\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\). The opposite inclusion is clear by the first part of (i). Now (ii)-(v) follows directly from Theorem \ref{thm.cofmain}.
\end{proof}
\begin{cor}\label{cor.flatCMapprox}
Let \(h:S\rightarrow T\) be an object in \(\cat{CM}\)\textup{.} Then\textup{:}
\begin{enumerate}
\item[(i)] \(\cat{D}_{h}=\cat{MCM}_{h}^{\perp}\cap\cat{MCM}_{h}\quad\text{and}\quad \hat{\cat{D}}{}^{\textnormal{fl}}_{h}=\cat{MCM}_{h}^{\perp}\cap\cat{mod}{}^{\textnormal{fl}}_{h}\)
\item[(ii)] The kernel of a surjective map in \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\) is contained in \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\)\textup{.}
\end{enumerate}
\end{cor}
\begin{proof}
(ii): Note that if \(N_{1}\rightarrow N_{2}\rightarrow N_{3}\) is a short exact sequence with \(N_{2}\) and \(N_{3}\) in \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\), in particular \(S\)-flat, then \(N_{1}\) has to be \(S\)-flat too. To show that \(N_{1}\) is in \(\hat{\cat{D}}_{h}\) we use the criterion in \cite[4.6]{aus/buc:89} to show that \(\check{\cat{D}}_{h}=\hat{\cat{D}}_{h}\):
Suppose that \(M\) is in \(\cat{MCM}_{h}\). Assume that \(M\) satisfies \(\injdim{\cat{MCM}_{h}}{M}= n<\infty\) which by \cite[4.3]{aus/buc:89} is equivalent to the existence of a coresolution of \(M\) of length \(n\) in \(\cat{D}_{h}\). The fibre at any \(s\in\Spec S\) gives a \(\cat{D}_{T_{s}}\)-coresolution of the MCM \(T_{s}\)-module \(M_{s}\). Since \(\check{\cat{D}}_{T_{s}}=\hat{\cat{D}}_{T_{s}}\) (\cite[6.3]{aus/buc:89}) it follows by \cite[4.6]{aus/buc:89} that \(M_{s}\) is contained in \(\cat{D}_{T_{s}}\). Since \(\cat{D}_{T_{s}}=\cat{MCM}_{T_{s}}\cap\cat{MCM}_{T_{s}}^{\perp}\) by \cite[3.7]{aus/buc:89} it follows from Corollary \ref{cor.xtdef} that \(\injdim{\cat{MCM}_{h}}{M}=0\) and so \(M\) is in \(\cat{MCM}_{h}\cap\cat{MCM}_{h}^{\perp}\). But by Theorem \ref{thm.flatCMapprox} we can invoke \cite[3.7]{aus/buc:89} again which gives \(\cat{MCM}_{h}\cap\cat{MCM}_{h}^{\perp}=\cat{D}_{h}\) so \(M\) is in \(\cat{D}_{h}\). By \cite[4.6(d)]{aus/buc:89} \(\check{\cat{D}}_{h}=\hat{\cat{D}}_{h}\) follows and \(N_{1}\) is in \(\hat{\cat{D}}_{h}\).
\end{proof}
Let \(\cat{P}\) denote the fibred subcategory of \(\cat{MCM}\) over \(\cat{CM}\) of pairs \((h,P)\) with \(h:S\rightarrow T\) in \(\cat{CM}\) and \(P\) a finite projective \(T\)-module. Let \(\hat{\cat{P}}{}^{\textnormal{fl}}\) denote the full subcategory of \(\cat{mod}{}^{\textnormal{fl}}\) of pairs \((h,Q)\) such that \(Q\) has a finite projective dimension. The inclusions of categories fibred in additive categories \(\cat{P}\subseteq\hat{\cat{P}}{}^{\textnormal{fl}}\subseteq\cat{mod}{}^{\textnormal{fl}}\) are closed under extensions over \(\cat{CM}\).
\begin{lem}\label{lem.proj}
There is an exact equivalence \(\hat{\cat{D}}{}^{\textnormal{fl}}\simeq\hat{\cat{P}}{}^{\textnormal{fl}}\) of categories fibred in additive categories defined by the functor \((h,L)\mapsto(h,\hm{}{T}{\omega_{h}}{L})\) with a quasi-inverse \((h,Q)\mapsto (h,Q{\otimes}_{T}\omega_{h})\)\textup{.} It induces an equivalence of fibred quotient categories \(\hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\simeq\hat{\cat{P}}{}^{\textnormal{fl}}/\cat{P}\)\textup{.}
\end{lem}
\begin{proof}
By base change we reduce to the absolute case where the equivalence is well known, cf.\ \cite[I 4.10.16]{has:00}. The functor \(\hm{}{-}{\omega_{-}}{-}\) takes a map \(\phi:L_{1}\rightarrow L_{2}\) over \(h_{1}\rightarrow h_{2}\) to the map \(\hm{}{T_{1}}{\omega_{h_{1}}}{L_{1}}\rightarrow\hm{}{T_{2}}{\omega_{h_{2}}}{L_{2}}\); \(\alpha\mapsto(\phi\alpha)^{\#}\), the unique map which pulls back to \(\phi\alpha\) by the cocartesian map \(\omega_{h_{1}}\rightarrow\omega_{h_{2}}\). It is a well defined functor since the dualising module is functorial.
If \(L\) is in \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\) then \(\xt{i}{T_{s}}{\omega_{T_{s}}}{L_{s}}=0\) for all \(i\neq 0\) and \(s\in \Spec S\). By Corollary \ref{cor.xtdef} the functor is exact and \(\hm{}{T}{\omega_{h}}{L}\) is \(S\)-flat. In particular \(\nd{}{T}{\omega_{h}}\cong T\). If \(D\) is in \(\cat{D}_{h}\) then \(\hm{}{T}{\omega_{h}}{D}\) is projective as a direct summand of a free module. If \(D^{*}\twoheadrightarrow L\) is a finite \(\cat{D}_{h}\)-resolution of \(L\), then \(\hm{}{T}{\omega_{h}}{D^{*}}\) gives a projective resolution of \(\hm{}{T}{\omega_{h}}{L}\) since \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\subseteq\cat{MCM}_{h}^{\perp}\) (Corollary \ref{cor.flatCMapprox}). The natural map \(\ev:\hm{}{T}{\omega_{h}}{L}{\otimes}\omega_{h}\rightarrow L\) commutes with base change to the fibres where it is an isomorphism (\cite[9.6.5]{bru/her:98}) and Nakayama's lemma and \(S\)-flatness implies that \(\ev\) is an isomorphism too. Let \(P^{*}\rightarrow Q\) be a \(\cat{P}\)-resolution of \(Q\) in \(\hat{\cat{P}}{}^{\textnormal{fl}}_{h}\) of length \(n\). Define covariant functors \(G^{i}:\cat{mod}_{S}\rightarrow\cat{mod}_{T}\) by \(G^{i}(V):=\cH^{i-n}(P^{*}{\otimes}_{T}\omega_{h}{\otimes}_{S}V)\). Since \(P^{*}{\otimes}_{T}\omega_{h}\) is \(S\)-flat, \(\{G^{i}\}\) defines a cohomological \(\delta\)-functor. Since \(P^{*}{\otimes} k(s)\) gives a resolution of \(Q{\otimes}_{S} k(s)\) for all \(s\in\Spec S\), \cite[9.6.5]{bru/her:98} and Proposition \ref{prop.nakayama} implies that \(G^{n}(S)=Q{\otimes}_{T}\omega_{h}\) is \(S\)-flat and \(P^{*}{\otimes}_{T}\omega_{h}\twoheadrightarrow Q{\otimes}_{T}\omega_{h}\) is a \(\cat{D}\)-resolution. Moreover, the natural map \(Q\rightarrow\hm{}{T}{\omega_{h}}{Q{\otimes}\omega_{h}}\) is an isomorphism by Nakayama's lemma again.
\end{proof}
\begin{ex}\label{ex.extiso}
Assume \(A\) is a Cohen-Macaulay ring with a canonical module \(\omega_{A}\).
Given \(L_{i}\in\hat{\cat{D}}_{A}\) and put \(Q_{i}=\hm{}{A}{\omega_{A}}{L_{i}}\) for \(i=1,2\). If \(I\) is an injective resolution of \(L_{2}\) and \(P\) is a projective resolution of \(Q_{1}\) then both spectral sequences of \(\hm{}{A}{P}{\hm{}{A}{\omega_{A}}{I}}\) collapse at page \(2\) (use \cite[I 4.10.19]{has:00}) to give canonical isomorphisms
\begin{equation}
\xt{*}{A}{L_{1}}{L_{2}}\cong\xt{*}{A}{Q_{1}}{Q_{2}}
\end{equation}
\end{ex}
\begin{rem}\label{rem.Buch}
In his unpublished manuscript Buchweitz gave a construction of Cohen-Macaulay approximation for finite \(A\)-modules if \(A\) is a not necessarily commutative Gorenstein ring, see \cite{buc:86}. The \(\cat{MCM}\)-approximation and the \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull in Theorem \ref{thm.flatCMapprox} can be given essentially by the same construction. Let \(N\) be a finite \(T\)-module which is \(S\)-flat where \(h:S\rightarrow T\) is a finite type Cohen-Macaulay map. Let \(P=P(N)\rightarrow N\) be a projective resolution of \(N\) (i.e.\ by finite projective \(T\)-modules). Then \(P^{\vee}=\hm{}{T}{P}{\omega_{h}}\) is a bounded below complex with bounded cohomology because \(\xt{i}{T}{N}{\omega_{h}}=0\) for \(i\) greater than the relative dimension \(d\) of \(h\) by Corollary \ref{cor.xtdef} (\(\Injdim \omega_{T_{s}}=\dim T_{s}\leq d\)). Then we can choose a projective resolution \(f:P(P^{\vee})\rightarrow P^{\vee}\) of \(P^{\vee}\) which is bounded above. Let \(C=C(f)\) be the mapping cone of \(f\). The modules in \(C\) are direct sums of projective modules and modules in \(\cat{D}_{h}\) and the (co)kernels in the acyclic \(C\) are modules in \(\cat{MCM}_{h}\). By Corollary \ref{cor.xtdef} it follows that \(C^{\vee}\) is acyclic too. There is a composition of natural maps \(P\cong P^{\vee\vee}\rightarrow P(P^{\vee})^{\vee}:=G\) which hence is a quasi-isomorphism. But then \(G\) is just the representing complex of \(N\) and the \(\cat{MCM}\)-approximation and the \(\hat{\cat{D}}\)-hull is obtained as in Section \ref{subsec.cplx}. In the the case of coherent rings with a cotilting module (a concept introduced by Y.\ Miyashita, see \cite[p.\ 142]{miy:86}) J.-i.\ Miyachi implicitly gave the same construction in \cite[3.2]{miy:96}.
Suppose \(h\) has pure relative dimension and \(N_{s}\) is a non-zero Cohen-Macaulay \(T_{s}\)-module with \(n=\dim T_{s}-\dim N_{s}\) constant for all \(s\in\Spec S\), i.e.\ \(N\) is a family of Cohen-Macaulay modules of codepth \(n\). Put \(N^{\vee}:=\xt{n}{T}{N}{\omega_{h}}\). Then \(P^{\vee}=P(N)^{\vee}\) is quasi-isomorphic to \(\cH^{n}(P^{\vee})=N^{\vee}\) by duality theory and Corollary \ref{cor.xtdef}. So if \((F,d)\rightarrow N^{\vee}\) is a projective resolution of \(N^{\vee}\), the representing complex is given as \(G=F^{\vee}\). If \(\syz{T}{i}N^{\vee}\) denotes the \(i^{\text{th}}\) syzygy module \(\im d_{i}\) and \(d_{i}^{\vee}\) denotes \(\hm{}{T}{d_{i}}{\omega_{h}}\) then the commutative diagram \eqref{eq.pullpush} with short exact sequences is given as
\begin{equation*}
\xymatrix@C-0pt@R-9pt@H-30pt{
\im (d_{n}^{\vee}) \ar[r]\ar@{=}[d] & \hm{}{T}{\syz{T}{n}N^{\vee}}{\omega_{h}} \ar[r]\ar[d]\ar@{}[dr]|{\Box} & N^{\vee\vee} \ar[d] \\
\im (d_{n}^{\vee}) \ar[r] & \hm{}{T}{F_{n}}{\omega_{h}} \ar[r]\ar[d] & \coker(d_{n}^{\vee}) \ar[d] \\
& \hm{}{T}{\syz{T}{n{+}1}N^{\vee}}{\omega_{h}} \ar@{=}[r] & \hm{}{T}{\syz{T}{n{+}1}N^{\vee}}{\omega_{h}}
}
\end{equation*}
with the \(\cat{MCM}_{h}\)-approximation and \(\hat{\cat{D}}_{h}\)-hull of \(N\) given by the upper horizontal and right vertical sequence, respectively, since \(N^{\vee\vee}\cong N\) by duality theory and Corollary \ref{cor.xtdef}.
Let \(C\) denote the mapping cone of a comparison map \(P(N)\rightarrow G\). The homology of the truncated short exact sequence of complexes \(G\rightarrow C\rightarrow P(N)[-1]\) gives the \(\cat{MCM}_{h}\)-approximation \(\coker d^{G}_{i+2}\rightarrow\coker d^{C}_{i+2}\rightarrow\syz{T}{i}N\) for \(i\geq 0\) (with \(i=0\) being the upper horizontal sequence in the diagram). In the absolute case (with \(N\) Cohen-Macaulay) this latter construction of the MCM approximation of \(\syz{}{i}N\) was given by J.\ Herzog and A.\ Martsinkovsky in \cite[1.1]{her/mar:93}.
\end{rem}
\subsection{Local cases}\label{subsec.loc}
We formulate local variants of the approximation theorem.
Fix a field \(k\). Let \(\cat{H}\) denote the category of local, henselian, noetherian rings \(S\) with residue field \(S/\fr{m}_{S}\cong k\) and with local ring homomorphisms. A map \(h:S\rightarrow T\) in \(\cat{H}\) is \emph{algebraic} (or \(T\) is an algebraic \(S\)-algebra) if there is a finite type \(S\)-algebra \(\tilde{T}\) and a maximal ideal \(\fr{m}\) in \(\tilde{T}\) with \(\tilde{T}/\fr{m}\cong k\) such that \(h\) is given by henselisation of \(\tilde{T}\) in \(\fr{m}\). Fix an algebraic \(k\)-algebra \(A\) which is supposed to be Cohen-Macaulay. Objects in the category \(\cat{hCM}\) are algebraic and flat \(S\)-algebras \(T\) with \(T{\otimes}_{S}k\cong A\). A morphism \(h_{1}\rightarrow h_{2}\) is a pair of commuting maps \(f:T_{1}\rightarrow T_{2}\) and \(g:S_{1}\rightarrow S_{2}\) in \(\cat{H}\) as for the finite type case, giving a cocartesian square. Fibre sum exists in \(\cat{H}\) and is given by the henselisation of the tensor product \(T=T_{1}{\otimes}_{S_{1}}S_{2}\) in the maximal ideal \(\fr{m}_{T_{1}}T+\fr{m}_{S_{2}}T\). We denote it by \(T_{1}{\tilde{\otimes}}_{S_{1}}S_{2}\). It has the same closed fibre as \(S_{1}\rightarrow T_{1}\) and it follows that the obvious functor \(\cat{hCM}\rightarrow\cat{H}\) is fibred in groupoids. The objects in \(\cat{hCM}\) will be called henselian Cohen-Macaulay (hCM) maps.
If \(\tilde{h}:\tilde{S}\rightarrow\tilde{T}\) is a CM map and \(t\) a \(k\)-point in \(\Spec \tilde{T}\) with image \(s\) in \(\Spec \tilde{S}\) then we get a map of the local rings \(S=\tilde{S}^{\text{h}}\rightarrow \tilde{T}^{\text{h}}=T\) for the \'etale topology at \(t\) and \(s\) given by henselisation. This is a hCM map. Conversely, every hCM map is obtained this way which follows from \cite[18.6.6 and 18.6.10]{EGAIV4} and \cite[15.4.3 and 12.1.1]{EGAIV3}. We will call an \(\tilde{h}:\tilde{S}\rightarrow\tilde{T}\) in \(\cat{CM}\) which induces \(h:S\rightarrow T\) in \(\cat{hCM}\) a (finite type) representative of \(h\). The dualising module \(\omega_{\tilde{h}}\) induces an \(S\)-flat finite \(T\)-module \(\omega_{h}\) called the dualising module for \(h\). Two representatives of \(h\) factor through a common \'etale neighbourhood contained in \(\cat{CM}\) and since the dualising module is functorial for \(\cat{CM}\) the dualising module \(\omega_{h}\) is functorial too.
Let \(\cat{mod}\) denote the category of pairs \((h:S\rightarrow T,\mc{N})\) with \(h\) in \(\cat{hCM}\) and \(\mc{N}\) a finite \(T\)-module. Morphisms are defined as for the finite type case and the forgetful functor \(\cat{mod}\rightarrow\cat{hCM}\) makes \(\cat{mod}\) fibred in abelian categories over \(\cat{hCM}\). Let \(\cat{mod}{}^{\textnormal{fl}}\) denote the full subcategory of pairs \((h,\mc{N})\) in \(\cat{mod}\) with \(\mc{N}\) \(S\)-flat and let \(\cat{MCM}\) denote the full subcategory of \(\cat{mod}{}^{\textnormal{fl}}\) where the closed fibre \(\mc{N}{\otimes}_{S}k\) is a maximal Cohen-Macaulay \(A\)-module. Let \(\cat{D}\) denote the full subcategory of \(\cat{MCM}\) of objects \((h,D)\) with \(D\) in \(\Add\{\omega_{h}\}\).
All three are FAd subcategories of \(\cat{mod}\) over \(\cat{hCM}\). Any finite \(T\)-module \(\mc{N}\) has finite presentation hence it is induced from a finite module over a representative of \(T\). If \(\mc{N}\) is contained in one of the subcategories the representative can be assumed to belong to the corresponding finite type subcategory ({\sl loc.\ sit.}). Similarly all maps in the various fibred categories over \(\cat{H}\) are induced from maps in the corresponding fibred categories over \(\cat{R}\).
Let \(\cat{L}\) (\(\cat{C}\)) denote the category of (complete) local, noetherian rings with residue field \(k\) and local ring homomorphisms. Let \(\cat{lCM}\) (\(\cat{cCM}\)) denote the category of local Cohen-Macaulay (lCM) maps (respectively complete Cohen-Macaulay or cCM maps) defined analogously as above with (completion of) \emph{essentially of finite type} replacing algebraic. Similar arguing as above makes \(\cat{lCM}\) (\(\cat{cCM}\)) fibred in groupoids over \(\cat{L}\) (\(\cat{C}\)). The definitions of the module categories apply in the local and the complete case too and we use the same notation in all three cases. Again objects and maps are induced from the finite type case.
Either arguing with representatives or applying the proofs for Theorem \ref{thm.flatCMapprox} and Corollary \ref{cor.flatCMapprox} (with only minor adjustments) we obtain the following.
\begin{cor}\label{cor.locCMapprox}
Let \(\cat{xCM}\) denote either \(\cat{hCM}\)\textup{,} \(\cat{lCM}\) or \(\cat{cCM}\)\textup{.} The pair \((\cat{mod},\cat{MCM})\) of fibred categories over \(\cat{xCM}\) satisfies \textup{BC1-2} and the triple of fibre categories \((\cat{mod}_{h},\cat{MCM}_{h},\cat{D}_{h})\) satisfies \textup{AB1-4} for all objects \(h\) in \(\cat{xCM}\)\textup{.}
Moreover\textup{;} the statements \textup{(i-v)} in \textup{Theorem \ref{thm.flatCMapprox}} and \textup{(i-ii) in Corollary \ref{cor.flatCMapprox}} are valid over \(\cat{xCM}\) too\textup{.}
\end{cor}
\section{Minimal approximations and semi-continuity of invariants}
We show that the Cohen-Macaulay approximation and the \(\hat{\cat{D}}\)-hull in Corollary \ref{cor.locCMapprox} can be chosen to be minimal. Upper semi-continuous invariants on \(\cat{MCM}_{A}\) or \(\cat{FID}_{A}\) extends to upper semi-continuous invariants on \(\cat{mod}_{A}\). An example is given by the \(\omega_{A}\)-ranks in the representing \(\cat{D}\)-complex.
\begin{lem}\label{lem.Aapprox}
Let \(S\rightarrow T\) be a homomorphism of noetherian rings and \(\fr{a}\) an ideal in \(S\) such that \(I=\fr{a}T\) is contained in the Jacobson radical of \(T\)\textup{.} Let \(M\) and \(N\) be finite \(T\)-modules\textup{.} Let \(T_{n}=T/I^{n+1}\)\textup{,} \(M_{n}=T_{n}{\otimes} M\) and \(N_{n}=T_{n}{\otimes} N\)\textup{.} Suppose there exists a tower of surjections \(\{\phi_{n}:M_{n}\rightarrow N_{n}\}\)\textup{.} Fix any non-negative integer \(n_{0}\)\textup{.} Then there exists a \(T\)-linear surjection \(\psi:M\rightarrow N\) such that \(T_{n_{0}}{\otimes}\psi=\phi_{n_{0}}\)\textup{.} If the \(\phi_{n}\) are isomorphisms and \(N\) is \(S\)-flat then \(\psi\) is an isomorphism\textup{.}
\end{lem}
\begin{proof}
Let \(\hat{T}=\varprojlim T_{n}\), \(\hat{M}=\varprojlim M_{n}\) and \(\hat{N}=\varprojlim N_{n}\). We have
\begin{equation}
\varprojlim\hm{}{T_{n}}{M_{n}}{N_{n}}\cong\hm{}{\hat{T}}{\hat{M}}{\hat{N}}\cong\hat{T}{\otimes}\hm{}{T}{M}{N}.
\end{equation}
Hence \(\varprojlim \phi_{n}=\Sigma r^{(i)}{\otimes}\beta^{(i)}\) with \(r^{(i)}\in \hat{T}\) and \(\beta^{(i)}\in\hm{}{T}{M}{N}\). Let \(r^{(i)}_{n_{0}}\) be the image of \(r^{(i)}\) under \(\hat{T}\rightarrow T_{n_{0}}\) and choose liftings \(t^{(i)}\) in \(T\) of \(r^{(i)}_{n_{0}}\). Put \(\psi=\Sigma t^{(i)}\beta^{(i)}\). Then \(T_{n_{0}}{\otimes}\psi=\phi_{n_{0}}\). Since \(T_{n_{0}}{\otimes}\coker\psi=\coker\phi_{n_{0}}=0\), Nakayama's lemma implies \(\coker\psi=0\). Since \(T_{n_{0}}{\otimes}_{T}N\) equals \(N{\otimes}_{S}S/\fr{a}^{n_{0}+1}\), \(S\)-flatness of \(N\) implies \(T_{n_{0}}{\otimes}\ker\psi=\ker\phi_{n_{0}}\) and if \(\ker\phi_{n_{0}}=0\) then Nakayama's lemma implies \(\ker\psi=0\).
\end{proof}
\begin{prop}\label{prop.minapprox}
Let \(\cat{xCM}\) denote either \(\cat{hCM}\)\textup{,} \(\cat{lCM}\) or \(\cat{cCM}\)\textup{,} let \(h:S\rightarrow T\) be an object in \(\cat{xCM}\) and let \(\xi:\mc{L}\rightarrow\mc{M}\xra{\pi}\mc{N}\) be an \(\cat{MCM}_{h}\)-approximation of \(\mc{N}\)\textup{.} Then the following statements are equivalent\textup{.}
\begin{enumerate}
\item[(i)] The sequence \(\xi\) is a right minimal \(\cat{MCM}_{h}\)-approximation\textup{.}
\item[(ii)] There are no surjections \(\mc{M}\rightarrow\omega_{h}\) which induces a surjection \(\mc{L}\rightarrow\omega_{h}\)\textup{.}
\item[(iii)] There are no common \(\omega_{h}\)-summand in \(\mc{L}\rightarrow\mc{M}\)\textup{.}
\item[(iv)] The closed fibre \(\xi{\otimes}_{S}k\) is a right minimal \(\cat{MCM}_{A}\)-approximation\textup{.}
\item[(v)] The completion \((\xi{\otimes}_{S}k)\hat{\,}\) of the closed fibre is a right minimal \(\cat{MCM}_{\hat{A}}\)-approximation\textup{.}
\end{enumerate}
The analogous statements \textnormal{(i')-(v')} for a \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\)-hull \(\xi':\mc{N}\rightarrow\mc{L}'\rightarrow\mc{M}'\) are equivalent\textup{.} In particular\textup{:}
\begin{enumerate}
\item[(i')] The sequence \(\xi'\) is a left minimal \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\)-hull\textup{.}
\item[(ii')] There are no surjections \(\mc{M}'\rightarrow\omega_{h}\) which induces a surjection \(\mc{L}'\rightarrow\omega_{h}\)\textup{.}
\end{enumerate}
\end{prop}
\begin{proof}
Suppose there is a surjection \(\mc{M}\rightarrow \omega_{h}\) such that the composition \(\mc{L}\rightarrow \omega_{h}\) is surjective too. Then the kernels of these maps give a new \(\cat{MCM}\)-approximation of \(\mc{N}\) by Corollary \ref{cor.locCMapprox}. Since a surjection \(\mc{L}\rightarrow\omega_{h}\) has to split by Corollary \ref{cor.locCMapprox}, \(\omega_{h}\) is a common summand in \(\mc{L}\rightarrow\mc{M}\) and \(\pi\) cannot be right minimal.
Let the closed fibre \(\xi{\otimes}_{S}k\) of the sequence \(\xi\) be denoted by \(L\rightarrow M\xra{p} N\).
Assume there is a non-surjective endomorphism \(\theta:\mc{M}\rightarrow\mc{M}\) with \(\pi\theta=\pi\). Then \(\theta_{0}=\theta{\otimes}_{S}k\) gives a non-surjective endomorphism of \(M\) with \(p\theta_{0}=p\). It follows that the completion \(\hat{L}\rightarrow\hat{M}\rightarrow\hat{N}\) is not a right minimal Cohen-Macaulay approximation of \(\hat{N}\). By \cite[1.12.8]{has:00} there is a common \(\omega_{\hat{A}}\)-summand in \(\hat{L}\rightarrow\hat{M}\). Let \(\phi:\hat{M}\rightarrow\omega_{\hat{A}}\) denote the projection. By Lemma \ref{lem.Aapprox} there exists a surjection \(\psi:M\rightarrow\omega_{A}\). The induced map \(L\rightarrow\omega_{A}\) is surjective too. The map \(\psi\) lifts to a surjection \(\mc{M}\rightarrow\omega_{h}\) (with \(\mc{L}\rightarrow\omega_{h}\) surjective) since the canonical map \(\hm{}{T}{\mc{M}}{\omega_{h}}\rightarrow\hm{}{A}{M}{\omega_{A}}\) is surjective by Corollary \ref{cor.xtdef}.
The \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-case is analogous.
\end{proof}
\begin{cor}\label{cor.minapprox}
Let \(\cat{xCM}\) denote either \(\cat{hCM}\)\textup{,} \(\cat{lCM}\) or \(\cat{cCM}\)\textup{.}
For any object \((h,\mc{N})\) in the fibred category \(\cat{mod}{}^{\textnormal{fl}}\) over \(\cat{xCM}\), \(\mc{N}\) admits a right minimal \(\cat{MCM}\)-approximation and a left minimal \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull which remain minimal after base change and which in particular are unique up to non-canonical isomorphism\textup{.}
\end{cor}
\begin{proof}
The existence of a right minimal \(\xi\) follows immediately from criterion (iii) in Proposition \ref{prop.minapprox}. Moreover \(\xi\) is right minimal if and only if the closed fibre \(\xi{\otimes}_{S}k\) is right minimal. Since any base change \(\xi_{1}\) of \(\xi\) has the same closed fibre as \(\xi\), \(\xi_{1}\) is right minimal if \(\xi\) is.
\end{proof}
\begin{rem}
In \cite[2.3]{has/shi:97} M.\ Hashimoto and A.\ Shida gave essentially the analog of Proposition \ref{prop.minapprox} in the absolute case of a Zariski local Cohen-Macaulay ring with a canonical module. They attributed the complete case to Y.\ Yoshino. Note Miyachi's proof in the cotilting semi-perfect case, see \cite[3.4]{miy:96} (cf.\ \cite[1.12.8]{has:00}). In \cite[3.1]{sim/str:02} A.-M.\ Simon and J.\ R.\ Strooker give a short independent proof. The proof of Proposition \ref{prop.minapprox} also works in the Zariski local case in full generality and is different from these (but depends on the complete case).
\end{rem}
Since minimal choices of \(\cat{MCM}_{A}\)-approximations and \(\cat{\hat{D}}_{A}\)-hulls exist and are unique up to isomorphism any invariant defined for MCM modules or for FID modules is extended to all finite \(A\)-modules. Upper semi-continuity of the invariants is also extended as we explain now. First some notation.
Let \(h:S\rightarrow T\) be ring homomorphism and \(\mc{N}\) a finite \(T\)-module. If \(t\in\Spec T\) has image \(s\in\Spec S\), let \(\mc{N}_{\fr{p}_{t}}\) denote the localisation of \(\mc{N}\) at the prime ideal \(t\), let \(h_{\fr{p}_{t}}:S_{\fr{p}_{s}}\rightarrow T_{\fr{p}_{t}}\) denote the ring homomorphism obtained by localising, and put \(\mc{N}(t)=\mc{N}_{\fr{p}_{t}}{\otimes}_{S_{\fr{p}_{s}}}k(s)\) which is a \(T(t)\)-module; indeed \(\mc{N}(t)\cong T(t){\otimes}_{T_{\fr{p}_{t}}}\mc{N}_{\fr{p}_{t}}\).
If \(h\) is a finite type Cohen-Macaulay map, \(h_{\fr{p}_{t}}\) is in \(\cat{lCM}\).
Suppose \(\mu\) is an invariant on \(\cat{MCM}_{A}\) where \(A\) is a Cohen-Macaulay local ring with canonical module. Let \({}_{\cat{MCM}}\mu\) denote the induced invariant on \(\cat{mod}_{A}\) defined by \({}_{\cat{MCM}}\mu(N) = \mu(M)\) where \(L\rightarrow M\rightarrow N\) is the minimal Cohen-Macaulay approximation of \(N\). Similarly \({}_{\cat{FID}}\mu(N)= \mu(L)\) for an invariant \(\mu\) defined on \(\cat{FID}_{A}\). Use the minimal hull \(N\rightarrow L'\rightarrow M'\) to define \({}_{\cat{FID}}\mu'\) and \({}_{\cat{MCM}}\mu'\).
The following theorem is a major application of what we have done so far.
\begin{thm}\label{thm.semicont}
Let \(\mu\) be an additive non-negative numerical invariant defined for maximal Cohen-Macaulay modules or for finite modules of finite injective dimension on a Cohen-Maculay local ring with canonical module\textup{.} Assume \(\mu\) is upper semi-continuous for finite type flat families \((h:S\rightarrow T,\mc{M})\) in \(\cat{MCM}\) \textup{(}or in \(\hat{\cat{D}}\)\textup{).} Then the induced invariants \({}_{\cat{MCM}}\mu\) and \({}_{\cat{MCM}}\mu'\) \textup{(}or \({}_{\cat{FID}}\mu\) and \({}_{\cat{FID}}\mu'\)\textup{)} are upper semi-continuous in finite type flat families \((h,\mc{N})\) in \(\cat{mod}{}^{\textnormal{fl}}\)\textup{.}
\end{thm}
\begin{proof}
Given \((h:S\rightarrow T,\mc{N})\) in \(\cat{mod}{}^{\textnormal{fl}}\) and \(t\in\Spec T\). By Corollary \ref{cor.minapprox} there exists open affines \(U=\Spec S_{1}\subseteq \Spec S\), \(V=\Spec T_{1}\subseteq \Spec T\) with \(t\in V\), \(h_{1}:S_{1}\rightarrow T_{1}\) induced from \(h\), and a \(\cat{MCM}_{h_{1}}\)-approximation \(\xi: \mc{L}\rightarrow\mc{M}\rightarrow\mc{N}_{\vert V}\) such that the localisation \(\xi_{\fr{p}_{t}}\) is minimal. By Corollary \ref{cor.minapprox} \(\xi(t)\) is minimal too. Put \(n=\mu(\mc{M}(t))\). Since \(\mu\) is upper semi-continuous there is an open \(V_{n}\subseteq V\) containing \(t\) such that \(\mu(\mc{M}(t'))\leq n\) for all \(t'\in V_{n}\). If \(L\rightarrow M\rightarrow \mc{N}(t')\) is the minimal MCM approximation of \(\mc{N}(t')\), \(M\) is a direct summand of \(\mc{M}(t')\) by Proposition \ref{prop.minapprox}, and hence \({}_{\cat{MCM}}\mu(\mc{N}(t'))=\mu(M)\leq\mu(\mc{M}(t'))\leq n\).
\end{proof}
\begin{ex}\label{ex.semicont1}
The Betti numbers \(\beta_{i}(M)=\dim\tor{A}{i}{M}{A/\fr{m}_{A}}\) are well known upper semi-continuous invariants of finite modules over local rings. By Theorem \ref{thm.semicont} the induced invariants \({}_{\cat{MCM}}\beta_{i}\), \({}_{\cat{MCM}}\beta_{i}'\), \({}_{\cat{FID}}\beta_{i}\) and \({}_{\cat{FID}}\beta_{i}'\) are upper semi-continuous too.
\end{ex}
We now consider some invariants defined in terms of Cohen-Macaulay approximation.
If \(h:S\rightarrow T\) is in one of the categories of local Cohen-Macaulay maps, a map \(\partial:D\rightarrow D'\) of objects in \(\cat{D}_{h}\) is said to be \emph{minimal} if \(k{\otimes}_{T}\partial=0\). Any module \(D\) in \(\cat{D}_{h}\) is isomorphic to some \(\omega_{h}^{\oplus n}\) and \(\nd{}{T}{\omega_{h}}\cong T\). Hence if \(\partial\) is not minimal then there is a surjection \(D'\rightarrow\omega_{h}\) inducing a surjection \(D\rightarrow\omega_{h}\). By Corollary \ref{cor.locCMapprox} the \(\omega_{h}\) splits off from \(\partial\). Hence any \(\cat{D}_{h}\)-complex is homotopy equivalent to one with all differentials being minimal, which is called a minimal \(\cat{D}\)-complex.
For any module \(\mc{N}\) in \(\cat{mod}{}^{\textnormal{fl}}_{h}\) over \(\cat{xCM}\) we choose a minimal \(\cat{MCM}\)-approximation \(\mc{L}\rightarrow\mc{M}\rightarrow\mc{N}\) and a minimal \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull \(\mc{N}\rightarrow\mc{L}'\rightarrow\mc{M}'\) which exist by Corollary \ref{cor.minapprox}. Spliced with a
minimal \(\cat{D}\)-resolution of \(\mc{L}\) and a minimal \(\cat{D}\)-coresolution of \(\mc{M}'\) we obtain complexes \({}^{-}C^{*}(\mc{N})\), \({}^{+}C^{*}(\mc{N})\) and \(D^{*}(\mc{N})\), as defined in Section \ref{subsec.cplx}, where no differential has any \(\omega_{h}\)-summand. We call such choices of these complexes for minimal. They are unique:
\begin{lem}\label{lem.uniqueC}
Suppose \(h\) is in \(\cat{hCM}\)\textup{,} \(\cat{lCM}\) or \(\cat{cCM}\) and \((h,\mc{N})\) is in \(\cat{mod}{}^{\textnormal{fl}}_{h}\)\textup{.} Then
minimal choices of \({}^{-}C^{*}(\mc{N})\), \({}^{+}C^{*}(\mc{N})\) and \(D^{*}(\mc{N})\) exist and are unique up to non-canonical isomorphisms\textup{.}
\end{lem}
\begin{proof}
Minimal choices \({}^{+}C^{*}_{1}\) and \({}^{+}C^{*}_{2}\) of coresolutions for \(\mc{N}\) are by Lemma \ref{lem.res} homotopic through chain maps \(\alpha\) and \(\beta\) starting with an isomorphism \(\mc{L}_{1}'\cong\mc{L}_{2}'\). If \(\rho_{i}\) are homotopies with \(\beta\alpha-\id=\partial\rho_{1}+\rho_{1}\partial\) and \(\alpha\beta-\id=\partial\rho_{2}+\rho_{2}\partial\) then tensoring down by \(k{\otimes}_{T}-\) makes the right hand side of these identities equal to zero by the minimality of the complexes. Hence \(\beta\alpha\) and \(\alpha\beta\) are surjective endomorphisms, i.e.\ isomorphisms. The same argument applies to \({}^{-}C^{*}\) and \(D^{*}\).
\end{proof}
For each module \(\mc{N}\) in \(\cat{mod}{}^{\textnormal{fl}}_{h}\) we fix a minimal \(\cat{D}\)-complex \((D^{*}(\mc{N}),\partial^{*})\) representing \(\mc{N}\). Let \(\mc{L}'=\coker \partial^{-1}\) and \(\mc{M}=\ker \partial^{0}\).
Put \(\syz{\omega_{h}}{i}\mc{L}':=\coker\{\partial^{-i{-}1}:D^{{-}i{-}1}\rightarrow D^{-i}\}\) for \(i\geq 0\) and \(\syz{\omega_{h}}{-i}\mc{M}:=\ker\{\partial^{i}:D^{i}\rightarrow D^{i{+}1}\}\) for \(i\geq 0\). For any finite \(T\)-module \(\mc{N}\) let the \emph{\(\omega_{h}\)-rank} of \(\mc{N}\), denoted \(\rk{\omega_{h}}{\mc{N}}\), be the largest number \(n\) with \(\omega_{h}^{{\oplus}n}{\oplus}\mc{N}'\cong \mc{N}\) for some \(T\)-module \(\mc{N}'\). Since \(\nd{}{T}{\omega_{h}}\cong T\) is a local ring, this is a well behaved invariant, cf.\ \cite[Sec.\ 1.1]{sim/str:02}.
\begin{defn}\label{defn.c}
Suppose \(h\) is in \(\cat{hCM}\)\textup{,} \(\cat{lCM}\) or \(\cat{cCM}\) and \(\mc{N}\) is in \(\cat{mod}{}^{\textnormal{fl}}_{h}\)\textup{.} Define the numbers\textup{:}
\begin{enumerate}
\item[(i)] \(d^{i}_{T}(\mc{N}):=\rk{\omega_{h}}{D^{i}(\mc{N})}\) for all \(i\)
\item[(ii)] \(\nu^{T}_{i}(\mc{N}):=\rk{\omega_{h}}{\syz{\omega_{h}}{i}\mc{L}'}\) for \(i\geq 0\)
\item[(iii)] \(\gamma_{T}(\mc{N}):=\rk{\omega_{h}}{\mc{M}}\)
\end{enumerate}
\end{defn}
The definition gives well defined invariants of \(\mc{N}\) by Lemma \ref{lem.uniqueC}. In particular we see that \(\nu^{T}_{0}(\mc{N})\) equals \(\rk{\omega_{h}}{\mc{L}'}\). We also notice that \(\rk{\omega_{h}}{\syz{\omega_{h}}{-i}\mc{M}}=0\) for all \(i>0\) by Proposition \ref{prop.minapprox}.
The same notation is used for the absolute counterparts of these invariants.
\begin{rem}
In the absolute case with \(A\) a Gorenstein local ring, the \(\gamma_{A}\)-invariant is called Auslander's \(\delta\) invariant and has been studied in particular by S.\ Ding. One has that \(\gamma_{A}(A/\fr{m}_{A}^{n})\leq 1\) with equality for \(n\gg0\). The smallest number \(n\) with \(\gamma_{A}(A/\fr{m}_{A}^{n})=1\) is called the \emph{index} of \(A\) and Ding has given results and conjectures concerning this invariant, cf.\ \cite{din:92}. See also \cite{has/shi:97}.
Let \((A,\fr{m},k)\) be a noetherian local ring and \(N\) a finite \(A\)-module. Simon and Strooker introduced the \emph{reduced Bass numbers}
\begin{equation}
\nu_{A}^{i}(N)=\dim_{k}\im\{\xt{i}{A}{k}{N}\rightarrow\cH^{i}_{\fr{m}}(N)\}
\end{equation}
and in the case \(A\) has a canonical module they showed in \cite[2.6 and 3.10]{sim/str:02} that
\begin{equation}
\nu^{A}_{0}(N)=\nu_{A}^{\dim A}(N)\quad\text{and}\quad \nu^{A}_{i}(N)=\nu_{A}^{\dim A{-}i}(L')\,\,\,\text{for } i\geq 0
\end{equation}
where \(L'\) is the minimal \(\cat{D}_{A}\)-hull of \(N\). The Canonical Element Conjecture and the Monomial Conjecture are equivalent to \(\nu_{A}^{\dim A}(\fr{b})\neq 0\) (or equivalently \(\gamma_{A}(A/\fr{b})=0\)) for certain ideals \(\fr{b}\) in any Gorenstein local ring \(A\), see \cite[6.4 and 6.6]{sim/str:02}.
\end{rem}
Let \(\cat{P}\) and \(\hat{\cat{P}}{}^{\textnormal{fl}}\) denote the FAds of finite projective modules, respectively modules of finite projective dimension over \(\cat{xCM}\) defined as for the finite type case.
\begin{lem}\label{lem.fri} With notation as above\textup{:}
\begin{enumerate}
\item[(i)] The functor \(F=\hm{}{-}{\omega_{-}}{-}\) gives exact equivalences of categories fibred in additive categories \(\cat{D}\simeq \cat{P}\) and \(\hat{\cat{D}}{}^{\textnormal{fl}}\simeq\hat{\cat{P}}{}^{\textnormal{fl}}\) over \(\cat{xCM}\)\textup{.}
\item[(ii)] In particular \(d^{-i}(\mc{N})=\beta_{i}(\hm{}{T}{\omega_{h}}{\mc{L}'})\) for all \(i\geq 0\)\textup{.}
\item[(iii)] \(\hm{}{T}{D^{*\geq 0}(\mc{N})}{\omega_{h}}\) gives a minimal free resolution of \(\hm{}{T}{\mc{M}}{\omega_{h}}\)\textup{.} In particular \(d^{i}(\mc{N})=\beta_{i}(\hm{}{T}{\mc{M}}{\omega_{h}})\) for all \(i\geq 0\)\textup{.}
\end{enumerate}
\end{lem}
\begin{proof}
(i) and (ii) are the local variants of Lemma \ref{lem.proj}. Breaking \(\mc{M}\hookrightarrow D^{*\geq 0}(\mc{N})\) into short exact sequences, (iii) follows from Corollary \ref{cor.locCMapprox}.
\end{proof}
\begin{cor}\label{cor.semicont}
Let \(h:S\rightarrow T\) be a finite type Cohen-Macaulay map and suppose \(\mc{N}\) is a \(T\)-module in \(\cat{mod}{}^{\textnormal{fl}}_{h}\)\textup{.} Then \(d^{i}(\mc{N}(t))\) are upper semi-continuous functions in \(t\in\Spec T\) for all \(i\)\textup{.}
\end{cor}
\begin{proof}
This follows from Theorem \ref{thm.semicont} and Lemma \ref{lem.fri}.
\end{proof}
\begin{rem}\label{rem.semicontnot}
The invariants \(\nu_{0}\) and \(\gamma\) are not semi-continuous either way, see Example \ref{ex.semicont}. Moreover, let \(\mc{L}\) be in \(\hat{\cat{D}}_{h}\) with \(L=\mc{L}{\otimes}_{S}k\). If \(\rho_{0}:\omega_{A}\rightarrow L\) is a direct summand, then \(\rho_{0}\) lifts to a map \(\rho:\omega_{h}\rightarrow\mc{L}\), but no lifting \(\rho\) of \(\rho_{0}\) has to split, even if \(A\) is a regular ring.
\end{rem}
\begin{rem}\label{rem.semicont}
One can also define functions of the base. E.g.\ if \(\phi:\Spec T\rightarrow \Spec S\) denotes the map induced from \(h\) and \(\mu(\mc{N}(t))\) is an upper semi-continuous function of \(t\in \Spec T\), then \(\phi_{*}\mu(\mc{N})\) defined by
\begin{equation}
\phi_{*}\mu(\mc{N})(s)=\sup_{t\in\phi^{-1}(s)}\!\!\mu(\mc{N}(t))
\end{equation}
is an upper semi-continuous function in \(s\in \Spec S\) since \(\phi\) is an open map.
\end{rem}
\section{The fundamental module of a Cohen-Macaulay map}\label{sec.fund}
\begin{ex}\label{ex.Fmod}
Let \((A,\fr{m}_{A},k)\) be a Cohen-Macaulay local ring with canonical module \(\omega_{A}\). Let \((G_{*},d_{*})\rightarrow k\) be a minimal \(A\)-free resolution of \(k\) and put \(\syz{}{i}=\syz{A}{i}k=\coker d_{i+1}\). Suppose \(\dim A=d\geq 2\). There are connecting isomorphisms \(\xt{1}{A}{\syz{}{d{-}1}}{\omega_{A}}\cong\xt{2}{A}{\syz{}{d{-}2}}{\omega_{A}}\cong\dots\cong\xt{d}{A}{k}{\omega_{A}}\) which is isomorphic to \(k\) by duality theory. To \(1\in k\) there is hence a non-split short exact sequence
\begin{equation}\label{eq.Fmod}
\theta:\,\,0\rightarrow\omega_{A}\longrightarrow E_{A}\xra{\,\,\pi\,\,} \syz{A}{d{-}1}k\rightarrow 0
\end{equation}
with \(E_{A}\) uniquely defined up to non-canonical isomorphism. We call \(E_{A}\) for the \emph{fundamental module} of \(A\). We claim that \(E_{A}\) is a maximal Cohen-Macaulay module which implies that \eqref{eq.Fmod} is the minimal MCM approximation of \(\syz{A}{d{-}1}k\). If we apply \(\hm{}{A}{-}{\omega_{A}}\) to \eqref{eq.Fmod} we obtain the exact sequence
\begin{equation}\label{eq.Fmod2}
\begin{aligned}
0\rightarrow\, &\hm{}{A}{\syz{A}{d{-}1}k}{\omega_{A}}\longrightarrow\hm{}{A}{E_{A}}{\omega_{A}}\longrightarrow\nd{}{A}{\omega_{A}}\xra{\,\,\partial\,\,}\\
&\xt{1}{A}{\syz{A}{d{-}1}k}{\omega_{A}}\longrightarrow\xt{1}{A}{E_{A}}{\omega_{A}}\rightarrow 0
\end{aligned}
\end{equation}
We have \(\partial(\id)=\theta\) so \(\partial\) is surjective and \(\xt{1}{A}{E_{A}}{\omega_{A}}=0\). By duality theory (e.g.\ \cite[3.5.11]{bru/her:98}) this excludes the possibility \(\depth E_{A}=d-1\) and we conclude that \(E_{A}\) is a maximal Cohen-Macaulay module. If \(N\) is a Cohen-Macaulay module of codimension \(c\) we denote \(\xt{c}{A}{N}{\omega_{A}}\) by \(N^{\vee}\). Since \(\nd{}{A}{\omega_{A}}\cong A\) we get from \eqref{eq.Fmod2} a short exact sequence:
\begin{equation}\label{eq.Fmod3}
0\rightarrow \hm{}{A}{\syz{A}{d{-}1}k}{\omega_{A}}\longrightarrow E_{A}^{\vee}\longrightarrow\fr{m}_{A}\rightarrow 0
\end{equation}
Since \(\xt{i}{A}{k}{\omega_{A}}=0\) for \(i\neq d\), \(0\rightarrow G_{0}^{\vee}\rightarrow\dots G_{d-2}^{\vee}\rightarrow\hm{}{A}{\syz{A}{d{-}1}k}{\omega_{A}}\rightarrow 0\) is a finite \(\omega_{A}\)-resolution and so \eqref{eq.Fmod3} gives the minimal MCM approximation of the maximal ideal. Auslander introduced the fundamental module in the case \(d=2\), see \cite{aus:86}.
\end{ex}
We can make a relative version of the fundamental module in much the same way. Let \({}^{(2)\!}\Delta:\cat{CM}\rightarrow\cat{CM}\) be the morphism of fibred categories over \(\cat{R}\) defined by taking the CM map \(h:S\rightarrow T\) to the composition \(h^{(2)}\) of \(h\) with \(\iota=1{\otimes}\id_{T}:T\rightarrow T{\otimes}_{S}T\) and taking a morphism \((g,f):h_{1}\rightarrow h_{2}\) to the composition of two cocartesian squares \((g,f^{\ot2})\) as follows:
\begin{equation}
\xymatrix@-0pt@C+6pt@R-4pt@H-0pt{
S_{1}\ar[r]^{h_{1}}\ar[d]_{g} & T_{1}\ar[d]^{f}\ar@{}[dr]|-{\mapsto} & S_{1}\ar[r]^{h_{1}}\ar[d]_{g} & T_{1}\ar[d]^{f}\ar[r]^(.4){\iota} & T_{1}{\otimes}_{S_{1}}T_{1}\ar[d]^{f^{\ot2}} \\
S_{2}\ar[r]^{h_{2}} & T_{2} & S_{2}\ar[r]^{h_{2}} & T_{2}\ar[r]^(.4){\iota} & T_{2}{\otimes}_{S_{2}} T_{2}
}
\end{equation}
There is also a functor \(\Delta:\cat{CM}\rightarrow\cat{CM}\) defined by mapping \((g,f)\) to the rightmost cocartesian square \((f,f^{\ot2})\), but it doesn't commute with the forgetful functor \(\cat{CM}\rightarrow\cat{R}\).
Let \({}^{d}\cat{CM}\) denote the full subcategory of CM maps of pure relative dimension \(d\). Then \({}^{d}\cat{CM}\) is a fibred subcategory of \(\cat{CM}\) over \(\cat{R}\) and \({}^{(2)\!}\Delta\) and \(\Delta\) restricts to a morphism \({}^{(2)\!}\Delta:{}^{d}\cat{CM}\rightarrow{}^{2d}\cat{CM}\) over \(\cat{R}\) and a functor \(\Delta:{}^{d}\cat{CM}\rightarrow{}^{d}\cat{CM}\).
Let \(h:S\rightarrow T\) be a finite type CM map of pure relative dimension \(d\geq 2\). Consider \(P\) in \(\cat{P}_{h}\) (see Lemma \ref{lem.proj}) as a \(T^{\ot2}\)-module by pullback along the multiplication map \(\mu:T^{\ot2}\rightarrow T\). By Corollary \ref{cor.xtdef} \(E=\xt{d}{T^{\ot2}}{P}{\omega_{\iota}}\) is flat and finite as \(T\)-module, i.e.\ \(T\)-projective. Let \(P^{*}\) denote \(\hm{}{T}{P}{T}\). By Corollary \ref{cor.xtdef}
\begin{equation}
\nd{}{T}{E} \cong \xt{d}{T^{{\otimes} 2}}{P}{\omega_{\iota}}{\otimes} E^{*}\cong \xt{d}{T^{\ot2}}{P}{\omega_{\iota}{\otimes} E^{*}}\,.
\end{equation}
Combined with the connecting isomorphisms the identity in \(\nd{}{T}{E}\) corresponds to a canonical extension of \(T^{{\otimes} 2}\)-modules:
\begin{equation}\label{eq.Fxt}
0\rightarrow \omega_{\iota}{\otimes}_{T}\xt{d}{T^{\ot2}}{P}{\omega_{\iota}}^{*}\longrightarrow E_{h}(P)\longrightarrow \syz{T^{\ot2}}{d{-}1}P\rightarrow 0\,.
\end{equation}
Let \({}^{d}\cat{P}\) and \({}^{d}\cat{MCM}\) denote the restriction of \(\cat{P}\) and \(\cat{MCM}\) to fibred categories over \({}^{d}\cat{CM}\).
\begin{prop}\label{prop.Fmod}
Let \(d\geq 2\)\textup{.} The association \((h,P)\mapsto E_{h}(P)\) in \eqref{eq.Fxt} induces
\begin{enumerate}
\item[(i)] a functor \(E:{}^{d}\cat{P}\rightarrow {}^{d}\cat{MCM}/{}^{d}\cat{P}\) which preserves cocartesian maps and lifts the functor \(\Delta:{}^{d}\cat{CM}\rightarrow{}^{d}\cat{CM}\)\textup{,} and
\item[(ii)] a morphism \({}^{(2)\!}E:{}^{d}\cat{P}\rightarrow {}^{2d}\cat{MCM}/{}^{2d}\cat{P}\) of fibred categories over \(\cat{R}\) which lifts \({}^{(2)\!}\Delta:{}^{d}\cat{CM}\rightarrow{}^{2d}\cat{CM}\)\textup{.}
\end{enumerate}
\end{prop}
\begin{proof}
As an extension of \(T\)-flat modules, \(E_{h}(P)\) is \(T\)-flat. Applying \(\hm{}{T^{\ot2}}{-}{\omega_{\iota}}\) to \eqref{eq.Fxt} with \(E=\xt{d}{T^{\ot2}}{P}{\omega_{\iota}}\) and \(\syz{}{d{-}1}=\syz{T^{\ot2}}{d{-}1}P\) gives an exact sequence
\begin{equation}\label{eq.Fmod4}
\begin{aligned}
0\rightarrow\, &\hm{}{T^{\ot2}}{\syz{}{d{-}1}}{\omega_{\iota}}\rightarrow\hm{}{T^{\ot2}}{E_{h}(P)}{\omega_{\iota}}\rightarrow\nd{}{T^{\ot2}}{\omega_{\iota}}{\otimes}_{T}E\xra{\partial}\\
&\xt{1}{T^{\ot2}}{\syz{}{d{-}1}}{\omega_{\iota}}\rightarrow\xt{1}{T^{\ot2}}{E_{h}(P)}{\omega_{\iota}}\rightarrow 0
\end{aligned}
\end{equation}
by Corollary \ref{cor.xtdef} and duality theory. In particular there is a canonical isomorphism \(\hm{}{T^{\ot2}}{\omega_{\iota}{\otimes}_{T}E^{*}}{\omega_{\iota}}\cong\nd{}{T^{\ot2}}{\omega_{\iota}}{\otimes}_{T}E\). We have that \(\nd{}{T^{\ot2}}{\omega_{\iota}}\) is canonically isomorphic to \(T^{\ot2}\) and \(\partial(t{\otimes}\xi)=\mu(t)\syz{T^{\ot2}}{d{-}1}(\xi)\in \xt{1}{T^{\ot2}}{\syz{}{d{-}1}}{\omega_{\iota}}\) where \(\syz{T^{\ot2}}{d{-}1}\) is the composition of the connecting isomorphisms. So \(\partial\) is surjective and \(\xt{1}{T^{\ot2}}{E_{h}(P)}{\omega_{\iota}}=0\) by Corollary \ref{cor.xtdef}. It follows that all fibres of \(E_{h}(P)\) are MCM modules and so \(E_{h}(P)\) is in \(\cat{MCM}_{\iota}\) and \(\eqref{eq.Fxt}\) is an \(\cat{MCM}_{\iota}\)-approximation of \(\syz{T^{\ot2}}{d{-}1}P\). Let \(\mc{I}_{h}\) denote the kernel of \(\mu:T^{\ot2}\rightarrow T\) and \((-)^{\vee}=\hm{}{T^{\ot2}}{-}{\omega_{\iota}}\). From \eqref{eq.Fmod4} we get another \(\cat{MCM}_{\iota}\)-approximation
\begin{equation}\label{eq.Fmod5}
0\rightarrow \hm{}{T^{\ot2}}{\syz{T^{{\otimes} 2}}{d{-}1}P}{\omega_{\iota}}\longrightarrow E_{h}(P)^{\vee}\longrightarrow\mc{I}_{h}{\otimes}_{T}\xt{d}{T^{\ot2}}{P}{\omega_{\iota}}\rightarrow 0\,.
\end{equation}
Dualising \eqref{eq.Fmod5} induces \eqref{eq.Fxt} since \(E_{h}(P)\cong E_{h}(P)^{\vee\vee}\) and \(\hm{}{T^{\ot2}}{\mc{I}_{h}}{\omega_{\iota}}\cong\omega_{\iota}\).
By Theorem \ref{thm.flatCMapprox} the image of \(E_{h}(P)^{\vee}\) in \(\cat{MCM}_{\iota}/\cat{D}_{\iota}\) is functorial in the \(T^{\ot2}\)-module \(\mc{I}_{h}{\otimes} E\) which again is contravariantly functorial in \(P\). Since \((-)^{\vee}\) induces an equivalence
\begin{equation}
\xymatrix@-0pt@C+6pt@R-4pt@H-0pt{
\vee:\cat{MCM}_{\iota}/\cat{P}_{\iota}\ar@{<->}[r]^(0.47){\simeq} & \cat{MCM}_{\iota}^{\text{op}}/\cat{D}_{\iota}^{\text{op}}:\vee
}
\end{equation}
we conclude that \(E_{h}(P)\) is functorial in \(\cat{MCM}/\cat{P}\) by our functorial choice of extension. This gives (i) and (ii).
\end{proof}
\begin{cor}\label{cor.Fmod}
For any Cohen-Macaulay map \(h:S\rightarrow T\) of pure relative dimension \(d\geq 2\) there is a finite \(T^{\ot2}\)-module \(E_{h}=E_{h}(T)\) which is faithfully flat along \(\iota:T\rightarrow T^{\ot2}\) with all fibres being maximal Cohen-Macaulay modules\textup{.} The association \(h\mapsto E_{h}\) defines a functor \({}^{d}\cat{CM}\rightarrow {}^{d}\cat{MCM}/{}^{d}\cat{P}\) lifting \(\Delta:{}^{d}\cat{CM}\rightarrow{}^{d}\cat{CM}\)\textup{.} In particular \(E_{h}\) gives \(\cat{MCM}_{\iota}\)-approximations
\begin{equation*}
0\rightarrow\omega_{\iota}\longrightarrow E_{h}\longrightarrow \syz{T^{\ot2}}{d{-}1}T\rightarrow 0\quad\text{and}
\end{equation*}
\begin{equation*}
0\rightarrow \hm{}{T^{\ot2}}{\syz{T^{\ot2}}{d{-}1}T}{\omega_{\iota}}\longrightarrow E_{h}^{\vee}\longrightarrow \mc{I}_{h}\rightarrow 0
\end{equation*}
where \(\mc{I}_{h}\) is the kernel of the multiplication map \(T^{\ot2}\rightarrow T\)\textup{.}
\end{cor}
\begin{proof}
This follows from Proposition \ref{prop.Fmod} and \eqref{eq.Fmod5} once we have proved the natural isomorphism \(\xt{d}{T^{\ot2}}{T}{\omega_{\iota}}\cong T\).
Choose a surjection of \(S\)-algebras \(P\rightarrow T\) with \(P=S[t_{1},\dots,t_{N}]\). Recall that \(\omega_{\iota}\) can be given as \(\xt{N-d}{P{\otimes} T}{T^{\ot2}}{\omega_{P{\otimes} T/T}}\) where \(\omega_{P{\otimes} T/T}=\wedge^{N}\Omega_{P{\otimes} T/T}\). There is a change of rings spectral sequence
\begin{equation}
\cE^{p,q}_{2}=\xt{q}{T^{\ot2}}{T}{\xt{p}{P{\otimes} T}{T^{\ot2}}{\omega_{P{\otimes} T/T}}}\Rightarrow\xt{p+q}{P{\otimes} T}{T}{\omega_{P{\otimes} T/T}}
\end{equation}
which by Corollary \ref{cor.xtdef} and duality theory collapses to the canonical isomorphism
\begin{equation}
\xt{d}{T^{\ot2}}{T}{\xt{N-d}{P{\otimes} T}{T^{\ot2}}{\omega_{P{\otimes} T/T}}}\cong\xt{N}{P{\otimes} T}{T}{\omega_{P{\otimes} T/T}}\,.
\end{equation}
By \cite[3.5.6]{con:00} \(\xt{N}{P{\otimes} T}{T}{\omega_{P{\otimes} T/T}}\) is canonically isomorphic to \(\omega_{T/T}=T\) as \(T^{\ot2}\)-module.
\end{proof}
We will call the module \(E_{h}\) given in Corollary \ref{cor.Fmod} for the \emph{fundamental module} of the Cohen-Macaulay map \(h\).
\begin{ex}\label{ex.semicont}
Let \(k\) be an algebraically closed field and \(A\) a finite type \(k\)-algebra which is Cohen-Macaulay of pure dimension \(2\). Then the fundamental module \(E=E_{h}\) of \(h:k\rightarrow A\) is the maximal Cohen-Macaulay approximation of \(I=\ker\{A^{{\otimes} 2}\rightarrow A\}\) in \(\cat{mod}{}^{\textnormal{fl}}_{\iota}\);
\begin{equation}\label{eq.semicont}
0\rightarrow\omega_{h}{\otimes} A\longrightarrow E\longrightarrow I\rightarrow 0
\end{equation}
where \(\iota=1{\otimes}\id:A\rightarrow A^{{\otimes} 2}\) and \(\omega_{h}\cong\omega_{A}\).
Let \(t\) in \(\Spec A^{{\otimes} 2}\) be a \(k\)-point, and \(t_{i}\in \Spec A\) be the image of \(t\) by the \(i^{\text{th}}\) projection. Let \(A_{i}\) denote \(A\) localised at \(t_{i}\). Let \(\fr{m}_{i}\) be the maximal ideal in \(A_{i}\). Localising gives a local Cohen-Macaulay map \(\iota_{\fr{p}_{t}}:A_{2}\rightarrow (A^{\ot2})_{\fr{p}_{t}}\) and a module \(E_{\fr{p}_{t}}\) in \(\cat{MCM}_{\iota_{\fr{p}_{t}}}\). Let \(E(t)\) denote base change of \(E_{\fr{p}_{t}}\) to \(k(t_{2})\). If \(t_{1}=t_{2}\) then \(I(t)\cong \fr{p}_{1}\) and \(E(t)\) equals the fundamental module \(E_{A_{1}}\) of \eqref{eq.Fmod}. If \(t_{1}=t_{2}\) is singular, then \(\rk{\omega}{E(t)}=0\) while if \(t_{1}=t_{2}\) is regular then \(E(t)\cong A_{1}^{{\oplus}2}\). If \(t_{1}\neq t_{2}\) then \(I(t)\cong A_{1}\cong E_{A_{1}}\) and \(E(t)\cong A_{1}^{\oplus 2}\). This shows that \(\gamma(I)(t)\) is \emph{not} upper semi-continuous as the \(d^{i}\)-invariants are.
In particular, if \(A\) equals \(k[x,y,z]/(x^{n{+}1}{-}yz)\) with a \(2\)-dimensional rational double point at \(\fr{m}_{0}=(x,y,z)\), similar considerations give the following table of invariants (note that \(\nu_{1}=d^{-1}\)):
\begin{tabular}[t]{| l | l | c | c | c | c |}
\hline
\(\iota:A\rightarrow A^{{\otimes} 2}\) & \(\gamma\) & \(\nu_{1}\) & \(d^{0}\) & \(\nu_{0}\) & \(I(t)\)
\\[0.5ex]\hline\hline
\(t_{1}=t_{2}=0\) singular point & \(0\) & \(1\) & \(4\) & \(1\) & \(\fr{m}_{0}A_{\fr{m}_{0}}\) \\ \hline
\(t_{1}=t_{2}\) non-singular point & \(2\) & \(1\) & \(2\) & \(0\) & \(\fr{m}_{1}A_{1}\) \\ \hline
\(t_{1}\neq t_{2}\) & \(1\) & \(0\) & \(1\) & \(1\) & \(A_{1}\) \\ \hline
\end{tabular}
\end{ex}
\section{Maps of deformation functors induced by\\ Cohen-Macaulay approximation}\label{sec.def}
We extend the Cohen-Macaulay approximation over henselian local rings to deformations and obtain maps between the associated deformation functors. We also introduce the appropriate Andr{\'e}-Quillen cohomology and links the various cohomologies in a fundamental long-exact sequence.
Fix an object \(\xi=(h:S\rightarrow T,\mc{N})\) in \(\cat{mod}{}^{\textnormal{fl}}\) over \(\cat{H}\). A \emph{deformation} of \(\xi\) is a cocartesian morphism \(\alpha_{1}:\xi_{1}\rightarrow \xi\) in \(\cat{mod}{}^{\textnormal{fl}}\). A \emph{map of deformations} \(\alpha_{1}\rightarrow\alpha_{2}\) is a cocartesian morphism \(\phi:\xi_{1}\rightarrow \xi_{2}\) in \(\cat{mod}{}^{\textnormal{fl}}\) such that \(\alpha_{2}\phi = \alpha_{1}\). Deformations and maps of deformations are objects and arrows in the comma category \(\cat{Def}_{\xi}:=\cat{mod}{}^{\textnormal{fl}}_{\text{coca}}/\xi\) which is fibred in groupoids over the comma category \(\cat{H}/S\), see Lemma \ref{lem.gpoid} and the proceeding comments. Let the \emph{deformation functor} \(\df{}{\xi}:\cat{H}/S\rightarrow \Sets\) be the functor corresponding to the associated groupoid of sets \(\ol{\cat{Def}}_{\xi}\). The comma category \(\cat{Def}_{h}:=\cat{hCM}/h\) of deformations of \(h:S\rightarrow T\) is also fibred in groupoids over \(\cat{H}/S\) and we have an obvious factorisation \(\cat{Def}_{\xi}\rightarrow\cat{hCM}/h\rightarrow \cat{H}/S\) which makes \(\cat{Def}_{\xi}\) fibred in groupoids over \(\cat{hCM}/h\). To ease readability (and by abuse of notation) we put \(\df{}{(T,\mc{N})}(S_{1})=\df{}{\xi}(S_{1}{\rightarrow} S)\) and \(\df{}{T}(S_{1})=\df{}{h}(S_{1}{\rightarrow} S)\). We also write a \emph{deformation of \((T,\mc{N})\)} meaning a deformation of \(\xi\) and likewise in similar situations.
For each object \(\xi_{i}=(h_{i},\mc{N}_{i})\) in \(\cat{mod}{}^{\textnormal{fl}}\) over \(\cat{H}\) we choose a minimal \(\cat{MCM}\)-approximation \(\pi_{i}:\mc{L}_{i}\rightarrow \mc{M}_{i}\xra{\pi_{i}}\mc{N}_{i}\) and a minimal \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull \(\iota_{i}:\mc{N}_{i}\xra{\iota_{i}}\mc{L}'_{i}\rightarrow \mc{M}_{i}'\) which exist by Corollary \ref{cor.locCMapprox} and Corollary \ref{cor.minapprox}. For each deformation \(\alpha_{i}:\xi_{i}\rightarrow \xi_{0}\) we choose extensions to commutative diagrams of deformations
\begin{equation}\label{eq.right2}
\xymatrix@C-0pt@R-8pt@H-30pt{
\mc{L}_{i}\ar[r]\ar@{.>}[d]_{\lambda_{i}} & \mc{M}_{i}\ar[r]^{\pi_{i}}\ar@{.>}[d]_{\mu_{i}} & \mc{N}_{i}\ar[d]_{\nu_{i}} \\
\mc{L}_{0}\ar[r] & \mc{M}_{0} \ar[r]^{\pi_{0}} & \mc{N}_{0}
}
\qquad\text{and}\qquad
\xymatrix@C-0pt@R-8pt@H-30pt{
\mc{N}_{i}\ar[r]^{\iota_{i}}\ar[d]_{\nu_{i}} & \mc{L}_{i}'\ar[r]\ar@{.>}[d]_{\lambda_{i}'} & \mc{M}_{i}'\ar@{.>}[d]_{\mu_{i}'} \\
\mc{N}_{0}\ar[r]^{\iota_{0}} & \mc{L}_{0}' \ar[r] & \mc{M}_{0}'
}
\end{equation}
as follows: By Corollary \ref{cor.minapprox} a base change of \(\pi_{i}\) by \(h_{i}\rightarrow h_{0}\) gives a minimal \(\cat{MCM}_{h_{0}}\)-approximation \(\mc{M}_{i}^{\#}\rightarrow\mc{N}_{i}^{\#}\xra{\simeq} \mc{N}_{0}\). By minimality it is isomorphic to \(\pi_{0}\). Choose an isomorphism. Let \(\mu_{i}\) be the composition \(\mc{M}_{i}\rightarrow \mc{M}_{i}^{\#}\xra{\simeq}\mc{M}_{0}\). It is cocartesian. Do similarly for the \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull. Let these choices be fixed.
The diagrams in \eqref{eq.right2} will be called an \(\cat{MCM}\)-approximation (denoted \(\pi_{*}\)) respectively a \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull (denoted \(\iota_{*}\)) of \(\nu_{i}\) (this terminology can be justified).
\begin{defn}\label{defn.defmap}
There are four maps of deformation functors
\begin{equation*}
\sigma_{X}:\df{}{(h_{0},\mc{N}_{0})}\longrightarrow\df{}{(h_{0},X)}:\cat{H}/S_{0}\longrightarrow\Sets
\end{equation*}
where \(X\) can be \(\mc{M}_{0},\mc{L}_{0},\mc{L}_{0}'\) and \(\mc{M}_{0}'\) given by \([(h_{i}\rightarrow h_{0}, \nu_{i})]\mapsto [(h_{i}\rightarrow h_{0}, x)]\) for \(x\) equal to \(\mu_{i},\lambda_{i},\lambda_{i}'\) and \(\mu_{i}'\) in \eqref{eq.right2} respectively.
\end{defn}
The following lemma implies that these maps are well defined and independent of choices.
\begin{lem}\label{lem.defCMapprox2}
Given two deformations \(\nu_{ij}:\mc{N}_{ij}\rightarrow\mc{N}_{0j}\) \textup{(}\(j=1,2\)\textup{)} in \(\cat{mod}{}^{\textnormal{fl}}\) over \(h_{ij}\rightarrow h_{0j}\) in \(\cat{hCM}\) and \(\cat{MCM}\)-approximations \(\pi_{*j}\) \textup{(}respectively \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hulls \(\iota_{*j}\)\textup{)} of \(\nu_{ij}\)\textup{.} Suppose we have a map of short exact sequences \(\pi_{01}\rightarrow\pi_{02}\) \textup{(}respectively \(\iota_{01}\rightarrow\iota_{02}\)\textup{)} and a map \(\alpha:\mc{N}_{i1}\rightarrow\mc{N}_{i2}\) lifting \(\mc{N}_{01}\rightarrow\mc{N}_{02}\)\textup{,} i\textup{.}e\textup{.}\ such that the following two diagrams of solid arrows are commutative\textup{:}
\begin{equation*}\label{eq.C}
\xymatrix@C-20pt@R-8pt@H-30pt{
& \mc{L}_{i2}\ar[rr]\ar[dd]^(0.3){\lambda_{i2}}|\hole && \mc{M}_{i2}\ar[rr]^(0.40){\pi_{i2}}\ar[dd]|\hole^(0.3){\mu_{i2}}
&& \mc{N}_{i2}\ar[dd]^(0.3){\nu_{i2}} \\
\mc{L}_{i1}\ar@{.>}[ur]\ar[rr]\ar[dd]^(0.3){\lambda_{i1}} && \mc{M}_{i1}\ar@{.>}[ur]^(0.4){\gamma}\ar[rr]^(0.4){\pi_{i1}}\ar[dd]^(0.3){\mu_{i1}}
&& \mc{N}_{i1}\ar[ur]^(0.4){\alpha}\ar[dd]^(0.3){\nu_{i1}} & \\
& \mc{L}_{02}\ar[rr]|\hole && \mc{M}_{02} \ar[rr]^(0.35){\pi_{02}}|\hole && \mc{N}_{02} \\
\mc{L}_{01}\ar[ur]\ar[rr] && \mc{M}_{01}\ar[ur]^(0.45){\beta} \ar[rr]^(0.4){\pi_{01}} && \mc{N}_{01}\ar[ur] &
}
\quad
\xymatrix@C-20pt@R-10pt@H-30pt{
& \mc{N}_{i2}\ar[rr]^(0.40){\iota_{i2}}\ar[dd]^(0.3){\nu_{i2}}|\hole && \mc{L}'_{i2}\ar[rr]\ar[dd]^(0.3){\lambda'_{i2}}|\hole
&& \mc{M}'_{i2}\ar[dd]^(0.3){\mu'_{i2}} \\
\mc{N}_{i1}\ar[ur]^(0.4){\alpha}\ar[rr]^(0.35){\iota_{i1}}\ar[dd]^(0.3){\nu_{i1}} && \mc{L}'_{i1}\ar@{.>}[ur]^(0.4){\gamma'}\ar[rr]\ar[dd]^(0.3){\lambda'_{i1}}
&& \mc{M}'_{i1}\ar@{.>}[ur]\ar[dd]^(0.3){\mu'_{i1}} & \\
& \mc{N}_{02}\ar[rr]^(0.35){\iota_{02}}|\hole && \mc{L}'_{02} \ar[rr]|\hole && \mc{M}'_{02} \\
\mc{N}_{01}\ar[ur]\ar[rr]^(0.4){\iota_{01}} && \mc{L}'_{01}\ar[ur]_(0.55){\beta'} \ar[rr] && \mc{M}'_{01}\ar[ur] &
}
\end{equation*}
Then there exist maps \(\gamma:\mc{M}_{i1}\rightarrow\mc{M}_{i2}\) and \(\gamma':\mc{L}'_{i1}\rightarrow\mc{L}'_{i2}\) such that the induced left \textup{(}respectively right\textup{)} diagram commutes\textup{.} If \(\alpha\) is cocartesian\textup{,} so are \(\gamma\) and \(\gamma'\)\textup{.}
\end{lem}
\begin{proof}
Consider the \(\cat{MCM}\)-approximation case. By applying base changes to the front diagram, we can reduce the problem to the case \(h_{i1}\rightarrow h_{01}\) equals \(h_{i2}\rightarrow h_{02}\). Then, by Corollary \ref{cor.locCMapprox}, there is a lifting \(\gamma_{1}:\mc{M}_{i1}\rightarrow\mc{M}_{i2}\) of \(\alpha\). We would like to adjust \(\gamma_{1}\) so that it lifts \(\beta\) too. We have that \(\mu_{i2}\gamma_{1}-\beta\mu_{i1}\) factors through \(\mc{L}_{02}\) by a map \(\tau_{i}:\mc{M}_{i1}\rightarrow\mc{L}_{02}\) which induces a unique map \(\tau_{0}:\mc{M}_{01}\rightarrow\mc{L}_{02}\) since \(\mu_{i1}\) is cocartesian. If \(\mc{D}_{*}\twoheadrightarrow\mc{L}_{i2}\) is a finite \(\cat{D}\)-resolution, then base change gives a finite \(\cat{D}\)-resolution \(\mc{D}_{*}^{\#}\twoheadrightarrow \mc{L}_{02}\) and \(\tau_{0}\) lifts to a \(\sigma_{0}:\mc{M}_{01}\rightarrow\mc{D}_{0}^{\#}\) by Corollary \ref{cor.locCMapprox}. By Corollary \ref{cor.xtdef} there is a \(\sigma:\mc{M}_{i1}\rightarrow\mc{D}_{0}\) lifting \(\sigma_{0}\) and subtracting the induced map \(\mc{M}_{i1}\rightarrow\mc{M}_{i2}\) from \(\gamma_{1}\) gives our desired \(\gamma\). If \(\alpha\) is an isomorphism so is \(\gamma\) by minimality of the approximations \(\pi_{ij}\).
The argument for the \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-case is similar.
\end{proof}
\begin{rem}\label{rem.defmap}
There are maps of fibred categories inducing the maps \(\sigma_{X}\) in Definition \ref{defn.defmap}. Two maps \(\alpha,\beta: (h_{1},\mc{N}_{1})\rightarrow (h_{2},\mc{N}_{2})\) in \(\cat{Def}_{(h_{0},\mc{N}_{0})}\) are stably equivalent if \(h_{1}=h_{2}\) and \(\alpha{-}\beta\) factors through an object in \(\cat{D}\). Let \(\ul{\cat{Def}}_{(h_{0},\mc{N}_{0})}\) denote the resulting quotient category which is fibred over \(\cat{hCM}/h_{0}\) and over \(\cat{H}/S_{0}\). Then Lemma \ref{lem.defCMapprox2} implies that there are well defined maps of categories fibred in groupoids \(\bs{\sigma}_{X}:\ul{\cat{Def}}_{(h_{0},\mc{N}_{0})}\rightarrow\ul{\cat{Def}}_{(h_{0},X)}\) for \(X\) equal to \(\mc{M}_{0}\), \(\mc{L}_{0}\), \(\mc{L}_{0}'\) and \(\mc{M}_{0}'\). The associated map of functors is \(\sigma_{X}\). Note that \(\ul{\cat{Def}}_{(h_{0},\mc{N}_{0})}\) is different from \(\underline{\cat{mod}}{}^{\textnormal{fl}}_{\text{coca}}/(h_{0},\mc{N}_{0})\). Stably isomorphic modules will in general have different deformation functors. E.g.\ let \(N=A\oplus\omega_{A}\). If \(A\) is not Gorenstein, then one can have \(\xt{1}{A}{N}{N}\neq 0\). But in the stable category \(N\) is isomorphic to \(A\) which is infinitesimally rigid.
\end{rem}
We have the following reformulation. To \((h:S\rightarrow T,\mc{N})\) in \(\cat{mod}{}^{\textnormal{fl}}\) consider \(\varGamma=T{\oplus}\mc{N}\) as a graded \(S\)-algebra with \(T\) in degree \(0\) and \(\mc{N}\) in degree \(1\). A deformation of graded algebras \({\varGamma}_{1}\rightarrow{\varGamma}\) over \(S_{1}\rightarrow S\) in \(\cat{H}/S\) is equivalent to a deformation \((T_{1},\mc{N}_{1})\) of \((T,\mc{N})\). More generally, given a homogeneous morphism of \(\BB{Z}\)-graded rings \(S\rightarrow T\) and a graded \(T\)-module \(M\), there are \emph{Andr{\'e}-Quillen cohomology groups of graded algebras} \(\gH{0}^{i}(S,T,M)=\cH^{i}{\hm{{\text{gr}}}{T}{L^{\text{gr}}_{T/S}}{M}}\). Here \(L^{\text{gr}}_{T/S}\) is the graded cotangent complex defined as \(\Omega_{P/S}{\otimes}_{P}T\) where \(P=P_{S}(T)\) is a graded simplicial degree-wise free \(S\)-algebra resolution of \(T\) and \(\Omega_{P/S}\) denotes the K{\"a}hler differentials, see \cite[IV]{ill:71} for more details (in a more general situation). See also \cite{kle:79}. If \(h:S\rightarrow T\) and \(p:S_{1}\rightarrow S\) are graded ring homomorphisms with \(p\) being surjective, then a \emph{lifting} of \(h\) (`of \(T\)') along \(p\) (`to \(S_{1}\)') is a commutative diagram of graded ring homomorphisms
\begin{equation}\label{eq.lift}
\xymatrix@C-0pt@R-12pt@H-30pt{
T & T_{1}\ar[l]_{q} \\
S\ar[u]^{h} & S_{1}\ar[l]_{p}\ar[u]_{h_{1}}
}
\quad
\textnormal{with \(q{\otimes} S:T_{1}{\otimes}_{S_{1}}S\cong T\) and \(\tor{S_{1}}{1}{T_{1}}{S}=0\).}
\end{equation}
Two liftings \(T_{1}\) and \(T'_{1}\) of \(T\) to \(S_{1}\) are equivalent if there is a graded \(S_{1}\)-algebra isomorphism \(T_{1}\cong T'_{1}\) commuting with \(q\) and \(q'\).
There is an obstruction theory for liftings of graded algebras in terms of graded Andr{\'e}-Quillen cohomology groups.
\begin{prop}[\cite{ill:71,kle:79}]\label{prop.grobs}
Given graded ring homomorphisms \(S\rightarrow T\) and \(p:S_{1}\rightarrow S\) with \(p\) surjective and \(I^{2}=0\) for \(I=\ker p\)\textup{.}
\begin{enumerate}
\item[(i)] There exists an element \(\ob(p,T)\in\gH{0}^{2}(S,T,T{\otimes} I)\) which is natural in \(p\) such that \(\ob(p,T)=0\) if and only if there exists a lifting of \(T\) to \(S_{1}\)\textup{.}
\item[(ii)] If \(\ob(p,T)=0\) then the set of equivalence classes of liftings of \(T\) to \(S_{1}\) is a torsor for \(\gH{0}^{1}(S,T,T{\otimes} I)\) which is natural in \(p\)\textup{.}
\end{enumerate}
\end{prop}
The element \(\ob(p,T)\) is called the \emph{obstruction class} of \((p,T)\). If the rings and modules are concentrated in degree \(0\) this equals the ungraded case and the cohomology groups equals the ungraded Andr{\'e}-Quillen cohomology \(\cH^{*}(S,T,T{\otimes} I)\).
\begin{prop}\label{prop.lang}
Suppose \(T\) is an \textup{(}ungraded\textup{)} \(S\)-algebra and \(N\) is a \(T\)-module\textup{.} Let \({\varGamma}=T{\oplus} N\) be the graded \(S\)-algebra with \(T\) in degree \(0\) and \(N\) in degree \(1\) and let \(J\) be a graded \({\varGamma}\)-module with graded components \(J=J_{0}{\oplus}J_{1}\) of degree \(0\) and \(1\)\textup{.} Then there is a natural long-exact sequence\textup{:}
\begin{align*}
0\rightarrow\, & \hm{}{T}{N}{J_{1}}\longrightarrow \gDer{0}_{S}({\varGamma},J)\longrightarrow \Der_{S}(T,J_{0})\xra{\,\,\partial\,\,} \\
& \xt{1}{T}{N}{J_{1}}\longrightarrow \gH{0}^{1}(S,{\varGamma},J)\longrightarrow \cH^{1}(S,T,J_{0})\xra{\,\,\partial\,\,} \xt{2}{T}{N}{J_{1}}\rightarrow\dots
\end{align*}
\end{prop}
\begin{proof}
To the graded ring homomorphisms \(S\rightarrow T\rightarrow {\varGamma}\) there is a distinguished triangle of transitivity
\begin{equation}\label{eq.triangle}
L^{\text{gr}}_{\varGamma/T/S}:\,\,L^{\text{gr}}_{T/S}{\otimes}_{T}{\varGamma}\longrightarrow L^{\text{gr}}_{{\varGamma}/S}\longrightarrow L^{\text{gr}}_{{\varGamma}/T}\longrightarrow L^{\text{gr}}_{T/S}{\otimes}_{T}{\varGamma}[1]
\end{equation}
in the graded derived category of \({\varGamma}\), see \cite[IV 2.3.4]{ill:71}. The (standard) simplicial resolution \(P_{T}({\varGamma})\) equals \(T\) in degree \(0\), the (standard) \(T\)-free resolution \(F_{T}(N)\) of the \(T\)-module \(N\) in degree \(1\), and higher degree terms, see \cite[IV 1.3.2.1]{ill:71}. It follows that \(\hm{{\text{gr}}}{{\varGamma}}{L^{\text{gr}}_{{\varGamma}/T}}{J}=\hm{}{T}{F_{T}(N)}{J_{1}}\). Since \(L^{\text{gr}}_{T/S}=L_{T/S}\) is consentrated in degree \(0\), \(\hm{{\text{gr}}}{{\varGamma}}{L^{\text{gr}}_{T/S}{\otimes}_{T}{\varGamma}}{J}= \hm{}{T}{L_{T/S}}{J_{0}}\).
\end{proof}
\begin{lem}\label{lem.cohmap}
Let \(h:S\rightarrow T\) be a finite type Cohen-Macaulay map and let \(\mc{N}\) be a \(T\)-module in \(\cat{mod}{}^{\textnormal{fl}}_{h}\)\textup{.} Let \(\mc{L}\rightarrow\mc{M}\xra{\pi}\mc{N}\) and \(\mc{N}\xra{\iota}\mc{L}'\rightarrow\mc{M}'\) be an \(\cat{MCM}_{h}\)-approximation and a \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\)-hull of \(\mc{N}\)\textup{.} Let \(X_{i}\) denote \(\mc{N}\)\textup{,} \(\mc{M}\) and \(\mc{L}'\) for \(i=0,1,2\) respectively\textup{,} and put \(\varGamma_{i}=T{\oplus}X_{i}\)\textup{.} Let \(I\) be any \(S\)-module\textup{.}
Then there are natural maps of short exact sequences of complexes \textup{(}see \eqref{eq.triangle}\textup{)}
\begin{equation*}
\hm{\textnormal{gr}}{\varGamma_{0}}{L^{\textnormal{gr}}_{\varGamma_{0}/T/S}}{\varGamma_{0}{\otimes} I}\xra{\pi^{*}}\hm{\textnormal{gr}}{\varGamma_{1}}{L^{\textnormal{gr}}_{\varGamma_{1}/T/S}}{\varGamma_{0}{\otimes} I}\xla{\pi_{*}}\hm{\textnormal{gr}}{\varGamma_{1}}{L^{\textnormal{gr}}_{\varGamma_{1}/T/S}}{\varGamma_{1}{\otimes} I}
\end{equation*}
and
\begin{equation*}
\hm{\textnormal{gr}}{\varGamma_{0}}{L^{\textnormal{gr}}_{\varGamma_{0}/T/S}}{\varGamma_{0}{\otimes} I}\xra{\iota_{*}}\hm{\textnormal{gr}}{\varGamma_{0}}{L^{\textnormal{gr}}_{\varGamma_{0}/T/S}}{\varGamma_{2}{\otimes} I}\xla{\iota^{*}}\hm{\textnormal{gr}}{\varGamma_{2}}{L^{\textnormal{gr}}_{\varGamma_{2}/T/S}}{\varGamma_{2}{\otimes} I}\textup{.}
\end{equation*}
The induced maps of graded Andr{\'e}-Quillen cohomology
\begin{align*}
\gH{0}^{n}(\pi_{*}) & :\gH{0}^{n}(S,\varGamma_{1},\varGamma_{1}{\otimes} I)\longrightarrow\gH{0}^{n}(S,\varGamma_{1},\varGamma_{0}{\otimes} I)\quad\text{and} \\
\gH{0}^{n}(\iota^{*}) & :\gH{0}^{n}(S,\varGamma_{2},\varGamma_{2}{\otimes} I)\longrightarrow\gH{0}^{n}(S,\varGamma_{0},\varGamma_{2}{\otimes} I)
\end{align*}
are isomorphisms for \(n>0\) and surjections for \(n=0\)\textup{.}
\end{lem}
\begin{proof}
There is a natural map \(L^{\text{gr}}_{\varGamma_{1}/T/S}{\otimes}_{\varGamma_{1}}\varGamma_{0}\rightarrow L^{\text{gr}}_{\varGamma_{0}/T/S}\) of short exact sequences of complexes (cf.\ \cite[II 2.1.1.6]{ill:71}) induced by the graded \(T\)-algebra map \(\varGamma_{1}\rightarrow\varGamma_{0}\). This gives \(\pi^{*}\). Covariance along \(\varGamma_{1}{\otimes} I\rightarrow\varGamma_{0}{\otimes} I\) gives \(\pi_{*}\). In each (cohomological) degree the rightmost terms are naturally identified with \(\hm{}{T}{L_{T/S}}{T{\otimes} I}\) as in the proof of Proposition \ref{prop.lang}. By Theorem \ref{thm.flatCMapprox} and Corollary \ref{cor.xtdef} one has \(\xt{n}{T}{\mc{M}}{\mc{L}{\otimes} I}=0\) for \(n>0\) and the \(\gH{0}^{n}(\pi_{*})\)-statement follows. The other case is similar.
\end{proof}
By Lemma \ref{lem.cohmap} (and Theorem \ref{thm.flatCMapprox}) we get induced natural maps for \(n>0\)
\begin{equation}\label{eq.sigma}
\sigma^{n}_{j}(I):\gH{0}^{n}(S,\varGamma_{0},\varGamma_{0}{\otimes}_{S} I)\longrightarrow\gH{0}^{n}(S,\varGamma_{j},\varGamma_{j}{\otimes}_{S} I)\,\, \text{for}\,\, j=1,\,2 \quad\textnormal{and}
\end{equation}
\begin{equation}\label{eq.tau}
\tau^{n}_{j}(I):\xt{n}{T}{X_{0}}{X_{0}{\otimes}_{S}I}\longrightarrow \xt{n}{T}{X_{j}}{X_{j}{\otimes}_{S}I}\,\, \text{for}\,\, j=1,\,2.
\end{equation}
\begin{ex}\label{ex.lang} By elementary diagram chase Lemma \ref{lem.cohmap} gives the following:
\begin{itemize}
\item[(i)] If \(\pi^{*}:\xt{n}{T}{\mc{N}}{\mc{N}{\otimes} I}\rightarrow\xt{n}{T}{\mc{M}}{\mc{N}{\otimes} I}\) is an isomorphism for \(n=1\) and injective for \(n=2\) then \(\sigma^{1}_{1}(I)\) is an isomorphism and \(\sigma^{2}_{1}(I)\) is injective.
\item[(ii)] If \(\iota_{*}:\xt{n}{T}{\mc{N}}{\mc{N}{\otimes} I}\rightarrow\xt{n}{T}{\mc{N}}{\mc{L}'{\otimes} I}\) is an isomorphism for \(n=1\) and injective for \(n=2\) then \(\sigma^{1}_{2}(I)\) is an isomorphism and \(\sigma^{2}_{2}(I)\) is injective.
\end{itemize}
\end{ex}
\section{Deforming hulls of finite injective dimension}
In order to use Artin's Approximation Theorem \cite{art:69} as extended by D.\ Popescu \cite{pop:86,pop:90} we fix an excellent ring \(\mathcal{O}\) (see \cite[7.8.2]{EGAIV2}). We consider the category of local henselian \(\mathcal{O}\)-algebras in \(\cat{H}\), denoted \({{}_{\varLambda}\cat{H}}\). Fibred categories \(\cat{hCM}\) and \(\cat{mod}{}^{\textnormal{fl}}\) over \({{}_{\varLambda}\cat{H}}\) and \(\cat{Def}_{h}\) and \(\cat{Def}_{\xi}\) over \({{{}_{\varLambda}\cat{H}}}/S\) are defined essentially as in Section \ref{sec.def}. Our previous constructions and results are valid in this context as well. In particular deformation functors \(\Def_{(T,\mc{N})}:{{}_{\varLambda}\cat{H}}/S\rightarrow\Sets\) are defined and the \(\cat{MCM}\)-approximation and \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull induce maps of deformation functors as in Definition \ref{defn.defmap}
\begin{defn}\label{def.modulobs}
Given a lifting diagram of ungraded ring homomorphisms as in \eqref{eq.lift} and a \(T\)-module \(N\). Then a \emph{lifting} of \(N\) to \(T_{1}\) is a \(T_{1}\)-module \(N_{1}\) with \(\tor{S_{1}}{1}{N_{1}}{S}=0\) and a map \(N_{1}\rightarrow N\) inducing an isomorphism \(N_{1}{\otimes} S\cong N\). Two liftings \(N_{1}\) and \(N_{1}'\) of \(N\) to \(T_{1}\) are \emph{equivalent} if there is an isomorphism of \(T_{1}\)-modules \(N_{1}\cong N_{1}'\) commuting with the maps to \(N\).
\end{defn}
There is an obstruction theory for liftings of modules in terms of \(\Ext\)-groups.
\begin{prop}[{\cite[IV 3.1.5]{ill:71}}]\label{prop.obsmodule}
Given \textup{(}ungraded\textup{)} ring homomorphisms as in \eqref{eq.lift} with \(I^{2}=0\) and a \(T\)-module \(N\)\textup{.}
\begin{enumerate}
\item[(i)] There exists an element \(\ob(q,N)\in\xt{2}{T}{N}{N{\otimes} I}\) which is natural in \(q\) such that \(\ob(q,N)=0\) if and only if there exists a lifting of \(N\) to \(T_{1}\)\textup{.}
\item[(ii)] If \(\ob(q,N)=0\) then the set of equivalence classes of liftings of \(N\) to \(T_{1}\) is a torsor for \(\xt{1}{T}{N}{N{\otimes} I}\) which is natural in \(q\)\textup{.}
\end{enumerate}
\end{prop}
The element \(\ob(q,N)\) is called the \emph{obstruction class} of \((q,N)\).
\begin{defn}\label{def.smooth}
Let \({{}_{\varLambda}\cat{A}}/k\) denote the subcategory of artin rings in \({{}_{\varLambda}\cat{H}}/k\).
Let \(F\) and \(G\) be set-valued functors on \({{}_{\varLambda}\cat{H}}/k\) (or \({{}_{\varLambda}\cat{A}}/k\)) with \(\#F(k)=1=\#G(k)\). A map \(\phi:F\rightarrow G\)
is \emph{smooth} (formally smooth) if the natural map of sets \(f_{\phi}:F(S)\rightarrow F(S_{0})\times_{G(S_{0})}G(S)\) is surjective for all surjections \(\pi:S\rightarrow S_{0}\) in \({{}_{\varLambda}\cat{H}}/k\) (in \({{}_{\varLambda}\cat{A}}/k\)).
An element \(\nu\in F(R)\) is \emph{versal} if the induced map \(\hm{}{{{}_{\varLambda}\cat{H}}/k}{R}{-}\rightarrow F\) is smooth and \(R\) is algebraic as \(\mathcal{O}\)-algebra. If the map is bijective then \(\nu\) is universal. An element \(\nu\in F(R)\) (or a formal element, i.e.\ a tower \(\{\nu_{n}\}\in \varprojlim F(R/\fr{m}_{R}^{n+1})\)) is \emph{formally versal} if the induced map \(\hm{}{{{}_{\varLambda}\cat{H}}/k}{R}{-}\rightarrow F\) of functors restricted to \({{}_{\varLambda}\cat{A}}/k\) is formally smooth.
See \cite{art:74}.
\end{defn}
\begin{thm}\label{thm.defgrade}
Let \(k\) be a field and let \(A\) be a Cohen-Macaulay local algebraic \(k\)-algebra\textup{.} Let \(N\) be a finite \(A\)-module\textup{.} Fix a minimal \(\cat{MCM}_{A}\)-approximation \(L\rightarrow M\rightarrow N\) and a minimal \(\hat{\cat{D}}_{A}\)-hull \(N\rightarrow L'\rightarrow M'\)\textup{.} Consider the map \(\sigma_{L'}:\df{}{(A,N)}\rightarrow\df{}{(A,L')}\) of functors \({{}_{\varLambda}\cat{H}}/k\rightarrow\Sets\) as in \textup{Definition \ref{def.modulobs}.}
\begin{enumerate}
\item[(i)] If \(\grade N\geq 1\) then \(\sigma_{L'}\) is injective\textup{.}
\item[(ii)] If \(\grade N\geq 2\) then \(\sigma_{L'}\) restricted to \({{}_{\varLambda}\cat{A}}/k\) is an isomorphism\textup{.}
\item[(iii)] If \(\grade N\geq 2\) and \(\df{}{(A,N)}\) has a versal element then \(\sigma_{L'}\) is an isomorphism\textup{.}
\end{enumerate}
The analogous statements hold for \(\sigma_{L}:\df{}{(A,N)}\rightarrow\df{}{(A,L)}\) if the grade conditions are strengthened by one\textup{.}
\end{thm}
\begin{proof}
(i): Given \(S\) in \({{}_{\varLambda}\cat{H}}/k\) and deformations \(({}^{i\!}h:S\rightarrow {}^{i\!}T,{}^{i\!}\mc{N})\) of \((A,N)\) to \(S\) for \(i=1,2\). Assume that the images \(({}^{i\!}h,{}^{i\!}\mc{L}')\) under \(\sigma_{L'}\) are isomorphic to \((h:S\rightarrow T,\mc{L}')\). Pullback along the isomorphisms of \({}^{i\!}h\) with \(h\) induce for all these modules deformations over \(h\). We show that the \({}^{i\!}\mc{N}\) are isomorphic as deformations over \(h\) which implies that \(\sigma_{L'}\) is injective. Let \(S_{n}=S/\fr{m}_{S}^{n+1}\), \(T_{n}=T{\otimes} S_{n}\) etc.. We construct a tower of isomorphisms \(\{\phi_{n}:{}^{1\!}\mc{N}_{n}\cong{}^{2\!}\mc{N}_{n}\}\) and conclude by Lemma \ref{lem.Aapprox} that the deformations are isomorphic. The case \(n=0\) is trivial. Given \(\phi_{n}\) and use it to identify the \({}^{i\!}\mc{N}_{n}\) and denote them by \(\mc{N}_{n}\). Let \(I=\ker(S_{n+1}\rightarrow S_{n})\). By Proposition \ref{prop.obsmodule} there exists an element \(\xi\) in \(\xt{1}{T_{n}}{\mc{N}_{n}}{\mc{N}_{n}{\otimes} I}\) giving the ``difference'' of the two deformations of \(\mc{N}_{n}\). But \(\mc{N}_{n}{\otimes} I\cong N{\otimes} I\) and by the edge map isomorphism of \eqref{eq.ss} we get \(\xt{i}{T_{n}}{\mc{N}_{n}}{\mc{N}_{n}{\otimes} I}\cong\xt{i}{A}{N}{N}{\otimes} I\) for all \(i\). If \(i>0\) then \(\xt{i}{A}{L'}{L'}\cong\xt{i}{A}{N}{L'}\) and \(\xt{i}{A}{N}{N}\rightarrow\xt{i}{A}{N}{L'}\) is injective if \(\grade N\geq i\) and an isomorphism if \(\grade N\geq i+1\).
The obtained injective map \(p:\xt{1}{A}{N}{N}{\otimes} I\rightarrow \xt{1}{A}{L'}{L'}{\otimes} I\) induces a map of the torsor actions in Proposition \ref{prop.obsmodule} on the liftings of \(\mc{N}_{n}\) and of \(\mc{L}'_{n}\) to \(T_{n+1}\). Since the \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h_{n+1}}\)-hulls of the \({}^{i\!}\mc{N}_{n+1}\) are isomorphic as deformations, \(p\) maps \(\xi\) to \(0\) and so \(\xi=0\) and by Proposition \ref{prop.obsmodule} the \({}^{i\!}\mc{N}_{n+1}\) are isomorphic by an isomorphism \(\phi_{n+1}\) compatible with \(\phi_{n}\).
(ii): By (i) we only need to prove surjectivity. Let \((h:S\rightarrow T,\mc{L}')\) be a deformation of \((A,L')\) to \(S\). We proceed by induction on the length of \(S\). Suppose \(\sigma_{L'}(T_{n},\mc{N}_{n})=(T_{n},\mc{L}'_{n})\). We find that \(\ob(T_{n+1}\rightarrow T_{n},\mc{N}_{n})\) in Proposition \ref{prop.obsmodule} maps to \(\ob(T_{n+1}\rightarrow T_{n},\mc{L}'_{n})\) under \(\xt{2}{T_{n}}{\mc{N}_{n}}{\mc{N}_{n}{\otimes} I}\rightarrow \xt{2}{T_{n}}{\mc{L}'_{n}}{\mc{L}'_{n}{\otimes} I}\) which by the assumption is injective. By assumption there is a lifting \((T_{n+1},\mc{L}'_{n+1})\) to \(S_{n+1}\) so \(\ob(T_{n+1}\rightarrow T_{n},\mc{L}'_{n})=0\). Hence there exists a lifting \({}^{1\!}\mc{N}_{n+1}\) of \(\mc{N}_{n}\) to \(T_{n+1}\). If \(\sigma_{L'}({}^{1\!}\mc{N}_{n+1})={}^{1\!}\mc{L}'_{n+1}\) the difference of \({}^{1\!}\mc{L}'_{n+1}\) and \(\mc{L}'_{n+1}\) gives by Proposition \ref{prop.obsmodule} a \(\xi\in \xt{1}{T_{n}}{\mc{L}_{n}'}{\mc{L}_{n}'{\otimes} I}\). By assumption \(\xt{1}{T_{n}}{\mc{L}_{n}'}{\mc{L}_{n}'{\otimes} I}\) is isomorphic to \(\xt{1}{T_{n}}{\mc{N}'_{n}}{\mc{N}'_{n}{\otimes} I}\). The corresponding element in the latter perturbs \({}^{1\!}\mc{N}_{n+1}\) to a lifting \(\mc{N}_{n+1}\) of \(\mc{N}_{n}\) with \(\sigma_{L'}(\mc{N}_{n+1})=\mc{L}_{n+1}'\).
(iii): Any \(S\) in \({{}_{\varLambda}\cat{H}}/k\) is a direct limit of a filtering system of algebraic \(\mathcal{O}\)-algebras in \({{}_{\varLambda}\cat{H}}/k\). Since \(\df{}{(A,L')}\) is locally of finite presentation (\(A\) is algebraic and \(L'\) has finite presentation) it is sufficient to prove surjectivity of \(\sigma_{L'}\) for \(S\) algebraic. Since \(\mathcal{O}\) is excellent, so is \(S\) by \cite[7.8.3]{EGAIV2} and \cite[18.7.6]{EGAIV4}.
We proceed as in (ii) and construct a tower of deformations \(\{\mc{N}_{n}\}\). If \(({}^{v}T,{}^{v\!}\mc{N})\in\df{}{(A,N)}(R)\) is a versal element, there is a corresponding tower of maps \(\{f_{n}:R\rightarrow S_{n}\}\) such that \(({}^{v}T,{}^{v\!}\mc{N})\) induces \(\{(T_{n},\mc{N}_{n})\}\). We obtain the algebra map \(f:R\rightarrow \hat{S}\) which induces a deformation \(({}^{*}T,{}^{*\!}\mc{N})\) of \((A,N)\) to \(\hat{S}\). We have \(\varprojlim {}^{*}T_{n}\cong\varprojlim T_{n}\) and hence also the completion in the maximal ideals gives an isomorphism \({}^{*}\hat{T}\cong\hat{T}\). By \cite[2.6]{art:69}, \cite[1.3]{pop:86} and \cite{pop:90} there is an isomorphism \({}^{*}T\cong T{\tilde{\otimes}}_{S}\hat{S}\) of deformations of \(A\) whereby \({}^{*}T\) is identified with \(T{\tilde{\otimes}}_{S}\hat{S}\). The tower of isomorphisms \(\{\sigma({}^{*\!}\mc{N}_{n})\cong\mc{L}'_{n}\}\) implies by Lemma \ref{lem.Aapprox} that there is an isomorphism of deformations \(({}^{*}T,\sigma({}^{*\!}\mc{N}))\cong ({}^{*}T,{}^{*\!}\mc{L}')\) where \({}^{*\!}\mc{L}'={}^{*}T{\otimes}_{T}\mc{L}'\).
To apply Artin's Approximation Theorem we define a functor of \(S\)-algebras \(F:{}_{S}\cat{H}/k\rightarrow \Sets\) as follows. If \(\bar{S}\) is in \({}_{S}\cat{H}/k\) let \(\bar{T}\) denote \(T{\tilde{\otimes}}_{S}\bar{S}\) and let \(\bar{\mc{L}}'\) denote \(\bar{T}{\otimes}_{T}\mc{L}'\). Then \(F(\bar{S})\) is defined as equivalence classes of pairs of maps of \(\bar{T}\)-modules \(\bar{\xi}=(\bar{\nu}:\bar{\mc{N}}\rightarrow N,\bar{\iota}:\bar{\mc{N}}\rightarrow\bar{\mc{L}}')\) such that \((\bar{S}\rightarrow\bar{T},\bar{\mc{N}})\) is a deformation of \((A,N)\) to \(\bar{S}\). A map \(\bar{S}\rightarrow\bar{S}'\) gives a map of pairs by base change. Two pairs, \({}^{1}\bar{\xi}\) and \({}^{2}\bar{\xi}\), are equivalent if there is an isomorphism of deformations \({}^{1\!}\bar{\mc{N}}\cong{}^{2\!}\bar{\mc{N}}\) commuting with the \({}^{j}\bar{\iota}\). We show that \(F\) is locally of finite presentation. Suppose \(\bar{S}=\varinjlim {}^{i\!}S\) for a filtered injective system of algebras in \({}_{S}\cat{H}/k\). Put \({}^{i}T=T{\tilde{\otimes}}{}^{i\!}S\). Then \(\varinjlim {}^{i}T\cong \bar{T}\) by \cite[7.8.3]{EGAIV2} and \cite[18.6.14]{EGAIV4}.
Given \(\bar{\xi}\in F(\bar{S})\) as above. Since \(\bar{\mc{N}}\) has finite presentation and since the maps \(\bar{\nu}\) and \(\bar{\iota}\) can be represented on the finite presentations, there is a finite \({}^{i}T\)-module \({}^{i\!}\mc{N}\) and \({}^{i}T\)-linear maps \({}^{i}\nu:{}^{i\!}\mc{N}\rightarrow N\) and \({}^{i}\iota:{}^{i\!}\mc{N}\rightarrow{}^{i\!}\mc{L}'={}^{i}T{\otimes}_{T}\mc{L}'\) inducing \(\bar{\xi}\) by base change. We may also assume that \({}^{i\!}\mc{N}\) is \({}^{i\!}S\)-flat. Hence \(\varinjlim F({}^{i\!}S)\rightarrow F(\bar{S})\) is surjective, and injectivity is similar. Let \({}^{*}\xi\) denote the element in \(F(\hat{S})\) given by \({}^{*\!}\mc{N}\rightarrow N\) and the \(\hat{\cat{D}}\)-hull \({}^{*\!}\mc{N}\rightarrow{}^{*\!}\mc{L}'\). By Artin's Approximation Theorem \cite[1.12]{art:69}, \cite[1.3]{pop:86}, \cite{pop:90} there exist a \(\xi=(\nu:\mc{N}\rightarrow N,\iota:\mc{N}\rightarrow \mc{L}')\) with \(\xi_{1}={}^{*}\xi_{1}\). In particular \(\iota_{0}:\mc{N}_{0}\rightarrow\mc{L}'_{0}\) is injective and by applying Proposition \ref{prop.nakayama} as in Example \ref{ex.nakcplx} \(\iota\) is injective and \(\coker \iota\) is \(S\)-flat. It follows that \(\iota\) is the \(\hat{\cat{D}}\)-hull of \(\mc{N}\) and hence \(\sigma_{L'}\) is surjective. The \(L\)-case is analogous.
\end{proof}
Consider the groupoid of finite type Cohen-Macaulay maps over noetherian rings \(\cat{CM}\rightarrow\cat{R}\) in Section \ref{subsec.ft}. Let \(\mathcal{O}\) be any noetherian ring. By abuse of notation let \(\cat{CM}\rightarrow{{}_{\varLambda}\cat{R}}\) denote the category fibred in groupoids obtained by restriction to the category of noetherian \(\mathcal{O}\)-algebras \({{}_{\varLambda}\cat{R}}\). If \(\mathcal{O}\rightarrow T^{\mspace{-1mu}\mathit{o}}\) is a Cohen-Macaulay map then there is a section \(s:{{}_{\varLambda}\cat{R}}\rightarrow\cat{CM}\) defined by \(s:S\mapsto (S\rightarrow T^{\mspace{-1mu}\mathit{o}}{\otimes} S)\) and we obtain a fibred subcategory \(\cat{T^{o}}\) fibred in sets over \({{}_{\varLambda}\cat{R}}\). We restrict \(\cat{mod}\), \(\cat{MCM}\) and \(\cat{D}\) to \(\cat{T^{o}}\) and obtain categories fibred in abelian and additive categories over \(\cat{T^{o}}\) respectively. These restricted fibred categories satisfy the axioms AB1-4 and BC1-2 and we obtain restricted versions of Theorem \ref{thm.flatCMapprox} and Corollary \ref{cor.flatCMapprox}.
For the local version, fix a field \(k\) and a Cohen-Macaulay local algebraic \(k\)-algebra \(A\). Let \(\mathcal{O}\rightarrow T^{\mspace{-1mu}\mathit{o}}\) be obtained by henselisation of a finite type Cohen-Macaulay map \(\tilde\mathcal{O}\rightarrow \tilde{T}^{\mspace{-1mu}\mathit{o}}\) where \(\tilde\mathcal{O}\) is assumed to be an excellent ring. In particular \(\mathcal{O}\) and \(T^{\mspace{-1mu}\mathit{o}}\) are excellent rings (\cite[7.8.3]{EGAIV2}, \cite[18.7.6]{EGAIV4}). Assume \(T^{\mspace{-1mu}\mathit{o}}/\fr{m}_{T^{\mspace{-1mu}\mathit{o}}}\cong k\) and \(T^{\mspace{-1mu}\mathit{o}}{\otimes}_{\mathcal{O}}k\cong A\). As above there is a section \({{}_{\varLambda}\cat{H}}\rightarrow\cat{hCM}\), \(\cat{T^{o}}\) is the fibred subcategory and we consider deformations in \(\cat{mod}{}^{\textnormal{fl}}_{\vert\cat{T^{o}}}\) of an object \(\xi=(S\rightarrow T^{\mspace{-1mu}\mathit{o}}{\tilde{\otimes}} S,\mc{N})\) and obtain the fibred category of deformations \(\cat{Def}{}^{T^{\mspace{-1mu}\mathit{o}}}_{\mc{N}}:=(\cat{mod}{}^{\textnormal{fl}}_{\vert\cat{T^{o}}})_{\text{coca}}/\xi\) over \({{}_{\varLambda}\cat{H}}/S\). The deformation functor \(\df{T^{\mspace{-1mu}\mathit{o}}}{\mc{N}}:{{}_{\varLambda}\cat{H}}/S\rightarrow\Sets\) is defined by the associated groupoid of sets \(\ol{\cat{Def}}{}^{T^{\mspace{-1mu}\mathit{o}}}_{\mc{N}}\). A special case is given by \(\mathcal{O}=k\) and \(T^{\mspace{-1mu}\mathit{o}}=A\).
\begin{cor}\label{cor.defgrade}
With these assumptions in addition to those in \textup{Theorem \ref{thm.defgrade}} consider \(\sigma_{L'}:\df{T^{\mspace{-1mu}\mathit{o}}}{N}\rightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{L'}\)\textup{.}
\begin{enumerate}
\item[(i)] If \(\grade N\geq 1\) then \(\sigma_{L'}\) is injective\textup{.}
\item[(ii)] If \(\grade N\geq 2\) then \(\sigma_{L'}\) restricted to \({{}_{\varLambda}\cat{A}}/k\) is an isomorphism\textup{.}
\item[(iii)] If \(\grade N\geq 2\) and \(\df{T^{\mspace{-1mu}\mathit{o}}}{N}\) has a versal element then \(\sigma_{L'}\) is an isomorphism\textup{.}
\end{enumerate}
The analogous statements hold for \(\sigma_{L}:\df{T^{\mspace{-1mu}\mathit{o}}}{N}\rightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{L}\) if the grade conditions are strengthened by one\textup{.}
\end{cor}
\begin{proof}
This is not a formal consequence of Theorem \ref{thm.defgrade}, but the proof is similar.
\end{proof}
\begin{prop}\label{prop.defequiv}
With general assumptions as in \textup{Theorem \ref{thm.defgrade}:}
\begin{enumerate}
\item[(i)] If \(Q'=\hm{}{A}{\omega_{A}}{L'}\) and \(Q=\hm{}{A}{\omega_{A}}{L}\) then \(Q'\) and \(Q\) have finite projective dimension and \(\df{}{(A,L')}\cong\df{}{(A,Q')}\) and \(\df{}{(A,L)}\cong\df{}{(A,Q)}\)\textup{.}
\item[(ii)] There are natural maps
\begin{equation*}
s:\df{}{(A,M)}\longrightarrow\df{}{(A,M')}\quad \text{and}\quad t:\df{}{(A,L')}\longrightarrow\df{}{(A,L)}
\end{equation*}
commuting with the maps
\(\sigma_{X}:\df{}{(A,N)}\rightarrow\df{}{(A,X)}\) for \(X\) equal to \(M\) and \(M'\)\textup{,} and to \(L'\) and \(L\)\textup{,} respectively\textup{.} If \(A\) is a Gorenstein ring\textup{,} then \(s\) is an isomorphism\textup{.}
\end{enumerate}
The analogous statements also hold for the deformation functors \(\df{T^{\mspace{-1mu}\mathit{o}}}{X}\)\textup{.}
\end{prop}
\begin{proof}
Lemma \ref{lem.fri} implies (i). For any \(\mc{M}\) in \(\cat{MCM}_{h}\) let \(\mc{M}^{\vee}\) denote \(\hm{}{T}{\mc{M}}{\omega_{h}}\). There is a short exact sequence \(M\rightarrow \omega_{A}^{\oplus n}\rightarrow M'\) giving the short exact sequence \(M^{\vee}\leftarrow A^{\oplus n}\leftarrow (M')^{\vee}\). The map \(s\) is the composition \(\df{}{(A,M)}\cong\df{}{(A,M^{\vee})}\rightarrow\df{}{(A,(M')^{\vee})}\cong\df{}{(A,M')}\), the middle map obtained by taking the syzygy of the deformation.
If \(A\) is a Gorenstein ring then \(\omega_{A}\cong A\) and there is an inverse \(\df{}{(A,M')}\rightarrow\df{}{(A,M)}\) given by the syzygy map.
Fix a minimal short-exact sequence \(L\rightarrow\omega_{A}^{\oplus r}\xra{\mu} L'\). For a deformation \(\lambda':\mc{L}'\rightarrow L'\) there is a lifting of \(\mu\) to a map \(\tilde{\mu}:\omega_{h}^{\oplus r}\rightarrow \mc{L}'\). If \(\mc{L}\) denotes the kernel of \(\tilde{\mu}\) then there is a cocartesian map \(\lambda:\mc{L}\rightarrow L\) commuting with \(\omega_{h}^{\oplus r}\rightarrow \omega_{A}^{\oplus r}\). By Lemma \ref{lem.defCMapprox2} \(\lambda'\mapsto \lambda\) gives a well defined map of deformation functors \(t:\df{}{(A,L')}\rightarrow\df{}{(A,L)}\).
There is a commutative diagram of deformations corresponding to \eqref{eq.pullpush}, with \(\omega_{h}^{\oplus r}\rightarrow \omega_{A}^{\oplus r}\) in the \(D\)-place, which gives the stated commutativity of maps of deformation functors.
\end{proof}
\begin{cor}\label{cor.defapprox}
Let \(A\) be an Cohen-Macaulay local algebraic \(k\)-algebra with residue field \(k\)\textup{.} Suppose \(\dim A\geq 2\)\textup{.} Then there exists finite \(A\)-modules \(L'\) and \(Q'\) with \(\Injdim L'=\dim A=\pdim Q'\) and universal deformations \(\mc{L}'\in\df{A}{L'}(A)\) and \(\mc{Q}'\in\df{A}{Q'}(A)\)\textup{.}
\end{cor}
\begin{proof}
Let \(h=1{\otimes}\id:A\rightarrow A{\tilde{\otimes}}_{k}A=T\) and \(\mc{N}=A\) be the cyclic \(T\)-module defined through the multiplication map. Then \(T{\otimes}_{A}k\cong A\) and \(\mc{N}{\otimes}_{A}k\cong k\) and this gives a deformation \(\mc{N}\rightarrow k\) of the residue field of \(A\) which is universal.
If \(L'\) is the minimal \(\hat{\cat{D}}_{A}\)-hull of the residue field \(k\) then \(\mc{L}'=\sigma_{L'}(\mc{N})\in \df{T^{\mspace{-1mu}\mathit{o}}}{L'}(A)\) is universal by Corollary \ref{cor.defgrade}. If \(Q'=\hm{}{A}{\omega_{A}}{L'}\) then \(\hm{}{T}{\omega_{T}}{\mc{L}'}\in\df{A}{Q'}(A)\) is universal by Proposition \ref{prop.defequiv}.
\end{proof}
The following result extends A.\ Ishii's \cite[3.2]{ish:00} somewhat, but the proof is essentially the same.
\begin{prop}\label{prop.Gor}
Let \(k\) be a field and let \(A\) be a Gorenstein local algebraic \(k\)-algebra. Suppose \(L\rightarrow M\rightarrow N\) is the minimal Cohen-Macaulay approximation of a finite \(A\)-module \(N\)\textup{.} If \(\depth N=\dim A -1\) then
\begin{equation*}
\sigma_{M}:\df{}{(A,N)}\longrightarrow\df{}{(A,M)}\quad\text{and}\quad \sigma_{M}^{A}:\df{A}{N}\longrightarrow\df{A}{M}\quad\text{are smooth\textup{.}}
\end{equation*}
\end{prop}
\begin{proof}
By assumption \(L\cong A^{\oplus r}\). Assume \((h_{1},\mc{M}_{1})\) in \(\df{}{(A,M)}(S_{1})\) maps to \((h,\mc{M})\) along the surjection \(S_{1}\rightarrow S\). Assume \(\sigma_{M}\) maps \((h',\mc{N})\) in \(\df{}{(A,N)}(S)\) to \((h,\mc{M})\). We can assume that \(h'=h\) and that the minimal \(\cat{MCM}_{h}\)-approximation of \(\mc{N}\) is \(\mc{L}\xra{\rho}\mc{M}\rightarrow\mc{N}\) where \(\mc{L}\cong T^{\oplus r}\). Let \(\mc{L}_{1}=T_{1}^{\oplus r}\) and choose a lifting \(\rho_{1}:\mc{L}_{1}\rightarrow\mc{M}_{1}\) of \(\rho\). Put \(\mc{N}_{1}:=\coker\rho_{1}\) with its natural map to \(\mc{N}\). Then \(\mc{N}_{1}\) is \(S_{1}\)-flat and \(\sigma_{M}(h_{1},\mc{N}_{1})=(h_{1},\mc{M}_{1})\).
\end{proof}
\begin{rem}
If \(\dim A\geq 1\) and an MCM \(A\)-module \(M\) has a rank, \(r=\Rk M\), then there is a short exact sequence \(A^{r}\rightarrow M\rightarrow N\) with \(N\) a codimension one Cohen-Macaulay module, cf.\ \cite[1.4.3]{bru/her:98}. Hence in the case \(A\) is a Gorenstein domain all MCM modules admits \(\cat{MCM}_{A}\)-approximations by CM modules in codimension one and Proposition \ref{prop.Gor} applies. However, it's not always possible to continue this reduction: If \(A\) is a normal Gorenstein complete local ring any MCM \(A\)-module \(M\) is the \(\cat{MCM}_{A}\)-approximation of a codimension \(2\) Cohen-Macaulay module up to stable isomorphism if and only if \(A\) is a unique factorisation domain, see \cite{yos/iso:00, kat:07}.
If \(A\) is a normal domain there is also a short exact sequence \(A^{r{-}1}\rightarrow M\rightarrow I\) where \(I\) is an ideal of grade one or two, see \cite[VII 4 Thm.\ 6]{bou:98}. Again Proposition \ref{prop.Gor} applies in the Gorenstein case. Let \(U\) denote the regular locus in \(X=\Spec A\). If \(T=A{\tilde{\otimes}}_{k}S\) for \(S\) in \({{}_{k}\cat{H}}/k\) there is a natural section \(A\rightarrow T\). Let \(U_{T}\) denote \(U\times_{X}\Spec T\). Consider the subfunctor \(\df{A,\wedge}{M}\subseteq\df{A}{M}\) of deformations \(\mc{M}\) with trivial induced deformation \(\wedge^{\Rk M}\mc{M}_{\vert U_{T}}\), cf.\ \cite{ish:00}. Assume \(\dim A=2\). Since \(\bar{I}:=\cH^{0}(U,I)\) is a reflexive module, one gets a non-trivial quotient functor \(\quot{\bar{I}}{I\subseteq \bar{I}}\) and a map \(\quot{\bar{I}}{I\subseteq \bar{I}}\rightarrow \df{A,\wedge}{M}\). The map is also smooth in the Gorenstein case by the argument in Proposition \ref{prop.Gor} (contained in \cite[3.2]{ish:00}). In particular, if \(E_{A}\) is the fundamental module and \(A/\fr{m}_{A}\cong k\) then \(\hm{}{{{}_{k}\cat{H}}/k}{A}{-}\cong\quot{A}{\fr{m}_{A}\subseteq A}\cong\df{A}{A/\fr{m}_{A}}\) gives a mini-versal family for \(\df{A,\wedge}{E_{A}}\) by the MCM approximation in \eqref{eq.semicont}, see \cite[3.4]{ish:00}. Any \(\fr{m}_{A}\)-primary ideal would give a similar result.
\end{rem}
\begin{ex}
Assume \(A/\fr{m}_{A}\cong k\) and let \(M\) denote the minimal MCM approximation of \(k\). It's given as \(M\cong \hm{}{A}{\syz{A}{d}(k^{\vee})}{\omega_{A}}\) where \(d=\dim A\), cf.\ Remark \ref{rem.Buch}. One has \(k^{\vee}=\xt{d}{A}{k}{\omega_{A}}\cong k\). We apply \(\hm{}{A}{-}{\omega_{A}}\) to the short exact sequence \(\syz{A}{}(\fr{m}_{A})\rightarrow A^{\oplus\beta_{1}}\xra{(\ul{x})}\fr{m}_{A}\). Assume \(\dim A = 2\). Since \(\xt{1}{A}{\fr{m}_{A}}{\omega_{A}}\cong k\) we obtain the MCM approximation of \(k\):
\begin{equation}
0\rightarrow\omega_{A}\xra{(\ul{x})^{\text{t}}}\omega_{A}^{\oplus\beta_{1}}\longrightarrow M\longrightarrow k\rightarrow 0
\end{equation}
In particular \(\Rk(M)=\beta_{1}-1\) and \(\mu(M)=t(A)\cdot\beta_{1}+1\) where \(t(A)\) is the Cohen-Macaulay type of \(A\). In the case \(A=A(m)=k[u^{m},u^{m-1}v,\dots,v^{m}]^{\text{h}}\), the vertex of the cone over the rational normal curve of degree \(m\), which has the indecomposable MCM modules \(M_{i}=(u^{i},u^{i-1}v,\dots,v^{i})\) for \(i=0,\dots,m{-}1\), one finds that \(M=M_{m-1}^{\oplus m}\). We have
\begin{equation}
\dim_{k}\df{A}{M}(k[\varepsilon])=\dim_{k}\xt{1}{A}{M}{M}= (m-1)\cdot m^{2}
\end{equation}
while \(\dim_{k}\df{A}{k}(k[\varepsilon])=m+1\). Even in the Gorenstein case (\(m=2\)) the tangent map isn't surjective and so Proposition \ref{prop.Gor} cannot in general be extended to \(\depth N=\dim A -2\). For a detailed description of the strata defined by Ishii in \cite{ish:00} of the reduced versal deformation space of \(M\), see \cite{gus/ile:04b}. Applying \(\hm{}{A}{k}{-}\) to \(\fr{m}_{A}\rightarrow A\rightarrow k\) gives an exact sequence
\begin{equation}
0\rightarrow\xt{1}{A}{k}{k}\longrightarrow\df{A}{\fr{m}_{A}}(k[\varepsilon])\longrightarrow k^{t(A)}\longrightarrow \xt{2}{A}{k}{k}
\end{equation}
since \(\xt{1}{A}{\fr{m}_{A}}{\fr{m}_{A}}\cong\xt{2}{A}{k}{\fr{m}_{A}}\) and \(\dim A = 2\). In the case \(A=A(m)\), \(N=\fr{m}_{A}\), the fundamental module \(E_{A}\) gives the MCM approximation and \(E_{A}\cong M_{m-1}^{\oplus 2}\) with \(\dim_{k}\df{A}{E_{A}}(k[\varepsilon])=4(m-1)\). The conclusion in Proposition \ref{prop.Gor} cannot hold in the non-Gorenstein case \(m>2\).
\end{ex}
\section{Existence of versal elements}\label{sec.versal}
Let \(k\) be a field, \(A\) an algebraic \(k\)-algebra with \(A/\fr{m}_{A}\cong k\) and \(N\) a finite \(A\)-module. Without any Cohen-Macaulay condition on \(A\) we define a deformation \((h:S\rightarrow T,\mc{N})\) of the pair \((A,N)\) to an \(S\) in \({{}_{k}\cat{H}}/k\) as before and obtain the deformation functor \(\df{}{(A,N)}:{{}_{k}\cat{H}}/k\rightarrow\Sets\) as equivalence classes of deformations of pairs.
We say that \(A\) is an \emph{isolated singularity over \(k\)} if there is a finite type \(k\)-algebra \(A^{\text{ft}}\) with a maximal ideal \(\fr{m}_{0}\) such that the henselisation \((A^{\text{ft}})^{\text{h}}_{\fr{m}_{0}}\) is isomorphic to \(A\) and which is smooth over \(k\) at all points in \(\Spec A^{\text{ft}} \setminus\{\fr{m}_{0}\}\). We say that the pair \((A,N)\) is an \emph{isolated singularity over \(k\)} if \(A\) is an isolated singularity over \(k\) and if \(N_{\fr{p}}\) is a free \(A_{\fr{p}}\)-module for all prime ideals \(\fr{p}\neq\fr{m}_{A}\). The following result is a consequence of a result of R.\ Elkik and an argument of H.\ von Essen.
\begin{thm}\label{thm.ExVers}
Let \((A,N)\) be an isolated singularity over the field \(k\) with \(A\) equidimensional\textup{.} Then \(\df{}{(A,N)}:{{}_{k}\cat{H}}/k\rightarrow\Sets\) has a versal element\textup{.}
\end{thm}
\begin{proof}
We apply \cite[3.2]{art:74} (cf.\ \cite{con/jon:02}) to show the existence of a formally versal element for \(\df{}{(A,N)}\). By the finiteness conditions it follows that \(\df{}{(A,N)}\) is locally of finite presentation. The condition (S1) holds in general by Proposition \ref{prop.grobs}. Let \((h:S\rightarrow T,\mc{N})\) be a deformation to \(S\) and \(I\) a finite \(S\)-module. If we can show that \(\xt{1}{T}{\mc{N}}{\mc{N}{\otimes} I}\) and \(\cH^{1}(S,T,T{\otimes} I)\) are finite \(S\)-modules, then condition (S2) holds by Proposition \ref{prop.grobs} and Proposition \ref{prop.lang}.
Let \(h^{\text{ft}}:S\rightarrow T^{\text{ft}}\) be a finite type representative of \(h:S\rightarrow T\). In particular \((T^{\text{ft}})^{\text{h}}_{\fr{m}}\cong T\) for a maximal ideal \(\fr{m}\) with \(k(\fr{m})\cong k\). Put \(h^{\text{ft}}_{0}=h^{\text{ft}}{\otimes}_{S}k:k\rightarrow A^{\text{ft}}\). We may assume that \(h_{0}^{\text{ft}}\) is smooth in the complement of \(\fr{m}_{0}=\fr{m}A^{\text{ft}}\). The non-smooth locus \(V(J)\) of \(h^{\text{ft}}\) is defined by an ideal \(J\subseteq T^{\text{ft}}\) such that the image \(J_{0}\) of \(J\) in \(A^{\text{ft}}\) is \(\fr{m}_{0}\)-primary. Put \(T_{N}^{\text{ft}}=T^{\text{ft}}/J^{N+1}\). Since \(A^{\text{ft}}/J_{0}^{N+1}\) has finite length, \(T_{N}^{\text{ft}}\) is finite over \(S\). By \cite[III 3.1.2]{ill:71} \(\cH^{i}(S,T^{\text{ft}},T^{\text{ft}}{\otimes} I)\) has support in \(V(J)\) for \(i>0\) and hence is finite over \(S\). Andr{\'e}-Quillen homology commutes with direct limits and the cotangent complex is trivial for {\'e}tale extensions. By \cite[III 21]{and:74} we get \(\cH^{i}(T^{\text{ft}},T,T{\otimes} I)\cong\hm{}{T}{\cH_{i}(T^{\text{ft}},T,T)}{T{\otimes} I}=0\) for all \(i\). From the transitivity sequence it follows that \(\cH^{i}(S, T,T{\otimes} I)\cong \cH^{i}(S, T^{\text{ft}},T^{\text{ft}}{\otimes} I){\otimes}_{T^{\text{ft}}}T\) for all \(i\). Put \(T_{N}=T_{N}^{\text{ft}} {\otimes}_{T^{\text{ft}}}T\). By \cite[18.5.10]{EGAIV4} \(T_{N}^{\text{ft}}\) is henselian. It follows that \(T_{N}\) is a quotient of \(T_{N}^{\text{ft}}\) and hence \(\cH^{1}(S, T,T{\otimes} I)\) is finite as \(S\)-module. Similarly (but easier) it follows that \(\xt{1}{T}{\mc{N}}{\mc{N}{\otimes} I}\) is finite as \(S\)-module.
For effectivity, let \(N^{\text{ft}}\) be a finite \(A^{\text{ft}}\)-module representing \(N\) such that \(N^{\text{ft}}\) is locally free at the smooth points. There is a deformation functor \(\df{}{(A^{\text{ft}},N^{\text{ft}})}:{{}_{k}\cat{H}}/k\rightarrow\Sets\) of base change maps of pairs \((S\rightarrow T^{\text{ft}},\mc{N}^{\text{ft}})\rightarrow (k\rightarrow A^{\text{ft}},N^{\text{ft}})\) where \(T^{\text{ft}}\) is a flat \(S\)-algebra of finite type and \(\mc{N}^{\text{ft}}\) is an \(S\)-flat finite \(T^{\text{ft}}\)-module. Base change is given by the standard tensor product. Similarly there is a \(\df{}{A^{\text{ft}}}:{{}_{k}\cat{H}}/k\rightarrow \Sets\). Restricted to \({{}_{k}\cat{A}}/k\) \(\df{}{(A^{\text{ft}},N^{\text{ft}})}\) satisfies (S1) and (S2). Let \(\{(T_{n}^{\text{ft}},\mc{N}^{\text{ft}}_{n})\}\in\varprojlim\df{}{(A^{\text{ft}},N^{\text{ft}})}(S_{n})\) be a formally versal formal element (where \(S_{n}=S/\fr{m}_{S}^{n+1}\)). Put \(S=\varprojlim S_{n}\). By \cite[Th\'{e}or\`{e}m 7, p.\ 595]{elk:73} (cf.\ \cite[II 5.1]{art:76}) there exists an element \(S\rightarrow T^{\text{ft}}\) in \(\df{}{A^{\text{ft}}}(S)\) which induces \(\{T_{n}^{\text{ft}}\}\). Let \(T'\) be the henselisation of \(T^{\text{ft}}\) in the maximal ideal \(\fr{m}=(T^{\text{ft}}{\rightarrow} A)^{{-}1}(\fr{m}_{A})\). Then \(S\rightarrow T'\) is a deformation of \(A\). Let \(T^{*}\) be the completion of \(T'\) at the ideal \(\fr{n}=\fr{m}_{S}T'\) and let \(\mc{N}^{*}=\varprojlim \mc{N}_{n}\). Then \(\mc{N}^{*}\) is an \(S\)-flat finite \(T^{*}\)-module. Let \(J^{*}\subseteq T^{*}\) denote the ideal \(I(\phi)\) where \(\phi\) is a minimal presentation of \(\mc{N}^{*}\). Then \(J^{*}\) defines the locus \(V(J^{*})\) where \(\mc{N}^{*}\) is not locally free. Let \(J=\ker(T'\rightarrow T^{*}{/}J^{*})\). Since \(T^{*}/J^{*}\) is finite as \(S=\hat{S}\)-module, \(T'/J\cong T^{*}/J^{*}\). The proof of \cite[2.3]{ess:90} works in this situation too (there is a typo in line 5: it should be a direct sum, not a tensor product) and shows that the completion of \(T'\) in the ideal \(\fr{a}=J\cap\fr{m}_{S}T'\) equals \(T^{*}\). Since \(\mc{N}^{*}\) is locally free on the complement of \(V(\fr{a}T^{*})\), the conditions in \cite[Th\'{e}or\`{e}m 3]{elk:73} hold. From this result we obtain a finite \(T'\)-module \(\mc{N}\) inducing \(\mc{N}^{*}\). In particular \(\mc{N}\) is \(S\)-flat.
We claim that the henselisation map \(\df{}{(A^{\text{ft}},N^{\text{ft}})}\rightarrow\df{}{(A,N)}\) is formally smooth. It follows that the element \((T',\mc{N})\) in \(\df{}{(A,N)}(S)\) is formally versal. For the claim, put \(\varGamma^{\text{ft}}=A^{\text{ft}}{\oplus}N^{\text{ft}}\) and \(\varGamma=A{\oplus} N\) and let \(\pi:S_{1}\rightarrow S_{0}=S_{1}/I\) be a small surjection in \({{}_{k}\cat{A}}/k\).
The obstruction \(\ob(\pi,\varGamma^{\text{ft}}_{0})\in\gH{0}^{2}(k,\varGamma^{\text{ft}},\varGamma^{\text{ft}}){\otimes}_{k}I\) for lifting a deformation \(\varGamma^{\text{ft}}_{0}\) of \(\varGamma^{\text{ft}}\) along \(\pi\) maps to the corresponding obstruction \(\ob(\pi,\varGamma_{0})\in\gH{0}^{2}(k,\varGamma,\varGamma){\otimes}_{k}I\).
The isomorphisms \(\cH^{i}(S, T,T)\cong \cH^{i}(S, T^{\text{ft}},T^{\text{ft}}){\otimes}_{T^{\text{ft}}}T\) for all \(i\) implies isomorphisms \(\gH{0}^{i}(k,\varGamma^{\text{ft}},\varGamma^{\text{ft}})\cong \gH{0}^{i}(k,\varGamma,\varGamma)\) for \(i=1,2\) as in the beginning of the proof. Smoothness follows by the standard obstruction argument. By \cite[3.2]{art:74} there is an algebraic \(k\)-algebra \(R\) and a formally versal element \((T,\mc{N})\) in \(\df{}{(A,N)}(R)\).
Finally we apply \cite[3.3]{art:74} to conclude that the formally versal element \((T,\mc{N})\) is versal. We already have (S1) and (S2). To check \cite[3.3(iii)]{art:74}, let \(S\) be algebraic in \({{}_{k}\cat{H}}/k\), \(I\) an ideal in \(S\) and put \(S^{*}=\varprojlim S_{n}\) where \(S_{n}=S/I^{n{+}1}\). Let \({}^{i}\xi=({}^{i}T,{}^{i}\mc{N})\) for \(i=1,2\) be two elements in \(\df{}{(A,N)}(S^{*})\) and \(\{\theta_{n}:{}^{1}\xi_{n}\cong{}^{2}\xi_{n}\}\) be a tower of isomorphisms between the \(S_{n}\)-truncations. There are finite type representatives \({}^{i}\xi^{\text{ft}}=({}^{i}T^{\text{ft}},{}^{i}\mc{N}^{\text{ft}})\) of the \({}^{i}\xi\). By the cohomology argument above one obtains by induction a tower of isomorphisms \(\{\theta_{n}^{\text{ft}}:{}^{1}\xi_{n}^{\text{ft}}\cong{}^{2}\xi_{n}^{\text{ft}}\}\) inducing \(\{\theta_{n}\}\). Since \(\varprojlim \hat{S}/I^{n{+}1}\hat{S}\cong S^{*}\) where \(\hat{S}\) is the completion of \(S\) in the maximal ideal, we can apply \cite[Lemme p.\ 600]{elk:73} to conclude that the henselisations of the \({}^{i}T^{\text{ft}}\) in \({}^{i}T^{\text{ft}}I\) are isomorphic by an isomorphism lifting \(\theta_{0}:{}^{1}T_{0}\cong{}^{2}T_{0}\).
Further henselisation in the maximal ideals gives an isomorphism of deformations \({}^{1}T\cong{}^{2}T\). By Lemma \ref{lem.Aapprox} the isomorphism is extended to an isomorphism of the pairs \(\psi:{}^{1}\xi\cong{}^{2}\xi\) which lifts \(\theta_{0}\). By \cite[1.3]{ess:90} condition \cite[3.3(ii)]{art:74} is unnecessary and we conclude that \((T,\mc{N})\) is versal.
\end{proof}
\begin{rem}\label{rem.isomod}
Let \(A\) be an Cohen-Macaulay local algebraic \(k\)-algebra and \(N\) a finite \(A\)-module. We say that \(N\) has an \emph{isolated singularity} if \(N_{\fr{p}}\) is a free \(A_{\fr{p}}\)-module for all prime ideals \(\fr{p}\neq\fr{m}_{A}\). In that case a similar, but easier argument gives that \(\df{A}{N}\) has a versal element. This is the result \cite[2.4]{ess:90} of von Essen, but for a slightly different fibred category of deformations where henselisation is taken along the closed fibre. However it implies the result in our case, essentially by henselisation at \(\fr{m}_{0}\). Corollary \ref{cor.ExVers} and \ref{cor.2dim} have obvious analogs for \(\df{A}{N}\) in this case.
\end{rem}
\begin{cor}\label{cor.ExVers}
Suppose \(A\) is an isolated Cohen-Macaulay singularity over the field \(k\) and \(N\) is a finite length \(A\)-module\textup{.} Let \(L\rightarrow M\rightarrow N\) and \(N\rightarrow L'\rightarrow M'\) be the minimal \(\cat{MCM}_{A}\)-approximation and \(\hat{\cat{D}}_{A}\)-hull of \(N\) respectively\textup{.}
Then\textup{:}
\begin{enumerate}
\item[(i)] \(\df{}{(A,N)}\) has a versal element\textup{.}
\item[(ii)] If \(\dim A\geq 2\) and \(Q'\) denotes \(\hm{}{A}{\omega_{A}}{L'}\) then
\begin{equation*}
\df{}{(A,N)}\cong\df{}{(A,L')}\cong\df{}{(A,Q')}\textup{.}
\end{equation*}
\item[(iii)] If \(\dim A\geq 3\) and \(Q\) denotes \(\hm{}{A}{\omega_{A}}{L}\) then
\begin{equation*}
\df{}{(A,N)}\cong\df{}{(A,L)}\cong\df{}{(A,Q)}\textup{.}
\end{equation*}
\end{enumerate}
\end{cor}
\begin{proof}
This is Theorem \ref{thm.ExVers}, Theorem \ref{thm.defgrade} and Proposition \ref{prop.defequiv}.
\end{proof}
\begin{cor}\label{cor.2dim}
Suppose \(A\) is a local algebraic \(k\)-algebra which is a Gorenstein normal domain with \(\dim A=2\) and \(N\) is a finite torsion-free \(A\)-module with an isolated singularity\textup{.} Let \(L\rightarrow M\rightarrow N\) be the minimal \(\cat{MCM}_{A}\)-approximation of \(N\)\textup{.} Assume \(k\) is perfect\textup{.} Then \(\df{}{(A,N)}\) and \(\df{}{(A,M)}\) both have versal elements and the map \(\df{}{(A,N)}\rightarrow\df{}{(A,M)}\) is smooth\textup{.}
\end{cor}
\begin{proof}
Since \(k\) is perfect it follows from \cite[6.7.7 and 6.8.6]{EGAIV2} that \((A,N)\) and \((A,M)\) are isolated singularities and hence Theorem \ref{thm.ExVers} applies. Proposition \ref{prop.Gor} gives smoothness since \(N\) torsion-free here is equivalent to \(N\) being a first syzygy.
\end{proof}
\begin{ex}
Indeed, if \(N\) is the syzygy module of a finite length module then it is torsion-free and locally free on the complement of the singularity.
\end{ex}
\section{Deforming maximal Cohen-Macaulay approximations\\ of Cohen-Macaulay modules}\label{sec.defCM}
Let \(h:S\rightarrow T\) be a homomorphism of noetherian rings and \(\mc{N}\) an \(S\)-flat finite \(T\)-module. If \(t\in\Spec T\) with image \(s\in\Spec S\) recall that \(\mc{N}_{\fr{p}_{t}}\) is the localisation of \(\mc{N}\) at the prime ideal \(t\) and \(\mc{N}(t)=\mc{N}_{\fr{p}_{t}}{\otimes}_{S_{\fr{p}_{s}}}k(s)\). An \emph{\((h,\mc{N})\)-sequence} (or just an \(h\)-sequence if \(\mc{N}=T\) and an \(h\)-regular element if \(n=1\)) is a sequence \(J=(f_{1},\dots,f_{n})\) in \(T\) such that the image \(\bar{J}\) in \(T(t)\) is a weak \(\mc{N}(t)\)-sequence for all closed \(t\) and such that \(J\mc{N}\neq\mc{N}\). Applying the Koszul complex \(K(J,\mc{N})\) as in Example \ref{ex.nakcplx} one sees that an \((h,\mc{N})\)-sequence is the same as a transversally \(\mc{N}\)-regular sequence relative to \(S\) as defined in \cite[19.2.1]{EGAIV4}.
In particular; \(J\) is an \((h,\mc{N})\)-sequence if and only if \(J\) is an \(\mc{N}\)-sequence and \(\mc{N}/J\mc{N}\) is \(S\)-flat.
\begin{thm}\label{thm.defMCM}
Let \(q:\mathcal{O}\rightarrow T^{\mspace{-1mu}\mathit{o}}\) denote the henselisation of a finite type Cohen-Macaulay map with \(T^{\mspace{-1mu}\mathit{o}}/\fr{m}_{T^{\mspace{-1mu}\mathit{o}}}=k\) and \(T^{\mspace{-1mu}\mathit{o}}{\otimes}_{\mathcal{O}}k=A\)\textup{.} Suppose \(J=(f_{1},\dots,f_{n})\) is a \(q\)-sequence\textup{.} Put \(\bar{T}^{\mspace{-1mu}\mathit{o}}=T^{\mspace{-1mu}\mathit{o}}/J\)\textup{,} \(B=\bar{T}^{\mspace{-1mu}\mathit{o}}{\otimes}_{\mathcal{O}}k\) and let \(\bar{J}\) be the image of \(J\) in \(A\)\textup{.}
Let \(N\) be a maximal Cohen-Macaulay \(B\)-module and \(L\rightarrow M\rightarrow N\) the minimal \(\cat{MCM}_{A}\)-approximation of \(N\)\textup{.}
If \(\ob{}{}(A/\bar{J}^{2}\rightarrow B,N)=0\)\textup{,} then the composition of maps
\begin{equation*}
\df{\bar{T}^{\mspace{-1mu}\mathit{o}}}{N}\longrightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{N} \longrightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{M}
\end{equation*}
of functors from \({{}_{\varLambda}\cat{H}}/k\) to \(\Sets\) is injective\textup{.}
\end{thm}
The existence of a splitting \(B\rightarrow A/\bar{J}^{2}\) implies that \(\ob{}{}(A/\bar{J}^{2},N)=0\)
for all \(B\)-modules \(N\) since \(A/\bar{J}^{2}{\otimes}_{B}N\) gives a lifting of \(N\) to \(A/\bar{J}^{2}\).
Let \(\cat{C}\) be a category. Then \(\Arr\cat{C}\) denotes the category with objects being arrows in \(\cat{C}\) and arrows being commutative diagrams of arrows in \(\cat{C}\). An endo-functor \(F\) on \(\cat{C}\) induces an endo-functor \(\Arr F\) on \(\Arr\cat{C}\). Let \(B\) be a local noetherian ring and \(\cat{P}_{\!B}\) the additive subcategory of projective modules in \(\cat{mod}_{B}\). Let \(\uhm{}{B}{N}{M}\) denote the homomorphisms from \(N\) to \(M\) in the quotient category \(\underline{\cat{mod}}_{B}=\cat{mod}_{B}/\cat{P}_{\!B}\) i.e.\ \(B\)-homomorphisms modulo the ones factoring through an object in \(\cat{P}_{\!B}\). For each \(N\) in \(\cat{mod}_{B}\) we fix a minimal \(B\)-free resolution and use it to define the syzygy modules of \(N\). Then the association \(N\mapsto\syz{B}{i}N\) induces an endo-functor on \(\underline{\cat{mod}}_{B}\) for each \(i\), considered by A.\ Heller \cite{hel:60}. Define \(\uxt{i}{B}{N}{M}\) as \(\uhm{}{B}{\syz{B}{i}N}{M}\) which turns out to be isomorphic to \(\xt{i}{B}{N}{M}\) for all \(i>0\).
\begin{lem}\label{lem.obssplit}
Let \(A\) be a local noetherian ring and \(I=(f_{1},\dots,f_{n})\) a regular sequence\textup{.} Put \(B=A/I\) and suppose \(N\)\textup{,} \(N_{1}\) and \(N_{2}\) are finite \(B\)-modules\textup{.} Let \(\bar{M}_{i}\) denote \(B{\otimes}\syz{A}{n}N_{i}\)\textup{.}
\begin{enumerate}
\item[(i)] There is an inclusion \(u_{N}:N\rightarrow \bar{M}_{N}\) of \(B\)-modules with \(\bar{M}_{N}\cong B{\otimes}\syz{A}{n}N\) which induces a functor \(u:\ul{\cat{mod}}_{B}\rightarrow\Arr\ul{\cat{mod}}_{B}\)\textup{.}
\item[(ii)] The functor \(u\) commutes with the \(B\)-syzygy functor\textup{:}
\begin{equation*}
\Arr\syz{B}{i}(u_{N})=u_{\syz{B}{i}N}
\end{equation*}
\item[(iii)] The endo-functor \(B{\otimes}\syz{A}{n}(-)\) induces a natural map \(\uxt{i}{B}{N_{1}}{N_{2}}\rightarrow\uxt{i}{B}{\bar{M}_{1}}{\bar{M}_{2}}\) which makes the following diagram commutative\textup{:}
\begin{equation*}
\xymatrix@-0pt@C-12pt@R-12pt@H-0pt{
\uxt{i}{B}{N_{1}}{N_{2}}\ar[rr]\ar[dr]_{(u_{N_{2}})_{*}} && \uxt{i}{B}{\bar{M}_{1}}{\bar{M}_{2}}\ar[dl]^{(u_{N_{1}})^{*}}\\
& \uxt{i}{B}{N_{1}}{\bar{M}_{2}} &
}
\end{equation*}
\item[(iv)] The inclusion \(u_{N}:N\hookrightarrow B{\otimes}\syz{A}{n}N\) splits \(\Longleftrightarrow\ob(A/J^{2}\rightarrow B,N)=0\).
\end{enumerate}
\end{lem}
\begin{rem}
Lemma \ref{lem.obssplit} (iv) strengthens \cite[3.6]{aus/din/sol:93} (in the commutative case).
\end{rem}
\begin{proof}
(i): Suppose \(F_{*}\rightarrow N\) is a minimal \(A\)-free resolution of \(N\). Tensoring the short exact sequence \(\syz{A}{n}N\xra{j} F_{n-1}\rightarrow\syz{A}{n{-}1}N\) with \(B\) gives the exact sequence
\begin{equation}
0\rightarrow\tor{A}{1}{B}{\syz{A}{n{-}1}N}\rightarrow B{\otimes}\syz{A}{n}N\rightarrow \bar{F}_{n-1}\rightarrow B{\otimes}\syz{A}{n{-}1}N\rightarrow 0\,.
\end{equation}
We have \(\tor{A}{1}{B}{\syz{A}{n{-}1}N}\cong \tor{A}{n}{B}{N}\cong N\). Let \(u_{N}\) be the inclusion \(N\cong\ker(B{\otimes}_{A}j)\rightarrow B{\otimes}\syz{A}{n}N\). Then \(N\mapsto u_{N}\) gives a functor of quotient categories.
(ii): Let \(p:Q\rightarrow N\) be the minimal \(B\)-free cover and \(P_{*}\rightarrow \syz{B}{}N\) the minimal \(A\)-free resolution of the \(B\)-syzygy \(\ker(p)\). Then there is an \(A\)-free resolution \(H_{*}\rightarrow Q\) which is an extension of \(F_{*}\) by \(P_{*}\). Since \(\syz{A}{n}B\cong A\), tensoring the short exact sequence of resolutions by \(B\) we obtain the commutative diagram with exact rows:
\begin{equation}
\xymatrix@C-4pt@R-12pt@H-0pt@M4pt{
0\ar[r] & \syz{B}{}N\ar@{_{(}->}[d]_{u_{\Syz N}}\ar[r] & Q\ar@<0.5ex>@{_{(}->}[d]\ar[r] & N\ar@<0.5ex>@{_{(}->}[d]_{u_{N}}\ar[r] & 0 \\
0\ar[r] & B{\otimes}\syz{A}{n}(\syz{B}{}N)\ar[r] & B^{r}{\oplus}Q\ar[r] & B{\otimes}\syz{A}{n}N\ar[r] & 0
}
\end{equation}
which proves the claim.
(iii): By (ii) it is enough to prove this for \(i=0\). The case \(i=0\) follows from the functoriality in (i).
(iv,\(\Leftarrow\)): For the case \(n=1\) see the proof of \cite[3.2]{aus/din/sol:93}. Assume \(n\geq 2\). We follow the proof of \cite[3.6]{aus/din/sol:93} closely. Let \(A_{1}=A/(f_{1})\). Then \(F_{*}^{(1)}=A_{1}{\otimes} F_{*\geq 1}[1]\) gives a minimal \(A_{1}\)-free resolution of \(A_{1}{\otimes}\syz{A}{}N\). We have \(\ob(A/J^{2}\rightarrow B,N)=0\Rightarrow \ob(A/(f_{1})^{2}\rightarrow A_{1},N)=0\) and hence \(N\) is a direct summand of \(A_{1}{\otimes}\syz{A}{}N\). Let \(G_{*}\rightarrow N\) be a minimal \(A_{1}\)-free resolution of \(N\). Then \(G_{*}\) is a direct summand of \(F_{*}^{(1)}\) and hence \(\syz{A_{1}}{n{-}1}N\) is a direct summand of \(\syz{A_{1}}{n{-}1}(A_{1}{\otimes}\syz{A}{}N)=A_{1}{\otimes} \syz{A}{n}N\). Tensoring this situation with \(B\) gives a commutative diagram:
\begin{equation}\label{eq.split}
\xymatrix@C-4pt@R-8pt@H-30pt{
N\ar[r]_(0.3){u}\ar@{=}[d] & B{\otimes}\syz{A}{n}N\ar[r]_(0.55){\bar{j}}\ar@<0.5ex>@{->>}[d] & \bar{F}_{n-1}\ar[r]\ar@<0.5ex>@{->>}[d] & \dots \ar[r] & \bar{F}_{1}\ar[r]\ar@<0.5ex>@{->>}[d] & \bar{F}_{0} \ar[r] & N \\
N\ar[r]^(0.3){u_{1}} & B{\otimes}\syz{A_{1}}{n{-}1}N\ar[r]^(0.6){\bar{j}_{1}}\ar@<0.5ex>[u] & \bar{G}_{n-2}\ar[r]\ar@<0.5ex>[u] & \dots\ar[r] & \bar{G}_{0}\ar[r]\ar@<0.5ex>[u] & N &
}
\end{equation}
Since \(\ob(A/J^{2}\rightarrow B,N)=0\Rightarrow \ob(A_{1}/(f_{2},\dots,f_{n})^{2}\rightarrow B,N)=0\) the map \(u_{1}\) splits by induction on \(n\). So \(u\) splits.
The other direction follows from \cite[3.6]{aus/din/sol:93}.
\end{proof}
\begin{prop}\label{prop.obssplit}
Let \(h:S\rightarrow T\) be a local Cohen-Macaulay map\textup{,} \(J=(f_{1},\dots,f_{n})\) an \(h\)-sequence\textup{,} \(\bar{h}:S\rightarrow\bar{T}=T/J\) the local Cohen-Macaulay map induced from \(h\)\textup{,} and \((\bar{h},\mc{N})\) an object in \(\cat{MCM}_{\bar{h}}\)\textup{.} Let \(\xi: \mc{L}\rightarrow \mc{M}\xra{\pi} \mc{N}\) be the minimal \(\cat{MCM}_{h}\)-approximation of \(\mc{N}\)\textup{.} Then tensoring \(\xi\) by \(\bar{T}\) gives a \(4\)-term exact sequence
\begin{equation*}
0\rightarrow \mc{N}{\otimes} J/J^{2}\longrightarrow \bar{\mc{L}}\longrightarrow \bar{\mc{M}}\xra{\,\,\bar{\pi}\,\,} \mc{N}\rightarrow 0
\end{equation*}
which represents the obstruction class \(\ob(T/J^{2}\rightarrow\bar{T},\mc{N})\in\xt{2}{\bar{T}}{\mc{N}}{\mc{N}{\otimes} J/J^{2}}\)\textup{.}
Moreover\textup{;} \(\ob(T/J^{2}\rightarrow\bar{T},\mc{N})=0 \Longleftrightarrow \ob(T/J^{2}\rightarrow\bar{T},\mc{N}^{\vee})=0 \Longleftrightarrow \bar{\pi}\) splits where \(\mc{N}^{\vee}=\xt{n}{T}{\mc{N}}{\omega_{h}}\)\textup{.}
\end{prop}
\begin{proof}
By Proposition \ref{prop.nakayama} \(\tor{T}{i}{\bar{T}}{\mc{M}}=\cH_{i}(K(\ul{f}){\otimes} \mc{M})=0\) for \(i>0\). There is a map from the defining short exact sequence \(\syz{T}{}\mc{N}\rightarrow F_{0}\rightarrow \mc{N}\) to \(\xi\) lifting \(\id_{\mc{N}}\). Tensoring with \(\bar{T}\) gives a map of \(4\)-term exact sequences with outer terms canonically identified. Hence they represent the same element \(\ob(T/J^{2}\rightarrow\bar{T},\mc{N})\) in \(\xt{2}{\bar{T}}{\mc{N}}{\mc{N}{\otimes} J/J^{2}}\).
By the argument in Remark \ref{rem.Buch} we can assume that \(\xi\) is given as \(\im(d_{n}^{\vee})\rightarrow(\syz{T}{n}\mc{N}^{\vee})^{\vee}\rightarrow \mc{N}^{\vee}{}^{\vee}\) where \((F_{*},d_{*})\) is a minimal \(T\)-free resolution of \(\mc{N}^{\vee}\). By Lemma \ref{lem.obssplit} \(\ob(T/J^{2}\rightarrow\bar{T},\mc{N}^{\vee})=0\) if and only if \(u:\mc{N}^{\vee}\rightarrow \bar{T}{\otimes}\syz{T}{n}\mc{N}^{\vee}\) splits. But applying \(\hm{}{\bar{T}}{-}{\omega_{\bar{h}}}\) to \(u\) gives \(\bar{\pi}\) since \(\mc{N}\cong\xt{n}{T}{\mc{N}^{\vee}}{\omega_{h}}\cong \hm{}{\bar{T}}{\mc{N}^{\vee}}{\omega_{\bar{h}}}\).
\end{proof}
\begin{rem}\label{rem.obssplit}
In the absolute Gorenstein case with \(n=1\) this is given in \cite[4.5]{aus/din/sol:93}.
\end{rem}
\begin{proof}[Proof of Theorem {\ref{thm.defMCM}}]
Given \(S\) in \({{}_{\varLambda}\cat{H}}/k\) and let \(h:S\rightarrow T\) and \(\bar{h}:S\rightarrow \bar{T}=T/JT\) be the induced hCM maps. Let \({}^{i\!}\mc{N}\) be deformations of \(N\) to \(\bar{h}\) for \(i=1,2\) and assume that the minimal \(\cat{MCM}_{h}\)-approximation modules \({}^{i\!}\mc{M}\) of \({}^{i\!}\mc{N}\) are isomorphic as deformations of \(M\). We proceed as in the proof of Theorem \ref{thm.defgrade} (i) with \(S_{n}=S/\fr{m}_{S}^{n{+}1}\), \(\mc{N}_{n}=\mc{N}{\otimes}_{S}S_{n}\) etc., construct a tower of isomorphisms \(\{\phi_{n}:{}^{1\!}\mc{N}\cong{}^{2\!}\mc{N}\}\), and conclude by Lemma \ref{lem.Aapprox} that \({}^{1\!}\mc{N}\) and \({}^{2\!}\mc{N}\) are isomorphic as deformations of \(N\). For the induction step we use that the map of torsor actions along \(\df{\bar{T}}{\mc{N}_{n}}(S_{n+1})\rightarrow\df{T}{\mc{M}_{n}}(S_{n+1})\) is induced by a natural map \(p:\xt{1}{B}{N}{N}\rightarrow \xt{1}{A}{M}{M}\) which is injective. The map \(p\) is given as follows.
Let \(\pi:M\rightarrow N\) denote the \(\cat{MCM}_{A}\)-approximation and \(\bar{\pi}:\bar{M}\rightarrow N\) be the \(B\)-quotient. Then \(\bar{\pi}\) splits by Proposition \ref{prop.obssplit}. Hence \(\bar{\pi}^{*}:\xt{1}{B}{N}{N} \hookrightarrow\xt{1}{B}{\bar{M}}{N}\) splits. Since \(J\) is an \(M\)-regular sequence, \(\xt{1}{B}{\bar{M}}{N}\cong\xt{1}{A}{M}{N}\). Since \(\xt{i}{A}{M}{L}=0\) for all \(i>0\), \(\pi_{*}:\xt{1}{A}{M}{M}\cong \xt{1}{A}{M}{N}\). Summarised:
\begin{equation}\label{eq.xt1}
\xymatrix@C-0pt@R-12pt@H-30pt{
& \xt{1}{A}{M}{N} & \xt{1}{A}{M}{M} \ar[l]_{\cong} \\
\xt{1}{B}{N}{N} \ar@<-0.5ex>@{^{(}->}[r]^{\bar{\pi}^{*}} & \xt{1}{B}{\bar{M}}{N} \ar[u]_{\cong} &
}
\end{equation}
\end{proof}
The technique used to prove Theorem \ref{thm.defMCM} also gives the following result.
\begin{thm}\label{thm.defsyz}
Let \(\mathcal{O}\) and \(T^{\mspace{-1mu}\mathit{o}}\) be henselian and noetherian local rings and \(q:\mathcal{O}\rightarrow T^{\mspace{-1mu}\mathit{o}}\) a local and flat ring homomorphism with \(T^{\mspace{-1mu}\mathit{o}}/\fr{m}_{T^{\mspace{-1mu}\mathit{o}}}=k\) and \(T^{\mspace{-1mu}\mathit{o}}{\otimes}_{\mathcal{O}}k=A\)\textup{.} Suppose \(J=(f_{1},\dots,f_{n})\) is a \(q\)-sequence\textup{.} Put \(\bar{T}^{\mspace{-1mu}\mathit{o}}=T^{\mspace{-1mu}\mathit{o}}/J\)\textup{,} \(B=\bar{T}^{\mspace{-1mu}\mathit{o}}{\otimes}_{\mathcal{O}}k\) and let \(\bar{J}\) be the image of \(J\) in \(A\)\textup{.} Let \(N\) be a finite \(B\)-module and let \(M\) denote the syzygy module \(\syz{A}{n}N\)\textup{.}
If \(\ob(A/\bar{J}^{2}\rightarrow B, N)=0\) then the natural map \(s:\df{\bar{T}^{\mspace{-1mu}\mathit{o}}}{N}\rightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{M}\) is injective\textup{.}
\end{thm}
\begin{proof}
We proceed as in the proof of Theorem \ref{thm.defgrade} and \ref{thm.defMCM}. Given deformations \({}^{i\!}\mc{N}\) of \(N\) to \(\bar{h}\) for \(i=1,2\). They map to \({}^{i\!}\mc{M}:=\syz{T}{n}({}^{i\!}\mc{N})\) which we suppose are isomorphic as deformations of \(M\) to \(h\).
Then the natural syzygy map \(s^{1}:\xt{1}{B}{N}{N}\rightarrow\xt{1}{A}{M}{M}\) induces the map of torsor actions along \(s\) of the inifinitesimal extensions. The composition of \(s^{1}\) with \(\xt{1}{A}{M}{M}\rightarrow\xt{1}{A}{M}{\bar{M}}\cong\xt{1}{B}{\bar{M}}{\bar{M}}\) commutes with the horizontal map in Lemma \ref{lem.obssplit} (iii). But Lemma \ref{lem.obssplit} (iv) implies that \((u_{N})_{*}\) is injective, hence \(s^{1}\) is injective too. Proceeding by induction on \(\fr{m}_{S}^{n+1}\)-truncations of the deformations we construct a tower of isomorphisms and conclude by Lemma \ref{lem.Aapprox}.
\end{proof}
\begin{rem}
Theorem \ref{thm.defsyz} resembles \cite[Thm.\ 1]{ile:07}. However Theorem \ref{thm.defsyz} makes a sounder statement in a more general setting and has a more transparent proof. Indeed, the various similar results in \cite{ile:07} can be changed and proved accordingly.
\end{rem}
\section{The Kodaira-Spencer map of Cohen-Macaulay approximations}
A modular family of objects is roughly speaking a family where the isomorphism class of the fibre changes non-trivially. The Kodaira-Spencer map makes this idea precise. We consider the Kodaira-Spencer classes and maps for families of pairs (algebra, module) and by invoking the long-exact transitivity sequence we relate them to the corresponding notions for the algebra and the module. Then we show that Cohen-Macaulay approximation of modular families under certain ``global'' conditions akin to those in Theorem \ref{thm.defgrade} and \ref{thm.defMCM} produce new modular families.
The following is a graded version of \cite[II 2.1.5.7]{ill:71}.
\begin{defn}\label{defn.KS}
Let \(\mathcal{O}\rightarrow S\) and \(S\rightarrow \varGamma\) be graded ring homomorphisms with \(\mathcal{O}\) and \(S\) concentrated in degree \(0\). The map \(L^{\text{gr}}_{\varGamma/S}\rightarrow L_{S/\mathcal{O}}{\otimes}_{S}\varGamma[1]\) in the corresponding distinguished transitivity triangle of (graded) cotangent complexes (see \eqref{eq.triangle}) is called the \emph{Kodaira-Spencer class} of \(\mathcal{O}\rightarrow S\rightarrow\varGamma\).
\end{defn}
Composing the Kodaira-Spencer class with the natural augmentation map
\begin{equation*}
L_{S/\mathcal{O}}{\otimes}_{S}\varGamma[1]\longrightarrow\Omega_{S/\mathcal{O}}{\otimes}_{S}\varGamma[1]
\end{equation*}
induces an element \(\kappa(\varGamma/S/\mathcal{O})\in\gH{0}^{1}(S,\varGamma,\Omega_{S/\mathcal{O}}{\otimes}_{S}\varGamma)\), the cohomological Kodaira-Spencer class, which is also given as follows. Let \(\mc{P}=\mc{P}_{S/\mathcal{O}}\) denote \(S{\otimes}_{\mathcal{O}}S/I^{2}\) where \(I\) is the kernel of the multiplication map \(S{\otimes}_{\mathcal{O}}S\rightarrow S\). There are two ring homomorphisms \(j_{1}\) and \(j_{2}\) from \(S\) to \(\mc{P}\) defined by \(j_{1}:s\mapsto s{\otimes} 1\) and \(j_{2}:s\mapsto 1{\otimes} s\). Let \(d_{S/\mathcal{O}}\) denote the universal derivation (induced by \(j_{2}-j_{1}\)). The principal parts of \(\varGamma\) is \(\mc{P}{\otimes}_{S}\varGamma\) (with the \(j_{2}\) tensor product), which gives an \(S\)-algebra extension (via \(j_{1}\)) representing the Kodaira-Spencer class:
\begin{equation}\label{eq.pp}
\kappa(\varGamma/S/\mathcal{O}):\quad0\rightarrow \Omega_{S/\mathcal{O}}{\otimes}_{S}\varGamma\longrightarrow \mc{P}_{S/\mathcal{O}}{\otimes}_{S}\varGamma\longrightarrow \varGamma \rightarrow 0
\end{equation}
see \cite[III 1.2.6]{ill:71}. Since \(\mc{P}{\otimes}_{S}\varGamma\) has a natural \(\mc{P}\)-algebra structure, \eqref{eq.pp} is also a (graded) algebra lifting of \(\varGamma\) along \(\mc{P}\rightarrow S\) as in Proposition \ref{prop.grobs}. The \(j_{1}\)-extension \(\varGamma{\otimes}_{S}\mc{P}\rightarrow \varGamma\) is a trivial lifting (split by \(\id_{\varGamma}{\otimes} 1_{\mc{P}} \)) and the difference in \(\gH{0}^{1}(S,\varGamma,\Omega_{S/\mathcal{O}}{\otimes}_{S}\varGamma)\) given by Proposition \ref{prop.grobs} (ii) equals \(\kappa(\varGamma/S/\mathcal{O})\), see \cite[III 2.1.5]{ill:71}.
Moreover, the difference \(1_{\mc{P}}{\otimes}\id_{\varGamma}-\id_{\varGamma}{\otimes} 1_{\mc{P}}\) induces \(d_{S/\mathcal{O}}{\otimes} 1_{T}\) (in degree \(0\)) which is mapped to \(\kappa(\varGamma/S/\mathcal{O})\) by the connecting homomorphism
\begin{equation}
\partial:\Der_{\mathcal{O}}(S,\Omega_{S/\mathcal{O}}{\otimes}_{S}T)\longrightarrow\gH{0}^{1}(S,\varGamma,\Omega_{S/\mathcal{O}}{\otimes}_{S}\varGamma)
\end{equation}
in the long-exact transitivity sequence, see \cite[III 1.2.6.5 and 1.2.7]{ill:71}.
In the special case of \(\varGamma=T{\oplus}\mc{N}\), \(\mc{N}\) a \(T\)-module and \(S=T\), the transitivity sequence of \(\mathcal{O}\rightarrow T\rightarrow\varGamma\) is given in Proposition \ref{prop.lang}. The Kodaira-Spencer class equals \(\partial(d_{T/\mathcal{O}})\in\xt{1}{T}{\mc{N}}{\Omega_{T/\mathcal{O}}{\otimes}_{T}\mc{N}}\) and is called the (cohomological) \emph{Atiyah class} and is denoted by \(\at_{T/\mathcal{O}}(\mc{N})\), cf.\ \cite[IV 2.3.6-7]{ill:71}. The class is represented by the short exact sequence
\begin{equation}
\at_{T/\mathcal{O}}(\mc{N}):\quad 0\rightarrow\Omega_{T/\mathcal{O}}{\otimes}_{T}\mc{N}\longrightarrow\mc{P}_{T/\mathcal{O}}{\otimes}_{T}\mc{N}\longrightarrow\mc{N}\rightarrow 0\,.
\end{equation}
The \emph{Kodaira-Spencer map of \(\mathcal{O}\rightarrow S\rightarrow\varGamma\)}
\begin{equation}\label{eq.KSmap}
g^{\varGamma}\!:\Der_{\mathcal{O}}(S)\longrightarrow\gH{0}^{1}(S,\varGamma,\varGamma)
\end{equation}
is defined by \(D\mapsto f^{D}_{*}\kappa(\varGamma/S/\mathcal{O})\) where \(f^{D}\!:\Omega_{S/\mathcal{O}}\rightarrow S\) corresponds to \(D\). Pushout of \eqref{eq.pp} by \(f^{D}{\otimes}\id_{\varGamma}\) gives the corresponding algebra lifting of \(\varGamma\) along \(S[\varepsilon]\rightarrow S\) given by \(g^{\varGamma}(D)\).
\begin{prop}\label{prop.at}
Let \(\varGamma\) denote the graded \(S\)-algebra \(T{\oplus}\mc{N}\) where \(\mathcal{O}\rightarrow S\) and \(S\rightarrow T\) are \textup{(}ungraded\textup{)} ring homomorphisms and \(\mc{N}\) is a \(T\)-module\textup{.} Consider the transitivity sequence of \(S\rightarrow T\xra{i} \varGamma\) in \textup{Proposition \ref{prop.lang}:}
\begin{align*}
\dots\rightarrow\Der_{S}(T,\Omega_{S/\mathcal{O}}{\otimes}_{S}T) & \xra{\partial}\xt{1}{T}{\mc{N}}{\Omega_{S/\mathcal{O}}{\otimes}_{S}\mc{N}}\xra{u}\gH{0}^{1}(S,\varGamma,\Omega_{S/\mathcal{O}}{\otimes}_{S}\varGamma)\xra{i^{*}}\\
\cH^{1}(S,T,\Omega_{S/\mathcal{O}}{\otimes}_{S}T) & \xra{\partial}\dots
\end{align*}
\begin{enumerate}
\item[(i)] The map \(i^{*}\) takes the Kodaira-Spencer class \(\kappa(\varGamma/S/\mathcal{O})\) to \(\kappa(T/S/\mathcal{O})\)\textup{.}
\item[(ii)] Assume \(\kappa(T/S/\mathcal{O})=0\) and choose an \(S\)-algebra splitting \(\sigma: T\rightarrow \mc{P}{\otimes}_{S}T\)\textup{.} Then there is a class \(\kappa(\sigma,\mc{N})=\kappa(T/S/\mathcal{O},\sigma,\mc{N})\in\xt{1}{T}{\mc{N}}{\Omega_{S/\mathcal{O}}{\otimes}_{S}\mc{N}}\) which maps to \(\kappa(\varGamma/S/\mathcal{O})\) by \(u\)\textup{.}
\item[(iii)] Let \(D(\sigma)\in\Der_{\mathcal{O}}(T,\Omega_{S/\mathcal{O}}{\otimes}_{S}T)\) be the derivation corresponding to the splitting \(\sigma\) and for each \(D_{1}\in\Der_{\mathcal{O}}(S)\) let \(X_{\sigma}(D_{1})\) denote \(f^{D_{1}}_{*}D(\sigma)\in\Der_{\mathcal{O}}(T)\)\textup{.} Then
\begin{equation*}
f^{D_{1}}_{*}\kappa(\sigma,\mc{N})=f^{X_{\sigma}(D_{1})}_{*}\at_{T/\mathcal{O}}(\mc{N})\quad\text{in}\quad \xt{1}{T}{\mc{N}}{\mc{N}}\,.
\end{equation*}
\end{enumerate}
\end{prop}
\begin{proof}
The degree zero part of \eqref{eq.pp} gives the image \(i^{*}\kappa(\varGamma/S/\mathcal{O})\) represented by the algebra extension
\begin{equation}
\kappa(T/S/\mathcal{O}):\quad 0\rightarrow\Omega_{S/\mathcal{O}}{\otimes}_{S} T\longrightarrow \mc{P}_{S/\mathcal{O}}{\otimes}_{S}T\longrightarrow T\rightarrow 0\,.
\end{equation}
The degree one part is the short exact sequence of \(\mc{P}{\otimes}_{S}T\)-modules
\begin{equation}\label{eq.pmod}
\tilde{\alpha}=\tilde{\alpha}(T/S/\mathcal{O},\mc{N}):\quad0\rightarrow \Omega_{S/\mathcal{O}}{\otimes}_{S}\mc{N}\longrightarrow\mc{P}_{S/\mathcal{O}}{\otimes}_{S}\mc{N}\longrightarrow\mc{N}\rightarrow 0\,.
\end{equation}
The splitting \(\sigma\) makes \(\tilde{\alpha}\) to a short exact sequence of \(T\)-modules which defines \(\kappa(T/S/\mathcal{O},\sigma,\mc{N})\).
For (iii) we have \(X_{\sigma}(D_{1})=f_{*}^{X_{\sigma}(D_{1})}d_{T/\mathcal{O}}\) and the result follows from the commutative diagram
\begin{equation}\label{eq.kappa}
\xymatrix@C-22pt@R-15pt@H-30pt{
D(\sigma)\in\Der_{\mathcal{O}}(T,\Omega_{S/\mathcal{O}}{\otimes}_{S}T)\ar@{.>}[rr]\ar[ddd]_{\partial}\ar[dr]_{f^{D_{1}}_{*}} && \Der_{\mathcal{O}}(T,\Omega_{T/\mathcal{O}})\ni d_{T/\mathcal{O}}\ar[dl]^{f^{X_{\sigma}(D_{1})}_{*}}\ar[ddd]^{\partial} \\
& \Der_{\mathcal{O}}(T,T)\ar[d]^{\partial} \\
& \xt{1}{T}{\mc{N}}{\mc{N}} \\
\kappa(\sigma,\mc{N})\in\xt{1}{T}{\mc{N}}{\Omega_{S/\mathcal{O}}{\otimes}_{S}\mc{N}}\ar[ur]^{f^{D_{1}}_{*}}\ar@{.>}[rr] && \xt{1}{T}{\mc{N}}{\Omega_{T/\mathcal{O}}{\otimes}_{T}\mc{N}}\ni\at_{T/\mathcal{O}}(\mc{N})\ar[lu]_{f^{X_{\sigma}(D_{1})}_{*}}
}
\end{equation}
where the two outer vertical maps are pointed.
\end{proof}
We call \(\kappa(\sigma,\mc{N})\) for the Kodaira-Spencer class of \((T/S/\mathcal{O},\sigma,\mc{N})\).
Define the \emph{Kodaira-Spencer map of \((T/S/\mathcal{O},\sigma,\mc{N})\)}
\begin{equation}\label{eq.modKS}
g^{(\sigma,\,\mc{N})}:\Der_{\mathcal{O}}(S)\longrightarrow\xt{1}{T}{\mc{N}}{\mc{N}}
\end{equation}
by \(g^{(\sigma,\,\mc{N})}(D):=(f^{D}{\otimes}\id)_{*}\kappa(\sigma,\mc{N})\).
In the case \(T=S{\otimes}_{\mathcal{O}} T^{\mspace{-1mu}\mathit{o}}\) we always choose the \(S\)-algebra splitting \(S{\otimes}_{\mathcal{O}} T^{\mspace{-1mu}\mathit{o}}\rightarrow \mc{P}_{S/\mathcal{O}}{\otimes}_{S}S{\otimes}_{\mathcal{O}}T^{\mspace{-1mu}\mathit{o}}\cong\mc{P}_{S/\mathcal{O}}{\otimes}_{\mathcal{O}}T^{\mspace{-1mu}\mathit{o}}\) given by \(s{\otimes} t\mapsto j_{1}(s){\otimes} t\). In particular \(\kappa(T/S/\mathcal{O})=0\) and we get a canonical Kodaira-Spencer class \(\kappa(\mc{N})\) and a corresponding Kodaira-Spencer map \(g^{\mc{N}}\).
\begin{rem}
There is no reason to believe that \(\kappa(\sigma,\mc{N})\) maps to \(\at_{T/\mathcal{O}}(\mc{N})\) in diagram \eqref{eq.kappa} for any choice of \(\sigma\). While there is a canonical map of short exact sequences (of \(S\)-modules)
\begin{equation*}
\xymatrix@C-0pt@R-12pt@H-30pt{
\;\;\;\kappa(\sigma,\mc{N}): & 0\ar[r] & \Omega_{S/\mathcal{O}}{\otimes}_{S}\mc{N}\ar[r]\ar[d] & \mc{P}_{S/\mathcal{O}}{\otimes}_{S}\mc{N}\ar[r]\ar[d]^{\tau} &\mc{N}\ar[r]\ar@{=}[d] & 0 \\
\at_{T/\mathcal{O}}(\mc{N}): & 0\ar[r] & \Omega_{T/\mathcal{O}}{\otimes}_{T}\mc{N}\ar[r] & \mc{P}_{T/\mathcal{O}}{\otimes}_{T}\mc{N}\ar[r] &\mc{N}\ar[r] & 0
}
\end{equation*}
\(\tau\) is in general not \(T\)-linear. However, in the case \(S\rightarrow T\) is smooth then \(\Omega_{S/\mathcal{O}}{\otimes} T\rightarrow \Omega_{T/\mathcal{O}}\rightarrow \Omega_{T/S}\) is split exact and hence there is a non-canonical lifting of the universal derivation in \(\Der_{\mathcal{O}}(T,\Omega_{T/\mathcal{O}})\) to \(\Der_{\mathcal{O}}(T,\Omega_{S/\mathcal{O}}{\otimes} T)\) and the corresponding choice of splitting \(\sigma\) makes \(\tau\) \(T\)-linear and so (or by \eqref{eq.kappa}) \(\kappa(\sigma,\mc{N})\) maps to \(\at_{T/\mathcal{O}}(\mc{N})\).
\end{rem}
\begin{ex}
Another special case is given by the base change of \(h:S\rightarrow T\) with itself to \(h{\otimes} T: T\rightarrow T{\otimes}_{S}T=T^{\ot2}\) and a \(T\)-flat \(T^{\ot2}\)-module \(\mc{N}\), cf.\ Section \ref{sec.fund}. Then \(\kappa(T^{\otimes 2}/T/S,\mc{N})\) in \(\xt{1}{T^{\ot2}}{\mc{N}}{\Omega_{T^{\ot2}/T}{\otimes}\mc{N}}\) equals \(\at_{T^{\ot2}/T}(\mc{N})\). The multiplication map \(\mu_{T^{\ot2}/T}:\mc{P}_{T^{\ot2}/T}\rightarrow T^{\ot2}\) equals \(\id_{T}{\otimes}\mu_{T/S}:T{\otimes}_{S}\mc{P}_{T/S}\rightarrow T^{\ot2}\). It follows that \(\at_{T^{\ot2}/T}(\mc{N})\) maps to \(\at_{T/S}(\mc{N})\) in \(\xt{1}{T}{\mc{N}}{\Omega_{T/S}{\otimes}\mc{N}}\) by the natural map. If \(\mc{N}=T\) then \(\at_{T/S}(T)=0\), but in general \(\at_{T^{\ot2}/T}(T)\neq0\).
\end{ex}
\begin{ex}
The transitivity sequence of \(\mathcal{O}\rightarrow T\xra{i}\varGamma\) and \(J\) with \(\varGamma=T{\oplus}\mc{N}\) and \(J=J_{0}{\oplus}J_{1}\) in Proposition \ref{prop.lang}
\begin{equation}\label{eq.deg0}
0\rightarrow \hm{}{T}{\mc{N}}{J_{1}}\rightarrow \gDer{0}_{\mathcal{O}}(\varGamma,J)\xra{i^{*}} \Der_{\mathcal{O}}(T,J_{0})\rightarrow \xt{1}{T}{\mc{N}}{J_{1}}\rightarrow\dots
\end{equation}
suggests the following characterisation. An element \(\mc{D}\in\gDer{0}_{\mathcal{O}}(\varGamma,J)\) is given by its degree \(0\) restriction \(D:=i^{*}(\mc{D})\in\Der_{\mathcal{O}}(T,J_{0})\) and its degree \(1\) restriction \(\nabla_{\!D}:=\mc{D}_{\vert\mc{N}}\in\hm{}{\mathcal{O}}{\mc{N}}{J_{1}}\) which should satisfy the following Leibniz rule: For all \(t\) in \(T\) and \(n\) in \(\mc{N}\)
\begin{equation}\label{eq.Leib}
\nabla_{\!D}(tn)=t\nabla_{\!D}(n)+D(t)n\,.
\end{equation}
With notation as in Proposition \ref{prop.at} recall
that \(\kappa(\varGamma/S/\mathcal{O})=\partial(d_{S/\mathcal{O}}{\otimes} 1_{T})\) in the transitivity sequence of \(\mathcal{O}\rightarrow S\rightarrow\varGamma\):
\begin{equation}\label{eq.deg0b}
\begin{split}
0\rightarrow{} &\gDer{0}_{S}(\varGamma,\Omega_{S/\mathcal{O}}{\otimes}\varGamma)\rightarrow\gDer{0}_{\mathcal{O}}(\varGamma,\Omega_{S/\mathcal{O}}{\otimes}\varGamma)\rightarrow\Der_{\mathcal{O}}(S,\Omega_{S/\mathcal{O}}{\otimes} T)\xra{\partial} \\
&\gH{0}^{1}(S,\varGamma,\Omega_{S/\mathcal{O}}{\otimes}\varGamma)\rightarrow\dots
\end{split}
\end{equation}
Hence \(\kappa(\varGamma/S/\mathcal{O})=0\) if and only if there exists a \(D\in\Der_{\mathcal{O}}(T,\Omega_{S/\mathcal{O}}{\otimes} T)\) which restricts to \(d_{S/\mathcal{O}}{\otimes} 1_{T}\) and a \(\nabla_{\!D} \in\hm{}{\mathcal{O}}{\mc{N}}{\Omega_{S/\mathcal{O}}{\otimes} \mc{N}}\) satisfying \eqref{eq.Leib}. As a well known special case (\(S=T\)) we get \(\at_{T/\mathcal{O}}(\mc{N})=0\) if and only if there exists a \(\nabla\in \hm{}{\mathcal{O}}{\mc{N}}{\Omega_{T/\mathcal{O}}{\otimes} \mc{N}}\) satisfying \eqref{eq.Leib} with \(D=d_{T/\mathcal{O}}\in\Der_{\mathcal{O}}(T,\Omega_{T/\mathcal{O}})\) (i.e.\ \(\nabla\) is a connection), or equivalently, a graded derivation \(\mc{D}\in \gDer{0}_{\mathcal{O}}(\varGamma,\Omega_{T/\mathcal{O}}{\otimes}_{T}\varGamma)\) restricting to \(d_{T/\mathcal{O}}\). Note that \eqref{eq.deg0} with \(J=\Omega_{T/\mathcal{O}}{\otimes}_{T}\varGamma\) equals \eqref{eq.deg0b} in this case.
\end{ex}
Recall the maps of cohomology groups \(\sigma^{1}_{j}(I)\) and \(\tau^{1}_{j}(I)\) in \eqref{eq.sigma} and \eqref{eq.tau}.
\begin{prop}\label{prop.KS}
In addition to the assumptions in \textup{Lemma \ref{lem.cohmap}} suppose \(\mathcal{O}\rightarrow S\) is a ring homomorphism\textup{.} For \(j=1,2\) the following holds\textup{:}
\begin{enumerate}
\item[(i)] The map \(\sigma^{1}_{j}(\Omega_{S/\mathcal{O}})\) takes \(\kappa(\varGamma_{0}/S/\mathcal{O})\) to \(\kappa(\varGamma_{j}/S/\mathcal{O})\) and the Kodaira-Spencer maps \(g^{\varGamma_{i}}:\Der_{\mathcal{O}}(S)\rightarrow\gH{0}^{1}(S,\varGamma_{i},\varGamma_{i})\) commute with \(\sigma^{1}_{j}\)\textup{,} i\textup{.}e\textup{.}\ \(\sigma^{1}_{j} g^{\varGamma_{0}}=g^{\varGamma_{j}}\)\textup{.}
\item[(ii)] Assume \(\kappa(T/S/\mathcal{O})=0\) and choose an \(S\)-algebra splitting \(\sigma: T\rightarrow \mc{P}{\otimes}_{S}T\)\textup{.}
Then \(\tau^{1}_{j}(\Omega_{S/\mathcal{O}})\) maps \(\kappa(\sigma,\mc{N})\) to \(\kappa(\sigma,X_{j})\) and the Kodaira-Spencer maps \(g^{(\sigma,X_{i})}:\Der_{\mathcal{O}}(S)\rightarrow\xt{1}{T}{X_{i}}{X_{i}}\) commute with \(\tau^{1}_{j}\), i.e.\ \(\tau^{1}_{j} g^{(\sigma,\mc{N})}\!=g^{(\sigma,X_{j})}\)\textup{.}
\end{enumerate}
\end{prop}
\begin{proof}
(i): Put \(\kappa_{j}=\kappa(\varGamma_{j}/S/\mathcal{O})\), \(\Omega=\Omega_{S/\mathcal{O}}\) and let \(\varGamma(\iota):\varGamma_{0}\rightarrow\varGamma_{2}\) denote the graded ring homomorphism induced from \(\iota\). Then \(\varGamma(\iota)\) induces a map of short exact sequences \(\kappa_{0}\rightarrow\kappa_{2}\), hence a map of short exact sequences \(\varGamma(\iota)_{*}\kappa_{0}\rightarrow\varGamma(\iota)^{*}\kappa_{2}\), i.e.\ \(\sigma_{2}(\Omega)(\kappa_{0})=(\varGamma(\iota)^{*})^{-1}\varGamma(\iota)_{*}\kappa_{0}=\kappa_{2}\). The maps \(\sigma_{2}(\Omega)\) and \(\sigma_{2}(S)\) commute with the covariant action of \(\Der_{\mathcal{O}}(S)\), hence the second assertion follows from the first. The arguments for the cases \(j=1\) and (ii) are similar.
\end{proof}
There are corresponding \emph{local} Kodaira-Spencer maps given as follows. Let \(t\in\Spec T\) map to \(s\in\Spec S\) and consider the localisations \(S_{\fr{p}_{s}}\rightarrow T_{\fr{p}_{t}}\) and \(S_{\fr{p}_{s}}\rightarrow\varGamma_{\fr{p}_{t}}\) and the induced map \(\mathcal{O}\rightarrow S_{\fr{p}_{s}}\). The localisation map \(\gH{0}^{1}(S,\varGamma,\Omega_{S/\mathcal{O}}{\otimes}_{S}\varGamma)\rightarrow\gH{0}^{1}(S_{\fr{p}_{s}},\varGamma_{\fr{p}_{t}},\Omega_{S_{\fr{p}_{s}}/\mathcal{O}}{\otimes}_{S_{\fr{p}_{s}}}\varGamma_{\fr{p}_{t}})\) maps \(\kappa(\varGamma/S/\mathcal{O})\) to \(\kappa(\varGamma_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\mathcal{O})\). Let \(\bar{\kappa}(\varGamma_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\mathcal{O})\) denote the image in \(\gH{0}^{1}(S_{\fr{p}_{s}},\varGamma_{\fr{p}_{t}},\Omega_{S_{\fr{p}_{s}}/\mathcal{O}}{\otimes}_{S_{\fr{p}_{s}}}\varGamma(t))\) by the map induced from \(\varGamma_{\fr{p}_{t}}\rightarrow\varGamma_{\fr{p}_{t}}/\fr{p}_{t}\varGamma_{\fr{p}_{t}}=\varGamma(t)\). Assume that \(\varGamma_{\fr{p}_{t}}\) is \(S_{\fr{p}_{s}}\)-flat. Then the natural base change map
\begin{equation}
\gH{0}^{1}(k(s),\varGamma(t),\Omega_{S_{\fr{p}_{s}}/\mathcal{O}}{\otimes}_{S_{\fr{p}_{s}}}\varGamma(t))\longrightarrow\gH{0}^{1}(S_{\fr{p}_{s}},\varGamma_{\fr{p}_{t}},\Omega_{S_{\fr{p}_{s}}/\mathcal{O}}{\otimes}_{S_{\fr{p}_{s}}}\varGamma(t))
\end{equation}
is an isomorphism, see \cite[II 2.2]{ill:71}. With this identification we define \(g^{\varGamma}\!(t)(D):=(f^{D}{\otimes} \id_{\varGamma(t)})_{*}\bar{\kappa}(\varGamma_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\mathcal{O})\) for any \(D\) in \(\Der_{{\mathcal{O}}}(S_{\fr{p}_{s}},k(s))\) and obtain local Kodaira-Spencer maps at \(t\) of \(\varGamma\), and (similarly) of \(T\), respectively:
\begin{align}
g^{\varGamma}\!(t):\,\,&\Der_{\mathcal{O}}(S_{\fr{p}_{s}},k(s))\longrightarrow\gH{0}^{1}(k(s),\varGamma(t),\varGamma(t)) \\
g^{T}\!(t):\,\,&\Der_{\mathcal{O}}(S_{\fr{p}_{s}},k(s))\longrightarrow\cH^{1}(k(s),T(t),T(t))
\end{align}
commuting with the natural map \(\gH{0}^{1}(k(s),\varGamma(t),\varGamma(t))\rightarrow \cH^{1}(k(s),T(t),T(t))\) in Proposition \ref{prop.lang}.
Note that if \(\mathcal{O}\) is an algebraically closed field then \(\Der_{\mathcal{O}}(S_{\fr{p}_{s}},k(s))\) is canonically isomorphic to the Zariski tangent space at any closed point \(s\in\Spec S\).
Let \(\mc{P}\) and \(\Omega\) denote \(\mc{P}_{S_{\fr{p}_{s}}/\mathcal{O}}\) and \(\Omega_{S_{\fr{p}_{s}}/\mathcal{O}}\) respectively. Assume \(\varGamma=T{\oplus}\mc{N}\). As in the global case we get a graded algebra extension representing \(\bar{\kappa}(\varGamma_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\mathcal{O})\) which in degree \(1\) is a short exact sequence \(\tilde{\alpha}(t): \Omega{\otimes}\mc{N}(t)\rightarrow k(s){\otimes}\mc{P}{\otimes}\mc{N}_{\fr{p}_{t}}\rightarrow\mc{N}(t)\).
If \(\bar{\kappa}(T_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\mathcal{O})=0\) we choose an \(S\)-algebra splitting \(\sigma:T(t)\rightarrow k(s){\otimes}\mc{P}{\otimes}\varGamma_{\fr{p}_{t}}\) and the local Kodaira-Spencer class \(\bar{\kappa}(\sigma,\mc{N}_{\fr{p}_{t}})\) in \(\xt{1}{T(t)}{\mc{N}(t)}{\Omega{\otimes}\mc{N}(t)}\) is represented by the obtained short exact sequence of \(T(t)\)-modules. Then we define the local Kodaira-Specer map of \((T_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\mathcal{O},\sigma,\mc{N}_{\fr{p}_{t}})\)
\begin{equation}
g^{(\sigma,\mc{N})}\!(t):\,\,\Der_{\mathcal{O}}(S_{\fr{p}_{s}},k(s))\longrightarrow\xt{1}{T(t)}{\mc{N}(t)}{\mc{N}(t)}
\end{equation}
by \(g^{(\sigma,\mc{N})}(t)(D):=(f^{D}{\otimes}\id)_{*}\bar{\kappa}(\sigma,\mc{N}_{\fr{p}_{t}})\).
Similarly, the class \(g^{\varGamma}\!(t)(D)\) is represented by a lifting of graded algebras \(\varGamma'\rightarrow \varGamma(t)\) along \(k(s)[\varepsilon]\rightarrow k(s)\). If the lifting \(g^{T}\!(t)(D):\, T'\rightarrow T(t)\) splits, a choice of splitting makes the short exact sequence \(\tilde{\alpha}(t)(D):\varepsilon\mc{N}(t)\rightarrow N'\rightarrow\mc{N}(t)\) \(T\)-linear and defines \(\tilde{\alpha}(t)(D)\) as an extension in the subspace \(\xt{1}{T(t)}{\mc{N}(t)}{\mc{N}(t)}\) of\, \(\xt{1}{T'}{\mc{N}(t)}{\mc{N}(t)}\).
We assume that \(\mathcal{O}\) is an algebraically closed field \(k\) for the rest of this section.
\begin{defn}\label{defn.modular}
Let \(h:S\rightarrow T\) be a local flat map of noetherian \(k\)-algebras and \(\mc{N}\) an \(S\)-flat \(T\)-module. Put \(\varGamma=T{\oplus}\mc{N}\), \(A=T{\otimes}_{S}k\), \(N=\mc{N}{\otimes}_{S}k\) and \(\varGamma(0)=\varGamma{\otimes}_{S}k=A{\oplus}N\).
We say that \((h,\mc{N})\) is \emph{locally modular} if the local Kodaira-Spencer map \(g^{\varGamma}\!(0):\Der_{k}(S,k)\rightarrow\gH{0}^{1}(k,\varGamma(0),\varGamma(0))\) is injective. If in addition \(T=S{\tilde{\otimes}}_{k}A\) then \(\mc{N}\) is locally modular if \(g^{\mc{N}}\!(0):\Der_{k}(S,k)\rightarrow\xt{1}{A}{N}{N}\) is injective.
If \(h^{\text{ft}}:S\rightarrow T\) is a faithfully flat finite type map of noetherian \(k\)-algebras with a \(k\)-point \(t\in\Spec T\) mapping to \(s\in \Spec S\) and \(\mc{N}\) is an \(S\)-flat finite \(T\)-module, we say that \((h^{\textnormal{ft}},\mc{N})\) is \emph{modular at \(t\)} if the henselisation of \((h^{\text{ft}},\mc{N})\) at \(t\) is locally modular.
If \(A\) is a finite type \(k\)-algebra and \(T=S{\otimes}_{k} A\), then \(\mc{N}\) is modular at \(t\) as \(T\)-module if its henselisation at \(t\) is locally modular. Let \(\nabla(h,\mc{N})\) (\(\nabla_{T}(\mc{N})\) if \(T=S{\otimes}_{k}A\)) denote the set of \(k\)-points \(t\in \Supp\mc{N}\) where \((h,\mc{N})\) (respectively \(\mc{N}\) as \(T\)-module) is modular.
\end{defn}
\begin{cor}\label{cor.modY}
Let \(h:S\rightarrow T\) be a finite type Cohen-Macaulay map of \(k\)-algebras and let \(\mc{N}\) be in \(\cat{mod}{}^{\textnormal{fl}}_{h}\)\textup{.} Suppose there is an \(h\)-regular element contained in \(\ann_{T}\mc{N}\)\textup{.} Let \(\iota:\,\mc{N}\rightarrow\mc{L}'\rightarrow\mc{M}'\) be a \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\)-hull for \(\mc{N}\)\textup{.} Then
\begin{equation*}
\nabla(h,\mc{N})=\nabla(h,\mc{L}')\quad\text{and}\quad \nabla_{T}(\mc{N})=\nabla_{T}(\mc{L}')\,\,\text{if}\,\, T=S{\otimes}_{k}A\,.
\end{equation*}
\end{cor}
\begin{proof}
Let \(t\) be a \(k\)-point in \(\Spec T\). By Theorem \ref{thm.flatCMapprox} \(\iota(t)\) is a \(\hat{\cat{D}}_{T(t)}\)-hull for \(\mc{N}(t)\) and by Proposition \ref{prop.minapprox} \(\iota(t)\) is minimal if and only if \(\iota_{\fr{p}_{t}}\) is minimal. In particular; the minimal hull of
\(\mc{N}_{\fr{p}_{t}}\) is a direct summand of \(\mc{L}'_{\fr{p}_{t}}\). We therefore assume that \(\iota_{\fr{p}_{t}}\) and hence \(\iota(t)\) is minimal.
By the grade condition on \(\mc{N}(t)\), the map \(\tau^{1}_{2}\) in \eqref{eq.tau} is injective. It implies that the map \(\sigma^{1}_{2}:\gH{0}^{1}(k(s),\varGamma_{0}(t),\varGamma_{0}(t))\rightarrow\gH{0}^{1}(k(s),\varGamma_{2}(t),\varGamma_{2}(t))\) in \eqref{eq.sigma} is injective and Proposition \ref{prop.KS} gives the statement.
\end{proof}
\begin{ex}\label{ex.modY}
Let \(A\) be a CM finite type \(k\)-algebra and domain of dimension \(\geq 2\). Let \(h:A\rightarrow T=A^{{\otimes} 2}\) be the base change by \(S=A\) and \(\mc{N}=A\) be the \(A\)-flat \(T\)-module defined by the multiplication map \(T\rightarrow A\). Let \(\Delta\subseteq \Spec T\) denote the closed points on the diagonal and let \(t\) be a closed point in \(\Spec T\) mapping to \(s\) in \(\Spec A\). If \(t\notin\Delta\) then \(\mc{N}(t)=0\). If \(t\in \Delta\) then \(T(t)\cong A_{\fr{p}_{s}}\) and \(\mc{N}(t)\cong k(s)\). The local Kodaira-Spencer class \(\bar{\kappa}(\mc{N}(t))\in \xt{1}{A_{\fr{p}_{s}}}{k(s)}{\Omega_{A_{\fr{p}_{s}}/k}{\otimes} k(s)}\) is represented by
\begin{equation}
\xymatrix@C-0pt@R-12pt@H-30pt{
0\ar[r] & \Omega_{A_{\fr{m}}/k}{\otimes} k(s)\ar[r] & k(s){\otimes} \mc{P}_{A_{\fr{m}}/k}\ar[r] & k(s)\ar[r] & 0 \\
0\ar[r] & \fr{m}/\fr{m}^{2}\ar[r]\ar[u]^{\cong}_{\delta} & A/\fr{m}^{2}\ar[r] \ar[u]^{\cong}_{\chi} & k(s)\ar@{=}[u]\ar[r] & 0
}
\end{equation}
(with \(\fr{m}=\fr{p}_{s}\)) where \(\delta(\bar{x})=d_{A_{\fr{m}}/k}(x){\otimes} 1\) and \(\chi\) is induced by \(1{\otimes} j_{2}\) (note that if \(x, y\in\fr{m}\) then \(1{\otimes} j_{2}(xy)=1{\otimes}([j_{2}(x)-j_{1}(x)][(j_{2}(y)-j_{1}(y)])\in k(s){\otimes} I^{ 2}\)). The local Kodaira-Spencer map \(g^{\mc{N}}(t)\) is given by the pushout
\begin{equation}
\phi\in\hm{}{k(s)}{\fr{m}/\fr{m}^{2}}{k(s)}\longrightarrow \xt{1}{A_{\fr{m}}}{k(s)}{k(s)}\ni\phi_{*}\bar{\kappa}(\mc{N}(t))
\end{equation}
which is an isomorphism. By Corollary \ref{cor.modY} we have \(\nabla_{T}(\mc{N})=\nabla_{T}(\mc{L}')=\Delta\). Put \(\mc{Q}'=\hm{}{T}{\omega_{h}}{\mc{L}'}\). By Proposition \ref{prop.defequiv} also the local Kodaira-Spencer map \(g^{\mc{Q}'}(t)\) is injective for \(t\in \Delta\). Hence \(\nabla_{T}(\mc{Q}')=\Delta\). Note that \(\mc{L}'(t)\) and \(\mc{Q}'(t)\) are rigid for \(t\notin\Delta\).
\end{ex}
\begin{cor}\label{cor.modCM}
Suppose \(T=S{\otimes}_{k}A\) for a finite type Cohen-Macaulay \(k\)-algebra \(A\)\textup{.} Let \(J=(f_{1},\dots,f_{n})\) be an \(A\)-sequence\textup{,} put \(B=A/J\) and let \(\bar{h}:S\rightarrow\bar{T}=S{\otimes}_{k}B\) be the induced Cohen-Macaulay map\textup{.} Suppose \(\mc{N}\) is in \(\cat{MCM}_{\bar{h}}\subseteq\cat{mod}{}^{\textnormal{fl}}_{h}\) and let \(\mc{L}\rightarrow\mc{M}\rightarrow\mc{N}\) be an \(\cat{MCM}_{h}\)-approximation of \(\mc{N}\)\textup{.} Assume \(\ob(T/(JT)^{2}\rightarrow \bar{T}, \mc{N})=0\)\textup{.} Then
\begin{equation*}
\nabla_{\bar{T}}(\mc{N})=\nabla_{T}(\mc{M})\cap\Supp\bar{T}\,.
\end{equation*}
\end{cor}
\begin{proof}
By Proposition \ref{prop.obsmodule} there is a lifting \(\mc{N}_{1}\rightarrow\mc{N}\) of \(\mc{N}\) to \(T_{1}=T/(JT)^{2}\). It induces liftings \(\mc{N}_{1}(t)\rightarrow\mc{N}(t)\) for all \(k\)-points \(t=(s,\fr{m})\) in \(\Supp\bar{T}\). The inclusion \(\bar{\tau}:\xt{1}{B_{\bar{\fr{m}}}}{\mc{N}(t)}{\mc{N}(t)}\rightarrow\xt{1}{A_{\fr{m}}}{\mc{M}(t)}{\mc{M}(t)}\) in \eqref{eq.xt1} commutes with the local Kodaira-Spencer maps. Proceed as in the proof of Corollary \ref{cor.modY}.
\end{proof}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
\section{Introduction}\label{sec.intro}
Axiomatic Cohen-Macaulay approximation was introduced by M. Auslander and R.-O.\ Buchweitz in \cite{aus/buc:89}. We define this theory in terms of fibred categories and obtain approximation results for various classes of flat families of modules.
Let \(A\) be a Cohen-Macaulay ring of finite Krull dimension with a canonical module \(\omega_{A}\). Let \(\cat{MCM}_{A}\) and \(\cat{FID}_{A}\) denote the categories of maximal Cohen-Macaulay modules and of finite modules with finite injective dimension, respectively. M.\ Auslander and R.-O.\ Buchweitz proved in \cite{aus/buc:89} that for any finite \(A\)-module \(N\) there exists short exact sequences
\begin{equation}\label{eq.MCMseq}
0\rightarrow L\longrightarrow M\longrightarrow N\rightarrow 0\qquad\text{and}\qquad 0\rightarrow N\longrightarrow L'\longrightarrow M'\rightarrow 0
\end{equation}
with \(M\) and \(M'\) in \(\cat{MCM}_{A}\) and \(L\) and \(L'\) in \(\cat{FID}_{A}\). The maps \(M\rightarrow N\) and \(N\rightarrow L'\) in \eqref{eq.MCMseq} are called a maximal Cohen-Macaulay approximation and a hull of finite injective dimension, respectively, of the module \(N\).
The association \(N\mapsto X\) for \(X\) equal to \(M,M',L\) and \(L'\) define functors of corresponding stable categories.
In this article we study the continuous properties of these functors.
Linear representations provided by (sheaves) of modules and the associated homological algebra plays an important role in algebra and algebraic geometry, e.g.\ as a means for classification by providing invariants.
Finite complexes have particular properties as seen in the Buchsbaum-Eisenbud acyclicity criterion and the intersection theorems of Peskine, Szpiro and Roberts. However, for a non-regular local ring \(A\), the standard homological invariants are given by the (generally) infinite minimal \(A\)-free resolutions, of which very little is known. To stay within finite complexes one can enlarge or change the category of resolving objects and Cohen-Macaulay approximation is a structured way of doing this.
Let \(\cat{D}_{A}\) denote the subcategory \(\Add\{\omega_{A}\}\) of modules \(D\) isomorphic to direct summands of the \(\omega_{A}^{\oplus r}\). A part of the approximation result says that all the modules in \(\cat{FID}_{A}\) have finite resolutions by objects in \(\cat{D}_{A}\). In particular the MCM approximation in \eqref{eq.MCMseq} can be extended to a finite resolution
\begin{equation}\label{eq.Dres}
0\rightarrow D^{-n}\longrightarrow D^{-n+1}\longrightarrow\dots \longrightarrow D^{-1}\longrightarrow M\longrightarrow N\rightarrow 0
\end{equation}
with the \(D^{i}\) in \(\cat{D}_{A}\). In the case \(A\) is Gorenstein, \(\cat{D}_{A}\) equals the category of finite projective modules \(\cat{P}_{\!A}\). This generalises: By a result of R.\ Y.\ Sharp \cite{sha:75b} the functor \(\hm{}{A}{\omega_{A}}{-}\) gives an exact equivalence \(\cat{D}_{A}\simeq\cat{P}_{\!A}\), hence a finite projective resolution is associated to \(N\). In the case \(A\) is local, the approximations and the complex can be chosen to be minimal and unique (with \(D^{i}\cong \omega_{A}^{\oplus d^{i}}\)) and in particular the \(d^{i}\) are invariants of \(N\).
The developments since Auslander and Buchweitz' fundamental work \cite{aus/buc:89} has included studies of invariants defined by Cohen-Macaulay approximation; \cite{din:92, aus/din/sol:93, has/shi:97} among several, `injectivity' and `surjectivity' properties of the approximation maps; \cite{kat:99, yos/iso:00, kat:07}, and characterisations of quasi-homogeneous isolated singularities; cf.\ \cite{her/mar:93, mar:00b}, all exclusively in the Gorenstein case. Noteworthy is \cite{sim/str:02} where A.-M.\ Simon and J.\ R.\ Strooker related some of these invariants with Hochster's Canonical Element Conjecture and the Monomial Conjecture. In particular these conjectures are equivalent to the vanishing of the \(\delta\)-invariant of certain cyclic modules over all Gorenstein rings.
S.\ P.\ Dutta applied the existence of a FID hull to prove a relationship between two of the Serre conjectures on intersection numbers: Failure of vanishing implies failure of higher non-negativity in the Gorenstein case under certain conditions, see \cite{dut:04}.
Buchweitz' unpublished manuscript \cite{buc:86}, a precursor to \cite{aus/buc:89}, contains homological ideas which have influenced subsequent developments (e.g.\ \cite{kra:05}). Auslander and I.\ Reiten elaborated in \cite{aus/rei:91} on \cite{aus/buc:89}, mainly with a view towards artin algebras, instigating several generalisations and analogies to Cohen-Macaulay approximation.
However, the `relative' and continuous aspects have received surprisingly little attention.
M.\ Hashimoto has given several new examples of Cohen-Macaulay approximation \cite{has:00}.
In \cite[IV 1.4.12]{has:00} an affine algebraic group \(G\) acts on a positively graded Cohen-Macaulay ring \(T\) which is flat over a regular base ring \(R\). Hashimoto considers graded maximal Cohen-Macaulay \(T\)-modules (which automatically are \(R\)-flat) and graded modules locally of finite injective dimension (not \(R\)-flat in general), all with \(G\)-action. His result (with trivial group) is hence different from our Theorem \ref{thm.flatCMapprox}. We also note some explicit \(1\)-parameter families of indecomposable finite length modules \(N_{t}\) (for many Gorenstein rings) such that the minimal MCM approximation module \(M_{t}\) is without free summands, see \cite{yos:99}.
A central part of the classification problem is to prove the existence of objects with certain properties and to estimate `how many' such objects there are.
A natural question is thus whether there is Cohen-Macaulay approximation for flat families of modules. In Theorem \ref{thm.flatCMapprox} we give a positive answer to this question. For a Cohen-Macaulay (CM) map \(h:S\rightarrow T\) and an \(S\)-flat and finite \(T\)-module \(\mc{N}\) there are short exact sequences of \(S\)-flat and finite \(T\)-modules
\begin{equation}\label{eq.flatMCMseq}
0\rightarrow \mc{L}\longrightarrow \mc{M}\longrightarrow \mc{N}\rightarrow 0\qquad\text{and}\qquad 0\rightarrow \mc{N}\longrightarrow \mc{L}'\longrightarrow \mc{M}'\rightarrow 0
\end{equation}
such that the fibres of these sequences give `absolute' approximations and hulls as in the two sequences \eqref{eq.MCMseq}. Note that \(T\) in general is \emph{not} a Cohen-Macaulay ring although the fibres of \(h\) are.
We consider a category \(\cat{mod}{}^{\textnormal{fl}}\) of pairs \(\xi=(h:S\rightarrow T,\mc{N})\) and subcategories \(\cat{MCM}\), \(\cat{FID}\) and \(\cat{D}\).
They are fibred over the category \(\cat{CM}\) of CM maps and also fibred over the base category of noetherian rings. The approximation and the hull \eqref{eq.flatMCMseq} induces functors of certain quotient categories fibred in additive categories over \(\cat{CM}\)
\begin{equation}
\cat{mod}{}^{\textnormal{fl}}/\cat{D}\rightarrow \cat{MCM}/\cat{D}\quad \text{and}\quad\cat{mod}{}^{\textnormal{fl}}/\cat{D}\rightarrow \cat{FID}/\cat{D}
\end{equation}
with analogous properties to the absolute case. If \(h:S\rightarrow T\) is a \emph{local} CM map, there is an approximation result with minimal (and hence unique) choices of the two sequences in \eqref{eq.flatMCMseq}, see Corollary \ref{cor.locCMapprox} and \ref{cor.minapprox}.
A major consequence of these results is that any numerical and additive upper semi-continuous invariant of MCM or FID modules by the minimal approximations and hulls induces upper semi-continuous invariants for all finite modules, see Theorem \ref{thm.semicont}. Examples of such invariants are given by the \(\omega_{A}\)-ranks in the minimal \emph{representing complex} \(D^{*}(N)\) which is an (infinite) extension to the right of the \(\cat{D}_{A}\)-complex in \eqref{eq.Dres}.
Auslander's fundamental module \(E_{A}\) for a normal \(2\)-dimensional singularity \(\Spec A\) is given by the MCM approximation of the maximal ideal;
\begin{equation}
0\rightarrow \omega_{A}\longrightarrow E_{A}\longrightarrow \fr{m}_{A}\rightarrow 0
\end{equation}
which in a certain sense generates all almost split sequences for \(A\), see \cite{aus:86}.
As a general example of flat Cohen-Macaulay approximation we define the fundamental module for any finite type CM map of pure relative dimension \(\geq 2\), see Corollary \ref{cor.Fmod}, and more generally a `fundamental' functor of projective modules in Proposition \ref{prop.Fmod}.
An attractive feature of Auslander and Buchweitz' theory is its axiomatic formulation with several applications besides the classical case described in the first paragraph, e.g.\ coherent rings with a cotilting module, the graded case, approximation with modules of Gorenstein dimension \(0\), and coherent sheaves on a projectively embedded Cohen-Macaulay scheme. See \cite{aus/buc:89} and \cite{has:00} for more examples. We formulate a relative Cohen-Macaulay approximation theory axiomatically in terms of categories \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\) fibred in abelian and additive subcategories over a base category \(\cat{C}\). In addition to the Auslander-Buchweitz axioms (AB1-4) for the fibre categories we formulate two axioms (BC1-2) regarding base change properties of the fibred categories. AB1-2 and BC1-2 imply the existence of an approximation and a hull which are preserved by any base change, see Theorem \ref{thm.cofapprox}. If AB3 holds too, we get functoriality and adjointness properties in suitable stable categories fibred in additive categories, see Theorem \ref{thm.cofmain}. In the case described above \(\cat{C}=\cat{CM}\), \(\cat{A}\) is the category \(\cat{mod}\) of pairs \((h:S\rightarrow T,\mc{N})\) where \(\mc{N}\) is a finite \(T\)-module (no \(S\)-flatness) and \(\cat{X}=\cat{MCM}\). Another application of this theory is given in \cite{ile:11xb}.
In the second half of the article we proceed to study properties of continuous families of MCM approximations and FID hulls by homological methods.
As a consequence of the existence of minimal approximations and hulls of local flat families there are induced natural maps of deformation functors of pairs of algebra and module:
\begin{equation}
\df{}{(A,N)}\longrightarrow\df{}{(A,X)}\quad\text{for}\quad X=M, M', L\,\,\text{and}\,\, L'
\end{equation}
There are corresponding maps \(\df{A}{N}\rightarrow\df{A}{X}\) of deformation functors of the modules where \(A\) only deforms trivially. Rather weak conditions on \(N\), e.g.\ \(\grade N\geq 1\), respectively \(\grade N\geq 2\), imply the injectivity and formal smoothness of these maps for \(X=L'\). If, in addition, there is a versal family in \(\df{}{(A,N)}\) (or \(\df{A}{N}\)) then the maps are smooth for the appropriate category of henselian rings, see Theorem \ref{thm.defgrade} and Corollary \ref{cor.defgrade}. As a consequence each CM algebraic \(k\)-algebra \(A\) with \(A/\fr{m}_{A}\cong k\) and \(\dim A\geq 2\) has a finite \(A\)-module \(Q'\) of finite \emph{projective} dimension with a \emph{universal} deformation in \(\df{A}{Q'}(A)\), see Corollary \ref{cor.defapprox}. There are analogous general results for \(X=M\), see Theorem \ref{thm.defgrade2} and Corollary \ref{cor.defgrade2}, with applications in Corollary \ref{cor.depth} and \ref{cor.2dim}. E.g.\ if there is a closed subscheme \(Z\) in \(\Spec A\) containing the singular locus and with complement \(U\) such that \(\tilde{N}_{\vert U}=0\) and \(\depth_{Z}N\geq 2\) then \(\sigma_{M}:\df{}{(A,N)}\rightarrow\df{}{(A,M)}\) is formally smooth.
Or if \(\Spec A\) is a \(2\)-dimensional normal Gorenstein singularity and \(N\) is torsion-free then the map \(\sigma_{M}\) is smooth. In this case both functors have versal elements by Theorem \ref{thm.ExVers}.
Consider a quotient ring \(B=A/I\) defined by a regular sequence \(I=(f_{1},\dots,f_{n})\) and an MCM \(B\)-module \(N\). Then \(N\) is also an \(A\)-module with an MCM approximation \(M\rightarrow N\). If \(N\) has a \emph{lifting} to \(A/I^{2}\), then the composition of natural maps \(\df{B}{N}\rightarrow \df{A}{N}\rightarrow\df{A}{M}\) is injective, see Theorem \ref{thm.defMCM}. It turns out that the lifting condition is equivalent to the splitting of \(B{\otimes}_{A}M\rightarrow N\) (this generalises \cite[4.5]{aus/din/sol:93}).
The second part of the article also contains some general deformation theory of a pair \((h:S\rightarrow T,\mc{N})\) of an algebra and a \(T\)-module. We define the graded algebra \(\varGamma:=T\oplus \mc{N}\) and consider the graded Andr{\'e}-Quillen cohomology \(\gH{0}^{*}(S,\varGamma, J)\) which govern the obstruction theory of the pair. In the case the graded \(\varGamma\)-module \(J\) is concentrated in degree \(0\) and \(1\) there is by Proposition \ref{prop.lang} a natural long-exact sequence which in the case \(J=\varGamma\) (with \(\gH{0}^{*}(S,\varGamma)=\gH{0}^{*}(S,\varGamma,\varGamma)\)) gives the suggestive
\begin{equation}
\begin{split}
0 &{} \rightarrow \nd{}{T}{\mc{N}}\rightarrow\gDer{0}_{S}(\varGamma)\rightarrow\Der_{S}(T)\rightarrow\xt{1}{T}{\mc{N}}{\mc{N}}\rightarrow\gH{0}^{1}(S,\varGamma)\rightarrow \\
&{}\cH^{1}(S,T) \rightarrow\xt{2}{T}{\mc{N}}{\mc{N}}\rightarrow\gH{0}^{2}(S,\varGamma)\rightarrow\cH^{2}(S,T)\rightarrow\dots
\end{split}
\end{equation}
It relates the cohomology of the pair with the cohomology groups governing the obstruction theory of the algebra \(T\) and of the module \(\mc{N}\).
The sequence is used in the proof of the existence of a versal element in \(\df{}{(A,N)}\) where \(\Spec A\) is an isolated equidimensional singularity and \(N\) is locally free on the smooth locus, see Theorem \ref{thm.ExVers}. It is also used to define and study the Kodaira-Spencer class \(\kappa(\varGamma/S/\varLambda)\) in \(\gH{0}^{1}(S,\varGamma,\Omega_{S/\varLambda}{\otimes}\varGamma)\) (where \(\varLambda\rightarrow S\) is a another ring homomorphism) which maps to the ungraded Kodaira-Spencer class \(\kappa(T/S/\varLambda)\). In the case the latter is zero we define a `secondary' Kodaira-Spencer class \(\kappa(\sigma,\mc{N})\) in \(\xt{1}{T}{\mc{N}}{\Omega_{S/\varLambda}{\otimes} \mc{N}}\) which depends on a choice of an \(S\)-algebra splitting \(\sigma\). This enables us to define `global' Kodaira-Spencer maps
\begin{equation}
g^{\varGamma}:\Der_{\varLambda}(S)\rightarrow\gH{0}^{1}(S,\varGamma,\varGamma)\quad \text{and}\quad g^{(\sigma,\,\mc{N})}:\Der_{\varLambda}(S)\rightarrow\xt{1}{T}{\mc{N}}{\mc{N}}\,.
\end{equation}
We also describe how classes and maps are related to the Atiyah class \(\at_{T/\varLambda}(\mc{N})\) in \(\xt{1}{T}{\mc{N}}{\Omega_{T/\varLambda}{\otimes}\mc{N}}\). These results might have a certain independent interest. The arguments are general and should be extendable to the setting of L.\ Illusie's \cite{ill:71}.
Injectivity of the corresponding local Kodaira-Spencer maps gives a criterion for a global flat family to be non-trivial. The Kodaira-Spencer maps commute with Cohen-Macaulay approximation and this is applied to show that injectivity of the Kodaira-Spencer map is preserved by Cohen-Macaulay approximation under conditions as in Theorem \ref{thm.defgrade}, \ref{thm.defgrade2} and \ref{thm.defMCM}.
To make the text more reader friendly we have included some background material, e.g.\ on Cohen-Macaualy approximation and Kodaira-Spencer maps, and some of the central technical tools such as a general `cohomology and base change' result and some language of fibred categories. Many results have analogous parts with similar arguments and the policy has been to give a fairly detailed proof of one case and leave the other cases to the reader.
\section{Preliminaries}\label{sec.CMapprox}
All rings are commutative. If \(A\) is a ring, \(\cat{Mod}_{A}\) denotes the category of \(A\)-modules and \(\cat{mod}_{A}\) denotes the full subcategory of finite \(A\)-modules. If \(A\) is local then \(\fr{m}_{A}\) denotes the maximal ideal. Subcategories are usually full and essential.
\subsection{Axiomatic Cohen-Macaulay approximation}\label{subs.CCM}
We briefly recall some of the main features of Cohen-Macaulay approximation as introduced by Auslander and Buchweitz in \cite{aus/buc:89}.
In this section let \(\cat{A}\) be an abelian category and \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\) additive subcategories.
Let \(\hat{\cat{X}}\) denote the subcategory of \(\cat{A}\) of objects \(N\) which have finite resolutions \(0\rightarrow M_{n}\rightarrow\dots\rightarrow M_{0}\rightarrow N\rightarrow 0\) with the \(M_{i}\) in \(\cat{X}\). If \(n\) is the smallest such number, then \(\resdim{\cat{X}}{N} = n\). Let \(\injdim{\cat{X}}{N}\) be the minimal \(n\) (possibly \(\infty\)) such that \(\xt{i}{\cat{A}}{M}{N}= 0\) for all \(i>n\) and all \(M\) in \(\cat{X}\). Let \(\cat{X}^{\perp}\) denote the subcategory of objects \(L\) in \(\cat{A}\) with \(\injdim{\cat{X}}{L}=0\); the right complement of \(\cat{X}\). The left complement \({}^{\perp}\cat{X}\) is defined analogously.
Let \(N\) be an object in \(\cat{A}\). An \emph{\(\cat{X}\)-approximation} and a \emph{\(\hat{\cat{D}}\)-hull}
of \(N\) are exact sequences as in \eqref{eq.MCMseq} with \(L\), \(L'\) in \(\hat{\cat{D}}\) and \(M\), \(M'\) in \(\cat{X}\).
In general any \(f:M\rightarrow N\) in \(\cat{A}\) is called a \emph{right \(\cat{X}\)-approximation of \(N\)} if \(M\) is in \(\cat{X}\) and any \(f':M'\rightarrow N\) with \(M'\) in \(\cat{X}\) factorises through \(f\). Dually, \(g:N\rightarrow L\) is called a \emph{left \(\cat{X}\)-approximation of \(N\)} if \(L\) is in \(\cat{X}\) and any \(g':N\rightarrow L'\) with \(L'\) in \(\cat{X}\) factorises through \(g\).
Consider the following conditions on the triple of categories \((\cat{A},\cat{X},\cat{D})\).
\begin{enumerate}
\item[(AB1)] \(\cat{X}\) is exact in \(\cat{A}\) (\(\cat{X}\) is closed under direct summands and extensions).
\item[(AB2)] \(\cat{D}\) is a \emph{cogenerator} for \(\cat{X}\), i.e.\ for each object \(M\) in \(\cat{X}\) there is an object
\(D\) in \(\cat{D}\) and a short exact sequence \(M\rightarrow D\rightarrow M'\) with \(M'\) in \(\cat{X}\).
\item[(AB3)] \(\cat{D}\) is \(\cat{X}\)-\emph{injective}, i.e.\ \(\cat{D}\subseteq\cat{X}^{\perp}\).
\item[(AB4)] \(\cat{A}\)-epimorphisms in \(\cat{X}\) are admissible (i.e.\ their kernels are contained in \(\cat{X}\)).
\end{enumerate}
If AB1 and AB2, there exist \(\cat{X}\)-approximations and \(\hat{\cat{D}}\)-hulls for all objects in \(\hat{\cat{X}}\) \cite[1.1]{aus/buc:89}. Assume AB1-3. Then any \(\cat{X}\)-approximation is a right \(\cat{X}\)-approximation and any \(\hat{\cat{D}}\)-hull is a left \(\hat{\cat{D}}\)-approximation. An \(\cat{X}\)-approximation determines a \(\hat{\cat{D}}\)-hull and vice versa through the following diagram of short exact sequences; the upper horizontal and right vertical being an \(\cat{X}\)-approximation and a \(\hat{\cat{D}}\)-hull of \(N\), \(D\) is in \(\cat{D}\). The boxed square is (co)cartesian (see \cite[1.4]{aus/buc:89}):
\begin{equation}\label{eq.pullpush}
\xymatrix@C-0pt@R-12pt@H-30pt{
L \ar[r]\ar@{=}[d] & M \ar[r]\ar[d]\ar@{}[dr]|{\Box} & N \ar[d] \\
L \ar[r] & D \ar[r]\ar[d] & L' \ar[d] \\
& M' \ar@{=}[r] & M'
}
\end{equation}
Moreover, the category \(\cat{D}\) is determined by \(\cat{X}\subset\cat{A}\). Indeed \(\cat{D}=\cat{X}\cap\cat{X}^{\perp}\). By \cite[3.9]{aus/buc:89} monomorphisms in \(\hat{\cat{D}}\) are admissible and \(\hat{\cat{D}}=\hat{\cat{X}}\cap\cat{X}^{\perp}\). Also \(\cat{X}={}^{\perp}\hat{\cat{D}}\cap\hat{\cat{X}}={}^{\perp}\cat{D}\cap\hat{\cat{X}}\). If \(\cat{X}/\cat{D}\) denotes the quotient category, the \(\cat{X}\)-approximation induces a right adjoint to the inclusion functor \(\cat{X}/\cat{D}\subseteq \hat{\cat{X}}/\cat{D}\) and the \(\hat{\cat{D}}\)-hull induces a left adjoint to the inclusion functor \(\hat{\cat{D}}/\cat{D}\subseteq\hat{\cat{X}}/\cat{D}\), see \cite[2.8]{aus/buc:89}.
A morphism \(f: M\rightarrow N\) in \(\cat{A}\) is called \emph{right minimal} if for any \(g:M\rightarrow M\) with \(fg=f\) it follows that \(g\) is an automorphism. Dually, \(f\) is called \emph{left minimal} if for any \(h:N\rightarrow N\) with \(hf=f\) it follows that \(h\) is an automorphism. Note that if \(f:M\rightarrow N\) and \(f: M'\rightarrow N\) both are right minimal then there exists an isomorphism \(g:M\rightarrow M'\) with \(f=f'g\), and similarly for left minimal morphisms.
We will simply call an \(\cat{X}\)-approximation (a \(\hat{\cat{D}}\)-hull) for minimal if it is right (left) minimal.
\begin{ex}\label{ex.MCMapprox}
Suppose \(A\) is a Cohen-Macaulay ring which posses a canonical module \(\omega_{A}\) in the sense that any localisation in a maximal ideal gives a maximal Cohen-Macaulay module of finite injective dimension and Cohen-Macaulay type \(1\), cf.\ \cite[3.3.16]{bru/her:98}. Let \(\cat{MCM}_{A}\) denote the category of maximal Cohen-Macaulay (MCM) \(A\)-modules and put \(\cat{D}_{A}:=\Add\{\omega_{A}\}\). Then the triple \((\cat{A},\cat{X},\cat{D})=(\cat{mod}_{A},\cat{MCM}_{A},\cat{D}_{A})\) satisfies properties AB1-4, cf.\ \cite[I 4.10.11]{has:00} and \(\hat{\cat{X}}=\cat{mod}_{A}\). If \(A\) in addition is a local ring, then the \(\cat{MCM}_{A}\)-approximation and the \(\hat{\cat{D}}_{A}\)-hull can be chosen to be minimal, cf.\ \cite[Sec.\ 3]{sim/str:02} or Corollary \ref{cor.minapprox}.
Let \(\cat{FID}_{A}^{l}\) denote the subcategory of finite \(A\)-modules \(E\) which have locally finite injective dimension, i.e.\ \(\Injdim_{A_{\fr{p}}} E_{\fr{p}}<\infty\) for all \(\fr{p}\in\Spec A\). The approximation result implies that \(\cat{FID}_{A}^{l}=\hat{\cat{D}}_{A}\): Let \(L\) be in \(\hat{\cat{D}}_{A}\). By induction on \(\resdim{\cat{D}_{A}}{L}\) \(L\) is in \(\cat{FID}_{A}^{l}\). Conversely let \(E\) be in \(\cat{FID}_{A}^{l}\). If \(L\rightarrow M\rightarrow E\) is an \(\cat{MCM}_{A}\)-approximation of \(E\) then \(M\) also has locally finite injective dimension. Let \(M^{\vee}\) denote \(\hm{}{A}{M}{\omega_{A}}\) and choose a surjection \(A^{{\oplus}n}\rightarrow M^{\vee}\). Both \(M^{\vee}\) and the kernel \(M_{1}\) are MCM. Applying \(\hm{}{A}{-}{\omega_{A}}\) gives (by duality theory) the short exact sequence \(M\xra{i} \omega_{A}^{{\oplus}n}\rightarrow M_{1}^{\vee}\). But \(i\) splits since \(\xt{1}{A}{M_{1}^{\vee}}{M}=0\) by \cite[3.3.3]{bru/her:98} and so \(M\) is in \(\cat{D}_{A}\) and \(E\) is in \(\hat{\cat{D}}_{A}\).
\end{ex}
\subsection{The representing complex}\label{subsec.cplx}
Consider an abelian category \(\cat{A}\) and additive subcategories \(\cat{D}\subseteq\cat{X}\subseteq \cat{A}\). A \emph{\(\cat{DX}\)-resolution} of an object \(N\) in \(\cat{A}\) is a finite resolution \({}^{-}C^{*}\twoheadrightarrow N\) with \({}^{-}C^{i}\in\cat{D}\) for \(i<0\) and \({}^{-}C^{0}\in \cat{X}\). If \(L:=\coker (d^{-2}:{}^{-}C^{-2}\rightarrow {}^{-}C^{-1})\), then the short exact sequence \(L\rightarrow {}^{-}C^{0}\rightarrow N\) is an \(\cat{X}\)-approximation. A \emph{\(\hat{\cat{D}}\cat{D}\)-coresolution} of \(N\) is a coresolution \(N\rightarrowtail {}^{+}C^{*}(N)\) such that \({}^{+}C^{0}\in\hat{\cat{D}}\), \({}^{+}C^{i}\in\cat{D}\) and \(\ker d^{i}\in \cat{X}\) for \(i>0\). If \(M':=\ker d^{1}\) then the short exact sequence \(N\rightarrow {}^{+}C^{0}\rightarrow M'\) is a \(\hat{\cat{D}}\)-hull.
Given AB1 and AB2, each \(N\) in \(\hat{\cat{X}}\) has a \(\cat{DX}\)-resolution and a \(\hat{\cat{D}}\cat{D}\)-coresolution.
Finally, a bounded below \(\cat{D}\)-complex \(D^{*}(N):\dots\rightarrow D^{-1}\rightarrow D^{0}\rightarrow D^{1}\rightarrow\dots\) with \(\ker d^{i}\in\cat{X}\) for all \(i\geq 0\) and its only non-trivial cohomology in degree zero with \(\cH^{0}(D^{*})\cong N\) is called a \emph{\(\cat{D}\)-complex representing \(N\)}. A representing complex splits into (and is (re)constructed from) a \(\cat{DX}\)-resolution given by \(\dots\rightarrow D^{-1}\rightarrow \ker d^{0}\twoheadrightarrow \cH^{0}(D^{*})=N\) and a \(\hat{\cat{D}}\cat{D}\)-coresolution \(N\rightarrowtail \coker d^{-1}\rightarrow D^{1}\rightarrow\dots\) where \(N\rightarrowtail \coker d^{-1}\) is induced by \(\ker d^{0}\rightarrowtail D^{0}\).
\begin{lem}\label{lem.res}
Assume \(\xt{1}{\cat{A}}{\cat{X}}{\hat{\cat{D}}}=0\).
Suppose \(f:N_{1}\rightarrow N_{2}\) is in \(\hat{\cat{X}}\)\textup{.} Assume \(F^{*}(N_{i})\) exists for \(i=1,2\) where \(F^{*}(N_{i})\) denotes one of the complexes \({}^{-}C^{*}(N_{i})\), \({}^{+}C^{*}(N_{i})\) or \(D^{*}(N_{i})\)\textup{.} Then \(f\) can be extended to an arrow of chain complexes \(f^{*}:F^{*}(N_{1})\rightarrow F^{*}(N_{2})\) which is uniquely defined up to homotopy\textup{.}
Assume \textup{AB1-3} for the triple of categories \((\cat{A},\cat{X},\cat{D})\)\textup{.} Then \(N\mapsto {}^{-}C^{*}(N)\)\textup{,} \(N\mapsto {}^{+}C^{*}(N)\) and \(N\mapsto D^{*}(N)\) induce functors to the homotopy categories of chain complexes as follows\textup{:}
\begin{equation*}
{}^{-}C^{*}:\hat{\cat{X}}\rightarrow \cat{K}^{\textnormal{b}}(\cat{X})\quad\quad
{}^{+}C^{*}:\hat{\cat{X}}\rightarrow \cat{K}^{+}(\hat{\cat{D}})\quad\quad
D^{*}:\hat{\cat{X}}/\cat{D}\rightarrow \cat{K}^{+}(\cat{D})
\end{equation*}
\end{lem}
\begin{proof}
The proof for \({}^{-}C^{*}(N)\) and \({}^{+}C^{*}(N)\) follows standard lines for constructing chain maps and homotopies. The assumption \(\xt{1}{\cat{A}}{\cat{X}}{\hat{\cat{D}}}=0\) is used every time a lifting or extension of an arrow is required.
Let \((D^{*}_{i}, d_{i}^{*})=D^{*}(N_{i})\) and let \(M_{i}=\ker d_{i}^{0}\) and \(L_{i}=\im d_{i}^{-1}\). Then there are short exact sequences \(L_{i}\rightarrow M_{i}\rightarrow N_{i}\) which by assumption are \(\cat{X}\)-approximations. Since \(\xt{1}{\cat{A}}{M_{1}}{L_{2}}=0\), the arrow \(N_{1}\rightarrow N_{2}\) extends to the \(\cat{X}\)-approximation and further on to the negative part of the complexes. If \(M'_{i}=\ker d^{1}_{i}\) then the \(M'_{i}\) are in \(\cat{X}\) by assumption and there are short exact sequences \(M_{i}\rightarrow D^{0}_{i}\rightarrow M'_{i}\). There is an extension of \(M_{1}\rightarrow D^{0}_{2}\) to \(D^{0}_{1}\rightarrow D^{0}_{2}\) and an induced arrow \(M'_{1}\rightarrow M'_{2}\) which again extends and so on to a chain map \(f^{*}:D^{*}_{1}\rightarrow D^{*}_{2}\).
Let \(g^{*}:D^{*}_{1}\rightarrow D^{*}_{2}\) be a chain map, put \(g=\cH^{0}(g^{*})\), \(s=f{-}g\) and \(s^{*}=f^{*}{-}g^{*}\). Suppose \(s\) factors through \(D\) in \(\cat{D}\); \(s=ab\) with \(a:D\rightarrow N_{2}\). Since \(\xt{1}{\cat{A}}{D}{L_{2}}=0\) there exist a lifting \(\tilde{a}:D\rightarrow M_{2}\) of \(a\). Put \(h_{N}=\tilde{a}b\) and continue similarly to construct a homotopy \(h\) for the extended negative part:
\begin{equation*}
\xymatrix@C+6pt@R-0pt@H+6pt{
& \dots \ar[r] & D_{1}^{-1} \ar[dl]_{h^{-1}}\ar[r]\ar[d]|-{s^{-1}} & M_{1} \ar[dl]|-{h_{M}}\ar[r]\ar[d]|-{\mr{Z}^{0}\!(s^{*})} & N_{1} \ar[dl]|-{h_{N}}\ar[d]^{s}\ar[r] & 0 \\
\dots \ar[r] & D_{2}^{-2}\ar[r] & D_{2}^{-1} \ar[r] & M_{2} \ar[r] & N_{2} \ar[r] & 0
}
\end{equation*}
In particular \(h_{M}:M_{1}\rightarrow D^{-1}_{2}\) can be extended to an \(h^{0}:D^{0}_{1}\rightarrow D^{-1}_{2}\) with \(s^{-1}=h^{0}d^{-1}_{1}{+}d^{-2}_{2}h^{-1}\). The construction of the \(h^{i}\) for \(i>0\) is standard.
\end{proof}
\begin{lem}\label{lem.xres}
Assume \textup{AB1-3} for the triple of categories \((\cat{A},\cat{X},\cat{D})\)\textup{.} Given an exact sequence \(\varepsilon:0\rightarrow N_{1}\rightarrow N_{2}\rightarrow N_{3}\rightarrow 0\) with objects in \(\hat{\cat{X}}\)\textup{.} Then there are exact sequences of complexes where \(\varepsilon\) equals the cohomology\textup{:}
\begin{enumerate}
\item[(i)] \(0\rightarrow {}^{-}C^{*}(N_{1})\longrightarrow {}^{-}C^{*}(N_{2})\longrightarrow {}^{-}C^{*}(N_{3})\rightarrow 0\)
\item[(ii)] \(0\rightarrow {}^{+}C^{*}(N_{1})\longrightarrow {}^{+}C^{*}(N_{2})\longrightarrow {}^{+}C^{*}(N_{3})\rightarrow 0\)
\item[(iii)] \(0\rightarrow D^{*}(N_{1})\longrightarrow D^{*}(N_{2})\longrightarrow D^{*}(N_{3})\rightarrow 0\) \textup{(}termwise split exact\textup{)}
\end{enumerate}
\end{lem}
\begin{proof}
Choose \(\cat{X}\)-approximations \(L_{i}\rightarrow M_{i}\rightarrow N_{i}\) for \(i=1,3\). There is an \(3{\times}3\) commutative diagram of \(6\) short exact sequences which extends the ``horseshoe'' diagram, cf.\ \cite[1.12.11]{has:00}. One obtains an \(\cat{X}\)-approximation of \(N_{2}\) and short exact sequences \(m:M_{1}\rightarrow M_{2}\rightarrow M_{3}\) and \(L_{1}\rightarrow L_{2}\rightarrow L_{3}\) in \(\cat{X}\) and \(\hat{\cat{D}}\) respectively since both categories are closed by extensions (by AB1 and \cite[3.8]{aus/buc:89}). If \(D^{*}_{i}[1]\xra{\eta_{i}} L_{i}\) are finite \(\cat{D}\)-resolutions then since \(\xt{1}{\cat{A}}{D^{-1}_{3}}{L_{1}}=0\) there is a lifting \(\tilde{\eta}_{3}:D^{-1}_{3}\rightarrow L_{2}\) of \(\eta_{3}\) which combined with \(\eta_{1}\) gives \(\eta_{2}:D^{-1}_{1}\coprod D^{-1}_{3}\rightarrow L_{2}\). The kernels of the resulting arrows between short exact sequences give a short exact sequence of objects in \(\hat{\cat{D}}(S)\). The argument is repeated. Splicing with \(m\) in degree zero the short exact sequence of \({}^{-}C^{*}\)-resolutions in (i) is obtained.
Choose short exact sequences \(M_{i}\rightarrow D^{0}_{i}\rightarrow M_{i}'\) for \(i=1,3\) as in AB2. Since \(\xt{1}{\cat{A}}{M_{3}}{D^{0}_{1}}=0\) there is an extension to an arrow of short exact sequences from \(m\) to \(D^{0}_{1}\rightarrow D^{0}_{2}\rightarrow D^{0}_{3}\) with \(D^{0}_{2}=D^{0}_{1}\coprod D^{0}_{3}\) and \(M_{2}':=\coker(M_{2}\rightarrow D^{0}_{2})\in \cat{X}\) by AB1. Repeated application of this argument gives a short exact sequence of \(\cat{D}\)-coresolutions and splicing with the sequences in (i) gives (iii). Pushout of \(M_{i}\rightarrow D^{0}_{i}\rightarrow M_{i}'\) along \(M_{i}\rightarrow N_{i}\) gives a short exact sequence of \(\hat{\cat{D}}\)-hulls and splicing with \(D^{1}_{i}\rightarrow D^{2}_{i}\rightarrow\dots\) gives (ii).
\end{proof}
\subsection{Base change}
The main tool for reducing properties to the fibres in a flat family will be the base change theorem.
We follow the quite elementary and general approach of A.\ Ogus and G.\ Bergman \cite{ogu/ber:72}.
\begin{defn}
Let \(h:S\rightarrow T\) be a ring homomorphism and \(I\) an \(S\)-module.
Let \(F\) be an \(S\)-linear functor of some additive subcategory of \(\cat{Mod}_{S}\) to \(\cat{Mod}_{T}\). Then the \emph{exchange map \(e_{I}\) for \(F\)} is defined as the \(T\)-linear map \(e_{I}:F(S){\otimes}_{S}I\rightarrow F(I)\) given by \(\xi{\otimes} u\mapsto F(u)(\xi)\) where we consider \(u\) as the multiplication map \(u:S\rightarrow I\). Let \(\mSpec T\) denote the set of closed points in \(\Spec T\).
\end{defn}
\begin{prop}\label{prop.nakayama}
Let \(h:S\rightarrow T\) be a ring homomorphism with \(S\) noetherian\textup{.} Suppose \(\{F^{q}:\cat{mod}_{S}\rightarrow \cat{mod}_{T}\}_{q\geq 0}\) is an \(h\)-linear cohomological \(\delta\)-functor\textup{.}
\begin{enumerate}
\item[(i)] If the exchange map \(e_{S/\fr{n}}^{q}:F^{q}(S){\otimes}_{S}S/\fr{n}\rightarrow F^{q}(S/\fr{n})\) is surjective for all \(\fr{n}\) in \(Z=\im\{\mSpec T\rightarrow\Spec S\}\)\textup{,} then \(e_{I}^{q}:F^{q}(S){\otimes}_{S}I\rightarrow F^{q}(I)\) is an isomorphism for all \(I\) in \(\cat{mod}_{S}\)\textup{.}
\item[(ii)] If \(e_{S/\fr{n}}^{q}\) is surjective for all \(\fr{n}\) in \(Z\)\textup{,} then \(e_{I}^{q-1}\) is an isomorphism for all \(I\) in \(\cat{mod}_{S}\) if and only if \(F^{q}(S)\) is \(S\)-flat\textup{.}
\end{enumerate}
\end{prop}
Note that if the \(F^{q}\) in addition extend to functors of all \(S\)-modules \(F^{q}:\cat{Mod}_{S}\rightarrow\cat{Mod}_{T}\) which commute with direct limits, then the conclusions are valied for all \(I\) in \(\cat{Mod}_{S}\).
\begin{ex}\label{ex.nakcplx}
Suppose \(S\) and \(T\) are noetherian. Let \(K^{*}:\, K^{0}\rightarrow K^{1}\rightarrow\dots\) be a complex of \(S\)-flat and finite \(T\)-modules. Define \(F^{q}:\cat{mod}_{S}\rightarrow\cat{mod}_{T}\) by \(F^{q}(I)=\cH^{q}(K^{*}{\otimes}_{S}I)\). Then \(\{F^{q}\}_{q\geq 0}\) is an \(h\)-linear cohomological \(\delta\)-functor which extends to all \(S\)-modules and commutes with direct limits.
\end{ex}
\begin{ex}\label{ex.nakayama}
Suppose \(S\) and \(T\) are noetherian. Let \(M\) and \(N\) be finite \(T\)-modules with \(N\) \(S\)-flat. Then the functors \(F^{q}:\cat{mod}_{S}\rightarrow\cat{mod}_{T}\) defined by \(F^{q}(I)=\xt{q}{T}{M}{N{\otimes}_{S}I}\) for \(q\geq 0\) give an \(h\)-linear cohomological \(\delta\)-functor which extends to all \(S\)-modules and commutes with direct limits.
\end{ex}
Let \(S\rightarrow T\) and \(S\rightarrow S'\) be ring homomorphisms, \(M\) a \(T\)-module, \(T'=T{\otimes}_{S}S'\) and \(N'\) a \(T'\)-module. Then there is a change of rings spectral sequence
\begin{equation}\label{eq.ss}
\cE_{2}^{p,q}=\xt{q}{T'}{\tor{S}{p}{M}{S'}}{N'}\,\Rightarrow\,\xt{p+q}{T}{M}{N'}
\end{equation}
which, in addition to the isomorphism \(\hm{}{T'}{M{\otimes}_{S}S'}{N'}\cong\hm{}{T}{M}{N'}\), gives edge maps \(\xt{q}{T'}{M{\otimes}_{S}S'}{N'}\rightarrow\xt{q}{T}{M}{N'}\) for \(q>0\) which are isomorphisms too if \(M\) (or \(S'\)) is \(S\)-flat. If \(I'\) is an \(S'\)-module we can compose the exchange map \(e^{q}_{I'}\) (regarding \(I'\) as \(S\)-module) with the inverse of this edge map for \(N'=N{\otimes}_{S}I'\) and obtain a map \(c^{q}_{I'}\) of \(T'\)-modules
\begin{equation}\label{eq.basechange}
c^{q}_{I'}:\xt{q}{T}{M}{N}{\otimes}_{S}I'\rightarrow\xt{q}{T'}{M{\otimes}_{S}S'}{N{\otimes}_{S}I'}\,.
\end{equation}
\begin{rem}
This is the base change map (in the affine case) considered by A.\ Altman and S.\ Kleiman, their conditions are slightly different, see \cite[1.9]{alt/kle:80}.
\end{rem}
We will use the following geometric notation. Suppose \(h:S\rightarrow T\) is a ring homomorphism, \(M\) is a \(T\)-module and \(s\) is a point in \(\Spec S\) with residue field \(k(s)\). Then \(M_{s}\) denotes the fibre \(M{\otimes}_{S}k(s)\) of \(M\) at \(s\) with its natural \(T_{s}=T{\otimes}_{S}k(s)\)-module structure. Now Proposition \ref{prop.nakayama} implies the following:
\begin{cor}\label{cor.xtdef}
Suppose \(S\rightarrow T\) and \(S\rightarrow S'\) are homomorphisms of noetherian rings\textup{,} \(M\) and \(N\) are finite \(T\)-modules\textup{,} \(Z=\im\{\mSpec T\rightarrow\Spec S\}\) and \(q\) is an integer\textup{.} Assume that \(M\) \textup{(}if \(q>0\)\textup{)} and \(N\) are \(S\)-flat\textup{.}
\begin{enumerate}
\item[(i)] If\, \(\xt{q+1}{T_{s}}{M_{s}}{N_{s}}=0\) for all \(s\) in \(Z\)\textup{,} then \(c^{q}_{I'}\) in \eqref{eq.basechange} is an isomorphism for all \(S'\)-modules \(I'\)\textup{.}
\item[(ii)] If in addition \(\xt{q-1}{T_{s}}{M_{s}}{N_{s}}=0\) for all \(s\in Z\)\textup{,} then \(\xt{q}{T}{M}{N}\) is \(S\)-flat\textup{.}
\end{enumerate}
\end{cor}
\section{Categories fibred in additive categories}\label{sec.cof}
We will phrase our results in the language of fibred categories\footnote{We have chosen to work with rings instead of (affine) schemes. Our definition of a fibred category \(p:\cat{F}\rightarrow\cat{C}\) reflects this choice and is equivalent to the functor of opposite categories \(p^{\text{op}}:\cat{F}^{\text{op}}\rightarrow\cat{C}^{\text{op}}\) being a fibred category as defined in \cite{FAG}.}. We therefore briefly recall some of the basic notions, taken mainly from A.\ Vistoli's article in \cite{FAG}. Then we define quotients of categories fibred in additive categories.
Consider a category \(\cat{C}\). Given a category over \(\cat{C}\), i.e.\ a functor \(p:\cat{F}\rightarrow\cat{C}\).
To an object \(T\) in \(\cat{C}\), let \(\cat{F}(T)\); the \emph{fiber of \(\cat{F}\) over \(T\)}, denote the subcategory of arrows \(\phi\) in \(\cat{F}\) such that \(p(\phi)=\id_{T}\).
An arrow \(\phi_{1}:\xi\rightarrow\xi_{1}\) in \(\cat{F}\) is \emph{cocartesian} if for any arrow \(\phi_{2}:\xi\rightarrow\xi_{2}\) in \(\cat{F}\) and any arrow \(f_{21}:p(\xi_{1})\rightarrow p(\xi_{2})\) in \(\cat{C}\) with \(f_{21} p(\phi_{1})=p(\phi_{2})\) there exists a unique arrow \(\phi_{21}:\xi_{1}\rightarrow\xi_{2}\) with \(p(\phi_{21})=f_{21}\) and \(\phi_{21}\phi_{1}=\phi_{2}\). If for any arrow \(f:T\rightarrow T'\) in \(\cat{C}\) and any object \(\xi\) in \(\cat{F}\) with \(p(\xi)=T\) there exists a cocartesian arrow \(\phi:\xi\rightarrow \xi'\) for some \(\xi'\) with \(p(\phi)=f\), then \(\cat{F}\) (or rather \(p:\cat{F}\rightarrow\cat{C}\)) is a \emph{fibred category}. Moreover, \(\xi'\) will be called a \emph{base change} of \(\xi\) by \(f\). If \(\xi''\) is another base change of \(\xi\) by \(f\) then \(\xi'\) and \(\xi''\) are isomorphic over \(T'\) by a unique isomorphism. We shall also say that a property \(P\) of objects in the fibres of \(\cat{F}\) is \emph{preserved by base change} if \(P(\xi)\) implies \(P(\xi')\) for any base change \(\xi'\) of \(\xi\).
A morphism of fibred categories is a functor \(F:\cat{F}_{1}\rightarrow\cat{F}_{2}\) with \(p_{2}F=p_{1}\) such that \(\phi\) cocartesian implies \(F(\phi)\) cocartesian. If \(F\) in addition is an inclusion of categories, \(\cat{F}_{1}\) is a fibred subcategory of \(\cat{F}_{2}\).
A category with all arrows being isomorphisms is a groupoid. A fibred category \(\cat{F}\) over \(\cat{C}\) is called a \emph{category fibred in groupoids} (often abbreviated to groupoid) if all fibres \(\cat{F}(T)\) are groupoids. Then all arrows in \(\cat{F}\) are cocartesian. If all fibres \(\cat{F}(T)\) only contain identities, then \(\cat{F}\) is called a \emph{category fibred in sets}.
\begin{lem}\label{lem.gpoid}
Given functors \(F:\cat{F}\rightarrow\cat{G}\) and \(q:\cat{G}\rightarrow\cat{C}\) and suppose \(q\) is fibred in sets\textup{.} Then \(F\) is fibred \textup{(}in groupoids\textup{/}sets\textup{)} if and only if \(qF\) is fibred \textup{(}in groupoids\textup{/}sets\textup{).}
\end{lem}
If \(T\) is an object in a category \(\cat{C}\) let \(\cat{C}/T\) denote the comma category of arrows to \(T\). Then the forgetful functor \(\cat{C}/T\rightarrow\cat{C}\) is fibred in sets. If \(p:\cat{F}\rightarrow\cat{C}\) is fibred (in groupoids/sets), \(\xi\) is an object in \(\cat{F}\) and \(T=p(\xi)\), then there is a natural functor \(p_{\xi}:\cat{F}/\xi\rightarrow\cat{C}/T\). The composition \(\cat{F}/\xi\rightarrow\cat{F}\rightarrow\cat{C}\) is clearly fibred (in groupoids/sets) and hence \(\cat{F}/\xi\rightarrow \cat{C}/T\) is fibred (in groupoids/sets) by Lemma \ref{lem.gpoid}.
If \(p:\cat{F}\rightarrow\cat{C}\) is a functor and \(\cat{C}'\) is a subcategory of \(\cat{C}\) we can define the \emph{restriction} \(p':\cat{F}_{\vert \cat{C}'}\rightarrow\cat{C}'\) of \(\cat{F}\) to \(\cat{C}'\) by picking for \(\cat{F}_{\vert \cat{C}'}\) the objects and morphisms in \(\cat{F}\) that \(p\) takes into \(\cat{C}'\). It follows that \(\cat{F}_{\vert \cat{C}'}\) is fibred (in groupoids/sets) if \(\cat{F}\) is.
The composition of two cocartesian arrows is cocartesian and isomorphisms are cocartesian. Hence the subcategory \(\cat{F}_{\text{coca}}\) of cocartesian arrows in a fibred category \(\cat{F}\) over \(\cat{C}\) is fibred in groupoids. If \(\cat{F}\) is fibred in groupoids there is an associated category fibred in sets \(\bar{\cat{F}}\rightarrow\cat{C}\) defined by identifying all isomorphic objects in all fibres \(\cat{F}(T)\) and identifying arrows accordingly. If \(\cat{F}\) is fibred in sets one defines a functor \(F:\cat{C}\rightarrow\Sets\) by \(F(T):=\cat{F}(T)\) and \(F(f):F(T)\rightarrow F(T')\) is defined by \(F(f)(\xi):=\eta_{\xi,f}\) where \(\phi_{\xi,f}:\xi\rightarrow \eta_{\xi,f}\) is the (in this case) unique cocartesian lifting of \(f\). From a functor \(G:\cat{C}\rightarrow\Sets\) one defines a category fibred in sets, and these two operations are inverse up to natural equivalences.
\begin{defn}\label{defn.addkof}
An \emph{additive \textup{(}abelian\textup{)} category \(\cat{F}\) over \(\cat{C}\)} is a functor \(p:\cat{F}\rightarrow\cat{C}\) such that:
\begin{enumerate}
\item[(i)] The fibre \(\cat{F}(T)\) is an additive (abelian) category for all objects \(T\) in \(\cat{C}\).
\item[(ii)] For all objects \(\xi_{1}\) and \(\xi_{2}\) in \(\cat{F}\) and arrows \(f:p(\xi_{1})\rightarrow p(\xi_{2})\) in \(\cat{C}\),
\begin{equation*}
\hm{}{f}{\xi_{1}}{\xi_{2}}:=\{\phi\in\hm{}{\cat{F}}{\xi_{1}}{\xi_{2}}\,\vert\,p(\phi)=f\}
\end{equation*}
is an abelian group, and composition of arrows
\begin{equation*}
\hm{}{f_{2}}{\xi_{2}}{\xi_{3}}\times\hm{}{f_{1}}{\xi_{1}}{\xi_{2}}\rightarrow\hm{}{f_{2}f_{1}}{\xi_{1}}{\xi_{3}}
\end{equation*}
is bilinear.
\end{enumerate}
A morphism \(F:\cat{F}_{1}\rightarrow\cat{F}_{2}\) of additive (abelian) categories over \(\cat{C}\) is a \emph{linear} functor \(F\) over \(\cat{C}\), i.e.\ which gives linear maps of \(\Hom\)-groups. If in addition \(F\) is an inclusion of categories then \(\cat{F}_{1}\) is an additive (abelian) subcategory of \(\cat{F}_{2}\) over \(\cat{C}\). A category \(\cat{F}\) over \(\cat{C}\) is \emph{fibred in additive \textup{(}abelian\textup{)} categories}, abbreviated by FAd (FAb), if \(\cat{F}\) is both fibred and additive (abelian) over \(\cat{C}\). Morphisms should be linear and preserve cocartesian arrows. A FAd subcategory is a morphism of FAds which is an inclusion of categories. For \(i=1,2\) let \(\cat{A}_{i}\) be a FAb over \(\cat{C}\) and \(\cat{X}_{i}\subseteq\cat{A}_{i}\) a FAd subcategory such that the fibre categories \(\cat{X}_{i}(T)\) are exact. Then a morphism of FAds \(F:\cat{X}_{1}\rightarrow\cat{X}_{2}\) is \emph{exact} if \(F\) preserves short exact sequences for all the fibre categories.
\end{defn}
Note that in a FAd finite (co)products in the fibres are preserved by base change.
Given a FAd subcategory \(\cat{D}\subseteq\cat{F}\). Two arrows \(\phi_{1}\) and \(\phi_{2}\) in \(\cat{F}\) are \(\cat{D}\)-equivalent if \(p(\phi_{1})=p(\phi_{2})\) and \(\phi_{1}-\phi_{2}\) factors through an object in \(\cat{D}\). Write \(\phi_{1}\sim\phi_{2}\).
Define the quotient category \(\cat{F}/\cat{D}\) over \(\cat{C}\) to have the same objects as \(\cat{F}\) and \(\hm{}{\cat{F}/\cat{D}}{\xi_{1}}{\xi_{2}}:=\hm{}{\cat{F}}{\xi_{1}}{\xi_{2}}/\sim\). The natural map to \(\cat{C}\) makes \(\cat{F}/\cat{D}\) an additive category over \(\cat{C}\) and the natural functor \(\cat{F}\rightarrow\cat{F}/\cat{D}\) is linear over \(\cat{C}\).
\begin{lem}\label{lem.keystable}
If \(\phi_{1}:\xi\rightarrow\xi_{1}\) is cocartesian in \(\cat{F}\) and \(\phi:\xi_{1}\rightarrow\xi_{2}\) is any arrow such that \(\phi\phi_{1}\sim 0\) then \(\phi\sim 0\)\textup{.}
\end{lem}
\begin{proof}
Suppose \(\phi\phi_{1}=\beta\alpha\) with \(\alpha:\xi\rightarrow\delta\) and with \(\delta\) in \(\cat{D}\). If \(p(\beta):T'\rightarrow T_{2}\) then since \(\cat{D}\) is a fibred subcategory there exists an arrow \(\delta\rightarrow\delta_{2}\) which is cocartesian in \(\cat{F}\) and with \(p(\delta_{2})=p(\xi_{2})=T_{2}\). Replacing \(\delta\) with \(\delta_{2}\) we assume \(p(\delta)=T_{2}\). Since \(\phi_{1}\) is cocartesian there exists a unique arrow \(\tau:\xi_{1}\rightarrow\delta\) with \(\tau\phi_{1}=\alpha\). Since \(\phi_{1}\) is cocartesian uniqueness implies that \(\beta\tau=\phi\).
\end{proof}
\begin{lem}\label{lem.stable}
Given a FAd subcategory \(\cat{D}\subseteq\cat{F}\) over \(\cat{C}\)\textup{,} then the quotient category \(\cat{F}/\cat{D}\) is FAd over \(\cat{C}\) and the quotient morphism \(\cat{F}\rightarrow\cat{F}/\cat{D}\) is a morphism of FAds\textup{.}
\end{lem}
\begin{proof}
We first show that if \(\phi_{1}:\xi\rightarrow\xi_{1}\) is cocartesian in \(\cat{F}\) then its image \([\phi_{1}]\) in \(\cat{F}/\cat{D}\) is cocartesian. Given \(\phi_{2}:\xi\rightarrow\xi_{2}\) and \(\theta:\xi_{1}\rightarrow\xi_{2}\) with \(\theta\phi_{1}=\phi_{2}\). Suppose \(\theta':\xi_{1}\rightarrow\xi_{2}\) with \(p(\theta')=p(\theta)\) satisfies \(\theta'\phi_{1}\sim\phi_{2}\). If \(\phi=\theta'-\theta\) then \(\phi\phi_{1}\sim 0\) so by Lemma \ref{lem.keystable} \(\phi\sim 0\). Now we show that \([\theta]\) is independent of the representations of the other maps. Let \(\phi_{i}':\xi\rightarrow\xi_{i}\) with \(\phi_{i}'\sim\phi_{i}\) and suppose (as we may) that \(\theta'\) satisfies \(\theta'\phi_{1}'=\phi_{2}'\) with \(p(\theta')=p(\theta)\). Again let \(\phi=\theta'-\theta\). Then \(0\sim\phi'_{2}-\phi_{2}=\theta'\phi'_{1}-\theta\phi_{1}=\theta'(\phi_{1}'-\phi_{1})+\phi\phi_{1}\sim\phi\phi_{1}\). By Lemma \ref{lem.keystable} \(\phi\sim 0\). Given \(f:T\rightarrow T_{1}\) and \(\xi\) in \(\cat{F}/\cat{D}\) with \(p(\xi)=T\) there exists a cocartesian \(\phi_{1}:\xi\rightarrow \xi_{1}\) in \(\cat{F}\) with \(p(\phi_{1})=f\) and by what we have done \([\phi_{1}]\) is cocartesian in \(\cat{F}/\cat{D}\).
\end{proof}
Note that there are in general more cocartesian arrows in \(\cat{F}/\cat{D}\) than those in the image of cocartesian arrows in \(\cat{F}\). The following lemma characterises the cocartesian arrows in the quotient category:
\begin{lem}\label{lem.stablecc}
If \(\rho\) and \(\theta\) are composeable arrows in \(\cat{F}\) with \(\rho\) cocartesian and \(\theta\) inducing an isomorphism in \(\cat{F}/\cat{D}\)\textup{,} then \([\theta\rho]\) is cocartesian in \(\cat{F}/\cat{D}\)\textup{.}
Conversely\textup{,} suppose \([\phi]:\xi_{1}\rightarrow\xi_{2}\) is cocartesian in \(\cat{F}/\cat{D}\) over \(f:T_{1}\rightarrow T_{2}\)\textup{.} Then for any base change
\(\rho:\xi_{1}\rightarrow \xi_{1}^{\#}\) of \(\xi_{1}\) over \(f\) in \(\cat{F}\)\textup{,} the induced arrow \(\phi^{\#}:\xi_{1}^{\#}\rightarrow\xi_{2}\) gives an isomorphism in \(\cat{F}/\cat{D}(T_{2})\)\textup{.}
\end{lem}
\begin{proof}
If \(\rho:\xi_{1}\rightarrow\xi_{2}\) is cocartesian and \([\theta]:\xi_{2}\rightarrow \xi_{3}\) is an isomorphism, let \(\phi=\theta\rho\). If \(\tau:\xi_{1}\rightarrow \xi_{4}\), \(p(\xi_{i})=T_{i}\) and there is a map \(f:T_{3}\rightarrow T_{4}\) with \(p(\tau)=f p(\theta) p(\rho)\), then there is a unique arrow \(\mu:\xi_{2}\rightarrow\xi_{4}\) above \(f p(\theta)\) with \(\mu\rho=\tau\). This gives the arrow \([\mu][\theta]^{-1}:\xi_{3}\rightarrow\xi_{4}\). If \(\mu_{i}:\xi_{3}\rightarrow\xi_{4}\) for \(i=1,2\) are two arrows with \([\mu_{i}][\phi]=[\tau]\), then \([\mu_{1}][\theta]=[\mu_{2}][\theta]\) since \([\rho]\) is cocartesian in \(\cat{F}/\cat{D}\) by Lemma \ref{lem.stable}. Since \([\theta]\) is an isomorphism, \([\mu_{1}]=[\mu_{2}]\).
Conversely, since \([\phi]\) is cocartesian there is a unique arrow \([\psi]:\xi_{2}\rightarrow\xi_{1}^{\#}\) in \(\cat{F}/\cat{D}(T_{2})\) with \([\psi\phi]=[\rho]\). By Lemma \ref{lem.stable} \([\rho]\) is cocartesian. It follows that \([\phi^{\#}]=[\psi]^{-1}\).
\end{proof}
\section{Cohen-Macaulay approximation in fibred categories}
Given a category \(\cat{C}\) and a category \(\cat{A}\) fibred in abelian categories over \(\cat{C}\). Base change by an \(f:T\rightarrow T'\) in \(\cat{C}\) applied to the objects in a complex \(\dots\rightarrow N_{d}\rightarrow N_{d-1}\rightarrow\dots\) in \(\cat{A}(T)\) can by Lemma \ref{lem.keystable} be uniquely extended to a complex and yield a commutative diagram where the vertical arrows are the cocartesian base change arrows:
\begin{equation*}
\xymatrix@C-0pt@R-12pt@H0pt{
\dots \ar[r] & N_{d+1} \ar[r]\ar[d] & N_{d} \ar[r]\ar[d] & N_{d-1} \ar[d]\ar[r] & \dots \\
\dots \ar[r] & N_{d+1}^{\#} \ar[r] & N_{d}^{\#} \ar[r] & N_{d-1}^{\#} \ar[r] & \dots
}
\end{equation*}
Similarly base change of a commutative diagram \(\Delta\) in \(\cat{A}(T)\) gives a commutative diagram \(\Delta^{\#}\) and the base change arrows give an arrow of diagrams \(\Delta\rightarrow\Delta^{\#}\).
Let \(\cat{X}\subseteq \cat{A}\) be a FAd subcategory.
Consider the following two conditions on the pair \((\cat{A},\cat{X})\) and an object \(T\) in \(\cat{C}\).
\begin{enumerate}
\item[(BC1)] If \(\alpha:A_{1}\rightarrow A_{2}\) is an epimorphism in \(\cat{A}(T)\) and \(f:T\rightarrow T'\) is an arrow in \(\cat{C}\) then any base change of \(\alpha\) by \(f\) is an epimorphism in \(\cat{A}(T')\).
\item[(BC2)] Let \(\xi:0\rightarrow A\rightarrow B\rightarrow M\rightarrow 0\) be an exact sequence in \(\cat{A}(T)\) with \(M\) in \(\cat{X}(T)\) and \(f:T\rightarrow T'\) is an arrow in \(\cat{C}\). Then any base change of \(\xi\) by \(f\) is an exact sequence in \(\cat{A}(T')\).
\end{enumerate}
The first condition would be satisfied if base change had a right adjoint. The second condition mimics flatness for all objects in \(\cat{X}(T)\).
The following is an elementary, but essential technical consequence of BC1.
\begin{lem}\label{lem.bs}
Let \(\cat{A}\) be a category fibred in abelian categories over \(\cat{C}\) which satisfies \textup{BC1} for \(T\) in \(\cat{C}\)\textup{.} Let \(c:\dots\rightarrow L_{n}\rightarrow L_{n-1}\rightarrow\dots\) be an acyclic complex in \(\cat{A}(T)\) which remains exact after a base change \(\dots\rightarrow L_{n}^{\#}\rightarrow L_{n-1}^{\#}\rightarrow\dots\) of \(c\) by \(f:T\rightarrow T'\)\textup{.} Then base change of \(K_{n}:=\ker\{d_{n-1}:L_{n-1}\rightarrow L_{n-2}\}\) by \(f\) is isomorphic to \(\ker d_{n-1}^{\#}\) for all \(n\)\textup{.}
\end{lem}
\begin{proof}
Let \(Q_{n}=\ker d_{n-1}^{\#}\). Since the composition \(K_{n}^{\#}\rightarrow L_{n-1}^{\#}\rightarrow L_{n-2}^{\#}\) by Lemma \ref{lem.keystable} is zero (as \(K_{n}\rightarrow K_{n}^{\#}\) is cocartesian), there is a factorisation \(\rho: K_{n}^{\#}\rightarrow Q_{n}\) of \(K_{n}^{\#}\rightarrow L_{n-1}^{\#}\). On the other hand the composition \(L_{n+1}^{\#}\rightarrow L_{n}^{\#}\rightarrow K_{n}^{\#}\) is zero too, hence there is an arrow from \(\coker d_{n+1}^{\#}\cong Q_{n}\) to \(K_{n}^{\#}\) which is a section of \(\rho\). By assumption \(L_{n}^{\#}\rightarrow K_{n}^{\#}\) is an epimorphism. It follows that \(Q_{n}\cong K_{n}^{\#}\).
\end{proof}
\begin{defn}\label{defn.f}
Given FAd subcategories \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\).
Let \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) denote the additive subcategory of \(\cat{A}(T)\) with objects \(N\) which have a finite \(\cat{X}\)-resolution \(M_{*}\rightarrow N\) which is preserved as resolution by any base change. Let \(\hat{\cat{X}}{}^{\textnormal{fl}}\subseteq \cat{A}\) denote the resulting FAd subcategory.
Let \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) denote the additive subcategory of \(\cat{A}(T)\) with objects \(L\) which have a \(\cat{D}(T)\)-resolution \(D^{*}\rightarrow L\) which is preserved as resolution by any base change. Let \(\hat{\cat{D}}{}^{\textnormal{fl}}\subseteq\cat{A}\) denote the resulting FAd subcategory.
\end{defn}
The reasoning in the beginning of this section combined with Lemma \ref{lem.res} gives the following.
\begin{lem}\label{lem.forall}
Let \(\eta:\dots \rightarrow E_{n}\rightarrow E_{n-1}\rightarrow\dots\) and \(\lambda:\dots\rightarrow F_{n}\rightarrow F_{n-1}\rightarrow\dots\) be complexes in \(\cat{A}(T)\) and \(\eta^{\#}\) and \(\lambda^{\#}\) the complexes resulting from base change over \(f:T\rightarrow T'\)\textup{.} If \(\eta\) is homotopic to \(\lambda\) then \(\eta^{\#}\) is homotopic to \(\lambda^{\#}\)\textup{.}
In particular\textup{;} if \(N\) in \(\cat{A}(T)\) has one \(\cat{DX}\)-resolution \textup{(}\(\cat{\hat{D}D}\)-coresolution\textup{)} which is preserved by base change then all \(\cat{DX}\)-resolutions \textup{(}\(\cat{\hat{D}D}\)-coresolutions\textup{)} are preserved by base change\textup{.}
\end{lem}
\begin{thm}\label{thm.cofapprox}
Let \(\cat{A}\) be a category fibred in abelian categories over \(\cat{C}\) and let \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\) be inclusion morphisms of categories fibred in additive categories\textup{.} Fix an object \(T\) in \(\cat{C}\)\textup{.} Assume \textup{BC1-2} for \((\cat{A},\cat{X})\) and \(T\)\textup{,} and \textup{AB1-2} for the triple of categories \((\cat{A}(T),\cat{X}(T),\cat{D}(T))\)\textup{.}
Then any object \(N\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) admits an \(\cat{X}(T)\)-approximation and a \(\hat{\cat{D}}(T)\)-hull\textup{;}
\begin{equation*}\label{eq.cofapprox}
0\rightarrow L\longrightarrow M\longrightarrow N\rightarrow 0 \quad \text{and} \quad 0\rightarrow N\longrightarrow L'\longrightarrow M'\rightarrow 0
\end{equation*}
with \(M\) and \(M'\) in \(\cat{X}(T)\) and \(L\) and \(L'\) in \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\)\textup{,} which are preserved by any base change\textup{.}
\end{thm}
\begin{proof}
The proof is a variation of the original proof of \cite[1.1]{aus/buc:89}.
For every \(N\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) let \(r(N)\) denote the minimal length of an \(\cat{X}(T)\)-resolution \(M_{*}\twoheadrightarrow N\) which is preserved by base change. The proof is by induction on \(r(N)\). If \(r(N)=0\) then \(N\) is in \(\cat{X}\) and so is its own \(\cat{X}\)-approximation, while AB2 provides a short exact sequence \(N\rightarrow D\rightarrow M'\) which is a \(\hat{\cat{D}}(T)\)-hull with \(D\) in \(\cat{D}(T)\subseteq\hat{\cat{D}}{}^{\textnormal{fl}}(T)\). The approximation is trivially preserved by base change, the hull because of BC2. Assume \(r=r(N)>0\) and let \(0\rightarrow M_{r}\rightarrow\dots\rightarrow M_{0}\twoheadrightarrow N\) be an \(\cat{X}(T)\)-resolution of minimal length preserved by base change. Then \(N_{1}=\ker(M_{0}\twoheadrightarrow N)\) is in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) by Lemma \ref{lem.bs} and \(r(N_{1})=r-1\). By induction there is a \(\hat{\cat{D}}(T)\)-hull \(N_{1}\rightarrow L\rightarrow M'_{1}\) with \(L\) in \(\hat{\cat{D}}{}^{\textnormal{fl}}\) which is preserved by base change. Pushout of \(e:N_{1}\rightarrow M_{0}\rightarrow N\) along \(N_{1}\rightarrow L\) gives an \(\cat{X}(T)\)-approximation \(L\rightarrow M\rightarrow N\) by AB1. In the commutative diagram obtained by a base change;
\begin{equation}
\xymatrix@C-0pt@R-12pt@H0pt{
N_{1}^{\#} \ar[r]\ar[d] & M_{0}^{\#} \ar[r]\ar[d] & N^{\#} \ar@{=}[d] \\
L^{\#} \ar[r]\ar[d] & M^{\#} \ar[r]\ar[d] & N^{\#} \\
(M'_{1})^{\#} \ar@{=}[r] & (M'_{1})^{\#}
}
\end{equation}
the upper row (by Lemma \ref{lem.bs}) and the columns (by BC2) are short exact sequences.
It follows that the middle row is a short exact sequence.
By AB2 there is a short exact sequence \(M\rightarrow D\rightarrow M'\) with \(D\) in \(\cat{D}(T)\) and \(M'\) in \(\cat{X}(T)\). Pushout of \(M\rightarrow D\rightarrow M'\) along \(M\rightarrow N\) gives a short exact sequence \(h:N\rightarrow L'\rightarrow M'\). Since the induced sequence \(L\rightarrow D\rightarrow L'\) is short exact, \(L'\) is contained in \(\hat{\cat{D}}(T)\). Applying a base change we obtain the following commutative diagram:
\begin{equation}
\xymatrix@C-0pt@R-12pt@H0pt{
L^{\#} \ar[r]\ar@{=}[d] & M^{\#} \ar[r]\ar[d] & N^{\#} \ar[d] \\
L^{\#} \ar[r] & D^{\#} \ar[r]\ar[d] & (L')^{\#} \ar[d] \\
& (M')^{\#} \ar@{=}[r] & (M')^{\#}
}
\end{equation}
The upper row and (by BC2) the two columns are short exact sequences. It follows that the middle row is a short exact sequence and hence that \(L'\) is contained in \(\hat{\cat{D}}{}^{\textnormal{fl}}\).
\end{proof}
Sequences as in Theorem \ref{thm.cofapprox} preserved by any base change will be called an \emph{\(\cat{X}\)-approximation} and a \emph{\(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull} of \(N\) respectively.
Lemma \ref{lem.stable} makes the following definition reasonable. Three categories fibred in additive categories (FAds) \(\cat{A}_{i}\), \(i=1,2,3\), an inclusion of FAds \(\cat{A}_{1}\subseteq \cat{A}_{2}\), and a morphism of FAds \(F:\cat{A}_{2}\rightarrow \cat{A}_{3}\) equivalent to the quotient morphism \(\cat{A}_{2}\rightarrow\cat{A}_{2}/\cat{A}_{1}\) is called a \emph{short exact sequence of categories fibred in additive categories} and is denoted by \(0\rightarrow \cat{A}_{1}\rightarrow\cat{A}_{2}\rightarrow\cat{A}_{3}\rightarrow 0\).
\begin{thm}\label{thm.cofmain}
Let \(\cat{A}\) be a category fibred in abelian categories over \(\cat{C}\) and let \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\) be inclusion morphisms of categories fibred in additive categories\textup{.} Assume \textup{BC1-2} for the pair \((\cat{A},\cat{X})\) and \textup{AB1-3} for the triple of categories \((\cat{A}(T),\cat{X}(T),\cat{D}(T))\)\textup{,} for all objects \(T\) in \(\cat{C}\)\textup{.}
Then\textup{:}
\begin{enumerate}
\item[(i)] The \(\cat{X}\)-approximation induces a morphism of categories fibred in additive categories \(j^{!}:\hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\rightarrow\cat{X}/\cat{D}\) which is a right adjoint to the full and faithful inclusion morphism \(j_{!}:\cat{X}/\cat{D}\rightarrow\hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\)\textup{.}
\item[(ii)] The \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull induces a morphism of categories fibred in additive categories \(i^{*}:\hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\rightarrow\hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\) which is a left adjoint to the full and faithful inclusion morphism \(i_{*}:\hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\rightarrow \hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\)\textup{.}
\item[(iii)] Together these maps give the following commutative diagram of short exact sequences of categories fibred in additive categories\textup{:}
\begin{equation*}
\xymatrix@C-0pt@R-8pt@H-30pt{
0\ar[r] & \, \hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D} \ar[r]^{i_{*}}\ar[d]_{\id} & \hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\ar[r]^{j^{!}}\ar@{=}[d] & \cat{X}/\cat{D} \ar[r] & 0 \\
0 & \hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\ar[l] & \hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\ar[l]_{i^{*}} & \, \cat{X}/\cat{D} \ar[l]_{j_{!}}\ar[u]_{\id} & 0 \ar[l]
}
\end{equation*}
\end{enumerate}
\end{thm}
\begin{proof}
In each fibre most of these statements are true by the arguments in the proof of \cite[2.8]{aus/buc:89} since we have Theorem \ref{thm.cofapprox}. The general cases are reduced to fibre cases by applying base change. First we have to establish the functors. Note that the quotient categories involved are FAds over \(\cat{C}\) by Lemma \ref{lem.stable}. Let \(p:\cat{A}\rightarrow\cat{C}\) denote the fibration. For each \(N_{i}\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}\) put \(T_{i}=p(N_{i})\) and choose a \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull \(\iota_{i}:N_{i}\rightarrow L_{i}\rightarrow M_{i}\) which exists by Theorem \ref{thm.cofapprox} and such that \(\iota_{i}=\id\) if \(N_{i}\) is in \(\hat{\cat{D}}{}^{\textnormal{fl}}\). For each arrow \(\psi:N_{1}\rightarrow N_{2}\) choose an arrow \(\lambda_{21}:L_{1}\rightarrow L_{2}\) commuting with \(\psi\). This arrow is obtained as a composition of a base change \(L_{1}\rightarrow L_{1}^{\#}\) over \(p(\psi):T_{1}\rightarrow T_{2}\) with an extension \(L_{1}^{\#}\rightarrow L_{2}\) of \(N_{1}^{\#}\rightarrow L_{2}\) obtained since \(\xt{1}{\cat{A}(T_{2})}{M_{1}^{\#}}{L_{2}}=0\) by \cite[2.5]{aus/buc:89}. If composeable it follows from \cite[2.8]{aus/buc:89} that \(\lambda_{32}\lambda_{21}\sim\lambda_{31}\).
There is a unique arrow \(\phi:N_{1}^{\#}\rightarrow N_{2}\) induced by \(\psi\).
If \(\lambda_{21}':L_{1}\rightarrow L_{2}\) is an extension of \(\psi':N_{1}\rightarrow N_{2}\) with \(p(\psi')=p(\psi)\) such that \(\delta_{1}:=\psi-\psi'\) is equivalent to \(0\), we have by Lemma \ref{lem.keystable} that \(\delta:N_{1}^{\#}\rightarrow N_{2}\) induced from \(\delta_{1}\) by base change factors through an object \(D\) in \(\cat{D}(T_{2})\). It follows that \(\delta\) factors through \(N_{1}^{\#}\rightarrow L_{1}^{\#}\). Let \(\tau\) denote the composition \(L_{1}^{\#}\rightarrow N_{2}\rightarrow L_{2}\) (so \(\tau\sim 0\)).
Let \(\eta\) be a base change over \(p(\psi)\) of the difference of the two extensions; \(\eta=(\lambda_{21}-\lambda_{21}')^{\#}\). One calculates that \((\eta-\tau)\iota_{1}^{\#}=0\), hence \(\eta-\tau\) is induced by an arrow \(M_{1}^{\#}\rightarrow L_{2}\) which lifts to an arrow \(M_{1}^{\#}\rightarrow D^{0}\) where \(D^{*}\twoheadrightarrow L_{2}\) is a finite \(\cat{D}\)-resolution of \(L_{2}\) (since \(\xt{1}{\cat{A}(T_{2})}{M_{1}^{\#}}{\hat{\cat{D}}(T_{2})}=0\)).
Hence \(\eta-\tau\sim 0\), so \(\eta\sim 0\) and \(\lambda_{21}\sim\lambda_{21}'\). We have shown that \(i^{*}:\hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\rightarrow \hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\) is a well defined functor. To show that \(i^{*}\) preserves cocartesian arrows we apply Lemma \ref{lem.stablecc}: If \([\psi]:N_{1}\rightarrow N_{2}\) is cocartesian in \(\hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\) then the induced map \([\phi]:N_{1}^{\#}\rightarrow N_{2}\) is an isomorphism and by \cite[2.8]{aus/buc:89} so is any extension \(L_{1}^{\#}\rightarrow L_{2}\) of \([\phi]\). Composed with the base change \(L_{1}\rightarrow L_{1}^{\#}\) we get a cocartesian arrow in \(\hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\) by Lemma \ref{lem.stablecc}.
A similar argument gives that the morphism \(j^{!}: \hat{\cat{X}}{}^{\textnormal{fl}}/\cat{D}\rightarrow \cat{X}/\cat{D}\) induced by (choices of) \(\cat{X}\)-approximation also is well defined as a map of fibred categories.
To prove adjointness for the pair \((j_{!},j^{!})\) consider the chosen \(\cat{X}\)-approximation \(L\rightarrow M\xra{\pi} N\) of \(N\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\). Given \(\phi_{1}:M_{1}\rightarrow N\) with \(M_{1}\) in \(\cat{X}(T_{1})\) and \(f=p(\phi_{1})\). Let \(\phi: M_{1}^{\#}\rightarrow N\) be induced by a base change of \(M_{1}\) by \(f\). Since \(\xt{1}{\cat{A}(T)}{M_{1}^{\#}}{L}=0\), \(\phi\) can be lifted to an arrow \(\psi:M_{1}^{\#}\rightarrow M\). Composing \(\psi\) with the base change \(M_{1}\rightarrow M_{1}^{\#}\) gives a lifting of \(\phi_{1}\) which shows surjectivity of the adjointness map \([\pi\circ -]\). To prove injectivity consider for \(i=2,3\) arrows \(\psi_{i}:M_{1}\rightarrow M\) in \(\cat{X}\) with \(\pi\psi_{2}=\pi\psi_{3}\). Since \(p(\pi)=\id\) we have \(p(\psi_{2})=p(\psi_{3})=f\) and we can define \(\psi_{1}=\psi_{2}-\psi_{3}\) with \(\pi\psi_{1}\sim 0\). Base change by \(f\) induces a \(\psi:M_{1}^{\#}\rightarrow M\) from \(\psi_{1}\). Lemma \ref{lem.keystable} gives \(\pi\psi\sim 0\). The argument in \cite[2.8]{aus/buc:89} implies that \(\psi\) and hence \(\psi_{1}\) factors through an object in \(\cat{D}(T)\). Analogous arguing gives the adjointness of the pair \((i^{*},i_{*})\).
The commutativity of the diagram in \textup{(iii)} follows by definition. For \(i^{*}j_{!}=0=j^{!}i_{*}\) see \cite[2.8]{aus/buc:89}. We prove exactness in the upper row. Given \(\phi:N_{1}\rightarrow N_{2}\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}\) with \(f=p(\phi):T_{1}\rightarrow T_{2}\) such that \(j^{!}[\phi]=0\). If \(\pi_{i}:M_{i}\rightarrow N_{i}\) are the chosen \(\cat{X}\)-approximations, \(j^{!}[\phi]\) is represented by a lifting \(\psi:M_{1}\rightarrow M_{2}\) and the assumption is that \(\psi\) factors through an object \(D\) of \(\cat{D}\). We claim that \(\phi\) factors through an object in \(\hat{\cat{D}}{}^{\textnormal{fl}}\). By base change it's sufficient to prove the special case \(f=\id_{T_{2}}\). If \(M\) is any object in \(\cat{X}(T)\) we have that the composition \(\xt{1}{\cat{A}(T)}{M}{N_{1}}\cong \xt{1}{\cat{A}(T)}{M}{M_{1}} \xra{\psi_{*}}\xt{1}{\cat{A}(T)}{M}{M_{2}}\cong \xt{1}{\cat{A}(T)}{M}{N_{2}}\) is \(\phi_{*}\) which hence equals \(0\). If \(e:N_{1}\rightarrow L_{1}'\rightarrow M_{1}'\) is a \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull of \(N_{1}\), the connecting takes \(\phi\) to \(\phi_{*}e\in \xt{1}{\cat{A}(T)}{M_{1}'}{N_{2}}\), i.e.\ to \(0\), and so there exists a \(\delta:L_{1}'\rightarrow N_{2}\) which induces \(\phi\). Exactness in the lower row is analogous.
\end{proof}
\begin{prop}\label{prop.cof}
Let \(\cat{A}\) be a category fibred in abelian categories over \(\cat{C}\) and let \(\cat{D}\subseteq\cat{X}\subseteq\cat{A}\) be inclusion morphisms of categories fibred in additive categories\textup{.} Fix an object \(T\) in \(\cat{C}\)\textup{.} Assume \textup{BC1-2} for \((\cat{A},\cat{X})\) and \(T\), and \textup{AB1-3} for the triple of categories \((\cat{A}(T),\cat{X}(T),\cat{D}(T))\)\textup{.} Then\textup{:}
\begin{enumerate}
\item[(i)] \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)=\hat{\cat{X}}{}^{\textnormal{fl}}(T)\cap \cat{X}(T)^{\perp}\) and \(\cat{D}(T)=\cat{X}(T)\cap \hat{\cat{D}}{}^{\textnormal{fl}}(T)\)\textup{.}
\item[(ii)] \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) and \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) are closed under extensions\textup{.}
\item[(iii)] Exact sequences \(\dots\rightarrow N_{n}\xra{d_{n}} N_{n-1}\rightarrow \dots\) with objects \(N_{i}\) and kernels \(\ker d_{i}\) in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) remain exact after base change\textup{.}
\end{enumerate}
If in addition \textup{AB4}\textup{,} then\textup{:}
\begin{enumerate}
\item[(iv)] Epimorphisms in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) are admissible\textup{.}
\item[(v)] \(\Add \hat{\cat{X}}{}^{\textnormal{fl}}(T) =\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) and \(\Add \hat{\cat{D}}{}^{\textnormal{fl}}(T) = \hat{\cat{D}}{}^{\textnormal{fl}}(T)\)\textup{.}
\end{enumerate}
\end{prop}
\begin{proof}
For (i) the proofs of \cite[3.6-7]{aus/buc:89} works with the fl-s too by Theorem \ref{thm.cofapprox}. By (i) it's sufficient to prove (ii) for \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\). Let \(e: N_{1}\rightarrow N_{2}\rightarrow N_{3}\) be a short exact sequence in \(\hat{\cat{X}}(T)\).
If \(N_{1}\) and \(N_{3}\) are in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\), Theorem \ref{thm.cofapprox} and Lemma \ref{lem.forall} implies that \({}^{-}C^{*}(N_{1})\) and \({}^{-}C^{*}(N_{3})\) in Lemma \ref{lem.forall} are preserved as resolutions by base change. Together with BC2 this implies that base change of the short exact sequence of resolutions in Lemma \ref{lem.xres} (i) gives a short exact sequence of \(\cat{DX}\)-resolutions.
For (iii) the long exact sequence is broken into short exact sequences \(\xi_{i}\) with objects in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\). By Lemma \ref{lem.bs} it is sufficient to prove the claim for short exact sequences. By Lemma \ref{lem.xres} the short exact sequence of \(\cat{DX}\)-resolutions in Lemma \ref{lem.xres} (i) is preserved by base change. It follows that the short exact sequences \(\xi_{i}\) remain exact after base change.
For (iv); if \(N_{2}\) and \(N_{3}\) in \(e\) are in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) then \(N_{1}\) is in \(\hat{\cat{X}}(T)\) by AB4, see \cite[3.5]{aus/buc:89}. The argument proceeds as for extensions.
By (i) it's sufficient to prove (v) for \(\hat{\cat{X}}{}^{\textnormal{fl}}\). If \(N_{1}\coprod N_{2}\) is an object in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\), then \(N_{i}\) is in \(\hat{\cat{X}}(T)\) for \(i=1,2\) by \cite[3.4]{aus/buc:89}. By Lemma \ref{lem.forall} \({}^{-}C^{*}(N_{1})\coprod {}^{-}C^{*}(N_{2})\) is preserved by base change as resolution of \(N_{1}\coprod N_{2}\). It follows that the resolution \({}^{-}C^{*}(N_{i})\) is preserved by base change for \(i=1,2\).
\end{proof}
\begin{cor}\label{cor.Cbs}
Assume \textup{BC1-2} and \textup{AB1-3} as in \textup{Proposition \ref{prop.cof}.}
If \(N\) is in \(\hat{\cat{X}}{}^{\textnormal{fl}}(T)\) then any \(\cat{DX}\)-resolution \({}^{-}C^{*}(N)\twoheadrightarrow N\) and any \(\hat{\cat{D}}\cat{D}\)-coresolution \(N\rightarrowtail {}^{+}C^{*}(N)\) as in \textup{Section \ref{subsec.cplx}} is preserved by base change\textup{.}
\end{cor}
\begin{proof}
By Theorem \ref{thm.cofapprox} there exist a \(\cat{DX}\)-resolution and a \(\cat{\hat{D}D}\)-coresolution in \(\hat{\cat{X}}{}^{\textnormal{fl}}\).
The result follows from Proposition \ref{prop.cof} (iii) and Lemma \ref{lem.forall}.
\end{proof}
\begin{defn}\label{defn.cof}
Let \(\cat{A}\) be a category fibred in abelian categories over \(\cat{C}\) and let \(\cat{D}\subseteq\cat{A}\) be an inclusion morphism of a category fibred in additive categories. For \(T\) in \(\cat{C}\) let \(\check{\cat{D}}{}^{\textnormal{fl}}(T)\) denote the full subcategory of \(\cat{A}(T)\) of objects \(K\) with a finite coresolution \(K\rightarrow L^{*}\) with objects \(L^{i}\) in \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) for \(i\geq 0\).
\end{defn}
\begin{lem}\label{lem.cof2}
With these notions we have\textup{:}
\begin{enumerate}
\item[(i)] Epimorphisms in \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) are admissible if and only if\, \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)=\check{\cat{D}}{}^{\textnormal{fl}}(T)\)\textup{.}
\end{enumerate}
Assume \textup{BC1-2} and \textup{AB1-4} for \((\cat{A}(T),\cat{X}(T),\cat{D}(T))\)\textup{.} Then\textup{:}
\begin{enumerate}
\item[(ii)] \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)=\hat{\cat{D}}(T)\cap\check{\cat{D}}{}^{\textnormal{fl}}(T)\)\textup{.}
\item[(iii)] Epimorphisms in \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) are admissible if epimorphisms in \(\hat{\cat{D}}(T)\) are admissible\textup{.}
\end{enumerate}
\end{lem}
\begin{proof}
(i) is trivially true. In (ii) \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\subseteq\hat{\cat{D}}(T)\cap\check{\cat{D}}{}^{\textnormal{fl}}(T)\) is obvious. For the other inclusion, suppose \(K\) is an object in \(\hat{\cat{D}}(T)\cap\check{\cat{D}}{}^{\textnormal{fl}}(T)\setminus \hat{\cat{D}}{}^{\textnormal{fl}}(T)\) with a \(\hat{\cat{D}}{}^{\textnormal{fl}}(T)\)-coresolution \(K\rightarrow L^{*}\) of length \(n>0\). Since monomorphisms are admissible in \(\hat{\cat{D}}(T)=\hat{\cat{X}}(T)\cap \cat{X}(T)^{\perp}\), all \(K^{i}=\ker(L^{i}\rightarrow L^{i+1})\) are contained in \(\hat{\cat{D}}(T)\), and we can assume \(n=1\). But then \(K\) has to be in \(\hat{\cat{X}}{}^{\textnormal{fl}}\cap\hat{\cat{D}}=\hat{\cat{D}}{}^{\textnormal{fl}}(T)\) by Proposition \ref{prop.cof}. Since (i) is true ``without the fl'' by \cite[4.1]{aus/buc:89}, (iii) follows immediately from (i) and (ii).
\end{proof}
\section{Cohen-Macaulay approximation of flat families}\label{sec.ft}
We define fibred categories of Cohen-Macaulay maps with flat modules and show that they allow Cohen-Macaulay approximation in the finite type case and the local, algebraic case.
\subsection{The finite type case}\label{subsec.ft}
Let \(h:S\rightarrow T\) be a ring homomorphism of noetherian rings. We say that \(h\) is a \emph{Cohen-Macaulay \textup{(}CM\textup{)} map} if it is of finite type, faithfully flat and all fibres are Cohen-Macaulay (cf.\ \cite[6.8.1]{EGAIV2}). In particular \(h\) is equidimensional (\cite[15.4.1]{EGAIV3}). B.\ Conrad has defined the \emph{dualising module} \(\omega_{h}\) for any CM \(h\), see \cite[Sec.\ 3.5]{con:00}.
Suppose \(h\) has pure relative dimension \(n\). For some \(N\geq n\) there is a surjective \(S\)-algebra map \(P\rightarrow T\) where \(P=S[t_{1},\dots,t_{N}]\). Let \(\omega_{P/S}:=\wedge^{N}\Omega_{P/S}\). Then there is an isomorphism \(\omega_{h}\cong\xt{N-n}{P}{T}{\omega_{P/S}}\) which is natural in the factorisation \(S\rightarrow P\rightarrow T\), see \cite[3.5.3-6]{con:00}. By (local) duality theory and Corollary \ref{cor.xtdef} \(\omega_{h}\) is \(S\)-flat
(or see \cite[Cor.\ 3.5.2]{con:00}).
If \(S\) is a field, we have that \(\omega_{h}\) is a canonical module of \(T\) as in Example \ref{ex.MCMapprox}, cf.\ \cite[3.3.7 and 16]{bru/her:98}.
Let \(\cat{CM}\) be the category with objects the CM maps and morphisms \((g,f):h_{1}\rightarrow h_{2}\) pairs of ring homomorphisms \(g:S_{1}\rightarrow S_{2}\) and \(f:T_{1}\rightarrow T_{2}\) such that \(h_{2}g=fh_{1}\) and such that the induced map \(f{\otimes} 1:T_{1}{\otimes} S_{2}\rightarrow T_{2}\) is an isomorphism:
\begin{equation*}
\xymatrix@C-0pt@R-8pt@H-30pt{
T_{1}\ar[r]^{f} & T_{2} & T_{1}{\otimes}_{S_{1}}S_{2} \ar[l]_(0.6){\simeq} \\
S_{1}\ar[u]^{h_{1}}\ar[r]^{g} & S_{2}\ar[u]_{h_{2}}
}
\end{equation*}
Let \(\cat{NR}\) denote the category of noetherian rings. The forgetful functor \(p:\cat{CM}\rightarrow\cat{NR}\); \((g,f)\mapsto g\), makes \(\cat{CM}\) fibred in groupoids over \(\cat{NR}\). The essential part is that \(\cat{CM}\) should allow base change, i.e.\ given \(g:S_{1}\rightarrow S_{2}\) and \(h_{1}:S_{1}\rightarrow T_{1}\) as above there should exist a \(T_{2}\), an \(h_{2}:S_{2}\rightarrow T_{2}\) and an \(f\) such that \((g,f)\) is a morphism \(h_{1}\rightarrow h_{2}\) in \(\cat{CM}\). This follows from \cite[15.4.3]{EGAIV3}.
Let \(\cat{mod}\) be the category of pairs \((h:S\rightarrow T,N)\) with \(h\) in \(\cat{CM}\) and \(N\) a finite \(T\)-module. A morphism \((h_{1},N_{1})\rightarrow (h_{2},N_{2})\) is a morphism \((g,f):h_{1}\rightarrow h_{2}\) in \(\cat{CM}\) and a \(f\)-linear map \(\alpha:N_{1}\rightarrow N_{2}\). Then \(\alpha\) is cocartesian with respect to the forgetful functor \(F:\cat{mod}\rightarrow \cat{CM}\) if \(1{\otimes} \alpha: T_{2}{\otimes} N_{1}\rightarrow N_{2}\) is an isomorphism. It follows that \(\cat{mod}\) is fibred in abelian categories over \(\cat{CM}\).
Adding the property that \(N\) is \(S\)-flat gives the full subcategory \(\cat{mod}{}^{\textnormal{fl}}\). Moreover, let \(\cat{MCM}\) be the full subcategory of \(\cat{mod}{}^{\textnormal{fl}}\) where the fibre \(N_{s}=N{\otimes}_{S}k(s)\) is a maximal Cohen-Macaulay \(T_{s}\)-module for all \(s\in \Spec S\). The inclusions \(\cat{MCM}\subseteq\cat{mod}{}^{\textnormal{fl}}\subseteq\cat{mod}\) are inclusion morphisms of categories fibred in additive categories (FAds) over \(\cat{CM}\). For \(\cat{MCM}\) this follows from \cite[15.4.3]{EGAIV3}.
If \(h\) is a CM map let \(\cat{mod}_{h}\), \(\cat{MCM}_{h}\), \dots denote the fibre categories of \(\cat{mod}\), \(\cat{MCM}\), \dots over \(h\). An object in \(\cat{MCM}_{h}\) is called an (\(h\)-)family of maximal Cohen-Macaulay modules.
Given a morphism \(h_{1}\rightarrow h_{2}\) in \(\cat{CM}\). By \cite[Thm.\ 3.6.1]{con:00} there is a natural isomorphism with base change \(T_{2}{\otimes}\omega_{h_{1}}\cong\omega_{h_{2}}\) which is compatible with localisation of \(T_{1}\) and is functorial with respect to composition \(h_{1}\rightarrow h_{2}\rightarrow h_{3}\). It follows that \(h\mapsto (h,\omega_{h})\) defines a morphism \(\omega:\cat{CM}\rightarrow\cat{MCM}\) of fibred categories over \(\cat{NR}\) which is a section of the forgetful \(F:\cat{MCM}\rightarrow\cat{CM}\). Let \(\cat{D}\) be the full subcategory of \(\cat{MCM}\) over \(\cat{CM}\) with the objects \((h,D)\) where \(D\) is an object in \(\Add\{\omega_{h}\}\). The inclusion \(\cat{D}\subseteq\cat{MCM}\) is an inclusion of FAds over \(\cat{CM}\).
If \(\cat{U}\) denotes any of these FAds over \(\cat{CM}\), let \(\ul{\cat{U}}\) denote the quotient (`stable') category \(\cat{U}/\cat{D}\). With this notation we have the following.
\begin{thm}\label{thm.flatCMapprox}
The pair \((\cat{mod},\cat{MCM})\) over \(\cat{CM}\) satisfies \textup{BC1-2} and the triple of fibre categories \((\cat{mod}_{h},\cat{MCM}_{h},\cat{D}_{h})\) satisfies \textup{AB1-4} for all objects \(h\) in \(\cat{CM}\)\textup{.}
Moreover\textup{:}
\begin{enumerate}
\item[(i)] The fibred categories \(\widehat{\cat{MCM}}{}^{\textnormal{fl}}\) and \(\hat{\cat{D}}{}^{\textnormal{fl}}\) equals \(\cat{mod}{}^{\textnormal{fl}}\) and \(\hat{\cat{D}}\cap\cat{mod}{}^{\textnormal{fl}}\) respectively\textup{.}
\item[(ii)] For any object \((h,N)\) in \(\cat{mod}{}^{\textnormal{fl}}\), \(N\) admits an \(\cat{MCM}\)-approximation and a \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull which in particular are preserved by any base change\textup{.}
\item[(iii)] The \(\cat{MCM}\)-approximation induces a morphism of categories fibred in additive categories \(j^{!}:\underline{\cat{mod}}{}^{\textnormal{fl}}\rightarrow\ul{\cat{MCM}}\) which is a right adjoint to the full and faithful inclusion morphism \(j_{!}:\ul{\cat{MCM}}\rightarrow\underline{\cat{mod}}{}^{\textnormal{fl}}\)\textup{.}
\item[(iv)] The \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull induces a morphism of categories fibred in additive categories \(i^{*}:\underline{\cat{mod}}{}^{\textnormal{fl}}\rightarrow{\underline{\hat{\cat{D}}}{}^{\textnormal{fl}}}\) which is a left adjoint to the full and faithful inclusion morphism \(i_{*}:\underline{\hat{\cat{D}}}{}^{\textnormal{fl}}\rightarrow \underline{\cat{mod}}{}^{\textnormal{fl}}\)\textup{.}
\item[(v)] Together these maps give the following commutative diagram of short exact sequences of categories fibred in additive categories\textup{:}
\begin{equation*}
\xymatrix@C-0pt@R-8pt@H-30pt{
0\ar[r] & \,\underline{\hat{\cat{D}}}{}^{\textnormal{fl}} \ar[r]^(0.45){i_{*}}\ar[d]_(0.45){\id} & \underline{\cat{mod}}{}^{\textnormal{fl}}\ar[r]^{j^{!}}\ar@{=}[d] & \ul{\cat{MCM}}\ar[r] & 0 \\
0 & \,\underline{\hat{\cat{D}}}{}^{\textnormal{fl}}\ar[l] & \underline{\cat{mod}}{}^{\textnormal{fl}}\ar[l]_{i^{*}} & \ul{\cat{MCM}}\ar[l]_(0.45){j_{!}}\ar[u]_{\id} & 0 \ar[l]
}
\end{equation*}
\end{enumerate}
\end{thm}
\begin{proof}
Since base change is given by the tensor product BC1 and BC2 follows. In particular, a short exact sequence \(e:N_{1}\rightarrow N_{2}\rightarrow N_{3}\) in \(\cat{mod}_{h}\) with \(N_{3}\) and either \(N_{1}\) or \(N_{2}\) in \(\cat{MCM}_{h}\) gives short exact sequences of MCMs after base change to each fibre \(T_{s}\), i.e.\ AB1 and AB4.
For AB2, suppose \(M\) is in \(\cat{MCM}_{h}\). Then \(M^{\vee}=\hm{}{T}{M}{\omega}\) is in \(\cat{MCM}_{h}\) too by Corollary \ref{cor.xtdef}. Since \(M^{\vee}\) is finite there is a short exact sequence
\(M^{\vee}\leftarrow T^{r}\leftarrow M_{1}\). By AB1 \(M_{1}\) is in \(\cat{MCM}_{h}\). Applying \(\hm{}{T}{-}{\omega_{h}}\) gives the desired short exact sequence since Corollary \ref{cor.xtdef} implies that \(\xt{1}{T}{M^{\vee}}{\omega_{h}}=0\) and that the natural map \(M\rightarrow M^{\vee\vee}\) is an isomorphism. AB3 also follows from Corollary \ref{cor.xtdef}.
Any \(N\) in \(\widehat{\cat{MCM}}{}^{\textnormal{fl}}_{h}\) has by definition a finite \(\cat{MCM}\)-resolution (say of length \(n\)) preserved by base change. Since objects in \(\cat{MCM}\) are \(S\)-flat it follows by induction on \(n\) that \(\tor{S}{1}{N}{k(s)}=0\) for all \(s\in\Spec S\). Hence \(N\) is \(S\)-flat. Conversely, if \(N\) is in \(\cat{mod}{}^{\textnormal{fl}}_{h}\) it follows that a sufficiently high syzygy of \(N\) is in \(\cat{MCM}_{h}\), i.e.\ \(N\) is in \(\widehat{\cat{MCM}}{}^{\textnormal{fl}}_{h}\).
With Proposition \ref{prop.cof} this gives \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}=\cat{mod}{}^{\textnormal{fl}}_{h}\cap\cat{MCM}_{h}^{\perp}\). By induction on the length of the resolution \(\hat{\cat{D}}_{h}\subseteq \cat{MCM}_{h}^{\perp}\) and so \(\cat{mod}{}^{\textnormal{fl}}_{h}\cap\hat{\cat{D}}_{h}\subseteq\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\). The opposite inclusion is clear by the first part of (i). Now (ii)-(v) follows directly from Theorem \ref{thm.cofmain}.
\end{proof}
\begin{cor}\label{cor.flatCMapprox}
Let \(h:S\rightarrow T\) be an object in \(\cat{CM}\)\textup{.} Then\textup{:}
\begin{enumerate}
\item[(i)] \(\cat{D}_{h}=\cat{MCM}_{h}^{\perp}\cap\cat{MCM}_{h}\quad\text{and}\quad \hat{\cat{D}}{}^{\textnormal{fl}}_{h}=\cat{MCM}_{h}^{\perp}\cap\cat{mod}{}^{\textnormal{fl}}_{h}\)
\item[(ii)] The kernel of a surjective map in \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\) is contained in \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\)\textup{.}
\end{enumerate}
\end{cor}
\begin{proof}
(ii): Note that if \(N_{1}\rightarrow N_{2}\rightarrow N_{3}\) is a short exact sequence with \(N_{2}\) and \(N_{3}\) in \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\), in particular \(S\)-flat, then \(N_{1}\) has to be \(S\)-flat too. To show that \(N_{1}\) is in \(\hat{\cat{D}}_{h}\) we use the criterion in \cite[4.6]{aus/buc:89} to show that \(\check{\cat{D}}_{h}=\hat{\cat{D}}_{h}\):
Suppose that \(M\) is in \(\cat{MCM}_{h}\). Assume that \(M\) satisfies \(\injdim{\cat{MCM}_{h}}{M}= n<\infty\) which by \cite[4.3]{aus/buc:89} is equivalent to the existence of a coresolution of \(M\) of length \(n\) in \(\cat{D}_{h}\). The fibre at any \(s\in\Spec S\) gives a \(\cat{D}_{T_{s}}\)-coresolution of the MCM \(T_{s}\)-module \(M_{s}\). Since \(\check{\cat{D}}_{T_{s}}=\hat{\cat{D}}_{T_{s}}\) (\cite[6.3]{aus/buc:89}) it follows by \cite[4.6]{aus/buc:89} that \(M_{s}\) is contained in \(\cat{D}_{T_{s}}\). Since \(\cat{D}_{T_{s}}=\cat{MCM}_{T_{s}}\cap\cat{MCM}_{T_{s}}^{\perp}\) by \cite[3.7]{aus/buc:89} it follows from Corollary \ref{cor.xtdef} that \(\injdim{\cat{MCM}_{h}}{M}=0\) and so \(M\) is in \(\cat{MCM}_{h}\cap\cat{MCM}_{h}^{\perp}\). But by Theorem \ref{thm.flatCMapprox} we can invoke \cite[3.7]{aus/buc:89} again which gives \(\cat{MCM}_{h}\cap\cat{MCM}_{h}^{\perp}=\cat{D}_{h}\) so \(M\) is in \(\cat{D}_{h}\). By \cite[4.6(d)]{aus/buc:89} \(\check{\cat{D}}_{h}=\hat{\cat{D}}_{h}\) follows and \(N_{1}\) is in \(\hat{\cat{D}}_{h}\).
\end{proof}
\begin{rem}
Let \(\varLambda\) be any noetherian ring. By abuse of notation let \(\cat{CM}\rightarrow{{}_{\varLambda}\cat{R}}\) denote the category fibred in groupoids obtained by restriction to the category of noetherian \(\varLambda\)-algebras \({{}_{\varLambda}\cat{R}}\). Fix a Cohen-Macaulay map \(\varLambda\rightarrow T^{\mspace{-1mu}\mathit{o}}\). There is a section \(s:{{}_{\varLambda}\cat{R}}\rightarrow\cat{CM}\) defined by \(s:S\mapsto (S\rightarrow T^{\mspace{-1mu}\mathit{o}}{\otimes} S)\). Let \(\cat{T^{o}}\) denote the resulting fibred subcategory of \(\cat{CM}\). We restrict \(\cat{mod}\), \(\cat{MCM}\) and \(\cat{D}\) to \(\cat{T^{o}}\) and obtain categories fibred in abelian and additive categories over \(\cat{T^{o}}\) respectively. These restricted fibred categories satisfy the axioms AB1-4 and BC1-2 and we obtain restricted versions of Theorem \ref{thm.flatCMapprox} and Corollary \ref{cor.flatCMapprox}.
\end{rem}
Let \(\cat{P}\) denote the fibred subcategory of \(\cat{MCM}\) over \(\cat{CM}\) of pairs \((h,P)\) with \(h:S\rightarrow T\) in \(\cat{CM}\) and \(P\) a finite projective \(T\)-module. Let \(\hat{\cat{P}}{}^{\textnormal{fl}}\) denote the full subcategory of \(\cat{mod}{}^{\textnormal{fl}}\) of pairs \((h,Q)\) such that \(Q\) has a finite projective dimension. The inclusions of categories fibred in additive categories \(\cat{P}\subseteq\hat{\cat{P}}{}^{\textnormal{fl}}\subseteq\cat{mod}{}^{\textnormal{fl}}\) are closed under extensions over \(\cat{CM}\).
\begin{lem}\label{lem.proj}
There is an exact equivalence \(\hat{\cat{D}}{}^{\textnormal{fl}}\simeq\hat{\cat{P}}{}^{\textnormal{fl}}\) of categories fibred in additive categories defined by the functor \((h,L)\mapsto(h,\hm{}{T}{\omega_{h}}{L})\) with a quasi-inverse \((h,Q)\mapsto (h,Q{\otimes}_{T}\omega_{h})\)\textup{.} It induces an equivalence of fibred quotient categories \(\hat{\cat{D}}{}^{\textnormal{fl}}/\cat{D}\simeq\hat{\cat{P}}{}^{\textnormal{fl}}/\cat{P}\)\textup{.}
\end{lem}
\begin{proof}
By base change we reduce to the absolute case where the equivalence is well known, cf.\ \cite[I 4.10.16]{has:00}. The functor \(\hm{}{-}{\omega_{-}}{-}\) takes a map \(\phi:L_{1}\rightarrow L_{2}\) over \(h_{1}\rightarrow h_{2}\) to the map \(\hm{}{T_{1}}{\omega_{h_{1}}}{L_{1}}\rightarrow\hm{}{T_{2}}{\omega_{h_{2}}}{L_{2}}\); \(\alpha\mapsto(\phi\alpha)^{\#}\), the unique map which pulls back to \(\phi\alpha\) by the cocartesian map \(\omega_{h_{1}}\rightarrow\omega_{h_{2}}\). It is a well defined functor since the dualising module is functorial.
If \(L\) is in \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\) then \(\xt{i}{T_{s}}{\omega_{T_{s}}}{L_{s}}=0\) for all \(i\neq 0\) and \(s\in \Spec S\). By Corollary \ref{cor.xtdef} the functor is exact and \(\hm{}{T}{\omega_{h}}{L}\) is \(S\)-flat. In particular \(\nd{}{T}{\omega_{h}}\cong T\). If \(D\) is in \(\cat{D}_{h}\) then \(\hm{}{T}{\omega_{h}}{D}\) is projective as a direct summand of a free module. If \(D^{*}\twoheadrightarrow L\) is a finite \(\cat{D}_{h}\)-resolution of \(L\), then \(\hm{}{T}{\omega_{h}}{D^{*}}\) gives a projective resolution of \(\hm{}{T}{\omega_{h}}{L}\) since \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\subseteq\cat{MCM}_{h}^{\perp}\) (Corollary \ref{cor.flatCMapprox}). The natural map \(\ev:\hm{}{T}{\omega_{h}}{L}{\otimes}\omega_{h}\rightarrow L\) commutes with base change to the fibres where it is an isomorphism (\cite[9.6.5]{bru/her:98}) and Nakayama's lemma and \(S\)-flatness implies that \(\ev\) is an isomorphism too. Let \(P^{*}\rightarrow Q\) be a \(\cat{P}\)-resolution of \(Q\) in \(\hat{\cat{P}}{}^{\textnormal{fl}}_{h}\) of length \(n\). Define covariant functors \(G^{i}:\cat{mod}_{S}\rightarrow\cat{mod}_{T}\) by \(G^{i}(V):=\cH^{i-n}(P^{*}{\otimes}_{T}\omega_{h}{\otimes}_{S}V)\). Since \(P^{*}{\otimes}_{T}\omega_{h}\) is \(S\)-flat, \(\{G^{i}\}\) defines a cohomological \(\delta\)-functor. Since \(P^{*}{\otimes} k(s)\) gives a resolution of \(Q{\otimes}_{S} k(s)\) for all \(s\in\Spec S\), \cite[9.6.5]{bru/her:98} and Proposition \ref{prop.nakayama} implies that \(G^{n}(S)=Q{\otimes}_{T}\omega_{h}\) is \(S\)-flat and \(P^{*}{\otimes}_{T}\omega_{h}\twoheadrightarrow Q{\otimes}_{T}\omega_{h}\) is a \(\cat{D}\)-resolution. Moreover, the natural map \(Q\rightarrow\hm{}{T}{\omega_{h}}{Q{\otimes}\omega_{h}}\) is an isomorphism by Nakayama's lemma again.
\end{proof}
\begin{ex}\label{ex.extiso}
Assume \(A\) is a Cohen-Macaulay ring with a canonical module \(\omega_{A}\).
Given \(L_{i}\in\hat{\cat{D}}_{A}\) and put \(Q_{i}=\hm{}{A}{\omega_{A}}{L_{i}}\) for \(i=1,2\). If \(I\) is an injective resolution of \(L_{2}\) and \(P\) is a projective resolution of \(Q_{1}\) then both spectral sequences of \(\hm{}{A}{P}{\hm{}{A}{\omega_{A}}{I}}\) collapse at page \(2\) (use \cite[I 4.10.19]{has:00}) to give canonical isomorphisms
\begin{equation}
\xt{*}{A}{L_{1}}{L_{2}}\cong\xt{*}{A}{Q_{1}}{Q_{2}}
\end{equation}
\end{ex}
\begin{rem}\label{rem.Buch}
In his unpublished manuscript Buchweitz gave a construction of Cohen-Macaulay approximation for finite \(A\)-modules if \(A\) is a not necessarily commutative Gorenstein ring, see \cite{buc:86}. The \(\cat{MCM}\)-approximation and the \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull in Theorem \ref{thm.flatCMapprox} can be given essentially by the same construction. Let \(N\) be a finite \(T\)-module which is \(S\)-flat where \(h:S\rightarrow T\) is a finite type Cohen-Macaulay map. Let \(P=P(N)\rightarrow N\) be a projective resolution of \(N\) (i.e.\ by finite projective \(T\)-modules). Then \(P^{\vee}=\hm{}{T}{P}{\omega_{h}}\) is a bounded below complex with bounded cohomology because \(\xt{i}{T}{N}{\omega_{h}}=0\) for \(i\) greater than the relative dimension \(d\) of \(h\) by Corollary \ref{cor.xtdef} (\(\Injdim \omega_{T_{s}}=\dim T_{s}\leq d\)). Then we can choose a projective resolution \(f:P(P^{\vee})\rightarrow P^{\vee}\) of \(P^{\vee}\) which is bounded above. Let \(C=C(f)\) be the mapping cone of \(f\). The modules in \(C\) are direct sums of projective modules and modules in \(\cat{D}_{h}\) and the (co)kernels in the acyclic \(C\) are modules in \(\cat{MCM}_{h}\). By Corollary \ref{cor.xtdef} it follows that \(C^{\vee}\) is acyclic too. There is a composition of natural maps \(P\cong P^{\vee\vee}\rightarrow P(P^{\vee})^{\vee}:=G\) which hence is a quasi-isomorphism. But then \(G\) is just the representing complex of \(N\) and the \(\cat{MCM}\)-approximation and the \(\hat{\cat{D}}\)-hull is obtained as in Section \ref{subsec.cplx}. In the the case of coherent rings with a cotilting module (a concept introduced by Y.\ Miyashita, see \cite[p.\ 142]{miy:86}) J.-i.\ Miyachi implicitly gave the same construction in \cite[3.2]{miy:96}.
Suppose \(h\) has pure relative dimension and \(N_{s}\) is a non-zero Cohen-Macaulay \(T_{s}\)-module with \(n=\dim T_{s}-\dim N_{s}\) constant for all \(s\in\Spec S\), i.e.\ \(N\) is a family of Cohen-Macaulay modules of codepth \(n\). Put \(N^{\vee}:=\xt{n}{T}{N}{\omega_{h}}\). Then \(P^{\vee}=P(N)^{\vee}\) is quasi-isomorphic to \(\cH^{n}(P^{\vee})=N^{\vee}\) by duality theory and Corollary \ref{cor.xtdef}. So if \((F,d)\rightarrow N^{\vee}\) is a projective resolution of \(N^{\vee}\), the representing complex is given as \(G=F^{\vee}\). If \(\syz{T}{i}N^{\vee}\) denotes the \(i^{\text{th}}\) syzygy module \(\im d_{i}\) and \(d_{i}^{\vee}\) denotes \(\hm{}{T}{d_{i}}{\omega_{h}}\) then the commutative diagram \eqref{eq.pullpush} with short exact sequences is given as
\begin{equation*}
\xymatrix@C-0pt@R-9pt@H-30pt{
\im (d_{n}^{\vee}) \ar[r]\ar@{=}[d] & \hm{}{T}{\syz{T}{n}N^{\vee}}{\omega_{h}} \ar[r]\ar[d]\ar@{}[dr]|{\Box} & N^{\vee\vee} \ar[d] \\
\im (d_{n}^{\vee}) \ar[r] & \hm{}{T}{F_{n}}{\omega_{h}} \ar[r]\ar[d] & \coker(d_{n}^{\vee}) \ar[d] \\
& \hm{}{T}{\syz{T}{n{+}1}N^{\vee}}{\omega_{h}} \ar@{=}[r] & \hm{}{T}{\syz{T}{n{+}1}N^{\vee}}{\omega_{h}}
}
\end{equation*}
with the \(\cat{MCM}_{h}\)-approximation and \(\hat{\cat{D}}_{h}\)-hull of \(N\) given by the upper horizontal and right vertical sequence, respectively, since \(N^{\vee\vee}\cong N\) by duality theory and Corollary \ref{cor.xtdef}.
Let \(C\) denote the mapping cone of a comparison map \(P(N)\rightarrow G\). The homology of the truncated short exact sequence of complexes \(G\rightarrow C\rightarrow P(N)[-1]\) gives the \(\cat{MCM}_{h}\)-approximation \(\coker d^{G}_{i+2}\rightarrow\coker d^{C}_{i+2}\rightarrow\syz{T}{i}N\) for \(i\geq 0\) (with \(i=0\) being the upper horizontal sequence in the diagram). In the absolute case (with \(N\) Cohen-Macaulay) this latter construction of the MCM approximation of \(\syz{}{i}N\) was given by J.\ Herzog and A.\ Martsinkovsky in \cite[1.1]{her/mar:93}.
\end{rem}
\subsection{Local cases}\label{subsec.loc}
We formulate local variants of the approximation theorem.
Fix a field \(k\). Let \(\cat{H}\) denote the category of noetherian, henselian, local rings \(S\) with residue field \(S/\fr{m}_{S}\cong k\) and with local ring homomorphisms. A map \(h:S\rightarrow T\) in \(\cat{H}\) is \emph{algebraic} (or \(T\) is an algebraic \(S\)-algebra) if there is a finite type \(S\)-algebra \(\tilde{T}\) and a maximal ideal \(\fr{m}\) in \(\tilde{T}\) with \(\tilde{T}/\fr{m}\cong k\) such that \(h\) is given by henselisation of \(\tilde{T}\) in \(\fr{m}\). Fix an algebraic \(k\)-algebra \(A\) which is supposed to be Cohen-Macaulay. Objects in the category \(\cat{hCM}\) are algebraic and flat \(S\)-algebras \(T\) with \(T{\otimes}_{S}k\cong A\). A morphism \(h_{1}\rightarrow h_{2}\) is a pair of commuting maps \(f:T_{1}\rightarrow T_{2}\) and \(g:S_{1}\rightarrow S_{2}\) in \(\cat{H}\) as for the finite type case, giving a cocartesian square. Base change exists for the forgetful functor \(\cat{hCM}\rightarrow\cat{H}\) and is given by the henselisation of the tensor product \(T=T_{1}{\otimes}_{S_{1}}S_{2}\) in the maximal ideal \(\fr{m}_{T_{1}}T+\fr{m}_{S_{2}}T\). We denote it by \(T_{1}{\tilde{\otimes}}_{S_{1}}S_{2}\). It follows that \(\cat{hCM}\rightarrow\cat{H}\) is fibred in groupoids. The objects in \(\cat{hCM}\) will be called henselian Cohen-Macaulay (hCM) maps.
If \(\tilde{h}:\tilde{S}\rightarrow\tilde{T}\) is a finite type CM map and \(t\) a \(k\)-point in \(\Spec \tilde{T}\) with image \(s\) in \(\Spec \tilde{S}\) then localisation for the \'etale topology at \(t\) and \(s\) gives a hCM map \(h:S=\tilde{S}^{\text{h}}\rightarrow \tilde{T}^{\text{h}}=T\). Conversely, every hCM map is obtained this way which follows from \cite[18.6.6 and 18.6.10]{EGAIV4} and \cite[15.4.3 and 12.1.1]{EGAIV3}. We will call such an \(\tilde{h}\) a (finite type) representative of \(h\). The dualising module \(\omega_{\tilde{h}}\) induces an \(S\)-flat finite \(T\)-module \(\omega_{h}\) called the dualising module for \(h\). Two representatives of \(h\) factor through a common \'etale neighbourhood contained in \(\cat{CM}\) and since the dualising module is functorial for \(\cat{CM}\) the dualising module \(\omega_{h}\) is functorial too.
Let \(\cat{mod}\) denote the category of pairs \((h:S\rightarrow T,\mc{N})\) with \(h\) in \(\cat{hCM}\) and \(\mc{N}\) a finite \(T\)-module. Morphisms are defined as for the finite type case and the forgetful functor \(\cat{mod}\rightarrow\cat{hCM}\) makes \(\cat{mod}\) fibred in abelian categories over \(\cat{hCM}\). Let \(\cat{mod}{}^{\textnormal{fl}}\) denote the full subcategory of objects \((h,\mc{N})\) in \(\cat{mod}\) with \(\mc{N}\) \(S\)-flat and let \(\cat{MCM}\) denote the full subcategory of objects \((h,\mc{N})\) in \(\cat{mod}{}^{\textnormal{fl}}\) where the closed fibre \(\mc{N}{\otimes}_{S}k\) is a maximal Cohen-Macaulay \(A\)-module. Let \(\cat{D}\) denote the full subcategory of \(\cat{MCM}\) of objects \((h,D)\) with \(D\) in \(\Add\{\omega_{h}\}\).
All three are FAd subcategories of \(\cat{mod}\) over \(\cat{hCM}\). Any finite \(T\)-module \(\mc{N}\) has finite presentation hence it is induced from a finite module over a representative of \(T\). If \(\mc{N}\) is contained in one of the subcategories the representative can be assumed to belong to the corresponding finite type subcategory ({\sl loc.\ sit.}). Similarly all maps in these fibred categories over \(\cat{H}\) are induced from maps in the corresponding fibred categories of finite type objects.
Let \(\cat{L}\) (\(\cat{C}\)) denote the category of noetherian, (complete) local rings with residue field \(k\) and local ring homomorphisms. Let \(\cat{lCM}\) (\(\cat{cCM}\)) denote the category of local Cohen-Macaulay (lCM) maps (respectively complete Cohen-Macaulay or cCM maps) defined analogously as above with (completion of) \emph{essentially of finite type} replacing algebraic. Similar arguing as above makes \(\cat{lCM}\) (\(\cat{cCM}\)) fibred in groupoids over \(\cat{L}\) (\(\cat{C}\)). The definitions of the module categories apply in the local and the complete case too and we use the same notation in all three cases. Again objects and maps are induced from the finite type case.
Either arguing with representatives or applying the proofs for Theorem \ref{thm.flatCMapprox} and Corollary \ref{cor.flatCMapprox} (with only minor adjustments) we obtain the following.
\begin{cor}\label{cor.locCMapprox}
Let \(\cat{xCM}\) denote either \(\cat{hCM}\)\textup{,} \(\cat{lCM}\) or \(\cat{cCM}\)\textup{.} The pair \((\cat{mod},\cat{MCM})\) of fibred categories over \(\cat{xCM}\) satisfies \textup{BC1-2} and the triple of fibre categories \((\cat{mod}_{h},\cat{MCM}_{h},\cat{D}_{h})\) satisfies \textup{AB1-4} for all objects \(h\) in \(\cat{xCM}\)\textup{.}
Moreover\textup{;} the statements \textup{(i-v)} in \textup{Theorem \ref{thm.flatCMapprox}} and \textup{(i-ii) in Corollary \ref{cor.flatCMapprox}} are valid over \(\cat{xCM}\) too\textup{.}
\end{cor}
\section{Minimal approximations and semi-continuity of invariants}
We show that the Cohen-Macaulay approximation and the \(\hat{\cat{D}}\)-hull in Corollary \ref{cor.locCMapprox} can be chosen to be minimal. Upper semi-continuous invariants on \(\cat{MCM}_{A}\) or \(\cat{FID}_{A}\) extends to upper semi-continuous invariants on \(\cat{mod}_{A}\). An example is given by the \(\omega_{A}\)-ranks in the representing \(\cat{D}\)-complex.
\begin{lem}\label{lem.Aapprox}
Let \(S\rightarrow T\) be a homomorphism of noetherian rings and \(\fr{a}\) an ideal in \(S\) such that \(I=\fr{a}T\) is contained in the Jacobson radical of \(T\)\textup{.} Let \(M\) and \(N\) be finite \(T\)-modules\textup{.} Let \(T_{n}=T/I^{n+1}\)\textup{,} \(M_{n}=T_{n}{\otimes} M\) and \(N_{n}=T_{n}{\otimes} N\)\textup{.} Suppose there exists a tower of surjections \(\{\phi_{n}:M_{n}\rightarrow N_{n}\}\)\textup{.} Fix any non-negative integer \(n_{0}\)\textup{.} Then there exists a \(T\)-linear surjection \(\psi:M\rightarrow N\) such that \(T_{n_{0}}{\otimes}\psi=\phi_{n_{0}}\)\textup{.} If the \(\phi_{n}\) are isomorphisms and \(N\) is \(S\)-flat then \(\psi\) is an isomorphism\textup{.}
\end{lem}
\begin{proof}
Let \(\hat{T}=\varprojlim T_{n}\), \(\hat{M}=\varprojlim M_{n}\) and \(\hat{N}=\varprojlim N_{n}\). We have
\begin{equation}
\varprojlim\hm{}{T_{n}}{M_{n}}{N_{n}}\cong\hm{}{\hat{T}}{\hat{M}}{\hat{N}}\cong\hat{T}{\otimes}\hm{}{T}{M}{N}.
\end{equation}
Hence \(\varprojlim \phi_{n}=\Sigma r^{(i)}{\otimes}\beta^{(i)}\) with \(r^{(i)}\in \hat{T}\) and \(\beta^{(i)}\in\hm{}{T}{M}{N}\). Let \(r^{(i)}_{n_{0}}\) be the image of \(r^{(i)}\) under \(\hat{T}\rightarrow T_{n_{0}}\) and choose liftings \(t^{(i)}\) in \(T\) of \(r^{(i)}_{n_{0}}\). Put \(\psi=\Sigma t^{(i)}\beta^{(i)}\). Then \(T_{n_{0}}{\otimes}\psi=\phi_{n_{0}}\). Since \(T_{n_{0}}{\otimes}\coker\psi=\coker\phi_{n_{0}}=0\), Nakayama's lemma implies \(\coker\psi=0\). Since \(T_{n_{0}}{\otimes}_{T}N\) equals \(N{\otimes}_{S}S/\fr{a}^{n_{0}+1}\), \(S\)-flatness of \(N\) implies \(T_{n_{0}}{\otimes}\ker\psi=\ker\phi_{n_{0}}\) and if \(\ker\phi_{n_{0}}=0\) then Nakayama's lemma implies \(\ker\psi=0\).
\end{proof}
\begin{prop}\label{prop.minapprox}
Let \(\cat{xCM}\) denote either \(\cat{hCM}\)\textup{,} \(\cat{lCM}\) or \(\cat{cCM}\)\textup{,} let \(h:S\rightarrow T\) be an object in \(\cat{xCM}\) and let \(\xi:\mc{L}\rightarrow\mc{M}\xra{\pi}\mc{N}\) be an \(\cat{MCM}_{h}\)-approximation of \(\mc{N}\)\textup{.} Then the following statements are equivalent\textup{.}
\begin{enumerate}
\item[(i)] The sequence \(\xi\) is a right minimal \(\cat{MCM}_{h}\)-approximation\textup{.}
\item[(ii)] There are no surjections \(\mc{M}\rightarrow\omega_{h}\) which induces a surjection \(\mc{L}\rightarrow\omega_{h}\)\textup{.}
\item[(iii)] There are no common \(\omega_{h}\)-summand in \(\mc{L}\rightarrow\mc{M}\)\textup{.}
\item[(iv)] The closed fibre \(\xi{\otimes}_{S}k\) is a right minimal \(\cat{MCM}_{A}\)-approximation\textup{.}
\item[(v)] The completion \((\xi{\otimes}_{S}k)\hat{\,}\) of the closed fibre is a right minimal \(\cat{MCM}_{\hat{A}}\)-approximation\textup{.}
\end{enumerate}
The analogous statements \textnormal{(i')-(v')} for a \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\)-hull \(\xi':\mc{N}\rightarrow\mc{L}'\rightarrow\mc{M}'\) are equivalent\textup{.} In particular\textup{:}
\begin{enumerate}
\item[(i')] The sequence \(\xi'\) is a left minimal \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\)-hull\textup{.}
\item[(ii')] There are no surjections \(\mc{M}'\rightarrow\omega_{h}\) which induces a surjection \(\mc{L}'\rightarrow\omega_{h}\)\textup{.}
\end{enumerate}
\end{prop}
\begin{proof}
Suppose there is a surjection \(\mc{M}\rightarrow \omega_{h}\) such that the composition \(\mc{L}\rightarrow \omega_{h}\) is surjective too. Then the kernels of these maps give a new \(\cat{MCM}\)-approximation of \(\mc{N}\) by Corollary \ref{cor.locCMapprox}. Since a surjection \(\mc{L}\rightarrow\omega_{h}\) has to split by Corollary \ref{cor.locCMapprox}, \(\omega_{h}\) is a common summand in \(\mc{L}\rightarrow\mc{M}\) and \(\pi\) cannot be right minimal.
Let the closed fibre \(\xi{\otimes}_{S}k\) of the sequence \(\xi\) be denoted by \(L\rightarrow M\xra{p} N\).
Assume there is a non-surjective endomorphism \(\theta:\mc{M}\rightarrow\mc{M}\) with \(\pi\theta=\pi\). Then \(\theta_{0}=\theta{\otimes}_{S}k\) gives a non-surjective endomorphism of \(M\) with \(p\theta_{0}=p\). It follows that the completion \(\hat{L}\rightarrow\hat{M}\rightarrow\hat{N}\) is not a right minimal Cohen-Macaulay approximation of \(\hat{N}\). By \cite[1.12.8]{has:00} there is a common \(\omega_{\hat{A}}\)-summand in \(\hat{L}\rightarrow\hat{M}\). Let \(\phi:\hat{M}\rightarrow\omega_{\hat{A}}\) denote the projection. By Lemma \ref{lem.Aapprox} there exists a surjection \(\psi:M\rightarrow\omega_{A}\). The induced map \(L\rightarrow\omega_{A}\) is surjective too. The map \(\psi\) lifts to a surjection \(\mc{M}\rightarrow\omega_{h}\) (with \(\mc{L}\rightarrow\omega_{h}\) surjective) since the canonical map \(\hm{}{T}{\mc{M}}{\omega_{h}}\rightarrow\hm{}{A}{M}{\omega_{A}}\) is surjective by Corollary \ref{cor.xtdef}.
The \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-case is analogous.
\end{proof}
\begin{cor}\label{cor.minapprox}
Let \(\cat{xCM}\) denote either \(\cat{hCM}\)\textup{,} \(\cat{lCM}\) or \(\cat{cCM}\)\textup{.}
For any object \((h,\mc{N})\) in the fibred category \(\cat{mod}{}^{\textnormal{fl}}\) over \(\cat{xCM}\), \(\mc{N}\) admits a right minimal \(\cat{MCM}\)-approximation and a left minimal \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull which remain minimal after base change and which in particular are unique up to non-canonical isomorphism\textup{.}
\end{cor}
\begin{proof}
The existence of a right minimal \(\xi\) follows immediately from criterion (iii) in Proposition \ref{prop.minapprox}. Moreover \(\xi\) is right minimal if and only if the closed fibre \(\xi{\otimes}_{S}k\) is right minimal. Since any base change \(\xi_{1}\) of \(\xi\) has the same closed fibre as \(\xi\), \(\xi_{1}\) is right minimal if \(\xi\) is.
\end{proof}
\begin{rem}
In \cite[2.3]{has/shi:97} M.\ Hashimoto and A.\ Shida gave essentially the analog of Proposition \ref{prop.minapprox} in the absolute case of a Cohen-Macaulay Zariski local ring with a canonical module. They attributed the complete case to Y.\ Yoshino. Note Miyachi's proof in the cotilting semi-perfect case, see \cite[3.4]{miy:96} (cf.\ \cite[1.12.8]{has:00}). In \cite[3.1]{sim/str:02} A.-M.\ Simon and J.\ R.\ Strooker give a short independent proof. The proof of Proposition \ref{prop.minapprox} also works in the Zariski local case in full generality and is different from these (but depends on the complete case).
\end{rem}
Since minimal choices of \(\cat{MCM}_{A}\)-approximations and \(\cat{\hat{D}}_{A}\)-hulls exist and are unique up to isomorphism any invariant defined for MCM modules or for FID modules is extended to all finite \(A\)-modules. Upper semi-continuity of the invariants is also extended as we explain now. First some notation.
Let \(h:S\rightarrow T\) be ring homomorphism and \(\mc{N}\) a finite \(T\)-module. If \(t\in\Spec T\) has image \(s\in\Spec S\), let \(\mc{N}_{\fr{p}_{t}}\) denote the localisation of \(\mc{N}\) at the prime ideal \(t\), let \(h_{\fr{p}_{t}}:S_{\fr{p}_{s}}\rightarrow T_{\fr{p}_{t}}\) denote the ring homomorphism obtained by localising, and put \(\mc{N}(t)=\mc{N}_{\fr{p}_{t}}{\otimes}_{S_{\fr{p}_{s}}}k(s)\) which is a \(T(t)\)-module; indeed \(\mc{N}(t)\cong T(t){\otimes}_{T_{\fr{p}_{t}}}\mc{N}_{\fr{p}_{t}}\).
If \(h\) is a finite type Cohen-Macaulay map, \(h_{\fr{p}_{t}}\) is in \(\cat{lCM}\).
Suppose \(\mu\) is an invariant on \(\cat{MCM}_{A}\) where \(A\) is a Cohen-Macaulay local ring with canonical module. Let \({}_{\cat{MCM}}\mu\) denote the induced invariant on \(\cat{mod}_{A}\) defined by \({}_{\cat{MCM}}\mu(N) = \mu(M)\) where \(L\rightarrow M\rightarrow N\) is the minimal Cohen-Macaulay approximation of \(N\). Similarly \({}_{\cat{FID}}\mu(N)= \mu(L)\) for an invariant \(\mu\) defined on \(\cat{FID}_{A}\). Use the minimal hull \(N\rightarrow L'\rightarrow M'\) to define \({}_{\cat{FID}}\mu'\) and \({}_{\cat{MCM}}\mu'\).
The following theorem is a major application of what we have done so far.
\begin{thm}\label{thm.semicont}
Let \(\mu\) be an additive non-negative numerical invariant defined for maximal Cohen-Macaulay modules or for finite modules of finite injective dimension on a Cohen-Maculay local ring with canonical module\textup{.} Assume \(\mu\) is upper semi-continuous for finite type flat families \((h:S\rightarrow T,\mc{M})\) in \(\cat{MCM}\) \textup{(}or in \(\hat{\cat{D}}\)\textup{).} Then the induced invariants \({}_{\cat{MCM}}\mu\) and \({}_{\cat{MCM}}\mu'\) \textup{(}or \({}_{\cat{FID}}\mu\) and \({}_{\cat{FID}}\mu'\)\textup{)} are upper semi-continuous in finite type flat families \((h,\mc{N})\) in \(\cat{mod}{}^{\textnormal{fl}}\)\textup{.}
\end{thm}
\begin{proof}
Given \((h:S\rightarrow T,\mc{N})\) in \(\cat{mod}{}^{\textnormal{fl}}\) and \(t\in\Spec T\). By Corollary \ref{cor.minapprox} there exists open affines \(U=\Spec S_{1}\subseteq \Spec S\), \(V=\Spec T_{1}\subseteq \Spec T\) with \(t\in V\), \(h_{1}:S_{1}\rightarrow T_{1}\) induced from \(h\), and a \(\cat{MCM}_{h_{1}}\)-approximation \(\xi: \mc{L}\rightarrow\mc{M}\rightarrow\mc{N}_{\vert V}\) such that the localisation \(\xi_{\fr{p}_{t}}\) is minimal. By Corollary \ref{cor.minapprox} \(\xi(t)\) is minimal too. Put \(n=\mu(\mc{M}(t))\). Since \(\mu\) is upper semi-continuous there is an open \(V_{n}\subseteq V\) containing \(t\) such that \(\mu(\mc{M}(t'))\leq n\) for all \(t'\in V_{n}\). If \(L\rightarrow M\rightarrow \mc{N}(t')\) is the minimal MCM approximation of \(\mc{N}(t')\), \(M\) is a direct summand of \(\mc{M}(t')\) by Proposition \ref{prop.minapprox}, and hence \({}_{\cat{MCM}}\mu(\mc{N}(t'))=\mu(M)\leq\mu(\mc{M}(t'))\leq n\).
\end{proof}
\begin{ex}\label{ex.semicont1}
The Betti numbers \(\beta_{i}(M)=\dim\tor{A}{i}{M}{A/\fr{m}_{A}}\) are well known upper semi-continuous invariants of finite modules over local rings. By Theorem \ref{thm.semicont} the induced invariants \({}_{\cat{MCM}}\beta_{i}\), \({}_{\cat{MCM}}\beta_{i}'\), \({}_{\cat{FID}}\beta_{i}\) and \({}_{\cat{FID}}\beta_{i}'\) are upper semi-continuous too.
\end{ex}
We now consider some invariants defined in terms of Cohen-Macaulay approximation.
If \(h:S\rightarrow T\) is in one of the categories of local Cohen-Macaulay maps, a map \(\partial:D\rightarrow D'\) of objects in \(\cat{D}_{h}\) is said to be \emph{minimal} if \(k{\otimes}_{T}\partial=0\). Any module \(D\) in \(\cat{D}_{h}\) is isomorphic to some \(\omega_{h}^{\oplus n}\) and \(\nd{}{T}{\omega_{h}}\cong T\). Hence if \(\partial\) is not minimal then there is a surjection \(D'\rightarrow\omega_{h}\) inducing a surjection \(D\rightarrow\omega_{h}\). By Corollary \ref{cor.locCMapprox} the \(\omega_{h}\) splits off from \(\partial\). Hence any \(\cat{D}_{h}\)-complex is homotopy equivalent to one with all differentials being minimal, which is called a minimal \(\cat{D}\)-complex.
For any module \(\mc{N}\) in \(\cat{mod}{}^{\textnormal{fl}}_{h}\) over \(\cat{xCM}\) we choose a minimal \(\cat{MCM}\)-approximation \(\mc{L}\rightarrow\mc{M}\rightarrow\mc{N}\) and a minimal \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull \(\mc{N}\rightarrow\mc{L}'\rightarrow\mc{M}'\) which exist by Corollary \ref{cor.minapprox}. Spliced with a
minimal \(\cat{D}\)-resolution of \(\mc{L}\) and a minimal \(\cat{D}\)-coresolution of \(\mc{M}'\) we obtain complexes \({}^{-}C^{*}(\mc{N})\), \({}^{+}C^{*}(\mc{N})\) and \(D^{*}(\mc{N})\), as defined in Section \ref{subsec.cplx}, where no differential has any \(\omega_{h}\)-summand. We call such choices of these complexes for minimal. They are unique:
\begin{lem}\label{lem.uniqueC}
Suppose \(h\) is in \(\cat{hCM}\)\textup{,} \(\cat{lCM}\) or \(\cat{cCM}\) and \((h,\mc{N})\) is in \(\cat{mod}{}^{\textnormal{fl}}_{h}\)\textup{.} Then
minimal choices of \({}^{-}C^{*}(\mc{N})\), \({}^{+}C^{*}(\mc{N})\) and \(D^{*}(\mc{N})\) exist and are unique up to non-canonical isomorphisms\textup{.}
\end{lem}
\begin{proof}
Minimal choices \({}^{+}C^{*}_{1}\) and \({}^{+}C^{*}_{2}\) of coresolutions for \(\mc{N}\) are by Lemma \ref{lem.res} homotopic through chain maps \(\alpha\) and \(\beta\) starting with an isomorphism \(\mc{L}_{1}'\cong\mc{L}_{2}'\). If \(\rho_{i}\) are homotopies with \(\beta\alpha-\id=\partial\rho_{1}+\rho_{1}\partial\) and \(\alpha\beta-\id=\partial\rho_{2}+\rho_{2}\partial\) then tensoring down by \(k{\otimes}_{T}-\) makes the right hand side of these identities equal to zero by the minimality of the complexes. Hence \(\beta\alpha\) and \(\alpha\beta\) are surjective endomorphisms, i.e.\ isomorphisms. The same argument applies to \({}^{-}C^{*}\) and \(D^{*}\).
\end{proof}
For each module \(\mc{N}\) in \(\cat{mod}{}^{\textnormal{fl}}_{h}\) we fix a minimal \(\cat{D}\)-complex \((D^{*}(\mc{N}),\partial^{*})\) representing \(\mc{N}\). Let \(\mc{L}'=\coker \partial^{-1}\) and \(\mc{M}=\ker \partial^{0}\).
Put \(\syz{\omega_{h}}{i}\mc{L}':=\coker\{\partial^{-i{-}1}:D^{{-}i{-}1}\rightarrow D^{-i}\}\) for \(i\geq 0\) and \(\syz{\omega_{h}}{-i}\mc{M}:=\ker\{\partial^{i}:D^{i}\rightarrow D^{i{+}1}\}\) for \(i\geq 0\). For any finite \(T\)-module \(\mc{N}\) let the \emph{\(\omega_{h}\)-rank} of \(\mc{N}\), denoted \(\rk{\omega_{h}}{\mc{N}}\), be the largest number \(n\) with \(\omega_{h}^{{\oplus}n}{\oplus}\mc{N}'\cong \mc{N}\) for some \(T\)-module \(\mc{N}'\). Since \(\nd{}{T}{\omega_{h}}\cong T\) is a local ring, this is a well behaved invariant, cf.\ \cite[Sec.\ 1.1]{sim/str:02}.
\begin{defn}\label{defn.c}
Suppose \(h\) is in \(\cat{hCM}\)\textup{,} \(\cat{lCM}\) or \(\cat{cCM}\) and \(\mc{N}\) is in \(\cat{mod}{}^{\textnormal{fl}}_{h}\)\textup{.} Define the numbers\textup{:}
\begin{enumerate}
\item[(i)] \(d^{i}_{T}(\mc{N}):=\rk{\omega_{h}}{D^{i}(\mc{N})}\) for all \(i\)
\item[(ii)] \(\nu^{T}_{i}(\mc{N}):=\rk{\omega_{h}}{\syz{\omega_{h}}{i}\mc{L}'}\) for \(i\geq 0\)
\item[(iii)] \(\gamma_{T}(\mc{N}):=\rk{\omega_{h}}{\mc{M}}\)
\end{enumerate}
\end{defn}
The definition gives well defined invariants of \(\mc{N}\) by Lemma \ref{lem.uniqueC}. In particular we see that \(\nu^{T}_{0}(\mc{N})\) equals \(\rk{\omega_{h}}{\mc{L}'}\). We also notice that \(\rk{\omega_{h}}{\syz{\omega_{h}}{-i}\mc{M}}=0\) for all \(i>0\) by Proposition \ref{prop.minapprox}.
The same notation is used for the absolute counterparts of these invariants.
\begin{rem}
In the absolute case with \(A\) a Gorenstein local ring, the \(\gamma_{A}\)-invariant is called Auslander's \(\delta\) invariant and has been studied in particular by S.\ Ding. One has that \(\gamma_{A}(A/\fr{m}_{A}^{n})\leq 1\) with equality for \(n\gg0\). The smallest number \(n\) with \(\gamma_{A}(A/\fr{m}_{A}^{n})=1\) is called the \emph{index} of \(A\) and Ding has given results and conjectures concerning this invariant, cf.\ \cite{din:92}. See also \cite{has/shi:97}.
Let \((A,\fr{m},k)\) be a noetherian local ring and \(N\) a finite \(A\)-module. Simon and Strooker introduced the \emph{reduced Bass numbers}
\begin{equation}
\nu_{A}^{i}(N)=\dim_{k}\im\{\xt{i}{A}{k}{N}\rightarrow\cH^{i}_{\fr{m}}(N)\}
\end{equation}
and in the case \(A\) has a canonical module they showed in \cite[2.6 and 3.10]{sim/str:02} that
\begin{equation}
\nu^{A}_{0}(N)=\nu_{A}^{\dim A}(N)\quad\text{and}\quad \nu^{A}_{i}(N)=\nu_{A}^{\dim A{-}i}(L')\,\,\,\text{for } i\geq 0
\end{equation}
where \(L'\) is the minimal \(\cat{D}_{A}\)-hull of \(N\). The Canonical Element Conjecture and the Monomial Conjecture are equivalent to \(\nu_{A}^{\dim A}(\fr{b})\neq 0\) (or equivalently \(\gamma_{A}(A/\fr{b})=0\)) for certain ideals \(\fr{b}\) in any Gorenstein local ring \(A\), see \cite[6.4 and 6.6]{sim/str:02}.
\end{rem}
Let \(\cat{P}\) and \(\hat{\cat{P}}{}^{\textnormal{fl}}\) denote the FAds of finite projective modules, respectively modules of finite projective dimension over \(\cat{xCM}\) defined as for the finite type case.
\begin{lem}\label{lem.fri} With notation as above\textup{:}
\begin{enumerate}
\item[(i)] The functor \(F=\hm{}{-}{\omega_{-}}{-}\) gives exact equivalences of categories fibred in additive categories \(\cat{D}\simeq \cat{P}\) and \(\hat{\cat{D}}{}^{\textnormal{fl}}\simeq\hat{\cat{P}}{}^{\textnormal{fl}}\) over \(\cat{xCM}\)\textup{.}
\item[(ii)] In particular \(d^{-i}(\mc{N})=\beta_{i}(\hm{}{T}{\omega_{h}}{\mc{L}'})\) for all \(i\geq 0\)\textup{.}
\item[(iii)] \(\hm{}{T}{D^{*\geq 0}(\mc{N})}{\omega_{h}}\) gives a minimal free resolution of \(\hm{}{T}{\mc{M}}{\omega_{h}}\)\textup{.} In particular \(d^{i}(\mc{N})=\beta_{i}(\hm{}{T}{\mc{M}}{\omega_{h}})\) for all \(i\geq 0\)\textup{.}
\end{enumerate}
\end{lem}
\begin{proof}
(i) and (ii) are the local variants of Lemma \ref{lem.proj}. Breaking \(\mc{M}\hookrightarrow D^{*\geq 0}(\mc{N})\) into short exact sequences, (iii) follows from Corollary \ref{cor.locCMapprox}.
\end{proof}
\begin{cor}\label{cor.semicont}
Let \(h:S\rightarrow T\) be a finite type Cohen-Macaulay map and suppose \(\mc{N}\) is a \(T\)-module in \(\cat{mod}{}^{\textnormal{fl}}_{h}\)\textup{.} Then \(d^{i}(\mc{N}(t))\) are upper semi-continuous functions in \(t\in\Spec T\) for all \(i\)\textup{.}
\end{cor}
\begin{proof}
This follows from Theorem \ref{thm.semicont} and Lemma \ref{lem.fri}.
\end{proof}
\begin{rem}\label{rem.semicontnot}
The invariants \(\nu_{0}\) and \(\gamma\) are not semi-continuous either way, see Example \ref{ex.semicont}. Moreover, let \(\mc{L}\) be in \(\hat{\cat{D}}_{h}\) with \(L=\mc{L}{\otimes}_{S}k\). If \(\rho_{0}:\omega_{A}\rightarrow L\) is a direct summand, then \(\rho_{0}\) lifts to a map \(\rho:\omega_{h}\rightarrow\mc{L}\), but no lifting \(\rho\) of \(\rho_{0}\) has to split, even if \(A\) is a regular ring.
\end{rem}
\begin{rem}\label{rem.semicont}
One can also define functions of the base. E.g.\ if \(\phi:\Spec T\rightarrow \Spec S\) denotes the map induced from \(h\) and \(\mu(\mc{N}(t))\) is an upper semi-continuous function of \(t\in \Spec T\), then \(\phi_{*}\mu(\mc{N})\) defined by
\begin{equation}
\phi_{*}\mu(\mc{N})(s)=\sup_{t\in\phi^{-1}(s)}\!\!\mu(\mc{N}(t))
\end{equation}
is an upper semi-continuous function in \(s\in \Spec S\) since \(\phi\) is an open map.
\end{rem}
\section{The fundamental module of a Cohen-Macaulay map}\label{sec.fund}
\begin{ex}\label{ex.Fmod}
Let \((A,\fr{m}_{A},k)\) be a Cohen-Macaulay local ring with canonical module \(\omega_{A}\). Let \((G_{*},d_{*})\rightarrow k\) be a minimal \(A\)-free resolution of \(k\) and put \(\syz{}{i}=\syz{A}{i}k=\coker d_{i+1}\). Suppose \(\dim A=d\geq 2\). There are connecting isomorphisms \(\xt{1}{A}{\syz{}{d{-}1}}{\omega_{A}}\cong\xt{2}{A}{\syz{}{d{-}2}}{\omega_{A}}\cong\dots\cong\xt{d}{A}{k}{\omega_{A}}\) which is isomorphic to \(k\) by duality theory. To \(1\in k\) there is hence a non-split short exact sequence
\begin{equation}\label{eq.Fmod}
\theta:\,\,0\rightarrow\omega_{A}\longrightarrow E_{A}\xra{\,\,\pi\,\,} \syz{A}{d{-}1}k\rightarrow 0
\end{equation}
with \(E_{A}\) uniquely defined up to non-canonical isomorphism. We call \(E_{A}\) for the \emph{fundamental module} of \(A\). We claim that \(E_{A}\) is a maximal Cohen-Macaulay module which implies that \eqref{eq.Fmod} is the minimal MCM approximation of \(\syz{A}{d{-}1}k\). If we apply \(\hm{}{A}{-}{\omega_{A}}\) to \eqref{eq.Fmod} we obtain the exact sequence
\begin{equation}\label{eq.Fmod2}
\begin{aligned}
0\rightarrow\, &\hm{}{A}{\syz{A}{d{-}1}k}{\omega_{A}}\longrightarrow\hm{}{A}{E_{A}}{\omega_{A}}\longrightarrow\nd{}{A}{\omega_{A}}\xra{\,\,\partial\,\,}\\
&\xt{1}{A}{\syz{A}{d{-}1}k}{\omega_{A}}\longrightarrow\xt{1}{A}{E_{A}}{\omega_{A}}\rightarrow 0
\end{aligned}
\end{equation}
We have \(\partial(\id)=\theta\) so \(\partial\) is surjective and \(\xt{1}{A}{E_{A}}{\omega_{A}}=0\). By duality theory (e.g.\ \cite[3.5.11]{bru/her:98}) this excludes the possibility \(\depth E_{A}=d-1\) and we conclude that \(E_{A}\) is a maximal Cohen-Macaulay module. If \(N\) is a Cohen-Macaulay module of codimension \(c\) we denote \(\xt{c}{A}{N}{\omega_{A}}\) by \(N^{\vee}\). Since \(\nd{}{A}{\omega_{A}}\cong A\) we get from \eqref{eq.Fmod2} a short exact sequence:
\begin{equation}\label{eq.Fmod3}
0\rightarrow \hm{}{A}{\syz{A}{d{-}1}k}{\omega_{A}}\longrightarrow E_{A}^{\vee}\longrightarrow\fr{m}_{A}\rightarrow 0
\end{equation}
Since \(\xt{i}{A}{k}{\omega_{A}}=0\) for \(i\neq d\), \(0\rightarrow G_{0}^{\vee}\rightarrow\dots G_{d-2}^{\vee}\rightarrow\hm{}{A}{\syz{A}{d{-}1}k}{\omega_{A}}\rightarrow 0\) is a finite \(\omega_{A}\)-resolution and so \eqref{eq.Fmod3} gives the minimal MCM approximation of the maximal ideal. Auslander introduced the fundamental module in the case \(d=2\), see \cite{aus:86}.
\end{ex}
We can make a relative version of the fundamental module in much the same way. Let \({}^{(2)\!}\Delta:\cat{CM}\rightarrow\cat{CM}\) be the morphism of fibred categories over \(\cat{NR}\) defined by taking the CM map \(h:S\rightarrow T\) to the composition \(h^{(2)}\) of \(h\) with \(\iota=1{\otimes}\id_{T}:T\rightarrow T{\otimes}_{S}T\) and taking a morphism \((g,f):h_{1}\rightarrow h_{2}\) to the composition of two cocartesian squares \((g,f^{\ot2})\) as follows:
\begin{equation}
\xymatrix@-0pt@C+6pt@R-4pt@H-0pt{
S_{1}\ar[r]^{h_{1}}\ar[d]_{g} & T_{1}\ar[d]^{f}\ar@{}[dr]|-{\mapsto} & S_{1}\ar[r]^{h_{1}}\ar[d]_{g} & T_{1}\ar[d]^{f}\ar[r]^(.4){\iota} & T_{1}{\otimes}_{S_{1}}T_{1}\ar[d]^{f^{\ot2}} \\
S_{2}\ar[r]^{h_{2}} & T_{2} & S_{2}\ar[r]^{h_{2}} & T_{2}\ar[r]^(.4){\iota} & T_{2}{\otimes}_{S_{2}} T_{2}
}
\end{equation}
There is also a functor \(\Delta:\cat{CM}\rightarrow\cat{CM}\) defined by mapping \((g,f)\) to the rightmost cocartesian square \((f,f^{\ot2})\), but it doesn't commute with the forgetful functor \(\cat{CM}\rightarrow\cat{NR}\).
Let \({}^{d}\cat{CM}\) denote the full subcategory of CM maps of pure relative dimension \(d\). Then \({}^{d}\cat{CM}\) is a fibred subcategory of \(\cat{CM}\) over \(\cat{NR}\) and \({}^{(2)\!}\Delta\) and \(\Delta\) restricts to a morphism \({}^{(2)\!}\Delta:{}^{d}\cat{CM}\rightarrow{}^{2d}\cat{CM}\) over \(\cat{NR}\) and a functor \(\Delta:{}^{d}\cat{CM}\rightarrow{}^{d}\cat{CM}\).
Let \(h:S\rightarrow T\) be a finite type CM map of pure relative dimension \(d\geq 2\). Consider \(P\) in \(\cat{P}_{h}\) (see Lemma \ref{lem.proj}) as a \(T^{\ot2}\)-module by pullback along the multiplication map \(\mu:T^{\ot2}\rightarrow T\). By Corollary \ref{cor.xtdef} \(E=\xt{d}{T^{\ot2}}{P}{\omega_{\iota}}\) is flat and finite as \(T\)-module, i.e.\ \(T\)-projective. Let \(P^{*}\) denote \(\hm{}{T}{P}{T}\). By Corollary \ref{cor.xtdef}
\begin{equation}
\nd{}{T}{E} \cong \xt{d}{T^{{\otimes} 2}}{P}{\omega_{\iota}}{\otimes} E^{*}\cong \xt{d}{T^{\ot2}}{P}{\omega_{\iota}{\otimes} E^{*}}\,.
\end{equation}
Combined with the connecting isomorphisms the identity in \(\nd{}{T}{E}\) corresponds to a canonical extension of \(T^{{\otimes} 2}\)-modules:
\begin{equation}\label{eq.Fxt}
0\rightarrow \omega_{\iota}{\otimes}_{T}\xt{d}{T^{\ot2}}{P}{\omega_{\iota}}^{*}\longrightarrow E_{h}(P)\longrightarrow \syz{T^{\ot2}}{d{-}1}P\rightarrow 0\,.
\end{equation}
Let \({}^{d}\cat{P}\) and \({}^{d}\cat{MCM}\) denote the restriction of \(\cat{P}\) and \(\cat{MCM}\) to fibred categories over \({}^{d}\cat{CM}\).
\begin{prop}\label{prop.Fmod}
Let \(d\geq 2\)\textup{.} The association \((h,P)\mapsto E_{h}(P)\) in \eqref{eq.Fxt} induces
\begin{enumerate}
\item[(i)] a functor \(E:{}^{d}\cat{P}\rightarrow {}^{d}\cat{MCM}/{}^{d}\cat{P}\) which preserves cocartesian maps and lifts the functor \(\Delta:{}^{d}\cat{CM}\rightarrow{}^{d}\cat{CM}\)\textup{,} and
\item[(ii)] a morphism \({}^{(2)\!}E:{}^{d}\cat{P}\rightarrow {}^{2d}\cat{MCM}/{}^{2d}\cat{P}\) of fibred categories over \(\cat{NR}\) which lifts \({}^{(2)\!}\Delta:{}^{d}\cat{CM}\rightarrow{}^{2d}\cat{CM}\)\textup{.}
\end{enumerate}
\end{prop}
\begin{proof}
As an extension of \(T\)-flat modules, \(E_{h}(P)\) is \(T\)-flat. Applying \(\hm{}{T^{\ot2}}{-}{\omega_{\iota}}\) to \eqref{eq.Fxt} with \(E=\xt{d}{T^{\ot2}}{P}{\omega_{\iota}}\) and \(\syz{}{d{-}1}=\syz{T^{\ot2}}{d{-}1}P\) gives an exact sequence
\begin{equation}\label{eq.Fmod4}
\begin{aligned}
0\rightarrow\, &\hm{}{T^{\ot2}}{\syz{}{d{-}1}}{\omega_{\iota}}\rightarrow\hm{}{T^{\ot2}}{E_{h}(P)}{\omega_{\iota}}\rightarrow\nd{}{T^{\ot2}}{\omega_{\iota}}{\otimes}_{T}E\xra{\partial}\\
&\xt{1}{T^{\ot2}}{\syz{}{d{-}1}}{\omega_{\iota}}\rightarrow\xt{1}{T^{\ot2}}{E_{h}(P)}{\omega_{\iota}}\rightarrow 0
\end{aligned}
\end{equation}
by Corollary \ref{cor.xtdef} and duality theory. In particular there is a canonical isomorphism \(\hm{}{T^{\ot2}}{\omega_{\iota}{\otimes}_{T}E^{*}}{\omega_{\iota}}\cong\nd{}{T^{\ot2}}{\omega_{\iota}}{\otimes}_{T}E\). We have that \(\nd{}{T^{\ot2}}{\omega_{\iota}}\) is canonically isomorphic to \(T^{\ot2}\) and \(\partial(t{\otimes}\xi)=\mu(t)\syz{T^{\ot2}}{d{-}1}(\xi)\in \xt{1}{T^{\ot2}}{\syz{}{d{-}1}}{\omega_{\iota}}\) where \(\syz{T^{\ot2}}{d{-}1}\) is the composition of the connecting isomorphisms. So \(\partial\) is surjective and \(\xt{1}{T^{\ot2}}{E_{h}(P)}{\omega_{\iota}}=0\) by Corollary \ref{cor.xtdef}. It follows that all fibres of \(E_{h}(P)\) are MCM modules and so \(E_{h}(P)\) is in \(\cat{MCM}_{\iota}\) and \(\eqref{eq.Fxt}\) is an \(\cat{MCM}_{\iota}\)-approximation of \(\syz{T^{\ot2}}{d{-}1}P\). Let \(\mc{I}_{h}\) denote the kernel of \(\mu:T^{\ot2}\rightarrow T\) and \((-)^{\vee}=\hm{}{T^{\ot2}}{-}{\omega_{\iota}}\). From \eqref{eq.Fmod4} we get another \(\cat{MCM}_{\iota}\)-approximation
\begin{equation}\label{eq.Fmod5}
0\rightarrow \hm{}{T^{\ot2}}{\syz{T^{{\otimes} 2}}{d{-}1}P}{\omega_{\iota}}\longrightarrow E_{h}(P)^{\vee}\longrightarrow\mc{I}_{h}{\otimes}_{T}\xt{d}{T^{\ot2}}{P}{\omega_{\iota}}\rightarrow 0\,.
\end{equation}
Dualising \eqref{eq.Fmod5} induces \eqref{eq.Fxt} since \(E_{h}(P)\cong E_{h}(P)^{\vee\vee}\) and \(\hm{}{T^{\ot2}}{\mc{I}_{h}}{\omega_{\iota}}\cong\omega_{\iota}\).
By Theorem \ref{thm.flatCMapprox} the image of \(E_{h}(P)^{\vee}\) in \(\cat{MCM}_{\iota}/\cat{D}_{\iota}\) is functorial in the \(T^{\ot2}\)-module \(\mc{I}_{h}{\otimes} E\) which again is contravariantly functorial in \(P\). Since \((-)^{\vee}\) induces an equivalence
\begin{equation}
\xymatrix@-0pt@C+6pt@R-4pt@H-0pt{
\vee:\cat{MCM}_{\iota}/\cat{P}_{\iota}\ar@{<->}[r]^(0.47){\simeq} & \cat{MCM}_{\iota}^{\text{op}}/\cat{D}_{\iota}^{\text{op}}:\vee
}
\end{equation}
we conclude that \(E_{h}(P)\) is functorial in \(\cat{MCM}/\cat{P}\) by our functorial choice of extension. This gives (i) and (ii).
\end{proof}
\begin{cor}\label{cor.Fmod}
For any Cohen-Macaulay map \(h:S\rightarrow T\) of pure relative dimension \(d\geq 2\) there is a finite \(T^{\ot2}\)-module \(E_{h}=E_{h}(T)\) which is faithfully flat along \(\iota:T\rightarrow T^{\ot2}\) with all fibres being maximal Cohen-Macaulay modules\textup{.} The association \(h\mapsto E_{h}\) defines a functor \({}^{d}\cat{CM}\rightarrow {}^{d}\cat{MCM}/{}^{d}\cat{P}\) lifting \(\Delta:{}^{d}\cat{CM}\rightarrow{}^{d}\cat{CM}\)\textup{.} In particular \(E_{h}\) gives \(\cat{MCM}_{\iota}\)-approximations
\begin{equation*}
0\rightarrow\omega_{\iota}\longrightarrow E_{h}\longrightarrow \syz{T^{\ot2}}{d{-}1}T\rightarrow 0\quad\text{and}
\end{equation*}
\begin{equation*}
0\rightarrow \hm{}{T^{\ot2}}{\syz{T^{\ot2}}{d{-}1}T}{\omega_{\iota}}\longrightarrow E_{h}^{\vee}\longrightarrow \mc{I}_{h}\rightarrow 0
\end{equation*}
where \(\mc{I}_{h}\) is the kernel of the multiplication map \(T^{\ot2}\rightarrow T\)\textup{.}
\end{cor}
\begin{proof}
This follows from Proposition \ref{prop.Fmod} and \eqref{eq.Fmod5} once we have proved the natural isomorphism \(\xt{d}{T^{\ot2}}{T}{\omega_{\iota}}\cong T\).
Choose a surjection of \(S\)-algebras \(P\rightarrow T\) with \(P=S[t_{1},\dots,t_{N}]\). Recall that \(\omega_{\iota}\) can be given as \(\xt{N-d}{P{\otimes} T}{T^{\ot2}}{\omega_{P{\otimes} T/T}}\) where \(\omega_{P{\otimes} T/T}=\wedge^{N}\Omega_{P{\otimes} T/T}\). There is a change of rings spectral sequence
\begin{equation}
\cE^{p,q}_{2}=\xt{q}{T^{\ot2}}{T}{\xt{p}{P{\otimes} T}{T^{\ot2}}{\omega_{P{\otimes} T/T}}}\,\Rightarrow\,\xt{p+q}{P{\otimes} T}{T}{\omega_{P{\otimes} T/T}}
\end{equation}
which by Corollary \ref{cor.xtdef} and duality theory collapses to the canonical isomorphism
\begin{equation}
\xt{d}{T^{\ot2}}{T}{\xt{N-d}{P{\otimes} T}{T^{\ot2}}{\omega_{P{\otimes} T/T}}}\cong\xt{N}{P{\otimes} T}{T}{\omega_{P{\otimes} T/T}}\,.
\end{equation}
By \cite[3.5.6]{con:00} \(\xt{N}{P{\otimes} T}{T}{\omega_{P{\otimes} T/T}}\) is canonically isomorphic to \(\omega_{T/T}=T\) as \(T^{\ot2}\)-module.
\end{proof}
We call the module \(E_{h}\) given in Corollary \ref{cor.Fmod} for the \emph{fundamental module} of the Cohen-Macaulay map \(h\).
\begin{ex}\label{ex.semicont}
Let \(k\) be an algebraically closed field and \(A\) a finite type \(k\)-algebra which is Cohen-Macaulay of pure dimension \(2\). Then the fundamental module \(E=E_{h}\) of \(h:k\rightarrow A\) is the maximal Cohen-Macaulay approximation of \(I=\ker\{A^{{\otimes} 2}\rightarrow A\}\) in \(\cat{mod}{}^{\textnormal{fl}}_{\iota}\);
\begin{equation}\label{eq.semicont}
0\rightarrow\omega_{h}{\otimes} A\longrightarrow E\longrightarrow I\rightarrow 0
\end{equation}
where \(\iota=1{\otimes}\id:A\rightarrow A^{{\otimes} 2}\) and \(\omega_{h}\cong\omega_{A}\).
Let \(t\) in \(\Spec A^{{\otimes} 2}\) be a \(k\)-point, and \(t_{i}\in \Spec A\) be the image of \(t\) by the \(i^{\text{th}}\) projection. Let \(A_{i}\) denote \(A\) localised at \(t_{i}\). Let \(\fr{m}_{i}\) be the maximal ideal in \(A_{i}\). Localising gives a local Cohen-Macaulay map \(\iota_{\fr{p}_{t}}:A_{2}\rightarrow (A^{\ot2})_{\fr{p}_{t}}\) and a module \(E_{\fr{p}_{t}}\) in \(\cat{MCM}_{\iota_{\fr{p}_{t}}}\). Let \(E(t)\) denote base change of \(E_{\fr{p}_{t}}\) to \(k(t_{2})\). If \(t_{1}=t_{2}\) then \(I(t)\cong \fr{p}_{1}\) and \(E(t)\) equals the fundamental module \(E_{A_{1}}\) of \eqref{eq.Fmod}. If \(t_{1}=t_{2}\) is singular, then \(\rk{\omega}{E(t)}=0\) while if \(t_{1}=t_{2}\) is regular then \(E(t)\cong A_{1}^{{\oplus}2}\). If \(t_{1}\neq t_{2}\) then \(I(t)\cong A_{1}\cong E_{A_{1}}\) and \(E(t)\cong A_{1}^{\oplus 2}\). This shows that \(\gamma(I)(t)\) is \emph{not} upper semi-continuous as the \(d^{i}\)-invariants are.
In particular, if \(A\) equals \(k[x,y,z]/(x^{n{+}1}{-}yz)\) with a \(2\)-dimensional rational double point at \(\fr{m}_{0}=(x,y,z)\), similar considerations give the following table of invariants (note that \(\nu_{1}=d^{-1}\)):
\begin{center}
\setlength{\extrarowheight}{2,5pt}
\begin{tabularx}{280pt}[t]{| X | l | c | c | c | c |}
\hline
\(\iota:A\rightarrow A^{{\otimes} 2}\) & \(\gamma\) & \(\nu_{1}\) & \(d^{0}\) & \(\nu_{0}\) & \(I(t)\)
\\[0.5ex]\hline\hline
\(t_{1}=t_{2}=0\) singular point & \(0\) & \(1\) & \(4\) & \(1\) & \(\fr{m}_{0}A_{\fr{m}_{0}}\) \\ \hline
\(t_{1}=t_{2}\) non-singular point & \(2\) & \(1\) & \(2\) & \(0\) & \(\fr{m}_{1}A_{1}\) \\ \hline
\(t_{1}\neq t_{2}\) & \(1\) & \(0\) & \(1\) & \(1\) & \(A_{1}\) \\ \hline
\end{tabularx}
\end{center}
\end{ex}
\section{Deformation functors and cohomology}\label{sec.def}
We extend the Cohen-Macaulay approximation over henselian local rings to deformations and obtain maps between the associated deformation functors. We also introduce the appropriate Andr{\'e}-Quillen cohomology and links the various cohomologies in a fundamental long-exact sequence.
Fix an object \(\xi=(h:S\rightarrow T,\mc{N})\) in \(\cat{mod}{}^{\textnormal{fl}}\) over \(\cat{H}\). A \emph{deformation} of \(\xi\) is a cocartesian morphism \(\alpha_{1}:\xi_{1}\rightarrow \xi\) in \(\cat{mod}{}^{\textnormal{fl}}\). A \emph{map of deformations} \(\alpha_{1}\rightarrow\alpha_{2}\) is a cocartesian morphism \(\phi:\xi_{1}\rightarrow \xi_{2}\) in \(\cat{mod}{}^{\textnormal{fl}}\) such that \(\alpha_{2}\phi = \alpha_{1}\). Deformations and maps of deformations are objects and arrows in the comma category \(\cat{Def}_{\xi}:=\cat{mod}{}^{\textnormal{fl}}_{\text{coca}}/\xi\) which is fibred in groupoids over the comma category \(\cat{H}/S\), see Lemma \ref{lem.gpoid} and the proceeding comments. Let the \emph{deformation functor} \(\df{}{\xi}:\cat{H}/S\rightarrow \Sets\) be the functor corresponding to the associated groupoid of sets \(\ol{\cat{Def}}_{\xi}\). The comma category \(\cat{Def}_{h}:=\cat{hCM}/h\) of deformations of \(h:S\rightarrow T\) is also fibred in groupoids over \(\cat{H}/S\) and we have an obvious factorisation \(\cat{Def}_{\xi}\rightarrow\cat{hCM}/h\rightarrow \cat{H}/S\) which makes \(\cat{Def}_{\xi}\) fibred in groupoids over \(\cat{hCM}/h\). To ease readability (and by abuse of notation) we put \(\df{}{(T,\mc{N})}(S_{1})=\df{}{\xi}(S_{1}{\rightarrow} S)\) and \(\df{}{T}(S_{1})=\df{}{h}(S_{1}{\rightarrow} S)\). We also write a \emph{deformation of \((T,\mc{N})\)} meaning a deformation of \(\xi\) and likewise in similar situations.
For each object \(\xi_{i}=(h_{i},\mc{N}_{i})\) in \(\cat{mod}{}^{\textnormal{fl}}\) over \(\cat{H}\) we choose a minimal \(\cat{MCM}\)-approximation \(\pi_{i}:\mc{L}_{i}\rightarrow \mc{M}_{i}\xra{\pi_{i}}\mc{N}_{i}\) and a minimal \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull \(\iota_{i}:\mc{N}_{i}\xra{\iota_{i}}\mc{L}'_{i}\rightarrow \mc{M}_{i}'\) which exist by Corollary \ref{cor.locCMapprox} and Corollary \ref{cor.minapprox}. For each deformation \(\alpha_{i}:\xi_{i}\rightarrow \xi_{0}\) we choose extensions to commutative diagrams of deformations
\begin{equation}\label{eq.right2}
\xymatrix@C-0pt@R-8pt@H-30pt{
\mc{L}_{i}\ar[r]\ar@{.>}[d]_{\lambda_{i}} & \mc{M}_{i}\ar[r]^{\pi_{i}}\ar@{.>}[d]_{\mu_{i}} & \mc{N}_{i}\ar[d]_{\nu_{i}} \\
\mc{L}_{0}\ar[r] & \mc{M}_{0} \ar[r]^{\pi_{0}} & \mc{N}_{0}
}
\qquad\text{and}\qquad
\xymatrix@C-0pt@R-8pt@H-30pt{
\mc{N}_{i}\ar[r]^{\iota_{i}}\ar[d]_{\nu_{i}} & \mc{L}_{i}'\ar[r]\ar@{.>}[d]_{\lambda_{i}'} & \mc{M}_{i}'\ar@{.>}[d]_{\mu_{i}'} \\
\mc{N}_{0}\ar[r]^{\iota_{0}} & \mc{L}_{0}' \ar[r] & \mc{M}_{0}'
}
\end{equation}
as follows: By Corollary \ref{cor.minapprox} a base change of \(\pi_{i}\) by \(h_{i}\rightarrow h_{0}\) gives a minimal \(\cat{MCM}_{h_{0}}\)-approximation \(\mc{M}_{i}^{\#}\rightarrow\mc{N}_{i}^{\#}\xra{\simeq} \mc{N}_{0}\). By minimality it is isomorphic to \(\pi_{0}\). Choose an isomorphism. Let \(\mu_{i}\) be the composition \(\mc{M}_{i}\rightarrow \mc{M}_{i}^{\#}\xra{\simeq}\mc{M}_{0}\). It is cocartesian. Do similarly for the \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull. Let these choices be fixed.
The diagrams in \eqref{eq.right2} will be called an \(\cat{MCM}\)-approximation (denoted \(\pi_{*}\)) respectively a \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull (denoted \(\iota_{*}\)) of \(\nu_{i}\) (this terminology can be justified).
\begin{defn}\label{defn.defmap}
There are four maps
\begin{equation*}
\sigma_{X}:\df{}{(h_{0},\mc{N}_{0})}\longrightarrow\df{}{(h_{0},X)}\text{ of functors }\cat{H}/S_{0}\longrightarrow\Sets
\end{equation*}
where \(X\) can be \(\mc{M}_{0},\mc{L}_{0},\mc{L}_{0}'\) and \(\mc{M}_{0}'\) given by \([(h_{i}\rightarrow h_{0}, \nu_{i})]\mapsto [(h_{i}\rightarrow h_{0}, x)]\) for \(x\) equal to \(\mu_{i},\lambda_{i},\lambda_{i}'\) and \(\mu_{i}'\) in \eqref{eq.right2} respectively.
\end{defn}
The following lemma implies that these maps are well defined and independent of choices.
\begin{lem}\label{lem.defCMapprox2}
Given two deformations \(\nu_{ij}:\mc{N}_{ij}\rightarrow\mc{N}_{0j}\) \textup{(}\(j=1,2\)\textup{)} in \(\cat{mod}{}^{\textnormal{fl}}\) over \(h_{ij}\rightarrow h_{0j}\) in \(\cat{hCM}\) and \(\cat{MCM}\)-approximations \(\pi_{*j}\) \textup{(}respectively \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hulls \(\iota_{*j}\)\textup{)} of \(\nu_{ij}\)\textup{.} Suppose we have a map of short exact sequences \(\pi_{01}\rightarrow\pi_{02}\) \textup{(}respectively \(\iota_{01}\rightarrow\iota_{02}\)\textup{)} and a map \(\alpha:\mc{N}_{i1}\rightarrow\mc{N}_{i2}\) lifting \(\mc{N}_{01}\rightarrow\mc{N}_{02}\)\textup{,} i\textup{.}e\textup{.}\ such that the following two diagrams of solid arrows are commutative\textup{:}
\begin{equation*}\label{eq.C}
\xymatrix@C-20pt@R-8pt@H-30pt{
& \mc{L}_{i2}\ar[rr]\ar[dd]^(0.3){\lambda_{i2}}|!{[dl];[dr]}\hole && \mc{M}_{i2}\ar[rr]^(0.40){\pi_{i2}}\ar[dd]^(0.3){\mu_{i2}}|!{[dl];[dr]}\hole
&& \mc{N}_{i2}\ar[dd]^(0.3){\nu_{i2}} \\
\mc{L}_{i1}\ar@{.>}[ur]\ar[rr]^(0.2){}\ar[dd]^(0.3){\lambda_{i1}} && \mc{M}_{i1}\ar@{.>}[ur]^(0.4){\gamma}\ar[rr]^(0.4){\pi_{i1}}\ar[dd]^(0.3){\mu_{i1}}
&& \mc{N}_{i1}\ar[ur]^(0.4){\alpha}\ar[dd]^(0.3){\nu_{i1}} & \\
& \mc{L}_{02}\ar[rr]|!{[ur];[dr]}\hole && \mc{M}_{02} \ar[rr]^(0.35){\pi_{02}}|!{[ur];[dr]}\hole && \mc{N}_{02} \\
\mc{L}_{01}\ar[ur]\ar[rr] && \mc{M}_{01}\ar[ur]^(0.45){\beta} \ar[rr]^(0.4){\pi_{01}} && \mc{N}_{01}\ar[ur] &
}
\quad
\xymatrix@C-20pt@R-10pt@H-30pt{
& \mc{N}_{i2}\ar[rr]^(0.40){\iota_{i2}}\ar[dd]^(0.3){\nu_{i2}}|!{[dl];[dr]}\hole && \mc{L}'_{i2}\ar[rr]\ar[dd]^(0.3){\lambda'_{i2}}|!{[dl];[dr]}\hole
&& \mc{M}'_{i2}\ar[dd]^(0.3){\mu'_{i2}} \\
\mc{N}_{i1}\ar[ur]^(0.4){\alpha}\ar[rr]^(0.35){\iota_{i1}}\ar[dd]^(0.3){\nu_{i1}} && \mc{L}'_{i1}\ar@{.>}[ur]^(0.4){\gamma'}\ar[rr]^(0.2){}\ar[dd]^(0.3){\lambda'_{i1}}
&& \mc{M}'_{i1}\ar@{.>}[ur]\ar[dd]^(0.3){\mu'_{i1}} & \\
& \mc{N}_{02}\ar[rr]^(0.35){\iota_{02}}|!{[ur];[dr]}\hole && \mc{L}'_{02} \ar[rr]^(0.2){}|!{[ur];[dr]}\hole && \mc{M}'_{02} \\
\mc{N}_{01}\ar[ur]\ar[rr]^(0.4){\iota_{01}} && \mc{L}'_{01}\ar[ur]_(0.55){\beta'} \ar[rr] && \mc{M}'_{01}\ar[ur] &
}
\end{equation*}
Then there exist maps \(\gamma:\mc{M}_{i1}\rightarrow\mc{M}_{i2}\) and \(\gamma':\mc{L}'_{i1}\rightarrow\mc{L}'_{i2}\) such that the induced left \textup{(}respectively right\textup{)} diagram commutes\textup{.} If \(\alpha\) is cocartesian\textup{,} so are \(\gamma\) and \(\gamma'\)\textup{.}
\end{lem}
\begin{proof}
Consider the \(\cat{MCM}\)-approximation case. By applying base changes to the front diagram, we can reduce the problem to the case \(h_{i1}\rightarrow h_{01}\) equals \(h_{i2}\rightarrow h_{02}\). Then, by Corollary \ref{cor.locCMapprox}, there is a lifting \(\gamma_{1}:\mc{M}_{i1}\rightarrow\mc{M}_{i2}\) of \(\alpha\). We would like to adjust \(\gamma_{1}\) so that it lifts \(\beta\) too. We have that \(\mu_{i2}\gamma_{1}-\beta\mu_{i1}\) factors through \(\mc{L}_{02}\) by a map \(\tau_{i}:\mc{M}_{i1}\rightarrow\mc{L}_{02}\) which induces a unique map \(\tau_{0}:\mc{M}_{01}\rightarrow\mc{L}_{02}\) since \(\mu_{i1}\) is cocartesian. If \(\mc{D}_{*}\twoheadrightarrow\mc{L}_{i2}\) is a finite \(\cat{D}\)-resolution, then base change gives a finite \(\cat{D}\)-resolution \(\mc{D}_{*}^{\#}\twoheadrightarrow \mc{L}_{02}\) and \(\tau_{0}\) lifts to a \(\sigma_{0}:\mc{M}_{01}\rightarrow\mc{D}_{0}^{\#}\) by Corollary \ref{cor.locCMapprox}. By Corollary \ref{cor.xtdef} there is a \(\sigma:\mc{M}_{i1}\rightarrow\mc{D}_{0}\) lifting \(\sigma_{0}\) and subtracting the induced map \(\mc{M}_{i1}\rightarrow\mc{M}_{i2}\) from \(\gamma_{1}\) gives our desired \(\gamma\). If \(\alpha\) is an isomorphism so is \(\gamma\) by minimality of the approximations \(\pi_{ij}\).
The argument for the \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-case is similar.
\end{proof}
\begin{rem}\label{rem.defmap}
There are maps of fibred categories inducing the maps \(\sigma_{X}\) in Definition \ref{defn.defmap}. Two maps \(\alpha,\beta: (h_{1},\mc{N}_{1})\rightarrow (h_{2},\mc{N}_{2})\) in \(\cat{Def}_{(h_{0},\mc{N}_{0})}\) are stably equivalent if \(h_{1}=h_{2}\) and \(\alpha{-}\beta\) factors through an object in \(\cat{D}\). Let \(\ul{\cat{Def}}_{(h_{0},\mc{N}_{0})}\) denote the resulting quotient category which is fibred over \(\cat{hCM}/h_{0}\) and over \(\cat{H}/S_{0}\). Then Lemma \ref{lem.defCMapprox2} implies that there are well defined maps of categories fibred in groupoids \(\bs{\sigma}_{X}:\ul{\cat{Def}}_{(h_{0},\mc{N}_{0})}\rightarrow\ul{\cat{Def}}_{(h_{0},X)}\) for \(X\) equal to \(\mc{M}_{0}\), \(\mc{L}_{0}\), \(\mc{L}_{0}'\) and \(\mc{M}_{0}'\). The associated map of functors is \(\sigma_{X}\). Note that \(\ul{\cat{Def}}_{(h_{0},\mc{N}_{0})}\) is different from \(\underline{\cat{mod}}{}^{\textnormal{fl}}_{\text{coca}}/(h_{0},\mc{N}_{0})\). Stably isomorphic modules will in general have different deformation functors. E.g.\ let \(N=A\oplus\omega_{A}\). If \(A\) is not Gorenstein, then one can have \(\xt{1}{A}{N}{N}\neq 0\). But in the stable category \(N\) is isomorphic to \(A\) which is infinitesimally rigid.
\end{rem}
We have the following reformulation. To \((h:S\rightarrow T,\mc{N})\) in \(\cat{mod}{}^{\textnormal{fl}}\) consider \(\varGamma=T{\oplus}\mc{N}\) as a graded \(S\)-algebra with \(T\) in degree \(0\) and \(\mc{N}\) in degree \(1\). A deformation of graded algebras \({\varGamma}_{1}\rightarrow{\varGamma}\) over \(S_{1}\rightarrow S\) in \(\cat{H}/S\) is equivalent to a deformation \((T_{1},\mc{N}_{1})\) of \((T,\mc{N})\). More generally, given a homogeneous morphism of \(\BB{Z}\)-graded rings \(S\rightarrow T\) and a graded \(T\)-module \(M\), there are \emph{Andr{\'e}-Quillen cohomology groups of graded algebras} \(\gH{0}^{i}(S,T,M)=\cH^{i}{\hm{{\text{gr}}}{T}{L^{\text{gr}}_{T/S}}{M}}\). Here \(L^{\text{gr}}_{T/S}\) is the graded cotangent complex defined as \(\Omega_{P/S}{\otimes}_{P}T\) where \(P=P_{S}(T)\) is a graded simplicial degree-wise free \(S\)-algebra resolution of \(T\) and \(\Omega_{P/S}\) denotes the K{\"a}hler differentials, see \cite[IV]{ill:71} for more details (in a more general situation). See also \cite{kle:79}.
\begin{defn}\label{def.grobs}
Given graded ring homomorphisms \(h:S\rightarrow T\) and \(p:S_{1}\rightarrow S\). Assume \(p\) is surjective. A \emph{lifting} of \(h\) (`of \(T\)') along \(p\) (`to \(S_{1}\)') is a commutative diagram of graded ring homomorphisms
\begin{equation*}\label{eq.lift}
\xymatrix@C-0pt@R-12pt@H-30pt{
T & T_{1}\ar[l]_{q} \\
S\ar[u]^{h} & S_{1}\ar[l]_{p}\ar[u]_{h_{1}}
}
\quad
\textnormal{with \(q{\otimes} S:T_{1}{\otimes}_{S_{1}}S\cong T\) and \(\tor{S_{1}}{1}{T_{1}}{S}=0\).}
\end{equation*}
Two liftings \(T_{1}\) and \(T'_{1}\) of \(T\) to \(S_{1}\) are equivalent if there is a graded \(S_{1}\)-algebra isomorphism \(T_{1}\cong T'_{1}\) commuting with \(q\) and \(q'\).
\end{defn}
There is an obstruction theory for liftings of graded algebras in terms of graded Andr{\'e}-Quillen cohomology groups.
\begin{prop}[\cite{ill:71,kle:79}]\label{prop.grobs}
Given graded ring homomorphisms \(S\rightarrow T\) and \(p:S_{1}\rightarrow S\) with \(p\) surjective and \(I^{2}=0\) for \(I=\ker p\)\textup{.}
\begin{enumerate}
\item[(i)] There exists an element \(\ob(p,T)\in\gH{0}^{2}(S,T,T{\otimes} I)\) which is natural in \(p\) such that \(\ob(p,T)=0\) if and only if there exists a lifting of \(T\) to \(S_{1}\)\textup{.}
\item[(ii)] If \(\ob(p,T)=0\) then the set of equivalence classes of liftings of \(T\) to \(S_{1}\) is a torsor for \(\gH{0}^{1}(S,T,T{\otimes} I)\) which is natural in \(p\)\textup{.}
\end{enumerate}
\end{prop}
The element \(\ob(p,T)\) is called the \emph{obstruction class} of \((p,T)\). If the rings and modules are concentrated in degree \(0\) this equals the ungraded case and the cohomology groups equals the ungraded Andr{\'e}-Quillen cohomology \(\cH^{*}(S,T,T{\otimes} I)\).
\begin{defn}\label{def.modulobs}
Given a lifting diagram of ungraded ring homomorphisms as in Definition \ref{def.grobs} and a \(T\)-module \(N\). Then a \emph{lifting} of \(N\) to \(T_{1}\) is a \(T_{1}\)-module \(N_{1}\) with \(\tor{S_{1}}{1}{N_{1}}{S}=0\) and a map \(N_{1}\rightarrow N\) inducing an isomorphism \(N_{1}{\otimes} S\cong N\). Two liftings \(N_{1}\) and \(N_{1}'\) of \(N\) to \(T_{1}\) are \emph{equivalent} if there is an isomorphism of \(T_{1}\)-modules \(N_{1}\cong N_{1}'\) commuting with the maps to \(N\).
\end{defn}
There is also an obstruction theory for liftings of modules in terms of \(\Ext\) groups.
\begin{prop}[{\cite[IV 3.1.5]{ill:71}}]\label{prop.obsmodule}
Given \textup{(}ungraded\textup{)} ring homomorphisms as in \textup{Definition \ref{def.grobs}} with \(I^{2}=0\) and a \(T\)-module \(N\)\textup{.}
\begin{enumerate}
\item[(i)] There exists an element \(\ob(q,N)\in\xt{2}{T}{N}{N{\otimes} I}\) which is natural in \(q\) such that \(\ob(q,N)=0\) if and only if there exists a lifting of \(N\) to \(T_{1}\)\textup{.}
\item[(ii)] If \(\ob(q,N)=0\) then the set of equivalence classes of liftings of \(N\) to \(T_{1}\) is a torsor for \(\xt{1}{T}{N}{N{\otimes} I}\) which is natural in \(q\)\textup{.}
\end{enumerate}
\end{prop}
The element \(\ob(q,N)\) is called the \emph{obstruction class} of \((q,N)\). The following long-exact sequence connects these three cohomologies.
\begin{prop}\label{prop.lang}
Suppose \(T\) is an \textup{(}ungraded\textup{)} \(S\)-algebra and \(N\) is a \(T\)-module\textup{.} Let \({\varGamma}=T{\oplus} N\) be the graded \(S\)-algebra with \(T\) in degree \(0\) and \(N\) in degree \(1\) and let \(J\) be a graded \({\varGamma}\)-module with graded components \(J=J_{0}{\oplus}J_{1}\) of degree \(0\) and \(1\)\textup{.} Then there is a natural long-exact sequence\textup{:}
\begin{align*}
0\rightarrow\, & \hm{}{T}{N}{J_{1}}\longrightarrow \gDer{0}_{S}({\varGamma},J)\longrightarrow \Der_{S}(T,J_{0})\xra{\,\,\partial\,\,} \\
& \xt{1}{T}{N}{J_{1}}\longrightarrow \gH{0}^{1}(S,{\varGamma},J)\longrightarrow \cH^{1}(S,T,J_{0})\xra{\,\,\partial\,\,} \xt{2}{T}{N}{J_{1}}\rightarrow\dots
\end{align*}
\end{prop}
\begin{proof}
To the graded ring homomorphisms \(S\rightarrow T\rightarrow {\varGamma}\) there is a distinguished triangle of transitivity
\begin{equation}\label{eq.triangle}
L^{\text{gr}}_{\varGamma/T/S}:\,\,L^{\text{gr}}_{T/S}{\otimes}_{T}{\varGamma}\longrightarrow L^{\text{gr}}_{{\varGamma}/S}\longrightarrow L^{\text{gr}}_{{\varGamma}/T}\longrightarrow L^{\text{gr}}_{T/S}{\otimes}_{T}{\varGamma}[1]
\end{equation}
in the graded derived category of \({\varGamma}\), see \cite[IV 2.3.4]{ill:71}. The (standard) simplicial resolution \(P_{T}({\varGamma})\) equals \(T\) in degree \(0\), the (standard) \(T\)-free resolution \(F_{T}(N)\) of the \(T\)-module \(N\) in degree \(1\), and higher degree terms, see \cite[IV 1.3.2.1]{ill:71}. It follows that \(\hm{{\text{gr}}}{{\varGamma}}{L^{\text{gr}}_{{\varGamma}/T}}{J}=\hm{}{T}{F_{T}(N)}{J_{1}}\). Since \(L^{\text{gr}}_{T/S}=L_{T/S}\) is consentrated in degree \(0\), \(\hm{{\text{gr}}}{{\varGamma}}{L^{\text{gr}}_{T/S}{\otimes}_{T}{\varGamma}}{J}= \hm{}{T}{L_{T/S}}{J_{0}}\).
\end{proof}
\begin{lem}\label{lem.cohmap}
Let \(h:S\rightarrow T\) be a finite type Cohen-Macaulay map and let \(\mc{N}\) be a \(T\)-module in \(\cat{mod}{}^{\textnormal{fl}}_{h}\)\textup{.} Let \(\mc{L}\rightarrow\mc{M}\xra{\pi}\mc{N}\) and \(\mc{N}\xra{\iota}\mc{L}'\rightarrow\mc{M}'\) be an \(\cat{MCM}_{h}\)-approximation and a \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\)-hull of \(\mc{N}\)\textup{.} Let \(X_{i}\) denote \(\mc{N}\)\textup{,} \(\mc{M}\) and \(\mc{L}'\) for \(i=0,1,2\) respectively\textup{,} and put \(\varGamma_{i}=T{\oplus}X_{i}\)\textup{.} Let \(I\) be any \(S\)-module\textup{.}
Then there are natural maps of short exact sequences of complexes \textup{(}see \eqref{eq.triangle}\textup{)}
\begin{equation*}
\hm{\textnormal{gr}}{\varGamma_{0}}{L^{\textnormal{gr}}_{\varGamma_{0}/T/S}}{\varGamma_{0}{\otimes} I}\xra{\pi^{*}}\hm{\textnormal{gr}}{\varGamma_{1}}{L^{\textnormal{gr}}_{\varGamma_{1}/T/S}}{\varGamma_{0}{\otimes} I}\xla{\pi_{*}}\hm{\textnormal{gr}}{\varGamma_{1}}{L^{\textnormal{gr}}_{\varGamma_{1}/T/S}}{\varGamma_{1}{\otimes} I}
\end{equation*}
and
\begin{equation*}
\hm{\textnormal{gr}}{\varGamma_{0}}{L^{\textnormal{gr}}_{\varGamma_{0}/T/S}}{\varGamma_{0}{\otimes} I}\xra{\iota_{*}}\hm{\textnormal{gr}}{\varGamma_{0}}{L^{\textnormal{gr}}_{\varGamma_{0}/T/S}}{\varGamma_{2}{\otimes} I}\xla{\iota^{*}}\hm{\textnormal{gr}}{\varGamma_{2}}{L^{\textnormal{gr}}_{\varGamma_{2}/T/S}}{\varGamma_{2}{\otimes} I}\textup{.}
\end{equation*}
The induced maps of graded Andr{\'e}-Quillen cohomology
\begin{align*}
\gH{0}^{n}(\pi_{*}) & :\gH{0}^{n}(S,\varGamma_{1},\varGamma_{1}{\otimes} I)\longrightarrow\gH{0}^{n}(S,\varGamma_{1},\varGamma_{0}{\otimes} I)\quad\text{and} \\
\gH{0}^{n}(\iota^{*}) & :\gH{0}^{n}(S,\varGamma_{2},\varGamma_{2}{\otimes} I)\longrightarrow\gH{0}^{n}(S,\varGamma_{0},\varGamma_{2}{\otimes} I)
\end{align*}
are isomorphisms for \(n>0\) and surjections for \(n=0\)\textup{.}
\end{lem}
\begin{proof}
There is a natural map \(L^{\text{gr}}_{\varGamma_{1}/T/S}{\otimes}_{\varGamma_{1}}\varGamma_{0}\rightarrow L^{\text{gr}}_{\varGamma_{0}/T/S}\) of short exact sequences of complexes (cf.\ \cite[II 2.1.1.6]{ill:71}) induced by the graded \(T\)-algebra map \(\varGamma_{1}\rightarrow\varGamma_{0}\). This gives \(\pi^{*}\). Covariance along \(\varGamma_{1}{\otimes} I\rightarrow\varGamma_{0}{\otimes} I\) gives \(\pi_{*}\). In each (cohomological) degree the rightmost terms are naturally identified with \(\hm{}{T}{L_{T/S}}{T{\otimes} I}\) as in the proof of Proposition \ref{prop.lang}. By Theorem \ref{thm.flatCMapprox} and Corollary \ref{cor.xtdef} one has \(\xt{n}{T}{\mc{M}}{\mc{L}{\otimes} I}=0\) for \(n>0\) and the \(\gH{0}^{n}(\pi_{*})\)-statement follows. The other case is similar.
\end{proof}
By Lemma \ref{lem.cohmap} (and Theorem \ref{thm.flatCMapprox}) we get induced natural maps for \(n>0\)
\begin{equation}\label{eq.sigma}
\sigma^{n}_{j}(I):\gH{0}^{n}(S,\varGamma_{0},\varGamma_{0}{\otimes}_{S} I)\longrightarrow\gH{0}^{n}(S,\varGamma_{j},\varGamma_{j}{\otimes}_{S} I)\,\, \text{for}\,\, j=1,\,2 \quad\textnormal{and}
\end{equation}
\begin{equation}\label{eq.tau}
\tau^{n}_{j}(I):\xt{n}{T}{X_{0}}{X_{0}{\otimes}_{S}I}\longrightarrow \xt{n}{T}{X_{j}}{X_{j}{\otimes}_{S}I}\,\, \text{for}\,\, j=1,\,2.
\end{equation}
\begin{ex}\label{ex.lang} By elementary diagram chase Lemma \ref{lem.cohmap} gives the following:
\begin{itemize}
\item[(i)] If \(\pi^{*}:\xt{n}{T}{\mc{N}}{\mc{N}{\otimes} I}\rightarrow\xt{n}{T}{\mc{M}}{\mc{N}{\otimes} I}\) is an isomorphism for \(n=1\) and injective for \(n=2\) then \(\sigma^{1}_{1}(I)\) is an isomorphism and \(\sigma^{2}_{1}(I)\) is injective.
\item[(ii)] If \(\iota_{*}:\xt{n}{T}{\mc{N}}{\mc{N}{\otimes} I}\rightarrow\xt{n}{T}{\mc{N}}{\mc{L}'{\otimes} I}\) is an isomorphism for \(n=1\) and injective for \(n=2\) then \(\sigma^{1}_{2}(I)\) is an isomorphism and \(\sigma^{2}_{2}(I)\) is injective.
\end{itemize}
\end{ex}
\section{Maps of deformation functors induced by \\Cohen-Macaulay approximation}
In order to use Artin's Approximation Theorem \cite{art:69} as extended by D.\ Popescu \cite{pop:86,pop:90} we fix an excellent ring \(\varLambda\) (see \cite[7.8.2]{EGAIV2}). We consider the category of henselian local \(\varLambda\)-algebras in \(\cat{H}\), denoted \({{}_{\varLambda}\cat{H}}\). Fibred categories \(\cat{hCM}\) and \(\cat{mod}{}^{\textnormal{fl}}\) over \({{}_{\varLambda}\cat{H}}\) and \(\cat{Def}_{h}\) and \(\cat{Def}_{\xi}\) over \({{{}_{\varLambda}\cat{H}}}/S\) are defined essentially as in Section \ref{sec.def}. Our previous constructions and results are valid in this context as well. In particular deformation functors \(\Def_{(T,\mc{N})}:{{}_{\varLambda}\cat{H}}/S\rightarrow\Sets\) are defined and the \(\cat{MCM}\)-approximation and \(\hat{\cat{D}}{}^{\textnormal{fl}}\)-hull induce maps of deformation functors as in Definition \ref{defn.defmap}
\begin{defn}\label{def.smooth}
Let \({{}_{\varLambda}\cat{A}}/k\) denote the subcategory of artin rings in \({{}_{\varLambda}\cat{H}}/k\).
Let \(F\) and \(G\) be set-valued functors on \({{}_{\varLambda}\cat{H}}/k\) (or \({{}_{\varLambda}\cat{A}}/k\)) with \(\#F(k)=1=\#G(k)\). A map \(\phi:F\rightarrow G\)
is \emph{smooth} (formally smooth) if the natural map of sets \(f_{\phi}:F(S)\rightarrow F(S_{0})\times_{G(S_{0})}G(S)\) is surjective for all surjections \(\pi:S\rightarrow S_{0}\) in \({{}_{\varLambda}\cat{H}}/k\) (in \({{}_{\varLambda}\cat{A}}/k\)).
An element \(\nu\in F(R)\) is \emph{versal} if the induced map \(\hm{}{{{}_{\varLambda}\cat{H}}/k}{R}{-}\rightarrow F\) is smooth and \(R\) is algebraic as \(\varLambda\)-algebra. If the map is bijective then \(\nu\) is universal. An element \(\nu\in F(R)\) (or a formal element, i.e.\ a tower \(\{\nu_{n}\}\in \varprojlim F(R/\fr{m}_{R}^{n+1})\)) is \emph{formally versal} if the induced map \(\hm{}{{{}_{\varLambda}\cat{H}}/k}{R}{-}\rightarrow F\) of functors restricted to \({{}_{\varLambda}\cat{A}}/k\) is formally smooth.
See \cite{art:74}.
\end{defn}
\begin{thm}\label{thm.defgrade}
Let \(k\) be a field, \(A\) a Cohen-Macaulay local algebraic \(k\)-algebra and \(N\) a finite \(A\)-module\textup{.} Let \(N\rightarrow L'\rightarrow M'\) be the minimal \(\hat{\cat{D}}_{A}\)-hull and \(L\rightarrow M\rightarrow N\) the minimal \(\cat{MCM}_{A}\)-approximation of \(N\)\textup{.} Consider the map
\begin{equation*}
\sigma_{L'}:\df{}{(A,N)}\longrightarrow\df{}{(A,L')}\text{ of functors }{{}_{\varLambda}\cat{H}}/k\longrightarrow\Sets
\end{equation*}
as in \textup{Definition \ref{defn.defmap}.}
\begin{enumerate}
\item[(i)] If\, \(\hm{}{A}{N}{M'}=0\) then \(\sigma_{L'}\) is injective\textup{.}
\item[(ii)] If\, \(\xt{1}{A}{N}{M'}=0\) then \(\sigma_{L'}\) is formally smooth\textup{.}
\item[(iii)] If\, \(\xt{1}{A}{N}{M'}=0\) and \(\df{}{(A,N)}\) has a versal element then \(\sigma_{L'}\) is smooth\textup{.}
\end{enumerate}
The analogous statements hold for \(\sigma_{L}:\df{}{(A,N)}\rightarrow\df{}{(A,L)}\) with \(\xt{1}{A}{N}{M}=0\) in \textup{(i)} and \(\xt{2}{A}{N}{M}=0\) in \textup{(ii)} and \textup{(iii).}
\end{thm}
\begin{ex}\label{ex.defgrade}
Note that \(\grade N\geq 1\) implies condition (i) and \(\grade N\geq 2\) implies both condition (i) and (ii).
\end{ex}
\begin{proof}
(i): Given \(S\) in \({{}_{\varLambda}\cat{H}}/k\) and deformations \(({}^{i\!}h:S\rightarrow {}^{i\!}T,{}^{i\!}\mc{N})\) of \((A,N)\) to \(S\) for \(i=1,2\). Assume that the images \(({}^{i\!}h,{}^{i\!}\mc{L}')\) under \(\sigma_{L'}\) are isomorphic to \((h:S\rightarrow T,\mc{L}')\). Pullback of all these modules along the isomorphisms of \(h\) with \({}^{i\!}h\) induce deformations over \(h\). We show that the \({}^{i\!}\mc{N}\) are isomorphic as deformations over \(h\) which implies that \(\sigma_{L'}\) is injective. Let \(S_{n}=S/\fr{m}_{S}^{n+1}\), \(T_{n}=T{\otimes} S_{n}\) etc.. We construct a tower of isomorphisms \(\{\phi_{n}:{}^{1\!}\mc{N}_{n}\cong{}^{2\!}\mc{N}_{n}\}\) and conclude by Lemma \ref{lem.Aapprox} that the deformations are isomorphic. The case \(n=0\) is trivial. Given \(\phi_{n}\) and use it to identify the \({}^{i\!}\mc{N}_{n}\) and denote them by \(\mc{N}_{n}\). Let \(I=\ker\{S_{n+1}\rightarrow S_{n}\}\). By Proposition \ref{prop.obsmodule} there exists an element \(\theta\) in \(\xt{1}{T_{n}}{\mc{N}_{n}}{\mc{N}_{n}{\otimes} I}\) giving the ``difference'' of the two deformations of \(\mc{N}_{n}\). But \(\mc{N}_{n}{\otimes} I\cong N{\otimes} I\) and by the edge map isomorphism of \eqref{eq.ss} we get \(\xt{i}{T_{n}}{\mc{N}_{n}}{\mc{N}_{n}{\otimes} I}\cong\xt{i}{A}{N}{N}{\otimes} I\) for all \(i\). If \(i>0\) then \(\xt{i}{A}{L'}{L'}\cong\xt{i}{A}{N}{L'}\) and \(\xt{1}{A}{N}{N}\rightarrow\xt{i}{A}{N}{L'}\) is injective by assumption.
The obtained injective map \(p:\xt{1}{A}{N}{N}{\otimes} I\rightarrow \xt{1}{A}{L'}{L'}{\otimes} I\) induces a map of the torsor actions in Proposition \ref{prop.obsmodule} on the liftings of \(\mc{N}_{n}\) and of \(\mc{L}'_{n}\) to \(T_{n+1}\). Since the \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h_{n+1}}\)-hulls of the \({}^{i\!}\mc{N}_{n+1}\) are isomorphic as deformations, \(p\) maps \(\theta\) to \(0\) and so \(\theta=0\) and by Proposition \ref{prop.obsmodule} the \({}^{i\!}\mc{N}_{n+1}\) are isomorphic by an isomorphism \(\phi_{n+1}\) compatible with \(\phi_{n}\).
(ii): Let \(S\rightarrow \bar{S}\) in \({{}_{\varLambda}\cat{A}}/k\) be surjective with kernel \(I\), \(\xi=(h:S\rightarrow T,\mc{L}')\) a deformation of \((A,L')\) to \(S\) and let \(\bar{\xi}=(\bar{h}:\bar{S}\rightarrow\bar{T},\bar{\mc{L}}')\) denote the base change of \(\xi\) to \(\bar{S}\). Suppose there is a deformation \((h^{(1)}:\bar{S}\rightarrow T^{(1)},\bar{\mc{N}})\) of \((A,N)\) which \(\sigma_{L'}\) maps to \(\bar{\xi}\). As above we can assume that \(h^{(1)}=\bar{h}\). By induction on the length of \(S\) we can assume that \(I\cdot\fr{m}_{S}=0\). We find that \(\ob(T\rightarrow \bar{T},\bar{\mc{N}})\) in Proposition \ref{prop.obsmodule} maps to \(\ob(T\rightarrow \bar{T},\bar{\mc{L}}')\) under \(\xt{2}{A}{N}{N}{\otimes} I\rightarrow \xt{2}{A}{L'}{L'}{\otimes} I\) which by the assumption is injective. Since \(\mc{L}'\) lifts \(\bar{\mc{L}'}\) to \(T\) we have \(\ob(T\rightarrow \bar{T},\bar{\mc{L}}')=0\). By Proposition \ref{prop.obsmodule} there exists a lifting \({}^{1\!}\mc{N}\) of \(\bar{\mc{N}}\) to \(T\). If \(\sigma_{L'}({}^{1\!}\mc{N})={}^{1\!}\mc{L}'\) the difference of \({}^{1\!}\mc{L}'\) and \(\mc{L}'\) gives by Proposition \ref{prop.obsmodule} a \(\theta\in \xt{1}{A}{L'}{L'}{\otimes} I\). By assumption \(\xt{1}{A}{N}{N}{\otimes} I\) maps surjectively to \(\xt{1}{A}{L'}{L'}{\otimes} I\) and a lifting of \(\theta\) perturbs \({}^{1\!}\mc{N}\) to a lifting \(\mc{N}\) of \(\bar{\mc{N}}\) with \(\sigma_{L'}(\mc{N})=\mc{L}'\).
(iii): Let \(S\rightarrow \bar{S}\) in \({{}_{\varLambda}\cat{H}}/k\) be surjective with kernel \(J\), \(\xi=(h:S\rightarrow T,\mc{L}')\) a deformation of \((A,L')\) to \(S\) and let \(\bar{\xi}=(\bar{h}:\bar{S}\rightarrow\bar{T},\bar{\mc{L}}')\) denote the base change of \(\xi\) to \(\bar{S}\). Suppose there is a deformation \((h^{(1)}:\bar{S}\rightarrow T^{(1)},\bar{\mc{N}})\) of \((A,N)\) which \(\sigma_{L'}\) maps to \(\bar{\xi}\). Again we can assume that \(h^{(1)}=\bar{h}\). We will find a deformation \(\mc{N}\) lifting \(\bar{\mc{N}}\) such that \(\sigma_{L'}(h,\mc{N})=(h,\mc{L}')\).
Any \(S\) in \({{}_{\varLambda}\cat{H}}/k\) is a direct limit of a filtering system of algebraic \(\varLambda\)-algebras in \({{}_{\varLambda}\cat{H}}/k\). Since \(\df{}{(A,L')}\) is locally of finite presentation (\(A\) is algebraic and \(L'\) has finite presentation) it is sufficient to prove the lifting property for \(S\) algebraic. Since \(\varLambda\) is excellent, so is \(S\) by \cite[7.8.3]{EGAIV2} and \cite[18.7.6]{EGAIV4}.
Put \(S_{n}=S/\fr{m}_{S}^{n}J\), \(\mc{L}'_{n}=\mc{L}'{\otimes}_{S}S_{n}\), \(T_{n}=T{\otimes}_{S}S_{n}\) and so on. We proceed by induction on \(n\) to construct a tower \(\{\mc{N}_{n}\}\) of deformations of \(\bar{\mc{N}}\) inducing \(\{\mc{L}'_{n}\}\). Each step is done as in (ii).
If \(({}^{v}T,{}^{v\!}\mc{N})\in\df{}{(A,N)}(R)\) is a versal element, there is a corresponding tower of maps \(\{f_{n}:R\rightarrow S_{n}\}\) such that \(({}^{v}T,{}^{v\!}\mc{N})\) induces \(\{(T_{n},\mc{N}_{n})\}\). We obtain the algebra map \(f:R\rightarrow {}^{*\!}S:=\varprojlim S_{n}\) which induces a deformation \(({}^{*}T,{}^{*\!}\mc{N})\) of \((\bar{T},\bar{\mc{N}})\) to \({}^{*\!}S\). Since \(\varprojlim {}^{*}T_{n}\cong\varprojlim T_{n}\) completion in the maximal ideals gives an isomorphism \({}^{*}\hat{T}\cong\hat{T}\). By \cite[2.6]{art:69}, \cite[1.3]{pop:86} and \cite{pop:90} there is an isomorphism \({}^{*}T\cong T{\tilde{\otimes}}_{S}{}^{*\!}S\) whereby \({}^{*}T\) is identified with \(T{\tilde{\otimes}}_{S}{}^{*\!}S\). The tower of isomorphisms \(\{\sigma_{L'}({}^{*\!}\mc{N}_{n})\cong\mc{L}'_{n}\}\) implies by Lemma \ref{lem.Aapprox} that there is an isomorphism of deformations \(\sigma_{L'}({}^{*\!}\mc{N})\cong {}^{*\!}\mc{L}'\) above \(\sigma_{L'}(\bar{\mc{N}})\cong \bar{\mc{L}}'\) where \({}^{*\!}\mc{L}'={}^{*}T{\otimes}_{T}\mc{L}'\).
To apply Artin's Approximation Theorem we define a functor of \(S\)-algebras \(F:{}_{S}\cat{H}\rightarrow \Sets\) as follows. If \(\tilde{S}\) is in \({}_{S}\cat{H}\) let \(\tilde{T}\) denote \(T{\tilde{\otimes}}_{S}\tilde{S}\) and let \(\tilde{\mc{L}}'\) denote \(\tilde{T}{\otimes}_{T}\mc{L}'\). Then \(F(\tilde{S})\) is defined as equivalence classes of pairs of maps of finite \(\tilde{T}\)-modules \(\tilde{\xi}=(\tilde{\nu}:\tilde{\mc{N}}\rightarrow \bar{\mc{N}},\tilde{\iota}:\tilde{\mc{N}}\rightarrow\tilde{\mc{L}}')\) such that \(\tilde{\mc{N}}\) is \(\tilde{S}\)-flat. A map \(\tilde{S}\rightarrow\tilde{S}'\) gives a map of pairs by base change. Two pairs, \({}^{1}\tilde{\xi}\) and \({}^{2}\tilde{\xi}\), are equivalent if there is an isomorphism \({}^{1\!}\tilde{\mc{N}}\cong{}^{2\!}\tilde{\mc{N}}\) commuting with the \({}^{j}\tilde{\iota}\) and the \({}^{j}\tilde{\nu}\). We show that \(F\) is locally of finite presentation. Suppose \(\tilde{S}=\varinjlim {}^{i\!}S\) for a filtered injective system of algebras in \({}_{S}\cat{H}\). Put \({}^{i}T=T{\tilde{\otimes}}{}^{i\!}S\). Then \(\varinjlim {}^{i}T\cong \tilde{T}\) by \cite[7.8.3]{EGAIV2} and \cite[18.6.14]{EGAIV4}.
Given \(\tilde{\xi}\in F(\tilde{S})\) as above. Since \(\tilde{\mc{N}}\) has finite presentation and since the maps \(\tilde{\nu}\) and \(\tilde{\iota}\) can be represented on the finite presentations, there is a finite \({}^{i}T\)-module \({}^{i\!}\mc{N}\) and \({}^{i}T\)-linear maps \({}^{i}\nu:{}^{i\!}\mc{N}\rightarrow \bar{\mc{N}}\) and \({}^{i}\iota:{}^{i\!}\mc{N}\rightarrow{}^{i\!}\mc{L}'={}^{i}T{\otimes}_{T}\mc{L}'\) inducing \(\tilde{\xi}\) by base change. We may also assume that \({}^{i\!}\mc{N}\) is \({}^{i\!}S\)-flat. Hence \(\varinjlim F({}^{i\!}S)\rightarrow F(\tilde{S})\) is surjective, and injectivity is similar. Let \({}^{*}\xi\) denote the element in \(F({}^{*\!}S)\) given by \({}^{*\!}\mc{N}\rightarrow \bar{\mc{N}}\) and the \(\hat{\cat{D}}\)-hull \({}^{*\!}\mc{N}\rightarrow{}^{*\!}\mc{L}'\). By Artin's Approximation Theorem \cite[1.12]{art:69}, \cite[1.3]{pop:86}, \cite{pop:90} there exists a \(\xi=(\nu:\mc{N}\rightarrow \bar{\mc{N}},\iota:\mc{N}\rightarrow \mc{L}')\) in \(F(S)\) with \(\xi_{1}={}^{*}\xi_{1}\). In particular \(\mc{N}\rightarrow\bar{\mc{N}}\) is a deformation. Since \(\iota_{0}:\mc{N}_{0}\rightarrow\mc{L}'_{0}\) equals the injective \(\bar{\mc{N}}\rightarrow\bar{\mc{L}}'\) Proposition \ref{prop.nakayama} applied as in Example \ref{ex.nakcplx} implies that \(\iota\) is injective and \(\coker \iota\) is \(S\)-flat. It follows that \(\iota\) is the \(\hat{\cat{D}}\)-hull of \(\mc{N}\), i.e.\ \(\sigma_{L'}(\mc{N})=\mc{L}'\). The \(L\)-case is analogous.
\end{proof}
\begin{thm}\label{thm.defgrade2}
With general assumptions as in \textup{Theorem \ref{thm.defgrade}} consider the map
\begin{equation*}
\sigma_{M}:\df{}{(A,N)}\longrightarrow\df{}{(A,M)}\text{ of functors }{{}_{\varLambda}\cat{H}}/k\longrightarrow\Sets
\end{equation*}
as in \textup{Definition \ref{defn.defmap}.}
\begin{enumerate}
\item[(i)] If\, \(\hm{}{A}{L}{N}=0\) then \(\sigma_{M}\) is injective\textup{.}
\item[(ii)] If\, \(\xt{1}{A}{L}{N}=0\) then \(\sigma_{M}\) is formally smooth\textup{.}
\item[(iii)] If\, \(\xt{1}{A}{L}{N}=0\) and \(\df{}{(A,N)}\) has a versal element then \(\sigma_{M}\) is smooth\textup{.}
\end{enumerate}
The analogous statements hold for \(\sigma_{M'}:\df{}{(A,N)}\rightarrow\df{}{(A,M')}\) with \(\xt{1}{A}{L'}{N}=0\) in \textup{(i)} and \(\xt{2}{A}{L'}{N}=0\) in \textup{(ii)} and \textup{(iii).}
\end{thm}
\begin{proof}
The proof is analogous to the proof of Theorem \ref{thm.defgrade}.
\end{proof}
Fix a field \(k\) and a Cohen-Macaulay local algebraic \(k\)-algebra \(A\). Let \(\varLambda\rightarrow T^{\mspace{-1mu}\mathit{o}}\) be obtained by henselisation of a finite type Cohen-Macaulay map \(\tilde\varLambda\rightarrow \tilde{T}^{\mspace{-1mu}\mathit{o}}\) where \(\tilde\varLambda\) is assumed to be an excellent ring. In particular \(\varLambda\) and \(T^{\mspace{-1mu}\mathit{o}}\) are excellent rings (\cite[7.8.3]{EGAIV2}, \cite[18.7.6]{EGAIV4}). Assume \(T^{\mspace{-1mu}\mathit{o}}/\fr{m}_{T^{\mspace{-1mu}\mathit{o}}}\cong k\) and \(T^{\mspace{-1mu}\mathit{o}}{\otimes}_{\varLambda}k\cong A\). There is a section \({{}_{\varLambda}\cat{H}}\rightarrow\cat{hCM}\) defined by \(S\mapsto T^{\mspace{-1mu}\mathit{o}}{\tilde{\otimes}} S\). Let \(\cat{T^{o}}\) denote the fibred subcategory of \(\cat{hCM}\). We consider deformations in \(\cat{mod}{}^{\textnormal{fl}}_{\vert\cat{T^{o}}}\) of an object \(\xi=(S\rightarrow T^{\mspace{-1mu}\mathit{o}}{\tilde{\otimes}} S,\mc{N})\) and obtain the fibred category of deformations \(\cat{Def}{}^{T^{\mspace{-1mu}\mathit{o}}}_{\mc{N}}:=(\cat{mod}{}^{\textnormal{fl}}_{\vert\cat{T^{o}}})_{\text{coca}}/\xi\) over \({{}_{\varLambda}\cat{H}}/S\). The deformation functor \(\df{T^{\mspace{-1mu}\mathit{o}}}{\mc{N}}:{{}_{\varLambda}\cat{H}}/S\rightarrow\Sets\) is defined by the associated groupoid of sets \(\ol{\cat{Def}}{}^{T^{\mspace{-1mu}\mathit{o}}}_{\mc{N}}\). A special case is given by \(\varLambda=k\) and \(T^{\mspace{-1mu}\mathit{o}}=A\).
\begin{cor}\label{cor.defgrade}
With general assumptions as in \textup{Theorem \ref{thm.defgrade}} consider the map \(\sigma_{L'}:\df{T^{\mspace{-1mu}\mathit{o}}}{N}\rightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{L'}\) of functors \({{}_{\varLambda}\cat{H}}/k\rightarrow\Sets\)\textup{.}
\begin{enumerate}
\item[(i)] If\, \(\hm{}{A}{N}{M'}=0\) then \(\sigma_{L'}\) is injective\textup{.}
\item[(ii)] If\, \(\xt{1}{A}{N}{M'}=0\) then \(\sigma_{L'}\) is formally smooth\textup{.}
\item[(iii)] If\, \(\xt{1}{A}{N}{M'}=0\) and \(\df{A}{N}\) has a versal element then \(\sigma_{L'}\) is smooth\textup{.}
\end{enumerate}
The analogous statements hold for \(\sigma_{L}:\df{T^{\mspace{-1mu}\mathit{o}}}{N}\rightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{L}\) with \(\xt{1}{A}{N}{M}=0\) in \textup{(i)} and \(\xt{2}{A}{N}{M}=0\) in \textup{(ii)} and \textup{(iii).}
\end{cor}
\begin{proof}
This is not a formal consequence of Theorem \ref{thm.defgrade}, but the proof is similar.
\end{proof}
\begin{cor}\label{cor.defgrade2}
With general assumptions as in \textup{Theorem \ref{thm.defgrade}} consider the map \(\sigma_{M}:\df{T^{\mspace{-1mu}\mathit{o}}}{N}\rightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{M}\) of functors \({{}_{\varLambda}\cat{H}}/k\rightarrow\Sets\)\textup{.}
\begin{enumerate}
\item[(i)] If\, \(\hm{}{A}{L}{N}=0\) then \(\sigma_{M}\) is injective\textup{.}
\item[(ii)] If\, \(\xt{1}{A}{L}{N}=0\) then \(\sigma_{M}\) is formally smooth\textup{.}
\item[(iii)] If\, \(\xt{1}{A}{L}{N}=0\) and \(\df{A}{N}\) has a versal element then \(\sigma_{M}\) is smooth\textup{.}
\end{enumerate}
The analogous statements hold for \(\sigma_{M'}:\df{T^{\mspace{-1mu}\mathit{o}}}{N}\rightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{M'}\) with \(\xt{1}{A}{L'}{N}=0\) in \textup{(i)} and \(\xt{2}{A}{L'}{N}=0\) in \textup{(ii)} and \textup{(iii).}
\end{cor}
\begin{prop}\label{prop.defequiv}
With general assumptions as in \textup{Theorem \ref{thm.defgrade}:}
\begin{enumerate}
\item[(i)] If \(Q'=\hm{}{A}{\omega_{A}}{L'}\) and \(Q=\hm{}{A}{\omega_{A}}{L}\) then \(Q'\) and \(Q\) have finite projective dimension and \(\df{}{(A,L')}\cong\df{}{(A,Q')}\) and \(\df{}{(A,L)}\cong\df{}{(A,Q)}\)\textup{.}
\item[(ii)] There are natural maps
\begin{equation*}
s:\df{}{(A,M)}\longrightarrow\df{}{(A,M')}\quad \text{and}\quad t:\df{}{(A,L')}\longrightarrow\df{}{(A,L)}
\end{equation*}
commuting with the maps
\(\sigma_{X}:\df{}{(A,N)}\rightarrow\df{}{(A,X)}\) for \(X\) equal to \(M\) and \(M'\)\textup{,} and to \(L'\) and \(L\)\textup{,} respectively\textup{.} If \(A\) is a Gorenstein ring\textup{,} then \(s\) is an isomorphism\textup{.}
\end{enumerate}
The analogous statements also hold for the deformation functors \(\df{T^{\mspace{-1mu}\mathit{o}}}{X}\)\textup{.}
\end{prop}
\begin{proof}
Lemma \ref{lem.fri} implies (i). There is a short exact sequence \(M\rightarrow \omega_{A}^{\oplus n}\rightarrow M'\) such that the last map is without a common \(\omega_{A}\)-summand, corresponding (through dualisation) to a short exact sequence \(M^{\vee}\leftarrow A^{\oplus n}\leftarrow (M')^{\vee}\) where \(n\) is minimal. The map \(s\) is the composition \(\df{}{(A,M)}\cong\df{}{(A,M^{\vee})}\rightarrow\df{}{(A,(M')^{\vee})}\cong\df{}{(A,M')}\) where the middle map is obtained by taking the syzygy of the deformation.
If \(A\) is a Gorenstein ring then \(\omega_{A}\cong A\) and there is an inverse \(\df{}{(A,M')}\rightarrow\df{}{(A,M)}\) given by the syzygy map.
Note that the pushout of \(M\rightarrow \omega_{A}^{\oplus n}\) with \(M\rightarrow N\) gives \(N\rightarrow L'\). Consider the induced short-exact sequence \(L\rightarrow\omega_{A}^{\oplus n}\xra{\mu} L'\). For a deformation \(\lambda':\mc{L}'\rightarrow L'\) there is a lifting of \(\mu\) to a map \(\tilde{\mu}:\omega_{h}^{\oplus n}\rightarrow \mc{L}'\). If \(\mc{L}\) denotes the kernel of \(\tilde{\mu}\) then there is a cocartesian map \(\lambda:\mc{L}\rightarrow L\) commuting with \(\omega_{h}^{\oplus n}\rightarrow \omega_{A}^{\oplus n}\). By Lemma \ref{lem.defCMapprox2} \(\lambda'\mapsto \lambda\) gives a well defined map of deformation functors \(t:\df{}{(A,L')}\rightarrow\df{}{(A,L)}\).
Given a deformation \((h,\mc{N})\) in \(\df{}{(A,N)}\), let \(\mc{L}\rightarrow\mc{M}\rightarrow\mc{N}\) and \(\mc{N}\rightarrow\mc{L}'\rightarrow\mc{M}'\) be the minimal sequences in Definition \ref{defn.defmap}.
There is a commutative diagram of deformations with (co)cartesian square
\begin{equation}\label{eq.pullpush2}
\xymatrix@C-0pt@R-12pt@H-30pt{
\mc{L} \ar[r]\ar@{=}[d] & \mc{M} \ar[r]\ar[d]\ar@{}[dr]|{\Box} & \mc{N} \ar[d] \\
\mc{L} \ar[r] & \omega_{h}^{\oplus n} \ar[r]\ar[d] & \mc{L}' \ar[d] \\
& \mc{M}' \ar@{=}[r] & \mc{M}'
}
\end{equation}
where \(\omega_{h}^{\oplus n}\rightarrow \mc{L}'\) is given as above. The stated commutativity of maps of deformation functors follows.
\end{proof}
\begin{cor}\label{cor.defapprox}
Let \(A\) be an Cohen-Macaulay local algebraic \(k\)-algebra with residue field \(k\)\textup{.} Suppose \(\dim A\geq 2\)\textup{.} Then there exists finite \(A\)-modules \(L'\) and \(Q'\) with \(\Injdim L'=\dim A=\pdim Q'\) and universal deformations \(\mc{L}'\in\df{A}{L'}(A)\) and \(\mc{Q}'\in\df{A}{Q'}(A)\)\textup{.}
\end{cor}
\begin{proof}
Let \(h=1{\otimes}\id:A\rightarrow A{\tilde{\otimes}}_{k}A=T\) and \(\mc{N}=A\) be the cyclic \(T\)-module defined through the multiplication map. Then \(T{\otimes}_{A}k\cong A\) and \(\mc{N}{\otimes}_{A}k\cong k\) and this gives a deformation \(\mc{N}\rightarrow k\) of the residue field of \(A\) which is universal.
If \(L'\) is the minimal \(\hat{\cat{D}}_{A}\)-hull of the residue field \(k\) then \(\mc{L}'=\sigma_{L'}(\mc{N})\in \df{T^{\mspace{-1mu}\mathit{o}}}{L'}(A)\) is universal by Corollary \ref{cor.defgrade}. If \(Q'=\hm{}{A}{\omega_{A}}{L'}\) then \(\hm{}{T}{\omega_{T}}{\mc{L}'}\in\df{A}{Q'}(A)\) is universal by Proposition \ref{prop.defequiv}.
\end{proof}
\begin{cor}\label{cor.depth}
With general assumptions as in \textup{Theorem \ref{thm.defgrade}} put \(X=\Spec A\)\textup{.} Let \(Z\) be a closed subscheme such that the complement \(U\) is contained in the regular locus\textup{.} Assume \(\tilde{N}_{\vert U}\) is locally free\textup{,} \(\depth_{Z}N\geq 2\) and \(\cH^{2}_{Z}(\hm{}{A}{L}{N})=0\)\textup{.} Then
\begin{equation*}
\sigma_{M}:\df{}{(A,N)}\longrightarrow\df{}{(A,M)}\quad\text{and}\quad \sigma_{M}^{T^{\mspace{-1mu}\mathit{o}}}:\df{T^{\mspace{-1mu}\mathit{o}}}{N}\longrightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{M}\quad\text{are formally smooth\textup{.}}
\end{equation*}
\end{cor}
\begin{proof}
We show that \(\xt{1}{A}{L}{N}=0\) and apply Theorem \ref{thm.defgrade2} and Corollary \ref{cor.defgrade2}.
By Th{\'e}or{\`e}me 1.6 in \cite[Expos{\'e} VI]{SGA2} there is a cohomological spectral sequence
\begin{equation}
\cE^{p,q}_{2}=\xt{q}{A}{L}{\cH^{p}_{Z}(N)}\,\Rightarrow\,\xt{p+q}{Z}{X;L}{N}\,.
\end{equation}
Since \(\cH^{i}_{Z}(N)=0\) for \(i=0,1\) the restriction map \(\xt{1}{A}{L}{N}\rightarrow\xt{1}{U}{X;L}{N}\) in the long exact sequence is injective. Since \(U\) is contained in the regular locus, \(\tilde{M}_{\vert U}\) and hence \(\tilde{L}_{\vert U}\) are locally free. It follows that \(\xt{1}{U}{X;L}{N}\) is isomorphic to
\begin{equation}
\xt{1}{\mathcal{O}_{U}}{\tilde{L}_{\vert U}}{\tilde{N}_{\vert U}}\cong\cH^{1}(U,\shm{}{\mathcal{O}_{X}}{\tilde{L}}{\tilde{N}}) \cong \cH^{2}_{Z}(\hm{}{A}{L}{N})
\end{equation}
which is zero by assumption.
\end{proof}
\begin{ex}\label{ex.depth}
The condition \(\cH^{2}_{Z}(\hm{}{A}{L}{N})=0\) is implied by \(\tilde{N}_{\vert U}=0\) or
\(\depth_{Z}(\hm{}{A}{L}{N})\geq 3\).
\end{ex}
The following result extends A.\ Ishii's \cite[3.2]{ish:00} to deformations of the pair.
\begin{prop}\label{prop.Gor}
Let \(k\) be a field and let \(A\) be a Gorenstein local algebraic \(k\)-algebra. Suppose \(L\rightarrow M\rightarrow N\) is the minimal Cohen-Macaulay approximation of a finite \(A\)-module \(N\)\textup{.} If \(\depth N=\dim A -1\) then
\begin{equation*}
\sigma_{M}:\df{}{(A,N)}\longrightarrow\df{}{(A,M)}\quad\text{and}\quad \sigma_{M}^{T^{\mspace{-1mu}\mathit{o}}}:\df{T^{\mspace{-1mu}\mathit{o}}}{N}\longrightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{M}\quad\text{are smooth\textup{.}}
\end{equation*}
\end{prop}
\begin{proof}
By assumption \(L\cong A^{\oplus r}\). Assume \((h_{1},\mc{M}_{1})\) in \(\df{}{(A,M)}(S_{1})\) maps to \((h,\mc{M})\) along the surjection \(S_{1}\rightarrow S\). Assume \(\sigma_{M}\) maps \((h',\mc{N})\) in \(\df{}{(A,N)}(S)\) to \((h,\mc{M})\). We can assume that \(h'=h\) and that the minimal \(\cat{MCM}_{h}\)-approximation of \(\mc{N}\) is \(\mc{L}\xra{\rho}\mc{M}\rightarrow\mc{N}\) where \(\mc{L}\cong T^{\oplus r}\). Let \(\mc{L}_{1}=T_{1}^{\oplus r}\) and choose a lifting \(\rho_{1}:\mc{L}_{1}\rightarrow\mc{M}_{1}\) of \(\rho\). Put \(\mc{N}_{1}:=\coker\rho_{1}\) with its natural map to \(\mc{N}\). Then \(\mc{N}_{1}\) is \(S_{1}\)-flat (\(\rho_{1}{\otimes} S=\rho\)) and \(\sigma_{M}(h_{1},\mc{N}_{1})=(h_{1},\mc{M}_{1})\).
\end{proof}
\begin{rem}
If \(\dim A\geq 1\) and an MCM \(A\)-module \(M\) has a rank, \(r=\Rk M\), then there is a short exact sequence \(A^{r}\rightarrow M\rightarrow N\) with \(N\) a codimension one Cohen-Macaulay module, cf.\ \cite[1.4.3]{bru/her:98}. Hence in the case \(A\) is a Gorenstein domain all MCM modules admits \(\cat{MCM}_{A}\)-approximations by CM modules in codimension one and Proposition \ref{prop.Gor} applies. However, it's not always possible to continue this reduction: If \(A\) is a normal Gorenstein complete local ring any MCM \(A\)-module \(M\) is the \(\cat{MCM}_{A}\)-approximation of a codimension \(2\) Cohen-Macaulay module up to stable isomorphism if and only if \(A\) is a unique factorisation domain, see \cite{yos/iso:00, kat:07}.
Let \(A\) be a Gorenstein normal domain of dimension \(2\) and \(A^{r{-}1}\rightarrow M\rightarrow I\) the minimal MCM approximation of a torsion-free rank \(1\) module \(I\). Let \(U\) denote the regular locus in \(X=\Spec A\). If \(T=A{\tilde{\otimes}}_{k}S\) for \(S\) in \({{}_{k}\cat{H}}/k\) there is a natural section \(A\rightarrow T\). Let \(U_{T}\) denote \(U\times_{X}\Spec T\). Consider the subfunctor \(\df{A,\wedge}{M}\subseteq\df{A}{M}\) of deformations \(\mc{M}\) with trivial induced deformation \(\wedge^{r}\mc{M}_{\vert U_{T}}\). Note that \(\cH^{0}(U,\wedge^{r}M)\) is isomorphic to the MCM module \(\bar{I}:=\cH^{0}(U,I)\). It follows from Propostion \ref{prop.Gor} that the resulting map from the quotient functor \(\quot{\bar{I}}{I\subseteq \bar{I}}\rightarrow \df{A,\wedge}{M}\) is also smooth, cf.\ \cite[3.2]{ish:00}. In particular, if \(E_{A}\) is the fundamental module and \(A/\fr{m}_{A}\cong k\) then \(\hm{}{{{}_{k}\cat{H}}/k}{A}{-}\cong\quot{A}{\fr{m}_{A}\subseteq A}\cong\df{A}{A/\fr{m}_{A}}\) gives a versal family for \(\df{A,\wedge}{E_{A}}\) by the MCM approximation in Corollary \ref{cor.Fmod}, see \cite[3.4]{ish:00}.
\end{rem}
\begin{ex}
Assume \(A/\fr{m}_{A}\cong k\) and let \(M\) denote the minimal MCM approximation of \(k\). It's given as \(M\cong \hm{}{A}{\syz{A}{d}(k^{\vee})}{\omega_{A}}\) where \(d=\dim A\), cf.\ Remark \ref{rem.Buch}. One has \(k^{\vee}=\xt{d}{A}{k}{\omega_{A}}\cong k\). We apply \(\hm{}{A}{-}{\omega_{A}}\) to the short exact sequence \(\syz{A}{}(\fr{m}_{A})\rightarrow A^{\oplus\beta_{1}}\xra{(\ul{x})}\fr{m}_{A}\). Assume \(\dim A = 2\). Since \(\xt{1}{A}{\fr{m}_{A}}{\omega_{A}}\cong k\) we obtain the MCM approximation of \(k\):
\begin{equation}
0\rightarrow\omega_{A}\xra{(\ul{x})^{\text{t}}}\omega_{A}^{\oplus\beta_{1}}\longrightarrow M\longrightarrow k\rightarrow 0
\end{equation}
In particular \(\Rk(M)=\beta_{1}-1\) and \(\mu(M)=t(A)\cdot\beta_{1}+1\) where \(t(A)\) is the Cohen-Macaulay type of \(A\). In the case \(A=A(m)=k[u^{m},u^{m-1}v,\dots,v^{m}]^{\text{h}}\), the vertex of the cone over the rational normal curve of degree \(m\), which has the indecomposable MCM modules \(M_{i}=(u^{i},u^{i-1}v,\dots,v^{i})\) for \(i=0,\dots,m{-}1\), one finds that \(M=M_{m-1}^{\oplus m}\). We have
\begin{equation}
\dim_{k}\df{A}{M}(k[\varepsilon])=\dim_{k}\xt{1}{A}{M}{M}= (m-1)\cdot m^{2}
\end{equation}
while \(\dim_{k}\df{A}{k}(k[\varepsilon])=m+1\). Even in the Gorenstein case (\(m=2\)) the tangent map isn't surjective and so Proposition \ref{prop.Gor} cannot in general be extended to \(\depth N=\dim A -2\). For a detailed description of the strata defined by Ishii in \cite{ish:00} of the reduced versal deformation space of \(M\), see \cite{gus/ile:04b}. Applying \(\hm{}{A}{k}{-}\) to \(\fr{m}_{A}\rightarrow A\rightarrow k\) gives an exact sequence
\begin{equation}
0\rightarrow\xt{1}{A}{k}{k}\longrightarrow\df{A}{\fr{m}_{A}}(k[\varepsilon])\longrightarrow k^{t(A)}\longrightarrow \xt{2}{A}{k}{k}
\end{equation}
since \(\xt{1}{A}{\fr{m}_{A}}{\fr{m}_{A}}\cong\xt{2}{A}{k}{\fr{m}_{A}}\) and \(\dim A = 2\). In the case \(A=A(m)\) the fundamental module \(E_{A}\) \eqref{eq.Fmod} is isomorphic to \(M_{m-1}^{\oplus 2}\) with \(\dim_{k}\df{A}{E_{A}}(k[\varepsilon])=4(m-1)\). The conclusion in Proposition \ref{prop.Gor} cannot hold in the non-Gorenstein case \(m>2\).
\end{ex}
\section{Existence of versal elements}\label{sec.versal}
We prove existence of a versal element for a pair (algebra, module) with isolated singularity. The following lemma is used in the proof.
\begin{lem}\label{lem.AQhens}
Let \(h^{\textnormal{ft}}:S\rightarrow T^{\textnormal{ft}}\) be a finite type homomorphism of noetherian rings\textup{.} Let \(\varGamma^{\textnormal{ft}}=T^{\textnormal{ft}}{\oplus} N\) be the graded \(S\)-algebra with a finite \(T^{\textnormal{ft}}\)-module \(N\) in degree \(1\) and let \(I={}^{0\!}I{\oplus}{}^{1\!}I\) be a graded \(\varGamma^{\textnormal{ft}}\)-module with \({}^{i\!}I\) in degree \(i\)\textup{.} Let \(T\) denote the henselisation of \(T^{\textnormal{ft}}\) in a maximal ideal \(\fr{m}\) and put \(\varGamma=T{\otimes}\varGamma^{\textnormal{ft}}\)\textup{.}
\begin{enumerate}
\item[(a)] There are natural isomorphisms of graded Andr{\'e}-Quillen cohomology
\begin{equation*}
\gH{0}^{i}(S,\varGamma,T{\otimes} I)\cong\gH{0}^{i}(S, \varGamma^{\textnormal{ft}},I){\otimes}_{T^{\textnormal{ft}}}T\,\,\text{ for all } i\,.
\end{equation*}
\end{enumerate}
Suppose in addition that \(h^{\textnormal{ft}}\) is flat\textup{,} \(I\) is finite as \(T^{\textnormal{ft}}\)-module\textup{,} \(S\) is local henselian and \(S/\fr{m}_{S}\cong T^{\textnormal{ft}}/\fr{m}\cong k\)\textup{.}
Let \(k\rightarrow A^{\textnormal{ft}}\) denote the central fibre of \(h^{\textnormal{ft}}\)\textup{.} Put \(\fr{m}_{0}=\fr{m}A^{\textnormal{ft}}\) and \(N_{0}=N{\otimes}_{S}k\)\textup{.} Assume \(V=\Spec A^{\textnormal{ft}} \setminus\{\fr{m}_{0}\}\) is smooth over \(k\) and \(N_{0}\) restricted to \(V\) is locally free\textup{.}
\begin{enumerate}
\item[(b)] For all \(i>0\) the graded Andr{\'e}-Quillen cohomology \(\gH{0}^{i}(S,\varGamma,T{\otimes} I)\) is finite as \(S\)-module and there is a natural \(T^{\textnormal{ft}}_{\fr{m}}\)-isomorphism
\begin{equation*}
\gH{0}^{i}(S,\varGamma,T{\otimes} I)\cong\gH{0}^{i}(S, \varGamma^{\textnormal{ft}},I)_{\fr{m}}\,.
\end{equation*}
\end{enumerate}
\end{lem}
\begin{proof}
(a) Note that there are natural maps \(\gH{0}^{i}(S,\varGamma,T{\otimes} I)\rightarrow \gH{0}^{i}(S,\varGamma^{\textnormal{ft}},T{\otimes} I)\) and \(\gH{0}^{i}(S, \varGamma^{\textnormal{ft}},I){\otimes}_{T^{\textnormal{ft}}}T\rightarrow\gH{0}^{i}(S,\varGamma^{\textnormal{ft}},T{\otimes} I)\). We consider the natural long-exact sequence in Proposition \ref{prop.lang} and prove that the analogous maps for the ungraded Andr{\'e}-Quillen and Ext cohomology are isomorphisms.
The cotangent complex is trivial for Zariski localisation and by the transitivity sequence \(\cH_{i}(T^{\textnormal{ft}},T,-)\rightarrow \cH_{i}(T^{\textnormal{ft}}_{\fr{m}},T,-)\) is an isomorphism for all \(i\). Andr{\'e}-Quillen homology commutes with direct limits in the second argument \cite[III 35]{and:74} and the cotangent complex is trivial for {\'e}tale extensions \cite[III 3.1.1]{ill:71}. Hence \(\cH_{i}(T^{\textnormal{ft}},T,T)=0\) for all \(i\). By \cite[III 21]{and:74}
\begin{equation}
\cH^{i}(T^{\textnormal{ft}},T,T{\otimes} {}^{0\!}I)\cong \hm{}{T}{\cH_{i}(T^{\textnormal{ft}},T,T)}{T{\otimes} {}^{0\!}I}=0 \text{ for all }i\,.
\end{equation}
From the transitivity sequence \(\cH^{i}(S,T,T{\otimes} {}^{0\!}I)\cong \cH^{i}(S,T^{\textnormal{ft}},T{\otimes} {}^{0\!}I)\) for all \(i\). Moreover \(T^{\textnormal{ft}}\rightarrow T\) is flat and \(T^{\text{ft}}\) is of finite type over the noetherian \(S\) and we obtain
\begin{equation}
\cH^{i}(S,T^{\textnormal{ft}},{}^{0\!}I){\otimes}_{T^{\textnormal{ft}}} T\cong \cH^{i}(S,T^{\textnormal{ft}},T{\otimes} {}^{0\!}I)\cong\cH^{i}(S,T,T{\otimes} {}^{0\!}I)
\end{equation}
for all \(i\) \cite[IV 58]{and:74}. Since \(T\) is \(T^{\textnormal{ft}}\)-flat and \(N\) finite
\begin{equation}
\xt{i}{T^{\textnormal{ft}}}{N}{{}^{1\!}I}{\otimes}_{T^{\textnormal{ft}}}T\cong
\xt{i}{T^{\textnormal{ft}}}{N}{T{\otimes} {}^{1\!}I}\cong
\xt{i}{T}{T{\otimes} N}{T{\otimes}{}^{1\!}I}\,.
\end{equation}
(b) The non-smooth locus of \(h^{\text{ft}}\) is closed, i.e.\ defined by an ideal \(J\subseteq T^{\text{ft}}\). Smooth is equivalent to flat with smooth fibres \cite[17.5.1]{EGAIV4}. Hence \(J_{0}=JA^{\textnormal{ft}}\) defines the non-smooth locus of \(k\rightarrow A^{\text{ft}}\) and \(J_{0}\) is \(\fr{m}_{0}\)-primary. Put \(\bar{T}^{\text{ft}}=T^{\text{ft}}/J\). Since \(A^{\text{ft}}/J_{0}\) has finite length, \(S\rightarrow \bar{T}^{\text{ft}}\) is quasi-finite at \(\fr{m}\) \cite[Err 20]{EGAIII2} by Chevalley's upper semi-continuity theorem \cite[13.1.3]{EGAIV3} and openness of \(\Spec T^{\text{ft}}\rightarrow \Spec S\) \cite[2.4.6]{EGAIV2}. Since \(S\) is henselian it follows that there is a ring \(T'\) such that \(\bar{T}^{\text{ft}}\) is isomorphic to \(\bar{T}^{\text{ft}}_{\fr{m}}\prod T'\) where \(\bar{T}^{\text{ft}}_{\fr{m}}\) is finite as \(S\)-module \cite[18.5.11]{EGAIV4}. Hence there is a Zariski neighborhood \(U\) of \(\fr{m}\) in \(\Spec T^{\text{ft}}\) such that non-smooth locus \(U\cap V(J)\) is finite over \(S\) and the support of \(\cH^{i}=\cH^{i}(S,T^{\text{ft}},{}^{1\!}I)\) in \(U\) is contained in \(U\cap V(J)\) for all \(i>0\) by \cite[III 3.1.2]{ill:71}. Since \(\cH^{i}_{\fr{m}}\) equals \(\cH^{i}\) restricted to \(U\), it follows that \(\cH^{i}_{\fr{m}}\) is finite as \(S\)-module. With the isomorphism in (a) we get
\begin{equation}
\cH^{i}(S,T,T{\otimes} {}^{0\!}I)\cong\cH^{i}(S,T^{\text{ft}},{}^{0\!}I)_{\fr{m}}{\otimes}_{T^{\text{ft}}_{\fr{m}}}T\cong \cH^{i}(S,T^{\text{ft}},{}^{0\!}I)_{\fr{m}} \,.
\end{equation}
The locus where \(N\) isn't locally free is closed, i.e.\ defined by an ideal \(J'\subseteq T^{\text{ft}}\). Locally free is equivalent to flat and locally free fibres. Hence \(J'_{0}=J'A^{\textnormal{ft}}\) defines the singular locus of \(N_{0}\) and \(J'_{0}\) is \(\fr{m}_{0}\)-primary. As for the Andr{\'e}-Quillen cohomology we get a Zariski neighborhoud \(U'\) of \(\fr{m}\) such that \(\cE^{i}=\xt{i}{T^{\textnormal{ft}}}{N}{{}^{1\!}I}\) restricted to \(U'\) equals \(\cE^{i}_{\fr{m}}\) and is finite as \(S\)-module for all \(i>0\). With (a) we get
\begin{equation}
\xt{i}{T}{T{\otimes} N}{T{\otimes} {}^{1\!}I}\cong \xt{i}{T^{\textnormal{ft}}}{N}{{}^{1\!}I}{\otimes}_{T^{\textnormal{ft}}}T\cong \xt{i}{T^{\textnormal{ft}}}{N}{{}^{1\!}I}_{\fr{m}} \,.
\end{equation}
We conclude by the long-exact sequence in Proposition \ref{prop.lang}.
\end{proof}
Let \(k\) be a field, \(A\) an algebraic \(k\)-algebra with \(A/\fr{m}_{A}\cong k\) and \(N\) a finite \(A\)-module. Without any Cohen-Macaulay condition on \(A\) we define a deformation \((h:S\rightarrow T,\mc{N})\) of the pair \((A,N)\) to an \(S\) in \({{}_{\varLambda}\cat{H}}/k\) as before and obtain the deformation functor \(\df{}{(A,N)}:{{}_{\varLambda}\cat{H}}/k\rightarrow\Sets\) as equivalence classes of deformations of pairs.
We say that \(A\) is an \emph{isolated singularity over \(k\)} if there is a finite type \(k\)-algebra \(A^{\text{ft}}\) with a maximal ideal \(\fr{m}_{0}\) such that the henselisation \((A^{\text{ft}})^{\text{h}}_{\fr{m}_{0}}\) is isomorphic to \(A\) and which is smooth over \(k\) at all points in \(\Spec A^{\text{ft}} \setminus\{\fr{m}_{0}\}\). We say that the pair \((A,N)\) is an \emph{isolated singularity over \(k\)} if \(A\) is an isolated singularity over \(k\) and if \(N_{\fr{p}}\) is a free \(A_{\fr{p}}\)-module for all prime ideals \(\fr{p}\neq\fr{m}_{A}\).
The following theorem is a consequence of results of R.\ Elkik and an argument of H.\ von Essen.
\begin{thm}\label{thm.ExVers}
Let \((A,N)\) be an isolated singularity over the field \(k\) with \(A\) equidimensional\textup{.} Then \(\df{}{(A,N)}:{{}_{\varLambda}\cat{H}}/k\rightarrow\Sets\) has a versal element\textup{.}
\end{thm}
\begin{proof}
We apply \cite[3.2]{art:74} with the extension to arbitrary excellent coefficients given by \cite[1.5]{con/jon:02} to show the existence of a formally versal element for \(\df{}{(A,N)}\). By the finiteness conditions it follows that \(\df{}{(A,N)}\) is locally of finite presentation. The condition (S1) holds in general by Proposition \ref{prop.grobs}. Let \((h:S\rightarrow T,\mc{N})\) be a deformation to \(S\).
Let \((T^{\text{ft}}, \fr{m})\) be an \(S\)-flat finite type representative for \(T\) such that \(A^{\text{ft}}=T^{\text{ft}}{\otimes}_{S}k\) has a single non-smooth closed point \(\fr{m}_{0}=\fr{m}A^{\text{ft}}\) and let \(\mc{N}^{\text{ft}}\) be an \(S\)-flat finite \(T^{\text{ft}}\)-module representing \(\mc{N}\). The singular locus of \(N^{\text{ft}}=\mc{N}^{\text{ft}}{\otimes}_{S}k\) is closed and equal to \(V(J_{0})\) for some ideal \(J_{0}\subseteq A^{\text{ft}}\). But \(\fr{m}_{0}\) is isolated in \(V(J_{0})\) so (possibly after inverting some element in \(T^{\text{ft}}\)) we may assume that \(N^{\text{ft}}\) is locally free away from \(\fr{m}_{0}\).
Let \(I\) be a finite \(S\)-module. By Lemma \ref{lem.AQhens} \(\gH{0}^{1}(S,\varGamma,\varGamma{\otimes}_{S} I)\) with \(\varGamma=T{\oplus}\mc{N}\) is a finite \(S\)-module, so by Proposition \ref{prop.grobs} condition (S2) holds.
For effectivity, there is a deformation functor \(\df{}{(A^{\text{ft}},N^{\text{ft}})}:{{}_{\varLambda}\cat{H}}/k\rightarrow\Sets\) of base change maps of pairs \((S\rightarrow T^{\text{ft}},\mc{N}^{\text{ft}})\rightarrow (k\rightarrow A^{\text{ft}},N^{\text{ft}})\) where \(T^{\text{ft}}\) is a flat \(S\)-algebra of finite type and \(\mc{N}^{\text{ft}}\) is an \(S\)-flat finite \(T^{\text{ft}}\)-module. Base change is given by the standard tensor product. Similarly there is a \(\df{}{A^{\text{ft}}}\). Restricted to \({{}_{\varLambda}\cat{A}}/k\) \(\df{}{(A^{\text{ft}},N^{\text{ft}})}\) satisfies (S1) and (S2). Hence there exists a formally versal formal element \(\{(T_{n}^{\text{ft}},\mc{N}^{\text{ft}}_{n})\}\) in \(\varprojlim\df{}{(A^{\text{ft}},N^{\text{ft}})}(S_{n})\) where \(S_{n}=S/\fr{m}_{S}^{n+1}\) for some \(S=\hat{S}\) in \({{}_{\varLambda}\cat{H}}/k\). By \cite[Th\'{e}or\`{e}m 7, p.\ 595]{elk:73} (cf.\ \cite[II 5.1]{art:76}) there exists an element \(S\rightarrow T^{\text{ft}}\) in \(\df{}{A^{\text{ft}}}(S)\) which induces \(\{T_{n}^{\text{ft}}\}\). Let \(T'\) be the henselisation of \(T^{\text{ft}}\) in the maximal ideal \(\fr{m}=(T^{\text{ft}}{\rightarrow} A)^{{-}1}(\fr{m}_{A})\). Then \(S\rightarrow T'\) is a deformation of \(A\). Let \(T^{*}\) be the completion of \(T'\) at the ideal \(\fr{n}=\fr{m}_{S}T'\) and let \(\mc{N}^{*}=\varprojlim \mc{N}_{n}\). Then \(\mc{N}^{*}\) is an \(S\)-flat finite \(T^{*}\)-module. Let \(J^{*}\subseteq T^{*}\) denote the ideal \(I(\phi)\) where \(\phi\) is a minimal presentation of \(\mc{N}^{*}\). Then \(J^{*}\) defines the locus \(V(J^{*})\) where \(\mc{N}^{*}\) is not locally free. Let \(J=\ker(T'\rightarrow T^{*}{/}J^{*})\). Since \(T^{*}/J^{*}\) is finite as \(S\)-module, \(T'/J\cong T^{*}/J^{*}\). The proof of \cite[2.3]{ess:90} works in this situation too (there is a typo in line 5: it should be a direct sum, not a tensor product) and shows that the completion of \(T'\) in the ideal \(\fr{a}=J\cap\fr{m}_{S}T'\) equals \(T^{*}\). Since \(\mc{N}^{*}\) is locally free on the complement of \(V(\fr{a}T^{*})\), the conditions in \cite[Th\'{e}or\`{e}m 3]{elk:73} hold. From this result we obtain a finite \(T'\)-module \(\mc{N}\) inducing \(\mc{N}^{*}\). In particular \(\mc{N}\) is \(S\)-flat.
We claim that the henselisation map \(\df{}{(A^{\text{ft}},N^{\text{ft}})}\rightarrow\df{}{(A,N)}\) is formally smooth. It follows that the element \((T',\mc{N})\) in \(\df{}{(A,N)}(S)\) is formally versal. For the claim, put \(\varGamma^{\text{ft}}=A^{\text{ft}}{\oplus}N^{\text{ft}}\) and \(\varGamma=A{\oplus} N\) and let \(\pi:S_{1}\rightarrow S_{0}=S_{1}/I\) be a small surjection in \({{}_{\varLambda}\cat{A}}/k\).
The obstruction \(\ob(\pi,\varGamma^{\text{ft}}_{0})\in\gH{0}^{2}(k,\varGamma^{\text{ft}},\varGamma^{\text{ft}}){\otimes}_{k}I\) for lifting a deformation \(\varGamma^{\text{ft}}_{0}\) of \(\varGamma^{\text{ft}}\) along \(\pi\) maps to the corresponding obstruction \(\ob(\pi,\varGamma_{0})\in\gH{0}^{2}(k,\varGamma,\varGamma){\otimes}_{k}I\).
The isomorphisms \(\cH^{i}(S, T,T)\cong \cH^{i}(S, T^{\text{ft}},T^{\text{ft}}){\otimes}_{T^{\text{ft}}}T\) for all \(i\) implies isomorphisms \(\gH{0}^{i}(k,\varGamma^{\text{ft}},\varGamma^{\text{ft}})\cong \gH{0}^{i}(k,\varGamma,\varGamma)\) for \(i=1,2\) as in the beginning of the proof. Smoothness follows by the standard obstruction argument. By \cite[3.2]{art:74} and \cite{con/jon:02} there is an algebraic \(k\)-algebra \(R\) and a formally versal element \((T,\mc{N})\) in \(\df{}{(A,N)}(R)\).
Finally we apply \cite[3.3]{art:74} (for general excellent coefficients) to conclude that the formally versal element \((T,\mc{N})\) is versal. We already have (S1) and (S2). To check \cite[3.3(iii)]{art:74}, let \(S\) be algebraic in \({{}_{\varLambda}\cat{H}}/k\), \(I\) an ideal in \(S\) and put \(S^{*}=\varprojlim S_{n}\) where \(S_{n}=S/I^{n{+}1}\). Let \({}^{i}\xi=({}^{i}T,{}^{i}\mc{N})\) for \(i=1,2\) be two elements in \(\df{}{(A,N)}(S^{*})\) and \(\{\theta_{n}:{}^{1}\xi_{n}\cong{}^{2}\xi_{n}\}\) be a tower of isomorphisms between the \(S_{n}\)-truncations. There are finite type representatives \({}^{i}\xi^{\text{ft}}=({}^{i}T^{\text{ft}},{}^{i}\mc{N}^{\text{ft}})\) of the \({}^{i}\xi\). By the cohomology argument above one obtains by induction a tower of isomorphisms \(\{\theta_{n}^{\text{ft}}:{}^{1}\xi_{n}^{\text{ft}}\cong{}^{2}\xi_{n}^{\text{ft}}\}\) inducing \(\{\theta_{n}\}\). Since \(\varprojlim \hat{S}/I^{n{+}1}\hat{S}\cong S^{*}\) where \(\hat{S}\) is the completion of \(S\) in the maximal ideal, we can apply \cite[Lemme p.\ 600]{elk:73} to conclude that the henselisations of the \({}^{i}T^{\text{ft}}\) in \({}^{i}T^{\text{ft}}I\) are isomorphic by an isomorphism lifting \(\theta_{0}:{}^{1}T_{0}\cong{}^{2}T_{0}\).
Further henselisation in the maximal ideals gives an isomorphism of deformations \({}^{1}T\cong{}^{2}T\). By Lemma \ref{lem.Aapprox} the isomorphism is extended to an isomorphism of the pairs \(\psi:{}^{1}\xi\cong{}^{2}\xi\) which lifts \(\theta_{0}\). By \cite[1.3]{ess:90} condition \cite[3.3(ii)]{art:74} is unnecessary and we conclude that \((T,\mc{N})\) is versal.
\end{proof}
\begin{rem}\label{rem.isomod}
Let \(A\) be an Cohen-Macaulay local algebraic \(k\)-algebra and \(N\) a finite \(A\)-module. We say that \(N\) has an \emph{isolated singularity} if \(N_{\fr{p}}\) is a free \(A_{\fr{p}}\)-module for all prime ideals \(\fr{p}\neq\fr{m}_{A}\). In that case a similar, but easier argument gives that \(\df{T^{\mspace{-1mu}\mathit{o}}}{N}\) has a versal element. This is the result \cite[2.4]{ess:90} of von Essen, but for a slightly different fibred category of deformations where henselisation is taken along the closed fibre. However it implies the result in our case, essentially by henselisation at \(\fr{m}_{0}\). Corollary \ref{cor.ExVers} and \ref{cor.2dim} have obvious analogs for \(\df{T^{\mspace{-1mu}\mathit{o}}}{N}\) in this case.
\end{rem}
\begin{cor}\label{cor.ExVers}
Suppose \(A\) is an isolated Cohen-Macaulay singularity over the field \(k\) and \(N\) is a finite length \(A\)-module\textup{.} Let \(L\rightarrow M\rightarrow N\) and \(N\rightarrow L'\rightarrow M'\) be the minimal \(\cat{MCM}_{A}\)-approximation and \(\hat{\cat{D}}_{A}\)-hull of \(N\) respectively\textup{.}
Then\textup{:}
\begin{enumerate}
\item[(i)] \(\df{}{(A,N)}\) has a versal element\textup{.}
\item[(ii)] If \(\dim A\geq 2\) and \(Q'\) denotes \(\hm{}{A}{\omega_{A}}{L'}\) then
\begin{equation*}
\df{}{(A,N)}\cong\df{}{(A,L')}\cong\df{}{(A,Q')}\textup{.}
\end{equation*}
\item[(iii)] If \(\dim A\geq 3\) and \(Q\) denotes \(\hm{}{A}{\omega_{A}}{L}\) then
\begin{equation*}
\df{}{(A,N)}\cong\df{}{(A,L)}\cong\df{}{(A,Q)}\textup{.}
\end{equation*}
\end{enumerate}
\end{cor}
\begin{proof}
This is Theorem \ref{thm.ExVers}, Theorem \ref{thm.defgrade} and Proposition \ref{prop.defequiv}.
\end{proof}
\begin{cor}\label{cor.2dim}
Suppose \(A\) is a local algebraic \(k\)-algebra which is a Gorenstein normal domain with \(\dim A=2\) and \(N\) is a finite torsion-free \(A\)-module\textup{.} Let \(L\rightarrow M\rightarrow N\) be the minimal \(\cat{MCM}_{A}\)-approximation of \(N\)\textup{.} Assume \(k\) is perfect\textup{.} Then \(\df{}{(A,N)}\) and \(\df{}{(A,M)}\) both have versal elements and the map \(\sigma_{M}:\df{}{(A,N)}\rightarrow\df{}{(A,M)}\) is smooth\textup{.}
\end{cor}
\begin{proof}
Since \(A\) is a domain \(N\) torsion-free implies that \(N\) is a first syzygy. It follows that \(N_{\fr{p}}\) is a MCM \(A_{\fr{p}}\)-module for all primes \(\fr{p}\neq\fr{m}_{A}\) and since \(A\) is \(2\)-dimensional and normal \(N_{\fr{p}}\) is free. As \(k\) is perfect it follows from \cite[6.7.7 and 6.8.6]{EGAIV2} that \((A,N)\) and \((A,M)\) are isolated singularities and hence Theorem \ref{thm.ExVers} applies. Since \(\depth N\geq 1\), \(L\) is projective and by Theorem \ref{thm.defgrade2} (iii) \(\sigma_{M}\) is smooth.
\end{proof}
\section{Deforming maximal Cohen-Macaulay approximations \\of Cohen-Macaulay modules}\label{sec.defCM}
Let \(h:S\rightarrow T\) be a homomorphism of local noetherian rings. An \emph{\(h\)-sequence} (or just an \(h\)-regular element if \(n=1\)) is a sequence \(J=(f_{1},\dots,f_{n})\) in \(T\) such that the image \(\bar{J}\) in \(A=T{\otimes}_{S}S/\fr{m}_{S}\) is an \(A\)-sequence. Applying the Koszul complex \(K(J)\) as in Example \ref{ex.nakcplx} one sees that an \(h\)-sequence is a transversally \(T\)-regular sequence relative to \(S\) as defined in \cite[19.2.1]{EGAIV4}. In particular; \(J\) is an \(h\)-sequence if and only if \(J\) is a \(T\)-sequence and \(T/J\) is \(S\)-flat.
\begin{thm}\label{thm.defMCM}
Let \(q:\varLambda\rightarrow T^{\mspace{-1mu}\mathit{o}}\) denote the henselisation of a finite type Cohen-Macaulay map with \(T^{\mspace{-1mu}\mathit{o}}/\fr{m}_{T^{\mspace{-1mu}\mathit{o}}}=k\) and \(T^{\mspace{-1mu}\mathit{o}}{\otimes}_{\varLambda}k=A\)\textup{.} Suppose \(J=(f_{1},\dots,f_{n})\) is a \(q\)-sequence\textup{.} Put \(\bar{T}^{\mspace{-1mu}\mathit{o}}=T^{\mspace{-1mu}\mathit{o}}/J\)\textup{,} \(B=\bar{T}^{\mspace{-1mu}\mathit{o}}{\otimes}_{\varLambda}k\) and let \(\bar{J}\) be the image of \(J\) in \(A\)\textup{.}
Let \(N\) be a maximal Cohen-Macaulay \(B\)-module and \(L\rightarrow M\rightarrow N\) the minimal \(\cat{MCM}_{A}\)-approximation of \(N\)\textup{.}
If \(\ob{}{}(A/\bar{J}^{2}\rightarrow B,N)=0\)\textup{,} then the composition of maps
\begin{equation*}
\df{\bar{T}^{\mspace{-1mu}\mathit{o}}}{N}\longrightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{N} \longrightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{M}
\end{equation*}
of functors from \({{}_{\varLambda}\cat{H}}/k\) to \(\Sets\) is injective\textup{.}
\end{thm}
\begin{ex}
The existence of a splitting \(B\rightarrow A/\bar{J}^{2}\) implies that \(\ob{}{}(A/\bar{J}^{2},N)=0\)
for all \(B\)-modules \(N\) since \(A/\bar{J}^{2}{\otimes}_{B}N\) gives a lifting of \(N\) to \(A/\bar{J}^{2}\).
\end{ex}
Let \(\cat{C}\) be a category. Then \(\Arr\cat{C}\) denotes the category with objects being arrows in \(\cat{C}\) and arrows being commutative diagrams of arrows in \(\cat{C}\). An endo-functor \(F\) on \(\cat{C}\) induces an endo-functor \(\Arr F\) on \(\Arr\cat{C}\). Let \(B\) be a noetherian local ring and \(\cat{P}_{\!B}\) the additive subcategory of projective modules in \(\cat{mod}_{B}\). Let \(\uhm{}{B}{N}{M}\) denote the homomorphisms from \(N\) to \(M\) in the quotient category \(\underline{\cat{mod}}_{B}=\cat{mod}_{B}/\cat{P}_{\!B}\) i.e.\ \(B\)-homomorphisms modulo the ones factoring through an object in \(\cat{P}_{\!B}\). For each \(N\) in \(\cat{mod}_{B}\) we fix a minimal \(B\)-free resolution and use it to define the syzygy modules of \(N\). Then the association \(N\mapsto\syz{B}{i}N\) induces an endo-functor on \(\underline{\cat{mod}}_{B}\) for each \(i\), considered by A.\ Heller \cite{hel:60}. Define \(\uxt{i}{B}{N}{M}\) as \(\uhm{}{B}{\syz{B}{i}N}{M}\) which turns out to be isomorphic to \(\xt{i}{B}{N}{M}\) for all \(i>0\).
\begin{lem}\label{lem.obssplit}
Let \(A\) be a noetherian local ring and \(I=(f_{1},\dots,f_{n})\) a regular sequence\textup{.} Put \(B=A/I\) and suppose \(N\)\textup{,} \(N_{1}\) and \(N_{2}\) are finite \(B\)-modules\textup{.} Let \(\bar{M}_{i}\) denote \(B{\otimes}\syz{A}{n}N_{i}\)\textup{.}
\begin{enumerate}
\item[(i)] There is an inclusion \(u_{N}:N\rightarrow \bar{M}_{N}\) of \(B\)-modules with \(\bar{M}_{N}\cong B{\otimes}\syz{A}{n}N\) which induces a functor \(u:\ul{\cat{mod}}_{B}\rightarrow\Arr\ul{\cat{mod}}_{B}\)\textup{.}
\item[(ii)] The functor \(u\) commutes with the \(B\)-syzygy functor\textup{:}
\begin{equation*}
\Arr\syz{B}{i}(u_{N})=u_{\syz{B}{i}N}
\end{equation*}
\item[(iii)] The endo-functor \(B{\otimes}\syz{A}{n}(-)\) induces a natural map \(\uxt{i}{B}{N_{1}}{N_{2}}\rightarrow\uxt{i}{B}{\bar{M}_{1}}{\bar{M}_{2}}\) which makes the following diagram commutative\textup{:}
\begin{equation*}
\xymatrix@-0pt@C-12pt@R-12pt@H-0pt{
\uxt{i}{B}{N_{1}}{N_{2}}\ar[rr]\ar[dr]_{(u_{N_{2}})_{*}} && \uxt{i}{B}{\bar{M}_{1}}{\bar{M}_{2}}\ar[dl]^{(u_{N_{1}})^{*}}\\
& \uxt{i}{B}{N_{1}}{\bar{M}_{2}} &
}
\end{equation*}
\item[(iv)] The inclusion \(u_{N}:N\hookrightarrow B{\otimes}\syz{A}{n}N\) splits \(\Longleftrightarrow\ob(A/J^{2}\rightarrow B,N)=0\).
\end{enumerate}
\end{lem}
\begin{rem}
Lemma \ref{lem.obssplit} (iv) strengthens \cite[3.6]{aus/din/sol:93} (in the commutative case).
\end{rem}
\begin{proof}
(i): Suppose \(F_{*}\rightarrow N\) is a minimal \(A\)-free resolution of \(N\). Tensoring the short exact sequence \(\syz{A}{n}N\xra{j} F_{n-1}\rightarrow\syz{A}{n{-}1}N\) with \(B\) gives the exact sequence
\begin{equation}
0\rightarrow\tor{A}{1}{B}{\syz{A}{n{-}1}N}\rightarrow B{\otimes}\syz{A}{n}N\rightarrow \bar{F}_{n-1}\rightarrow B{\otimes}\syz{A}{n{-}1}N\rightarrow 0\,.
\end{equation}
We have \(\tor{A}{1}{B}{\syz{A}{n{-}1}N}\cong \tor{A}{n}{B}{N}\cong N\). Let \(u_{N}\) be the inclusion \(N\cong\ker(B{\otimes}_{A}j)\rightarrow B{\otimes}\syz{A}{n}N\). Then \(N\mapsto u_{N}\) gives a functor of quotient categories.
(ii): Let \(p:Q\rightarrow N\) be the minimal \(B\)-free cover and \(P_{*}\rightarrow \syz{B}{}N\) the minimal \(A\)-free resolution of the \(B\)-syzygy \(\ker(p)\). Then there is an \(A\)-free resolution \(H_{*}\rightarrow Q\) which is an extension of \(F_{*}\) by \(P_{*}\). Since \(\syz{A}{n}B\cong A\), tensoring the short exact sequence of resolutions by \(B\) we obtain the commutative diagram with exact rows:
\begin{equation}
\xymatrix@C-4pt@R-12pt@H-0pt@M4pt{
0\ar[r] & \syz{B}{}N\ar@{_{(}->}[d]_{u_{\Syz N}}\ar[r] & Q\ar@<0.5ex>@{_{(}->}[d]\ar[r] & N\ar@<0.5ex>@{_{(}->}[d]_{u_{N}}\ar[r] & 0 \\
0\ar[r] & B{\otimes}\syz{A}{n}(\syz{B}{}N)\ar[r] & B^{r}{\oplus}Q\ar[r] & B{\otimes}\syz{A}{n}N\ar[r] & 0
}
\end{equation}
which proves the claim.
(iii): By (ii) it is enough to prove this for \(i=0\). The case \(i=0\) follows from the functoriality in (i).
(iv,\(\Leftarrow\)): For the case \(n=1\) see the proof of \cite[3.2]{aus/din/sol:93}. Assume \(n\geq 2\). We follow the proof of \cite[3.6]{aus/din/sol:93} closely. Let \(A_{1}=A/(f_{1})\). Then \(F_{*}^{(1)}=A_{1}{\otimes} F_{*\geq 1}[1]\) gives a minimal \(A_{1}\)-free resolution of \(A_{1}{\otimes}\syz{A}{}N\). We have \(\ob(A/J^{2}\rightarrow B,N)=0\Rightarrow \ob(A/(f_{1})^{2}\rightarrow A_{1},N)=0\) and hence \(N\) is a direct summand of \(A_{1}{\otimes}\syz{A}{}N\). Let \(G_{*}\rightarrow N\) be a minimal \(A_{1}\)-free resolution of \(N\). Then \(G_{*}\) is a direct summand of \(F_{*}^{(1)}\) and hence \(\syz{A_{1}}{n{-}1}N\) is a direct summand of \(\syz{A_{1}}{n{-}1}(A_{1}{\otimes}\syz{A}{}N)=A_{1}{\otimes} \syz{A}{n}N\). Tensoring this situation with \(B\) gives a commutative diagram:
\begin{equation}\label{eq.split}
\xymatrix@C-4pt@R-8pt@H-30pt{
N\ar[r]_(0.3){u}\ar@{=}[d] & B{\otimes}\syz{A}{n}N\ar[r]_(0.55){\bar{j}}\ar@<0.5ex>@{->>}[d] & \bar{F}_{n-1}\ar[r]\ar@<0.5ex>@{->>}[d] & \dots \ar[r] & \bar{F}_{1}\ar[r]\ar@<0.5ex>@{->>}[d] & \bar{F}_{0} \ar[r] & N \\
N\ar[r]^(0.3){u_{1}} & B{\otimes}\syz{A_{1}}{n{-}1}N\ar[r]^(0.6){\bar{j}_{1}}\ar@<0.5ex>[u] & \bar{G}_{n-2}\ar[r]\ar@<0.5ex>[u] & \dots\ar[r] & \bar{G}_{0}\ar[r]\ar@<0.5ex>[u] & N &
}
\end{equation}
Since \(\ob(A/J^{2}\rightarrow B,N)=0\Rightarrow \ob(A_{1}/(f_{2},\dots,f_{n})^{2}\rightarrow B,N)=0\) the map \(u_{1}\) splits by induction on \(n\). So \(u\) splits.
The other direction follows from \cite[3.6]{aus/din/sol:93}.
\end{proof}
\begin{prop}\label{prop.obssplit}
Let \(h:S\rightarrow T\) be a local Cohen-Macaulay map\textup{,} \(J=(f_{1},\dots,f_{n})\) an \(h\)-sequence\textup{,} \(\bar{h}:S\rightarrow\bar{T}=T/J\) the local Cohen-Macaulay map induced from \(h\)\textup{,} and \((\bar{h},\mc{N})\) an object in \(\cat{MCM}_{\bar{h}}\)\textup{.} Let \(\xi: \mc{L}\rightarrow \mc{M}\xra{\pi} \mc{N}\) be the minimal \(\cat{MCM}_{h}\)-approximation of \(\mc{N}\)\textup{.} Then tensoring \(\xi\) by \(\bar{T}\) gives a \(4\)-term exact sequence
\begin{equation*}
0\rightarrow \mc{N}{\otimes} J/J^{2}\longrightarrow \bar{\mc{L}}\longrightarrow \bar{\mc{M}}\xra{\,\,\bar{\pi}\,\,} \mc{N}\rightarrow 0
\end{equation*}
which represents the obstruction class \(\ob(T/J^{2}\rightarrow\bar{T},\mc{N})\in\xt{2}{\bar{T}}{\mc{N}}{\mc{N}{\otimes} J/J^{2}}\)\textup{.}
Moreover\textup{;} \(\ob(T/J^{2}\rightarrow\bar{T},\mc{N})=0 \Longleftrightarrow \ob(T/J^{2}\rightarrow\bar{T},\mc{N}^{\vee})=0 \Longleftrightarrow \bar{\pi}\) splits where \(\mc{N}^{\vee}=\xt{n}{T}{\mc{N}}{\omega_{h}}\)\textup{.}
\end{prop}
\begin{proof}
By Proposition \ref{prop.nakayama} \(\tor{T}{i}{\bar{T}}{\mc{M}}=\cH_{i}(K(\ul{f}){\otimes} \mc{M})=0\) for \(i>0\). There is a map from the defining short exact sequence \(\syz{T}{}\mc{N}\rightarrow F_{0}\rightarrow \mc{N}\) to \(\xi\) lifting \(\id_{\mc{N}}\). Tensoring with \(\bar{T}\) gives a map of \(4\)-term exact sequences with outer terms canonically identified. Hence they represent the same element \(\ob(T/J^{2}\rightarrow\bar{T},\mc{N})\) in \(\xt{2}{\bar{T}}{\mc{N}}{\mc{N}{\otimes} J/J^{2}}\).
By the argument in Remark \ref{rem.Buch} we can assume that \(\xi\) is given as \(\im(d_{n}^{\vee})\rightarrow(\syz{T}{n}\mc{N}^{\vee})^{\vee}\rightarrow \mc{N}^{\vee}{}^{\vee}\) where \((F_{*},d_{*})\) is a minimal \(T\)-free resolution of \(\mc{N}^{\vee}\). By Lemma \ref{lem.obssplit} \(\ob(T/J^{2}\rightarrow\bar{T},\mc{N}^{\vee})=0\) if and only if \(u:\mc{N}^{\vee}\rightarrow \bar{T}{\otimes}\syz{T}{n}\mc{N}^{\vee}\) splits. But applying \(\hm{}{\bar{T}}{-}{\omega_{\bar{h}}}\) to \(u\) gives \(\bar{\pi}\) since \(\mc{N}\cong\xt{n}{T}{\mc{N}^{\vee}}{\omega_{h}}\cong \hm{}{\bar{T}}{\mc{N}^{\vee}}{\omega_{\bar{h}}}\).
\end{proof}
\begin{rem}\label{rem.obssplit}
In the absolute Gorenstein case with \(n=1\) this is given in \cite[4.5]{aus/din/sol:93}.
\end{rem}
\begin{proof}[Proof of Theorem {\ref{thm.defMCM}}]
Given \(S\) in \({{}_{\varLambda}\cat{H}}/k\) and let \(h:S\rightarrow T\) and \(\bar{h}:S\rightarrow \bar{T}=T/JT\) be the induced hCM maps. Let \({}^{i\!}\mc{N}\) be deformations of \(N\) to \(\bar{h}\) for \(i=1,2\) and assume that the minimal \(\cat{MCM}_{h}\)-approximation modules \({}^{i\!}\mc{M}\) of \({}^{i\!}\mc{N}\) are isomorphic as deformations of \(M\). We proceed as in the proof of Theorem \ref{thm.defgrade} (i) with \(S_{n}=S/\fr{m}_{S}^{n{+}1}\), \(\mc{N}_{n}=\mc{N}{\otimes}_{S}S_{n}\) etc., construct a tower of isomorphisms \(\{\phi_{n}:{}^{1\!}\mc{N}\cong{}^{2\!}\mc{N}\}\), and conclude by Lemma \ref{lem.Aapprox} that \({}^{1\!}\mc{N}\) and \({}^{2\!}\mc{N}\) are isomorphic as deformations of \(N\). For the induction step we use that the map of torsor actions along \(\df{\bar{T}}{\mc{N}_{n}}(S_{n+1})\rightarrow\df{T}{\mc{M}_{n}}(S_{n+1})\) is induced by a natural map \(p:\xt{1}{B}{N}{N}\rightarrow \xt{1}{A}{M}{M}\) which is injective. The map \(p\) is given as follows.
Let \(\pi:M\rightarrow N\) denote the \(\cat{MCM}_{A}\)-approximation and \(\bar{\pi}:\bar{M}\rightarrow N\) be the \(B\)-quotient. Then \(\bar{\pi}\) splits by Proposition \ref{prop.obssplit}. Hence \(\bar{\pi}^{*}:\xt{1}{B}{N}{N} \hookrightarrow\xt{1}{B}{\bar{M}}{N}\) splits. Since \(J\) is an \(M\)-regular sequence, \(\xt{1}{B}{\bar{M}}{N}\cong\xt{1}{A}{M}{N}\). Since \(\xt{i}{A}{M}{L}=0\) for all \(i>0\), \(\pi_{*}:\xt{1}{A}{M}{M}\cong \xt{1}{A}{M}{N}\). Summarised:
\begin{equation}\label{eq.xt1}
\xymatrix@C-0pt@R-12pt@H-30pt{
& \xt{1}{A}{M}{N} & \xt{1}{A}{M}{M} \ar[l]_{\cong} \\
\xt{1}{B}{N}{N} \ar@<-0.5ex>@{^{(}->}[r]^{\bar{\pi}^{*}} & \xt{1}{B}{\bar{M}}{N} \ar[u]_{\cong} &
}
\end{equation}
\end{proof}
The technique used to prove Theorem \ref{thm.defMCM} also gives the following result.
\begin{thm}\label{thm.defsyz}
Let \(\varLambda\) and \(T^{\mspace{-1mu}\mathit{o}}\) be henselian and noetherian local rings and \(q:\varLambda\rightarrow T^{\mspace{-1mu}\mathit{o}}\) a local and flat ring homomorphism with \(T^{\mspace{-1mu}\mathit{o}}/\fr{m}_{T^{\mspace{-1mu}\mathit{o}}}=k\) and \(T^{\mspace{-1mu}\mathit{o}}{\otimes}_{\varLambda}k=A\)\textup{.} Suppose \(J=(f_{1},\dots,f_{n})\) is a \(q\)-sequence\textup{.} Put \(\bar{T}^{\mspace{-1mu}\mathit{o}}=T^{\mspace{-1mu}\mathit{o}}/J\)\textup{,} \(B=\bar{T}^{\mspace{-1mu}\mathit{o}}{\otimes}_{\varLambda}k\) and let \(\bar{J}\) be the image of \(J\) in \(A\)\textup{.} Let \(N\) be a finite \(B\)-module and let \(M\) denote the syzygy module \(\syz{A}{n}N\)\textup{.}
If \(\ob(A/\bar{J}^{2}\rightarrow B, N)=0\) then the natural map \(s:\df{\bar{T}^{\mspace{-1mu}\mathit{o}}}{N}\rightarrow\df{T^{\mspace{-1mu}\mathit{o}}}{M}\) is injective\textup{.}
\end{thm}
\begin{proof}
We proceed as in the proof of Theorem \ref{thm.defgrade} and \ref{thm.defMCM}. Given deformations \({}^{i\!}\mc{N}\) of \(N\) to \(\bar{h}\) for \(i=1,2\). They map to \({}^{i\!}\mc{M}:=\syz{T}{n}({}^{i\!}\mc{N})\) which we suppose are isomorphic as deformations of \(M\) to \(h\).
Then the natural syzygy map \(s^{1}:\xt{1}{B}{N}{N}\rightarrow\xt{1}{A}{M}{M}\) induces the map of torsor actions along \(s\) of the inifinitesimal extensions. The composition of \(s^{1}\) with \(\xt{1}{A}{M}{M}\rightarrow\xt{1}{A}{M}{\bar{M}}\cong\xt{1}{B}{\bar{M}}{\bar{M}}\) commutes with the horizontal map in Lemma \ref{lem.obssplit} (iii). But Lemma \ref{lem.obssplit} (iv) implies that \((u_{N})_{*}\) is injective, hence \(s^{1}\) is injective too. Proceeding by induction on \(\fr{m}_{S}^{n+1}\)-truncations of the deformations we construct a tower of isomorphisms and conclude by Lemma \ref{lem.Aapprox}.
\end{proof}
\begin{rem}
Theorem \ref{thm.defsyz} resembles \cite[Thm.\ 1]{ile:07}. However Theorem \ref{thm.defsyz} makes a sounder statement in a more general setting and has a more transparent proof. Indeed, the various similar results in \cite{ile:07} can be changed and proved accordingly.
\end{rem}
\section{The Kodaira-Spencer map of Cohen-Macaulay approximations}
A modular family of objects is roughly speaking a family where the isomorphism class of the fibre changes non-trivially. The Kodaira-Spencer map makes this idea precise. We consider the Kodaira-Spencer classes and maps for families of pairs (algebra, module) and by invoking the long-exact transitivity sequence we relate them to the corresponding notions for the algebra and the module. Then we show that Cohen-Macaulay approximation of modular families under conditions as in Theorem \ref{thm.defgrade}, \ref{thm.defgrade2} and \ref{thm.defMCM} produce new modular families.
The following is a graded version of \cite[II 2.1.5.7]{ill:71}.
\begin{defn}\label{defn.KS}
Let \(\varLambda\rightarrow S\) and \(S\rightarrow \varGamma\) be graded ring homomorphisms with \(\varLambda\) and \(S\) concentrated in degree \(0\). The map \(L^{\text{gr}}_{\varGamma/S}\rightarrow L_{S/\varLambda}{\otimes}_{S}\varGamma[1]\) in the corresponding distinguished transitivity triangle of (graded) cotangent complexes (see \eqref{eq.triangle}) is called the \emph{Kodaira-Spencer class} of \(\varLambda\rightarrow S\rightarrow\varGamma\).
\end{defn}
Composing the Kodaira-Spencer class with the natural augmentation map
\begin{equation*}
L_{S/\varLambda}{\otimes}_{S}\varGamma[1]\longrightarrow\Omega_{S/\varLambda}{\otimes}_{S}\varGamma[1]
\end{equation*}
induces an element \(\kappa(\varGamma/S/\varLambda)\in\gH{0}^{1}(S,\varGamma,\Omega_{S/\varLambda}{\otimes}_{S}\varGamma)\), the cohomological Kodaira-Spencer class, which is also given as follows. Let \(\mc{P}=\mc{P}_{S/\varLambda}\) denote \(S{\otimes}_{\varLambda}S/I^{2}\) where \(I\) is the kernel of the multiplication map \(S{\otimes}_{\varLambda}S\rightarrow S\). There are two ring homomorphisms \(j_{1}\) and \(j_{2}\) from \(S\) to \(\mc{P}\) defined by \(j_{1}:s\mapsto s{\otimes} 1\) and \(j_{2}:s\mapsto 1{\otimes} s\). Let \(d_{S/\varLambda}\) denote the universal derivation (induced by \(j_{2}-j_{1}\)). The principal parts of \(\varGamma\) is \(\mc{P}{\otimes}_{S}\varGamma\) (with the \(j_{2}\) tensor product), which gives an \(S\)-algebra extension (via \(j_{1}\)) representing the Kodaira-Spencer class:
\begin{equation}\label{eq.pp}
\kappa(\varGamma/S/\varLambda):\quad0\rightarrow \Omega_{S/\varLambda}{\otimes}_{S}\varGamma\longrightarrow \mc{P}_{S/\varLambda}{\otimes}_{S}\varGamma\longrightarrow \varGamma \rightarrow 0
\end{equation}
see \cite[III 1.2.6]{ill:71}. Since \(\mc{P}{\otimes}_{S}\varGamma\) has a natural \(\mc{P}\)-algebra structure, \eqref{eq.pp} is also a (graded) algebra lifting of \(\varGamma\) along \(\mc{P}\rightarrow S\) as in Proposition \ref{prop.grobs}. The \(j_{1}\)-extension \(\varGamma{\otimes}_{S}\mc{P}\rightarrow \varGamma\) is a trivial lifting (split by \(\id_{\varGamma}{\otimes} 1_{\mc{P}} \)) and the difference in \(\gH{0}^{1}(S,\varGamma,\Omega_{S/\varLambda}{\otimes}_{S}\varGamma)\) given by Proposition \ref{prop.grobs} (ii) equals \(\kappa(\varGamma/S/\varLambda)\), see \cite[III 2.1.5]{ill:71}.
Moreover, the difference \(1_{\mc{P}}{\otimes}\id_{\varGamma}-\id_{\varGamma}{\otimes} 1_{\mc{P}}\) induces \(d_{S/\varLambda}{\otimes} 1_{T}\) (in degree \(0\)) which is mapped to \(\kappa(\varGamma/S/\varLambda)\) by the connecting homomorphism
\begin{equation}
\partial:\Der_{\varLambda}(S,\Omega_{S/\varLambda}{\otimes}_{S}T)\longrightarrow\gH{0}^{1}(S,\varGamma,\Omega_{S/\varLambda}{\otimes}_{S}\varGamma)
\end{equation}
in the long-exact transitivity sequence, see \cite[III 1.2.6.5 and 1.2.7]{ill:71}.
In the special case of \(\varGamma=T{\oplus}\mc{N}\), \(\mc{N}\) a \(T\)-module and \(S=T\), the transitivity sequence of \(\varLambda\rightarrow T\rightarrow\varGamma\) is given in Proposition \ref{prop.lang}. The Kodaira-Spencer class equals \(\partial(d_{T/\varLambda})\in\xt{1}{T}{\mc{N}}{\Omega_{T/\varLambda}{\otimes}_{T}\mc{N}}\) and is called the (cohomological) \emph{Atiyah class} and is denoted by \(\at_{T/\varLambda}(\mc{N})\), cf.\ \cite[IV 2.3.6-7]{ill:71}. The class is represented by the short exact sequence
\begin{equation}
\at_{T/\varLambda}(\mc{N}):\quad 0\rightarrow\Omega_{T/\varLambda}{\otimes}_{T}\mc{N}\longrightarrow\mc{P}_{T/\varLambda}{\otimes}_{T}\mc{N}\longrightarrow\mc{N}\rightarrow 0\,.
\end{equation}
The \emph{Kodaira-Spencer map of \(\varLambda\rightarrow S\rightarrow\varGamma\)}
\begin{equation}\label{eq.KSmap}
g^{\varGamma}\!:\Der_{\varLambda}(S)\longrightarrow\gH{0}^{1}(S,\varGamma,\varGamma)
\end{equation}
is defined by \(D\mapsto f^{D}_{*}\kappa(\varGamma/S/\varLambda)\) where \(f^{D}\!:\Omega_{S/\varLambda}\rightarrow S\) corresponds to \(D\). Pushout of \eqref{eq.pp} by \(f^{D}{\otimes}\id_{\varGamma}\) gives the corresponding algebra lifting of \(\varGamma\) along \(S[\varepsilon]\rightarrow S\) given by \(g^{\varGamma}(D)\).
\begin{prop}\label{prop.at}
Let \(\varGamma\) denote the graded \(S\)-algebra \(T{\oplus}\mc{N}\) where \(\varLambda\rightarrow S\) and \(S\rightarrow T\) are \textup{(}ungraded\textup{)} ring homomorphisms and \(\mc{N}\) is a \(T\)-module\textup{.} Consider the transitivity sequence of \(S\rightarrow T\xra{i} \varGamma\) in \textup{Proposition \ref{prop.lang}:}
\begin{align*}
\dots\rightarrow\Der_{S}(T,\Omega_{S/\varLambda}{\otimes}_{S}T) & \xra{\partial}\xt{1}{T}{\mc{N}}{\Omega_{S/\varLambda}{\otimes}_{S}\mc{N}}\xra{u}\gH{0}^{1}(S,\varGamma,\Omega_{S/\varLambda}{\otimes}_{S}\varGamma)\xra{i^{*}}\\
\cH^{1}(S,T,\Omega_{S/\varLambda}{\otimes}_{S}T) & \xra{\partial}\dots
\end{align*}
\begin{enumerate}
\item[(i)] The map \(i^{*}\) takes the Kodaira-Spencer class \(\kappa(\varGamma/S/\varLambda)\) to \(\kappa(T/S/\varLambda)\)\textup{.}
\item[(ii)] Assume \(\kappa(T/S/\varLambda)=0\) and choose an \(S\)-algebra splitting \(\sigma: T\rightarrow \mc{P}{\otimes}_{S}T\)\textup{.} Then there is a class \(\kappa(\sigma,\mc{N})=\kappa(T/S/\varLambda,\sigma,\mc{N})\in\xt{1}{T}{\mc{N}}{\Omega_{S/\varLambda}{\otimes}_{S}\mc{N}}\)\textup{,} natural in the argument\textup{,} which maps to \(\kappa(\varGamma/S/\varLambda)\) by \(u\)\textup{.}
\item[(iii)] Let \(D(\sigma)\in\Der_{\varLambda}(T,\Omega_{S/\varLambda}{\otimes}_{S}T)\) be the derivation corresponding to the splitting \(\sigma\) and for each \(D_{1}\in\Der_{\varLambda}(S)\) let \(X_{\sigma}(D_{1})\) denote \(f^{D_{1}}_{*}D(\sigma)\in\Der_{\varLambda}(T)\)\textup{.} Then
\begin{equation*}
f^{D_{1}}_{*}\kappa(\sigma,\mc{N})=f^{X_{\sigma}(D_{1})}_{*}\at_{T/\varLambda}(\mc{N})\quad\text{in}\quad \xt{1}{T}{\mc{N}}{\mc{N}}\,.
\end{equation*}
\end{enumerate}
\end{prop}
\begin{proof}
The degree zero part of \eqref{eq.pp} gives the image \(i^{*}\kappa(\varGamma/S/\varLambda)\) represented by the algebra extension
\begin{equation}
\kappa(T/S/\varLambda):\quad 0\rightarrow\Omega_{S/\varLambda}{\otimes}_{S} T\longrightarrow \mc{P}_{S/\varLambda}{\otimes}_{S}T\longrightarrow T\rightarrow 0\,.
\end{equation}
The degree one part is the short exact sequence of \(\mc{P}{\otimes}_{S}T\)-modules
\begin{equation}\label{eq.pmod}
\tilde{\alpha}=\tilde{\alpha}(T/S/\varLambda,\mc{N}):\quad0\rightarrow \Omega_{S/\varLambda}{\otimes}_{S}\mc{N}\longrightarrow\mc{P}_{S/\varLambda}{\otimes}_{S}\mc{N}\longrightarrow\mc{N}\rightarrow 0\,.
\end{equation}
The splitting \(\sigma\) makes \(\tilde{\alpha}\) to a short exact sequence of \(T\)-modules which defines \(\kappa(T/S/\varLambda,\sigma,\mc{N})\).
For (iii) we have \(X_{\sigma}(D_{1})=f_{*}^{X_{\sigma}(D_{1})}d_{T/\varLambda}\) and the result follows from the commutative diagram
\begin{equation}\label{eq.kappa}
\xymatrix@C-22pt@R-15pt@H-30pt{
D(\sigma)\in\Der_{\varLambda}(T,\Omega_{S/\varLambda}{\otimes}_{S}T)\ar@{.>}[rr]\ar[ddd]_{\partial}\ar[dr]_{f^{D_{1}}_{*}} && \Der_{\varLambda}(T,\Omega_{T/\varLambda})\ni d_{T/\varLambda}\ar[dl]^{f^{X_{\sigma}(D_{1})}_{*}}\ar[ddd]^{\partial} \\
& \Der_{\varLambda}(T,T)\ar[d]^{\partial} \\
& \xt{1}{T}{\mc{N}}{\mc{N}} \\
\kappa(\sigma,\mc{N})\in\xt{1}{T}{\mc{N}}{\Omega_{S/\varLambda}{\otimes}_{S}\mc{N}}\ar[ur]^{f^{D_{1}}_{*}}\ar@{.>}[rr] && \xt{1}{T}{\mc{N}}{\Omega_{T/\varLambda}{\otimes}_{T}\mc{N}}\ni\at_{T/\varLambda}(\mc{N})\ar[lu]_{f^{X_{\sigma}(D_{1})}_{*}}
}
\end{equation}
where the two outer vertical maps are pointed.
\end{proof}
We call \(\kappa(\sigma,\mc{N})\) for the Kodaira-Spencer class of \((T/S/\varLambda,\sigma,\mc{N})\).
Define the \emph{Kodaira-Spencer map of \((T/S/\varLambda,\sigma,\mc{N})\)}
\begin{equation}\label{eq.modKS}
g^{(\sigma,\,\mc{N})}:\Der_{\varLambda}(S)\longrightarrow\xt{1}{T}{\mc{N}}{\mc{N}}
\end{equation}
by \(g^{(\sigma,\,\mc{N})}(D):=(f^{D}{\otimes}\id)_{*}\kappa(\sigma,\mc{N})\).
In the case \(T=S{\otimes}_{\varLambda} T^{\mspace{-1mu}\mathit{o}}\) we always choose the \(S\)-algebra splitting \(S{\otimes}_{\varLambda} T^{\mspace{-1mu}\mathit{o}}\rightarrow \mc{P}_{S/\varLambda}{\otimes}_{S}S{\otimes}_{\varLambda}T^{\mspace{-1mu}\mathit{o}}\cong\mc{P}_{S/\varLambda}{\otimes}_{\varLambda}T^{\mspace{-1mu}\mathit{o}}\) given by \(s{\otimes} t\mapsto j_{1}(s){\otimes} t\). In particular \(\kappa(T/S/\varLambda)=0\) and we get a canonical Kodaira-Spencer class \(\kappa(\mc{N})\) and a corresponding Kodaira-Spencer map \(g^{\mc{N}}\).
\begin{rem}
There is no reason to believe that \(\kappa(\sigma,\mc{N})\) maps to \(\at_{T/\varLambda}(\mc{N})\) in diagram \eqref{eq.kappa} for any choice of \(\sigma\). While there is a canonical map of short exact sequences (of \(S\)-modules)
\begin{equation*}
\xymatrix@C-0pt@R-12pt@H-30pt{
\;\;\;\kappa(\sigma,\mc{N}): & 0\ar[r] & \Omega_{S/\varLambda}{\otimes}_{S}\mc{N}\ar[r]\ar[d] & \mc{P}_{S/\varLambda}{\otimes}_{S}\mc{N}\ar[r]\ar[d]^{\tau} &\mc{N}\ar[r]\ar@{=}[d] & 0 \\
\at_{T/\varLambda}(\mc{N}): & 0\ar[r] & \Omega_{T/\varLambda}{\otimes}_{T}\mc{N}\ar[r] & \mc{P}_{T/\varLambda}{\otimes}_{T}\mc{N}\ar[r] &\mc{N}\ar[r] & 0
}
\end{equation*}
\(\tau\) is in general not \(T\)-linear. However, in the case \(S\rightarrow T\) is smooth then \(\Omega_{S/\varLambda}{\otimes} T\rightarrow \Omega_{T/\varLambda}\rightarrow \Omega_{T/S}\) is split exact. There is a non-canonical lifting of the universal derivation in \(\Der_{\varLambda}(T,\Omega_{T/\varLambda})\) to \(\Der_{\varLambda}(T,\Omega_{S/\varLambda}{\otimes} T)\) and the corresponding choice of splitting \(\sigma\) makes \(\tau\) \(T\)-linear and so (or by \eqref{eq.kappa}) \(\kappa(\sigma,\mc{N})\) maps to \(\at_{T/\varLambda}(\mc{N})\).
\end{rem}
\begin{ex}
Another special case is given by the base change of \(h:S\rightarrow T\) with itself to \(h{\otimes} T: T\rightarrow T{\otimes}_{S}T=T^{\ot2}\) and a \(T\)-flat \(T^{\ot2}\)-module \(\mc{N}\), cf.\ Section \ref{sec.fund}. Then \(\kappa(T^{\otimes 2}/T/S,\mc{N})\) in \(\xt{1}{T^{\ot2}}{\mc{N}}{\Omega_{T^{\ot2}/T}{\otimes}\mc{N}}\) equals \(\at_{T^{\ot2}/T}(\mc{N})\). The multiplication map \(\mu_{T^{\ot2}/T}:\mc{P}_{T^{\ot2}/T}\rightarrow T^{\ot2}\) equals \(\id_{T}{\otimes}\mu_{T/S}:T{\otimes}_{S}\mc{P}_{T/S}\rightarrow T^{\ot2}\). It follows that \(\at_{T^{\ot2}/T}(\mc{N})\) maps to \(\at_{T/S}(\mc{N})\) in \(\xt{1}{T}{\mc{N}}{\Omega_{T/S}{\otimes}\mc{N}}\) by the natural map. If \(\mc{N}=T\) then \(\at_{T/S}(T)=0\), but in general \(\at_{T^{\ot2}/T}(T)\neq0\).
\end{ex}
\begin{ex}
The transitivity sequence of \(\varLambda\rightarrow T\xra{i}\varGamma\) and \(J\) with \(\varGamma=T{\oplus}\mc{N}\) and \(J=J_{0}{\oplus}J_{1}\) in Proposition \ref{prop.lang}
\begin{equation}\label{eq.deg0}
0\rightarrow \hm{}{T}{\mc{N}}{J_{1}}\rightarrow \gDer{0}_{\varLambda}(\varGamma,J)\xra{i^{*}} \Der_{\varLambda}(T,J_{0})\rightarrow \xt{1}{T}{\mc{N}}{J_{1}}\rightarrow\dots
\end{equation}
suggests the following characterisation. An element \(\mc{D}\in\gDer{0}_{\varLambda}(\varGamma,J)\) is given by its degree \(0\) restriction \(D:=i^{*}(\mc{D})\in\Der_{\varLambda}(T,J_{0})\) and its degree \(1\) restriction \(\nabla_{\!D}:=\mc{D}_{\vert\mc{N}}\in\hm{}{\varLambda}{\mc{N}}{J_{1}}\) which should satisfy the following Leibniz rule: For all \(t\) in \(T\) and \(n\) in \(\mc{N}\)
\begin{equation}\label{eq.Leib}
\nabla_{\!D}(tn)=t\nabla_{\!D}(n)+D(t)n\,.
\end{equation}
With notation as in Proposition \ref{prop.at} recall
that \(\kappa(\varGamma/S/\varLambda)=\partial(d_{S/\varLambda}{\otimes} 1_{T})\) in the transitivity sequence of \(\varLambda\rightarrow S\rightarrow\varGamma\):
\begin{equation}\label{eq.deg0b}
\begin{split}
0\rightarrow{} &\gDer{0}_{S}(\varGamma,\Omega_{S/\varLambda}{\otimes}\varGamma)\rightarrow\gDer{0}_{\varLambda}(\varGamma,\Omega_{S/\varLambda}{\otimes}\varGamma)\rightarrow\Der_{\varLambda}(S,\Omega_{S/\varLambda}{\otimes} T)\xra{\partial} \\
&\gH{0}^{1}(S,\varGamma,\Omega_{S/\varLambda}{\otimes}\varGamma)\rightarrow\dots
\end{split}
\end{equation}
Hence \(\kappa(\varGamma/S/\varLambda)=0\) if and only if there exists a \(D\in\Der_{\varLambda}(T,\Omega_{S/\varLambda}{\otimes} T)\) which restricts to \(d_{S/\varLambda}{\otimes} 1_{T}\) and a \(\nabla_{\!D} \in\hm{}{\varLambda}{\mc{N}}{\Omega_{S/\varLambda}{\otimes} \mc{N}}\) satisfying \eqref{eq.Leib}. As a well known special case (\(S=T\)) we get \(\at_{T/\varLambda}(\mc{N})=0\) if and only if there exists a \(\nabla\in \hm{}{\varLambda}{\mc{N}}{\Omega_{T/\varLambda}{\otimes} \mc{N}}\) satisfying \eqref{eq.Leib} with \(D=d_{T/\varLambda}\in\Der_{\varLambda}(T,\Omega_{T/\varLambda})\) (i.e.\ \(\nabla\) is a connection), or equivalently, a graded derivation \(\mc{D}\in \gDer{0}_{\varLambda}(\varGamma,\Omega_{T/\varLambda}{\otimes}_{T}\varGamma)\) restricting to \(d_{T/\varLambda}\). Note that \eqref{eq.deg0} with \(J=\Omega_{T/\varLambda}{\otimes}_{T}\varGamma\) equals \eqref{eq.deg0b} in this case.
\end{ex}
Recall the maps of cohomology groups \(\sigma^{1}_{j}(I)\) and \(\tau^{1}_{j}(I)\) in \eqref{eq.sigma} and \eqref{eq.tau}.
\begin{prop}\label{prop.KS}
In addition to the assumptions in \textup{Lemma \ref{lem.cohmap}} suppose \(\varLambda\rightarrow S\) is a ring homomorphism\textup{.} For \(j=1,2\) the following holds\textup{:}
\begin{enumerate}
\item[(i)] The map \(\sigma^{1}_{j}(\Omega_{S/\varLambda})\) takes \(\kappa(\varGamma_{0}/S/\varLambda)\) to \(\kappa(\varGamma_{j}/S/\varLambda)\) and the Kodaira-Spencer maps \(g^{\varGamma_{i}}:\Der_{\varLambda}(S)\rightarrow\gH{0}^{1}(S,\varGamma_{i},\varGamma_{i})\) commute with \(\sigma^{1}_{j}\)\textup{,} i\textup{.}e\textup{.}\ \(\sigma^{1}_{j} g^{\varGamma_{0}}=g^{\varGamma_{j}}\)\textup{.}
\item[(ii)] Assume \(\kappa(T/S/\varLambda)=0\) and choose an \(S\)-algebra splitting \(\sigma: T\rightarrow \mc{P}{\otimes}_{S}T\)\textup{.}
Then \(\tau^{1}_{j}(\Omega_{S/\varLambda})\) maps \(\kappa(\sigma,\mc{N})\) to \(\kappa(\sigma,X_{j})\) and the Kodaira-Spencer maps \(g^{(\sigma,X_{i})}:\Der_{\varLambda}(S)\rightarrow\xt{1}{T}{X_{i}}{X_{i}}\) commute with \(\tau^{1}_{j}\), i.e.\ \(\tau^{1}_{j} g^{(\sigma,\mc{N})}\!=g^{(\sigma,X_{j})}\)\textup{.}
\end{enumerate}
\end{prop}
\begin{proof}
(i): Put \(\kappa_{j}=\kappa(\varGamma_{j}/S/\varLambda)\), \(\Omega=\Omega_{S/\varLambda}\) and let \(\varGamma(\iota):\varGamma_{0}\rightarrow\varGamma_{2}\) denote the graded ring homomorphism induced from \(\iota\). Then \(\varGamma(\iota)\) induces a map of short exact sequences \(\kappa_{0}\rightarrow\kappa_{2}\), hence a map of short exact sequences \(\varGamma(\iota)_{*}\kappa_{0}\rightarrow\varGamma(\iota)^{*}\kappa_{2}\), i.e.\ \(\sigma_{2}(\Omega)(\kappa_{0})=(\varGamma(\iota)^{*})^{-1}\varGamma(\iota)_{*}\kappa_{0}=\kappa_{2}\). The maps \(\sigma_{2}(\Omega)\) and \(\sigma_{2}(S)\) commute with the covariant action of \(\Der_{\varLambda}(S)\), hence the second assertion follows from the first. The arguments for the cases \(j=1\) and (ii) are similar.
\end{proof}
There are corresponding \emph{local} Kodaira-Spencer maps given as follows. Let \(t\in\Spec T\) map to \(s\in\Spec S\) and consider the localisations \(S_{\fr{p}_{s}}\rightarrow T_{\fr{p}_{t}}\) and \(S_{\fr{p}_{s}}\rightarrow\varGamma_{\fr{p}_{t}}\) and the induced map \(\varLambda\rightarrow S_{\fr{p}_{s}}\). The localisation map \(\gH{0}^{1}(S,\varGamma,\Omega_{S/\varLambda}{\otimes}_{S}\varGamma)\rightarrow\gH{0}^{1}(S_{\fr{p}_{s}},\varGamma_{\fr{p}_{t}},\Omega_{S_{\fr{p}_{s}}/\varLambda}{\otimes}_{S_{\fr{p}_{s}}}\varGamma_{\fr{p}_{t}})\) maps \(\kappa(\varGamma/S/\varLambda)\) to \(\kappa(\varGamma_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\varLambda)\). Let \(\bar{\kappa}(\varGamma_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\varLambda)\) denote the image in \(\gH{0}^{1}(S_{\fr{p}_{s}},\varGamma_{\fr{p}_{t}},\Omega_{S_{\fr{p}_{s}}/\varLambda}{\otimes}_{S_{\fr{p}_{s}}}\varGamma(t))\) by the map induced from \(\varGamma_{\fr{p}_{t}}\rightarrow\varGamma_{\fr{p}_{t}}{\otimes}_{S_{\fr{p}_{s}}}k(s)=\varGamma(t)\). Assume that \(\varGamma_{\fr{p}_{t}}\) is \(S_{\fr{p}_{s}}\)-flat. Then the natural base change map
\begin{equation}
\gH{0}^{1}(k(s),\varGamma(t),\Omega_{S_{\fr{p}_{s}}/\varLambda}{\otimes}_{S_{\fr{p}_{s}}}\varGamma(t))\longrightarrow\gH{0}^{1}(S_{\fr{p}_{s}},\varGamma_{\fr{p}_{t}},\Omega_{S_{\fr{p}_{s}}/\varLambda}{\otimes}_{S_{\fr{p}_{s}}}\varGamma(t))
\end{equation}
is an isomorphism, see \cite[II 2.2]{ill:71}. With this identification we define \(g^{\varGamma}\!(t)(D):=(f^{D}{\otimes} \id_{\varGamma(t)})_{*}\bar{\kappa}(\varGamma_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\varLambda)\) for any \(D\) in \(\Der_{{\varLambda}}(S_{\fr{p}_{s}},k(s))\) and obtain local Kodaira-Spencer maps at \(t\) of \(\varGamma\), and (similarly) of \(T\), respectively:
\begin{align}
g^{\varGamma}\!(t):\,\,&\Der_{\varLambda}(S_{\fr{p}_{s}},k(s))\longrightarrow\gH{0}^{1}(k(s),\varGamma(t),\varGamma(t)) \\
g^{T}\!(t):\,\,&\Der_{\varLambda}(S_{\fr{p}_{s}},k(s))\longrightarrow\cH^{1}(k(s),T(t),T(t))
\end{align}
commuting with the natural map \(\gH{0}^{1}(k(s),\varGamma(t),\varGamma(t))\rightarrow \cH^{1}(k(s),T(t),T(t))\) in Proposition \ref{prop.lang}.
Note that if \(\varLambda\) is an algebraically closed field then \(\Der_{\varLambda}(S_{\fr{p}_{s}},k(s))\) is canonically isomorphic to the Zariski tangent space at any closed point \(s\in\Spec S\).
Let \(\mc{P}\) and \(\Omega\) denote \(\mc{P}_{S_{\fr{p}_{s}}/\varLambda}\) and \(\Omega_{S_{\fr{p}_{s}}/\varLambda}\) respectively. Assume \(\varGamma=T{\oplus}\mc{N}\). As in the global case we get a graded algebra extension representing \(\bar{\kappa}(\varGamma_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\varLambda)\) which in degree \(1\) is a short exact sequence \(\tilde{\alpha}(t): \Omega{\otimes}\mc{N}(t)\rightarrow k(s){\otimes}\mc{P}{\otimes}\mc{N}_{\fr{p}_{t}}\rightarrow\mc{N}(t)\).
If \(\bar{\kappa}(T_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\varLambda)=0\) we choose an \(S\)-algebra splitting \(\sigma:T(t)\rightarrow k(s){\otimes}\mc{P}{\otimes}\varGamma_{\fr{p}_{t}}\) and the local Kodaira-Spencer class \(\bar{\kappa}(\sigma,\mc{N}_{\fr{p}_{t}})\) in \(\xt{1}{T(t)}{\mc{N}(t)}{\Omega{\otimes}\mc{N}(t)}\) is represented by the obtained short exact sequence of \(T(t)\)-modules. Then we define the local Kodaira-Specer map of \((T_{\fr{p}_{t}}/S_{\fr{p}_{s}}/\varLambda,\sigma,\mc{N}_{\fr{p}_{t}})\)
\begin{equation}
g^{(\sigma,\mc{N})}\!(t):\,\,\Der_{\varLambda}(S_{\fr{p}_{s}},k(s))\longrightarrow\xt{1}{T(t)}{\mc{N}(t)}{\mc{N}(t)}
\end{equation}
by \(g^{(\sigma,\mc{N})}(t)(D):=(f^{D}{\otimes}\id)_{*}\bar{\kappa}(\sigma,\mc{N}_{\fr{p}_{t}})\).
Similarly, the class \(g^{\varGamma}\!(t)(D)\) is represented by a lifting of graded algebras \(\varGamma'\rightarrow \varGamma(t)\) along \(k(s)[\varepsilon]\rightarrow k(s)\). If the lifting \(g^{T}\!(t)(D):\, T'\rightarrow T(t)\) splits, a choice of splitting makes the short exact sequence \(\tilde{\alpha}(t)(D):\varepsilon\mc{N}(t)\rightarrow N'\rightarrow\mc{N}(t)\) \(T\)-linear and defines \(\tilde{\alpha}(t)(D)\) as an extension in the subspace \(\xt{1}{T(t)}{\mc{N}(t)}{\mc{N}(t)}\) of\, \(\xt{1}{T'}{\mc{N}(t)}{\mc{N}(t)}\).
We assume that \(\varLambda\) is an algebraically closed field \(k\) for the rest of this section.
\begin{defn}\label{defn.modular}
Let \(h:S\rightarrow T\) be a local flat map of noetherian \(k\)-algebras and \(\mc{N}\) an \(S\)-flat \(T\)-module. Put \(\varGamma=T{\oplus}\mc{N}\), \(A=T{\otimes}_{S}k\), \(N=\mc{N}{\otimes}_{S}k\) and \(\varGamma(0)=\varGamma{\otimes}_{S}k=A{\oplus}N\).
We say that \((h,\mc{N})\) is \emph{locally modular} if the local Kodaira-Spencer map \(g^{\varGamma}\!(0):\Der_{k}(S,k)\rightarrow\gH{0}^{1}(k,\varGamma(0),\varGamma(0))\) is injective. If in addition \(T=S{\tilde{\otimes}}_{k}A\) then \(\mc{N}\) is locally modular if \(g^{\mc{N}}\!(0):\Der_{k}(S,k)\rightarrow\xt{1}{A}{N}{N}\) is injective.
Let \(h^{\text{ft}}:S\rightarrow T\) be a faithfully flat finite type map of noetherian \(k\)-algebras with a \(k\)-point \(t\in\Spec T\) mapping to \(s\in \Spec S\) and \(\mc{N}\) is an \(S\)-flat finite \(T\)-module. We say that \((h^{\textnormal{ft}},\mc{N})\) is \emph{modular at \(t\)} if the henselisation of \((h^{\text{ft}},\mc{N})\) at \(t\) is locally modular.
If \(A\) is a finite type \(k\)-algebra and \(T=S{\otimes}_{k} A\), then \(\mc{N}\) is modular at \(t\) as \(T\)-module if its henselisation at \(t\) is locally modular. Let \(\nabla(h,\mc{N})\) (\(\nabla_{T}(\mc{N})\) if \(T=S{\otimes}_{k}A\)) denote the set of \(k\)-points \(t\in \Supp\mc{N}\) where \((h,\mc{N})\) (respectively \(\mc{N}\) as \(T\)-module) is modular.
\end{defn}
\begin{cor}\label{cor.modY}
Let \(h:S\rightarrow T\) be a finite type Cohen-Macaulay map of \(k\)-algebras and let \(\mc{N}\) be in \(\cat{mod}{}^{\textnormal{fl}}_{h}\)\textup{.} Let \(\mc{N}\xra{\iota}\mc{L}'\rightarrow\mc{M}'\) and \(\mc{L}\rightarrow\mc{M}\rightarrow\mc{N}\) be a \(\hat{\cat{D}}{}^{\textnormal{fl}}_{h}\)-hull and a \(\cat{MCM}_{h}\)-approximation for \(\mc{N}\) respectively\textup{.}
\begin{enumerate}
\item[(i)] Suppose \(\hm{}{T(t)}{\mc{N}(t)}{\mc{M}'(t)}=0\) for all \(t\in \mSpec T\) \textup{(}e.g.\ there is an \(h\)-regular element contained in \(\ann_{T}\mc{N}\)\textup{).}
Then
\begin{equation*}
\nabla(h,\mc{N})=\nabla(h,\mc{L}')\quad\text{and}\quad \nabla_{T}(\mc{N})=\nabla_{T}(\mc{L}')\,\,\text{if}\,\, T=S{\otimes}_{k}A\,.
\end{equation*}
\item[(ii)] Suppose \(\hm{}{T(t)}{\mc{L}(t)}{\mc{N}(t)}=0\) for all \(t\in \mSpec T\)\textup{.}
Then
\begin{equation*}
\nabla(h,\mc{N})=\nabla(h,\mc{M})\quad\text{and}\quad \nabla_{T}(\mc{N})=\nabla_{T}(\mc{M})\,\,\text{if}\,\, T=S{\otimes}_{k}A\,.
\end{equation*}
\end{enumerate}
\end{cor}
\begin{proof}
(i) Let \(t\) be a \(k\)-point in \(\Spec T\). By Theorem \ref{thm.flatCMapprox} \(\iota(t)\) is a \(\hat{\cat{D}}_{T(t)}\)-hull for \(\mc{N}(t)\) and by Proposition \ref{prop.minapprox} \(\iota(t)\) is minimal if and only if \(\iota_{\fr{p}_{t}}\) is minimal. In particular; the minimal hull of
\(\mc{N}_{\fr{p}_{t}}\) is a direct summand of \(\mc{L}'_{\fr{p}_{t}}\). We therefore assume that \(\iota_{\fr{p}_{t}}\) and hence \(\iota(t)\) is minimal.
By the condition, the map \(\tau^{1}_{2}\) in \eqref{eq.tau} is injective. It implies that the map \(\sigma^{1}_{2}:\gH{0}^{1}(k(s),\varGamma_{0}(t),\varGamma_{0}(t))\rightarrow\gH{0}^{1}(k(s),\varGamma_{2}(t),\varGamma_{2}(t))\) in \eqref{eq.sigma} is injective and Proposition \ref{prop.KS} gives the statement. (ii) is similar.
\end{proof}
\begin{ex}\label{ex.modY}
Let \(A\) be a CM finite type \(k\)-algebra and domain of dimension \(\geq 2\). Let \(h:A\rightarrow T=A^{{\otimes} 2}\) be the base change by \(S=A\) and \(\mc{N}=A\) be the \(A\)-flat \(T\)-module defined by the multiplication map \(T\rightarrow A\). Let \(\Delta\subseteq \Spec T\) denote the closed points on the diagonal and let \(t\) be a closed point in \(\Spec T\) mapping to \(s\) in \(\Spec A\). If \(t\notin\Delta\) then \(\mc{N}(t)=0\). If \(t\in \Delta\) then \(T(t)\cong A_{\fr{p}_{s}}\) and \(\mc{N}(t)\cong k(s)\). The local Kodaira-Spencer class \(\bar{\kappa}(\mc{N}(t))\in \xt{1}{A_{\fr{p}_{s}}}{k(s)}{\Omega_{A_{\fr{p}_{s}}/k}{\otimes} k(s)}\) is represented by
\begin{equation}
\xymatrix@C-0pt@R-12pt@H-30pt{
0\ar[r] & \Omega_{A_{\fr{m}}/k}{\otimes} k(s)\ar[r] & k(s){\otimes} \mc{P}_{A_{\fr{m}}/k}\ar[r] & k(s)\ar[r] & 0 \\
0\ar[r] & \fr{m}/\fr{m}^{2}\ar[r]\ar[u]^{\cong}_{\delta} & A/\fr{m}^{2}\ar[r] \ar[u]^{\cong}_{\chi} & k(s)\ar@{=}[u]\ar[r] & 0
}
\end{equation}
(with \(\fr{m}=\fr{p}_{s}\)) where \(\delta(\bar{x})=d_{A_{\fr{m}}/k}(x){\otimes} 1\) and \(\chi\) is induced by \(1{\otimes} j_{2}\) (note that if \(x, y\in\fr{m}\) then \(1{\otimes} j_{2}(xy)=1{\otimes}([j_{2}(x)-j_{1}(x)][(j_{2}(y)-j_{1}(y)])\in k(s){\otimes} I^{ 2}\)). The local Kodaira-Spencer map \(g^{\mc{N}}(t)\) is given by the pushout
\begin{equation}
\phi\in\hm{}{k(s)}{\fr{m}/\fr{m}^{2}}{k(s)}\longrightarrow \xt{1}{A_{\fr{m}}}{k(s)}{k(s)}\ni\phi_{*}\bar{\kappa}(\mc{N}(t))
\end{equation}
which is an isomorphism. By Corollary \ref{cor.modY} we have \(\nabla_{T}(\mc{N})=\nabla_{T}(\mc{L}')=\Delta\). Put \(\mc{Q}'=\hm{}{T}{\omega_{h}}{\mc{L}'}\). By Proposition \ref{prop.defequiv} also the local Kodaira-Spencer map \(g^{\mc{Q}'}(t)\) is injective for \(t\in \Delta\). Hence \(\nabla_{T}(\mc{Q}')=\Delta\). Note that \(\mc{L}'(t)\) and \(\mc{Q}'(t)\) are rigid for \(t\notin\Delta\).
\end{ex}
\begin{cor}\label{cor.modCM}
Suppose \(A\) is a finite type Cohen-Macaulay \(k\)-algebra and \(S\) a noetherian \(k\)-algebra and put \(h:S\rightarrow T=S{\otimes}_{k}A\)\textup{.} Let \(J=(f_{1},\dots,f_{n})\) be an \(A\)-sequence\textup{,} put \(B=A/J\) and let \(\bar{h}:S\rightarrow\bar{T}=S{\otimes}_{k}B\) be the induced Cohen-Macaulay map\textup{.} Suppose \(\mc{N}\) is in \(\cat{MCM}_{\bar{h}}\subseteq\cat{mod}{}^{\textnormal{fl}}_{h}\) and let \(\mc{L}\rightarrow\mc{M}\rightarrow\mc{N}\) be an \(\cat{MCM}_{h}\)-approximation of \(\mc{N}\)\textup{.} Assume \(\ob(T/(JT)^{2}\rightarrow \bar{T}, \mc{N})=0\)\textup{.} Then
\begin{equation*}
\nabla_{\bar{T}}(\mc{N})=\nabla_{T}(\mc{M})\cap\Supp\bar{T}\,.
\end{equation*}
\end{cor}
\begin{proof}
By Proposition \ref{prop.obsmodule} there is a lifting \(\mc{N}_{1}\rightarrow\mc{N}\) of \(\mc{N}\) to \(T_{1}=T/(JT)^{2}\). It induces liftings \(\mc{N}_{1}(t)\rightarrow\mc{N}(t)\) for all \(k\)-points \(t\) in \(\Supp\bar{T}\). The inclusion \(\bar{\tau}:\xt{1}{\bar{T}(t)}{\mc{N}(t)}{\mc{N}(t)}\rightarrow\xt{1}{T(t)}{\mc{M}(t)}{\mc{M}(t)}\) in \eqref{eq.xt1} commutes with the local Kodaira-Spencer maps. Proceed as in the proof of Corollary \ref{cor.modY}.
\end{proof}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction and Summary}
The mirror Thermodynamic Bethe Ansatz (TBA) approach is the only tool currently available to determine exact energies of ${\rm AdS}_5\times {\rm S}^5$ string states and, thanks to the AdS/CFT conjecture \cite{M}, scaling dimensions of ${\cal N}=4$ SYM local gauge-invariant composite operators. In essence, the TBA is a set of coupled non-linear integral equations
for the so-called Y-functions, whose solutions are expected to yield the spectrum of the corresponding string/gauge theory.
\smallskip
Although the TBA approach\footnote{We will not describe it here, referring the interested reader to the original literature \cite{Zamolodchikov90,Kuniba1} and recent
reviews \cite{Kuniba2,Arutyunov:2009ga,Bajnok:2010ke}. } has been successfully used in the case of two-dimensional relativistic integrable models for quite some time \cite{Zamolodchikov90}, its application
to the ${\rm AdS}_5\times {\rm S}^5$ superstring\footnote{There is currently much evidence in favor of integrability of the ${\rm AdS}_5\times {\rm S}^5$ superstring and ${\cal N}=4$ SYM, see the recent collection of reviews \cite{Brev} and references therein.} is not straightforward and requires a careful thought. Importantly, the string sigma model is not Lorentz invariant on the two-dimensional world-sheet and, therefore, under
the double Wick rotation, which is in the heart of the TBA construction, it transforms into another model, termed in \cite{AF07} a {\it mirror}. The ground state
energy of the original string model is then related to the free energy (or Witten's index, depending on the boundary conditions for fermions) of its mirror.
In turn, the free energy and the TBA equations for the ground state is derivable from the so-called string hypothesis \cite{Takahashi72}, which for the ${\rm AdS}_5\times {\rm S}^5$ mirror
model has been formulated in \cite{AF09a}. In this way the ground state TBA equations were obtained \cite{AF09b}-\cite{AF09d},\footnote{The final missing piece of the TBA which was the mirror dressing phase was provided by \cite{AFdp}.} and in \cite{FS} it was shown that
the corresponding solution correctly reproduces the vanishing energy of the ground state that corresponds to the protected half-BPS operator of the gauge theory.
\smallskip
The importance of the ground state TBA equations lies in the fact that they admit a generalization to excited states by means of a
contour deformation trick which is similar to the analytic continuation procedure of \cite{DT96}, see also \cite{BLZe}. The contour deformation relies on an assumption that the set of TBA equations is universal for any state of the model; excited states TBA equations may differ from each other only by a choice of integration contours of convolution terms, and by analytic properties of Y-functions which determine driving terms in the TBA equations once the integration contours are taken back to the real line of the mirror model. TBA equations for string excited states in the $\sl(2)$ sector have been studied along these lines in \cite{GKV09b}-\cite {Frolov:2010wt}.
In general, Y-functions have quite intricate analytic properties which are currently under investigation \cite{AFS09}, \cite{CFT10}-\cite{AFT11}.
\smallskip
In this paper we continue studies of the mirror TBA approach and make three new observations.
\smallskip
The first observation is on the origin of the large $J$ asymptotic solution. As was emphasized in \cite{AJK},
in the large $J$ or small $g$ limit\footnote{Here $J$ is a charge of a string state and $g$ is the string tension which is related to the 't Hooft coupling
$\lambda$ as
$\lambda=4\pi^2 g^2$.
}, the leading exponential correction to energies of string states should be given by a proper generalization of
L\"uscher's formula \cite{Luscher85}. Such a generalization to the case of non-Lorentz invariant string sigma model and to string states containing many particles
was proposed in \cite{BJ08} and used there to compute the four-loop anomalous dimension of the Konishi operator\footnote{In the context of the string sigma model L\"uscher's approach received recently a considerable attention, see the review \cite{Janik:2010kd} and references therein. The four-loop result for the anomalous dimension of the Konishi operator agrees with the field-theoretic computation
\cite{Sieg,Vel}.}.
The corresponding energy correction is
given in terms of $Y_Q$-functions that are in turn expressed via transfer matrices $T_{Q,1}(v)$, see eq.(\ref{YQasympt}). These transfer matrices are associated with
a scattering matrix which scatters a mirror theory $Q$-particle bound state of rapidity $v$ with string theory fundamental particles. In what follows we refer to eq.(\ref{YQasympt})
as to the Bajnok-Janik formula.
We further recall that L\"uscher's formulae provide an approximation to the exact TBA equations when $Y_Q$-functions are small. Indeed, recently a perfect agreement has been found between
L\"uscher's formulae at five loops \cite{BJ09,LRV09} and the corresponding predictions of the mirror TBA \cite{AFS10,BH10a}.
\smallskip
In addition to the main $Y_Q$-functions, the TBA equations also involve $Y_{\pm}$-, $Y_{M|vw}$- and $Y_{M|w}$-functions, as implied by the string hypothesis
\cite{AF09b}. Thus, to know the whole asymptotic solution, one has to also find the asymptotic expressions for $Y_{\pm}$, $Y_{M|vw}$ and $Y_{M|w}$.
In this paper we show that the corresponding expressions immediately follow from the Bajnok-Janik formula and the Y-system.
We recall that the Y-system is a set of functional relations between Y-functions which is obtained from the ground state TBA equations by applying a certain projection operation \cite{ZY}.
In the context of the string sigma model the corresponding Y-system and the asymptotic solution were
conjectured in \cite{GKV09}. The emphasis of our consideration here is on the fact that the asymptotic solution is {\it derivable} in a straightforward manner from the Y-system and
the Bajnok-Janik formula, see Figure 1. In particular, in the process of the derivation, the Bazhanov-Reshetikhin formula \cite{BR} and Hirota equations \cite{Hirota}
relating various transfer matrices emerge naturally.
\begin{figure}
\begin{center}
\includegraphics[width=0.9\textwidth]{TBA-Y.pdf}
\end{center}
\caption{The ground state TBA equations imply the Y-system. The Bajnok-Janik formula together with the Y-system
leads to determination of the whole asymptotic solution. The ground state TBA equations together with the asymptotic solution allow one to engineer excited states TBA equations via the contour deformation trick.}
\label{fig:TBA-Y}
\end{figure}
Our second observation concerns the construction of excited TBA equations beyond the $\sl(2)$ sector. As was mentioned above, obtaining excited state TBA equations requires the detailed knowledge of the asymptotic solution. For generic states, the asymptotic solution is constructed in terms of transfer matrices that in addition to physical momenta
$p_1,\ldots , p_N$ of string theory particles also involve auxiliary variables (roots) which satisfy the so-called auxiliary Bethe equations. We argue that for $J$ finite, a physical state
is completely characterized by a set of charges it carries under the global symmetry group and by a set of momenta
$p_1,\ldots , p_N$; auxiliary roots, as well as their Bethe equations, are invisible in the mirror TBA approach. Our arguing is based on the fact that all the transfer matrices and, therefore, the asymptotic Y-functions do not exhibit any singular behavior at locations of auxiliary Bethe roots. These roots satisfy auxiliary Bethe equations which guarantee regularity of transfer matrices. As a matter of fact, it is the main rational behind the {\it analytic} Bethe ansatz \cite{Kuniba2} that Bethe equations are derivable from the requirement of analyticity of the corresponding transfer matrices.
Of course, the presence of auxiliary Bethe roots affects analytic properties of Y-functions, but only in an indirect way. More precisely, as a first step towards constructing excited states TBA equations, auxiliary roots corresponding to a given asymptotic state
must be solved in terms of physical momenta and
substituted into asymptotic Y-functions, so the later become functions of $p_1,\ldots, p_N$ and of rapidity $v$. Further, the physical momenta are determined from the main Bethe equations.
Finally, upon substituting $p_1,\ldots, p_N$ into asymptotic Y-functions,
one has to determine their analytic properties as functions of $v$ and use them to engineer exact TBA equations via the contour deformation trick. We also point out, that the main (asymptotic)
Bethe equation for some $p_k$ implies the condition $Y_{1*}^o(p_k)=-1$, where $Y_{1*}^o$ is an asymptotic $Y_1$-function analytically continued to the string theory region. For finite $J$
these asymptotic conditions are replaced by {\it exact} Bethe equations $Y_{1*}(p_k)=-1$, where $Y_{1*}$ is an exact Y-function \cite{BLZe,DT96}. However, there are no any ``exact equations" for determining auxiliary
Bethe roots, such equations are not there already for the asymptotic solution because of the above-mentioned analyticity of the asymptotic Y-functions.
\smallskip
Our third observation deals with the issue of the ${\rm PSU}(2,2|4)$ symmetry. This symmetry was an important ingredient for conjecturing the all-loop asymptotic Bethe ansatz \cite{BS,BES} based on the $\alg{psu}(2|2)\oplus \alg{psu}(2|2)$-invariant S-matrix \cite{B05}. In the present paper
we argue that ${\rm PSU}(2,2|4)$ is automatically built in into the mirror TBA approach, for a simple reason of being also the symmetry of the
asymptotic solution. It appears that asymptotic Y-functions are merely invariant under the action of supersymmetry generators, {\it i.e.}, they are the same for any member of a given superconformal multiplet. Since the asymptotic solution is used to build exact TBA equations, the latter should be associated to the whole superconformal multiplet as well. We also point out, that the relation $L_{TBA}=J+2$ between the TBA length parameter $L_{TBA}$ and the charge $J$
established in our previous work \cite{AFS09} has a interesting interpretation -- $L_{TBA}$ equals to the maximal $J$-charge in the supersymmetry multiplet under consideration.
\smallskip
The paper is organized as follows. In section 2 we give a derivation of the asymptotic solution from the Y-system and the Bajnok-Janik formula.
We further explain how to engineer excited states TBA equations by means of the contour deformation trick. We also discuss the issue of auxiliary Bethe roots.
In section 3 we discuss implementations of the ${\rm PSU}(2,2|4)$ symmetry in the mirror TBA approach. In two appendices we collect the ground state TBA equations and discuss in detail duality transformation for transfer matrices, whose explicit expressions in various
gradings are also provided.
\section{TBA equations and asymptotic solution}
The mirror TBA approach is based on a bold assumption that excited states TBA equations have the universal form for any state. The TBA equations\footnote{For reader's convenience, the ground state TBA equations and following from them Y-system equations are collected in appendix \ref{appYsys}.} may differ from each other only by a choice of integration contours of convolution terms, and by analytic properties of Y-functions which determine driving terms in the TBA equations once the integration contours are taken back to the real line of the mirror model. The driving terms are of the form $\pm\log S(u_*,v)$ where $u_*$ is either a zero, pole or $-1$ of a Y-function. One can check that all the driving terms appearing via the procedure are annihilated by the operator $s^{-1}$: $\log S\cdot s^{-1}(u_*,v)=0$ for $|v|\le2$, and, therefore,
Y-functions which solve the excited states TBA equations also solve the associated Y-system equations.\footnote{The inverse is not true. Most solutions of the Y-system do not solve the TBA equations.}
\subsection*{ Asymptotic solution}
For large $J$ the energy of a string state can be found by solving the Bethe-Yang (BY) equations and taking into account the leading L\"uscher's exponential corrections. The BY equations depend on momenta of fundamental particles the string state is composed of (which could be complex if there are bound state particles), and on auxiliary roots. The auxiliary roots, however, are functions of the momenta, and, as a result, an $N$-particle string state is completely characterized by the momenta $p_k$ or, equivalently, by the $u$-plane rapidity variable $u_k$, or by the $z$-torus variables $z_k$. Moreover, the imaginary parts of the momenta are determined by their real parts, and, therefore, the number of independent parameters is equal to the number of fundamental and bound state particles the string state is made of.
\smallskip
Then, at large $J$ the $Y_Q$-functions become exponentially small
and can be written in terms of the $\alg{su}(2|2)$ transfer matrices defined in appendix \ref{appT} as follows \cite{BJ08}
\begin{eqnarray}\label{YQasympt}
Y_Q^o(v)&=&{\Upsilon}_Q(v)\,T_{Q,-1}(v)\,T_{Q,1}(v)\,, ~~~\\\nonumber
{\Upsilon}_Q(v)&=&{e^{-J\widetilde{\cal E}_Q(v) }\over \prod_{i=1}^{N} S_{\sl(2)}^{1_*Q}(u_i,v)}=e^{-J\widetilde{\cal E}_Q(v)}\, \prod_{i=1}^{N} S_{\sl(2)}^{Q1_*}(v,u_i)\,, ~~~\end{eqnarray}
where $v$ is the rapidity variable of the mirror $v$-plane and $\widetilde{\cal E}_Q$ is the energy of a mirror $Q$-particle. Here $S_{\sl(2)}^{1_*Q}$ denotes the $\sl(2)$-sector S-matrix with the first and second arguments in the string and mirror regions, respectively. The ${\rm AdS}_5\times {\rm S}^5$ string world-sheet S-matrix is a tensor product of two $\alg{su}(2|2)$-invariant ones, and this results in the appearance of two (in general different) transfer matrices
$T_{Q,\pm1}(v)$ in \eqref{YQasympt}. The transfer matrices
$T_{Q,\pm1}$ obviously depend on $u_k$, and the normalization of $T_{1,\pm1}$
is chosen so that if $u_k$ solve the BY equations then $ Y_{1_*}^o(u_k)=-1$.
\smallskip
The asymptotic $Y_Q$-functions \eqref{YQasympt} together with the Y-system equations completely fix the form of all the other Y-functions in the large $J$ limit. Indeed, we first notice that for $|v|<2$ the prefactors ${\Upsilon}_Q$ in
\eqref{YQasympt} satisfy the equations \cite{AF09d}
\begin{eqnarray}\label{dL}
{{\Upsilon}_{Q}^{+}\, {\Upsilon}_{Q}^{-} = {\Upsilon}_{Q-1}{\Upsilon}_{Q+1}} \,.
\end{eqnarray}
Thus, for large $J$ the Y-system equation \eqref{YQ1} takes the form
\begin{eqnarray}
{T_{1,-1}^+\,T_{1,-1}^-\over T_{2,-1}}\,
{T_{1,1}^+\,T_{1,1}^-\over T_{2,1}}
&=& {\left(1 - {1\over Y_{-}^{(-)}} \right)
\left(1 - {1\over Y_{-}^{(+)}} \right) }\,,~~~~~~
\end{eqnarray}
and one finds the asymptotic solution for $Y_-$-functions
\begin{eqnarray}\label{Ymasym}
Y_{-}^{(\pm)}=-{T_{2,\pm1}\over T_{1,\pm2}}\,,\quad T_{1,\pm2}\equiv T_{1,\pm1}^+\,T_{1,\pm1}^--T_{2,\pm1}\,.
\end{eqnarray}
In the same way eq.\eqref{YQQ} allows one to get the asymptotic solution for $Y_{Q|vw}$-functions
\begin{eqnarray}
Y_{Q|vw}^{(\pm)}&=&{T_{Q,\pm1}T_{Q+2,\pm1}\over T_{Q+1,\pm2}}\,,\quad
T_{Q+1,\pm2}\equiv T_{Q+1,\pm1}^+\,T_{Q+1,\pm1}^- -T_{Q,\pm1}T_{Q+2,\pm1}\,.\label{Yvwasym}
\end{eqnarray}
Next, moving to eq.\eqref{Ym} for $Y_-$, one determines the asymptotic
$Y_{1|w}$-functions
\begin{eqnarray}\label{Ywasym}
Y_{1|w}^{(\pm)}(v) &=&
{T_{1,\pm2}^+T_{1,\pm2}^--T_{2,\pm2}\over T_{2,\pm2}}={T_{1,\pm1}T_{1,\pm3}\over T_{2,\pm2}} \,,\quad
T_{1,\pm3} \equiv \left|
\begin{array}{ccc}
T_{1,\pm1}^{++}&1&0\\
T_{2,\pm1}^+&T_{1,\pm1}&1\\
T_{3,\pm1}&T_{2,\pm1}^-&T_{1,\pm1}^{--}
\end{array}
\right|.~~~~~~
\end{eqnarray}
Finally, eq.\eqref{Yvw1} for $Y_{1|vw}$-functions and
eqs.\eqref{Yw1} and \eqref{YwM} for $Y_{Q|w}$-functions allow one to find
$Y_{+}$ and the remaining $Y_{Q|w}$-functions
\begin{eqnarray}\label{Ypasym}
Y_{+}^{(\pm)} =-\frac{T_{2,\pm3}T_{2,\pm1}}{T_{3,\pm2}T_{1,\pm2}}\,,\quad Y_{Q|w}^{(\pm)} =\frac{T_{1,\pm Q\pm2}T_{1,\pm Q}}{T_{2,\pm Q\pm1}}\,,
\end{eqnarray}
where the transfer matrices $T_{a,\pm s}$ are expressed through $T_{a,\pm1}$ by means of the Bazhanov-Reshetikhin formula \cite{BR}
\begin{eqnarray}\label{BRf}
T_{a,\pm s}(v)=\hbox{det}_{1\leq i,j\leq
s}T_{a+i-j,\pm 1}\Big(v+\frac{i}{g}(s+1-i-j)\Big)\, .
\end{eqnarray}
Here one assumes that $T_{a,\pm1}$ satisfy
the conditions: $T_{0,\pm1} =1$ and $T_{a<0,\pm1}=0$.
\smallskip
It is worth mentioning that by construction the asymptotic T-functions satisfy the T-system or Hirota equations \cite{Hirota}
\begin{eqnarray}\label{Tasym}
T_{a,s}^+T_{a,s}^-=T_{a+1,s}T_{a-1,s}+T_{a,s+1}T_{a,s-1}\,,\quad a>0, s\neq 0\,,
\end{eqnarray}
where $T_{a,0}=1$. Clearly, the functions $T_{a,s}$ were constructed from $T_{a,1}$ in a purely algebraic manner, and it remains to be proven that they actually coincide with the transfer matrices corresponding to the representations $(a,s)$ of the centrally-extended $\alg{psu}(2|2)$. In the appendix we derive independently an explicit expression for $T_{1,s}$ which can be used to check numerically that it agrees with the one obtained from $T_{a,1}$ by means of \eqref{BRf}.
\smallskip
Obviously, in the large $J$ limit the negative $s$ T-functions are decoupled from the positive $s$ ones. For finite $J$ one takes the exact $Y_Q$-functions to be of the form \cite{Kuniba1,Kuniba2}
\begin{eqnarray}\label{YQexact}
Y_Q&=&{\Upsilon}_Q\,{T_{Q,-1}\,T_{Q,1}\over T_{Q-1,0}T_{Q+1,0}}\,, \end{eqnarray}
where $T_{Q,\pm1}$ reduce to the asymptotic ones and $T_{Q,0}$ reduce to 1 in the large $J$ limit. Then one assumes that in addition to the Hirota equations \eqref{Tasym} the T-functions satisfy the following equations\footnote{Since ${\Upsilon}_Q$ obey the discrete Laplace equations \eqref{dL}, they can be absorbed in the transfer matrices by a proper rescaling, see {\it e.g.} \cite{Ryo}. }
\begin{eqnarray}\label{Ts0}
T_{a,0}^+T_{a,0}^-=T_{a+1,0}T_{a-1,0}+{\Upsilon}_a\, T_{a,1}T_{a,-1}\,,\quad a>0\,,
\end{eqnarray}
and expresses all the other Y-functions in terms of $T_{a,\pm1}$ and $T_{a,0}$, or, equivalently, in terms of $T_{1,s}$.
Introducing the functions
\begin{eqnarray}\nonumber
&&Y_{Q,0}=Y_Q\,, \quad Y_{1,-1}=-\frac{1}{Y_-^{(-)}}\,, \quad Y_{1,1}=-\frac{1}{Y_-^{(+)}}\,,
\qquad Y_{2,-2}=-Y_+^{(-)}\, ,\quad Y_{2,2}=-Y_+^{(+)}\, ,\\
&&Y_{Q+1,-1}=\frac{1}{Y_{Q|vw}^{(-)}}\,,\quad
Y_{Q+1,1}=\frac{1}{Y_{Q|vw}^{(+)}}\, ,\qquad
Y_{1,-Q-1}=Y_{Q|w}^{(-)}\,,\quad
Y_{1,Q+1}=Y_{Q|w}^{(+)}\,,~~~~~
\end{eqnarray}
and performing the rescaling mentioned in footnote 8, one can write the relations in the standard form \cite{Kuniba1,Kuniba2}
\begin{eqnarray}
Y_{a,s}={T_{a,s-1}T_{a,s+1}\over T_{a-1,s}T_{a+1,s}}\,.
\end{eqnarray}
It is important to stress that the existence of T-functions $T_{a,\pm1}$ and $T_{a,0}$ which satisfy \eqref{Ts0} does not follow from anything we know about the ${\rm AdS}_5\times {\rm S}^5$ superstring or ${\cal N}=4$ SYM.
Fortunately, this assumption plays no role in constructing excited states TBA equations via the contour deformation trick.
\subsection*{ Engineering TBA equations for any state}
In this subsection we formulate the general strategy for engineering excited state TBA equations via the contour deformation trick.
It is basically the same as the one we used in \cite{AFS09} to analyze the states from the $\sl(2)$ sector. The main new ingredient now is that the description of a generic state
by means of the BY equations requires using not only momenta carrying Bethe roots but also a number of auxiliary ones.
\begin{enumerate}
\item Start with the BY equations in any grading, {\it e.g.}, the $\sl(2)$ or $\alg{su}(2)$ ones, and choose a charge $J$, a number of fundamental particles $N=K^{\rm I}$ and a set of auxiliary roots numbers $K^{\rm II}_\alpha$ and $K^{\rm III}_\alpha$ for the state/operator under consideration. All the other 4 charges of the state are determined by $J$ and the auxiliary roots numbers. The canonical dimension of the primary operator or the energy of the string state at $g=0$ is given by
$E_0=\Delta_0 = J+N$.
Solve the BY equations and choose from many solutions the one which
corresponds to the state of interest.
This state is characterized by a definite set of
$g$-dependent momenta, and by the set of auxiliary roots numbers $K^{\rm II}_\alpha$ and $K^{\rm III}_\alpha$. Auxiliary roots are completely fixed by the momenta, and play no independent role in the description of the state.
In practice a solution of the BY equations can be found only numerically for small $g$.
\item Compute asymptotic T- and Y-functions by using the set of momenta and auxiliary roots for the state, and find the location of zeroes and poles of
$1+Y_{a,s}$ and $Y_{a,s}$ functions in the mirror and string regions. The asymptotic solution can be trusted for finite $J$ and small $g$.
Note that the grading of the transfer matrices should match the grading of the BY equations used.
\item Choose integration contours and
engineer TBA equations for the state so that the asymptotic TBA equations obtained by dropping the terms with $\log(1+Y_Q)$ are solved by the asymptotic solution for Y-functions.
\item The exact momenta of fundamental particles are found from the exact Bethe equations $Y_{1*}(p_k)=-1$ which are derived by analytically continuing the excited state TBA equation for $Y_1$.
\end{enumerate}
Following the steps one can write down (at least in principle) excited states TBA equations for an arbitrary string state or ${\cal N}=4$ primary operator.
Some important comments are in order.
\begin{itemize}
\item There are no equations to determine the exact location of the auxiliary roots. As we discuss in the next subsection, Y-functions are regular at the locations of auxiliary roots, and do not have neither zeroes, poles or $-1$'s there.
\item Asymptotic Y-functions have zeroes, poles or $-1$'s at other locations, and exact Y-functions have similar properties too.
Exact locations of these zeroes, poles and $-1$'s are determined by analytically continuing the TBA equations for corresponding Y-functions to their (approximate) locations, and setting the Y-functions there to $0\,,\ \infty$ or $-1$.
\item Some zeroes, poles or $-1$'s of asymptotic Y-functions might be spurious and be absent in exact Y-functions.
For example, the analysis of a state from the $\alg{su}(2)$-sector composed of a fundamental particle and a two-particle bound state shows \cite{AFT11} that auxiliary Y-functions may have some zeroes, poles or -1's whose contributions to the TBA equations should not be included. These spurious zeroes, poles or $-1$'s are outside the physical strip and related to asymptotic $Y_\pm$ functions and probably to $T_{2,3}$ and $T_{3,2}$ which are on the boundary of the T-hook.
\item The energy spectrum computed by using the asymptotic
Bethe equations is very degenerate. Most of this degeneracy is however lifted as soon as the leading exponential corrections are taken into account.
As an example, let us consider the states which have $K^{\rm I}=4$, $K^{\rm II}_-+K^{\rm II}_+=2$ and
$K^{\rm III}_\alpha=0$. There are obviously three possible $y$-root distributions:
(i) $K^{\rm II}_-=0, K^{\rm II}_+=2$, (ii) $K^{\rm II}_-=2, K^{\rm II}_+=0$, (iii) $K^{\rm II}_-=1, K^{\rm II}_+=1$.
Since $K^{\rm III}_\alpha=0$ all the $y$-roots satisfy one and the same equation \eqref{BA1}. There are two solutions, $y_1$ and $y_2$, to this equation not equal to $0$ or $\infty$. Let one of the $y$-roots be equal to $y_1$, and the other to $y_2$.
The next step is to find a solution to the main Bethe equation \eqref{BAmain}. This equation involves all the $y$-roots, and, therefore, the solution is the same for any of the three states. Thus, the asymptotic energies of these states are the same.
The states (i) and (ii) correspond to conjugate representations
and therefore have equal energies also for finite $J$ and $g$. On the other hand, computing the $Y_Q$-functions for the third state one finds that they are different from those for the first two states. Thus, already the leading exponential correction would lift the asymptotic degeneracy of the spectrum.
\end{itemize}
We conclude this discussion with a word of caution. The procedure described above has so far been tested only for states composed of fundamental particles with real momenta. If some of the momenta are complex, {\it e.g.} there are bound states,
then in the large $J$ limit some of the asymptotic $Y_Q$-functions constructed via eq.\eqref{YQasympt} develop singularities on the real line of the mirror theory. Therefore,
strictly speaking, eq.\eqref{YQasympt} does not provide a large $J$ asymptotic solution of TBA equations.
We believe however that for finite $J$ and small $g$ eq.\eqref{YQasympt} or its mild modification can still be used to construct TBA and exact Bethe equations.
This issue is currently under investigation \cite{AFT11}.
\subsection*{Auxiliary roots and T- and Y-functions}
In this subsection we discuss whether the locations of auxiliary roots $y_k^{(\alpha)}$ and $w_k^{(\alpha)}$ are encoded in any way in analytic properties of T- and Y-functions. A naive expectation would be that some of the auxiliary Y-functions should be equal to $-1$ just as $Y_{1_*}(u_k)=-1$ at the locations of the main Bethe roots. We will see that this is not the case and all T- and Y-functions are regular for any value of $v$ related to the auxiliary roots $y_k^{(\alpha)}\,, w_k^{(\alpha)}$ by the shifts of the form $in/g$ for integer $n$. Regularity of transfer matrices at locations of auxiliary Bethe roots is a well-known fact, and our discussion below is given mainly to illustrate our conclusions.
\medskip
It is sufficient to consider only the roots with $\alpha=+$ which will be denoted as $y_k\,, w_k$. For definiteness we discuss transfer matrices in the mirror $v$-plane. The same conclusions are reached by considering the analytic continuation of the transfer matrices to the string $u$-plane.
We begin the discussion with $T_{1,1}$. It is written in the form, see appendix \ref{appT}
\begin{eqnarray}\label{t11}
T_{1,1}(v)={\mathscr N}_1(v)\Omega_1(v)\, ,
\end{eqnarray}
where the normalization factor
\begin{eqnarray}
{\mathscr N}_1(v)=
\prod_{i=1}^{K^{\rm{II}}}{\textstyle{\frac{y_i-x^-}
{y_i-x^+}\sqrt{\frac{x^+}{x^-}}} }
\end{eqnarray}
is necessary to provide a proper normalization of $Y_1$-functions,
and
\begin{eqnarray}\label{eqn;FullEignvalue2}
\Omega_1(v)=1 &+& \prod_{i=1}^{K^{\rm{II}}}{\textstyle
\frac{v-\nu_i+\frac{i}{g}}{v-\nu_i-\frac{i }{g}}}
\prod_{i=1}^{K^{\rm{I}}} {\textstyle{\frac{(x^--x^-_i)(1-x^-
x^+_i)}{(x^+-x^-_i)(1-x^+
x^+_i)}\frac{x^+}{x^-} }}-\\
& - & \prod_{i=1}^
{K^{\rm{I}}}{\textstyle{\frac{x^+-x^+_i}{x^+-x^-_i}
\sqrt{\frac{x^-_i}{x^+_i}} }} \times\nonumber
\left\{\prod_{i=1}^{K^{\rm{III}}}{\textstyle{\frac{w_i-v
-\frac{2i}{g}}{w_i-v}}}+ \prod_{i=1}^{K^{\rm{II}}}{\textstyle
\frac{v-\nu_i+\frac{i}{g}}{v-\nu_i-\frac{i
}{g}}}\prod_{i=1}^{K^{\rm{III}}}{\textstyle{\frac{w_i-v
+\frac{2i}{g}}{w_i-v}}}\right\}\, \nonumber
\end{eqnarray}
is normalized so that $\Omega_1(u_k)=1$. This guarantees that the asymptotic $Y_Q$-functions satisfy the conditions $Y_{1_*}(u_k)=-1$ if the main Bethe roots $u_k$ solve the BY equations.
\smallskip
Since $\nu_k=y_k+1/y_k$, potential singularities of $T_{1,1}$ may occur either at $v=\nu_k\pm{i\over g}$ or at $v=w_k$.
One can readily show that $T_{1,1}$ is regular at $v=w_k$ due to the auxiliary equations for the roots $w_k$.
As to $v=\nu_k\pm{i\over g}$,
there are two cases $y_k=x(\nu_k)$ and $y_k=1/x(\nu_k)$ to be discussed. In the first case one can easily see that a potential pole at $v=\nu_k+{i\over g}$ cancels out due to the normalization factor ${\mathscr N}_1$, and $T_{1,1}(v)$ is regular at $v=\nu_k+{i\over g}$. Then, one can check that the potential pole in the normalization factor at $v=\nu_k-{i\over g}$ cancels out due to the auxiliary Bethe equations for $\nu_k$. In the second case the normalization factor is regular for any $v$ but there still exists a potential pole at $v=\nu_k+{i\over g}$. A careful analysis of the residue of $T_{1,1}$ at $v=\nu_k+{i\over g}$ shows that it vanishes due to the auxiliary Bethe equations for $\nu_k$, and, therefore $T_{1,1}$ is regular at $v=\nu_k+{i\over g}$. One can also show in a similar way that $1/T_{1,1}$ is regular at all these points.
Thus, both $T_{1,1}$ and $1/T_{1,1}$ are regular for any value of $v$ related to the auxiliary roots $y_k\,, w_k$. The same consideration shows that any asymptotic transfer matrix $T_{a,1}$ and its inverse are also regular. Since $T_{a,s}$ are expressed via $T_{a,1}$ by means of the BR formula, we conclude that any $T_{a,s}$ is regular for any value of $v$ related to the auxiliary roots $y_k\,, w_k$. One can also check that $1/T_{a,s}$ is regular for these values of $v$.
\smallskip
The regularity of $T_{a,s}$ and $1/T_{a,s}$ immediately implies that no Y-function can have either zero or pole or be equal to $-1$ at these $y_k\,, w_k$ related locations. Hence, we are to conclude that there are no exact Bethe equations for auxiliary roots, and they are just invisible in the TBA equations. Another evidence why this should be the case is that the number of auxiliary $y$-roots depends on the grading used, and, {\it e.g.}, an $N$-particle state from the $\alg{su}(2)$ sector has no auxiliary roots if one uses the $\alg{su}(2)$ grading and $2N$ $y$-roots if one uses the $\sl(2)$ grading. The asymptotic Y-functions are independent of the grading because as we show in appendix \ref{appT} the transfer matrices in the $\sl(2)$ and $\alg{su}(2)$ grading can be obtained one from another by a duality transformation, and therefore, they are just equal on the solutions of the corresponding auxiliary Bethe equations. Thus, it would be unclear why one should have some extra equations for the auxiliary roots if one uses the $\alg{su}(2)$ grading.
\smallskip
To summarize, a physical state is completely characterized by a set of momenta of particles it is made of, and a set of auxiliary roots numbers $K^{\rm II}_\alpha$ and $K^{\rm III}_\alpha$, while the auxiliary roots are definite functions of the momenta. As a result, the eigenvalues of transfer matrices and Y-functions are also determined by the values of the momenta, and the location of the auxiliary roots is not reflected in their analytic properties.
\section{Implementation of the ${\rm PSU}(2,2|4)$ symmetry}\label{sec:psu}
Here we discuss the issue of the ${\rm PSU}(2,2|4)$ symmetry
in the mirror TBA approach. We start with the asymptotic Bethe Ansatz equations in the $\sl(2)$ grading \cite{BS}.
The main Bethe equations have the form
\begin{eqnarray} \label{BAmain}
&&1=e^{iJp_k}\prod_{l\neq
k}^{K^{\rm I}}S_{\sl(2)}(u_k,u_l)\prod_{l=1}^{K^{\rm II}_{-}}\frac{x_k^- -
y_l^{(-)} }{x_k^+-y_l^{(-)}}\sqrt{\frac{x_k^+}{x_k^-}}\,
\prod_{l=1}^{K^{\rm II}_{+}}\frac{x_k^- - y_l^{(+)}
}{x_k^+-y_l^{(+)}}\sqrt{\frac{x_k^+}{x_k^-}}\, .
\end{eqnarray}
These equations are supplied with auxiliary Bethe equations for the roots
$y^{(\alpha)}$ and $w^{(\alpha)}$, $\alpha=\pm $,
\begin{eqnarray}
\label{BA1}
\prod_{i=1}^
{K^{\rm{I}}}\frac{y_k^{(\alpha)}-x^-_i}{y_k^{(\alpha)}-x^+_i}\sqrt{\frac{x^+_i}{x^
-_i}} &=&\prod_{i=1}^{K^{\rm{III}}_{\alpha}}
\frac{w_i^{(\alpha)}-\nu_k^{(\alpha)}-\frac{i}{g}}{w_i^{(\alpha)}-\nu_k^{(\alpha)}+\frac{i}{g}}\, ,
\\ \label{BA2}
\prod_{i=1}^{K^{\rm II}_{\alpha}}\frac{w_k^{(\alpha)}-\nu_i^{(\alpha)}+\frac{i}{g}}
{w_k^{(\alpha)}-\nu_i^{(\alpha)}-\frac{i}{g}}
&=&-\prod_{i=1}^{K^{\rm III}_{\alpha}}\frac{w_k^{(\alpha)}-w_i^{(\alpha)}+\frac{2i}{g}}{w_k^{(\alpha)}-w_i^{(\alpha)}-\frac{2i}{g}}\, .
\end{eqnarray}
Here we introduced a concise notation $\nu_k^{(\alpha)}=y_k^{(\alpha)}+\frac{1}{y_k^{(\alpha)}}$.
Solutions are therefore characterized by the
following five excitation numbers
$$
(K^{\rm III}_-,K^{\rm II}_-,K^{\rm I},K^{\rm II}_+,K^{\rm III}_+)\, .
$$
The number $K^{\rm I}$ is a number of momentum-carrying particles, while $K^{\rm II}_{\alpha}$ and
$K^{\rm III}_{\alpha}$ give the weights of four ${\rm SU}(2)$ subgroups which represent a manifest symmetry of the string sigma model in the light-cone gauge.
The ${\rm SU}(4)$ weights $[q_1,p,q_2]$ and the spins $[s_1,s_2]$
of the corresponding excited state are \begin{eqnarray}
\begin{array}{lll}
q_1=K^{\rm II}_--2K^{\rm III}_- & ~~~~~~~~~~ & s_1=K^{\rm I}-K^{\rm II}_- \\
p=J-\frac{1}{2}(K^{\rm II}_-+K^{\rm II}_+)+K^{\rm III}_-+K^{\rm III}_+ & & s_2=K^{\rm I}-K^{\rm II}_+
\\
q_2=K^{\rm II}_+-2K^{\rm III}_+ & &
\end{array}
\end{eqnarray}
Instead of weights of $\alg{su}(4)$
one can use the weights $(J_1,J_2,J_3)$ of ${\rm SO}(6)$ and the relation between the two is
\begin{eqnarray} J\equiv J_1=\frac{1}{2}(q_1+2p+q_2) \, , ~~~~~~~
J_2=\frac{1}{2}(q_1+q_2)\, , ~~~~~~
J_3=\frac{1}{2}(q_2-q_1) \, . \end{eqnarray}
Now we are ready to discuss the realization of the ${\rm PSU}(2,2|4)$
symmetry on asymptotic solutions. First, we recall that this symmetry can be consistently realized only on physical states \cite{BS}, {\it i.e.} those which satisfy the level-matching condition. As can be seen from eqs.(\ref{BA1})
and ({\ref{BA2}}), such states might have a certain number of roots $\nu^{(\alpha)}$
and $w^{(\alpha)}$ located at infinity\footnote{Obviously, the corresponding roots $y^{(\alpha)}$ can be located at either zero or infinity.}.
Second, the states belonging to the same multiplet must have the one and the same anomalous dimension and canonical dimensions which might differ from each other by a half-integer only.
For the light-cone string sigma model the dispersion relation is
\begin{eqnarray}
E=J+\sum_{k=1}^{K^{\rm I}}\sqrt{1+4g^2\sin^2\frac{p_k}{2}}\, .
\end{eqnarray}
For $g=0$ the last formula gives the canonical dimension of the corresponding gauge theory operator, while the difference $E-J-K^{\rm I}$ corresponds to the anomalous dimension. Thus, adding
particles with zero momentum changes the canonical dimension, but does not influence the
anomalous one and therefore should correspond to passing to a different member of a supersymmetry multiplet. Note that $p=0$ corresponds to $u=\infty$ in the standard
$u$-plane parametrization of the momentum.
\smallskip
What was said above motivates the following treatment of the ${\rm PSU}(2,2|4)$
symmetry on the asymptotic solutions. We assume that every multiplet of ${\rm PSU}(2,2|4)$ has a {\it unique
regular} representative among the solutions of the Bethe ansatz
equations. Regularity means that all Bethe roots
$(\vec{u},\vec{\nu}^{(\alpha)},\vec{w}^{(\alpha)})$ of the
corresponding solution are finite. The regular representative\footnote{Which primary state is
regular depends on the grading chosen.} is a primary state of
the bosonic subgroup ${\rm SU}(2)\times {\rm SU}(2)\times {\rm
SU}(2)\times {\rm SU}(2)$ and it carries excitation numbers $K^{\rm I}$,
$K^{\rm II}_{\alpha}$ and $K^{\rm III}_{\alpha}$. All the other states in the multiplet are obtained by adding roots at infinity. Obviously, the regular representative has a minimal number of Bethe roots.
\smallskip
In spite of the fact that irregular states in a supermultiplet have excitation numbers different from the regular ones, they must be nevertheless described by the same Bethe equations as for the regular state. From the main equation (\ref{BAmain}),
one sees that adding a particle with $p=0$ does not influence this equation at all, since for this value of momentum $x^+/x^-=1$. The situation is different for $y$-roots: Adding a single $y=\infty$ root will leave eq.(\ref{BAmain}) invariant only
if the original charge $J$ will be replaced as $J\to J-\frac{1}{2}$. Oppositely, adding a single $y=0$ root requires a shift $J\to J+\frac{1}{2}$. Hence, there is a correlation between the number of irregular $y$-roots and the charge $J$ of the corresponding state.
Finally, we note that adding any number of irregular $y$- or $w$-roots does not influence the auxiliary Bethe equations for a regular physical state.
\smallskip
It is of interest to determine the Bethe root content and weights (Dynkin labels) of the superconformal primary state corresponding to a given regular state. The structure of a generic superconformal multiplet is such that primary states with respect to the conformal group are in correspondence with states which are obtained by acting on the superconformal primary with
all possible combinations of 16 Poincar\'e supercharges\footnote{Conformal supercharges annihilate a superconformal primary state.}. A superconformal primary state has obviously the lowest canonical dimension in the multiplet.
To understand the action of the Poincar\'e supercharges, it is convenient to split them into two groups. We recall \cite{AFPZ} that in the uniform light-cone gauge $x_+=\tau$, where $\tau$ is the world-sheet time, the supercharges are naturally divided with respect to their dependence on
the unphysical field $x_-$: they are either {\it kinematical} (independent of $x_-$) or {\it dynamical} (dependent on $x_-$). Kinematical supercharges do depend on $x_+$ and, for this reason, they do not commute with the world-sheet Hamiltonian
${\rm H}=E-J$, while dynamical supercharges are independent of $x_+$ and commute with ${\rm H}$. In the Tables 1 and 2 we presented the weights of the kinematical and dynamical supercharges, respectively, as well as their action on the excitation numbers.
\vskip 0.5cm
\begin{center}
{\small
{
\renewcommand{\arraystretch}{1.5}
\renewcommand{\tabcolsep}{0.2cm}
\begin{tabular}{||c|l|c|c|c|c|}
\hline Charge & Weights & $\Delta K^{\rm II}_-$ & $\Delta K^{\rm II}_+$ & $\DeltaK^{\rm III}_-$ & $\DeltaK^{\rm III}_+$\\
\hline $Q^3_{\alpha} $ & $[0,-1,1]_{(\pm \frac{1}{2},0)}$ & $0_{+\frac{1}{2}}$, $2_{-\frac{1}{2}}$ & $1_{+\frac{1}{2}} $, $1_{-\frac{1}{2}}$ & $0_{+\frac{1}{2}}$, $1_{-\frac{1}{2}}$ & $ 0_{\pm \frac{1}{2}} $ \\
\hline $Q^4_{\alpha} $ & $ [0,0,-1]_{(\pm\frac{1}{2},0)}$ & $0_{+\frac{1}{2}}$, $2_{-\frac{1}{2}}$ & $1_{\pm \frac{1}{2}} $ & $0_{+\frac{1}{2}}$, $1_{-\frac{1}{2}}$ & $ 1_{\pm \frac{1}{2}} $ \\
\hline $\bar{Q}_{1\dot{a}}$ & $[-1,0,0]_{(0,\pm \frac{1}{2})}$ & $1_{\pm \frac{1}{2}} $ & $0_{+\frac{1}{2}}$, $2_{-\frac{1}{2}}$ & $ 1_{\pm \frac{1}{2}} $ & $0_{+\frac{1}{2}}$, $1_{-\frac{1}{2}}$ \\
\hline $\bar{Q}_{2\dot{a}}$ & $ [1,-1,0]_{(0,\pm \frac{1}{2})}$ & $1_{+\frac{1}{2}} $, $1_{-\frac{1}{2}}$ & $0_{+\frac{1}{2}}$, $2_{-\frac{1}{2}}$ &$ 0_{\pm \frac{1}{2}} $ & $0_{+\frac{1}{2}}$, $1_{-\frac{1}{2}}$ \\
\hline
\end{tabular}
}
}
\vspace{0.5cm}
\parbox{13cm}
{\small Table 1. Kinematical Poincar\'e supercharges. These supercharges decrease $J$ by $-1/2$ and increase $K^{\rm I}$ by 1, they never decrease $K^{\rm II}_{\alpha}$ and $K^{\rm III}_{\alpha}$. Here $\Delta K^{\rm II}_{\alpha}$ and $\Delta K^{\rm III}_{\alpha}$ denote the change of the corresponding excitation numbers under the action of a supercharge.}
\end{center}
\vspace{0.3cm}
Concerning kinematical supercharges, one can see that any such supersymmetry generator raises the number of zero momentum particles by one and, on the other hand, never lowers all the other excitation numbers.
More specifically, acting with one of these supercharges adds in total either one or three irregular $y$-roots, depending on a supercharge under consideration.
Thus, when applied to a superconformal primary state
these supercharges always add further irregular roots and, therefore, can never generate the corresponding regular state.
\smallskip
Note that we can easily determine the location of irregular $y$-roots created by the action of a kinematical supercharge.
Since acting with such a charge leads to the shift $J\to J-\frac{1}{2}$, we conclude from our consideration of the main Bethe equation
that for a charge generating a single $y$-root
the latter must be at infinity, while for one generating three $y$-roots, two of them must be at infinity and one at zero.
\vskip 0.5cm
\begin{center}
{\small
{
\renewcommand{\arraystretch}{1.5}
\renewcommand{\tabcolsep}{0.2cm}
\begin{tabular}{||c|l|c|c|c|c|}
\hline Charge & Weights & $\DeltaK^{\rm II}_-$ & $\DeltaK^{\rm II}_+$ & $\DeltaK^{\rm III}_-$ & $\DeltaK^{\rm III}_+$\\
\hline $Q^1_{\alpha} $ & $[1,0,0]_{(\pm \frac{1}{2},0)}$ & $-1_{+\frac{1}{2}}$, $1_{-\frac{1}{2}}$ & $0_{\pm \frac{1}{2}} $ & $-1_{+\frac{1}{2}}$, $0_{-\frac{1}{2}}$ & $ 0_{\pm \frac{1}{2}} $ \\
\hline $Q^2_{\alpha} $ & $ [-1,1,0]_{(\pm \frac{1}{2},0)}$ & $-1_{+\frac{1}{2}}$, $1_{-\frac{1}{2}}$ & $0_{\pm \frac{1}{2}} $ & $0_{+\frac{1}{2}}$, $1_{-\frac{1}{2}}$ & $ 0_{\pm \frac{1}{2}} $ \\
\hline $\bar{Q}_{3\dot{a}}$ & $[0,1,-1]_{(0,\pm \frac{1}{2})}$ & $0_{\pm \frac{1}{2}} $ & $-1_{+\frac{1}{2}}$, $1_{-\frac{1}{2}}$ & $ 0_{\pm \frac{1}{2}} $ & $0_{+\frac{1}{2}}$, $1_{-\frac{1}{2}}$ \\
\hline $\bar{Q}_{4\dot{a}}$ & $ [0,0,1]_{(0,\pm \frac{1}{2})}$ & $0_{\pm \frac{1}{2}} $ & $-1_{+\frac{1}{2}}$, $1_{-\frac{1}{2}}$ & $ 0_{\pm \frac{1}{2}} $ & $-1_{+\frac{1}{2}}$, $0_{-\frac{1}{2}}$ \\
\hline
\end{tabular}
}
}
\vspace{0.5cm}
\parbox{13cm}
{\small Table 2. Dynamical Poincar\'e supercharges. These supercharges do not change the value of $K^{\rm I}$, but they raise both $J$ and the canonical dimension by $1/2$. There exists four supercharges which lower
$K^{\rm II}$ by 1.}
\end{center}
\vspace{0.3cm}
The situation with dynamical generators is a bit different. They do not change the value of $K^{\rm I}$, as they simultaneously raise $E$ and $J$ by $1/2$.
As one can see from Table 2, there are four supercharges which have non-negative excitation numbers, and, for the same reason as for kinematical supercharges,
they cannot generate a regular state from its superconformal primary. Any of the remaining four supercharges lowers $K^{\rm II}$ by one and leaves intact or lowers
$K^{\rm III}$ by one.
It is these generators we are most interested in, because their application to a superconformal primary state decreases the number of irregular roots and, therefore, can produce a regular state.
It is also clear that we should apply all these four generators to a superconformal primary to get the regular one, as in the opposite case, there remains a possibility to further lower the number of irregular roots.
Denote by $E_{hws}$ and $J_{hws}$ the charges of a superconformal primary, which is the highest weight state of a supersymmetry multiplet, and by $E_{reg}$ and $J_{reg}$ the charges of the corresponding regular state.
Hence, we conclude that\footnote{The reader can verify this formula for the case of the Konishi multiplet. We recall that the Konishi superconformal primary operator has the canonical dimension $\Delta=2$ and $J=0$. It has a regular descendent in the $\sl(2)$ sector
which has $\Delta=4$ and $J=2$.}
\begin{eqnarray}
\label{EJhws}
E_{hws}=E_{reg}-2\, ,~~~~~~ J_{hws}=J_{reg}-2\, .
\end{eqnarray}
Also, the relationship between the corresponding excitation numbers is
\begin{eqnarray}
K^{\rm I}_{reg}=K^{\rm I}_{hws}\, ,~~~~~
K^{\rm II}_{\alpha,reg}= K^{\rm II}_{\alpha, hws}-2 \, , ~~~~~ K^{\rm III}_{\alpha,reg}= K^{\rm III}_{\alpha, hws}-1\end{eqnarray}
for both $\alpha$. Our discussion of the main Bethe equation together with the relation $J_{hws}=J_{reg}-2$ implies that four irregular $y$-roots which distinguish a superconformal primary from its regular state must all be located at infinity.
\smallskip
Similar to what has been done for kinematical generators, we can now establish locations of irregular $y$-roots created by the action of dynamical generators. Since application of dynamical generators shifts $J\to J+\frac{1}{2}$,
four of them generate four roots at $y=0$, while the other four remove four roots at $y=\infty$. This leads to the following description of an arbitrary state generated by a superconformal primary $|{\rm hws}\rangle$
\begin{eqnarray}
\prod (Q_{-\infty}^{d})^{n_{\infty}^d}
(Q_{+0}^{d})^{n_{0}^d}(Q_{+\infty}^{k})^{n_{\infty}^k}(Q_{+2\infty,+0}^{k})^{n_{\infty,0}^k}|{\rm
hws}\rangle\, ,\end{eqnarray}
where the integers $n_{\infty}^d,\dots, n_{\infty,0}^k$ can take any value
from zero to four. Here the upper subscript $``d"$ or $``k"$ in the definition of a supercharge $Q$ specifies if it is dynamical or kinematical, respectively.
The subscript of $Q$ points the location and the number of $y$-roots with $+$ and $-$ sign signifying if the corresponding root is added or removed, respectively.
Taking into account how each supercharge
increases or decreases the value of $J$, one finds that such a state
has \begin{eqnarray}
J=J_{hws}+\frac{1}{2}(n_{\infty}^d+n_{0}^d-n_{\infty}^k-n_{\infty,0}^k)\,
\end{eqnarray}
and its energy is
\begin{eqnarray} E=E_{hws}
+\frac{1}{2}(n_{\infty}^d+n_{0}^d+n_{\infty}^k+n_{\infty,0}^k)\, .
\end{eqnarray}
Also, this state has the following number of main particles
\begin{eqnarray}
K^{\rm I} =K^{\rm I}_{reg}+n_{\infty}^k+n_{\infty,0}^k \, .\end{eqnarray}
As a check, we get
\begin{eqnarray}
E-J=E_{hws}&-&J_{hws}+n_{\infty}^k+n_{\infty,0}^k=\\ \nonumber
&=&E_{reg}-J_{reg}+ n_{\infty}^k+n_{\infty,0}^k=K^I_{reg}+n_{\infty}^k+n_{\infty,0}^k=K^{\rm I}\, ,\end{eqnarray}
{\it i.e.} as expected for the free dispersion relation for any member of the multiplet.
Further, we see that for the state under consideration a number of {\it irregular} roots, denoted by ${\cal K}^{\rm II}$, is
\begin{eqnarray}
{\cal K}^{\rm II}_{\infty}&=&4-n_{\infty}^d+n_{\infty}^k+2n_{\infty,0}^k \, ,\\
{\cal K}^{\rm II}_{0}&=&n_{0}^d+n_{\infty,0}^k\, \end{eqnarray}
so that the total number of $y$-roots is $K^{\rm II}=K^{\rm II}_{reg}+{\cal K}^{\rm II}_{0}+ {\cal K}^{\rm II}_{\infty}$.
This allows one to express $J$ in terms of $J_{reg}$ and
the number of irregular $y$-roots
\begin{eqnarray}
\label{JvsJhws}
J=J_{hws}+2+\sfrac{1}{2}({\cal K}^{\rm II}_{0}-{\cal K}^{\rm II}_{\infty})=J_{reg}+\sfrac{1}{2}({\cal K}^{\rm II}_{0}-{\cal K}^{\rm II}_{\infty})\, .\end{eqnarray}
Obviously, this formula is in complete agreement with the fact that the state we consider must have the same Bethe equations as the corresponding regular state.
\smallskip
Now we turn our attention to the large $J$ asymptotic solution of the mirror TBA equations.
The asymptotic $Y_Q$-functions for a state of charge $J$ and excitation numbers $K^{\rm I}$, $K^{\rm II}$ and $K^{\rm III}$ are given by eqs.(\ref{YQasympt}).
It is important to realize that this expression for $Y_Q$ is valid not only for a regular state, but for any state
in the corresponding superconformal multiplet. As we now show, all states in a multiplet have in fact the one and same asymptotic $Y_Q$-functions.
First, from the explicit expression (\ref{eqn;FullEignvalue1}) for transfer matrices $T_{Q,\pm 1}$, one can see that they remain unchanged if one adds any number of
infinite roots $u$ or $w^{(\alpha)}$. Second, if a state has a number of $y$-roots ${\cal K}^{\rm II}_{0}$ at zero and a number of $y$-roots ${\cal K}^{\rm II}_{\infty}$
at infinity, then the expression for transfer matrices shows that they are related to the transfer matrices of the corresponding regular state
in a very simple fashion
\begin{eqnarray}
T_{Q,+1}T_{Q,-1}=\Big(\frac{x^+}{x^-}\Big)^{\sfrac{1}{2}({\cal K}^{\rm II}_{\infty}-{\cal K}^{\rm II}_{0}) }T_{Q,+1}^{reg}T_{Q,-1}^{reg}\, .
\end{eqnarray}
Accordingly, the $Y_Q$-functions of this state can be written as
\begin{eqnarray}
Y_Q^o=\Big(\frac{x^+}{x^-}\Big)^{J-J_{reg}}\Big(\frac{x^+}{x^-}\Big)^{\sfrac{1}{2}({\cal K}^{\rm II}_{\infty}-{\cal K}^{\rm II}_{0}) }Y_Q^{o,reg}=Y_Q^{o,reg}\, ,\end{eqnarray}
where in the last formula we used eq.(\ref{JvsJhws}). Thus, for all states in a multiplet the $Y_Q^o$-functions are the same as for the regular state.
Since the knowledge of $Y_Q^o$ allows one to find all the other asymptotic Y-functions, we conclude that
all states in a superconformal multiplet must share the one and the same asymptotic solution. Moreover, since the excited states TBA equations are engineered
by using the analytic properties of the asymptotic Y-functions, we conclude that these equations are constructed not for a particular state but rather for a whole supersymmetry multiplet. This implies that the ${\rm PSU}(2,2|4)$ symmetry is in some sense built in into the mirror TBA approach, in a similar way as it is for
the asymptotic Bethe ansatz.
\smallskip
Now we point out an interesting interpretation of the TBA length parameter $L_{TBA}$. Applying all kinematical generators to a highest weight state of charge $J_{hws}$
, in a generic situation $J>3$, we obtain a state of the lowest $J$-charge, $J=J_{hws}-4$. Oppositely, acting on $|{\rm hws}\rangle$
with all dynamical generators produces a state of the highest $J$-charge, $J=J_{hws}+4$. Thus, the $J$-charge of any state in a generic multiplet obeys inequalities
\begin{eqnarray} J_{hws}-4\leq J\leq
J_{hws}+4=J_{reg}+2\, . \end{eqnarray}
As was found in \cite{AFS09}, the length parameter $L_{TBA}$ must be related to the $J$-charge of a regular state as
$$L_{TBA}=J_{reg}+2\,.$$
Thus, $L_{TBA}$ simply coincides with the maximal $J$-charge in a supersymmetry multiplet.
\smallskip
One remark is in oder. A generic supersymmetry multiplet is obtained by free action of 16 Poincar\'e supercharges on a highest weight state (a state of the lowest dimension in the multiplet) and arising representations of the maximal bosonic subgroup have Dynkin labels which are obtained by adding the weights of the corresponding supercharges to
Dynkin labels of the highest weight state. In the case where the resulting labels turn out to be negative (non-generic multiplets), special rules must be applied to find the corresponding Dynkin content\footnote{For ${\rm PSU}(2,2|4)$ these rules can be found, for instance, in \cite{Dolan:2002zh}. }. The Konishi multiplet, which has $J_{hws}=0$, provides an example of a non-generic multiplet; it has $2^{16}$ states which are organized in 532 representations of the maximal
bosonic subgroup of the conformal group. Nevertheless, certain formulas we discussed in this section remain valid for the Konishi multiplet, for instance,
the relations (\ref{EJhws}), and also $L_{TBA}=4$.
\section*{Acknowledgements}
We are grateful to Niklas Beisert for discussions concerning superconformal symmetry.
G.A. acknowledges support by the Netherlands Organization for Scientific Research (NWO) under the VICI grant 680-47-602.
The work of S.F. was supported in part by the Science Foundation Ireland under Grant 09/RFP/PHY2142.
\section{Appendix}
\subsection{ Simplified TBA equations and Y-system}\label{appYsys}
The simplified TBA equations for the ground state derived in \cite{AF09b,AF09d} can be written in the following form\footnote{We set the regularization parameters $h_\alpha$ to 0, where the parameter $\alpha$ takes the values $\pm$. In our previous papers $\alpha=1,2$.}
\noindent
$\bullet$\ $Q=1$-particle
\begin{eqnarray}
\nonumber
\log Y_{1}(v)&=&
\log\left(1-{1\over Y_{-}^{(-)}} \right)\left(1-{1\over Y_{-}^{(+)}} \right)Y_2\, \hat{\star}\, s- \log(1 + Y_{2})\star s \\\nonumber
&-& \log\left(1-{1 \over Y_{-}^{(-)}} \right)\left(1-{1 \over Y_{+}^{(-)}}\right)\left(1-{1 \over Y_{-}^{(+)}} \right)\left(1-{1 \over Y_{+}^{(+)}}\right)
Y_2^2\, \hat{\star}\, \check{K}\,\check{\star}\, s
\\\label{YforQ1}
&-& \log\left(1+Y_{Q} \right)\star \big( 2\check{K}_Q^\Sigma +
\check{K}_Q +\check{K}_{Q-2}\big)\,\check{\star}\, s +\log{Y_1}\star \check{K}_1\,\check{\star}\, s-L\, \check{\cal E}\,\check{\star}\, s\,.~~~~~~~
\end{eqnarray}
Here and in what follows we use the definitions and conventions from
\cite{AFS09}.
It has been found \cite{AFS09} that for all the excited states analyzed in the TBA approach the length parameter $L$ is related to the charge $J$ carried by a string state as $L=J+2\,.$
We argue in section \ref{sec:psu} that the relation between length and charge is universal and holds for any excited state with regular Bethe roots.
Assuming that $|v|\le2$ and acting on \eqref{YforQ1} by the operator $s^{-1}$, one derives the first Y-system equation
\begin{eqnarray}\label{YQ1}
{Y_{1}^{+}\, Y_{1}^{-}\over Y_{2}} &=& {\left(1 - {1\over Y_{-}^{(-)}} \right)
\left(1 - {1\over Y_{-}^{(+)}} \right)\over 1+Y_2 }\, ,\end{eqnarray}
where we use the notation $f^\pm=f(v\pm{i\over g}\mp i0)$ and take into account that for $|v|\le2$ only the first line in \eqref{YforQ1} contributes.
\bigskip
\noindent
$\bullet$\ $Q$-particles
\begin{eqnarray}
\log Y_{Q+1}&=&\log{\left(1 + {1\over Y_{Q|vw}^{(-)}} \right)\left(1 + {1\over Y_{Q|vw}^{(+)}} \right)\over (1 + {1\over Y_{Q} })(1 + {1\over Y_{Q+2} }) }\star s\label{YforQ}
\,,\quad Q\ge 1\,.~~~~~~~
\end{eqnarray}
This equations lead to the following Y-system equations valid for any $v$
\begin{eqnarray}\label{YQQ}
{Y_{Q+1}^{+}\, Y_{Q+1}^{-}\over Y_{Q}Y_{Q+2}} &=& {\left(1 + {1\over Y_{Q|vw}^{(-)}} \right)
\left(1 + {1\over Y_{Q|vw}^{(+)}} \right)\over \left( 1+Y_{Q} \right)\left( 1+Y_{Q+2} \right) } \,.
\end{eqnarray}
\bigskip
\noindent
$\bullet$\ $y$-particles, $\alpha=\pm$
\begin{eqnarray}
\label{Yfory1}
\log {Y_+^{(\alpha)} \over Y_-^{(\alpha)} }(v)&=&\log(1 + Y_{Q})\star K_{Qy}\,,~~~~~~~ \\
\label{Yfory2}
\log {Y_+^{(\alpha)} Y_-^{(\alpha)} }(v) &=&2\log{1 + Y_{1|vw}^{(\alpha)} \over 1 + Y_{1|w}^{(\alpha)} }\star s
\\\nonumber
&-&
\log\left(1+Y_Q \right)\star K_Q+ 2 \log(1 + Y_{Q})\star K_{xv}^{Q1} \star s\,.~~~~
\end{eqnarray}
The AdS/CFT Y-system is incomplete and contains equations only for $Y_-^{(\alpha)} $-functions.
Their derivation is not straightforward and can be found in \cite{AF09b}.
Assuming that $Y_\pm^{(\alpha)} $-functions are analytic in the vicinity of
the interval $|v|\le 2$, one gets
\begin{eqnarray}\label{Ym}
Y_{-}^{(\alpha)+}\,Y_{-}^{(\alpha)-} &=& {1+ Y_{1|vw}^{(\alpha)} \ov1+ Y_{1|w}^{(\alpha)}} \, {1\over 1+Y_1 }
\,.
\end{eqnarray}
This assumption is compatible with the kernel $K_{Qy}$ where the cut is chosen to be for $|v|\ge2$. One can, however, choose the cut to be for $|v|\le2$, which is in fact more natural if one thinks about Y-functions as being defined on a $z$-torus. To discuss the asymptotic solution it is easier to use the cut $|v|\ge2$.
\bigskip
\noindent
$\bullet$\ $M|vw$-strings: $\ M\ge 1\ $, $Y_{0|vw}=0$
\begin{eqnarray}\label{Yforvw}
\hspace{-0.3cm}\log Y_{M|vw}^{(\alpha)} (v)&=&- \log(1 + Y_{M+1})\star s~~~~~\\\nonumber
&+& \log(1 + Y_{M-1|vw}^{(\alpha)} )(1 + Y_{M+1|vw}^{(\alpha)} )\star s+\delta_{M1} \log{1-Y_-^{(\alpha)} \over 1-Y_+^{(\alpha)} }\,\hat{\star}\, s\,,~~~~~
\end{eqnarray}
and the Y-system equations
\begin{eqnarray}
\label{Yvw1}
Y_{1|vw}^{(\alpha)+}\, Y_{1|vw}^{(\alpha)-} &=& { 1+Y_{2|vw}^{(\alpha)} \over 1+Y_{2}} \, {1-Y_-^{(\alpha)}\over 1-Y_+^{(\alpha)}} \\
\label{YvwM}
Y_{M|vw}^{(\alpha)+}\, Y_{M|vw}^{(\alpha)-} &=& \left( 1+Y_{M-1|vw}^{(\alpha)} \right)\left( 1+Y_{M+1|vw}^{(\alpha)} \right)\, {1\over 1+Y_{M+1}}\,, \quad M\ge 2\,,~~~~~~~
\end{eqnarray}
\bigskip
\noindent
$\bullet$\ $M|w$-strings: $\ M\ge 1\ $, $Y_{0|w}=0$
\begin{eqnarray}\label{Yforw}
\log Y_{M|w}^{(\alpha)} = \log(1 + Y_{M-1|w}^{(\alpha)} )(1 + Y_{M+1|w}^{(\alpha)} )\star s
+\delta_{M1}\, \log{1-{1\over Y_-^{(\alpha)} }\over 1-{1\over Y_+^{(\alpha)} } }\,\hat{\star}\, s\,.~~~~~
\end{eqnarray}
and the Y-system equations
\begin{eqnarray}
\label{Yw1}
Y_{1|w}^{(\alpha)+}\,Y_{1|w}^{(\alpha)-} &=& \left( 1+Y_{2|w}^{(\alpha)} \right){1-{1\over Y_-^{(\alpha)}}\over 1-{1\over Y_+^{(\alpha)}} }\\
\label{YwM}
Y_{M|w}^{(\alpha)+}\,Y_{M|w}^{(\alpha)-} &=& \left( 1+Y_{M-1|w}^{(\alpha)} \right)\left( 1+Y_{M+1|w}^{(\alpha)} \right)\,, \quad M\ge 2\,,
\end{eqnarray}
Let us stress again that the equations above are valid only for $|v|\le2$. For other values of $v$ one should use an analytic continuation, and the resulting equations depend on it.
\subsection{ Duality transformation and transfer matrices }\label{appT}
The duality transformation of the $y$-roots can be implemented on the level of transfer matrices and used to obtain explicit expressions for
$T_{a,1}$ and $T_{1,s}$ in the $\sl(2)$ and $\alg{su}(2)$ gradings. This consideration seems to give support to the statement that auxiliary roots do not play any role for T-functions because their number and their values change under a duality transformation.
\smallskip
The eigenvalues of the transfer matrix $T_{a,1}$ in the $\sl(2)$-grading conjectured in \cite{B06} were obtained in \cite{ALST} by using previously found
scattering matrices for string bound states. The transfer matrix $T_{a,1}$ depends on the rapidities $u_1,\ldots, u_N$ of $N\equiv K^{\rm I}$ physical particles and on the rapidity
$v$ of an auxiliary particle that transforms in the bound state representation $(a,1)$ of $\alg{psu}(2|2)$. Eigenvalues of $T_{a,1}$
can be parametrized by a set $(y^{(+)},w^{(+)})$
of auxiliary roots which satisfy the set (\ref{BA1}), (\ref{BA2}) of auxiliary Bethe equations. The Y-functions also involve the transfer matrices $T_{a,-1}$, which have the same structure as $T_{a,1}$ with the replacement $(y^{(+)},w^{(+)})$ for $(y^{(-)},w^{(-)})$. Thus, in what follows we consider $T_{a,1}$ only
and, for the sake of simplicity\footnote{We will also not distinguished between a transfer matrix and its eigenvalues.}, denote the auxiliary roots as $(y,w)$.
The transfer matrix $T_{a,1}$ in the $\sl(2)$ grading reads
\begin{eqnarray}\label{eqn;FullEignvalue1}
&&T_{a,1}^{\sl(2)}(v)=\prod_{i=1}^{K^{\rm{II}}}{\textstyle{\frac{y_i-x^-}
{y_i-x^+}\sqrt{\frac{x^+}{x^-}} \, }}\left[1+
\prod_{i=1}^{K^{\rm{II}}}{\textstyle{
\frac{v-\nu_i+\frac{i}{g}a}{v-\nu_i-\frac{i}{g}a}}}\prod_{i=1}^{K^{\rm{I}}}
{\textstyle{\left[\frac{(x^--x^-_i)(1-x^-
x^+_i)}{(x^+-x^-_i)(1-x^+
x^+_i)}\frac{x^+}{x^-} \right]}} \right.\\
&&{\textstyle{+}}
\sum_{k=1}^{a-1}\prod_{i=1}^{K^{\rm{II}}}{\textstyle{
\frac{v-\nu_i+\frac{i}{g}a}{v-\nu_i+\frac{i}{g}(a-2k)}}} \Big[
\prod_{i=1}^{K^{\rm{I}}}{\textstyle{\frac{x(v+(a-2k)\frac{i}{g})-x_i^-}{x(v+(a-2k)\frac{i}{g})-x_i^+}}}+
\prod_{i=1}^{K^{\rm{I}}}{\textstyle{\frac{1-x(v+(a-2k)\frac{i}{g})x_i^-}{1-x(v+(a-2k)\frac{i}{g})x_i^+}
}}\Big]\prod_{i=1}^{K^{\rm{I}}}{\textstyle{\frac{x^+-x_i^+}{x^+-x_i^-}\frac{v-v_i-(2k+1-a)\frac{i}{g}}{v-v_i+(a-1)\frac{i}{g}
}}}\nonumber\\
&& -\sum_{k=0}^{a-1}\prod_{i=1}^{K^{\rm{II}}} {\textstyle{
\frac{v-\nu_i+\frac{i}{g}a}{v-\nu_i+\frac{i}{g}(a-2k)}}}\prod_{i=1}^
{K^{\rm{I}}}{\textstyle{\frac{x^+-x^+_i}{x^+-x^-_i}\sqrt{\frac{x^-_i}{x^
+_i}} \frac{v-v_i-(2k+1-a)\frac{i}{g}}{v-v_i+(a-1)\frac{i}{g}
}}}\prod_{i=1}^{K^{\rm{III}}}{\textstyle{\frac{w_i-v
+\frac{i(2k-1-a)}{g}}{w_i-v+\frac{i(2k+1-a)}{g}} }} \nonumber\\
&&\left. -\sum_{k=0}^{a-1}\prod_{i=1}^{K^{\rm{II}}} {\textstyle{
\frac{v-\nu_i+\frac{i}{g}a}{v-\nu_i+\frac{i
}{g}(a-2k-2)}}}\prod_{i=1}^
{K^{\rm{I}}}{\textstyle{\frac{x^+-x^+_i}{x^+-x^-_i}\sqrt{\frac{x^-_i}{x^
+_i}} \frac{v-v_i-(2k+1-a)\frac{i}{g}}{v-v_i+(a-1)\frac{i}{g}
}}}\prod_{i=1}^{K^{\rm{III}}}{\textstyle{\frac{w_i-v+\frac{i}{g}(2k+3-a)}{
w_i-v+\frac{i}{g}(2k+1-a)}}}\right]. \nonumber
\end{eqnarray}
In this formula
$$
v=x^++\frac{1}{x^+}-\frac{i}{g}a=x^-+\frac{1}{x^-}+\frac{i}{g}a\, .
$$
The variable $v$ takes values in the mirror theory $v$-plane, so that $x^{\pm}=x(v\pm \frac{i}{g}a)$ with $x(v)$ being the mirror theory $x$-function.
Similarly, $x^{\pm}_j=x_s(u_j\pm \frac{i}{g})$, where $x_s$ is the string theory $x$-function. The overall factor
\begin{eqnarray}
\mathscr{N}_a(v)=\prod_{i=1}^{K^{\rm{II}}}{\textstyle{\frac{y_i-x^-}
{y_i-x^+}\sqrt{\frac{x^+}{x^-}} \, }}\end{eqnarray}
satisfies $\mathscr{N}^+_a\mathscr{N}^-_a=\mathscr{N}_{a-1}\mathscr{N}_{a+1}$ and is, therefore, a gauge transformation which drops from the auxiliary Y-functions.
\smallskip
Concerning the structure of $T_{a,1}^{\sl(2)}$,
we point out that the unity occurring in the first line of (\ref{eqn;FullEignvalue1}) can be considered as
coming from the first product (in the square brackets) in the second line with $k=0$, while
the second term in the first line can be considered as coming from
the first product in the second line with $k=a$.
\smallskip
To discuss the duality transformation, see e.g. \cite{Essler, BS,B06} and reference therein,\footnote{The derivation of the dual form of the Bethe equations basically repeats the one performed in section 3 of \cite{B06}, and is presented here just for completeness.} we introduce the following polynomial in the variable $y$ of degree\footnote{To simplify the derivation, we assume that the level-matching condition is not imposed. We impose the level-matching condition after the duality transformation is done.}
$K^{\rm I}+2K^{\rm III}$ \begin{eqnarray} P(y)&=& \prod_{i=1}^{K^{\rm I }}
(y-x_i^-)\sqrt{\frac{x^+_i}{x^-_i}}\prod_{i=1}^{K^{\rm III}} y\Big(w_i-y-\frac{1}{y}+\frac{i}{g}\Big)\nonumber\\
&-& \prod_{i=1}^{K^{\rm I }} (y-x_i^+)\prod_{i=1}^{K^{\rm III}} y\Big(w_i-y-\frac{1}{y}-\frac{i}{g}\Big) \, .
\label{pol}
\end{eqnarray}
This polynomial has $K^{\rm II}$ roots $y_i$ that are
solutions of the Bethe equations (\ref{BA1}). Therefore, this polynomial can be
written in the form
\begin{eqnarray} P(y)=c \prod_{i=1}^{K^{\rm II }}(y-y_i) \prod_{i=1}^{\widetilde{K}^{\rm II }}(y-\tilde{y}_i) \,.\end{eqnarray}
Here $\widetilde{K}^{\rm II}=K^{\rm I}-K^{\rm II}+2K^{\rm III}$ is the number of
the {\it dual} roots $\tilde{y}_i$. Thus, the ratio \begin{eqnarray}
R(y)=\frac{P(y)}{ \prod_{i=1}^{K^{\rm II }}(y-y_i) \prod_{i=1}^{\widetilde{K}^{\rm II }}(y-\tilde{y}_i)}=c \end{eqnarray} is a
constant independent of $y$. As a result, one has $R(a)=R(b)$ for any $a$ and $b$,
that is \begin{eqnarray}\label{maindual} \prod_{i=1}^{K^{\rm II }}
\frac{y_i-a}{y_i-b}&=&\frac{P(a)}{P(b)} \prod_{i=1}^{\widetilde{K}^{\rm II }}\frac{\tilde{y}_i-b}{\tilde{y}_i-a}\,
.\end{eqnarray}
In particular, \begin{eqnarray} R(x^+)=R(x^-)\,
,~~~~~~~~~~~R(1/x^+)=R(1/x^-)\, , \end{eqnarray} yielding \begin{eqnarray} \hspace{-0.3cm} \prod_{i=1}^{K^{\rm II }}
\frac{y_i-x^-}{y_i-x^+}=\frac{P(x^-)}{P(x^+)} \prod_{i=1}^{\widetilde{K}^{\rm II }}\frac{\tilde{y}_i-x^+}{\tilde{y}_i-x^-}\,
,~~~~
\prod_{i=1}^{K^{\rm II }}\frac{y_i-\frac{1}{x^-}}{y_i-\frac{1}{x^+}}=
\frac{P(\frac{1}{x^-})}{P(\frac{1}{x^+})} \prod_{i=1}^{\widetilde{K}^{\rm II }}\frac{\tilde{y}_i-\frac{1}{x^+}}{\tilde{y}_i-\frac{1}{x^-}}\,
.
\end{eqnarray}
We start with showing how the auxiliary Bethe equations (\ref{BA1}) and (\ref{BA2}) are dualized. By construction, $\tilde{y}_k$ is a root of $P(y)$, {\it i.e. }
$P(\tilde{y}_k)=0$, which is nothing else but the Bethe equations for $\tilde{y}_k$:
\begin{eqnarray} \prod_{i=1}^
{K^{\rm{I}}}\frac{\tilde{y}_k-x^-_i}{\tilde{y}_k-x^+_i}\sqrt{\frac{x^+_i}{x^
-_i}} =\prod_{i=1}^{K^{\rm{III}}}\frac{w_i-\tilde{\nu}_k
-\frac{i}{g}}{w_i-\tilde{\nu}_k+\frac{i}{g}}\, , \end{eqnarray}
where $k=1,\ldots, \widetilde{K}^{\rm II}$.
Next, introducing $x^{\pm}(w)$
\begin{eqnarray}
w\pm \frac{i}{g}=x^{\pm}(w)+\frac{1}{x^{\pm}(w)}\, ,
\end{eqnarray}
we factorize
\begin{eqnarray}
w_k-y-\frac{1}{y}\pm\frac{i}{g}=(x^{\pm}(w_k)-y)\Big(1-\frac{1}{y x^{\pm}(w_k)}\Big)\, .
\end{eqnarray}
Thus,
\begin{eqnarray}
\prod_{i=1}^{K^{\rm II}}\frac{w_k-\nu_i+\frac{i}{g}}{w_k-\nu_i-\frac{i}{g}}&=&\prod_{i=1}^{K^{\rm II}}\frac{(y_i-x^+(w_k))\Big(y_i-\frac{1}{x^+(w_k)}\Big)}
{(y_i-x^-(w_k))\Big(y_i-\frac{1}{x^-(w_k)}\Big)}=\\
\nonumber
&&\hspace{2cm}=\frac{P(x^+(w_k))P\left(\frac{1}{x^+(w_k)}\right)}{P(x^-(w_k))P\left(\frac{1}{x^-(w_k)}\right)}
\prod_{i=1}^{\tilde{K}^{\rm{II}} }\frac{w_k-\tilde{\nu}_i-\frac{i}{g}}{w_k-\tilde{\nu}_i+\frac{i}{g}}\, .
\end{eqnarray}
Therefore, the Bethe equations (\ref{BA2}) acquire the form
\begin{eqnarray}
\label{BA2inter}
\prod_{i=1}^{\tilde{K}^{\rm{II}} }\frac{w_k-\tilde{\nu}_i+\frac{i}{g}}{w_k-\tilde{\nu}_i-\frac{i}{g}}=-
\frac{P(x^+(w_k))P\left(\frac{1}{x^+(w_k)}\right)}{P(x^-(w_k))P\left(\frac{1}{x^-(w_k)}\right)} \prod_{i=1}^{K^{\rm III}}
\frac{w_k-w_i-\frac{2i}{g}}{w_k-w_i+\frac{2i}{g}}\, .\end{eqnarray}
By using eq.(\ref{pol}), it is easy to compute
\begin{eqnarray}
\nonumber
\frac{P(x^+(w_k))P\left(\frac{1}{x^+(w_k)}\right)}{P(x^-(w_k))P\left(\frac{1}{x^-(w_k)}\right)} &=&
\prod_{i=1}^{K^{\rm I}}\frac{(x^+(w)-x^+_i)\Big(1-\frac{1}{x^+_ix^+(w_k)}\Big)}{(x^-(w)-x^-_i)\Big(1-\frac{1}{x^-_ix^-(w_k)}\Big)}
\prod_{i=1}^{K^{\rm III}}\left(\frac{w_k-w_i+\frac{2i}{g}}{w_k-w_i-\frac{2i}{g}}\right)^2=\nonumber \\
&&\hspace{3cm}=\prod_{i=1}^{K^{\rm III}}\left(\frac{w_k-w_i+\frac{2i}{g}}{w_k-w_i-\frac{2i}{g}}\right)^2\, .\end{eqnarray}
Substituting this formula into (\ref{BA2inter}), we obtain the dualized auxiliary Bethe equations
\begin{eqnarray}
\prod_{i=1}^{\tilde{K}^{\rm{II}} }\frac{w_k-\tilde{\nu}_i+\frac{i}{g}}{w_k-\tilde{\nu}_i-\frac{i}{g}}= \prod_{i=1}^{K^{\rm III}}
\frac{w_k-w_i+\frac{2i}{g}}{w_k-w_i-\frac{2i}{g}}\, . \end{eqnarray}
Let us now discuss the duality transformation of $T_{a,1}^{\sl(2)}$. For the simplest case of $T_{1,1}$ the transformation was used in section 5 of \cite{B06}.
To perform this transformation, we combine the term with
$k=0$ from the third line with the unit in the first line,
also, the term with $k=a-1$ from the forth line will be combined
with the second term in the first line; in the remaining sum in
the forth line we make the change of the summation index $k\to
k+1$. This rearrangement yields
\begin{eqnarray}\label{eqn;FullEignvalue3}
&&T_{a,1}^{\sl(2)}(v)=\prod_{i=1}^{K^{\rm{II}}}{\textstyle{\frac{y_i-x^-}
{y_i-x^+}\sqrt{\frac{x^+}{x^-}} \,
}}\left[1-\prod_{i=1}^{K^{\rm{I}}}{\textstyle{\frac{x^+-x^+_i}{x^+-x^-_i}\sqrt{\frac{x^-_i}{x^
+_i}} }}\prod_{i=1}^{K^{\rm III}}
{\textstyle{\frac{w_i-v-\frac{i}{g}(a+1)}{w_i-v-\frac{i}{g}(a-1)}}}
\right. \\
&& +\prod_{i=1}^{K^{\rm{II}}}{\textstyle{
\frac{v-\nu_i+\frac{i}{g}a}{v-\nu_i-\frac{i}{g}a}}}\Big[\prod_{i=1}^{K^{\rm{I}}}
{\textstyle{\frac{(x^--x^-_i)(1-x^- x^+_i)}{(x^+-x^-_i)(1-x^+
x^+_i)}\frac{x^+}{x^-}}} -\prod_{i=1}^
{K^{\rm{I}}}{\textstyle{\frac{x^+-x^+_i}{x^+-x^-_i}\sqrt{\frac{x^-_i}{x^
+_i}} \frac{v-v_i-(a-1)\frac{i}{g}}{v-v_i+(a-1)\frac{i}{g}
}}}\prod_{i=1}^{K^{\rm{III}}}{\textstyle{\frac{w_i-v
+\frac{i(a+1)}{g}}{w_i-v+\frac{i(a-1)}{g}} }} \Big]\nonumber\\
\nonumber &&+\sum_{k=1}^{a-1}\left[
\prod_{i=1}^{K^{\rm{II}}}{\textstyle{
\frac{v-\nu_i+\frac{i}{g}a}{v-\nu_i+\frac{i}{g}(a-2k)}}} \Big[
\prod_{i=1}^{K^{\rm{I}}}{\textstyle{\frac{x(v+(a-2k)\frac{i}{g})-x_i^-}{x(v+(a-2k)\frac{i}{g})-x_i^+}}}
{\textstyle{\frac{x^+-x_i^+}{x^+-x_i^-}\frac{v-v_i+(a-2k-1)\frac{i}{g}}{v-v_i+(a-1)\frac{i}{g}
}}}+ \right.\\
\nonumber &&~~~~~~+
\prod_{i=1}^{K^{\rm{I}}}{\textstyle{\frac{1-x(v+(a-2k)\frac{i}{g})x_i^-}{1-x(v+(a-2k)\frac{i}{g})x_i^+}
}}{\textstyle{\frac{x^+-x_i^+}{x^+-x_i^-}\frac{v-v_i+(a-2k-1)
\frac{i}{g}}{v-v_i+(a-1)\frac{i}{g} }}}\\
\nonumber &&~~~~~~-\prod_{i=1}^
{K^{\rm{I}}}{\textstyle{\frac{x^+-x^+_i}{x^+-x^-_i}\sqrt{\frac{x^-_i}{x^
+_i}} \frac{v-v_i+(a-2k-1)\frac{i}{g}}{v-v_i+(a-1)\frac{i}{g}
}}}\prod_{i=1}^{K^{\rm{III}}}{\textstyle{\frac{w_i-v
+\frac{i(2k-1-a)}{g}}{w_i-v+\frac{i(2k+1-a)}{g}} }}\\
\nonumber &&~~~~~~\left.-\prod_{i=1}^
{K^{\rm{I}}}{\textstyle{\frac{x^+-x^+_i}{x^+-x^-_i}\sqrt{\frac{x^-_i}{x^
+_i}} \frac{v-v_i+(a-2k+1)\frac{i}{g}}{v-v_i+(a-1)\frac{i}{g}
}}}\prod_{i=1}^{K^{\rm{III}}}{\textstyle{\frac{w_i-v
+\frac{i(2k+1-a)}{g}}{w_i-v+\frac{i(2k-1-a)}{g}} }} \Big]\right]\,
.
\end{eqnarray}
By using the polynomial $P$ defined in (\ref{pol}) and performing tedious but straightforward computation, one can show that the last formula can be written in the following more compact form
\begin{eqnarray}
\label{Ta1sl2}
T_{a,1}^{\sl(2)}(v)&=&{\textstyle{P(x^+)}}\displaystyle{\prod_{i=1}^{K^{\rm II}} \textstyle{\frac{y_i-x^-}{y_i-x^+}\sqrt{\frac{x^+}{x^-}}}}\,
\prod_{i=1}^{K^{\rm I}}{\textstyle\frac{ 1}{x^+-x_i^-}\sqrt{\frac{x^-_i}{x^+_i}}}\, \displaystyle{\prod_{i=1}^{K^{\rm III}}}{\textstyle\frac{1}{x^+(w_i-v-\frac{i}{g}(a-1))}} \\
\nonumber
&-&{\textstyle{P\left(\frac{1}{x^-}\right)}}\displaystyle{\prod_{i=1}^{K^{\rm II}}{\textstyle\frac{y_i-\frac{1}{x^+}}{y_i-\frac{1}{x^-}}\sqrt{\frac{x^+}{x^-}}}
\prod_{i=1}^{K^{\rm I}}{\textstyle \frac{x^--x_i^-}{(x^+-x_i^-)(\frac{1}{x^+}-x_i^+)}}
}\displaystyle{\prod_{i=1}^{K^{\rm III}}{\textstyle\frac{x^-}{w_i-v+\frac{i}{g}(a-1)}}}\\
\nonumber
&-&\sum_{k=1}^{a-1}{\textstyle{P\Big(\frac{1}{x(v+\frac{i}{g}(a-2k))}\Big)P\Big(x(v+\frac{i}{g}(a-2k))\Big)}}
\displaystyle{\prod_{i=1}^{K^{\rm I}}}{\textstyle\frac{1}{(x^+-x_i^-)(\frac{1}{x^+}-x_i^+)}\sqrt{\frac{x^-_i}{x^+_i}}}\times\\
\nonumber
&\times &
\frac{\displaystyle{\prod_{i=1}^{K^{\rm II}}}{\textstyle \frac{(y_i-x^-)(y_i-\frac{1}{x^+})}{(y_i-x(v+\frac{i}{g}(a-2k) ))\big(y_i-\frac{1}{x(v+\frac{i}{g}(a-2k) )}\big)}\sqrt{\frac{x^+}{x^-}} } }
{
\displaystyle{\prod_{i=1}^{K^{\rm III}}}(w_i-v-\frac{i}{g}(a-2k+1))(w_i-v-\frac{i}{g}(a-2k-1))}
\, .
\end{eqnarray}
\normalsize
\noindent
Here the first and the second line correspond to the first and the second line in eq.(\ref{eqn;FullEignvalue3}), respectively. In such a representation dualization of
$T_{a,1}^{\sl(2)}$ is straightforward, one has just to use eq.(\ref{maindual}) with proper variables $a$ and $b$. In this way we find the expression for the transfer matrix $T_{a,1}^{\sl(2)}$
in terms of dual variables $\tilde{y}_i$, $i=1,\ldots, \widetilde{K}^{\rm II}$. We further change $\tilde{y}_i\to y_i$, $\widetilde{K}^{\rm II}\to K^{\rm II}$ and denote the corresponding transfer
matrix as $T_{a,1}^{\alg{su}(2)}$. It reads\footnote{Note that eq.(\ref{Ta1sl2}) contains a factor $\left(\frac{x^+}{x^-}\right)^{\frac{1}{2}K^{\rm II}}$. When dualizing, one should express
$K^{\rm II}$ via $\widetilde{K}^{\rm II}$, which gives $\left(\frac{x^+}{x^-}\right)^{K^{\rm II}}=\left(\frac{x^-}{x^+}\right)^{\widetilde{K}^{\rm II}}\left(\frac{x^+}{x^-}\right)^{\frac{1}{2}K^{\rm I}+K^{\rm III}}$.}
\begin{eqnarray}
\begin{aligned}
T_{a,1}^{\alg{su}(2)}(v)&= {\textstyle \left(\frac{x^+}{x^-}\right)^{\sfrac{K^{\rm I}}{2}} }
\displaystyle{ \prod_{i=1}^{K^{\rm II}}} {\textstyle \frac{y_i-x^+}{y_i-x^-} \sqrt{\frac{x^-}{x^+}} }\left[
{\textstyle P(x^-)} \prod_{i=1}^{K^{\rm I}}{\textstyle{\frac{1}{x^+-x_i^-}\sqrt{\frac{x^-_i}{x^+_i}} } }\, \displaystyle{\prod_{i=1}^{K^{\rm III}}}{\textstyle{\frac{1}
{x^-(w_i-v-\frac{i}{g}(a-1))}}}
\right.
\\
&-{\textstyle P\left(\frac{1}{x^+}\right)}
\prod_{i=1}^{K^{\rm I}}{\textstyle\frac{x^--x_i^-}{(x^+-x_i^-)(\frac{1}{x^+}-x_i^+)}}
\displaystyle{\prod_{i=1}^{K^{\rm II}}}{\textstyle\frac{v-\nu_i-\frac{i}{g} a}{v-\nu_i+\frac{i}{g} a}}\displaystyle{\prod_{i=1}^{K^{\rm III}}}{\textstyle \frac{x^+}{w_i-v+\frac{i}{g}(a-1)} }\\
&-{\textstyle{P(x^-)}} {\textstyle{P\left(\frac{1}{x^+}\right)}}\displaystyle{\prod_{i=1}^{K^{\rm I}}}{\textstyle \frac{1}{(x^+-x_i^-)\big(\frac{1}{x^+}-x_i^+\big)}\sqrt{\frac{x^-_i}{x^+_i}}}\, \times\\
& \times
\left.
\sum_{k=1}^{a-1}\frac{\displaystyle{\prod_{i=1}^{K^{\rm II}}}{\textstyle \frac{v-\nu_i+\frac{i}{g}(a-2k)}{v-\nu_i+\frac{i}{g}a}} }
{\, \, \, \, \displaystyle{\prod_{i=1}^{K^{\rm III}}} (w_i-v-\frac{i}{g}(a-2k+1))(w_i-v-\frac{i}{g}(a-2k-1))\frac{x^-}{x^+}
}\right]\, .\end{aligned}
\end{eqnarray}
The transfer matrix above is $T_{a,1}$ in the $\alg{su}(2)$ grading. The number $K^{\rm II}$ should be understood here as a number of $y$-roots which the corresponding state
exhibits in the $\alg{su}(2)$ grading.
\smallskip
Now we turn our attention to the transfer matrix $T_{1,s}$. To obtain the $\sl(2)$ graded version of this transfer matrix, one can apply the complex conjugation
to $T_{a,1}^{\alg{su}(2)}$ regarding the roots $u,y,w$ as real and then pick up a certain normalization factor. For real $u$ complex conjugation
transforms $x^+_i$ into $x^-_i$ and vice versa. This motivates us to introduce a polynomial \begin{eqnarray} P_c(y)&=& \prod_{i=1}^{K^{\rm I }}
(y-x_i^+)\sqrt{\frac{x^-_i}{x^+_i}}\prod_{i=1}^{K^{\rm III}} y\Big(w_i-y-\frac{1}{y}-\frac{i}{g}\Big)\nonumber\\
&-& \prod_{i=1}^{K^{\rm I }} (y-x_i^-)\prod_{i=1}^{K^{\rm III}} y\Big(w_i-y-\frac{1}{y}+\frac{i}{g}\Big) \,
\label{pol1} \end{eqnarray}
which is complex conjugation of $P$ defined in eq.(\ref{pol}). Using this polynomial and replacing $a\to s$, we obtain the following expression for the transfer matrix
$T_{1,s}^{\sl(2)}$
\begin{eqnarray}
\begin{aligned}
T_{1,s}^{\sl(2)}(v)&= \mathscr{M}_{s}\, \displaystyle{ \prod_{i=1}^{K^{\rm II}}} {\textstyle \frac{y_i-x^-}{y_i-x^+} \sqrt{\frac{x^+}{x^-}} }\left[
{\textstyle P_c(x^+)} \prod_{i=1}^{K^{\rm I}}{\textstyle{\frac{1}{x^--x_i^+}\sqrt{\frac{x^+_i}{x^-_i}} } }\, \displaystyle{\prod_{i=1}^{K^{\rm III}}}{\textstyle{\frac{1}
{x^+(w_i-v+\frac{i}{g}(s-1))}}}
\right.
\\
&-{\textstyle P_c\left(\frac{1}{x^-}\right)}
\prod_{i=1}^{K^{\rm I}}{\textstyle\frac{x^+-x_i^+}{(x^--x_i^+)(\frac{1}{x^-}-x_i^-)}}
\displaystyle{\prod_{i=1}^{K^{\rm II}}}{\textstyle\frac{v-\nu_i+\frac{i}{g} s}{v-\nu_i-\frac{i}{g} s}}\displaystyle{\prod_{i=1}^{K^{\rm III}}}{\textstyle \frac{x^-}{w_i-v-\frac{i}{g}(s-1)} }\\
&-{\textstyle{P_c(x^+)}} {\textstyle{P_c\left(\frac{1}{x^-}\right)}}\displaystyle{\prod_{i=1}^{K^{\rm I}}}{\textstyle \frac{1}{(x^--x_i^+)\big(\frac{1}{x^-}-x_i^-\big)}\sqrt{\frac{x^+_i}{x^-_i}}}\times\\
& \times
\left.
\sum_{k=1}^{s-1}\frac{\displaystyle{\prod_{i=1}^{K^{\rm II}}}{\textstyle \frac{v-\nu_i-\frac{i}{g}(s-2k)}{v-\nu_i-\frac{i}{g}s}} }
{\, \, \, \, \displaystyle{\prod_{i=1}^{K^{\rm III}}(w_i-v+\frac{i}{g}(s-2k+1))(w_i-v+\frac{i}{g}(s-2k-1))\frac{x^+}{x^-}}
}\right]\, .\end{aligned}
\end{eqnarray}
Here the overall normalization factor \begin{eqnarray} \nonumber
\mathscr{M}_{s}={\textstyle{(-1)^s }} \prod_{i=1}^{K^{\rm I }} {\textstyle{\left(\frac{x^-_i}{x^+_i}\right)^{\sfrac{s}{2}}\frac{x^--x^+_i}{x^+-x^-_i} }}
\prod_{k=1}^{s-1}
{\textstyle{\frac{x(v+\frac{i}{g}(s-2k))-x^+_i}{x(v-\frac{i}{g}(s-2k))-x^-_i}}}\, \end{eqnarray}
has been found by requiring that $T_{1,s}^{\sl(2)}$ reproduces $T_{a,1}^{\sl(2)}$ through Bazhanov-Reshetikhin formula. Note that $\mathscr{M}_{s}$ is also
a gauge transformation.
\smallskip
Finally, we present for completeness the transfer matrix $T_{1,s}$ in the $\alg{su}(2)$ grading. It is obtained by complex conjugation of $T_{a,1}^{\sl(2)}$ and picking up
a proper normalization factor :
\begin{eqnarray}
\nonumber
T_{1,s}^{\alg{su}(2)}(v)&=&\mathscr{M}_s{\textstyle \left(\frac{x^+}{x^-}\right)^{\frac{K^{\rm I}}{2}}}\displaystyle{\prod_{i=1}^{K^{\rm II}} \textstyle{\frac{y_i-x^+}{y_i-x^-}\sqrt{\frac{x^-}{x^+}}}}\left[{\textstyle{P_c(x^-)}}\,
\prod_{i=1}^{K^{\rm I}}{\textstyle\frac{ 1}{x^--x_i^+}\sqrt{\frac{x^+_i}{x^-_i}}}
\, \displaystyle{\prod_{i=1}^{K^{\rm III}}}{\textstyle\frac{1}{x^-(w_i-v+\frac{i}{g}(s-1))}} \right.\\
\label{T1ssu2}
&-&{\textstyle{P_c\left(\frac{1}{x^+}\right)}}
\displaystyle{\prod_{i=1}^{K^{\rm II}}{\textstyle\frac{v-\nu_i-\frac{i}{g}s}{v-\nu_i+\frac{i}{g}s}}
\prod_{i=1}^{K^{\rm I}}{\textstyle \frac{x^+-x_i^+}{(x^--x_i^+)(\frac{1}{x^-}-x_i^-)}}
}\displaystyle{\prod_{i=1}^{K^{\rm III}}{\textstyle\frac{x^+}{w_i-v-\frac{i}{g}(s-1)}}}\\
\nonumber
&-&\sum_{k=1}^{s-1}{\textstyle{P_c\Big(\frac{1}{x(v-\frac{i}{g}(s-2k))}\Big)P_c\Big(x(v-\frac{i}{g}(s-2k))\Big)}}
\displaystyle{\prod_{i=1}^{K^{\rm I}}}{\textstyle\frac{1}{(x^--x_i^+)(\frac{1}{x^-}-x_i^-)}\sqrt{\frac{x^+_i}{x^-_i}}}\times\\
\nonumber
&\times & \left.
\frac{\displaystyle{\prod_{i=1}^{K^{\rm II}}}{\textstyle \frac{v-\nu_i-\frac{i}{g}s}{v-\nu_i-\frac{i}{g}(s-2k) }} }
{
\displaystyle{\prod_{i=1}^{K^{\rm III}}}(w_i-v+\frac{i}{g}(s-2k+1))(w_i-v+\frac{i}{g}(s-2k-1))}\right]
\, .
\end{eqnarray}
This completes our discussion of the duality transformation for transfer matrices.
|
\section{Introduction}
Several prominent anomalies in the large-angle, low-$\ell$ cosmic microwave
background (CMB) have been identified, starting with pioneering
observations by the \satellite{Cosmic Background
Explorer}~(\satellite{COBE})~\citep{DMR4}, and confirmed and extended with the high
precision observations from the \satellite{Wilkinson Microwave Anisotropy
Probe}~(\satellite{WMAP})~\citep{WMAP1-maps}. These anomalies include the
unexpectedly low correlations at scales above 60 degrees \citep{DMR4,
WMAP1-maps,CHSS-review,SHCSS2011},the alignments of the largest
multipoles with each other and the Solar System
\citep{deOliveira-Costa2004,SSHC2004,Land-MPV,CHSS-anomalies}, a parity
asymmetry at low multipoles \citep{KN2010a,KN2010b,KN2010c,KN2010d}, and
the spatial asymmetries in the distribution of power observed at smaller
scales~\citep{EHBGL2004, EHBGL2004Erratum, Hansen2009}. Numerous attempts
have been made to explain or explain away these anomalies
\citep{Slosar2004,Hajian2007, Afshordi-extradim,WMAP7-anomalies} -- none of
them successful \citep[see][and references therein, for a
review]{CHSS-review}.
The most peculiar and robust CMB anomaly is arguably the lack of correlation on
large angular scales first observed by \satellite{COBE}~\citep{DMR4} and confirmed and
further quantified through the $S_{1/2}$ statistic by
\satellite{WMAP}~\citep{WMAP1-cosmology}. Subsequent study of the two point angular
correlation function, $C(\theta)$, has found further oddities; the large
angle correlation is mainly missing outside of the Galactic region, there
being essentially no correlation on large angles. The large-angle
correlation that is observed comes from the foreground removed Galactic
region of the reconstructed full-sky map~\citep{CHSS-WMAP5}. From the
internal linear combination (ILC) map,\footnote{The ILC map and all data
from the \satellite{WMAP}{} mission is freely available on-line at
\texttt{http://lambda.gsfc.nasa.gov/}.} the full-sky map created from the
individual frequency bands which provides our best picture of the full
sky microwave background radiation, it is found that the lack of
correlation is unlikely at the approximately $95$ per cent level. However,
when solely the region outside the Galaxy of the individual frequency or
ILC maps are analysed the lack of correlation is rare at the approximately
$99.975$ percent level~\citep{CHSS-WMAP5}.
The study of the large-angle CMB presents special problems that must be
treated carefully. Since there is only one Universe to observe and few
independent modes at low-$\ell$, large sky coverage is needed, and even with
this coverage, very little independent information about the ensemble is available.
Further, given the observed low quadrupole power, $C_2^{\mathrm obs}\sim
100$--$200\unit{(\mathrm{\umu K})^2}$, compared to the best fit $\Lambda$CDM model,
$C_2^{\Lambda\mathrm{CDM}}\sim1300\unit{(\mathrm{\umu K})^2}$, large-angle studies are
particularly sensitive to assumptions and unintended biases.
One suggestive example of this is provided by the ILC map itself. If we
use a pixel based estimator for the $C_\ell$ as implemented in
\textsc{SpICE}~\citep{polspice} we can easily determine the quadrupole power inside
and outside the \satellite{WMAP}{} provided analysis mask KQ75y7 to be
\begin{equation}
C_2^{\mathrm{inside}} \approx 610\unit{(\mathrm{\umu K})^2}, \qquad
C_2^{\mathrm{outside}} \approx 80\unit{(\mathrm{\umu K})^2}.
\end{equation}
The KQ75y7 mask cuts out approximately $25$ percent of the sky. Taking the
weighted average of these values produces the intriguing result
\begin{equation}
0.25 C_2^{\mathrm{inside}} + 0.75 C_2^{\mathrm{outside}} \approx
200\unit{(\mathrm{\umu K})^2},
\end{equation}
a value consistent with the \satellite{WMAP}{} reported
$C_2$~\citep{WMAP7-angular-ps}. Again we stress this is a suggestive
example, not a careful analysis; the pseudo-$C_\ell$ (PCL) estimator employed
here is suboptimal, we have not include errors on the estimates, etc. It
does, however, show the wide discrepancy between the Galactic region and
the rest of the sky, a common theme for the ILC map. Further it shows how
a large value mixed in from a small region of the sky significantly impacts
the final result.
In a recent paper \cite{Efstathiou2010}, the authors claimed that the
low $S_{1/2}$ results are due to the use of a suboptimal estimator
(the pixel based estimator) of $C(\theta)$ and proposed an alternative
based on reconstructing the full sky. This proposal avoids addressing
the question of \textit{why} the partial sky contains essentially no
correlations on large angular scales and instead focuses on a new
question that centre on the issue of how the full sky is reconstructed. In
this work we carefully study full-sky reconstruction algorithms and
their effects on the low-$\ell$ CMB\@.
It is well known that contamination affects the reconstruction of the low
multipoles \citep{Bielewicz2004,Naselsky2008,Liu2009,Aurich2010}. In
particular \cite{Aurich2010} have found that smoothing of full sky map
prior to analysis, as required by a reconstruction algorithm (see
\cite{Efstathiou2010} and our discussion below) leaks information from from
the region inside the mask to pixels outside the mask. They showed that
the pixels outside the mask have errors that are a significant fraction of
the mean CMB temperature. They further find that it is safest to calculate
the two point angular correlation function on the cut-sky. Here we confirm
and extend these results.
Alternative analyses such as that suggested in \cite{Efstathiou2010}, must
be performed with care. In this work we carefully study the full-sky
reconstruction, based on the cut-sky data, in a Universe with low
quadrupole power. In Sec.~\ref{sec:formalism} we briefly present the
formalism typically employed in CMB studies. Sec.~\ref{sec:results}
contains our results and we conclude in Sec.~\ref{sec:conclusions}.
Ultimately we find that if a full-sky map, such as the ILC, is a faithful
representation of the true CMB sky, then a reconstruction algorithm can
reproduce its properties. This is not surprising: if the full-sky map is
already trusted, there is no need to perform a reconstruction and nothing
is gained by doing so. However, if part of the full sky is not trusted or is
known to be contaminated, then, by reconstructing without properly
accounting for the assumptions implicit in the algorithm, the final results
will be biased toward the full-sky values. Again this is not surprising,
if information from the questioned region is allowed to leak into the rest
of the map then it will affect the final results and nothing will be
learned about the validity of the reconstruction.
In any reconstruction of unknown values from the properties of existing
data assumptions must be made. Often these assumptions are not explicitly
stated. For the work presented here we take the observed microwave sky
outside of the Galactic region as defined through the KQ75y7 mask to be a
fair sample of the CMB\@. This partial-sky region is known to have
essentially no correlations on large angular scales; it is unlikely in the
best fit $\Lambda$CDM model at the $99.975$ per cent level \citep{CHSS-WMAP5}.
Our study shows the bias introduced into full-sky reconstructions when
an admixture of a region with larger angular correlations is included prior
to reconstruction. We stress that results of the partial-sky analysis are
\textit{not} being questioned, instead a new question is being asked; how
should the full sky be reconstructed when there is a wide disparity between
the statistical properties of the region outside the Galaxy and that inside.
\section{Reconstruction Formalism}
\label{sec:formalism}
Optimal, unbiased estimators for both the $C_\ell$ and $a_{\ell m}$ are well known
and discussed extensively in the literature \citep[see, for
example,][]{Tegmark1997-quadratic-estimators, Efstathiou2004-hybrid,
CMB-mapmaking, Efstathiou2010}. Here we provide a brief overview of the
maximum likelihood estimator~(MLE) technique and introduce our notation.
For details including discussions of invertability of the matrices, proofs
of optimality, etc., see the references.
The microwave temperature fluctuations on the sky can be represented by the
vector $\vec x(\unitvec e_j)$,
\begin{equation}
\vec x = \mat Y\vec a+\vec n,
\end{equation}
where $\mat Y$ is the matrix of the $Y_{\ell,m}(\unitvec{e}_j)$, $j$ runs
over all pixels on the sky, $\unitvec{e}_j$ is the radial unit vector in
the direction of pixel $j$, $\vec a$ is the vector of $a_{\ell m}$ coefficients,
and $\vec n$ is the noise in each pixel. For the work considered here we
are only interested in the large-angle, low-$\ell$ behaviour so we assume
that $\vec n$ can be ignored and set $\vec n=0$ in what follows. When
working with the \satellite{WMAP}{} data at low resolution this is justified, for
example the W band maps at $\textsc{Nside}=16$ have pixel noise
$\sigma_{\mathrm{pix}} < 3\unit{\mathrm{\umu K}}$. At higher resolution this is not as
clearly justified. In this work we study reconstruction bias independent
of pixel noise so we may ignore $\vec n$ for our simulations. When
setting $\vec n=0$ are further assuming that the region we are analysing is
free of foregrounds. This is a standard, though implicit, assumption when
reconstructions are performed. The covariance matrix is then given by
\begin{equation}
\mat C = \langle \vec x \transpose{\vec x} \rangle = \mat S.
\label{eq:Ctheta-ensemble}
\end{equation}
Here the angle brackets, $\langle\cdot\rangle$ represent an ensemble
average. This is the expectation value of the theoretical two point
angular correlation function, not its measured value. As is customary, we
call $\mat S$ the signal matrix despite the fact that it is \textit{not}
the two point angular correlation measured on the sky. We do not include a
noise matrix, $\mat N$, in our covariance since we are neglecting noise.
\subsection{Reconstructing the $\bmath{a_{\ell m}}$}
\label{subsec:estimator-alm}
To reconstruct the $a_{\ell m}$ we define the signal matrix
as the two point
angular correlation function of \textit{the unreconstructed modes}
\begin{equation}
\mat C = \mat S = \sum_{\ell=\ell_{\mathrm{recon}}+1}^{\ellmax} C_\ell \mat P^\ell.
\label{eq:C-alm}
\end{equation}
Here $\mat P^\ell$ is the matrix of the weighted Legendre polynomials,
\begin{equation}
\mat P^\ell_{i,j} \equiv \frac{2\ell+1}{4\pi}
P_\ell(\unitvec{e}_i\cdot\unitvec{e}_j),
\end{equation}
and we assume all modes with $2\le\ell\le\ell_{\mathrm{recon}}$ are to be
reconstructed. Here $\ellmax$ is the maximum multipole considered. We
have chosen $\ellmax=4\textsc{Nside}+2$ for this work. The optimal, unbiased
estimator is then given by \citep{CMB-mapmaking}
\begin{equation}
\hat{\vec a} = \mat W\vec x, \qquad \mat W\equiv [\transpose{\mat Y}\mat{C}^{-1}
\mat Y]^{-1} \transpose{\mat Y}\mat{C}^{-1}.
\label{eq:estimator-alm}
\end{equation}
Note that here and throughout we work in the real spherical harmonic basis,
so $\mat Y$ is a real matrix. The covariance matrix of our estimator is
\begin{equation}
\Sigma \equiv \langle \hat{\vec a}\transpose{\hat{\vec a}} \rangle -
\langle\hat{\vec a}\rangle \transpose{\langle\hat{\vec a}\rangle}
= [\transpose{\mat Y}\mat{C}^{-1}\mat Y]^{-1}.
\label{eq:covariance-alm}
\end{equation}
The signal matrix, $\mat C$, need not include all pairs of pixels on the
sky. When it does, a reconstruction will produce precisely the spherical
harmonic decomposition. Conversely, when a sky is masked, we only include
the unmasked pixels in $\mat C$. The process of `masking' is thus
performed by removing the masked pixels from the signal matrix, and this
process is equivalent to assigning infinite noise to the masked pixels.
\subsection{Reconstructing the $\bmath{C_\ell}$}
\label{subsec:estimator-Cl}
To reconstruct the $C_\ell$ we define the signal matrix as the two point
angular correlation of \textit{all the modes};
\begin{equation}
\mat C = \mat S = \sum_{\ell=2}^{\ellmax} C_\ell \mat P^\ell.
\label{eq:C-Cl}
\end{equation}
Notice that this differs from our previous definition (\ref{eq:C-alm}).
The optimal, unbiased estimator for
the $C_\ell$ is then constructed from an unnormalized estimator, $\vec
y_\ell$. Let
\begin{equation}
\vec y_\ell \equiv \transpose{\vec x}\mat E^\ell\vec x, \qquad
\mat E^\ell \equiv \frac12 \mat{C}^{-1} \mat P^\ell\mat{C}^{-1}.
\label{eq:estimator-yl}
\end{equation}
The correlation matrix of this estimator is the Fisher matrix,
\begin{equation}
\mat F_{\ell,\ell'} = \langle\vec y_\ell\transpose{\vec
y_{\ell'}}\rangle - \langle\vec y_\ell\rangle \transpose{\langle \vec
y_{\ell'}\rangle} = \frac12 \mathop{\mathrm{Tr}}\nolimits\bigl[\mat{C}^{-1}\mat P^\ell\mat{C}^{-1}\mat
P^{\ell'}\bigr].
\label{eq:Fisher-matrix}
\end{equation}
Finally, this gives the optimal, unbiased estimator of the $C_\ell$,
\begin{equation}
\hat{C}_\ell = \sum_{\ell'} \mat F^{-1}_{\ell,\ell'} \vec y_{\ell'}.
\label{eq:estimator-Cl}
\end{equation}
Though the full Fisher matrix can be calculated, it turns out to be nearly
diagonal for reasonably small masks such as the \satellite{WMAP}{} KQ75y7 mask. In this
case the approximations
\begin{equation}
\mat F_{\ell,\ell'} \approx \frac{2\ell+1}{2 \hatC_\ell^2} \delta_{\ell,\ell'},
\qquad \mat F_{\ell,\ell'}^{-1} \approx \frac{2 \hatC_\ell^2}{2\ell+1}
\delta_{\ell,\ell'}
\label{eq:Fisher-approximations}
\end{equation}
may be employed. We have confirmed the validity of this approximation and
have employed it when applicable in our subsequent analyses.
\subsection{Relating the Estimators}
\label{subsec:estimators-relation}
The optimal, unbiased estimators for $a_{\ell m}$ and $C_\ell$ are related to each
other. If we define the weighted harmonic coefficients by
\begin{equation}
\vec\beta \equiv \Sigma^{-1}\hat{\vec a},
\label{eq:definition-beta}
\end{equation}
then
\begin{equation}
\vec y_\ell = \frac12\sum_m |\beta_{\ell m}|^2
\label{eq:estimator-beta}
\end{equation}
is identical to~(\ref{eq:estimator-yl}) from which we may calculate $\hatC_\ell$
\citep{CMB-mapmaking,Efstathiou2010}.
In our discussion we have been careful to note that $\mat C$ is defined
differently when used as an estimator for the $a_{\ell m}$ versus the $C_\ell$. In
practise when the signal-to-noise is large the estimator for the $a_{\ell m}$ is
not sensitive to the precise values and range of the $C_\ell$ employed.
However, to find $\hatC_\ell$ from $\hat{\vec a}$ through the weighted
harmonic coefficients~(\ref{eq:definition-beta}), the \textit{full} signal
matrix~(\ref{eq:C-Cl}) must be used when calculating the covariance
matrix~(\ref{eq:covariance-alm}) and Fisher
matrix~(\ref{eq:Fisher-matrix}).
The above discussion shows that Eq.~(\ref{eq:estimator-Cl}) is the optimal,
unbiased estimator for the $C_\ell$. Even so, given
$\hat{\vec a}$ from~(\ref{eq:estimator-alm}) it is tempting
to define a naive estimator for the $C_\ell$ via
\begin{equation}
C_\ell^e \equiv \frac1{2\ell+1}\sum_m |\hat{\vec a}_{\ell m}|^2
\label{eq:estimator-naive-Cl}
\end{equation}
and use this to reconstruct $C(\theta)$~\cite[see fig.~5
of][]{Efstathiou2010}. In general this is a poor definition for the
estimator as clearly an optimal, unbiased estimator for some quantity does
\textit{not} provide an optimal, unbiased estimator for the square of that
quantity. Its use leads to a biased estimator for the $C_\ell$ and a biased
reconstruction of $C(\theta)$. We will explore both this estimator and the
optimal, unbiased one below.
\subsection{Two Point Angular Correlation Function}
The two point angular correlation function is defined as a sky average,
that is by a sum over all pixels on the sky separated by the angle
$\cos\theta_{i,j} = \unitvec{e}_i\cdot \unitvec{e}_j$,
\begin{equation}
C(\theta_{i,j}) \equiv \sum_{i,j} \vec x_i \vec x_j.
\end{equation}
Ideally the two point angular correlation function would also contain an
ensemble average over realisations of the underlying model. Since we only
have one Universe, this ensemble average cannot be calculated. However,
for a statistically isotropic Universe the sky average and ensemble average
are equivalent. This definition has the additional benefit that it can be
calculated on a fraction of the sky.
Alternatively the two point angular correlation function may be expanded in
a Legendre series,
\begin{equation}
C(\theta_{i,j}) = \sum_{\ell} \frac{2\ell+1}{4\pi} C_\ell
P_\ell(\cos\theta_{i,j}).
\label{eq:Ctheta-vs-Cl}
\end{equation}
Note that for partial sky coverage or lack of statistical isotropy the
$C_\ell$ in this this expression are \textit{not} the same as the $\hatC_\ell$
obtained from the $a_{\ell m}$; see~\cite{CHSS-WMAP3} for a discussion. This
subtlety will not be important for the following work.
\subsection{$\bmath{S_{1/2}}$ Statistic}
\label{sec:S12}
To quantify the lack of large-angle correlations the $S_{1/2}$ statistic
has been defined by \cite{WMAP1-cosmology} to be
\begin{equation}
S_{1/2}\equiv \int_{-1/2}^1 \left[ C(\theta) \right]^2
\mathrm{d}(\cos\theta).
\label{eq:S12-def}
\end{equation}
Expanding $C(\theta)$ in terms of the $C_\ell$ as
above~(\ref{eq:Ctheta-vs-Cl}) we find
\begin{equation}
S_{1/2} = \sum_{\ell,\ell'} C_\ell \mathcal{I}_{\ell,\ell'} C_{\ell'},
\label{eq:S12-Cl}
\end{equation}
where
\begin{equation}
\mathcal{I}_{\ell,\ell'} \equiv \frac{(2\ell+1)(2\ell'+1)}{(4\pi)^2}
\int_{-1/2}^1 P_\ell(\cos\theta) P_{\ell'}(\cos\theta) \mathrm{d}(\cos\theta)
\end{equation}
is a known matrix (see \citealt{CHSS-review}) that can be evaluated.
The estimator generally employed for $S_{1/2}$ is
\begin{equation}
\hat S_{1/2} = \sum_{\ell,\ell'} \hat C_\ell \mathcal{I}_{\ell,\ell'}
\hat C_{\ell'}.
\label{eq:estimator-S12}
\end{equation}
Even with $\hat C_\ell$ itself an optimal, unbiased estimator of $C_\ell$,
this does \textit{not} produce an optimal, unbiased estimator for $S_{1/2}$
\citep{Pontzen2010}. For the unbiased estimator~(\ref{eq:estimator-Cl})
we have
\begin{eqnarray}
\langle \hat C_\ell \hat C_{\ell'}\rangle & = &
\sum_{\tilde\ell,\tilde\ell'} \mat F_{\ell,\tilde\ell}^{-1}
\mat F_{\tilde\ell',\ell'}^{-1} \langle y_{\tilde\ell} \transpose{y_{\tilde\ell'}}
\rangle \nonumber \\
& = &
\sum_{\tilde\ell,\tilde\ell'} \mat F_{\ell,\tilde\ell}^{-1}
\mat F_{\tilde\ell',\ell'}^{-1}
\left(
\langle y_{\tilde\ell} \rangle \transpose{\langle
y_{\tilde\ell'}\rangle}
+ \mat F_{\tilde\ell,\tilde\ell'} \right) \nonumber \\
& = &
\sum_{\tilde\ell} \left( \sum_{\tilde\ell'} \langle
\mat F_{\ell,\tilde\ell}^{-1} y_{\tilde\ell}\rangle
\transpose{\langle \mat F_{\ell',\tilde\ell'}^{-1} y_{\tilde\ell'}\rangle}
+ \mat F_{\ell,\tilde\ell}^{-1}\delta_{\ell',\tilde\ell}
\right)\nonumber \\
& \approx & C_\ell C_{\ell'} + \frac{2}{2\ell+1} C_\ell^2 \delta_{\ell,\ell'}.
\end{eqnarray}
In the second line we have used the definition of the Fisher
matrix~(\ref{eq:Fisher-matrix}), the third line is an algebraic
simplification, and in the final line we have again
used~(\ref{eq:estimator-Cl}), the fact that $\hat C_\ell$ is
unbiased, and the approximation from~(\ref{eq:Fisher-approximations}).
With this it now straightforward to see that
\begin{eqnarray}
\langle\hat S_{1/2}\rangle & = &
\sum_{\ell,\ell'} \langle \hat C_\ell \hat C_{\ell'}\rangle
\mathcal{I}_{\ell,\ell}
= S_{1/2} + \sum_{\ell} \frac{2 C_\ell^2}{2\ell+1}
\mathcal{I}_{\ell,\ell}
\nonumber \\
& \neq & S_{1/2}.
\end{eqnarray}
It is thus clear that~(\ref{eq:estimator-S12}) is a biased estimator and, in
fact, is biased toward larger values of $S_{1/2}$.
As noted by \cite{Pontzen2010} this is of `pedagogical interest' but does
not affect the studies of low $S_{1/2}$. The Monte Carlo simulations
employed \citep[see][for example]{CHSS-WMAP5} account for this bias. It
does suggest that an alternative measure of the lack of large-angle
correlations is desirable.
\subsection{Assumptions}
\cite{Efstathiou2010} claim that the full-sky, large-angle CMB can be
reconstructed solely from the harmonic structure of the CMB outside the
masked, Galactic region, and independent of the contents of the
masked portion of the sky. We will demonstrate in what follows that this
claim does not hold up to closer scrutiny.
It is clear that without assumptions regarding the harmonic structure inside
the masked region nothing can be said about it. In principle the low-$\ell$
harmonic structure inside the masked region could be anything, ranging from no
power, to large power, to wild oscillations, making the full-sky
reconstruction impossible.
Assuming a cosmological origin for the observed microwave signal outside
the masked region, it seems reasonable to assume it will be consistent with
the signal inside the masked region. With that assumption, the harmonic
structure outside the masked region can be extended into the masked region.
For actual, full-sky maps there is a further assumption: the region inside
the mask is well enough determined and statistically close enough to the
region outside the mask that it does not bias the reconstruction. This
latter assumption turns out to not be true as we demonstrate below.
Note also that if the region inside the mask is trusted, then there is no
need to perform either masking or the reconstruction at all, the full-sky
map can be analysed directly. Therefore, validity of the stringent
assumptions required for the reconstruction obviates the very need for the
reconstruction.
When the reconstruction formalism described above is applied to actual
data, further assumptions are implicit. In our development we have
assumed that the temperature fluctuations contain pure CMB signal. In
practise, besides pixel noise (which we have not included as described
above) the data may contain unknown foregrounds. To avoid contamination
by foregrounds it is common to analyse a foreground-cleaned map, such as
the ILC map, and to mask the most contaminated regions of the sky.
In following this approach, care must be taken not to reintroduce
contamination in the data prior to reconstruction. As we will show below,
the standard process of preparing data for reconstruction, in particular
smoothing the full-sky map, violates this requirement.
\section{Results}
\label{sec:results}
To explore how data handling prior to reconstruction affects the results,
we have performed a series of Monte Carlo simulations of $\Lambda$CDM based
on reconstruction procedures suggested in the literature. We have employed
the simplest best-fitting $\Lambda$CDM model from \satellite{WMAP}{} based solely on
the \satellite{WMAP}{} data. This is model ``lcdm+sz+lens'' with ``wmap7'' data from
the lambda site. Our results are insensitive to the exact details of the
model since we are performing a theoretical study examine relative
differences between reconstructions and not performing parameter
estimation. Our simulations are performed at $\textsc{Nside}=128$ unless otherwise
noted and we will focus on the reconstruction of $a_{2 m}$ and $C_2$.
Further, our simulations only consider $\ell_{\mathrm{recon}}=10$, reconstruct from the
pixels outside the KQ75y7 mask provided by \satellite{WMAP}{} and degraded to the
appropriate resolution, and use the data from the \satellite{WMAP}{} seven year
release.
A collection of realisations of the full sky are created as follows:
\begin{enumerate}
\item Generate a random sky at $\textsc{Nside}=512$ from the best-fitting
$\Lambda$CDM model.
\item Extract the $a_{2 m}$ and calculate the power in the quadrupole,
denote this value by $C_2$.
\item Rescale the $a_{2 m}$ so that the $C_2$ in the map has a fixed value,
for example, rescale so that $C_2=100\unit{(\mathrm{\umu K})^2}$ by replacing the
$a_{2 m}$ with $a_{2 m} \rightarrow a_{2 m}
\sqrt{100\unit{(\mathrm{\umu K})^2}/C_2}$. Notice that this does not change
the phase structure of the $a_{2 m}$.
\item Smooth the map with a $10\degr$ Gaussian beam, if desired.
\item Degrade the map to the desired resolution ($\textsc{Nside}=128$ or $\textsc{Nside}=16$).
\item Repeat the rescaling of the $a_{2 m}$ for each value of $C_2$
that we wish to consider. In our simulations we consider
$C_2=10\mbox{--}10^4\unit{(\mathrm{\umu K})^2}$.
This ensures that the same map realisation is used with only the quadrupole
power changed.
\end{enumerate}
This procedure constitutes a single realisation. The results in this work
are based on at least $20,000$ realisations.
Degrading masks requires an extra processing step. Pixels near mask
boundaries turn from the usual $1$ or $0$ to denote inclusion or exclusion
from the analysis, respectively, to fractional values. We redefine our
degraded masks by setting all pixels with a value greater than $0.7$ to $1$
and all others to $0$. For the KQ75y7 mask this process leaves about $70$
per cent of the pixels for analysis. To be precise, at $\textsc{Nside}=128$ this
leaves $136,828$ unmasked pixels and at $\textsc{Nside}=16$ there are $2,157$
pixels left.
A map with a modest angular resolution contains all the low-$\ell$ CMB
information, so it may seem surprising that we employ $\textsc{Nside}=128$ in our
studies. Instead it is common in low-$\ell$ studies to employ a map at
$\textsc{Nside}=16$, corresponding to pixels of approximately $3\degr$ in size
\citep[see][for a recent example of this]{Efstathiou2010}. The effects of
the choice of resolution, the need for smoothing a map prior to analysis, and
the leaking of information this causes will now be explored.
\subsection{Choice of Map Resolution}
\begin{figure}
\includegraphics[width=0.5\textwidth]{a2m_LCDM_r7_contours_nside_128}
\caption{The $95$ and $5$ percentile lines for the $a_{2 m}$
reconstructed from the pixels outside the KQ75y7 mask at $\textsc{Nside}=128$ (and
thus $\ellmax=514$) of $\Lambda$CDM realisations with
$C_2=100\unit{(\mathrm{\umu K})^2}$, as discussed in the text. The red, solid
lines are for the real part of the $a_{2 m}$ and the blue, dashed lines
are for the imaginary part. The black, solid line shows the expected
result for a perfect reconstruction. We see that the reconstruction is
unbiased, that is, it tracks the true value.}
\label{fig:recon-a2m-C2-100-nside-128}
\end{figure}
\begin{figure}
\includegraphics[width=0.5\textwidth]{a2m_LCDM_r7_contours_nside_16}
\caption{The same as Fig.~\ref{fig:recon-a2m-C2-100-nside-128} now
reconstructed from the pixels outside the KQ75y7 mask at $\textsc{Nside}=16$
(and thus $\ellmax=66$). Here we see that the reconstruction is
biased.}
\label{fig:recon-a2m-C2-100-nside-16}
\end{figure}
The study of large-angle, low-$\ell$ properties of the CMB appears naively
not to require high resolution maps. Maps degraded to the resolution
corresponding to $\textsc{Nside}=16$ are commonly employed
\citep{CMB-mapmaking,Efstathiou2010}. When a map is degraded by averaging
over pixels, high frequency noise is introduced as may be seen in
Figs.~\ref{fig:recon-a2m-C2-100-nside-128} and
\ref{fig:recon-a2m-C2-100-nside-16}. These figures show the reconstructed
$a_{2 m}$ using the optimal, unbiased estimator from
Eq.~(\ref{eq:estimator-alm}) for realisations with
$C_2=100\;\unit{(\mathrm{\umu K})^2}$. The solid, red lines (dashed, blue lines) show
the $5$ and $95$ percentile lines from our realisations for the
reconstructed real (imaginary) parts of each $a_{2 m}$, using maps degraded
to $\textsc{Nside}=128$ (Fig.~\ref{fig:recon-a2m-C2-100-nside-128}) and $\textsc{Nside}=16$
(Fig.~\ref{fig:recon-a2m-C2-100-nside-16}) and pixels outside the KQ75y7
mask. As expected from an unbiased estimator the reconstructed values
track the true values (Fig.~\ref{fig:recon-a2m-C2-100-nside-128}). Further
we see that the $a_{2 1}$ are best determined and the $a_{2 0}$ and
$a_{2 2}$ have larger variances due to the mask which produces greater
admixture of ambiguous modes for these cases. However, for $\textsc{Nside}=16$
(Fig.~\ref{fig:recon-a2m-C2-100-nside-16}) we see that the reconstruction
does not track the true values and is instead biased. This bias is due to
the averaging done to degrade the maps and becomes more significant the
more the map is degraded. From this we conclude that the coupling of the
small-scale modes to the large-scale modes caused by using maps with
resolution that is too coarse can be at least partly responsible for
reconstruction bias.
\subsection{Smoothing the Map}
\begin{figure}
\includegraphics[width=0.5\textwidth]{a2m_LCDM_r7_contours_nside_128_10deg}
\caption{The same as Fig.~\ref{fig:recon-a2m-C2-100-nside-128} now
reconstructed from maps smoothed with a $10\degr$ Gaussian beam
applied to the full sky map. As in
Fig.~\ref{fig:recon-a2m-C2-100-nside-128}, the reconstruction is
unbiased.}
\label{fig:recon-a2m-C2-100-nside-128-10deg}
\end{figure}
\begin{figure}
\includegraphics[width=0.5\textwidth]{a2m_LCDM_r7_contours_nside_16_10deg}
\caption{The same as Fig.~\ref{fig:recon-a2m-C2-100-nside-128-10deg} now
reconstructed from maps with $\textsc{Nside}=16$. The reconstruction is now
also unbiased.}
\label{fig:recon-a2m-C2-100-nside-16-10deg}
\end{figure}
In practise raw degraded maps are not used for the reasons shown in the
previous section, instead the maps are smoothed with a Gaussian beam with
FWHM of at least the size of the pixels and then degraded. In this work we
employ a smoothing scale of $10\degr$, consistent with
\cite{Efstathiou2010}. Smoothing the maps studied in the previous section
prior to reconstructing the $a_{\ell m}$ produces the results shown in
Figs.~\ref{fig:recon-a2m-C2-100-nside-128-10deg} and
\ref{fig:recon-a2m-C2-100-nside-16-10deg}. With smoothing we see that the
$a_{\ell m}$ estimator is unbiased for both resolutions, $\textsc{Nside}=128$ and
$\textsc{Nside}=16$. Smoothing is thus an essential step when working with low
resolution maps.
In Figs.~\ref{fig:recon-a2m-C2-100-nside-128-10deg} and
\ref{fig:recon-a2m-C2-100-nside-16-10deg} we also see that the variance in
the reconstructed values is resolution dependent with the smaller variance
provided by the higher resolution maps. Again this is not surprising, and
can be understood as follows. Our covariance matrix in
Eq.~(\ref{eq:covariance-alm}) does not include a noise term yet we have
introduced noise by degrading. Smoothing does a good job at reducing the
noise to a level where the reconstruction is unbiased, however, there is
still residual noise that affects the covariance of the estimator. The
higher the resolution the smaller this noise. The best results are
obtained by working at the highest resolution that is feasible. For this
reason we work at $\textsc{Nside}=128$ in our simulations. See
Appendix~\ref{app:high-resolution} for technical details.
\subsection{Reconstructing the $\bmath{a_{\ell m}}$}
We have now seen that the estimator in Eq.~(\ref{eq:estimator-alm}) is an
optimal, unbiased estimator for the $a_{\ell m}$ when we work at high resolution
and/or smooth the maps prior to reconstruction
(Figs.~\ref{fig:recon-a2m-C2-100-nside-128},
\ref{fig:recon-a2m-C2-100-nside-128-10deg},
\ref{fig:recon-a2m-C2-100-nside-16-10deg}). Although
this has only been shown for $C_2=100\unit{(\mathrm{\umu K})^2}$ we have verified that
this is true independent of the quadrupole power.
\begin{figure}
\includegraphics[width=0.5\textwidth]{a2m_LCDM_r7_contours_100_with_ilc}
\caption{The same as Fig.~\ref{fig:recon-a2m-C2-100-nside-128-10deg} now
with the masked region filled in with the ILC map \textit{prior} to
smoothing and rescaling. We clearly see the reconstructed $a_{2 m}$
are \textit{not} unbiased. The bias in reconstructing $a_{20}$ and
$a_{22}$ is particularly apparent. This is due to the leakage of
information from inside the masked region.}
\label{fig:recon-a2m-C2-100-with-ILC}
\end{figure}
\begin{figure}
\includegraphics[width=0.5\textwidth]{a2m_LCDM_r7_contours_1000_with_ilc}
\caption{The same as Fig.~\ref{fig:recon-a2m-C2-100-with-ILC} now with
$C_2=1000\unit{(\mathrm{\umu K})^2}$.}
\label{fig:recon-a2m-C2-1000-with-ILC}
\end{figure}
As noted above, the fact that we are smoothing the maps prior to masking
imposes assumptions on the maps. For the realisations discussed above the
assumptions are met; the region inside the mask is, statistically, identical
to the region outside. However, for real data the Galaxy is a bright
foreground that must be removed. The \satellite{WMAP}{} ILC procedure attempts to do this
and produce a full-sky CMB map. Even so, masking is often performed to avoid
relying on the information inside this region since it may still be
contaminated by Galactic foregrounds.
Unfortunately, when the map is smoothed information leaks out of the masked
region and biases the reconstruction as shown in
Figs.~\ref{fig:recon-a2m-C2-100-with-ILC} and
\ref{fig:recon-a2m-C2-1000-with-ILC}. For this analysis, for each synthetic
map we filled the
masked region with the corresponding portion (i.e.\ the masked region)
taken from the ILC map. We then smoothed and degraded the resulting
synthetic map. In these two figures we show the true and reconstructed
values of the coefficients $a_{2 m}$; we also show the ILC map's $a_{2
m}$ for reference. We clearly see the bias in the reconstructed
$a_{2 m}$ and its correlation with the ILC values. If $a_{2
m}^{\mathrm{rec}}<a_{2 m}^{\mathrm{ILC}}$, then $a_{2
m}^{\mathrm{rec}}$ is biased upwards, and vice versa. For example,
the ILC $a_{2 2}$ values are large and negative which leads to the
reconstruction being skewed to agree better at large, negative values
than at large positive values. This trend continues for the other
$a_{2 m}$ and clearly shows that the smoothing has mixed information
from the masked region.
We can also recognise other details in the quality of the reconstruction
that are specifically due to the orientation of the KQ75y7 mask in Galactic
coordinates. For example, we see that the variance in the reconstructed
real part of $a_{22}$ is larger than that for the imaginary part
of $a_{22}$; the reason is that the real part of $Y_{22}$ has an extremum in
the centre of the Milky Way where the mask `bulges' while the imaginary
part has a node at this location. Therefore, more information relevant to
the real part of $a_{22}$ is missing than for the imaginary part, and the
former has a larger reconstruction error. Moreover, it is also the case
that the $Y_{20}$ and $Y_{22}$ have extrema in the Galactic plane whereas
$Y_{21}$ has nodes. Due to this the variances of $a_{20}$ and $a_{22}$ are
expected to be larger than that of $a_{21}$, as our reconstruction plots
show.
Notice also that the reconstruction bias we find is \textit{not} an artifact
of the sharp transition introduced in the process of filling the masked
regions of simulated maps with the ILC contents. The smoothing procedure,
for one, completely removes the sharp feature in the map. Moreover, we have
explicitly checked that the reconstructed $a_{\ell m}$ are not biased when the
cut is filled with contents of \textit{another} statistically isotropic
map. Therefore, the reconstruction bias seen in our plots is real, and is
caused to the specific structure of the ILC map behind the Galactic plane
which `leaks' into the unmasked region.
The question, then, is how to fill the masked region before smoothing.
In principle anything could be used to fill the Galactic region, but
then the information about this fill would leak outside the mask due
to the smoothing. If the map were masked prior to smoothing then
`zero' would be leaked and bias the reconstruction. Alternatively, if
the Galactic region were filled with Gaussian noise with
root-mean-squared value consistent with the region outside the Galaxy
then the estimator would be unbiased similar to the results in
Fig.~\ref{fig:recon-a2m-C2-100-nside-128}, but this would rely on the
\textit{assumption} that the true CMB inside the mask has precisely
the same statistical properties as the CMB in the region outside.
Filling with the ILC values would make sense if we could be completely
confident that the ILC reconstruction of the region inside the mask is
accurate. However, in the ILC the region inside the Galactic mask has
different statistical properties than the region outside, particularly
for the large-angle behaviour. This alone raises concerns that the
ILC reconstruction is not entirely accurate. Further, if we knew how
to properly treat the region inside the mask, either by accepting the
ILC values or filling it with appropriate statistical values, there
would be no need for a reconstruction as we would have a full sky map
to analyse!
The challenge is that there is no, or at least no unique, compelling choice
of how to fill the masked region before smoothing. In the face of this,
the approach we take below is to study how the admixture of the large-angle
behaviour of the Galactic region from the ILC map affects the
reconstruction of the low-$\ell$ CMB, particularly when the region outside
the Galaxy has low quadrupole power and lack of large-angle correlations.
We show how this particular choice biases the reconstruction.
\subsection{Reconstructing the $\bmath{C_\ell}$}
Since we are interested in reconstructing $C(\theta)$ we next need to
reconstruct the $C_\ell$. From the $a_{\ell m}$ we first proceed using the naive
estimator~(\ref{eq:estimator-naive-Cl}), denoted $C_\ell^e$ (as used to
generate fig.~5 of \citealt{Efstathiou2010}).
\begin{figure}
\includegraphics[width=0.5\textwidth]{C2_LCDM_r7_contours}
\caption{The $95$, $50$, and $5$ percentile lines of the reconstructed
$C_2$ (top to bottom, respectively) from our realisations. The maps
have \textit{not} been smoothed prior to reconstruction. The pixel based
(blue, solid line) comes from \textsc{SpICE}; where as, the reconstructed (red,
dashed line) is the estimator $C_\ell^e$. We see that this estimator is
clearly biased toward larger reconstructed values for small, true
$C_2$, such as the values extracted from \satellite{WMAP}{} using either the PCL
or MLE procedures. For a value of $C_2$ near the $\Lambda$CDM value
the reconstruction method is a good estimator. The pixel based method
produces values of $C_2$ with a median much closer to the true values,
though with larger error bars.}
\label{fig:recon-C2-r7}
\end{figure}
The results for this estimator are shown in Fig.~\ref{fig:recon-C2-r7}.
For these realisations the maps were \textit{not} smoothed. The
reconstruction is shown as the dashed, red lines representing the $5$,
$50$, and $95$ percentile values as a function of the true $C_2$ used to
generate the maps. The solid, blue lines are the equivalent values from
the reconstruction based on the pixel estimator from \textsc{SpICE}.\@ Again the
solid, black line is the reconstructed=true relation plotted to guide the
eye. At large $C_2$ we see the desired behaviour: the reconstructed values
from both estimators are centred around the true value, and $C_2^e$ does
have a smaller variance, as an optimal estimator should (however, this does
not mean it \textit{is} optimal). At low $C_2$, in particular near the
\satellite{WMAP}{} PCL and MLE values, the pixel based estimator is
still centred around the true value, though with large variance; however,
the $C_2^e$ is now biased toward larger values.
\begin{figure}
\includegraphics[width=0.5\textwidth]{C2_LCDM_r7_contours_10deg}
\caption{The same as Fig.~\ref{fig:recon-C2-r7} now with the realisations
smoothed to $10\degr$ prior to reconstruction. It appears that the
estimator~(\ref{eq:estimator-naive-Cl}) does a better job of
reproducing $C_2$ for a $\Lambda$CDM model, though, see
Fig.~\ref{fig:recon-C2-r7-10deg-ILC}.}
\label{fig:recon-C2-r7-10deg}
\end{figure}
The results in Fig.~\ref{fig:recon-C2-r7} were for unsmoothed maps. The
usual approach is to smooth the maps which suppresses power on small
scales (high-$\ell$). Fig.~\ref{fig:recon-C2-r7-10deg} shows the results
when the maps are smoothed prior to reconstruction; they are encouraging.
Both estimators now track the true values much more closely. Even the
median of $C_2^e$ remains close to the true value for values near the \satellite{WMAP}{}
PCL value. This shows that with smoothing the correlations are reduced due
to the lack of high frequency noise. It suggests that smoothing the map,
reconstructing the $a_{\ell m}$, and employing $C_\ell^e$ as our estimator is
sufficient and nearly optimal.
\begin{figure}
\includegraphics[width=0.5\textwidth]{C2_LCDM_r7_contours_ilc_10deg}
\caption{The same as Fig.~\ref{fig:recon-C2-r7-10deg} now with the region
inside the masked replaced by the ILC prior to smoothing to $10\degr$
and reconstructing. Clearly the ILC information from inside the masked
region has leaked out biasing the reconstruction. Not surprisingly the
reconstruction now is only accurate near the \satellite{WMAP}{} MLE value; the value
consistent with this region of the ILC. At lower $C_2$ values the
reconstruction plateaus to this value as it is the main contribution to
quadrupole power. At higher values of $C_2$ the quadrupole power is
suppressed by the leakage.}
\label{fig:recon-C2-r7-10deg-ILC}
\end{figure}
Unfortunately this is not the case. As noted above, smoothing makes
assumptions about the validity of the region inside the mask. We saw that
even for the $a_{\ell m}$ this leads to a bias (see
Fig.~\ref{fig:recon-a2m-C2-100-with-ILC}). When the corresponding ILC
portion is placed into the masked region prior to smoothing the $C_2^e$ is
also biased as shown in Fig.~\ref{fig:recon-C2-r7-10deg-ILC}. We see that
the masked region drives $C_2^e$ to be near the value inside the mask
(approximately the \satellite{WMAP}{} MLE value). The $C_2^e$ results are biased
upward for very small $C_2$ and downward for large $C_2$. Thus, even
though smoothing helps in removing the correlation bias in the $C_\ell^e$
estimator it introduces its own bias. How the masked region is filled
determines how the distribution of $C_2^e-C_2^{\mathrm{true}}$ will be
skewed. Roughly speaking the values inside the mask will be favoured,
raising the reconstructed values that are lower than the masked region
values, and lowering values that are higher than those from the masked
region.
We have seen that the naive estimator, $C_\ell^e$, provides an unbiased
estimate of $C_2$ when the true value is near the expected, $\Lambda$CDM
value. However, when the true value is low this estimator tends to
\textit{overestimate} $C_2$. Further, when smoothing is applied the
reconstruction skews the values towards those consistent with the region
inside the mask. This is to be expected. In fact, if the region inside
the mask were believed then there would be no need to reconstruct at all, a
full-sky map would already exist and it could be used for analysis without
this extra effort.
\begin{figure}
\includegraphics[width=0.5\textwidth]{C2_LCDM_r7_unbiased_contours}
\caption{Similar to Fig.~\ref{fig:recon-C2-r7-10deg} now comparing
$C_2^e$, the harmonic coefficient
estimator~(\ref{eq:estimator-naive-Cl}) again as the dashed, red
lines to $\hat C_2$, the weighted harmonic coefficient
estimator~(\ref{eq:estimator-Cl}) as the green, solid lines. We see
that the weighted harmonic coefficient estimator is unbiased over the
full true $C_2$ range.}
\label{fig:recon-unbiased-C2-r7}
\end{figure}
\begin{figure}
\includegraphics[width=0.5\textwidth]{C2_LCDM_r7_unbiased_ilc_contours}
\caption{The same as Fig.~\ref{fig:recon-unbiased-C2-r7} now for
the Galactic region filled with the ILC values. We see that both
estimators are now biased to agree best near the \satellite{WMAP}{} MLE value as
we saw in Fig.~\ref{fig:recon-C2-r7-10deg-ILC}.}
\label{fig:recon-unbiased-C2-r7-ILC}
\end{figure}
\subsection{Optimal, Unbiased $\bmath{C_\ell}$ Estimator}
The general behaviour found for the naive estimator, $C_\ell^e$, carry over to
the optimal, unbiased estimator, $\hatC_\ell$, based on the weighted harmonic
coefficients~(\ref{eq:estimator-Cl}), as we now see. Calculating $\hatC_\ell$
for the realisations considered above we find the results in
Fig.~\ref{fig:recon-unbiased-C2-r7} for Gaussian smoothed maps. This
figure should be compared to Fig.~\ref{fig:recon-C2-r7-10deg}. We see that
$\hat C_2$ is nearly unbiased over the full range of true $C_2$ as expected.
The effect of smoothing when the ILC is inserted into the masked region is
shown in Fig.~\ref{fig:recon-unbiased-C2-r7-ILC}. Again we see the bias
introduced by smoothing when the two regions do not contain the same
structure. These results are qualitatively similar to those found
in Fig.~\ref{fig:recon-C2-r7-10deg-ILC} and the same discussion applies.
\begin{table}
\caption{$S_{1/2}$ values for the ILC map calculated for $2\le\ell\le
10$. The map is unprocessed, Gaussian smoothed with a $10\degr$ beam,
or had the Galactic region filled with a Gaussian random, statistically
isotropic sky realisation with the same power spectrum as the region
outside this region prior to smoothing. The values are calculated for
the full sky and for the KQ75y7 masked sky at $\textsc{Nside}=128$ using a
pixel based estimator or the optimal, unbiased $C_\ell$
estimator~(\ref{eq:estimator-Cl}) from maps at $\textsc{Nside}=128$ and
$\textsc{Nside}=16$. The last row refers to the map whose mask area has been
filled with Gaussian random field whose power is consistent with power
measured outside the mask. }
\label{tab:S12-ilc}
\begin{tabular}{ccccc} \hline
& \multicolumn{4}{c}{\boldmath $S_{1/2}\;(\mu \mathbf{K})^4$}
\\ \cline{2-5}
\textbf{ILC Map} & \textbf{Full Sky}& \textbf{Cut Sky}&
\multicolumn{2}{c}{\textbf{Reconstructed Sky}} \\ \cline{4-5}
\textbf{Processing} && \textbf{Pixel-}
&$\mathbf{N_{side}}$&$\mathbf{N_{side}}$\\
& & \textbf{based}&
\textbf{=128} &\textbf{=16} \\ \hline
Unsmoothed & 8835 & 1275 & 5390 & 2300 \\
$10^\circ$ smoothing & 8835 & 1270 & 2230 & 1670 \\
Filled with consistent & 1020 & 1290 & 1020 & 950 \\
power and smoothed & & & & \\
\hline
\end{tabular}
\end{table}
\begin{figure}
\includegraphics[width=0.5\textwidth]{C2_LCDM_r7_unbiased_raw_multires_contours}
\caption{The same as Fig.~\ref{fig:recon-unbiased-C2-r7} now comparing
the weighted harmonic coefficient estimator~(\ref{eq:estimator-Cl}) for
$\textsc{Nside}=128$ as the green, solid lines and $\textsc{Nside}=16$ as the red,
dashed lines without smoothing the map prior to reconstruction. Since
there is no smoothing, the results do not depend on the contents of the
masked galactic region. We see that the reconstruction without
smoothing is unbiased for most of the $C_2$ range, however see
Sec.~\ref{sec:nosmoothing} for a discussion of its inapplicability to
real data.}
\label{fig:recon-unbiased-multires-C2}
\end{figure}
\subsection{Reconstructing Without Smoothing}
\label{sec:nosmoothing}
The reconstruction of the $a_{2m}$ without smoothing showed that for
$\textsc{Nside}=128$ the reconstruction was unbiased
(Fig.~\ref{fig:recon-a2m-C2-100-nside-128}) but for $\textsc{Nside}=16$ there was a
resolution dependent bias (Fig.~\ref{fig:recon-a2m-C2-100-nside-16}).
Calculating the weighted harmonic coefficient
estimator~(\ref{eq:estimator-Cl}) from these realisations produces the
results in Fig.~\ref{fig:recon-unbiased-multires-C2}. At first glance
these results are surprising and encouraging. The green, solid lines for
$\textsc{Nside}=128$ and red, dashed lines for $\textsc{Nside}=16$ nearly overlap and the
central value very closely follows the true value. This is surprising
since the $a_{2m}$ at $\textsc{Nside}=16$ are biased and have smaller variance than
the corresponding $\textsc{Nside}=128$ (see
Figs.~\ref{fig:recon-a2m-C2-100-nside-128} and
\ref{fig:recon-a2m-C2-100-nside-16}). Even so, when combined to determine
$C_2$ these differences average out and lead to nearly identical
predictions.
Based on Fig.~\ref{fig:recon-unbiased-multires-C2} we may think we have
solved the reconstruction problem; just reconstruct using the optimal,
unbiased $C_\ell$ estimator~(\ref{eq:estimator-Cl}) without smoothing!
Unfortunately we cannot draw this conclusion from the results presented
here. Recall that the reconstructions have been performed on noise-free,
pure CMB maps. Real maps contain noise and potentially residual, unmasked
foregrounds. In particular uncorrected, diffuse foregrounds are known to
contaminate the low-$\ell$ reconstruction~\citep{Naselsky2008}. A careful
study of the issues faced when applying the reconstruction to real data is
beyond the scope of this work and will be reserved for future
study. However, naive application of this method to real data yields highly
biased reconstructions.
\subsection{$\bmath{S_{1/2}}$ Estimator}
The study of the $S_{1/2}$ statistic is a large project in its own right
and will not be pursued in detail here. Our Universe as encoded in the ILC
map contains a somewhat small full-sky $S_{1/2}$ and an extremely small cut
sky $S_{1/2}$\@. If we are to perform such a statistical study of
$S_{1/2}$ we could enforce this structure, that is, only choose skies that
have somewhat low full-sky and very low cut-sky $S_{1/2}$ values.
Alternatively we could choose from an ensemble based on the best-fitting
$\Lambda$CDM model. In the latter case it has already been shown that the
ILC map is a rare realisation, unlikely at the $99.975\%$ level
\citep{CHSS-WMAP5}. The assumptions made in any study will determine the
statistical questions that can be asked. Conversely, the statistical
questions asked will implicitly contain the assumptions imposed.
In Table~\ref{tab:S12-ilc} we show the $S_{1/2}$ for the ILC map calculated
from~(\ref{eq:estimator-S12}) under various assumptions. Note that these
values all contain the bias discussed in Sec.~\ref{sec:S12} as is standard
in the literature. Shown in the table are the values calculated for the
full sky and for the partial sky where the KQ75y7 mask is employed to cut
out the Galactic region. The cut-sky results are calculated using the
pixel based estimator of \textsc{SpICE}{} and the optimal, unbiased $C_\ell$
estimator~(\ref{eq:estimator-Cl}) from reconstructed maps at $\textsc{Nside}=128$
and $\textsc{Nside}=16$. Further, the results are shown for different map
processing, including no processing (the unsmoothed entry where the map has
only been degraded as required for the reconstruction), employing a
$10\degr$ Gaussian smoothing, and filling the Galactic region with a
realisation that has the same power in each $\ell$-mode as the region
outside the mask but with the phases randomised.
The results in Table~\ref{tab:S12-ilc} are consistent with what we have
found for the $C_\ell$ reconstructions. For the unsmoothed map the full-sky
and pixel based estimators calculated at $\textsc{Nside}=128$ show the usual
result, the large discrepancy between the full and cut-sky values. This
holds true for the smoothed map also. Further, the reconstructed values
show the large discrepancy between the unsmoothed and smoothed maps (see,
for example, Figs.~\ref{fig:recon-a2m-C2-100-nside-16} and
\ref{fig:recon-a2m-C2-100-nside-16-10deg}). We also see that the
reconstructed values are systematically larger than the pixel based
estimator showing that the reconstruction is more sensitive to leakage for
information from inside the masked region. Finally the last line of the
table shows the expected behaviour for a map where the full sky has power
consistent with that from the cut-sky. Notice that of the cut-sky
pixel-based results are consistent with each other since information
leakage is unimportant. The small difference between the $\textsc{Nside}=128$ and
$\textsc{Nside}=16$ reconstructions shows the residual sensitivity on resolution.
\begin{figure}
\includegraphics[width=0.5\textwidth]{a3m_LCDM_r7_contours_100_with_ilc}
\caption{The $95$ and $5$ percentile lines for the $a_{3 m}$
reconstructed from the pixels outside the KQ75y7 mask at $\textsc{Nside}=128$
(and thus $\ellmax=514$) of $\Lambda$CDM realisations with
$C_2=100\unit{(\mathrm{\umu K})^2}$ and the masked region filled in with the ILC
map \textit{prior} to smoothing and rescaling, as discussed in the
text. The red, solid lines are for the real part of the $a_{3 m}$ and
the blue, dashed lines are for the imaginary part. The black, solid
line shows the expected result for a perfect reconstruction. We
clearly see the reconstructed $a_{3 m}$ are \textit{not} unbiased. Now
the bias is most prominent for $a_{33}$, the octopole mode with all its
extrema within the Galactic plane. This figure should be compared to
Fig.~\ref{fig:recon-a2m-C2-100-with-ILC}.}
\label{fig:recon-a3m-C2-100-with-ILC}
\end{figure}
\begin{figure}
\includegraphics[width=0.5\textwidth]{a4m_LCDM_r7_contours_100_with_ilc}
\caption{The same as Fig.~\ref{fig:recon-a3m-C2-100-with-ILC} now for
$\ell=4$.}
\label{fig:recon-a4m-C2-100-with-ILC}
\end{figure}
\subsection{Higher Multipoles}
In this work we have focused on how data handling affects the
reconstruction of the quadrupole. The quadrupole serves as an example
of the general behaviour. As show in
Figs.~\ref{fig:recon-a3m-C2-100-with-ILC} and
\ref{fig:recon-a4m-C2-100-with-ILC} we see the same results for $\ell=3$
and $\ell=4$. These figures were generated from the same realisations
employed in making Fig.~\ref{fig:recon-a2m-C2-100-with-ILC}. Again we
see that the reconstruction is biased toward the values from the ILC map.
\section{Conclusions}
\label{sec:conclusions}
It has been argued that the large-angle CMB can be reliably reconstructed
from partial-sky data and that when this is done the lack of large-angle
correlation is not significantly deviant from the expectation
\citep{Efstathiou2010}. At first glance the argument appears sound. The
large-angle modes extend over large fractions of the sky, thus knowing
their values on one region of the sky allows us to extrapolate them into
the masked regions. However, in practise and under close scrutiny this
argument fails. Implicit assumptions built in to the reconstruction
process enforce agreement between the reconstruction and the previously
constructed full sky (the ILC map in this case) through mixing of
information from inside the masked region to that outside. Due to this the
reconstruction has no value independent of the original full-sky map. It
neither confirms nor denies the validity of that map.
To study the large-angle CMB a choice must be made on what data to take as
a fair representation of the CMB sky One choice is to accept a cleaned,
full-sky map, such as the ILC map produced by \satellite{WMAP}, to accurately
represent the primordial CMB sky. In this case the full-sky map may be
analysed with no reconstruction required. In \cite{CHSS-WMAP5} and in this
work, however, we have taken the region outside the Galaxy as defined by
the \satellite{WMAP}{} KQ75y7 mask to be a fair representation. We have shown that the
large-angle CMB can be reconstructed using unbiased estimators for the
$a_{\ell m}$ and $C_\ell$, however the standard approach requires processing the
original map by degrading and smoothing it. Unfortunately it is precisely
the smoothing process that mixes the region we have taken as a fair
representation of the CMB with the region we are trying to exclude. When
the excluded region has the same statistical properties as the region we
are including then no biases are introduced. On the other hand, when, as
is the case with the ILC map, the properties are significantly different
the reconstruction is biased to agree with the full map. This is not
surprising. Through this process one is trusting the full-sky map, mixing
information from it into the rest of the sky, then reconstructing it. This
is a circular process and is unnecessary. If the full-sky map is already
trusted then there is no point in performing a reconstruction to produce a
poorer version of the original map.
The important point is that even \textit{in principle} reconstructing
following the standard approach leads to biased results unless the full-sky
CMB is already known. We have shown for noise free, pure CMB maps that
smoothing mixes information and biases the results. When applied to real
data the problems only get worse. Encouragingly we also found that
\textit{in principle} reconstructing without smoothing leads to unbiased
results. Unfortunately, directly applying this to real data with noise and
residual, unmasked foregrounds yields highly biased reconstructions
requiring further care to apply this method successfully to real-world CMB.
Overall the question of how to perform an unbiased reconstruction of the
full large angle CMB sky remains an interesting one. Previous work
\citep{Bielewicz2004,Naselsky2008,Liu2009,Aurich2010} has shown that
contamination significantly affects the reconstruction of the large angle
multipole moments. \cite{Aurich2010} studied the case most similar to that
considered in this work. They showed that smoothing of full sky map leaks
information from the pixels not used in the reconstruction (those in a
mask) to the pixels that will be used. In this work we have extended their
result and shown how a reconstruction such as that performed by
\cite{Efstathiou2010} is biased due to this leakage of information. This
shows the fundamental problem with trying to reconstruct the full sky from
a partial sky.
Fortunately large-angle CMB studies are not dependent on reconstructed
full-sky maps. The partial sky when used consistently (see
\citealt{CHSS-WMAP5}, for example) has been shown to be a robust
representation of the large scale CMB by \citet{Aurich2010} and in this
work. Despite the fact that such an approach is suboptimal in the sense
that the inferred $C_\ell$ do not have the smallest possible variance, it is
far less biased than the `optimal' $C_\ell$ inferred through the
maximum-likelihood reconstruction. More robust statements about the
large-angle CMB behaviour may therefore be made with the partial sky
pixel-based $C_\ell$.
We conclude that the lack of large-angle correlation, particularly on the
region of the sky outside the Galaxy, remains a matter of serious
concern.
\section*{Acknowledgments}
We thank Devdeep Sarkar for collaboration during initial stages of this
work. DH is supported by DOE OJI grant under contract DE-FG02-95ER40899,
and NSF under contract AST-0807564. DH and CJC are supported by NASA under
contract NNX09AC89G; DJS is supported by Deutsche Forschungsgemeinschaft
(DFG); GDS is supported by a grant from the US Department of Energy; both
GDS and CJC were supported by NASA under cooperative agreement
NNX07AG89G. This research was also supported in part by the NSF Grant
No.\ NSF PHY05-51164. This work made extensive use of the \textsc{HEALPix}{}
package~\citep{healpix}. The numerical simulations were performed on the
facilities provided by the Case ITS High Performance Computing Cluster.
\bibliographystyle{mn2e_new}
|
\section{Introduction}
Considerable interest has recently been devoted to the study of quantum
plasmas, see e.g.\ Refs \cit
{Shukla-Eliasson,Manfredi2005,Padma-Nature,Garcia2005,Misra2006,Haas2008,Cro-2008,haas2005
. Much of the research has been motivated by applications to e.g.\ quantum
wells \cite{Manfredi-quantum-well}, plasmonics \cite{Atwater-Plasmonics},
spintronics \cite{Spintronics}, astrophysics \cite{Astro} and ultra-cold
plasmas \cite{Ultracold}. Two of the most basic and much studied quantum
effects are those of the Fermi pressure and the particle dispersive effects
(directly associated with the Bohm De Broglie potential), see e.g.\ Refs.\ \cite{Shukla-Eliasson,Manfredi2005,Padma-Nature,Garcia2005,Misra2006,Haas2008,Cro-2008,haas2005}. Other studies
\cit
{Padma-Nature,Asenjo,Vladimirov,Lundin2010,brodin2007-1,brodin2008,brodin2010,zamanian2010,zamanian2010-2}
focus on the electron spin properties that result in the magnetic dipole
force and a magnetization current, in addition to some more complex aspects
of the spin dynamics. Although most quantum effects has a tendency to be
more important in plasmas of high density and low temperature, the regimes
of relevance differ to some extent for the various quantum effects, see
Ref.\ \cite{Lundin2010} for a discussion of this issue. A consequence is
that it is possible to focus on certain of the quantum effects and ignore
the others. In this paper we will make use of this fact and concentrate on
the physics associated with the spin coupling in the Pauli Hamiltonian. Our
starting point is a recenty presented spin-fluid model \cite{zamanian2010-2
, derived from kinetic theory \cite{zamanian2010}, which in addition to the
most basic spin precession dynamics includes effects of spin-velocity
correlations. Evaluating this model in the MHD regime, we make a conjecture
based on certain teoretical arguments; That there are two equivalent ways to
model spin-MHD dynamics, either by a one-fluid model including spin-velocity
correlations, or by a two-fluid model without spin-velocity correlations. In
the latter case the spin-up and spin-down states relative to the magnetic
field are regarded as different fluids \cite{brodin2008}. The conjectured
equivalence of these models in the MHD regime is tested by considering a
specific problem of three-wave interaction. For this purpose we calculate
the coupling coefficients between two shear Alfv\'en waves and one
compressional Alfv\'en wave (fast magnetosonic wave) in a magnetized plasma.
The coupling coefficients turns indeed out to be identical in the two cases,
and the coefficients are also seen to obey the Manley-Rowe relations, which
give further support to the soundness of the models used. The applicability
of the two-fluid model \textit{without spin-velocity correlations in the
MHD-regime, }which is strongly supported by our findings, is a very useful
result. The reason is that as this model can be easily adopted into standard
Particle-In-Cell schemes with only small modifications, as will be discussed
in the final section.
\section{Model equations}
Starting from a scalar kinetic equation for a spin-1/2 particle \cit
{zamanian2010}, spin fluid equations can be derived \cite{zamanian2010-2}.
These are given by the continuity equation
\begin{equation}
\partial _{t}n^{(s)}+\nabla \cdot (n^{(s)}\mathbf{v}^{(s)})=0
\label{eq:Continuity}
\end{equation
and the fluid momentum equation
\begin{equation}
m^{(s)}\frac{Dv_{i}^{(s)}}{Dt}=q^{(s)}\left( E_{i}+\varepsilon
_{ijk}v_{j}^{(s)}B_{k}\right) +\mu ^{(s)}S_{j}^{(s)}\frac{\partial B_{j}}
\partial x_{i}}-\frac{1}{n^{(s)}}\frac{\partial P_{ij}^{(s)}}{\partial x_{j}
, \label{eq:Momentum}
\end{equation
where the superscript $s=e,i$ denotes the species (electrons or ions),
D/Dt\equiv \partial _{t}+\mathbf{v}^{(s)}\cdot \nabla $, and summation over
repeated indices $i,j,k=x,y,z$ is implied. Here $m^{(s)}$ is the mass,
q^{(s)}$ is the charge, $\mu ^{(s)}$ the magnetic dipole moment, $n^{(s)}$
is the number density, and $\mathbf{v}^{(s)}$ is the fluid velocity of
species $s$. Furthermore, $\mathbf{S}$ is the spin vector normalized to
unity, $P_{ij}$ is the pressure tensor, and $\varepsilon_{ijk}$ is the
Levi-Civita symbol. Since the ions normally have a much smaller magnetic
moment than the electrons \cite{proton-note}, the spin contribution due to
the ions can be neglected compared to the electron contribution, i.e. we may
let $\mu ^{(i)}\approx 0$. We have here neglected contributions to the force
from the Fermi pressure and particle dispersive effects (the so called Bohm
de Broglie potential), see \cite{haas2005}. For a discussion of the
parameter regime of importance of various quantum effects, see e.g.\ Refs.\
\cite{Quant-classical,Lundin2010}. In this approximation, the pressure moment
satisfies the evolution equation
\begin{align}
\frac{DP_{ij}^{(s)}}{Dt}=& -P_{ik}^{(s)}\frac{\partial v_{j}^{(s)}}{\partial
x_{k}}-P_{jk}^{(s)}\frac{\partial v_{i}^{(s)}}{\partial x_{k}}-P_{ij}^{(s)
\frac{\partial v_{k}^{(s)}}{\partial x_{k}}+\frac{q^{(s)}}{m^{(s)}
\varepsilon _{imn}P_{jm}^{(s)}B_{n}+\frac{q^{(s)}}{m^{(s)}}\varepsilon
_{jmn}P_{im}^{(s)}B_{n} \notag \\
& +\frac{\mu ^{(s)}}{m^{(s)}}\Sigma _{ik}\frac{\partial B_{k}}{\partial x_{j
}+\frac{\mu ^{(s)}}{m^{(s)}}\Sigma _{jk}\frac{\partial B_{k}}{\partial x_{i}
, \label{Pressure}
\end{align
where again the last two terms can be dropped for the ion species.
Furthermore, to describe the spin dynamics we need the electron spin
evolution equation, which is given by
\begin{equation}
\frac{DS_{i}^{(e)}}{Dt}=\frac{2\mu ^{(e)}}{\hbar }\varepsilon
_{ijk}S_{j}^{(e)}B_{k}-\frac{1}{m^{(e)}n^{(e)}}\frac{\partial \Sigma
_{ij}^{(e)}}{\partial x_{j}} \label{eq:Spin}
\end{equation
where $\Sigma _{ij}^{(e)}$ is the spin-velocity correlation tensor. Finally
the evolution of the spin-velocity moment is described by
\begin{align}
\frac{D\Sigma _{ij}^{(e)}}{Dt}=& -\Sigma _{ij}^{(e)}\frac{\partial
v_{k}^{(e)}}{\partial x_{k}}-\Sigma _{ik}^{(e)}\frac{\partial v_{j}^{(e)}}
\partial x_{k}}-P_{jk}^{(e)}\frac{\partial S_{i}^{(e)}}{\partial x_{k}}
\notag \\
& +\frac{q^{(e)}}{m^{(e)}}\varepsilon _{jkl}\Sigma _{ik}^{(e)}B_{l}+\frac
2\mu ^{(e)}}{\hbar }\varepsilon _{ikl}\Sigma _{kj}^{(e)}B_{l}+\mu
^{(e)}n^{(e)}\frac{\partial B_{i}}{\partial x_{j}}-\mu
^{(e)}n^{(e)}S_{i}^{(e)}S_{k}^{(e)}\frac{\partial B_{k}}{\partial x_{j}}
\label{eq:SigmaA}
\end{align
In Eq.\ (\ref{Pressure}) we have neglected the heat flux tensor $Q_{ijk}$ to
obtain a closed set of equations. Similarly we have neglected the higher
order tensor $\Lambda _{ijk}$ in the evolution equation for the
spin-velocity tensor Eq.\ (\ref{eq:SigmaA}). The validity of the truncation
has been investigated in Refs.\ \cite{zamanian2010,stefan2011-1}, and the
truncation seem to be an accurate approximation in the low-temperature
limit. The equations above together with Maxwell's equations constitute a
closed system, where a magnetization current density $\mathbf{j}_M = \nabla
\times \mathbf{M}$, due to the spin, should be added to the free current
density, and where naturally all species contribute in the latter term. The
set of equations (\ref{eq:Continuity})-(\ref{eq:SigmaA}) has been studied by
Refs.\ \cite{stefan2011-1,zamanian2010-2}, but without inclusion of the ion
dynamics. The aim of the current paper is to apply the above set of
equations to the MHD regime where the ion dynamics is essential, at the same
time as making a careful evaluation of the electron spin magnetization.
\subsection{MHD-limit}
As concluded in the previous section, we will primarily consider wave
dynamics in the MHD regime where the frequencies are smaller than the
ion-cyclotron frequency and the wavelengths are longer than the Larmor
radius. Furthermore, we will consider the low-temperature (i.e. low-beta)
limit, where the pressure terms are dropped. Under these assumptions, the
system will be described by the magnetohydrodynamic equation \cit
{brodin2007-1}
\begin{equation}
\rho \left( \frac{\partial }{\partial t}+\mathbf{u}\cdot \nabla \right)
\mathbf{u}=-\nabla \left( \frac{B^{2}}{2\mu _{0}}-\mathbf{M}\cdot \mathbf{B
\right) +(\mathbf{B}\cdot \nabla )\mathbf{M}-\nabla P_{e} \label{eq:MHD}
\end{equation
together with the equations
\begin{equation}
\partial _{t}\mathbf{B}=\nabla \times \left( \mathbf{u}\times \mathbf{B
\right) . \label{eq:B-MHD}
\end{equation
and
\begin{equation}
\frac{\partial \rho }{\partial t}+\nabla \cdot \left( \rho \mathbf{u}\right)
= 0 \label{eq:Cont-MHD}
\end{equation
Here we have neglected the electron contribution to the fluid mass density
by setting $\rho \approx m^{(i)}n^{(i)}$, and the fluid velocity can be
written as $\mathbf{u}=(n^{(e)}m^{(e)}\mathbf{v}^{(e)}+n^{(i)}m^{(i)}\mathbf
v}^{(i)})/\rho \approx \mathbf{v}^{(i)}$. We have neglected the magnetic
moment of the ions such that the magnetization $\mathbf{M}=\mu ^{(e)}n^{(e)
\mathbf{S}^{(e)}$ is purely due to the electron spin. The derivation of Eq.\
(\ref{eq:MHD}) made in Ref.\ \cite{brodin2007-1} was done starting from a
somewhat less elaborate set of equations, not including the spin-velocity
correlations. However, the derivation does not involve the Eqs.\ (\re
{eq:Spin})-(\ref{eq:SigmaA}) describing the spin dynamics, and hence we may
adopt this result within the current model.
Without the magnetization $\mathbf{M}$ we obtain standard ideal MHD
equations, and thus (\ref{eq:MHD}), (\ref{eq:B-MHD}) and (\ref{eq:Cont-MHD})
constitute a closed system. The magnetization can be easily included with
the spin determined by Eqs.\ (\ref{eq:Spin}) together with (\ref{eq:SigmaA
), where the terms containing derivatives of the velocity turns out to be
negligible. The aim is to solve for the magnetization in terms of the
magnetic field, in which case Eqs.\ (\ref{eq:MHD}), (\ref{eq:B-MHD}) and
\ref{eq:Cont-MHD}) are sufficient to produce a closed spin-MHD theory. This
can be achieved in two different ways; either by considering the electrons
in spin up and spin down states relative the magnetic field as two separate
fluids, or by treating them as a single fluid with a macroscopic spin that
is proportional to the difference in population density of the two spin
states. We will now go on to discuss this in further detail.
\subsection{One-fluid vs two-fluid}
In this sub-section we will discuss how to determine the magnetization, in
order to use Eqs.\ (\ref{eq:MHD}), (\ref{eq:B-MHD}) and (\ref{eq:Cont-MHD}).
To find the magnetization we first need to solve (\ref{eq:Spin}) to
determine the spin. The first term of the right hand side of Eq.\ (\re
{eq:Spin}) is the basic spin precession. If the spin-velocity correlations
in (\ref{eq:Spin}) can be neglected, the solutions for $\mathbf{S}$ are
particularly simple in the MHD regime. This is because the left hand side
term of Eq.\ (\ref{eq:Spin}) is smaller than the spin \ precession term by a
factor of the order $O(\omega ^{(\mathrm{ch})}/\omega _{cg}^{(\mathrm{ch})})
, where $\omega ^{(\mathrm{ch})}\sim \partial _{t}$ is a characteristic
frequency scale of the problem, and $\omega _{cg}^{(\mathrm{ch})}\sim 2\mu
_{e}B/\hbar $ is the characteristic spin precession frequency (which is
close to the characteristic cyclotron frequency $\omega _{c}^{(\mathrm{ch
)}\sim qB/m$). Assuming that spin-velocity correlations can be omitted, we
note that spin evolution equation in the MHD regime reduces to
\begin{equation}
\varepsilon _{ijk}S_{j}^{(e)}B_{k}=0. \label{Eq-spin-precession}
\end{equation
This has two solutions, where $\mathbf{S}$ is either parallel or
antiparallel to $\mathbf{B}\,$, that is $S_{i}=\pm b_{i}$, where
b_{i}=B_{i}/B$ is a unit vector in the direction of $\mathbf{B}$. A
comparatively general way to deal with spins obeying $S_{i}=\pm b_{i}$ is to
consider a two-fluid model of electrons, where for one of the species the
electron spin state is parallel to $\mathbf{B}$, and for the other species
antiparallel. Eq.\ (\ref{Eq-spin-precession}) then implies that these spin
states are conserved. However, as seen from the above discussion this is
only an adequate approximation if the spin-velocity correlations give a
small contribution in Eq.\ (\ref{eq:Spin}). Thus our next step is to outline
the solutions of Eq.\ (\ref{eq:SigmaA}) using MHD approximations, in order
to determine the contribution from $\Sigma _{ij}^{(e)}$ in (\ref{eq:Spin}).
Firstly we note that the three first terms in (\ref{eq:SigmaA}) are at most
of order $\omega ^{(\mathrm{ch})}\Sigma _{ij}$, whereas the fifth and sixth
are of order $\omega _{c}^{(\mathrm{ch})}\Sigma _{ij}$ $\sim \omega _{cg}^{
\mathrm{ch})}\Sigma _{ij}$. Thus neglecting the three first terms we can
formally write Eq.\ (\ref{eq:SigmaA}) on the form $\overleftrightarrow{O
\cdot \overrightarrow{\Sigma }=\overrightarrow{\sigma }$, where
\overleftrightarrow{O}$ is a $9\times 9$-matrix where all coefficients are
\pm \omega _{c\alpha }$ or $\pm \omega _{cg\alpha }$. Here $\overrightarrow
\Sigma }$ is a $9$-component vector containing all elements of $\Sigma _{ij}
, and $\overrightarrow{\sigma }$ is a $9$-component vector containing the
source terms, i.e.\ $P_{jk}(\partial S_{i}/\partial x_{k})$, $\mu n(\partial
B_{i}/\partial x_{j})$ and $\mu nS_{i}S_{k}(\partial B_{k}/\partial x_{j})$,
$\omega _{c\alpha }=qB_{\alpha }/m$ and $\omega _{cg\alpha }=2\mu
_{e}B_{\alpha }/\hbar $ with $\alpha =x,y,z$. Note that the full field
strength is used and not the linearized field when defining $\omega
_{c\alpha }$ and $\omega _{cg\alpha }$. Since $\overleftrightarrow{O}$
contains no operators, we can do a simple matrix inversion to find
\overrightarrow{\Sigma }=\overleftrightarrow{O}^{-1}\overrightarrow{\sigma }
. This turns out to be sufficient to determine all components of $\Sigma
_{ij} $, except for a component directed as $\mathbf{b}\otimes \mathbf{b}$.
Thus this approximation scheme allows us to compute $\Sigma _{ij}$ apart
from a contribution that can be expressed as $\Sigma _{ij}=\Phi b_{i}b_{j}$,
where $\Phi $ is a scalar field. Using the solution $\overrightarrow{\Sigma
=\overleftrightarrow{O}^{-1}\overrightarrow{\sigma }$ we can easily check
that the determined components of $\Sigma _{ij}$ are of order $\Sigma _{ij}$
$\sim \mu _{B}nkB/\omega _{c}^{(\mathrm{ch})}$. This means that the
contributions from $\Sigma _{ij}$ are sufficiently small to be neglected in
\ref{eq:Spin}). However, in general we must also account for the
contribution $\Sigma _{ij}=\Phi b_{i}b_{j}$ whose magnitude is unknown.
Since the scalar field $\Phi $ cannot be determined if the three first terms
of (\ref{eq:SigmaA}) are omitted, we must extend our model to the solve the
full case of Eq.\ (\ref{eq:SigmaA}). We will do so within a one-fluid model
in the section III A, and it turns out that $\Phi $ becomes sufficiently
large for this component of $\Sigma _{ij}$ to signficantly influence the
solutions to (\ref{eq:Spin}), also within the MHD-regime. However, it also
turns out that in order to get a large value of $\Phi $, we must have a spin
vector that is different from $\pm \mathbf{b}$. This is the normal case in a
one-fluid theory, where the macroscopic spin results from averaging over all
spin states. However, within a two-fluid MHD model (treating spin-up and
down states as different species) without spin-velocity correlations Eq.\
\ref{Eq-spin-precession}) can be applied leading to $S_{i}=\pm b_{i}$, and
the situation would then again be modified. Indeed, contracting Eq.\ (\re
{eq:SigmaA}) with $b_{i}b_{j}$ to compute the source terms for $\Phi $, we
find that all the source terms for $\Sigma _{ij}$ vanishes if $S_{i}=\pm
b_{i}$, as the fourth term in Eq.\ (\ref{eq:SigmaA}) becomes
b_{i}b_{j}P_{jk}(\partial S_{i}/\partial x_{j})$, which is zero as
b_{i}(\partial S_{i}/\partial x_{j})=\pm (1/2)\partial (b_{i}b_{i})/\partial
x_{j}$, whereas terms 7 and 8$\,$together become
\begin{equation*}
b_{i}b_{j}\frac{\partial B_{i}}{\partial x_{j}}-b_{i}b_{j}S_{i}S_{k}\frac
\partial B_{k}}{\partial x_{j}}=0
\end{equation*
where $S_{i}=\pm b_{i}$ was used in the last step. This result provides the
theoretical basis for adopting a two-fluid model of electrons in the
MHD-regime and omitting spin-velocity correlations in (\ref{eq:Spin}),
leading to $S_{i}=\pm b_{i}$. The division into two fluids leaves the rest
of the basic equations structurally unaffected, but we now obatin two
contributions such that the magnetization is calculated as $\mathbf{M}=\mu
n_{\uparrow }\mathbf{s}_{\uparrow }+\mu n_{\downarrow }\mathbf{s
_{\downarrow }$ due to the difference in density perturbations of the two
spin states. We will consider this in more detail within perturbation theory
in our model problem below. The conclusions of this section is then
confirmed, since the one-fluid models that keeps the spin-velocity
correlations in Eq.\ (\ref{eq:Spin}) give indeed an identical expression for
the magnetization as the two-fluid model with up and down spins $S_{i}=\pm
b_{i}$. The allowance for independent density variations of the two species
in the latter model provides the physical mechanism that reproduces the
effects of spin-velocity correlations in the one-fluid model. It should
however be stressed that this conclusion is limited to the MHD regime.
\section{Three wave interaction - a model problem}
We will now consider a model problem with the purpose of testing our
conclusions about the similarities between the one-fluid and two-fluid
models outlined in the previous section. Specifically, we consider three
wave interaction between two shear Alfv\'en waves ($A,A^{\prime }$) and one
compressional Alfv\'en wave ($MS$); $MS\rightarrow A+A^{\prime }$. Using
three-wave interaction as a model problem has the advantage that an
unphysical assumption (or an incorrect calculation) is likely to result in a
broken Manley-Rowe symmetry \cite{Manley-Rowe}, in which case one gets a
clear indication that something needs to be revised.
The waves are assumed to be small perturbations on a homogeneous background,
and we write $\mathbf{B}=B_{0}\hat{\mathbf{z}}+\mathbf{B}_{1}$, $\rho =\rho
_{0}+\rho _{1}$, etc., but omit the index 1 on variables whose backgound
values are zero. Furthermore, we omit index 1 whenever the cartesian
components are specified for notational convenience, i.e. we write $\mathbf{
}_{1}=B_{x}\hat{\mathbf{x}}+B_{y}\hat{\mathbf{y}}+B_{z}\hat{\mathbf{z}}$.
Moreover, we assume that there is no drift so that $\mathbf{u}_{0}=0$ and
also that there is no spin-velocity correlation in the background
distribution, i.e. $\Sigma _{0}=0$. For simplicity we further assume that
the temperature is sufficiently low such that the equilibrium pressure can
be neglected, $P_{0}=0 $ (i.e. that we have a low-beta plasma with the
ion-acoustic velocity much smaller than the Alfv\'en velocity). Furthermore,
assuming that $\mu _{B}B_{0}/(k_{B}T)\ll 1$ we can make the approximation
that the equilibrium spin up and spin down populations are equal so that
n_{0\uparrow }=n_{0\downarrow }$ in the two-fluid model which implies that
the total zeroth order magnetization vanishes \cite{Magnetization-note}. For
consistency between the one-fluid and two-fluid models we should
consequently pick $\mathbf{M}_{0}=0$ also in the latter case. The difference
in the model equations between the one-fluid and two-fluid approach is then
primarily that in the two-fluid model we have $\mathbf{S}_{0}=\pm \hat
\mathbf{z}}$ for the two spin states (in which case we obtain a finite zero
order magnetization if only one of the electron fluids are counted) whereas
in the one-fluid model $\mathbf{M}_{0}=0$ and $\mathbf{S}_{0}=0$.
Next we make a harmonic decomposition $\partial _{t}\rightarrow -i\omega $
and $\partial _{\mathbf{x}}\rightarrow i\mathbf{k}$ for each wave, where the
frequencies and wave vectors satisfy the conditions
\begin{align}
\omega ^{MS}& =\omega ^{A}+\omega ^{A^{\prime }} \label{matching1} \\
\mathbf{k}^{MS}& =\mathbf{k}^{A}+\mathbf{k}^{A^{\prime }}. \label{matching2}
\end{align
with the index $MS$ denoting the compressional Alfv\'en (or fast magnetosonic)
wave, and the index $A$ and $A^{\prime }$ denoting the shear Alfv\'en waves.
The coordinate system is defined so that the $z$-direction points in the
direction of the unperturbed magnetic field, $\mathbf{B}=B_{0}\hat{\mathbf{z
}$, and for simplicty we assume all wave vectors to lie in the $xz$-plane.
Throughout the calculation we will use $\omega /\omega _{c}$ and
kC_{A}/\omega _{c}$ as small expansion parameters (where $\omega
_{c}=qB_{0}/m$ is the cyclotron frequency), in accordance with standard MHD
theory. Here $C_{A}=(B_{0}^{2}/\mu _{0}m_{i}n_{0})^{1/2} =
(B_0^2/\mu_0\rho_0)^{1/2}$ is the Alfv\'en velocity, and $\omega $ and $k$
represents any of the wave frequencies or wave vector components. We also
note that $\omega _{cg}\simeq \omega _{c}$, where $\omega _{cg}=$ $2\mu
_{e}B_{0}/\hbar $ is the spin precession frequency. Furthermore, $\omega
_{cg}-\omega _{c}$ is of the same order as the ion-cyclotron frequency, and
is therefore much larger than wave frequencies within the MHD regime \cit
{Frequency-note}. We will therefore drop terms proportional to $(\omega
_{cg}-\omega _{c})^{-1}$ compared to $\omega ^{-1}$ in our final results.
It should be pointed out that unlike the classical case, the pressure tensor
$P_{ij}$ does not necessarily vanish in the limit of zero temperature.
However, we note that we need not be concerned about the contribution from
the pressure term in this particular case. This is because $P_{ij}$ vanishes
linearly in the MHD limit and thereby enters as a cubic nonlinearity (which
does not affect the three-wave interaction) in the evolution equation for
\Sigma _{ij}$. The pressure tensor, however, also gives a contribution in
the MHD equation \eqref{eq:MHD}, but it turns out that this is a nonlinear
contribution proportional to $(\omega _{cg}-\omega _{c})^{-1}$ which is
small compared to leading terms proportional to $\omega ^{-1}$. We may
therefore neglect the contribution from $P_{ij}$ altogether.
\subsection{One-fluid calculation}
We start by considering the one fluid model for which, as mentioned above,
the unperturbed spin-density is zero, $\mathbf{S}_{0}=0$. Our first aim is
to find the linear dispersion relation as well as the linear eigenvectors
(polarizations) of the shear Alfv\'en wave and the compressional Alfv\'en wave.
We note that in the MHD equation \eqref{eq:MHD} we need an expression for
the magnetization. We therefore begin by solving the spin-velocity evolution
equation to find the $\Sigma $-tensor. Linearly, this is straightforward and
we find
\begin{equation}
\Sigma _{ij}=in_{0}\mu \left(
\begin{array}{ccc}
-\frac{k_{x}B_{y}\omega _{cg}}{\omega _{c}^{2}-\omega _{cg}^{2}} & -\frac
k_{x}B_{x}\omega _{c}}{\left( \omega _{c}^{2}-\omega _{cg}^{2}\right) } &
\frac{k_{z}B_{y}}{\omega _{cg}} \\
\frac{k_{x}B_{x}\omega _{cg}}{\left( \omega _{c}^{2}-\omega _{cg}^{2}\right)
} & -\frac{k_{x}B_{y}\omega _{c}}{\omega _{c}^{2}-\omega _{cg}^{2}} & -\frac
k_{z}B_{x}}{\omega _{cg}} \\
0 & -\frac{k_{x}B_{z}}{\omega _{c}} & i\frac{k_{z}B_{z}}{\omega
\end{array
\right) . \label{eq:linear-sigma-tensor}
\end{equation
As can be seen, the components in (\ref{eq:linear-sigma-tensor}) have
different magnitudes, but we keep all of them at this stage in the
calculation. Next we use the linear $\Sigma $-tensor in the spin evolution
equation \eqref{eq:Spin} to find an expression for the linear spin $\mathbf{
}$ and thereby the linear magnetization $\mathbf{M}=\mu n_{0}\mathbf{S}$.
\begin{equation}
\mathbf{M}=\frac{n_{0}\mu ^{2}}{m}\left(
\begin{array}{c}
\frac{k_{x}^{2}B_{x}}{(\omega _{c}^{2}-\omega _{cg}^{2})} \\
\frac{k_{x}^{2}B_{y}}{(\omega _{c}^{2}-\omega _{cg}^{2})} \\
-\frac{k_{z}^{2}B_{z}}{\omega ^{2}
\end{array
\right) \label{linear-magn}
\end{equation
Here we have dropped contributions to components of $\mathbf{M}$ that are
smaller by factors $\left\vert \omega _{c}^{2}-\omega _{cg}^{2}\right\vert
/\omega _{cg}^{2}$ and/or $\omega /\omega _{cg}$. \ Substituting (\re
{linear-magn}) into (\ref{eq:MHD}), the linear dispersion relations are
obtained from (\ref{eq:MHD}), (\ref{eq:B-MHD}) and (\ref{eq:Cont-MHD}).
Similarly to the classical ideal MHD case, the modes decouple into the shear
Alfv\'en wave described b
\begin{equation}
D_{A}(\omega ,\mathbf{k})\equiv \omega ^{2}-k_{z}^{2}C_{A}^{2}\left( 1-\frac
n_{0}\mu _{0}\mu ^{2}}{m}\frac{k_{x}^{2}}{(\omega _{c}^{2}-\omega _{cg}^{2})
\right) =0 \label{Shear-DR}
\end{equation
and the compressional Alfv\'en (fast magnetosonic) wave with the dispersion
relation
\begin{equation}
D_{MS}(\omega ,\mathbf{k})\equiv \omega ^{2}-k_{x}^{2}C_{A}^{2}\left( 1
\frac{n_{0}\mu _{0}\mu ^{2}}{m}\frac{k_{z}^{2}}{\omega ^{2}}\right)
-k_{z}^{2}C_{A}^{2}\left( 1-\frac{n_{0}\mu _{0}\mu ^{2}}{m}\frac{k_{x}^{2}}
(\omega _{c}^{2}-\omega _{cg}^{2})}\right) =0 . \label{Compressional-DR}
\end{equation
Note that the last terms in (\ref{Shear-DR}) and (\ref{Compressional-DR})
are smaller than the first spin-modification in (\ref{Compressional-DR}) as
\omega ^{2}\ll \left\vert \omega _{c}^{2}-\omega _{cg}^{2}\right\vert $.
Furthermore, the linear eigenvector components for the shear Alfv\'en wave are
$u_{x}^{A}=u_{z}^{A}=0$, $B_{x}^{A}=B_{z}^{A}=0$, $\rho _{1}^{A}=0$, and
\begin{equation}
B_{y}^{A}=-\frac{k_{z}^{A}B_{0}}{\omega ^{A}}u_{y}^{A} \label{linear-1}
\end{equation
For the compressional Alfv\'en wave (index $MS$) we instead obtain
u_{y}^{MS}=u_{z}^{MS}=0$, $B_{y}^{MS}=0$, and
\begin{eqnarray}
B_{x}^{MS} &=&-\frac{k_{z}^{MS}B_{0}}{\omega ^{MS}}u_{x}^{MS}
\label{linear-ms-1} \\
B_{z}^{MS} &=&\frac{k_{\bot }^{MS}B_{0}}{\omega ^{MS}}u_{x}^{MS}
\label{linear-ms-2} \\
\rho _{1}^{MS} &=&\rho _{0}\frac{k_{\bot }^{MS}}{\omega ^{MS}}u_{x}^{MS}
\label{linear-ms-3}
\end{eqnarray
Next we aim to calculate the three wave coupling coefficients due to the
quadratic nonlinearities. We have calculated the nonlinear contribution to
the coupling coefficients including all terms proportional to $(\omega
_{cg}-\omega _{c})^{-1}$. However, our results show that these terms only
give rise to small corrections to the leading terms proportional to $\omega
^{-1}$. Since the full analysis is rather tedious we will therefore only
write out the leading terms in the NL contribution to the $\Sigma $-tensor
as well as to the magnetization $\mathbf{M}$. Under the given
approximations, keeping the resonant terms, we find the components with a
nonzero nonlinear contribution to the $\Sigma $-tensor to be
\begin{equation}
\Sigma _{yz}^{A^{\prime }}=-\frac{k_{z}B_{x}^{A^{\prime }}}{\omega _{cg}}
\frac{2i\mu }{\hbar }\frac{k_{z}^{MS}B_{z}^{MS}B_{y}^{A^{\ast }}}{\omega
_{cg}\omega ^{MS}} \label{Sigma-NLA1}
\end{equation
an
\begin{equation}
\Sigma _{zy}^{A^{\prime }}=-\frac{k_{x}B_{z}}{\omega _{c}}+\frac{iq}{m}\frac
k_{z}^{MS}B_{z}^{MS}B_{y}^{A^{\ast }}}{\omega _{c}\omega ^{MS}}
\label{Sigma-NLA2}
\end{equation
for the Alfv\'en wave. Here $*$ denotes complex conjugation. For the
magnetosonic wave the component with a nonzero nonlinear contribution i
\begin{equation}
\Sigma _{33}^{MS}=i\frac{k_{z}B_{z}^{MS}}{\omega }+\frac{2i\mu }{\hbar
\frac{1}{\omega _{cg}\omega ^{MS}}\left( k_{z}^{A}+k_{z}^{A^{\prime
}}\right) B_{y}^{A^{\prime }}B_{y}^{A} \label{Sigma-NL-MS}
\end{equation
Solving the spin evolution equation with the sources from $\Sigma $ given
above, we find the a nonlinear contribution to the $z$-component of the
magnetization
\begin{equation}
M_{z}^{MS}=-\frac{n_{0}\mu ^{2}}{m}\left( \frac{k_{z}^{2}B_{z}^{MS}}{\omega
^{2}}+\frac{2\mu }{\hbar }\frac{k_{z}^{MS}\left( k_{z}^{A}+k_{z}^{A^{\prime
}}\right) }{\omega _{cg}\omega ^{2(MS)}}B_{y}^{A}B_{y}^{A^{\prime }}\right)
\label{NL-magn1}
\end{equation
for the MS wave, and a nonlinear contribution to the $y$-component of the
magnetization
\begin{equation}
M_{y}^{A^{\prime }}=\frac{n_{0}\mu ^{2}}{m}\left( \frac{k_{x}^{2}B_{y}}
(\omega _{c}^{2}-\omega _{cg}^{2})}-\frac{2\mu }{\hbar }\frac{k_{z}^{2(MS)}}
\omega _{cg}\omega ^{2(MS)}}B_{z}^{MS}B_{y}^{A\ast }\right) \label{NL-magn2}
\end{equation
for the Alfv\'en wave. Now that we have expressed the magnetization in terms
of the magnetic field, correct to second order in the amplitude, we may
substitute these results into (\ref{eq:MHD}), and perform the rest of the
calculations using (\ref{eq:MHD}), (\ref{eq:B-MHD}) and (\ref{eq:Cont-MHD})
as in standard MHD theory. Accounting for time dependent amplitudes with the
substitions $D_{A}(\omega ,\mathbf{k})\rightarrow \lbrack \partial
D_{A}/\partial \omega ]i\partial /\partial t$ and $D_{MS}(\omega ,\mathbf{k
)\rightarrow \lbrack \partial D_{MS}/\partial \omega ]i\partial /\partial t
, doing successive elimination keeping the velocity variables as the wave
amplitudes, we find the following coupled equations for the different wave
modes
\begin{equation}
\frac{\partial u_{y}^{A^{\prime }}}{\partial t}=-i\frac{\omega ^{2(A^{\prime
})}}{\partial D_{A^{\prime }}/\partial \omega }Cu_{y}^{A^{\ast }}u_{x}^{MS}
\label{Wave-NL-1}
\end{equation
and
\begin{equation}
\frac{\partial u_{x}^{MS}}{\partial t}=-i\frac{\omega ^{2(MS)}}{\partial
D_{MS}/\partial \omega }Cu_{y}^{A}u_{y}^{A^{\prime }} \label{Wave-NL-2}
\end{equation
with the coupling coefficient
\begin{equation}
C=\frac{k_{x}^{MS}}{\omega ^{MS}}\left( 1+\frac{n_{0}\mu _{0}\mu ^{2}}{m
\frac{k_{z}^{2(MS)}}{\omega ^{2MS}}\right) . \label{Coeff}
\end{equation
Due to the symmetry between the two shear Alfv\'en waves, the equation for
\partial u_{y}^{A}/\partial t$ is obtained by exchanging $A$ and $A^{\prime
} $ in Eq.\ (\ref{Wave-NL-1}). The appearance of the common factor $C$ in
the three coupled equations is a reflection of the Manley-Rowe symmetry \cit
{Manley-Rowe}. The first term of $C$ is a purely classical contribution,
that agrees with e.g.\ Refs.\ \cite{Chin1972,Classical-C} in the cold limit.
For the spin contribution in Eq.\ (\ref{Coeff}) \ to be important as
compared to the classical one, a rather dense plasma is required. By
contrast, other MHD phenomena exists that require less extreme parameters
for the electron spin to be important. Nevertheless, as will be discussed in
the final section, the results derived here have a number of interesting
theroretical consequences. It should be stressed that the contribution to
the magnetization in this one-fluid model stems from the $\Sigma $-tensor.
This is in contrast to the two-fluid model as we will see below.
\subsection{Two-fluid calculation}
We now consider the problem of three wave coupling using the two-fluid
model. The spin is then determined from (\ref{eq:Spin}) with the
contribution from $\Sigma $ omitted, as described in section II, but we now
have two species of electrons, which have the unperturbed spin $\mathbf{S
_{0\uparrow }=\hat{\mathbf{z}}$ and $\mathbf{S}_{0\downarrow }=-\hat{\mathbf
z}}$, respectively. The total magnetization is then written as
\begin{equation}
\mathbf{M}=\mu \left( n^{(\uparrow )}\mathbf{S}^{(\uparrow)}+n^{(\downarrow)
}\mathbf{S}^{(\downarrow) }\right) \label{eq:Magn-2f}
\end{equation
which gives $\mathbf{M}_{0}=0$ in agreement with the previous section,
provided we let $n_{0\uparrow }=n_{0\downarrow }=n_{0}/2$ which will be used
henceforth. Next we find the linear spin-vector to be
\begin{equation}
\mathbf{S}_{1}=\left(
\begin{array}{c}
\frac{2\mu }{\hbar \omega _{cg}}S_{0}B_{x}+\frac{\mu }{m}\frac{k_{x}^{2}B_{x
}{\omega _{c}^{2}-\omega _{cg}^{2}} \\
\frac{2\mu }{\hbar \omega _{cg}}S_{0}B_{y}+\frac{\mu }{m}\frac{k_{x}^{2}B_{y
}{\omega _{c}^{2}-\omega _{cg}^{2}} \\
\end{array
\right) . \label{spin-two-fluid}
\end{equation
Note here that although the terms $\propto S_0$ in (\ref{spin-two-fluid})
are larger than the terms $\propto (\omega _{c}^{2}-\omega _{cg}^{2})^{-1}$,
the former has opposite signs for the up- and down species, and hence give
no contribution to the linear magnetization. It turns out that the terms in
\ref{spin-two-fluid}) $\propto (\omega _{c}^{2}-\omega _{cg}^{2})^{-1}$ are
needed to get agreement with the linear magnetization obtained with the
one-fluid model \eqref{linear-magn}. However, it can be noted that these
terms have been dropped in the expression for the coupling coefficient
\eqref{Coeff} where, in the end, only the leading term is kept. Next we need
to find an expression for the fluid densities of the electron spin fluids. From
the continuity equation \eqref{eq:Continuity} we have
\begin{align}
n^{(s)} = n_0^{(s)} + n_0^{(s)}\frac{\mathbf{k}\cdot\mathbf{v}^{(s)}}{\omega}
+ n^{(s)\text{NL}}
\end{align}
where $n^{(s)\text{NL}}$ is a non-linear contribution that can be shown not
to contribute to the magnetization after summation of the spin states
s=\uparrow,\downarrow$. An expression for the electron velocities is
obtained by solving the fluid momentum equation \eqref{eq:Momentum} together
with $-\partial _{t}\mathbf{B}=\nabla \times \mathbf{E}$. To close this
system we may in general need to use a full fluid description. However,
within the MHD regime and for this specific problem, it suffices to
determine the magnetization. For our case with $n_{0\uparrow
}=n_{0\downarrow }=n_{0}/2$ several terms vanish in Eq.\ (\ref{eq:Magn-2f})
after the summation over up and down species. In MHD we make the
approximation that we may write the electric field as $\mathbf{E}=-\mathbf{v
^{(i)}\times \mathbf{B\simeq }-\mathbf{u}\times \mathbf{B}$. This allows us
to again make use of Eqs.\ (\ref{linear-1})-(\ref{linear-ms-3}). Under these
assumptions we note that $E_{z}$ vanishes linearly, and also that $E_{z}^{NL}
$ is only proportional to quadratic combinations of $B$-field components and
will therefore not contribute to the magnetization. Thus $E_{z}$ may
therefore be set to zero from now on. Under these assumptions, it is easy to
show that the fluid velocities may be written as
\begin{subequations}
\label{eq:v}
\begin{align}
v_{x}& =\frac{q}{m}\frac{\omega }{\omega _{c}}\frac{B_{z}}{k_{x}}+\frac{q}{m
\frac{1}{\omega _{c}}\left( -v_{x}B_{z}+v_{z}B_{x}\right) \\
v_{y}& =-\frac{q}{m}\frac{\omega }{\omega _{c}}\frac{B_{y}}{k_{z}}+i\frac
\mu }{m\omega _{c}}k_{x}B_{z}S_{0}-\frac{q}{m}\frac{1}{\omega _{c}}\left(
v_{y}B_{z}-v_{z}B_{y}\right) -i\frac{\mu }{m\omega _{c}}k_{x}\left(
B_{x}S_{x}-B_{y}S_{y}\right) \\
v_{z}& =-\frac{\mu }{m\omega }k_{z}B_{z}S_{0}+i\frac{q}{m}\frac{1}{\omega
\left( v_{x}B_{y}-v_{y}B_{x}\right) -\frac{\mu }{m\omega }k_{z}\left(
B_{x}S_{x}-B_{y}S_{y}\right)
\end{align
\end{subequations}
From Eqs.\ \eqref{eq:Magn-2f}--\eqref{eq:v} we find the linear
magnetization, and it agrees identically with the expression obtained from
the one-fluid model. Furthermore, an extended analysis gives agreement for
the NL terms of the magnetization as well. \ Consequently, the coupling
coefficients remain the same regardless if the one-fluid or two-fluid model
are used to determine the magnetization. This corroborates the usefulness of
the two fluid model in the MHD regime.
\section{Discussion}
In the present paper we have studied a recently presented fluid model
accounting for the electron spin \cite{zamanian2010-2}, and adopted it to
the MHD regime. The main feature of the original model is that in addition
to basic spin effects as the magnetic dipole force, spin precession, and the
magnetization current it incorporates spin-velocity correleations. The
spin-velcoity correlations have been shown to be important for a number of
spin plasma phenomena \cite{zamanian2010-2,stefan2011-1}. Introducing the
approximations appropriate for the MHD regime, it turns out that essentially
the ordinary MHD equations are recovered, but with a magnetization that
needs to be determined. This can be done in different ways. Firstly from a
single fluid model of electrons, that besides the basic spin precession
contains spin-velocity correlations. Or, secondly, from a two-fluid model
where spin-up and spin-down electrons constitute different species. A
theoretical argument is presented in Section II B suggesting that these two
models are equivalent in the MHD regime. However, the equivalence argument
depends on certain assumptions which is difficult to justify rigorously, and
thus practical tests of the equivalence is valuable. For this purpose we
have evaluated the different models using a nonlinear three wave interaction
as a test problem. Specifically, we have computed the coupling coefficients
between two shear Alfv\'en waves and a compressional Alfv\'en wave. A classical
as well as quantum mechanical (spin) contribution to the coupling
coefficients are found, and the coupling coefficients are indeed identical
in the two models. Furthermore, the coupling cofficients obey the
Manley-Rowe symmetries \cite{Manley-Rowe}. The Manley-Rowe relations is a
reflection of the underlying Hamiltonian structure \cite{Larsson99} of the
model. The fact that the coefficients preserve these relations strongly
suggests that the approximations made when deriving the models are sound, as
otherwise it is highly likely that the Manly-Rowe symmetries would be broken.
Since we have put forward two somewhat different models in this paper, one
may ask which one that is most easy to use. For the analytical calculations
made here, the degree of complexity is found to be roughly the same.
However, the two-fluid model has a great advantage in case one would like to
do Particle-In-Cell (PIC) simulations. In that case, the only modification
of a standard code would be to have two species of electrons, and add a
force proportional to $\pm \mu _{e}\nabla B$ in the momentum equation, as
well as to compute $\mathbf{M}=\mu _{e}\mathbf{b}(n_{\uparrow
}-n_{\downarrow })$ to find the contribution from the Magnetization current
in Amperes law. As the variables for the density and magnetic field are
monitored throughout the PIC-simulations anyway, no new equations and only
little extra complexity is added to the general concept. Developing a
PIC-scheme incorporating spin-velocity correlations, however, is a much more
cumbersome project, as the evolution equation for this object is not present
in present PIC-schemes, and also such equations are considerably more
complex. Furthermore, it is not clear that it is at all possible to model
spin-velocity correlations as a single-particle property, which makes the
adaption of this model to the PIC-scheme questionable conceptually. Thus we
conclude that the two-fluid model is valuable for the purpose of
incorporating electron spin effects in PIC-simulations, although we stress
that the applicability of such an approach will be limited to the MHD regime.
|
\section{Introduction}\label{sec:intro}
This paper is a short review of recent work on constructing smooth Calabi-Yau
threefolds with interesting topological properties, such as small cohomology
groups and non-trivial fundamental group. In practice, these two properties often
go hand-in-hand, as emphasised in \cite{Triadophilia,Candelas:2008wb}.
The majority of known three-dimensional Calabi-Yau manifolds are constructed as
complete intersections in higher-dimensional toric varieties \cite{Green:1986ck,
Candelas:1987kf,Green:1987cr,Batyrev:1994pg,Kreuzer:1995cd,Kreuzer:2000xy}.
Most of the new examples found in recent years are in fact obtained from these via
one of two techniques. The first is to take the quotient by a holomorphic action of
some finite group. As explained in \sref{sec:quotients}, when the group action is
fixed-point-free, this is guaranteed to yield another Calabi-Yau manifold, and many
Calabi-Yau threefolds with non-trivial fundamental group have been constructed in
this way. Several early examples can be found in \cite{Candelas:1985en,Yau1,
Strominger:1985it,Candelas:1987du,Schimmrigk:1987ke,Beauville}, but recent
efforts have brought to light many more \cite{Candelas:2008wb,GrossPopescuI,
Braun:2004xv,Batyrev:2005jc,Hua,Bouchard:2007mf,BorisovHua,GrossPopescuII,
Braun:2009qy,Braun:2010vc,Candelas:2010ve,Braun:2011hd,Bini:2011dp}, some
of which will be discussed later. In the case that the group action has fixed points, it
is often possible to resolve the resulting orbifold singularities in such a way as to
again obtain a Calabi-Yau manifold. Examples can be found in
\cite{Stapledon:2010mo,Freitag:2010st,Freitag:2011st}. The second technique is to
vary either the complex structure or K\"ahler moduli of a known space until it
becomes singular, and then desingularise it by varying the other type of moduli.
Topologically, such a process is a surgery, and yields a Calabi-Yau manifold
topologically distinct from the original. Two classes of such topological transitions,
the conifold and hyperconifold transitions, are discussed in \sref{sec:transitions},
and explicit examples of each are given. Conifold transitions have been known
for some time to connect many Calabi-Yau threefolds
\cite{Green:1988bp,Green:1988wa,Candelas:1988di,Candelas:1989ug}, and have
been used to construct new manifolds in
\cite{Candelas:2008wb,Candelas:2010ve,Batyrev:2008rp,Filippini:2011rf}.
Hyperconifold transitions were described in \cite{Davies:2009ub}, and the first
examples of new manifolds discovered this way were given in \cite{Davies:2011is}.
The examples of \sref{sec:hyperconifolds} yield two more new manifolds, one of
which is the first known with fundamental group $S_3$, the smallest non-Abelian
group.
The fruit of these labours is that there are now many more known
Calabi-Yau threefolds with small Hodge numbers (defined arbitrarily in this paper by
$h^{1,1}+h^{2,1}~\leq~24$) than were known to the authors of \cite{Triadophilia}. The
number with non-trivial fundamental group has also increased dramatically,
thanks largely to Braun's classification of free group actions on complete intersections
in products of projective spaces \cite{Braun:2010vc}.\footnote{The Hodge
numbers of many of these quotients are yet to be calculated, but some are likely to
be ``small" as defined above.}
The physical motivation for studying such manifolds comes predominantly from heterotic
string theory. In this context, a non-trivial fundamental group is necessary to be able to
turn on discrete Wilson lines and thus obtain a realistic four-dimensional gauge group.
The requirement of small Hodge numbers is not so clear-cut, but it seems advantageous
if one wants to appeal to the methods of \cite{Anderson:2010mh,Anderson:2011cz} to
stabilise the moduli, and this is currently the only known way to stabilise all (geometric)
moduli in heterotic Calabi-Yau backgrounds. Although heterotic model building is not
the theme of this review, other recent developments will be mentioned sporadically.
Throughout the paper, an arbitrary Calabi-Yau and its universal cover will be
denoted by $X$ and $\widetilde X$ respectively, while a particular Calabi-Yau
threefold with Hodge numbers $(h^{1,1},\,h^{2,1})$ will be denoted by
$X^{h^{1,1},h^{2,1}}$. $X^\sharp$ will denote a singular member of the family $X$,
and $\widehat X^\sharp$ a resolution of such a singular variety.
\section{Quotients by group actions}\label{sec:quotients}
\subsection{The Calabi-Yau condition}
It is an elementary fact of topology that every manifold $X$ has a simply-connected
universal covering space $\widetilde X$, from which it can be obtained as a
quotient by the free action of a group $G \cong \pi}\renewcommand{\P}{\Pi}\newcommand{\vp}{\varpi_1(X)$. We will write this relationship
as $X = \widetilde X/G$. Although our interest is in (complex) threefolds, we will allow
the dimension $n$ of $X$ to be arbitrary through much of this section.
If $X$ is a Calabi-Yau manifold, it is easy to see that its universal cover
$\widetilde X$ is too, by pulling back the complex structure, K\"ahler form $\o$, and
holomorphic $(n,0)$-form $\O$ under the covering map (for this reason we will often
abuse notation by using the same symbols for these objects on $X$ and $\widetilde X$).
Only a little more difficult is the converse: under what conditions is $X = \widetilde X/G$
a Calabi-Yau manifold, given that $\widetilde X$ is? There are several points to
consider (we assume always that $G$ is a finite group):
\begin{itemize}
\item
$X$ will be a manifold as long as the action of $G$ is fixed-point free. Otherwise it
will have orbifold singularities.
\item
It will furthermore be a complex manifold if and only if $G$ acts by biholomorphic maps.
In this case, $X$ simply inherits the complex structure of $\widetilde X$.
\item
To see that $X$ is K\"ahler, pick any K\"ahler form $\o$ on $\widetilde X$. Now note
that for any element $g \in G$, $g^*\o$ is also a K\"ahler form, since
$d(g^*\o) = g^*(d\o) = 0$, and for any $k$-dimensional complex submanifold $M_k$,
\begin{equation*}
\int_{M_k} g^*\o^k ~=~ \int_{g(M_k)} \o^k ~>~ 0~.
\end{equation*}
We can use this to construct a K\"ahler form which is invariant under $G$, and
therefore descends to a K\"ahler form on $X$:
\begin{equation*}
\o^G ~:=~ \frac{1}{|G|}\sum_{g\in G} g^*\o~.
\end{equation*}
\item
Finally, we must check whether $X$ supports a nowhere-vanishing holomorphic
$(n,0)$-form. This can only descend from the one on $\widetilde X$, so we need to
check whether $\O$ is $G$-invariant. Note that $\O$ is the unique (up to scale) element
of $H^{n,0}(\widetilde X)$, so since $G$ acts freely, the Atiyah-Bott fixed point formula
(\cite{AtiyahBott,AtiyahBott2}) for any $g \in G\setminus e$ reduces to
\begin{equation*}
0 ~=~ \sum_{q=0}^n (-1)^q\, \mathrm{Tr} \big( H^{n,q}(g) \big) ~=~
\mathrm{Tr}\big( H^{n,0}(g) \big) + (-1)^n\, \mathrm{Tr}\big( H^{n,n}(g) \big)~,
\end{equation*}
where $H^*(g)$ denotes the induced action of $g$ on the cohomology $H^*$.
The group $H^{n,n}(\widetilde X)$ is generated by $\big(\o^G \big)^n$, which is
invariant, so we conclude that $\mathrm{Tr}\big(H^{n,0}(g)\big) = (-1)^{n+1}$, and therefore
\begin{equation*}
g^*\O = (-1)^{n+1}\O ~~\forall~~ g\in G\setminus e~.
\end{equation*}
In odd dimensions, therefore, $\O$ is automatically invariant under free group actions.
In even dimensions, on the other hand, this simple calculation shows that there are no
multiply-connected Calabi-Yau manifolds.
\end{itemize}
In summary, if $\widetilde X$ is a smooth Calabi-Yau threefold, then $X = \widetilde X/G$ is
a Calabi-Yau manifold if and only if $G$ acts freely and holomorphically.
\subsubsection{Smoothness}
Above, we have simply assumed that the covering space $\widetilde X$ is smooth. In
practice, $\widetilde X$ usually belongs to a family of spaces of which only a sub-family
admits a free $G$ action. It may be the case that, although a generic member of
$\widetilde X$ is smooth, members of the symmetric sub-family are all singular, so we
never get a smooth quotient. Although this seems to be rare, it does occur for
$\mathbb{Z}_8{\times}\mathbb{Z}_8$-symmetric complete intersections of four quadrics in $\mathbb{P}^7$
\cite{Strominger:1985it,Hua}, $\mathbb{Z}_5{\times}\mathbb{Z}_5$-symmetric complete intersections of five
bilinears in $\mathbb{P}^4{\times}\mathbb{P}^4$ \cite{Candelas:2008wb}, and a number
of other examples, including some found in \cite{Braun:2010vc}. The extra condition, that
a generic \emph{symmetric} member is smooth, must therefore be checked on a
case-by-case basis.
Let us start with the special case of complete intersection Calabi-Yau manifolds in smooth
ambient spaces. This includes what have traditionally been called the `CICY' manifolds,
where the ambient space is a product of projective spaces
\cite{Green:1986ck,Candelas:1987kf,Green:1987cr}, and certain of the toric hypersurfaces
\cite{Kreuzer:1995cd,Kreuzer:2000xy}. The complete intersection condition means that if
the ambient space has dimension $n+k$, then the Calabi-Yau $\widetilde X$ is given by the
intersection of $k$ hypersurfaces, each given by a single polynomial equation $f_a = 0$.
In other words, the number of equations needed to specify $\widetilde X$ is equal to its
codimension.
When the ambient space is smooth, it can be covered in affine patches each isomorphic to
$\mathbb{C}^{n+k}$, and the condition for $\widetilde X$ to be smooth is that
$df_1\wedge\ldots\wedge df_k$ is non-zero at every point on $\widetilde X$. This is a very
intuitive condition --- if it holds, then at any point of $\widetilde X$ we can choose local
coordinates $x_1, \ldots,x_{n+k}$ such that $f_a = x_a + \mathcal{O}(x^2)$. Locally, then,
$\widetilde X$ projects biholomorphically onto the linear subspace
$x_1 = x_2 = \ldots = x_k = 0$, and is therefore smooth. On any affine coordinate patch, the
components of the differential form $df_1\wedge\ldots\wedge df_k$ are just the $k{\times}k$
minors of the Jacobian matrix $J = \left(\partial f_a/\partial x_i\right)$, so the condition is that
this matrix has rank $k$ everywhere on $\widetilde X$. It is therefore necessary to check
that there is no simultaneous solution to the equations $f_a = 0$ along with the vanishing
of all $k{\times}k$ minors of $J$, which is equivalent to the algebraic statement that the
ideal generated by the polynomials and the minors is the entire ring
$\mathbb{C}[x_1,\ldots,x_{k+n}]$. This is checked by calculating a Gr\"obner basis for the ideal,
algorithms for which are implemented in a variety of computer algebra packages
\cite{Singular,Mathematica,Gray:2008zs}; a Gr\"obner basis for the entire ring is just a
constant (usually given as $1$ or $-1$ by software).
The more general case of singular ambient spaces or non-complete intersections is not
much harder than the above. Suppose $\widetilde X$ is not a complete intersection, so
that it is given by $l$ equations in an $n+k$-dimensional ambient space, where now we
allow $l > k$.\footnote{A typical example is the Veronese embedding of $\mathbb{P}^2$ in $\mathbb{P}^5$.
If we take homogeneous coordinates $z_i$ for $\mathbb{P}^2$, and $w_{ij}$ for $\mathbb{P}^5$, where
$j \geq i$, then the embedding is given by $w_{ij} = z_i z_j$. The equations needed to
specify the image of this map are $w_{ij}w_{kl} - w_{il}w_{kj} = 0$, which amount to six
independent equations, whereas the codimension of the embedded surface is only
three.} Then the condition for $\widetilde X$ to be smooth is still that the rank of the
Jacobian be equal to $k$ (the codimension) everywhere on $\widetilde X$
\cite{Hartshorne}. The reasoning is the same as before --- if this is true, $k$ of the
polynomials will provide good local coordinates on the ambient space, allowing us to
define a smooth coordinate patch on $\widetilde X$.
If some affine patch on the ambient space is singular, it
can still be embedded in $\mathbb{C}^N$ for some $N$, by polynomial equations
$F_1 = \ldots = F_K = 0$. The Calabi-Yau is then given by
$F_1 = \ldots = F_K = f_1 = \ldots = f_l = 0$, and the condition for smoothness is once
again that the Jacobian has rank equal to the codimension, $N - n$, at all points.
For examples of interest, Gr\"obner basis calculations are often very computationally
intensive, since at intermediate stages the number of polynomials, as well as their
coefficients, can become extremely large. It is therefore convenient to choose integer
coefficients for all polynomials, and perform the calculation over a finite field $\mathbb{F}_p$. As
explained in \cite{Candelas:2008wb}, if a collection of polynomial equations are
inconsistent over $\mathbb{F}_p$, then they are also inconsistent over $\mathbb{C}$, so the corresponding
variety is smooth.
Note that there does exist a slight variation on the above procedure which still leads to
smooth quotient manifolds. It may be the case that although the symmetric manifolds
admit a free group action, they are all singular. If, however, these singularities can be
resolved in a group-invariant way, the resolved space still admits a free group action,
with a smooth quotient. Examples can be found in
\cite{GrossPopescuI,Hua,GrossPopescuII}.
The final possibility is that the symmetric manifolds are smooth, but the group action
always has fixed points, in which case the quotient space has orbifold singularities.
It is frequently possible to resolve these in such a way as to again obtain a
Calabi-Yau manifold, but this will not be discussed in detail here.
\subsection{Notable Examples}
\subsubsection{New Three-Generation Manifolds} \label{sec:(8,44)}
Calabi-Yau threefolds with Euler number $\chi = \pm 6$ were of particular
interest in the early days of string phenomenology, since these give physical models with
three generations of fermions via the `standard embedding' compactification of the heterotic
string \cite{Candelas:1985en}. This typically gives an $E_6$ grand unified theory, and
although the gauge symmetry can be partially broken by Wilson lines, it is impossible to obtain
exactly the standard model gauge group in this way \cite{McInnes:1989rg}. Nevertheless, it
was argued by Witten that deformations of the standard embedding, combined with Wilson
lines, can give realistic models \cite{Witten:1985bz}, and this was put on firmer mathematical
foundations by Li and Yau \cite{Li:2004hx}.
The archetypal example of a three-generation manifold is Yau's manifold, with fundamental
group $\mathbb{Z}_3$ \cite{Yau1}, but recently two new promising three-generation manifolds were
constructed in \cite{Braun:2009qy}. These are quotients of a manifold $X^{8,44}$ by
groups of order twelve, which are the cyclic group $\mathbb{Z}_{12}$ and the non-Abelian group
$\mathrm{Dic}_3 \cong \mathbb{Z}_3{\rtimes}\mathbb{Z}_4$ (this is generated by two elements, one of order three
and one of order four, satisfying $g_4 g_3 g_4^{-1} = g_3^2$), and each has Hodge
numbers $(h^{1,1},\,h^{2,1}) = (1,4)$. Unfortunately, it was shown in \cite{Davies:DPhil} that
the physical model on the non-Abelian quotient does not admit a deformation which yields
exactly the field content of the minimal supersymmetric standard model (MSSM) in four
dimensions. However, the $\mathbb{Z}_{12}$ quotient allows many more distinct deformations, and
the analysis of the corresponding physical models has not been completed.
The covering space $X^{8,44}$ is an anti-canonical hypersurface in $\dP_6{\times}\dP_6$,
where $\dP_6$ is the del Pezzo surface of degree six, which is $\mathbb{P}^2$ blown up in
three generic points. This surface is rigid and toric, and its fan is shown in \fref{fig:dP6fan}.
\begin{figure}
\begin{center}
\includegraphics[width=.35\textwidth]{fig_dP6fan.pdf}
\parbox{.8\textwidth}{\caption{\label{fig:dP6fan}
\small
The fan for the toric surface $\dP_6$. Removing the dashed rays corresponds to the
projection to $\mathbb{P}^2$. All graphics were produced in Mathematica~\cite{Mathematica}.
}}
\end{center}
\vspace{-15pt}
\end{figure}
As well as the action of the torus $\big(\mathbb{C}^*\big)^2$, $\dP_6$ also admits an action
by the dihedral group $D_6$, as suggested by its fan. This can be realised as a
group of lattice morphisms preserving the fan, generated by an order-six rotation
$\r$ and a reflection $\sigma}\renewcommand{\S}{\Sigma}\newcommand{\vs}{\varsigma$, with matrix representations
\begin{equation*}
\r = \left(\begin{array}{rr} 1 & -1 \\ 1 & 0 \end{array} \right)
~,~~
\sigma}\renewcommand{\S}{\Sigma}\newcommand{\vs}{\varsigma = \left(\begin{array}{rr} 0 & 1 \\ 1 & \phantom{-}0 \end{array} \right) ~.
\end{equation*}
The product $\dP_6{\times}\dP_6$ therefore has symmetry group
$(D_6{\times}D_6)\rtimes\mathbb{Z}_2$, where the extra $\mathbb{Z}_2$ factor swaps the two copies
of the surface. The quotient groups $\mathrm{Dic}_3$ and $\mathbb{Z}_{12}$ are both order-twelve
subgroups of this which act transitively on the vertices of the fan. Many more details
can be found in \cite{Braun:2009qy}.
\subsubsection{Quotients of the (19,19) Manifold} \label{sec:(19,19)}
The Euler number of a three-dimensional Calabi-Yau manifold is given by the
simple formula $\chi = 2(h^{1,1} - h^{2,1})$. If a group $G$ acts freely, then
$\chi(\widetilde X/G) = \chi(\widetilde X)/G$, so this gives a simple necessary condition
for the existence of such an action: the order of the group must divide $\chi/2$. This
usually gives a fairly strong restriction on the groups which can act freely on any given
manifold. The only time it gives no restriction is when $\chi = 0$.
In this section we will look at a particular manifold, $X^{19,19}$, which admits free
actions by a number of disparate groups, including groups of order five, eight, and
nine. For a Calabi-Yau threefold with $\chi \neq 0$, this would imply $|\chi| \geq 720$.
The manifold $X^{19,19}$ can be represented in a number of different ways. Abstractly,
it is the fibre product of two rational surfaces, each elliptically fibred over $\mathbb{P}^1$
\cite{Braun:2004xv,Bouchard:2007mf}. Such
a surface is given by blowing up $\mathbb{P}^2$ at the nine points given by the intersection of
two cubic curves; if we take homogeneous coordinates $t_0, t_1$ on $\mathbb{P}^1$ and
$z_0, z_1, z_2$ on $\mathbb{P}^2$, the corresponding equation is
\begin{equation}\label{eq:ellipticsurface}
f(z) t_0 + g(z) t_1 = 0 ~,
\end{equation}
where $f$ and $g$ are homogeneous cubic polynomials. We can easily see that this
corresponds to $\mathbb{P}^2$ blown up at the nine points given by $f = g = 0$. Indeed, for any
point of $\mathbb{P}^2$ where $f\neq 0$ or $g\neq 0$, we get a unique solution for
$[t_0 : t_1]$, whereas for $f = g = 0$, the equation is satisfied identically, giving
a whole copy of $\mathbb{P}^1$. To see that it is also an elliptic fibration over $\mathbb{P}^1$, note
that for any fixed value of $[t_0 : t_1] \in \mathbb{P}^1$, we get a cubic equation in $\mathbb{P}^2$, which
defines an elliptic curve.
To get the fibre product of two such surfaces, we introduce another $\mathbb{P}^2$, with
homogeneous coordinates $w_0, w_1, w_2$, and another equation of the form
\eqref{eq:ellipticsurface} over the same $\mathbb{P}^1$. The resulting threefold is Calabi-Yau,
and has a projection to $\mathbb{P}^1$, with typical fibre which is a product of two elliptic curves.
In \cite{Bouchard:2007mf}, Bouchard and Donagi studied group actions which
preserve the elliptic fibration, and found free actions by the groups $\mathbb{Z}_3{\times}\mathbb{Z}_3$,
$\mathbb{Z}_4{\times}\mathbb{Z}_2$, $\mathbb{Z}_6 \cong \mathbb{Z}_3{\times}\mathbb{Z}_2$ and $\mathbb{Z}_5$ (as well as all
subgroups of these, of course).
Certain of these quotient manifolds have in fact played crucial roles in the heterotic string
literature. A model with the spectrum of the $U(1)_{B-L}$-extended supersymmetric
standard model was constructed on a quotient by $\mathbb{Z}_3{\times}\mathbb{Z}_3$ and studied in
\cite{Braun:2005nv,Braun:2006me} (a similar model on the $\mathbb{Z}_3{\times}\mathbb{Z}_3$ quotient
of the `bicubic', which is related to this manifold by a conifold transition, was found in
\cite{Anderson:2009mh}), while a model with the exact MSSM spectrum exists on a
$\mathbb{Z}_2$ quotient, and was described in \cite{Bouchard:2005ag,Bouchard:2006dn}.
In \cite{Braun:2007xh,Braun:2007vy}, the quotient by a different $\mathbb{Z}_3{\times}\mathbb{Z}_3$
action was used as a test case for calculating instanton corrections on manifolds with
torsion curves.
There are in fact further (relatively large) groups which act freely on $X^{19,19}$, which
can be easily described using its representation(s) as a CICY. First, we note that the
fibre product construction above is equivalent to the rather more prosaic statement that
the manifold is a complete intersection of two hypersurfaces in
$\mathbb{P}^1{\times}\mathbb{P}^2{\times}\mathbb{P}^2$, of multi-degrees $(1,3,0)$ and $(1,0,3)$. In the notation
of $\cite{Green:1986ck,Candelas:1987kf}$, $X^{19,19}$ can therefore be specified by the
`configuration matrix'
\vskip-10pt
\begin{equation*}
\cicy{\mathbb{P}^1 \\ \mathbb{P}^2 \\ \mathbb{P}^2}{
1 & 1 \\
3 & 0 \\
0 & 3}\quad.
\end{equation*}
By utilising various \emph{splittings} and \emph{contractions} (see e.g.
\cite{Candelas:2008wb,Candelas:1987kf,Green:1988bp,Green:1988wa}), and checking
that the Euler number remains constant, it is easy to show that $X^{19,19}$ can also be
specified by the configuration matrices
\begin{equation*}
\cicy{\mathbb{P}^1\\ \mathbb{P}^1 \\ \mathbb{P}^1 \\ \mathbb{P}^1 \\ \mathbb{P}^1}{
1 & 1 \\
2 & 0 \\
2 & 0 \\
0 & 2 \\
0 & 2} \qquad, \qquad
\cicy{\mathbb{P}^1\\ \mathbb{P}^2\\ \mathbb{P}^2\\ \mathbb{P}^2\\ \mathbb{P}^2}{
1 & 1 & 0 & 0 & 0 & 0\\
1 & 0 & 1 & 1 & 0 & 0 \\
1 & 0 & 1 & 1 & 0 & 0 \\
0 & 1 & 0 & 0 & 1 & 1\\
0 & 1 & 0 & 0 & 1 & 1 }\quad.
\end{equation*}
It was shown in \cite{Candelas:2008wb} that in the first form, $X^{19,19}$ admits a
free action of the order-eight quaternion group (denoted in \cite{Candelas:2008wb} by
$\mathbb{H}$, but more conventionally by $Q_8$), with elements $\{\pm 1, \pm \ii, \pm j, \pm k\}$,
induced by a linear action of this group on the ambient space.
In the second form, $X^{19,19}$ admits free actions by two groups of order twelve. One
is the cyclic group $\mathbb{Z}_{12}$, and the other is the dicyclic group
$\mathrm{Dic}_3 \cong \mathbb{Z}_3{\rtimes}\mathbb{Z}_4$ (introduced in \sref{sec:(8,44)}) \cite{Davies:DPhil}. These
were in fact discovered via conifold transitions from the corresponding quotients of
$X^{8,44}$, an idea reviewed in \sref{sec:transitions}.
\begin{table}[ht]
\begin{center}
\begin{tabular}{| c c c |}
\hline
\vrule height16pt width0pt depth10pt
$~h^{1,1} = h^{2,1}~$ & ~Fundamental Group~ & ~Reference~ \\
\hline
\hline
11 & $\mathbb{Z}_2$ & \cite{Candelas:2008wb,Bouchard:2007mf} \\
\hline
7 & $\mathbb{Z}_3~,~\mathbb{Z}_2{\times}\mathbb{Z}_2$ & \cite{Candelas:2008wb,Bouchard:2007mf} \\
\hline
5 & $\mathbb{Z}_4$ & \cite{Bouchard:2007mf} \\
\hline
3 & $\mathbb{Z}_5~,~\mathbb{Z}_6~,~\mathbb{Z}_4{\times}\mathbb{Z}_2~,~Q_8~,~\mathbb{Z}_3{\times}\mathbb{Z}_3$ & \cite{Candelas:2008wb,Bouchard:2007mf} \\
\hline
2 & $\mathbb{Z}_{12}~,~\mathrm{Dic}_3$ & \cite{Braun:2009qy,Davies:DPhil} \\
\hline
\end{tabular}\\
\parbox{.75\textwidth}{\caption{\label{tab:1919hodgenos}
\small The Hodge numbers for known quotients of $X^{19,19}$.
}}
\end{center}
\vskip-10pt
\end{table}
In summary, $X^{19,19}$ is rather exceptional in that it admits free actions by the groups
$\mathbb{Z}_{12}, \mathrm{Dic}_3, \mathbb{Z}_3{\times}\mathbb{Z}_3, \mathbb{Z}_4{\times}\mathbb{Z}_2, Q_8, \mathbb{Z}_6$ and $\mathbb{Z}_5$. The
Hodge numbers for the quotients by all these groups and their subgroups are collected
in \tref{tab:1919hodgenos}.
\subsubsection{Manifolds with Hodge numbers (1,1)} \label{sec:(1,1)}
For a long time, the smallest known Hodge numbers of a Calabi-Yau threefold
satisfied $h^{1,1} + h^{2,1} = 4$. This record has now been overtaken by Braun's
examples of manifolds with $(h^{1,1},\,h^{2,1}) = (1,1)$ \cite{Braun:2011hd} (as well
as Freitag and Salvati Manni's manifold with $(h^{1,1},\,h^{2,1}) = (2,0)$
\cite{Freitag:2011st}).
The covering space of Braun's $(1,1)$ manifolds is a self-mirror manifold
$X^{20,20}$. This is realised as an anti-canonical hypersurface in the toric fourfold
determined by the face fan over the 24-cell, which is a self-dual regular
four-dimensional polytope.
There are three different groups of order 24 which act freely on particular smooth
one-parameter sub-families of $X^{20,20}$; these are
$\mathbb{Z}_3{\rtimes}\mathbb{Z}_8, \mathbb{Z}_3{\times}Q_8$ and $SL(2,3)$. The first two are
self-explanatory, while the third is the group of two-by-two matrices of determinant one
over the field with three elements. All the groups act via linear transformations on the
lattice in which the polytope lives, and act transitively on its vertices.
Full details can be found in \cite{Braun:2011hd}.
\subsubsection{Complete intersections of four quadrics in
\texorpdfstring{$\mathbb{P}^7$}{P7}} \label{sec:fourquadrics}
A particularly fertile starting point for finding new Calabi-Yau manifolds has been the
complete intersection of four quadrics in $\mathbb{P}^7$. A smooth member of this family is
a Calabi-Yau manifold with Hodge numbers $(h^{1,1},\,h^{2,1}) = (1,65)$. Hua classified
free group actions on smooth sub-families in \cite{Hua}, finding groups of order 2, 4,
8, 16 and 32. The quotients all have $h^{1,1} = 1$, and $h^{2,1} =$ 33, 17, 9, 5 and
3 respectively.
Certain nodal families allow free actions of groups of order 64, and furthermore have
equivariant small resolutions \cite{GrossPopescuI,Hua}. The resolutions have Hodge
numbers $(h^{1,1},\,h^{2,1}) = (2,2)$, and inherit the free group actions. Remarkably, in
this case all the quotients have the same Hodge numbers as the covering space.
The quotient by $\mathbb{Z}_8{\times}\mathbb{Z}_8$ was investigated as a background for heterotic
string theory in \cite{Bak:2008ey}, but unfortunately no realistic models were found.
Freitag and Salvati Manni have also constructed a large number of new manifolds by
starting with a particular complete intersection $X^\sharp$ of four quadrics which has
96 nodes and a very large symmetry group \cite{Freitag:2010st,Freitag:2011st}. They
show that the quotients by many subgroups admit crepant projective resolutions,
thereby giving rise to a large number of new Calabi-Yau manifolds. Some of the
subgroups of order 2, 4, 8 and 16 act freely on a small resolution of $X^\sharp$, and
the corresponding quotient manifolds are connected to some of Hua's examples by
conifold transitions. The manifolds from \cite{Freitag:2011st} with small Hodge
numbers are listed in \aref{app:new}, including one with $(h^{1,1},\,h^{2,1}) = (2,0)$, which
is therefore equal with Braun's manifolds for smallest known Hodge numbers. Note
that the theoretical minimum is $(h^{1,1},\,h^{2,1}) = (1,0)$.
\section{New manifolds from topological transitions} \label{sec:transitions}
One fascinating feature of Calabi-Yau threefolds is the inter-connectedness of moduli
spaces of topologically distinct manifolds. Generally speaking, there are two ways to
pass from one smooth Calabi-Yau to another. We may deform the complex structure
until a singularity develops, and then `resolve' this singularity using the techniques of
algebraic geometry, which involves replacing the singular set with new embedded
holomorphic curves or surfaces. Alternatively, we may allow certain embedded curves
or surfaces to collapse to zero size, and then `smooth' the resulting singular space by
varying its complex structure. Obviously these two processes are inverses of each
other.
Our main interest here is in constructing new smooth Calabi-Yau threefolds via
such topological transitions, but first I will indulge in a few comments about the
connectedness of the space of all Calabi-Yau threefolds.
The suggestion that all Calabi-Yau threefolds might be connected by topological
transitions goes back to \cite{ReidFantasy}. Early work showed
that this was true for nearly all examples known at the time \cite{Green:1988bp,
Green:1988wa}. These papers considered conifold transitions, in which the intermediate
variety has only nodal singularities; the smoothing process replaces these singular points
by three-spheres, while the `small' resolution replaces them by two-spheres
(holomorphically embedded). Such singularities were shown to be at finite distance in
moduli space \cite{Candelas:1988di,Candelas:1989ug}, and conifold transitions were
later shown to be smooth processes in type II string theory \cite{Strominger:1995cz,
Greene:1995hu}.
If we wish to connect all Calabi-Yau threefolds, conifold transitions are not sufficient,
because they cannot change the fundamental group. To see this we note that,
topologically, a conifold transition consists of removing neighbourhoods of some
number of copies of $S^3$, each with boundary $S^3{\times}S^2$, and replacing
them with similar neighbourhoods of $S^2$. Since all these spaces are
simply-connected, a simple application~of van Kampen's theorem (see e.g.
\cite{Hatcher}) shows that the fundamental group does not change.
There do exist relatively mild topological transitions which can change the
fundamental group; these are known as \emph{hyperconifold transitions}, and
were described by the author in \cite{Davies:2009ub,Davies:2011is}. Here the
singularities of the intermediate variety are finite quotients of a node, and
their resolutions are no longer `small'. It is an interesting question whether all
Calabi-Yau threefolds can be connected by conifold and hyperconifold transitions.
In the following sections, we will consider these two types of transition separately,
mostly through examples. The examples in \sref{sec:hyperconifolds} actually yield
previously unknown manifolds, with Hodge numbers $(h^{1,1},\,h^{2,1}) = (2,5)$ and
$(2,3)$ and fundamental groups $\mathbb{Z}_5$ and $S_3$ respectively.
\subsection{Conifold transitions} \label{sec:conifolds}
In \cite{Candelas:2008wb,Candelas:2010ve}, free group actions were followed through
conifold transitions, leading to webs of conifold transitions between smooth quotients
with the same fundamental group (conifold transitions were also used in
\cite{Batyrev:2008rp,Filippini:2011rf} to construct new simply-connected manifolds).
Here we will just consider a simple example (taken from \cite{Candelas:2008wb}) which
exemplifies the idea.
Consider the well-known family of quintic hypersurfaces in $\mathbb{P}^4$, with Hodge numbers
$(h^{1,1},\,h^{2,1}) = (1,101)$ and hence Euler number $\chi = -200$. If we take
homogeneous coordinates $z_0,\ldots,z_4$ on $\mathbb{P}^4$, then an action of $\mathbb{Z}_5$
can be defined by the generator
\begin{equation*}
g_5 : z_i \to z_{i+1} ~.
\end{equation*}
Then there is a smooth family of invariant quintics, given by
\begin{equation} \label{eq:symquintic}
f = \sum_{ijklm} \, A_{j-i,k-i,l-i,m-i}\, z_i z_j z_k z_l z_m = 0~.
\end{equation}
For generic coefficients, $\mathbb{Z}_5$ acts freely, so we get a family of smooth quotients with
Hodge numbers $(h^{1,1},\,h^{2,1}) = (1,21)$.
Now let us consider a non-generic choice for the coefficients in \eqref{eq:symquintic},
such that $f$ is the determinant of some $5\times5$ matrix $M$ which is linear in the
homogeneous coordinates. If we take the entries
of $M$ to be
\begin{equation*}
M_{ik} = \sum_j \, a_{j-i,k-i}\, z_j ~,
\end{equation*}
then the induced $\mathbb{Z}_5$ action is $M_{ik} \to M_{i+1, k+1}$, so the determinant
does indeed correspond to an invariant quintic. The action of $\mathbb{Z}_5$ is still generically
fixed-point-free on the family given by $\det M = 0$, but the hypersurfaces are no longer
smooth. Indeed, using a computer algebra package, it can be checked that the rank of
$M$ drops to three at exactly fifty points on a typical such hypersurface, and that these
points are nodes. Furthermore, they fall into ten orbits of five nodes under the $\mathbb{Z}_5$
action.
We now ask whether these nodes can be resolved in a group-invariant way; if so, the
group will still act freely on the resolved manifolds, and we will have constructed a
conifold transition between the quotient manifolds. In fact this is easy to do. Introduce
a second $\mathbb{P}^4$, with homogeneous coordinates $w_0,\ldots,w_4$, and consider the
equations
\begin{equation} \label{eq:splitquintic}
f_i := \sum_k \, M_{ik}\, w_k = \sum_{j,k} \, a_{j-i, k-i}\, z_j\, w_k = 0 ~.
\end{equation}
These are five bilinears in $\mathbb{P}^4{\times}\mathbb{P}^4$, and it can be checked that they generically
define a smooth Calabi-Yau threefold $X^{2,52}$. Since we cannot have
$w_i = 0 ~\forall~ i$, there are only simultaneous solutions to these equations when
$\det M = 0$, so this gives a projection from $X^{2,52}$ to nodal members of $X^{1,101}$.
At most points, this is one-to-one, but at the fifty points where the rank of $M$ drops to
three, we get a whole copy of $\mathbb{P}^1 \subset \mathbb{P}^4$ projecting to a (nodal) point of
$X^{1,101}$. In this way we see that we have constructed a conifold transition
$X^{1,101} \rightsquigarrow X^{2,52}$.
To see that the free $\mathbb{Z}_5$ action is preserved by the conifold transition above, it
suffices to note that if we extend the action by defining $g_5 : w_i \to w_{i+1}$, then
this induces $f_i \to f_{i+1}$, implying that the manifolds defined by \eqref{eq:splitquintic}
are $\mathbb{Z}_5$-invariant. The absence of fixed points follows from the absence of fixed
points on the nodal members of $X^{1,101}$, although this can also be checked directly.
Since the conifold transition from $X^{1,101}$ to $X^{2,52}$ can be made
$\mathbb{Z}_5$-equivariant, it descends to a conifold transition between their quotients,
$X^{1,21} \rightsquigarrow X^{2,12}$, where the intermediate variety has ten nodes.
\subsection{Hyperconifold transitions} \label{sec:hyperconifolds}
The conifold transition in the last section illustrates two completely general features of
such transitions: the fundamental group does not change, for reasons explained
previously, and the intermediate singular variety has multiple nodes
\cite{Candelas:1989js}. We now turn our attention to a class of transitions for which
neither of these statements hold --- the so-called hyperconifold transitions introduced
in \cite{Davies:2009ub}. Here the intermediate space typically has only one singular
point, which is a quotient of a node by some finite cyclic group
$\mathbb{Z}_N$.\footnote{Quotients by non-Abelian groups can also occur, but these do not
admit a toric description, and their resolutions have not been studied.} These arise
naturally when a generically-free group action is allowed to develop a fixed point. A
$\mathbb{Z}_N$-hyperconifold transition changes the Hodge numbers according to the general
formula
\vspace{-5pt}
\begin{equation} \label{eq:hchodge}
\d(h^{1,1},\,h^{2,1}) = (N-1, -1) ~.
\end{equation}
\vspace{-25pt}
The resolution of a hyperconifold singularity replaces the singular point with a
simply-connected space, and in this way we see that the transitions can change the
fundamental group. It is worth pausing here to consider this in more detail than has
been done in previous papers.
Suppose we have a smooth quotient $X = \widetilde X/G$, and deform the complex
structure until some order-$N$ element $g_N$, which generates a subgroup
$\langle g_N \rangle \cong \mathbb{Z}_N\! < G$, develops a fixed point $p \in \widetilde X$.
Then, as described in \cite{Davies:2009ub}, this point will be singular, and generically
a node. In some cases, the group structure implies that other elements will
simultaneously develop fixed points, which we can see by taking a group element
$g' \in G\setminus\langle g_N\rangle$ and performing an elementary calculation,
\begin{equation*}
g'\,g_N\,g'^{-1} \cdot (g'\cdot p) = g'\cdot(g_N\cdot p) = g'\cdot p ~.
\end{equation*}
So the point $g' \cdot p \in \widetilde X$ is fixed by $g'\,g_N\,g'^{-1}$. We see that every
subgroup conjugate to $\langle g_N \rangle$ also develops a fixed point. All such
points are identified by $G$, so the singular quotient $X^\sharp$ has only one
$\mathbb{Z}_N$-hyperconifold singularity.
What is the fundamental group of the resolution $\widehat X^\sharp$? To calculate this,
excise a small ball around each fixed point of $\widetilde X^\sharp$, to obtain a smooth
space $\widetilde X'$ on which the whole group $G$ acts freely. We can then quotient
by $G$ to obtain $X'$, with fundamental group $G$. Finally, we glue in a
neighbourhood $\S$ of the exceptional set of the resolution of the hyperconifold. $\S$
is simply-connected. We now have $\widehat X^\sharp = X' \cup \S$, and can use van
Kampen's theorem to calculate $\pi_1(\widehat X^\sharp)$. Note that the intersection
of the two subspaces $X'$ and $\S$ is homotopy-equivalent to $S^3{\times}S^2/\mathbb{Z}_N$,
since the stabiliser of each point on the covering space was isomorphic to $\mathbb{Z}_N$. So
we have the data
\begin{equation*}
\widehat X^\sharp = X' \cup \S ~,~~ X' \cap \S \stackrel{\mathrm{hom.}}{\simeq} S^3{\times}S^2/\mathbb{Z}_N
~,~~ \pi_1(\S) \cong {\bf 1} ~,~~ \pi_1(X') \cong G ~,
\end{equation*}
which by van Kampen's theorem immediately implies that
$\pi_1(\widehat X^\sharp) \cong G/\langle g_N \rangle^G$, where $\langle g_N \rangle^G$
is the smallest normal subgroup of $G$ which contains $\langle g_N \rangle$, usually
called the normal closure.
Trivial examples arise when $G = \mathbb{Z}_N{\times}H$ or $\mathbb{Z}_N{\rtimes}H$, and
the generator of $\mathbb{Z}_N$ develops a fixed point. In this case, the corresponding
hyperconifold transition changes the fundamental group from $G$ to $H$.
\subsubsection{Example 1:
\texorpdfstring{$X^{1,6} \rightsquigarrow X^{2,5}$}{X(1,6) to X(2,5)}} \label{sec:ex1}
We will first consider an example related to that in \sref{sec:conifolds}. If we demand that
the matrix appearing in \eqref{eq:splitquintic} is symmetric, $a_{jk} = a_{kj}$, then the
resulting family of threefolds are invariant under a further order-two symmetry, generated
by $g_2 : z_i \longleftrightarrow w_i$. As shown in \cite{Candelas:2008wb}, this family
is still generically smooth, and the entire group $\mathbb{Z}_5{\times}\mathbb{Z}_2~\cong~\mathbb{Z}_{10}$
acts freely, so we get a smooth quotient family $X^{1,6}$.
Suppose now that we ask for $g_2$ to develop a fixed point. In the ambient space,
it fixes an entire copy of $\mathbb{P}^4$, given by $w_i = z_i ~\forall~i$. Choose a single
point on this locus (as long as it is not also a fixed point of $g_5$), say
$w_i = z_i = \d_{i0}$. The evaluation of the defining polynomials at this point is
$f_i = c_{-i,-i}$, so it lies on the hypersurface if $c_{i,i} = 0 ~\forall~i$. One can check
that for arbitrary choices of the other coefficients, this point is a node on the covering
space, and there are no other singularities. This therefore corresponds to a sub-family
of $X^{1,6}$ with an isolated $\mathbb{Z}_2$-hyperconifold singularity. Such a singularity has a
crepant projective resolution, as described in \cite{Davies:2009ub}, obtained by a simple
blow-up of the singular point. This introduces an irreducible exceptional divisor, thus
increasing $h^{1,1}$ by one, and since we imposed a single
constraint\footnote{Na\"ively, it seems that we have imposed five constraints,
$c_{i,i} = 0$. However, we had the freedom to choose a generic point on the fixed
$\mathbb{P}^4$, corresponding to a four-parameter choice of possible conditions, so the number
of complex structure parameters is actually only reduced by one.} on the complex
structure of $X^{1,6}$, the resolved space has Hodge numbers $(h^{1,1},\,h^{2,1}) = (2,5)$,
and its fundamental group is $\mathbb{Z}_5$.\footnote{Note that this is in fact a new
`three-generation' manifold, with $\chi = -6$. Unfortunately, $\mathbb{Z}_5$-valued Wilson
lines cannot perform the symmetry breaking required for a realistic model.}
So we have constructed a $\mathbb{Z}_2$-hyperconifold transition
$X^{1,6} \stackrel{\mathbb{Z}_2}{\rightsquigarrow} X^{2,5}$, where the fundamental group of
the first space is $\mathbb{Z}_{10}$, and that of the second space is $\mathbb{Z}_5$.
\subsubsection{Example 2:
\texorpdfstring{$X^{1,4} \rightsquigarrow X^{2,3}$}{X(1,4) to X(2,3)}} \label{sec:ex2}
For a second example, which will also yield an interesting new manifold, consider the
$\mathrm{Dic}_3$ quotient of $X^{8,44}$, described in \sref{sec:(8,44)}. As shown in
\cite{Braun:2009qy}, there is a codimension-one locus in moduli space where the
unique order-two element of the group develops a fixed point. It is easy to check that
on the covering space, this is the only singular point, and is a node. As such, the
quotient space $X^{1,4}$ develops a $\mathbb{Z}_2$-hyperconifold singularity. Blowing up
this point yields a new Calabi-Yau manifold, with Hodge numbers $(h^{1,1},\,h^{2,1}) = (2,3)$,
as per the general formula \eqref{eq:hchodge}.
The $\mathbb{Z}_2$ subgroup of $\mathrm{Dic}_3$ is actually the centre, so it is trivially normal, and the
fundamental group of the new manifold $X^{2,3}$ is $\mathrm{Dic}_3/\mathbb{Z}_2$, which is isomorphic
to $S_3$, the symmetric group on three letters. To see this, recall that $\mathrm{Dic}_3$ is
generated by two elements, $g_3$ and $g_4$, of orders three and four respectively,
subject to the relation $g_4 g_3 g_4^{-1} = g_3^2$. So the $\mathbb{Z}_2$ subgroup is
generated by $g_4^2$, meaning that in $\mathrm{Dic}_3/\mathbb{Z}_2$, $g_4^2 \sim e$. To reflect
this, we rename $g_4$ to $g_2$, and obtain
\begin{equation*}
\mathrm{Dic}_3/\mathbb{Z}_2 ~\cong~ \langle\, g_2, g_3 ~\big\vert~ g_2^2 = g_3^3 = e ~,~ g_2 g_3 g_2 = g_3^2 \,\rangle~,
\end{equation*}
which is the standard presentation of $S_3$.
So in summary, we have constructed a $\mathbb{Z}_2$-hyperconifold transition
$X^{1,4} \stackrel{\mathbb{Z}_2}{\rightsquigarrow} X^{2,3}$, where the fundamental group of the
first space is $\mathrm{Dic}_3$, and that of the second space is $S_3$. This is the first known
Calabi-Yau threefold with fundamental group $S_3$ \cite{Braun:2010vc}.
\subsection*{Acknowledgements}
I would like to thank Eberhard Freitag for useful correspondence, and Philip Candelas
for helpful comments on a draft of this paper. This work was supported by the
Engineering and Physical Sciences Research Council [grant number EP/H02672X/1].
|
\section{ \textbf {Introduction}}
The classical laws of the iterated logarithm (LIL) as fundamental
limit theorems in probability theory play an important role in the
development of probability theory and its applications. The original
statement of the law of the iterated logarithm obtained by Khinchine
(1924) is for a class of Bernoulli random variables. Kolmogorov
(1929) and Hartman-Wintner (1941) extended Khinchine's result to
large classes of independent random variables. L\'{e}vy (1937)
extended Khinchine's result to martingales, an important class of
dependent random variables; Strassen (1964) extended
Hartman-Wintner's result to large classes of functional random
variables. After that, the research activity of LIL has enjoyed
both a rich classical period and a modern resurgence ( see, Stout
1974 for details). To extend the LIL, a lot of fairly neat methods
have been found (see, for example, De Acosta 1983), however, the key
in the proofs of LIL is the additivity of the probabilities and the
expectations. In practice, such additivity assumption is not
feasible in many areas of applications because the uncertainty
phenomena can not be modeled using additive probabilities or
additive expectations. As an alternative to the traditional
probability/expectation, capacities or nonlinear
probabilities/expectations have been studied in many fields such as
statistics, finance and economics. In statistics, capacities have
been applied in robust statistics (Huber,1981). For example, under
the assumption of 2-alternating capacity, Huber and Strassen (1973)
have generalized the Neyman-Pearson lemma. Similarly Wasserman and
Kadane (1990) have generalized the Bayes theorem for capacities. It
is well-known that, in finance, an important question is how to
calculate the price of contingent claims. The famous Black-Shores's
formula states that, if a market is complete and self-financial,
then there exists a neutral probability measure $P$ such that the
pricing of any discounted contingent claim $\xi$ in this market is
given by $E_P[\xi].$ In this case, by Kolmogorov's strong law of
large number and LIL, one can obtain the estimates of the mean
$\mu:=E_P[\xi]$ and the variance $\sigma^2 :=E_P[|\xi-\mu|^2]$ with
probability one by
$$
\mu=\lim_{n\to\infty} \frac 1n S_n, \quad \sigma=\limsup
\limits_{n\to\infty}(2n\log \log n)^{-1/2} |S_n-n\mu|
$$
where $S_n$ is the sum of the first $n$ of a sample $\{X_i\}$ with
mean $\mu$ and variance $\sigma^2.$
Statistically, an important feature of strong LLN and LIL is to
provide a frequentist perspective for mean $\mu$ and standard
variance $\sigma.$ However, if the market is incomplete, such a
neutral probability measure is no longer unique, it is a set ${\cal P}$
of probability measures. In that case, one can give sub-hedge
pricing and super-hedge pricing by ${\cal E}[\xi]:=\inf\limits_{Q\in{\cal P}}
E_Q[\xi]$ and $\mathbb{E}[\xi]:=\sup\limits_{Q\in{\cal P}}E_Q[\xi].$ Obviously,
both ${\cal E}[\cdot]$ and $\mathbb{E}[\cdot]$ as functional operators of random
variables are nonlinear. Statistically, how to calculate sub-super
hedge pricing is of interest. Motivated by sub-hedge and
super-hedge pricing and model uncertainty in finance, Peng
(2006-2009) initiated the notion of IID random variables and the
definition of $G$-normal distribution. He further obtained new
central limit theorems (CLT) under sub-linear expectations. Chen
(2009) also obtained strong laws of large numbers in this
framework. A natural question is the following: Can the classical
LIL be generalized for capacities? In this paper, adapting the
Peng's IID notion and applying Peng's CLT under sub-linear
expectations, we investigate LIL for capacities. Our result shows
that in the nonadditive setting, the supremum limit points of $
\{(2n\log \log n)^{-1/2} |S_n|\}_{n\ge 3}$ lie, with probability
(capacity) one, between the lower and upper standard variances, the
others lie, with probability (capacity) one, between zero and the
lower standard variance. This becomes the Kolmogorov and the
Hartman-Wintner law of the iterated logarithm if capacity is
additive, since in this case lower and upper variances coincide.
\section{ \textbf {Notations and Lemmas}}
In this section, we introduce some basic notations and lemmas. For
a given set ${\cal P}$ of multiple prior probability measures on
$(\Omega,{\cal F}),$ let $\mathcal {H}$ be the set of random variables
on $(\Omega,{\cal F}).$
For any $\xi\in \mathcal {H},$ we define a pair of so-called
maximum-minimum expectations $(\mathbb{E},{\cal E})$ by
$$\mathbb{E}[\xi]:=\sup_{P\in{\cal P}}E_P[\xi], \quad {\cal E}[\xi]:=\inf_{P\in{\cal P}}
E_{P}[\xi].$$ Without confusion, here and in the sequel,
$E_P[\cdot]$ denotes the classical expectation under probability
measure $P.$ We use $\mathbb{E}[\cdot]$ to denote supremum expectation
over ${\cal P}.$
Let $\xi=I_A $ for $A\in{\cal F},$ immediately, a pair $({\cal V},v)$ of
capacities is given by
$${\cal V}(A):=\sup_{P\in {\cal P}}P(A),\;\quad v(A):=\inf_{P\in{\cal P}}P(A), \quad \forall A\in{\cal F}.$$
Obviously, $\mathbb{E}$ is a sub-linear expectation in the sense that\\
\noindent{\bf Definition 1} A functional $\mathbb{E}$ on $\mathcal
{H} \mapsto(-\infty,+\infty)$ is called a sub-linear expectation,
if it satisfies the following properties: for all $X,Y\in\mathcal
{H},$
\begin{description}
\item{({\bf a})} Monotonicity: $X\geq Y$ implies
$\mathbb{E}[X]\geq\mathbb{E}[Y]$.
\item{({\bf b})} Constant preserving:
$\mathbb{E}[c]=c$, $\forall c\in\mathbb{R}$. \item{({\bf c})}
Sub-additivity: $\mathbb{E}[X+Y]\leq\mathbb{E}[X]+\mathbb{E}[Y]$.
\item{({\bf d})} Positive homogeneity: $\mathbb{E}[\lambda
X]=\lambda\mathbb{E}[X]$, $\forall\lambda\geq0$.
\end{description}
\noindent{\bf Remark} Artzner, Delbaen, Eber and Heath (1997) showed
that a sub-linear expectation indeed is a supremum expectation. That
is, if $\hat{\mathbb{E}}$ is a sub-linear expectation on $\mathcal {H}$;
then there exists a set (say $\hat{{\cal P}}$) of probability measures
such that $$\hat{\mathbb{E}}[\xi]=\sup_{P\in\hat{{\cal P}}}E_P[\xi], \quad
-\hat{\mathbb{E}}[-\xi]=\inf_{P\in\hat{{\cal P}}} E_{P}[\xi].$$ Moreover, a
sub-linear expectation $\hat{\mathbb{E}}$ can generate a pair
$(\hat{{\cal V}},\hat{v})$ of capacities denoted by
$$\hat{{\cal V}}(A):=\hat{\mathbb{E}}[I_A], \quad \hat{v}(A):=-\hat{\mathbb{E}}[-I_A], \quad \forall A\in{\cal F}.$$ Therefore,
without confusion, we sometimes call the supremum expectation as the
sub-linear expectation.
It is easy to check that the pair of capacities satisfies
$$
{\cal V}(A)+v(A^c)=1,\quad \forall A\in{\cal F}
$$
where $A^c$ is the complement set of $A.$
For ease of exposition, in this paper, we suppose that ${\cal V}$ and
$v$ are continuous in the sense that\\
\noindent{\bf Definition 2} A set function $V$:
${\cal F}\rightarrow[0,1]$ is called a continuous capacity if it
satisfies
\begin{description}
\item {({\bf 1})} $V(\phi)=0$, $V(\Omega)=1$. \item {({\bf 2})}
$V(A)\leq V(B),$ whenever $A\subset B$ and $A,B\in{\cal F}$. \item
{({\bf 3})} $V(A_{n})\uparrow V(A)$, if $A_{n}\uparrow A$. \item
{({\bf 4})} $V(A_{n})\downarrow V(A)$, if $A_{n}\downarrow A$,
where $A_{n}, A\in {\cal F}$.
\end{description}
The following is the notion of IID random variables under
sub-linear expectations introduced by Peng
\cite{peng2,peng3,peng4,peng5,peng6}.\\
\noindent{\bf Definition 3} ({\bf IID under sublinear
expectations/capacities})
\begin{description}
\item{\bf Independence:} Suppose that $Y_1,Y_2,\cdots,Y_n$ is a
sequence of random variables such that $Y_i \in\mathcal {H}.$ Random
variable $Y_n$ is said to be independent of
$X:=(Y_1,\cdots,Y_{n-1})$ under $\mathbb{E}$, if for each measurable
function $\varphi$ on $\mathbb{R}^n$ with $\varphi(X,Y_n)\in \mathcal {H}$
and $\varphi(x,Y_n)\in\mathcal {H}$ for each $x\in
\mathbb{R}^{n-1},$ we have
$$\mathbb{E}[\varphi(X,Y_n)]=\mathbb{E}[\overline{\varphi}(X)],$$
where $\overline{\varphi}(x):=\mathbb{E}[\varphi(x,Y_n)]$ and
$\overline{\varphi}(X)\in\mathcal {H}$.
\item{\bf Identical distribution:} Random variables $X$ and $Y$
are said to be identically distributed, denoted by
$X\overset{d}{=}Y$, if for each measurable function $\varphi$ such
that $\varphi(X), \; \varphi(Y)\in \mathcal {H}$,
$$\mathbb{E}[\varphi(X)]=\mathbb{E}[\varphi(Y)].$$
\item{\bf IID random variables:} A sequence of random variables
$\{X_i\}_{i=1}^\infty$ is said to be IID, if $X_i\overset{d}{=}X_1$
and $X_{i+1}$ is independent of $Y:=(X_1,\cdots,X_i)$ for each $i\ge
1.$
\item{\bf Pairwise independence:} Random variable $X$ is said to
be pairwise independent to $Y$ under capacity $V,$ if for all
subsets $D$ and $G\in{\cal B}(\mathbb{R}),$
$$
V(X\in D, Y\in G)=V(X\in D)V(Y\in G).
$$
\end{description}
The following lemma shows the relation between Peng's independence
and pairwise independence.\\
\noindent{\bf Lemma 1} Suppose that $X,Y \in \mathcal {H}$ are two
random variables. $\mathbb{E}$ is a sub-linear expectation and $({\cal V},v)$ is
the pair of capacities generated by $\mathbb{E}.$ If random variable $X$
is independent of $Y$ under $\mathbb{E}$, then $X$ also is pairwise
independent of $Y$ under capacities ${\cal V}$ and $v$. \\
\noindent{\bf Proof} If we choose $\varphi(x,y)=I_D(x)I_G(y)$ for
$\mathbb{E}, $ by the definition of Peng's independence, it is easy to
obtain
$$
{\cal V}(X\in D, Y\in G)={\cal V}(X\in D){\cal V}(Y\in G).
$$
Similarly, if we choose $\varphi(x,y)=-I_D(x)I_G(y)$ for $\mathbb{E},$
it is easy to obtain
$$
v(X\in D, Y\in G)=v(X\in D)v(Y\in G).
$$
The proof is complete.\\
Borel-Cantelli Lemma is still true for capacity under some
assumptions.\\
\noindent{\bf Lemma 2} Let $\{A_n,n\geq1\}$ be a sequence of
events in ${\cal F}$ and $({\cal V},v)$ be a pair of capacities generated by
sub-linear expectation $\mathbb{E}$.
\begin{description}
\item{(1)} If $\sum\limits_{n=1}^\infty{\cal V}(A_n)<\infty,$ then
${\cal V}\left(\bigcap\limits_{n=1}^\infty\bigcup\limits_{i=n}^\infty
A_i\right)=0.$
\item{(2)} Suppose that $\{A_n,n\geq1\}$ are pairwise independent
with respect to ${\cal V}$ , i.e.,
$${\cal V}\left(\bigcap\limits_{i=1}^\infty
A_i^c\right)=\prod_{i=1}^\infty {\cal V}(A_i^c).$$ If
$\sum\limits_{n=1}^\infty{v}(A_n)=\infty $, then
$$v\left(\bigcap\limits_{n=1}^\infty\bigcup\limits_{i=n}^\infty A_i\right)=1.$$
\end{description}
\noindent{\bf Proof}
$$\begin{array}{lcl}
0&\leq& {\cal V}(\bigcap\limits_{n=1}^\infty\bigcup\limits_{i=n}^\infty
A_i)\\
&\leq& {\cal V}(\bigcup\limits_{i=n}^\infty
A_i)\\
&\leq &\sum\limits_{i=n}^\infty{\cal V}(A_i) \to 0,\; \hbox{ as}\ \; n\to
\infty.
\end{array}
$$
The proof of (1) is complete.
$$
\begin{array}{lcl}
0&\leq &1-v(\bigcap\limits_{n=1}^\infty\bigcup\limits_{i=n}^\infty A_i)\\
&=& 1-\lim\limits_{n\to \infty}v(\bigcup\limits_{i=n}^\infty A_i)\\
&=& \lim\limits_{n\to \infty}[1-v(\bigcup\limits_{i=n}^\infty A_i)]\\
& =&\lim\limits_{n\to \infty}{\cal V}(\bigcap\limits_{i=n}^\infty A_i^c)\\
&=&\lim\limits_{n\to \infty}\prod\limits_{i=n}^\infty
{\cal V}(A_i^c)\\
&=&\lim\limits_{n\to
\infty}\prod\limits_{i=n}^\infty(1-v(A_i))\\
&\leq&\lim\limits_{n\to \infty}\prod\limits_{i=n}^\infty{\rm
exp}(-v(A_i))\\
&=&\lim\limits_{n\to \infty}{\rm exp}(-\sum\limits_{i=n}^\infty
v(A_i))=0.
\end{array}.$$ We complete the proof of (2).\\
\noindent{\bf Definition 4} ({\bf $G$-normal distribution}, see
Definition 10 in Peng \cite{peng3}) Given a sub-linear expectation
$\mathbb{E}$, a random variable $\xi\in\mathcal {H}$ with
$$\overline{\sigma}^2=\mathbb{E}[\xi^2],\ \ \ \ \underline{\sigma}^2={\cal E}[\xi^2]$$ is called a $G$-normal distribution,
denoted by ${\cal N}(0;[{\underline\sigma}^2,{\overline\sigma}^2])$,
if for any bounded Lipschtiz function $\phi,$ writting
$u(t,x):=\mathbb{E}[\phi(x+\sqrt t \xi)],$ $(t,x)\in[0,\infty)\times \mathbb{R}$,
then $u$ is the viscosity solution of PDE:
$$
\partial_t u -G(\partial^2_{xx} u)=0,\quad u(0,x)=\phi(x),
$$
where $G(x):=\frac 12(\overline \sigma^2 x^+-\underline \sigma^2
x^-)$ and $x^+:=\max\{x,0\}$, $x^-:=(-x)^+$.\\
The following lemma can be found in Denis, Hu and Peng
\cite{Denis}.\\
\noindent{\bf Lemma 3} Suppose that $\xi$ is $G$-normal
distributed by ${\cal
N}(0;[{\underline\sigma}^2,{\overline\sigma}^2])$. Let $P$ be a
probability measure and $\phi$ be a bounded continuous function.
If $\{B_t\}_{t\geq0}$ is a $P$-Brownian motion, then
$$\mathbb{E}[\phi(\xi)]=\sup_{\theta\in
\Theta}E_P\left[\phi\left(\int_0^1\theta_sdB_s\right)\right],\quad
{\cal E}[\phi(\xi)]=\inf_{\theta\in
\Theta}E_P\left[\phi\left(\int_0^1\theta_sdB_s\right)\right],
$$
where $$\Theta:=\left\{\{\theta_t\}_{t\ge 0}: \theta_t \;
\hbox{is
${\cal F}_t$-adaped process such that} \;\; \underline \sigma \leq
\theta_t\leq \overline \sigma\right\},$$
$${\cal F}_t:=\sigma\{B_s,0\leq s\leq t\}\vee{\cal N},\ \ \ \ {\cal N}\ \hbox{is the collection of P-null subsets}.$$
For the sake of completeness, the sketched proof of Lemma 3 is given in Appendix A.\\
With the notion of IID under sub-linear expectations, Peng shows the
central limit theorem under sub-linear expectations (see Theorem
5.1 in Peng \cite{peng5}).\\
\noindent{\bf Lemma 4} ({\bf Central limit theorem under
sub-linear expectations}) Let $\{X_i\}_{i=1}^\infty$ be a sequence
of IID random variables. We further assume that
$$\mathbb{E}[X_1]={\cal E}[X_1]=0.$$ Then the sequence $\{\overline{S}_n\}_{n=1}^\infty$
defined by $$\overline{S}_n:=\frac{1}{\sqrt{n}}\sum\limits_{i=1}^nX_i$$ converges in law to $\xi$, i.e.,
$$\lim\limits_{n\rightarrow\infty}\mathbb{E}[\varphi(\overline{S}_n)]=\mathbb{E}[\varphi(\xi)],$$
for any continuous function $\varphi$ satisfying the linear growth
condition, where $\xi$ is a $G$-normal distribution.\\
\noindent{\bf Remark 1} Suppose that
$\mathbb{E}[X_1^2]={\overline{\sigma}}^2$, $\overline{\sigma}>0$ and
$\varphi$ is a convex function, then, we have,
$$\mathbb{E}[\varphi(\xi)]=\frac{1}{\sqrt{2\pi{\overline{\sigma}}^2}}
\int_{-\infty}^\infty\varphi(y){\rm
exp}\left(-\frac{y^2}{2{\overline{\sigma}}^2}\right){\rm d}y.$$
\section{Main results}
In this section, we will prove the following LIL for capacities:\\
\noindent{\bf Theorem 1} Let $\{X_n\}_{n=1}^\infty$ be a sequence
of bounded IID random variables for sub-linear expectation $\mathbb{E}$
with zero means and bounded variances, i.e.,
\begin{description}
\item{(A.1)} $\mathbb{E}[X_1]={\cal E}[X_1]=0$, \item{(A.2)}
$\mathbb{E}[X_1^2]={\overline{\sigma}}^2$,
${\cal E}[X_1^2]={\underline{\sigma}}^2$, where
$0<\underline{\sigma}\leq\overline{\sigma}<\infty$.\\
\end{description}
Denote
$S_n=\sum\limits_{i=1}^nX_i$. Then
\begin{description}
\item {(I)}$$v\left({\underline
{\sigma}}\leq\limsup\limits_{n\rightarrow\infty}\frac{S_n}{\sqrt{2n{\rm
loglog}n}} \leq{\overline{\sigma}}\right)=1.$$
\item{(II)}$$v\left({-\overline
{\sigma}}\leq\liminf\limits_{n\rightarrow\infty}\frac{S_n}{\sqrt{2n{\rm
loglog}n}} \leq-{\underline{\sigma}}\right)=1.$$ \item{(III)}
Suppose that $C(\{x_n\})$ is the cluster set of a sequence of
$\{x_n\}$ in $\mathbb{R},$ then
$$
v\left (C\left(\left\{\frac{S_n}{\sqrt{2n{\rm
loglog}n}}\right\}\right)\backslash
\left\{\limsup\limits_{n\rightarrow\infty}\frac{S_n}{\sqrt{2n{\rm
loglog}n}},\liminf\limits_{n\rightarrow\infty}\frac{S_n}{\sqrt{2n{\rm
loglog}n}}\right\}=(-\underline\sigma, \underline\sigma)\right)=1.
$$
\end{description}
In order to prove Theorem 1, we need the following lemmas.\\
\noindent{\bf Lemma 5} Suppose $\xi$ is distributed to $G$ normal
${\cal N}(0;[{\underline\sigma}^2,{\overline\sigma}^2])$, where
$0<\underline{\sigma}\leq\overline{\sigma}<\infty$. Let $\phi$ be
a bounded continuous function. Furthermore, if $\phi$ is a
positively even function, then, for any $b\in \mathbb{R}$,
$$
e^{-\frac {b^2}{2{\underline\sigma}^2}}{\cal E}[\phi(\xi)]\leq
{\cal E}[\phi(\xi-b)].
$$
\noindent{\bf Proof} Let $P$ be a probability measure,
$\{B_t\}_{t\geq0}$ be a $P$-Brownian motion. Since $\xi$ is
distributed to $G$-normal, by Lemma 3, we have
$$
{\cal E}[\phi(\xi)]=\inf_{\theta\in
\Theta}E_P\left[\phi\left(\int_0^1\theta_s dB_s\right)\right].
$$
For any $\theta \in\Theta,$ write $\tilde{B_t}:=B_t-\int_0^t\frac
b{\theta_s}ds$. By Girsanov's theorem, $\{\tilde B_t\}_{t\geq0}$ is a
$Q$-Brownian motion under $Q$ denoted by
$$
\frac{dQ}{dP}:=e^{-\frac 12 \int_0^1(\frac
b{\theta_s})^2ds+\int_0^1\frac b{\theta_s}dB_s}.
$$
That is
$$
\frac{dP}{dQ}=e^{-\frac 12 \int_0^1(\frac
b{\theta_s})^2ds-\int_0^1\frac b{\theta_s}d{\tilde {B}_s}}.
$$
Thus
\begin{equation}\label{IE1}
\begin{array}{lcl}
E_P\left [\phi\left(\int_0^1\theta_s dB_s-b\right)\right]
&=&E_P\left[\phi\left(\int_0^1\theta_t d(B_t-\int_0^t\frac b{\theta_s}ds)\right)\right]\\
&=&E_P\left[\phi\left(\int_0^1\theta_s d\tilde B_s\right)\right] \\
&=&E_Q\left[\phi\left(\int_0^1\theta_s d\tilde B_s\right)\cdot
e^{-\frac
12\int_0^1(\frac b{\theta_s})^2ds-\int_0^1\frac b{\theta_s}d\tilde B_s}\right]\\
&\ge & e^{-\frac 12\left(\frac
b{\underline\sigma}\right)^2}E_Q\left[\phi\left(\int_0^1\theta_s
d\tilde B_s\right)\cdot e^{-\int_0^1\frac b{\theta_s}d\tilde
B_s}\right].
\end{array}
\end{equation}
We now prove that if $\phi$ is even, then
$$
E_Q\left[\phi\left(\int_0^1\theta_s d\tilde B_s\right)\cdot
e^{-\int_0^1\frac b{\theta_s}d\tilde B_s}\right]\ge
E_Q\left[\phi\left(\int_0^1\theta_s d\tilde B_s\right)\right].
$$
In fact, let $\overline B_t:=-\tilde B_t,$ then $\{\overline
B_t\}_{t\geq0}$ is also a $Q$-Brownian motion. Note that the
assumption that function $\phi$ is even, therefore
$$
\begin{array}{lcl}
E_Q\left[\phi\left(\int_0^1\theta_s d\tilde B_s\right)\cdot
e^{-\int_0^1\frac b{\theta_s}d\tilde B_s}\right]
&=&E_Q\left[\phi\left(\int_0^1\theta_s d\overline
B_s\right)\cdot e^{-\int_0^1\frac b{\theta_s}d\overline B_s}\right]\\
&=& E_Q\left[\phi\left(-\int_0^1\theta_s d\tilde
B_s\right)\cdot e^{\int_0^1\frac b{\theta_s}d \tilde B_s}\right]\\
&=&
E_Q\left[\phi\left(\int_0^1\theta_s d\tilde B_s\right)\cdot
e^{\int_0^1\frac b{\theta_s}d \tilde B_s}\right].
\end{array}
$$
Since $ \frac {e^{\int_0^1b/\theta_sd\tilde B_s}+
e^{-\int_0^1b/\theta_sd\tilde B_s}}2\ge 1,$ we have
$$
\begin{array}{lcl}
E_Q\left[\phi\left(\int_0^1\theta_s d\tilde B_s\right)\cdot
e^{-\int_0^1\frac b{\theta_s}d\tilde B_s}\right] &=&\frac 12
E_Q\left[\phi\left(\int_0^1\theta_s d\tilde B_s\right)\cdot\left(
e^{\int_0^1b/\theta_sd\tilde B_s}+
e^{-\int_0^1b/\theta_sd\tilde B_s}\right)\right]\\
&\ge & E_Q\left[\phi\left(\int_0^1\theta_s d\tilde
B_s\right)\right].
\end{array}
$$
From (1), we have
$$
\begin{array}{lcl}
{\cal E}[\phi(\xi-b)]=\inf\limits_{\theta\in\Theta}E_P\left[\phi\left(\int_0^1\theta_s
d B_s-b\right)\right] &\ge & e^{-\frac 12(b/\underline\sigma)^2}
\inf\limits_{\theta\in\Theta} E_Q\left[\phi\left(\int_0^1\theta_s d\tilde B_s\right)\right]\\
&=& e^{-\frac
12(b/\underline\sigma)^2}\inf\limits_{\theta\in\Theta}
E_P\left[\phi\left(\int_0^1\theta_s d B_s\right)\right]\\
&=& e^{-\frac 12(b/\underline\sigma)^2} {\cal E}[\phi(\xi)].
\end{array}
$$
The proof of Lemma 5 is complete.\\
\noindent{\bf Lemma 6} Under the assumptions of Theorem 1, then,
for each $r>2$, there exists a positive constant $K_r$ such that
$$\mathbb{E}[ \max\limits_{i\leq n}|S_{m,i}|^r]\leq
K_rn^\frac{r}{2}\ \ \ \hbox{for all }\ \ m\geq0,$$ where
$S_{m,n}=\sum\limits_{i=m+1}^{m+n}X_i$. \\
\noindent{\bf Proof.} First, we prove that there exists a positive
constant $C_r$ such that \begin{equation}\label{R3}
\sup\limits_{m\geq0}\mathbb{E}|S_{m,n}|^r\leq C_rn^\frac{r}{2}.
\end{equation}
By Lemma 4 and Remark 1, it is easy to check that
$$\lim\limits_{n\rightarrow\infty}\mathbb{E}[|S_{m,n}/\sqrt{n}|^r]
=\mathbb{E}[|\xi|^r]=\frac{1}{\sqrt{2\pi{\overline{\sigma}}^2}}
\int_{-\infty}^\infty|y|^r{\rm
exp}\left(-\frac{y^2}{2{\overline{\sigma}}^2}\right){\rm
d}y<\infty.$$ So, we can choose
$$D_r>\frac{1}{\sqrt{2\pi{\overline{\sigma}}^2}}
\int_{-\infty}^\infty|y|^r{\rm
exp}\left(-\frac{y^2}{2{\overline{\sigma}}^2}\right){\rm d}y,$$
then there exists $n_0$ such that $\forall n\geq n_0$,
$$\mathbb{E}|S_{m,n}|^r\leq D_rn^\frac{r}{2}.$$
Note that $\{X_n\}_{n=1}^\infty$ is a bounded sequence, then there
exists a constant $M>0$, such that, for each $n$, $|X_n|\leq M$. So
we can obtain (2) holds. Hence, in a manner similar to Theorem 3.7.5
of Stout \cite{stout}, we can obtain
$$\mathbb{E}[ \max\limits_{i\leq n}|S_{m,i}|^r]\leq
K_rn^\frac{r}{2}\ \ \ \hbox{for all }\ \ m\geq0.$$
\noindent{\bf Lemma 7} Under the assumptions of Theorem 1, if
$$v\left(\limsup\limits_{n\rightarrow\infty}\frac{|S_n|}{\sqrt{2n{\rm
loglog}n}} \leq{\overline{\sigma}}\right)=1,$$ then, for any $b\in
\mathbb{R}$ satisfying $|b|<\underline{\sigma}$,
$$v\left(\liminf\limits_{n\rightarrow\infty}\left|\frac{S_n}{\sqrt{2n{\rm
loglog}n}}-b\right|=0 \right)=1.
$$
\noindent{\bf Proof} We only need to prove that for any
$\epsilon>0$,
$$v\left(\liminf\limits_{n\rightarrow\infty}\left|\frac{S_n}{\sqrt{2n{\rm
loglog}n}}-b\right|\leq \epsilon \right)=1.$$ To do so, we only
need to prove that there exists an increasing subsequence
$\{n_k\}$ of $\{n\}$ such that
\begin{equation}\label{R3}
v\left(\bigcap_{m=1}^{\infty}\bigcup_{k=m}^{\infty}\{|S_{n_k}/\sqrt{2n_k{\rm
loglog}n_k}-b|\leq\epsilon\}\right)=1.
\end{equation} Indeed, let us choose $n_k:=k^k$ for $k\ge 1.$
For each $t>0$, write $$N_k:=[(n_{k+1}-n_k)^2t^2/2n_{k+1}{\rm
loglog}n_{k+1}],$$
$$m_k:=[2n_{k+1}{\rm loglog}n_{k+1}/t^2(n_{k+1}-n_k)],$$ $$r_k:=\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}/tm_k.$$ Since $\{X_n\}_{n=1}^\infty$ is a
sequence of IID random variables under sub-linear expectation
$\mathbb{E}$, we have
\begin{equation}\label{R3}\begin{array}{lcl}
&&v\left( \left |\frac {S_{n_{k+1}}-S_{n_k}}{\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}}-b\right|\leq \epsilon\right)=
v\left(b-\epsilon\leq
\frac{S_{n_{k+1}-n_k}}{\sqrt{2n_{k+1}{\rm loglog}n_{k+1}}}\leq b+\epsilon\right)\\
&\geq& v\left(b-\epsilon/2\leq \frac{S_{N_k
m_k}}{\sqrt{2n_{k+1}{\rm loglog}n_{k+1}}}\leq
b+\epsilon/2\right)\cdot v\left(-\epsilon/2\leq
\frac{S_{n_{k+1}-n_k}-S_{N_k m_k}}{\sqrt{2n_{k+1}{\rm loglog}n_{k+1}}}\leq\epsilon/2\right)\\
&\geq& v\left(tm_k(b-\epsilon/2)\leq \frac{S_{N_k m_k}}{r_k}\leq
tm_k(b+\epsilon/2)\right)\cdot v\left(-\epsilon/2\leq
\frac{S_{n_{k+1}-n_k-N_k m_k}}{\sqrt{2n_{k+1}{\rm loglog}n_{k+1}}}\leq\epsilon/2\right)\\
&\geq& \left(v\left(bt-\epsilon t/2\leq \frac{S_{N_k}}{r_k}\leq
bt+\epsilon t/2\right)\right)^{m_k}\cdot v\left(-\epsilon/2\leq
\frac{S_{n_{k+1}-n_k-N_k m_k}}{\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}}\leq\epsilon/2\right)\\
&\geq& \left({\cal E}\left[\phi(S_{N_k}/r_k
-bt)\right]\right)^{m_k}\cdot v\left(-\epsilon/2\leq
\frac{S_{n_{k+1}-n_k-N_k m_k}}{\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}}\leq\epsilon/2\right),
\end{array}
\end{equation} where $\phi(x)$ is a even function defined by
$$
\phi(x):=\left\{
\begin{array}{l}
1-e^{|x|-\epsilon t/2},\quad\quad |x|\leq \epsilon t/2;\\
0, \quad\quad\quad\quad \quad\;\; |x|>\epsilon t/2.
\end{array}
\right.
$$
Note the fact that $N_k\rightarrow\infty$, as $k\rightarrow\infty$
and applying Lemmas 4 and 5,
$$
\log {\cal E}[\phi(S_{N_k}/r_k -bt)]\to \log {\cal E}[\phi(\xi -bt)]\ge
-\frac {b^2t^2}{2\underline\sigma^2} +\log {\cal E}[\phi(\xi)],\ \ \
\hbox{as}\ \ \ k\to\infty. $$ Thus
\begin{equation}\label{R3}\begin{array}{lcl} &&{\rm
log}\left({\cal E}[\phi(S_{N_k}/r_k -bt)]\right)^{m_k}\cdot(n_{k+1}-n_k)/2n_{k+1}{\rm loglog}n_{k+1}\\
&=&(n_{k+1}-n_k)/2n_{k+1}{\rm loglog}n_{k+1}\cdot m_k{\rm
log}{\cal E}[\phi(S_{N_k}/r_k -bt)]\\
&\to&t^{-2}\log {\cal E}[\phi(\xi -bt)]\ge-\frac
{1}{2}(b/\underline{\sigma})^2+t^{-2}\log
{\cal E}[\phi(\xi)].\end{array}\end{equation}
However,
\begin{equation}\label{R3}\liminf\limits_{t\rightarrow\infty}t^{-2}\log
{\cal E}[\phi(\xi -bt)]\geq-\frac
{1}{2}(b/\underline{\sigma})^2.\end{equation} So, from (5) and
(6), we have, for any $\delta>0$ and large enough $t$,
\begin{equation}\label{R3}\lim\limits_{k\rightarrow\infty}{\rm
log}\left({\cal E}[\phi(S_{N_k}/r_k
-bt)]\right)^{m_k}\cdot(n_{k+1}-n_k)/2n_{k+1}{\rm
loglog}n_{k+1}\geq-\frac
{1}{2}(|b|/\underline{\sigma}+\delta/2)^2.\end{equation} On the
other hand, by Chebyshev's inequality,
$${\cal V}\left(\left|\frac{S_{n_{k+1}-n_k-N_k m_k}}{\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}}\right|>\epsilon/2\right)\leq2(n_{k+1}-n_k-{N_k
m_k){\overline{\sigma}}^2/\epsilon^2n_{k+1}{\rm
loglog}n_{k+1}}\rightarrow0,\ \ \ \hbox{as}\ \ \
k\rightarrow\infty.$$ So, as $k\rightarrow\infty$,
\begin{equation}\label{R3}\begin{array}{lcl} &&\left(n_{k+1}-n_k\right)/2n_{k+1}{\rm
loglog}n_{k+1}\cdot{\rm log}v\left(-\epsilon/2\leq
\frac{S_{n_{k+1}-n_k-N_k m_k}}{\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}}\leq\epsilon/2\right)\\
&=&(n_{k+1}-n_k)/2n_{k+1}{\rm loglog}n_{k+1}\cdot{\rm
log}\left(1-{\cal V}\left(\left|\frac{S_{n_{k+1}-n_k-N_k
m_k}}{\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}}\right|>\epsilon/2\right)\right)\rightarrow0.\end{array}\end{equation}
Therefore, from (4), (7) and (8), we have
$$\liminf\limits_{k\rightarrow\infty}\left(n_{k+1}-n_k\right)/2n_{k+1}{\rm
loglog}n_{k+1}\cdot{\rm log}v\left( \left |\frac
{S_{n_{k+1}}-S_{n_k}}{\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}}-b\right|\leq \epsilon\right)\geq-\frac
{1}{2}(|b|/\underline{\sigma}+\delta/2)^2.$$ Now we choose
$\delta>0$ such that $|b/\underline{\sigma}|+\delta<1$. Then, for
given $\delta>0$, there exists $k_0$ such that $\forall k\geq
k_0$,
$$\begin{array}{lcl} &&v\left( \left |\frac
{S_{n_{k+1}}-S_{n_k}}{\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}}-b\right|\leq \epsilon\right)\\
&\geq&{\rm exp}\left\{-2n_{k+1}{\rm
loglog}n_{k+1}/\left(n_{k+1}-n_k\right)\cdot\left(\left(|b/\underline{\sigma}|+\delta\right)/2\right)\right\}\\
&\sim&{\rm exp}(-(|b/\underline{\sigma}|+\delta){\rm
loglog}n_{k+1}).\end{array}$$ Thus
$$\sum\limits_{k=1}^\infty v\left( \left |\frac {S_{n_{k+1}}-S_{n_k}}{\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}}-b\right|\leq \epsilon\right)=\infty.$$ Using the
second Borel-Cantelli Lemma, we have
$$
\liminf_{k\to\infty}\left|\frac{ S_{n_k}-S_{n_{k-1}} }
{\sqrt{2n_{k}{\rm loglog}n_{k}}}-b\right|\leq \epsilon,\quad
\hbox{a.s.} \; v.
$$
But
\begin{equation}\label{ED}
\left|\frac{S_{n_k}}{\sqrt{2n_{k}{\rm loglog}n_{k}}}-b\right|\leq
\left|\frac{ S_{n_k}-S_{n_{k-1}} }{\sqrt{2n_{k}{\rm
loglog}n_{k}}}-b\right|+\frac{|S_{n_{k-1}}|}{\sqrt{2n_{k-1}{\rm
loglog}n_{k-1}}}\frac{\sqrt{2n_{k-1}{\rm
loglog}n_{k-1}}}{\sqrt{2n_{k}{\rm loglog}n_{k}}}.
\end{equation}
Note the following fact
$$
\frac{{n_{k-1}}}{n_k}\to 0,\; \hbox{as}\ \ \; k\to \infty
$$
and
$$
\limsup_{n\to\infty} |S_n|/\sqrt{2n{\rm loglog}n}\leq \overline
\sigma,\quad \hbox{a.s.} \; v.
$$
Hence, from inequality (9), for any $\epsilon >0,$
$$
\liminf_{k\to\infty}\left|\frac{S_{n_k}}{\sqrt{2n_{k}{\rm
loglog}n_{k}}}-b\right|\leq \epsilon, \hbox{ a.s.}\; v,
$$ therefore,
$$v\left(\liminf\limits_{n\rightarrow\infty}\left|\frac{S_n}{\sqrt{2n{\rm
loglog}n}}-b\right|\leq \epsilon \right)=1.
$$ Since $\epsilon$ is arbitrary, we have
$$v\left(\liminf\limits_{n\rightarrow\infty}\left|\frac{S_n}{\sqrt{2n{\rm
loglog}n}}-b\right|=0\right)=1.$$
We complete the proof of Lemma 7.\\
\noindent{\bf The Proof of Theorem 1} (I) First, we prove that
$$v\left(\limsup\limits_{n\rightarrow\infty}\frac{S_n}{\sqrt{2n{\rm
loglog}n}} \leq{\overline{\sigma}}\right)=1.$$
For each $\epsilon>0$ and $\lambda>0$, by Markov's inequality,
\begin{equation}\label{R3}
\begin{array}{lcl}
{\cal V}\left(\frac{S_n}{\sqrt{2n{\rm
loglog}n}}>(1+\epsilon)\overline{\sigma}\right)&=&{\cal V}\left(\frac{S_n}{\sqrt{n\overline{\sigma}^2}}
>(1+\epsilon)\sqrt{2{\rm loglog}n}\right)\\
&\leq &{\rm exp}\left(-2(1+\epsilon)^2\lambda{\rm
loglog}n\right)\mathbb{E}\left[{\rm
exp}\left(\lambda\left(\frac{S_n}{\sqrt{n\overline{\sigma}^2}}\right)^2\right)\right].
\end{array}
\end{equation}
On the other hand, by Lemma 4 and Remark 1, we have, if
$\lambda<\frac{1}{2}$,
$$\lim\limits_{n\rightarrow\infty}\mathbb{E}\left[{\rm
exp}\left(\lambda\left(\frac{S_n}{\sqrt{n\overline{\sigma}^2}}\right)^2\right)\right]=\frac{1}{\sqrt{2\pi}}
\int_{-\infty}^\infty{\rm exp}(\lambda y^2){\rm
exp}\left(-\frac{y^2}{2}\right){\rm d}y<\infty.$$ Fixing $\beta>1$,
for each $\epsilon>0$, we can choose
$\lambda_\epsilon\in(0,\frac{1}{2})$ such that
$\beta=2(1+\epsilon)^2\lambda_\epsilon>1$. So, we can choose
$$C_\epsilon>\frac{1}{\sqrt{2\pi}}\int_{-\infty}^\infty{\rm exp}\left(\lambda_\epsilon
y^2\right){\rm exp}\left(-\frac{y^2}{2}\right){\rm d}y,$$ then
there exists $n_0$ such that $\forall n\geq n_0$,
\begin{equation}\label{R3} E\left[{\rm
exp}\left(\lambda_\epsilon\left(\frac{S_n}{\sqrt{n\overline{\sigma}^2}}\right)^2\right)\right]\leq
C_\epsilon.
\end{equation}
From (10) and (11), we can obtain, $\forall n\geq n_0$,
$${\cal V}\left(\frac{S_n}{\sqrt{2n{\rm loglog}n}}>(1+\epsilon)\overline{\sigma}\right)\leq C_\epsilon{\rm
exp}(-\beta{\rm loglog}n).$$
Choose $0<\alpha<1$ such that $\alpha\beta>1$. Let
$n_k:=[e^{k^\alpha}]$ for $k\geq1$. Then
$$\sum_{n_k\geq n_0}{\cal V}\left(\frac{S_{n_k}}{\sqrt{2n_k{\rm loglog}{n_k}}}
>(1+\epsilon)\overline{\sigma}\right)\leq D_\epsilon\sum_{n_k\geq n_0}k^{-\alpha\beta}<\infty,$$ where $D_\epsilon$ is a positive
constant. By the first Borel-Cantelli Lemma, we can get
$${\cal V}\left(\bigcap\limits_{m=1}^\infty\bigcup_{k=m}^\infty\left\{\frac{S_{n_k}}{\sqrt{2n_k{\rm loglog}{n_k}}}
>(1+\epsilon)\overline{\sigma}\right\}\right)=0.$$ Also
$$v\left(\limsup\limits_{k\rightarrow\infty}\frac{S_{n_k}}{\sqrt{2n_k{\rm
loglog}n_k}} \leq(1+\epsilon){\overline{\sigma}}\right)=1.$$
Let $M_k:=\max\limits_{n_k\leq
n<n_{k+1}}\frac{|S_n-S_{n_k}|}{\sqrt{2n_k{\rm loglog}n_k}}$ for
$k\geq1$. For each $k\geq1$, $$\frac{S_n}{\sqrt{2n{\rm
loglog}n}}\leq\frac{S_{n_k}}{\sqrt{2n_k{\rm
loglog}n_k}}\frac{\sqrt{2n_k{\rm loglog}n_k}}{\sqrt{2n{\rm
loglog}n}}+M_k\frac{\sqrt{2n_k{\rm loglog}n_k}}{\sqrt{2n{\rm
loglog}n}},$$ for $n_k\leq n<n_{k+1}$. For given $\alpha$, we
choose $p>2$ such that $p(1-\alpha)\geq2$. By Lemma 6, we get
\begin{equation}\label{R3}\sum\limits_{k=1}^\infty\mathbb{E}[M_k^p]\leq
K_p\sum\limits_{k=1}^\infty\frac{(n_{k+1}-n_k)^\frac{p}{2}}
{(2n_k{\rm loglog}n_k)^\frac{p}{2}}\leq
D_p^{'}\sum\limits_{k=1}^\infty k^{-\frac{p(1-\alpha)}{2}}({\rm
log}k)^{-\frac{p}{2}}<\infty,\end{equation} where $D_p^{'}$ is a
positive constant. From (12) and by Chebyshev's inequality, for
each $\epsilon>0$,
$$\sum\limits_{k=1}^\infty{\cal V}\left(M_k>\epsilon\right)\leq\sum\limits_{k=1}^\infty\frac{\mathbb{E}[M_k^p]}{\epsilon^p}<\infty.$$ Hence, by the first
Borel-Cantelli Lemma again,
$$v\left(\limsup\limits_{k\rightarrow\infty}M_k\leq\epsilon\right)=1.$$ Since $\epsilon$ is arbitrary, we have
$$v\left(\lim\limits_{k\rightarrow\infty}M_k=0\right)=1.$$ Noting
that $$\frac{\sqrt{2n_k{\rm loglog}n_k}}{\sqrt{2n_{k+1}{\rm
loglog}n_{k+1}}}\rightarrow1,\ \ \ \hbox{as}\ \ \
k\rightarrow\infty,$$ we have
$$\begin{array}{lcl}&&v\left(\limsup\limits_{n\rightarrow\infty}\frac{S_{n}}{\sqrt{2n{\rm
loglog}n}} \leq(1+\epsilon){\overline{\sigma}}\right)\\
&\geq&
v\left(\left\{\limsup\limits_{k\rightarrow\infty}\frac{S_{n_k}}{\sqrt{2n_k{\rm
loglog}n_k}}
\leq(1+\epsilon){\overline{\sigma}}\right\}\bigcap\left\{\lim\limits_{k\rightarrow\infty}M_k=0\right\}\right)\\
&=&1,\end{array}$$ which implies
\begin{equation}\label{R3} v\left(\limsup\limits_{n\rightarrow\infty}\frac{S_{n}}{\sqrt{2n{\rm
loglog}n}}\leq{\overline{\sigma}}\right)=1.\end{equation}
Similarly, considering the sequence $\{-X_n\}_{n=1}^\infty$, it
suffices to obtain
\begin{equation}\label{R3}v\left(\limsup\limits_{n\rightarrow\infty}\frac{-S_{n}}{\sqrt{2n{\rm
loglog}n}}\leq{\overline{\sigma}}\right)=1.$$ Also
$$v\left(\liminf\limits_{n\rightarrow\infty}\frac{S_{n}}{\sqrt{2n{\rm
loglog}n}}\geq-{\overline{\sigma}}\right)=1.\end{equation}
Now we prove that
$$v\left(\limsup\limits_{n\rightarrow\infty}\frac{S_n}{\sqrt{2n{\rm
loglog}n}}\ge {\underline {\sigma}}\right)=1.$$ Indeed, from (13)
and (14), it is easy to obtain
$$v\left(\limsup\limits_{n\rightarrow\infty}\frac{|S_{n}|}{\sqrt{2n{\rm
loglog}n}}\leq{\overline{\sigma}}\right)=1.$$ For any number
$b\in(0, \underline \sigma),$ noting the fact that
$|b|<\underline\sigma$, by Lemma 7, we have
$$v\left(\liminf\limits_{n\rightarrow\infty}\left|\frac{S_n}{\sqrt{2n{\rm
loglog}n}}-b\right|=0\right)=1,$$ which implies that
\begin{equation}\label{ED}v\left(\limsup\limits_{n\rightarrow\infty}\frac{S_n}{\sqrt{2n{\rm
loglog}n}}\ge {\underline {\sigma}}\right)=1.\end{equation}
So, from (13) and (15), we can obtain
$$v\left({\underline
{\sigma}}\leq\limsup\limits_{n\rightarrow\infty}\frac{S_n}{\sqrt{2n{\rm
loglog}n}} \leq{\overline{\sigma}}\right)=1.$$
The proof of (I) is complete.
(II) Considering the sequence $\{-X_n\}_{n=1}^\infty$, by (I), it
suffices to obtain $$v\left({\underline
{\sigma}}\leq\limsup\limits_{n\rightarrow\infty}\frac{-S_n}{\sqrt{2n{\rm
loglog}n}} \leq{\overline{\sigma}}\right)=1.$$ Thus
$$v\left({-\overline
{\sigma}}\leq\liminf\limits_{n\rightarrow\infty}\frac{S_n}{\sqrt{2n{\rm
loglog}n}} \leq-{\underline{\sigma}}\right)=1.$$
To prove (III). We only need to prove that for any number
$b\in(-\underline \sigma, \underline \sigma),$
\begin{equation}\label{ED}v\left(\liminf\limits_{n\rightarrow\infty}\left|\frac{S_n}{\sqrt{2n{\rm
loglog}n}}-b\right|=0 \right)=1.\end{equation} Noting the fact
that $|b|<\underline\sigma$, by Lemma 7, we can easily obtain
(16).
The proof of (III) is complete.
|
\section{Introduction}
It has been known for many years that maximal supergravity theories have hidden rigid symmetry groups that increase in dimension as the dimension of spacetime decreases \cite{Cremmer:1979up}. The $D=3$ case is special in the sense that the symmetry group $E_8$ is the largest finite one in the $E$ series; in $D=2$ one has $E_9$ \cite{Nicolai:1988jb}. More recently it has been suggested that these symmetries might be extended to $E_{10}$ \cite{Kleinschmidt:2004dy} or $E_{11}$ \cite{West:2001as}. In addition, in the $D=3$ theory \cite{Marcus:1983hb}, the $128+128$ on-shell degrees of freedom are entirely non-gravitational, so that in a sense the theory is really an $SO(16)\backslash E_8$ non-linear sigma model, although there is an induced geometrical structure.
In this paper we describe this model in a superspace with a local $SL(2,\mbox{\bbbold R})\times SO(16)$ structure group. We begin with a brief review of the geometry that describes the off-shell superconformal multiplet and then impose further constraints to accommodate the sigma model fields. Because the scalars transform non-linearly under $E_8$ and the fermions transform according to a spinor representation of $SO(16)$ it follows that the former can only appear covered by derivatives in the geometrical tensors, and that the latter cannot appear linearly. This circumstance simplifies the geometry somewhat, especially due to the fact that the dimension one-half torsion tensor must vanish which is not the case in higher dimensions (except $D=11$). This might lead one to believe that there could be generalised chiral (CR) structures involving more mutually anti-commuting odd derivatives than in higher-dimensional spacetimes, but this is not the case due to obstructions arising from the curvature.
In recent years there have been several studies of the systematics of form fields in supergravity theories, starting with \cite{Cremmer:1997ct,Cremmer:1998px}. It was realised that this could be formalised in terms of Borcherds algebras \cite{HenryLabordere:2002dk}, and also in terms of $E_{11}$ \cite{Julia:1997cy,Riccioni:2007au,Bergshoeff:2007qi,Riccioni:2009xr}. (See \cite{Henneaux:2010ys} for a discussion of the relation between the two). In a separate, but related development, it has been shown that the same sets of forms contribute to the hierarchies found in gauged supergravities, see \cite{deWit:2008gc} and references therein. A key feature is that these forms fall into representations of the duality groups. In addition to the physical forms and their duals there are also $(D-1)$-form potentials, related to gaugings, and $D$-form potentials, related to space-filling branes. We discuss these fields explicitly and show that all of the coupled Bianchi identities for the associated field strengths are satisfied. Our construction uses only supersymmetry and $E_8$ symmetry, although the allowed representations that the forms transform under agree with those predicted from $E_{11}$. The superspace method has some advantages, especially for the top forms (three-form potentials in $D=3$). This is because it makes sense to consider four- (and indeed higher)-form field strengths in superspace due to the fact that the odd basis differential forms are commutative. This point of view was advocated previously in the context of maximal supersymmetry in ten dimensions \cite{Bergshoeff:2007ma,Bergshoeff:2010mv}. An additional feature of the formalism is that it is manifestly supersymmetric, and indeed, if one concentrates on the field strengths, manifestly covariant under all symmetries.
Three dimensions is special for gauged maximal supergravity \cite{Nicolai:2000sc,Nicolai:2001sv,deWit:2008ta} because there are no independent vector degrees of freedom. Gauging is very natural in superspace and the constraints on the embedding tensor can be seen as a direct consequence of the gauged Maurer-Cartan equation. Essentially, the dimension-one constraints are modified by a set of functions that fit together in the $1+3875$ representations of $E_8$. The higher-rank forms, discussed in detail in \cite{deWit:2008ta}, can also be deformed, and we choose to do this in a covariant fashion, i.e. by deforming the Bianchi identities. Discussions of the $E_{11}$ approach to gauged supergravity in $D=3$ can be found in \cite{Bergshoeff:2007qi,Bergshoeff:2008qd}.
The organisation of the paper is as follows: in section 2 we describe the geometrical set-up and review the off-shell superconformal constraints for $N$-extended supergravity in $D=3$. In section 3 we introduce the $E_8$ sigma model in the context of this supergravity background and show how the latter can accommodate it by making appropriate identifications. In section 4 we analyse the geometrical Bianchi identities up to dimension two. This allows us to identify all components of the torsion and curvature tensors and to verify explicitly that the constraints are consistent. It is important to do this as the initial set of constraints include conventional ones for the $SO(16)$ connection, whereas the sigma model fixes this connection in terms of the physical fields, so that one needs to check that these choices are compatible. In section 5 we turn our attention to the additional form fields. These consist of the duals to the scalars (two-form field strengths) as well as three- and four-forms whose potentials have no physical degrees of freedom, and we also give a brief discussion of possible five-forms. The analysis of the associated Bianchi identities is considerably simplified by the use of superspace cohomology. In section 6 we discuss the gauging of the theory in a superspace setting. We start by modifying the geometry by gauging the Maurer-Cartan equation and then deform the Bianchi identities for the form fields. In the final section we make some concluding remarks. There are appendices on our conventions, superspace cohomology, some details of a calculation, and a final one in which we show, using harmonic superspace, that there are no higher-rank (Lorentzian) CR structures in this theory than in higher dimensional maximal supergravity.
\section{Geometrical set-up}
We consider a supermanifold $M$ with (even$|$odd)-dimension $(3|32)$. The basic structure is determined by a choice of odd tangent bundle $T_1$ such that the Frobenius tensor, which maps pairs of sections of $T_1$ to the even tangent bundle, $T_0$, generates the latter. We shall also suppose that there is a preferred basis $E_{\alpha}\def\adt{\dot \alpha i},\,\alpha}\def\adt{\dot \alpha=1,2;\,i=1,\ldots 16$ for $T_1$ such that the components of the Frobenius tensor, which we shall also refer to as the dimension-zero torsion, are
\begin{equation}}\def\ee{\end{equation}
T_{\alpha}\def\adt{\dot \alpha i\beta}\def\bdt{\dot \beta j}{}^c=-i\delta}\def\D{\Delta}\def\ddt{\dot\delta_{ij} (\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^c)_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta}\,;\qquad c=0,1,2\ .
\la{2.1}
\ee
At this stage $T_0$ is defined as the quotient, $T/T_1$, but we can make a definite choice for $T_0$ by imposing some suitable dimension one-half constraint. When this has been done, the structure group will be reduced to $SL(2,\mbox{\bbbold R})\times SO(16)$, with the Lorentz vector indices being acted on by the local $SO(1,2)$ associated with $SL(2,\mbox{\bbbold R})$.\footnote{The dimension-zero torsion \eq{2.1} is also invariant under local Weyl rescalings, but we shall not include this factor in the structure group. It is broken in the on-shell Poincar\'e theory.} With respect to this structure we have preferred basis vector fields $E_A=(E_a,E_{\underline{\alpha}})=(E_a,E_{\alpha}\def\adt{\dot \alpha i})$ with dual one-forms $E^A=(E^a,E^{\underline{\alpha}})=(E^a, E^{\alpha}\def\adt{\dot \alpha i})$, the latter being related to the coordinate basis forms $dz^M=(dx^m, d\theta}\def\Th{\Theta}\def\vth{\vartheta^{{\underline m}})$ by the supervielbein matrix $E_M{}^A$, i.e. $E^A=dz^M E_M{}^A$. Here, coordinate indices are taken from the middle of the alphabet, preferred basis indices from the beginning, while even (odd) indices are latin and greek respectively. Underlined odd indices run from 1 to 32, and $SO(16)$ vector indices are denoted $i,j$ etc.
We now introduce a set of connection one-forms, $\O_A{}^B$, for the above structure group. We have
\begin{eqnarray}
\O_a{}^{\underline{\phantom{\alpha}}\!\!\!\beta}&=& \O_{\underline{\alpha}}{}^b=0\nonumber\w1
\O_{\alpha}\def\adt{\dot \alpha i}{}^{\beta}\def\bdt{\dot \beta j}&=& \delta}\def\D{\Delta}\def\ddt{\dot\delta_i{}^j \O_\alpha}\def\adt{\dot \alpha{}^\beta}\def\bdt{\dot \beta + \delta}\def\D{\Delta}\def\ddt{\dot\delta_\alpha}\def\adt{\dot \alpha{}^\beta}\def\bdt{\dot \beta \O_i{}^j\nonumber\w1
\O_a{}^b&=& - (\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a{}^b)_\alpha}\def\adt{\dot \alpha{}^\beta}\def\bdt{\dot \beta \O_\beta}\def\bdt{\dot \beta{}^\alpha}\def\adt{\dot \alpha\ .
\la{2.2}
\end{eqnarray}
Spinor indices $\alpha}\def\adt{\dot \alpha,\beta}\def\bdt{\dot \beta$ are raised and lowered by the epsilon tensor, while Lorentz and $SO(16)$ vector indices are raised by the corresponding metrics $\eta_{ab}, \delta}\def\D{\Delta}\def\ddt{\dot\delta_{ij}$. We have $\O_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta}=\O_{\beta}\def\bdt{\dot \beta\alpha}\def\adt{\dot \alpha}$ while $\O_{ab}$ and $\O_{ij}$ are antisymmetric. The torsion and curvature are defined in the usual way
\begin{eqnarray}
T^A&=& DE^A:=d E^A + E^B \O_B{}^A\nonumber\w1
R_A{}^B&=& d\O_A{}^B + \O_A{}^C \O_C{}^B\ .
\la{2.3}
\end{eqnarray}
The Bianchi identities are
\begin{eqnarray}
DT^A&=& E^B R_B{}A\nonumber\w1
D R_A{}^B&=&0\ .
\la{2.4}
\end{eqnarray}
Equation \eq{2.1} does not simply determine the structure group, it is also a constraint. With an appropriate choice of dimension one-half connections and of $T_0$, and making use of the dimension one-half Bianchi identity, one finds that all components of the dimension one-half torsion may be set to zero:
\begin{equation}}\def\ee{\end{equation}
T_{\underline{\alpha}\underline{\phantom{\alpha}}\!\!\!\beta}{}^{\underline{\phantom{\alpha}}\!\!\!\gamma}=T_{a \underline{\phantom{\alpha}}\!\!\!\beta}{}^c=0\ .
\la{2.5}
\ee
Imposing further conventional constraints corresponding to the dimension-one connection components we find that the dimension-one torsion can be chosen to have the form
\begin{eqnarray}
T_{ab}{}^c&=&0\nonumber\w1
T_{a \beta}\def\bdt{\dot \beta j}{}^{\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k}&=& (\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a)_\beta}\def\bdt{\dot \beta{}^\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma K_j{}^k + (\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^b)_\beta}\def\bdt{\dot \beta{}^\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma L_{ab j}{}^k\ ,
\la{2.6}
\end{eqnarray}
where $K_{ij}$ is symmetric and $L_{abij}$ is antisymmetric on both pairs of indices. The dimension-one curvatures are
\begin{eqnarray}
R_{\alpha}\def\adt{\dot \alpha i\beta}\def\bdt{\dot \beta j, cd}&=& -2i(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_{cd})_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} K_{ij} -2i\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} L_{cdij}\nonumber\w1
R_{\alpha}\def\adt{\dot \alpha i\beta}\def\bdt{\dot \beta j,kl}&=& i\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta}(M_{ijkl} + 4 \delta}\def\D{\Delta}\def\ddt{\dot\delta_{[i[k} K_{j]l]})-i(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a)_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} (4\delta}\def\D{\Delta}\def\ddt{\dot\delta_{(i[k} L_{a j) l]}-\delta}\def\D{\Delta}\def\ddt{\dot\delta_{ij} L_{a kl})\ ,
\la{2.7}
\end{eqnarray}
where $L_{ab}=\ve_{abc} L^c$, and $M_{ijkl}$ is totally antisymmetric. This geometry is valid for any $N$, not just $N=16$, and describes an off-shell superconformal multiplet \cite{Howe:1995zm}. The interpretation of the dimension-one fields, $K,L,M$, is as follows. The geometry is determined by the basic constraint \eq{2.1} which is invariant under Weyl rescalings where the parameter is an unconstrained scalar superfield. This means that some of the fields that appear in the geometry do not belong to the Weyl supergravity multiplet. At dimension one $K$ and $L$ are of this type, so that we could set them to zero if we were only interested in the superconformal multiplet. The field $M_{ijkl}$, on the other hand, can be considered as the field strength superfield for the Weyl supermultiplet \cite{Howe:1995zm}.\footnote{This was discussed explicitly in for the case of $N=8$ in \cite{Howe:2004ib}.} The fact that $M$ is not expressible in terms of the torsion is due to a lacuna in Dragon's theorem \cite{Kuzenko:2011xg,Cederwall:2011pu} which in higher-dimensional spacetimes states that the curvature is so determined \cite{Dragon:1978nf}. We recall that in three-dimensional spacetime there is no Weyl tensor but that its place is taken by the dimension-three Cotton tensor. This turns out to be a component of the superfield $M_{ijkl}$ so that we could refer to the latter as the super Cotton tensor. Using the notation $[k,l]$ to denote fields that have $k$ antisymmetrised $SO(N)$ indices and $l$ symmetrised spinor indices, one can see that the component fields of the superconformal multiplet fall into two sequences starting from $M$. The first has fields of the type $[4-p,p]$, where the top ($[4,0]$) component is the supersymmetric Cotton tensor, while the second has fields of the type $(4+p,p)$ and therefore includes higher spin fields for $N>8$. There is also a second scalar $(4,0)$ at dimension two. Fields with two or more spinor indices obey covariant conservation conditions so that each field in the multiplet has two degrees of freedom multiplied by the dimension of the $SO(N)$ representation, provided that we count the dimension one and two scalars together. It is easy to see that the number of bosonic and fermionic degrees of freedom in this multiplet match.
The other components of the curvature and torsion can be derived straightforwardly from here, but we shall postpone this until we have introduced the physical fields. For the conformal case, the dimension three-halves Bianchi identities were solved explicitly in \cite{Kuzenko:2011xg}, while a detailed discussion of the $N=8$ case has been given in \cite{Cederwall:2011pu}.
\section{The sigma model}
As we have seen the structure group contains an $SO(16)$ factor which is associated with an $SO(16)$ principal bundle. The sigma model can be introduced via the requirement that this bundle can be lifted to a flat $E_8$ bundle. Our conventions for the Lie algebra, $\ge_8$, are as follows: the generators are $(M_{ij}, N_I)$ where $M_{ij}=-M_{ji}$ are the generators for $\gs\go(16)$ and the remaining generators, $N_I,\,I=1,\ldots 128$, transform under one of the two Weyl spinor representations of $\gs\gp\gi\gn(16)$. We shall denote the other representation by primed indices, e.g. $I'$. The algebra of $\ge_8$ is
\begin{eqnarray}
[M_{ij}, M^{kl}]&=& -4\delta}\def\D{\Delta}\def\ddt{\dot\delta_{[i}{}^{[k} M_{j]}{}^{l]}\nonumber\w1
[M_{ij},N_I]&=& -\frac{1}{2}}\def\qu{\frac{1}{4}(\S_{ij})_{IJ} N_J\nonumber\w1
[N_I, N_J]&=& (\S^{ij})_{IJ} M_{ij}\ ,
\la{3.1}
\end{eqnarray}
where the $SO(16)$ sigma matrices are denoted by $\S$ (see appendix for conventions).
The sigma model field can be viewed as a section ${\cal V}$ of the $E_8$ bundle. It is acted on to the right by $E_8$ and to the left by the local $SO(16)$ and therefore corresponds to an $SO(16)\backslash E_8$ sigma model superfield. The Maurer-Cartan form is
\begin{equation}}\def\ee{\end{equation}
\F:=d{\cal V} {\cal V}^{-1}:=P+ Q\ ,
\la{3.2}
\ee
where $Q=\frac{1}{2}}\def\qu{\frac{1}{4} \O^{ij} M_{ij}$ and where $P$ takes its values in the quotient algebra, i.e. $P=P^I N_I$. From the Maurer-Cartan equation (vanishing $E_8$ curvature), $d\F+\F^2=0$, we find
\begin{eqnarray}
DP&=&0\w1
R&=&-P^2\ ,
\la{3.3}
\end{eqnarray}
where $R:= \frac{1}{2}}\def\qu{\frac{1}{4} R^{ij} M_{ij}$ is the $SO(16)$ curvature, while $D$ is the $SO(16)$-covariant exterior derivative. In indices, the above equations are
\begin{eqnarray}
2 D_{[A} P_{B]} + T_{AB}{}^C P_C&=& 0\w1
\la{3.4}
R_{AB,ij}&=& 2P_A \S_{ij} P_B\ .
\la{3.5}
\end{eqnarray}
We shall need to impose a constraint on the dimension one-half component of $P$ to ensure that we have the correct number of degrees of freedom, namely 128 bosonic and fermionic. We therefore set
\begin{equation}}\def\ee{\end{equation}
P_{\alpha}\def\adt{\dot \alpha i I}=i(\S_i \L_\alpha}\def\adt{\dot \alpha)_I\ ,
\la{3.6}
\ee
where $\L_{\alpha}\def\adt{\dot \alpha I'}$ describes the physical 128 spin one-half fields. The dimension-one component of \eq{3.4} is then satisfied if
\begin{equation}}\def\ee{\end{equation}
D_{\alpha}\def\adt{\dot \alpha i} \L_{\beta}\def\bdt{\dot \beta I'}=\frac{1}{2}}\def\qu{\frac{1}{4}(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a)_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} (\S_i P_a)_{I'}\ .
\la{3.7}
\ee
We can think of $P_{a I}$ as essentially the spacetime derivative of the physical scalar fields. In order to see this more explicitly, it is perhaps useful to look at the linearised limit. In the physical gauge we can put ${\cal V}=\exp (\phi}\def\F{\Phi}\def\vf{\varphi^I N_I)$ where $\phi}\def\F{\Phi}\def\vf{\varphi^I$ denotes the 128 scalars. If we now keep only terms linear in the fields we find
\begin{eqnarray}
D_{\alpha}\def\adt{\dot \alpha i} \phi}\def\F{\Phi}\def\vf{\varphi_I&=& i (\S_i \L_\alpha}\def\adt{\dot \alpha)_I \nonumber\w1
D_{\alpha}\def\adt{\dot \alpha i}\L_{\beta}\def\bdt{\dot \beta J'}&=& \frac{1}{2}}\def\qu{\frac{1}{4}(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a)_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} (\S_i P_a)_{J'}=\frac{1}{2}}\def\qu{\frac{1}{4}(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a)_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} (\S_i \partial}\def\delb{\bar\partial_a \phi}\def\F{\Phi}\def\vf{\varphi)_{J'}\ ,
\la{3.7.1}
\end{eqnarray}
where $D_{\alpha}\def\adt{\dot \alpha i}$ is now the usual supercovariant derivative in flat superspace. It follows from \eq{3.7.1} that both $\phi}\def\F{\Phi}\def\vf{\varphi_I$ and $\L_{\alpha}\def\adt{\dot \alpha I'}$ satisfy free field equations of motion.
Note that we have now specified the $SO(16)$ connection in two ways, by choosing corresponding conventional constraints on the torsion, and explicitly in terms of ${\cal V}$. We therefore need to verify that these are compatible. We can easily see that they are by making use of the dimension-one component of \eq{3.5}. Comparing with \eq{2.6}, \eq{2.7} we find agreement provided that
\begin{eqnarray}
K_{ij}&=& -\frac{i}{2}\delta}\def\D{\Delta}\def\ddt{\dot\delta_{ij} B :=-\frac{i}{2}\delta}\def\D{\Delta}\def\ddt{\dot\delta_{ij} \L\L\nonumber\w1
L_{a ij}&=&\, iA_{a ij}:= i\L \gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a \S_{ij} \L\nonumber\w1
M_{ijkl}&=& -iB_{ijkl}:= -i\L\S_{ijkl}\L\ ,
\la{3.8}
\end{eqnarray}
where, on the right-hand-side, the spacetime and internal spinor indices are contracted in the natural way (see appendix). The non-zero dimension-one torsion therefore becomes
\begin{equation}}\def\ee{\end{equation}
T_{a\beta}\def\bdt{\dot \beta j \gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k}=-\frac{i}{2}(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a)_{\beta}\def\bdt{\dot \beta\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma}\delta}\def\D{\Delta}\def\ddt{\dot\delta_{jk} B -i (\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a{}^b)_{\beta}\def\bdt{\dot \beta\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma} A_{b jk}\ .
\la{3.9}
\ee
With this, we now have a solution to the coupled Maurer-Cartan equations and geometrical Bianchi identities up to dimension one expressed entirely in terms of the physical fields. In terms of the sigma model fields the dimension-one curvatures are
\begin{eqnarray}
R_{\alpha}\def\adt{\dot \alpha i\beta}\def\bdt{\dot \beta j, cd}&=& -\delta}\def\D{\Delta}\def\ddt{\dot\delta_{ij}(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_{cd})_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} B +2 \ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} A_{cdij}\ \Rightarrow\nonumber\w1
R_{\alpha}\def\adt{\dot \alpha i,\beta}\def\bdt{\dot \beta j,\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma\delta}\def\D{\Delta}\def\ddt{\dot\delta}&=&\frac{1}{2}}\def\qu{\frac{1}{4} \delta}\def\D{\Delta}\def\ddt{\dot\delta_{ij} (\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a)_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} (\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a)_{\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma\delta}\def\D{\Delta}\def\ddt{\dot\delta} B-
\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta}(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a)_{\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma\delta}\def\D{\Delta}\def\ddt{\dot\delta} A_{a ij}\ ,\nonumber\w1
R_{\alpha}\def\adt{\dot \alpha i\beta}\def\bdt{\dot \beta j,kl}&=& \ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta}(B_{ijkl} + 2 \delta}\def\D{\Delta}\def\ddt{\dot\delta_{[i[k} \delta}\def\D{\Delta}\def\ddt{\dot\delta_{j]l]}B)+(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a)_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} (4\delta}\def\D{\Delta}\def\ddt{\dot\delta_{(i[k} A_{a j) l]}-\delta}\def\D{\Delta}\def\ddt{\dot\delta_{ij} A_{a kl})\ .
\la{3.9.1}
\end{eqnarray}
Note that there is an interesting feature of this solution that does not occur in higher-dimensional maximal supergravity theories (except for $D=11$), namely the fact that the dimension one-half torsion tensor is zero. This is easily understood in terms of group representations because in $D=3$ the spinor field transforms as a spinor under the internal symmetry group whereas the geometrical tensors can only accommodate tensor representations. If we move up to $N=8$ supergravity in $D=4$, for example, the internal symmetry group is $SU(8)$ and the spin one-half fermions transform under the $56$-dimensional representation. They can therefore be accommodated in the dimension one-half torsion as follows \cite{Brink:1979nt}:
\begin{equation}}\def\ee{\end{equation}
T_{\alpha}\def\adt{\dot \alpha i,\beta}\def\bdt{\dot \beta j, \cdt k}=\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} \bar\L_{\cdt ijk} \ ,
\la{3.10}
\ee
where we have used two-component spinor notation, where $i,j,k=1,\ldots 8$ and where $\bar\L_{\cdt ijk}$ is totally antisymmetric on its internal indices. Its leading component in a $\theta}\def\Th{\Theta}\def\vth{\vartheta$-exansion is the physical spin one-half fields in the 56 of $SU(8)$.
\section{Torsion and curvature up to dimension two}
\subsection{Dimension three-halves}
In this section we shall check the various identities up to dimension two. This will enable us to confirm the consistency of the solution and also to compute the dimension three-halves and two components of the torsion and curvature. As expected, these turn out to be functions of the physical fields, there being no gravitational degrees of freedom in three dimensions.
There are two relevant Bianchi identities, as well as the dimension three-halves components of $DP=0$ and $R=-P^2$. They are
\begin{eqnarray}
2 R_{[a \underline{\alpha},b]c} &=& -T_{ab}{}^{\underline{\phantom{\alpha}}\!\!\!\beta} T_{\underline{\alpha}\underline{\phantom{\alpha}}\!\!\!\beta c} \la{4.1}\w1
2 R_{a (\underline{\alpha},\underline{\phantom{\alpha}}\!\!\!\beta) \underline{\phantom{\alpha}}\!\!\!\gamma}&=& -2D_{(\underline{\alpha}} T_{a \underline{\phantom{\alpha}}\!\!\!\beta) \underline{\phantom{\alpha}}\!\!\!\gamma} - T_{\underline{\alpha}\underline{\phantom{\alpha}}\!\!\!\beta}{}^b T_{ab \underline{\phantom{\alpha}}\!\!\!\gamma}\la{4.2}\w1
D_a P_{\underline{\phantom{\alpha}}\!\!\!\beta}-D_{\underline{\phantom{\alpha}}\!\!\!\beta} P_a + T_{a\underline{\phantom{\alpha}}\!\!\!\beta}{}^{\underline{\phantom{\alpha}}\!\!\!\gamma} P_{\underline{\phantom{\alpha}}\!\!\!\gamma}&=& 0\la{4.3}\w1
R_{a \beta}\def\bdt{\dot \beta j, kl}&=& + 2iP_a \S_{kl} \S_j \L_\alpha}\def\adt{\dot \alpha \ .
\la{4.4}
\end{eqnarray}
Equation \eq{4.1} allows us to solve for the dimension three-halves Lorentz curvature in terms of the dimension three-halves torsion. The $\theta}\def\Th{\Theta}\def\vth{\vartheta=0$ component of the latter can be identified as the gravitino field strength, so we shall give it a new notation $T_{ab}{}^{\underline{\phantom{\alpha}}\!\!\!\gamma}:=\Psi_{ab}{}^{\underline{\phantom{\alpha}}\!\!\!\gamma}=\ve_{abc}\Psi^{c\,\underline{\phantom{\alpha}}\!\!\!\gamma}$. From \eq{4.2} we find that
\begin{equation}}\def\ee{\end{equation}
\Psi_{a i}=2\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^b \gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a P_b \S_i \L\ ,
\la{4.5}
\ee
confirming that the gravitino field strength is completely determined by the matter fields, as promised. The dimension-three-halves Lorentz curvature is
\begin{equation}}\def\ee{\end{equation}
R_{a\beta}\def\bdt{\dot \beta j,b}=i(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a \Psi_{b i}-\frac{1}{2}}\def\qu{\frac{1}{4} \eta_{ab}\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^c \Psi_{c i})_{\beta}\def\bdt{\dot \beta}\ .
\la{4.5.1}
\ee
From \eq{4.3} we can determine the supersymmetry variation of $P_a$,
\begin{equation}}\def\ee{\end{equation}
D_{\alpha}\def\adt{\dot \alpha i} P_a =i(\S_i D_a \L_\alpha}\def\adt{\dot \alpha)+ i T_{a\alpha}\def\adt{\dot \alpha i}{}^{\beta}\def\bdt{\dot \beta j} (\S_j \L_\beta}\def\bdt{\dot \beta)\ ,
\la{4.6}
\ee
where the second term on the right-hand-side gives terms that are cubic in $\L$.
At this stage we see that we have determined the geometric tensors in terms of the physical fields, but there are several other equalities that arise. One of these is the equation of motion for $\L$, and the others turn out to be identically satisfied even though this is not at all obvious at first sight. The full details of this are relegated to appendix D; here we simply state the equation of motion:
\begin{equation}}\def\ee{\end{equation}
\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a D_a \L=-\frac{i}{2} B\L +\frac{i}{6} \S^{ij} A_{a ij}\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a\L.
\la{4.11}
\ee
\subsection{Dimension two}
There are two Bianchi identities at dimension two, the first of which simply tells us that the Riemann tensor $R_{ab,cd}$ has the usual symmetries in the absence of torsion. In three dimensions it can be written in the form
\begin{equation}}\def\ee{\end{equation}
R_{ab,cd}=\ve_{abe}\ve_{cdf} G^{ef}
\la{4.16}
\ee
where $G_{ab}:=R_{ab}-\frac{1}{2}}\def\qu{\frac{1}{4} \eta_{ab} R$ is the Einstein tensor. The second Bianchi identity is
\begin{equation}}\def\ee{\end{equation}
R_{ab,\underline{\phantom{\alpha}}\!\!\!\gamma\underline\delta}=2D_{[a} T_{b]\underline{\phantom{\alpha}}\!\!\!\gamma\underline\delta} + D_{\underline{\phantom{\alpha}}\!\!\!\gamma} T_{ab\underline\delta} + 2 T_{[a \underline{\phantom{\alpha}}\!\!\!\gamma}{}^{\underline\epsilon} T_{b]\underline\epsilon \underline\delta}\ .
\la{4.17}
\ee
In addition, we have the dimension-two component of the Maurer-Cartan equation which gives
\begin{equation}}\def\ee{\end{equation}
R_{ab ij} = 2 P_a \S_{ij} P_b\ .
\la{4.18}
\ee
It is a lengthy computation to analyse the content of these equations. Clearly, the $SO(16)$ curvature is immediately found from \eq{4.18}, while the Lorentz curvature is obtained from \eq{4.17}. But then there are a lot other components of \eq{4.17} which must be satisfied identically. It is indeed the case that this is so, but to prove it requires further Fierz rearrangement.
For the Lorentz curvature we find
\begin{equation}}\def\ee{\end{equation}
G_{ab}=-4(P_a P_b -\frac{1}{2}\eta_{ab} P^c P_c) + 4i(\L \gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_{(a} D_{b)}\L-\eta_{ab} (B^2-\frac{1}{3} A_{a ij} A^{a ij}\ )) ,
\la{4.19}
\ee
where the right-hand side is essentially the on-shell energy-momentum tensor for the sigma model.
Finally, the equation of motion for the scalars can be found by acting on the fermion equation of motion with a spinorial derivative. It is
\begin{equation}}\def\ee{\end{equation}
D^a P_a=\frac{i}{32} \S^{ij} P_a A^a_{ij}-\frac{i}{16. 6!}\S^{i_1\ldots i_6} P_a A^a_{i_1\ldots i_6}\ .
\la{4.20}
\ee
\section{Forms}
In this section we discuss the various form field strengths that can arise in this theory. We shall concentrate on these rather than the potentials as this approach is gauge-invariant. This is particularly advantageous for the four-forms we shall discuss because the superspace field strengths can be non-vanishing even though the purely even components are identically zero. We shall derive the full set of forms up to degree four using only supersymmetry and $E_8$. In addition to the physical one-forms $P$ we are allowed to introduce their dual two-forms which transform under the 248 of $E_8$. Beyond these, we can in principle have three- and four-forms in arbitrary $E_8$ representations as the bosonic potentials do not introduce any new independent degrees of freedom; the problem is therefore to determine which representations are allowed. We also briefly mention the possible five-forms. The analysis of this problem is facilitated by the use of superspace cohomology techniques which we review in appendix C. The Bianchi identities for an $n$-form field strength have the following schematic form:
\begin{equation}}\def\ee{\end{equation}
I_{n+1}:=(d F_n - \sum F_p F_q)=0\ ; \quad p+q=n+1\ .
\la{5.1}
\ee
For the two-forms we have
\begin{equation}}\def\ee{\end{equation}
d F_2^R=0\ ,
\la{5.2}
\ee
where $R,S,T=1,\ldots 248$ denote adjoint representation indices. We shall see shortly that this identity is indeed satisfied, and that the $(2,0)$ component of $F_2$ is given by the dual of $P_a$ together with a bilinear fermion contribution. The components of the forms in an $E_8$ basis generically involve the scalar field matrix ${\cal V}$ which we write, in the adjoint representation as,
\begin{equation}}\def\ee{\end{equation}
{\cal V}_{\bar R}{}^R=\left(V_{ij}{}^R, V_I{}^R\right)\ ,
\la{5.3}
\ee
where the barred index is to be acted on by $SO(16)$ and therefore splits into the appropriate representations determined by the branching rules. The dimension-zero component of $F_2^R$ is given by
\begin{equation}}\def\ee{\end{equation}
F_{\alpha}\def\adt{\dot \alpha i\beta}\def\bdt{\dot \beta j}^R=-2i\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} V_{ij}{}^R\ .
\la{5.4}
\ee
Now suppose we have a set of three-forms $F_3^{\cal X}$ transforming under some representation labelled by ${\cal X}$. The Bianchi identity has the form
\begin{equation}}\def\ee{\end{equation}
d F_3^{{\cal X}}=F_2^S F_2^R t_{RS}{}^{{\cal X}}\ ,
\la{5.5}
\ee
where $t_{RS}{}^{\cal X}$ is an $E_8$-invariant tensor. Because of \eq{5.2}, such a Bianchi is automatically consistent, in the sense that $d$ acting on both sides gives zero. The Bianchi identity is itself a four-form $I_4$ transforming in the representation ${\cal X}$. Now any superspace $n$-form can be decomposed into a sum of $(p,q)$-forms, $p=n-q$, where $p\,(q)$ denotes the number of even (odd) indices carried by the form. The lowest-dimensional component of the four-form $I_4$ , $I_{0,4}$, has dimension zero, and since $d I_4=0$, it must satisfy $t_0 I_{0,4}=0$ ($t_0$ is defined in appendix C). Suppose that this equation is satisfied, then we will have $t_0 I_{1,3}=0$ at dimension one-half. But the cohomology group $H_t^{p,q}=0$ for $p\geq 1$ (see appendix C), so we have $I_{1,3}=t_0 J_{2,1}$. The equation $J_{2,1}=0$ can obviously be satisfied by an appropriate choice of $F_{2,1}$ because it contains exactly the right number of components, and so there can be no obstruction at dimension one-half. Similar arguments show that there are no higher-dimensional obstructions so that we conclude that the Bianchi identity \eq{5.5} is satisfied provided that its dimension-zero component is.
The symmetric product of two $248$s is $1+ 3875 + 27000$. For the singlet the Bianchi identity is
\begin{equation}}\def\ee{\end{equation}
dF_3 = F_2^R F_2^S a_{RS} \ ,
\la{5.6}
\ee
where $a_{RS}$ is the $E_8$ metric (given in appendix B). The dimension-zero component is
\begin{equation}}\def\ee{\end{equation}
F_{a\beta}\def\bdt{\dot \beta j\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k}=-i\delta}\def\D{\Delta}\def\ddt{\dot\delta_{jk} (\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a)_{\beta}\def\bdt{\dot \beta\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma}\ .
\la{5.7}
\ee
It is not difficult to see that $I_{0,4}=0$, and so we conclude that there is a singlet three-form.
Next consider the $3875$. The branching rule is $3875\rightarrow 135+1820 +1920'$. The 135 is a symmetric traceless tensor which we shall write as $t_{i,j}$, the $1820$ is a fourth-rank antisymmetric tensor, and the $1920'$ is a sigma-traceless primed vector-spinor. The dimension-zero component of this three-form is
\begin{equation}}\def\ee{\end{equation}
F_{a\beta}\def\bdt{\dot \beta j\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k}^U=-i(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a)_{\beta}\def\bdt{\dot \beta\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma} V_{j,k}{}^U\ ,
\la{5.8}
\ee
where $U,V,W=1,\ldots 3875$ and where ${\cal V}_{\bar U}{}^U$ is the scalar field matrix in the 3875 representation, so that $V_{i,j}{}^U$ is the projection onto the 135 in $\bar U$. If we write the Bianchi identity as $I_4^U=dF_3^U-F^S_2 F_2^R b_{RS}{}^U=0$, we can see that its dimension-zero component is indeed satisfied. This is because the symmetrised product of two $120$s coming from $F_{0,2}^S F_{0,2}^R b_{RS}{}^U$ can give both 135 and 1820 (since these are both contained in 3875), but the 1820 drops out in the Bianchi identity because the symmetrisation over the four odd indices would require antisymmetrisation over the four two-component Lorentz spinor indices. We can thus conclude without any further calculation that this Bianchi identity is satisfied.
The final possibility for three-forms is the 27000. The branching rule is $27000\rightarrow 1+1820+6435+5304+128+13312$, the last two being spinorial representations. The most significant one for us is the 5304; this is a tensor with the symmetries of the Weyl tensor (in sixteen dimensions). The only possibility for the dimension-zero component is
\begin{equation}}\def\ee{\end{equation}
F^{{\cal X}}_{a\beta}\def\bdt{\dot \beta j\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k}=-i\delta}\def\D{\Delta}\def\ddt{\dot\delta_{jk} (\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a)_{\beta}\def\bdt{\dot \beta\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma} V_0{}^{\cal X}\ ,
\la{5.9}
\ee
where $V_0{}^{\cal X}$ denotes the singlet projection of the scalar matrix in the 27000. However, the dimension-zero $(F_2)^2$ term now has a contribution in the 5304 that cannot be balanced in the Bianchi identity, and so we conclude that the 27000 is not allowed.
Next, we consider the four-forms. They obey Bianchi identities of the form
\begin{equation}}\def\ee{\end{equation}
d F_4^{\cal X}= F_3^U F_2^R t_{RU}{}^{\cal X}\ ,
\la{5.10}
\ee
where $t_{RU}{}^{\cal X}$ is an invariant tensor in the indicated representations. The possible representations are therefore contained in the tensor product of 248 and 3875 which is 779247+147250+30380+ 3875+248. In order for the Bianchi itself to be consistent we must have
\begin{equation}}\def\ee{\end{equation}
b_{(RS}{}^U t_{T)U}{}^{\cal X}=0
\la{5.11}
\ee
This will be true if the symmetrised triple product of 248 does not contain the representation ${\cal X}$. Of the possible representations, only the 3875 and the 147250 have this property and so we can discard the others. The Bianchi identities for the four-forms are five-forms, $I_5$, so that the lowest possible non-trivial components are $I_{1,4}$ (dimension zero). We must have $t_0 I_{1,4}=0$ which implies that $I_{1,4}=t_0 J_{2,2}$. Setting $J_{2,2}=0$ simply allows us to solve for the dimension-zero components of the four-forms, and the argument can be repeated at dimension one-half. There are no dimension-one components as we are in three-dimensional spacetime. We therefore conclude that there are no obstructions to solving any consistent (i.e. closed) Bianchi identities for four-forms.
There is a possible singlet four-form $F_4$ but it has to be gauge-trivial, i.e. $d F_4=0$. However, this is trivial in the sense that one can write $F_4=d G_3$, where the only non-zero component of $G$ is $G_{abc}\propto \ve_{abc}$.
Finally, we comment on the possible five-forms that can arise in the theory. The corresponding potentials for these do not have purely even components, but there are three-form gauge parameters that can have non-zero $(3,0)$-components. Indeed, as has been pointed out \cite{deWit:2008ta}, these can play a r\^{o}le in the gauged theory. The five-form Bianchi identities have the schematic form
\begin{equation}}\def\ee{\end{equation}
dF_5^{{\cal X}}= (F_4 F_2)^{\cal X} + (F_3 F_3)^{\cal X}\ .
\la{5.11.1}
\ee
The possible representations are therefore contained in the products $248 \times 3875=248+3875+30380+147250+779247$, $248\times 147250=3875+30380+147250+779247+2450240+6696000+2641100$ and the antisymmetric product of two $3875$s which gives $248+30380+779247+6696000$. The only component of a five-form that can be non-zero in supergravity is $F_{3,2}$ and this must be proportional to $\ve_{abc}\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta}$ multiplied by a function of the form $F_{ij}^{\cal X}$ that is antisymmetric on $ij$. It therefore follows that the only representations that can be non-zero must contain $120$ in the branching down to $SO(16)$. This leaves only four possibilities:$\,248, 30380,779247$ and $6696000$. It is straightforward to verify, using group theory, that all four of these, which have contributions from at least two terms on the right-hand side, have consistent Bianchi identities and that they all occur with multiplicity one, i.e. there are no free parameters. Given that the Bianchi identities are consistent, it follows from cohomology that there will be no obstructions to solving them.
To see this in more detail, consider an arbitrary five-form with the Bianchi identity \eq{5.11.1}. Applying another $d$ to this equation we find that both terms give rise to terms of the form
\begin{equation}}\def\ee{\end{equation}
F_3^U F_2^R F_2^S t_{U,RS,}{}^{\cal X}\ ,
\la{5.11.2}
\ee
where $t$ is an invariant tensor, and the sum of these terms must vanish for consistency. The pair $RS$ is symmetric and can therefore be in any of the representations $1+3875+27000$. So to check whether a given term can be non-zero we have only to check if this representation multiplied by the $3875$ contains the representation ${\cal X}$. Clearly the singlet cannot occur, but the other two can, at least in principle. As an example consider ${\cal X}=R$, the 248. There are two terms in the Bianchi identity and the one coming from two three-forms can only give rise to $RS$ in the $3875$. So if the other term, coming from $F_4(3875)\times F_2$ were to allow the 27000 the Bianchi identity would not be consistent. But $3875\times 27000$ does not contain the 248, so the two $3875$ contributions can cancel and we are left with precisely one consistent Bianchi identity. The argument can easily be repeated for the other three representations that can tolerate non-zero five-forms in supergravity and one finds that they are indeed all consistent with multiplicity one.
The same argument can be extended to the representations corresponding to the five-form field strengths that must be zero in supergravity. They are $3875, 147250, 2450240$ and $26411008$. The last two are definitely not consistent and it is unlikely that other two are either, although this has not been checked in detail.
In summary, the dual two-forms are in the adjoint representation of $E_8$, the allowed three-forms transform under the singlet and 3875 representations and the allowed four-forms transform under the 3875 and 147250 representations together with a trivial singlet. The Bianchi identities are
\begin{eqnarray}
dF_2^R&=&0 \la{5.12}\w1
dF_3&=& F_2^S F_2^R a_{RS}\la{5.13}\w1
dF_3^U&=&F_2^R F_2^S b_{RS}{}^U\la{5.14} \w1
dF_4&=&0 \la{5.15}\w1
dF_4^U&=& F_3^V F_2^R c_{RV}{}^U\la{5.16}\w1
dF_4^X&=& F_3^V F_2^R d_{RV}{}^X\ ,
\la{5.17}
\end{eqnarray}
where $a,b,c,d$ are $E_8$-invariant tensors, and where $X,Y,Z=1,\dots 147250$.
The components of $F_2^R$ are
\begin{eqnarray}
F_{\alpha}\def\adt{\dot \alpha i\beta}\def\bdt{\dot \beta j}^R&=& -2i\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} V_{ij}{}^R\nonumber\w1
F_{a\beta}\def\bdt{\dot \beta j}^R&=& -i(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a\S_j\L)_{\beta}\def\bdt{\dot \beta }{}^I V_I{}^R\nonumber\w1
F_{ab}^R&=& \ve_{ab}{}^c (P_c{}^I V_I{}^R -2i A_c{}^{ij} V_{ij}{}^R)
\la{5.18}
\end{eqnarray}
The components of the singlet $F_3$ are
\begin{eqnarray}
F_{a\beta}\def\bdt{\dot \beta j\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k}&=&-i\delta}\def\D{\Delta}\def\ddt{\dot\delta_{jk}(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a)_{\beta}\def\bdt{\dot \beta\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma}\nonumber\w1
F_{ab\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k}&=&0\nonumber\w1
F_{abc}&=& 4i\ve_{abc} B\ .
\la{5.19}
\end{eqnarray}
The components of $F_3^U$ are
\begin{eqnarray}
F_{a\beta}\def\bdt{\dot \beta j\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k}^U&=&-i(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a)_{\beta}\def\bdt{\dot \beta\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma} V_{j,k}{}^U\nonumber\w1
F_{ab\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k}^U&=& 2iV_{kI'} {}^U(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_{ab}\L)_{\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma I'} \nonumber\w1
F_{abc}^U&=& 2i\ve_{abc} B^{ijkl} V_{ijkl}{}^U\ .
\la{5.20}
\end{eqnarray}
\black
The components of $F_4^U$ are
\begin{eqnarray}
F_{ab \gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k\delta}\def\D{\Delta}\def\ddt{\dot\delta l}^U&=&-i(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_{ab})_{\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma\delta}\def\D{\Delta}\def\ddt{\dot\delta} V_{k,l}{}^U\nonumber\w1
F_{abc \delta}\def\D{\Delta}\def\ddt{\dot\delta l}^U&=& a\ve_{abc} V_{l I'} {}^U \L_{\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma I'} \ ,
\la{5.21}
\end{eqnarray}
while the components of $F_4^X$ are
\begin{eqnarray}
F_{ab \gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k\delta}\def\D{\Delta}\def\ddt{\dot\delta l}^X&=&-i(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_{ab})_{\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma\delta}\def\D{\Delta}\def\ddt{\dot\delta} V_{k,l}{}^X\nonumber\w1
F_{abc \delta}\def\D{\Delta}\def\ddt{\dot\delta l}^X&=& b \ve_{abc} V_{l I'} {}^X \L_{\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma I'} \ ,
\la{5.22}
\end{eqnarray}
where $a,b$ are real, calculable constants.
It is easy to see that the singlet four-form $F_4$ is exact as the only non-zero component is
\begin{equation}}\def\ee{\end{equation}
F_{ab\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma k\delta}\def\D{\Delta}\def\ddt{\dot\delta l}=-i\delta}\def\D{\Delta}\def\ddt{\dot\delta_{kl} (\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_{ab})_{\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma\delta}\def\D{\Delta}\def\ddt{\dot\delta}\ .
\la{5.23}
\ee
Clearly $F_4=d G_3$, where the only non-vanishing component of $G_3$ is $G_{abc}=\ve_{abc}$.
In addition there can be non-zero five-forms in the representations 248,30380,779247 and 6696000, obeying Bianchi identities of the form \eq{5.11.1}. The five-forms can only be non-vanishing at dimension zero where they have expressions of the form
\begin{equation}}\def\ee{\end{equation}
F_{abc\alpha}\def\adt{\dot \alpha i\beta}\def\bdt{\dot \beta j}^{\cal X}=i c\ve_{abc}\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} V_{ij}{}^{\cal X}\ ,
\la{5.24}
\ee
where ${\cal X}$ can be one of the above representations, $ij$ denotes the 120 of $SO(16)$ and $c$ is some real constant.
The forms can equally well be discussed in an $SO(16)$ basis. We shall distinguish this basis by barring quantities or indices. The Bianchi identities can be written
\begin{equation}}\def\ee{\end{equation}
D \bar F_n =- \bar F_n\wedge P + \bar F_{n-1} \wedge \bar F_2
\la{5.25}
\ee
where $\bar F= F {\cal V}$, $P$ is considered as being Lie-algebra-valued in the appropriate representation (with barred indices) and where the last term is understood as involving the appropriate invariant tensor. For each $F$ this equation can be split into various representations of $SO(16)$ according to the branching rules. The components of the $F$s in this basis can be read off from those of the $E_8$ basis straightforwardly. A key point is that they do not contain any explicit scalars; in particular, the dimension-zero components are just given by $SO(16)$-invariant tensors.
\section{Gauging}
\subsection{Geometry}
The gauging of maximal $D=3$ supergravity has been discussed in \cite{Nicolai:2000sc,Nicolai:2001sv}, and the differential forms were subsequently discussed in \cite{deWit:2008ta}. The key tool is the embedding tensor, ${\cal E}_R{}^S$.\footnote{The embedding tensor is usually called $\Th$ but we have chosen a different notation to avoid confusion with the superspace coordinates.} The embedding tensor allows one to present the results in a way which looks $E_8$ covariant but which is actually only covariant with respect to the local $SO(16)$ and the gauge group $G_0\subset E_8$ that we shall not need to specify explicitly (see \cite{Nicolai:2001sv} for a list of the possible gauge groups). The embedding tensor is essentially given by a sum of projectors onto the irreducible subspaces of $\ge_8$ corresponding to the simple factors of $\gg_0$ \cite{Nicolai:2001sv}. It can be taken to be symmetric, ${\cal E}_{RS}:={\cal E}_R{}^T a_{TS}={\cal E}_{SR}$ and there is also a quadratic constraint on ${\cal E}$ that follows from demanding that it be invariant under gauge transformations. It is
\begin{equation}}\def\ee{\end{equation}
{\cal E}_R{}^P {\cal E}_{(M}{}^Q f_{N)PQ}=0\ ,
\la{6.0}
\ee
where $f_{PQR}$ denotes the $\ge_8$ structure constants. The discussion is best approached via the gauged Maurer-Cartan form \cite{de Wit:1982ig} (see \cite{Howe:1981tp} for the superspace version) which can be written
\begin{equation}}\def\ee{\end{equation}
\F={\cal D}{\cal V} {\cal V}^{-1}=P+ Q\ ,
\la{6.1}
\ee
where ${\cal D}$ is a gauge-covariant derivative that acts on the $E_8$ index carried by ${\cal V}_{\bar R}{}^R$, i.e. the superscript. The gauged Maurer-Cartan equation, which follows directly from \eq{6.1}, is
\begin{equation}}\def\ee{\end{equation}
R+ DP+ P^2 = g{\cal F}:=g{\cal V} {\cal F}{\cal V}^{-1}\ .
\la{6.2}
\ee
Here, $g$ is a constant with dimensions of mass which characterises the deformation and $D$ is covariant with respect to both $SO(16)$ and $G_0$. The theory has both of these groups as local symmetries, but the rigid $E_8$ is broken. The technique we shall use in the following analysis is to work with $SO(16)$ indices, so that the gauge group is hidden from view.
The original geometrical constraint in superspace \eq{2.1}, i.e. taking the dimension-zero torsion to be the same as in flat space, together with the allowed conventional constraints, leads to the dimension-one torsion and curvatures given in equations \eq{2.6} and \eq{2.7}. Since the deformation parameter $g$ has dimension one it follows that we can expect changes to the tensors $K_{ij}, L_{a ij}$ and $M_{ijkl}$. These can only be proportional to $g$ multiplied by functions of the scalars and so $L_{a ij}$ must be unchanged. This leaves $K$ and $M$ which together fall into the $1+135+1820$ representations of $SO(16)$. Anticipating a little we can see that these can be combined into the $E_8$ representations $1+3875$ if we also have the $1920'$. This representation can be found in the scalar part of $D_{\alpha}\def\adt{\dot \alpha i}\L_{\beta}\def\bdt{\dot \beta I'}$, which vanishes in the non-gauged case but which will be modified when the gauging is turned on as one can see from \eq{6.2}.
To implement the gauging explicitly we first need to solve for the two-form field strength. This should be projected along $\gg_0$ which leads us to propose that it should have the form
\begin{equation}}\def\ee{\end{equation}
{\cal F}^R=F^S{\cal E}_S{}^R\ .
\la{6.2.1}
\ee
It is easy to see, using the fact that ${\cal D}{\cal E}_R{}^S=0$, that the Bianchi identity for ${\cal F}^R$ will be solved if we take the components of $F^R$ to have the same form as in the ungauged case. In fact, the only $g$-dependence could be at dimension one, but since this component of $F$ is a spacetime two-form this cannot occur. Lowering the index on ${\cal F}$ and converting to an $SO(16)$ basis we find, at dimension zero,
\begin{eqnarray}
{\cal F}_{\alpha}\def\adt{\dot \alpha i\beta}\def\bdt{\dot \beta j,kl}&=&-2i\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} V_{ij}{}^R V_{kl}^S\, {\cal E}_{RS}\w1
\la{6.2.2}
{\cal F}_{\alpha}\def\adt{\dot \alpha i\beta}\def\bdt{\dot \beta j,I}&=&-2i\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} V_I{}^R V_{kl}^S\, {\cal E}_{RS}\ ,
\la{6.3}
\end{eqnarray}
and at dimension one-half,
\begin{equation}}\def\ee{\end{equation}
{\cal F}_{a\beta}\def\bdt{\dot \beta j,I}=-i(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a \S_j\L)_\beta}\def\bdt{\dot \beta{}^J V_J{}^R V_{I}^S\, {\cal E}_{RS}\ .
\ee
Since ${\cal E}_{RS}$ is symmetric it can contain the $1+3875+27000$ representations of $E_8$, but we can see directly from the 120 component of \eq{6.2} that the $27000$ must be absent due to the fact that it cannot be accommodated in the dimension-one curvature. This is the basic constraint on ${\cal E}$ derived in \cite{Nicolai:2000sc,Nicolai:2001sv}. It then follows that the only representation in \eq{6.3} will be the 1920'. The functions appearing in the dimension-zero and one-half ${\cal F}_2$'s can therefore be written
\begin{eqnarray}
V_{ij}{}^R V_{kl}^S\, {\cal E}_{RS}&=&f_{ij,kl}:=\delta}\def\D{\Delta}\def\ddt{\dot\delta_{i[k}\delta}\def\D{\Delta}\def\ddt{\dot\delta_{l]j} f_0 + \delta}\def\D{\Delta}\def\ddt{\dot\delta_{[i[k} f_{j],l]} + f_{ijkl}\nonumber\w1
V_I{}^R V_{kl}^S {\cal E}_{RS}\,&=& (\S_{[k})_{IJ'} f_{l] J'}\nonumber\w1
V_{I}{}^R V_{J}^S\, {\cal E}_{RS}&=&\frac{1}{4!}(\S_{ijkl})_{IJ}f^{ijkl}-2\delta}\def\D{\Delta}\def\ddt{\dot\delta_{IJ}f_0\ ,
\la{6.4}
\end{eqnarray}
where the functions $f_0,f_{i,j},f_{ijkl}$ and $f_i$ exhibit the $1+135+1820+1920'$ split explicitly. The deformation feeds into the geometrical tensors at dimension one via the gauged Maurer-Cartan equation from which we find
\begin{eqnarray}
K_{ij}&=& -\frac{i}{2}\delta}\def\D{\Delta}\def\ddt{\dot\delta_{ij}B +2g( f_{i,j}+\delta}\def\D{\Delta}\def\ddt{\dot\delta_{ij} f_0)\nonumber\w1
M_{ijkl}&=& -i B_{ijkl} +8g f_{ijkl}\nonumber\w1
D_{\alpha}\def\adt{\dot \alpha i}\L_{\beta}\def\bdt{\dot \beta I'}&=&(\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a)_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} (\S_i P_a)_{I'} +g\ve_{\alpha}\def\adt{\dot \alpha\beta}\def\bdt{\dot \beta} f_{i I'}\ .
\la{6.5}
\end{eqnarray}
It is easy to check that the geometrical Bianchi identities at dimension three-halves are satisfied. To do this one needs the following easily derivable identities
\begin{eqnarray}
D_{\alpha}\def\adt{\dot \alpha i}f_{0}&=&0\nonumber\w1
D_{\alpha}\def\adt{\dot \alpha i}f_{j,k}&=& 2i\L_{\alpha}\def\adt{\dot \alpha}\S_i\S_{(j}f_{k)}\nonumber\w1
D_{\alpha}\def\adt{\dot \alpha i}f_{jklm}&=& -i\L_{\alpha}\def\adt{\dot \alpha}\S_i\S_{[jkl}f_{m]}\nonumber\w1
D_{\alpha}\def\adt{\dot \alpha i}f_{j}&=& i\L_{\alpha}\def\adt{\dot \alpha}\S_i\S^k f_{j,k} +
\frac{i}{48}\L_{\alpha}\def\adt{\dot \alpha}\S_i( \S_{j}{}^{klmn} f_{klmn} +12 \S^{klm}f_{jklm})\ .
\la{6.5.0}
\end{eqnarray}
There is a modification to the gravitino field strength given by
\begin{equation}}\def\ee{\end{equation}
\Psi_{a i}(g)=4g f_{i} \gamma}\def\C{\Gamma}\def\cdt{\dot\gamma_a \L\ ,
\la{6.5.1}
\ee
as well as a $g$-dependent term in the fermion equation of motion,
\begin{equation}}\def\ee{\end{equation}
\gamma}\def\C{\Gamma}\def\cdt{\dot\gamma^a D_a \L(g)=4g(f_0 \L -\frac{1}{48}f_{ijkl} \S^{ijkl}\L)\ .
\la{6.5.2}
\ee
Finally, at dimension two, there are changes to the curvature scalar and the scalar equation of motion given by
\begin{eqnarray}
R(g)&=&-\frac{2ig}{3}(48f_0 B - f_{ijkl} B^{ijkl})\nonumber\\
&&-3g^2(f_if^i-2f_{i,j}f^{i,j}-32f_0f_0)\nonumber\w1
D^a P_a(g)&=& \frac{5ig}{4.4!}\S_{ijk} f_l B^{ijkl}\nonumber \w1
&&-\frac{3g^2}{4}(\S_if_jf^{i,j}-\frac{1}{9}\S_{jkl}f_i f^{ijkl})\, .
\la{6.5.3}
\end{eqnarray}
\subsection{Forms}
We now consider the hierarchy of forms. In the geometrical discussion above we have only used the true non-abelian gauge fields, but in order to accommodate all of the forms it will be necessary to include the other two-form gauge fields which we could think of as being abelian, although they do transform under the gauge group. In other words we have a set of 248 gauge fields $F^R$, where ${\cal F}^R=F^S{\cal E}_S{}^R$. The Bianchis for the forms can then be written, in the $E_8$ basis,
\begin{equation}}\def\ee{\end{equation}
{\cal D} F_n =(F F)_{n+1}+ g F_{n+1}Y_{n+1,n}\ ,
\la{6.6}
\ee
where $(FF)$ denotes the bilinear term of the same form as in the ungauged case and $Y_{n+1,n}$ denotes a mapping from the representation space ${\cal R}_{n+1}$ of the $(n+1)$-forms to that of the $n$-forms, ${\cal R}_n$. In order to determine the $Y$-matrices one must compute the effect of applying ${\cal D}$ to \eq{6.6}; clearly one will require, in agreement with the general discussion in \cite{deWit:2008gc}, that
\begin{equation}}\def\ee{\end{equation}
Y_{n+1,n}\ Y_{n,n-1}=0
\la{6.7}
\ee
in order for the $g^2$ terms to cancel. The presence of the $g$-dependent term on the right-hand side of a deformed Bianchi identity implies that the $(n-1)$-form potential will transform under the $(n-1)$-form gauge transformation of the $n$-form potential, which is the way the hierarchy has been derived previously \cite{deWit:2008ta}. Before discussing this system in more detail we note that the $F$s themselves are hardly changed from the abelian case. Since $g$ has dimension one, it is only the purely even components of the $F$s that can get deformed and these only by the $f$-functions of the previous section. So only the $(3,0)$ components of the three-forms can receive corrections, which for the 3875 take the form
\begin{equation}}\def\ee{\end{equation}
F_{abc}^U=\ve_{abc} (a' f_0 V_0{}^U + b' f^{i,j} V_{i,j}{}^U + c' f^{ijkl} V_{ijkl}{}^U)\ ,
\la{6.8}
\ee
where $a',b'.c'$ are constants.
We now give an example of the hierarchy computation in this covariant language. The two-form Bianchi identity is ${\cal D} F_2=gF_3 Y_{3,2}$. Applying a second ${\cal D}$ to this we get
\begin{equation}}\def\ee{\end{equation}
gF_2^T F_2^S X_{ST}{}^R=gF_2^S F_2^T b_{ST}{}^U Y_U{}^R + g^2 F_4^{{\cal X}} Y_{{\cal X}}{}^U Y_U{}^R\ ,
\la{6.9}
\ee
where $X_{ST}{}^R={\cal E}_S{}^P f_{PT}{}^R$ is the generator of the gauge group within the 248 representation \cite{deWit:2008gc}.
The second term on the right must vanish in order to satisfy \eq{6.7}, and we can easily satisfy the part of the equation linear in $g$ by taking
\begin{equation}}\def\ee{\end{equation}
Y_{ST}{}^R={\cal E}_{(S{}}{}^P f_{T)P}{}^R
\la{6.10}
\ee
in agreement with \cite{deWit:2008ta}. Here, we have replaced $U$ by a symmetrised pair of $248$ indices on $Y$, and \eq{6.10} is correct as it stands because the singlet and the 27000 vanish on the right. One can continue in this way for the higher forms for which we find $Y_{4,3}$s that agree with those of \cite{deWit:2008ta}, although we have not checked beyond this. The complete hierarchy for the gauged forms in supergravity requires the five-forms and their Bianchi identities, and the Bianchi identities for the six-forms, even though the latter vanish in supergravity. This is because the seven-form right-hand side of the six-form Bianchi identities involve terms of the form $(F_5,F_2)$ and $(F_4,F_3)$ and these expressions can in principle be non-zero at dimension zero. However, provided that the identities are themselves consistent, these equations will automatically be satisfied for cohomological reasons.
The fact that the Bianchi identities are consistent suggest that the field strengths should be expressible in terms of potentials, and this is indeed the case. For the two- and three-forms we find
\begin{eqnarray}
F_2^R&=& d A_1^R + \frac{g}{2} A_1^T A_1^S X_{ST}{}^R + A_2^U Y_U{}^R\la{6.11}\w1
F_3^{RS}&=& {\cal D} A_2^{RS} + g A_1^R( dA_1^S + \frac{1}{3} A_I^Q A_1^P X_{PQ}{}^S) + g A_4^{\cal X} Y_{\cal X}{}^{RS}\ ,
\la{6.12}
\end{eqnarray}
where, in the second equation, the indices $RS$ are symmetrised and projected onto the $3875$, while ${\cal X}$ denotes the 3875 and the 147250 representations.
In the above we have ignored the singlet three-form, but its Bianchi identity can also be deformed:
\begin{equation}}\def\ee{\end{equation}
dF_3=F_2^R F_2^S a_{RS} + g F_4^U Y_U\ ,
\la{6.13}
\ee
where $Y_U$ is proportional to the 3875 component of the embedding matrix. It is not difficult to verify the consistency of this Bianchi identity.
\section{Conclusions}
In this paper we have presented the maximal supergravity theory in three dimensions in a superspace setting. Starting from the off-shell superconformal constraints we have shown how one can accommodate the supergravity sigma model by introducing the scalar fields in the coset $SO(16)\backslash E_8$. The sigma model fields enter the geometry via the dimension-one components of the torsion and curvature tensors, that is the functions $K_{ij}, L_{a ij}$ and $M_{ijkl}$. The spin one-half fields $\L$ cannot appear in the dimension one-half torsion owing to the fact that they transform under a spinor representation rather than a tensor one, although this does not imply that there are Lorentz-covariant CR structures of a higher rank than there are in more spacetime dimensions (appendix E). We solved for all of the components of the torsion and curvature tensors and derived the equations of motion.
The theory can be gauged by introducing a non-abelian gauged subgroup $G_0$ of $E_8$ and making use of the gauged Maurer-Cartan form. There are terms in the gauge-deformed Maurer-Cartan equation involving the two-form gauge field strength that are proportional to the parameter $g$ and that modify the dimension-one scalar functions in the theory. We computed these, the changes induced in all the components of the geometrical tensors and the modifications to the equations of motion. The dimension-one functions fall into the representations $1+135+1820+1920'$ of $SO(16)$, the representations that appear in the $1+3875$ representations of $E_8$. This gives a nice derivation of the fact that the embedding tensor, which is in the symmetric product of two adjoint representations of $E_8$, cannot contain the 27000 representation. This picture extends to higher dimensions. For example, in $D=4$ there are dimension-one scalars that appear in the torsion in the $36+420$ representations of $SU(8)$ \cite{Howe:1981gz} that can be combined into the $912$ representation of $E_7$, as one would expect \cite{deWit:1983gs}.
The ungauged theory admits a set of differential forms that transform in various representations of $E_8$, including one-, two- and three-form potentials. We examined which representations can appear using supersymmetry and consistency of the Bianchi identities. The analysis is made easier in superspace owing to the fact that one can study the problem covariantly using the field strengths even in the case of the three-form potentials, because a four-form makes sense in superspace even when there are only three even dimensions. In addition, explicitly checking that the Bianchi identities are satisfied is facilitated by the use of superspace cohomology; only a few components of the Bianchi identities need to be checked due to the fact that $H_t^{p,q}=0$ for $p\neq 0$ in $D=3,N=16$ superspace.
The differential forms can also be studied in the gauged case by deforming the Bianchi identities. The field-strengths transform under the gauge group so that the exterior derivative must be made gauge-covariant. The Bianchi identity for $F_n$ develops a term $gF_{n+1}Y_{n+1,n}$, where $Y_{n+1,n}$ maps the representation space of $(n+1)$-forms to that of $n$-forms and depends linearly on the embedding tensor. This means that all of the forms become related by a sequence of such maps that must be exact in order for the Bianchi identities to be consistent. The closure of the full system of forms requires the introduction of five-forms in the supergravity limit, and it was shown that there are such objects in the ungauged case. These are unmodified in the gauged case, and indeed the only forms that can be are the three-forms at dimension zero.
\pagebreak
{\bf\Large{Acknowledgements}}
We thank Eric Bergshoeff, Fabio Riccioni, Duncan Steele and Elias Lousseief for helpful comments.
JG thanks Tekn. Dr Marcus Wallenbergs Stiftelse and the STFC for financial support.
\vskip .5cm
|
\section{Introduction}
Thermodynamic properties of fluids at very low temperatures are of
significant interest. For example, the current International
Temperature Scale~\cite{its90} makes use of volumetric properties and
vapor pressures of helium isotopes below the triple point of neon
(24.5561~K); below the triple point of hydrogen (13.8033~K), the scale
is based entirely on properties of ${}^3$He and ${}^4$He. The
theoretical analysis of relevant properties at these conditions, such
as the virial coefficients that describe the fluid's departure from
ideal-gas behavior, is complicated by the presence of quantum effects.
The inclusion of quantum effects in the calculation of virial
coefficients was one of the first numerical applications of the
Path-Integral Monte Carlo (PIMC) method.~\cite{FH} In a series of
pioneering works published in the 1960's, Fosdick and Jordan
showed how to calculate the second and third virial coefficient of a
monatomic gas using computer
simulations.~\cite{fosdick-jordan66,jordan-fosdick68,fosdick68} Given
the limited computational resources available at that time, they were
able to calculate the third virial coefficient only
in the case of two-body interactions, using a model potential of the
Lennard-Jones form and assuming distinguishable particles (Boltzmann
statistics). They argued that their method could be extended to
include the proper quantum statistics, but they were able to compute
exchange effects only in the case of the second virial coefficient.
Recently, the exponential increase in computational power
has enabled use of the path-integral method to
calculate the properties of quantum degenerate systems, notably
superfluid helium.~\cite{Ceperley}
At the same time, progress in the computation of {\em ab initio}
electronic properties of interacting atoms resulted in the availability of
very precise two- and three-body interparticle potentials, at least for
the lightest particles such as helium
atoms~\cite{Jez07,HBV07,He-3body,CPS09,He2-2010} or hydrogen
molecules.~\cite{diep-johnson00,virial-H2}
A natural application for these potentials is the calculation of
virial coefficients. As is well known, the second virial coefficient
depends only on the two-body potential, the third virial coefficient
depends only on two-body and three-body interactions, etc. The second
virial coefficient for a monatomic gas can be rigorously obtained at
the fully quantum level from the calculation of the phase shifts due
to the pair potential, and previous work has shown that a completely
{\em ab initio} calculation of second virial coefficients for helium
can have uncertainties comparable to and in many cases smaller than
those of the most precise
experiments.~\cite{phi07,hur00,BHV07,Mehl09,Cencek11}
In the case of the third virial coefficient, no closed-form solution
of the quantum statistical mechanics problem is known. First-order
semiclassical approaches have been derived~\cite{yokota60,ram-singh73}
and show that, in the case of helium, quantum diffraction effects
result in significant modifications of the classical result, even at
room temperature. However, there is no rigorous way to evaluate the
accuracy or uncertainty of the semiclassical result, especially at low
temperatures.
In recent work,~\cite{Garberoglio2009b} we extended the methodology
pioneered by Fosdick and Jordan, deriving a set of formulae allowing a
path-integral calculation of the third virial coefficient $C(T)$ of
monatomic species for arbitrary two- and three-body potentials. Our
results were limited to Boltzmann statistics (i.e., distinguishable
particles) and we did not present results for temperatures lower than
the triple point of neon ($24.5561$~K), which we deemed to be a
reasonable lower bound so that exchange effects could be neglected.
Nevertheless, we were able to compute the value of the third virial
coefficient of ${}^4$He with an uncertainty one order of magnitude
smaller than that of the best experiments.
Recent experimental results overlapping with our temperature
range,~\cite{Gaiser09,Gaiser10} although mostly consistent with our
calculations, seemed to indicate a systematic deviation which the
authors speculated could originate from our neglect of the proper
quantum statistics of helium atoms.
In this paper, we extend our computational methodology to calculate
the quantum statistical contributions to the third virial coefficient,
and compute $C(T)$ for both isotopes of helium in the temperature
range $2.6 - 24.5561$~K, extending the temperature range considered in
our previous work down into the range where exchange effects are
important.
We show that quantum statistical effects are significant only for
temperatures smaller than about $7$~K, and compare our results to
low-temperature experimental data.
In a subsequent publication~\cite{paper2}, we will present results
covering the entire temperature range (improving on our previous
results for ${}^4$He at 24.5561~K and above) with rigorously derived
uncertainties. We will also extend our methodology to include
acoustic virial coefficients, and compare those calculations to
available data. In the present work, our focus is on low temperatures
and specifically on the effect of non-Boltzmann statistics.
\section{Path-integral calculation of the virial coefficients}
The second and third virial coefficients, $B(T)$ and $C(T)$ respectively, are
given by~\cite{hcb}
\begin{eqnarray}
B(T) &=& -\frac{1}{2 V} \left( Z_2 - Z_1^2 \right) \label{eq:B} \\
C(T) &=& 4 B^2(T) - \frac{1}{3 V} \left[ Z_3 - 3 Z_2 Z_1 + 2 Z_1^3
\right], \label{eq:C}
\end{eqnarray}
where $V$ is the integration volume (with the limit $V
\rightarrow \infty$ taken at the end of the calculations), and the functions
$Z_N$ are given by:
\begin{eqnarray}
Z_3 &=& \Lambda^9 \int \mathrm d1 \mathrm d2 \mathrm d3 ~
\langle 1 2 3 | \mathrm e^{-\beta \hat H_3} \sum_{\pi_3} {\cal P}_{\pi_3} | 1 2 3
\rangle \label{eq:Z3} \\
Z_2 &=& \Lambda^6 \int \mathrm d1 \mathrm d2 ~
\langle 1 2 | \mathrm e^{-\beta \hat H_2 } \sum_{\pi_2} {\cal P}_{\pi_2} | 1 2 \rangle
\label{eq:Z2} \\
Z_1 &=& \Lambda^3 \int \mathrm d1 ~
\langle 1 | \mathrm e^{-\beta \hat H_1} | 1 \rangle = V, \label{eq:Z1}
\end{eqnarray}
where $\hat H_N$ is the $N$-body Hamiltonian, $\beta = 1/(k_\mathrm B
T)$, ${\cal P}$ is a permutation operator (multiplied by the sign of
the permutation in the case of Fermi--Dirac statistics), the index
$\pi_3$ runs over the 6 permutations of 3 objects (i.e., $123$, $132$,
$213$, $321$, $231$ and $312$), and $\pi_2$ runs over the 2
permutations of 2 objects (i.e., $12$ and $21$). $\Lambda = h /
\sqrt{2 \pi m k_\mathrm B T}$ is the thermal de~Broglie wavelength of
a particle of mass $m$ at temperature $T$. For the sake of
conciseness, we denote by $| i \rangle$ an eigenvector of the position
operator relative to particle $i$ and by $\mathrm d i$ ($i=1,2,3$) the
integration volume relative to the Cartesian coordinates of the $i$-th
particle. Note that, in order to produce the molar units used by
experimenters and in our subsequent comparisons with data, the right
side of Eq.~(\ref{eq:B}) and the second term in the right side of
Eq.~(\ref{eq:C}) must be multiplied by Avogadro's number and its
square, respectively.
In the following, we will derive a path-integral expression for the
calculation of the virial coefficients with Eqs.~(\ref{eq:B}) and
(\ref{eq:C}). We perform the derivation in detail in the case of
$B(T)$ to establish the notation, and then extend the results to the
more interesting case of $C(T)$.
\subsection{Second virial coefficient}
In this paper, we adopt Cartesian coordinates to describe the atomic
positions. This differs from the approach developed in
Refs.~\onlinecite{jordan-fosdick68} and \onlinecite{Garberoglio2009b},
where Jacobi coordinates were used. This choice allows the exchange
contribution to be computed in a much simpler manner than would be the
case if Jacobi coordinates were used, especially in the case of three
or more particles.
From Eqs.~(\ref{eq:B}) and (\ref{eq:Z2}), it can be seen that there
are two contributions to $B(T)$. The first one comes from considering
the identity permutation only, and takes into account only quantum
diffraction effects. This is the only contribution that gives a
nonzero result at high temperatures, where the particles can be treated
as distinguishable (Boltzmann statistics).
The second contribution to $B(T)$, which we will call {\it exchange}
(xc), comes from the only other permutation involved in the definition
of the quantity $Z_2$ above.
The expression of these two contributions in Cartesian coordinates is:
\begin{widetext}
\begin{eqnarray}
B_\mathrm{Boltzmann}(T) &=& -\frac{\Lambda^6}{2 V}
\int \mathrm d \boldsymbol{r}_1 \mathrm d \boldsymbol{r}_2 ~
\langle \boldsymbol{r}_1 \boldsymbol{r}_2 |
\exp\left[ -\beta (\hat K_2 + \hat U_2(|\boldsymbol{r}_2 - \boldsymbol{r}_1|)
) \right] - \exp\left[ -\beta \hat K_2 \right]
| \boldsymbol{r}_1 \boldsymbol{r}_2 \rangle \label{eq:Bdir} \\
B_\mathrm{xc}(T) &=& \mp \frac{\Lambda^6}{2 V}
\int \mathrm d \boldsymbol{r}_1 \mathrm d \boldsymbol{r}_2 ~
\langle \boldsymbol{r}_1 \boldsymbol{r}_2 |
\exp\left[ -\beta (\hat K_2 + \hat U_2(|\boldsymbol{r}_2 - \boldsymbol{r}_1|)
) \right]| \boldsymbol{r}_2 \boldsymbol{r}_1 \rangle, \label{eq:Bx}
\end{eqnarray}
\end{widetext}
where we denote by $K_N$ the total kinetic energy of $N$ bodies and by $\hat
U_2(r)$ the two-body potential energy operator.
The upper (lower) sign in Eq.~(\ref{eq:Bx}) corresponds to
Bose--Einstein (Fermi--Dirac) statistics.
Equations~(\ref{eq:Bdir}) and (\ref{eq:Bx}) can be rewritten by using the
Trotter identity
\begin{equation}
\mathrm e^{\hat K_2 + \hat U_2} = \lim_{P \rightarrow \infty} \left( \mathrm e^{\hat
K_2/P} \mathrm e^{\hat U_2/P} \right)^P
\label{eq:Trotter}
\end{equation}
with a positive integer value of the Trotter index $P$.
Following the procedure outlined in Ref.~\onlinecite{Garberoglio2009b}, one can
then write $B_\mathrm{Boltzmann}(T)$ as
\begin{eqnarray}
B_\mathrm{Boltzmann}(T) = - 2 \pi^2 \int_0^\infty r^2 \mathrm dr ~
\left(\exp\left[-\beta \overline{U}_2(r)\right] - 1 \right),
\label{eq:BBoltz}
\end{eqnarray}
where the two-body effective potential $\overline U_2(r)$ is given by
\begin{widetext}
\begin{eqnarray}
\exp\left[-\beta \overline{U}_2(r)\right] &=&
\int \prod_{i=1}^{P-1} \mathrm d\Delta\boldsymbol{x}_1^{(i)}
\mathrm d\Delta\boldsymbol{x}_2^{(i)} ~
\exp\left[
-\frac{\beta}{P} \sum_{i=1}^{P} U_2(|\boldsymbol{r} + \boldsymbol{x}_2^{(i)} -
\boldsymbol{x}_1^{(i)}|) \right] \times \nonumber \\
& &
F_\mathrm{ring}(\Delta\boldsymbol{x}_1^{(1)},\ldots,\Delta\boldsymbol{x}_1^{(P)})
F_\mathrm{ring}(\Delta\boldsymbol{x}_2^{(1)},\ldots,\Delta\boldsymbol{x}_2^{(P)})
\label{eq:U2_eff}
\\
&\underset{P \rightarrow \infty}{=}&
\oint {\cal D} \boldsymbol x_1 {\cal D} \boldsymbol x_2 ~
\exp\left[ -\frac{1}{\hbar} \int_0^{\beta \hbar}
\frac{m}{2} \left| \frac{\mathrm d \boldsymbol x_1(\tau)}{\mathrm d \tau}
\right|^2 +
\frac{m}{2} \left| \frac{\mathrm d \boldsymbol x_2(\tau)}{\mathrm d \tau} \right|^2 ~
+ U_2(|\boldsymbol r + \boldsymbol x_1(\tau) - \boldsymbol x_2(\tau)|)
\mathrm d \tau \right],
\label{eq:U2_pi}
\end{eqnarray}
\end{widetext}
where
\begin{equation}
F_\mathrm{ring}
=
\Lambda^3 \left( \frac{P^{3/2}}{\Lambda^3} \right)^P
\exp\left[ -\frac{\pi P}{\Lambda^2} \sum_{i=1}^P
\left| \Delta\boldsymbol{x}_1^{(i)} \right|^2 \right].
\label{eq:Fring}
\end{equation}
In the previous equations, we have defined
$\Delta\boldsymbol{x}_k^{(i)} = \boldsymbol{r}_k^{(i+1)} -
\boldsymbol{r}_k^{(i)}$, where $\boldsymbol{r}_k^{(i)}$ is the
coordinate of particle $k$ ($k=1,2$) in the $i$-th ``imaginary time
slice''. These ``slices'' are obtained by inserting $P$ completeness
relations of the form
\begin{equation}
1 = \int \mathrm d\boldsymbol{r}_1^{(i)} \mathrm d\boldsymbol{r}_2^{(i)} ~
| \boldsymbol{r}_1^{(i)} \boldsymbol{r}_2^{(i)} \rangle \langle
\boldsymbol{r}_1^{(i)} \boldsymbol{r}_2^{(i)} |
\end{equation}
between the factors $\mathrm e^{\hat K_2/P}$ and $\mathrm e^{\hat U_2/P}$ of
the Trotter expansion of Eq.~(\ref{eq:Trotter}). We used the overall
translation invariance of the system to remove the factor $V$ in
Eq.~(\ref{eq:Bdir}) and fix the $\tau=0$ slice of particle 2 at the
origin of the coordinate system. We also denoted by
$\boldsymbol{x}_1^{(i)}$ and $\boldsymbol{x}_2^{(i)}$ the coordinates
of two ring polymers having one of their endpoints fixed at the origin
($\boldsymbol{x}_1^{(1)} = \boldsymbol{x}_2^{(1)} = \mathbf 0$), and
we introduced the variable $\boldsymbol{r}$ denoting the distance
between the $\tau = 0$ time slice of the two ring polymers. In the
classical limit, where the paths $\boldsymbol x_1(\tau)$ and $\boldsymbol
x_2(\tau)$ shrink to a point, the coordinate $r$ reduces to the distance
between the particles and one has $\overline U_2(r) = U_2(r)$.
Note that the effect of the identity permutation is to set
$\boldsymbol{r}_k^{(P+1)} = \boldsymbol{r}_k^{(1)}$. The path-integral
formalism allows one to map the quantum statistical properties of a
system with $N$ distinguishable particles (Boltzmann statistics) onto
the classical statistical properties of a system of $N$ ring polymers,
each having $P$ beads (sometimes called imaginary-time slices), which
are distributed according to the function $F_\mathrm{ring}$ of
Eq.~(\ref{eq:Fring}).~\cite{boltzmann-bias} The mapping is exact in
the $P \rightarrow \infty$ limit, although convergence is usually
reached with a finite (albeit large) value of $P$.
In the calculation of the second virial coefficient,
Eq.~(\ref{eq:BBoltz}) shows that the second virial coefficient at the
level of Boltzmann statistics is obtained from an expression similar
to that for the classical second virial coefficient, using an
effective two-body potential. This effective potential,
$\overline{U}_2(r)$, is obtained by averaging the intermolecular
potential $U_2(r)$ over the coordinates of two ring polymers,
corresponding to the two interacting particles entering the definition
of $B(T)$.
Equation~(\ref{eq:BBoltz}) is equivalent to Eq.~(19) of
Ref.~\onlinecite{Garberoglio2009b}. The only difference is that the
current approach uses Cartesian coordinates, and therefore we are left
with an average over two ring polymers of mass $m$ instead of one ring
polymer of mass $\mu = m/2$, corresponding to the relative coordinate
of the two-particle system. The two approaches are of course
equivalent, and in fact it can be shown that Eqs.~(\ref{eq:BBoltz})
and (\ref{eq:U2_eff}) reduce to the form derived in
Refs.~\onlinecite{fosdick-jordan66} and \onlinecite{Garberoglio2009b}.
Equation~(\ref{eq:BBoltz}) is the same expression previously
derived by Diep and Johnson for spherically symmetric potentials on
the basis of heuristic arguments,~\cite{diep-johnson00} and later
generalized by Schenter to the case of rigid bodies and applied to a
model for water.~\cite{Schenter02}
Equation~(\ref{eq:U2_eff}) is actually the discretized version of a
path integral, as shown in Eq.~(\ref{eq:U2_pi}).
The circled integral is defined as
\begin{eqnarray}
\oint {\cal D} \boldsymbol x ~ \exp\left[ -\frac{1}{\hbar} \int_0^{\beta \hbar}
\frac{m}{2}
\left| \frac{\mathrm d \boldsymbol x(\tau)}{\mathrm d \tau} \right|^2 ~
\mathrm d \tau \right] &\equiv& \nonumber \\
\lim_{P \rightarrow \infty}
\int \prod_{i=1}^{P-1} \mathrm d\Delta\boldsymbol{x}^{(i)}
F_\mathrm{ring}(\Delta\boldsymbol{x}^{(1)},\ldots,\Delta\boldsymbol{x}^{(P)})
&=& 1,
\label{eq:pi}
\end{eqnarray}
and it indicates that one has to consider all the cyclic paths with
ending points at the origin, that is $\boldsymbol x(0) = \boldsymbol
x(\beta \hbar) = \boldsymbol 0$. The normalization of the path
integral is also indicated in Eq.~(\ref{eq:pi}).
We can perform on Eq.~(\ref{eq:Bx}), describing the exchange
contribution to the second virial coefficient, the same steps leading
from Eq.~(\ref{eq:Bdir}) to Eq.~(\ref{eq:BBoltz}). The only difference
is the presence of the permutation operator, whose main consequence is
fact that $\boldsymbol r^{(P+1)}_1 = \boldsymbol r^{(1)}_2$ and
$\boldsymbol r^{(P+1)}_2 = \boldsymbol r^{(1)}_1$. In this case,
defining $\boldsymbol X^{(i)} = \boldsymbol r^{(i)}_1$ and
$\boldsymbol X^{(i+P)} = \boldsymbol r^{(i)}_2$, one obtains
\begin{widetext}
\begin{eqnarray}
B_\mathrm{xc}(T) &=& \mp \frac{\Lambda^6}{2 V} \int
\mathrm d \boldsymbol X^{(1)} \ldots \mathrm d \boldsymbol X^{(2P)} ~
\exp\left[-\frac{\beta}{P}
\sum_{i=1}^P U_2(|\boldsymbol X^{(P+i)} - \boldsymbol X^{(i)}|) \right]
\times \nonumber \\ & &
\left(
\frac{P^{3/2}}{\Lambda^3} \right)^{2P} \exp\left[-\frac{\pi
P}{\Lambda^2} \sum_{i=1}^{2P} \left( \boldsymbol X^{(i+1)} - \boldsymbol
X^{(i)}\right)^2 \right]
\label{eq:Bxc} \\
&=&
\mp \frac{\Lambda^3}{2 V} \int \mathrm d \boldsymbol X^{(1)} \ldots \mathrm d \boldsymbol X^{(2P)} ~
\exp\left[-\frac{\beta}{P}
\sum_{i=1}^P U_2(|\boldsymbol X^{(P+i)} - \boldsymbol X^{(i)}|) \right]
\times \nonumber \\ & &
\frac{\Lambda_\mu^3}{2^{3/2}} \left( \frac{(2P)^{3/2}}{\Lambda_\mu^3}
\right)^{2P}
\exp\left[-\frac{\pi ~ 2 P}{\Lambda_\mu^2} \sum_{i=1}^{2P}
\left( \boldsymbol X^{(i+1)} - \boldsymbol X^{(i)}\right)^2 \right] \\
&=&
\mp \frac{1}{2} \frac{\Lambda^3}{2^{3/2}} \left\langle
\exp\left[-\frac{\beta}{P}
\sum_{i=1}^P U_2(|\boldsymbol X^{(P+i)} - \boldsymbol X^{(i)}|) \right]
\right\rangle
\label{eq:Bexchange} \\
&\underset{P \rightarrow \infty}{=}&
\mp \frac{\Lambda^3}{2^{5/2}} ~ \oint {\cal D}
\boldsymbol X ~ \exp\left[
-\frac{1}{\hbar} \int_0^{\beta \hbar} \frac{\mu}{2} \left|
\frac{\mathrm d \boldsymbol X(\tau)}{\mathrm d \tau} \right|^2 +
U_2(|\boldsymbol X(\tau + \beta \hbar /2) - \boldsymbol X(\tau)|)
~ \mathrm d\tau
\right],
\label{eq:Bexch_pi}
\end{eqnarray}
\end{widetext}
where we have defined $\Lambda_\mu =
\sqrt{2}\Lambda$. The exchange contribution to the second virial
coefficient is given simply as an average of the two-body potential
taken on ring polymers corresponding to particles of mass $\mu = m/2$.
In the discretized version of the path integral, one has to consider
$2P$ beads.
In Eq.~(\ref{eq:Bexchange}), we have used the overall translation
invariance of the integral to remove the factor of $V$ in the
denominator.
The effect of the various permutations can be visualized as generating paths
with a larger number of beads, which are obtained by coalescing the
ring polymers corresponding to the particles that are exchanged by the
permutation operator.
\subsection{Third virial coefficient}
We now discuss the third virial coefficient, starting from the
expression given in Eq.~(\ref{eq:C}). Since $4 B^2(T)$ can be
calculated by the methods of the previous section, we concentrate on
the second term, whose summands can be written as follows:
\begin{eqnarray}
Z_3 &=& \Lambda^9 \int \mathrm d1 \mathrm d2 \mathrm d3 ~
\langle 1 2 3 | \mathrm e^{-\beta \hat H_3} \sum_{\pi_3} {\cal P}_{\pi_3} | 1 2 3
\rangle \\
Z_2 Z_1 &=& \Lambda^9 \int \mathrm d1 \mathrm d2 \mathrm d3 ~
\langle 1 2 3 | \mathrm e^{-\beta (\hat H_2 + \hat T_3)} \sum_{\pi_2} {\cal P}_{\pi_2}
| 1 2 3 \rangle \label{eq:Z2Z1} \\
Z_1^3 &=& \Lambda^9 \int \mathrm d1 \mathrm d2 \mathrm d3 ~
\langle 1 2 3 | \mathrm e^{-\beta \hat K_3} | 1 2 3 \rangle,
\label{eq:Z13}
\end{eqnarray}
where $\hat T_i = -\frac{\hbar^2}{2m} \nabla_i^2$ is the kinetic energy
operator of particle $i$.
We can simplify the expression in square brackets on the right-hand side of
Eq.~(\ref{eq:C}) by writing the three $Z_2 Z_1$ terms choosing each time a
different particle for $Z_1$ (in Eq.~(\ref{eq:Z2Z1}) we have chosen particle 3
as coming from $Z_1$). After considering all the permutations of two and
three particles, we end up with $6 + 3 \times 2 + 1 = 13$ terms
building the term in square brackets of Eq.~(\ref{eq:C}). It is useful to
collect these 13 terms as follows:
\begin{enumerate}
\item{{\em Term 1 (identity term)}: we sum together permutation $123$ from $Z_3$, the
identity permutations from the three $Z_2 Z_1$ and the whole $2 Z_1^3$ term.
Adding $4 B^2_{\rm Boltzmann}(T)$, one obtains the
Boltzmann expression for $C(T)$, already discussed in
Ref.~\onlinecite{Garberoglio2009b}. In the present formulation based on Cartesian
coordinates, the value $C(T)$ in the case of Boltzmann statistics involves
an average over three independent ring polymers, which correspond to the
three particles. In the following, this contribution to $C(T)$ will be
referred to as $C_{\rm Boltzmann}(T)$ and is made by $1 + 3 + 1 = 5$ of the
13 terms described above.}
\item{{\em Term 2 (odd term)}: we take permutations $132$, $213$ and $321$
from $Z_3$ and the three exchange permutations from the $Z_2 Z_1$
terms. These permutations are all odd, and we consider them with a positive
sign (Bose--Einstein statistics). In the case of Fermi--Dirac statistics,
this term has to be multiplied by an overall minus sign. All of these
permutations correspond to configurations where two of the three particles
are exchanged. The sum of these 6 terms will be referred to as $C_{\rm
odd}(T)$.}
\item{{\em Term 3 (even term)}: we take the permutations $231$ and $312$ from
$Z_3$. These are the remaining two terms from the 13, and
are both even permutations, hence the name. Both of these terms
correspond to a cyclic exchange of the three particles, and their
sum will be referred to as $C_{\rm even}(T)$.}
\end{enumerate}
Using these definitions, the full $C(T)$, including quantum statistical
effects, can be written as
\begin{equation}
C(T) = C_{\rm Boltzmann}(T) \pm C_{\rm odd}(T) + C_{\rm even}(T) +
C_{\rm B}(T),
\label{eq:C_full}
\end{equation}
where the last term in the right-hand sum is given by
\begin{equation}
C_{\rm B}(T) = \pm 8 B_{\rm Boltzmann}(T) B_\mathrm{xc}(T) + 4
B^2_\mathrm{xc}(T),
\label{eq:CB}
\end{equation}
since the contribution of $4 B^2_\mathrm{Boltzmann}(T)$ to $C(T)$ is
already included in $C_{\rm Boltzmann}(T)$. In Eqs.~(\ref{eq:C_full})
and (\ref{eq:CB}), the upper (lower) sign corresponds to Bose--Einstein
(Fermi--Dirac) statistics.
Using the same procedure outlined above in the case of $B(T)$, one can
write the Boltzmann contribution to the third virial coefficient as
\begin{widetext}
\begin{eqnarray}
C_\mathrm{Boltzmann}(T) &=& 4 B^2_\mathrm{Boltzmann}(T) - \nonumber
\\
& & \frac{1}{3}
\int \mathrm d\boldsymbol{r}_1 \mathrm d\boldsymbol{r}_2
\left[
\mathrm e^{-\beta \overline{V}_3(\boldsymbol{r}_1,\boldsymbol{r}_2)} -
\mathrm e^{-\beta \overline{U}_2(|\boldsymbol{r}_1|)} -
\mathrm e^{-\beta \overline{U}_2(|\boldsymbol{r}_2|)} -
\mathrm e^{-\beta \overline{U}_2(|\boldsymbol{r}_1 - \boldsymbol{r}_2|)}
+ 2 \right],
\label{eq:CBoltz} \\
\exp\left[-\beta \overline{V}_3(\boldsymbol{r}_1,\boldsymbol{r}_2)\right]
&=&
\int \prod_{i=1}^{P-1}
\Delta \boldsymbol x_1^{(i)}
\Delta \boldsymbol x_2^{(i)}
\Delta \boldsymbol x_3^{(i)} ~
F_\mathrm{ring}^{(1)}
F_\mathrm{ring}^{(2)}
F_\mathrm{ring}^{(3)} ~
\exp\left[-\beta
\overline{V}_3^{\mathrm B}(\boldsymbol{r}_1,\boldsymbol{r}_2)
\right]
\label{eq:V3disc}
\\
&\underset{P \rightarrow \infty}{=}&
\oint {\cal D} \boldsymbol x_1
{\cal D} \boldsymbol x_2
{\cal D} \boldsymbol x_3 ~
\exp\left[-\frac{1}{\hbar} \int_0^{\beta \hbar}
\frac{m}{2} \left(
\left| \frac{\mathrm d \boldsymbol x_1(\tau)}{\mathrm d \tau} \right|^2 +
\left| \frac{\mathrm d \boldsymbol x_2(\tau)}{\mathrm d \tau} \right|^2 +
\left| \frac{\mathrm d \boldsymbol x_3(\tau)}{\mathrm d \tau} \right|^2
\right) + \right.
\nonumber \\
& & \left.
V_3(\boldsymbol r_1 + \boldsymbol x_1(\tau),
\boldsymbol r_2 + \boldsymbol x_2(\tau), \boldsymbol x_3(\tau))
~ \mathrm d \tau
\right], \label{eq:V3}
\end{eqnarray}
\end{widetext}
where $F_\mathrm{ring}^{(k)}$ denotes the probability distribution of
the path relative to particle $k$, as defined in Eq.~(\ref{eq:Fring}).
In Eq.~(\ref{eq:V3disc}), the three-body effective potential energy
$\overline{V}_3$ is obtained as an average performed over {\em three}
independent ring polymers of the total three-body potential energy:
\begin{eqnarray}
V_3(\boldsymbol{x}, \boldsymbol{y}, \boldsymbol{z}) &=&
U_3(\boldsymbol{x}, \boldsymbol{y}, \boldsymbol{z}) +
U_2(|\boldsymbol x - \boldsymbol y|) +
\nonumber \\
& &
U_2(|\boldsymbol x - \boldsymbol z|) +
U_2(|\boldsymbol y - \boldsymbol z|),
\end{eqnarray}
where $U_3(\boldsymbol{x}, \boldsymbol{y}, \boldsymbol{z})$ is the
non-additive three-body potential of three atoms.
In Eq.~(\ref{eq:V3disc}) the total three-body
potential energy for the Boltzmann contribution to the third virial
coefficient is
\begin{eqnarray}
\overline{V}_3^{\mathrm B}(\boldsymbol{r}_1,\boldsymbol{r}_2) &=&
\frac{1}{P} \sum_{i=1}^P U_3(\boldsymbol{r}_1 + \boldsymbol{x}_1^{(i)},
\boldsymbol{r}_2 + \boldsymbol{x}_2^{(i)},
\boldsymbol{x}_3^{(i)}) + \nonumber \nonumber \\ & &
U_2(|\boldsymbol{r}_1 + \boldsymbol{x}_1^{(i)} -
\boldsymbol{r}_2 - \boldsymbol{x}_2^{(i)}|) +
\nonumber \nonumber \\ & &
U_2(|\boldsymbol{r}_1 + \boldsymbol{x}_1^{(i)} -
\boldsymbol{x}_3^{(i)}|) +\nonumber \nonumber \\ & &
U_2(|\boldsymbol{r}_2 + \boldsymbol{x}_2^{(i)} - \boldsymbol{x}_3^{(i)}|),
\end{eqnarray}
where the variables with superscript $(i)$ denote the coordinates of
three ring polymers with one of the beads at the origin. Notice that
in passing from Eq.~(\ref{eq:C}) to Eq.~(\ref{eq:CBoltz}) we have used
the translation invariance of the integrand to perform the integration
over $\boldsymbol r_3$, which removed the factor of $V$ in the
denominator. As a consequence, the paths corresponding to particle 3
have their endpoints at the origin of the coordinate system (or,
equivalently, the third particle is fixed at the origin when the
classical limit is performed.) In the same limit, the variables
$\boldsymbol r_1$ and $\boldsymbol r_2$ appearing in Eq.~(\ref{eq:V3})
reduce to the positions of particles 1 and 2, respectively, and one has
$\overline V_3(\boldsymbol r_1,\boldsymbol r_2) = V_3(\boldsymbol
r_1,\boldsymbol r_2)$.
The term $C_\mathrm{odd}(T)$ is obtained by exchanging the positions
of two particles. This operation reduces the number of ring polymers
to two: one having $2P$ beads, corresponding to the exchanged
particles, and the other having $P$ beads, corresponding to the
remaining one. The odd contribution is given by
\begin{widetext}
\begin{eqnarray}
C_\mathrm{odd}(T) &=& - \frac{\Lambda^9}V
\int \mathrm d1 \mathrm d2 \mathrm d3 ~
\langle 1 2 3 |
\exp\left[-\beta \hat H_3 \right] - \exp\left[-\beta (\hat
K_2 + \hat U_2(\boldsymbol{r}_2 - \boldsymbol{r}_1)) \right]
| 2 1 3 \rangle \label{eq:Codd1} \\
&=&
- \frac{\Lambda^3}{2^{3/2}} \int \mathrm d \boldsymbol r_3 ~ \left\langle
\exp\left[ -\beta \overline{V}^\mathrm{odd}_3 \right] - \exp\left[ -\beta
\overline{U}^\mathrm{odd}_2 \right]
\right\rangle
\label{eq:Codd} \\
&=&
- \frac{\Lambda^3}{2^{3/2}} \int \mathrm d \boldsymbol r_3 ~
\left\{
\oint {\cal D}\boldsymbol x {\cal D}\boldsymbol y ~
\exp\left[-\frac{1}{\hbar}
\int_{0}^{\beta \hbar}
\frac{m}{4} \left| \frac{\mathrm d \boldsymbol x(\tau)}{\mathrm d
\tau} \right|^2 +
\frac{m}{2} \left| \frac{\mathrm d \boldsymbol y(\tau)}{\mathrm d
\tau} \right|^2
+ V_3\left(\boldsymbol x\left(\tau + {\beta \hbar}/{2}\right),
\boldsymbol x(\tau), \boldsymbol r_3 + \boldsymbol y(\tau)\right)
\mathrm d \tau \right] \right. \nonumber \\
& & \left.
- \oint {\cal D}\boldsymbol x ~ \exp\left[-\frac{1}{\hbar}
\int_{0}^{\beta \hbar}
\frac{m}{4} \left| \frac{\mathrm d \boldsymbol x(\tau)}{\mathrm d
\tau} \right|^2
+ U_2(|\boldsymbol x(\tau + \beta \hbar /2) - \boldsymbol x(\tau)|)
\mathrm d \tau \right]
\right\},
\end{eqnarray}
\end{widetext}
where we have defined
\begin{eqnarray}
\overline{V}^\mathrm{odd}_3(\boldsymbol r_3) &=& \frac{1}{P} \sum_{i=1}^{P}
U_3(\boldsymbol X^{(i)}, \boldsymbol X^{(i+P)},
\boldsymbol r_3 + \boldsymbol x_3^{(i)}) +
\nonumber \\ & &
U_2(|\boldsymbol X^{(i)} - \boldsymbol X^{(i+P)}|) +
\nonumber \\ & &
U_2(|\boldsymbol X^{(i)} - \boldsymbol r_3 - \boldsymbol x_3^{(i)}|) +
\nonumber \\ & &
U_2(|\boldsymbol X^{(i+P)} - \boldsymbol r_3 - \boldsymbol x_3^{(i)}|)
\label{eq:V_odd} \\
\overline{U}^\mathrm{odd}_2 &=&
\frac{1}{P} \sum_{i=1}^{P}
U_2(|\boldsymbol X^{(i)} - \boldsymbol X^{(i+P)}|).
\label{eq:U2_odd}
\end{eqnarray}
The $2P$ variables $\boldsymbol X^{(i)}$ have been defined analogously
to what has been done in Eq.~(\ref{eq:Bxc}). Notice that in the
discretized version, the average defining the odd exchange term in
Eq.~(\ref{eq:Codd}) is performed over two different kinds of ring
polymers: the first has $2P$ beads of mass $m/2$ and connects
particles 1 and 2 whose coordinates are exchanged by the permutation
operator, whereas the second -- corresponding to the third particle of
mass $m$ -- has $P$ beads.
A similar derivation holds for the even contribution to the third virial
coefficient, which is given by
\begin{widetext}
\begin{eqnarray}
C_\mathrm{even}(T) &=& -\frac{2 \Lambda^9}{3 V}
\int \mathrm d1 \mathrm d2 \mathrm d3 ~
\left\langle 1 2 3 \left|
\exp\left( -\beta \hat H_3 \right) \right| 3 1 2 \right\rangle
=
-\frac{2}{3} \frac{\Lambda^6}{3^{3/2}} \left\langle
\exp\left( - \beta \overline{V}^\mathrm{even}_3 \right)
\right\rangle
\label{eq:Ceven} \\
&=&
-\frac{2}{3} \frac{\Lambda^6}{3^{3/2}}
\oint {\cal D}\boldsymbol{x} ~ \exp\left[
-\frac{1}{\hbar} \int_0^{\beta \hbar} \frac{m}{6}
\left| \frac{\mathrm d \boldsymbol x(\tau)}{\mathrm d \tau} \right|^2
+ V_3(\boldsymbol x(\tau + 2 \beta \hbar /3),
\boldsymbol x(\tau + \beta \hbar /3),
\boldsymbol x(\tau))
\mathrm d \tau \right],
\end{eqnarray}
\end{widetext}
where we have defined
\begin{eqnarray}
\overline{V}^\mathrm{even}_3 &=& \frac{1}{P} \sum_{i=1}^{P}
U_3(\boldsymbol Y^{(i)}, \boldsymbol Y^{(i+P)}, \boldsymbol Y^{(i+2P)}) +
\nonumber \\ & &
U_2(|\boldsymbol Y^{(i)} - \boldsymbol Y^{(i+P)}|) +
\nonumber \\ & &
U_2(|\boldsymbol Y^{(i)} - \boldsymbol Y^{(i+2P)}|) +
\nonumber \\ & &
U_2(|\boldsymbol Y^{(i+P)} - \boldsymbol Y^{(i+2P)}|),
\label{eq:V_even}
\end{eqnarray}
together with
$\boldsymbol Y^{(i)} = \boldsymbol r_1^{(i)}$,
$\boldsymbol Y^{(i+P)} = \boldsymbol r_2^{(i)}$, and
$\boldsymbol Y^{(i+2P)} = \boldsymbol r_3^{(i)}$.
In the discretized version, the even contribution to the third virial
coefficient is an average over the coordinates of the $3P$ beads of a
single ring polymer corresponding to a particle of mass $m/3$.
Notice that, from a computational point of view, the evaluation of the
exchange contributions to the third virial coefficient is much less
demanding than the calculation of the Boltzmann part, which is
given as an integral over the positions of two particles. In fact, the
odd contribution is calculated as an integration over the position of
one particle only, whereas the even contribution is given by a simple
average over ideal-gas ring-polymer configurations. In particular,
the full calculation of $C(T)$ at the lowest temperature with 2.5~GHz
processors required $\sim 2400$ CPU hours, only 15\% of which was
needed to calculate the exchange contributions.
\section{Results and discussion}
\subsection{Details of the calculation}
We have calculated $C(T)$ for both isotopes of helium with the
path-integral method described above. We used the highly accurate
two-body potential of Przybytek {\em et al.},~\cite{He2-2010} which
includes the most significant corrections (adiabatic, relativistic,
and quantum electrodynamics) to the Born--Oppenheimer result. We also
used the three-body {\em ab initio} potential of Cencek {\em et
al.},~\cite{CPS09} which was derived at the Full Configuration
Interaction level and has an uncertainty approximately one-fifth that of
the three-body potential~\cite{He-3body} used in our
previous work.~\cite{Garberoglio2009b}
We generated ring-polymer configurations using the interpolation
formula of Levy.~\cite{levy54,fosdick-jordan66}
The number of beads was chosen as a function of the temperature $T$
according to the formulae $P = \mathrm{int}[(1200~\mathrm{K})/T] + 7$ for ${}^4$He and $P = \mathrm{int}[(1800~\mathrm{K})/T] + 7$ for ${}^3$He, where $\mathrm{int}[x]$ indicates the integer closest to $x$. These values of $P$ were
enough to reach convergence in the path-integral results at all the temperatures considered in the present study.
The spatial integrations were performed with the VEGAS
algorithm~\cite{vegas1}, as implemented in the GNU Scientific
Library,~\cite{gsl} with 1 million integration points and cutting off
the interactions at 4~nm.
The three-body interaction was pre-calculated on a three-dimensional
grid and interpolated with cubic splines.
The values of the virial coefficient and their statistical uncertainty
were obtained by averaging over the results of 256 independent
runs.
First of all, we checked that our methodology was able to reproduce
well-converged fully quantum $B(T)$ calculations for helium, which
were obtained using the same pair potential as the present
work.~\cite{Cencek11} Our results agree within mutual uncertainties
with these independent calculations, and confirm the observation,
already made when analyzing theoretical $B(T)$ calculations performed
using Lennard-Jones potentials, that exchange effects are significant
only for temperatures lower than about 7~K.~\cite{boyd69} The exchange
contribution to the second virial coefficient is negative in the case
of Bose--Einstein statistics and positive in the case of Fermi--Dirac
statistics, as one would expect.
\subsection{The third virial coefficient of ${}^4$He}
\begin{table*}
\begin{center}
\begin{tabular}{d|dd|dd|dd|dd|dd}
\multicolumn{1}{c|}{Temperature} &
\multicolumn{2}{c|}{$C$} &
\multicolumn{2}{c|}{$C_\mathrm{Boltzmann}$} &
\multicolumn{2}{c|}{$C_\mathrm{odd}$} &
\multicolumn{2}{c|}{$C_\mathrm{even}$} &
\multicolumn{2}{c}{$C_\mathrm{B}$} \\
(\mathrm K) &
\multicolumn{2}{c|}{$(\mathrm{cm}^6~\mathrm{mol}^{-2})$} &
\multicolumn{2}{c|}{$(\mathrm{cm}^6~\mathrm{mol}^{-2})$} &
\multicolumn{2}{c|}{$(\mathrm{cm}^6~\mathrm{mol}^{-2})$} &
\multicolumn{2}{c|}{$(\mathrm{cm}^6~\mathrm{mol}^{-2})$} &
\multicolumn{2}{c}{$(\mathrm{cm}^6~\mathrm{mol}^{-2})$} \\
\hline
2.6 & 266 & \pm 21 & 245 & \pm 21 & -863 & \pm 3 & -88.1 & \pm
0.5 & 972 & \pm 1 \\
2.8 & 631 & \pm 21 & 607 & \pm 20 & -504 & \pm 2 & -49.0 & \pm 0.3 &
577.5 & \pm 0.7 \\
3 & 848 & \pm 17 & 828 & \pm 17 & -301.7 & \pm 1.4 & -28.0 & \pm 0.2 &
349.67 & \pm 0.5 \\
3.2 & 937 & \pm 14 & 923 & \pm 14 & -184.9 & \pm 0.8 & -16.29 & \pm 0.12 &
215.0 & \pm 0.3 \\
3.5 & 1061 & \pm 10 & 1050 & \pm 10 & -88.9 & \pm 0.5 & -7.35 & \pm 0.06 &
106.89 & \pm 0.15 \\
3.7 & 1070 & \pm 9 & 1062 & \pm 9 & -55.5 & \pm 0.4 & -4.50 & \pm 0.05 &
67.97 & \pm 0.12 \\
4 & 1082 & \pm 8 & 1077 & \pm 8 & -27.9 & \pm 0.2 & -2.14 & \pm 0.02
& 35.14 & \pm 0.07 \\
4.2 & 1074 & \pm 7 & 1070 & \pm 7 & -17.79 & \pm 0.16 &
-1.352 & \pm 0.017 & 22.83 & \pm 0.05 \\
4.5 & 1049 & \pm 6 & 1047 & \pm 6 & -9.156 & \pm 0.09 &
-0.663 & \pm 0.008 & 12.18 & \pm 0.03 \\
5 & 986 & \pm 5 & 985 & \pm 5 & -3.28 & \pm 0.05 & -0.227 & \pm 0.004
& 4.50 & \pm 0.02 \\
6 & 861 & \pm 3 & 861 & \pm 3 & -0.361 & \pm 0.014 & -0.027 & \pm
0.001 & 0.682 & \pm 0.004 \\
7 & 746& \pm 2 & 746 & \pm 2 &
-0.029 & \pm 0.003 &
-0.004 & \pm 0.0003 &
0.115 & \pm 0.001 \\
8.5 & 620.7 & \pm 1.6 & 620.7 & \pm 1.6 & & & & & & \\
10 & 532.3 & \pm 0.8 & 532.3 & \pm 0.8 & & & & & & \\
12 & 449.7 & \pm 0.8 & 449.7 & \pm 0.8 & & & & & & \\
13.8033 & 401.0 & \pm 0.4 & 401.0 & \pm 0.4 & & & & & & \\
15 & 375.1 & \pm 0.5 & 375.1 & \pm 0.5 & & & & & & \\
17 & 342.2 & \pm 0.4 & 342.2 & \pm 0.4 & & & & & & \\
18.689 & 321.2 & \pm 0.2 & 321.2 & \pm 0.2 & & & & & & \\
20 & 307.7 & \pm 0.3 & 307.7 & \pm 0.3 & & & & & & \\
24.5561 & 274.2 & \pm 0.2 & 274.2 & \pm 0.2 & & & & & &
\end{tabular}
\caption{Values of the third virial coefficient of ${}^4$He and its
components at selected temperatures.
The $\pm$ values reflect only the standard uncertainty of the Monte
Carlo integration; see Ref.~\onlinecite{paper2} for complete
uncertainty analysis.}
\label{tab:CHe4}
\end{center}
\end{table*}
\begin{figure}
\includegraphics[width=0.95\linewidth]{C-He4-lowT}
\caption{The third virial coefficient of ${}^4$He. The black circles
are the results of the present calculations, with error bars
representing expanded uncertainties with coverage factor $k=2$. The
gray area shows the results of the recent low-temperature
experiments by Gaiser and collaborators.~\cite{Gaiser09,Gaiser10}}
\label{fig:CHe4}
\end{figure}
We report in Table~\ref{tab:CHe4} the values of the third virial
coefficient of ${}^4$He, together with the various contributions of
Eq.~(\ref{eq:C_full}), for temperatures in the range from $2.6$~K to
$24.5561$~K, which is the lowest temperature studied in our
previous work.~\cite{Garberoglio2009b}
The same data are plotted in Figure~\ref{fig:CHe4}, where they are
compared with the recent experimental measurements by Gaiser and
collaborators.~\cite{Gaiser09,Gaiser10}
More extensive comparison with available data over a wide range of
temperatures will be presented elsewhere.~\cite{paper2}
In Fig.~\ref{fig:CHe4}, our results are plotted with expanded
uncertainties with coverage factor $k=2$ as derived in
Ref.~\onlinecite{paper2}; the uncertainty at the same expanded level
for the experimental results was estimated from a figure in
Ref.~\onlinecite{Gaiser09}.
First, we notice that exchange effects are completely
negligible in the calculation of the third virial coefficient for
temperatures larger than 7~K, where their contribution to the overall
value is close to one thousandth of that of the Boltzmann part.
This is analogous to what has already been observed for the second virial
coefficient.
When the temperature is lower than 7~K, the various exchange terms
have contributions of similar magnitude and opposite sign, but their
overall contribution to $C(T)$ is positive at all the temperatures
that have been investigated. The exchange contribution to $C(T)$
is comparable to the statistical uncertainty of the calculation, which
progressively increases as the temperature is lowered.
In Fig.~\ref{fig:CHe4}, it can be seen that our theoretical values of
$C(T)$ are compatible with those of recent
experiments~\cite{Gaiser09,Gaiser10} down to the temperature of
10~K. For lower temperatures, the experimental results are
somewhat larger than the calculated values, even though agreement
is found again for temperatures below 4~K, where $C(T)$ passes through
a maximum.
\subsection{The third virial coefficient of ${}^3$He}
Similar behavior is observed in the case of the third virial
coefficient for ${}^3$He, whose calculated values are reported in
Table~\ref{tab:CHe3}. Also in this case the exchange contributions are
of opposite signs, but their combined effect is to reduce the value
obtained with Boltzmann statistics, which is the opposite trend to
that observed for ${}^4$He.
\begin{figure}
\includegraphics[width=0.95\linewidth]{BarChart}
\caption{The magnitude and sign of the various contributions to $C(T)$
at $T=3$~K.}
\label{fig:BarChart}
\end{figure}
The effects of the various contributions to the third virial
coefficient, in both the Bose--Einstein and Fermi--Dirac case, are
summarized in Fig.~\ref{fig:BarChart} for the representative
temperature of $T=3$~K. First, we notice that the largest
contribution to the third virial coefficient comes from the Boltzmann
term. The even exchange term has only a minor contribution, whereas
the two remaining terms ($C_\mathrm{odd}$ and $C_\mathrm B$)
have almost equal magnitudes and opposite signs. In the case of
Bose--Einstein statistics, the contribution to $C$ from
$C_\mathrm{odd}$ is negative, while that from $C_\mathrm B$ is
positive; the opposite situation is observed in the case of
Fermi--Dirac statistics. The overall sum of the exchange
contributions is positive for ${}^4$He and negative in the case of
${}^3$He.
The magnitude of each exchange contribution at a given temperature is
significantly greater for ${}^3$He; this reflects the larger de~Broglie
wavelength, which not only appears directly in the exchange terms but
also affects the range of space sampled by the ring polymers.
In the case of ${}^3$He, the exchange contribution is significantly
larger than the uncertainty of our calculations, at least at the
lowest temperatures that we have investigated. Similarly to the case
of ${}^4$He, quantum statistical effects on $C(T)$ contribute less
than one part in a thousand for temperatures higher than 7~K.
Even in the case of ${}^3$He, we observe $C(T)$ pass through a
maximum, at a temperature around 3~K, which is 1~K lower than the
temperature where $C(T)$ reaches a maximum for the ${}^4$He isotope.
\begin{table*}
\begin{center}
\begin{tabular}{d|dd|dd|dd|dd|dd}
\multicolumn{1}{c|}{Temperature} &
\multicolumn{2}{c|}{$C$} &
\multicolumn{2}{c|}{$C_\mathrm{Boltzmann}$} &
\multicolumn{2}{c|}{$C_\mathrm{odd}$} &
\multicolumn{2}{c|}{$C_\mathrm{even}$} &
\multicolumn{2}{c}{$C_\mathrm{B}$} \\
(\mathrm K) &
\multicolumn{2}{c|}{$(\mathrm{cm}^6~\mathrm{mol}^{-2})$} &
\multicolumn{2}{c|}{$(\mathrm{cm}^6~\mathrm{mol}^{-2})$} &
\multicolumn{2}{c|}{$(\mathrm{cm}^6~\mathrm{mol}^{-2})$} &
\multicolumn{2}{c|}{$(\mathrm{cm}^6~\mathrm{mol}^{-2})$} &
\multicolumn{2}{c}{$(\mathrm{cm}^6~\mathrm{mol}^{-2})$} \\
\hline
2.6 & 1338 &\pm 29 & 1857 &\pm 28 & -1803 &\pm 4 &
-274.8 &\pm 0.8 & -2047 &\pm 2 \\
2.8 & 1477 &\pm 24 & 1817 &\pm 23 & -1164 &\pm 3 &
-167.9 &\pm 0.6 & -1336.6 &\pm 1.3 \\
3 & 1480 &\pm 17 & 1712 &\pm 17 & -760.2 &\pm 2.2 &
-105.7 &\pm 0.4 & -886.4 &\pm 0.9 \\
3.2 & 1463 &\pm 17 & 1621 &\pm 17 & -503.5 &\pm 1.7 & -66.8 &\pm 0.3 & -594.2 &\pm 0.6 \\
3.5 & 1395 &\pm 13 & 1487 &\pm 13 & -277.2 &\pm 1.2 & -34.89 &\pm 0.15 & -333.7 &\pm 0.4 \\
3.7 & 1376 &\pm 11 & 1439 &\pm 11 & -189.2 &\pm 0.8 & -23.12 &\pm 0.11 & -229.6 &\pm 0.3 \\
4 & 1303 &\pm 9 & 1342 &\pm 9 & -107.0 &\pm 0.5 & -12.69 &\pm
0.07 & -133.43 &\pm 0.16 \\
4.2 & 1245 &\pm 9 & 1273 &\pm 9 & -73.9 &\pm 0.5 & -8.65 &\pm 0.05
& -93.82 &\pm 0.13 \\
4.5 & 1173 &\pm 7 & 1190 &\pm 7 & -43.7 &\pm 0.3 & -4.86 &\pm 0.03
& -55.94 &\pm 0.09 \\
5 & 1071 &\pm 6 & 1079 &\pm 6 & -18.41 &\pm 0.16 &
-1.963 &\pm 0.016 & -24.42 &\pm 0.05 \\
6 & 895 &\pm 4 & 897 &\pm 4 & -3.56 &\pm 0.06 & -0.353 &\pm 0.005 &
-5.143 &\pm 0.013 \\
7 & 772 &\pm 3 & 773 &\pm 3 & -0.78 &\pm 0.02 &
-0.0784 &\pm 0.002 & -1.196 &\pm 0.005 \\
8.5 & 645 &\pm 2 & 645 &\pm 2 &
-0.059 &\pm 0.006 & -0.0087 &\pm 0.0003 & -0.155 &\pm 0.001 \\
10 & 558.3 &\pm 1.6 & 558.3 &\pm 1.6 & & & & & & \\
12 & 475.5 &\pm 1.1 & 475.5 &\pm 1.1 & & & & & & \\
13.8033 & 426.2 &\pm 0.8 & 426.2 &\pm 0.8 & & & & & & \\
15 & 402.0 &\pm 0.7 & 402.0 &\pm 0.7 & & & & & & \\
17 & 369.6 &\pm 0.5 & 369.6 &\pm 0.5 & & & & & & \\
18.689 & 347.8 &\pm 0.4 & 347.8 &\pm 0.4 & & & & & & \\
20 & 333.4 &\pm 0.4 & 333.4 &\pm 0.4 & & & & & & \\
24.5561 & 297.8 &\pm 0.3 & 297.8 &\pm 0.3 & & & & & &
\end{tabular}
\caption{Values of the third virial coefficient of ${}^3$He and its
components at selected temperatures. Note that the odd contribution
$C_\mathrm{odd}$ contributes with a negative sign to the overall
value of the third virial coefficient $C$ (see
Eq.~(\ref{eq:C_full})). The $\pm$ values reflect only the standard
uncertainty of the Monte Carlo integration; see
Ref.~\onlinecite{paper2} for complete uncertainty analysis.}
\label{tab:CHe3}
\end{center}
\end{table*}
\begin{figure}
\includegraphics[width=0.95\linewidth]{Helium-3}
\caption{The third virial coefficient of ${}^3$He.}
\label{fig:CHe3}
\end{figure}
There are only a few sources of experimental data for $C(T)$ for ${}^3$He.
Keller~\cite{Keller55} measured five pressure-volume isotherms at temperatures
below 4~K; these were later reanalyzed by Roberts {\em et
al.}~\cite{Roberts64} and meaningful values of $C$ were obtained only for
the two highest temperatures. A later analysis of the Keller data was
performed by Steur (unpublished), whose equation for temperatures below 3.8~K
was reported by Fellmuth and Schuster~\cite{Fellmuth92}.
Some points were also extracted from volumetric data by Karnatsevich {\em et
al.}~\cite{Karna88} Recently, Gaiser and Fellmuth~\cite{Gaiser08,Gaiser_pc}
extracted virial coefficients from their measurements of two isotherms for
${}^3$He with dielectric-constant gas thermometry.
Figure~\ref{fig:CHe3} compares our calculated values to the available
experimental data, where the error bars represent expanded
uncertainties with coverage factor $k=2$. Error bars are not drawn
for our values above 5~K because they would be smaller than the size
of the symbol. As was the case in our previous
work,~\cite{Garberoglio2009b} the uncertainty of our values of $C(T)$
is determined by the statistical uncertainty of our Monte Carlo
calculations (shown in Tables~\ref{tab:CHe4} and \ref{tab:CHe3}) and
by the uncertainty in the two- and three-body potentials. At the
temperatures considered here, the statistical uncertainty is the
dominant contribution to the overall uncertainty. The full uncertainty
analysis is presented elsewhere.~\cite{paper2}
For the experimental points, these expanded uncertainties were taken as
reported in the original sources; we note that in some cases (notably
Ref.~\onlinecite{Karna88}) this appears to be merely the scatter of a fit and
therefore underestimates the total uncertainty.
Our results are qualitatively similar to the rather scattered
experimental data. We are quantitatively consistent with the values
based on analysis of the data of Keller, but values from the other
experimental sources are more positive than our results. We note that
a similar comparison for ${}^4$He~\cite{paper2}, where the
experimental data situation is much better, shows the $C(T)$ values of
Ref.~\onlinecite{Karna88} for ${}^4$He to deviate in a very similar
way not only from our results but from other experimental data we
consider to be reliable.
\section{Conclusions}
We used path-integral methods to derive an expression for the third
virial coefficient of monatomic gases, including the effect of
quantum statistics. We applied this formalism to the case of helium
isotopes, using state-of-the-art two- and three-body potentials.
We showed that exchange effects make no significant contribution
to the third virial coefficient above a temperature of approximately
7~K for both the fermionic and bosonic isotope. This is the same
behavior observed in the calculation of the second virial coefficient.
For temperatures lower than $7$~K, the sign of the contribution to
$C(T)$ from exchange effects depends on the bosonic or fermionic
nature of the atom.
In the case of ${}^4$He, the exchange contribution to $C(T)$ increases
its value compared to the value obtained with Boltzmann
statistics, although in our simulations the total exchange
contribution has the same order of magnitude as the statistical
uncertainty of the PIMC integration. In the case of ${}^3$He, the
exchange contribution is negative, and its magnitude is much larger
than the statistical uncertainty.
The range of temperatures that we have investigated covers the
low-temperature maximum of $C(T)$ for both isotopes. The third virial
coefficient of ${}^4$He reaches its maximum close to $4$~K, whereas in
the case of ${}^3$He the maximum is attained at a lower temperature.
For both helium isotopes, the uncertainty in our calculated third
virial coefficients is much smaller than that of the limited and
sometimes inconsistent experimental data. For ${}^4$He, we obtain good
agreement with the most recent experimental results, except for
some temperatures below 10~K. A full comparison with available
experimental data for ${}^4$He, including the higher temperatures of
importance for metrology, will be presented elsewhere.\cite{paper2} For
${}^3$He, we are qualitatively consistent with the sparse and scattered
experimental values; in this case especially our calculations provide
results that are much less uncertain than experiment. In both cases,
at the temperatures considered here, the uncertainty is dominated by
the statistical uncertainty of the Monte Carlo integration, meaning
that the uncertainty of $C(T)$ could be reduced somewhat with greater
expenditure of computer resources.
We note two directions in which extension of the present work could be
fruitful. One is the calculation of higher-order virial coefficients,
which is a straightforward extension of the method presented here.
This would be much more computationally demanding, but the fourth
virial coefficient $D(T)$ may be feasible, at least at higher
temperatures where the number of beads in the ring polymers would not
be large. Second, the method can be extended to calculate temperature
derivatives such as $\mathrm dC/ \mathrm dT$; such derivatives are of
interest in interpreting acoustic measurements. Work on the
evaluation of acoustic virial coefficients is in
progress.~\cite{paper2}
\begin{acknowledgments}
We thank C. Gaiser for providing information on low-temperature data
for $C(T)$ of helium isotopes, and M. R. Moldover and J. B. Mehl for
helpful discussions on various aspects of this work. The calculations
were performed on the KORE computing cluster at Fondazione Bruno
Kessler.
\end{acknowledgments}
\bibliographystyle{aipnum4-1}
|
\section{Introduction}
The possibility of discriminating the number of impinging photons on a
detector is a fundamental tool in many different fields of optical
science and technology \cite{had09}, including nanopositioning and the
redefinition of candela unit in quantum metrology \cite{pol09,zwi10},
foundations of quantum mechanics \cite{gen05}, quantum imaging
\cite{bri10} and quantum information \cite{lan09,obr09,gis07,ben10}, e.g
for communication and cryptography. As a matter of fact, conventional
single-photon detectors can only distinguish between zero and one (or
more) detected photons, with photon number resolution that can be
obtained by spatially \cite{multiSpatial} or temporally
\cite{multiTemporal} multiplexing this kind of on/off detectors.
\par
Genuine Photon Number Resolving (PNR) detectors needs a process
intrinsically able to produce a pulse proportional to the number of
absorbed photons. In fact, detectors with PNR capability are few, e.g.
photo-multiplier tubes \cite{burle}, hybrid photo-detectors \cite{NIST}
and quantum-dot field-effect transistors \cite{gan07}. At the moment,
the most promising genuine PNR detectors are the visible light photon
counters \cite{yam99} and Transition Edge Sensors (TESs)
\cite{fuk11,pre09,cab08,lit08,ros05,bandler06}, i.e. microcalorimeters
based on a superconducting thin film working as a very sensitive
thermometer \cite{irw05}.
\par
For a practical application of these detectors it is crucial to achieve
their precise characterization
\cite{lss99,jar01,dem03,mit03,dar04,zam05,lob08,rah11}. In particular,
it is generally assumed that TESs are linear photon counters, with a
detection process corresponding to a binomial convolution. It is also
expected that dark counts are not present in TESs. Taken together, these
assumptions allow one to characterize a TES by a single number assessing
the quantum efficiency of the detector, i.e. the probability $0\leq \eta
\leq 1$ that a photon impinging onto the detector is actually revealed.
In this paper, we present the first experimental reconstruction of the
POVM describing the operation of a TES and, in turn, the first
demonstration of the linearity. In section 2 we illustrate the method used for POVM
reconstruction, while in section 3 we describe the
experimental implementation. In section 4 we discuss the results and
close the paper with some concluding remarks.
\section{POVM reconstruction technique}
As TESs are microcalorimeters, they are intrinsically phase insensitive
detectors. In the following we thus assume that the elements of the
positive operator-value measurement (POVM) $\{\Pi_n\}$ are diagonal
operators in the Fock basis, i.e. \begin{equation} \Pi_n=\sum_m
\Pi_{nm} |m\rangle \langle m|, \end{equation} with completeness relation
$\sum_n \Pi_n = \mathbf{I}$. Matrix elements $\Pi_{nm}=\langle m| \Pi_n|
m\rangle$ describe the detector response to $m$ incoming photons, i.e.
the probability of detecting $n$ photons with $m$ photons at the
input\footnote{This corresponds to consider our TES as a \emph{grey box}
(instead of a black box), on the basis of this solid physics assumption,
i.e. the fact that they are microbolometers. On the other hand, trying
to find an experimental evidence of this phase-insensitiveness
assumption is pointless, as there is not a phase reference (e.g. from
the TES itself) to modify the phase of our probe states with respect to
it.}. A reconstruction scheme for $\Pi_{nm}$, i.e. a tomography of the
POVM, provides the characterization of the detector at the quantum
level. In order to achieve the tomography of the TES POVM, we exploit an
effective and statistically reliable technique \cite{h1,lun09,h2} based
on recording the detector response for a known and suitably chosen set
of input states, e.g. an ensemble of coherent signals providing a sample
of the Husimi Q-function of the elements of the POVM. \par Let us
consider a set of $K$ coherent states of different amplitudes
$|\alpha_j\rangle$, $j=1,...,K$. The probability of obtaining the
outcome $n$ from the TES, i.e. of detecting $n$ photons, with the $j$-th
state as input is given by
\begin{equation}
\label{eq:stat_model}
p_{nj} =\hbox{Tr}[|\alpha_j\rangle\langle\alpha_j| \Pi_n] =\sum_m \Pi_{nm}\, q_{mj}
\end{equation}
where $q_{mj}=\exp(-\mu_j)\mu_j^m/m!$ is the ideal photon statistics of
the coherent state $|\alpha_j\rangle$, $\mu _j = |\alpha_j|^2$ being the
average number of photons. In order to reconstruct the matrix elements
$\Pi_{nm}$, we sample the probabilities $p_{nj}$ and invert the
statistical model composed by the set of Eqs. (\ref{eq:stat_model}).
Since the Fock space is infinite dimensional, this estimation problem
contains, in principle, an infinite number of unknowns.
\par
A suitable truncation at a certain dimension $M$ should be performed, with
the constraint that the probability of having $m\geq M$ photons in the
states $|\alpha_j\rangle$ is not too large. In other words, given the
set of probing coherent states, we have a little amount of data for
the entries with $m \geq M$ and we cannot
investigate the performances of the detector above the corresponding
energy regimes.
\par
The distributions $p_{nj}$ in Eq. (\ref{eq:stat_model}) provide a sample
of the Q-functions $\langle\alpha_j| \Pi_n|\alpha_j\rangle$ of the POVM
elements, and any reconstruction scheme for the $\Pi_{nm}$ basically
amounts to recover the Fock representation of the $\Pi_n$'s from their
phase space Q-representation. In general, this cannot be done exactly
due to singularity of the antinormal ordering of Fock number projectors
$|n\rangle\langle n|$ \cite{bal83}. On the other hand, upon exploiting
the truncation described above, we deal with POVM elements expressed as
finite mixture of Fock states, which are amenable to reconstruction
\cite{qsm,dmr}. The statistical model in (\ref{eq:stat_model}) may be
solved using maximum likelihood (ML) methods or a suitable approximation
of ML. We found that reliable results are obtained already with a least
squares fit, i.e we have effectively estimated $\Pi_{nm}$ by minimization
of a regularized version of the square difference
$\sum_{nj} (\sum_{m=0}^{M-1} q_{mj}\, \Pi_{nm} - p_{nj})^2$
where the physical constraints of smoothness is implemented by exploiting
a convex, quadratic and device-independent function \cite{lun09}. We
also force normalization $\sum_{n=0}^{11} \Pi_{nm}=1$, $\forall m$,
where the last POVM element is defined as $\Pi_{11}=1-\sum_{n=0}^{10}\Pi_{n})$.
\section{Experiment}
The TES we have characterized is composed by a $\sim$ \mbox {90 nm}
thick Ti/Au film \cite{por08,tar07}, fabricated by e-beam deposition on
silicon nitride substrates. The effective sensitive area, obtained by
lithography and chemical etching, is $20\times20$ $\mu$m. The
superconducting wirings of Al, with thicknesses between 100 nm and 150
nm, have been defined by a lift-off technique combined with RF
sputtering of the superconducting films. Upon varying the top Ti film
thickness, the critical temperatures of these TESs can range between 90
mK and 130 mK, showing a sharp transition (1-2 mK).
\par
The characterization of TES has been carried out in a dilution
refrigerator with a base temperature of 30 mK. Furthermore, the
detector is voltage biased, in order to take advantage of the negative
Electro-Thermal Feedback, providing the possibility to obtain a
self-regulation of the bias point without a fine temperature control and
reducing the detector response time. The read-out operations on our TES
is performed with a DC-SQUID current sensor \cite{dru07}. Using room
temperature SQUID electronics, we bias our device and read out the
current response. Finally, the SQUID output is addressed to a LeCroy 400
MHz oscilloscope, performing the data acquisition, first elaboration and
storage. In our experiment, we have illuminated the TES with a
power-stabilized fiber coupled pulsed laser at $\lambda=1570$ nm (with a
pulse duration of 37 ns and a repetition rate of 9 kHz), whose pulse is
also used to trigger the data acquisition for a temporal window of 100
ns. The laser pulse energy $(365\pm 2)$ pJ is measured by a calibrated
power meter, and then attenuated to photon counting regime exploiting
two fiber coupled calibrated attenuators in cascade. The attenuated
laser pulses are then sent to the TES detection surface by a single mode
optical fiber. The set of coherent states needed to perform the POVM
reconstruction has been generated by lowering the initial laser pulse
energy from an initial attenuation of 63.5 dB {(corresponding to
an average of 130 photons per pulse)}, to 76.5 dB {(mean photon
number per pulse: 6.5)}, to obtain 20 different states $|\alpha_j
\rangle =|\sqrt{\tau_j} \alpha\rangle$ where $\tau_j$ is the channel
transmissivity, $j=1, ..., 20$.
\begin{figure}[h]
\begin{center}
\includegraphics[width=\columnwidth]{f1hist2}
\caption{{Dots represent the TES counts for two different values of
$|\alpha_j \rangle$: each point corresponds to a binning of an amplitude
interval of $1.3$ mV. Solid lines are the Gaussian fits on the
experimental data, while the dotted vertical lines are the thresholds.
Figure (a) is obtained with a coherent state characterized by a mean
photon number per pulse $\mu=31$, while for figure (b) the state used
had $\mu=87$. The insets of both figures compare the experimental
probability distribution (black bars), obtained from measurements binned
according to the drawn thresholds, with the corresponding Poisson
distributions of mean value $ \eta\mu$ (with $\eta= 5.1 \%$) (yellow
bars): as evident from the plots, the experimental results are in
remarkable agreement with the theoretical predictions, showing
respectively a fidelity of $99.994\%$ and $99.997\%$ .} \label{f1_TES}}
\end{center}
\end{figure}
\par
We work at fixed wavelength $\lambda=1570$ nm and thus, in ideal
conditions, we would expect a discrete energy distribution with outcomes
separated by a minimum energy gap $\Delta E=\frac{hc}{\lambda}$.
Experimentally, we observe a distribution with several peaks, whose
variances represent the energy resolution of the whole detection device.
In a first calibration run, after a binning on the oscilloscope
channels, we fit the data with a sum of independent Gaussian functions
(Fig. \ref{f1_TES} shows that the fitting functions are in excellent
agreement with experimental data); the first peak on the left is the
``0-peak", corresponding to no photon detection. These fits allowed us
to fix the amplitude thresholds (located close to the local minima)
corresponding to $n$ detected photons: this way, the histogram of counts
is obtained just binning on the intervals identified by these
thresholds. The distributions $p_{nj}$ are finally evaluated upon
normalizing the histogram bars to the total number of events for the
given state \footnote{Remarkably, the reconstructions obtained by
binning data using thresholds are almost indistinguishable from the ones
obtained by evaluating the number of events in the $n-th$ peak by
integrating the corresponding Gaussian of the fit reported in Fig.
\ref{f1_TES}.}. This threshold-based counts binning may introduce some
bias or fluctuations since the tails of the $n$-th Gaussian peak fall
out of the $n$ counts interval. On the other hand, the effects in
neighbouring peaks compensate each other and, overall, do not affect the
tomographic reconstruction.
\section{Results}
The POVM of our TES detection system has been reconstructed up to
$M=140$ incoming photons and considering $N=12$ POVM elements $\Pi_n$,
$n=0,...,N-1$, with $\Pi_{N-1}= \mathtt{1}- \sum_{n=0}^{N-2}\Pi_{n} $
describing the probability operator for the detection of more than $N-2$
photons. In Fig. \ref{f2_POVM} we show the matrix elements $\Pi_{nm}$ of
the first 9 POVM operators ($n=0,..,8$), for $0\leq m\leq 100$. The bars
represent the reconstructed $\Pi_{nm}$, while the solid lines denote the
matrix elements of a linear detector. In fact, as mentioned above, the
POVM of a linear photon counter can be expressed as a binomial
distribution
\begin{equation}
\Pi_n=\sum_{m=n}^{\infty} B_{nm} |m\rangle\langle m|
\end{equation}
of the the ideal photon number spectral measure with $B_{nm}
=\left(\begin{array}{c} m \\ n \end{array}\right) \eta^n (1-\eta)^{m-n}
$, where $\eta$ is the quantum efficiency of the detector. In order to
compare the POVM elements of the linear detector, i.e. $B_{nm}$, with
the reconstructed POVM elements $\Pi_{nm}$ we have first to estimate the
value of the quantum efficiency $\eta$.
\par
This can be done on the sole basis of the experimental data using ML
estimation, i.e. we average the values of $\eta$ which maximize the
log-likelihood functions
\begin{equation}
L_j=\sum_{n} N_{nj} \log\left(\sum_m B_{nm} q_{mj}\right)
\end{equation}
where $N_{nj}$ is the number of $n$-count events obtained with the
$j$-th input state $|\sqrt{\tau_j}\alpha\rangle$. The overall procedure
leads to an estimated value of the quantum efficiency $\eta=(5.10 \pm
0.04) \%$, where the uncertainty accounts for the statistical
fluctuations (for each signal probe we estimated the value of $\eta$,
and then we averaged over the ensemble).
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.75\columnwidth]{f2povm2}
\caption{(Color Online) Reconstructed POVM of our TES photon counting
systems. Bars represent the matrix elements $\Pi_{nm}$ as a function of
$m=0,100$ for $n=0,1,2$ (main plot), $n=3,4,5$ (b), $n=6,7,8$ (c).
Continuous lines represent the POVM elements of a linear photon counter
with quantum efficiency $\eta=5.10\%$.} \label{f2_POVM}
\end{center}
\end{figure}
\par
As it is apparent from Fig. \ref{f2_POVM}, we have an excellent
agreement between the reconstructed POVM and the linear one with the
estimated quantum efficiency. In particular, the elements of the
POVM are reliably reconstructed for $m \leq 100$, whereas for higher
values of $m$ the quality of the reconstructions degrades.
In the regime $m \leq 100$ the fidelity $F_m=\sum_n \sqrt{\Pi_{n m} B_{n
m}}$ is larger than 0.99 (see the right inset of Fig. \ref{hus}), while
it degrades to 0.95 for $100 \leq m \leq 140$. In order to investigate
the effects of experimental uncertainties, we performed a sensitivity
analysis taking into account the uncertainties on the energy of the
input state and on the attenuators, obtaining fidelities always greater
than $98.35\%$ for the $12$ entries. In order to further confirm the
linearity hypothesis, as well as to assess the reliability of the
reconstruction, we have compared the measured distributions $p_{nj}$
with those obtained for a linear detector, i.e.
\begin{equation}
l_{nj}= \eta^n \exp(-\eta \mu_j) \mu_j^n/n!
\end{equation}
and with those obtained using the reconstructed
POVM elements, i.e.
\begin{equation}
r_{nj}=\sum_{m=n}^{M} \Pi_{nm} q_{mj} .
\end{equation}
In Fig. \ref{hus} we report the three distributions for the whole set of
probing coherent states, whereas in the left inset we show the (absolute)
differences $|p_{nj}-l_{nj}|$ and $|p_{nj}-r_{nj}|$ between those
distributions and the measured ones.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.6\columnwidth,angle=270]{f3all2}
\vspace{15mm}
\caption{(Color Online) Comparison of the measured distributions
$p_{nj}$ (green bars, on the left of each group) of the coherent states
$|\alpha_j\rangle$ used for POVM reconstruction with those obtained
using the reconstructed POVM elements $r_{nj}$ (yellow central bars).
and with those obtained under the linearity hypothesis $l_{nj}$ (blue
right bars) The left inset shows the absolute differences
$|p_{nj}-r_{nj}|$ (yellow left bars) and $|p_{nj}-l_{nj}|$ (blue right
bars). The right inset shows the fidelity $F_m$ between the
reconstructed POVM elements at fixed $m$ and those of a linear photon
counter with quantum efficiency $\eta=5.10\%$. }\label{hus} \end{center}
\end{figure}
\par
As it is apparent from the plots,
we have an excellent agreement between the different determinations of
the distributions. This confirms the linear behavior of the detector,
and proves that the reconstructed POVM provides a reliable description
of the detection process.
We have also modified the detection model to take into account the
possible presence of dark counts. In this case, upon assuming a
Poissonian background, the matrix elements of the POVM are given by
$\Pi_{nm}= \exp(-\gamma) \sum_j\gamma^j/j!\,B_{(n-j)m}$ and we have
developed a ML procedure to estimate both the quantum efficiency $\eta$
and the mean number of dark counts per pulse $\gamma$. We found that the
value for $\eta$ is statistically indistiguishable from the one obtained
with the linear-detector model, whereas the estimated dark counts per
pulse are $\gamma=(-0.03
\pm 0.04)$, in excellent agreement with the direct measurement performed
on our TES detector using the same fitting technique discussed above, providing a substantially negligible dark count level
$\gamma=(1.4\pm0.6)\times10^{-6}$. The same conclusion is obtained for
any other model, e.g. super-Poissonian, of the background.
\par
In conclusion, we have performed the first tomographic reconstruction of
the POVM describing a TES photon detector. Our results clearly validate
the description of TES detectors as linear photon counters and, together with the precise estimation of the
quantum efficiency, pave the way for practical applications of TES
photon counters in quantum technology.
\section*{Acknowledgements}
We acknowledge the support of the EC FP7 program ERA-NET Plus, under
Grant Agreement No. 217257. MGAP thanks Stefano Olivares for useful
discussions.
\section*{References}
|
\section{Introduction}
During the last three decades, the Berry phase \cite{Berry} has become
an important concept in condensed matter physics, \cite{S-W} playing a
fundamental role in various phenomena such as electric polarization,
orbital magnetism, anomalous Hall effects, {\it etc.} \cite{Niu2010} The
Berry phase (or geometrical phase) in solids is determined by
topological characteristics of the energy bands in the Brillouin zone
(BZ) and represents a fundamental property of the system. For example, a
non-zero Berry phase, which can be measured directly in the
magnetotransport experiments, reflects the existence of a singularity in
the energy bands such as a band-contact line in three-dimensional (3D)
bulk states or a Dirac point in a two-dimensional (2D) surface state.
\cite{Mikitik1999} Also, the Berry phase of $\pi$ is responsible for the
peculiar ``anti-localization" effects in carbon nanotubes or graphene.
\cite{TAndo} Recently, the $\pi$ Berry phase has been observed in the
Shubnikov-de Haas (SdH) oscillations in graphene, \cite{N-G2005,Kim2005}
giving one of the key evidences for the Dirac nature of quasiparticles
in the 2D carbon sheet.
The 3D topological insulator (TI) also supports spin polarized 2D Dirac
fermions on its surface,\cite{HK} which can be distinguished from
ordinary charge carriers by a non-zero Berry phase. Recently, several
groups have reported observations of the SdH oscillations coming from
the 2D surface states of TIs.
\cite{BiSb_amro,Ong2010,Fisher_np,BTS_Rapid,Xiong,Morpurgo,HgTe,Bi2Te3nano}
In those studies, a finite Berry phase has been reported, but it usually
deviates from the exact $\pi$ value. For example, in the new TI material
Bi$_{2}$Te$_{2}$Se (BTS), \cite{BTS_Rapid} where a large contribution of
the surface transport to the total conductivity has been observed, the
apparent Berry phase extracted from the SdH-oscillation data was
0.44$\pi$. So far, the Zeeman coupling of the spin to the magnetic field
has been considered \cite{Fisher_np} as a possible source of such a
discrepancy. Here, we show that in addition to the
Zeeman term, the deviation of the dispersion relation $E(k)$ from an
ideal linear dispersion \cite{DasSarma2010} can shift the Berry phase
from $\pi$. We further show how the real experimental data for non-ideal
Dirac fermions could be understood by taking into account those
additional factors.
\section{energy dispersion of surface states}
\begin{figure}\includegraphics*[width=8.0cm]{Fig1.ps}
\caption{(Color online)
Experimental band dispersions (symbols) in Bi$_{2}$Te$_{2}$Se measured by ARPES in Ref.
\onlinecite{BTS_Hasan} and the fitting of Eq. (1) to the surface state (solid line).
Large symbols depict the bulk state.
}
\label{fig1}
\end{figure}
The energy dispersion of the surface states in TIs can be directly measured in
angle-resolved photoemission spectroscopy (ARPES) experiments.
As an example, Fig. 1 shows the dispersion of the surface state (together with
the bulk state) in BTS reported by
Xu {\it et al.} \cite{BTS_Hasan}
One can easily recognize that $E(k)$ is not an ideal
Dirac-like dispersion, but it can be fitted reasonably well for the
two high-symmetry axes with
\begin{equation}
E(k) = v_{F} \hbar \, k + \frac{\hbar^{2}}{2m} k^{2},
\end{equation}
with a single Fermi velocity $v_{F}$ = 3.4$\times$10$^{5}$ m/s and the
effective mass $m$ which slightly varies with the direction in the
surface BZ as shown by the solid lines in Fig. 1 [$m/m_{0}$ = 0.15
(0.125) for the $\bar{\Gamma} \rightarrow \bar{M}$ ($\bar{\Gamma}
\rightarrow \bar{K}$) direction with $m_0$ the free electron mass].
Similar fittings can be obtained for other TIs owing to the progress in
the ARPES studies of these materials. \cite{HK,HgTe,H5,Chen}
\section{Berry phase in quantum oscillations}
It is commonly accepted that quantum oscillations observed in 3D metals
can be well understood within Lifshits-Kosevich \cite{L-K} (the de
Haas-van Alphen effect) and Adams-Holstein \cite{A-H} (the SdH effect)
theories. Recently this approach has been generalized to describe
magnetic oscillations in graphene, which is a 2D system with a
Dirac-like spectrum of charge carriers. \cite{Sh-G-B, G-Sh} There are
two most prominent features that distinguish such systems from materials
with a parabolic spectrum: First, rather weak magnetic fields are
sufficient to bring the system into a regime where only a few Landau
levels are occupied. Second, Dirac quasiparticles acquire the Berry
phase of $\pi$ in the cyclotron motion, changing the phase of quantum
oscillations.
In the SdH effect, the oscillating part of $\rho_{xx}$ follows
\begin{equation}
\Delta \rho_{xx} \sim \cos[2\pi (\frac{F}{B}-\gamma)],
\end{equation}
where $F$ is the oscillation frequency and 2$\pi \gamma$ is the phase factor
($0 \le \gamma < 1$).
This is the same $\gamma$ as in the Onsager's semiclassical quantization
condition \cite{Shoenberg1984}
\begin{equation}
A_{N}=\frac{2\pi e}{\hbar}B(N+\gamma),
\end{equation}
when the $N$-th Landau level (LL) is crossing the Fermi energy $E_F$
($A_{N}$ is the area of an electron orbit in the $k$-space).
$\gamma$ is directly related to the Berry phase through\cite{Mikitik1999}
\begin{equation}
\gamma - \frac{1}{2} = -\frac{1}{2\pi} \oint_{\Gamma} \vec{\Omega}\,d\vec{k},
\end{equation}
where $\vec{\Omega}(\vec{k})$=i$\int d\vec{k} \, u_{\vec{k}}^{*}(\vec{r}) \,
\vec{\nabla}_{\vec{k}} u_{\vec{k}}(\vec{r})$ is the Berry connection,
$u_{\vec{k}}(\vec{r})$ is the amplitude of the Bloch
wave function, $\Gamma$ is a closed electron orbit (the intersection of the Fermi
surface $E(\vec{k}) = E_F$ with the plane $k_{z} = const$). For spinless quasiparticles,
it is known \cite{Mikitik1999,Shoenberg1984} that the Berry phase is zero for
a parabolic energy dispersion ($\gamma$ = $\frac{1}{2}$) and $\pi$ for a linear
energy dispersion ($\gamma$ = 0).
Experimentally, $\gamma$ can be obtained from an analysis of the
Landau-level (LL) fan diagram. There are three quantities which are
often used as abscissa for plotting a LL fan diagram: (i) Landau level
index $N$, which determines the energy $E_{N}$ of the $N$-th LL. (ii)
Filling factor $\nu$ ($ \equiv \frac{N_{s}S}{N_{\phi}}$, where $N_{s}$
is the density of charge carriers, $S$ is the area of the sample,
$N_{\phi}$ = $\frac{BS}{\Phi_{0}}$ is the number of flux quanta, and
$\Phi_{0}$=$\frac{h}{e}$ is the flux quantum). (iii) An integer number
$n$ which marks the $n$-th minimum of the oscillations in $\rho_{xx}$.
Although all three quantities are related to each other, the most
straightforward way to plot a LL fan diagram from the $\rho_{xx}$
oscillations in a 2D system \cite{Ong2010} is to assign an integer $n$
to a minimum of $\rho_{xx}$ (or a half-integer to a maximum of
$\rho_{xx}$). From Eq. (2), one can see that the first minimum in
$\rho_{xx}$ is always in the range of $0 < \frac{F}{B_{1}} \leq 1$.
Thus, the plot of $F/B_{n}$ vs $n$, which makes a straight line with a
unit slope for periodic oscillations, is uniquely defined and cuts the
$n$-axis between 0 and 1 depending on the phase of the oscillations,
$\gamma$.
The ordinate 1/$B_{n}$ in a LL fan diagram is determined by the Landau
quantization of the cyclotron motion of electrons in a magnetic field.
In 2D systems, upon sweeping $B$, $\rho_{xx}$ shows a maximum (or a
sharp peak in the quantum Hall effect \cite{Ong2010}) each time when
$E_{N}(B)$ crosses the Fermi level. Thus, the position of the maximum in
$\rho_{xx}$ that corresponds to the $N$-th LL, 1/$B_{N}$, is given by
\begin{equation}
2\pi \Big(\frac{F}{B_{N}}-\gamma \Big) = 2\pi N.
\end{equation}
On the other hand, the $n$-th minimum in $\rho_{xx}$ occurs at 1/$B_n$ when
$2\pi(\frac{F}{B_n}-\gamma) = 2\pi n - \pi$, so the positions of the maxima
and minima are shifted by $\frac{1}{2}$ on the $n$-axis.
The Onsager's relation \cite{Shoenberg1984} gives $F$ in terms of the Fermi wave
vector $k_{F}$ as
$F = (\hbar / 2\pi e)\pi k_{F}^{2}$, and this $k_{F}$ can be calculated from Eq. (1) as
\begin{equation}
k_{F}^{2} = 2\Big(\frac{m v_{F}}{\hbar}\Big)^{2}\Bigg(1+\frac{E_{F}}{m v_{F}^{2}} -
\sqrt{1+\frac{2E_{F}}{m v_{F}^{2}}} \Bigg).
\end{equation}
Also, when $E_F$ is at the $N$-th LL, there is a relation
\begin{equation}
E_{N}(B_{N}) = E_{F}.
\end{equation}
From Eqs. (5)--(7), one obtains
\begin{equation}
\gamma = \frac{m v_{F}^{2}}{\hbar \, \omega_{c}}\Bigg(1+\frac{E_{N}}{m v_{F}^{2}} -
\sqrt{1+\frac{2E_{N}}{m v_{F}^{2}}} \Bigg) - N,
\end{equation}
where $\omega_{c}$=$eB/m$ is the cyclotron frequency.
In general case, $\gamma$ is a function of $B$, meaning that
oscillations in $\rho_{xx}$ are quasi-periodic in $1/B$.
In order to calculate $\gamma$ one needs to find the eigenvalues $E_{N}$
for a given Hamiltonian.
\section{Model Hamiltonian}
For the (111) surface state of the Bi$_{2}$Se$_{3}$-family TI compounds,
the Hamiltonian for non-ideal Dirac quasiparticles in perpendicular
magnetic fields can be written as \cite{MQOsc}
\begin{equation}
\hat{H} = v_{F}(\Pi_{x}\sigma_{y}-\Pi_{y}\sigma_{x}) +
\frac{{\bf\Pi}^{2}}{2m} - \frac{1}{2} \,g_{s} \mu_{B}B \sigma_{z} ,
\end{equation}
where the Landau gauge ${\bf A} = (0,By,0)$ for the vector potential is
used, ${\bf \Pi}$=$\hbar \, {\bf k}$+$e {\bf A}$, $\sigma_{i}$ are the Pauli
matrices, $\mu_{B}$ is the Bohr magneton, and $g_{s}$ is the surface $g$-factor.
The LL energies are given by \cite{MQOsc,PP2009}
\begin{equation}
E_{N}^{(\pm)} = \hbar \omega_{c} N \pm \sqrt{2 \hbar \, v_{F}^{2} e B N +
\Big(\frac{1}{2} \hbar \omega_{c} - \frac{1}{2} \,g_{s} \mu_{B} B\Big)^{2}} ,
\end{equation}
where ``$+$" and ``$-$" branches are for electrons and holes, respectively.
The obtained eigenvalues $E_{N}$ define the exact positions of maxima in
$\rho_{xx}$ and, thus, the phase of oscillations through Eq. (8).
In two extreme cases, for non-magnetic fermions ($g_{s}$ = 0),
Eq. (8) gives the expected results. First, for a
linear dispersion (ideal Dirac fermions), $m \rightarrow \infty$ leads to
$E_{N}$=$\pm \sqrt{2\hbar \, e v_{F}^{2} B N}$ and
$\gamma \rightarrow \frac{E_{N}^2}{2\hbar \, e v_{F}^{2} B} - N$,
giving $\gamma$ = 0 (Berry phase is $\pi$).
Second, for a parabolic dispersion, $v_{F} \rightarrow$ 0 leads to
$E_{N}$=$\hbar \,\omega_{c}(N+\frac{1}{2})$ and
$\gamma \rightarrow \frac{E_{N}}{\hbar \, \omega_{c}} - N$,
giving $\gamma$ = $\frac{1}{2}$ (Berry phase is zero).
This gives confidence that the expression for $\gamma$ given in Eq. (8) is
generally valid for the topological surface state with a non-ideal Dirac
cone described by Eq. (1).
\section{Landau-level fan diagram for non-ideal Dirac fermions}
\begin{figure}\includegraphics*[width=6.5cm]{Fig2.ps}
\caption{(Color online)
(a) Landau level fan diagram calculated for $F$ = 60 T,
$v_{F}$ = 3$\times$10$^{5}$ m/s, $g_{s}$ = 0, and different $m/m_{0}$.
Arrows show the direction of decreasing $m/m_{0}$. The dashed and dotted
lines are the expected behaviors for an ideal Dirac dispersion and a
parabolic dispersion, respectively.
(b) Landau level fan diagram calculated for $F$ = 60 T, $m/m_{0}$ = 0.1,
$g_{s}$ = 0, and different $v_{F}$. Arrows show the direction of
decreasing $v_{F}$. Insets show the calculated
$\gamma(N)$.
}
\label{fig2}
\end{figure}
Let us first consider how the LL fan diagram will be modified, when both
linear and parabolic terms are present in the Hamiltonian [Eq. (9)]. For
the moment, the Zeeman coupling of the electron spin to the magnetic
field is assumed to be negligible ($g_{s}$ = 0). Figure 2 (a) shows the
calculated positions of maxima and minima in $\rho_{xx}$ for
oscillations with $F$ = 60 T and $v_{F}$ = 3$\times$10$^{5}$ m/s as
$m/m_{0}$ is varied. One can see that upon decreasing $m/m_{0}$, the
calculated lines on the LL fan diagram are gradually shifting upward
from the ideal Dirac line that crosses the $n$-axis at exactly
$\frac{1}{2}$. Moreover, the lines are not straight anymore, which is
clearly inferred in the dependence of $\gamma$ vs $N$ shown in the
inset. With decreasing $N$ (increasing $B$), $\gamma$ becomes larger,
reflecting the change in the phase of oscillations at high fields.
Similar change in the LL fan diagram occurs if we modify another parameter,
$v_{F}$. As shown in Fig. 2 (b), the calculated lines are gradually shifting
upward from the ideal Dirac line as $v_{F}$ is decreased.
The results shown in Figs. 2 can be understood as a competition
between linear and quadratic terms in the Hamiltonian [Eq. (9)].
Note that for the whole range of the parameters $v_{F}$ and $m/m_{0}$, the positions
of maxima and minima in $\rho_{xx}$ lie between two straight lines
(shown as dotted and dashed lines in Figs. 2)
corresponding to $\gamma$ = 0 and $\gamma$ = $\frac{1}{2}$.
\begin{figure}\includegraphics*[width=6.5cm]{Fig3.ps}
\caption{(Color online)
Landau level fan diagram calculated for $F$ = 60 T and different $g_{s}$,
keeping $v_{F}$ = 3$\times$10$^{5}$ m/s and $m/m_{0}$ = 0.1 constant.
Arrows show the direction of changing $g_{s}$.
The dotted line is the expected behavior for a parabolic dispersion.
Inset shows the calculated $\gamma(N)$.
}
\label{fig3}
\end{figure}
Let us now take the Zeeman term into considerations. Figure 3 shows the
LL fan diagram calculated with $F$ = 60 T, $v_{F}$ = 3$\times$10$^{5}$
m/s, and $m/m_{0}$ = 0.1, while $g_{s}$ is varied. To understand the
effect of the Zeeman coupling, it is important to recognize the
following two points: (i) The Zeeman term in Eq. (10) would tend to
cancel the $\frac{1}{2}\hbar\omega_c$ term when $g_s$ is positive. In
fact, when $\frac{1}{2}\hbar\omega_c$ = $\frac{1}{2} g_s \mu_B B$ ({\it
i.e.}, $g_s = 2 m_0/m$) is satisfied, the effect of the finite effective
mass is canceled and the LL fan diagram becomes identical to that for
the linear dispersion (ideal Dirac) case. In the present simulations, we
use $m/m_0$ = 0.1, so that this cancellations occurs when $g_s$ = 20.
(ii) A pair of $g_s$ values that give the same
$|\frac{1}{2}\hbar\omega_c - \frac{1}{2} g_s \mu_B B|$ are effectively
the same in determining the behavior of the LL fan diagram. The result
of our calculations shown in Fig. 3 is a demonstration of these two
points. Since the Zeeman effect is more pronounced at higher fields, the
LL fan diagram in Fig. 3 is strongly modified from a straight line when
the quantum limit is approached, {\it i.e.}, close to $N$ = 0.
\section{The case of BTS}
\begin{figure}\includegraphics*[width=8.7cm]{Fig4.ps}
\caption{(Color online)
Landau level fan diagram for oscillations in $d\rho_{xx}/dB$ measured at
$T$ = 1.6 K and $\theta \simeq$ 0$^{\circ}$ reported in Ref.
\onlinecite{BTS_Rapid} for BTS. Minima and maxima in $d\rho_{xx}/dB$
correspond to $n+\frac{1}{4}$ and $n+\frac{3}{4}$, respectively. Solid
(dark gray) line is the calculated diagram for an ideal Dirac cone with
$v_{F}$ = 3.4$\times$10$^{5}$ m/s and $F$ = 62 T; dashed (blue) line
includes the effect of the actual dispersion with $m/m_{0}$ = 0.13;
dotted (red) line further includes the Zeeman effect, where $g_{s}$ = 76
or $-45$ was determined from a least-square fitting to the data. Inset
shows the experimental data and calculations after subtracting the contribution
from an ideal Dirac cone, $(1/B)_{\rm Dirac}$, where
$\Delta (1/B) \equiv (1/B) - (1/B)_{\rm Dirac}$.
}
\label{fig4}
\end{figure}
Let us examine the real data measured in the BTS sample,
\cite{BTS_Rapid} in the light of the above considerations.
Figure 4 shows the
LL fan diagram for oscillations in $d\rho_{xx}/dB$ measured at $T$ = 1.6
K in magnetic fields perpendicular to the (111) plane. \cite{BTS_Rapid}
In Ref. \onlinecite{BTS_Rapid}, the data were simply fitted with a
straight line, and the least-square fitting gave a slope of $F$ = 64 T
with the intersection of the $n$-axis at 0.22$\pm$0.12; this result
implies a finite Berry phase, but it was not exactly equal to $\pi$,
which remained a puzzle. \cite{BTS_Rapid} Now, we analyze this LL fan
diagram by considering the non-ideal Dirac dispersion as well as the
Zeeman effect. The ARPES data \cite{BTS_Hasan} for the surface state of
BTS (Fig. 1) gives $v_{F}$ = 3.4$\times$10$^{5}$ m/s and the averaged
effective mass $m/m_{0}$ = 0.13. We fix the oscillation frequency $F$ at
62 T obtained from the Fourier-transform analysis of the $d\rho_{xx}/dB$
oscillations.\cite{BTS_Rapid}
In Fig. 4, the calculated diagram for an
ideal Dirac cone is shown by the solid (dark gray) line, whereas
that for the non-ideal Dirac cone with the effective-mass term is shown
by the dashed (blue) line. One can see that the difference is small,
which indicates that the effective mass of 0.13$m_0$ is not light enough
to significantly alter the LL fan diagram. One may also see that these
two lines undershoot the actual data points at smaller $n$, which is even more
clearly seen in the inset, where the experimental data and the calculations are
shown after subtracting the contribution from an ideal Dirac cone.
By further including the Zeeman effect, we can greatly improve the analysis,
as shown by the dotted (red) line; here, $g_s$ is taken as the only fitting
parameter and a least-square fitting to the data was performed. The best
value of $g_s$ is 76 or $-45$.
The inset of Fig. 4 makes it clear that it is the slight deviation of
the experimental points from the ideal Dirac line that causes a simple
straight-line fitting of the LL fan diagram to intersect the $n$-axis
not exactly at 0.5. Since the Berry phase in real situations is not a
fixed value but is dependent on the magnetic field, the simple
straight-line analysis of the LL fan diagram should not be employed for
the determination of the Berry phase. Obviously, the SdH oscillations of
the topological surface states are best understood by the analysis which
considers both the the deviation of the energy spectrum of the
Dirac-like charge carriers from the ideal linear dispersion and their
strong coupling with an external magnetic field.
\section{Other materials}
\begin{table}[b]
\centering
\begin{tabular}{l c c c c}
\hline\hline
Material & $v_{F}$ (m/s) & $m/m_{0}$ & Ref. & remark \\ [0.5ex]
\hline
Bi$_{2}$Se$_{3}$ & 3.0 $\times$10$^{5}$ & 0.25 & [\onlinecite{H5}] & averaged \\
Bi$_{2}$Te$_{2}$Se & 3.4 $\times$10$^{5}$ & 0.13 & [\onlinecite{BTS_Rapid}] & averaged\\
Bi$_{2}$Te$_{3}$ & 3.7 $\times$10$^{5}$ & 3.8 & [\onlinecite{Chen}] & near Dirac point\\
graphene & 1 $\times$10$^{6}$ & $\infty$ & [\onlinecite{N-G2005}] & calculations \\ [1ex]
\hline
\end{tabular}
\caption{Parameters of the surface states from ARPES.}
\end{table}
\begin{figure}\includegraphics*[width=8.7cm]{Fig5.ps}
\caption{(Color online)
Landau level fan diagrams for SdH oscillations observed in various TIs
and graphene. Symbols are obtained from the published experimental data
in the literature.
Solid lines are calculations taking into account the non-ideal
dispersions of the surface states (determined by $m/m_{0}$) and the
Zeeman coupling to an external magnetic field (determined by $g_{s}$).
Dashed lines are calculations for ideal Dirac fermions ($m/m_{0}$
= $\infty$ and $g_{s}$ = 0). Open diamonds are $(d\rho_{yx}/dB)_{min,max}$
in Bi$_{2}$Te$_{3}$ from Ref. \onlinecite{Ong2010}; filled circles are
$(\Delta R_{xx})_{min}$ in Bi$_{2}$Se$_{3}$ from
Ref. \onlinecite{Fisher_np}; open circles are $(R_{xx})_{min,max}$ in
graphene from Ref. \onlinecite{N-G2005}; filled squares are $(\Delta
R_{xx})_{min, max}$ in a Bi$_{2}$Te$_{3}$ nanoribbon from
Ref. \onlinecite{Bi2Te3nano}; open squares are
$(d\rho_{xx}/dB)_{min,max}$ in BTS from Ref. \onlinecite{BTS_Rapid}.
}
\label{fig5}
\end{figure}
Similar analysis can be performed for other TIs in which the quantum
oscillations coming from the 2D topological surface states have been
observed. Figure 5 shows the LL fan diagrams for the SdH oscillations
published to date for TI materials,
\cite{Fisher_np,Ong2010,BTS_Rapid,Bi2Te3nano,note1} together with the
data obtained in graphene, \cite{N-G2005} which provides a good
reference for studies of Dirac fermions. We digitized the published
experimental data in the literature and determined ourselves the
positions of minima $1/B_{min}$ and maxima $1/B_{max}$ of the
oscillating parts of resistivity (resistance), Hall resistivity, or
their derivatives with respect to $B$. The obtained data
for various materials are plotted as functions of $n$ in Fig. 5. Note
that, to avoid ambiguities, we considered only those data that show
oscillations with a single frequency. \cite{note1}
\begin{table}[t]
\centering
\begin{tabular}{l c c c c}
\hline\hline
Material & Ref. & $F$ (T) & $E_{F}$ (eV) & $g_{s}$ \\ [0.5ex]
\hline
Bi$_{2}$Se$_{3}$ & [\onlinecite{Fisher_np}] & 30.7 & 0.074 & 55 or -39 \\
Bi$_{2}$Se$_{3}$ & [\onlinecite{Fisher_np}] & 88.6 & 0.143 & 55 or -39 \\
Bi$_{2}$Te$_{2}$Se & [\onlinecite{BTS_Rapid}] & 62.0 & 0.152 & 76 or -45 \\
Bi$_{2}$Te$_{3}$ & [\onlinecite{Ong2010}] & 27.3 & 0.074 &65 or -65 \\
Bi$_{2}$Te$_{3}$, nanoribbon & [\onlinecite{Bi2Te3nano}] & 54.7 & 0.101 &65 or -65 \\
graphene & [\onlinecite{N-G2005}] & 43.3 & 0.239 & 0 \\
\hline
\end{tabular}
\caption{Parameters used for the calculations shown in Fig. 5.}
\end{table}
The parameters of the surface states used in our fan-diagram analyses
have been obtained from the published ARPES data by fitting them in the
same way as for BTS (see Fig. 1). Table I shows $v_{F}$ and $m/m_{0}$
for the Bi$_{2}$Se$_{3}$/Bi$_{2}$Te$_{3}$ family and graphene. These
parameters were fixed during the fitting of the data shown in Fig. 5.
The only parameter that could vary in our calculations was $g_{s}$. Note
that the frequency of oscillations $F$ (and, thus, the Fermi energy
$E_{F}$) is essentially determined by the periodicity of the observed
oscillations. Table II summarizes the parameters thus obtained. The
results of our calculations are shown in Fig. 5 by solid lines. Dashed
lines depict the behavior expected for ideal Dirac cones
($m/m_{0}$=$\infty$) and negligible Zeeman coupling ($g_{s}$ = 0) for
the TI data .
One can clearly see in Fig. 5 that only graphene shows the ideal behavior in
the LL fan diagram: a straight line that crosses the $n$-axis at 0.5.
All TI materials, despite their essentially Dirac-like nature of the
surface state, present the LL fan diagrams that deviate from the ideal
behavior. (The deviations from the dashed lines are most clearly
seen in strong magnetic fields.)
In view of the good agreements between the data and the fittings for
all the materials analyzed in Fig. 5, one may conclude that the advanced
analysis considering both the curvature of the Dirac cone and the Zeeman
effect can reasonably describe the SdH-oscillation data obtained for
TIs and confirm the Dirac nature in their surface states.
\section{Summary}
We derived the formula for the phase $\gamma$ of the SdH oscillations
coming from the surface Dirac fermions of realistic topological
insulators with a non-ideal dispersion given by Eq. (1). We also
calculated how the curvature in the dispersion as well as the effect of
Zeeman coupling affect the Landau-level fan diagram of the SdH
oscillations for realistic parameters. Finally, we demonstrate that the
Landau-level fan diagrams obtained from recently reported SdH
oscillations in topological insulators can actually be understood to
signify the essentially Dirac nature of the surface states, along with a
relatively large Zeeman effect in those narrow-gap materials.
\begin{acknowledgments}
We thank G.P. Mikitik for helpful discussions.
This work was supported by JSPS (NEXT Program), MEXT
(Innovative Area ``Topological Quantum Phenomena" KAKENHI 22103004),
and AFOSR (AOARD 10-4103).
\end{acknowledgments}
|
\section{Introduction}
\indent
The microarray technology allows simultaneous measurements of
messenger RNA level of thousands of genes, and its adoption dramatic
changes the way biological and biomedical research
is carried out \citep{schena,young,butte,slonim,stoughton,trevino}.
In particular, the more labor-extensive real-time PCR can be
replaced by microarray profiling in a preliminary round,
as the general agreement between the two methods is considered to be good
\citep{etienne,dallas,morey}. As an emerging technology,
there are still many issues to be worked out, such as the
consistency among different platforms
\citep{park,larkin,irizarry,draghici,kuo,patterson,chen},
batch effect \citep{churchill,baggerly,kitchen},
level of noise \citep{ioannidis,eindor}, limit of dynamic range \citep{sharov}, etc.
However, with better probe design \citep{yang}, better data quality
control \citep{shi,shi06}, better data reporting requirement \citep{ioannidis09},
better normalization scheme \citep{quack,vande,fujita,steinhoff,stafford,astola},
and better understanding of the study goals, these are not insurmountable problems.
Analyzing large amount of expression data from microarray
experiments was thought as a major challenge in early days,
but this problem was over-estimated. First, the amount the data
from thousands of genes and a hundred or so samples is still much smaller
than, e.g., the data generated by whole-genome association studies
\citep{estrada}
or next generation sequencing \citep{schadt}, and a moderately sized computer
might handle the data without problems. Second, no brand new statistical
learning methods had to be re-invented and existing machine techniques
could already extract meaningful information from the data \citep{hastie}.
Third, the problem of larger number of false positives due
to the large number of genes being profiled has been addressed
and properly handled \citep{storey,storey-q,reiner,pawitan}.
Fourth, in using multiple genes in constructing classifier,
the well known ``large $p$, small $n$" problem (large number of
variables with small number of sample size) can be solved
by the variable/subset/feature/model selection techniques
\citep{xing,liyang,ambroise,guyon,li,liao,zhao}
One of the most common applications of microarray is
``differential expression" profiling: finding mRNAs/genes
whose expression level to be very different under two conditions,
e.g., with disease and being healthy. Not only could differentially
expressed genes provide insight to the biological processes
involved in disease etiology, but also these can be used
as biomarkers for diagnosis \citep{golub,hedenfalk,dhan,adib,yeatman}
or prognosis \citep{pomeroy,vij,colman,kim}.
The phrase ``differential expression" means that the {\sl averaged} expression
level of a mRNA/gene in one phenotype-specific group is much
{\sl larger} or {\sl smaller} than that in another group. However, the
terms ``{\sl average}" and ``{\sl larger/smaller}" are up to various
interpretations.
There are at least two definitions of average: arithmetic mean
or geometric mean. For a random variable $x$, arithmetic mean
can be represented by $E[x]$, $\langle x \rangle$, or $\overline{x}$, which
is equal to $\frac{1}{n} \sum_{i=1}^n x_i$ (where $n$ is the
sample size). Geometric mean is defined by
$(x_1 x_2 \cdots x_n)^{1/n}$. For fluorescence-light-intensity
based microarray data $x$, it is a common practice to
logarithmically transform the data $x'=\log(x)$, because
$x'$ fits better than $x$ to a normal distribution. Then arithmetic
mean of $x'$ is actually equal to the logarithm of geometric
mean of $x$: $E[x']= \frac{1}{n} \sum_{i=1}^n
\log(x_i) = \log (x_1 x_2 \cdots x_n)^{1/n}$.
Deciding ``how larger one group's average is compared to
the other" is no less trivial. Fold change and $t$-statistic
are the two main choices for measure of differential expression.
In microarray analysis field, these two measures have been
in and out of favor at various time. Fold change had been
commonly used before it was pointed out that it did not
take the noise into account \citep{chen97,baldi}. $t$-statistic
enjoyed its acceptance until another round of papers suggesting
that genes selected by fold-change are more consistent among
different microarray platforms than those selected by $t$-statistics
\citep{shi05,shi06,guo}. This result triggered more comments
on the relationship between reproducibility and accuracy,
and between biological and statistical signal \citep{witten}.
If we recognize that both fold change and $t$-statistic
have advantages and shortcomings, then both pieces of
information should be used in an analysis. The problem with
fold change is that the same fold change value will be
less impressive if the variance is large. Although $t$-statistic
aims at taking the noise level into account, the practical
problem is that the variance may not be estimated reliably,
especially when the sample size is small. Volcano plot,
the topic of this review, is a visual tool to display both
fold change and $t$-statistic.
This article is organized as follows:
Section 2 establishes a relationship between the fold-change
and $t$-statistic; Section 3 introduces volcano plots
which simultaneously display both fold-change and
$t$-statistic; Section 4 introduces the modified $t$-statistic
which tends to reduce the gene-to-gene variation of variance
estimation; and Section 5 is the discussion section.
One microarray dataset is used throughout this paper,
which consists of 37 chronic lymphocytic leukemia (CLL)
samples and 17 control samples. The expression profiling
has been carried out on Illumina platform with 48804 probesets.
\begin{figure}[ht]
\begin{center}
\begin{turn}{-90}
\epsfig{file=fig1-histogram-illustrate-log.eps, height=15cm}
\end{turn}
\end{center}
\caption{
\label{fig1}
Histogram of expression levels of a microarray experiment
(this unpublished dataset contains 37 case samples, 18 control
samples, and 48804 probesets in Illumina platform, normalized
by ``quantile normalization"):
(A) in linear scale.
(B) $x$-axis in a log scale.
(C) for log-transformed expression.
}
\end{figure}
\begin{figure}[ht]
\begin{center}
\begin{turn}{-90}
\epsfig{file=fig2-fc1-fc2.eps, height=13cm}
\end{turn}
\end{center}
\caption{
\label{fig2}
Comparison of two definitions of fold changes. The $x$ is
FC defined in Eq.(\ref{eq-fc1}), in log scale. The $y$
is the $\log(FC')$ defined in Eq.(\ref{eq-fc2}).
}
\end{figure}
\section{Fold change and $t$-statistic: signal and signal-to-noise ratio}
\indent
Fold change (FC) and $t$-statistic seem to be two very different quantities:
one is intuitive and a straightforward measure of differences, another
is rooted deeply in field of statistics. However, with logarithm transformation
there is a relationship between the two.
The need for logarithmic transformation can be illustrated by Fig.\ref{fig1}.
Fig.\ref{fig1} shows the three histograms of fluorescence-light intensity
$E$ of a microarray experiment which is indicative of
the number of mRNA copies hybridized to the probe, thus
a measure of mRNA expression level: (A) in regular scale, (B) in log-transformed
$x$-axis scale, and (C) of $\log(E)$ itself. Without the logarithmic
transformation, the distribution of $E$ is very long-tailed, and very
skewed (asymmetric). With the log transformation (or other similar
transformations), even though the distribution is still not a perfect
normal distribution, it is much more normal-like.
There are other advantages of a log transformation, e.g. variance is more stablized
and does not tend to increase with the mean; it is consistent with
a psycho-physics law relating human sensation to the logarithm of
the stimulus level \citep{fechner}. Note that in a future technology
where the number of copies of mRNA can be read directly without
fluorescence light intensity as the intermediate \citep{geiss,robinson}, the
role of log transformation might be reconsidered. However, the decision
on whether to log transform or not is still based on the histogram of
$E$ vs. that of $\log(E)$.
The simplest definition of FC is:
\begin{equation}
\label{eq-fc1}
FC = \frac{\langle E_1 \rangle}{ \langle E_0 \rangle},
\end{equation}
where the arithmetic average is over the fluorescence-light intensity of samples
in group 1 (e.g. diseased group) and group 0 (e.g. control group).
The logarithm of FC is:
\begin{equation}
\log(FC)= \log \frac{\langle E_1 \rangle}{ \langle E_0 \rangle}
= \log \langle E_1 \rangle - \log \langle E_0 \rangle
\approx \langle \log E_1 \rangle - \langle \log E_0 \rangle.
\end{equation}
Switching the order of averaging and log-transformation usually
does not lead to identical values, so the above expression is
only approximately true. We can have a second definition of FC
called FC':
\begin{equation}
\label{eq-fc2}
\log(FC') = \langle \log E_1 \rangle - \langle \log E_0 \rangle
\end{equation}
Fig.\ref{fig2} shows that FC is mostly similar to FC' and
we do not distinguish the two definitions. The same conclusion
is also reached in \citep{witten}.
The $t$-test is an example of statistical testing whose goal is
to compare any observed result with chance events. The statistic
used in $t$-test (e.g. \citep{snedecor}) is the difference of
arithmetic means in two groups divided (``standardized") by
the estimated standard deviation of that difference. Standard
deviation of parameters (e.g., sample mean, sample variance)
is often called ``standard error" (SE) \citep{snedecor}. One requirement
for using $t$-test is that values in two groups roughly follow
normal distributions (with different means). As discussed above,
we need to log transform the fluorescence light intensity $E$ to have a
normal-like distribution, so $t$-statistic is:
\begin{equation}
\label{eq-t}
t= \frac{ \langle \log E_1 \rangle - \langle \log E_0 \rangle}
{SE_{\langle \log E_1 \rangle -\langle \log E_0 \rangle}}.
\end{equation}
The commonly used estimation of $SE$, due to Welsh \citep{welsh},
assumes different variances in group 1 and group 0:
\begin{equation}
\label{eq-welsh}
t_{welsh} = \frac{ \langle \log E_1 \rangle - \langle \log E_0 \rangle}
{ \sqrt{\frac{s_1^2}{n_1} + \frac{s_0^2}{n_0} }}
\end{equation}
where $s_1^2$ and $s_0^2$ are the estimated variances (of $\log(E)$)
of group 1 and 0, and $n_1$, $n_0$ are number of samples in the two groups.
For readers who are not familiar with statistics, the following
points can be used to understand the pooled estimation of
$SE$ in the denominator of Eq.\ref{eq-welsh}:
(1) $SE$ is the square root of variance and variance is the square of $SE$;
(2) variance of sum or difference of two variables, $Var[x_1 \pm x_2]$,
is the sum of individual variances $Var[x_1] + Var[x_2] $;
(3) $SE$ of sample means is sample standard deviation divided by $\sqrt{n}$
\citep{snedecor}.
The last point can be particularly hard to grasp for biologists,
but we are dealing with two different types of mean and standard
deviations here. For a dataset with $n$ samples, the mean,
standard deviation, variance are $E[x]$ (or $\langle x \rangle$, $\mu$),
$Sd[x]$ (or $\sqrt{Var[x]}$, $s$), $Var[x]$ (or $s^2$). When
the dataset is hypothetically replicated many times, we can
talk about mean, standard deviation, variance of the sample-mean $\mu$,
as each replicate of the dataset may not be exactly the same.
The dataset-mean is the same as sample-mean, but the dataset-standard-error
is $s/\sqrt{n}$ and dataset-variance is $s^2/n$ \citep{snedecor}.
This section establishes a relationship between $\log(FC)$
and $t$-statistic: $t$ is $\log(FC)$ (or more accurately, $\log(FC')$)
standardized by the noise level as measured by the pooled standard
error. In the field of statistical behavioral science,
quantitative psychology, epidemiology, and meta-analysis,
there is a similar theme of unstandardized vs. standardized effect
size \citep{cohenbook}. In the field of engineering, quantities
like $t$ can be called a signal-to-noise ratio (another definition
of signal-to-noise ratio is based on power ratio, thus a square
operation is applied). In the field of applied probability,
mean divided by standard deviation is the inverse of coefficient of variation.
\begin{figure}[th]
\begin{center}
\begin{turn}{-90}
\epsfig{file=fig3-volcano-plot.eps, height=13cm}
\end{turn}
\end{center}
\caption{
\label{fig3}
(A) $x$-axis: $t$-statistic, $y$-axis: $-\log_{10}(p$-value)
of $t$-test.
(B) Volcano plot using $t$-statistic in $y$-axis ($x$-axis is
log$_{10}$FC.
(C) Volcano plot using $-\log_{10}(p$-value) in $y$-axis.
}
\end{figure}
\section{Volcano plot and its basic use}
\indent
If the noise level is known or can be reliably estimated, we of
course prefer the measure of differential expression that
takes the noise level into account, such as $t$-statistic.
In reality, not only is smaller sample sizes an issue
for variance estimation, but also, if systematic error exists,
we may not improve the situation by increasing the sample size.
For example, it is observed that noise level during the hybridization
stage is much higher than that during the sample preparation
or amplification stage \citep{tu}. If a probe sequence for
an mRNA is highly represented in the genome, cross-hybridization
can be a cause of error and variation. However, the probability
of this error does not seem to decrease with large sample sizes.
Facing this reality, we might just display and use both FC and
$t$-statistic, and this is the volcano plot. Volcano plot
most often refers to the scatter-plot with $-\log_{10}(p$-value)
from the $t$-test as the $y$-axis and (log$_{10}$)FC as the $x$-axis
\citep{gibson,cui1}. However, $t$-statistic and $-\log_{10}(p$-value)
is (see Fig.\ref{fig3}(A)) is highly correlated, and whether
the $t$ (Fig.\ref{fig3}(B)) or $-\log_{10}(p$-value) (Fig.\ref{fig3}(C))
is used in the $y$-axis, the outcome is very similar.
The reason why $t$ and $p$-value from $t$-test is not
one-to-one corresponding (Fig.\ref{fig3}(A)) is because
in determining $p$-value, Welsh's $t$ distribution has
a degree of freedom parameter which also depends on the data
\citep{pan}.
\begin{figure}[th]
\begin{center}
\begin{turn}{-90}
\epsfig{file=fig4-volcano-with-lines.eps, height=13cm}
\end{turn}
\end{center}
\caption{
\label{fig4}
Illustration of the double filtering criterion (upper-left and upper-right
corners delineated by black lines), FC-only single-gene criterion
(lower-left and lower-right coners delineated by red lines),
and $t$-test-only single-gene criterion (``football goalpost" in the middle
delineated by red lines).
}
\end{figure}
\begin{figure}[th]
\begin{center}
\begin{turn}{-90}
\epsfig{file=fig5-illustration.eps, height=13cm}
\end{turn}
\end{center}
\caption{
\label{fig5}
(A) a gene with very good $t$-test result ($p$-value = $7.7 \times 10^{-17}$)
but only moderate fold-change (FC=1.38).
(B) a gene with large fold-change (FC=2.66) but moderate $t$-test result
($p$-value= 3 $\times 10^{-3}$).
}
\end{figure}
The basic use of volcano plots is to check genes that could be
selected by one differential expression criterion but not the other.
The familiar ``double filtering" \citep{zhang} used by many
groups is to set the gene selection criterion by: (i)
$|\log_{10}FC| > \log_{10}FC_0$; and (ii) $t > t_0$.
Equivalently, it can be defined as (i) $|\log_{10}FC| > \log_{10}FC_0$;
and (ii) $p-$value $< p_0$. FC$_0$, $t_0$, $p_0$ are preset
threshold values for fold change, $t$-statistic, and $t$-test $p$-value.
The double filtering criterion corresponds to a cutting of two
outer rectangular regions in the volcano plot (Fig.\ref{fig4}).
The single filtering criterion, after removing the
double criterion selected genes, corresponds to rectangular
regions along the two axes (Fig.\ref{fig4}). These are
often the genes not selected for reasonable arguments:
(i) genes with large fold change but nevertheless insignificant
test result may be caused a few outliers with very large
values in one group. (ii) genes with good test result
(large $t$'s and small $t$-test $p$-values) but low fold
change could be false signal due to low variance, which
can be caused by batch effect \citep{leek}, or low expression
level (to be discussed later).
The goal of using double filtering criterion is to
obtain a more robust result. The cost we pay is that
some real differentially expressed genes might be missed.
Volcano plot allows us to pick some genes from the
single filtering region for further examination.
Fig.\ref{fig5} shows two examples of genes selected by
single filtering criterion.
Fig.\ref{fig5}(A) is gene selected by $t$-test $p$-value only
($p= 7.7 \times 10^{-17}$) while FC is lower than 2
(FC=1.379). If the true variance is indeed low and we estimated
it correctly from 17 control samples, then we trust that
this gene is significantly differentially expressed.
Fig.\ref{fig5}(B) is selected by FC only (FC=2.66) whereas
the $p$-value is only $3 \times 10^{-3}$. This gene can
still be a significantly differential-expression if
the large variance in the case group is due to something
else, e.g. sub-disease types. Statistical analysis alone
should not be the only foundation for selecting potentially
relevant genes, and volcano plot is a way to pick those
genes which otherwise might be missed.
\section{Robust variance estimation, regularization, SAM, and joint filtering}
\indent
The essential difference between FC and $t$-statistic
is the consideration of statistical noise (variance), but
the real challenge is how to estimate the variance from
a small number of samples. Since variance is calculated around the
mean which is also estimated, one idea for robust variance
estimation is to iteratively remove outliers then calculate
mean and variance \citep{igor}. The drawback of this approach
is that the number of samples used is gradually reduced.
Another idea for robust variance estimation is motivated by
the typical ``large $p$ small $n$" situation for a microarray
experiment \citep{liyang}. Though the sample size $n$ could
be small, the number of genes $p$ is nevertheless large,
and that large number of genes make it possible for a reliable
estimation of common variance cross all genes \citep{pan,cui},
at least for the control group.
One main worry about variance estimation is that its value can
be low due to the low expression level. To avoid the estimated
variance being too low, we may add a constant ``penalty"
term $s_0$ to the sample-estimated standard deviation \citep{sam}:
\begin{equation}
\label{eq-sam}
t_{sam} = \frac{ \langle \log E_1 \rangle - \langle \log E_0 \rangle}
{ \sqrt{\frac{s_1^2}{n_1} + \frac{s_0^2}{n_0} } +s_0}.
\end{equation}
The penalty is also called ``regularization", reflecting the
prior belief (in the Bayesian framework) that variance estimation
across different genes should exhibit certain smooth behavior
\citep{baldi,hastie}.
A popular software package called SAM (Statistical Analysis of Microarray)
({\sl http://www-stat.stanford.edu/\~{}tibs/SAM/}) is based on Eq.(\ref{eq-sam}).
Another {\sl R} ({\sl R} is a free software environment for statistical computing:
{\sl http://www.r-project.org/}) implementation of the same idea, {\sl siggen}, can
be found at {\sl http://www.bioconductor.org/packages/2.3/bioc/html/siggenes.html}.
In SAM, detailed procedures are proposed to determine the $s_0$ value
from the data. It is not clear whether this procedure is unique
or it is just one of many options. In practice,
any small value of $s_0$, such as the 5\% percentile of standard
deviations of all genes, can stabilize the variance estimation.
A Bayesian derivation of the extra term in variance estimation
is derived in \citep{baldi}. In this framework, mean, variance
of a normal distribution (of $\log(x)=x'$) has a prior distribution,
as well as a posterior distribution after data are observed. For convenience,
they pick a functional form of prior distribution so that posterior
will have the same functional form (inverse Gamma distribution for
the variance, normal distribution for the mean). It can be shown
that (the mean of) posterior variance is a weighted sum
of prior variance ($\sigma_0^2$) and the sample-estimated
of variance $s^2$ \citep{baldi}:
\begin{equation}
\label{max_post_var}
\sigma^2_{mean.of.posterior} = w s^2 + (1-w) \sigma_0^2
\end{equation}
where weight $w$ tend to close to 1 for larger sample size
($w= (n-1)/(\nu_0 +n-2)$, $\nu_0$ is the prior degree of freedom
for the inverse Gammar distribution).
The modified or regularized variance $\sigma^2_{mean.of.posterior}$
in Eq.(\ref{max_post_var}) has the effect of drawing gene-specific
variance towards the middle, since the change from the estimated
variance:
\begin{equation}
\sigma^2_{mean.of.posterior}- s^2 = w s^2 + (1-w) \sigma_0^2 - s^2
= -(1-w) (s^2- \sigma_0^2),
\end{equation}
is negative when $s^2 > \sigma_0^2$ and positive when
$s^2 < \sigma_0^2$.
Note that variance is added in Eq.(\ref{max_post_var}),
as versus standard deviations being added in the denominator
in Eq.(\ref{eq-sam}). However, the idea of adding an extra
constant term is the same in Eq.(\ref{eq-sam}) and
Eq.(\ref{max_post_var}).
In fact, there is a second extra term in variance estimation
if the sample-estimated mean is not a good estimate of the
true mean \citep{baldi}. For this reason, it is reasonable
to consider removing outliers to make sure the mean is
estimated robustly \citep{igor}.
What is the relationship between robust variance estimation
and volcano plots? FC can be considered to be the special case
when variances of all genes are equal, $t$-statistic of course
contains gene-specific variance, and $t_{sam}$ in Eq.(\ref{eq-sam})
is somewhere in-between. Rewrite
$|\langle \log E_1 \rangle - \langle \log E_0 \rangle|$
as $\delta$, $\sqrt{ s_1^2/n_1 + s_0^2/n_0}$ as $s$,
the regularized $t$-statistic in Eq.(\ref{eq-sam}) can be
split into two terms \citep{zhang}:
\begin{equation}
\label{eq-sam-weighted-sum}
t_{sam}= \frac{\delta}{s+s_0}
= \frac{1}{2(s+s_0)} \delta + \frac{s}{2(s+s_0)} \frac{\delta}{s}
= \frac{1}{2(s+s_0)} |\log(FC')| + \frac{s}{2(s+s_0)} t_{welsh}.
\end{equation}
In other words, $t_{sam}$ is a weighted sum of $\log(FC')$ and $t$-statistic.
\begin{figure}[th]
\begin{center}
\begin{turn}{-90}
\epsfig{file=fig6-sam-lines.eps, height=14cm}
\end{turn}
\end{center}
\caption{
\label{fig6}
All lines correspond to the constant $t_{sam}=5$ value (Eq.(\ref{eq-sam}))
with $s_0=0.0266$ being the 5\% percentile standard deviation
of all 48804 probes/genes. The six lines are for genes with different
standard deviations: $s=0.0238, 0.0266, 0.0283, 0.0425, 0.0871, 0.137$
(1\%, 5\%, 10\%, 50\%, 75\%, 90\% percentiles, pink, red, brown, purple, green
and blue).
}
\end{figure}
A constant $t_{sam}$ value corresponds a line in the volcano plot:
$w |x| + v y = t_{0,sam}$, where $w= 0.5/(s+s_0)$, $v=0.5s/(s+s_0)$.
Gene selection criterion by SAM (Eq.(\ref{eq-sam}) and Eq.(\ref{eq-sam-weighted-sum}))
is $t_{sam} > t_{0,sam}$. Because each gene has its own standard
deviation value $s$, the threshold can be gene-specific. We illustrate
this important property of SAM in Fig.\ref{fig6}. The $s_0$ is
set at 0.0266 which is the 5\% percentile value of $s$'s of all genes
in our dataset. For a gene with standard deviation of $s=0.0238, 0.0266,
0.0283, 0.0425, 0.0871, 0.137$ (1\%, 5\%, 10\%, 50\%, 75\%, 90\%
percentiles), the $t_{sam} = 5 $ threshold is represented by lines
with various slopes (pink, red, brown, purple, green, blue in
Fig.\ref{fig6}).
The lines for low-variance genes have steeper slopes, indicating
that FC plays a more important role in differential expression
gene selection. On the other hand, for high-variance genes,
the threshold lines have flatter slope, indicating that $t$-test
result is more important. As discussed in the previous section
(and Fig.\ref{fig5}), low-variance genes tend to have low FC values
and high-variance genes tend to have less significant test
result, so the consequence of using SAM is to counter-balance this
trend and to obtain a more robust outcome. We also note that the SAM-based gene selection
regions in Fig.\ref{fig6} (complementary to triangles) are
very different from those by double-filtering criterion
(rectangles in Fig.\ref{fig5}). This can also be called a ``joint
filtering" criterion \citep{zhang}.
\begin{figure}[th]
\begin{center}
\begin{turn}{-90}
\epsfig{file=fig7-top100.eps, height=17cm}
\end{turn}
\end{center}
\caption{
\label{fig7}
Green, red, green, black dots are the top 100 probes/genes selected
by $t_{sam}$, $FC'$, $t$-statistic, and $p$-value of $t$-test.
(A) on volcano plot, $x$: log(FC'), $y$: $-\log_{10}(p$-value).
(B) $x$: mean of all samples, $y$: standard deviation of all samples.
(C) $x$: mean of control samples, $y$: standard deviation of control samples.
(D) $x$: mean of diseased samples, $y$: standard deviation of diseased samples.
}
\end{figure}
Fig.\ref{fig7}(A) compares the top 100 genes selected by
SAM (regularized $t$) (blue) with those selected by FC (red),
$t$-test $p$-value (black), and $t$-statistic itself (green).
Although there are certain overlaps among different selection
criteria, SAM is able to pick up genes that are not selected
by either FC or $t$-test $p$-value alone.
To address another question on whether $t$-test criterion
tends to select genes with low variance and low expression level.
Fig.\ref{fig7}(B)(C)(D) show the standard deviation ($y$-axis)
vs. mean ($x$-axis) for all samples, control samples only,
and diseased (CLL) samples only. Indeed, FC-based criterion
tend to select genes with high variances, whereas $t$-test based
criterion selects relatively low variance genes. SAM achieves
a balance between the two criteria, and selects genes with
intermediate variance values. On the other hand, there is
no strong evidence that any selection criterion tends to select low
expression level genes.
\begin{figure}[th]
\begin{center}
\begin{turn}{-90}
\epsfig{file=fig8-stratify.eps, height=17cm}
\end{turn}
\end{center}
\caption{
\label{fig8}
Stratified volcano plot: probes/genes on chromosome 6 are marked
by red, and those with ``cytokine" in gene annotation is marked
by blue.
}
\end{figure}
\section{Discussion}
\indent
The idea and the use of volcano plots can be expanded in several directions.
First of all, as a 2-dimensional plot, with potentially interesting
genes scattered outward, one can examine any external information
by introducing colors. If that external piece of information is
relevant to differential expression, we can easily recognize
the fact by a visual impression of the plot. This coloring of
a volcano plot can be called ``stratified volcano plot". One
example is to label all probes/genes that belong to a particular
pathway, cellular component, function, or process coded in
GO (gene ontology) categories \citep{go}.
Fig.\ref{fig8} illustrates a stratified volcano plot by marking
1614 probes/genes that are located on chromosome 6 (red),
and 31 probes/genes whose annotation contains the word ``cytokine".
From the stratified volcano plot, we can easily identify
interesting candidate genes involving cytokines such
as CLCF1 (cardiotrophin-like cytokine factor 1, $p$-value= 1.4$\times 10^{-16}$,
FC'=0.22, down-regulated), SOCS2 (suppressor of cytokine signaling 2,
FC'= 0.11, $p$-value= 3.8$\times 10^{-8}$, down-regulated),
SOCS3 (suppressor of cytokine signaling 3, FC'=0.28,
$p$-value= 6.2 $\times 10^{-8}$, down-regulated), etc.
In a work-in-progress, we have developed an $R$ package to color
a volcano plots using the average expression levels \citep{hua}.
In the program, we introduced an interactive feature for users
to click a probe/gene on the volcano plot to show the gene names
or other information.
Secondly, the idea of simultaneously display of noise-level-standardized
signal and unstandardized one can be useful beyond the microarray field.
In genetic association studies, the association signal of
a single-nucleotide polymorphism (SNP) is usually measured by
two quantities. One is the odds-ratio (OR) of the 2-by-2 count table
with disease status as row and two alleles as column. OR is not
standardized by the noise level or sample size, though the 95\%
confidence interval of OR does become narrower for larger sample
sizes thus lower level of chance events \citep{li-bib}. On the
other hand, the chi-square statistic or the $p$-value of the
chi-square ($\chi^2$) test strongly dependent on sample size,
thus chance event probability. In fact, the chi-square statistics
is proportional to the total number of samples for a SNP that
contains association signal.
Besides using OR in $x$-axis (in log scale), another choice is
to use the allele frequency difference in case and control group.
Denote the four counts in the 2-by-2 table in case-control
association analysis are $a,b,c,d$, $\log_{10}$OR is
$\log_{10}(ad)-\log_{10}(bc)$, whereas allele frequency difference
is $a/(a+b)- c/(c+d)= (ad-bc)(a+b)^{-1}(c+d)^{-1}$. In other
words, the difference between the two choices is whether
$ad$ and $bc$ are compared in the logarithmic or regular scale.
It is rare to find a genetic association paper that applies
volcano plots \citep{sirota,miclaus}. We believe that many extensions
and applications of volcano plots in microarray analysis
can be equally useful in genetic association analysis.
For example, the joint filtering criterion, the stratified
volcano plot coloring external pieces of information,
and uncovering of systematic patterns when the colorings
are on other statistical information. We have found that
the location of a SNP on the volcano plot is intrinsically
related to its minor allele frequency. This will provide further
insight on how one should balance the chi-square test result
and odds-ratio in selecting genetically associated genes.
In conclusion, volcano plot displays both noise-level-standardized
and unstandardized signal concerning differential expression
of mRNA levels. Joint filtering has a simple geometric
interpretation in volcano plot, and its advantage over
double filter criterion of genes can be easily understood.
As a scattering plot, volcano plot can incorporate other
external information, such as gene annotation, to aid
the hypothesis generating process concerning a disease.
\section*{Acknowledgements}
I would like to thank Frank Batliwalla, Percio Gulko, Max Brenner,
Peter Gregersen for asking about differences between genes selected
by fold-change and those by $t$-test, which provided the initial
incentive for writing this review, Joy Yan, Sophia Yancopoulos
for discussion on volcano plots and allowing me to use the CLL data,
Yaning Yang, Xing Hua for collaboration on software development.
\large
|
\section{Introduction}
\label{sec:intro}
The bulk of the stellar mass observed today in galaxies is built up at
high redshift, where star formation and mass assembly are very
efficient. This makes observations at high redshift extremely
important in understanding galaxy evolution in general. For example,
the maximum value of the mean cosmological star-formation rate (SFR)
density is achieved at redshift 1--3 (e.g. Hopkins \& Beacom 2006).
Although high redshift observations are naturally plagued with larger
uncertainties than local data, it is clear that a successful model of
galaxy formation and evolution should match them within realistic
error margins. Current models are often mainly tuned to reproduce the
low redshift universe, and it is therefore very instructive to test
them against high redshift observations. For example, Fontanot et al.
(2009) and Guo et al. (2010) point out that the low mass end of the
stellar mass function is built too quickly at high redshift in current
semi-analytic models (SAMs hereafter), while Khochfar et al. (2007)
claim that their SAM can reproduce the evolution of the faint end of
the luminosity function. It seems however likely that the evolution of
the stellar mass function over time alone is insufficient to fully
constrain the models, as was argued by Neistein \& Weinmann (2010).
Consequently, it is a crucial next step to compare the observed SFR of
high redshift galaxies to models in more detail than previously done,
where the main focus was in trying to match the global star formation
rate density.
Pioneering observational estimates of the SFR ($\dot{m}_{\rm star}$)
and stellar mass ($m_{\rm star}$) indicate that the specific SFR
(sSFR, $\dot{m}_{\rm star}/m_{\rm star}$) is roughly constant in time
throughout the redshift range $z=2-7$, for galaxies with roughly the
same mass $m_{\rm star} \sim$ $(0.2-1) \cdot 10^{10}M_\odot$, at a
level ${\rm sSFR} \sim 1-2$ ${\rm Gyr}^{-1}$. (e.g. Stark et al.
2009a; Gonz\'{a}lez et al. 2010; Labb\'{e} et al. 2010a, 2010b).
There are indications that this sSFR plateau is associated with a
rather constant sSFR within each galaxy as it grows (Papovich et
al. 2010; Stark et al. 2009a). The plateau is hard to reconcile with
the current theoretical wisdom for several reasons.
In current models of galaxy formation the high-$z$ SFR is assumed to a
large extent to be driven by the fresh gas supply (see
Kere\v{s} et al. 2005;
Dekel et
al. 2009; Bouch\'{e} et al.~2010; and basically all the SAM; but
see also Narayanan et al. 2010 who discuss a potential merger origin
for the population of submillimeter galaxies at $\emph{z} \sim 2$).
The
specific cosmological accretion rate of baryons is steeply declining
with time, $\dot{M}/M \propto (1+z)^{2.5}$ (Neistein \& Dekel 2008;
Dekel et al. 2009). In particular, at $z \sim 7$ the specific
accretion rate is higher than the observed sSFR by a factor of a few,
and at $z \sim 2$ it is lower than the sSFR by a similar factor. This
is in marked contrast to the observed plateau in the sSFR. A constant
sSFR during the evolution of each galaxy (main progenitor) would
require either an exponential growth in time of $m_{\rm star}$ and
$\dot{m}_{\rm star}$ (this is if most stars are formed in situ to the
main progenitor), or a non-trivial combination of effective SFR and
stellar assembly rate as a function of time and mass. An obvious
related difficulty is introduced by the fact that a low SFR at high
$z$ could make it hard to produce enough massive galaxies by
$z\!\sim\!2$ to match the bright end of the observed galaxy mass
function at that epoch.
Given the marked contrast between the indicated observation and the
current models of galaxy formation, we appeal to a special
semi-analytic tool. Traditional SAMs (e.g Kauffmann et al. 1993; Cole
et al. 2000; De Lucia \& Blaizot 2007) describe the processes that are
responsible for galaxy evolution by physically motivated recipes that
are fixed a priori. They are thus geared to solve the `forward
problem', i.e., test to what extent the assumed set of physical
recipes provides a match to the observed properties of the galaxy
population. This methodology is not ideal for exploring a large
variety of physical recipes, some of which may need to deviate
significantly from the standard assumptions. Our approach here is to
solve the `inverse problem', where we investigate how the observations
at high \emph{z} constrain the basic recipes, especially those
associated with the processes of mergers, star formation, and
feedback. This is achievable using the method of Neistein \& Weinmann
(2010, NW10 hereafter), which allows freedom in choosing the recipes
of interest, and a very efficient exploration of a broad parameter
space.
In order to reproduce the sSFR plateau, we will try to either lower
the SFR efficiency after a very short period of high efficiency at $z
\geq 7$, or to enhance the suppression of SFR by feedback at $z \geq
4$, especially in high-mass haloes, to be followed by an enhanced SFR
in the retained or reincorporated gas at $z \sim 2-3$. In both cases,
the low sSFR at high $z$ makes it difficult to form enough massive
galaxies at $z \sim 1-3$, unless the rate of mass assembly due to
mergers and the associated starbursts are pushed to their limits.
The models presented in this paper are not a priori physically
motivated. However, we have tried to keep them simple, and to
minimize deviations both from the standard model and from a monotonic
dependence on halo mass and time. We also try to keep the models as
physically plausible as possible. Our aim in this paper is not to
find the ``right'' model. The main goal is to investigate the
approximate nature of the needed changes, and how they impact on the
properties of the galaxy population other than the sSFR.
The measurements of SFR and stellar mass at high $z$ are still at
their infancy, and therefore their interpretation, in particular the
sSFR plateau, is uncertain and highly controversial. The main source
of uncertainty is the obscuration by dust, where several authors bring
convincing arguments for little or no dust at $z \geq 4$ (e.g. Bouwens
et al. 2009; Finkelstein et al. 2010), while others do apply dust
corrections and obtain higher values of sSFR.
Stellar masses may also be affected by systematic errors. For example,
uncertainties in the treatment of the TP AGB-phase may lead to errors
in the stellar mass estimates at $\emph{z}=2-3$ by factors of $\sim$ 2 (e.g.
Maraston et al. 2006; Magdis et al. 2010a).
The disagreement
between a constant sSFR and our current theoretical wisdom is so
pronounced that it motivates a study of the theoretical implications
despite the observational uncertainties, adopting the validity of the
sSFR plateau as a working assumption. An alternative way out from the
puzzle might be to assume a time-dependent stellar initial mass
function (IMF) (e.g. Dav\'{e} 2010), but this is kept beyond the scope
of the current paper.
The outline of the paper is as follows. In section \ref{sec:problem},
we summarize the observed sSFR at high redshift and explain the points
of tension with theory. In section \ref{sec:formalism}, we explain
the NW10 method that we use here. In section \ref{sec:explore}, the
main part of this paper, we demonstrate how a simple standard SAM
fails to reproduce the sSFR plateau, and proceed by making controlled
changes to this model in order to better reproduce the observations.
In sections \ref{sec:change_quiescent} and \ref{sec:change_mergers} we
discuss possible physical motivations for these changes. In section
\ref{sec:dust} we comment on the observational uncertainties regarding
obscuration by dust. In section \ref{sec:previous_work} we compare
our results to previous work. Finally, in section
\ref{sec:conclusions}, we present our conclusions.
Throughout the paper, we refer to the redshift range $\emph{z}=2-3$ as
``intermediate" redshift, to $\emph{z}=4-6$ as ``high" redshift, and to
$\emph{z}>6$ as ``very high" redshift. Our models are based on dark-matter
merger trees from the Millennium $N$-body simulation (Springel et
al. 2005), and we thus use WMAP1 cosmological parameters. In
particular, we quote masses assuming $h=0.73$.
\section{The problem}
\label{sec:problem}
\subsection{Observations}
\begin{figure}
\centerline{\psfig{figure=fig1_wnd.eps,width=3.5in}}
\caption{The observed sSFR plateau.
Shown are measurements of specific star-formation rate as a function of
redshift for galaxies in a similar stellar
mass range $\sim (0.2-1) \cdot 10^{10} M_\odot$.
The references are marked and listed in the text.
The grey belt captures most of the measurements, and reflects
the uncertainty of $\pm 0.3$ dex estimated by Gonz\'{a}lez et al. (2010).
It indicates a constant sSFR $\sim 2$ ${\rm Gyr}^{-1}$ in the range $z = 2-7$,
followed by a steep decline toward $z=0$.
All references given in
\emph{square brackets} refer to a rather high median mass of $\sim 10^{10}
M_{\odot}$.
}
\label{fig:obs}
\end{figure}
{\bf A} correlation between stellar mass and SFR has
been observed at various redshifts
up to $z \sim$ 4 (e.g. Daddi et al. 2007; Stark et al. 2009; Karim et al.
2011).
Figure \ref{fig:obs} shows a compilation of observational estimates
of sSFR as a function of redshift for star-forming galaxies\footnote{typically
selected at high redshift as Lyman-break galaxies (LBGs, Steidel et al. 1999),
which possibly excludes a population of low SFR galaxies
(e.g. Richards et al. 2011). This will however
only increase differences between
model and basic theoretical predictions.
}
of a similar stellar mass $\sim (0.2-1) \cdot 10^{10} M_{\odot}$.
The data reveal a rather constant sSFR $\sim$ 2 ${\rm Gyr}^{-1}$
in the redshift range $z=2-7$ (Feulner et al. 2005; Yan et al. 2006; Eyles et al. 2007; Daddi et al. 2007; Stark et al. 2009; Gonz\'{a}lez et al. 2010;
Labb\'{e} et al. 2010a,b; Rodighiero et al. 2010;
Gonz\'{a}lez et al. 2010; Magdis et al. 2010a,b;
McLure et al. 2011).
except for three
higher estimates (Yabe et al. 2009; Schaerer \& de Barros
2010, Shim et al. 2011)\footnote{We obtained part of the estimates by dividing the median
SFR by the median stellar mass (for Yan et al. 2006; Eyles et al. 2007;
Yabe et al. 2009). In some other cases, the estimates are based on an
extrapolation of the sSFR-stellar mass relation to a stellar mass
of $\sim 0.5 \cdot 10^{10} M_{\odot}$ (Daddi et al. 2007; Rodighiero et al.
2010).}.
The main reason for these higher estimates is the larger correction
for dust extinction assumed by these authors (to be discussed in
section \ref{sec:dust}), as well as different treatments of nebular emission
lines and different assumptions concerning star-formation histories
(to be discussed in section \ref{sec:individual}). Additionally,
Shim et al. (2011) only include galaxies with indications for H$\alpha$
emission, which will
bias the estimate of the sSFR high. We note that
all the estimates above do not include submillimeter galaxies, which are outliers
to the relation between stellar mass and SFR, simply because these tend
to have stellar
masses above the limit we consider here (e.g. Daddi et al. 2007).
At $z<2$, the
sSFR declines steeply (Noeske et al. 2007; Elbaz et al. 2007; Dunne et al. 2009; Oliver et al. 2010;
Rodighiero et al. 2010).
The grey belt in Fig. \ref{fig:obs} tries to capture the overall trend,
reflecting an uncertainty of $\pm 0.3$ dex as estimated by
Gonz\'{a}lez et al. (2010), and ignoring the two high estimates.
As will be discussed below, this observed sSFR plateau is puzzling ---
its level is surprisingly high at $z\sim 2$ and surprisingly low at $z >4$.
For the purpose of the theoretical analysis of the current paper,
we adopt the sSFR plateau as marked by the grey belt.
\subsection{Tension with theory}
Here, we outline the main potential points of tension between the
observed sSFR and theoretical predictions both from relatively
detailed SAMs and simple analytical arguments.
\subsubsection{Tension with SAMs}
\begin{figure}
\centerline{\psfig{figure=fig2_wnd.eps,width=3.5in}}
\caption{
Evolution of sSFR in the SAMs of NW10,
and of the specific dark matter accretion rate.
Shown are four different models (I, II, III and V, with results
of models II and V being indistinguishable and thus represented
by one line) by the SAM of NW10
for galaxies in the mass range [$2\cdot10^9$, $10^{10}$] $M_\odot$
(curves in colour).
The completeness limits trying to mimic the observed ones are described in
the text.
In all models the sSFR is steeply declining in time, not reproducing the
observed sSFR plateau marked by the grey belt from Fig.~1.
Also shown is the specific dark matter
accretion rate onto haloes
of log$(M_{\rm halo}) \sim 10^{12} M_{\odot}$
according to Neistein \& Dekel (2008) (dashed black line),
and the same quantity multiplied by a factor of 2, to account for the effect
of instantaneous mass loss from newly formed stars, as assumed in the
models (solid black line).
}
\label{fig:nw10_ssfr}
\end{figure}
\begin{figure}
\centerline{\psfig{figure=fig3_wnd.eps,width=3.5in}}
\caption{Evolution of sSFR in variations
of the standard SAM used in this work, and of the
specific dark matter accretion rate.
The predicted dark matter accretion rate onto haloes
with log$(M_{\rm halo}) \sim 10^{12} M_{\odot}$
according to Neistein \& Dekel (2008) (dashed black line) and multiplied
by a factor of 2 (black line), compared to the standard
model (blue dashed line,
described in section \ref{sec:standard}), the standard model
without feedback (red solid line), and the standard model
without feedback and stellar mass loss
(green dot-dashed line).
The sSFR in the raw model without feedback and mass loss matches the total
specific accretion rate. The mass loss adds a factor of two.
}
\label{fig:DMrate}
\end{figure}
In Fig.~\ref{fig:nw10_ssfr}, we show the sSFR as a function of redshift
at a fixed mass for four of the models presented in NW10, in comparison
with the
observed sSFR plateau.
To account for the observational
completeness limit as indicated by Stark et al. (2009), we only take
into account model galaxies with log(SFR) $>$ 0.25, 0.45 and 0.5
$\rm{yr}^{-1}$ at z=4, 5 and 6 respectively. The
four models are
described in detail in NW10 with the same numbering used here
and can be summarized as follows: I) model
without SN feedback, II) model without ejective SN feedback, III)
model including cold accretion [which is the model most similar to
other standard SAMs], and
V)
model in
which cooling and star formation shuts down after major mergers.
All of those models have been tuned to reproduce key observables like
the stellar mass functions at different redshifts, and star formation
rate at z=0.
Remarkably, all models show an extremely similar
behaviour despite their fundamental differences, all in disagreement
with observations. They overestimate the sSFR by about an order of
magnitude at $z \sim 6$, and underestimate it by around 0.3 dex at $z
\sim 2$.
Other current SAMs show a similar behaviour. For example, Lacey et al. (2010) show that the
sSFR of galaxies in the Baugh et al. (2005) model have
sSFR $>10$ ${\rm Gyr}^{-1}$ at z=6, an order of magnitude higher
than observational results. Daddi et al. (2007)
indicate that their observed SFR at $z \sim 2$ is significantly higher
than the values predicted by the model of Kitzbichler \& White (2007).
Finally, we have confirmed ourselves that the model of
De Lucia \& Blaizot (2007) predicts results very similar to our model
predictions shown in Fig.~2 (see also Guo \& White 2008, their Fig. 3).
\subsubsection{Tension with basic theoretical considerations}
\label{sec:tension1}
The observed sSFR plateau is in disagreement with the standard
wisdom concerning galaxy evolution.
First, the average specific
accretion rate into dark-matter haloes of a given
mass is rapidly increasing with redshift, roughly in proportion to
$(1+z)^{2.5}$ (Neistein \& Dekel 2008).
Second, galaxies are more dense and gas rich at high-redshift,
which is expected to lead to higher SFR (e.g. Dutton et al. 2010). In
agreement with these studies, Bouch\'{e} et al. (2010) find
in their idealized model that the sSFR of
individual galaxies is indeed monotonically decreasing with time.
To illustrate the first point, we show
in Fig. \ref{fig:nw10_ssfr} the approximation for the
average specific dark matter accretion rate $\dot{M}/M$
for haloes with log$(M_{\rm halo}) \sim 10^{12} M_{\odot}$
(Neistein \& Dekel 2008), and compare it to the standard model,
described in section \ref{sec:standard}, in Fig. \ref{fig:DMrate}.
Remarkably, when we remove feedback from the standard model,
it
predicts a sSFR evolution in excellent agreement with twice
the specific dark matter accretion rate. As shown in the figure,
the remaining factor of two difference
is fully explained by instantaneous
stellar mass loss.
Given that the ratio between stellar mass
and halo mass is only about 5\%
for the galaxies
considered here, this agreement is noteworthy.
We see in the figure that the feedback as implemented in the standard
model does not have a significant effect on the sSFR at $z>4$, while
it gradually reduces the sSFR at lower redshifts. Compared
to the models of NW10 shown in Fig. \ref{fig:nw10_ssfr}, the standard
model that we use in this paper
has a low sSFR at $z<2$, due to efficient feedback at late times.
This is not relevant for studying the plateau.
In what follows, we will demonstrate that despite this
serious tension with theory, the sSFR plateau can in
principle be reproduced by models of galaxy evolution, but it takes
non-negligible modifications to common ingredients of these models.
\section{The Formalism}
\label{sec:formalism}
In this section we describe the formalism we use for modeling the
evolution of galaxies. For more details about the methodology the
reader is referred to NW10. It is shown there that the results of our
model are very similar to those given by a standard SAM, although the
recipes are simplified and schematic. In the context of this work, the
simplicity of the model allows us to tune it easily, without losing
the complex interplay between different process like dark-matter
growth, cooling, SF, feedback, and merging. The code is available for
public usage through the Internet (see
\texttt{http://www.mpa-garching.mpg.de/galform/sesam})
\subsection{Merger trees}
We use merger trees extracted from the Millennium $N$-body simulation
(Springel et al. 2005). This simulation was run using the cosmological
parameters $(\Omega_{\rm m},\,\Omega_{\Lambda},\,h,\,\sigma_8)=(0.25,\,0.75,\,0.73,\,0.9)$,
with a particle mass of $8.6\cdot 10^8 h^{-1}M_{\odot}$ and a box size
of 500 $h^{-1}$Mpc. The merger trees used here are based on
\emph{subhaloes} identified using the \textsc{subfind} algorithm
(Springel et al. 2001). They are defined as the bound density peaks
inside {\scshape fof~} groups (Davis et al. 1985). More details on the
simulation and the subhalo merger-trees can be found in Springel et
al. (2005) and Croton
et al. (2006). The mass of each subhalo
(referred to as $M_{\rm h}$ in what follows) is determined according
to the number of particles it contains. Within each {\scshape fof~} group the
most massive subhalo is termed the central subhalo of this group.
Throughout this paper we will use the term `haloes' for both subhaloes
and the central (sub)halo of {\scshape fof~} groups.
\subsection{Quiescent evolution}
\label{sec:model_quiescent}
Each galaxy is modeled by a 4-component vector, %
\begin{eqnarray}
{\bf m} = \left( \begin{array}{c} m_{\rm star} \\ m_{\rm cold} \\ m_{\rm fil} \\ m_{\rm fb} \end{array}
\right) \, ,
\label{eq:m_vec}
\end{eqnarray}
where $m_{\rm star}$ is the mass of stars, $m_{\rm cold}$ is the mass of cold gas
within the disk, $m_{\rm fil}$ is the mass within cold filaments streaming
within the host halo into the central galaxy, and $m_{\rm fb}$ is the mass
currently made unavailable for star formation by stellar feedback. We
use the term `quiescent evolution' to mark all the evolutionary
processes of a galaxy, except those related to mergers.
All the models in this work assume that fresh gas is added
to a galaxy only by cold filaments, increasing the mass of
$m_{\rm fil}$. The infall rate into
filaments is assumed to be proportional to
the dark-matter growth rate, %
\begin{equation}
\left[\dot{m}_{\rm fil}\right]_{\rm accretion} = 0.17 \, \dot{M}_{\rm
h} \\
\label{eq:smooth_acc}
\end{equation}
Here 0.17 is the cosmic baryonic fraction, and $\dot{M}_{\rm h}$ is
the rate of dark-matter smooth accretion which does not include
mergers with resolved progenitors (if $\dot{M}_{\rm h}<0$ we use a gas
accretion rate of zero).
The mass of cold gas within the disk is increased due to the
free infall of cold filaments from the outer parts of the host
halo. We mimic this effect by assuming that gas joins the disk with a
specific rate $f_{\rm c}$, %
{\begin{equation} \left[\dot{m}_{\rm cold}\right]_{\rm ff}
= -\left[\dot{m}_{\rm fil} \right]_{\rm ff} = f_{\rm c} \cdot m_{\rm fil} \,.
\end{equation}
The efficiency $f_{\rm c}=f_{\rm c}(M_{\rm h},t)$ is a function of the host
halo mass $M_{\rm h}$ and the cosmic time $t$ only, and is given in
units of Gyr$^{-1}$.
We assume that the SF rate is proportional to the amount of cold
gas, %
\begin{equation}
\left[\dot{m}_{\rm star}\right]_{\rm SF} = -\left[\dot{m}_{\rm cold}\right]_{\rm SF} =
f_{\rm s}\cdot m_{\rm cold} \,,
\label{eq:sf}
\end{equation}
where $f_{\rm s}=f_{\rm s}(M_{\rm h},t)$ is a function of the halo mass and
time, in units of Gyr$^{-1}$. For each SF episode we assume that a
constant fraction of the mass is returned back to the cold gas
component due to SN events and stellar winds. This recycling is
assumed to be instantaneous, and contributes %
\begin{equation}
\left[\dot{m}_{\rm cold}\right]_{\rm recycling} = -\left[\dot{m}_{\rm star}\right]_{\rm
recycling} = R \left[\dot{m}_{\rm star}\right]_{\rm SF} \,.
\end{equation}
Following NW10, we use $R=0.5$ for all models. This
is the recycled fraction for a Chabrier (2003) IMF at 13.5 Gyr after a star burst
according to the Bruzual \& Charlot (2003) stellar population models. We note that
this is the only point where the assumption on the IMF enters our model.
Cold gas can be affected by feedback, which
means that it becomes unavailable for star formation and moves from
the cold phase to the feedback phase. Assuming that stellar feedback
immediately follows star formation, this feedback should be in
proportion to the SF rate, %
\begin{eqnarray}
\lefteqn{ \left[\dot{m}_{\rm fb}\right]_{\rm feedback} =
-\left[\dot{m}_{\rm cold}\right]_{\rm feedback} = } \\ \nonumber & &
f_{\rm d}\left[\dot{m}_{\rm star}\right]_{\rm SF} = f_{\rm d} f_{\rm s} m_{\rm cold} \,.
\end{eqnarray}
We model feedback
by a function of halo mass and time,
$f_{\rm d}=f_{\rm d}(M_{\rm h},t)$.
Once gas has been made unavailable for star
formation due to feedback, we allow it to return to the cold phase,
with a re-incorporation efficiency: %
\begin{eqnarray}
\left[\dot{m}_{\rm cold}\right]_{\rm rc} = -\left[\dot{m}_{\rm fb}\right]_{\rm rc}
= f_{\rm rc} m_{ \rm fb}.
\end{eqnarray}
Note that the feedback mechanism we use is moving gas from the cold
phase to the feedback phase. The mass within filaments is not
participating in the feedback and re-incorporation processes.
If the cold gas mass becomes negative in a given timestep due to
strong feedback, the star formation rates are adjusted such that
a cold gas mass of zero is produced.
To conclude, each process is described by one function which depends
on the host halo mass and time only. All processes discussed in this
section can be written in a compact form by using the following
differential equations: %
\begin{equation}
\dot{{\bf m}} = {\bf A}{\bf m} + {\bf B}\dot{M}_{\rm h} \, ,
\label{eq:m_evolve}
\end{equation}
where
\begin{eqnarray}
{\bf A} = \left( \begin{array}{cccc} 0 & (1-R)f_{\rm s} & 0 & 0 \\ 0 & -(1-R)f_{\rm s}
-f_{\rm d} f_{\rm s} & f_{\rm c} & f_{\rm rc} \\ 0 & 0 & -f_{\rm c} & 0 \\ 0 & f_{\rm d} f_{\rm s} & 0 &
-f_{\rm rc}
\end{array} \right)
\end{eqnarray}
\begin{eqnarray}
{\bf B} = \left( \begin{array}{c} 0 \\ 0 \\ 0.17 \\ 0
\end{array} \right) \,\, .
\label{eq:AB_defs}
\end{eqnarray}
Photoionization heating of the intergalactic medium is assumed to
suppress the amount of cold gas available for SF within low mass
haloes. This effect is critical for modeling the formation of dwarf
galaxies. The minimum halo mass of $\sim 2\cdot 10^{10}
h^{-1}M_{\odot}$ in the Millennium simulation, which we use here, does
however
not allow a detailed modeling of small mass galaxies.
Thus, instead of implementing a detailed treamtment of reionization,
we simply assume that all the gas is kept hot until redshift 9,
where cooling and SF are allowed to start. This is a
higher redshift than in NW10, which
we found is needed
to produce a high enough number of galaxies at very high
redshifts.
\subsection{Mergers and satellite galaxies}
\label{sec:model_mergers}
Satellite galaxies are defined as all galaxies inside a {\scshape fof~} group
except the main galaxy inside the central (most massive)
subhalo. Once the subhalo corresponding to a given galaxy cannot
be resolved anymore, it is considered as having merged with the
central halo. Due to the effect of dynamical friction, the galaxy is
then assumed to spiral towards the center of the {\scshape fof~} group and merge
with the galaxy in the central halo after a significant delay time.
At the last time the dark matter subhalo of a satellite galaxy is
resolved we compute its distance from the central halo ($r_{\rm
sat}$), and estimate the dynamical friction time using the formula
of Binney (1987), %
\begin{equation}
t_{\rm df} = \alpha_{\rm df} \cdot \, \frac{1.17 V_v r_{\rm
sat}^{2}}{G m_{\rm sat}\ln\left( 1+ M_{\rm h}/m_{\rm sat} \right)
} \, .
\label{eq:t_df}
\end{equation}
For $m_{\rm sat}$ we use the baryonic (stars + cold gas) mass of the
satellite galaxy plus the minimum subhalo mass which can be resolved
by the Millennium simulation. $V_v,\,M_{\rm h}$ are the virial
velocity and mass of the central subhalo. If a satellite falls into a
larger halo together with its central galaxy we update $t_{\rm df}$
for both objects according to the new central galaxy.
While satellite galaxies move within their {\scshape fof~} group, they suffer
from loss of their extended gas reservoir
due to tidal stripping. We assume that all satellite
galaxies are losing their reservoir of filament gas exponentially, on
a time scale of a few Gyr. In order to properly model this stripping
we modify ${\bf A}$ by subtracting a constant $\alpha_h$ from one of its
elements:
\begin{eqnarray}
{\bf A}_{\rm sat}(3,3) = -f_{\rm c}-\alpha_h \,.
\label{eq:ABsat_defs}
\end{eqnarray}
Note that a constant in the diagonal of ${\bf A}$ gives an exponential
time dependence. However, the actual dependence of $m_{\rm fil}$ on time for
satellite galaxies is more complicated due to contributions from
accretion. In general the parameter $\alpha_h$ should depend on the
dynamical time of the host halo. For simplicity we consider it to be a
constant here. We assume that the gas which is in the feedback
phase is not stripped.
When galaxies finally merge we assume that a SF burst is triggered. We
follow Mihos et al. (1994), Somerville et al. (2001) and Cox et
al. (2008) and model the amount of stars produced by %
\begin{equation}
\Delta m_{\rm star} = f_{\rm burst} (m_{1,{\rm cold}}+m_{2,{\rm cold}}) \, ,
\label{eq:sf_burst}
\end{equation}
where
\begin{equation}
f_{\rm burst} = \alpha_b \left( \frac{m_1}{m_2} \right)^{\alpha_c} \,
.
\end{equation}
Here $m_i$ are the baryonic masses of the progenitor galaxies (cold
gas plus stars), $m_{i,{\rm cold}}$ is their cold gas mass, and
$\alpha_b$, $\alpha_c$ are constants. %
The burst duration has been shown to vary in hydrodynamical
simulations between tens of Myr to a few Gyr depending on the merger
mass ratio, and whether multiple bursts are considered or just the
main peak (Cox et al. 2008). We use a timescale of 10 Myr in all our
models, following
De Lucia \& Blaizot (2007).
Merger-induced bursts cause feedback in the same way as quiescent star
formation.
\section{Modifying a simple model}
\label{sec:explore}
Below, we present a simple standard SAM, which we use as a starting
point for our tuning procedure. This simple model includes features
common to many current models of galaxy formation. It is kept as
simple as possible in order to facilitate tuning and
to simplify interpretation of changes to the model. In section
\ref{sec:tuning}, we present 6 alternative models which reproduce the
sSFR plateau.
\begin{figure}
\psfig{file=fig4_wnd.eps,width=220pt}
\caption{The cooling efficiency as a function of halo mass and redshift,
as used in all the models described in sec. \ref{sec:explore}.
}
\label{fig:cooling}
\end{figure}
\begin{figure}
\begin{tabular}{cc}
\psfig{file=fig5a_wnd.eps,width=220pt}\\ \psfig{file=fig5b_wnd.eps,width=220pt}\\
\end{tabular}
\caption{Star formation efficiency, $\epsilon_{\rm s}$, as a function of redshift and halo
mass, in the different models as indicated.
Top: in the standard model and in models FB1 and FB2, $\epsilon_{\rm s}$ is constant
at all times and halo masses. In models SF1, SF2, and SF3,
it is a function of redshift only.
Bottom: in model SF4, $\epsilon_{\rm s}$ is a function of both redshift and halo mass.}
\label{fig:sf_eff}
\end{figure}
\begin{figure*}
\begin{tabular}{cc}
\psfig{file=fig6a_wnd.eps,width=220pt} &
\psfig{file=fig6b_wnd.eps,width=220pt}\\ \psfig{file=fig6c_wnd.eps,width=220pt}&
\psfig{file=fig6d_wnd.eps,width=220pt}\\
\end{tabular}
\caption{
Feedback as a function of halo mass and redshift in our models.
Top left: feedback efficiency $f_{\rm d}$ in the standard model and all SF models.
Top right: $f_{\rm d}$ in model FB1.
Bottom left: $f_{\rm d}$ in model FB2.
Contours mark log$(f_{\rm d})=$ -0.5, 0, 0.5, 1, 1.5 and 2.
Bottom right: reincorporation efficiency $f_{\rm rc}$ in units of $1/t_{\rm dyn}$
in model FB2.
The main new feature in the FB models is strong feedback at $z\geq 4$,
and especially so in massive galaxies.
Reincorporation is effective at $z<3$ and in massive galaxies.
}
\label{fig:fb_eff}
\end{figure*}
\subsection{The standard model}
\label{sec:standard}
The ingredients
of this model were chosen based on the simplicity of the physical
processes involved and consistency with observational constraints.
\begin{itemize}
\item We assume that all accretion is cold, with $f_{\rm c} = 1/t_{\rm
dyn}$, and that accretion is quenched at $z<3$ above $M_{\rm h}= 1.6
\cdot 10^{12} M_{\odot}$ (e.g. Dekel \& Birnboim 2006; Cattaneo et al. 2006;
Ocvirk et al. 2008; Dekel et al. 2009). To avoid sharp breaks
in the stellar mass function, we smoothed the transition both in
redshift and in mass by hand, as shown in Fig. \ref{fig:cooling}.
\item The rate at which cold gas is turned into stars is
\begin{equation}
f_{\rm s}=\epsilon_{\rm s}/t_{\rm dyn} \,,
\end{equation}
where $\epsilon_{\rm s}$ is the star-formation efficiency.
For $t_{\rm dyn, disk}$, we assume
\begin{equation}
t_{\rm dyn, disk} = \frac{3\lambda \cdot R_{\rm vir}}{\sqrt{2} \cdot
V_{\rm vir}} \sim 0.0072 \cdot t_{\rm Hubble} \,,
\end{equation}
where the
halo spin parameter $\lambda = 0.03$ (according to
the mean value found by Mu\~{n}oz-Cuartas et al. 2010 in N-body
simulations).
We use a
constant star formation efficiency $\epsilon_{\rm s}=0.01$, as shown in Fig. \ref{fig:sf_eff} as blue line.
This is comparable to the estimates
of Krumholz \& Tan (2007), and similar to the efficiencies used in
other recent models (e.g. Krumholz \& Dekel 2010; Bouch\'{e} et al. 2010 and Guo et
al. 2010).
\item We include a simple prescription for stellar feedback
which is the usual way to
reproduce the low mass end of the stellar mass function
(Dekel \& Silk 1986). The
formation of stars affects the surrounding cold gas, making part of
it unavailable for star formation for a certain amount of time. The
rate at which gas is made unavailable for star formation is equal to
the SFR times an efficiency $f_{\rm d}$. As in most
other SAMs, $f_{\rm d}$ is a combination of a constant term (which is
called the ``reheating'' part of feedback in De Lucia \& Blaizot
2007), and a term that is inversely proportional to the virial
velocity to some power, which should roughly mimic ejection of
material due to SN explosions (e.g. Bower et al. 2006; De Lucia \&
Blaizot 2007; Guo et al. 2010). The feedback
efficiency is thus given by %
\begin{equation}
f_{\rm d} = \delta + (\frac{V_{\rm vir}}{\gamma})^{-\alpha}.
\end{equation}
We use $\alpha$=3.5, $\gamma$=161 km/s, $\delta=1$\footnote{A
non-zero delta is necessary in order to have some feedback even in
massive galaxies -- otherwise, even in the absence of cooling, the gas
coming from stellar recycling and mergers
is enough to keep up relatively high star formation
rates in massive galaxies.}. We do not let log($f_{\rm d}$) be higher than 2.5,
and we enforce a constant value for $f_{\rm d}$ at $z \geq 5.7$.
The value of
$f_{\rm d}$ as a function of halo mass and time is shown in
Fig.~\ref{fig:fb_eff}, top panel.
\item There is no reincorporation, i.e. we assume that
cold gas that has been made unavailable for star formation by
feedback once never returns back to the cold gas reservoir.
\item We treat dynamical friction as in NW10, with a prefactor of
$\alpha_{\rm df}=$3.
\item No merger-induced star bursts are included.
\item The gas residing in the filaments of satellites is stripped with an exponential
timescale of 4 Gyr.
\end{itemize}
Overall our model has similar basic scalings as other SAMs, although
the parameter values are slightly different, and the recipes are
simplified.
We ran the standard model (as well as all following models) on the
full volume of the Millennium Simulation (Springel et al. 2005). The
resulting stellar mass functions (SMFs) at different redshifts and the
sSFR-$\emph{z}$ relation (defined in what follows as the sSFR as a function of
redshift at stellar mass $2 \cdot 10^9 - 10^{10} M_{\odot}$)
are shown in Fig.~\ref{fig:mass_funs} and
\ref{fig:ssfrz} as thin blue lines.
To account for the observational completeness limit as indicated by
Stark et al. (2009), we only take into account galaxies with log(SFR)
$>$ 0.25, 0.45 and 0.5 $\rm{yr}^{-1}$ at $z$=4, 5 and 6 respectively in
Fig.~\ref{fig:ssfrz}. Clearly, the observed relation in the mass bin
$2 \cdot 10^9 - 10^{10} M_{\odot}$ is not reproduced, very similar to
the models from NW10 shown in Fig.~\ref{fig:nw10_ssfr}. On the
other hand, note that the the evolution of the SMF is reproduced well
(Fig.\ref{fig:mass_funs}),
which is important for the following discussion of non-standard models.
We show the relation between sSFR and stellar mass at $z$=4 and $z$=6
in Fig. \ref{fig:sSFR_mass4}, top left panels.
The evolution of the mean cosmological SFR density is shown
as thin blue line in Fig. \ref{fig:res_madau}.
We have
checked that including merger-induced bursts according to Croton et
al. (2006) does not have a significant impact on any of the results shown
here.
\begin{figure*}
\centerline{\psfig{file=fig7_wnd.eps,width=130mm,bbllx=30mm,bblly=90mm,bburx=188mm,bbury=210mm,clip=}}
\caption{
Stellar mass function (SMF) of galaxies at different
redshift bins, for our different models as indicated
and as summarized in table \ref{tab:models}. Observational
data are marked by grey symbols.
At $z>0$, the model stellar mass is convolved with a Gaussian error
distribution of standard deviation 0.25 dex, which includes the
differences in the IMFs assumed by the different observers.
Data at $z=0$ are by Li et al. (2009, circles), Baldry et al. (2008,
crosses), and Panter et al. (2007, pluses). Data at $z>0$
are from Bundy et al. (2006, $\emph{z}=0.75-1$,
circles), Borch et al. (2006, $\emph{z}=0.8-1$, crosses),
P\'{e}rez-Gonz\'{a}lez et al. (2008, $\emph{z}=0.8-1$, $\emph{z}=1.6-2$,
$\emph{z}=2.5-3$, plus signs), Fontana et al. (2006,$\emph{z}=0.8-1$,
$\emph{z}=1.6-2$, $\emph{z}=2-3$, stars), Drory et al. (2004, $\emph{z}=0.8-1$,
upward-pointing triangles), Drory et al. (2005, $\emph{z}=0.75-1.25$,
$\emph{z}=1.75-2.25$, $z=2.25-3$ , diamonds and squares), Marchesini et
al. (2009, $\emph{z}=1.3-2$, $\emph{z}=2-3$, right-pointing triangles).
Model SMFs are plotted at $\emph{z}=0$ (top left panel),
$\emph{z}=0.8, 1, 1.2$ (top right panel), $z=1.2, 1.5, 2$ (bottom left
panel) and \emph{z}=2, 2.5, 3 (bottom right panel).
}
\label{fig:mass_funs}
\end{figure*}
\begin{figure*}
\centerline{\psfig{file=fig8_wnd.eps,width=100mm}}
\caption{Evolution of sSFR for galaxies with stellar mass in the fixed bin
$2 \cdot 10^{9} - 10^{10} M_{\odot}$ for the models indicated and
listed in table \ref{tab:models}. At $z \geq 4$, only galaxies with SFR above
the completeness limits given by Stark et al. (2009) are included.
Errorbars denote $\pm 1\sigma$ at $z>2$.
Each of the redshift bins contains more than 50 galaxies.
All the models shown reproduce the sSFR plateau at $z=2-6$.}
\label{fig:ssfrz}
\end{figure*}
\begin{figure*}
\begin{tabular}{c}
\centerline{\psfig{file=fig9a_wnd.eps,width=300pt}}\\
\centerline{\psfig{file=fig9b_wnd.eps,width=300pt}}
\end{tabular}
\caption{
sSFR as a function of stellar mass, at
$z$=4 (top panels) and $z=6$ (bottom panels). The results of model FB1 are
very similar to model FB2.
The greyscale refers to log number density of galaxies in units
of ${\rm Mpc}^{-3}{\rm dex}^{-2}.$}
\label{fig:sSFR_mass4}
\end{figure*}
\subsection{Tuning}
\label{sec:tuning}
We next gradually modify selected ingredients of the standard model.
Our first goal is to bring the median sSFR at $z=2-6$
and in the mass bin $(0.2-1)\cdot 10^{10} M_{\odot}$
into better agreement with the observed plateau,
the grey band in Fig.~\ref{fig:obs}. We also
consider the relation between sSFR and stellar mass at $z=4$ and $z=6$,
shown in Fig. \ref{fig:sSFR_mass4}, in order to verify
that the median sSFR
tends to decrease with increasing stellar mass.
We do not attempt
a fit with the observed sSFR-mass relation
outside the mass bin $(0.2-1)\cdot 10^{10} M_{\odot}$,
because the lower mass
bin is strongly affected by incompleteness (e.g. Stringer et al. 2011),
while results for the higher mass bin seem controversial
(compare Stark et al. 2009 to Lee et al. 2010a).
Our secondary goal is to avoid severe deviations from the
observed SMF at intermediate redshifts, and from the
observed evolution of mean SFR density.
We do not attempt to fit the observed sSFR at $z<2$.
This is because we believe the mismatch in the sSFR between observations and models
at $z<2$, where the SFR in the universe declines rapidly,
is a separate problem that deserves a dedicated study. It is discussed in more detail
elsewhere (e.g. Fontanot et al. 2009; Karim et al. 2011). Additionally, we find fitting the
$z>2$ sSFR challenging enough even prior to adding
the additional constraints imposed by the low redshift data.
On a separate note, we point out that observations of sSFR in galaxies
with masses $(0.2-1)\cdot 10^{10} M_{\odot}$ at $z=1-2$ are rather uncertain, probably more
so than at higher redshifts, as they
tend to fall below the current stellar mass completeness limits at these redshifts
(e.g.
Rodighiero et al. 2010, Dunne et al. 2009). This is due to the decline of the global
star formation rate density and the increasing
importance of dust, which decrease the intrinsic luminosity
of galaxies at $z<2$.
For tuning the models, we focus on one model ingredient at the time,
i.e. the star formation efficiency or the feedback efficiency.
Using a table with discrete values of the efficiency
in 8 bins in time, and 10 bins in halo mass, we start by fitting the plateau at the highest redshift
and then subsequently continue to lower redshifts.
We try to keep dependencies on halo mass and time as monotonic as possible, to limit
the number of possible models, and also because monotonic dependencies are easier to
motivate physically.
Once we have found a solution for the plateau from $\emph{z}=2-7$, we compare the model results
with the other observational
constraints. Depending on the outcome,
we discard the model or improve the
fit to the stellar mass functions and the sSFR-stellar mass relation by making additional changes to the efficiencies.
Part of our models (SF2, SF3, FB2) are then further improved by simple changes to one or two other
model ingredients.
We point out that we
have not carried out a systematic study covering all the parameter space,
but a process of trial and error that is geared towards finding successful solutions which fit the constraints
in question.
Fitting the sSFR plateau and the evolution of the
stellar mass function together is not trivial, and a substantial number of
iterations were needed for each model to
arrive at the solutions presented below.
The different models are summarized in table \ref{tab:models}.
\begin{table}
\caption{A summary of the models discussed in this work. {$\epsilon_{\rm s}$ is the
star formation efficiency, $f_{\rm d}$ the feedback efficiency,
$f_{\rm rc}$ the reincorporation efficiency, $f_{\rm burst}$ the efficiency
of merger-induced star bursts, and $\alpha_{\rm df}$ is the dynamical
friction prefactor. $m_1$ and $m_2$ is the sum of the stellar and
cold gas mass in the merger progenitors.
If not listed, elements are kept at the standard
model values.}}
\begin{center}
\begin{tabular}{lcccccc}
\hline Models & Modifications & line type \\ \hline standard & -- &
thin blue \\ SF1 & $\epsilon_{\rm s}(t)$
& thin dashed red \\ SF2 & $\epsilon_{\rm s}(t)$, $f_{\rm
burst}(t, m_1/m_2)$ & dotted-dashed green \\ SF3 & $\epsilon_{\rm s}(t)$, $f_{\rm burst} (t, m_1/m_2)$,
$\alpha_{\rm df}(t)$ & dotted brown \\ SF4 & $\epsilon_{\rm s}(t, M_{\rm halo}) $ & dashed cyan \\ FB1 &
$f_{\rm d} (t, M_{\rm halo})$ & dashed pink \\ FB2 & $f_{\rm d}(t, M_{\rm halo})$, $f_{\rm rc}(t, M_{\rm halo})$ & thick black \\ \hline
\end{tabular}
\end{center}
\label{tab:models}
\end{table}
\subsection{Reproducing the sSFR plateau}
In this section, we explore two alternative modifications to the
standard model that aim at reproducing the sSFR plateau:
(i) models ``SF", in which $\epsilon_{\rm s}$ is reduced
at high redshifts, and
(ii) models ``FB", where the feedback efficiency is
enhanced at high redshifts.
Results for all the models are shown in Fig.~\ref{fig:mass_funs},
\ref{fig:ssfrz}, \ref{fig:sSFR_mass4},
and \ref{fig:res_madau}.
\subsubsection{Model SF1 - tune $\epsilon_{\rm s}$}
In model SF1, we tune the SFE parameter, $\epsilon_{\rm s}$, while all other model
parameters are kept fixed as in the standard model.
With $\epsilon_{\rm s}$ varying in time as shown in Fig.~\ref{fig:sf_eff} (top panel),
the model reproduces the sSFR plateau at $z=2-6$ as shown in
Fig.~\ref{fig:ssfrz}.
While $\epsilon_{\rm s}$ does not need to depend on halo mass, it is not monotonic with
time. In order to have an sSFR plateau starting from redshift $z_{\rm p}$
(chosen here to be $z_{\rm p} \sim 7$), this model requires an earlier epoch
where $\epsilon_{\rm s}$ is well above its fiducial value 0.01.
This early star formation is needed in order to produce enough galaxies
of large-enough stellar mass by $z_{\rm p}$, after which the much-lower SFR
adds only slowly to the stellar mass.
At the onset of the plateau, $\epsilon_{\rm s}$ has to drop to values well below $0.01$,
and it should continue to gradually decline till $z \sim 4$,
in order to permit the observed low sSFR values at $4<z<z_{\rm p}$.
After $z \sim 4$, $\epsilon_{\rm s}$ is gradually increasing in order to match the
high observed sSFR in this regime. It catches up with the fiducial value
$\epsilon_{\rm s} \sim 0.01$ only at low redshifts.
We note that model SF1 predicts that the sSFR plateau does not extend all
the way to the epoch of the emergence of the first galaxies; it is preceded
by a starburst epoch.
Unfortunately, Fig.~\ref{fig:mass_funs} shows that
model SF1 does not produce enough massive galaxies
to match the observed mass function at intermediate redshifts,
despite the high initial $\epsilon_{\rm s}$.
Model SF1 thus reveals a
nontrivial tension
between two sets of data, namely the low values of sSFR
at high redshifts
versus the high-mass end of the SMF at intermediate redshifts.
This indicates that the high values of sSFR obtained in SAMs
at high redshifts are needed there for the purpose of building up massive
enough galaxies by $z \sim 2$, and are therefore not easy to avoid.
\begin{figure}
\begin{tabular}{cc}
\psfig{file=fig10_wnd.eps,width=220pt} \\
\end{tabular}
\caption{
Cosmological evolution of SFR density for models as indicated
versus observations.
The upper grey belt represents the compilation by Hopkins \& Beacom
(2006, 3$\sigma$ confidence level) from measurements of SFR. More
recent determination by Bouwens et al. (2009) and by Kistler et al. (2009)
from gamma-ray bursts are consistent with the upper grey belt.
The lower grey belt is from Wilkins et al. (2008; 1$\sigma$
confidence level), derived from the growth curve of the
stellar mass density.
The model predictions lie in between.
}
\label{fig:res_madau}
\end{figure}
\subsubsection{Model FB1 - tune feedback}
In model FB1, we only tune the stellar feedback parameter $f_{\rm d}$,
while all other model parameters are kept fixed as in the standard model.
The values of $f_{\rm d}$ as a function of mass and redshift are shown in
Fig.~\ref{fig:fb_eff} (top-right panel), and the resultant sSFR evolution
and SMF are shown in Fig.~\ref{fig:mass_funs} and \ref{fig:ssfrz}.
We find that fitting the sSFR plateau requires very high
stellar feedback efficiencies at early times ($z \lower.7ex\hbox{\gtsima} 3$),
and especially so for massive haloes.
This is needed in order to balance the high accretion rates of these haloes.
High feedback efficiency is not needed in the massive haloes at
later times because the cooling in them is set to zero at $z \geq 3$.
At lower
halo masses $M_{\rm h} \lower.7ex\hbox{\ltsima} 10^{11} M_{\odot}$, moderately strong
feedback is needed at all redshifts. Note that stellar feedback is not
entirely monotonic here with either mass or time; we improve this in
the following model FB2 by adding reincorporation as an additional
model ingredient.
As in model SF1, we do not reproduce the high mass end of the SMF with
this model, again indicating that suppressing the efficiency of
star formation at high redshift leads to an underproduction
of massive galaxies, if no other changes to the model are made.
As can be seen in Fig. \ref{fig:mass_funs},
this shortage is more severe than in the SF models,
and more so at higher redshifts.
We did not manage to improve this aspect of the FB1 model so far,
since tuning the FB models
is more difficult than tuning the SF models. The reason for this is that
the star formation rate of galaxies is directly proportional
to $\epsilon_s$, while its dependence on the feedback efficiency
is non-linear.
All our models that reproduce the sSFR plateau at $z< 7$ require an
earlier phase of high SFR, with a sSFR higher
than the plateau level by a factor of a few. This provides a high
enough stellar mass at $z \sim 6$,
which allows the desired low sSFR at the plateau level while the SFR
is driven to high values by the high accretion rate.
For example, in order to reach by $z \sim 7$ a mass of $2 \cdot 10^9
M_{\odot}$,
the minimum mass of galaxies on the plateau, a
galaxy needs an average SFR $\sim$ 10 $M_{\odot}{\rm yr}^{-1}$ between
the onset of star formation at $z \sim 9$
and $z \sim 7$, say. This implies a sSFR $\sim 5\, {\rm Gyr}^{-1}$ at
$z \sim 7$, and even higher values
at earlier times, when the stellar mass is smaller.
Such an early phase of high SFR has to be introduced by hand when $
\epsilon_s$ is set to low values at $z<7$,
as in model SF1.
In model FB1, on the other hand, the high SFR at $z>7$ occurs
naturally, because the gas available for SFR in a
growing galaxy is proportional to the instantaneous accretion rate,
which is only slowly increasing with time. This is in contrast to the SF
models, where the SFR is proportional to the
accumulated gas in the galaxy and therefore tends to increase faster
with time.
\subsubsection{Model FB2 -
tune feedback and reincorporation}
In model FB2, we improve on model FB1 by adding reincorporation from
the feedback phase back to the cold phase as an additional model
ingredient. This makes it possible to keep the
variation of the feedback efficiency with time monotonic
(although the trend
is of opposite sense at low and high halo masses).
With model FB2, we can
reproduce the sSFR plateau very well down to $z \sim 2$
(Fig.~\ref{fig:ssfrz}, black thick solid line), but again fail to
simultaneously reproduce the high mass end of the SMF at $\emph{z} \lower.7ex\hbox{\gtsima}$ 2
(Fig.~\ref{fig:mass_funs}, black thick solid line). Feedback and
reincorporation efficiencies in model FB2 are shown in
Fig.~\ref{fig:fb_eff}, left and right bottom panel
respectively. Reincorporation is needed in order to keep up high sSFR
at $\emph{z} < 3$ for haloes with masses $M_{\rm h} \lower.7ex\hbox{\gtsima} 10^{11}
M_{\odot}$. We set it to very low values outside this range. The
efficiency of reincorporation in the range of halo mass and time when
it is needed then quite similar at all redshifts when expressed in
units of $1/t_{\rm dyn}$, with a slight peak at around $\emph{z} \sim 2$.
\subsection{Reproducing the sSFR and the evolution of the SMF}
While we have found in the previous section that it is
possible to reproduce the sSFR plateau once the
star-formation or feedback efficiency is allowed to vary in time
and with mass,
all the models discussed so far underproduce
the high mass end of the SMF at $z \lower.7ex\hbox{\gtsima} 2$. In this section, we
attempt to improve the fit to the SMF at
intermediate redshifts by an additional modification to the model.
For simplicity, we focus on modifications to model SF1.
Figure \ref{fig:mass_funs} shows that these models, SF2-SF4,
reproduce the high
mass end of the intermediate-redshift SMF better than the previous models.
Figure \ref{fig:sSFR_mass4} demonstrates that this success is associated
with a population of galaxies more massive than $10^9 M_{\odot}$ with high SFR
at $z \sim 4$,
which were missing in the previous models.
We explore here three different ways for producing this population.
In model SF2, we boost the efficiency of
merger-induced star bursts.
In model SF3, we speed up the merger
rate.
In model SF4, we introduce a rather
involved variation of $\epsilon_s$ both with halo mass and time,
making star formation more efficient in relatively high mass haloes while
keeping it inefficient in lower mass haloes.
\subsubsection{Model SF2 - tune $\epsilon_{\rm s}$ and mergers}
In model SF2,
we reproduce the required massive galaxies at $z \sim 2$ by boosting up
merger-induced starbursts.
We first tried
the starburst prescription of Croton et al. (2006)
where $f_{\rm burst}=0.56(m_1/m_2)^{0.7}$, but this had only
a little effect on the SMF.
However, with $f_{\rm burst}=(m_1/m_2)^{0.3}$ at $z>1$,
followed by the Croton et al. (2006) prescription at $z<1$,
model SF2 produces a higher
abundance of high mass galaxies in better
agreement with the SMF at intermediate redshift
(Fig.~\ref{fig:mass_funs}),
while still reproducing the sSFR plateau
(Fig.~\ref{fig:ssfrz}).
An additional parameter governing merger-induced starbursts is the burst
duration, which is set to 10 Myr. We find that smearing the bursts over a much longer duration
(e.g. 500 Myr) makes very little difference to the results.
As we only consider the median and the 68\% range, a small population
of starbursting galaxies can be present without ruining the sSFR plateau.
Strongly star bursting galaxies might also not make it into the Stark
et al. (2009) or Gonz\'{a}lez et al. (2010) samples, as the highly
star forming regions in galaxies tend to be strongly dust attenuated
(e.g Charlot \& Fall 2000), and are therefore a good way
of building up stellar mass in a hidden mode.
\subsubsection{Model SF3 - tune $\epsilon_{\rm s}$ and mergers}
Model SF3 is a variation of model SF2, where the high starburst
efficiency at $\emph{z}>1$ is replaced by a shorter characteristic time for dynamical
friction. The fiducial dynamical friction prefactor of
$\alpha_{\rm df}=3$ is replaced by
$\alpha_{\rm df}=0.1$ at $z>1$ and $\alpha_{\rm df}=5$ at $z<1$,
while the starbursts are moderate, $f_{\rm burst}=0.2(m_1/m_2)^{0.7}$.
The rapid stellar assembly
in this model boosts up the buildup of stellar mass, and we need to lower the efficiency
of star formation accordingly (Fig.~\ref{fig:sf_eff}, top panel).
This model provides a sensible fit to the SMF (Fig.~\ref{fig:mass_funs})
and it reproduces the sSFR plateau (Fig.~\ref{fig:ssfrz}).
\subsubsection{Model SF4 - tune $\epsilon_{\rm s}$}
Model SF4 demonstrates a third way to build up high mass
galaxies by moderate redshifts without enhancing the contribution of mergers.
In this model, the star-formation efficiency, $\epsilon_s$, is varied
as a function of both halo mass and time as shown in Fig. \ref{fig:sf_eff}.
This allows for a simultaneous fit to the SMF (Fig.~\ref{fig:mass_funs})
and the sSFR plateau (Fig.~\ref{fig:ssfrz}).
We note that in this model, the sSFR does not in general decrease with
increasing stellar mass, as shown in Fig. \ref{fig:sSFR_mass4}, for the z=6
results, bottom right panel. This is contrary to all observations of
this relation we are aware of, which tend to find that the median sSFR of galaxies
decreases with increasing stellar mass.
\subsection{Summary of model results}
In this section, we summarize the successes and failures of the models,
guided by Fig. ~\ref{fig:mass_funs}, \ref{fig:ssfrz} and ~\ref{fig:res_madau}.
From Fig. \ref{fig:mass_funs}, which shows the stellar mass functions
up to $z \sim 3$, it becomes clear that
only models SF3 and SF4 manage to reproduce the SMF at $z>1.2$.
In Fig. \ref{fig:ssfrz}, we show that all our models except the
standard models provide a reasonable fit to the sSFR plateau from $z \sim 2-6$.
At $z<2$, the sSFR is lower than the observational estimates.
This may be partially
due to incompleteness in the observations, which misses galaxies
with low SFR especially at low redshift, where the
global SF in the Universe has declined. The mismatch at $z \sim 0$, however,
is probably real and already apparent in the standard model. It is likely
connected to a similar underproduction of the sSFR present in
other recent SAMs (e.g. Fontanot et al. 2009, Guo et al. 2011). We note
that our efforts to improve the match to the sSFR plateau lead to an
increase in the
sSFR at $z<2$ and thus to a better agreement with observations.
In Fig. \ref{fig:res_madau} we show the global star formation
rate density as a function redshift in our models. The upper grey belt represents
the directly observed SFR density from Hopkins \& Beacom (2006),
while the lower belt is derived
from the growth curve of the stellar mass (Wilkins et al. 2008).
More recent estimates like from Bouwens et al. (2009) and
from Kistler et al. (2009) are in agreement with the upper belt.
The observational results that are directly measured from star formation,
and those that are obtained
indirectly from the
evolution of the stellar mass density
are thus clearly inconsistent, which is the reason
that SAMs (like e.g. Guo et al. 2011) usually have a SF density somewhere in between
these two constraints.
Results of our
various models start to deviate considerably at $z>2$, but the discrepancy between
the observational results makes it impossible to use them to constrain models.
The highest global SFR density
at $z>3$ is reached by the standard model, followed
by the models which have the best fit to the SMF at high redshift,
namely model SF3 and SF4.
\subsection{The sSFR of individual galaxies}
\label{sec:individual}
\begin{figure*}
\centerline{\psfig{figure=fig11_wnd.epsi,width=6in}}
\caption{
Evolution in a single galaxy within the same halo according to different
models.
Top row: standard model.
Middle row: model SF2.
Bottom row: model FB2.
Left: mass in dark matter (same in all models), stellar mass, and cold gas.
Right: SFR and sSFR.
Horizontal dashed lines indicate the stellar mass range
relevant for the median sSFR shown in Fig.~\ref{fig:ssfrz}.
The peaks in the SF history of model SF2 are due to starbursts.
}
\label{fig:connect}
\end{figure*}
In Fig.~\ref{fig:connect}, we follow the growth of one individual
model galaxy following the main branch of its merger history.
Shown as a function of redshift is the mass in the dark-matter, cold gas and
stars, the SFR and sSFR, for the standard model, SF2 and FB2.
The resulting stellar mass at $z=0$ is similar in the three models.
However, the star-formation history and the growth of stellar mass
are different --- in both model SF2 and FB2, the buildup of stellar mass
is delayed compared to the standard model.
We note that at $z>2$ the SFR in individual galaxies
is rising with time for all models.
This is a common feature to all our model galaxies, and is consistent with
the predictions
of Finlator et al. (2006, 2011), as well as with the observational
finding of Papovich et al. (2010), Maraston et al. (2010),
Lee et al. (2010a,b).
It is emerging naturally from the fact that in an individual galaxy,
the accretion rate is expected
to be growing in time in its rapid growth phase at $z \geq 2$ (Neistein
\& Dekel 2008).
We see in Fig. \ref{fig:connect}
that as the galaxy evolves from $z=6$ to $z=2$,
the sSFR declines in a rather slow pace under models SF2 and FB2
compared to the steeper decline in the standard model.
The sSFR plateau discussed in the previous sections, which refers to the
median value for galaxies of a fixed mass at different redshifts,
does not necessarily imply a constant sSFR in individual galaxies as
they grow. However, the observations of Stark et al. (2009) and Papovich
et al. (2010, Figs. 2 and 3) do argue for a constant sSFR in individual
galaxies as well. It is thus encouraging that our models that reproduce
the sSFR plateau also come close to reproducing the constancy of the sSFR
in individual galaxies.
\section{Discussion}
\label{sec:discuss}
In subsections \ref{sec:change_quiescent} and \ref{sec:change_mergers}
of this discussion section, we comment on the physical plausibility
of the ad-hoc modifications suggested by our models.
This discussion is not complete or conclusive --- it is only meant
to raise some of the relevant issues and to trigger further discussion.
The limited goal of the current paper remains to point out the possible
nature of the modifications to the standard models that are required
for reproducing the sSFR plateau together with the SMF at moderate redshift.
\subsection{Changes to the quiescent evolution of galaxies}
\label{sec:change_quiescent}
\begin{enumerate}
\item \emph {Star formation efficiency:}
In the SF models, $\epsilon_{\rm s}$ has to vary in time, with a very high
efficiency at $z \geq 7$, and a low efficiency at $z = 3-6$. The latter
is consistent with the prediction of Gnedin \& Kravtsov (2010) that the
low metallicity and high UV radiation at $z \sim 3$ should significantly
lower the normalization of the Kennicutt-Schmidt relation.
This effect is particularly strong for low gas surface densities.
An effective suppression of star formation in high redshift galaxies and at low masses turns out
to be a natural outcome of the low metallicity in these galaxies (Krumholz \& Dekel 2011).
The low dust content enables the UV radiation from stars to heat the gas (while dissociating H$_{2}$ molecules)
and to prevent further star formation, until enough metallicity is built up and the SFR regions are shielded.
Observational studies at $z \sim$ 3 seem to agree with this
prediction (e.g. Wolfe \& Chen 2006; Rafelski et al. 2010).
Furthermore, both SAMs and hydrodynamical simulations have suggested
that a low $\epsilon_{\rm s}$ at high redshift improves the match with observations
in a number of ways (Baugh et al. 2005; Lacey et al. 2010; Agertz et al. 2011).
It is more difficult to explain the high $\epsilon_{\rm s}$ at $z \geq 7$,
necessary for producing enough galaxies with a high stellar mass by
$z \sim 6$. What might help is a feedback mechanism that causes an
abrupt change in the mode of star formation after a very active initial phase.
Model SF4 suggests in addition an increase of $\epsilon_{\rm s}$ with halo mass.
Such a variation might result from faster metal enrichment or less efficient
feedback in more massive haloes.
\item \emph {Stellar feedback:}
The FB models suggest strong early stellar feedback that is not limited to
small galaxies. This is in potential disagreement with conventional models
of SN feedback, in which the effect of feedback at a fixed halo mass is
expected to be weaker at high redshift, where the deeper potential
well makes it harder to eject gas from the galaxy (Dekel \& Silk 1986;
Dekel \& Birnboim 2006).
One should therefore consider additional feedback mechanisms that may
vary in the required way with time and mass.
For example, a metallicity-dependent stellar feedback (Nishi \& Tashiro 2000;
Krumholz
\& Dekel 2011)
will be especially effective at high
redshifts.
The required strong feedback at high halo masses ($\log(M_{\rm h})>11.5$
at $t_{\rm Hubble} < 2$) may be achieved by AGN feedback from massive
black holes rather than stellar feedback.
\item \emph{Reincorporation:}
Model FB2 requires a high rate of reincorporation only at $z \lower.7ex\hbox{\ltsima} 3$.
In other SAMs, the reincorporation rates scale with $1/t_{\rm dyn}$,
and therefore gradually increase towards higher redshift, making the effective
stellar feedback less efficient at high $z$.
A weaker time dependence, more consistent with FB2,
is found in simulations that incorporate
strong momentum-driven winds (Oppenheimer et al. 2010).
Another option is ``reincorporation" that utilizes a gas reservoir
in the galaxy rather than gas that has been ejected earlier.
For example, this occurs naturally for metallicity-dependent feedback.
As long as the metal content in a galaxy is low, gas is not well shielded
from UV radiation by dust, leading to efficient heating and $H_2$ dissociation.
Once the metal content reaches a certain threshold value, the gas is shielded and star formation
becomes efficient.
The abrupt appearance at $z \sim 3$ of reincorporation in FB2 is
an oversimplification, but the general increase of reincorporation
with halo mass is plausible both for ejective and non-ejective feedback.
\end{enumerate}
\subsection{Changes to the recipes for mergers}
\label{sec:change_mergers}
\begin{enumerate}
\item \emph{Enhanced starburst efficiency at high redshift:}
In model SF2, the starburst efficiency, defined according
to eq. \ref{eq:sf_burst}, is higher at higher \emph{z}.
This is not to be confused with the SF efficiency $\epsilon_s$,
which is relevant for quiescent star formation.
An increased star burst efficiency at high $\emph{z}$
is plausible due to the shorter dynamical times.
The gas-rich mergers at high redshift may in principle be very different
from today's more familiar gas-poor mergers. High-resolution gas-rich merger
simulations indeed indicate an effective star formation efficiency
that is higher than in non-merger situations (Teyssier, Chapon \&
Bournaud 2010).
\item \emph{Enhanced merger rate at high redshift:}
In model SF3, the dynamical friction prefactor $\alpha_{\rm df}$ is reduced
to 0.1 at $z>1$, which speeds up the merger process by a factor $\sim 20-30$
compared to the standard model and other SAMs (Croton et al. 2006;
De Lucia et al. 2007). At high redshift, the merging galaxies flow from
the virial radius to the central galaxy along narrow radial streams
associated with the cosmic-web filaments in about one half of a halo
crossing time, $R_{\rm vir}/V_{\rm vir}$, which is $\sim 0.1 t_{\rm Hubble}$
(Dekel et al. 2009). This implies a merging time substantially shorter
then the dynamical friction time estimated for gradual spiraling in.
Hopkins et al. (2010) point out that the standard estimate of merger
time based on dynamical friction is indeed an overestimate
when the approach of the satellite is along a radial orbit,
or when the dynamical friction estimate starts at a distance larger
than 0.1 - 0.2 $R_{\rm vir}$, where the satellite is no longer properly
resolved in the simulation. Both of these conditions tend to be valid at
high redshift (Wetzel 2010, Hopkins et al. 2010), and together they may
lead to an order of magnitude overestimate of the merger time in the standard
model. We used the Millennium simulation to verify that with
$\alpha_{\rm df}=0.1$, the merging time is comparable to the halo crossing
time.
\end{enumerate}
\subsection{Dust at high redshift}
\label{sec:dust}
The validity of the sSFR plateau at high redshifts crucially depends
on the correction adopted for dust extinction by the different authors.
At $z \sim 2$, the estimates for dust extinction in LBGs from the UV
slopes are confirmed by radio estimates (Pannella et al. 2009) and
from comparison with a local sample of Lyman break analogs
(Overzier et al. (2011). At higher redshifts, the estimates
are more uncertain.
Stark et al. (2009), Gonz\'{a}lez et al. (2010)
and Labb\'{e} et al. (2010a,b) all assume practically
no dust extinction at $z>4$. This assumption is
supported by simple theoretical considerations and several
observational studies. It is expected that high-redshift galaxies should
have a lower dust content than today's galaxies simply because they had
less time to produce metals and dust. Indeed, theoretical models
(e.g. Guo \& White 2009) often assume a lower dust
extinction at higher redshift. Bouwens et al. (2009) estimated
the dust extinction at high redshift based on the observed UV
continuum slope, and found low values at $z \lower.7ex\hbox{\gtsima} 4$, especially for
the dominant population of galaxies with relatively low UV luminosities.
Labb\'{e} et al. (2010b), Finkelstein et al. (2010) and
Bouwens et al. (2010) also report on low or zero dust extinction at
$\emph{z} \sim 6-8$. Brammer \& van Dokkum (2007) tested whether the usual
LBG selection criteria miss a significant number of dusty galaxies at
$\emph{z} \sim 4$ by selecting according to Balmer breaks instead, and found
that this is not the case, again indicating that strong dust
obscuration at high redshift is rare. This may however not
be the case for massive galaxies, which potentially already have
high dust content at $z \sim 3-4$ (Mancini et al. 2009, Marchesini et al. 2011)
and thus may be missed by the usual LBG selection criteria.
Also,
Schaerer \& de Barros (2010) and Yabe et
al. (2009) argue for high correction factors of $\sim$ 9 when translating
the UV flux to SFR at $z=5-7$. We note that in order to reconcile the
plateau in the sSFR with current SAMs by the effect of dust-extinction
alone, the dust
extinction would have to increase with redshift, a trend which is
opposite to basic theoretical expectations.
While the ``dust" has not settled yet on this debate,
it seems that the evidence for a sSFR plateau is intriguing enough
to justify a serious theoretical consideration, but it is left for
future observations to tell whether this is indeed a valid strong constraint
or a fluke.
\subsection{Comparison to previous work}
\label{sec:previous_work}
Previous efforts to understand the properties of star forming galaxies
at $z>4$ include Finlator et al. (2006, 2011),
Night et al. (2006), Mao et al. (2007), Nagamine et al. (2008), Stark
et al. (2009), Lee et al. (2009), Khochfar \& Silk (2011), Lacey et
al. (2010), Lo Faro et al. (2009), and Stringer et al. (2010).
Here we compare our findings to
some of those studies.
Lee et al. (2009) concluded based on clustering analysis that the duty
cycle of star formation in high-redshift galaxies must be short, consisting
of episodes of $<0.35$ Gyr in each galaxy.
Stark et al. (2009) came to similar conclusions.
This would be in apparent conflict with the need to build up the massive
end of the SMF by $z \sim 2-3$.
Our model SF1 that reproduces the sSFR plateau assumes a continuous mode of
star formation, and even then, it underpredicts the massive end of the SMF.
If the low sSFR were associated instead with short-lived
bursts of star formation, it would have been even harder to build up the
stellar mass quickly enough. A way to reconcile a short duty cycle
with the required fast build-up of stellar mass might be that the short-lived
episodes of SF are only the final phase of a much more active
and dust-obscured starbursts, enough for building up the high mass end of
the SMF but undetectable in their violent phase.
Finlator et al. (2011) suggest that the clustering data by Lee et al. (2009)
can also be explained without a bursty star formation history, if there
are strong outflows that lead to an increased scatter in the relation
between baryonic mass and host halo mass.
Khochfar \& Silk (2011) present a model that reproduces the
sSFR plateau at $z>4$ with (i) a very high
frequency of star-bursts at high redshift and (ii) a burst efficiency
that scales with the inverse of the halo circular velocity to the
third power.
In this way, star formation in galaxies in the mass
range relevant for the plateau is inefficient at high redshift, but stellar mass
grows efficiently during mergers in low mass haloes.
Their model thus resembles a combination of our models SF2 and SF3.
The simultaneous need for a low sSFR at high redshift and
a fast build-up of stellar mass seems to be in odds
with the notion that the observed growth rate in stellar mass
density is low compared to that predicted by integrating
the observed SFR density (Wilkins et al. 2008).
These seemingly inconsistent problems may reflect a difference between
the overall evolution of stellar mass density and the evolution of
the population dealt with here, involving relatively massive galaxies in
a fixed mass bin.
Lo Faro et al. (2009), comparing a SAM for LBGs at $z \sim 4-6$ to
observations, found that their model produces an excess of star-forming
galaxies with stellar masses $\sim 10^{9} - 10^{10} M_{\odot}$ at
$z \sim 4$, within the range relevant for our study of the sSFR plateau.
They conclude that some feedback mechanism must suppress star formation
at early times in the corresponding haloes of
$\sim 10^{11} -10^{12} M_{\odot}$,
but it should become ineffective at later times in haloes
of similar mass. As Lo Faro et al. (2009) point out, such a behaviour of the feedback
efficiency is not part of the standard feedback models.
Their conclusions are thus
similar to ours, albeit they originate from a
different observational constraint.
\section{Conclusions}
\label{sec:conclusions}
We explored possible ad-hoc modifications to standard semi-analytic models
that reproduce a constant sSFR at $z>2$ while growing enough massive
galaxies by $z \sim 2$, and came to the following conclusions.
\begin{itemize}
\item
We confirm that a sSFR plateau at $z=2-6$ is in robust disagreement with
current models, as claimed in Bouch\'{e} et al. (2010).
We have demonstrated that, in a fixed stellar mass bin of
$2\cdot 10^9-10^{10} M_{\odot}$, the common feature of a variety of standard
SAMs is a gradual decline of the sSFR with time, associated
with the decline of the total specific accretion rate.
\item
We find that it is possible to reproduce the sSFR plateau together
with the stellar mass function at $z \sim 2$ via non-trivial modifications
to the standard SAMs. Three different modifications seem necessary,
related to three different observational features:
(1) the low sSFR at $z>4$, (2) the high sSFR at $z \sim 2-3$, and (3)
the abundance of high mass galaxies at $z \sim 2-3$.
The low sSFR at $z>4$ is reproduced either by strong stellar feedback at high
redshift in all masses, or by inefficient star formation at high redshift
following a phase of very efficient star formation at very high redshift.
The high sSFR at $z=2-3$ could emerge either by a drop
in the feedback efficiency at this epoch, or by a corresponding enhancement of
star formation efficiency, or by efficient reincorporation of gas
that was previously prevented from forming stars.
Finally, the high mass end of the SMF at $z > 1$ can be generated
despite the low SFR at high redshift by an additional modification of
the star formation in a sub-population of massive galaxies at high redshift.
This is achieved in our models by either speeding up the mergers, or enhancing
the merger-induced starbursts, or by introducing a non-trivial
dependence of star formation efficiency on halo mass.
However, none of our modified models matches the stellar
mass function at $\emph{z} \sim 3$ as well as our standard model. This
reflects the fact that the low sSFR at high redshift
and the presence of massive galaxies by $\emph{z} \sim 3$ are not easily
reconciled.
\item
Our models predict that the SFR in individual galaxies
is monotonically increasing with time at $z >2$.
This is in agreement with the theoretical predictions of
Finlator et al. (2006, 2011) and Bouch\'{e} et al. (2010), as well as
the observational results by Papovich et al. (2010).
We note in particular that an SFR that grows exponentially with time
implies a constant sSFR. It should be mentioned that
the decreasing SFR sometimes assumed in SED fitting
(e.g. in Stark et al. 2009) is incorrect.
\end{itemize}
We have demonstrated that the simple SAM of NW10 is useful for
exploring how current SAMs should be modified in order to
match new observational constraints, and for pointing out
apparently conflicting observational constraints. A similar method will be useful in addressing
other puzzling observations such as
the tilt in the relation of sSFR and stellar mass
(e.g. Somerville et al. 2008), the fraction of passive galaxies
as a function of
stellar mass at $z=0$ (e.g. Weinmann et al. 2010), or the fraction of
AGNs as a function of stellar mass (Fontanot et al. 2010).
We learn that the observed sSFR at high redshift has the potential for
posing powerful constraints on the physical processes of star formation and
feedback. If the sSFR plateau, as observed by Stark et al. (2009),
Labb\'{e} et al. (2010a, 2010b), and Gonz\'{a}lez et al. (2010), is
confirmed, it will provide invaluable information on the baryonic physical
processes at high $z$, indicating that it could be different from those
at low redshift. One should be eagerly waiting for new developments in the
observations of SFR and stellar mass at high redshift.
While we started a discussion of the physical plausibility of the
required modifications to the standard recipes of galaxy-formation models,
the modifications suggested above are primarily of an ad-hoc nature.
They deserve a thorough theoretical study of physical mechanisms that could
be responsible for the required variation with time and mass.
\section*{Acknowledgments}
We thank Rychard Bouwens, Sadegh Khochfar and the anonymous referee
for helpful comments on the draft,
and Rachel Somerville, Ben Oppenheimer, Ivo Labb\'{e}, Niv Drory,
Emmanuele Daddi, Marcel Haas, Raanan Nordon and Giulia Rodighiero
for useful discussion. EN was partially supported by the Minerva fellowship
during this project.
AD was partially supported by ISF grant 6/08,
by GIF grant G-1052-104.7/2009,
by a DIP grant, and by NSF grant AST-1010033.
SQL databases containing the Millennium simulations are publicly released at
\texttt{http://www.mpa-garching.mpg.de/millennium}. The code used to
generate semi-analytic models based on the NW10 method
is publicly available under
\texttt{http://www.mpa-garching.mpg.de/galform/sesam}
The Millennium site was created as part of the activities of the German
Astrophysical Virtual Observatory.
|
\section{Introduction}
The aim of this paper is to consider a regression model, where the response $Y$ is to be predicted by covariates $\mbox{${\mathbf z}$}$, with $Y$ real-valued and with $\mbox{${\mathbf z}$}$ a real explanatory vector. Relaxing the usual assumption of normality, we consider a generalized framework. The response value $y$ is drawn from a one-parameter exponential family of distributions, with a probabilistic density of the form: \begin{equation}\label{modele}\exp\left(\frac{y\eta(\mbox{${\mathbf z}$})-b(\eta(\mbox{${\mathbf z}$}))}{\phi}+c(y,\phi)\right).\end{equation} In this expression, $b(\cdot)$ and $c(\cdot)$ are known functions, which determine the specific form of the distribution. The parameter $\phi$ is a dispersion parameter and is also supposed to be known in what follows. The unknown function $\eta(\cdot)$ is the natural parameter of the exponential family, which carries information from the explanatory variables. Given a random sample of size $n$ drawn independently from a generalized regression model, the aim is to predict the function $\eta(\cdot)$. Such a model gives a large scope of applications because observations can result from many distribution families such as Gaussian, Poisson, Binomial, Gamma, {\it etc}. For a more thorough description of generalized regression modelling, we refer to \cite{MacCullaghNelder} or \cite{FahrmeirTutz}.
Let $(Y_i,\mbox{${\mathbf z}$}_i)_{i=1,\ldots,n}$ be an independent random sample drawn from a generalized regression model. The conditional mean and variance of the $\text{i}^{\text{th}}$ response $Y_i$ are given by:
\begin{eqnarray}
\mathbb{E}[Y_i|\mbox{${\mathbf z}$}_i]&=&\dot b(\eta(\mbox{${\mathbf z}$}_i))\,=\,\mu(\mbox{${\mathbf z}$}_i),\\
Var[Y_i|\mbox{${\mathbf z}$}_i]&=&\phi\,\ddot b(\eta(\mbox{${\mathbf z}$}_i)),
\end{eqnarray} where $\dot b(\cdot)$ and $\ddot b(\cdot)$ denote respectively the first and second derivatives of $b(\cdot)$. The function $G=\dot b^{-1}$ is called link function and one have $G(\mathbb{E}(Y_i|\mbox{${\mathbf z}$}_i))=\eta(\mbox{${\mathbf z}$}_i)$.
A linear model consists in assuming that the dependence from the covariate $\mbox{${\mathbf z}$}$ is linear, meaning $\eta(\mbox{${\mathbf z}$})$ can be written on the form $\eta(\mbox{${\mathbf z}$})=\mbox{${\mathbf z}$}^T\mbox{\boldmath$\beta$}$ ; the superscript $T$ denotes the transpose of a vector or matrix. Yet, in some applications, the linear model is insufficient to explain the relationship between the response variable and its associate predictors. The generalized functional model relax this linearity, considering a nonparametric form for the canonical parameter, say $\eta(\mbox{${\mathbf z}$})=f(\mbox{${\mathbf z}$})$. Nevertheless in such a model appears the well-known {\it curse of dimensionality}. To avoid this drawback, we allow most predictors to be modelled linearly, while a small number is modelled nonparametrically. To this aim we decompose the covariate $\mbox{${\mathbf z}$}$ in two components: in the following $\mbox{${\mathbf z}$}=(\mbox{${\mathbf X}$},t)$, with $\mbox{${\mathbf X}$}$ a $p$-dimensional vector and $t$ a real-valued covariate. The function $\eta(\cdot)$ is given by:
\begin{equation}\label{GPLM}
G(\mu(\mbox{${\mathbf X}$},t))=\eta(\mbox{${\mathbf X}$},t)=\mbox{${\mathbf X}$}^T\mbox{\boldmath$\beta$}+f(t),
\end{equation}
where $\mbox{\boldmath$\beta$}$ is an unknown $p$-dimensional real parameter vector and $f(\cdot)$ is an unknown real-valued function; such a model is called a generalized partially linear model (GPLM). Given the observed data $(Y_i,\mbox{${\mathbf X}$}_i,t_i)_{i=1,\ldots,n}$, the aim is then to estimate from the data the vector $\mbox{\boldmath$\beta$}$ and the function $f(\cdot)$.
In a Gaussian modelisation, \cite{Rice} and \cite{Speckman} put in evidence that the rates of the estimates for linear and nonlinear parts could not be both optimized without a control of the correlation between the explanatory variable of the linear part and the functional part of the model. With such a control, \cite{Speckman} proves that it is possible to obtain both optimal linear rate and nonparametric rate for the estimates. To my knowledge, the only paper establishing such a result in GPLM is \cite{MammenVanderGeer}.
Many papers focus on the asymptotic behaviour of the estimator of the parametric part $\mbox{\boldmath$\beta$}$ in generalized partially linear models (see {\it e.g.}~\cite{Chen87} by a penalized {\it quasi-least squares} or \cite{SeveriniStaniswalis} by {\it profile likelihood} methods). A recent article of \cite{BoenteHeZou} establishes a uniformly convergent estimation for $f$ using a robust {\it profile likelihood} similar to \cite{SeveriniStaniswalis}. But few works consider simultaneously the parametric and the nonparametric part of the model. The paper of \cite{MammenVanderGeer} shows minimax optimality for the estimations of both $f$ and $\mbox{\boldmath$\beta$}$ with a penalized {\it quasi-least squares} procedure. Note that the authors use a Sobolev type penalty. The conditions given there for attaining optimality for both parametric and nonparametric estimators appear to be more restrictive than those given in Gaussian partially linear models in the literature.
This paper proposes a new estimation procedure based on wavelet decompositions. Wavelet based estimators for the nonparametric component of a Gaussian partially linear model have been investigated by \cite{Meyer03}, \cite{ChangQu}, \cite{FadiliBullmore} or \cite{gannaz} and \cite{gannaz_these} more recently. But it has not been studied in the content of GPLM. Estimation methods encountered in literature need the choice of a smoothing parameter, which optimal value depends on the regularity of the functional part $f$. A cross-validation procedure is then necessary to evaluate this parameter. As noted by \cite{Rice} or \cite{Speckman}, cross-validation can present much instability in partially linear models. The use of wavelets here leads to a procedure where no cross-validation is needed. This adaptivity is the main novelty of our estimation scheme in such models. We moreover establish the near-minimax optimality of the estimation for both the linear predictor $\mbox{\boldmath$\beta$}$ and the nonparametric part $f$, under usual assumptions of correlation between the two parts. The correlation condition appears to be similar to what is classically for the Gaussian case and is weaker than in \cite{MammenVanderGeer}. Finally, we present an algorithm for computing the estimates.
The paper is organized as follows: Section 2 presents the assumptions and the estimation procedure. It also gives the main properties of our estimators. We distinguish two cases: non adaptive penalties and a $\ell^1$ type penalty, which leads to adaptivity. In Section 3, we propose a computational algorithm of the adaptive estimation procedure and we present a small simulation study for the numerical implementation. Proofs of our results are given in the Appendix.
\section{Assumptions and estimation scheme}
We consider a generalized regression model, where the response $Y$ depends on covariates $(\mbox{${\mathbf X}$},t)$, where $Y$ is real-valued, $\mbox{${\mathbf X}$}$ is a $p$-dimensional vector and $t$ is a real-valued covariate. The value $y$ is drawn from an exponential family of distributions, with a probabilistic density of the form:
\begin{equation*}
\exp\left(\frac{y\eta(\mbox{${\mathbf X}$},t)-b(\eta(\mbox{${\mathbf X}$},t))}{\phi}+c(y,\phi)\right),
\end{equation*}
where functions $b(\cdot)$ and $c(\cdot)$ as well as the dispersion parameter $\phi$ are supposed to be known. The aim is to evaluate the unknown generating function $\eta(\cdot)$. As noted above, we are interested in this paper by a partially linear modelisation of the function $\eta(\cdot)$ and hence we suppose that it has the semiparametric expression:
\begin{equation*}
\eta(\mbox{${\mathbf X}$},t)=\mbox{${\mathbf X}$}^T\mbox{\boldmath$\beta$}+f(t).
\end{equation*}
The vector $\mbox{\boldmath$\beta$}$ and the function $f$ are respectively the parametric and nonparametric components of the generalized partially linear model (GPLM). In the following, $(Y_i,\mbox{${\mathbf X}$}_i,t_i)_{i=1,\ldots,n}$ will denote an independent random sample drawn from the GPLM described here.
\subsection{Penalized maximum likelihood}
The aim of the paper is to estimate simultaneously the parameter $\mbox{\boldmath$\beta$}$ and the function $f$, given the observed data $(Y_i,\mbox{${\mathbf X}$}_i,t_i)_{i=1,\ldots,n}$. We propose a penalized maximum loglikelihood estimation. Let $\ell$ denotes the loglikelihood function: $\ell(y,\eta)=y\eta-b(\eta).$ We consider throughout the paper that the estimators $\hat f_n$ and $\hat \mbox{\boldmath$\beta$}_n$ are solutions of :
\begin{equation}\label{critere}
(\hat f_n,\hat\mbox{\boldmath$\beta$}_n)=\displaystyle \mathop{argmax}_{\{f,\,\|f\|_\infty\leq C_\infty\},\mbox{\boldmath$\beta$}\in\mathbb{R}^p} K_n(f,\mbox{\boldmath$\beta$}) \quad\text{ with~~} K_n(f,\mbox{\boldmath$\beta$}) = \sum_{i=1}^n \ell\left(y_i, \mbox{${\mathbf X}$}_i^T\mbox{\boldmath$\beta$}+f(t_i)\right) \, - \, Pen(f).
\end{equation}
We refer to \cite{AnestisGijbelsNikolova} for the conditions of existence of such a maximization problem. In what follows, we will assume the penalty is convex and the likelihood function is bounded in order to ensure the existence of maxima (unicity is not acquired but there is no local maxima). We did not succeed in getting rid of the constraint $\|f\|_\infty\leq C_\infty$ in the proofs but such a condition does not seem too restrictive in practice.
The computation of the maximization problem is done in two step. Some studies, such as \cite{Speckman}, incite to estimate first the functional part. Thus, we will proceed as follows:
\begin{enumerate}
\item $\widetilde f_{n,\mbox{\boldmath$\beta$}}=\displaystyle \mathop{argmax}_{\{f,\,\|f\|_\infty\leq C_\infty\}} K_n(f,\mbox{\boldmath$\beta$})$.
\item $\hat\mbox{\boldmath$\beta$}_n=\displaystyle \mathop{argmax}_{\mbox{\boldmath$\beta$}\in\mathbb{R}^p} K_n(\widetilde f_{\mbox{\boldmath$\beta$}},\mbox{\boldmath$\beta$})$. Actually, a classical procedure is here to maximise a modified criterion called {\it profile likelihood} (see among others \cite{SeveriniWong} or \cite{BoenteHeZou}). The criterion maximized is then $\sum_{i=1}^n \ell(y_i,\dot b(\mbox{${\mathbf X}$}_i^T\mbox{\boldmath$\beta$}+\widetilde f_{\mbox{\boldmath$\beta$}}(t_i)))- Pen(f)$. The expression of $\dot b(\mbox{${\mathbf X}$}_i^T\mbox{\boldmath$\beta$}+\widetilde f_{\mbox{\boldmath$\beta$}}(t_i))$ can indeed be simplified using first order conditions of step 1. Due to the non-linearity of our procedure, we choose here to keep a loglikelihood approach.
\item $\hat f_n=\widetilde f_{\hat\mbox{\boldmath$\beta$}_n}$.
\end{enumerate}
Note also that an usual estimation procedure used in generalized models is quasi-likelihood maximization. For details, we refer among others to \cite{SeveriniStaniswalis} or \cite{MacCullaghNelder}. In GPLM, quasi-likelihood estimation was developed for example by \cite{Chen87} and \cite{MammenVanderGeer}.
\subsection{Discrete wavelet transform}
The aim of the present work is to introduce a wavelet penalty in estimation. The idea is to use the wavelet representation of the functional part we wish to estimate through this penalty. For more precision on wavelets, the reader is referred to \cite{Daubechies}, \cite{Meyer92} or \cite{Mallat99}.
Let $\left(L^2[0,1],\mbox{$\langle$} \cdot, \cdot \mbox{$\rangle$}\right)$ be the space of squared-integrable functions on $[0,1]$ endowed with the inner product $\mbox{$\langle$} f,g \mbox{$\rangle$}=\int_{[0,1]}f(t)g(t)\,dt$. Throughout the paper we assume that we are working within an $r$-regular ($r\geq 0$) multiresolution analysis of $\left(L^2[0,1],\mbox{$\langle$} \cdot, \cdot \mbox{$\rangle$}\right)$, associated with an orthonormal basis generated by dilatations and translations of a compactly supported scaling function, $\varphi(t)$, and a compactly supported mother wavelet, $\psi(t)$. For simplicity reasons, we will consider periodic wavelet bases on $[0,1]$.
For any $j\geq 0$ and $k=0,1,\ldots,2^j-1$, let us define $\varphi_{j,k}(t)=2^{j/2}\varphi(2^j t-k)$ and $\psi_{j,k}(t)=2^{j/2}\psi(2^j t-k)$. Then for any given resolution level $j_0\geq 0$ the family
$$ \left\{\varphi_{j_0,k},\,k=0,1,\ldots,2^{j_0}-1;\;\psi_{j,k},\,j\geq j_0;\,k=0,1,\ldots,2^j-1\right\} $$ is an orthonormal basis of $L^2[0,1]$. Let $f$ be a function of $L^2[0,1]$ ; if we denote by $c_{j_0,k}=\mbox{$\langle$} f,\varphi_{j_0,k}\mbox{$\rangle$}$ ($k=0,1,\ldots,2^{j_0}-1$) the scaling coefficients and by $d_{j,k}=\mbox{$\langle$} f,\psi_{j,k}\mbox{$\rangle$}$ ($j\geq j_0;\,k=0,1,\ldots,2^j-1$) the wavelet coefficients of $f$, the function $f$ can then be decomposed as follows:
$$ f(t)=\sum_{k=0}^{2^{j_0}-1}c_{j_0,k}\varphi_{j_0,k}(t)+\sum_{j=j_0}^\infty\sum_{k=0}^{2^j-1}d_{j,k}\psi_{j,k}(t),\quad t\in [0,1].$$
Yet in practice, we are more concerned with discrete observation samples rather than continuous. Consequently we are more interested by the discrete wavelet transform (DWT). Given a vector of real values $\mbox{${\mathbf e}$}=(e_1,\ldots,e_n)^T$, the discrete wavelet transform of $\mbox{${\mathbf e}$}$ is given by $\mbox{\boldmath$\theta$}=\Psi_{n\times n}\mbox{${\mathbf e}$}$, where $\mbox{\boldmath$\theta$}$ is an $n\times 1$ vector comprising both discrete scaling coefficients, $\theta_{j_0,k}^S$, and discrete wavelet coefficients, $\theta_{j,k}^W$. The matrix $\Psi_{n\times n}$ is an orthogonal $n\times n$ matrix associated with the orthonormal periodic wavelet basis chosen, where one can distinguish the {\it Blocs} spanned respectively by the scaling functions and the wavelets.
Note that if $\mbox{${\mathbf F}$}$ is a vector of function values $\mbox{${\mathbf F}$}=(f(t_1),\ldots,f(t_n))^T$ at equally spaced points $t_i$, then the corresponding empirical coefficients $\theta_{j_0,k}^S$ and $\theta_{j,k}^W$ are related to their continuous counterparts $c_{j_0,k}$ and $d_{j,k}$ with a factor $n^{-1/2}$. It is worthy to remark also that because of orthogonality of $\Psi_{n\times n}$, the inverse DWT is simply given by $\mbox{${\mathbf F}$}=\Psi_{n\times n}^T\mbox{\boldmath$\theta$}$. If $n=2^J$ for some positive integer $J$, \cite{Mallat} propose a fast algorithm, that requires only order $n$ operations, to compute the DWT and the inverse DWT.
\subsection{Assumptions and asymptotic minimaxity}
Let $\|.\|$ denotes the euclidean norm on $\mathbb{R}^p$ and $\|h\|_n^2=\frac 1 n \sum_{i=1}^n h(t_i)^2$ for any function $h$. To ameliorate the comprehension of the results and the assumptions, the subscript $0$ will identify in the following the true values of the model.
Due to the use of the Discrete Wavelet Transform described above, we will consider in the following that the functional part is observed on an equidistant sample $t_i=\frac i n$, and that the sample size satisfy $n=2^J$ for some positive integer $J$.
We first introduce assumption (A1) which ensures the identifiability of the model:
\vspace{-\baselineskip}
\begin{description}
\item[(A1)] $\frac{1}{n} X^T X$ converges to a strictly positive matrix when $n$ goes to infinity, and $\frac 1 n X^T F_0$ goes to 0 when $n$ goes to infinity, with $F_0=\left(f_0(t_1),\ldots,f_0(t_n)\right)^T$.
\end{description}
\vspace{-\baselineskip}
Define $H=X(X^TX)^{-1}X^T$ the projection matrix on the space generated by the columns of $X$. The matrix $H$ admits a rank and thus a trace equal to $p$. If $h_i=\mbox{${\mathbf X}$}_i(X^TX)^{-1}\mbox{${\mathbf X}$}_i^T$ denotes the $i^{th}$ diagonal term of $H$, this means that $\sum h_i=p$. We moreover suppose that:
\vspace{-\baselineskip}
\begin{description}
\item[(A2)] $h=\max_{i=1\dots n} h_i\to 0$
\end{description}
\vspace{-\baselineskip}
This assumption is very usual (see {\it e.g.} \cite{Huber}).
Concerning the loglikelihood function, we assume that:
\vspace{-\baselineskip}
\begin{description}
\item[(A3)] $\sup_{\|\eta-\eta_0\|_n\leq 2 (C_\infty+\|f\|_\infty)}\sup_i \dddot b(\eta_{i})\,\leq\,\dddot b_{\infty}\,<\,\infty$. We also suppose that $\min_i \ddot b(\mbox{${\mathbf X}$}_i^T\beta_0+f_0(t_i))\,\geq\,\ddot b_0\,>\,0$ and $\max_i \ddot b(\mbox{${\mathbf X}$}_i^T\beta_0+f_0(t_i))\,\leq\,\ddot b_\infty\,<\,\infty$. Moreover, $\frac 1 n \displaystyle \sum_{i=1}^n \ddot b(\mbox{${\mathbf X}$}_i^T\beta_0+f_0(t_i))\mbox{${\mathbf X}$}_i^T \mbox{${\mathbf X}$}_i$ converges to a positive matrix $\Sigma_0$.
\end{description}
Recall that the function $\ddot b(\cdot)$ is associated to the variance of the observations and that on have $\ddot b>0$.
Then some more restrictive assumptions are made on the form of the
distribution:
\begin{description}
\item[(A4.1)] There exist a constant $a>0$ such that $\max_{i=1,\dots,n}\mathbb{E}\left[\exp(\dot \ell(Y_i,\mbox{${\mathbf X}$}_i^T\beta+f(t_i))^2/a)\right]\leq a$,
\item[(A4.2)] There exist constants $K,\,\sigma_0^2\,>0$ such that $$\max_{i=1,\dots,n} K^2\left(\mathbb{E}[\exp\left(\left|\dot \ell(Y_i,\mbox{${\mathbf X}$}_i^T\beta+f(t_i))\right|/K^2\right)]-1\right)\leq \sigma_0^2.$$
\end{description}
\vspace{-\baselineskip}
Assumption (A4.1) corresponds to exponential tails and is weaker than assumption (A4.2), which corresponds to sub-Gaussian tails.
We aim to control the correlation between the linear and the nonparametric parts of the model. Following \cite{Rice} or \cite{Speckman} we decompose the components of the design matrix $X$ into a sum of a deterministic function of $L^2[0,1]$ and a noise term. More precisely, the $(i,j)$-component of $X$, say $x_{i,j}$, is supposed to take the form $x_{i,j}=g_i(t_j)+\xi_{i,j}$ with functions $(g_i)_{i=1,\ldots,p}$ forming an orthogonal family on $\left(L^2[0,1],\mbox{$\langle$}\cdot,\cdot\mbox{$\rangle$}\right)$ and with $\xi_{i,j}$ denoting a realization of a random variable $\mbox{\boldmath$\xi$}_i$. The variables $(\mbox{\boldmath$\xi$}_i)_{i=1,\ldots,n}$ are supposed to be centered, independent, with finite variance. We can easily see that the orthogonality of the family $(g_i)_{i=1,\ldots,p}$ ensures that the matrix $\frac{1}{n}X^T X$ goes to a strictly positive matrix when $n$ goes to infinity. If in addition the system $(g_i)_{i=1,\ldots,p}$ satisfies $\int_{[0,1]}f(t)g_i(t)\,dt=0$, then assumption (A1) and consequently identifiability are guaranteed.
We also make an assumption on the distribution of the random variables $\mbox{\boldmath$\xi$}_i$, and of course we suppose we control the regularity of the functions $g_i$.
\vspace{-\baselineskip}
\begin{description}
\item[($A_{corr}$)] $\forall j=1,\dots,p,\,i=1,\dots,n,\,\,X_{i,j}=g_j(t_i)+\xi_{i,j}$, with polynomial functions $g_j$ of degree less or equal than the number of vanishing moments of the wavelet considered. For all $j=1,\dots,p$, $(\xi_{i,j})_{i=1,\dots,n}$ is a $n$-sample such that $\max_{i=1,\dots,n}\mathbb{E}\left[\exp(\mbox{\boldmath$\xi$}_{i,j}^2/a_j)\right]\leq a_j$,
for given constants $a_j>0$.
\end{description}
\subsubsection{Nonadaptive case} \label{sec:nonadapt}
Assume $J(\cdot)$ is a given criterion on the functions from $[0,1]$ to the positive real line. We introduce the function class $\mathcal A=\left\{g,\,J(g)\leq C\right\}$. Let us recall the definition of the entropy of a subspace:
\begin{Definition} Let $\mathcal F$ be a subset of a metric space $(\mathcal L,d)$ of real-valued functions. The $\delta$-covering number $N(\delta,\mathcal F,d)$ of the space $\mathcal F$ is the smallest number $N$ such that their exist $a_1,\ldots,a_N$ such that for each $a\in\mathcal F$ one have $d(a,a_i)\leq\delta$ for some $i\in\{1,\ldots,N\}$. The $\delta$-entropy $\mathcal H(\delta,\mathcal F,d)$ of the space $\mathcal F$ is defined as $\mathcal H(\delta,\mathcal F,d)=\log\,N(\delta,\mathcal F,d)$.
\end{Definition}
We here suppose that the $\delta$-entropies of the subspace $\mathcal A$ for the distance associated to the norm $\|\cdot\|_n$ behave like $$\lim\sup_{n\to\infty}\,\sup_{\delta>0}\;\delta^\nu\,\mathcal H(\delta,\mathcal A,\|\cdot\|_n)\;<\;\infty,$$ for a given $0<\nu<2$.
The penalty in equation~(\ref{critere}) is chosen according to the two assumptions (A6) and (A7):
\vspace{-\baselineskip}
\begin{description}
\item[(A5)] For any function $h$, $\frac{\lambda v_n^2}{n}Pen(h)\geq J(h)$ with $v_n=n^{1/(2+\nu)}$.
\item[(A6)] $f\mapsto K_n(f,\mbox{\boldmath$\beta$})= \sum_{i=1}^n \ell(y_i,\mbox{${\mathbf X}$}_i^T\mbox{\boldmath$\beta$}+f(t_i))-\lambda Pen(f)$ is concave.
\end{description}
\vspace{-\baselineskip}
When the special structure given in $(A_{corr})$ is assumed, we are willing to exploit it through a penalty on wavelet coefficients :
\vspace{-\baselineskip}
\begin{description}
\item[(A7)] The penalty $Pen(h)$ applies only to the wavelet decomposition coefficients $(\theta^W_{j,k})_{j\geq j_S, \,k\in\mathbb{Z}}$ of the function~$h$.
\end{description}
\vspace{-\baselineskip} We are now in position to give our first asymptotic result.
\begin{Theoreme}\label{th1}
Suppose assumptions (A1) to (A3), (A4.1), (A5) and (A6) hold and $J(f_0)<\infty$. Let $\mbox{\boldmath$\beta$}$ be a given $p$-vector such that $\sqrt{n}\|\mbox{\boldmath$\beta$}-\mbox{\boldmath$\beta$}_0\|\leq c$. Define $\widetilde f_{\mbox{\boldmath$\beta$}}=\displaystyle \mathop{argmax}_{\{f,\,\|f\|_\infty\leq C_\infty\},\, \mbox{\boldmath$\beta$}\in\mathbb{R}^p} K_n(f,\mbox{\boldmath$\beta$}).$ Then,
\begin{eqnarray*}
v_n\|\widetilde f_{\mbox{\boldmath$\beta$}}-f_0\|_n&=&\bigcirc_\mathbb{P}(1)\\
J(\widetilde f_{\mbox{\boldmath$\beta$}}) &=&\bigcirc_\mathbb{P}(1).
\end{eqnarray*}
Define $\hat\mbox{\boldmath$\beta$}_n=\displaystyle \mathop{argmax}_{\mbox{\boldmath$\beta$}\in\mathbb{R}^p} K_n(\widetilde f_{\mbox{\boldmath$\beta$}},\mbox{\boldmath$\beta$})$. Then, $$v_n\|\hat \mbox{\boldmath$\beta$}_n-\mbox{\boldmath$\beta$}_0\|=\bigcirc_\mathbb{P}(1),$$
If in addition the covariates of the linear part admit a representation of the form given in assumption $(A_{corr})$ and if the penalty satisfies (A7), then: $$\sqrt{n}\|\hat \mbox{\boldmath$\beta$}_n-\mbox{\boldmath$\beta$}_0\|=\bigcirc_\mathbb{P}(1).,$$
The results still hold if the number $p$ of regression covariates goes to infinity provided the sequence $n^{(\nu-2)/(\nu+2)}p$ goes to 0 when $n$ goes to infinity and the sequence $n^{-\nu/(2+\nu)}p$ is bounded.
\end{Theoreme}
The proof is given in Appendix. The main keys are controls given by \cite{VanderGeer00}, relying on the entropy.
Note that minimax optimality is obtained both for the linear predictor and the nonparametric estimator. The condition of correlation ($A_{corr}$) under which the optimality is attained is similar to the one given in \cite{gannaz_these}. This condition on design covariates appears to be more flexible than Rice's (1986) or Speckman's (1988) in the sense that the maximal degree of the polynomial functions intervening in the covariates depends on the number of vanishing moments of the wavelet, instead of depending on the regularity of the function $f$. Compared to \cite{MammenVanderGeer}, the assumptions seem much more weaker. Note that without correlation conditions both estimators attain nonparametric convergence rates.
The fact that the results hold for a number of covariate $p$ going to infinity allows to have many covariates in the linear part. This remark can be useful for dimension reduction modelling, where the number of covariates is large. However the rate of convergence for $p$ may be poor when the regularity of the function $f$ is poor.
In order to illustrate the general framework in which we gave the asymptotic behaviour, let us consider the penalty proposed in \cite{Antoniadis96}. To exploit the sparsity of wavelet representations, we will assume that $f$ belongs to a Besov space on the unit interval, $\mathcal B^s_{\pi,r}([0,1])$, with $s+1/\pi-1/2>0$. The last condition ensures in particular that an evaluation of $f$ at a given point makes sense. For a detailed overview of Besov spaces we refer to \cite{DonJohn98}.
\begin{Corollaire}
Suppose $f$ belongs to a Besov ball $\mathcal B^s_{\pi,r}(R)$ with $s>1/2$, $0<s-1/2+1/\pi$, $\pi>2/(1+2S)$ and $1/\pi<s<\min(R,N)$, where $N$ denotes the number of vanishing moments of the wavelet $\psi$.
Take the penalty: $Pen(f)=\sum_{j=j_S}^{n}2^{2js}\sum_k|\theta_{j,k}^W|^2$ where $\theta_{j,k}^W$ are the wavelet coefficients of $f$.
Assume conditions of Theorem~\ref{th1}, (A7) and $(A_{corr})$ hold.\\
If $\lambda\sim n^{-2s/(1+2s)}$, we can deduce from Theorem~\ref{th1} that \begin{eqnarray*}
\|\widehat f_{n}\|_{s,2,\infty}&=&\bigcirc_\mathbb{P}(1),\\
n^{s/(1+2s)}\|\widehat f_{n}-f_0\|_{n}&=&\bigcirc_\mathbb{P}(1)\\
\sqrt{n}\|\hat \mbox{\boldmath$\beta$}_n-\mbox{\boldmath$\beta$}_0\|&=&\bigcirc_\mathbb{P}(1).
\end{eqnarray*}
where $\|f\|_{s,2,\infty}={\displaystyle \sup_{j\geq j_S}} 2^{j(s-1/2)}\left(\sum_k |\theta_{j,k}^W|^2\right)^{1/2}.$
\end{Corollaire}
\begin{proof}
\cite{BirgeMassart} establish the entropy of Besov balls $\mathcal B^s_{\pi,\infty}(1)$, with $2/(1+2s)<\pi$, is $\nu=1/s$.
One can see that $\frac{\lambda v_n^2}{n}Pen(f)\sim\|f\|_{s,2,\infty}^2=J(f)$. Consequently the $\delta$-entropy of the functional set $\{f,J(f)\leq c\}$ can be bounded up to a constant by $\delta^{-1/s}$. We thus can apply Theorem~\ref{th1}.
\end{proof}
One drawback of this estimation procedure is its non adaptivity: the optimal value of the smoothing parameter $\lambda$ depends on the regularity $s$ of the function, which is unknown. Actually the way the penalty term is used cannot lead to adaptivity since it is closely linked to the norm of the functional space where the function $f$ lies. Another penalty type may be introduced.
\subsubsection{Adaptive case}
This section deals with the introduction of an adaptive penalty. We choose to use a $\ell^1$-penalty on the wavelet coefficients. It is well known that such a penalty leads to soft-thresholding wavelet estimators, which are adaptive. We refer to \cite{AntoniadisFan} for general theory on penalization on wavelets coefficients and to \cite{LoubesVanderGeer} for the use of the $\ell^1$ penalty in functional models.
Our asymptotic results for $\ell^1$ penalty is the following:
\begin{Theoreme}\label{th2}
Suppose assumptions (A0) to (A3) and assumptions (A4.2) and (A6) hold.
Let $\mbox{\boldmath$\beta$}$ be a given $p$-vector such that $\sqrt{n}\|\mbox{\boldmath$\beta$}-\mbox{\boldmath$\beta$}_0\|\leq c$.
Let $K_n$ be given by equation~(\ref{critere}), with the penalty: $Pen(f)=\lambda \sum_{i=i_S}^n|\theta_i^W|$ where $(\theta_{i}^W)$ are the wavelet coefficients of $f$. Define $\widetilde f_{\mbox{\boldmath$\beta$}}=\displaystyle \mathop{argmax}_{\{f,\,\|f\|_\infty\leq C_\infty\},\, \mbox{\boldmath$\beta$}\in\mathbb{R}^p} K_n(f,\mbox{\boldmath$\beta$}).$ Then, for $\lambda\sim \sqrt{\log(n)}$, we have $$\left(\frac{n}{\log(n)}\right)^{s/(1+2s)}\|\widetilde f_{\mbox{\boldmath$\beta$}}-f_0\|_n=\bigcirc_\mathbb{P}(1).$$
Define $\hat\mbox{\boldmath$\beta$}_n=\displaystyle \mathop{argmax}_{\mbox{\boldmath$\beta$}\in\mathbb{R}^p} K_n(\widetilde f_{\mbox{\boldmath$\beta$}},\mbox{\boldmath$\beta$})$. Then, $$v_n\|\hat \mbox{\boldmath$\beta$}_n-\mbox{\boldmath$\beta$}_0\|=\bigcirc_\mathbb{P}(1),$$
If in addition the covariates of the linear part admit a representation of the form given in assumption $(A_{corr})$, then:
$$\sqrt{n}\|\hat \mbox{\boldmath$\beta$}_n-\mbox{\boldmath$\beta$}_0\|=\bigcirc_\mathbb{P}(1).$$
The results still hold if the number $p$ of regression covariates goes to infinity provided the sequence $\left(\frac{n}{\log(n)}\right)^{(2s-1)/(1+2s)}p$ goes to 0 when $n$ goes to infinity and the sequence $\left(\frac{n}{\log(n)}\right)^{1/(1+2s)}p$ is bounded.
\end{Theoreme}
The proof is given in Appendix and relies on M-estimation techniques of \cite{VanderGeer00}.
As noted before, assumption (A4.2) is more restrictive than assumption (A4.1). The results of the Theorem could probably be extended to exponential tails distributions, weakening assumption (A4.2). We refer to discussion on page 134 of \cite{VanderGeer00} (Corollaries 8.3 and 8.8) for a discussion on the price to pay to release assumption (A4.2).
The adaptivity is acquired. As explained before, it gives the possibility of a computable procedure, without need of a cross-validation step. Yet, the parameter $\lambda$ is only chosen among an asymptotic condition and a finite sample application arises that the exact choice of this parameter is important. This will be discussed hereafter. Note also that the link with soft-thresholding is not evident, but appears in the iterative implementation of the estimators in next section.
The optimality of the functional part estimation is of course still available in a generalized functional model. In such models \cite{AntoniadisBesbeasSapatinas} and \cite{AntoniadisSapatinas} propose an adaptive estimation when the variance is respectively cubic or quadratic. These assumption have to be compared with (A3) and (A4.2).
Recently, \cite{BrownCaiZhou} introduced a method which consists in a transformation on the observations, based on the central limit theorem, in order to be able to use Gaussian framework's results. Yet, even if the asymptotic results are satisfactory, the implementation needs an important number of observations. We will see in numerical study that our procedure is easier to compute. Note also that it is available with the presence of the linear part.
\subsection{Algorithm}
This section is only devoted to the adaptive $\ell^1$-type penalty. Similar algorithms are available for other penalties but a cross-validation procedure should be elaborated because of the lack of adaptivity. It is worthy to see that the proposed estimators can be easily computed, by the way of iterative algorithms. A short simulation study is also given to evaluate the performance of the estimation with finite samples.
The paper of \cite{Muller01} concludes that a loglikelihood maximisation is preferable to a {\it profile likelihood} estimation (we recall this procedure consists in the maximisation of a modified loglikelihood criterion when estimating the parametric part). According to the author, {\it backfitting} computation gives in general better numerical values in estimation than others methods in GPLM. In a Gaussian framework, the same conclusion was observed in \cite{gannaz_these} (note that time differences with \cite{gannaz} are due to an improvement of the {\it backfitting} algorithm used). We therefore implement a {\it backfitting} algorithm.
{\bf Backfitting:} Let $\mbox{\boldmath$\beta$}^{(0)}$ be a given $p$-dimensional vector. For each iteration $k$, do:
\vspace{-\baselineskip}
\begin{description}
\item[Step 1] $f^{(k+1)}=\displaystyle \mathop{argmax}_f K(f,\mbox{\boldmath$\beta$}^{(k)})$
\item[Step 2] $\mbox{\boldmath$\beta$}^{(k+1)}=\displaystyle \mathop{argmax}_{\mbox{\boldmath$\beta$}} K(f^{(k+1)},\mbox{\boldmath$\beta$})$
\end{description}
\vspace{-0.5\baselineskip}
The algorithm is stopped either when a maximal number of iterations $\kappa$ is attained or when the algorithm is stabilized, {\it i.e.} when $\|\mbox{\boldmath$\beta$}^{(k)}-\mbox{\boldmath$\beta$}^{(k-1)}\|\leq \delta\|\mbox{\boldmath$\beta$}^{(k-1)}\|$ for a given tolerance value $\delta$. The returned values are $\hat\mbox{\boldmath$\beta$}_n=\mbox{\boldmath$\beta$}^{(K)}$ and $\hat f_n=f^{(K)}$ with $K$ maximal number of iterations of the algorithm.
To compute each of the two steps we apply a classical Fisher-scoring algorithm, detailed among others page 40 of \cite{MacCullaghNelder}. Usual in generalized models, this algorithm consists in building new variables of interest by a gradient descending method, in order to apply a ponderate regression on these new variables. Recall the notations of the GPLM given in equation~(\ref{GPLM}): one have $\eta(X,t)=X^T\mbox{\boldmath$\beta$}+f(t)$ and $\mu(X,t)=\dot b(\eta(X,t))$. We will omit the dependence to $(X,t)$ for the sake of simplicity.
\vspace{-\baselineskip}
\begin{description}
\item[Step 1:] ~\\ Note $f^{(k,0)}=f^{(k)}$.
Repeat the following iteration for $j=0\dots J_1-1$:\\
\vspace{-\baselineskip}
\begin{itemize}
\item[] Let $\eta^{(k,j)}=X\mbox{\boldmath$\beta$}^{(k)}+f^{(k,j)}$.
\item[] We introduce $Y^{(k,j)}=f^{(k,j)}+(y-\mu^{(k,j)})\left.\frac{d \eta}{d \mu}\right|_{\mu=\mu^{(k,j)}}$ and $W^{(k,j)}=\text{diag}\left(\left.\frac{d\eta}{d\mu}\right|_{\mu=\mu^{(k,j)}}\right)$.
\item[] With an $\ell^1$-penalty, we establish that $f^{(k,j+1)}$ is a nonlinear wavelet estimator for the observations $Y^{(k,j)}$, obtained by soft-thresholding of the wavelet coefficients. The threshold levels are $\lambda\Psi {W^{(k,j)-1}}\Psi^T\mbox{1\hspace{-.35em}1}_{n\times 1}$ where $\Psi$ denotes the forward wavelet transform and $\Psi^T$ the inverse wavelet transform.
\end{itemize}
Take $f^{(k+1)}=f^{(k,J_1)}$.
\item[Step 2:]~\\
Define $\mbox{\boldmath$\beta$}^{(k,0)}=\mbox{\boldmath$\beta$}^{(k)}$. Repeat the following iteration for $j=0\dots J_2-1$:\\
\vspace{-\baselineskip}
\begin{itemize}
\item[] Let $\widetilde \eta^{(k,j)}=X\mbox{\boldmath$\beta$}^{(k,j)}+f^{(k+1)}$.
\item[] We introduce $\widetilde Y^{(k,j)}=X\mbox{\boldmath$\beta$}^{(k)}+(y-\widetilde \mu^{(k,j)})\left.\frac{d \eta}{d \mu}\right|_{\mu=\widetilde \mu^{(k,j)}}$ and $\widetilde W^{(k,j)}=\left(\left.\frac{d\eta}{d\mu}\right|_{\mu=\widetilde \mu^{(k,j)}}\right)$.
\item[] Then $\mbox{\boldmath$\beta$}^{(k,j+1)}$ is the regression parameter of $Y^{(k,j)}$ on $X$ with ponderations $\widetilde W^{(k,j)}$, {\it i.e.} $\mbox{\boldmath$\beta$}^{(k,j+1)}=(X^T {\widetilde W}^{(k,j){-1}}X)^{-1}X^T {\widetilde W}^{(k,j){-1}}\widetilde Y^{(k,j)}$.
\end{itemize}
Take $\mbox{\boldmath$\beta$}^{(k+1)}=\mbox{\boldmath$\beta$}^{(k,J_2)}$.
\end{description}
\vspace{-\baselineskip}
Maximal number of iterations $J_1$ and $J_2$ can be fixed to $1$ to simplify the algorithm (this is what is proposed actually in \cite{Muller01}). In computation studies, no main difference is observed while modifying parameters $J_1$ and $J_2$. We therefore also decide to fix them equal to 1.
The initialization values are set as follows: for all $i=1,\ldots,n$, $f^{(0)}(t_i)=G(y_i)$, with $G$ the link function associated to the model (with a slight modification if the value does not exist) and for all~$j=1,\ldots,p$, $\mbox{\boldmath$\beta$}^{(0)}_j=0$.
In the particular Gaussian case, the two steps are non iterative. We explicit how the estimators are implemented in a Gaussian framework to ameliorate the comprehension of the algorithm:
\vspace{-\baselineskip}
\begin{description}
\item[Step 1:] The iterate $f^{(k+1)}$ is the wavelet estimator for the observations $y-X^T\mbox{\boldmath$\beta$}^{(k)}$, with a thresholding on wavelet coefficients with an uniform threshold level $\lambda$.
\item[Step 2:] We obtain $\mbox{\boldmath$\beta$}^{(k+1)}$ by a maximum likelihood estimation on $y-f^{(k+1)}$; this means that one have $\mbox{\boldmath$\beta$}^{(k+1)}=(X^T X)^{-1}X^T (Y-f^{(k+1)})$.
\end{description}
\vspace{-\baselineskip}
We can recognize the {\it backfitting} algorithm studied in \cite{ChangQu}, \cite{FadiliBullmore} and \cite{gannaz}. In a Gaussian framework, the variance in observations is constant and consequently the threshold level is uniform. In a generalized framework, the matrix $W$ ponderates the threshold level to take into account the inhomogeneity of the variance.
In generalized functional models, {\it i.e.}~without the presence of a linear part, the numerical implementation of the estimators proposed here has already been explored by \cite{Sardy}. The authors propose an interior point algorithm based on the dual maximisation problem. Comparing to this resolution scheme, our procedure has the advantage of an easy interpretation of the different steps in the algorithm. As noted previously, wavelet estimators have also been explored by \cite{AntoniadisBesbeasSapatinas}, \cite{AntoniadisSapatinas} and \cite{BrownCaiZhou}. These papers need to aggregate the data into a given number of bins. If the two first cited papers allow to choose small size of bins, it appears to be quite a constraint for the third one. Actually, it is worthy noticing the simplicity of our algorithm, {\it e.g.}~comparing with those implementations.
\subsection{Simulations}
In this subsection, we give some simulation results. All the calculations were carried out in MATLAB 7.6 on a Unix environment. For the DWT, we used the WaveLab toolbox developed by Donoho and his collaborators at the Statistics Department of Stanford University (see \url{http://www-stat.stanford.edu/~wavelab}).
We simulate $n = 2^8$ observations. The covariates are written according to assumption $(A_{corr})$: $x_i=g(i/n)+\mbox{\boldmath$\xi$}_i,$ with $\mbox{\boldmath$\xi$}_i$ independent and identically distributed variables following a standard normal distribution, and with $g(x)=30\,(x-0.5)^4-6\,(x-0.5)^2+(x-0.5)$. Three functional parts~$f$ were considered : we will refer at \begin{itemize}\vspace{-\baselineskip}\item the Sinus function, for a smooth function, combination of sinusoidal functions, \item the {\it Blocs} function for a piecewise constant function, \item and the {\it Pics} function for a function presenting high variations.\end{itemize} \vspace{-\baselineskip} These functions are given in Figure~\ref{functions}.
\figs{functions}{15.5cm}{Functional part of the generalized partially linear model. Figure~(a) corresponds to the {\it Sinus} function, Figure~(b) to the {\it Blocs} function and Figure~(c) to the {\it Pics} function.}
We will more precisely study the estimation quality for
\vspace{-\baselineskip}
\begin{itemize}
\item a Gaussian distribution; observations are $y_i\sim\mathcal N(\eta_i, \sigma^2)$, with $\sigma^2=\phi$
\item a Binomial distribution; observations are $y_i$ such that $y_i\times m\sim\mathcal B(\mu_i, m)$ with $m=\phi^{-1}$ and the logit link function $\mu_i=\frac{\exp(\eta_i)}{1+\exp(\eta_i)}$
\item a Poisson distribution; observations are $y_i\sim\mathcal P(\mu_i)$, with the link function $\mu_i=\exp(\eta_i)$,
\end{itemize}
\vspace{-\baselineskip}
as these distributions seem to be the most frequently encountered in modelisation.
To conclude with respect to the quality of estimation, we will give the mean value estimated for the parameter $\beta$ as well as the standard deviation of the estimation. For the nonparametric part, we will evaluate the MISE, which is the empirical estimate for $\mathbb{E}\left[\int_{[0,1]} \left(\hat f_n(t)-f_0(t)\right)^2dt\right]^{1/2}$. All results were obtained on $500$ simulations with the same covariates $x_i$ and the same functional parts~$f$. A maximal number of $\kappa=5000$ iterations was taken and the tolerance value defined above was equal to $\delta=10^{-20}$ when applying the algorithm.
\subsubsection{Preliminary study : choice of the threshold level}
The asymptotic behaviour of the estimators only defines the threshold level $\lambda$ up to a constant. Yet, numerical implementation needs to determine the exact threshold level $\lambda$ in the algorithm. In practice, one can see it has an important impact on the quality of estimation. Figure~\ref{min_gplm} gives the evolution of the MISE with respect to the threshold level in the GPLM with different distributions. One can observe that the threshold levels attaining the minima are very different among the distribution.
\figs{min_gplm}{15cm}{Evolution of the MISE with respect to the threshold level in a GPLM with the {\it Sinus} function. Figure (a) corresponds to a Gaussian distribution, Figure (b) to a binomial distribution, Figure (c) to a Poisson distribution and Figure (d) to an Exponential distribution.}
With a Gaussian distribution, following \cite{DonJohnKerkPic}, we choose $\lambda=\sqrt{2\phi\log(n)}$. This choice overevaluates the optimal threshold level in many cases but is the most often encountered in practice.
\newpage
\vspace{-0.2cm}
{\bf Binomial distribution}
Due to homogeneity reasons, it seems well-adapted to fix a threshold level of the form $\lambda=c\sqrt{\phi\log(n)}$, where $\phi$ corresponds to the dispersion parameter in distribution (\ref{modele}) and $c$ denotes a positive constant.
To ensure this conjecture in a Binomial setup, we plot the optimal threshold obtained for different value of the Binomial parameter $m$. The study was done on a generalized functional model (GFM) and on a GPLM. Figure~\ref{seuils_binomial} confirms that the chosen form seems appropriated.
\figs{seuils_binomial}{16cm}{Evolution of the threshold minimizing the MISE when estimating the {\it Sinus}, {\it Blocs} and {\it Pics} functions in a Binomial setup, with respect to $\sqrt{\phi}$. In a GFM on Figure (a) and in a GPLM in Figure(b). Calculations were done on $100$ samples of size $n=2^8$.}
\begin{table}[htbp]
\vspace{\baselineskip}
\centering
\begin{tabular}{@{} lccc @{}}
\toprule
\multicolumn{4}{c}{Generalized functional model}\\ \hline
Function & {\it Sinus} & {\it Blocs} & {\it Pics}\\
$R^2$ coefficient & 0.984 & 0.977 & 0.961\\
constant $c$ & 0.483 & 0.488 & 0.512\\
\midrule \\ [0.3cm]
\midrule
\multicolumn{4}{c}{Generalized partially linear model}\\ \hline
Function & {\it Sinus} & {\it Blocs} & {\it Pics}\\
$R^2$ coefficient & 0.947 & 0.981 & 0.989\\
constant $c$ & 0.510 & 0.507 & 0.498\\
\bottomrule
\end{tabular}
\caption{{\small Numerical indexes for the regression of the threshold level that minimizes the MISE with respect to $\sqrt{\phi \log(n)}$ with a Binomial distribution. Calculations were done on $100$ samples of size $n=2^8$.}}
\label{tab:seuils_binomial}
\end{table}
Linear fittings are given in Table~\ref{tab:seuils_binomial}. The linear fitting is coherent provided the $R^2$ coefficients. The constant $c$ obtained by linear fitting is varying with respect to the simulated samples, but the observed variations are small. The mean value is $0.4997$ and the standard deviation is $0.011$. The conjecture of a uniform constant seems acceptable. We therefore choose to take the value $c=0.5$ in the following for Binomial distributions. Note that due to the central limit theorem, one would have expected to take $c=\sqrt{2}$ for large values of the parameter $m$, but our simulation study show that this would lead to an oversmoothing for small values of $m$.
Note all the calculations were done with a fixed sample size $n=2^8$. To better evaluate the form of the optimal threshold in practice, one could also study the evolution of the threshold level with respect to the sample size $n$.
{\bf Poisson distribution}
Estimation in a Poisson functional model has been more intensively explored. Note that \cite{Sardy} propose a threshold level. The main drawback is that the level given depends on the estimated function. Yet, the choice is based on an universal large deviation inequality which does not seem to be well-adapted in this procedure. Indeed, due to the iterative interpretation of the estimation, the inhomogeneity of the variance of the observations is taken into account within the estimation.
Recently, \cite{ReynaudRivoirard} have developed a procedure based on wavelet hard-thresholding estimation for Poisson regression. In their estimation the thresholding step is defined directly and not through a penalization procedure like here. The authors then present a detailed numerical study showing the high instability of the optimal threshold level with respect to the estimated function.
In our procedure, we can hope for a better stability of the threshold level, considering the fact that it takes into account the variance of the pseudo-wavelet coefficients at each iteration.
Figure~\ref{seuils_poisson} represents the evolution of the threshold level which minimizes the MISE with respect to $\sqrt{\log(n)}$. Table~\ref{tab:seuils_poisson} gives some numerical results associated with Figure~\ref{seuils_poisson}.
\vspace{0.5\baselineskip}
\figs{seuils_poisson}{11cm}{Evolution of the optimal threshold with respect to $\sqrt{\log(n)}$ with a Poisson distribution. In a GFM on Figure (a) and in a GPLM in Figure(b). Calculations were done on $100$ samples of size $n=2^8$.}
\begin{table}[htbp]
\vspace{\baselineskip}
\centering
\begin{tabular}{@{} lcccc @{}}
\toprule
\multicolumn{4}{c}{Generalized functional model}\\ \hline
Sample size $n$ & 7 & 8 & 9 & 10\\
Mean value & 1.699 & 1.783 & 2.005 & 2.094\\
Standard deviation & 0.086 & 0.110 & 0.099 & 0.072 \\ [0.3cm]
\midrule \\ [0.3cm]
\midrule
\multicolumn{4}{c}{Generalized partially linear model}\\ \hline
Sample size $n$ & 7 & 8 & 9 & 10\\
Mean value & 1.584& 1.675 & 1.944 & 2.094\\
Standard deviation & 0.056 & 0.053 & 0.028 & 0.027\\
\bottomrule
\end{tabular}
\caption{{\small Mean value and standard deviation of the ratio between the threshold level that minimizes the MISE and $\sqrt{\log(n)}$ with a Poisson distribution. Calculations were done on $100$ samples for each function {\it Sinus}, {\it Blocs} and {\it Pics}.}}
\label{tab:seuils_poisson}
\end{table}
~
\vspace{-\baselineskip}
We observe that the optimal threshold level does not vary significantly with respect to the function estimated. This is an advantage on the estimation developed by \cite{ReynaudRivoirard}. This also confirms that the threshold level proposed by \cite{Sardy} is not adapted here. It seems in fact that the threshold level advised by \cite{Sardy} is more convenient for a uniform procedure like in \cite{ReynaudRivoirard} than for the iterative estimation scheme implemented here. Moreover it appears that the presence of a linear part in the model does not change the threshold level. This stability of the estimation procedure is a interesting property. Recall it can be explained by the evaluation of the variance of the pseudo-variables in the iterative algorithm.
Note that we do not obtain an optimal threshold level proportional to $\sqrt{\log(n)}$. As the theoretical result is asymptotic, we should study larger values of the sample size $n$. According to the results in Table~\ref{tab:seuils_poisson}, one may choose a constant approximatively equal to $2$ in a functional model and in a GPLM. We therefore choose to take the value $c=2$ in the following for Poisson distribution.
{\bf Remark:} In a Gaussian or a Binomial regression, one may need the dispersion parameter~$\phi$. Actually in literature, it is classically estimated at each iteration by $$\phi^{(k)}=\frac{1}{n}\sum_{i=1}^n\frac{(y_i-\mu_i^{(k)})^2}{\ddot b(\eta^{(k)})}.$$ Due to the bad quality of this estimator in GPLM, we prefer to consider in this paper that the dispersion parameter is known. In a Gaussian model, \cite{gannaz} proposed an efficient QR-based estimator for $\phi$. It would be worthy to explore whether it could be extended to generalized models.
\subsubsection{Example 1: Gaussian distribution}
Example 1 deals with a Gaussian model. The Gaussian distribution implementation is not a novelty for this estimation procedure, and we refer to \cite{ChangQu}, \cite{FadiliBullmore} and \cite{gannaz} for detailed studies on simulated or real values data. We briefly consider this case in order to have a comparison base for other distributions.
\vspace{0.5cm}
\figs{obs_gplm_gaussian}{16cm}{An example of a simulated data set in Example 1, with the function {\it Sinus} in Figure (a), the function {\it Blocs} in Figure (b) and the function {\it Pics} in Figure (c).}
The signal-to-noise-ratio (SNR) of a signal is defined as the norm of the ratio of the mean value with respect to the standard deviation. In GPLM the SNR for the nonparametric part, noted $SNR_f$ and the SNR for the linear part of the model, noted $SNR_{\mbox{\boldmath$\beta$}}$, are respectively equal to \begin{eqnarray*} SNR_f^2&=&\frac{1}{n}\sum_{i=1}^n \frac{f_0(t_i)^2}{\phi\,\ddot b\left(\mbox{${\mathbf X}$}_i^T\mbox{\boldmath$\beta$}_0+f_0(t_i)\right)},\\
\text{and~~~}SNR_{\mbox{\boldmath$\beta$}}^2&=&\frac{1}{n}\sum_{i=1}^n \frac{(\mbox{${\mathbf X}$}_i^T\mbox{\boldmath$\beta$}_0)^2}{\phi\,\ddot b\left(\mbox{${\mathbf X}$}_i^T\mbox{\boldmath$\beta$}_0+f_0(t_i)\right)}.
\end{eqnarray*}
With a high SNR, say approximatively $9$, one can expect a good quality of estimation, while with a small value, like $2$, the quality of estimation cannot be satisfying
Table~\ref{tab:table_gauss} gives quality indexes for the estimation of the model on 500 simulations. As already noted in the different papers using this estimation procedure, the numerical results appear to be of fairly good quality.
\vspace{0.5cm}
\begin{table}[htbp]
\centering
\begin{tabular}{@{} l@{\hspace{1cm}} cc @{\hspace{1cm}} cc @{\hspace{1cm}}r @{}}
\toprule
Function & \multicolumn{2}{l}{Estimation of $\mbox{\boldmath$\beta$}$} & \multicolumn{2}{l}{Estimation of $f$} & Time\\
\hlin
& Mean SNR & Mean estimation & Mean SNR & MISE & \\
\midrule
{\it Sinus} & 8.7 & 0.9993(0.0098) & 9.1 & 0.1639 & 1.31(326)\\
{\it Blocs} & 6.3 & 0.9991(0.0168) & 9.0 & 0.2197 & 0.49(118)\\
{\it Pics} & 2.2 & 1.0002(0.0296) & 1.6 & 0.3367 & 0.64(153)\\
\bottomrule
\end{tabular}
\caption{{\small Measures of quality the estimates over the 500 simulations in Example 1 with $n=2^{8}$.
For the parameter $\mbox{\boldmath$\beta$}$, the true value is $1$ and the value given is the mean value and standard deviation appears in brackets. In the column Time, the mean of numbers of iterations in the algorithm is given in brackets
}}
\label{tab:table_gauss}
\end{table}
\vspace{0.5cm}
\figs{gplm_gaussian}{16cm}{An example of estimation of the nonparametric part in Example 1, with the function {\it Sinus} in Figure (a), the function {\it Blocs} in Figure (b) and the function {\it Pics} in Figure (c).}
For the value of the signal-to-noise ratio ($SNR_f=9$) adopted in our simulations for the nonparametric part, the estimator is nearly able to detect the discontinuity of the {\it Sinus} function, as shown in Figure~\ref{gplm_gaussian}~(a). Results for the nonparametric part are very similar to those obtained in a nonparametric signal denoising, without a linear part. The asymptotic behaviour of the estimates effectively states that the presence of the linear part does not affect the estimation of the nonparametric part, under assumption $(A_{corr})$. See \cite{gannaz} for a more argued discussion on the influence of the linear part on estimation, explained through a parallel established with robust M-estimates.
\vspace{-0.5cm}
\subsubsection{Example 2: Binomial distribution}
\vspace{-0.3cm}
In Example 2, we consider a Binomial distribution: observations $Y_i$ are such that $Y_i\times m$ are independently drawn from Binomial distributions $\mathcal B(\mu_i, m)$ with the parameter $m$ equal to $m=\phi^{-1}$. The link function considered is the logistic. The mean is thus equal to $$\mu_i=\frac{\exp(\mbox{${\mathbf X}$}_i^T\mbox{\boldmath$\beta$}_0+f_0(t_i))}{1+\exp(\mbox{${\mathbf X}$}_i^T\mbox{\boldmath$\beta$}_0+f_0(t_i))}.$$
The logistic link makes sense if the canonical parameter $\eta(\cdot)$ belongs to the interval $[-4,4]$, as one can see for example page 28 of \cite{FahrmeirTutz}. Consequently, to get a SNR of order $9$ one may choose a parameter $m$ in the binomial distribution equal approximatively to $200$. Note that Binomial distribution is often chosen to classify the data. A number of $200$ classes is not adapted for real data applications ; actually, a classical number of classes is 2, 3 or 4, which will lead to small SNRs and thus to a bad quality of estimation.
\figs{obs_gplm_binomial}{16cm}{An example of the observations obtained on a simulation in Example 2, with the function {\it Sinus} in Figure(a), the function {\it Blocs} in Figure (b) and the function {\it Pics} in Figure (c).}
Due to this remark, we choose here to make a compromise and to apply the algorithm with a parameter $m$ equal to $24$, which corresponds to $25$ classes. This choice is not coherent with a classification problem but allows to consider much reasonable SNRs. Indeed, with small values of $m$ one is not in capacity of identifying the functional part in a GPLM.
The observations of a simulated sample are represented in Figure~\ref{obs_gplm_binomial}. The results for this example are summarized in Table~\ref{tab:table_bino}.
\vspace{0.5cm}
\begin{table}[htbp]
\centering
\begin{tabular}{@{} l@{\hspace{1cm}} cc @{\hspace{1cm}} cc @{\hspace{1cm}}r @{}}
\toprule
Function & \multicolumn{2}{l}{Estimation of $\mbox{\boldmath$\beta$}$} & \multicolumn{2}{l}{Estimation of $f$} & Time\\
\hlin
& Mean SNR & Mean estimation & Mean SNR & MISE & \\
\midrule
{\it Sinus} & 2.1 & 0.8559(0.0597) & 3.0 & 0.7656 & 17.0(3922)\\
{\it Blocs} & 2.0 & 0.8581(0.0607) & 2.9 & 0.7285 & 18.1(4142)\\
{\it Pics} & 1.9 & 0.9487(0.0426) & 1.4 & 0.4831 & 9.8(1886)\\
\bottomrule
\end{tabular}
\caption{{\small Measures of quality the estimates over the 500 simulations in Example 2 with $n=2^{8}$.
For the parameter $\mbox{\boldmath$\beta$}$, the true value is $1$ and the value given is the mean value and standard deviation appears in brackets. In the column Time, the mean of numbers of iterations in the algorithm is given in brackets
}}
\label{tab:table_bino}
\end{table}
\vspace{0.5cm}
The bad quality of estimation is due to the bad SNRs of the model. In order to better analyse the results, we simulate a GPLM with a Gaussian distribution and the same SNRs. This was done by modifications on the dispersion parameter or choosing proportional covariates in the linear part. We obtain results given in Table~\ref{tab:table_gauss_bino}.
Comparing Table~\ref{tab:table_bino} and Table~\ref{tab:table_gauss_bino}, it appears that the estimation quality is better with a Gaussian framework than with a Binomial. Yet, the estimates in the Binomial GPLM seems satisfying for such SNRs. Figure~\ref{gplm_binomial} confirms that the estimate of the functional part only renders the shape of the function but does not allow to recover the function. Again this is in coherence with the small SNR.
\begin{table}[htbp]
\centering
\begin{tabular}{@{} l@{\hspace{1cm}} cc @{\hspace{1cm}} cc @{\hspace{1cm}}r @{}}
\toprule
Function & \multicolumn{2}{l}{Estimation of $\mbox{\boldmath$\beta$}$} & \multicolumn{2}{l}{Estimation of $f$} & Time\\
\hlin
& Mean SNR & Mean estimation & Mean SNR & MISE & \\
\midrule
{\it Sinus} & 2.3 & 1.0008(0.0333) & 3.0 & 0.4219 & 1.04(253)\\
{\it Blocs} & 1.9 & 1.0033(0.0373) & 2.9 & 0.4547 & 1.87(451)\\
{\it Pics} & 2.2 & 1.0002(0.0296) & 1.6 & 0.3367 & 0.64(153)\\
\bottomrule
\end{tabular}
\caption{{\small Measures of quality the estimates over the 500 simulations with a Gaussian distribution with similar SNRs than in Example 2 with $n=2^{8}$.
For the parameter $\mbox{\boldmath$\beta$}$, the true value is $1$ and the value given is the mean value and standard deviation appears in brackets. In the column Time, the mean of numbers of iterations in the algorithm is given in brackets
}}
\label{tab:table_gauss_bino}
\end{table}
\figs{gplm_binomial}{16cm}{An example of estimation of the nonparametric part in Example 2, with the function {\it Sinus} in Figure(a), the function {\it Blocs} in Figure (b) and the function {\it Pics} in Figure (c).}
\newpage
When comparing the time costs, one can see that the algorithm is faster in a Gaussian framework. The reason is the fact that each step of the {\it backfitting} algorithm described above are exact calculation in a Gaussian case. While with a Binomial distribution each of these steps has to be solved iteratively. The number of iterations necessary to stabilize the algorithm is thus higher with the Binomial distribution
This can also explain the lower quality of estimation in the Binomial distribution : the mean number of iterations in the algorithm is higher than 3900 with the {\it Sinus} and the {\it Blocs} functions, where the difference with the Gaussian distribution is the most important. Recall the maximal number of iterations is 5000. Thus perhaps sometimes the algorithm is not stabilized when it stops, explaining the lower quality.
\subsubsection{Example 3: Poisson distribution}
In Example 3, we consider a Poisson distribution: $y_i\sim\mathcal P(\mu_i)$ with $\mu_i=\exp(\eta_i)$. Observations of a simulated data sample are represented in Figure~\ref{obs_gplm_poisson}.
The results for this example are summarized in Table~\ref{tab:table_poisson}.
\vspace{0.5cm}
\figs{obs_gplm_poisson}{16cm}{An example of the observations obtained on a simulation in Example 3, with the function {\it Sinus} in Figure(a), the function {\it Blocs} in Figure (b) and the function {\it Pics} in Figure (c).}
\vspace{0.5cm}
\vspace{0.5cm}
\begin{table}[htbp]
\centering
\begin{tabular}{@{} l@{\hspace{1cm}} cc @{\hspace{1cm}} cc @{\hspace{1cm}}r @{}}
\toprule
Function & \multicolumn{2}{l}{Estimation of $\mbox{\boldmath$\beta$}$} & \multicolumn{2}{l}{Estimation of $f$} & Time\\
\hlin
& Mean SNR & Mean estimation & Mean SNR & MISE & \\
\midrule
{\it Sinus} & 9.1 & 1.0971(0.0763) & 7.9 & 0.5548 & 11.4(2701)\\
{\it Blocs} & 3.1 & 1.2020(0.1201) & 8.2 & 0.4156 & 18.8(4479)\\
{\it Pics} & 3.0 & 1.1876(0.1751) & 7.1 & 0.4809 & 19.6(4528)\\
\bottomrule
\end{tabular}
\caption{{\small Measures of quality the estimates over the 500 simulations in Example 3 with $n=2^{8}$.
For the parameter $\mbox{\boldmath$\beta$}$, the true value is $1$ and the value given is the mean value and standard deviation appears in brackets. In the column Time, the mean of numbers of iterations in the algorithm is given in brackets
}}
\label{tab:table_poisson}
\end{table}
\vspace{0.5cm}
Similarly to what was done in Example 2, we also simulate data sets with similar SNRs with a Gaussian distribution, in order to be able to compare the qualities of the estimates to a given reference. Results obtained in a Gaussian case are given in Table~\ref{tab:table_gaussian_poisson}.
\begin{table}[htbp]
\centering
\begin{tabular}{@{} l@{\hspace{1cm}} cc @{\hspace{1cm}} cc @{\hspace{1cm}}r @{}}
\toprule
Function & \multicolumn{2}{l}{Estimation of $\mbox{\boldmath$\beta$}$} & \multicolumn{2}{l}{Estimation of $f$} & Time\\
\hlin
& Mean SNR & Mean estimation & Mean SNR & MISE & \\
\midrule
{\it Sinus} & 9.3 & 0.9990(0.0101) & 7.9 & 0.1875 & 0.61(147)\\
{\it Blocs} & 3.2 & 0.9943(0.0311) & 8.2 & 0.2261 & 1.39(336)\\
{\it Pics} & 3.0 & 0.9957(0.0332) & 7.1 & 0.1416 & 0.74(178)\\
\bottomrule
\end{tabular}
\caption{{\small Measures of quality the estimates over the 500 simulations with a Gaussian distribution with similar SNRs than in Example 3 with $n=2^{8}$.
For the parameter $\mbox{\boldmath$\beta$}$, the true value is $1$ and the value given is the mean value and standard deviation appears in brackets. In the column Time, the mean of numbers of iterations in the algorithm is given in brackets
}}
\label{tab:table_gaussian_poisson}
\end{table}
\newpage
The quality of estimation seems good, even if lower than in the Gaussian distribution framework. Like for the Binomial distribution, each step of the {\it backfitting} algorithm is solved by Fisher-scoring. Consequently, the time of calculation is higher than in the Gaussian model.
Like with the Binomial distribution, the lowest qualities are observed when the mean of iteration numbers is high. Clearly, the fact that the mean of iteration numbers is more than 4400, with still a maximum at 5000, means that the algorithm often would have need more iterations to get stabilized. Thus the quality should probably be better with more iterations
\figs{gplm_poisson}{16cm}{An example of estimation of the nonparametric part in Example 3, with the function {\it Sinus} in Figure (a), the function {\it Blocs} in Figure (b) and the function {\it Pics} in Figure (c).}
Figure~\ref{gplm_poisson} gives an example of the estimation of the functional part obtained. One can see that visually the quality is satisfying. Note that one cannot compare with the Binomial case because of the difference in SNRs.
\section*{Conclusion}
\vspace{-0.5cm}
This paper proposes a penalized loglikelihood estimation in generalized partially linear models. With appropriate penalties, we establish the asymptotic near-minimaxity of the estimators for both the linear and the nonparametric parts of the model. The result holds for a large class of functions, possibly non smooth and the conditions of correlation between the covariate design of the linear part and the functional part are similar to literature's. Moreover with an $\ell^1$ penalty on the wavelet coefficients of the function (leading to soft thresholding), the procedure appears to be adaptive relatively to the smoothness of the function. In this particular case we develop an implementation and observed satisfactory results on simulation studies.
Our ongoing research may deal with the estimation of the dispersion parameter in generalized partially linear models. Developments in non-equidistant designs for the nonparametric part should also be explored.
\section*{Acknowledgements}
\vspace{-0.5cm}
Pr. Antoniadis is gratefully acknowledged for his constructive comments and fruitful discussions. The author would also like to thank Pr. Gadat for giving some helpful references.
{
\bibliographystyle{stylebib2
|
\section{Introduction}
Quantum state estimation \cite{PR2004b,H1997,B-K2010} remains one of the hot topics in the field of quantum information processing. The hope to recover each element in the density matrix, however, is impeded by the exponential growth of the number of matrix elements with the number of qubits, and the concomitant exponential growth in time and memory required to compute and store the density matrix. The task can become intimidating when 14 qubits are involved \cite{MSBCNCHHHB2010}, and so efforts have been made to simplify quantum state tomography. One such effort focused on states that have high purity \cite{GLFBE2010} so that the size of the state space shrinks significantly (from ${\cal O}(D^2)$ to ${\cal O}(D)$ for a system described by a $D$ dimensional Hilbert space). Given that the measurement record is used to verify the assumptions made initially, this method avoids the trap of simplification through imposing {\em a priori} assumptions merely by {\em fiat}. Another recent effort \cite{CPFSGBL-CPL2010} in the same spirit considered multi-qubit states that are well represented by matrix product states \cite{KSZ1991,FNW1992,P-GVWC2007} (which require a number of parameters growing only polynomially with the number of qubits). Many states of interest, such as ground states of certain model Hamiltonians in condensed-matter physics, are of that form.
Crucially, the particular form of the state can be verified by the data.
Here we go one step further, and we will allow, tentatively, {\em any} parametrized form for the density matrix of the quantum system to be tested, possibly containing just a few parameters. In fact, we may have several different tentative ideas of how our quantum state is best parameterized. The questions are then, how the data reveal which of those descriptions work sufficiently well, and which description is the best.
This idea corresponds to a well-developed field in statistics: model selection \cite{J1961b,BA2002b,Z2000}. All mathematical descriptions of reality are in fact models (and a quantum state, pure or mixed, is an excellent example of a model), and they can be evaluated by judging their performance relative to that of the true model (assuming it exists). In order to quantify this relative performance, we will make use of the Kullback-Leibler divergence (aka mutual information, aka cross entropy, aka relative entropy) \cite{KL1951}, which has the interpretation of the amount of information lost when a specific model is used rather than the true model.
Based on the minimization of the Kullback-Leibler divergence over different models, the Akaike Information Criterion (AIC) \cite{A1974} was developed as a ranking system so that models are evaluated with respect to each other, given measurement data. The only quantities appearing in the criterion are the maximum likelihood obtainable with a given model (i.e., the probability the observed data would occur according to the model, maximized over all model parameters), and the number of independent parameters of the model.
The minimization does not require any knowledge of the true model, only that the testing model is sufficiently close to the true model. The legitimate application of AIC should, therefore, in principle be limited to ``good'' models, ones that include the true model (in our case, the exact quantum state that generated the data), at least to a very good approximation. However this does not prevent one from resorting to the AIC for model evaluation when there is no such guarantee. In fact, Takeuchi studied the case where the true model does not belong to the model set and came up with a more general criterion, named the Takeuchi Information Criterion, TIC \cite{T1976}. However the estimation of the term introduced by Takeuchi to counterbalance the bias of the maximum likelihood estimator used in the AIC, requires estimation of a $K\times K$ matrix ($K$ being the number of independent parameters used by a model) from the data, which, unfortunately, is prone to significant error. This reduces the overall charm and practical use of the TIC. Since in most cases the AIC is still a good approximation to the TIC \cite{BA2002b}, especially in the case of many data, we stick to the simpler and more robust criterion here.
Information criteria are designed to produce a relative (rather than absolute) ranking of models, so that fixing a reference model is convenient. Throughout this paper we choose the ``full-parameter model'' (FPM) as reference, that is, a model with just enough independent variables to fully parameterize the measurement on our quantum system. For tomographically complete measurements (discussed in detail in Sec.~\ref{sec_tomography}) the number of independent variables is given by the number of free parameters in the density matrix ($2^{2D}-1$ for a $D$-dimensional Hilbert space). For tomographically incomplete measurements (see Sec.~\ref{sec_witness}), the number of independent variables of FPM is smaller, and equals the number of independent observables. We will, in fact, not even need the explicit form of the FPM (which may be hard to construct for tomographically incomplete measurements), as its maximum possible likelihood can be easily upper-bounded.
We should note an important distinction between maximum likelihood estimation (MLE) \cite{BDPS1999}, a technique often used in quantum tomography, and the method of information criteria and model selection. MLE produces the state that fits the data best. Now the data inevitably contains (statistical) noise, and the MLE state predicts, incorrectly, that same noise to appear in future data. Information criteria, on the other hand, have been designed to find the model that best predicts future data, and tries, in particular, to avoid overfitting to the data, by limiting the number of model parameters. This is how a model with a few parameters can turn out to be the best predictive model, even if, obviously, the MLE state will fit the (past) data better.
We also note that information criteria have been applied mostly in areas of research outside of physics. This is simply due to the happy circumstance that in physics we tend to know what the ``true'' model underlying our observations is (or should be), whereas this is much less the case in other fields. Within physics, information criteria have been applied to astrophysics \cite{L2007}, where one indeed may not know the ``true'' model (yet), but also to the problem of entanglement estimation \cite{LvE2009}. In the latter case (and in quantum information theory in general) the problem is not that we do not know what the underlying model is, but that that model may contain far too many parameters. Hence the potential usefulness of information criteria. And as we recently discovered, the AIC has even been applied to quantum state estimation, not for the purpose of making it more efficient, but making it more accurate, by avoiding overfitting \cite{UNTMN2003}.
\section{The Akaike Information Criterion - A Schematic Derivation}
\label{sec_AIC}
Suppose we are interested in measuring certain variables, summarized as a vector $\mathbf{x}$, and their probability of occurrence as outcome of our measurement. We denote $f(\mathbf{x})$ as the probabilistic model that truthfully reflects reality (assuming for convenience that such a model exists) and $g(\mathbf{x}|\vec{\theta})$ as our (approximate) model characterized by one or more parameters, summarized as a vector $\vec{\theta}$. The models satisfy the normalization condition $\int {\rm d}\mathbf{x}f(\mathbf{x})=\int{\rm d}\mathbf{x}g(\mathbf{x}|\vec{\theta})=1$ for all $\vec{\theta}$. By definition, we say there is no information lost when $f(\mathbf{x})$ is used to describe reality. The amount of information lost when $g(\mathbf{x}|\vec{\theta})$ is used instead of the true model is defined to be the Kullback-Leibler divergence \cite{KL1951} between the model $g(\mathbf{x}|\vec{\theta})$ and the true model $f(\mathbf{x})$:
\begin{align}
I(f,g_{\vec{\theta}})=&\int {\rm d}\mathbf{x} f(\mathbf{x})\log(f(\mathbf{x}))\nonumber\\
&-\int {\rm d}\mathbf{x} f(\mathbf{x})\log(g(\mathbf{x}|\vec{\theta})).
\label{eq_Ifg_int}
\end{align}
Eq.~(\ref{eq_Ifg_int}) can be conveniently rewritten as
\begin{align}
I(f,g_{\vec{\theta}})=E_\mathbf{x}\left[\log(f(\mathbf{x}))\right]-E_\mathbf{x}\left[\log(g(\mathbf{x}|\vec{\theta}))\right],
\label{eq_Ifg_exp}
\end{align}
where $E_\mathbf{x}[\cdot]$ denotes an estimate with respect to the true distribution $f(\mathbf{x})$.
We see that $\mathbf{x}$ is no longer a variable in the above estimator, as we integrated it out. The only variable that affects $I(f,g_{\vec{\theta}})$ is $\vec{\theta}$. Since the first term in Eq.~(\ref{eq_Ifg_int}) is irrelevant to the purpose of rank-ordering different models $g$ (not to mention we cannot evaluate it when $f$ is not known), we only have to consider the second term. Suppose there exists $\vec{\theta}_0$ such that $g(\mathbf{x}|\vec{\theta}_0)=f(\mathbf{x})$ for every $\mathbf{x}$, that is, the true model is included in the model set. Note that for this to hold, $\vec{\theta}$ does not necessarily contain the same number of parameters as the dimension of the system. To use a simpler notation without the integration over $\mathbf{x}$ we denote the second term in Eq.~(\ref{eq_Ifg_int}) (without the minus sign) as
\begin{align}
S(\vec{\theta}_0:\vec{\theta})=\int {\rm d}\mathbf{x} g(\mathbf{x}|\vec{\theta}_0)\log(g(\mathbf{x}|\vec{\theta})),
\label{eq_S_int}
\end{align}
where we have used $g(\mathbf{x}|\vec{\theta}_0)$ to represent the true model $f(\mathbf{x})$. The advantage of this estimator is that it can be approximated without knowing the true distribution $f(\mathbf{x})$. To do that we first consider the situation where $\vec{\theta}$ is close to $\vec{\theta}_0$. This assumption can be justified in the limit of large $N$, $N$ being the number of measurement records, since the model $\vec{\theta}$ ought to approach $\vec{\theta}_0$ asymptotically (assuming, for simplicity, $\vec{\theta}_0$ is unique). We know that $S(\vec{\theta}_0:\vec{\theta})$ must have a maximum when $\vec{\theta}=\vec{\theta}_0$, and we may then symbolically expand $S(\vec{\theta}_0:\vec{\theta})$ in the vicinity of $\vec{\theta}_0$ by
\begin{align}
S(\vec{\theta}_0:\vec{\theta})= S(\vec{\theta}_0:\vec{\theta}_0)&-\frac{1}{2}||\vec{\theta}-\vec{\theta}_0||_{\vec{\theta}_0}^2\nonumber\\
&+{\cal O}\left(||\vec{\theta}-\vec{\theta}_0||_{\vec{\theta}_0}^{3/2}\right),
\label{eq_S_theta0-theta}
\end{align}
where
\begin{align}
||\vec{\theta}-\vec{\theta}_0||_{\vec{\theta}_0}^2=\left(\vec{\theta}-\vec{\theta}_0\right)'\cdot\left.\frac{\partial^2S(\vec{\theta}_0:\vec{\theta})}{\partial\vec{\theta}^2}\right|_{\vec{\theta}=\vec{\theta}_0}\cdot\left(\vec{\theta}-\vec{\theta}_0\right)
\end{align}
denoting a \emph{squared length} derived from a metric defined at $\vec{\theta}_0$. It can be proved that when $N$ is sufficiently large
$||\vec{\theta}-\vec{\theta}_0||_{\vec{\theta}_0}^2$ can be approximated by the $\chi_K^2$ distribution, with $K$ equal to the number of independent parameters used by the model $\vec{\theta}$. From the properties of the $\chi_K^2$ distribution, we know the average value of $||\vec{\theta}-\vec{\theta}_0||_{\vec{\theta}_0}^2$ will approach $K$.
The next step is to evaluate the estimator $S(\vec{\theta}_0:\vec{\theta}_0)$, where $\vec{\theta}_0$ is now considered a variable. Suppose we find the maximum likelihood estimate $\vec{\theta}_M$ from the measurement outcomes such that $S(\vec{\theta}_M:\vec{\theta}_M)$ is the maximum. Now $\vec{\theta}_M$ should also be close to the true model $\vec{\theta}_0$, when $N$ is sufficiently large. Therefore we can similarly expand $S(\vec{\theta}_0:\vec{\theta}_0)$ in the vicinity of $\vec{\theta}_M$ as
\begin{align}
S(\vec{\theta}_0:\vec{\theta}_0)= S(\vec{\theta}_M:\vec{\theta}_M)&-\frac{1}{2}||\vec{\theta}_0-\vec{\theta}_M||_{\vec{\theta}_M}^2\nonumber\\
&+{\cal O}\left(||\vec{\theta}-\vec{\theta}_0||_{\vec{\theta}_M}^{3/2}\right),
\label{eq_S_theta0-theta0}
\end{align}
$||.||_{\vec{\theta}_M}$ is a length similarly defined as in Eq.~(\ref{eq_S_theta0-theta}) and has the same statistical attributes
as $||\vec{\theta}-\vec{\theta}_0||_{\vec{\theta}_0}^2$ since $\vec{\theta}_0$ is related to $\vec{\theta}_M$ the same way $\vec{\theta}$ is related to $\vec{\theta}_0$ and $\vec{\theta}_M$ is very close to $\vec{\theta}_0$. Its average value, therefore, approaches again $K$, according to the $\chi_K^2$ distribution. Thus we are able to rewrite Eq.~(\ref{eq_S_int}) as
\begin{align}
S(\vec{\theta}_0:\vec{\theta})\approx S(\vec{\theta}_M:\vec{\theta}_M)-K.
\end{align}
We see that now our target estimator $S(\vec{\theta}_0:\vec{\theta})$ is evaluated by the MLE solution $\vec{\theta}_M$ only (plus the number of parameters $K$ of the model), with no knowledge of what the true model $f$ is. The assumption that underlies this convenience is constituted by two parts: estimating $S(\vec{\theta}_0:\vec{\theta})$ with its maximum $\vec{\theta}_0$ and estimating $S(\vec{\theta_0}:\vec{\theta}_0)$ from the data by its optimum $\vec{\theta}_M$. The deviations from their respective maxima are equal and result simply in the appearance of the constant $K$.
We now denote $\L_M=S(\vec{\theta}_M:\vec{\theta}_M)$, which is the maximum likelihood obtainable by our model, with respect to a given set of measurement records. The AIC is then defined by
\begin{align}
{\rm AIC}=-2\L_M+2K.\label{eq_AIC}
\end{align}
Apart from the conventional factor 2, and a constant independent of the model $\vec{\theta}$, AIC is an estimator of the quantity in Eq.~(\ref{eq_Ifg_int}) we originally considered, that is, the Kullback-Leibler divergence between a model that is used to describe the true model and the true model itself. Therefore a given model is considered better than another if it has a {\em lower} value of AIC.
Finally, in the case that $N$ is not so large yet that asymptotic relations hold to a very good approximation, one
can include a correction factor to the AIC taking the deviation from asymptotic values into account. The corrected AIC gives rise to a slightly different criterion \cite{Sug1978}:
\begin{align}
{\rm AICc}=-2\L_M+2K+\frac{2K(K+1)}{N-K-1}.\label{eq_AICc}
\end{align}
\section{Results}
\label{sec_results}
\subsection{Dicke states}
\label{sec_Dicke}
We will apply the AIC to measurements on a popular family of entangled states, the Dicke states of four qubits \cite{D1954,HHRBCCKRRSBGDB2005,KSTSW2007,PCDLvEK2009,CGPvEK2010}. We simulate two different experiments, one tomographically complete experiment, another measuring an entanglement witness. We include imperfections of a simple type, and we investigate how model selection, according to the AIC, would work. We consider cases where we happen to guess the correct model, as well as cases where our initial guess is, in fact, incorrect.
We consider the four-qubit Dicke states with one or two excitations $\ket{D_4^{1,2}}$ (with the state $\ket{1}$ representing an excitation):
\begin{subequations}
\begin{align}
\ket{D_4^1}=&\left(\ket{0001}+\ket{0010}+\ket{0100}+\ket{1000}\right)/2,\\
\ket{D_4^2}=&\left(\ket{0011}+\ket{0101}+\ket{0110}+\ket{1001}\right.\nonumber\\
&+\left.\ket{1010}+\ket{1100}\right)\sqrt{6}.
\end{align}
\end{subequations}
For simplicity, let us suppose that white noise is the only random noise in the state generation,
and that it corresponds to mixing of the ideal state with the maximally mixed state of the entire space (instead of the subspace with exactly one or two excitations, which could be a reasonable choice, too, depending on the actual implementation of the Dicke states). We thus write the states under discussion as
\begin{align}
\rho^{1,2}(\alpha)=(1-\alpha)\ket{D_4^{1,2}}\bra{D_4^{1,2}}+\alpha\openone/D,
\end{align}
where $\openone/D$ is the maximally mixed state for dimension $D=2^4$, and $0\leq \alpha\leq 1$. We will fix the actual state generating our data to be
\begin{align}
\rho_{\rm actual}^{1,2}=\rho^{1,2}(\alpha=0.2).
\end{align}
This choice is such that the mixed state is entangled (as measured by our multi-qubit version of the negativity, see below), even though the entanglement witness whose measurement we consider later in Sec.~\ref{sec_witness}, just fails to detect it.
For our first model (to be tested by AIC) we wish to pick a one-parameter model (so, $K=1$) that also includes a wrong guess.
A straightforward model choice, denoted by $M_{1\phi}$, is
\begin{align}
M_{1\phi}: \rho_\phi^{1,2}(q)=(1-q)\ket{\Psi_{\rm target}^{1,2}(\phi)}&\nonumber\\
\bra{\Psi_{\rm target}^{1,2}(\phi)}+&q\openone/D.
\label{eq_M1_SICPOVM}
\end{align}
We refer to the pure states appearing here as the {\em target} states $\ket{\Psi_{\rm target}^{1,2}(\phi)}$, simulating the case where we (possibly incorrectly) think we would be creating a pure state of that form, if only the white noise were absent ($q=0$). The phase $\phi$ is included not as a (variable) parameter of the model but as an inadvertently mis-specified property. In this case, it stands for us being wrong about a single relative phase in one of the qubits in state $|1\rangle$.
Without loss of generality we assume the first qubit in our representation to carry the wrong phase, and we write
\begin{subequations}
\begin{align}
\ket{\Psi_{\rm target}^1(\phi)}=&\frac{1}{2}\left(\ket{0001}+\ket{0010}+\ket{0100}\right.\nonumber\\
&+\left.e^{i\phi}\ket{1000}\right),\label{eq_tilde1}\\
\ket{\Psi_{\rm target}^2(\phi)}=&\frac{1}{\sqrt{6}}\left[\ket{0011}+\ket{0100}+\ket{0110}\right.\nonumber\\
+&\left.e^{i\phi}\left(\ket{1001}+\ket{1010}+\ket{1100}\right)\right].\label{eq_tilde2}
\end{align}
\label{eq_tilde}
\end{subequations}
Alternatively, if we {\em do} consider this a two-parameter model (changing $K=1$ to $K=2$), then $\phi$ is variable, and we would optimize over $\phi$. In our case, this optimum value should always be close to $\phi=0$.
\begin{figure}[t]
\begin{center}
\subfigure[$M_{1\phi}$.]
{\includegraphics[width=195pt]{D41}}
\\
\subfigure[$M_{2\phi}$.]
{\includegraphics[width=195pt]{D41_2bitflip_1e4}}
\end{center}
\caption{{\em How AIC ranks the one- and two-parameter models vs. the full-parameter model (FPM):} \\
Plot of the difference between AIC values of our models and the FPM, \textsl{i.e.}, $-\Delta{\rm AIC}={\rm AIC(FPM)}-{\rm AIC}(M_{1\phi})$ or $-\Delta{\rm AIC}={\rm AIC(FPM)}-{\rm AIC}(M_{2\phi})$, for various numbers of SIC-POVM measurements, $N$, with $\ket{\Psi_{\rm target}^1}$ as the target state, as functions of the angle $\phi$. The horizontal line demarcates $\Delta{\rm AIC}=0$: points above (below) that line correspond to cases where the model with fewer (more) parameters is preferred. The figures with $\ket{\Psi_{\rm target}^2}$ as the target state look very similar (see FIG.~\ref{fig_AIC_M1_1e4_SICPOVM} for an example of this similarity).}
\label{fig_AIC_M1M2_SICPOVM}
\end{figure}
\subsection{Tomographically complete measurement}
\label{sec_tomography}
We first consider a tomographically complete measurement, in which a so-called SIC-POVM (symmetric informationally complete positive operator values measure \cite{RK-BSC2004}) with 4 outcomes is applied to each qubit individually. We first test our one-parameter model, and compare it to the FPM, which contains 255 ($=4^4-1$) parameters, which is the number of parameters needed to fully describe a general state of 4 qubits. With definition Eq.~(\ref{eq_AIC}) we have
\begin{align}
{\rm AIC}(M_{1\phi})=-2\L_M(M_{1\phi})+2,
\label{eq_AIC_M1_SICPOVM}
\end{align}
since $K=1$ for $M_{1\phi}$. For the FPM we have
\begin{align}
{\rm AIC(FPM)} =-2\L_M({\rm FPM})+2\times 255,
\label{eq_AIC_FPM_SICPOVM}
\end{align}
where $\L_M({\rm FPM})$ is the log of the maximum likelihood obtainable by the FPM. The latter can be bounded from above by noting that the best possible FPM would generate probabilities that exactly match the actual observed frequencies of all measurement outcomes. In the following we will always use that upper bound, rather than the actual maximum likelihood. Even though it is possible to find the maximum likelihood state in principle (and even in practice for small enough Hilbert spaces), we are only concerned with the FPM's ranking according to the AIC, which does not require its density matrix representation. For $M_{1\phi}$ to beat the FPM we require
\begin{align}
-\Delta{\rm AIC}:={\rm AIC}({\rm FPM})-{\rm AIC}(M_{1\phi})>0.
\end{align}
This is a sufficient but not necessary requirement, as we use the above-mentioned upper bound to the FPM likelihood.
\begin{figure}[t]
\begin{center}
\includegraphics[width=195pt]{N1e4}
\end{center}
\caption{{\em Comparing single- and double-excitation Dicke states}: The difference between AICs of $M_{1\phi}$ and the FPM, \textsl{i.e.}, $-\Delta{\rm AIC}={\rm AIC(FPM)}-{\rm AIC}(M_{1\phi})$ for both target states, when $N=10000$, as functions of $\phi$. The horizontal line demarcates $\Delta{\rm AIC}=0$. }
\label{fig_AIC_M1_1e4_SICPOVM}
\end{figure}
We plot the difference $\Delta{\rm AIC}$ between the two rankings in FIG.~\ref{fig_AIC_M1M2_SICPOVM}(a) for various values of the number of measurements, and for various values of the phase $\phi$. We observe the following: The simple model
is, correctly, judged better than the FPM when the phase $\phi$ is sufficiently small.
The more measurements one performs, the smaller $\phi$ has to be for AIC to still declare the model superior to the FPM (i.e., for the points to stay above the solid line, at $\Delta{\rm AIC}=0$).
Although the correction to the AIC mentioned in Eq.~(\ref{eq_AICc}) is not very small for the FPM for $N=1000$, applying that correction still does not shift the second and third point below zero: that is, $N=1000$ measurements is still not sufficiently large for the AICc to recognize that $\phi=\pi/4$ and $\phi=\pi/2$ are incorrect guesses. One can argue about what the cause of this is: it could be that $N$ is just too small for the derivation of the AIC (or even the AICc) to be correct. Or it could be that the AIC ranking is unreliable because the assumption that the true model is included in the model, is violated. Or it could be that, even with a perfectly valid criterion (perhaps the TIC), the statistical noise present in the data would still be too large.
If we consider the phase $\phi$ as a second (variable) parameter (thus creating a two-parameter model), then we can give FIG.~1 a different interpretation: we would pick $\phi=0$ as the best choice, and we would increase $K$ by 1. The latter correction is small on the scale of the plots, and so we find the two-parameter model to be superior to the one-parameter model for any nonzero plotted value of $\phi$, and to the FPM.
This is a good illustration of
the following rather obvious fact:
even if one has the impression that a particular property of one's quantum source is (or ought to be) known, it still might pay off to represent that property explicitly as a variable parameter (at the small cost of increasing $K$ by 1), and let the data determine its best value.
\subsection{Cross modeling}
Suppose one picked a one-parameter model with a wrong (nonzero) value of $\phi$, and the AIC has declared the model to be worse than the FPM. How can one improve the model in a systematic way when one lacks a good idea of which parameters to add to the model (we assume we already incorporated all parameters deemed important {\em a priori}). Apart from taking more and different measurements, one could use a hint from the existing data.
One method making use of the data is to apply ``cross modeling,'' where half the data is used to construct a modification to the model, and the remaining half is used for model validation, again by evaluating AIC on just that part of the data. So suppose $N$ measurements generate a data sequence $\mathcal{F}=\{f_1,f_2,...,f_N\}$. One takes, \textsl{e.g.}, the first $N/2$ data points, $\{f_1,...,f_{N/2}\}$, as the training set, and acquires the MLE state $\rho_{\rm MLE}$, or a numerically feasible approximation thereof, with respect to the training set. We then create a model with two parameters like so:
\begin{align}
M_{2\phi}:\rho_\phi(\epsilon,q)=(1-\epsilon)\left[(1-q)\rho_{\rm MLE}\right.&\nonumber\\
\left.+q\ket{\Psi_{\rm target}^{1,2}(\phi)}\bra{\Psi_{\rm target}^{1,2}(\phi)}\right]+\epsilon\openone/D.&
\label{eq_M2_SICPOVM}
\end{align}
For practical reasons $\rho_{\rm MLE}$ does not need to be strictly the MLE state, in particular when the dimension of the full parameter space is large. One would only require it to explain $\{f_1,...,f_{N/2}\}$ well enough to make sure that part of the data is properly incorporated in the model. Thus, one could, for example, use one of the numerical shortcuts described in \cite{KJ2009}. The rest of the data $\{f_{N/2+1},...,f_N\}$ is used to evaluate $M_{2\phi}$ against the FPM.
We note the resemblance of this procedure with the method of ``cross-validation'' \cite{SEJS1986}. In cross-validation one tries to find out how well a {\em given} predictive model performs by partitioning the data set into training set and validation set (exactly the same idea as given above). One uses multiple different partitions, and the results are averaged and optimized over those partitions. It can be shown \cite{S1977} that under certain conditions cross-validation and the AIC are asymptotically equivalent in model selection. This virtually exempts one from having to check multiple partitions of the data set, by applying the AIC to the whole data set.
It is worth emphasizing that what we do here is different in two ways. First, our model is not fixed but modified, based on information obtained from one half of the data. Second, we partition the data set only once, and the reason is, that it would be cheating to calculate the (approximate) MLE state of the full set of data (or, similarly, check many partitions and average), and then consider the resulting MLE state a parameter-free model.
FIG.~\ref{fig_AIC_M1M2_SICPOVM}(b) shows results for $M_{2\phi}$ and SIC-POVM measurements. When the number of measurements is $N=1000$, all $M_{2\phi}$ models
are considered better than the FPM, regardless of the phase error $\phi$ assumed for the target state. The reason is that around $\phi=\pi/2$ the approximate MLE state obtained from the first half of the data is able to ``predict'' the measurement outcomes (including their large amount of noise!) on the second half better than the 1-parameter model with the wrong phase.
On the other hand, when $N=10000$ the AIC recognizes only the simple models with small phase errors ($\phi=0, \pi/8, \pi/4$) as better than the FPM.
So, neither the approximate MLE state, nor the 1-parameter model with wrong phase are performing well. This indicates how many measurements are needed to predict a single phase to a given precision.
\begin{figure}[t]
\begin{center}
\includegraphics[width=195pt]{borges_angle_N.pdf}
\end{center}
\caption{{\em How the AIC ranks our one-parameter model vs. the FPM for an entanglement witness measurement}:
The difference between AICs of $M_{1\phi}$ with $\ket{\Psi_{\rm target}^2(\phi)}$ and the FPM, \textsl{i.e.}, $-\Delta{\rm AIC}={\rm AIC(FPM)}-{\rm AIC}(M_{1\phi})$, for different numbers of witness measurements, as functions of $\phi$. The horizontal line demarcates $\Delta{\rm AIC}=0$.}
\label{fig_AIC_M1_witness}
\end{figure}
\subsection{Witness measurement}
\label{sec_witness}
For states that are close to symmetric Dicke states $\ket{D_N^{N/2}}$, their entanglement can be verified by using measurements that require only two different local settings, \textsl{e.g.}, spins (or polarizations) either {\em all} in the $x$-direction or {\em all} in the $y$-direction. In particular, when $N=4$, an efficient witness is $W_{Jxy}=7/2+\sqrt{3}-J_x^2-J_y^2$ \cite{T2007}, where $J_{x,y}=\sum_j\sigma_{x,y}^{(j)}/2$, with $\sigma_{x,y}^{(j)}$ the Pauli matrices for the $j$-th subsystem. This witness detects (by having a negative expectation value) Dicke states with a white noise background, \textsl{i.e.}, $\rho(\alpha)=(1-\alpha)\ket{D_4^2}\bra{D_4^2}+\alpha\openone/D$ whenever $0\leq\alpha<0.1920$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=195pt]{WJxy.pdf}
\caption{{\em Does the witness $W_{J_{xy}}$ detect entanglement if there is a phase error?}: Witness performance $\langle W_{J_{xy}}\rangle$ for different states (defined as in Eq.~(\ref{eq_M1_witness})) as a function of $\phi$. A negative expectation value detects entanglement.}
\label{fig_WJxy}
\end{center}
\end{figure}
So we suppose we perform $N/2$ measurements on all of the four spins in the $x$-direction simultaneously, and another $N/2$ similar measurements in the $y$-direction. Instead of calculating the witness $W_{J_{xy}}$ and ending up with one single value determining entanglement, we make use of the full record of all individual outcomes in order to evaluate (and then maximize) likelihoods. For example, for the measurement of all four spins in the $x$-direction simultaneously, we can count the number of times they are projected onto the $\ket{x+x+x+x+}$ state, the $\ket{x+x+x+x-}$ state, etc. In both $x$- or $y$-directions, the number of independent observables (i.e., the number of independent joint expectation values) is 15, which can be seen as follows: Any density matrix of $M$ qubits can be expressed in terms of the expectation values of $4^M$ tensor products of the 3 Pauli operators and the identity $\openone$, but the expectation value of the product of $M$ identities equals 1 for any density matrix, thus leading to $4^M-1$ independent parameters encoded in a general density matrix. From having measured just $\sigma_x$ on all $M$ qubits, we can evaluate all expectation values of all operators that are tensor products of $\sigma_x$ and the identity. There are $2^M$ such products, and subtracting the trivial expectation value for $\openone^{\otimes M}$ leaves $2^M-1$ independent expectation values.
This means it only takes $2\times 15=30$ independent parameters to form the FPM, and we have $K=30$. Similar to the tomographically complete case, we do not need the concrete form of the whole 255-30 dimensional manifold of MLE states, nor do we need to explicitly parameterize the 30-parameter FPM states, as we can simply upper bound the maximum likelihood for this model, $L_M({\rm FPM})$, by noting the best one could possibly do is reproduce exactly the observed frequencies of all possible measurement outcomes.
\begin{figure}[t]
\begin{center}
\subfigure[$N=100$]{\includegraphics[width=195pt]{neg_dist1e2_1}}
\subfigure[$N=1000$]{\includegraphics[width=195pt]{neg_dist1e3_1}}
\caption{{\em How one quantifies entanglement from a witness measurement}:
The posterior probability distributions of $\mathcal{N}_0$ for different numbers of witness measurements from model $M_1(q)$ with target state $\ket{\Psi_{\rm target}^{2(1)}}$, where $\phi=0, \pi/6, \pi/3$. $\mathcal{N}_0(\rho_{\rm actual})=0.4770$. The prior distribution is assumed to be uniform on [0,1] for both $\epsilon$ and $q$. The distributions of $\mathcal{N}_1$ and $\mathcal{N}_2$ are similar (up to a simple shift).}
\label{fig_negdist_M1_witness}
\end{center}
\end{figure}
\subsection{Estimating entanglement}
Our state $\rho_{\rm actual}=\rho(\alpha=0.2)$ is just not detected by the witness $W_{Jxy}$, but still contains a considerable amount of entanglement.
We choose to quantify this entanglement by means of three entanglement monotones (of which only two are independent), simply constructed from all bipartite negativities.
If the four parties are denoted $A$, $B$, $C$ and $D$, the generalized negativities \cite{VW2002,SG-A2008,YvE2011} are defined as
\begin{subequations}
\begin{align}
\mathcal{N}_1=&\left(\mathcal{N}_{AB-CD}\mathcal{N}_{AC-BD}\mathcal{N}_{AD-BC}\right)^{1/3},\\
\mathcal{N}_2=&\left(\mathcal{N}_{A-BCD}\mathcal{N}_{B-CDA}\mathcal{N}_{C-DAB}
\mathcal{N}_{D-ABC}\right)^{1/4},\\
\mathcal{N}_0=&\left(\mathcal{N}_1^3 \mathcal{N}_2^4\right)^{1/7},
\end{align}
\end{subequations}
where $\mathcal{N}_{AB-CD}$ denotes the negativity with respect to partition $AB$ against $CD$, etc. The main advantage of the generalized negativities is that they are all efficiently computable directly from the density matrix.
We have for our state
$ \mathcal{N}_1=0.6293$, $ \mathcal{N}_2=0.3875$, and $ \mathcal{N}_0=0.4770$.
Similarly to the tomographically complete case, we first consider the following one-parameter model:
\begin{align}
M_{1\phi}:~\rho_\phi(q)=(1-q)\ket{\Psi_{\rm target}^2(\phi)}&\nonumber\\
\bra{\Psi_{\rm target}^2(\phi)}+q&\openone/D,\label{eq_M1_witness}
\end{align}
where $\ket{\Psi_{\rm target}^2(\phi)}$ is defined in Eg.~(\ref{eq_tilde1}). The AICs for $M_{1\phi}$ and FPM are
\begin{align}
{\rm AIC}(M_{1\phi})=&-2\L_M(M_{1\phi})+2,\label{eq_AIC_M1_witness}\\
{\rm AIC}({\rm FPM})=&-2\L_M({\rm FPM})+2\times30.\label{eq_AIC_FPM_witness}
\end{align}
FIG.~\ref{fig_AIC_M1_witness} shows that, as before, the marks above the horizontal solid line correspond to models deemed better than FPM. Compared to the case of full tomography (FIG.~\ref{fig_AIC_M1M2_SICPOVM}(a)), here the value of ${\rm AIC}(M_{1\phi})$ is larger than ${\rm AIC}({\rm FPM})$ by a much smaller amount, even when the phase term is correct ($\phi=0$). The absolute value of the difference is not relevant, though, and what counts is its sign. The obvious reason for the smaller difference is that the number of independent parameters for the FPM has dropped from 255 to 30. In addition, the FPM in this case does not refer to a specific 30-parameter model. On the contrary, since the number of degrees of freedom of the quantum system is still 255, there is a whole subspace of states, spanning a number of degrees of freedom equal to 225 (=255-30), all satisfying the maximum likelihood condition.
The witness measurement is very sensitive to the phase error, even when the number of measurements is still small. When $N=1000$, the estimation of $\phi$ is within an error of $\pi/6$, as the second point plotted is already below the line $\Delta{\rm AIC}=0$.
Compared to FIG.~\ref{fig_AIC_M1M2_SICPOVM}(a), this precision is only reached when $N=10000$.
An interesting comparison can be made between AIC and the entanglement-detecting nature of witness $W_{J_{xy}}$. FIG.~\ref{fig_WJxy} shows the performance of $\langle W_{J_{xy}}\rangle$ for the pure state $\ket{\Psi_{\rm target}^2(\phi)}$ ($\rho_\phi(q=0)$, solid curve) and the mixed with 20\% of identity mixed in ($\rho_\phi(q=0.2)$, dot-dashed curve). Even when the state is pure, $\langle W_{J_{xy}}\rangle$ will not be able to witness any entanglement if the phase error is larger than $\pi/3$, just about when AIC declares such a model deficient. Entanglement in the mixed $\rho_\phi(q=0.2)$ of course is never witnessed. This means $\langle W_{J_{xy}}\rangle$ is only an effective witness in the vicinity of $\ket{D_4^2}$, with limited tolerance of either white noise or phase noise in even just one of the four qubits. (Of course, one would detect the entanglement in the pure state by appropriately rotating the axes in the spin measurement on the first qubit over an angle $\phi$.)
To test whether a few-parameter model correctly quantifies entanglement if that model is preferred over the FPM by AIC, we estimate a (posterior, Bayesian) probability distribution over the generalized negativities (defined above). We see that the first three curves in
FIG.~\ref{fig_negdist_M1_witness}(a) and the first two curves in FIG.~\ref{fig_negdist_M1_witness}(b), which correspond to the data points above the horizontal line in FIG.~\ref{fig_AIC_M1_witness}, all give consistent estimates of $\mathcal{N}_0$, compared to the actual value of $\mathcal{N}_0$ for the true state (and the same holds for $\mathcal{N}_{1,2}$ (not shown)). Conversely, the estimate cannot be trusted when AIC deems the simple model inferior to the FPM (of course, it may still happen to be a correct estimate, but one could not be sure).
This gives additional evidence for the success of AIC.
\subsection{Cross modeling for a witness measurement}
We now construct a two-parameter model $M_{2\phi}$ similar in spirit to that discussed for tomographically complete measurements: half the data [on which half the time $(\sigma_x)^{\otimes 4}$ is measured, and half the time $(\sigma_y)^{\otimes 4}$] are used to generate a better model, which is then tested on the other half of the data (also containing both types of measurements equally). We write
\begin{align}
\rho(\epsilon,q)=&(1-\epsilon)\left[(1-q)\rho_{\rm observation}\right.\nonumber\\
&\left.+q\ket{\Psi_{\rm target}^{2}(\phi)}\bra{\Psi_{\rm target}^{2}(\phi)}\right]+\epsilon\openone/D.\label{eq_M2_witness}
\end{align}
To find a $\rho_{\rm observation}$---there are many equivalent ones for predicting the outcomes of the witness measurements---we recall that a generic four-qubit state can be expressed as
\begin{align}
\rho=\sum_{jklm}c_{jklm}\sigma_j\otimes\sigma_k\otimes\sigma_l\otimes\sigma_m,
\end{align}
where $j,k,l,m=1,2,3,4$ where
$\sigma_{1,2,3}$ denote the Pauli matrices $\sigma_{x,y,z}$ and $\sigma_4=\openone$. The witness measures the coefficients $c_{jklm}$ where $j,k,l,m$ can be combinations of only 1 and 4 or combinations of only 2 and 4 (\textsl{e.g.}, $c_{1441}$ or $c_{4222}$). We label the $c_{jklm}$'s that can be recovered from witness measurement as $c_{jklm}^w$ ($w$ as in \emph{witness}). We do not include in $c_{jklm}^w$ the coefficient $c_{4444}$, which always equals $1/16$, so that it does not depend on measurement outcomes. We define
\begin{align}
\rho_{\rm observation}=\sum_{jklm}c_{jklm}^w\sigma_j\otimes\sigma_k\otimes\sigma_l
\otimes\sigma_m+\openone/16.
\end{align}
\begin{figure}[t]
\begin{center}
\includegraphics[width=165pt]{unphysical_e_q.pdf}
\caption{{\em What fraction of the
model Eq. (\ref{eq_M2_witness}) describes physical states?}: The lower left part separated by the curves is where $\rho(\epsilon,q)$ of Eq. (\ref{eq_M2_witness}) is unphysical (and so is not actually included in the model), for different number of measurements $N$. }
\label{fig_unphysical_e_q}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=195pt]{L_M_D42_UNPHYSICAL.pdf}
\caption{{\em How the AIC ranks our two-parameter model vs. the FPM, for a witness measurement}: The difference between AICs of $M_{2\phi}$ and the FPM, \textsl{i.e.}, $-\Delta{\rm AIC}={\rm AIC(FPM)}-{\rm AIC}(M_{2\phi})$, for different numbers of witness measurements as a function $\phi$. The target state is $\ket{\Psi_{\rm target}^{2(1)}}$. The horizontal line demarcates $\Delta{\rm AIC}=0$.}
\label{fig_AIC_M2_witness}
\end{center}
\end{figure}
Note that $\rho_{\rm observation}$ can be considered as a trace-one \emph{pseudostate}, since it is not necessarily positive semi-definite. But the most attractive property of $\rho_{\rm observation}$ is that it preserves the measurement outcomes. It is in fact the unique \emph{pseudostate} that reproduces the exact frequencies of all measurement outcomes {\em and} that has vanishing expectation values for all other {\em un}performed collective Pauli measurements. As a component of $\rho(\epsilon,q)$, we allow $\rho_{\rm observation}$ to be unphysical, but we only keep those $\rho(\epsilon,q)$ that are positive semi-definite.
We checked numerically for what values of $\epsilon$ and $q$ the states end up being physical, and how this depends on the number of measurements performed.
Physical states are located in the upper right part of the square in FIG.~\ref{fig_unphysical_e_q}. That is, only if $\epsilon$ and/or $q$ are sufficiently large, so that a sufficiently large amount of $\ket{\Psi_{\rm target}^2}$ and/or $\openone/16$ has been mixed in, does $\rho(\epsilon,q)$ become physical. Depending on the number of measurements, the area of the upper right part is about 69\%-77\% of the whole square. The physical/unphysical boundary shifts closer to the origin as the number of measurements increases.
\begin{figure*}[t]
\begin{center}
\subfigure[$N=100, \ket{\Psi_{\rm target}^2(\phi)}$]{\includegraphics[width=162pt]{posteriorN2_D42_1e2_1_UNPHYSICAL.pdf}}
\subfigure[$N=1000, \ket{\Psi_{\rm target}^2(\phi)}$]{\includegraphics[width=162pt]{posteriorN2_D42_1e3_1_UNPHYSICAL.pdf}}
\subfigure[$N=10000, \ket{\Psi_{\rm target}^2(\phi)}$]{\includegraphics[width=162pt]{posteriorN2_D42_1e4_1_UNPHYSICAL.pdf}}
\caption{{\em Quantifying entanglement from a witness measurement}: The posterior distributions for $\mathcal{N}_2$, for different numbers of measurements, using model $M_{2\phi}$ with $\rho_{\rm observation}$ and target state $\ket{\Psi_{\rm target}^2}$, where $\phi=0, \pi/6, \pi/3$. $\mathcal{N}_2(\rho_{\rm actual})=0.3875$. The same prior is used as in FIG.~\ref{fig_negdist_M1_witness}. Whenever the AIC declares a model superior to the FPM, the estimated entanglement agrees, within error bars, with the actual value, but may be wrong otherwise. }
\label{fig_negdist_M2_witness}
\end{center}
\end{figure*}
\begin{figure}[t]
\begin{center}
\includegraphics[width=195pt]{M1vsM2}
\caption{{\em Comparing one- and two-parameter models directly}: The difference between AICs of $M_{2\phi}$ and $M_{1\phi}$ for 20 different sets of witness measurements ($N=50$) as functions of $\phi$. The target state is $\ket{\Psi_{\rm target}^{2}}$. The horizontal line demarcates $\Delta{\rm AIC}=0$. The dotted-dashed line is the average of all 20 points at each different $\phi$.}
\label{fig_AIC_M1vsM2_witness}
\end{center}
\end{figure}
We test the two-parameter model (the physical part of it), and show the results in FIG.~\ref{fig_AIC_M2_witness}. We find that for $N=100$ the AIC ranks $\rho_{2\phi}(\epsilon,q)$ better than the FPM, even when the guess about $\phi$ is very imprecise: 100 witness measurements are, unsurprisingly, not enough for a correct reconstruction of the state. When $N=1000$, AIC only prefers the models with a value for $\phi$ within $\pi/6$ of the correct value. And when $N=10000$, the accepted values of $\phi$ are even closer to the true value.
The corresponding posterior distributions of negativities ${\cal N}_2$ are plotted in FIG.~\ref{fig_negdist_M2_witness} for the three better guesses, $\phi=0, \pi/6, \pi/3$.
When $N=100$ all three give decent predictions of $\mathcal{N}_2$ (and indeed, AIC ranks those models highly). For $N=1000$ and $N=10000$, we would only trust the estimates arising from the lower two values of $\phi$, or just the correct value of $\phi$, respectively. This trust is rewarded in FIG.~\ref{fig_AIC_M2_witness}(b) and FIG.~\ref{fig_AIC_M2_witness}(c), as those estimates are indeed correct, within the error bars. In addition, the untrusted estimate for $\phi=\pi/6$ for $N=10000$ still happens to be correct, too.
\subsection{Comparing one- and two-parameter models directly}
Finally, the AIC can compare the one- and two-parameter models $M_{1\phi}$ and $M_{2\phi}$ directly.
For that purpose one needs to use the {\em same} validation set of data, which implies that the two-parameter model needs {\em additional} data to generate $\rho_{{\rm observation}}$. Here we display results for just 50 witness measurements, and an additional set of 50 measurements for $M_{2\phi}$. FIG.~\ref{fig_AIC_M1vsM2_witness} shows that even such a small number of additional data is useful if the angle $\phi$ is wrong, and, similarly, it shows that the same small number suffices to detect a wrong single-qubit phase when it is larger than $\pi/3$.
\section{Conclusions}
\label{sec_conclusion}
We applied information criteria, and the Akaike Information Criterion (AIC) developed in Ref.~\cite{A1974} in particular, to quantum state estimation. We showed it to be a powerful method, provided one has a reasonably good idea of what state one's quantum source actually generates.
For each given model, which may include several parameters describing error and noise, as well as some parameters---call them the ideal-state parameters--- describing the state one would like to generate in the ideal (noiseless and error-free) case, the AIC determines a ranking from the observed data. One can construct multiple models, for instance, models where some ideal-state parameters and some noise parameters are fixed (possibly determined by previous experiments in the same setup), with others still considered variable.
Crucially, the AIC also easily ranks the full-parameter model (FPM), which uses in principle all exponentially many parameters in the full density matrix, and which is, therefore, the model one would use in full-blown quantum state tomography. This ranking of the FPM can be accomplished without actually having to find the maximum-likelihood state (or its likelihood)---which quickly would run into insurmountable problems for many-qubit systems---by using a straightforward upper bound.
This way, observed data is used to justify {\em a posteriori} the use of the few-parameter models---namely, if the AIC ranks that model above the FPM---and thus our method is in the same spirit as several other recent proposals \cite{GLFBE2010,CPFSGBL-CPL2010} to simplify quantum tomography, by tentatively introducing certain assumptions on the quantum state generated, after which data is used to certify those assumptions (and if the certification fails, one at least knows the initial assumptions were incorrect).
We illustrated the method on (noisy and mis-specified) four-qubit members of the family of Dicke states, and demonstrated its effectiveness and efficiency. For instance, we showed that one can detect mis-specified ideal-state parameters and determine noise and error parameters. We also showed by example the successful application of the method to a specific and useful subtask, that of quantifying multi-qubit entanglement.
\section*{Acknowledgement}
This work was supported by NSF grant PHY-1004219.
|
\section{Introduction}
\label{sec:introduction}
Geometry and topology play an increasing role in the modern language
used to describe physical problems \cite{AbMaRa1988,Frankel2004}. A
large part of this language is \emph{exterior calculus} which
generalizes vector calculus to smooth manifolds of arbitrary
dimensions. The main objects are differential forms (which are
anti-symmetric tensor fields), general tensor fields, and vector
fields defined on a manifold. In addition to physical applications,
differential forms are also used in cohomology theory in
topology~\cite{BoTu1982}.
Once the domain of interest is discretized, it may not be smooth and
so the objects and operators of exterior calculus have to be
reinterpreted in this context. For example, a surface in $\mathbb{R}^3$ may be
discretized as a two dimensional simplicial complex embedded in
$\mathbb{R}^3$, i.e., as a triangle mesh. Even when the domain is a simple
domain in space, such as an open subset of the plane or space the
discretization is usually in the form of some mesh. The various
objects of the problem then become defined in a piecewise varying
fashion over such a mesh and so a discrete calculus is required there
as well. After discretizing the domains, objects, and operators, one
can compute numerical solutions of partial differential equations
(PDEs), and compute some topological invariants using the same
discretizations. Both these classes of applications are considered
here.
There have been several recent developments in the discretization of
exterior calculus and in the clarification of the role of algebraic
topology in computations. These go by various names, such as covolume
methods \cite{NiTr2006}, support operator methods \cite{Shashkov1996},
mimetic discretization \cite{HySh1997a,BoHy2006}, discrete exterior
calculus (DEC) \cite{Hirani2003,DeHiLeMa2005, GiBa2010}, compatible
discretization, finite element exterior calculus \cite{ArFaWi2006,
ArFaWi2010, HoSt2011}, edge and face elements or Whitney forms
\cite{Bossavit1988a,Hiptmair2002a}, and so on. PyDEC provides an
implementation of discrete exterior calculus and lowest order finite
element exterior calculus using Whitney forms.
Within pure mathematics itself, ideas for discretizing exterior
calculus have a long history. For example the de Rham map that is
commonly used for discretizing differential forms goes back at least
to \cite{Rham1955}. The reverse operation of interpolating discrete
differential forms via the Whitney map appears in
\cite{Whitney1957}. A combinatorial (discrete) Hodge decomposition
theorem was proved in \cite{Dodziuk1976} and the idea of a
combinatorial Hodge decomposition dates to \cite{Eckmann1945}. More
recent work on discretization of Hodge star and wedge product is in
\cite{Wilson2007,Wilson2008}. Discretizations on other types of
complexes have been developed as well \cite{GrHi1999,Sen2003}.
\subsection{Main contributions}
In this paper we describe the algorithms and design of PyDEC, a Python
software library implementing various complexes and operators for
discretization of exterior calculus, and the algorithms and data
structures for those. In PyDEC all the discrete operators are
implemented as sparse matrices and we often reduce algorithms to a
sequence of standard high-level operations, such as sparse
matrix-matrix multiplication~\cite{BaDo1993}, as opposed to more
specialized techniques and ad hoc data structures. Since these
high-level operations are ultimately carried out by efficient,
natively-compiled routines (e.g. C or Fortran implementations) the
need for further algorithmic optimization is generally unnecessary.
As is commonly done, in PyDEC we implement discrete differential forms
as real valued cochains which will be defined in
Section~\ref{sec:overview}. PyDEC has been used in a
thesis~\cite{Bell2008}, in classes taught at University of Illinois,
in experimental parts of some computational topology
papers~\cite{DeHiKr2010, DuHi2011, HiKaWaWa2011}, in Darcy
flow~\cite{HiNaCh2011}, and in least squares ranking on
graphs~\cite{HiKaWa2011}. The PyDEC source code and examples are
publicly available~\cite{BeHi2008a}. We summarize here our
contributions grouped into four areas.
\medskip\noindent{\bf Basic objects and methods:}
\begin{inparaenum}[(1)]
\item Data structures for : simplicial complexes of dimension $n$
embedded in $\mathbb{R}^N$, $N\ge n$; abstract simplicial complexes;
Vietoris-Rips complexes for points in any dimension; and regular
cube complexes of dimension $n$ embedded in $\mathbb{R}^n$;
\item Cochain objects for the above complexes;
\item Discrete exterior derivative as a coboundary operator,
implemented as a method for cochains on various complexes.
\end{inparaenum}
\medskip\noindent{\bf Finite element exterior calculus:}
\begin{inparaenum}[(1)]
\item Fast algorithm to construct sparse mass matrices for Whitney
forms by eliminating repeated computations; and
\item Assembly of stiffness matrices for Whitney forms from mass
matrices by using products of boundary and stiffness matrices. Note
that only the lowest order ($\mathcal{P}_1^{-}$) elements of finite
element exterior calculus are implemented in PyDEC.
\end{inparaenum}
\medskip\noindent{\bf Discrete exterior calculus:}
\begin{inparaenum}[(1)]
\item Diagonal sparse matrix discrete Hodge star for well-centered
(circumcenters inside simplices) and Delaunay simplicial complexes
(with an additional boundary condition);
\item Circumcenter calculation for $k$-simplex in an $n$-dimensional
simplicial complex embedded in $\mathbb{R}^N$ using a linear system in
barycentric coordinates; and
\item Volume calculations for primal simplices and circumcentric dual
cells.
\end{inparaenum}
\medskip\noindent{\bf Examples:}
\begin{inparaenum}[(1)]
\item Resonant cavity curl-curl problem;
\item Flow in porous medium modeled as Darcy flow, i.e., Poisson's
equation in first order (mixed) form;
\item Cohomology basis calculation for a simplicial mesh, using
harmonic cochain computation using Hodge decomposition;
\item Finding sensor network coverage holes by modeling an abstract,
idealized sensor network as a Rips complex; and
\item Least squares ranking on graphs using Hodge decomposition of
partial pairwise comparison data.
\end{inparaenum}
\section{Overview of PyDEC}
\label{sec:overview}
One common type of discrete domain used in scientific computing is
triangle or tetrahedral mesh. These and their higher dimensional
analogues are implemented as $n$-dimensional simplicial complexes
embedded in $\mathbb{R}^N$, $N \ge n$. Simplicial complexes are useful even
without an embedding and even when they don't represent a manifold,
for example in topology and ranking problems. Such abstract simplicial
complexes without any embedding for vertices are also implemented in
PyDEC. The other complexes implemented are regular cube complexes and
Rips complexes. Regular cubical meshes are useful since it is easy to
construct domains even in high dimensions whereas simplicial meshing
is hard enough in 3 dimensions and rarely done in 4 or larger
dimensions. Rips complexes are useful in applications such as
topological calculations of sensor network coverage
analysis~\cite{SiGh2007}. The representations used for these four
types of complexes are described in
Section~\ref{sec:simplicial_complex}-\ref{sec:abstrct_smplcl_cmplx}.
A complex that is a manifold (i.e., locally Euclidean) will be
referred to as \emph{mesh}.
The definitions here are given for simplicial complexes and generalize
to the other types of complexes implemented in PyDEC. In PyDEC we only
consider integer valued chains and real-valued cochains. Also, we are
only interested in finite complexes, that is, ones with a finite
number of cells. Let $K$ be a finite simplicial complex and denote its
underlying space by $\abs{K}$. Give $\abs{K}$ the subspace topology as
a subspace of $\mathbb{R}^N$ (a set $U$ in $\abs{K}$ is open iff $U \cap
\abs{K}$ is open in $\mathbb{R}^N$). For a finite complex this is the same as
the standard way of defining topology for $\abs{K}$ \cite[pages
8-9]{Munkres1984} and $\abs{K}$ is a closed subspace of $\mathbb{R}^N$.
An oriented simplex with vertices $v_0,\dots,v_p$ will be written as
$\simplex{p}$ and given names like $\sigma_i^p$ with the superscript
denoting the dimension and subscript denoting its place in some
ordering of $p$-simplices. Sometimes the dimensional superscript
and/or the indexing subscript will be dropped. The orientation of a
simplex is one of two equivalence classes of vertex orderings. Two
orderings are equivalent if one is an even permutation of the
other. For example $[v_0,v_1,v_2]$ and $[v_1, v_2, v_0]$ denote the
same oriented triangle while $[v_0,v_2,v_1]$ is the oppositely
oriented one.
A \emph{$p$-chain of $K$} is a function $c$ from oriented
$p$-simplices of $K$ to the set of integers $\mathbb{Z}$, such that
$c(-\sigma) = -c(\sigma)$ where $-\sigma$ is the simplex $\sigma$
oriented in the opposite way. Two chains are added by adding their
values. Thus $p$-chains are formal linear combinations (with integer
coefficients) of oriented $p$-dimensional simplices. The space of
$p$-chains is denoted $C_p(K)$ and it is a free abelian group. See
\cite[page 21]{Munkres1984}. Free
abelian groups have a basis and one does not need to impose a vector
space structure. For example, a basis for $C_p(K)$ is the set of
integer valued functions that are 1 on a $p$-simplex and 0 on the
rest, with one such basis element corresponding to each $p$-simplex.
These are called \emph{elementary chains} and the one corresponding to
a $p$-simplex $\sigma^p$ will also be referred to as $\sigma^p$. The
existence of this basis and the addition and negation of chains is the
only aspect that is important for this paper. The intuitive way to
think of chains is that they play a role similar to that played by the
domains of integration in the smooth theory. The negative sign allows
one to talk about orientation reversal and the integer coefficient
allows one to say how many times integration is to be done on that
domain.
Sometimes we will need to refer to a \emph{dual mesh} which will in
general be a cell complex obtained from a subdivision of the given
complex $K$. We'll refer to the dual complex as $\dual K$. For a
discrete Hodge star diagonal matrix of DEC, the dual mesh is the one
obtained from circumcentric subdivision of a well-centered or Delaunay
simplicial complex and such a Hodge star is described in
Section~\ref{sec:metric}.
Homomorphisms from the $p$-chain group $C_p(K)$ to $\mathbb{R}$ are called
\emph{$p$-cochains of $K$} and denoted $C^p(K;\mathbb{R})$. This set is an
abelian group and also a vector space over $\mathbb{R}$. Similarly the dual
$p$-cochains are denoted $C^p(\dual K; \mathbb{R})$ or $D^p(\dual K; \mathbb{R})$.
The discretization map from space of smooth $p$-forms to $p$-cochains
is called the \emph{de Rham} map $\deRham : \Omega^p(K) \to C^p(K;\mathbb{R})$
or $\deRham : \Omega^p(K) \to C^p(\dual K;\mathbb{R})$. See \cite{Rham1955,
Dodziuk1976}. For a smooth $p$-form $\alpha$, the de Rham map is
defined as $\deRham : \alpha \mapsto (c \mapsto \int_c \alpha)$ for
any chain $c \in C_p(K)$. We will denote the evaluation of the cochain
$\deRham(\alpha)$ on a chain $c$ as $\eval{\deRham(\alpha)}{c}$. A
basis for $C^p(K;\mathbb{R})$ is the set of \emph{elementary cochains}. The
elementary cochain $(\sigma^p)^\ast$ is the one that takes value 1 on
elementary chain $\sigma^p$ and 0 on the other elementary chains. Thus
the vector space dimension of $C^p(K;\mathbb{R})$ is the number of
$p$-simplices in $K$. We'll denote this number by $N_p$. Thus $N_0$
will be the number of vertices, $N_1$ the number of edges, $N_2$ the
number of triangles and so on.
Like most of the numerical analysis literature mentioned in
Section~\ref{sec:introduction} we assume that the smooth forms are
either defined in the embedding space of the simplicial complex, or on
the complex itself, or can be obtained by pullback from the manifold
that the complex approximates. In contrast, most mathematics
literature quoted including \cite{Whitney1957,Dodziuk1976} uses
simplicial complex defined on the smooth manifold as a ``curvilinear''
triangulation. In the applied literature, the complex approximates
the manifold. Many finite element papers deal with open subsets of the
plane or $\mathbb{R}^3$ so they are working with triangulations of a manifold
with piecewise smooth boundaries. Surface finite element methods have
been studied outside of exterior calculus~\cite{DeDz2007}. A
variational crimes methodology is used for finite element exterior
calculus on simplicial approximations of manifolds
in~\cite{HoSt2011}. In the computer graphics literature,
piecewise-linear triangle mesh surfaces embedded in $\mathbb{R}^3$ are common
and convergence questions for operators on such surfaces have been
studied \cite{HiPoWa2006}. In light of all of these, PyDEC's framework
of using simplicial or other approximations of manifolds is
appropriate.
Operators such as the discrete exterior derivative ($\operatorname{d}$) and Hodge
star ($\hodge$) can be implemented as sparse matrices. At each
dimension, the exterior derivative can be easily determined by the
incidence structure of the given simplicial mesh. For DEC the Hodge
star is a diagonal matrix whose entries are determined by the ratios
of primal and dual volumes. Care is needed for dual volume calculation
when the mesh is not well-centered. For finite element exterior
calculus we implement Whitney forms. The corresponding Hodge star is
the mass matrix which is sparse but not diagonal. One of the stiffness
matrices can be obtained from it by combining it with the exterior
derivative.
Once the matrices implementing the basic operators have been
determined, they can be composed together to obtain other operators
such as the codifferential ($\codiff$) and Laplace-deRham
($\laplacian$). While this composition could be performed manually,
i.e. the appropriate set of matrices combined to form the desired
operation, it is prone to error. In PyDEC this composition is handled
automatically. For example, the function $\operatorname{d}(.)$ which implements the
exterior derivative, looks at the dimension of its argument to
determine the appropriate matrix to apply. The same method can be
applied to the codifferential function $\codiff(.)$, which then makes
their composition $\codiff(\operatorname{d}(.))$ work automatically. This
automation eliminates a common source of error and makes explicit
which operators are being used throughout the program.
PyDEC is intended to be fast, flexible, and robust. As an interpreted
language, Python by itself is not well-suited for high-performance
computing. However, combined with numerical libraries such as NumPy
and SciPy one can achieve fast execution with only a small amount of
overhead. The NumPy array structure, used extensively throughout
SciPy and PyDEC, provides an economical way of storing N-dimensional
arrays (comparable to C or Fortran) and exposes a C API for
interfacing Python with other, potentially lower-level
libraries~\cite{WaCoVa2011}. In this way, Python can be used to
efficiently ``glue'' different highly-optimized libraries together
with greater ease than a purely C, C++, or Fortran implementation
would permit~\cite{Oliphant2007}. Indeed, PyDEC also makes extensive
use of the \texttt{sparse} module in SciPy which relies on
natively-compiled C++ routines for all performance-sensitive
operations, such as sparse matrix-vector and matrix-matrix
multiplication. PyDEC is therefore scalable to large data sets and
capable of solving problems with millions of elements~\cite{Bell2008}.
Even large-scale, high-performance libraries such as Trilinos provide
Python bindings showing that Python is useful beyond the prototyping
stage. We also make extensive use of Python's built-in unit testing
framework to ensure PyDEC's robustness. For each non-trivial
component of PyDEC, a number of examples with known results are used
to check for consistency.
\subsection{Previous work}
Discrete differential forms now appear in several finite element
packages such as FEMSTER \cite{CaRiWh2005}, DOLFIN \cite{LoWe2010} and
{deal.II}~\cite{BaHaKa07}. These libraries support arbitrary order
conforming finite element spaces in two and three dimensions. In
contrast, for finite elements PyDEC supports simplicial and cubical
meshes of arbitrary dimension, albeit with lowest order elements. In
addition, PyDEC also supports the operators of discrete exterior
calculus and complexes needed in topology. We note that
Exterior~\cite{LoMa2008}, an experimental library within the
FEniCS~\cite{LoMaWe2012} project, realizes the framework developed by
Arnold et al.~\cite{ArFaWi2009} which generalizes to arbitrary order
and dimension. Exterior uses symbolic methods and supports integration
of forms on the standard simplex. PyDEC supports mass and stiffness
matrices on simplicial and cubical complexes. The discovery of lower
dimensional faces in a complex and the computation of all the boundary
matrices is also implemented in PyDEC.
The other domain where PyDEC is useful is in computational
topology. There are several packages in this domain as well, and again
PyDEC has a different set of features and aims from these.
In~\cite{KaMiKoMr2004} efficient techniques are developed for finding
meaningful topological structures in cubical complexes, such as
digital images. In addition to simplicial and cubical manifolds, PyDEC
also provides support for abstract simplicial complexes such as the
Rips complex of a point set. The Applied and Computational Topology
group at Stanford University has been the source for several packages
for computational topology. These include various versions of PLEX
such as JPlex and javaPlex which are designed for persistent homology
calculations. Another package from the group is Dionysus, a C++
library that implements persistent homology and
cohomology~\cite{EdLeZo2002, ZoCa2005} and other interesting
topological and geometric algorithms. In contrast, we view the role of
PyDEC in computational topology as providing a tool to specify and
represent different types of complexes, compute their boundary
matrices, and compute cohomology representatives with or without
geometric information.
\section{Simplicial Complex Representation}
\label{sec:simplicial_complex}
Before detailing the algorithms used to implement discretizations of
exterior calculus, we discuss the representation of various complexes,
starting in this section with simplicial complexes. Consider the
triangle mesh shown in Figure~\ref{fig:mesh_example_filled} with
vertices and faces enumerated as shown. This example mesh is
represented by arrays
\begin{align*}
\mathbb{V} = \begin{bmatrix}
0 & 0 \\
1 & 0 \\
2 & 0 \\
1 & 1 \\
2 & 1
\end{bmatrix},
&&
\mathbb{S}_2 = \begin{bmatrix}
0 & 1 & 3 \\
1 & 2 & 3 \\
2 & 4 & 3
\end{bmatrix},
\end{align*}
where the subscript $2$ denotes the dimension of the simplices. The
$i$-th row of $\mathbb{V}$ contains the spatial coordinates of the
$i$-th vertex. Likewise the $i$-th row of simplex array
$\mathbb{S}_2$ contains the indicies of the vertices that form the
$i$-th triangle. The indices of each simplex in $\mathbb{S}_2$ in
this example are ordered in a manner that implies a counter-clockwise
orientation for each. For an $n$-dimensional discrete manifold, or
mesh, arrays $\mathbb{V}$ and $\mathbb{S}_n$ suffice to describe
the computational domain.
\begin{figure}
\begin{center}
\includegraphics[height=1in]{pydec/simplex_example_filled}
\end{center}
\caption{Simplicial mesh with enumerated vertices and simplices.}
\label{fig:mesh_example_filled}
\end{figure}
In addition to $\mathbb{V}$ and $\mathbb{S}_n$, an
$n$-dimensional simplicial complex is comprised by its $p$-dimensional
faces, $\mathbb{S}_p, 0 \leq p < n$. In the case of
Figure~\ref{fig:mesh_example_filled}, these are
\begin{align*}
\mathbb{S}_0 = \begin{bmatrix}
0\\
1\\
2\\
3\\
4
\end{bmatrix},
&&
\mathbb{S}_1 = \begin{bmatrix}
0 & 1\\
0 & 3\\
1 & 2\\
1 & 3\\
2 & 3\\
2 & 4\\
3 & 4
\end{bmatrix},
\end{align*}
which correspond to the vertices (0-simplices) and oriented edges
(1-simplices) of the complex. A graphical representation of this
simplicial complex is shown in Figure~\ref{fig:mesh_example_oriented}.
Since the orientation of the lower ($< n$) dimensional faces is
arbitrary, we use the convention that face indices will be in sorted
order. Furthermore, we require the rows of $\mathbb{S}$ to be
sorted in lexicographical order. As pointed out in
Section~\ref{sec:exterior_derivative} and~\ref{sec:whtny_innrprdct},
these conventions facilitate efficient construction of differential
operators and stiffness matrices.
\begin{figure}
\begin{center}
\includegraphics[height=1in]{pydec/simplex_example_oriented}
\end{center}
\caption{Simplicial complex with oriented edges and triangles.}
\label{fig:mesh_example_oriented}
\end{figure}
\section{Regular Cube Complex Representation}
\label{sec:cube_complex}
PyDEC provides a regular cube complex of dimension $n$ embedded in
$\mathbb{R}^n$ for any $n$. As mentioned earlier, in dimension higher than 3,
constructing simplicial manifold complexes is hard. In fact, even
construction of good tetrahedral meshes is still an active area in
computational geometry. This is one reason for using regular cube
complexes in high dimensions. Moreover, for some applications, like
topological image analysis or analysis of voxel data, the regular cube
complex is a very convenient framework~\cite{KaMiKoMr2004}.
A regular cube complex can be easily specified by an $n$-dimensional
array of binary values (bitmap) and a regular $n$-dimensional cube is
placed where the bit is on. For example the cube complex shown in
Figure~\ref{fig:cube_mesh} can be created by specifying the bitmap
array
\[
\begin{bmatrix}
0 & 1\\ 1 & 1
\end{bmatrix}\, .
\]
A bitmap suffices to describe the top level cubes, but a cube array
(like simplex array) is used during construction of differential
operators and for computing faces. In this paper we describe the
construction of exterior derivative, Hodge star and Whitney forms on
simplicial complexes. For cube complexes we describe only the
construction of exterior derivative and lower dimensional
faces. However, the other operators and objects are also implemented
in PyDEC for such complexes. For example, Whitney-like elements for
hexahedral grids are described in~\cite{BoRo2002} and are implemented
in PyDEC.
\begin{figure}
\begin{center}
\includegraphics[height=1.25in]{pydec/cube_example_filled}
\end{center}
\caption{Regular cube mesh with enumerated vertices and faces.}
\label{fig:cube_mesh}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[height=1.25in]{pydec/cube_example_oriented}
\end{center}
\caption{Regular cube complex with oriented edges and faces.}
\label{fig:cube_oriented}
\end{figure}
Converting a bitmap representation of a mesh into a cube array
representation is straightforward. For example, the cube array
representation of the mesh in Figure~\ref{fig:cube_mesh} is
\[
\mathbb{C}_2 =
\begin{bmatrix}
0 & 0 & & 0 & 1 \\
1 & 0 & & 0 & 1 \\
1 & 1 & & 0 & 1
\end{bmatrix}.
\]
As with the simplex arrays, the rows of $\mathbb{C}_2$ correspond to
individual two-dimensional cubes in the mesh. The two left-most
columns of $\mathbb{C}_2$ encode the origins of each two-dimensional
cube, namely $(0,0)$, $(1,0)$ and $(1,1)$. The remaining two columns
encode the coordinate directions spanned by the cube. Since
$\mathbb{C}_2$ represents the top-level elements, all cubes span both
the $x$ (coordinate $0$) and $y$ (coordinate $1$) dimensions. In
general, the first $n$ columns of $\mathbb{C}_k$ encode the origin or
corner of a cube while the remaining $k$ columns identify the
coordinate directions swept out by the cube. We note that the cube
array representation is similar to the cubical notation used by
Sen~\cite{Sen2003}.
The edges of the mesh in Figure~\ref{fig:cube_mesh} are represented by
the cube array
\[
\mathbb{C}_1 =
\begin{bmatrix}
0 & 0 & & 0 \\
0 & 0 & & 1 \\
0 & 1 & & 0 \\
1 & 0 & & 0 \\
1 & 0 & & 1 \\
1 & 1 & & 0 \\
1 & 1 & & 1 \\
1 & 2 & & 0 \\
2 & 0 & & 1 \\
2 & 1 & & 1
\end{bmatrix}\,
\]
where again the first two columns encode the origin of each edge and
the last column indicates whether the edge points in the $x$ or $y$
direction. For example, the row $[0,0,0]$ corresponds to edge $0$
in Figure~\ref{fig:cube_oriented} which begins at $(0,0)$ and
extends one unit in the $x$ direction. Similarly the row $[2, 1, 1]$
encodes an edge starting at $(2,1)$ extending one unit in the $y$
direction. Since zero-dimensional cubes (points) have no spatial extent
their cube array representation
\[
\mathbb{C}_0 =
\begin{bmatrix}
0 & 0 \\
0 & 1 \\
1 & 0 \\
1 & 1 \\
1 & 2 \\
2 & 0 \\
2 & 1 \\
2 & 2 \\
\end{bmatrix}\,
\]
contains only their coordinate locations.
The cube array provides a convenient representation for regular cube
complexes. While a bitmap representation of the top-level cubes is
generally more compact, the cube array representation generalizes
naturally to lower-dimensional faces and is straightforward to
manipulate.
\section{Rips Complex Representation}
\label{sec:rips_complex}
The \emph{Rips complex}, or Vietoris-Rips complex of a point set is
defined by forming a simplex for every subset of vertices with
diameter less than or equal to a given distance $r$. For example, if
pair of vertices $(v_i,v_j)$ are no more than distance $r$ apart, then
the Rips complex contains an edge (1-simplex) between the vertices.
In general, a set of $p \geq 2$ vertices forms a $(p-1)$-simplex when
all pairs of vertices in the set are separated by at most $r$.
\begin{figure}
\begin{center}
\includegraphics[width=0.7\textwidth]{pydec/rips/illustration}
\end{center}
\caption{Broadcast radii and Rips complex for a sensor network.}
\label{fig:sensor_network_illustration}
\end{figure}
In recent work, certain sensor network coverage problems have been
shown to reduce to finding topological properties of the network's
Rips complex, at least for an abstract model of sensor
networks~\cite{SiGh2007}. Such coordinate-free methods
rely only on pairwise communication between nodes and do not require
the use of positioning devices. These traits are especially important
in the context of ad-hoc wireless networks with limited per-node
resources. Figure \ref{fig:sensor_network_illustration} depicts a
planar sensor network and its associated Rips complex.
In this section we describe an efficient method for computing the Rips
complex for a set of points. Although we consider only the case of
points embedded in Euclidean space, our methodology applies to more
general metric spaces. Indeed, only the construction of the
$1$-skeleton of the Rips complex requires metric information. The
higher-dimensional simplices are constructed directly from the
$1$-skeleton.
We compute the $1$-skeleton of the Rips complex with a kD-Tree data
structure. Specifically, for each vertex $v_i$ we compute the set of
neighboring vertices $\{ v_j : \norm{v_j - v_i} \leq r\}$. The
hierarchical structure of the kD-Tree allows such queries to be
computed efficiently.
The $1$-skeleton of the Rips complex is stored in an array
$\mathbb{S}_1$, using the convention discussed in Section
\ref{sec:simplicial_complex}. Additionally, the (oriented) edges of
the $1$-skeleton are used to define $\mathbb{E}$, a directed graph
stored in a sparse matrix format. Specifically, $\mathbb{E}(i,j) = 1$
if $[i,j]$ is an edge of the Rips complex, and zero otherwise. For
the Rips complex depicted in Figure
\ref{fig:simple_rips_complex_example},
\begin{align*}
\mathbb{S}_1 = \begin{bmatrix}
0 & 1 \\
0 & 2 \\
0 & 3 \\
1 & 3 \\
2 & 3
\end{bmatrix},
&&
\mathbb{E} = \begin{bmatrix}
0 & 1 & 1 & 1 \\
0 & 0 & 0 & 1 \\
0 & 0 & 0 & 1 \\
0 & 0 & 0 & 0
\end{bmatrix},
\end{align*}
are the corresponding simplex array and directed graph respectively.
\begin{figure}
\begin{center}
\includegraphics[width=0.7\textwidth]
{pydec/rips/small_rips_complex_example}
\end{center}
\caption{Five directed edges form the $1$-skeleton of the Rips
complex.}
\label{fig:simple_rips_complex_example}
\end{figure}
The arrays of higher dimensional simplices $\mathbb{S}_2,
\mathbb{S}_3, \ldots$ can be computed as follows. Let
$\mathbb{F}_p$ denote the (sparse) matrix whose rows are identified
with the $p$-simplices as specified by $\mathbb{S}_p$. Each row of
$\mathbb{F}_p$ encodes the vertices which form the corresponding
simplex. Specifically, $\mathbb{F}_p(i,j)$ takes the value $1$ if the
$i$-th simplex contains vertex $j$ and zero otherwise. For the
example shown in Figure \ref{fig:simple_rips_complex_example},
\begin{align*}
\mathbb{F}_1 = \begin{bmatrix}
1 & 1 & 0 & 0 \\
1 & 0 & 1 & 0 \\
1 & 0 & 0 & 1 \\
0 & 1 & 0 & 1 \\
0 & 0 & 1 & 1
\end{bmatrix},
\end{align*}
encodes the edges stored in $\mathbb{S}_1$. Once $\mathbb{F}_p$ is
constructed we compute the sparse matrix-matrix product $\mathbb{F}_p
\,\mathbb{E}$. For our example the result is
\begin{align*}
\mathbb{F}_1\, \mathbb{E} =
\begin{bmatrix}
0 & 1 & 1 & 2\\
0 & 1 & 1 & 2\\
0 & 1 & 1 & 1\\
0 & 0 & 0 & 1\\
0 & 0 & 0 & 1
\end{bmatrix}.
\end{align*}
Like $\mathbb{F}_p$, the product $\mathbb{F}_p\, \mathbb{E}$ is a
matrix that relates the $p$-simplices to the vertices: the matrix
entry $(i,j)$ of $\mathbb{F}_p\, \mathbb{E}$ counts the number of
directed edges that exist from the vertices of simplex $i$ to vertex
$j$. When the value of $(i,j)$ entry of $\mathbb{F}_p\, \mathbb{E}$ is
equal to $p+1$, we form a $p$-simplex of the Rips complex by
concatenating simplex $i$ with vertex $j$. In the example, matrix
entries $(0,3)$ and $(1,3)$ of $\mathbb{F}_1\,\mathbb{E}$ are equal to
$2$ which implies that the $2$-skeleton of the Rips complex contains
two simplices, formed by appending vertex $3$ to the $1$-simplices
$[0, 1]$ and $[0, 2]$, or
\begin{align*}
\mathbb{S}_2 = \begin{bmatrix}
0 & 1 & 3 \\
0 & 2 & 3
\end{bmatrix},
\end{align*}
in array format. This process may be applied recursively to develop
higher dimensional simplices $\mathbb{S}_3, \mathbb{S}_4,
\ldots$ as required by the application. Thus our algorithm computes
simplices of the Rips complex with a handful of sparse and dense
matrix operations.
\section{Abstract Simplicial Complex Representation}
\label{sec:abstrct_smplcl_cmplx}
In Section~\ref{sec:rips_complex} we saw an example of a simplicial
complex which was not a manifold complex
(Figure~\ref{fig:sensor_network_illustration}). Rips complexes
described in Section~\ref{sec:rips_complex} demonstrate one way to
construct such complexes in PyDEC, starting from locations of
vertices. There are other applications, for example in topology, where
we would like to create a simplicial complex which is not necessarily
a manifold. In addition we would like to do this without requiring
that the location of vertices be given. For example, in topology,
surfaces are often represented as a polygon with certain sides
identified. One way to describe such an object is as an \emph{abstract
simplicial complex} \cite[Section 3]{Munkres1984}. This is a
collection of finite nonempty sets such that if a set is in the
collection, so are all the nonempty subsets of it.
Figure~\ref{fig:asc_examples} shows two examples of abstract simplicial
complexes created in PyDEC.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.4,trim=1.5in 1in 1.5in 2in, clip]
{cat/nonorntbl/moebius_6}
\includegraphics[scale=0.4,trim=1.5in 1in 1.5in 1in, clip]
{cat/nonorntbl/prjctvpln_10}
\caption{Examples of abstract simplicial complexes. The one of the
left represents the triangulation of a M\"obius strip and the one
on the right that of a projective plane.}
\label{fig:asc_examples}
\end{figure}
In PyDEC, abstract simplicial complexes are created by specifying a
list of arrays. Each array contains simplices of a single dimension,
specified as an array of vertex numbers. Lower dimensional faces of a
simplex need not be specified explicitly. For example the M\"obius
strip triangulation shown in Figure~\ref{fig:asc_examples} can be
created by giving the array
\[
\begin{bmatrix}
0 & 1 & 3\\0 & 3 & 5\\3 & 2 & 5\\5 & 2 & 4\\2 & 0 & 4\\0 & 1 & 4
\end{bmatrix}
\]
as input to PyDEC. Abstract simplicial complexes need not be a
triangulation of a manifold. For example one consisting of 2 triangles
with an extra edge attached and a standalone vertex may be created
using a list consisting of the arrays
\[
\begin{bmatrix}
5
\end{bmatrix}\qquad
\begin{bmatrix}
1 & 4
\end{bmatrix}\qquad
\begin{bmatrix}
0 & 1 & 2\\ 1 & 2 & 3
\end{bmatrix}
\]
as input.
The boundary matrices of a simplicial complex encode the connectivity
information and can be computed from a purely combinatorial
description of a simplicial complex. The locations of the vertices are
not required. Thus the abstract simplicial complex structure is all
that is required to compute these matrices as will be described in the
next section.
\section{Discrete Exterior Derivative}
\label{sec:exterior_derivative}
Given a manifold $M$, the exterior derivative $\operatorname{d} : \Omega^p(M) \to
\Omega^{p+1}$ which acts on differential $p$-forms, generalizes the
derivative operator of calculus. When combined with metric dependent
operators Hodge star, sharp, and flat appropriately, the vector
calculus operators div, grad and curl can be generated from $\operatorname{d}$. But
$\operatorname{d}$ itself is a metric independent operators whose definition does
not require any Riemannian metric on the
manifold. See~\cite{AbMaRa1988} for details. The discrete exterior
derivative (which we will also denote as $\dd$) in PyDEC is defined as
is usual in the literature, as the coboundary operator of algebraic
topology~\cite{Munkres1984}. Thus
\begin{equation*}
\eval{\dd_p a}{c} = \eval{a}{\boundary_{p+1} c}\, ,
\end{equation*}
for arbitrary $p$-cochain $a$ and $(p+1)$-chain $c$. Recall that the
boundary operator on cochains, $\boundary_p : C_p(K) \to C_{p-1}(K)$
is defined by extension of its definition on an oriented simplex.
The boundary operator on a $p$-simplex $\sigma^p = \simplex{p}$ is
given in terms of its $(p-1)$-dimensional faces ($p+1$ in number) as
\begin{equation}\label{eq:boundary}
\boundary_p \sigma^p = \sum_{i=0}^p
(-1)^i \bigl[v_0,\ldots,\widehat{v_i},\ldots,v_p\bigr]\, ,
\end{equation}
where $\widehat{v_i}$ means that $v_i$ is omitted. Therefore, given
an $n$-dimensional simplicial complex represented by
$\mathbb{S}_0,\dots,\mathbb{S}_n$, for the discrete exterior
derivative, it suffices to compute $\boundary_0, \dots, \boundary_n$.
As is usual in algebraic topology, in PyDEC we compute matrix
representations of these in the elementary chain basis. Boundary
matrices are useful in finite elements since their transposes are the
coboundary operators. They are also useful in computational topology
since homology and cohomology groups are the quotient groups of kernel
and image of boundary matrices~\cite{Munkres1984}.
For the complex pictured in Figure~\ref{fig:mesh_example_oriented}
the boundary operators are
\begin{align*}
\boundary_0 = \begin{bmatrix}
0 & 0 & 0 & 0 & 0
\end{bmatrix}\, ,
&&
\boundary_1 = \begin{bmatrix}
-1 & -1 & \p0 & \p0 & \p0 &
\p0 & \p0\\
\p1 & \p0 & -1 & -1 & \p0 &
\p0 & \p0\\
\p0 & \p0 & \p1 & \p0 &
-1 & -1 & \p0\\
\p0 & \p1 & \p0 & \p1 &
\p1 & \p0 & -1\\
\p0 & \p0 & \p0 & \p0 &
\p0 & \p1 & \p1
\end{bmatrix}\, ,
\\
\boundary_2 = \begin{bmatrix}
\p1 & \p0 & \p0\\
-1 & \p0 & \p0\\
\p0 & \p1 & \p0\\
\p1 & -1 & \p0\\
\p0 & \p1 & -1\\
\p0 & \p0 & \p1\\
\p0 & \p0 & -1
\end{bmatrix}.
&&
\end{align*}
In the following we describe an algorithm that takes as input
$\mathbb{S}_n$ and computes both $\mathbb{S}_{n-1}$ and
$\boundary_{n-1}$. This procedure is applied recursively to produce
all faces of the complex and their boundary operator at that
dimension.
The first step of the algorithm converts a simplex array
$\mathbb{S}_n$ into a canonical format. In the canonical format
each simplex (row of $\mathbb{S}_n$) is replaced by the simplex
with sorted indices and the relative parity of the original ordering
to the sorted order. For instance, the simplex $(1,3,2)$ becomes
$((1,2,3), -1)$ since an odd number of transpositions (namely one) are
needed to transform $(1,3,2)$ into $(1,2,3)$. Similarly, the
canonical format for simplex $(2,1,4,3)$ is $((1,2,3,4),+1)$ since an
even number of transpositions are required to sort the simplex
indices. Since the complex dimension $n$ is typically small (i.e. $<
10$), a simple sorting algorithm such as insertion sort is employed at
this stage. We denote the aforementioned process
\texttt{canonical\_format} $( \mathbb{S}_n ) \to \mathbb{S}^+_n$
where the rightmost column of $\mathbb{S}^+_n$ contains the simplex
parity. Applying \texttt{canonical\_format} to $\mathbb{S}_2$ in
our example yields
\begin{equation*}
\mathbb{S}_2 = \begin{bmatrix}
0 & 1 & 3 \\
1 & 2 & 3 \\
2 & 4 & 3
\end{bmatrix}
\to
\begin{bmatrix}
0 & 1 & 3 & & \p1 \\
1 & 2 & 3 & & \p1\\
2 & 3 & 4 & & -1
\end{bmatrix}
= \mathbb{S}^+_2
\end{equation*}
Once a simplex array $\mathbb{S}_n$ has been transformed into
canonical format, the $(n-1)$-dimensional faces $\mathbb{S}_{n-1}$
and boundary operator $\boundary_n$ are readily obtained. We denote
this process
\[
\texttt{boundary\_faces}(\mathbb{S}^+_n) \to
\mathbb{S}_{n-1},\boundary_n \, .
\]
In order to establish the correspondence between the $n$-dimensional
simplices and their faces, we first enumerate the simplices by adding
another column to $\mathbb{S}^+_n$ to form $\mathbb{S}^{++}_n$.
For example,
\begin{equation*}
\mathbb{S}^+_2 = \begin{bmatrix}
0 & 1 & 3 & & \p1 \\
1 & 2 & 3 & & \p1\\
2 & 3 & 4 & & -1
\end{bmatrix}
\to
\mathbb{S}^{++}_2 = \begin{bmatrix}
0 & 1 & 3 & & \p1 & & 0 \\
1 & 2 & 3 & & \p1 & & 1 \\
2 & 3 & 4 & & -1 & & 2
\end{bmatrix}.
\end{equation*}
The formula~\eqref{eq:boundary} is applied to $\mathbb{S}^{++}_n$
in a columnwise fashion by excluding the $i$-th column of simplex
indices, multiplying the parity column by $(-1)^i$, and carrying the
last column over unchanged. For example,
\begin{equation*}
\mathbb{S}^{++}_2 = \begin{bmatrix}
0 & 1 & 3 & & \p1 & & 0 \\
1 & 2 & 3 & & \p1 & & 1 \\
2 & 3 & 4 & & -1 & & 2
\end{bmatrix}
\to
\begin{bmatrix}
1 & 3 & & \p1 & & 0 \\
2 & 3 & & \p1 & & 1 \\
3 & 4 & & -1 & & 2 \\
0 & 3 & & -1 & & 0 \\
1 & 3 & & -1 & & 1 \\
2 & 4 & & \p1 & & 2 \\
0 & 1 & & \p1 & & 0 \\
1 & 2 & & \p1 & & 1 \\
2 & 3 & & -1 & & 2
\end{bmatrix}
\end{equation*}
The resultant array is then sorted by the first $n$ columns in
lexicographical order, allowing the unique faces to then be extracted.
\begin{equation*}
\begin{bmatrix}
1 & 3 & & \p1 & & 0 \\
2 & 3 & & \p1 & & 1 \\
3 & 4 & & -1 & & 2 \\
0 & 3 & & -1 & & 0 \\
1 & 3 & & -1 & & 1 \\
2 & 4 & & \p1 & & 2 \\
0 & 1 & & \p1 & & 0 \\
1 & 2 & & \p1 & & 1 \\
2 & 3 & & -1 & & 2
\end{bmatrix}
\to
\begin{bmatrix}
0 & 1 & & \p1 & & 0 \\
0 & 3 & & -1 & & 0 \\
1 & 2 & & \p1 & & 1 \\
1 & 3 & & \p1 & & 0 \\
1 & 3 & & -1 & & 1 \\
2 & 3 & & \p1 & & 1 \\
2 & 3 & & -1 & & 2 \\
2 & 4 & & \p1 & & 2 \\
3 & 4 & & -1 & & 2 \\
\end{bmatrix}
\to
\begin{bmatrix}
0 & 1 \\
0 & 3 \\
1 & 2 \\
1 & 3 \\
2 & 3 \\
2 & 4 \\
3 & 4 \\
\end{bmatrix}.
\end{equation*}
Furthermore, a Compressed Sparse Row (CSR)~\cite{Saad2003} sparse
matrix representation of $\boundary_n$ as
\[
\boundary_n = (\texttt{ptr},\texttt{indices},\texttt{data})
\]
is obtained from the sorted matrix. For $\boundary_2$ these are
\begin{align*}
\texttt{ptr} &=
\begin{bmatrix}
0 & 1 & 2 & 3 & 5 & 7 & 8 & 9
\end{bmatrix}\, , \\
\texttt{indices} &=
\begin{bmatrix}
\p0 & \p0 & \p1 & \p0 & \p1 & \p1 & \p2 & \p2 & \p2
\end{bmatrix}\, , \\
\texttt{data} &=
\begin{bmatrix}
\p1 & -1 & \p1 & \p1 & -1 & \p1 & -1 & \p1 & -1
\end{bmatrix}\,
\end{align*}
where \texttt{indices} and \texttt{data} correspond to the fourth and
third rows of the sorted matrix.
This process is then applied to $\mathbb{S}_{n-1}$ and so on down
the dimension. Since the lower dimensional simplex array rows are
already sorted, those arrays are already in canonical format.
Thus a single algorithm generates the lower dimensional faces as well
as the boundary matrices at all the dimensions. The boundary matrices,
and hence the coboundary operators, are generated in a convenient
sparse matrix format.
\subsection{Generalization to Abstract Complexes}
The boundary operators and faces of an abstract simplicial complex are
computed with a straightforward extension of the
\texttt{boundary\_faces} algorithm. Recall from
Section~\ref{sec:abstrct_smplcl_cmplx} that an abstract simplicial
complex is specified by a \emph{list} of simplex arrays of different
dimensions, where the lower-dimensional simplex arrays represent
simplices that are not a face of any higher-dimensional simplex.
Generalizing the previous scheme to the case of abstract simplicial
complexes is accomplished by (1) augmenting the set of computed faces
with the user-specified simplices and (2) modifying the computed
boundary operator accordingly.
Consider the abstract simplicial complex represented by the simplex
arrays
\[
\mathbb{S}_0 = [5] \qquad\qquad
\mathbb{S}_1 = [1, 4] \qquad\qquad
\mathbb{S}_2 = \begin{bmatrix} 0 & 1 & 2 \\ 1 & 2 &3 \end{bmatrix}
\]
which consists of two triangles, an edge, and an isolated vertex.
Applying \texttt{boundary\_faces} to $\mathbb{S}_{2}$ produces
an array of face edges and corresponding boundary operator
\[
\texttt{boundary\_faces}(\mathbb{S}_2) \to
\mathbb{S}_{1},\boundary_2 \ =
\begin{bmatrix}
0 & 1 \\
0 & 2 \\
1 & 2 \\
1 & 3 \\
2 & 3 \\
\end{bmatrix},
\begin{bmatrix}
\p1 & \p0 \\
-1 & \p0 \\
\p1 & \p1 \\
\p0 & -1 \\
\p0 & \p1 \\
\end{bmatrix},
\]
which includes all but the user-specified edge $[1,4]$.
User-specified simplices are then incorporated into the simplex array
in a three-stage process:
\begin{inparaenum}[(1)]
\item user-specified simplices are concatenated to the computed face
array;
\item the rows of the combined simplex array are sorted
lexicographically;
\item redundant simplices (if any) are removed from the sorted array.
\end{inparaenum}
Upon completion, the augmented simplex array contains the union of the
face simplices and the user-specified simplices. Continuing the
example, the edge $[1,4]$ is incorporated into $\mathbb{S}_1$ as
follows
\[
\begin{bmatrix}
0 & 1 \\
0 & 2 \\
1 & 2 \\
1 & 3 \\
2 & 3 \\
\end{bmatrix},
\begin{bmatrix}
1 & 4 \\
\end{bmatrix}
\to
\begin{bmatrix}
0 & 1 \\
0 & 2 \\
1 & 2 \\
1 & 3 \\
2 & 3 \\
1 & 4 \\
\end{bmatrix}
\to
\begin{bmatrix}
0 & 1 \\
0 & 2 \\
1 & 2 \\
1 & 3 \\
1 & 4 \\
2 & 3 \\
\end{bmatrix}
\to
\begin{bmatrix}
0 & 1 \\
0 & 2 \\
1 & 2 \\
1 & 3 \\
1 & 4 \\
2 & 3 \\
\end{bmatrix} = \mathbb{S}_1
\]
In the final stage of the procedure, the computed boundary operator
($\boundary_2$ in the example) is updated to reflect the newly
incorporated simplices. Since the new simplices do not lie in the
boundary of any higher-dimensional simplex, we may simply add empty
rows into the sparse matrix representation of the boundary operator
for each newly added simplex. Therefore, the boundary operator update
procedure amounts to a simple remapping of row indices. In the
example, the addition of the edge $[1,4]$ into the fifth row of the
simplex array requires the addition of an empty row into the boundary
operator at the corresponding position,
\[
\begin{bmatrix}
\p1 & \p0 \\
-1 & \p0 \\
\p1 & \p1 \\
\p0 & -1 \\
\p0 & \p1 \\
\end{bmatrix}
\to
\begin{bmatrix}
\p1 & \p0 \\
-1 & \p0 \\
\p1 & \p1 \\
\p0 & -1 \\
\p0 & \p0 \\
\p0 & \p1 \\
\end{bmatrix} = \boundary_2.
\]
The Rips complex of Section~\ref{sec:rips_complex} does have the
location information for the vertices. However, ignoring those, such a
complex is an abstract simplicial complex. Thus the boundary matrices
for a Rips complex can be computed as described above. In practice
some efficiency can be obtained by ordering the computation
differently, so that the matrices are built as the complex is being
built from the edge skeleton of the Rips complex. That is how it is
implemented in PyDEC.
\subsection{Boundary operators and faces for cubical complexes}
\label{subsec:cbbndry}
The algorithm used to compute the faces and boundary operator of a
given cube array ($\mathbb{C}_p \rightarrow \mathbb{C}_{p-1},
\boundary_p$) is closely related to the procedure discussed in
Section~\ref{sec:exterior_derivative} for simplex arrays. Consider a
general $p$-cube in $n$-dimensions, denoted by the pair
$\bigl[(c_0,\ldots, c_{n-1}) (d_0,\ldots,d_{p-1})\bigr]$
where $(c_0,\ldots, c_{n-1})$ are the coordinates of cube's origin and
$(d_0,\ldots,d_{p-1})$ are the directions which the $p$-cube spans.
Note that the values $[c_0, \ldots, c_{n-1}, d_0, \ldots, d_{p-1}]$
correspond exactly to a row of the cube array representation
introduced in Section~\ref{sec:cube_complex}. Using this notation,
the boundary of a $p$-cube is given by the expression
\begin{multline}\label{eq:cube_boundary}
\boundary_p \bigl[(c_0,\ldots, c_{n-1}) (d_0,\ldots,d_{p-1})\bigr] =
\sum_{i=0}^{p-1} (-1)^i \bigl(\bigl[(c_0, \ldots, c_{d_i} + 1,
\ldots,
c_{n-1}) (d_0,\ldots,\widehat{d_i},\ldots,d_{p-1})\bigr]
- \\
\bigl[(c_0, \ldots, c_{d_i} + 0, \ldots, c_{n-1})
(d_0,\ldots,\widehat{d_i},\ldots,d_{p-1})\bigr]\bigr)\,
\end{multline}
where $\widehat{d_i}$ denotes the omission of the $i$-th spanning
direction and $c_{d_i}$ is the corresponding coordinate. For example,
the boundary of a square centered at the location $(10,20)$ is
\begin{equation}
\boundary_2 \bigl[(10, 20) (0, 1)\bigr] =
\bigl[(11,20) (1)\bigr] - \bigl[(10,20) (1)\bigr] -
\bigl[(10,21) (0)\bigr] + \bigl[(10,20) (0)\bigr].
\end{equation}
The canonical format for a $p$-cube is the one where the spanning
directions are specified in ascending order. For instance, the
$2$-cube $\bigl[(10, 20) (0, 1)\bigr]$ is in the canonical format
because $d_0 < d_1$. As with simplices, each cube has a unique
canonical format, through which duplicates are easily identified.
Since the top-level cube array $\mathbb{C}_n$ is generated from a
bitmap it is already in the canonical format and no reordering of
indices or parity tracking is necessary.
Applying Equation~\ref{eq:cube_boundary} to a $p$-cube array with $N$
members generates a collection $2 N$ oriented faces. In the mesh
illustrated in Figure~\ref{fig:cube_mesh} the three squares in
$\mathbb{C}_2$ are initially expanded into
\[
\mathbb{C}_2 =
\begin{bmatrix}
0 & 0 & & 0 & 1 \\
1 & 0 & & 0 & 1 \\
1 & 1 & & 0 & 1
\end{bmatrix}
\to
\begin{bmatrix}
0 & 0 & & 1 & & -1 & & 0\\
1 & 0 & & 1 & & -1 & & 1\\
1 & 1 & & 1 & & -1 & & 2\\
1 & 0 & & 1 & & \p1 & & 0\\
2 & 0 & & 1 & & \p1 & & 1\\
2 & 1 & & 1 & & \p1 & & 2\\
0 & 0 & & 0 & & \p1 & & 0\\
1 & 0 & & 0 & & \p1 & & 1\\
1 & 1 & & 0 & & \p1 & & 2\\
0 & 1 & & 0 & & -1 & & 0\\
1 & 1 & & 0 & & -1 & & 1\\
1 & 2 & & 0 & & -1 & & 2\\
\end{bmatrix} = \mathbb{C}_1^+
\]
where the fourth column of $\mathbb{C}_1^+$ encodes the orientation of
the face and the fifth column records the $2$-cube to which each face
belongs. Sorting the rows of $\mathbb{C}_1^+$ in lexicographical
order
\[
\begin{bmatrix}
0 & 0 & & 1 & & -1 & & 0\\
1 & 0 & & 1 & & -1 & & 1\\
1 & 1 & & 1 & & -1 & & 2\\
1 & 0 & & 1 & & \p1 & & 0\\
2 & 0 & & 1 & & \p1 & & 1\\
2 & 1 & & 1 & & \p1 & & 2\\
0 & 0 & & 0 & & \p1 & & 0\\
1 & 0 & & 0 & & \p1 & & 1\\
1 & 1 & & 0 & & \p1 & & 2\\
0 & 1 & & 0 & & -1 & & 0\\
1 & 1 & & 0 & & -1 & & 1\\
1 & 2 & & 0 & & -1 & & 2\\
\end{bmatrix}
\to
\begin{bmatrix}
0 & 0 & & 0 & \p1 & & 0\\
0 & 0 & & 1 & -1 & & 0\\
0 & 1 & & 0 & -1 & & 0\\
1 & 0 & & 0 & \p1 & & 1\\
1 & 0 & & 1 & -1 & & 1\\
1 & 0 & & 1 & \p1 & & 0\\
1 & 1 & & 0 & \p1 & & 2\\
1 & 1 & & 0 & -1 & & 1\\
1 & 1 & & 1 & -1 & & 2\\
1 & 2 & & 0 & -1 & & 2\\
2 & 0 & & 1 & \p1 & & 1\\
2 & 1 & & 1 & \p1 & & 2\\
\end{bmatrix}
\to
\begin{bmatrix}
0 & 0 & & 0 \\
0 & 0 & & 1 \\
0 & 1 & & 0 \\
1 & 0 & & 0 \\
1 & 0 & & 1 \\
1 & 1 & & 0 \\
1 & 1 & & 1 \\
1 & 2 & & 0 \\
2 & 0 & & 1 \\
2 & 1 & & 1
\end{bmatrix}
=
\mathbb{C}_1
\]
allows the unique faces to be extracted. Lastly, a sparse matrix
representation of the boundary operator is obtained from the sorted
cube array in the same manner as for simplices.
\section{Review of Whitney Map and Whitney Forms}
\label{sec:whitney}
In this section we review and collect some material, most of which is
well-known in DEC and finite element exterior calculus research
communities. It is included here partly to fix notation. In this
section we also give the monomials based definition of inner product
of differential forms. This is not the way inner product of forms is
usually defined in most textbooks, \cite{Morita2001} being one
exception we know of. The monomial form leads to an efficient
algorithm for computation of stiffness and mass matrices for Whitney
forms given in Section~\ref{sec:whtny_innrprdct}.
The basic function spaces that are useful with exterior calculus are
the space of square integrable $p$-forms on a manifold and Sobolev
spaces derived from that. Let $M$ be a \emph{Riemannian manifold}, a
manifold on which a smoothly varying inner product is defined on the
tangent space at each point. Let $g$ be its \emph{metric}, a smooth
tensor field that defines the inner product on the tangent space at
each point on $M$.
For differential forms on such a manifold $M$, the space of square
integrable forms is denoted $L^2 \Omega^p(M)$. One can then define the
spaces $H\Omega^p(M)$ which generalize the spaces $H(\operatorname{div})$ and
$H(\curl)$ used in mixed finite element methods~\cite{ArFaWi2010}. To
define $L^2 \Omega^p(M)$ one has to define an inner product on the
space of forms which is our starting point for this section. All these
function spaces have been discussed in~\cite{ArFaWi2006}
and~\cite{ArFaWi2010}. The definitions and properties of Whitney map
and Whitney forms is in \cite{Dodziuk1976}, the geometric analysis
background is in~\cite{Jost2005} and the basic definition of inner
products on forms is in~\cite[page 411]{AbMaRa1988}.
\subsection{Inner product of forms}
To define the spaces $L^2\Omega^p(M)$ and $H\Omega^p(M)$ more
precisely, we recall the definitions related to inner products of
forms. We will need the exterior calculus operators wedge product and
Hodge star which we recall first. For a manifold $M$ the \emph{wedge
product} $\wedge: \Omega^p(M) \times \Omega^q(M) \to
\Omega^{p+q}(M)$ is an operator for building $(p+q)$-forms from
$p$-forms and $q$-forms. It is defined as the skew-symmetrization of
the tensor product of the two forms involved. For a Riemannian
manifold of dimension $n$, the Hodge star operator $\hodge:\Omega^p(M)
\to \Omega^{n-p}(M)$ is an isomorphism between the spaces of $p$ and
$(n-p)$-forms. For more details, see \cite[page 394]{AbMaRa1988} for
wedge products and \cite[page 411]{AbMaRa1988} for Hodge star.
\begin{definition}
\label{def:wdghdg_ip}
Given two smooth $p$-forms $\alpha$, $\beta \in \Omega^p(M)$ on a
Riemannian manifold $M$, their \emph{pointwise inner product} at
point $x \in M$ is defined by
\begin{equation} \label{eq:wdghdg_ip}
\ainnerproduct{\alpha(x)}{\beta(x)}\,\mu = \alpha(x) \wedge \hodge
\beta(x) \, ,
\end{equation}
where $\mu = \hodge 1$ is the volume form associated with the metric
induced by the inner product on $M$.
\end{definition}
The pointwise inner products of forms can be defined in another way,
which will be more useful to us in computations. The second definition
given below in Definition~\ref{def:monomials_ip} is equivalent to the
one given above in Definition~\ref{def:wdghdg_ip}. The operator
$\sharp$ (the \emph{sharp} operator) used below is an isomorphism
between 1-forms and vector fields and is defined by $g(\alpha^\sharp,
X) = \alpha(X)$ for given 1-form $\alpha$ and all vector fields
$X$. See~\cite{AbMaRa1988} for details.
\begin{definition}
\label{def:monomials_ip}
Let $\alpha_1,\dots,\alpha_p$ and $\beta_1,\dots,\beta_p$ be 1-forms
on a Riemannian manifold $M$. By analogy with polynomials we'll call
$p$-forms of the type $\alpha_1 \wedge \dots \wedge \alpha_p$ and
$\beta_1 \wedge \dots \wedge \beta_p$ \emph{monomial
$p$-forms}. Define the following operator at a point $x \in M$:
\begin{equation} \label{eq:monomials_ip}
\aInnerproduct{\alpha_1 \wedge\dots\wedge\alpha_p}
{\beta_1\wedge\dots\wedge\beta_p} := \det\bigl[
g\bigl(\alpha_i^\sharp,\beta_j^\sharp\bigr)\bigr]\, ,
\end{equation}
where $\bigl[g\bigl(\alpha_i^\sharp,\beta_j^\sharp\bigr)\bigr]$ is the
matrix obtained by taking $1 \le i,j \le p$. In the equation above,
all the 1-forms are evaluated at the point $x \in M$. Extend the
operation in \eqref{eq:monomials_ip} bilinearly pointwise to the space
of all $p$-forms. It can be shown that this defines a \emph{pointwise
inner product} of $p$-forms equivalent to the one defined in
\eqref{eq:wdghdg_ip}. Note that if $\alpha_i = \beta_i$ for all $i$,
the expression on the right in~\eqref{eq:monomials_ip} is the Gram
determinant.
\end{definition}
\begin{remark}\label{rem:whch_ip_dfntn}
Note that unlike~\eqref{eq:wdghdg_ip} the definition
in~\eqref{eq:monomials_ip} does not involve wedge product and Hodge
star explicitly. This is an advantage of the latter form since a
discrete wedge product is not available in PyDEC. The RHS
of~\eqref{eq:monomials_ip} does involve the sharp operator, but as
we will see in the next section, this is easy to interpret for the
purpose of discretization in this context.
\end{remark}
\begin{definition} \label{def:L2innrprdctfrms} The pointwise
innerproduct in~\eqref{eq:wdghdg_ip}, or equivalently
in~\eqref{eq:monomials_ip}, induces an $L^2$ \emph{inner product on
$M$} as
\begin{equation} \label{eq:forms_ip}
\pinnerproduct{\alpha}{\beta}_{L^2} =
\int_M \ainnerproduct{\alpha(x)}{\beta(x)}\,\mu \, .
\end{equation}
The space of $p$-forms obtained by completion of $\Omega^p(M)$ under
this inner product is the Hilbert space of \emph{square integrable
$p$-forms} $L^2\Omega^p(M)$. The other useful space mentioned at
the beginning of this section is the Sobolev space $H \Omega^p(M) :=
\{\alpha \in L^2 \Omega^p(M) \;|\; \operatorname{d}\alpha \in L^2
\Omega^{p+1}(M)\}$.
\end{definition}
\subsection{Whitney map and Whitney forms}
Let $K$ be an $n$-dimensional manifold simplicial complex embedded in
$\mathbb{R}^N$ and $\abs{K}$ the underlying space. The metric on the
interiors of the top dimensional simplices of $K$ will be the one
induced from the embedding Euclidean space $\mathbb{R}^N$. As is usual in
finite element methods, finite dimensional subspaces of the function
spaces described in the previous paragraph are used in the numerical
solution of PDEs. The finite dimensional spaces can be obtained by
``embedding'' the space of cochains into these spaces by using an
interpolation. For example, to embed $C^p(K;\mathbb{R})$ into $\Lforms2p$ one
can use the Whitney map $\whitney:C^p(K;\mathbb{R})\to \Lforms2p$, which will
be reviewed in this subsection. The image
$\whitney\bigl(C^p(K;\mathbb{R})\bigr)$ is a linear vector subspace of
$\Lforms2p$ and is the space of \emph{Whitney $p$-forms}
\cite{Whitney1957,Dodziuk1976,Bossavit1988a} and is denoted
$\mathcal{P}_1^- \Omega^p(\abs{K})$ in~\cite{ArFaWi2009,
ArFaWi2010}. (We use $\Omega^p$ instead of $\Lambda^p$ used
in~\cite{ArFaWi2010}.) The embedding of cochains is analogous to how
scalar values at discrete sample points would be interpolated to get a
piecewise affine function. In the scalar case also, the space of such
functions is a vector subspace of square integrable functions. In
fact, $\whitney\bigl(C^0(K;\mathbb{R})\bigr)$, the space of Whitney $0$-forms
is the space of continuous piecewise affine functions on
$\abs{K}$. The Whitney map for $p>0$ is actually built from
barycentric coordinates which are the building blocks of piecewise
linear interpolation. Thus the embedding of $C^p(K;\mathbb{R})$ into
$\Lforms2p$ for $p>0$ can be considered to be a generalization of the
embedding of $C^0(K;\mathbb{R})$ into $\Lforms20$. Thus Whitney forms enable
only low order methods. However, arbitrary degree polynomial spaces
suitable for use in finite element exterior calculus have been
discovered~\cite{ArFaWi2006, ArFaWi2009}. These however are not yet a
part of PyDEC.
The space of Whitney $p$-forms is the space of piecewise smooth
differential $p$-forms obtained by applying the Whitney map to
$p$-cochains. It can be thought of as a method for interpolating
values given on $p$-simplices of a simplicial complex. For example,
inside a tetrahedron Whitney forms allow the interpolation of numbers
on edges or faces to a smooth 1-form or 2-form respectively. As
mentioned above, for 0-cochains, i.e. scalar functions sampled at
vertices, the interpolation is the one obtained using the standard
scalar piecewise affine basis functions on each simplex, that is the
barycentric coordinates corresponding to each vertex of the
simplex. We recall the definition of barycentric coordinates followed
by the definition of the Whitney map.
\begin{definition}\label{def:barycentric_coords}
Let $\sigma^p = \simplex{p}$ be a $p$-simplex embedded in
$\mathbb{R}^N$. The affine functions $\mu_i:\mathbb{R}^N \to \mathbb{R}$, $i=0,\ldots,p$,
which when restricted to $\sigma^p$ take the value 1 on vertex $v_i$
and 0 on the others, are called the \emph{barycentric coordinates}
in $\sigma^p$.
\end{definition}
\begin{definition}
Let $\sigma^p$ be an oriented $p$-simplex $[v_{i_0},\ldots,v_{i_p}]$
in an $n$-dimensional manifold complex $K$, and
$\bigl(\sigma^p\bigr)^\ast$ the corresponding elementary
$p$-cochain. We define
\begin{equation} \label{eq:whitney_map}
\whitney\left(\left(\sigma^p\right)^{\ast}\right) :=
p! \sum_{k=0}^p (-1)^k \mu_{i_k}\,d\mu_{i_0}\wedge\cdots
\wedge\widehat{d\mu_{i_k}}\wedge\cdots\wedge d\mu_{i_p}\, ,
\end{equation}
where $\mu_{i_{k}}$ is the barycentric coordinate function with
respect to vertex $v_{i_{k}}$ and the notation
$\widehat{d\mu_{i_{k}}}$ indicates that the term $d\mu_{i_{k}}$ is
omitted from the wedge product. The \emph{Whitney map}
$\whitney:C^p(K;\mathbb{R})\to\Lforms2p$ is the above map $\whitney$
extended to all of $C^p(K;\mathbb{R})$ by requiring that $\whitney$ be a
linear map. $\whitney(\left(\sigma^p\right)^{\ast})$ is called the
\emph{Whitney form} corresponding to $\sigma^p$, and for a general
cochain $c$, $\whitney(c)$ is called the Whitney form
corresponding to $c$.
\end{definition}
For example, the Whitney form corresponding to the edge $[v_0,v_1]$ is
$\whitney([v_0,v_1]^\ast) = \mu_0 \operatorname{d} \mu_1 - \mu_1 \operatorname{d} \mu_0 \, , $ and
the Whitney form corresponding to the triangle $[v_1, v_2, v_3]$ in a
tetrahedron $[v_0, v_1, v_2, v_3]$ is
\[
\whitney([v_1, v_2, v_3]^\ast)
= 2\,(\mu_1 \operatorname{d} \mu_2 \wedge \operatorname{d} \mu_3 - \mu_2 \operatorname{d} \mu_1 \wedge \operatorname{d} \mu_3
+ \mu_3 \operatorname{d} \mu_1 \wedge \operatorname{d} \mu_2)\, .
\]
\begin{remark}\label{rem:mnml_cmbntns}
If we were using local coordinate charts on a manifold then at any
point a $p$-form would be a linear combination of monomials. Note
from \eqref{eq:whitney_map} that the Whitney form
$\whitney(\sigma^\ast)$ is a sum of monomials with
coefficients. Thus Whitney forms allow us to treat forms at a point as
a linear combination of monomials even though we are not using local
coordinate charts.
\end{remark}
We emphasize again that this section was a review of known
material. We have tried to present this material in a manner which
makes it easier to explain the examples of Section~\ref{sec:examples}
and the construction of mass matrix for Whitney forms described in the
next section.
\section{Whitney Inner Product of Cochains}
\label{sec:whtny_innrprdct}
Given a manifold simplicial complex $K$, an inner product between two
$p$-cochains $a$ and $b$ can be defined by first embedding these
cochains into $\Lforms2p$ using Whitney map and then taking the $L^2$
inner product of the resulting Whitney forms \cite{Dodziuk1976}.
\begin{definition}\label{def:whtnyip}
Given two $p$-cochains $a,b \in C^p(K;\mathbb{R})$, their \emph{Whitney inner
product} is defined by
\begin{equation}\label{eq:cochains_ip}
\pinnerproduct{a}{b} := \pinnerproduct{\whitney a}{\whitney b}_{L^2}
= \int_{\abs{K}} \ainnerproduct{\whitney a}{\whitney b} \mu\, ,
\end{equation}
using the $L^2$ inner product on forms given
in~\eqref{eq:forms_ip}. The matrix for Whitney inner product of
$p$-forms in the elementary $p$-cochain basis will be denoted $M_p$.
That is, $M_p$ is a square matrix of order $N_p$ (the number of
$p$-simplices in $K$) such that the entry in row $i$ and column $j$ is
$M_p(i,j) = \pInnerproduct{(\sigma_i^p)^\ast}{(\sigma_j^p)^\ast}$,
where $(\sigma_i^p)^\ast$ and $(\sigma_j^p)^\ast$ are the elementary
$p$-cochains corresponding to the $p$-simplices $\sigma_i^p$ and
$\sigma_j^p$ with index number $i$ and $j$ respectively.
\end{definition}
The integral in~\eqref{eq:cochains_ip} is the sum of integrals over
each top dimensional simplex in $K$. Inside each such simplex the
inner product of smooth forms applies since the Whitney form in each
simplex is smooth all the way up to the boundary of the simplex. The
interior of each top dimensional simplex is given an inner product
that is induced from the standard inner product of the embedding space
$\mathbb{R}^N$.
\begin{remark}
Given cochains $a,b\in C^p(K;\mathbb{R})$ we will refer to their
representations in the elementary cochain basis also as $a$ and
$b$. Then the matrix representation of the Whitney inner product of
$a$ and $b$ is $\pinnerproduct{a}{b} = a^T M_p b$.
\end{remark}
The inner product of cochains defined in this way is a key concept
that connects exterior calculus to finite element methods and
different choices of the inner product lead to different
discretizations of exterior calculus. This is because the inner
product matrix $M_p$ is the mass matrix of finite element methods
based on Whitney forms. The details of the efficient computation of
$M_p$ for any $p$ and $n$ will be given in Section~\ref{subsec:mssmtrx}
and~\ref{subsec:mssmtrx_algrthm}.
Recall that for a Riemannian manifold $M$, if $\codiff_{p+1} :
\Omega^{p+1}(M) \to \Omega^p(M)$ is the codifferential, then the
Laplace-deRham operator on $p$-forms is $\laplacian_p :=
\operatorname{d}_{p-1}\codiff_p + \codiff_{p+1} \operatorname{d}_p$. For a boundaryless $M$ the
codifferential $\codiff_{p+1}$ is the adjoint of the exterior
derivative $d_p$. In case $M$ has a boundary, we have instead that
\begin{equation} \label{eq:adjoint}
\pinnerproduct{\operatorname{d}_p\alpha}{\beta} =
\pinnerproduct{a}{\codiff_{p+1}\beta} + \int_{\partial
M}\alpha\wedge\hodge\beta \, .
\end{equation}
See \cite[Exercise 7.5E]{AbMaRa1988} for a derivation of the
above. Now consider Poisson's equation $\laplacian_p u = f$ on
$p$-forms defined on a $p$-dimensional simplicial manifold complex
$K$. For simplicity, we'll consider the weak form of this using smooth
forms rather than Sobolev spaces of forms. See~\cite{ArFaWi2006,
ArFaWi2010} for a proper functional analytic treatment. We will also
assume that the correct boundary conditions are satisfied, so that the
boundary term in~\eqref{eq:adjoint} is 0. In our simple treatment, the
weak form of the Poisson's equation is to find a $u \in
\Omega^p(\abs{K})$ such that $\pinnerproduct{\codiff_p u}{\codiff_p
v}_{L^2} + \pinnerproduct{\operatorname{d}_p u}{\operatorname{d}_p v}_{L^2} =
\pinnerproduct{f}{v}_{L^2}$. Thus it is clear that a Galerkin
formulation using Whitney forms $\mathcal{P}_1^- \Omega^p$ will
require the computation of a term like $\pinnerproduct{\operatorname{d}_p \whitney
a}{\operatorname{d}_p \whitney b}_{L^2}$ for cochains $a$ and $b$. By the
commuting property of Whitney forms $\operatorname{d}_p W = W \operatorname{d}_p$ (where the
second $\operatorname{d}_p$ is the coboundary operator on cochains) we have that the
above inner product is equal to $\pinnerproduct{\whitney \operatorname{d}_p
a}{\whitney \operatorname{d}_p b}_{L^2}$. (See~\cite{Dodziuk1976} for a proof of
the commuting property.) Now, by definition of the Whitney inner
product of cochains in~\eqref{eq:cochains_ip} this is equal to
$\pinnerproduct{\dd_p a}{\dd_p b}$ in the inner product on
$(p+1)$-cochains. The matrix form of this inner product can be
obtained from the mass matrix $M_{p+1}$ as $\operatorname{d}_p^T\, M_{p+1}\,
\operatorname{d}_p$. This is what we mean when we say that the stiffness matrix can
be computed easily from the mass matrix. The term on the right in the
weak form will use the mass matrix $M_p$. Since codifferential of
Whitney forms is 0, the first term in the weak form has to be handled
in another way, as described in~\cite{ArFaWi2010}.
\begin{remark}\label{rem:stffnss}
Exploiting the aforementioned commutativity of Whitney forms to
compute the stiffness matrix represents a significant simplification
to our software implementation. While computing the stiffness
matrix directly from the definition is possible, it is a complex
operation which requires considerable programmer effort, especially
if the performance of the implementation is important. In contrast,
our formulation requires no additional effort and has the
performance of the underlying sparse matrix-matrix multiplication
implementation, an optimized and natively-compiled routine. All of
the complex indexing, considerations of relative orientation,
mappings between faces and indices, etc. is reduced to a simple
linear algebra expression. The lower dimensional faces are oriented
lexicographically and the orientation information required in
stiffness matrix assembly is implicit in the boundary matrices.
\end{remark}
\subsection{Computing barycentric differentials}
Given that the Whitney form $\whitney(\sigma^\ast)$ in
\eqref{eq:whitney_map} is built using wedges of differentials of
barycentric coordinates, it is clear that the algorithm for computing
an inner product of Whitney forms involves computation of the
gradients or differentials of the barycentric coordinates. The
following lemma shows how these are computed using simple linear
algebra operations.
\begin{lemma}\label{lem:barycentric_grads}
Let $\sigma^p = \simplex{p}$ be a $p$-simplex embedded in $\mathbb{R}^N$, $p
\le N$ where the vertices $v_i \in \mathbb{R}^N$ are given in some basis for
$\mathbb{R}^N$. Let $X\in \mathbb{R}^{N \times p}$ be a matrix whose $j$-th column
consists of the components of $\operatorname{d} \mu_j$ in the dual basis, for $j =
1,\ldots,p$ . Let $V_0\in \mathbb{R}^{N \times p}$ be a matrix whose $j$-th
column is $v_j - v_0$, for $j = 1,\ldots,p$. Then $X^T =
\bigl(V_0^TV_0\bigr)^{-1}V_0^T = V_0^+$, the pseudoinverse of $V_0$.
\end{lemma}
\begin{proof}
Let $\zeta = [\mu_1,\dots,\mu_p]^T$ be the vector of barycentric
coordinates (other than $\mu_0$) with respect to $\sigma^p$, for a
point $x = [x_1,\dots,x_N]^T \in \mathbb{R}^N$. Then by definition of
barycentric coordinates and simplices, $V_0 \zeta = x - v_0$ is the
linear least squares system for the barycentric coordinates. Thus
$\zeta = V_0^+ (x -v_0)$ which implies that $\operatorname{d} \zeta = X^T =
V_0^+$.
\end{proof}
\begin{remark}
The use of normal equations in the solution of the least squares
problem in the above proposition suffers from the well-known
condition squaring problem. This is only likely to be a problem if
the simplices are nearly degenerate. In that case one can just use
an orthogonalization method to compute a QR factorization and use
that to solve the least squares problem. Notice that in typical
physical problems $V_0$ will typically be $2\times 2$, $3\times 2$
or $3 \times 3$ matrix so any of these methods are easy to
implement.
\end{remark}
Once the components for $\operatorname{d} \mu_i$ have been obtained for $i =
1,\dots,p$, the components of $\operatorname{d} \mu_0$ can be obtained by noting
that $\operatorname{d} \mu_0 + \cdots + \operatorname{d} \mu_p = 0$ which follows from the fact
that the barycentric coordinates sum to 1. Also note that the
components of the gradients $\nabla \mu_i$ will be the same as those
of $\operatorname{d} \mu_i$ if the standard metric of Euclidean space is used for
the embedding space $\mathbb{R}^N$ which is the case in all of PyDEC.
\subsection{Whitney inner product matrix}
\label{subsec:mssmtrx}
We will now use the inner product of forms in~\eqref{eq:monomials_ip}
and the cochains inner product defined in~\eqref{eq:cochains_ip} to
give a formula for the computation of $M_p$, the Whitney inner product
matrix for $p$-cochains. We will also refer to this as the Whitney
mass matrix.
\begin{notation}
Given simplices $\sigma$ and $\tau$ the notation $\sigma \succeq
\tau$ or $\tau \preceq \sigma$ means $\tau$ is a face of
$\sigma$. Note that this means $\tau$ can be equal to $\sigma$ since
any simplex is its own face. For proper inclusion we
use $\tau \prec \sigma$ or $\sigma \succ \tau$ to
indicate that $\tau$ is a proper face of $\sigma$. The use of this
notation simplifies the expression of summations over various classes
of simplices in a complex. For example, given two fixed
$p$-simplices $\sigma_i^p$ and $\sigma_j^p$
\[
\underset{\sigma^n \succeq \sigma^p_i, \sigma^p_j}
{\sum_{\sigma^n}}
\]
is read as ``sum over all $n$-simplices $\sigma^n$ which have
$\sigma_i^p$ and $\sigma_j^p$ as faces'' . Another notation used in
the proof of the proposition below is the \emph{star} of a simplex
$\sigma$, written $\St(\sigma)$ (not to be confused with the dual
star $\star \sigma$). This star $\St(\sigma)$ is the union of the
interiors of all simplices of the complex that have $\sigma$ as a
face. That includes $\sigma$ also. The closure of this open set is
called the \emph{closed star} and written $\ClSt{\sigma}$. This is
the union of simplices that contain~$\sigma$.
\end{notation}
\begin{proposition} \label{prop:whtny_ip} Let $\sigma_i^p =
[v_{i_0},\dots,v_{i_p}]$ and $\sigma_j^p = [v_{j_0},\dots,v_{j_p}]$
be oriented $p$-simplices in an $n$-dimensional manifold simplicial
complex $K$, with $0 \le p \le n$. Then the row $i$, column
$j$ entry of the Whitney inner product matrix $M_p$ is given by
\[
M_p(i,j) = (p!)^2 \underset{\sigma^n \succeq \sigma^p_i, \sigma^p_j}
{\sum_{\sigma^n}}
\sum_{k,l=0}^p(-1)^{k+l} c_{kl}\int_{\sigma^n}\mu_{i_k}\mu_{j_l} \, \mu\, ,
\]
where $c_{kl} = 1$ for $p=0$, and for $p>0$
\[
c_{kl} = \det
\begin{bmatrix}
\aInnerproduct{\operatorname{d} \mu_{i_0}}{\operatorname{d} \mu_{j_0}} & \dots &
\widehat{\aInnerproduct{\operatorname{d} \mu_{i_0}}{\operatorname{d} \mu_{j_l}}} &
\dots & \aInnerproduct{\operatorname{d} \mu_{i_0}}{\operatorname{d} \mu_{j_p}}\\
\vdots & & \vdots & & \vdots\\
\widehat{\aInnerproduct{\operatorname{d} \mu_{i_k}}{\operatorname{d} \mu_{j_0}}} &
\dots &
\widehat{\aInnerproduct{\operatorname{d} \mu_{i_k}}{\operatorname{d} \mu_{j_l}}} &
\dots &
\widehat{\aInnerproduct{\operatorname{d} \mu_{i_k}}{\operatorname{d} \mu_{j_l}}}\\
\vdots & & \vdots & & \vdots\\
\aInnerproduct{\operatorname{d} \mu_{i_p}}{\operatorname{d} \mu_{j_0}} & \dots &
\widehat{\aInnerproduct{\operatorname{d} \mu_{i_p}}{\operatorname{d} \mu_{j_l}}} &
\dots & \aInnerproduct{\operatorname{d} \mu_{i_p}}{\operatorname{d} \mu_{j_p}}
\end{bmatrix}\, ,
\]
the hats indicating the deleted terms. Here $\mu$ is the volume form
corresponding to the standard inner product in $\mathbb{R}^N$ and
$\mu_{i_k}$ and $\mu_{j_l}$ are the barycentric coordinates
corresponding to vertices $i_k$ and $j_l$.
\end{proposition}
\begin{proof} See Appendix. \end{proof}
For the $p=n$ case a simpler formulation is given in
Proposition~\ref{prop:p_equal_n}. The above proposition shows that
computation of the Whitney inner product matrix involves computations
of inner products of differentials of barycentric coordinates. Since
the only metric implemented in PyDEC is the standard one inherited
from the embedding space $\mathbb{R}^N$,
\[
\aInnerproduct{\operatorname{d} \mu_i}{\operatorname{d} \mu_j} = g((\operatorname{d} \mu_i)^\sharp,(\operatorname{d}
\mu_j)^\sharp) = \nabla\mu_i \cdot \nabla\mu_j\, .
\]
\begin{example} \label{ex:ttrhdrn_whtny} Consider the simplicial
complex corresponding to a tetrahedron $\sigma^3$ embedded in $\mathbb{R}^3$
for which we want to compute $M_2$, the Whitney inner product matrix
for 2-cochains. Here $N=n=3$ and $p=2$ and $M_2$ is of order $N_2 =
4$, the number of triangles in the complex. Label the vertices as
$0, 1, 2, 3$. Then by PyDEC's lexicographic numbering scheme, the
edges numbered 0 to 5 are $[0,1]$, $[0,2]$, $[0,3]$, $[1,2]$,
$[1,3]$, $[2,3]$ and the triangles numbered 0 to 3 are $[0,1,2]$,
$[0,1,3]$, $[0,2,3]$, and $[1,2,3]$. We will describe the
computation of the row 0, column 3 entry and the row 0, column 1
entry of $M_2$. The $(0,3)$ entry corresponds to the inner product
of cochains $[v_0,v_1,v_2]^\ast$ and $[v_1,v_2,v_3]^\ast$ since, in
the lexicographic ordering and naming convention of PyDEC these are
$\sigma^2_0$ and $\sigma^2_3$ respectively. Thus we are computing
\[
\pInnerproduct{\cochainBasis 2 0}{\cochainBasis 2 3} =
\pInnerproduct{\whitney \cochainBasis 2 0}
{\whitney \cochainBasis 2 3}_{L^2}\, .
\]
The corresponding Whitney forms are
\begin{align*}
\whitney \cochainBasis{2}{0} &= 2! \,
\bigl(\mu_0\, \operatorname{d} \mu_1 \wedge \operatorname{d} \mu_2 -
\mu_1 \operatorname{d} \mu_0 \wedge \operatorname{d} \mu_2 +
\mu_2 \operatorname{d} \mu_0 \wedge \operatorname{d} \mu_1\bigr)\\
\whitney \cochainBasis{2}{3} &= 2! \,
\bigl(\mu_1\, \operatorname{d} \mu_2 \wedge \operatorname{d} \mu_3 -
\mu_2 \operatorname{d} \mu_1 \wedge \operatorname{d} \mu_3 +
\mu_3 \operatorname{d} \mu_1 \wedge \operatorname{d} \mu_2\bigr)\, .
\end{align*}
Then $\pInnerproduct {\whitney \cochainBasis 2 0}{\whitney
\cochainBasis 2 3}_{L^2}/(2!)^2$ is
\begin{equation}\label{expr:ip03full}
\int_{\sigma^3} \mu_0\mu_1
\aInnerproduct{\operatorname{d} \mu_1 \wedge \operatorname{d} \mu_2}
{\operatorname{d}\mu_2\wedge \operatorname{d}\mu_3}\, \mu -\\
\int_{\sigma^3} \mu_0\mu_2
\aInnerproduct{\operatorname{d} \mu_1 \wedge \operatorname{d} \mu_2}
{\operatorname{d}\mu_1\wedge \operatorname{d}\mu_3}\, \mu +
\dots\, ,
\end{equation}
where $\mu$ is just $\operatorname{d} x\, \wedge\, \operatorname{d} y\, \wedge\, \operatorname{d} z$, the
standard volume form in $\mathbb{R}^3$. Each term like
$\aInnerproduct{\operatorname{d}\mu_1 \wedge \operatorname{d} \mu_2} {\operatorname{d}\mu_2\wedge \operatorname{d}\mu_3}$
is
\[
\det
\begin{bmatrix}
\ainnerproduct{\operatorname{d}\mu_1}{\operatorname{d} \mu_2} &
\ainnerproduct{\operatorname{d}\mu_1}{\operatorname{d} \mu_3} \\
\ainnerproduct{\operatorname{d}\mu_2}{\operatorname{d} \mu_2} &
\ainnerproduct{\operatorname{d}\mu_2}{\operatorname{d} \mu_3}
\end{bmatrix}\, ,
\]
which in the notation of Prop.~\ref{prop:whtny_ip} is
\[
\det
\begin{bmatrix}
\widehat{\ainnerproduct{\operatorname{d}\mu_0}{\operatorname{d}\mu_1}} &
\widehat{\ainnerproduct{\operatorname{d}\mu_0}{\operatorname{d}\mu_2}} &
\widehat{\ainnerproduct{\operatorname{d}\mu_0}{\operatorname{d}\mu_3}} \\
\widehat{\ainnerproduct{\operatorname{d}\mu_1}{\operatorname{d}\mu_1}} &
\ainnerproduct{\operatorname{d}\mu_1}{\operatorname{d} \mu_2} &
\ainnerproduct{\operatorname{d}\mu_1}{\operatorname{d} \mu_3} \\
\widehat{\ainnerproduct{\operatorname{d}\mu_2}{\operatorname{d}\mu_1}} &
\ainnerproduct{\operatorname{d}\mu_2}{\operatorname{d} \mu_2} &
\ainnerproduct{\operatorname{d}\mu_2}{\operatorname{d} \mu_3}
\end{bmatrix}\, .
\]
Using a shorthand notation for matrices like above, the $2\times 2$
matrices whose determinants need to be computed for calculating the
$(0,3)$ entry of $M_2$ are given below.
\[
\begin{array}{ccc}
\begin{bmatrix}
\widehat{01} & \widehat{02} & \widehat{03}\\
\widehat{11} & 12 & 13\\
\widehat{21} & 22 & 23
\end{bmatrix} &
\begin{bmatrix}
\widehat{02} & \widehat{01} & \widehat{03}\\
\widehat{12} & 11 & 13\\
\widehat{22} & 21 & 23
\end{bmatrix} &
\begin{bmatrix}
\widehat{03} & \widehat{01} & \widehat{02}\\
\widehat{13} & 11 & 12\\
\widehat{23} & 21 & 22
\end{bmatrix}\\
& & \\
\begin{bmatrix}
\widehat{11} & \widehat{12} & \widehat{13}\\
\widehat{01} & 02 & 03\\
\widehat{21} & 22 & 23
\end{bmatrix} &
\begin{bmatrix}
\widehat{12} & \widehat{11} & \widehat{13}\\
\widehat{02} & 01 & 03\\
\widehat{22} & 21 & 23
\end{bmatrix} &
\begin{bmatrix}
\widehat{13} & \widehat{11} & \widehat{12}\\
\widehat{03} & 01 & 02\\
\widehat{23} & 21 & 22
\end{bmatrix}\\
& & \\
\begin{bmatrix}
\widehat{21} & \widehat{22} & \widehat{23}\\
\widehat{01} & 02 & 03\\
\widehat{11} & 12 & 13
\end{bmatrix} &
\begin{bmatrix}
\widehat{22} & \widehat{21} & \widehat{23}\\
\widehat{02} & 01 & 03\\
\widehat{12} & 11 & 13
\end{bmatrix} &
\begin{bmatrix}
\widehat{23} & \widehat{21} & \widehat{22}\\
\widehat{03} & 01 & 02\\
\widehat{13} & 11 & 12
\end{bmatrix}
\end{array}
\]
Removing the deleted rows and columns the above matrices are given
below as the actual $2\times 2$ matrices.
\begin{equation} \label{expr:ip03}
\begin{array}{ccc}
\begin{bmatrix}
12 & 13\\
22 & 23
\end{bmatrix} &
\begin{bmatrix}
11 & 13\\
21 & 23
\end{bmatrix} &
\begin{bmatrix}
11 & 12\\
21 & 22
\end{bmatrix}\\
& & \\
\begin{bmatrix}
02 & 03\\
22 & 23
\end{bmatrix} &
\begin{bmatrix}
01 & 03\\
21 & 23
\end{bmatrix} &
\begin{bmatrix}
01 & 02\\
21 & 22
\end{bmatrix}\\
& & \\
\begin{bmatrix}
02 & 03\\
12 & 13
\end{bmatrix} &
\begin{bmatrix}
01 & 03\\
11 & 13
\end{bmatrix} &
\begin{bmatrix}
01 & 02\\
11 & 12
\end{bmatrix}
\end{array}
\end{equation}
The $2\times 2$ matrices whose determinants are needed in computing
$\pInnerproduct{\cochainBasis 2 0}{\cochainBasis 2 1}$, i.e., entry
$(0,1)$ of $M_2$ are given below.
\begin{equation}\label{expr:ip01}
\begin{array}{ccc}
\begin{bmatrix}
11 & 13\\
21 & 23
\end{bmatrix} &
\begin{bmatrix}
10 & 13\\
20 & 23
\end{bmatrix} &
\begin{bmatrix}
10 & 11\\
20 & 21
\end{bmatrix}\\
& & \\
\begin{bmatrix}
01 & 03\\
21 & 23
\end{bmatrix} &
\begin{bmatrix}
00 & 03\\
20 & 23
\end{bmatrix} &
\begin{bmatrix}
00 & 01\\
20 & 21
\end{bmatrix}\\
& & \\
\begin{bmatrix}
01 & 03\\
11 & 13
\end{bmatrix} &
\begin{bmatrix}
00 & 03\\
10 & 13
\end{bmatrix} &
\begin{bmatrix}
00 & 01\\
10 & 11
\end{bmatrix}
\end{array}
\end{equation}
\end{example}
Recall that each number in these matrices is shorthand for an inner
product of two barycentric differentials. For example, the entry 12
stands for $\aInnerproduct{\operatorname{d} \mu_1}{\operatorname{d} \mu_2} = g((\operatorname{d}
\mu_1)^\sharp,(\operatorname{d} \mu_2)^\sharp) = \nabla\mu_1 \cdot \nabla\mu_2$.
\begin{proposition} \label{prop:p_equal_n}
For $p=n$, $M_p$ is a diagonal matrix with $M_p(i,i) =
1/\abs{\sigma_i^n}$, where $\abs{\sigma_i^n}$ is the volume of the
simplex.
\end{proposition}
\begin{proof}
For any $n$-simplex $\sigma^n$, the Whitney form $\whitney
(\sigma^n)^\ast$ is 0 on other $n$-simplices and so $M_p$ is
diagonal. Furthermore, it is a constant coefficient volume form on
$\sigma^n$ with $\int_{\sigma^n} \whitney (\sigma^n)^\ast =
1$. See~\cite{Dodziuk1976} for proofs of these properties. Thus it
must be that $\whitney (\sigma^n)^\ast = \mu / \abs{\sigma^n}$ where
$\mu$ is the volume form on the simplex. Thus
\[
\aInnerproduct{\whitney (\sigma^n)^\ast}{\whitney (\sigma^n)^\ast}
\, \mu = \whitney (\sigma^n)^\ast \wedge \hodge \whitney
(\sigma^n)^\ast = \frac{\mu}{\abs{\sigma^n}^2} \, .
\]
Thus $\pInnerproduct{\whitney (\sigma^n)^\ast}{\whitney
(\sigma^n)^\ast}_{L^2}$ is $\int_{\sigma^n} \mu /
\abs{\sigma^n}^2$ which is $1/\abs{\sigma^n}$.
\end{proof}
\subsection{Algorithm for Whitney inner product matrix}
\label{subsec:mssmtrx_algrthm}
We motivate our algorithm for Whitney mass matrix computation by
making some observations about Example~\ref{ex:ttrhdrn_whtny}. The
first, and obvious observation is that matrix $M_p$ is symmetric,
being an inner product matrix. Thus only the diagonal entries and
those above (or below) the diagonal need be computed. A more
interesting efficiency comes from the structure of the entries of the
matrix collections, such as ones shown in~\eqref{expr:ip03}
and~\eqref{expr:ip01}. Note that many entries repeat in the shorthand
collection of matrices in~\eqref{expr:ip03} and~\eqref{expr:ip01}. For
example the entry 12 appears 4 times by itself in the matrix
collection~\eqref{expr:ip03}. Moreover, due to the symmetry of inner
product, the entry 12 corresponds to the same result as the entry 21
and 21 appears 4 times as well. That entry also appears 4 times in the
collection~\eqref{expr:ip01}. Thus it is clear that a saving in
computational time can be achieved by doing such calculations only
once. That is, $\ainnerproduct{\operatorname{d} \mu_1}{\operatorname{d} \mu_2} = \ainnerproduct{\operatorname{d}
\mu_2}{\operatorname{d} \mu_1}$ need only be computed once for the tetrahedron.
The determinants of all the matrices in a collection such
as~\eqref{expr:ip03} are needed to plug into an expression
like~\eqref{expr:ip03full} to obtain a single entry (in this case row
0, column 3) of the Whitney inner product matrix for $p$-cochains
($p=2$ in this case), whose size ($4\times 4$ in this case) depends on
the number of $p$-simplices in the simplicial complex. Thus reusing
repeated inner products of barycentric differentials can add up to a
substantial saving in computational expense when all the unique
entries of $M_p$ are computed. These savings are quantified later in
this subsection.
Another useful point to note in the example calculation is that the
collection~\eqref{expr:ip01} of matrices can be obtained from the
collection~\eqref{expr:ip03} by keeping the first digit in each entry
same and making the substitutions $1\rightarrow 0$; $2\rightarrow 1$;
and $3\rightarrow 3$ in the second digit. The first digits in the two
collections are the same because both correspond to the triangle
$\sigma^2_0$. The substitution above works for the second digit
because $\sigma^2_3 = [v_1, v_2, v_3]$ and $\sigma^2_1 = [v_0, v_1,
v_3]$. This suggests the use of a template simplex for creating a
template collection of matrices whose determinants are needed. The
actual instances of the collections can then be obtained by using the
vertex numbers in a given simplex. This is another idea that is used
in the algorithm implemented in PyDEC. The algorithm takes as input
a manifold simplicial $n$-complex $K$, embedded in $\mathbb{R}^N$ and $0
\le p \le n$. The output is $M_p$, an $N_p \times N_p$ matrix
representation of inner product on $C^p(K;\mathbb{R})$ using elementary
cochain basis. If a naive algorithm, which does not take into
account the duplications in determinant calculations were to be used,
the number of operations required in the mass matrix calculation are
\[
N_n \times \dfrac{\dbinom{n+1}{p+1}^2+\dbinom{n+1}{p+1}}{2}
\times \binom{n}{p}^2 \times Np^2 \times (O(p!) \text{ or } O(p^3)) \, .
\]
The last term is written as $O(p!)$ or $O(p^3)$ because a determinant
can be computed using the formula for determinant or by LU
factorization. For low values of $p$ (i.e. $\le$ about 5) the formula
will likely be better.
According to the above formula, for example, for $n=3, p=2$, the
number of determinants required in a naive implementation of mass
matrix calculation would be
\[
\dfrac{\dbinom{4}{3}^2+\dbinom{4}{3}}{2}
\times \binom{3}{2}^2 = 10 \times 9 = 90\, .
\]
But there are only 21 unique determinants needed for $n=3, p=2$. Our
algorithm computes the unique
determinants first and the operation count is
\[
N_n \times \dfrac{\dbinom{n+1}{p}^2+\dbinom{n+1}{p}}{2}
\times Np^2 \times (O(p!) \text{ or } O(p^3)) \, .
\]
\begin{figure}[p]
\centering
\begin{tabular}[h]{ccc}
\includegraphics[scale=0.3]{pydec/whitney_ip/numdets_n3} &
\includegraphics[scale=0.3]{pydec/whitney_ip/numdets_n5}\\
\includegraphics[scale=0.3]{pydec/whitney_ip/numdets_n7} &
\includegraphics[scale=0.3]{pydec/whitney_ip/numdets_n9}\\
\includegraphics[scale=0.3]{pydec/whitney_ip/numdets_n11} &
\includegraphics[scale=0.3]{pydec/whitney_ip/numdets_n13}
\end{tabular}
\caption{Comparison between PyDEC and a naive algorithm for
computing Whitney mass matrix. The figure shows the number of
determinant computations needed by the two algorithms, for various
values of $p$, the Whitney form dimension, and $n$, the simplicial
complex dimension. The embedding dimension is not relevant in
these calculations. For $p=n$ case we use the shortcut described
in Proposition~\ref{prop:p_equal_n} so that case is not shown.}
\label{fig:dtrmnnt_cnts}
\end{figure}
Figure~\ref{fig:dtrmnnt_cnts} shows a comparison of determinant
counts for our algorithm compared with a naive algorithm that does the
duplicate work that our PyDEC algorithm avoids. Note that for any $n$,
the most advantage is gained intermediate values of $p$. The savings
that the PyDEC implementation provides over a naive algorithm are
several orders of magnitude, especially for moderately large $n$ and
higher. For $p=n$ case, in PyDEC we use the shortcut described in
Proposition~\ref{prop:p_equal_n}.
\section{Metric Dependent Operators}
\label{sec:metric}
We now describe the PyDEC implementations of some metric dependent
exterior calculus operators. The simplicial complex $K$ is now
supposed to be an approximation of a Riemannian $n$-manifold $M$. The
metric implemented in PyDEC is the one induced from an embedding space
$\mathbb{R}^N$. The main metric dependent operator is the Hodge star which
enables the discretization of codifferential and Laplace-deRham
operators. The sharp and flat, which are isomorphisms between 1-forms
and vector fields, are not implemented.
For the DEC Hodge star, the implementation is using the circumcentric
dual as in~\cite{Hirani2003, DeHiLeMa2005} and the other operators are
then simply defined in terms of the exterior derivative and the Hodge
star. For PyDEC's implementation of low order finite element exterior
calculus, we define the Hodge star to be the Whitney mass matrix
described in Section~\ref{sec:whtny_innrprdct}. The other operators
are defined by analogy with DEC even though the dual mesh concept is
not part of finite element exterior calculus. Extensive experimental
justification for this approach can be seen in its effectiveness in
numerical experiments in~\cite{Bell2008} and in~\cite{HiKaWaWa2011}.
For the definitions in this section we will need two cochain complexes
of real-valued cochains. One will be on the simplicial complex $K$ and
for brevity we'll call this space of $p$-cochains $C^p(K)$ instead of
$C^p(K;\mathbb{R})$. The other cochain complex is on the circumcentric dual
cell complex $\dual K$ and we'll denote the $(n-p)$-dimensional
cochains as $D^{n-p}(\dual K)$. At each dimension, these will be
connected by discrete Hodge star operators to be defined below. Since
the exterior derivative is the coboundary operator, the matrix
representation for the exterior derivative on the dual mesh is the
boundary operator. The matrix form for the DEC Hodge star
on $p$-cochains will be denoted $\hodge_p: C^p(K) \to D^{n-p}(\dual
K)$. One box of the primal and dual complexes is shown below.
\begin{equation*}
\begin{CD}
C^p(K) @>\dd_p>> C^{p+1}(K) \\
@VV\hodge_p V @VV\hodge_{p+1} V \\
D^{n-p}(\dual K) @<\dd_p^T< < D^{n-p-1}(\dual K)
\end{CD}
\end{equation*}
As described in~\cite{Hirani2003, DeHiLeMa2005} and other references,
the DEC Hodge star is defined by
\[
\frac{\eval{\hodge_p \sigma^\ast_i}{\dual \sigma_j}}{\abs{\dual\sigma_j}} =
\frac{\eval{\sigma^\ast_i}{\sigma_j}}{\abs{\sigma_j}}\, ,
\]
for $p$-simplices $\sigma_i$ and $\sigma_j$. Here $\sigma^\ast$ is the
elementary cochain corresponding to $\sigma$ and
$\eval{\sigma^\ast}{\tau}$ stands for the evaluation of the cochain
$\sigma^\ast$ on the elementary chain $\tau$. Thus the matrix
representation of the DEC Hodge star $\hodge_p$ is as a diagonal
matrix with $\hodge_p(i,i) =
\abs{\dual\sigma_i}/\abs{\sigma_i}$. In~\cite{Hirani2003,
DeHiLeMa2005} this was defined for well-centered meshes. For the
codimension 1 Hodge star the definition extends to Delaunay meshes
with a slight additional condition for boundary simplices. This
extension involves computing the volume of $\dual\sigma$ taking into
account signs. Consider a codimension 1 simplex $\sigma$ shared by
simplices $L$ and $R$. For the portion of $\dual\sigma$ corresponding
to $L$, the sign is positive if the circumcenter and remaining vertex
of $L$ are on the same side of $\sigma$. Similarly for $R$. (For
surface meshes and higher dimensional analogs the circumcenter
condition above is one way to define a Delaunay-like condition.) If
$\sigma$ is a codimension 1 face of top dimensional $\tau$ and is on
domain boundary then the circumcenter of $\tau$ and vertex opposite to
$\sigma$ should be on the same side. The smooth Hodge star on
$p$-forms satisfies $\hodge\hodge = (-1)^{p(n-p)}$. In the discrete
setting $\hodge\hodge$ is written as $\hodge_p^{-1}\hodge_p$ or
$\hodge_p\hodge_p^{-1}$ and this is defined to be $(-1)^{p(n-p)}\, I$
where $I$ is the identity matrix.
In the smooth theory, the codifferential $\codiff_{p+1} :
\Omega^{p+1}(M) \to \Omega^p(M)$ is defined as $\codiff_{p+1} =
(-1)^{np+1} \hodge \operatorname{d} \hodge$ and so we define the discrete
codifferential $\dcodiff_{p+1} : C^p(K) \to C^{n-p}(K)$ as
$\dcodiff_{p+1} := (-1)^{np+1} \hodge_p^{-1} \dd_p^T \hodge_{p+1}$.
For finite element exterior calculus implemented in PyDEC, we take
this to be the definition, without reference to a dual mesh. If we now
take $\hodge_p$ to be the Whitney mass matrix then $\dd_p$ and
$\dcodiff_{p+1}$ are adjoints (up to sign) with respect to the Whitney
inner product on cochains as shown in~\cite{HiKaWaWa2011}. We will
call the use of Whitney mass matrix as $\hodge_p$ to be a
\emph{Whitney Hodge star} matrix.
In the discrete setting the Laplace-deRham operator is implemented in
the weak form. For $0<p<n$ the discrete definition is $\laplacian_p :=
\operatorname{d}_p\hodge_{p+1}\operatorname{d}_p + (-1)^{(p-1)(n-p+1)}
\hodge_p\operatorname{d}_{p-1}\hodge^{-1}_{p-1}\operatorname{d}^T_{p-1}\hodge_p$, with the
appropriate term dropped for the $p=0$ and $p=n$ cases. The above
expression involves inverses of the Hodge star, which is easy to
compute for DEC Hodge star since that is a diagonal matrix. For a
Whitney Hodge star see~\cite{Bell2008, HiKaWaWa2011} for various
approaches to avoiding explicitly forming the inverse Whitney mass
matrix in computations.
\subsection{Circumcenter Calculation}
\label{subsec:circumcenter}
Circumcentric duality is used in DEC. To compute the DEC Hodge star, a
basic computational step is the computation of the circumcenter of a
simplex. We give here a linear system for computing the circumcenter
using barycentric coordinates.
The circumcenter of a simplex is the unique point that is equidistant
from all vertices of that simplex. In the case that a simplex (or
face) is not of the same dimension as the embedding (e.g. a triangle
embedded in $\mathbb{R}^4$), we choose the point that lies in the affine space
spanned by the vertices of the simplex. In either case we can write
the circumcenter in terms of barycentric coordinates of the simplex.
Let $\sigma^p$ be the $p$-simplex defined by the points $\{ v_0, v_1,
\ldots v_p\}$ in $\mathbb{R}^N$. Let $R$ denote the circumradius and $c$ the
circumcenter of simplex, which can be written in barycentric
coordinates as $c = \sum_j b_j v_j$ where $b_j$ is the barycentric
coordinate for the circumcenter corresponding to $v_j$. For each
vertex $i$ we have
\[
\left\Vert v_i - \sum_{j=0}^p b_j v_j \right\Vert^2 - R^2 = 0 \, ,
\]
which can be rewritten as
\[
v_i \cdot v_i - 2 v_i \cdot \left(\sum_{j=0}^p b_j v_j\right) +
\left\Vert \sum_{j=0}^p b_j v_j\right\Vert^2 - R^2 = 0\, .
\]
Here the norm and the dot product are the standard ones on $\mathbb{R}^N$.
Rearranging the above yields
\[
2 v_i \cdot \left(\sum_{j=0}^p b_j v_j\right) -
\left(\left\Vert\sum_{j=0}^p b_j v_j\right\Vert^2 - R^2\right) =
v_i \cdot v_i\, .
\]
The second term on the left hand side is some scalar which is unknown,
but is the same for every equation. So we can replace it by the
unknown $Q$ and write
\[
2 v_i \cdot \left(\sum_{j=0}^p b_j v_j\right) + Q = v_i \cdot v_i \, .
\]
With the additional constraint that barycentric coordinates sum to
one, we have a linear system with $p + 2$ unknowns ($b_0 \ldots b_p$
and $Q$) and $p + 2$ equations with the following matrix form
\[
\begin{pmatrix}
2 v_0 \cdot v_0 & 2 v_0 \cdot v_1 & \dots & 2 v_0 \cdot v_p &1\\
2 v_1 \cdot v_0 & 2 v_1 \cdot v_1 & \dots & 2 v_1 \cdot v_p &1\\
\vdots & \vdots & \ddots & \vdots & \vdots \\
2 v_p \cdot v_0 & 2 v_p \cdot v_1 & \dots & 2 v_p \cdot v_p &1\\
1 & 1 & \dots & 1 & 0\\
\end{pmatrix}
\begin{pmatrix}
b_0 \\
b_1 \\
\vdots \\
b_p \\
Q \\
\end{pmatrix}
=
\begin{pmatrix}
v_0 \cdot v_0 \\
v_1 \cdot v_1 \\
\vdots \\
v_p \cdot v_p \\
1 \\
\end{pmatrix}
\]
The solution to this yields the barycentric coordinates from which the
circumcenter $c$ can be located. Another quantity required for DEC
Hodge star is the unsigned volume of a simplex. This can be computed
by the well-known formula $\sqrt{\det V^T V}/p!$ where $V$ is the $p$
by $N$ matrix with rows formed by the vectors $\{ v_1 - v_0, v_2 -
v_0, \ldots v_p - v_0\}$.
\section{Examples} \label{sec:examples}
In the domains for which PyDEC is intended, it is often possible to
easily translate the mathematical formulation of a problem into a
working program. To make this point, and to demonstrate a variety of
applications of PyDEC, we give 5 examples from different fields. The
first example (Section~\ref{subsec:cavity}) is a resonant cavity
eigenvalue problem in which Whitney forms work nicely while the nodal
piecewise linear Lagrange vector finite element $\mathcal{P}_1^2$
fails when directly applied. The second is Darcy flow
(Section~\ref{subsec:darcy}), which is an idealization of the steady
flow of a fluid in a porous medium. We solve it here using DEC. The
third problem (Section~\ref{subsec:cohomology}) is computation of a
basis for the cohomology group of a mesh with several holes. This is
achieved in our code here by Hodge decomposition of cochains, again
using DEC. Next example (Section~\ref{subsec:sensor}) is an
idealization of the sensor network coverage problem. Some randomly
located idealized sensors in the plane are connected into a Rips
complex based on their mutual distances. Then a harmonic cochain
computation reveals the possibility of holes in coverage. The last
example (Section~\ref{subsec:ranking}) involves the ranking of
alternatives by a least squares computation on a graph.
None of these problem is original and they all have been treated in
the literature by a variety of techniques. We emphasize that we are
including these just to demonstrate the capabilities of PyDEC. We have
included the relevant parts of the Python code in this paper. The full
working programs are available with the PyDEC
package~\cite{BeHi2008a}.
\subsection{Resonant cavity curl-curl problem}
\label{subsec:cavity}
An electromagnetic resonant cavity is an idealized box made of a
perfect conductor and containing no enclosed charges in which
Maxwell's equations reduce to an eigenvalue problem. Several authors
have popularized this example as one of many striking examples that
motivate finite element exterior calculus. See for
instance~\cite{ArFaWi2010}. The use of $\mathcal{P}_1^2$ finite
element space, i.e. piecewise linear, Lagrange finite elements with 2
components, yields a corrupted spectrum. On the other hand, the use
of $\mathcal{P}_1^-$ elements, i.e., Whitney 1-forms yields the
qualitatively correct spectrum. For detailed analysis and background
see~\citep{BoFeGaPe1999,ArFaWi2010}.
Let $M \subset \mathbb{R}^2$ be a square domain with side length $\pi$. We
first give the equation in vector calculus notation and then in the
corresponding exterior calculus notation. In the former, the resonant
cavity problem is to find vector fields $E$ and eigenvalues $\lambda
\in \mathbb{R}$ such that
\[
\vcurl\, \curl E = \lambda E \quad \text{ on } M \quad \text{ and }
\quad E_\parallel = 0
\text{ on } \partial M \, ,
\]
where $E_\parallel$ is the tangential component of $E$ on the
boundary. Here $\vcurl \phi = (\partial \phi / \partial y, -\partial
\phi / \partial x)$ and $\curl v = \partial v_2 / \partial x
- \partial v_1 / \partial y$ for scalar function $\phi$ and vector
field $v = (v_1, v_2)$. Note that for $\lambda \ne 0$ this equation is
equivalent to the pair of equations $\Delta E = \lambda E$ and $\operatorname{div} E
= 0$. This is because the vector Laplacian $\Delta = \vcurl \circ
\curl - \grad \circ \operatorname{div}$ and $\operatorname{div} \circ \curl = 0$.
Now we give the equation in exterior calculus notation so the
transition to PyDEC will be easier. Let $u \in \Omega^1(M)$ be the unknown
electric field 1-form and $i : \partial M \hookrightarrow M$ the
inclusion map. Then the above vector calculus equation is equivalent
to
\begin{alignat*}{2}
\delta_2 \operatorname{d}_1 u & = \lambda\, u &&\quad \text{in } M\\
i^\ast u &= 0 && \quad \text{on } \boundary M\, .
\end{alignat*}
The pullback $i^\ast u$ by inclusion map means restriction of $u$ to
the boundary, i.e., allowing only vectors tangential to the boundary
as arguments to $u$. As usual, we will seek $u$ not in $\Omega^1(M)$
but in $H\Omega^1(M)$ subject to boundary conditions. Define the
vector space $V = \{v\,\vert\, v \in H\Omega^1(M), i^\ast v = 0 \text{
on } \partial M\}$.
To express the PDE in weak form, we seek a $(u,\lambda)$ in $V\times
\mathbb{R}$ such that $\pinnerproduct{\codiff_2 \operatorname{d}_1 u}{v}_{L^2} = \lambda
\pinnerproduct{u}{v}_{L^2}$ for all $v \in V$. By the properties of
the codifferential, the expression on the left is equal to
$\pinnerproduct{\operatorname{d}_1 u}{\operatorname{d}_1 v}_{L^2} - \int_{\partial M} u \wedge
\hodge \operatorname{d}_1 v$. But the boundary term is 0 because $u$ is in $V$. Thus
the weak form is to find a $(u,\lambda) \in V\times \mathbb{R}$ such that
$\pinnerproduct{\operatorname{d}_1 u}{\operatorname{d}_1 v}_{L^2} = \lambda
\pinnerproduct{u}{v}_{L^2}$ for all $v \in V$.
Taking the Galerkin approach of looking for a solution in a finite
dimensional subspace of $V$ here we pick the space of Whitney 1-forms,
that is, $\mathcal{P}_1^-\Omega^1$ as the finite dimensional
subspace. We define these over a triangulation of $M$ which we will
call $K$. The Whitney map $\whitney : C^1(K;\mathbb{R}) \to
L^2\Omega^1(\abs{K})$ is an injection with its image
$\mathcal{P}_1^-(K)$. Thus an equivalent formulation is over
cochains. Using the same names for the variables, we seek a
$(u,\lambda) \in C^1(K;\mathbb{R}) \times \mathbb{R}$ such that $\pinnerproduct{\operatorname{d}_1
\whitney u}{\operatorname{d}_1 \whitney v}_{L^2} = \lambda \pinnerproduct{\whitney
u}{\whitney v}_{L^2}$ for all 1-cochains $v\in C^1(K;\mathbb{R})$. Since the
Whitney map commutes with the exterior derivative and coboundary
operator, and using the definition of cochain inner product, the above
is same as $\pinnerproduct{\operatorname{d}_1 u}{\operatorname{d}_1 v} = \lambda
\pinnerproduct{u}{v}$ where now the inner product is over cochains and
$\operatorname{d}_1$ is the coboundary operator. In matrix notation, using
$\hodge_1$ and $\hodge_2$ to stand for the Whitney mass matrices $M_1$
and $M_2$, the generalized eigenvalue problem is to find $(u,\lambda)
\in C^1(K;\mathbb{R}) \times \mathbb{R}$ such that
\[
\operatorname{d}_1^T \, \hodge_2 \, \operatorname{d}_1 u = \lambda \, \hodge_1 u \, .
\]
We now translate this equation into PyDEC code. Once the appropriate
modules have been imported, a simplicial complex object \texttt{sc} is
created after reading in the mesh files. Now the main task is to find
matrix representations for the stiffness matrix $\operatorname{d}_1^T \, \hodge_2 \,
\operatorname{d}_1$and the mass matrix $\hodge_1$. This is accomplished by the
following two lines, where \texttt{K} is the stiffness matrix :
\lstinputlisting[firstline=25,lastline=26]{code/cavity/driver.py} The
boundary conditions can be imposed by simply removing the edges that
lie on the boundary. The indices of such edges is easily determined
and stored in the list \texttt{non\_boundary\_indices} which is used
below to impose the boundary conditions :
\lstinputlisting[firstline=34,lastline=35]{code/cavity/driver.py} Now
all that remains is to solve the eigenvalue problem. To simplify the
code and because the matrix size is small, we use the dense eigenvalue
solver \texttt{scipy.linalg.eig}
\lstinputlisting[firstline=39,lastline=39]{code/cavity/driver.py} Some
of the resulting eigenvalues are displayed in the left part of
Figure~\ref{fig:cavity}. The 1-cochain $u$ which is the eigenvector
corresponding to one of these eigenvalues is shown as a vector field
in the right part of Figure~\ref{fig:cavity}. The visualization as a
vector field is achieved by interpolating the 1-cochain $u$ using the
Whitney map and then sampling the vector field $(\whitney u)^\sharp$
at the barycenter. This is achieved by the PyDEC command:
\lstinputlisting[firstline=54,lastline=54]{code/cavity/driver.py}
where \texttt{all\_values} contains both the known and the computed
values of the 1-cochain. There is no sharp operator in PyDEC. But
since PyDEC only implements the Riemannian metric from the embedding
space of simplices the transformation from 1-form to vector field just
involves using the components of the Whitney 1-form as the vector
field components.
\begin{figure}[t]
\centering
\includegraphics[height=2in,trim=0.25in 0.4in 0.25in 0.5in, clip]
{pydec/cavity/eigenvalues50}
\includegraphics[height=2.2in,trim=1.1in 0.8in 1in 0.75in, clip]
{pydec/cavity/eigenvector3}
\caption{The first 50 nonzero eigenvalues for the resonant cavity
problem of Section~\ref{subsec:cavity} and the eigenvector
corresponding to one of these eigenvalues. The eigenvector is a
1-cochain which is visualized as a vector field by first
interpolating it using a Whitney map. See
Section~\ref{subsec:cavity} for details.}
\label{fig:cavity}
\end{figure}
\subsection{Darcy flow or Poisson's in mixed form}
\label{subsec:darcy}
We give here a brief description of the equations of Darcy flow and
their PyDEC implementation. For more details
see~\cite{HiNaCh2011}. The resonant cavity example in
Section~\ref{subsec:cavity} was implemented using finite element
exterior calculus. For variety we use a DEC implementation for Darcy
flow.
Darcy flow is a simple model of steady state flow of an incompressible
fluid in a porous medium. It models the statement that flow is from
high to low pressure. For a fixed pressure gradient, the velocity is
proportional to the permeability $\kappa$ of the medium and inversely
proportional to the viscosity $\mu$ of the fluid. Let the domain be
$M$, a polygonal planar domain. Assuming that there are no sources of
fluid in $M$ and there is no other force acting on the fluid, the
equations of Darcy flow are
\begin{equation}\label{eq:vfdrcy}
v + \frac{\kappa}{\mu} \nabla p = 0 \quad\text{and}
\quad \operatorname{div} v = 0 \quad \text{in } M \quad \text{with}\quad
v \cdot \hat{n} = \psi \quad \text{on } \partial M.
\end{equation}
where $\kappa > 0$ is the coefficient of permeability of the medium
$\mu > 0$ is the coefficient of (dynamic) viscosity of the fluid,
$\psi:\partial M \rightarrow \mathbb{R}$ is the prescribed normal
component of the velocity across the boundary, and $\hat{n}$ is the
unit outward normal vector to $\partial M$. For consistency
$\int_{\partial M} \psi \, d\Gamma = 0$, where $d\Gamma$ is the
measure on $\partial M$. Since $\operatorname{div} \circ \grad = \Delta$, the
simplified Darcy flow equations above are equivalent to Laplace's
equation.
Let $K$ be a simplicial complex that triangulates $M$. Instead of
velocity and pressure, we will use flux and pressure as the primary
unknowns. The flux through the edges is $f=\hodge v^\flat$ and thus it
will be a primal 1-cochain. Although PyDEC does not implement a flat
operator, this is not an issue here because we never solve for $v$,
and make $f$ itself one of the unknowns. This implies that the
pressure $p$ will be a dual 0-cochain since $\hodge \operatorname{d} p$ has to be of
the same type as $f$. The choice to put flux on primal edges and
pressures on circumcenters can be reversed, as shown in a dual
formulation in~\cite{GiBa2010a}. In exterior calculus notation, the
PDE in~\eqref{eq:vfdrcy} is
$-(\mu/k)\hodge f + \operatorname{d} p = 0$ and $\operatorname{d} f = 0$, which, when discretized,
translates to the matrix equation
\[
\begin{bmatrix}
-(\mu/k)\hodge_1 & \operatorname{d}_1^T\\
\operatorname{d}_1 & 0
\end{bmatrix}
\begin{bmatrix}
f \\ p
\end{bmatrix} =
\begin{bmatrix}
0\\0
\end{bmatrix}\, .
\]
In PyDEC, the construction of this matrix is straightforward. Once a
simplicial complex \texttt{sc} has been constructed, the following 3
lines construct the matrix in the system above :
\lstinputlisting[firstline=40,lastline=42]{code/darcy/driver.py}
After computing the boundary condition in terms of flux through the
boundary edges, the linear system is adjusted for the known values and
then solved for the fluxes and pressures. Figure~\ref{fig:darcy} shows
the solution for the case of constant horizontal velocity and linear
pressure gradient.
\begin{figure}[t]
\centering
\includegraphics[scale=0.4,trim=0in 0in 0in 0.5in, clip]{pydec/darcy/flux}
\includegraphics[scale=0.4]{pydec/darcy/pressure}
\caption{Darcy flow using discrete exterior calculus. The boundary
condition is that fluid is coming in from left and leaving from
right with velocity 1. The velocity inside should be constant and
pressure should be linear. The flux and pressure are computed in a
mixed formulation. The flux is taken to be a primal 1-cochain
associated with primal edges, and the pressure is a dual 0-cochain
on dual vertices, which are circumcenters of the triangles. The
velocity is obtained by Whitney interpolation of the flux, which
is sampled at the barycenters. See Section~\ref{subsec:darcy} for
more details.}
\label{fig:darcy}
\end{figure}
\subsection{Cohomology basis using Hodge decomposition}
\label{subsec:cohomology}
The Hodge Decomposition Theorem~\cite[page 539]{AbMaRa1988} states
that for a compact boundaryless smooth manifold $M$, for any $p$-form
$\omega \in \Omega^p(M)$, there exists an $\alpha\in\Omega^{p-1}(M)$,
$\beta \in \Omega^{p+1}(M)$, and a harmonic form $h \in \Omega^p(M)$
such that $\omega = d \alpha + \delta \beta + h$. Here harmonic means
that $\Delta h = 0$, where $\Delta$ is the Laplace-deRham operator
$\operatorname{d}\delta +\delta \operatorname{d}$. Moreover $\operatorname{d}\alpha$, $\delta \beta$ and $h$ are
mutually $L^2$-orthogonal, which makes them uniquely determined. In
case of a manifold with boundary, the decomposition is similar, with
some additional boundary conditions. See~\cite{AbMaRa1988} for
details.
The Hodge-deRham theorem~\cite{AbMaRa1988}, relates the analytical
concept of harmonic forms with the topological concept of
cohomology. For any topological space, the cohomology groups or vector
spaces of various dimension capture essential topological information
about the space~\cite{Munkres1984}. For the manifold $M$ above, the
$p$-dimensional cohomology group with real coefficients, which is a
finite-dimensional space, is denoted $H^p(M;\mathbb{R})$ or just $H^p(M)$. For
example, for a torus, $H^1$ has dimension 2. For a square with 4 holes
used in this example, which does have boundaries, $H^1$ has dimension
4. The elements of $H^p(M)$ are equivalence classes of closed forms
(those whose $\operatorname{d}$ is 0). Two closed forms are equivalent if their
difference is exact (that is, is $\operatorname{d}$ of some form). While the
representatives of 1-homology spaces can be visualized as loops around
holes, handles, and tunnels, those of 1-cohomology should be
visualized as fields. If the space of harmonic forms is denoted
$\mathcal{H}^p(M)$, then the Hodge-deRham theorem says that it is
isomorphic, as a vector space, to the $p$-th cohomology space $H^p(M)$
in the case of a closed manifold. See~\cite{Jost2005} for
details. Again, the case of $M$ with boundary requires some
adjustments in the definitions, as given in~\cite{AbMaRa1988}.
For finite dimensional spaces, Hodge decomposition follows from very
elementary linear algebra. If $U$, $V$ and $W$ are finite-dimensional
inner product vector spaces and $A : U \to V$ and $B: V \to W$ are
linear maps such that $B \circ A = 0$ then middle vector space $V$
splits into 3 orthogonal components, which are $\im A$, $\im B^T$, and
$\ker A^T \cap \ker B$. In this example, we find a basis for $H^1$ for
a square. This is done by finding a basis of harmonic 1-cochains.
Thus given a 1-cochain $\omega$, its discrete Hodge decomposition
exists and is $\omega = \dd_0 \alpha + \dcodiff_2 \beta + h$. In this
example, the cochains $\alpha$ and $\beta$ are obtained by solving the
linear systems
$\dcodiff_1 \dd_0 \alpha = \dcodiff_1 \omega$ and $\dd_1 \dcodiff_2
\beta = \dd_1 \omega$. The harmonic component can then be computed
by subtraction.
In the example code, the main function is the one that computes the
Hodge decomposition of a given cochain \texttt{omega}. First empty
cochains for \texttt{alpha} and \texttt{beta} are created:
\lstinputlisting[firstline=30,lastline=33]{code/hodge/driver.py} Now
the solution for \texttt{alpha} and \texttt{beta} closely follows the
above equations for $\alpha$ and $\beta$:
\lstinputlisting[firstline=36,lastline=38]{code/hodge/driver.py}
\lstinputlisting[firstline=41,lastline=43]{code/hodge/driver.py} Even
though the matrices \texttt{A} above are singular, the solutions
exist, and since conjugate gradient is used, the presence of the
nontrivial kernels does not pose any problems~\cite{BoLe2005}.
\begin{figure}[t]
\centering
\includegraphics[height=1.5in,trim=0.75in 0.75in 0.75in 0.75in,
clip]{pydec/hodge/harmonic0}
\includegraphics[height=1.5in,trim=0.75in 0.75in 0.75in 0.75in,
clip]{pydec/hodge/harmonic1} \\
\includegraphics[height=1.5in,trim=0.75in 0.75in 0.75in 0.75in,
clip]{pydec/hodge/harmonic2}
\includegraphics[height=1.5in,trim=0.75in 0.75in 0.75in 0.75in,
clip]{pydec/hodge/harmonic3}
\caption{Four harmonic cochains form a basis for the first
cohomology space $H^1$ for a mesh with four holes. Each cochain is
visualized above as a vector field by interpolating it from the
edge values using Whitney interpolation. See
Section~\ref{subsec:cohomology} for details.}
\label{fig:hodge_harmonic_basis}
\end{figure}
The harmonic $1$-forms shown in Figure~\ref{fig:hodge_harmonic_basis}
are obtained by decomposing random $1$-forms and retaining their
harmonic components. Since the initial basis has no particular
spatial structure, an ad hoc orthogonalization procedure is then
applied. For each basis vector, the algorithm identifies the
component with the maximum magnitude and applies Householder
transforms to force the other vectors to zero at that same component.
\subsection{Sensor network coverage}
\label{subsec:sensor}
As discussed in Section~\ref{sec:rips_complex}, sensor network
coverage gaps can be identified with coordinate-free methods based on
topological properties of the Rips complex. This is for an idealized
abstraction of a sensor network. The following example constructs a
\texttt{rips\_complex} object from a set of 300 points randomly
distributed over the unit square, as illustrated in top left in
Figure~\ref{fig:sensor_network}. Recall that the Rips complex is
constructed by adding an edge between each pair each pair of points
within a given radius. Top right of Figure~\ref{fig:sensor_network}
illustrates the edges of the Rips complex produced by a cut-off radius
of 0.15. The triangles of the Rips complex, illustrated in bottom
left of Figure~\ref{fig:sensor_network}, represent triplets of
vertices that form a clique in the edge graph of the Rips complex. The
Rips complex is created by the following two lines of code
\lstinputlisting[firstline=15,lastline=16]{code/rips/driver.py}
The sensor network is tested for coverage holes by inspecting the
kernel of the matrix $\Delta_1 = \boundary_1^T \boundary_1 +
\boundary_2 \boundary_2^T$ \cite{SiGh2007}. Specifically,
null-vectors of $\Delta_1$, which are called harmonic 1-cochains (by
analogy with the definition of harmonic cochains used in the previous
subsection), reveal the presence of holes in the sensor network. In
this example we explore the kernel of $\Delta_1$ by generating a
random 1-cochain \texttt{x} and extracting its harmonic part using a
discrete hodge decomposition as outlined in the previous
subsection. If the harmonic component of \texttt{x} is (numerically)
zero then we may conclude with high confidence that $\Delta_1$ is
nonsingular and that no holes are present. However, in this case the
hodge decomposition of \texttt{x} produces a nonzero harmonic
component \texttt{h}. Indeed, plotting \texttt{h} on the edges of the
Rips complex localizes the coverage hole, as the bottom right of
Figure~\ref{fig:sensor_network} demonstrates.
To set up the linear systems, the boundary matrices are obtained from
the Rips complex \texttt{rc} created above:
\lstinputlisting[firstline=20,lastline=22]{code/rips/driver.py}
Then the random cochain is created and the Hodge decomposition
computed, to find the harmonic cochain which is then normalized:
\lstinputlisting[firstline=24,lastline=29]{code/rips/driver.py}
\begin{figure}
\centering
\includegraphics[width=0.3\textwidth]{pydec/rips/points_300dpi}
\includegraphics[width=0.3\textwidth]{pydec/rips/edges_300dpi}\\
\includegraphics[width=0.3\textwidth]{pydec/rips/complex_300dpi}
\includegraphics[trim=1.9in 1in 1.8in 1in, clip, scale=0.48]
{pydec/rips/harmonic_300dpi}
\caption{Hodge decomposition for finding coverage holes in an
idealized sensor network. Top left shows a sample sensor network
with 300 randomly distributed points. Pairs of points within a
fixed distance of one another are connected by an edge and this is
shown on top right. Triangles are added to the Rips complex when
three points form a clique (complete graph). These are shown in
bottom left. The existence of a harmonic 1-cochain indicates a
potential hole in the sensor network coverage. In the bottom right
figure, edge thickness reflects the magnitude of the harmonic
cochain on each edge. See Section~\ref{subsec:sensor} for more
details.}
\label{fig:sensor_network}
\end{figure}
\subsection{Least squares ranking on graphs} \label{subsec:ranking}
This is a formulation for ranking alternatives based on pairwise data.
Given is a collection of alternatives or objects that have to be
ranked, by computing a ranking score that sorts them. The ranking
scores are to be computed starting from some pairwise
comparisons. Some examples of objects to be ranked are basketball
teams, movies, candidates for a job. Typically, the given data will
not have pairwise comparisons for all the possible pairs. There is no
geometry in this application, hence no exterior calculus is
involved. PyDEC however still proves useful because the full version
of this example~\cite{HiKaWa2011} uses an abstract simplicial
2-complex. Thus PyDEC is useful in forming the complex and for
determining its boundary matrix. In this simplified example, only an
abstract simplicial 1-complex is needed.
Form a simple graph $G$, with the objects to be ranked being the
vertices and with an edge between any two which have pairwise
comparison data given. If there are $n$ objects, possibly only a
sparse subset out of all possible $O(n^2)$ pairs may have comparison
data associated with them. Here we'll only require that the graph be
connected. This condition can be dropped with the consequence that the
rankings of separate components become independent of each other. The
comparison values are real numbers.
Since $G$ is a simple graph, by orienting the edges arbitrarily it
becomes an oriented 1-dimensional abstract simplicial complex. The
vector $\omega$ of pairwise comparison values is a 1-cochain since if
A is preferred over B by, say, $4$ points, then B is preferred over A
by $-4$ points.
The ranking scores $\alpha$ which are to be computed on vertices form
a 0-cochain. For any edge $e = (u,v)$ from vertex $u$ to vertex $v$,
the difference of vertex values $\alpha(v) - \alpha(u)$ should match
$\omega(e)$ as much as possible, for example, in a least squares
sense. This idea is from~\cite{Leake1976} who proposed it as a method
for ranking football teams. By including the 3-cliques as triangles,
$G$ becomes a 2-dimensional simplicial complex. This was used
in~\cite{JiLiYaYe2011} to extend this ranking
idea. In~\cite{JiLiYaYe2011} the computation of the scores $\alpha$ is
interpreted as one part of the Hodge decomposition of $\omega$. See
Section~\ref{subsec:cohomology} above for a basic discussion of Hodge
decomposition where it is used for computing harmonic cochains on a
mesh. Here we will just compute the ranking score $\alpha$. This is
done by solving the least squares problem
$\boundary_1^T \, \alpha \simeq \omega$.
The graph in this example is used for ranking basketball teams, using
real data for a small subset of American Men's college basketball
games from 2010-2011 season. Each team is a node in the graph and has
been given a number as a name. An edge between two teams indicates
that one of more games have been played between them. The score
difference from these games becomes the input 1-cochain $\omega$, with
one value on each edge. If multiple games were played by a pair the
score differences were added to create this data. The data is stored
as a matrix in which the first two columns are the teams and the third
column is the value of the 1-cochain on that edge.
Once this data is loaded from file, an abstract simplicial complex is
created from the first two columns which form the edges of the
graph. The loading and complex creation is done by the following few
lines of code
\lstinputlisting[firstline=17,lastline=21]{code/ranking/driver.py}
In PyDEC, the simplices that are given as input to construct a complex
are preserved as is. Lower dimensional simplices that are derived from
them are stored and oriented in sorted order. Thus in the above data,
the edge between node 8 and 1 will be oriented from 8 to 1. The above
example data may mean, for example, that team labelled 1 lost to team
labelled 8 by 9 points.
The 1-cochain $\omega$ is now extracted from the data array and the
boundary matrix needed is obtained from the complex and the least
squares problem solved. All this is accomplished in the following
lines
\lstinputlisting[firstline=23,lastline=25]{code/ranking/driver.py} The
resulting alpha values computed are given below.
\[
\begin{array}{lccccccccccc}
\text{Team number} & 0 &1 &2 &3 &4 &5 &6 &7 &8 &9 &10\\
\alpha \text{ value} & 14.5&2.0&0.0&7.2&5.9&9.0&2.3&23.6&11.0&23.8&21.7
\end{array}
\]
The team with $\alpha=0$ is the worst team according to these rankings
and the one with the largest $\alpha$ value (23.8 here) is the best
team. Note that the score difference from the game between 8 and 1
happens to be exactly the difference in $\alpha$ values between
them. This won't always be true. See for example teams 3 and 9. Such
discrepancies come from having a residual in the least squares
solution, and that is a direct result of the presence of cycles in the
graph. In fact it may even happen that team A beats B, which beats C,
which in turn beats A. Thus no assignment of $\alpha$ values will
resolve this inconsistency. This is where a second least squares
problem and hence a Hodge decomposition plays a role. The second
problem is not considered in this example. See~\cite{HiKaWa2011} for
details on the second least squares problem and the role of Hodge
decomposition and harmonic cochains in the ranking context.
\begin{figure}[t]
\centering
\includegraphics[scale=0.4, trim=1in 0in 1in 0in, clip]
{pydec/ranking/games}
\caption{A typical graph of a subset of basketball games. Each node
is a team, labelled by a number. An edge represents one or more
games played by the two teams connected by it. The actual graph
used in the example in Section~\ref{subsec:ranking} is a complete
graph.}
\label{fig:ranking}
\end{figure}
\section{Conclusions}
PyDEC is intended to be a tool for solving elliptic PDEs formulated in
terms of differential forms and for exploring computational topology
problems. It has been used for numerical experiments for Darcy
flow~\cite{HiNaCh2011}, computation of harmonic cochains on two and
three dimensional meshes~\cite{HiKaWaWa2011}, and least squares
ranking on graphs~\cite{HiKaWa2011}. It has also proved valuable in
computational topology work~\cite{DeHiKr2010,DuHi2011}, for creating
complexes and computing boundary matrices. The design goals for PyDEC
have been efficiency and ability to express mathematical formulations
easily. Section~\ref{sec:examples} which described some examples
should give an idea of how close the PyDEC code is to the mathematical
formulation of the problems considered. Many packages exist and are
being created for numerical PDE solutions using differential
forms. There are also many excellent computational topology packages.
PyDEC can handle a large variety of complexes and provides
implementations of discrete exterior calculus and lowest order finite
element exterior calculus using Whitney forms. These qualities make it
a convenient tool to explore the interrelationships between topology,
geometry, and numerical PDEs.
\section*{Acknowledgment}
This work was funded in part by NSF CAREER Award Grant DMS-0645604.
\bibliographystyle{acmdoi}
|
\section{Introduction}
Spectacular progress both in theories and experiments has recently
been made on ultracold atomic and molecular gases in optical
lattices. Owing to the remarkable control over physical
parameters, ultracold gases offer opportunities to simulate the
physics of strongly correlated systems in regimes which are not
easily accessible to solid-state materials. As a result, they
provide very clean and tunable systems in the search for exotic
quantum phases and in probing quantum critical behaviors around
these phases.~\cite{BEC_reviews} For instance, successful
experimental realization of the superfluid-Mott transition for
ultracold bosons in an optical lattice~\cite{BEC_exps} has paved
the way for studying other strongly correlated phases in various
lattice models.
Very recently, it was suggested that intriguing quantum critical
behaviors could occur in attractive bosonic lattice gases with
three-body on-site constraint.~\cite{Diehl10,Lee10} The on-site
constraint can arise naturally due to large three-body loss
processes,~\cite{Daley09,Roncaglia10} and it stabilizes the
attractive bosonic systems against collapse. Such three-body
constrained systems can be realized also in Mott insulating states
of ultracold spin-one atoms at unit filling.~\cite{Mazza10} As
found by the authors in Ref.~\onlinecite{Diehl10}, a dimer
superfluid (DSF) phase consisting of the condensation of boson
pairs can be realized under sufficiently strong attraction.
According to their analysis, the transitions between the DSF phase
and the conventional atomic superfluid (ASF) state are proposed to
be of Ising-like at unit filling and driven first-order by
fluctuations via the Coleman-Weinberg
mechanism~\cite{Coleman-Weinberg} at other fractional fillings.
Later investigation focuses on the nature of the
superfluid-insulator transitions.~\cite{Lee10} It is shown that,
while the Mott-insulator (MI) to DSF transitions are always of
second order, the continuous MI-ASF transitions can be preempted
by first-order ones and interesting tricritical points can thus
appear on the MI-ASF phase boundaries. The conclusions obtained in
Refs.~\onlinecite{Diehl10} and \onlinecite{Lee10} are partly
supported by a recent numerical study employing stochastic series
expansion (SSE) quantum Monte Carlo (QMC) method implemented with
a generalized directed loop algorithm.~\cite{Bonnes11} In
particular, the existence of a tricritical point along the
saturation transition line is verified. However, the nature of
MI-DSF transitions is not examined in their QMC work.
In the present work, the effect of the nearest-neighbor mutual
repulsion on the ground-state phase diagrams of three-body
constrained attractive lattice bosons is investigated by means of
exact diagonalizations (ED).
The systems under consideration are described by the extended
Bose-Hubbard model with a three-body constraint
$a_i^{\dag\,3}\equiv 0$ on square lattices,
\begin{align}
H &= H_{\rm EBH} - \mu \sum_{i} n_{i} \; , \\
H_{\rm EBH} &= - t \sum_{\langle i,j\rangle} a_{i}^{\dagger }a_{j}
+ \frac{U}{2} \sum_{i} n_{i}(n_{i} -1)
+ V \sum_{\langle i,j \rangle} n_i n_j \; . \nonumber
\label{eqn:H}
\end{align}
Here, $a_i (a_i^\dag)$ is the bosonic annihilation (creation)
operator at site $i$, $t$ is the nearest-neighbor hopping
integral, $U<0$ is the on-site two-body attraction ($|U|\equiv 1$
as the energy unit), and $\mu$ is the chemical potential. $V>0$
denotes the nearest-neighbor repulsion, which can come from the
dipole-dipole interactions of the dipolar bosons polarized
perpendicularly to the lattice plane by truncating it at the
nearest-neighbor distance.
For the discussions of possible experimental realizations on the
present model, we refer to Ref.~\onlinecite{Dalmonte11} and
references therein.
The averaged particle density is denoted by $n$ and periodic
boundary conditions are assumed.
Our main results are summarized in Fig.~\ref{fig:phase_diag}. When
$V\neq0$, in addition to the MI and two superfluid phases
mentioned above, a dimer checkerboard solid (DCS) state consisting
of the checkerboard arrangement of \emph{boson pairs} emerges and
occupies the middle part of the phase diagram in the small-$t$
limit. The DSF states now appear only in between the uniform MI
and the DCS states.
Moreover, finite repulsion has interesting effects on the
MI-superfluid transitions also. Because a modest nearest-neighbor
repulsion can avoid cluster formation and then will suppress phase
separation, we observe that the segments of the first-order MI-ASF
transitions on phase boundaries of either the $n=0$ or the $n=2$
MI states (denoted by thick blue dashed lines in
Fig.~\ref{fig:phase_diag}) can shrink to zero as $V$ is increased.
Same conclusion has been reached in other
context.~\cite{Schmidt06} Besides, our findings show that the
phase boundaries of the MI-DSF transitions (denoted by thin red
solid lines in Fig.~\ref{fig:phase_diag}) and therefore the
stability region of the DSF phase in the low-density limit can be
extended to modest hopping parameters $t$ upon tuning $V$, such
that experimental exploration of this interesting state becomes
more feasible.
Our work hence provides a useful guide to the experimental search
of the DSF phase and the associated quantum phase transitions in
ultracold Bose gases in optical lattices.
\begin{figure}[tb]
\includegraphics[clip,width=\columnwidth]{Nfig_1.eps}
\caption{(Color online)
Ground-state phase diagram obtained by ED with (a) $V=0$, (b)(c)
$V=1/2$, and (d)(e) $V=1$.
The thin red solid lines indicate the continuous MI-DSF
transitions, and the thick blue solid (dashed) lines show the
continuous (first-order) MI-ASF transitions.
The phase boundaries of the DCS state are denoted by the green
solid circles (lines are guides to the eye).
The error bar of the determined phase boundaries is smaller than
the width of the lines or the size of the symbols.
The schematic phase boundaries between the ASF and the DSF phases
are added as the thin dotted lines for clarity. }
\label{fig:phase_diag}
\end{figure}
The details of our analysis are explained below. According to the
discussions in Ref.~\onlinecite{Schmidt06}, one can determine the
phase boundaries of either the $n=0$ MI state (i.e., the empty
state) or the $n=2$ MI state (i.e., the completely filled state)
for a given $t$ as follows. For the transitions out of the $n=0$
MI state, we calculate the excitation energies (relative to the
$n=0$ state) $E(1)-\mu$, $E(2)-2\mu$, and $E(4)-4\mu$ of the
lowest states within the subspaces of fixed total particle numbers
$N_{\rm p}=1$, 2, and 4, respectively. Here $E(N_{\rm p})$ denotes
the lowest excitation energy of $H_{\rm EBH}$ within the fixed
$N_{\rm p}$ subspace. From these quantities, the two- and the
four-particle binding energies, $\Delta_{\rm 2p}=E(2)-2E(1)$ and
$\Delta_{\rm 4p}=E(4)-2E(2)$, can be obtained. For a given value
of the hopping parameter $t$, if both $\Delta_{\rm 2p}$ and
$\Delta_{\rm 4p}$ are positive, there exists no bound states and
the energy gap of the single-particle state closes first upon
increasing the chemical potential $\mu$. This leads to a
continuous MI-ASF transition at $\mu_{c,\textrm{0MI-ASF}} =
E(1)$.~\cite{note1} When $\Delta_{\rm 2p}<0$ but $\Delta_{\rm
4p}>0$, instead, a bound state of two particles appears and its
gap closes first. Thus a continuous MI-DSF transition will occur
at $\mu_{c,\textrm{0MI-DSF}} = E(2)/2$. Aside from these two
possibilities, the condition of $\Delta_{\rm 4p}<0$ gives a
precursor of instability of boson pairs towards cluster formation.
That is, phase separation emerges and a first-order transition
will be observed in varying $\mu$. In the present ED analysis, we
estimate the first-order transition point by $E(4)/4$, which
provides an upper bound of the exact value. Similar discussions
apply to the transitions out of the $n=2$ MI state, where the
holes created from the $n=2$ state take the role played by the
particles discussed previously. We remind that the first-order
transition points estimated by the four-hole excitation energies
will be instead lower bounds of the exact values along the
saturation transition line.
\begin{figure}[tb]
\includegraphics[clip,width=\columnwidth]{Nfig_2.eps}
\caption{(Color online)
(a) Two- and four-hole binding energies, $\Delta_{\rm 2h}$ and
$\Delta_{\rm 4h}$, around the $n=2$ state and (b) two- and
four-particle binding energies, $\Delta_{\rm 2p}$ and $\Delta_{\rm
4p}$, around the $n=0$ state as functions of $t$ for $V=1/2$ and
$N_{\rm s}=10\times10$. The dashed lines separate the transitions
of different characters. } \label{fig:binding_energies}
\end{figure}
Here distinct MI-superfluid transitions are determined by the
above method for systems of $N_{\rm s}=10\times10$ sites. The
analysis for the $V=1/2$ case is described below for illustration.
The results of various binding energies as functions of hopping
parameter $t$ for this $V$ are presented in
Fig.~\ref{fig:binding_energies}. As seen from
Fig.~\ref{fig:binding_energies}(b), there is a finite region of
$t$ within which $\Delta_{\rm 4p}<0$. It implies that, on the
phase boundary of the $n=0$ state, the continuous MI-DSF (for
$\Delta_{\rm 2p}<0$ but $\Delta_{\rm 4p}>0$) and continuous MI-ASF
(for both $\Delta_{\rm 2p}$, $\Delta_{\rm 4p}>0$) transitions are
separated by first-order MI-ASF transitions for $0.15 \lesssim t
\lesssim 0.26$. On the contrary, the two continuous MI-superfluid
transitions on the phase boundary of the $n=2$ state should meet
directly at $t\approx 0.068$. The transition points $\mu_c$ for a
given $t$ can be determined by using the excitation energies
$E(N_{\rm p})$ as explained above.~\cite{note2}
According to our ED calculations, for $V=1/2$, the continuous
$n=2$ MI-DSF transitions end at
$(t_{\rm E}, \mu_{\rm E}) \simeq (0.068, 3.55)$,
while $(t_{\rm E}, \mu_{\rm E}) \simeq (0.15, -0.64)$
for that of the continuous $n=0$ MI-DSF transition line. There
exists also a tricritical point
$(t_{\rm T}, \mu_{\rm T}) \simeq (0.26, -1.04)$
on the $n=0$ MI-ASF transition line separating the continuous from
the first-order ones.
\begin{figure}[tb]
\includegraphics[clip,width=\columnwidth]{Nfig_3.eps}
\caption{(Color online)
Total particle numbers $N_{\rm p}$ as functions of chemical
potential $\mu$ for (a) $t=0.1$ and (b) $t=0.3$ with $V=1/2$ and
system sizes $N_{\rm s}=4\times4$.
The insets illustrates the complete changes in particle density
$n$ from the $n=0$ to the $n=2$ states as $\mu$ increases. }
\label{fig:n_rho_vs_mu}
\end{figure}
We now turn to the discussions on the phase boundaries of the DCS
state. Due to the effect of the nearest-neighbor repulsion $V$,
\emph{boson pairs} can form a solid with checkerboard structure at
unit filling $n=N_{\rm p}/N_{\rm s}=1$ and lead to the DCS state.
In the zero-hopping limit, this DCS state can be stabilized when
$-1/2 < \mu < -1/2 + 2zV$ with the coordination number $z=4$ for
square lattices. The DCS state can melt into the DSF or the ASF
states under increasing hopping, and its stability region in $\mu$
is expected to reduce to zero as $t$ increases. This picture is
supported by our numerical calculations as seen in
Fig.~\ref{fig:phase_diag}. Due to the limitation in numerics, the
phase boundaries of the DCS state are estimated from the $n=1$
plateaus in the $\mu$-$n$ plots for different $t$'s with systems
sizes of $N_{\rm s}=4\times4$. For illustration, total particle
numbers $N_{\rm p}$ and the particle density $n$ as functions of
chemical potential $\mu$ for two different values of $t$'s with
$V=1/2$ are presented in Fig.~\ref{fig:n_rho_vs_mu}. The large
plateaus at $n=1$ in the $\mu$-$n$ plots clearly indicate the
presence of the DCS state. Two ends of the plateau give the
melting transition points in $\mu$ for a given $t$, as depicted in
Fig.~\ref{fig:phase_diag}.
Moreover, distinct characters in various transitions can be
revealed by the way in which the total particle number $N_{\rm p}$
changes upon varying the chemical potential
$\mu$.~\cite{Schmidt06} In the conventional ASF phase, the number
of particles will increase by 1 when the chemical potential is
increased. However, in the DSF phase, only pairs of bosons appear
in the system due to the presence of a pairing gap. Thus adding a
single boson is forbidden and the jumps in the particle number by
2 will be observed as $\mu$ is varied. Besides forming pairs,
bosons can become unstable towards cluster formation such that the
particle number jumps by a finite amount in the course of tuning
$\mu$. This corresponds to a first-order transition under the
change in the chemical potential. As shown in
Fig.~\ref{fig:n_rho_vs_mu}, for $t=0.1$, the system evolves across
a continuous MI-DSF transition from the $n=0$ state to the DSF
state, and then follows a first-order transition to the DCS state.
For larger $t$ (say, $t=0.3$), the state after the continuous
transition from the $n=0$ MI state becomes the ASF one instead.
Adding more bosons by further increasing $\mu$, the system can
again follows a first-order transition to the DCS state.
\begin{figure}[tb]
\includegraphics[clip,width=\columnwidth]{Nfig_4.eps}
\caption{(Color online)
Static structure factor $S(\mathbf{k})$ of the $n=1$ states for
various $t$'s with $V=1/2$ and $N_{\rm s}=4\times4$. }
\label{fig:structure_factor}
\end{figure}
To provide support on the nature of the DCS states within the
$n=1$ plateaus in the $\mu$-$n$ curves, the static structure
factors $S(\mathbf{k})=(1/N_{\rm s}^2) \langle |\sum_j n_j
e^{i\mathbf{k}\cdot\mathbf{r}_j}|^2 \rangle$ of the $n=1$ states
for several hopping parameters $t$'s with $V=1/2$ and system size
$N_{\rm s}=4\times4$ are shown in Fig.~\ref{fig:structure_factor}.
It is found that the static structure factors do have peaks at the
wave vector $\mathbf{k}=(\pi,\pi)$ and thus signal the
checkerboard pattern of the boson pairs. As observed from
Fig.~\ref{fig:structure_factor}, when $t$ is increased, the peak
value $S(\pi,\pi)$ of the structure factor will decrease from its
classical value $S(\pi,\pi)=1$ at $t=0$. This indicates the
quantum melting of the DCS state into the DSF or the ASF states
under increasing hopping.
\begin{figure}[t]
\includegraphics[clip,width=\columnwidth]{Nfig_5.eps}
\caption{(Color online)
Root-mean-square separation $r_{\rm rms}$ of the (a) particle (b)
hole pairs on $N_{\rm s}=10\times10$ square lattices as functions
of $t$ for various repulsions $V$.
Insets: $r_{\rm rms}$ and maximal extent $t_{\rm max}$ in hopping
integral for the MI-DSF transitions as functions of $V$.}
\label{fig:coherence_length}
\end{figure}
Some further microscopic details can be uncovered by the ED
calculations. Similar to the counterpart of fermion pairing, a
smooth crossover from the on-site pairs to the loosely bound pairs
may occur also in the present attractive boson systems as the
hopping parameter $t$ increases. In
Fig.~\ref{fig:coherence_length}, evidences supporting this
expectation are presented. Here the coherence length of the boson
pairs is estimated by the root-mean-square separation $r_{\rm
rms}\equiv \sqrt{\langle r^2 \rangle}$,~\cite{Ohta1995,Leung02}
which is evaluated under the ground state of a single pair within
the two-particle or the two-hole subspaces. Our results show rapid
but smooth crossovers from the tightly bound molecules to the
loosely bound pairs, as long as the DSF phase can be stabilized up
to modest values of $t$ (say, the $V=1$ case for the particle
pairs and the $V=1/2$ case for the hole pairs). The general
dependence on the repulsion $V$ of the maximal extent $t_{\rm
max}$ in hopping integral for the stable DSF state in the
low-density limit and the corresponding $r_{\rm rms}$ is shown in
the insets of Fig.~\ref{fig:coherence_length}.
We find that larger $t_{\rm max}$'s in general lead to longer
$r_{\rm rms}$'s. The dependence of $t_{\rm max}$ and $r_{\rm rms}$
on $V$ is found to be nonmonotonic, and their functional forms
shows asymmetry between the cases of the particle and the hole
pairs.
Our conclusions presented in Figs.~\ref{fig:phase_diag} and
\ref{fig:coherence_length} should be of help in determining
optimal experimental settings in the search of the DSF phase in
ultracold Bose gases in optical lattices.
In summary, the ground-state phase diagrams of the three-body
constrained extended Bose-Hubbard model for various repulsions are
investigated. Large plateaus at $n=1$ in the $\mu$-$n$ curves
which show the DCS states are observed. Finite repulsions modify
the MI-superfluid transitions also. We find that the repulsion $V$
can change the first-order MI-ASF transitions into the continuous
ones and the stability regions of the DSF phase can be tuned by
$V$. Therefore, carefully adjusting system parameters into the
suggested parameter regime are necessary to find experimentally
the interesting DSF phase in real ultracold Bose gas in optical
lattices.
Y.-C.C., K.-K.N., and M.-F.Y. thank the National Science Council
of Taiwan for support under Grant No. NSC 99-2112-M-029-002-MY3,
NSC 97-2112-M-029-003-MY3, and NSC 99-2112-M-029-003-MY3,
respectively.
|
\section{Introduction}
Recently, a quite complete mathematical thermodynamic structure for general stochastic processes has been proposed, for both discrete Markov jump processes and continuous Langevin-Fokker-Planck systems \cite{Ge:2010fk,esposito_prl:2010,qian_decomp}. The entropy production rate $e_p$ of a Markov dynamics can be mathematically decomposed into two non-negative terms: free energy dissipation rate $-\dot{F}$, corresponding to Boltzmann's original theory on irreversibility of spontaneous change, and house-keeping heat $Q_{hk}$, corresponding to Brussels school's notion of irreversibility in nonequilibrium steady states (NESS) \cite{NP_book,zqq,gqq}. For almost all applications of stochastic dynamic theories in physics, chemistry and biology, there will be multiple time scales, and often with a significant separation. When a dynamical system is highly nonlinear, and its interaction network includes feedbacks, multistability with several attractors is often the rule rather than exception. On the other hand, the concept of ``landscape'' has become a highly popular metaphor as well as a useful analytical device \cite{wolynes,ge_qian_lc}. When stochastic nonlinear dynamical systems of populations of individuals become large, a time scale separation between inter- and intra-attractoral dynamics becomes almost guaranteed. In cellular biology, they have been called {\em biochemical network} and {\em cellular evolution} time scales
respectively \cite{qian_iop}.
In chemistry, a separation of time scales has lead to a fundamental understanding of chemical reactions in terms of discrete states of molecules, in complementary to the full mechanical description of constitutive atoms in terms of bond lengths and bond angles. In fact, one of the most significant, novel chemical concepts is ``transition state'', which in terms of modern nonlinear dynamical systems is the saddle point on a separatrix that divides two attractors \cite{wolynes_sfi}. Recall also that in applications of Gibbs' formalism of statistical mechanics to chemical equilibrium, the conditional free energy plays a central role \cite{ben_naim_book,bwt_book}. One usually does not work with the pure mechanical energy of a system; rather, one works with a {\em conditional free energy} from a coarse-grained representation and develops a partition function thereafter. An essential notion in this approach is the {\em consistency across scales}. We shall expand on these ideas more precisely in the following section.
In the present work we address the question of ``whether the mathematical thermodynamic structure of a given continuous stochastic nonlinear dynamical system is consistent with the one associated with the emergent discrete Markov jump process.'' In other words: whether the formal mathematical relations between state functions and process variables remain unchanged when the system is viewed at either a finer- or a coarse-grained scale.
It is important to point out, at the onset, that the ``state'' of a stochastic dynamical system has always had two distinctly different meanings: ($a$) a state of a single, stochastically fluctuating, system; and ($b$) a state in terms of the distribution over an ensemble. In more precise mathematical terms, ($a$) are functions of a stochastic process, while ($b$) are functionals of the solution to a Fokker-Planck equation. The deep insight from the theory of
probability is that these are two complementary, yet mathematically
identical, descriptions of a same dynamical process. With this distinction in mind, entropy and free energies are state functionals of the second type, while energy is a state function of the first type. A state function of the first type naturally has fluctuations. On the other hand, most classical thermodynamic functions are the second type.
Attempting to introduce entropy as a function of the first type, \citet{Qian:2001fk} defined a trajectory based entropy $\Upsilon_t = -\ln f_X^s(X_t)$ where $X_t$ is a diffusion process, and $f_X^s(x)$ is the stationary solution to the corresponding Fokker-Planck (Kolmogorov forward) equation. One immediately sees that entropy is really a population-based concept. For irreversible diffusion processes (i.e., without detailed balance), $f_X^s(x)$ is non-local \cite{Qian:2001fk}. However, for reversible diffusion processes with detailed balance, since $f^s_X(x)\propto e^{-\phi(x)}$ where $\phi(x)$ is potential energy, fluctuating $\Upsilon_t$ and fluctuating energy $\phi(X_t)$ are the same.
\section{Equilibrium statistical thermodynamic consistency across scales}
In equilibrium statistical mechanics, the concept of \emph{consistency}---or invariance---has a fundamental importance in the study of realistic physical systems at an appropriate scale \cite{bwt_book,ben_naim_book}. In a continuous system, the conditional free energy is known as the {\em potential of mean force} \cite{Kirkwood:1935fk}. The conditional free energy can do work just as the Newtonian mechanical energy; the concept of {\em entropic force} is well understood in physical chemistry \cite{dill_book}.
For an investigator working on certain level of description, with discrete states ($i=1,2,\cdots$) and conditional free energy ($A_i$), the canonical partition function of the statistical thermodynamic system is \cite{bwt_book,ben_naim_book,dill_book}
\begin{equation}
Z(T) = \sum_{i=1} e^{-A_i/k_BT}.
\label{eq_01}
\end{equation}
Note that, since $A_i$ is a conditional free energy, it can be decomposed into $A_i = E_i-TS_i$, where $E_i =\partial (A_i/T)/\partial (1/T)$ and $S_i=-\partial A_i/\partial T$. In general, both $E_i$ and $S_i$ are themselves functions of the temperature.
Now, for another investigator who works at a much more refined level, with a continuos variable $\mathbf{x}$, each state $i$ corresponds to a unique region of the phase space $\omega_i$, with $\omega_i\bigcap\omega_j=\emptyset$ for $i\neq j$, and $\bigcup_{i=1} \omega_i = \Omega$ covering the entire phase-space region available to the system. Let $V(\mathbf{x})$ ($\mathbf{x}\in\Omega$) be the potential of mean force at this level. Then, his/her canonical partition function is
\begin{equation}
\widetilde{Z}(T) = \int_{\Omega} d\mathbf{x} e^{-V(\mathbf{x})/k_BT}.
\label{eq_02}
\end{equation}
We see that $Z(T)$ and $\widetilde{Z}(T)$ are equal if the $A_i$ in Eq. (\ref{eq_01}) are such that
\begin{equation}
A_i(T)=-k_BT\ln\left(\int_{\omega_i} d\mathbf{x} e^{-V(\mathbf{x})/k_BT}
\right).
\label{eq_03}
\end{equation}
The equality $Z(T)=\widetilde{Z}(T)$ following Eqs. (\ref{eq_01}) and (\ref{eq_03}) is exact in equilibrium statistical mechanics. Nonetheless, as we demonstrate in this work, its generalization to include dynamics requires a separation of time scales for the dynamics within each $\omega_i$ and the dynamics between $\omega$'s (this is well understood in physical chemistry as ``rapid equilibrium'' averaging). Here, we choose the $\omega$'s according to the basins of attraction of the underlying nonlinear dynamics. In this case, the separation of time scales for intra- and inter-attractoral dynamics is widely accepted.
The mathematical origin of the consistency discussed above relies in fact upon the concepts of \emph{conditional probability}, \emph{marginal probability}, and the \emph{law of total probability}! The free energy of a system, or a sub-system, is directly related to its probability. The same cannot be said for the entropy \cite{noyes61,Qian:1996uq}, which increses with more detailed descriptions and is also coordinate-system dependent for continuous variabels:
\begin{eqnarray}
S(T) &=& -k_B\sum_i\left(\frac{e^{-A_i/k_BT}}{Z(T)}\right)
\ln\left(\frac{e^{-A_i/k_BT}}{Z(T)}\right)
\nonumber\\
&\leq& -k_B\int_{\Omega} d\mathbf{x} \left(\frac{e^{-V(\mathbf{x})/k_BT}}
{\widetilde{Z}(T)}\right)
\ln\left(\frac{e^{-V(\mathbf{x})/k_BT}}{\widetilde{Z}(T)}\right)
= \widetilde{S}(T).
\label{EntRef}
\end{eqnarray}
The proof for the inequality can be found in any text on information theory \cite{cover_book}. Also see Appendix \ref{appendix}.
Note that since internal energy is the sum $-k_BT\ln Z(T) + TS(T)=E(T)$, one immediately has $\widetilde{E}(T)\ge E(T)$ across scales as well. This leads to a type of entropy-energy compensation \cite{Qian:1996uq, Qian:1998kx, Qian:2001fk, Santillan:2011vn}.
\section{Open System concepts and definitions}
\subsection{Fokker-Planck equation, stationary distribution, and detailed balance}
Consider a system whose state is represented by variable $\mathbf{x}$, and assume that $\mathbf{x}$ is a stochastic variable following a continuous-space continuous-time diffusion process. Let $P(\mathbf{x},t)$ denote the probability density of finding the system in state $\mathbf{x}$ at time $t$. In what follows we shall assume that the master equation (Chapman-Kolomogorov equation) governing the dynamics of $P(\mathbf{x},t)$ can be represented by the following Fokker-Planck equation \footnote{Following \citeauthor{Kubo:1973fk} we write the Fokker-Planck equation for the probability density function of a continuous diffusion in the divergence form. Other choices such as Ito or Stratonovich forms can be readily transformed into the present one.}:
\begin{equation}
\frac{\partial P(\mathbf{x},t)}{\partial t} = -\nabla \cdot \mathbf{J},
\label{fpe}
\end{equation}
where
\begin{equation}
\mathbf{J}(\mathbf{x},t) = -D(\mathbf{x})\left[\epsilon\nabla P(\mathbf{x},t)
+ \mathbf{F}(\mathbf{x}) P(\mathbf{x},t)\right]
\label{probcurr}
\end{equation}
is the probability current. In equation (\ref{probcurr}), $D(\mathbf{x})$ is the diffusion coefficient, $\mathbf{F}(\mathbf{x})$ is the force (not necessarily conservative) acting upon the system, and $\epsilon$ is a parameter which will serve as our ``temperature''. For fluctuations of isothermal molecular systems in equilibrium at temperature $T$, Einstein's relation dictates that $\epsilon= k_BT$, where $k_B$ is Boltzmann's constant. However, in the present work, the notion of temperature does not exist.
We shall assume that the system can be driven and approaches to a nonequilibrium steady state in infinite time \cite{zqq,gqq}. The nonequilibrium driving force comes from a ``chemical driving force'' in $\nabla\times\mathbf{F}(\mathbf{x})\neq 0$ \cite{qian_jpc_06, qian_arpc}. When $\mathbf{F}(\mathbf{x})$ is conservative,
\[
\nabla\times \mathbf{F}(\mathbf{x})= 0 \
\Rightarrow \
\mathbf{F}(\mathbf{x})=-\nabla V(\mathbf{x}).
\]
Then, the stationary $P^{s}(\mathbf{x})=e^{-V(\mathbf{x})/\epsilon}$, while the stationary $\mathbf{J}(\mathbf{x})=0$. Furthermore, the stationary distribution $P^{s}(\mathbf{x})$ complies with detailed balance. That is, $P^{s}(\mathbf{x})$ is analogous to thermodynamic equilibrium \citep{Qian:2002fk}. Hence, a stationary system (\ref{fpe}) is also mathematically called equilibrium in this case \cite{zqq,gqq}.
Let us assume that Eq. (\ref{fpe}) has one single ergodic stationary solution $P^s(\mathbf{x})$ with the corresponding stationary current $\nabla\cdot\mathbf{J}^s(\mathbf{x})=0$; but usually, $\mathbf{J}^s(\mathbf{x})\neq 0$. We shall again write the stationary probability density as
\begin{equation}
P^s(\mathbf{x}) = C \exp \left(-\Psi_{\epsilon}(\mathbf{x})/\epsilon \right),
\label{IrrevPot}
\end{equation}
where function $\Psi_{\epsilon}(\mathbf{x})$ is known as the \emph{non-equilibrium} potential \cite{Kubo:1973fk}, $C = \left[ \int_{\Omega} d\mathbf{x} \exp \left(- \Psi_{\epsilon}(\mathbf{x}) /\epsilon\right) \right]^{-1}$ is a normalization constant, and $\Omega$ represents the region of the state space available to the system. Note that in general $\Psi_{\epsilon}(\mathbf{x})$ is actually also a function of $\epsilon$. However, for many interesting applications, $\Psi_{\epsilon}(\mathbf{x})$ is a function of $\mathbf{x}$ alone in the limit of $\epsilon\rightarrow 0$. The probability current $\mathbf{J}^s(\mathbf{x})$ is a time-invariant, divergence-free vector field for the
stationary distribution $P^s(\mathbf{x})$ \cite{qian_decomp}.
\subsection{Small $\epsilon$ limit}
Let us consider first the case where $\mathbf{F}(\mathbf{x})$ is conservative. When $\epsilon=0$, the system dynamic behavior is dictated, in a deterministic fashion, by the potential $V(\mathbf{x})$. That is, depending on the initial condition, the system state will evolve towards one of the local minima of $V(\mathbf{x})$ and will remain there indefinitely. In that sense, every local minimum of $V(\mathbf{x})$ corresponds to a stable steady state. Moreover, each stable steady state has a basin of attraction associated to it. Whenever the initial condition lies within a given basin of attraction, the system will eventually reach the state corresponding to the local minimum $V(\mathbf{x})$ point. Finally, all neighboring basins of attraction are separated by saddle points and separatrices which the system has to surpass in order to go from one basin to the other.
If $\epsilon$ is not zero, but very small as compared to the height of the saddle points and separatrices between basins of attraction, the stationary probability distribution $P^s(\mathbf{x})$ will present high narrow peaks around the stationary states, and will attain very low values at the saddle nodes separating neighboring attractive basins. This further implies that the transition rates between every two attractive basins are small as well, as compared with the probability relaxation-rates within each basin.
In the case of a non-conservative force, the stationary distribution depends on the non-equilibrium potential $\Psi_{\epsilon}(\mathbf{x})$, when it exists, in an analogous way as $P^s(\mathbf{x})$ depends on $V(\mathbf{x})$ for conservative forces \cite{fw_book}. This means that the above considerations could be still valid when $\mathbf{F}(\mathbf{x})$ is non conservative. In particular, $\Psi_{\epsilon}(\mathbf{x})$ defines a landscape in the state space \cite{ge_qian_lc}, a basin of attraction can be identified around each of the local minima of $\Psi_{\epsilon}(\mathbf{x})$ and, in the small $\epsilon$ limit, $P^s(\mathbf{x})$ presents high narrow peaks around each minimum of $\Psi_{\epsilon}(\mathbf{x})$ and takes very low values at the saddle points and separatrices that separate neighboring attractive basins. See \cite{ge_qian_lc} for systems with limit cycles.
\section{Probability discretization}
\subsection{Discretization of the state space}
Consider a system whose non-equilibrium potential $\Psi_{\epsilon}(\mathbf{x})$ has $N$ local minima with the corresponding basins of attraction in the state space. Let $\omega_i$ be the region of the state space delimited by the attractive basin of the $i$th local minimum of $\Psi_{\epsilon}(\mathbf{x})$, and let $\Xi_i$ denote the boundary of $\omega_i$. In Appendix \ref{boundary} we demonstrate that $\Xi_i$ can always be written as
\begin{equation}
\Xi_i = \bigcup_{j=0}^N \Xi_{ij},
\label{boundaries}
\end{equation}
where $\Xi_{ij}=\Xi_{ji}$ ($j=1,2\dots N$) represents the common boundary between $\omega_i$ and $\omega_j$, while $\Xi_{i0}$ denotes the part of the $\omega_i$ boundary not shared with any other region. In case that $\omega_i$ and $\omega_j$ share no boundary, $\Xi_{ij} = \emptyset$.
From the above considerations, the probability $P_i$ that the system state is in region $\omega_i$ is
\begin{equation}
P_i(t) = \int_{\omega_i} d\mathbf{x} P(\mathbf{x},t).
\label{pi}
\end{equation}
Furthermore, it follows from (\ref{fpe}) and Stokes' theorem that
\begin{equation}
\frac{d P_i(t)}{dt} = \int_{\omega_i} d\mathbf{x} \frac{\partial P(\mathbf{x},t)}{\partial t}
= - \sum_{j=1}^N \int_{\Xi_{ij}} d\mathbf{s} \cdot \mathbf{J}(\mathbf{x},t) .
\label{dpidtaux}
\end{equation}
In the derivation of the above equation we have assumed that the probability current $\mathbf{J}$ is zero along $\Xi_{i0}$.
Let us analyze the integral $\int_{\Xi_{ij}} d\mathbf{s} \cdot \mathbf{J}$. From (\ref{probcurr}), it can be rewritten as
\begin{eqnarray}
\int_{\Xi_{ij}} d\mathbf{s} \cdot \mathbf{J} & = & - \int_{\Xi_{ij}} d\mathbf{s} \cdot [\epsilon D(\mathbf{x}) \nabla P(\mathbf{x},t)] \; \mathrm{H}(- d\mathbf{s} \cdot [\epsilon D(\mathbf{x}) \nabla P(\mathbf{x},t)]) \nonumber \\
& & - \int_{\Xi_{ij}} d\mathbf{s} \cdot [\epsilon D(\mathbf{x}) \nabla P(\mathbf{x},t)] \nonumber \; \mathrm{H}(d\mathbf{s} \cdot [\epsilon D(\mathbf{x}) \nabla P(\mathbf{x},t)]) \\
& & + \int_{\Xi_{ij}} d\mathbf{s} \cdot [D(\mathbf{x}) \mathrm{F}(\mathbf{x}) P(\mathbf{x},t)] \; \mathrm{H}( d\mathbf{s} \cdot [D(\mathbf{x}) \mathrm{F}(\mathbf{x}) P(\mathbf{x},t)]) \nonumber \\
& & + \int_{\Xi_{ij}} d\mathbf{s} \cdot [D(\mathbf{x}) \mathrm{F}(\mathbf{x}) P(\mathbf{x},t)] \; \mathrm{H}( - d\mathbf{s} \cdot [D(\mathbf{x}) \mathrm{F}(\mathbf{x}) P(\mathbf{x},t)]) , \nonumber
\end{eqnarray}
with $\mathrm{H}(\cdot)$ being Heaviside's function. Given that $\mathrm{H}(x)>0$ if and only if $x>0$, just one of the first two terms in the right hand side of the previous equation is positive or zero, while the other is negative or zero; the same is true for the last two terms. Let us define $J_{ij}$ as the sum of the two positive terms, and $J_{ji}$ as minus the sum of the two negative terms. Hence,
\begin{equation}
\int_{\Xi_{ij}} d\mathbf{s} \cdot \mathbf{J} = J_{ij} - J_{ji}.
\label{sumcurr}
\end{equation}
From its definition, $J_{kl} \geq 0$ for all $k,l = 1,2\dots N$. Furthermore, from Eq. (\ref{sumcurr}), $J_{kl}$ can be interpreted as the net probability flux from $\omega_k$ into $\omega_l$. Finally, by substituting Eq. (\ref{sumcurr}) into Eq. (\ref{dpidtaux}) we obtain
\begin{equation}
\frac{d P_i(t)}{dt} = \sum_{j=1}^N \left(J_{ji} - J_{ij}\right).
\label{dpidt}
\end{equation}
\subsection{Adiabatic approximation and Kramers theory}
Following \citet{Kampen:2007kx}, \citet{Risken:1996uq} and
Freidlin-Wentzell \cite{fw_book} we make use of the assumption that $\epsilon$ is much smaller than the height of the saddle nodes between every two attractive basins so that the corresponding transition rates are very small, as compared with the probability dynamics inside each basin. In consequence, the probability distribution within any $\omega_i$ can be approximated by the quasi-stationaty distribution
\begin{equation}
P(\mathbf{x},t) \approx C_i(t)\exp(-\Psi_{\epsilon}(\mathbf{x})/\epsilon),
\label{qsdist}
\end{equation}
with $C_i(t)$ given by
\begin{equation}
C_i(t) = \frac{ P_i(t) }{\int_{\omega_i} d \mathbf{x} \exp(-\Psi_{\epsilon}(\mathbf{x})/\epsilon) },
\label{normconst}
\end{equation}
so that $\int_{\omega_i} d \mathbf{x} P(\mathbf{x},t) = P_i(t) $. From the approximation above, and a theorem from \cite{fw_book} that justifies the application of Kramers' theory to any pair of
adjacent $i$ and $j$ \citep{Kampen:2007kx,Risken:1996uq}, it follows that
\begin{equation}
J_{ij}(t) = \gamma_{ij} P_i(t),
\label{linearflux}
\end{equation}
where the transition rates $\gamma_{ij}$ are determined by the
so-called local pseudo-potential; particularly, by the height
of the saddle points between neighbouring $i$ and $j$ attractors \cite{weinan,ge_qian_lc}.
Finally, by substituting Eq. (\ref{linearflux}) into Eq. (\ref{dpidt}) we obtain the following master equation for $P_i(t)$:
\begin{equation}
\frac{d P_i(t)}{dt} = \sum_{j=1}^N \gamma_{ji} P_j(t) - \gamma_{ij} P_i(t).
\label{me}
\end{equation}
Note that, since in the stationary state $J_{ij}^s=P^s_i\gamma_{ij}$, but $J^s_{ij}\neq J^s_{ji}$ in general, $P^s_i\gamma_{ij} \neq P^s_j\gamma_{ji}$. Therefore the emergent master equation in (\ref{me}) is not necesarilly detail balanced.
\section{Thermodynamic state functionals}
\subsection{Internal Energy}
Under the assumptions that the system modeled by Eq. (\ref{fpe}) has a unique stationary distribution one can mathematically define, following Kubo \cite{Kubo:1973fk} and many others including Ge and Qian \cite{Ge:2010fk}, the energy function associated to state $\mathbf{x}$ via the stationary distribution $P^s(\mathbf{x})$ as
\begin{equation}
\phi(\mathbf{x}) = -\epsilon \ln P^s(\mathbf{x}).
\label{udef}
\end{equation}
In systems with detailed balance, $P^s(\mathbf{x})$ equals the thermodynamic-equilibrium probability distribution $P^e(\mathbf{x})$ and Eq. (\ref{udef}) is equivalent to Boltzmann's law---provided we choose the zero level of free energy such that the partition function equals one. When detailed balance is not fulfilled, Kubo et. al. called $\phi(\mathbf{x})$ a \emph{stochastic potential} \citep{Kubo:1973fk}. Finally, from (\ref{udef}), the mean ``energy'' of the mesoscopic state $P(\mathbf{x},t)$ can be written as
\begin{equation}
U(t) = \int_{\Omega} d \mathbf{x} P(\mathbf{x},t)\phi(\mathbf{x}) = -\epsilon \int_{\Omega} d \mathbf{x} P(\mathbf{x},t) \ln P^s(\mathbf{x}).
\label{IntEner}
\end{equation}
Given the definition of the attractive basins, $\Omega = \bigcup_{i=1}^N \omega_i$, while $\omega_i \bigcap \omega_j = \emptyset$ for all $i \neq j$. Then, Eq. (\ref{IntEner}) can be rewritten as
\begin{equation}
U(t) = -\epsilon\sum_{i=1}^N P_i(t) \ln P_i^s -\epsilon \sum_{i=1}^N P_i(t) \int_{\omega_i} d \mathbf{x}\frac{P(\mathbf{x},t)}{P_i(t)} \ln \frac{P^s(\mathbf{x})}{P_i^s},
\label{udecomp}
\end{equation}
with $ P_i^s = \int_{\omega_i} d \mathbf{x} P^s(\mathbf{x})$. Finally, substitution of Eqs. (\ref{qsdist}) and (\ref{normconst}) into Eq. (\ref{udecomp}) leads to
\begin{equation}
U(t) = -\epsilon\sum_{i=1}^N P_i(t) \ln P_i^s +\sum_{i=1}^N P_i(t) \tilde{s}_i,
\label{udecomp2}
\end{equation}
where
\begin{equation}
\tilde{s}_i = -\epsilon \int_{\omega_i} d \mathbf{x} \frac{\exp(-\phi(\mathbf{x}/\epsilon))}{Z_i} \ln \frac{\exp(-\phi(\mathbf{x}/\epsilon))}{Z_i},
\label{tildes}
\end{equation}
and ${Z_i} = \int_{\omega_i} d \mathbf{x} \exp(-\phi(\mathbf{x}/\epsilon))$ . The first term in the right hand side of Eq. (\ref{udecomp2}) can be interpreted as a coarse-grained contribution to the system's internal energy, arising from the distribution of probability among the $N$ available attractive basins. On the other hand, the second term in the right hand side of Eq. (\ref{udecomp2}) corresponds to the fine-grained contribution to the system internal energy, due to the distribution of probability density $P(\mathbf{x},t)$ within each basin. The Boltzmann-like form of the terms within the integral originates from the adiabatic approximation we have made.
\subsection{Entropy and Free Energy}
The Gibbs entropy is defined as usual:
\begin{equation}
S(t) = -\epsilon \int_{\Omega} d \mathbf{x} P(\mathbf{x},t) \ln P(\mathbf{x},t) .
\label{Entropy}
\end{equation}
By following an analogous procedure to the one in the previous subsection, known as the chain rules for entropy and
relative entropy, Eq. (\ref{Entropy}) can be rewritten as
\begin{eqnarray}
S(t) & = & -\epsilon\sum_{i=1}^N P_i(t) \ln P_i^s -\epsilon\sum_{i=1}^N P_i(t) \int_{\omega_i} d \mathbf{x} \frac{P(\mathbf{x},t)}{P_i(t)} \ln \frac{P(\mathbf{x},t)}{P_i(t)} , \nonumber \\
& = &-\epsilon\sum_{i=1}^N P_i(t) \ln P_i(t) + \sum_{i=1}^N P_i(t) \tilde{s}_i,
\label{sdecomp}
\end{eqnarray}
Once again, the entropy can be decomposed into a coarse-grained contribution---due to the distribution of probability among the $N$ available attractive basins---as well as a fine-grained contribution due to the distribution of the probability density $P(\mathbf{x},t)$ within each basin. This result is in complete agreement with that in Eq. (\ref{EntRef}). Notice that, because of the adiabatic approximation, the fine-grained contributions to both $U$ and $S$ happen to be equal.
From its definition, the mean Helmholtz free energy is
\begin{equation}
F(t) = U(t)-S(t) = \epsilon \int_{\Omega} d \mathbf{x} P(\mathbf{x},t) \ln \frac{P(\mathbf{x},t)}{P^s(\mathbf{x})} .
\label{FreeEnergy}
\end{equation}
Furthermore, after performing the separation into coarse- and fine-grained contributions we obtain
\begin{equation}
F(t) = \epsilon \sum_{i=1}^N P_i(t) \ln \frac{P_i(t)}{P_i^s}.
\label{fdecomp}
\end{equation}
Observe that, in this case, the fine-grained contribution is absent, the reason being that the corresponding terms in $U$ and $S$ cancel at the time of subtracting.
\subsection{Further thermodynamic significance underlying the scale separation}
In the coarse-grained perspective one can define
\begin{equation}
u_i = - \epsilon \ln P_i^s.
\label{ui}
\end{equation}
Hence, from Eq. (\ref{udecomp2})
\begin{equation}
U(t) = \sum_{i=1}^N P_i(t) (u_i + \tilde{s}_i).
\label{fi1}
\end{equation}
Since $U(t)$ is the average internal energy, we must have
\begin{equation}
U(t) = \sum_{i=1}^N P_i(t) \tilde{u}_i,
\label{fi2}
\end{equation}
with $\tilde{u}_i$ the mean internal energy associated to the basin $\omega_i$. Therefore, it follows from Eqs. (\ref{fi1}) and (\ref{fi2}) that
\[
u_i = \tilde{u}_i - \tilde{s}_i.
\]
We see from this last expression, and the fact that $\tilde{s}_i$ is an entropic term (\ref{tildes}), that $u_i$ takes the form of a conditional free energy. Finally, we have from (\ref{pi}), (\ref{udef}), and (\ref{ui}) that
\[
u_i = -\epsilon \ln \left( \int_{\omega_i} d \mathbf{x} P^s(\mathbf{x},t) \right)= -\epsilon \ln \left( \int_{\omega_i} d \mathbf{x} e^{- u(\mathbf{x})/k_B T} \right).
\]
in agreement with Eq. (\ref{eq_03}).
\section{Time evolution and thermodynamic process variables}
\subsection{General case}
By differentiating Eqs. (\ref{udecomp2}), (\ref{sdecomp}), and (\ref{fdecomp}) and making use of Eq. (\ref{me}) we obtain the following expressions for $\dot{U}$, $\dot{S}$, and $\dot{F}$:
\begin{eqnarray}
\dot{U} & = & \frac{\epsilon}{2} \sum_{i,j=1}^N (P_j \gamma_{ji} - P_i \gamma_{ij}) \left( \ln \frac{P_j^s}{P_i^s} - \frac{\tilde{s}_j - \tilde{s}_i}{k_B} \right), \label{dotu} \\
\dot{S} & = & \frac{\epsilon}{2} \sum_{i,j=1}^N (P_j \gamma_{ji} - P_i \gamma_{ij}) \left( \ln \frac{P_j}{P_i} - \frac{\tilde{s}_j - \tilde{s}_i}{k_B} \right), \label{dots} \\
\dot{F} & = & \frac{\epsilon}{2} \sum_{i,j=1}^N (P_j \gamma_{ji} - P_i \gamma_{ij}) \ln \frac{P_j^s P_i}{P_i^s P_j}, \label{dotf}
\end{eqnarray}
Before proceeding any further, notice that the formulas for $\dot{U}$ and $\dot{S}$ possess both coarse- and fine-grained terms. Nonetheless, because of the adiabatic approximation, the fine-grained terms in $\dot{U}$ and $\dot{S}$ are equal. Hence, they cancel in $U - S$ and, in consequence, the time derivative for the free energy ($\dot{F}$) is the same no matter wether the system has a fine-grained structure or not \citep{Qian:2001fk,Ge:2010fk}.
Following a procedure completely analogous to that in \citep{Ge:2010fk} we introduce the following definitions for the entropy production rate ($e_p$), the heat dissipation rate ($Q_d$), and the housekeeping heat ($Q_{hk}$):
\begin{eqnarray}
e_p &=& \frac{\epsilon}{2} \sum_{i,j=1}^N (P_j \gamma_{ji} - P_i \gamma_{ij}) \ln \frac{P_j \gamma_{ji}}{P_i \gamma_{ij}}, \label{sigma} \\
Q_d &=& \frac{\epsilon}{2} \sum_{i,j=1}^N (P_j \gamma_{ji} - P_i \gamma_{ij}) \left(\ln \frac{\gamma_{ji}}{\gamma_{ij}} + \frac{\tilde{s}_j - \tilde{s}_i}{k_B} \right), \label{qd} \\
Q_{hk} &=& \frac{\epsilon}{2} \sum_{i,j=1}^N (P_j \gamma_{ji} - P_i \gamma_{ij}) \ln \frac{P_j^s \gamma_{ji}}{P_i^s \gamma_{ij}}. \label{qhk}
\end{eqnarray}
It is straightforward to prove from the definitions above and Eqs. (\ref{dotu})-(\ref{dotf}) that
\begin{equation}
\dot{U} = Q_{hk} - Q_d, \quad
\dot{S} = e_p - Q_d, \quad
\dot{F} = Q_{hk} - e_p.
\label{flows}
\end{equation}
As discussed elsewhere \citep{Oono:1998ly,Ge:2010fk}, and mentioned above, $e_p$ represents the entropy production rate of the system, $Q_d$ the heat dissipation rate, and $Q_{hk}$ the energy influx rate necessary to keep the stationary distribution away from thermodynamic equilibrium.
Given that the stationary distribution satisfies (Eq. \ref{me})
\[
\sum_{j=1}^N (P_j^s \gamma_{ji} - P_i^s \gamma_{ij}) = 0,
\]
it is not hard to prove, see \citep{Santillan:2011vn} for a detailed demonstration, that
\[
\sum_{i,j=1}^N (P_j \gamma_{ji} - P_i \gamma_{ij}) \ln \frac{P_j^s }{P_i^s} = 0
\quad \text{and} \quad \sum_{i,j=1}^N (P_j \gamma_{ji} - P_i \gamma_{ij}) (\tilde{s}_j - \tilde{s}_i) = 0.
\]
These results further mean that, in the steady state,
\begin{equation}
e_p = Q_d = Q_{hk} = \frac{\epsilon}{2} \sum_{i,j=1}^N (P_j^s \gamma_{ji} - P_i^s \gamma_{ij}) \ln \frac{\gamma_{ji}}{\gamma_{ij}} > 0.
\label{fluxessd}
\end{equation}
That is, all fluxes are larger than zero, but they balance in such a way that $\dot{U}, \dot{S}, \dot{F} = 0$ in the steady state.
We point out that only $Q_d$ possesses a fine grained term. However, the corresponding discussion is delayed to the next subsection, in connection with the imposed adiabatic approximation and detailed balance.
\subsection{Detailed balance}
In the particular case where the stationary distribution $P^s(\mathbf{x})$ complies with detailed balance, the probability flux is null ($\mathbf{J}^s = 0$) for every $\mathbf{x}$ \citep{Qian:2002fk,Kampen:2007kx,Risken:1996uq}. This, together with Eq. (\ref{sumcurr}), further implies that $J_{ij}^s=J_{ji}^s$ for all $i\neq j$. And so (Eq. \ref{linearflux}) that
\begin{equation}
\gamma_{ij} P_i^s = \gamma_{ji} P_j^s.
\label{detbal}
\end{equation}
Substitution of this last equation into Eqs. (\ref{sigma})-(\ref{qhk}) gives
\begin{eqnarray}
e_p &=& \frac{\epsilon}{2} \sum_{i,j=1}^N (P_j \gamma_{ji} - P_i \gamma_{ij}) \ln \frac{P_j P_i^s}{P_i P_j^s}, \label{sigmadb} \\
Q_d &=& \frac{\epsilon}{2} \sum_{i,j=1}^N (P_j \gamma_{ji} - P_i \gamma_{ij}) \left(\ln \frac{P_i^s}{P_j^s} + \frac{\tilde{s}_j - \tilde{s}_i}{k_B} \right), \label{qddb} \\
Q_{hk} &=& 0. \label{qhkdb}
\end{eqnarray}
It then follows from Eq. (\ref{flows}) that
\begin{equation}
\dot{U} = - Q_d, \quad
\dot{S} = e_p - Q_d, \quad
\dot{F} = - e_p.
\label{flowsdb}
\end{equation}
By substituting $P_i = P_i^s$ into Eqs. (\ref{sigmadb})-(\ref{qhkdb}) and taking into account Eq. (\ref{detbal}), we have that
\[
e_p = Q_d = Q_{hk} = 0.
\]
That is, when detailed balance is satisfied (or equivalently, when the system is in thermodynamic equilibrium), all state variables remain constant in time because all fluxes are null.
Consider again the adiabatic approximation. We can see from Eq. (\ref{qsdist}) that it is equivalent to assuming that the probability distribution immediately evolves, within each $\omega_i$, to a local quasi-stationary distribution compatible with thermodynamic equilibrium. This last fact explains why neither $e_p$ nor $Q_{kh}$ | Eqs. (\ref{sigma}) and (\ref{qhk}) | possess fine-grained terms.
\subsection{Emergent coordinate and mean-field approximation}
The coordinate system of the phase space of a given stochastic dynamics, $(x,y)$, usually is not the most natural one in terms of the multiscale dynamics. Fig. \ref{fig:1} illustrates how a dynamically natural coordinate system $(r,s)$ can emerge from slow and fast manifolds. The slow manifold is widely known in chemical reaction dynamics as the ``reaction coordinate''. The potential of mean force along the slow manifold, $A(r)$, is widely called the ``energy landscape''.
\begin{figure}[htb]
\includegraphics[width=2in]{SlowFastMF.pdf}
\caption{Schematic representation of how a natural coordinate system $(r,s)$ can emerge
from the slow and fast manifolds of a given dynamical system originally described in the phase space $(x,y)$.}
\label{fig:1}
\end{figure}
In terms of the emergent dynamic coordinates $(r,s)$, the partition function is
\begin{equation}
Z(\epsilon) = \int dx \int dy\ e^{-V(x,y)/\varepsilon}
= \int dr\ e^{-A_r(r)/\epsilon},
\label{eq_42}
\end{equation}
in which
\begin{equation}
A_r(r) = -\epsilon\ln \int ds\
\Big\|\frac{D(x,y)}{D(r,s)}\Big\|\
e^{-\widetilde{V}(r,s)/\epsilon},
\end{equation}
with $\widetilde{V}(r,s)=V\left(x(r,s),y(r,s)\right)$. Furthermore $(r,s)$ is a coordinate transformation of $(x,y)$: $x=x(r,s),y=y(r,s)$, with a non-singular Jacobian$ \big\|{D(x,y)} / {D(r,s)}\big\|\ne 0$.
How to discover the dynamically natural slow coordinate? One of the most widely used approaches is the {\em mean field method}. To illustrate this approach, consider the conditional free energy
\begin{equation}
A_x(x) = -\epsilon\ln\int dy\ e^{-V(x,y)/\epsilon},
\end{equation}
and also the conditional mean value for $y$, $\langle y\rangle_x = E\left[y|x\right]$:
\begin{equation}
\langle y\rangle_x =
\frac{\int dy y\ e^{-V(x,y)/\epsilon}}
{\int dy e^{-V(x,y)/\epsilon}}
= e^{A_x(y) / \epsilon}\int dy y\ e^{-A_x(y)/\epsilon} .
\end{equation}
The curve $\langle y\rangle_x$ can be considered as an emergent reaction coordinate. Then, using $x$ as a parameter, $A_x(x)$ and $\langle y\rangle_x$ give an ``energy function'' along the reaction coodinate. One can in fact choose a new coordinate $r$ along the curve $y=\langle y\rangle_x$.
It is easy to verify that (see Eq. \ref{eq_42}):
\begin{equation}
\int\ dx \; e^{-A_x(x)/\epsilon} = \int dr \;
e^{-A_r(r)/\epsilon} = Z(\epsilon).
\label{eq_46}
\end{equation}
All the equations so far are exact. However, in studies of real chemical and biophysical problems, one often chooses not to compute the last integral in (\ref{eq_46}). Rather, one finds the local or global minima of $A_r(r)$. The reasons for this practice are twofold:
\begin{itemize}
\item First, it is often analytically impossible
to carry out the integration. In this case, finding
the global minimum ($r^*$) is a reasonable approximation,
especailly for small $\epsilon$ \cite{bender_book}:
\begin{equation}
-\epsilon\ln \int dr e^{-A_r(r)/\epsilon} =
A_r\left(r^*\right) -\frac{\epsilon}{2} \ln
\left(\frac{2\pi \epsilon}{A_r''(r^*)}\right)
+ \cdots.
\end{equation}
Since the approximation neglects the fluctuations in $r$ around $r^*$, the method is widely called {\em mean field approximation.} In applied mathematics, this is known as Laplace's method for integrals \cite{bender_book}.
\item Second, $A_r(r)$ might have multiple minima, say two. In that case, carrying out the integration is not as insightful as to identifying the bistability of the system, and the associated transitions. They can be visualized by the potential of mean force
$A_r(r)$. In that case, a slow, emergent stochastic dynamics on $A_r(r)$ arises. Both Flory-Huggins theory of polymer solutions \cite{fht_book} and the Bragg-Williams approximation for nonequilibrium steady-stat \cite{bwt_book} are successful
examples.
\end{itemize}
\section{Concluding remarks}
In this work we have studied the thermodynamic consistency, or invariance, across scales of a continuous-state continuous-time system, undergoing a Markovian stochastic process. In particular, we tackled the question of how the system thermodynamic variables, as well as the relations among them, transform when the system is described, in a coarse-grained fashion, by means of discrete variables. In that respect, we proved that, in the Helmholtz free-energy perspective, the \emph{thermodynamics} derived from the continuous underlying detailed dynamics is \emph{exact}. I.e. it is the same as if one only takes a middle-road and starts with a discrete description, with the transition rates $\gamma_{ij}$ either directly measured, or estimated by parameter fitting of experimental data. Below we further discuss some interesting consequences from these results.
\subsection{Energy and thermodynamics across scales}
Consider a stochastic dynamical system with two levels of descriptions: an upper coarse-grained level and and a lower refined level with well-separated dynamic time scales. Then, following the analysis in the present paper, one has $F_1 = U_1-S_1$ and $F_2=U_2-S_2$, where subscripts ``1'' and ``2'' denote upper and lower levels. Furthermore, $U_1<U_2$. Their difference is considered to be ``heat'' dissipated from the upper level to the lower level.
The relationship between the classical Newtonian mechanics with $S_1\approx 0$ and the molecular description of matter is an example. The energy difference $U_2-U_1$ is entropic; it can not be fully used to ``do work'' at the upper level. The dynamics on the fast time scale at the lower level is considered to be ``fluctuations'' for the upper level. In a spontaneous transient at the upper level, ${d(U_2-U_1)}/{dt}$ is the rate of the amount of energy being passed to the lower level. Energy conservation can only be understood from the description of the lowest level. Conversely, entropy is the concept required to characterize the changing $U$ across scales.
The thermodynamics across scales in stochastic dynamics, in particular the energy dissipation from a upper scale into a lower scale, has been a central unresolved issue in the theory of turbulence \cite{frisch_book}. Whether the newly developed thermodynamic framework of stochastic dynamics can shed some light on the problem remains to be seen.
\subsection{Coarse-graining as conditional probability}
In a recent study \cite{Santillan:2011vn} we have shown that the conditional free energy, which corresponds to the potential of mean force in continuous stochastic systems, plays an essential role in the invariance of mathematical irreversible thermodynamics of multiscale stochastic systems. Furthermore, in \cite{moy_qian_12}, we have proved that Legendre transforms between different thermodynamic potentials for different Gibbs ensembles can be derived in terms of conditional probability for a pair of random variables. As a matter of fact, one can consider coarse-graining as a special form of conditional probability. Treating $\{(\mathbf{x},i)|\mathbf{x}\in\mathbb{R}^M,1\le i\le N\}$ as a pair of random variables, the coase-graining means
\begin{equation}
f(\mathbf{x}|\ell) = \frac{\Pr\{ x\le\mathbf{x}\le x+d\mathbf{x}|i=\ell\}}
{d\mathbf{x}} =\left\{
\begin{array}{ccc}
{f_{\ell}(\mathbf{x})} / {P_{\ell}}
&& \mathbf{x}\in\omega_{\ell}, \\
0 && \mathbf{x}\notin\omega_{\ell},
\end{array} \right.
\end{equation}
in which $f_{\ell}(\mathbf{x})$ is the joint probability, $f(\mathbf{x}|\ell)$ is the conditional probability, and
\begin{equation}
P_{\ell} = \int_{\Omega} d \mathbf{x} f_{\ell}(\mathbf{x}) = \int_{\omega_{\ell}}d \mathbf{x}
f_{\ell}(\mathbf{x}).
\end{equation}
Then, the standard chain rule for free energy (i.e. relative entropy) \cite{cover_book},
\begin{equation}
\sum_{\ell=1}^N\int _{\Omega} d \mathbf{x} f_{\ell}(\mathbf{x})\ln
\frac{f_{\ell}(\mathbf{x})}{f_{\ell}^{s}(\mathbf{x})}
= \sum_{\ell=1}^N P_{\ell}\ln\frac{P_{\ell}}{P^s_{\ell}}
+ \sum_{\ell=1}^N P_{\ell}
\left(\int_{\Omega} d \mathbf{x} f(\mathbf{x}|\ell)\ln
\frac{f(\mathbf{x}|\ell)}{f^{s}(\mathbf{x}|\ell)}
\right),
\label{total_f}
\end{equation}
takes an interesting, equivalent form:
\begin{equation}
\sum_{\ell=1}^N P_{\ell}\ln\frac{P_{\ell}}{P^s_{\ell}}
+ \sum_{\ell=1}^N P_{\ell}
\left(\int_{\omega_{\ell}} d \mathbf{x} f(\mathbf{x}|\ell)\ln
\frac{f(\mathbf{x}|\ell)}{f^{s}(\mathbf{x}|\ell)}
\right),
\label{eq_51}
\end{equation}
in which
\[
\int_{\omega_{\ell}} d \mathbf{x} f(\mathbf{x}|\ell)\ln
\frac{f(\mathbf{x}|\ell)}{f^{s}(\mathbf{x}|\ell)}
\]
is the ``conditional free energy'' of the sub-system $\ell$. For subsystems with rapid steady state, this term is
zero. Thus, a system's total free energy (\ref{total_f}) is the free energy of the coase-grained system.
\acknowledgments
The authors are in debt with Prof. Eduardo S. Zeron for the proof in Appendix \ref{boundary}.
|
\section{Introduction}
The primordial gas after recombination was composed of neutral
hydrogen and helium atoms with traces of electrons, deuterium and
lithium, exposed to the radiation field of the cosmic
background radiation (hereafter CBR), and in adiabatic expansion. In these
conditions, only a limited number of gas-phase reactions
was possible. The first molecules started to form only
ten thousand years after the Big Bang, at a resdhift $z\approx 2000$,
reaching a freeze-out abundance at $z\approx 100$: about $10^{-6}$ for
H$_2$, followed by HD, H$_2^+$, HeH$^+$ and several less abundant species
(see Lepp et al.~2002 for a review).
Despite the simple chemical composition, the kinetics of elementary
processes in the early Universe is very complex, due to the presence of
a large number of quantum states for atoms and molecules involved in
the network processes, including the interaction with
the CBR. Detailed calculations of molecule formation
in the early Universe have been recently performed by
Hirata \& Padmanabhan~(2006, hereafter HP06), Puy \& Signore~(2007),
Glover \& Jappsen~(2007), Schleicher et al.~(2008,
hereafter S08), and Vonlanthen et al.~(2009,
hereafter V09). In addition to these
studies, comprehensive collections of analytic fits of chemical
reaction rates were also given by Anninos \& Norman~(1996), Stancil et
al.~(1998), Galli \& Palla~(1998, hereafter GP98) and,
more recently, by Glover \& Abel~(2008).
In studies of primordial chemistry, the rovibrational
levels manifold of the molecules has been usually
ignored, mainly because of the lack of state-to-state resolved cross
sections (or, equivalently, reaction rates) of the relevant chemical
processes. Early estimates of the vibrational distribution function
of H$_2$ attempted by means of physical
order-of-magnitude arguments (Khersonskii~1982) led to the conclusion
that the vibrationally excited states of H$_2$ are rapidly deactivated
through radiative decay. This is because the H-H$_2$ collision time is
much longer than the radiative lifetime of vibrational
levels, so that the population of the vibrational levels is essentially
determined by the Einstein coefficients and the probability of
formation of vibrationally excited $\mathrm{H_2}$ molecules. HP06 were
the first to compute in detail the vibrational level populations of
H$_2^+$. However, they did not consider collisional processes inducing
vibrational relaxation.
Because of these simplifications, chemical reaction rates are usually
evaluated under the hypothesis of local thermal equilibrium (hereafter
LTE). However, non-equilibrium distribution function can significantly
affect the results of such calculations, as shown explicitly by
Capitelli et al.~(2007) for the dissociative attachment process of
H$_2$. This evidence suggests that the evaluation of the exact
vibrational distribution function is necessary to model the plasma
kinetics, as pointed out by Lepp et.~(2002). The relatively simple
internal structure of atomic and molecular hydrogen, and its molecular
cation H$_2^+$, makes possible such a {\it
state-to-state} approach at least for these species.
In the present work, we follow for the first time the
chemical composition of the primordial gas, determining the population
of vibrational levels of H$_2$ and H$_2^+$, adopting a detailed network
of formation and destruction channels and relaxation
processes, such as radiative and atom/ion-molecule
collisions. We ignore any contribution due to subsequent astrophysical
processes that affect the chemical composition of the gas such as
primordial massive supernovae, which, besides providing heavy elements,
may also contribute to the formation of H$_2$
in the early Universe (Cherchneff \& Lilly~2008).
The paper is organized as follows: in Section~2 we discuss the relevant
vibrationally resolved reactions and we compute state-to-state reaction
rate coefficients; in Section~3 we describe our chemical network; in
Section~4 we present our results for the evolution of the vibrational
level populations of H$_2$ and H$_2^+$; finally, in Section~5 we
summarize our conclusions. In the Appendix, we give a
list of fitting formulae for specific rate coefficients
under LTE conditions.
\section{Vibrationally resolved reaction rates for $\mathrm{H_2}$ and
$\mathrm{H_2^+}$}
In a hydrogen plasma,
H$_2$ and H$_2^+$ form
by various collision processes, generally in vibrationally
excited states. For example, in the case of H$_2^+(v)$, the electron
impact ionization process produces
a known population of vibrational states
given by the corresponding Franck-Condon factors
for the transitions. For other formation processes,
H$_2^+(v)$ is created with different populations of vibrational
levels. Additionally, inelastic collisions with other plasma constituents
strongly affect the evolution and thermalization of vibrational
levels of H$_2$ and H$_2^+$.
This fact underlines the importance
of such a vibrationally resolved kinetics, since a priori assumptions and approximations could
fail in evaluating the vibrational level distribution.
Typically, collisional processes in a low-temperature plasma involve
all molecular degrees of freedom (electronic, vibrational and
rotational); however, in order to model the kinetics of vibrational
levels of H$_2$ and H$_2^+$, rotational equilibrium
has been assumed.
For processes not involving H$_2$ and H$_2^+$, we have adopted the
chemical network recently developed by S08, updating the value of the
rate coefficient for the H$^-$ photodetachment to include the
contribution of non-thermal photons due to cosmological H and He
recombination (see HP06). The corresponding rate
coefficient for this process is given in Table~\ref{fits}. For H$_2$ and H$_2^+$, we computed
state-to-state rate coefficients from available cross sections,
including all vibrational levels of their fundamental electronic states (15 and 19 states, respectively).
A list of the chemical processes included in our calculation is
shown in Table~\ref{vibsolvedchemprocess}. Each process will be
described separately in the following subsections.
\begin{table}
\caption{Vibrationally resolved reactions included in this work}
\begin{tabular*}{\columnwidth}{ll}
\hline
Chemical process & Reference \\
\hline
& \\
$1] ~~\mathrm{H}+\mathrm{H_2}(v) \rightarrow \mathrm{H}+\mathrm{H_2}(v^\prime)$ & (1) \\
$2] ~~\mathrm{H}+\mathrm{H_2}(v) \rightarrow \mathrm{H}+\mathrm{H}+\mathrm{H}$ & (1) \\
$3] ~~\mathrm{H}+\mathrm{H^-} \rightarrow \mathrm{H_2}(v) + \mathrm{e^-}$ & (2) \\
$4] ~~\mathrm{H}+\mathrm{H^+} \rightarrow \mathrm{H_2^+}(v)+h\nu$ & detailed balance from (6) \\
$5] ~~\mathrm{H_2^+}(v)+\mathrm{H} \rightarrow \mathrm{H_2}(v^\prime)+\mathrm{H^+}$ & (3) \\
$6] ~~\mathrm{H_2}(v)+\mathrm{H^+} \rightarrow \mathrm{H_2^+}(v^\prime)+\mathrm{H}$ & (3) \\
$7] ~~\mathrm{H_2^+}(v)+\mathrm{H} \rightarrow \mathrm{H_2^+}(v^\prime)+\mathrm{H}$ & (3) \\
$8] ~~\mathrm{H_2}(v)+\mathrm{H^+} \rightarrow \mathrm{H_2}(v^\prime)+\mathrm{H^+}$ & (3) \\
$9] ~~\mathrm{H_2^+}(v)+\mathrm{H} \rightarrow \mathrm{H^+}+\mathrm{H}+\mathrm{H}$ & (4) \\
$10] ~~\mathrm{H_2}(v)+\mathrm{H^+} \rightarrow \mathrm{H}+\mathrm{H}+\mathrm{H^+}$ & (4) \\
$11] ~~\mathrm{H_2}(v)+h\nu \rightarrow \mathrm{H_2^+}(v^\prime)+\mathrm{e^-}$ & (5) \\
$12] ~~\mathrm{H_2^+}(v) + h\nu \rightarrow \mathrm{H}+\mathrm{H^+}$ & (6) \\
$13] ~~\mathrm{H_2}(v)+h\nu \rightarrow \mathrm{H}+\mathrm{H}$ & direct + indirect (7) \\
$14] ~~\mathrm{H_2^+}(v)+\mathrm{e^-} \rightarrow \mathrm{H}+\mathrm{H}$ & (8) \\
$15] ~~\mathrm{H_2}(v)+\mathrm{e^-} \rightarrow \mathrm{H^-}+\mathrm{H}$ & (9) \\
$16] ~~\mathrm{H_2}(v) + \mathrm{e^-} \rightarrow \mathrm{H_2}(v^\prime)+\mathrm{e^-}+h\nu$ & (9) \\
$17] ~~\mathrm{H_2}(v) \rightarrow \mathrm{H_2}(v^\prime)+h\nu$ & (10) \\
$18] ~~\mathrm{H_2^+}(v) \rightarrow \mathrm{H_2^+}(v^\prime)+h\nu$ & (11) \\
$19] ~~\mathrm{HeH^+} + \mathrm{H} \rightarrow \mathrm{H_2^+}(v)+\mathrm{He}$ & (12) \\
& \\
\hline
\end{tabular*}
{References:
(1) Esposito et al.~(1999, 2001); (2) \u{C}\'{\i}\v{z}ek et al.~(1998);
(3) Krsti\'{c} et al.~(1999, 2002), Krsti\'{c}~(2002); (4) Krsti\'{c}~(2003); (5) Flannery et
al.~(1977); (6) Dunn~(1968); (7) Allison \& Dalgarno~(1969, 1970), Dalgarno
\& Stephens~(1970); (8) Takagi~(2002); (9) Celiberto et al.~(2001); (10)
Wolniewicz et al.~(1998); (11) Posen et al.~(1983); (12) Linder et al.~(1995) (see S08).}
\label{vibsolvedchemprocess}
\end{table}
\subsection{Associative detachment}
\label{subcizek}
The importance of the reaction of associative detachment for H$^-$,
\[
\mathrm{H}+\mathrm{H^-} \rightarrow \mathrm{H_2}(v)+\mathrm{e^-},
\]
was first pointed out by Dalgarno (quoted by Pagel~1959).
Various studies have been published on the vibrationally resolved cross
sections and/or rate coefficients for this process (Bieniek \&
Dalgarno~1979, Launay et al.~1991), underlying the importance of
this channel of H$_2$ formation (e.g. Dalgarno~2005;
Flower~2000).
Our calculations are based on the
detailed data by
\u{C}\'{\i}\v{z}ek et al.~(1998). These authors developed an improved
nonlocal resonance model for the description of the nuclear dynamics of
the H$_2^-$ collision complex, and calculated both associative
detachment and dissociative attachment cross sections. They provide
data for different values of the angular momentum $J$ of the H$^-$+H
collision\footnote{Data in electronic format are available at {\tt
http://utf.mff.cuni.cz/$^\sim$cizek/AD$\_$H2/}.}. For each initial
angular momentum, the H$_2$ molecule can be produced
in states with the
rotational quantum number equal to $J+1$ or $J-1$ due to selection
rules; states with the same $J$ are not allowed by symmetry conditions
on the total wavefunction and higher changes in $J$ are neglected in
their model. The total rate coefficient, obtained as a sum of the
vibrationally resolved rate coefficients,
\beq
k_{\rm tot}(T_{\rm gas})=\sum_v k_v(T_{\rm gas}),
\label{ad_tot}
\eeq
is shown in Fig.~\ref{cizekcomp}, compared with
the rate coefficient computed by Launay et al.~(1991),
as fitted by GP98. The new fit is given in Table~\ref{fits}.
Using this rate coefficient in a
non-vibrationally resolved chemical network of the primordial Universe
has a negligible impact on the H$_2$ abundance. Indeed, this reaction is also one of the most important H$^-$ loss channel.
Therefore increasing the associative detachment rate coefficient cannot lead to a production rate of H$_2$ faster than the H$^-$
production rate. For the same reasons,
H$^-$ fraction is reduced accordingly, by a factor of 4.
\begin{figure}
\includegraphics[width=6cm,angle=-90]{f1.eps}
\caption{H--H$^-$ associative detachment rate coefficient.
{\it Triangles}, total rate coefficient obtained from the cross sections of
\u{C}\'{\i}\v{z}ek et al.~(1998), according to Eq.~(\ref{ad_tot}); {\it
dashed curve}, fit by GP98 of the total rate coefficient from Launay et al.~(1991); {\it
solid curve}, fit reported in Table~\ref{fits}.}
\label{cizekcomp}
\end{figure}
\subsection{Photodissociation of $\mathrm{H_2^+}(v)$}
The photodissociation of H$_2^+(v)$,
\[
\mathrm{H_2^+}(v)+h\nu\rightarrow \mathrm{H}+\mathrm{H^+},
\]
is a typical example of direct process of photodestruction. Lebedev et
al.~(2000, 2003) provided a significant analysis of bound-free
(including also photodissociation) and free-free transitions for
H$_2^+(v)$, limited however to sums on the rotational and vibrational
quantum numbers. In this work, we have adopted the vibrationally
resolved cross sections calculated by Dunn~(1968),
evaluated fixing the rotational quantum number $J=1$.
The comparison between the LTE rate coefficient obtained using the vibrationally resolved data by Dunn
and the fit by GP98 of data by
Argyros~(1974) for
$2500$~K$<T_{\rm rad}<26000$~K and by Stancil~(1994) for
$3150~$K$<T_{\rm rad}<25200$~K is shown in Fig.~\ref{h2pphotodissociation}.
\begin{figure}
\includegraphics[width=6cm,angle=-90]{f2.eps}
\caption{H$_2^+$ photodissociation rate coefficient. {\it Solid curve},
total rate in LTE, summed over all possible initial,
from the cross sections by Dunn~1968; {\it dashed curve},
fit by GP98 based on data by Argyros~(1974) and Stancil~(1994).}
\label{h2pphotodissociation}
\end{figure}
\subsection{Radiative association of $\mathrm{H}$ and $\mathrm{H^+}$}
\label{radasssection}
The reaction of
radiative association of H and H$^+$,
\[
\mathrm{H}+\mathrm{H^+} \rightarrow \mathrm{H_2^+}(v)+h\nu,
\]
has been the subject of several studies
(e.g., Ramaker \& Peek~1976, Shapiro \& Kang~1987,
Stancil et al.~1993). However, no vibrationally resolved data are
available in literature. For this reason, we have applied
the principle of detailed balance to the inverse H$_2^+(v)$ photodissociation
reaction,
\beq
k_v^{\rm rad. ass.}=
\frac{Z_{{\rm H}_2^+(v)}}{Z_{\rm H} Z_{{\rm H}^+}}
e^{h\nu_v/kT_{\rm rad}}\,k_v^{\rm ph.}
\label{rcradass}
\eeq
where $T_{\rm rad}$ is the temperature of the CBR,
$h\nu_v$ is the thermal threshold for the $v$-th level.
The partition functions $Z$ of reagents and product
are
\[
Z=\frac{(2\pi kT_{\rm gas})^{3/2}}{h^3}
\left\{ \begin{array}{l}
m_{\rm H}^{3/2} g_{\rm H} Q_{\rm H}~~~\mbox{for H},\\
m_{{\rm H}^+}^{3/2} g_{{\rm H}^+}~~~\mbox{for H$^+$},\\
m_{{\rm H}_2^+}^{3/2} g_{{\rm H}_2^+}Q_{{\rm H}_2^+}Z_{\rm rot}(v,T_{\rm gas})~\mbox{for H$_2^+(v)$},\\
\end{array} \right.
\]
where $g_{\rm H}$, $g_{{\rm H}^+}$ and $g_{{\rm H}_2^+}$ are the
multiplicities of the nuclear spin contributions for H, H$^+$ and
H$_2^+$, equal to $g_{\rm H}=g_{{\rm H}^+}=2\times (1/2)+1=2$ and
$g_{{\rm H}_2^+}=4$; $Q_{\rm H}$ and $Q_{{\rm H}_2^+}$ are the electron
spin degeneracies, equal to 2 both for H and H$_2^+$ under the
hypothesis that only their ground electronic states contribute to the
partition function and $Z_{\rm rot}(v,T_{\rm gas})$ is the rotational
partition function, that takes into account the rotational structure of
each vibrational level of the cation hydrogen molecule,
given by
\[
Z_{\rm rot}(v,T_{\rm gas})=\sum_J g_J e^{-(E_0-E_{vJ})/kT_{\rm gas}},
\]
where $E_{vJ}$ is the binding energy of the $(v,J)$-th level and $E_0$
the binding energy of the lowest rovibrational level.
The symbol $g_J$ represents the multiplicity of each electronic-rotational level,
\[
g_J=\left\{
\begin{array}{l}
\frac{1}{4}(2J+1) \qquad \mbox{for $J$ even}, \\
\frac{3}{4}(2J+1) \qquad \mbox{for $J$ odd}.
\end{array}
\right.
\]
A comparison between the present LTE calculation and the
fit by GP98 of data by Ramaker \& Peek~(1976) is shown
in Fig.~\ref{radassfigure}.
\begin{figure}
\includegraphics[width=6cm,angle=-90]{f3.eps}
\caption{H--H$^+$ radiative association rate coefficient. {\it Solid
curve}, total rate coefficient obtained using the detailed balance
principle from the photodissociation data by Dunn~(1968), according to
Eq.~(\ref{rcradass}); {\it dashed curve}, fit by GP98 of the total rate
coefficient from Ramaker \& Peek~(1976).}
\label{radassfigure}
\end{figure}
\subsection{Charge transfer, collisional excitation, and dissociation}
The H$^+$--H$_2$ and the H--H$_2^+$ systems represent one of the most
important collision complexes in the hydrogen plasma, called in short
the ${\rm H}_3^+$ collision system. In addition to elastic
collisions H$_2$--H$^+$ and H$_2^+$--H, the relevant reactions are:
({\it i}\/) charge transfer,
\[
\begin{array}{l}
\mathrm{H_2}(v)+\mathrm{H}^+\rightarrow \mathrm{H_2^+}(v^\prime)+\mathrm{H}~~({\it a}\/),\\
\mathrm{H_2^+}(v)+\mathrm{H}\rightarrow \mathrm{H_2}(v^\prime)+\mathrm{H}^+~~({\it b}\/),\\
\end{array}
\]
and ({\it ii}\/) collisional dissociation,
\[
\begin{array}{l}
\mathrm{H_2}(v)+\mathrm{H}^+\rightarrow \mathrm{H}+\mathrm{H}+\mathrm{H}^+~~({\it a}\/),\\
\mathrm{H_2^+}(v)+\mathrm{H}\rightarrow \mathrm{H}+\mathrm{H}+\mathrm{H}^+~~({\it b}\/).\\
\end{array}
\]
The most complete theoretical work describing these chemical reaction
pathways has been recently developed by Krsti\'{c} and collaborators
(Krsti\'c~2002, 2003, 2005; Krsti\'c et al.~1999, 2002) within the
\emph{infinite order sudden approximation} method (IOSA)\footnote{Data
in electronic form are available at {\tt
http://www-cfadc.phy.ornl.gov/astro/ps/data/home.html}.}. As described
by Krsti\'{c} et al.~(2002), charge transfer ({\it a}\/) is endoergic
for $v\le 3$, although with low threshold energies, and is strictly
competitive with the vibrational excitation channel, especially for
those states than can overcome the kinetic barrier ($v\ge 4$). For
higher vibrational levels, the process is exoergic, as the process of
charge transfer ({\it b}), for all $v$. The LTE rate
coefficient for the charge exchange reaction ({\it a}\/) is shown in
Fig.~\ref{cth2hp}, and compared with the fit by Savin et al. (2004) for H$_2(v=0)$. For the
reaction ({\it b}\/), cross-sections have been extrapolated at low
energies using a Langevin-type power law; the resulting rate
coefficients have been scaled in order to obtain an LTE value
corresponding to the experimental measurement by Karpas et al.~1979 at
$T_{gas}=300~$K. The rate coefficients for the collisional
dissociation reactions ({\it a}\/) and ({\t b}\/) are shown in
Fig.~\ref{dissh3+}. The fits are given in Table~\ref{fits}.
\begin{figure}
\includegraphics[width=9cm]{f4.eps}
\caption{H$_2$--H$^+$ charge transfer rate coefficient. {\it Triangles}, total rate in LTE, summed over all possible initial and final
states, from the cross sections of Krsti\'c~(2002) and Krsti\'c et
al.~(2002, 2003); {\it dashed curve}, fit by Savin et al.~(2004) for
H$_2(v=0)$; {\it
solid curve}, fit listed in Table~\ref{fits}.}
\label{cth2hp}
\end{figure}
\begin{figure}
\includegraphics[width=9cm]{f5.eps}
\caption{H$_2^+$--H ({\it filled triangles}) and H$_2$--H$^+$ ({\it
open triangles}) total dissociation rate coefficients in LTE, summed
over all possible initial and final states, from the cross sections of
Krsti\'c~(2002) and Krsti\'c et al.~(2002, 2003). The fits
listed in Table \ref{fits} are also shown as solid and
dashed lines, respectively.}
\label{dissh3+}
\end{figure}
\subsection{$\mathrm{H_2}/\mathrm{H}$
vibrational-translational transfer}
Elastic collisions among hydrogen atoms and molecules deeply modify the
initial vibrational distribution function of molecular components in a
hydrogen plasma. They are generally classified as V-T
(vibrational-translational) and V-V (vibrational-vibrational)
processes. In the former case, energy transfer between the
translational degree of freedom of the colliding atom and the
vibrational state of the target molecule occurs; the latter represents
the energy exchange between the vibrational manifolds of the colliding
molecules. In both cases, multi-quantum transitions can occur, coupling
the entire set of vibrational levels. Recently,
Krsti\'{c} et al.~(2002) have calculated V-T cross sections for
collisions between H and H$_2$,
\[
\mathrm{H}+\mathrm{H_2}(v) \rightarrow \mathrm{H}+\mathrm{H_2}(v^\prime),
\]
using the quantum mechanical IOSA method. However, while
data are available for the collision system
H$^+$--H$_2(v)$, no state-to-state information on collisions among
hydrogen atoms and molecules is available for all
$(v,v')$ pairs. The V-T rate coefficients used in the present work have
been obtained from cross sections computed by Esposito et
al.~(1999, 2001) using the \emph{quasiclassical
trajectory} (QCT) method.
Collisions of H$_2$ with He atoms have been studied by
e.g. Clark~(1997), Flower et al.~(1998), Balkrishnan et al.~(1999), Lee
et al.~(2005), but no data for each $(v,v^\prime)$
pair are available. In this work we have neglected V-T
processes with He. This assumption is based on the observation that
vibrating nuclei in H$_2$ are lighter than the colliding He atom;
moreover, the relative abundance of He is small: for these reasons,
the V-T transfer with He can be considered less
efficient than V-T transfer with H.
Also molecule-molecule collisions have been neglected,
owing to the small fractional abundance of H$_2$ with respect to H.
\subsection{Collisional dissociation of $\mathrm{H_2}$}
The cross sections for the dissociation of H$_2$ induced by collisions
with H,
\[
\mathrm{H}+\mathrm{H_2}(v)\rightarrow \mathrm{H}+\mathrm{H}+\mathrm{H},
\]
have been computed by Esposito et al.~(1999, 2001) with the QCT method.
The whole vibrational manifold of H$_2$ has been explored. The total
LTE rate coefficient is shown in Fig.~\ref{dissfab},
and the fit is given in Table~\ref{fits}.
\begin{figure}
\includegraphics[width=9cm]{f6.eps}
\caption{H$_2$ dissociation by H impact:
{\it Triangles}: rate coefficient computed from the cross sections of
Esposito et al.~(1999); {\it solid line}: fit reported in Table \ref{fits}.}
\label{dissfab}
\end{figure}
\subsection{Photoionization of $\mathrm{H_2}$}
Ionization of H$_2$ can proceed through collisions
with particles and photons. Data on ionization caused by electron
impact are described by Liu \&
Shemansky~(2004), and compared with photoionization data. Because
threshold energies for the electron-impact ionization cross sections
are much higher than thermal energies available at the redshift range
considered in the present work ($E>16$~eV vs. $E<2$~eV), we have
considered only photoionization of H$_2$,
\[
\mathrm{H_2}(v)+h\nu\rightarrow \mathrm{H_2^+}(v^\prime)+\mathrm{e^-}.
\]
Several theoretical calculations of H$_2$ photoionization have been
proposed during the years (Lewis Ford et al.~1975, ONeil \&
Reinhardt~1978). This chemical reaction couples each vibrational level
of H$_2$ with each vibrational level of H$_2^+$. Available data on
vibrationally resolved $(v,v^\prime)$ cross sections have been
published by Flannery et al.~(1977), following the theoretical and
computational procedures described by Tai \& Flannery~(1977).
Fig.~\ref{photoionization} shows the comparison between the total
photoionization rate coefficient calculated in the hypothesis of LTE
using the cross sections by Flannery et al.~(1977), whose fit is given in Table \ref{fits},
and the fit by GP98
based on data by ONeil \& Reinhardt~(1978).
\begin{figure}
\includegraphics[width=9cm]{f7.eps}
\caption{H$_2$ photoionization rate coefficient. {\it Triangles},
total rate in LTE, summed over all possible initial and final states,
from the cross sections by Flannery et al.~(1977); {\it dashed curve},
fit by GP98 based on data by ONeil \& Reinhardt~(1978); {\it solid curve}, fit given in Table \ref{fits}.}
\label{photoionization}
\end{figure}
\subsection{Photodissociation of $\mathrm{H_2}$}
Photodissociation of H$_2$,
\[
\mathrm{H_2}(v)+h\nu\rightarrow \mathrm{H}+\mathrm{H},
\]
can occur both directly (absorption from a lower level into the continuum)
or indirectly (the so-called ``Solomon process''), a process consisting in the
absortpion into an individual level of a bound upper state, followed by
dissociation. In the present work, we considered both
destruction pathways. For the direct process,
vibrationally resolved rate coefficients are given by
\beq
k_v^{\rm ph.}(T_{\rm rad})=4\pi\int_0^\infty \frac{\sigma_v^{\rm ph.}(\nu)}{h\nu}
J_{\nu}(T_{\rm rad})\,d\nu,
\label{rcrad}
\eeq
where the cross sections $\sigma_v^{\rm ph.}$ are taken from the work
by Allison \& Dalgarno~(1969). More recently, new calculations of
photodissociation cross sections have been published by
Glass-Maujean~(1986) and Zucker \& Eyler~(1986), including excited
electronic states; however, no significant changes have been introduced
for the Lyman and Werner systems. A thorough analysis of the direct
process with related fits and reaction rates for the kinetics of the
early Universe is given in Coppola et al.~(2011).
For the calculation of the rate coefficient of
the indirect process, we have used the discrete version of Eq.~(\ref{rcrad})
\beq
k_v^{\rm ph.}(T_{\rm rad})=\frac{\pi e^2}{m_{\rm e}c^2} \lambda_{uv}^2 f_{uv}
\epsilon_u J_{\lambda_{uv}}(T_{\rm rad}),
\label{rcrad2}
\eeq
where $u$ is the vibrational level of both Lyman and Werner
electronically excited intermediate systems, whose dissociation
efficiencies $\epsilon_u$ have been taken from Dalgarno \&
Stephens~(1970). The oscillator strengths $f_{uv}$ for the same
molecular states have been calculated by Allison \& Dalgarno~(1970),
and the vibrational energies of both Lyman and Werner states by Fantz
\& W\"{u}nderlich~(2006). In Fig.~\ref{solomon} we show the
comparison between the fit of the rate coefficient by S08 of the
data by Glover \& Jappsen~(2007)
and the total LTE rate coefficient obtained using Eq.~(\ref{rcrad2}).
The present calculation includes both Lyman and Werner
systems, and the contribution of the total vibrational manifold has
been considered.
\begin{figure}
\includegraphics[width=9cm]{f8.eps}
\caption{H$_2$ indirect photodissociation rate coefficient. {\it Triangles},
total rate in LTE, both considering the Lyman and the Werner system, obtained using Eq.~(\ref{rcrad2}) and
data by Dalgarno \&
Stephens~(1970) and Allison \& Dalgarno~(1970); {\it dashed curve},
fit by S08 on Glover \& Jappsen~(2007) data; {\it solid curve}, fit given in Table \ref{fits}.}
\label{solomon}
\end{figure}
\subsection{Dissociative excitation and dissociative recombination of
$\mathrm{H_2^+}$}
The fate of the dissociative scattering processes between electrons and
H$_2^+$ strictly depends on the energy of the collision.
The effectiveness of these chemical reactions is connected to the energy
thresholds of the vibrational states of the intermediate excited states:
at high collision energies, the favored dissociative channel
is dissociative excitation (DE):
\[
\mathrm{H_2^+}(v)+\mathrm{e}^-\rightarrow
\left\{ \begin{array}{l}
\mathrm{H_2^+}(2p\sigma_u)+\mathrm{e}^- \\
\mathrm{H_2^+}(2p\pi_u)+\mathrm{e}^- \\
\end{array}
\right\}
\rightarrow
\mathrm{H}+\mathrm{H}^++\mathrm{e}^-.\\
\]
However, experimental evidence (Yousif \& Mitchell~1995) has shown that
large DE cross sections are operative at energies $\sim 0.01$~eV. For
this reason, other reaction pathways become important at
these energies, involving electron capture into
excited electronic states of H$_2$. In particular,
$\mathrm{H_2}(^1\Sigma^+_g)$ or auto-ionizing dissociative Rydberg
states lying below the $\mathrm{H_2^+}(2p\sigma_u)$, both auto-ionizing
in the continuum of $\mathrm{H_2^+}$, have been considered in the
dynamical mechanisms of dissociative excitation.
At low energies, DE processes are competitive with
dissociative recombination (DR), in which the incident electron is
temporally captured into the Rydberg state by transferring its kinetic
energy to vibrational motion. The unstable molecule can dissociate or go
back to the ionizing state by an elastic or vibrational
transition. This mechanism is sometimes called ``indirect process'' of
DR. In the direct process, excited hydrogen molecules
dissociate in the vibrational continuum of H$_2$. In this case,
two H atoms are produced, one in the ground electronic
state, the other in states with principal quantum number $n \geq 2$:
\[
\mathrm{H_2^+}(v)+\mathrm{e}^-\rightarrow
\mathrm{H}+\mathrm{H}(n\ge2).
\]
Different studies have focused on DE and/or DR of H$_+^2$ (e.g., Takagi~2002; Stroe \&
Fifirig~2009; Fifirig \& Stroe 2008; Motapon et al.~2008; Takagi et al.~2009).
We have used data from the work by Takagi~(2002), where
DR and DE cross sections have been calculated using the
\emph{multi-quantum defect theory} (MQDT). For each vibrational level,
Takagi~(2002) provided both DR and DE cross sections.
All excited dissociative states are included (all
Rydberg states of 5 different symmetries) and for the lowest
dissociative state, the electronic interaction is fully taken into
account. Fig.~\ref{takagifitcs} shows the comparison between the total
dissociative recombination rate coefficient calculated in the
hypothesis of LTE using the cross sections by Takagi~(2002), and the
fit by GP98 based on data by Schneider et al.~(1994). Our fit is given
in Table~\ref{fits}. Although the cross sections of this process have
a threshold component which corresponds to a repulsive channel, the
threshold feature is located at too high energy to affect results in
the present application. The main features relevant here are instead
the presence of many resonances at low energy which convolute
into a strongly decreasing trend for energies below a few eV, and a
global increase of this low energy portion of the cross sections with
$v$. As a result, the LTE rate coefficient decreases with
$T_{\rm gas}$ until a few thousand K, where the
raising population of vibrational levels reverses this trend. The fit
by GP98, reported in Fig.~\ref{takagifitcs} for reference,
was based on data by Schneider et al.~(1994), given
up to $T_{\rm gas}=4000$~K and
therefore showed a monotonic trend.
\begin{figure}
\includegraphics[width=6cm,angle=90]{f9.eps}
\caption{H$_2^+$ dissociative recombination rate coefficient. {\it
Triangles}, total rate in LTE, summed over all initial states, from the
cross sections by Takagi ~(2002); {\it dashed curve}, fit by GP98 based
on data by Schneider et al.~(1994), given in the temperature range
20~K$<T_{\rm gas}<4000$~K; {\it solid curve}, fit
reported in Table \ref{fits}.}
\label{takagifitcs}
\end{figure}
\subsection{Dissociative attachment}
Electron-impact inelastic processes of vibrationally excited
H$_2$ molecules,
\[
\mathrm{H_2}(v)+\mathrm{e}^-\rightarrow \mathrm{H}+\mathrm{H^-},
\]
play an important role in the kinetics of a low-temperature hydrogen
plasma. Especially for inelastic collisions involving an electronic
transition, two important features can be noted:
if a vibrationally excited molecule is involved, the threshold
decreases and the cross section of the process
increases (for some processes dramatically) with increasing vibrational
excitation of the molecule. In this work we have used
data calculated by Celiberto et al.~(2001), based on an improved
version of the resonant scattering model originally developed by
Fano~(1961). The dynamics of this route provides for the formation of a
temporary negative molecular ion, whose fate is either ionization or
dissociation. The fit of the LTE rate coefficient can be found in
Capitelli et al.~(2007).
\subsection{Electron collisional excitation of $\mathrm{H_2}$}
The process of vibrational excitation of molecules by
electron collisions,
\[
\mathrm{H_2}(v)+\mathrm{e^-} \rightarrow \mathrm{H_2}(v^\prime)+\mathrm{e^-}+h\nu,
\]
also known as the E-V process, selectively populates high
vibrational levels and represents one of the main non-equilibrium pathways
in the vibrational kinetics of hydrogen plasma. It is a two-step process
that links a vibrational state of the electronic ground state to another
vibrational state of the same manifold, via an intermediate singlet
state that can radiatively decay. As in the case of
dissociative attachment, we have used data from Celiberto
et al.~(2001).
\subsection{Formation of $\mathrm{H_2^+}$ via $\mathrm{HeH^+}$}
The $\mathrm{HeH^+}$ channel for the formation of $\mathrm{H_2^+}(v)$,
\[
\mathrm{HeH^+} + \mathrm{H} \rightarrow \mathrm{H_2^+}(v)+\mathrm{He},
\]
has been considered in several studies (e.g. Dalgarno~2005, HP06,
S08). The rate coefficient usually adopted in these kinetics models
is estimated from the experimental work by Karpas et
al.~(1979) or data by Linder et al.~(1995). However,
no vibrationally resolved information is available. In order
to overcome this limitation, we follow here an approach
similar to that of Hirata \& Padmanabhan~(2006), where this reaction is
assumed to equally populate all levels compatible with energy balance,
i.e. all levels with energy $E<-1.844$~eV with respect to the H$_2^+$
dissociation limits. Since in the present work individual
rovibrational levels are not considered, we follow
a simplified approach by assuming that this reaction channel produces
H$_2^+$ molecules in the first three vibrational levels with equal rates.
\subsection{Radiative transitions of $\mathrm{H_2}$}
Quadrupolar transition probabilities for
$\mathrm{H_2}(v)$ have been calculated by Wolniewicz et
al.~(1998)\footnote{Data in electronic format are avalaible at
{\tt http://cfa-www.harvard.edu/$^\sim$simbotin/4pole.html}.}.
The authors improved earlier calculations by Turner
et al.~(1977) by adopting a more accurate potential
energy curve, providing tables for the spontaneous decay from an
initial to a final vibrational level for $\Delta
J=0,\pm 2$. The rovibrational energy levels used to perform the sum
of available transition rates have been calculated
using the WKB method (Esposito~2010, priv. comm.). They are in very
good agreement with classical calculations by Kolos \&
Wolniewicz~(1964).
\subsection{Radiative transitions of $\mathrm{H_2^+}$}
Calculations of quadrupole transition probabilities for H$_2^+(v)$
follow the method used by Wolniewicz et al.~(1998)
for H$_2$. Radiative transitions for H$_2^+$
have been calculated by Posen et al.~(1983) for
vibrational-vibrational transitions at given $J$ and $\Delta J$.
The energies of the rovibrational levels used to perform
the sum have been calculated by Hunter et al.~(1974).
\section{Chemical network}
\label{bozomath}
To describe the time evolution of the chemical species during the
expansion of the Universe, we solve numerically
the system of ODEs
\beq
\begin{split}
\frac{dn_v}{dt}=&-n_v\sum_{v'}(R_{vv'}+P_{vv'}+C_{vv'}n_{v'})+\\
&+\sum_{v'}R_{vv'}n_{v'}+\sum_{v'}\sum_{v''}\mathbf{C}^{v'v''}_vn_{v'}n_{v''}. \\
\end{split}
\label{ode}
\eeq
where $R_{vv^\prime}$ are the spontaneous and stimulated
radiative rate coefficients; $P_{vv^\prime}$ are the
destruction rate coefficients of the $v$-th species by photons; $C_{vv^\prime}$ are the destruction rate
coefficients for the $v$-th species for collisions
with the $v^\prime$-th chemical partner; and ${\bf
C}^{v^\prime v^{\prime\prime}}_v$ are the formation
rate coefficients of the $v$-th species due to collisions between the
$v^\prime$-th and $v^{\prime\prime}$-th species, photons included.
The radiative rate coefficients are
related to the Einstein coefficients by
\[
\begin{split}
R_{vv^\prime} = & \left\{
\begin{array}{ll}
A_{vv^\prime}B_{vv^\prime}u(\nu_{vv^\prime},T_{\rm rad}) & \mbox{if $v^\prime<v$} \\
B_{vv^\prime}u(\nu_{vv^\prime},T_{\rm rad}) & \mbox{if $v^\prime>v$,} \\
\end{array} \right. \\
= & \left\{
\begin{array}{ll}
A_{vv^\prime}[1+\eta(\nu_{vv^\prime},T_{\rm rad})] & \mbox{if $v^\prime<v$} \\
g_{v^\prime}A_{v^\prime v}\eta(\nu_{vv^\prime},T_{\rm rad})/g_v & \mbox{if $v^\prime>v$} \\
\end{array} \right.
\end{split}
\]
where $\nu_{vv^\prime}$ is the frequency of the transition $v\rightarrow
v^\prime$, $g_v=1$ is statistical weight of the $v$-th level,
$u(\nu_{vv^\prime},T_{\rm rad})$ is the Planck photon distribution,
and $\eta(\nu_{vv^\prime},T_{\rm rad})=[\exp(h\nu_{vv^\prime}/kT_{\rm
rad})-1]^{-1}$.
The chemical network is completed by the
equations for the gas and radiation temperature evolutions
and the equation for the redshift,
\beq
\frac{dt}{dz}=-\frac{1}{(1+z)H(z)},
\eeq
where
\beq
H(z)=H_0 \sqrt{\Omega_{\rm r}(1+z)^4+\Omega_{\rm m}(1+z)^3+\Omega_{\rm K}(1+z)^2+\Omega_\Lambda}.
\eeq
The number density of hydrogen atoms is
\beq
n_{\rm H}=\Omega_b \frac{3 H_0^2}{8\pi G m_{\rm H}}(1-Y_p)(1+z)^3,
\label{nb}
\eeq
where $\Omega_b$ is the baryon fraction, $Y_p$ is the helium mass
fraction, $G$ is the constant of gravitation and $m_{\rm H}$ the atomic hydrogen mass.
Numerical values of the cosmological parameters $H_0$, $T_0$, $\Omega_r$,
$\Omega_m$, $\Omega_b$, $\Omega_\Lambda$ and $Y_p$ have been
obtained
from WMAP5 data (Komatsu et al.~2009, see Table~\ref{tablecosmo}). The initial fractional abundances for
H, He and D are also listed in the same table. The hydrogen
ionization fraction was computed using the routine RECFAST (Seager et
al.~1999, Seager et al.~2000, Wong et al.~2008). Helium
was assumed to be fully neutral in the redshift range of interest
($z\le 2000$). The electron density was determined by imposing
the condition of charge neutrality. Atomic and molecular weights
were taken from the NIST webpage \footnote{\tt
http://webbook.nist.gov/chemistry/name-ser.html}.
\begin{table}[ht]
\caption{Cosmological parameters}
\begin{tabular*}{\columnwidth}{ll}
\hline
Cosmological parameter & Numerical value\\
\hline
$H_0$ & $100\,h$~km~s$^{-1}$~Mpc$^{-1}$ \\
$h$ & $0.705$\\
$z_{\rm eq}$ & $3141$\\
$T_0$ & $2.725$~K\\
$\Omega_{\rm dm}$ & $0.228$\\
$\Omega_{\rm b}$ & $0.0456$\\
$\Omega_{\rm m}$ & $\Omega_{\rm dm}+\Omega_{\rm b}$\\
$\Omega_{\rm r}$ & $\Omega_{m}/(1+z_{\rm eq})$\\
$\Omega_\Lambda$ & $0.726$\\
$\Omega_{\rm K}$ & $1-\Omega_{\rm r}-\Omega_{\rm m}-\Omega_\Lambda$ \\
$Y_p$ & $0.24$\\
$f_{\rm H}$ & 0.924\\
$f_{\rm He}$ & 0.076\\
$f_{\rm D}$ & $2.38\times 10^{-5}$\\
& \\
\hline
\end{tabular*}
\label{tablecosmo}
\end{table}
Data tables of vibrationally resolved rate coefficients
have been inserted as input of the routine LSODE \footnote{\tt
https://computation.llnl.gov/casc/odepack/download/lsode\_agree.html}
used to integrate the chemical system, which consists of 47 differential
equations, one for each chemical species introduced in the model and one
for the gas temperature. An implicit method has been used to perform
the integration, being the kinetic problem stiff. Linear interpolation
of the rate coefficients and of the ionization fraction is performed in
logarithmic scale at each step of integration.
\section{Results and discussion}
\begin{figure}
\includegraphics[width=8cm]{f10.eps}
\caption{Vibrational level populations of H$_2$ ({\it top panel}, from
$v=0$ to $v=14$) and H$_2^+$ ({\it bottom panel}, from $v=0$ to $v=18$)
as function of redshift. In both cases, the {\it solid curve} shows
the total fractional abundance.}
\label{case0}
\end{figure}
The resulting fractional abundances of $\mathrm{H_2}$ and
$\mathrm{H_2^+}$ for each vibrational level are shown as
a function of redshift in Fig.~\ref{case0}. The total
fractional abundance of H$_2$ shows a rapid increase at three epochs:
at $z\approx 1500$ by the charge transfer channel,
dominant at high temperatures, at $z\approx 300$
by H$_2^+$ radiative association formation, that modulates the
charge transfer channel, and at $z\approx 100$ by the
associative detachment process.
At $z=10$ the fractional abundances of
H$_2$ and H$_2^+$ are $5.76\times 10^{-7}$ and $6.56\times 10^{-14}$, respectively.
Our value of the H$_2$ abundance is in good agreement with that obtained by
HP06 ($f({\rm H}_2)=6.0\times 10^{-7}$ at $z=20$) but about twice the value
of V09 ($2.7\times 10^{-7}$ at $z=10$).
As for H$_2^+$, our abundance is in good agreement with the result of V09
($f({\rm H}_2^+)=6.7\times 10^{-14}$ at $z=10$), whereas HP06 obtain three
different abundances depending on the value of the H$_2^+$--H charge exchange reaction.
For H$_2$, a marked non-equilibrium distribution of
populations is established at $z\approx 300$, followed
by a plateau involving levels from $v=1$ to $9$. Levels above $v=9$
are excluded from this last plateau as their populations
drop dramatically due to the endoergic character of
the process of associative detachment (see Section~\ref{subcizek}).
Therefore, when the thermal kinetic energy is much lower than the energy
thresholds, the formation of highly excited ($v\geq10$) $\mathrm{H_2}$
molecules is strongly suppressed. Such a strong splitting of level
histories is not observed at higher $z$, since the main H$_2$ formation
mechanism ($\mathrm{H+H_2^+}$ charge transfer) is exoergic for all $v$
(in constrast with the charge transfer $\mathrm{H^++H_2}$, that is a
threshold process).
The $\mathrm{H_2}$ vibrational distribution functions obtained at
different values of the redshift are shown in Fig.~\ref{vdfcase0},
compared with the equilibrium curves corresponding to the
Boltzmann vibrational distributions at the corresponding
value of the gas temperatures. The shape of the level population
distribution of H$_2$ at low $z$ can be understood and even roughly
evaluated on the basis of a balance between radiative and formation rates.
Indeed, the bunching of population of excited levels for low $z$ is due to
the fact that each population is determined essentially from the ratio
of associative detachment and radiative rates.
The ratios for all exothermic channels ($v<$10) have comparable values
on a enlarged logarithmic scale like that used here.
\begin{figure}
\includegraphics[width=8cm]{f11.eps}
\caption{Vibrational level populations of H$_2$, normalized to the
total H$_2$ fractional abundance, at various redshifts. The Boltzmann
distributions at each redshift are indicated with the label ``B''.}
\label{vdfcase0}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{f12.eps}
\caption{Same as Fig.~\ref{vdfcase0} for H$_2^+$.}
\label{vdfh2pcase0}
\end{figure}
For H$_2^+$, the most relevant formation process is radiative
association. The vibrational resolution of this species shows that
the formation process, being faster for highly excited vibrational
levels of the products, leads to a pronounced tail in the vibrational
distribution (Fig. \ref{vdfh2pcase0}). The latter is characterized by a
high population of the first few vibrational levels followed by a
suprathermal tail, in qualitative agreement with the
results of HP06 (see their Fig.~4). However, for HP06 the dominant
formation process of H$_2^+$ is via HeH$^+$. This
result supports the claim of HP06 that the radiative
association channel plays a minor role among formation pathways in their
model. Also in this case, we observe a strong grouping of the fractional
abundance of excited levels which is explained by the balance between
state specific level production rates and radiative
transitions to $v=0$.
These findings suggest that the processes leading to redistribution
of the vibrational quanta need further attention. This study is also appropriate
here because such processes cannot be modeled in the
framework of the usual, not state-resolved approach. In our case,
such redistributing channels are V-T and spontaneous/stimulated
radiative processes. In order to understand how these pathways effect
the vibrational distribution of H$_2$ and H$_2^+$, further numerical
experiments have been performed in various regimes,
corresponding to particular physical conditions. Fig.~\ref{case1}
shows the results obtained for the case in which radiative processes
have been omitted. The figure shows that, although
radiative processes are not fast enough to thermalize level populations,
they are essential to produce the actual vibrational distribution,
and their neglect leads to complete different results
especially at low $z$.
On the other hand, removal of V-T processes does
not produce any appreciable variation with respect to the results of
Fig.~\ref{case0}, since the rates of V-T processes
are much smaller than the radiative ones. The role of V-T processes
is better appreciated by looking at Fig.~\ref{case3}, where
V-T processes and radiative decay are both ignored. Some variations with respect to Fig.~\ref{case1} are observed: a higher
vibrational temperature, a different high $v$ plateau for H$_2$,
especially at low $z$, and some differences of the total H$_2^+$
fraction.
The main destruction channel (dissociative
recombination) active at lower $z$ is deeply affected by radiative
processes: when these are ignored,
the destruction channel becomes more efficient and the
fractional abundance at the second peak (at $z \sim 300$) is reduced by
about a factor of 2. As a further
check, removing radiative processes together with the loss channel leads
to small variations of the fractional abundance.
Thus, redistribution among levels is able
to affect even the results for the total $\mathrm{H_2^+}$, confirming
the relevance of the non-equilibrium vibrational distributions.
No rotationally resolved data exists for most of the formation and destruction processes included
in the present work, so a state-to-state kinetics cannot at the present stage go beyond the resolution
of molecular vibrations. For H$_2$ and H$_2^+$, the typical rotational energies are of the order of 0.01 eV:
this suggests that the way in which rotation is included in the kinetic model can affect the
low $z$ trend of molecular abundance.
To test the sensitivity of the chemical abundances to the rotational structure of H$_2$ and H$_2^+$,
we have computed the H$_2^+$ radiative association rate coefficient (Subsection \ref{radasssection}) assuming that
the rotational manifold for each vibrational level reduces to only one J, either the lowest or the highest.
Even if this hypothesis implies a very strong rotational
non-equilibrium, it has negligible effect on the freeze-out
value of H$_2^+$, although the abundance of this species at $z \simeq 1000$ is changed. Furthermore,
since the photodissociation
cross section is not rotationally resolved, the rotational degree of freedom only enters in Eq.(\ref{rcradass}) through $Z_{rot}$.
As a consequence, the nascent vibrational distribution function, which is the most important quantity here, is only marginally modified.
\begin{figure}
\includegraphics[width=8cm]{f13.eps}
\caption{Same as Fig.~\ref{case0}, including formation
and destruction channels, V-T processes but no radiative transition.}
\label{case1}
\end{figure}
\begin{figure}
\includegraphics[width=8cm]{f14.eps}
\caption{Same as Fig.~\ref{case0}, including formation
and destruction channels, but no V-T process and radiative transition.}
\label{case3}
\end{figure}
\section{Conclusions}
In this work we have performed calculations of the vibrational
distribution of both H$_2$ and H$_2^+$ for the conditions expected in
the early Universe and based on a comprehensive state-to-state chemical
kinetics. The vibrational level distribution for these two species are
reported for a wide and continuous range of the redshift parameter.
The results can be summarized as follows:
\begin{enumerate}
\item The vibrational distribution function of H$_2$ and H$_2^+$ assumes a
quasi-equilibrium shape at redshift $z \sim 1500$; after that, extended
plateau in the vibrational level distributions form, underlying the
presence of pumping phenomena for the intermediate vibrational levels;
full thermalization is not observed, because vibrational relaxation
processes are not fast enough to balance the strong vibrational
selectivity of formation rates.
\item Radiative processes play a fundamental role in the redistribution of
vibrational quanta, affecting the vibrational distribution function and
the overall fractional abundance (as in the case of H$_2^+$);
\item All these features can not be described in terms of the LTE
distributions which are usually assumed in chemical networks for the
primordial Universe.
\end{enumerate}
\acknowledgments
We are grateful to the following colleagues for their precious aid in
retrieving vibrationally resolved cross sections: Roberto Celiberto,
Martin \u{C}\'{\i}\v{z}ek, Gordon Dunn, Fabrizio Esposito, Predrag
Krsti\'{c}, Ratko Janev, Ralph Jaquet, Susanta Mahapatra, Donald
Shemansky, Hidekazu Takagi, Xavier Urbain. C.M.C. would also acknowledge Jonathan Tennyson and Ioan F. Schneider for useful discussions on electron$-$molecule collisions, and MIUR
and Universit\`{a} degli Studi di Bari, that partially supported this work (\textquotedblleft fondi di Ateneo 2010\textquotedblright).
|
\section{Introduction}
Scientific advances over the past 150 years, particularly in the medical field, have allowed the extension of life expectancy in western countries and this trend seems to increase in future years. Conservative estimates suggest that by 2030 in EU countries the proportion of people over 60 years regard the entire population will be around 50$\%$; this means that we will see a gradual increase in the number of those subjects with chronic diseases (ie diseases not involving healing), that will therefore increase the cost and effort over health care facilities \cite{CarmineZoccali06012010}.
As consequence of the exponential growth of hardware and software infrastructure it is possible to rethink the whole approach to the treatment of complex chronic disease, by limiting the hospitalization only to a severe worsening of patient's condition. This was the original idea behind the CHRONIOUS project: constructing a generic platform to monitor, in an unobtrusive way, a chronic disease patient with two goals\cite{Vitacca01022009}:
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{Improve the patients quality of life, by reducing as much as possible the hospitalizations.}
\item{Allow the clinician a continuous monitoring the patients, both in standard and potential risk situations.}
\end{itemize}
\indent
To gain this two goals, the CHRONIOUS platform has to integrate different technologies and hardware and software modules that need to interact among themselves. This paper is organized as follows: in the first section, the general structure of CHRONIOUS hardware and software modules are described. A deep analysis of the preprocessing algorithms covers the entire second section. Section three is dedicated to illustrate the machine learning algorithms. In last section we will evaluate possible improvements to the Chronious intelligence system.
\section{The CHRONIOUS system: an overview}
CHRONIOUS system deals with COPD and CKD: for these different types of chronic diseases, \hyphenation{ac-cor-ding}according to the medical guidelines, it is important to monitor different data in a way to check patient health status and to activate suitable emergency alarms for the clinician (for the COPD: ECG, SPO$2$ and respiratory rate; for CKD: glucose level, body weight and blood pressure) in case of critical event.
The CHRONIUOS platform consists of many modules that act together.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\textwidth]{chronious.png}
\caption{Chronious modules}
\label{}
\end{figure}
We can organize them in three frameworks
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{Patient Sensor Framework}
\item{Communication Framework.}
\item{Monitor Framework.}
\end{itemize}
\indent
The Patient Sensor Framework consists of hardware devices used to grab data from the patients. The hardware equipments, installed at patient's home, are
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{Wearable and external devices.}
\item{Touch screen Home Patient Monitor(HPM).}
\item{Personal Digital Assistant (PDA).}
\end{itemize}
\indent
The data collected are of two types:
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{silent, when the data recording is automated and it does not involve the patient interaction, as respiratory frequency measurement for COPD patients from the wearable device}
\item{no-silent, when it is requested a direct patient or caregiver interaction with the device, as for the blood pressure and questionnaires inputted by the HPM for diet, activity and food intake monitoring. The no-silent data acquisition is particularly important for monitoring CDK patient lifestyle.
}
\end{itemize}
\indent
During the day there are several measureaments, with different time intervals and different frequencies; only one data transmission is done, if there is no worsening in patient parameters. For COPD patient all the data are collected in a silent mode from a wereable t-shirt and trasmitted to the PDA via Bluetooth; for CDK patient all the measurements, including a lifestyle questionnarie are stored in the HPM and trasmitted in silent mode to the PDA.
PDA is in charge of doing the following action
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{Collect all the data from the devices.}
\item{Use a set of machine learning algorithms to determinate if it is needed to force a non scheduled trasmission to the Monitoring Framework.}
\item{Trasmit the collected and analysed information to the Monitoring Framework and receive back the changes, that will affect the interaction between the patient and the other devices.}
\end{itemize}
\indent
The Communication Framework is in charge of transmitting data among \hyphenation{dif-fe-rent} devices and from PDA to Central DB. This transmission is done using messages in a predefined xml format. The device that is in charge of doing the transmission of the xmls is the PDA.
\newline
\indent
The Monitor framework is the principal interface between the clinician and the patient. It receives data from the remote PDAs and it transmits back the therapy for food intake, drug intake, activities and scheduled measurements. The PDA is equipped also with a rule based decision support system, that contains an updatable set of rules created from the literature and validated by clinicians.
The rule based decision support system is in charge of analysing the data arrived from the PDAs and deciding if there is a grave or mild worsening of patient's conditions.
It is also able to alert the clinician and propose the suggestions about the action to do in several cases.
In the next section we will analyse the preprocessing phase needed to activate the intelligence in the PDA.
\section{The PDA CHRONIOUS preprocessing phase}
As we pointed out above, the Personal Digital Assistant is a smartphone equipped with WINDOWS MOBILE 6.5 Operating system, a SQL SERVER 2005 COMPACT database and a .NET FRAMEWORK 3.5.
The data registered by the PDA are the following
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{Data from the wereable jacket.}
\item{Answers to questionnaries concerning dietary habits, drug intake and lifestyle of the patient.}
\item{Data from the home patient monitor like blood pressure, glucometer measurements and body weight.}
\end{itemize}
\indent
Once these data are collected they are saved to the PDA database and a set of algorithms are triggered to analyse these data.
Since the PDA analyzes data for two different diseases, two sets of different algorithms are used.
The fundamental data needed by COPD treatment are ECG signals and Respiration data, so in case of a COPD patient we have a first processing Electrocardiogram Pre-processing. After that a Feature extraction phase is needed and at the end an Evaluation phase of the extracted features is done to determinate if an alert must be triggered to the Central System.
For external devices used in particular by CDK patients, there is no need of a preprocessing phase beacuse foundamental measures are the ones provided by the glucometer, the weight scale and the blood pressure measure; so they are discrete time and directly used by the set of machine learning. Combined with these data, the answers to a set of queries concerning food intake, drug intake, lifestyle and mental status are passed to a set of machine learning algorithms to evaluate the whole patient condition.
In next subsections we will analyse first the COPD set of algorithms used.
\subsection{Preprocessing of COPD signals}
The aim of Preprocessing Phase is to improve the general quality of the ECG, for more accurate analysis and measurement, because there's the possibilty to have some noises on the signals.
Possible noises in the signal include
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{Low frequency Base Line Wandering (BW) caused by respiration and body movements.}
\item{High frequency random noises caused by mains interference (50 or 60Hz).}
\item{Muscular activity and random shifts of the ECG signal amplitude caused by poor electrode contact and body movements.}
\end{itemize}
\indent
The preprocessing comprises:
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{Removal of base line wandering.}
\item{Removal of high frequency noise.}
\item{QRS detection.}
\end{itemize}
\indent
The BW which is is an extragenoeous low-frequency activity which may interfere with the signal analysis, rendering its clinical interpretation inaccurate and misleading.
Two major techniques are employed for BW removal:
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{Linear filtering: involves the design of a LTI high pass filter with
cut off in way that the clinical information in the ECG is preserved
and the BW is removed as much as possible.}
\item{Polynomial fitting: includes the fitting of polynomials to
representative points (knots) in the ECG, with one knot for each
beat. Knots are selected from a silent segment, e.g. the PQ
interval. A polynomial is fitted so that it passes through every
knot in a smooth fashion.}
\end{itemize}
\indent
The High Frequency Noise can be caused by the high frequency as well
as power supply interference from the ECG signal. It's removal is done using:
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{The Daubechies (DB4) wavelet employed on the
basis of the resemblance and similar frequency
response characteristics of the db4 basis function
with the ECG waveform}.
\item{ Using wavelets to remove noise from a signal
requires identifying which components contain the noise,
using optimal methods to threshold them, and then reconstructing
the signal using the thresholded coefficients.}
\end{itemize}
\indent
The prepocessing phase finally deals with the QRS detection. The main features that should be calculated: the Inter-beat (RR) interval and the Heart Rate Variability (variation in the beat-to-beat interval).
For the Inter-beat (RR) interval, two methods have been explored
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{Filtering the ECG signal with continuous (CWT) and fast wavelet\cite{Chesnokov}
transforms (FWT)\footnote{The reconstructed ECG signal after denoising contains only spikes with non-zero
values at the location of QRS complexes. From this signal, the PQ junction and J point
can be located as the boundaries of the spike. If the length of the spike is more or
less than a predefined QRS length range it is annotated as noise and if the voltage is
below a certain threshold, it is annotated as an artifact.
The next stage is the detection of the T wave, and the P wave in the PQ interval. The
peaks of Q, R and S waves are identified in the annotated part of the ECG signal from
the PQ junction to J point.}.}
\item{Following Pan-Tompkins\cite{Pan}, wavelets are used to remove noise from a signal
requires identifying which component or
components contain the noise, using optimal
methods to threshold them, and then reconstructing
the signal using the thresholded coefficients\footnote{
The algorithm includes a series of filters and
methods that perform lowpass, high-pass, derivative, squaring, integration, adaptive
thresholding and search procedures.
}.}
\end{itemize}
\indent
All the previuos features are extracted from ECG signals. For COPD patient also the Respiratory Rate is a fundamental parameter that need to be analysed.
In order to calculate the respiration rate using the reference respiration signal, a dominant frequency detection algorithm, based on short-time Fourier transform (STFT) \cite{0967-3334-26-5-R01}, is applied.
The STFT is a localized Fourier transform, utilizing a Hamming window:
\begin{equation}
STFT(f(t)) = STFT(\omega, \tau) = \int_{\infty}^{\infty} f(t) w(t-\tau) e^{-j\omega t} dt
\end{equation}
were $w(t)$ is the window function, commonly a Hann window or Gaussian hill centered around zero, and f(t) is the signal to be transformed.
Because frequency components of the respiration signal are very low (<$2$Hz), a window size of $60$ seconds is selected.Every 60s, the hamming window is multiplied to the respiration
signal, and the result is transformed to the frequency domain using
Fourier transform. The dominant frequency is then detected by
finding the maximum amplitude of the spectrums. When the
dominant frequency components are found, inverse numbers are
calculated in order to obtain the respiration rate.
After this first preprocessing phase for COPD patients we wil now analyse the which kind of Features are extracted.
\subsection{Features Extraction for COPD patients }
From the Inter-beat (RR) interval and the Heart Rate Variability,
several features can be extracted, either in time or in frequency
domain.
\newline
\indent
Dealing with Time domain the values extracted are
\begin{enumerate}
\item{ SDNN(msec): Standard deviation of all normal RR intervals in
the entire ECG recording using the following
\begin{equation}
sdnn = \sqrt{\frac{1}{n} \sum_{i=1}^n(NN_i - m)^2 }
\end{equation}
where $NN_i$ is the duration of the $i$-th $NN$ interval in the analyzed
ECG, $n$ is the number of all $NN$ intervals, and $m$ is their mean duration.
}
\item{
SDANM(msec): Standard deviation of the mean of the normal
RR intervals for each $5$ minutes period of the ECG recording.
}
\item{
SDNNIDX (msec): Mean of the standard deviations of all
normal RR intervals for all $5$ minutes segments of the ECG
recording.
}
\item{
pNN50
intervals that are greater than $50$ msec, computed over the
entire ECG recording.
}
\item{
r-MSSD (msec): Square root of the mean of the sum of the
squares of differences between adjacent normal RR intervals
over the entire ECG recording the formula is
\begin{equation}
rMSSD = \sqrt{\frac{1}{n-1} \sum_{i=1}^n (NN_{i+1}-NN_{i})^2}
\end{equation}
where NNi is the duration of the $i$-th $NN$ interval in the analyzed
ECG and $n$ is the number of all $NN$ intervals.
}
\end{enumerate}
If we now move to the frequency domain the Feature Extraction on the PDA studies two bands:
\begin{enumerate}
\item{ The Low – Frequency band (LF), which includes frequencies in
the area $[0.03 – 0.15]$ Hz.
}
\item{
The High – Frequency band (HF), which includes frequencies in
the area $[0.15 – 0.40]$ Hz.
}
\end{enumerate}
If we now move to the Respiration signal several features can be extracted either directly or indirectly we focused on:
\begin{enumerate}
\item{ Respiration Rate: The number of breaths per minute.}
\item{Tidal Volume (VT): The normal volume of the air inhaled after an
exhalation.}
\item{Vital capacity (VC): The volume of a full expiration. This metric
depends on the size of the lungs, elasticity, integrity of the
airways and other parameters, therefore it is highly variable
between subjects.}
\item{Residual volume (VR): The volume that remains in the lungs
following maximum exhalation.}
\end{enumerate}
After all the preprocessing phase of the data gathered by the weareable devices, all these information are passed to a Classification System. The classification system is responsible for the analys of the outcome from the preprocessing phase for the COPD patients and of the data gathered by the external devices and questionnaries inputted by the patient himself.
Below we will see how the software is used to transform all these rich set of data in an information to be, in case, transmitted real-time or scheduled to the Monitoring Framework.
\section{The CHRONIOUS Classification System}
After the collected data have been preprocessed for COPD patient and all the CKD patient input have been acquired, a set of machine learning algorithms are fired up to decide if a potentially risky situation is present. The aim of these tools is alerting the Central System that contains a rule based decision support system, for a better evaluation of the message triggered by the PDA. In case the message containg an alarm for life risk danger, the Central Decision Support System is able to alert the emergency staff or to suggest the clinician to modify the therapy approach.
Most of these tools need a preprocessing phase for identifying the correct parameters that need to be validated by the clinicians. This means that in the first validation of the CHRONIOUS project a large amount of efforts has been dedicated to gather feedback from the clinicians about the correctness of the rules / parameters that have been inferred by the algorithms. In these phase another important effort has been dedicated by the technicians to evaluate some probability index for fuzzy data measures.
The CHRONIOUS Classification System is composed of the following part:
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{A light rule based expert system.}
\item{A supervised classification system.}
\end{itemize}
\indent
The light rule based expert system is an xml parser that is able to extract from an xml a set of "`if then"' rules created an validated by the clinician.
With these rules combined with the data collected, the rule base system is able to decide if a patient is in a potentially life-risk situation.
For example the following rules will generate an immediate alarm to the central system:
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{
The Hearth Rate is above 120 bpm for both COPD and CKD patient.
}
\item{
If the weight increase by $2\%$ in the last 24 hour for CDK patient .
}
\end{itemize}
\indent
These rules are most for alarm triggering. It means that they aren't use for light monitor alerting. In the CHRONIOUS PDA system the Supervised Classification System is composed of the following machine algorithms:
\begin{itemize}
\renewcommand{\labelitemi}{$\bullet$}
\item{Support Vector Machines \cite{Cristianini}}
\item{Random Forest \cite{Brieman}}
\item{Multi-Layer Perceptron \cite{Haykin}}
\item{Decision Tree \cite{Quinlan}}
\item{Na\"ive Bayes \cite{Friedman}}
\item{Partial decision Trees \cite{Eibe}}
\item{Bayesian Network \cite{holmes}}
\end{itemize}
\indent
Apart from Bayesian Network, the other algorithms have been trained with a dataset to generate a set of rule that have been validated by clinician. The Bayesian Network have been used to identified possible rules about the mental and stress evaluator of the patient.
This means that once trained, the rules generate can be used to identify is some stressfull condition can alter some parameter and leading to a worsening of the general state of the patient.
The use of these algorithms was needed beacuse for the CKD patient the diet covers the most part of the medical treatment, so any factor that can influence a changement on the diet intake, would potentially and indirectly lead to a worsening of the patient condition. For example, if a female CKD patient feeld sad, these condition could lead her to eat a bigger piece of pie for satisfatcion purporse. In general a bad feeling could lead chronic patients to be uncompliance with the medical treatment. However the kind of rules aren't liked the vital signs so their fuzzyness could be identified by these type of algorithms. Clearly while diet is important to avoid worsening on CDK patient conditions, on the other side the lifestyle could be and indirect cause of COKD worsening condition.
Nevertheless for both of them we need static rules to have alarms sending, because for both for example having a body temperature above $38$ could be a risky situation were an hospitalization is needed for both COPD and CKD patients.
Considering the training dataset a set of $41$ attributes have been identified.
These data comes from the $2$ sets of $2$ hours health recordings ($11$ attributes), food input module ($12$ attributes), drug intake module ($1$
attribute), activity input module ($2$ attributes), questionnaire ($13$ attributes) and external device ($2$ attributes).
Some of the results of some classifier are shown in table $1$.
\begin{table}[htbp]
\centering
\caption{Classification System: MAE: Mean absolute error, RMSE: Root mean squared error, RAE : Relative absolute error, CI: Correctly/Instances}
\begin{tabular}{|c|ccc|r|}
\hline
Method & MAE & RMSE & RAE & CI \\
\hline
PART & 0.1336 & 0.312 & 57.81 \% & 2.67 \\
J48 & 0.1336 & 0.321 & 57.81 \% & 2.67 \\
Forest & 0.2 & 0.341 & 86.52 \% & 1.75 \\
Na\"ive & 0.127 & 0.343 & 54.95 \% & 2.67 \\
\hline
\end{tabular}
\label{tab:table1}
\end{table}
Again we point out that even if there are some errors due to false positive matches, the PDA system in these case would only generate a rule that will send a message to the clinician that most of the times would only say that a light worsening is present. The core intelligence that deals with the central system would be the real suggestion system that would indicate to clinician a suggestion on how to act to possibly revert the trend.
The Bayesian Network is the fundament algorithm for the Mental support tool. It uses a set of attributes that affect a stress index and they weight based on the clinician's feedback as shown in table of Figure 2. When the total stress indicator is above a certain value a light alarm is triggered to the Central Database to inform clinician of a potentially worsening of patient conditions.
In the same way a module in the PDA is in charge of the rules concerning the lifestyle od the patient: the lifestyle tool. It collects data inputted by the patient or caregiver in a validation phase and using a Bayesian Network is able to compute an index of good or poor lifestyle of the patient also in this case if the poor lifestyle is find a light alarm is sended to the clinician for monitoring purporse.
\begin{figure}[ht]
\centering
\includegraphics[width=0.50\textwidth]{mental.png}
\caption{Attributes}
\label{fig:mental}
\end{figure}
\section{Conclusions an improvements}
In these paper we presented a set of machine learning algorithms store in a smart phone that, combined with some external devices and patient inputted data, can be used for a first monitor/alert system for treatment of patient affected by chronic diseases.
Dealing with telemedicine application these kind of software could help to improve patient life quality and could be also be a valid help for clinician to allow a more precise monitoring of patient conditions without need of the physical presence of the clinician. Apart these potentially advantage a PDA equipped with these kind of application can suffer of some limitations.
During developing phase we face these problems:
\begin{itemize}
\item{Heavy resource consumption of preprocessing algorithms.}
\item{Updating a trained supervised algorithm. }
\end{itemize}
\indent
The preprocessing algorithm for COPD parameter denoising is the most memory/CPU consumption. This potentially can became a problem when we deal with life risk situations, because in the time that the algorithm denoise the ECG signal, the patient can became unconscius and so precious time can be lost in these phase. Other time is losed due to the huge amount of signal trasmitted from the wereable, this because we deal with an ECG signal that is composed of a mean of ten measurements per second so this means that 5 minute of ECG signal became nearly $3000$ sql commands to a SQL SERVER 2005 COMPACT that is not so performant in these case. Increasing hardware requirements of the PDA can be a first solution however it would be interesting to understand if a relational database is the best solution for storing these type of data or other structure would be more confortable for storing purpose.
The other important issue is that once the supervised learning algorithms are trained also a little change in one parameter will need to retrain the algorithms and most important, it would need a new validation of the outputs by the clinicians. These can lead to difficults when after these phase the new trained algorithms need to be updated on the pda, this because the Chronious Communication Framework is only able to transmits value from the PDA to Monitor Framework and back. In these case an interesting solution could be also to allow remote updating of the structure validated. For example a trained neural network could read the weights matrix from a structure upgradable by the communication framework. Apart from these improvements, and many others that could lead to a software system closer as much as possbile to the clinician and patient needs, it is our opinion that with the smart devices that are closer to a normal PC, the algorithms presented in this paper could become an important part of the telemedicine platforms.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The subdwarf B (sdB) stars are generally known to be core-helium-burning
stars with inert hydrogen-dominated envelopes of very low mass.
This composition places them on the hot extension of the
horizontal branch,
the extreme horizontal branch \citep[EHB;][]{heber86,heber09}.
Their low-mass envelopes prevent them from reaching a second giant stage,
and instead, after their core helium is exhausted, they
expand only briefly, never exceeding more than about a third of a solar
radius, before contracting and passing on
towards higher effective temperatures and gravities.
Thus, an sdB star evolves into the hotter sdO population before reaching
degeneracy and the associated white dwarf (WD) cooling track \citep{dorman93}.
While the future evolution of sdB stars is unproblematic, exactly how
they lose their envelopes during their first red giant (RGB) stage is
more complicated.
If the star has a close low-mass companion, mass transfer on the tip of
the RGB will quickly pull it into the envelope.
As friction transfers angular momentum from the orbit to the envelope,
it will spin up and be ejected from the system.
If the core has become massive enough for helium to ignite ($\sim$0.47\,\ensuremath{\rm{M}_{\odot}}),
the result will be a very close sdB+dM system. These binaries have spectacular
reflection effects, and a total of 14 such objects are described in the literature;
13 are listed in \citet{for10} and the most recent one in \citet{twom1938}.
While such sdB+dM binaries make out only a tiny fraction of the total
sdB population, cases where the companion responsible for ejecting the
envelope is a low-mass WD, appear to be much more common.
\citet{maxted01} found that 21 of 36 stars in their radial velocity study are in
short period systems, most likely with WD companions.
If the companion is more massive than the subdwarf (at least at the
end of mass transfer), the orbit will have expanded to more than
a hundred days. Such orbits are hard to measure, but the companion
is easily detectable spectroscopically or from infrared excess.
\citet{napiwotzki04} found that more than a third of their sdB sample
show the spectroscopic signature of main sequence companions, while
\citet{reed04}, using {\sc 2mass}\ photometry, inferred that about half
of the sdBs in the field are likely to be of this type.
Thus, the vast majority of sdB stars in the field must be in binary systems.
However, a few pulsating sdBs have been studied in such detail that any
companion close enough to have interacted during the RGB stage can be
ruled out. A possible explanation for such single sdB stars can be the merger
of two low-mass WDs.
These three scenarios are all investigated in detail in the binary population
study of \citet{han02,han03}. They find that while the first two scenarios
must have sdBs with masses close to the He-flash mass of 0.47\,\ensuremath{\rm{M}_{\odot}}, the
merger scenario can produce sdBs with a broad distribution of masses
ranging from 0.4 to 0.7\,\ensuremath{\rm{M}_{\odot}}.
\begin{figure*}
\includegraphics[width=\hsize]{lc_all.eps}
\caption{
Light curves of J20136+0928. The top panel represents
the discovery light curve.
In the online version the different data sets are shown in
different colors (Mercator; red, Baker; green, SPM; blue).
}
\label{fig:lc_all}
\end{figure*}
While most sdB stars form a band in the \ensuremath{T_{\rm{eff}}}--\ensuremath{\log g}\ diagram that is
consistent with a mass close to the canonical 0.47\,\ensuremath{\rm{M}_{\odot}}, and various
envelope thicknesses up to the H-burning limit, a few exceptions are
found. One such is \object{V338\,Ser}, discovered to be a high
amplitude sdB pulsator by \citet{koen98_pg1605}.
A multi-site campaign devoted to this pulsator \citep{kilkenny99} found more
than fifty pulsation frequencies in the range 365--529\,s, which now defines the
long-period edge for the V361\,Hya group of short-period sdB variables (sdBVs).
Time-resolved spectroscopy of V338\,Ser was analysed by \citet{tillich07},
and most recently \citet{v338Ser} performed high-resolution time-resolved
spectroscopy.
Its unusually low gravity was discussed briefly by \citet{ostensen09},
proposing that it may be a core-He-burning EHB star with a mass larger
than the canonical, rather than a shell-He-burning post-EHB star
as suggested by \citet{koen98_pg1605}.
Asteroseismic solutions supporting this hypothesis were recently
presented by \citet{vanGrootel10_V338Ser}, implying a mass of
0.76\,\ensuremath{\rm{M}_{\odot}}, and an envelope mass fraction of 0.2\%.
Since V338\,Ser remained a unique object in spite of a decade
of searches for V361\,Hya pulsators, we have recently been
paying particular attention to candidates in this low-gravity region.
Here we present the discovery of pulsations in J20136+0928\
(Figure~\ref{fig:lc_all}), which has an
even lower surface gravity than V338\,Ser, and the lowest of any
sdB pulsator found to date.
\section{Target selection}
Since 2008 we have observed UV-excess targets selected based on
UV photometry from the {\em Galaxy Evolution Explorer} ({\sc galex})
satellite \citep{GALEX}.
An estimate based on targets with NUV\,$<$\,14.0 and
FUV\,--\,B\,$<$\,+0.5 is that only about two-thirds
were already cataloged.
UV-excess objects in this cut include a host of interesting compact stars.
All types of pulsating WDs and hot subdwarf stars are
included, as are most types of cataclysmic variables
and planetary nebula nuclei. Our first cut based on {\sc galex}\ data
release 4 (GR4), contained 649 objects of which only about 400 had
even approximate classes in the literature. We started observing
targets from this list as a backup programme for various observing
runs, and have by now mostly completed the sample. Full details of
this survey will be given in a forthcoming paper. A similar survey
is described by \citet{vennes11}.
As we have accumulated spectroscopic
observations, new sdBV candidates have entered our observing
lists for photometric follow-up, and new discoveries are emerging.
The first result was the discovery of pulsations in \object{J08069+1527},
recently presented by \citet{baran11}, and establishing it as a being
a high-amplitude member of the rare DW\,Lyn (hybrid) type of sdBVs.
The object presented here appears in {\sc galex}\ data release 6
(GR6) as GALEX J201337.6+092801, with UV magnitudes
FUV\,=\,12.263(3), NUV\,=\,13.079(3).
Photographic magnitudes for this star from the {\sc nomad}\ database are
$B$\,=\,12.01, $V$\,=\,12.22, $R$\,=\,12.34, and from the {\sc 2mass}\
near-IR survey;
$J$\,=\,13.00, $H$\,=\,13.11, $K$\,=\,13.23. The magnitudes
are all perfectly consistent with an sdB star that has no
main sequence companion.
J20136+0928\ also appears in the Tycho-2 catalog as TYC 1077-218-1,
where it has $B$\,=\,12.16, $V$\,=\,11.78.
The {\sc nomad}\ database also lists the target with a small but
significant proper motion (\ensuremath{\mu_{\alpha}},\,\ensuremath{\mu_{\delta}}\,=\,12.1,\,2.8 mas/yr).
\begin{figure}
\centering
\includegraphics[width=6cm]{tgfit.eps}
\caption{LTE model fit of J20136+0928.}
\label{fig:tgfit}
\end{figure}
The target was observed spectroscopically during a run at the
Isaac Newton Telescope (INT) at
Observatorio Roque de los Muchachos on the island of La Palma (Spain).
After the observing run we processed all the data and extracted the spectra
with {\sc iraf} in the regular way. We fitted the hot subdwarfs using the
pure H+He LTE model grids of \citet{heber00}, and
we show the resulting fit for J20136+0928\ in Figure~\ref{fig:tgfit}.
The formal fitting errors (shown on the figure) do not account for
systematic effects inherent in the models, so we generously increase
the errors by a factor five when stating
\ensuremath{T_{\rm{eff}}}\,=\,32\,100$\pm$1000, \ensuremath{\log g}\,=\,5.15$\pm$0.20,
and \ensuremath{\log \left(N_{\mathrm{He}}/N_{\mathrm{H}}\right)}\,=\,--2.8$\pm$0.2. Note that the typical shift of
when going from LTE to NLTE models is $\sim$+1000\,K, and
can become larger when metalicity effects are taken into account
\citep[see][and references therein]{heber09}.
However, the low surface gravity of this object is clear.
\begin{figure}
\centering
\includegraphics[width=8cm]{ft_pw.eps}
\caption{FT of the five consecutive nights taken together
(top panel), and the prewhitening sequence after fitting
and subtracting five, ten and sixteen frequencies from
the original light curve. The window function is shown
as an inset to the top panel.
The arrows indicate the positions
of the frequencies identified and removed in each step.
The scale in the bottom panel has been expanded
to show the residual level, and the line at which we
stop is at 3.7 times the mean residual level.
}
\label{fig:ft_pw}
\end{figure}
\section{Photometric observations}
The discovery run was made on October 6, 2010 with the
1.2-m Mercator Telescope, sited just next to the INT.
Although the conditions were poor, the main period of $\sim$520\,s was
clearly seen, and recognised to be longer than normal for a V361\,Hya star.
Since it was already clear from the spectroscopy that the gravity of the
target was unusually low, the connection with V338\,Ser was immediately made,
and we urgently proceeded to confirm the discovery.
As the target was already setting early in the
night, we pooled observing time at several telescopes in order to
collect enough photometry for a frequency analysis.
Photometric data were collected at three different sites within one
month of the discovery run.
At the Mercator telescope we used the Merope CCD camera,
which was recently upgraded with a new E2V frame transfer CCD with
2048$\times$3074 illuminated pixels \citep{ostensen10}.
The observations were done with a Geneva-$B$ filter
($\lambda$\,=\,4201$\pm$286\,\AA), and a cadence period of $\sim$15 sec.
With the 0.4-m telescope at Baker Observatory (Missouri, USA) we used
a Photometrics RS-1340 CCD with a BG40 filter
($\lambda$\,=\,4750$\pm$1500), and a cadence of $\sim$30\,s.
We also obtained three consecutive nights with the 0.84-m telescope
at San Pedro M\'artir (SPM) observatory in Mexico.
The CCD camera is equipped with an E2V 4240 CCD, and we used
a Bessel-$B$ filter ($\lambda$\,=\,4350$\pm$980\,\AA),
and a cadence of 38\,s.
The light curves are all shown in Figure~\ref{fig:lc_all},
with the discovery run
on top. Poor weather at Mercator left us unable to collect
a significant amount of data at our prime site.
Differential photometry was made with respect to the sum of three or four of the
brightest reference stars found within the field of view of the CCDs.
Most of this reference signal comes from TYC 1077-402-1, about 1.7' SW of
the target, which is comparable in brightness to J20136+0928.
The five bottom panels are consecutive nights, but the coverage is rather
poor due to the target only being visible for a few hours before setting.
On both October 27 and 28, there was a few hours of overlap between data
from Baker and SPM.
This allowed us to confirm that there is no significant difference in
amplitude between observations made with the two telescopes, in spite of
a significantly broader bandpass at Baker.
We only use the last five nights for the frequency analysis described
in the next section.
\newcommand{\,\ensuremath{\pm}\,}{\,\ensuremath{\pm}\,}
\begin{table}
\caption[]{List of Frequencies}
\label{tbl:freqs}
\centering
\begin{tabular}{lcccr} \hline
& Frequency & Period & Amplitude & Phase \\
ID & \multicolumn{1}{c}{(\ensuremath{\mu{\rm{Hz}}})}& \multicolumn{1}{c}{(s)} &
\multicolumn{1}{c}{(mma)} & \multicolumn{1}{c}{(s)} \\ \hline
$f_{9}$ & 1767.3\tpm0.3 & 565.84\tpm0.09 & 2.18\tpm0.38 & 194\tpm27 \\
$u_{15}$& 1774.5\tpm0.6 & 563.53\tpm0.19 & 1.28\tpm0.43 & 12\tpm56 \\
$f_{4}$ & 1810.7\tpm0.2 & 552.26\tpm0.06 & 3.91\tpm0.42 & 123\tpm18 \\
$f_{7}$ & 1894.8\tpm0.4 & 527.75\tpm0.11 & 3.07\tpm0.87 & 246\tpm40 \\
$u_{12}$& 1908.1\tpm0.7 & 524.08\tpm0.18 & 1.84\tpm0.71 & 344\tpm88 \\
$f_{1}$ & 1933.8\tpm0.1 & 517.12\tpm0.04 & 8.51\tpm0.82 & 139\tpm19 \\
$u_{11}$& 1946.9\tpm0.5 & 513.65\tpm0.14 & 2.11\tpm1.03 & 466\tpm54 \\
$f_{3}$ & 2059.3\tpm0.1 & 485.60\tpm0.03 & 4.59\tpm0.36 & 82\tpm11 \\
$f_{10}$& 2142.8\tpm0.3 & 466.67\tpm0.06 & 2.11\tpm0.36 & 97\tpm22 \\
$f_{2}$ & 2214.1\tpm0.3 & 451.66\tpm0.05 & 5.44\tpm1.15 & 70\tpm29 \\
$u_{14}$& 2215.2\tpm1.0 & 451.42\tpm0.19 & 1.45\tpm1.18 & 31\tpm107 \\
$f_{8}$ & 2328.8\tpm0.2 & 429.41\tpm0.04 & 2.88\tpm0.35 & 103\tpm15 \\
$f_{13}$& 2392.9\tpm0.4 & 417.91\tpm0.07 & 1.51\tpm0.35 & 405\tpm28 \\
$f_{6}$ & 2521.7\tpm0.4 & 396.56\tpm0.06 & 3.16\tpm0.99 & 31\tpm37 \\
$f_{5}$ & 2523.3\tpm0.3 & 396.30\tpm0.05 & 3.86\tpm0.99 & 0\tpm30 \\
$u_{16}$& 2659.5\tpm0.7 & 376.02\tpm0.10 & 0.75\tpm0.35 & 26\tpm50 \\
\hline
\end{tabular}\\
{\bf Notes.}---Phases are times of maxima since HJD 2455495.5.\\
Frequencies with large uncertainties are identified with $u$.
\end{table}
\begin{figure*}[t]
\centering
\includegraphics[width=12cm]{tgplot.eps}
\caption{The \ensuremath{T_{\rm{eff}}}--\ensuremath{\log g}\ diagram as observed by the Bok--Green survey
\citep{green08} with V338\,Ser highlighted, and J20136+0928\ added.
The evolutionary tracks are from \citet{kawaler05}.}
\label{fig:tgplot}
\end{figure*}
\section{Frequency analysis}
The Fourier transform (FT) of all runs longer than 1\,h
is far too short to resolve the complexity of the frequency
spectrum displayed by J20136+0928. Joining the five consecutive nights and
taking the FT of the combined dataset produces the complex frequency
spectrum with strong 1-day aliasing
shown in the top panel of Figure~\ref{fig:ft_pw}.
The tools used for the frequency determination were developed particularly for
multi-site data, and are described in \citet{vuckovic06}.
We proceed to prewhiten the most significant well-separated peaks, five or
six at a time, until we reach a level of 0.71\,mma (straight line in the
bottom panel of the same figure). This level is 4.2 times the noise level
when computed in the high-frequency region where no pulsational power
is seen (although we do detect the drive frequency of some of our
telescopes), and 3.7 times the mean residual level of the 1000 to
3000\,\ensuremath{\mu{\rm{Hz}}}\ region shown in the last panel.
In each step the most significant peaks are used as input for a non-linear
least-squares fitting procedure, and all 16 frequencies, amplitudes and phases
are left free in the final step. Several frequencies close to the
most significant peaks do not converge to reliable values, and we consider
these as uncertain (IDs labeled $u$ in Table~\ref{tbl:freqs}).
It is not clear whether this is due to the poor coverage of our data, or caused
by real amplitude variability of the target. Concern can also be raised
about some frequencies that are barely resolved ({\em e.g.} $f_5$ and $f_6$).
\section{Discussion and Conclusions}
A new, bright sdBV with an exceptionally low gravity has been found in a sample
of UV-excess stars identified from {\sc galex}\ photometry.
In Figure~\ref{fig:tgplot} we show its
position with respect to V338\,Ser and the bulk of the sdB population, which
falls between the zero and terminal age EHB lines. J20136+0928\ is
located close to minimum \ensuremath{\log g}\ during the post-EHB stage of the uppermost
evolutionary track,
which corresponds to
a 0.470\,\ensuremath{\rm{M}_{\odot}}~helium burning star with an H-envelope mass of 0.004\,\ensuremath{\rm{M}_{\odot}}.
Evolutionary models of more massive EHB stars are not shown, but can
be found in \citet{han02}, from which one may infer that models with
a mass of $\sim$0.7\,\ensuremath{\rm{M}_{\odot}}\ and an envelope of $\sim$0.01\,\ensuremath{\rm{M}_\star}\
would approximate the position of J20136+0928\ reasonably well.
As mentioned in
the introduction, such an overmassive EHB star cannot form through the regular
binary mass-transfer scenarios that are responsible for producing the
bulk of the EHB stars, but can form through the merger of two helium
core WDs. Theoretical models for these scenarios,
combined with future observational efforts, may allow us to determine
the origins of both V338\,Ser and J20136+0928.
This discovery brings the total number of short period sdBV stars (both V361\,Hya
and hybrid DW\,Lyn stars) up to 52 (49 are summarised in \citealt{sdbnot}, one
was recently discovered in the field of the {\em Kepler} spacecraft by
\citealt{kawaler10a}, and the last one by \citealt{baran11}).
The discovery of pulsations in a star with such a low gravity as J20136+0928\ is well
explained by the classical $\kappa$ mechanism for pulsations in sdBVs \citep{charpinet01}.
However, the exceptionally low gravity of J20136+0928\ is surprising in evolutionary
terms, and the presence of pulsations allows the mass and internal structure to
be explored.
The encouraging brightness of J20136+0928\ makes it an excellent target for
time-resolved spectroscopy studies, and arranging a multi-site campaign with small
telescopes when the target is visible throughout the night will be a priority for
the coming season. By accurately determining the frequencies present in the star,
asteroseismology will be able to reliably determine its mass and internal
structure, and allow us to determine the most likely evolutionary
origins of this object.
\acknowledgements
The research leading to these results has received funding from the European
Research Council under the European Community's Seventh Framework Programme
(FP7/2007--2013)/ERC grant agreement N$^{\underline{\mathrm o}}$\,227224
({\sc prosperity}), as well as from the Research Council of K.U.Leuven grant
agreement GOA/2008/04.
RO acknowledges financial support from the Spanish grants AYA2009-08481-E,
AYA2010-14840 and AYA2009-14648-C02-02.
AB gratefully appreciates funding from Polish Ministry of Science and Higher
Education under project N$^{\underline{\mathrm o}}$\,554/MOB/2009/0.
ACQ, JTG and LLH were supported by the Missouri Space Grant, funded by NASA.
LFM acknowledges financial support from PAPIIT IN114309.
Based on observations made with the Mercator Telescope, operated on the
island of La Palma by the Flemish Community, at the Spanish Observatorio
del Roque de los Muchachos of the Instituto de Astrof\'{i}sica de Canarias.
\newcommand{\apjournal}[3]{#1, #2, #3}
|
\section{Introduction}
Recently, there has been an apparent anomaly in the top sector: the observation by the CDF experiment of a top forward-backward asymmetry (${A_{FB}}$)~\cite{Aaltonen:2011kc}, where the forward-backward asymmetry in a particular invariant mass bin, $M_{t\bar{t},i}$, is defined by
\begin{equation}
A^{t\bar{t}}(M_{t\bar{t},i}) = \frac{N(\Delta y>0,M_{t\bar{t},i})-N(\Delta y<0,M_{t\bar{t},i})}{N(\Delta y>0,M_{t\bar{t},i})+N(\Delta y<0,M_{t\bar{t},i})},
\label{AFBdef}
\end{equation}
with $\Delta y$ the rapidity difference between a top and an anti-top. The recent CDF anlaysis shows ${A_{FB}} = 0.475 \pm 0.114$ for $M_{t\bar{t}} > 450$ GeV~\cite{Aaltonen:2011kc}, while the Next-to-Leading Order (NLO) Standard Model (SM) predicts much lower values $0.088 \pm 0.013$~\cite{Kuhn:1998jr,Kuhn:1998kw,Bowen:2005ap,Almeida:2008ug}, corresponding to a 3.4$\sigma$ deviation. The D0 collaboration also observes a larger than predicted asymmetry~\cite{Abazov:2007qb}.
Since the SM prediction for the top-pair production cross section is in relatively good agreement with observation, a new physics model must generate the large forward-backward asymmetry without disturbing the total cross-section or observed invariant mass spectrum of $t\bar{t}$ production. In order to do this, many models assume an additional tree-level contribution from the exchange of a new particle in a way that maximizes the effect on the forward-backward asymmetry while minimizing the effect on the overall production cross section. Models proposed thus far in the literature that generate the observed Tevatron ${A_{FB}}$ at tree level while not grossly disrupting the top pair production cross section fall into two categories according to the nature of the new particle exchange: (i) $s$-channel exchange of vector mediators with axial couplings ({\it e.g.} axigluon models)~\cite{Sehgal:1987wi,Bagger:1987fz,Ferrario:2009bz,Frampton:2009rk,Chivukula:2010fk,Djouadi:2009nb,Bauer:2010iq,Alvarez:2010js,Chen:2010hm, Delaunay:2011vv, Bai:2011ed, Zerwekh:2011wf,Barreto:2011au} or (ii) $t$-channel exchange of flavor-violating mediators~\cite{Jung:2009jz,Cheung:2009ch,Shu:2009xf,Dorsner:2009mq,Arhrib:2009hu,Barger:2010mw, Gupta:2010wt, Xiao:2010hm, Gupta:2010wx,Cheung:2011qa,Cao:2011ew,Berger:2011ua,Barger:2011ih,Grinstein:2011yv, Patel:2011eh}. Comparative studies of these models have also been carried out \cite{Jung:2009pi,Cao:2009uz,Cao:2010zb,Jung:2010yn,Choudhury:2010cd,Jung:2010ri,Blum:2011up,Ahrens:2011mw, Craig:2011an, Delaunay:2011gv,Foot:2011xu}. The $s$-channel mediators tend to have maximally axial couplings, while the $t$-channel mediators tend to be maximally flavor violating, connecting a light quark to the top quark. Recently, it has been pointed out that such maximal flavor violation can also explain anomalies in the $B_s$ and $B_d$ systems when the $b$-quark is coupled as well~\cite{Shelton:2011hq,note}.
The recent CDF analysis \cite{Aaltonen:2011kc} greatly extends the experimental information about the asymmetry. Due to much improved statistics of top-pair production at the Tevatron, the analysis shows event distributions in more detail. In particular, CDF has now collected enough data to give the forward-backward asymmetry with respect to the invariant mass, $M_{t\bar{t}}$, and the rapidity, $y$, of reconstructed tops, and to compare these distributions for various data subsets, such as the 4- and 5-jet samples. This alows us to reassess the viability of previously suggested theoretical models that address the $A_{FB}$ anomaly. For example, in the high invariant mass bin, $M_{t\bar{t}}>$ 450 GeV, the asymmetry is very large, $\sim 50\%$, and hence some models that marginally explain the previous $A_{FB}$ will be challenged. Diquark models, for example, have great difficulty in producing a large enough asymmetry without generating an overly large correction to the total production cross-section.
In addition, although the top forward-backward asymmetry and related variables are straight-forwardly defined observables, the interpretation of experimental observation to confirm or falsify models must be done with care due to the indirect nature of top quark identification. In spite of the effort of the CDF group to show an ``unfolded'' $A_{FB}$ for direct comparison with theory predictions, a broad coverage of models is needed to justify its model independence, especially where selection effects can come into play. Therefore, it is necessary to compare direct experimental signatures to expectations within a wide variety of models.
The purpose of this paper is to give a comprehensive analysis of a variety of models that have been proposed in the literature to explain the ${A_{FB}}$ anomaly in order to (1) reassess the viability of such models given the new experimental data and (2) investigate subtleties associated with unfolding the ``raw'' $A_{FB}$ to a parton level $A_{FB}$ for comparison to models.\footnote{We use ``raw'' to refer to the measurement uncorrected for detector effects, but with background subtracted, as presented in the CDF paper.} In order to do the latter, we must decay the tops, sending the events through a detector simulation, and reconstruct the tops utilizing a $\chi^2$-based top reconstruction algorithm. Because, unlike previous studies, we do fully reconstruct the tops and simulate detector effects, we are able to make comparisons to the raw CDF asymmetries.
Comparing our results against the raw asymmetries and invariant mass spectra also enables us to assess the model dependence of event selection efficiencies. We find that these efficiencies can dramatically affect the extracted parton level total and differential cross-sections.\footnote{We use ``parton level'' as in the CDF $A_{FB}$ paper, where the term refers to de-convolving event selection efficiencies, detector efficiencies, jet algorithms, background, etc., from the underlying physics \cite{Aaltonen:2011kc}.} For example, $Z'_H$ and $W'$ models that produce large asymmetries have event selection efficiencies in the high invariant mass bins that are about half or less than what a Standard Model $t\bar{t}$ Monte Carlo would predict. This must be taken into account when comparing a model against parton level results. For example, while the parton level invariant mass spectra of $Z'_H$ and $W'$ models appear badly in conflict with the reconstructed invariant mass spectra from CDF at high invariant mass, we find that once the selection efficiencies are taken into account, the agreement between the data and model is good.
This paper is complementary to our earlier paper in which we examined the reach of the LHC at 7 TeV to discover the mediators of $t$-channel physics generating the Tevatron asymmetry. We found there that a top-flavor violating state light enough and with large enough couplings to generate the asymmetry will be rapidly discoverable at the LHC via a search for $tj$ resonances in $t\bar{t}j$ events \cite{Gresham:2011dg}.
The outline of this paper is as follows. In the next section we summarize and discuss the qualitative features of the classes of models that have been shown to be capable of generating the Tevatron top $A_{FB}$.
In the following section, we first show our parton-level comparison of various models to the unfolded Tevatron results. Then we show detector-level comparison utilizing a reconstructed top sample. Next we present a lepton asymmetry from fully leptonic $t\bar{t}$ events, which has recently been discussed in a CDF note \cite{CDFLeptons}. Lastly, we discuss the LHC reach for discovering such states, based on the analysis of \cite{Gresham:2011dg}.
\section{Models}
\begin{figure}
\centering
\subfigure[~s-channel $q\bar{q}$]{\includegraphics[width=0.2\textwidth]{pics/feynSqq.pdf}}\qquad
\subfigure[~s-channel $gg$]{\includegraphics[width=0.2\textwidth]{pics/feynSgg.pdf}}~~~
\subfigure[~t-channel $q\bar{q}$]{\includegraphics[width=0.2\textwidth]{pics/feynTqq.pdf}}~~~
\caption[$t \bar{t}$ production diagrams.]{Tree level $t \bar{t}$ production diagram with mediator $M$ exchange.}
\label{feynmandiagram}
\end{figure}
The Leading Order (LO) SM tree-level amplitude for $t\bar{t}$ production does not generate a forward-backward asymmetry. In the SM, a small positive top forward-backward asymmetry is generated through interference between a one-loop box diagram and a LO tree level diagram, $A_{FB} (M_{t\bar{t}} < 450 {\rm GeV}) = 0.040 \pm 0.006$, $A_{FB} (M_{t\bar{t}} > 450~{\rm GeV}) = 0.088\pm 0.013$.\footnote{ Interference between initial state gluon radiation and final state gluon radiation makes a very small negative contribution to the asymmetry.}
Since the SM contribution is generated at NLO, if there is an additional LO tree-level contribution from new physics, it can easily dominate.
Such LO diagrams are of the form of those in Fig.~(\ref{feynmandiagram}). They can be either $s$-channel (Fig.~(\ref{feynmandiagram}a) and (\ref{feynmandiagram}b)) or $t$-channel (Fig.~\ref{feynmandiagram}c). $s$-channel mediators couple directly to light flavors and gluons, and therefore the mediator masses must be large enough to evade dijet resonance search constraints \cite{Chivukula:2010fk,Bai:2011ed}. To maximize the contribution to $A_{FB}$, such a model must have a big axial coupling.
On the other hand, $t$-channel models should have large flavor violation between the light and the top generations, as can be seen in Fig.~(\ref{feynmandiagram}c). Large flavor violation
is experimentally allowed even for low mass mediators, $M$, as long as new couplings between light generations and left-handed quarks is suppressed; then strong limits on flavor violation and from dijet resonance searches are avoided. Additionally, the same-sign top signature search limit prefers $M$ to be a non-self-conjugate state \cite{Jung:2009jz}. Therefore, ordinary $Z'$ models run into difficulty. Here, to avoid same-sign top constraints, we consider horizontal $Z_H'$s with flavor charge. Color exotic states and $W'$s can also satisfy the requirement.
In the following sections, we summarize the defining Lagrangian of $t$-channel $W'$, $Z_H'$, triplet scalar, sextet scalar and $s$-channel axigluon models and present the tree-level differential cross sections, ${d \sigma(q \bar{q} \rightarrow t \bar{t}) \over d {\cos \theta}}$.
\subsection{Flavor-Changing $W'$, $Z'$}
The Lagrangian for a flavor-violating $Z'$ interaction is
\beq\label{eq:FVZprimeLagrangian}
{\cal L} = \frac{1}{\sqrt{2}}\bar{t} \gamma^\mu (g_L P_L +g_R P_R) u Z'_\mu + \mbox{h.c.},
\eeq
giving rise to a scattering cross-section
\beq
\frac{d \sigma}{d\cos\theta} =\frac{\beta}{32 \pi \hat{s}} \left({\cal A}_{SM} + {\cal A}_{int} + {\cal A}_{sq}\right),
\eeq
where
\beq
{\cal A}_{SM} = \frac{2 g_s^4}{9}(1+c_\theta^2 + \frac{4 m_t^2}{\hat{s}}),
\eeq
%
with $c_\theta = \beta \cos \theta$ and $\beta = \sqrt{1-4 m_t^2/\hat{s}}$. The new physics contributions are \cite{Cao:2010zb}
\begin{eqnarray}
{\cal A}_{int} = \frac{2 g_s^2 }{9} \frac{(g_L^2+g_R^2)}{\hat{s} \hat{t}_{Z'}}\left[2 \hat{u}_t^2+2 \hat{s} m_t^2+\frac{m_t^2}{m_{Z'}^2}(\hat{t}_t^2+\hat{s} m_t^2)\right],
\label{Z'int}
\end{eqnarray}
%
\begin{eqnarray}
{\cal A}_{sq} = \frac{1}{2 \hat{t}_{Z'}^2}&&\left[(g_L^4+g_R^4)\hat{u}_t^2+2g_L^2g_R^2 \hat{s} (\hat{s}-2 m_t^2) +\frac{m_t^4}{4 m_{Z'}^4}(g_L^2+g_R^2)^2(\hat{t}_{Z'}^2+4 \hat{s} m_{Z'}^2)\right],
\label{Z'sq}
\end{eqnarray}
with $\hat{t}_i \equiv \hat{t} - m_i^2$ and $\hat{u}_i \equiv \hat{u} - m_i^2$. The Mandelstam variables are related to the scattering angles via $\hat{t} = -\hat{s}(1-c_\theta)/2 + m_t^2$ and $\hat{u} = -\hat{s}(1+c_\theta)/2 + m_t^2$. Note that the Lagrangian has been defined with a $\sqrt{2}$ with respect to some other conventions in the literature. Similar expressions hold for the flavor-violating $W'$ via the interaction Lagrangian
\beq\label{eq:FVWprimeLagrangian}
{\cal L} = \frac{1}{\sqrt{2}}\bar{d} \gamma^\mu (g_L P_L +g _R P_R) t {W'}_\mu + \mbox{h.c.}
\eeq
\subsection{Color Triplets and Sextets}
The quantum numbers of the color triplet and sextet are
\beq
(\bar{3},1)_{4/3}~~~~(6,1)_{4/3},
\eeq
and their interactions with up and top quarks is given by
\beq\label{eq:TripSextetLagrangian}
{\cal L}_\phi = \phi^a \bar{t^c}T_r^a(g_L P_L + g_R P_R) u.
\eeq
This gives rise to a scattering cross-section \cite{Shu:2009xf}
\beq
{\cal A}_{int} + {\cal A}_{sq} = \frac{2 g^2 g_S^2 C_{(0)}}{9}\frac{\hat{u}_t^2 + \hat{s} m_t^2}{\hat{s} \hat{u}_\phi}+\frac{g^4 C_{(2)}}{9}\frac{\hat{u}_t^2}{\hat{u}_\phi^2}
\label{tripintsq}
\eeq
where $C_{(0)} = 1 (-1)$ for triplets (sextets) \cite{Shu:2009xf,Arhrib:2009hu} is a color factor that comes from the interference of new $t$-channel physics with the $s$-channel gluon. The color factor $C_{(2)}$ comes from the squared new $t$-channel physics term and is equal to $C_{(2)} = 3/2$ for sextets and $C_{(2)} = 3/4$ for triplets. We have also defined
$
g \equiv \sqrt{(g_L^2 + g_R^2)/2}.
$
\subsection{Color Octet}
The exotic gluon couples to light quarks through
\begin{eqnarray}\label{eq:ExoticGluonLagrangian}
{\cal L}_{axi} =
g_s \left(\bar{q} \, T^A \gamma^\mu(g_L^q P_L+g_R^q P_R) q + \bar{t} \, T^A \gamma^\mu(g_L^t P_L+g_R^t P_R) t \right) {G'}^A_\mu.
\end{eqnarray}
Note the inclusion of the QCD coupling constant, $g_s$, in the interaction. The scattering cross-sections calculated through these interactions are \cite{Cao:2010zb}
\begin{eqnarray}
{\cal A}_{int} & = & \frac{g_s^4}{9}\frac{\hat{s}(\hat{s}-m_{G'}^2)}{(\hat{s}-m_{G'}^2)^2+m_{G'}^2\Gamma_{G'}^2}(g^q_L+g^q_R)(g^t_L+g^t_R) \nonumber \\
&&\left[(2 - \beta^2)+2\frac{(g_L^q-g_R^q)(g_L^t-g_R^t)}{(g_L^q+g_R^q)(g_L^t+g_R^t)} c_\theta +c_\theta^2\right],
\end{eqnarray}
\begin{eqnarray}
{\cal A}_{sq} & = & \frac{g_s^4}{18}\frac{\hat{s}^2}{(\hat{s}-m_{G'}^2)^2+m_{G'}^2\Gamma_{G'}^2}({g_L^q}^2+{g_R^q}^2)({g_L^t}^2+{g_R^t}^2) \nonumber \\
&&\left[1+(1 - \beta^2)\frac{2 g_L^t g_R^t}{{g_L^t}^2+{g_R^t}^2}+2\frac{({g_L^q}^2-{g_R^q}^2)({g_L^t}^2-{g_R^t}^2)}{({g_L^q}^2+{g_R^q}^2)({g_L^t}^2+{g_R^t}^2)}c_\theta+c_\theta^2\right].
\end{eqnarray}
As does CDF, we consider the case where the couplings of the vector color octets are purely axial, so $g_V^q = (g_R^q + g_L^q)/2 = 0$ and $g_V^t = (g_L^t + g_R^t)/2 = 0$, and where the axial coupling of the boson to light quarks is positive and opposite the coupling of the boson to tops, so $g_A^t = (g_R^t - g_L^t)/2 = - g_A^q = (-g_R^q + g_L^q)/2$. This axigluon case leads to the largest positive contribution to the asymmetry per contribution to the total cross-section.
\section{Parton Level Tevatron Top Forward-Backward Asymmetry}
We now simulate the models described in the previous section, and we will use the formulae presented there to discuss the qualitative features.
\begin{figure}
\includegraphics[width = 0.95\textwidth]{pics/modelScatter.pdf}
\caption[Scatter plot of $A_{FB}(m_{t \bar{t}})$ versus LO cross-section.]{Parton level forward-backward asymmetry for events with $t \bar{t}$ invariant mass greater than $450$ GeV versus {\em leading-order} cross-section in picobarns for various models. The mass (in GeV) and coupling of the mediator are indicated by ${ \text{mass} \choose \text{coupling} } $. A comprehensive scan of models was carried out, and only representative points are shown. The coupling shown for axigluons is $g_A^q = -g_A^t$ (supposing $g_V^q = g_V^t = 0$) while the coupling shown for $Z_H'$ and $W'$ models is $g_R$ (assuming $g_L = 0$) and the quoted couplings for triplets and sextets is $g = \sqrt{(g_L^2 + g_R^2)/2}$. For comparison against observations, the horizontal shaded area lies in the $\pm 1\sigma$ region of the measured (parton-level) value of $A_{FB}(M_{t \bar{t}} > 450~\text{GeV})$. The vertical lines lie at the central value of the CDF $t\bar{t}$ production cross-section ($7.5 \pm 0.48$ pb), divided by a K-factor of 1.3 or 1. The SM marker lies at the value of the LO Standard Model cross-section and the NLO value for the SM forward-backward asymmetry. Note that care must be taken when comparing the new physics cross-sections against the Standard Model cross-section, as the selection efficiencies for the new physics models can be lower. This is discussed in more detail in the main text.}\label{modelScatterPlot}
\end{figure}
We begin by analyzing models for top asymmetry generation at the parton level. ``Parton level'' as used in the CDF $A_{FB}$ paper refers to de-convolving event selection efficiencies, detector efficiencies, jet algorithms, background, etc., from the underlying physics \cite{Aaltonen:2011kc}. Thus for the parton level analysis, we simulate and compare our results after showering in {\tt PYTHIA} but before folding in detector effects. To do this, we use {\tt MadGraph/MadEvent} with matrix element / parton shower (ME/PS) matching in the MLM scheme, which was implemented in the {\tt MadGraph/MadEvent} package using {\tt PYTHIA}. The events were generated using a fixed renormalization scale and factorization scale of 200 GeV.\footnote{The current version of {\tt MadGraph} uses $\alpha_s(m_Z) = 0.13$, which is substantially larger than the current measured value of $\alpha_s = 0.118$. Therefore our choice of the renormalization scale is effectively lower than the nominal scale, assuming $\alpha_s(m_Z) = 0.118$. Our choice reproduces well the known theoretical LO Standard Model value for the $t \bar{t}$ cross-section at Tevatron. } {\tt MadGraph5 v0.6.1 / MadEvent 4.4.44} with {\tt QCUT} = 30 and {\tt xqcut} = 20 was used to generate all signal and standard model events.
Showering and matching are done in order to improve accuracy, including the effects of single mediator production in inclusive $t \bar{t}$ events.
\begin{figure}
\includegraphics[height = 0.3\textwidth]{pics/efficiencies.pdf}
\includegraphics[height = 0.3\textwidth]{pics/efficiencies2bin.pdf}
\caption[Efficiencies by mass bins.]{ Efficiencies after showering events with {\tt PYTHIA} and making the cuts detailed in section~\ref{Sec:reconstructed} on final state leptons and jets, with the hatched regions corresponding to $1\sigma$ errors based on the limited statistics of the sample. No detector effects have been taken into account here. Black bars are for LO Standard Model. All samples are LO matched.}\label{efficiencies}
\end{figure}
\subsection{Cross-sections and Efficiencies}
We make a few observations about total cross-sections and invariant mass distributions before delving into a detailed analysis of the asymmetry. Fig.~(\ref{modelScatterPlot}) shows the parton level $A^{t \bar{t}}_{FB}$ for events with $m_{t \bar{t}} > 450$ GeV versus total leading order $p \bar{p} \rightarrow t \bar{t}$ + 0 or 1 jet cross-section for a swath of flavor-changing $W'$, $Z'_H$, triplet, sextet models, and axigluon models. A comprehensive scan of models was carried out, and only representative points are shown. The horizontal shaded band lies at $\pm 1\sigma$ values of the observed asymmetry in the high invariant mass bin, $M_{t\bar{t}} > 450 \mbox{ GeV}$. The vertical dashed lines correspond to the combined $t \bar{t}$ cross-section from CDF with 4.6 fb$^{-1}$; CDF measures a cross-section of $7.5 \pm 0.31 (\mbox{stat}) \pm 0.34 (\mbox{syst}) \pm 0.15(\mbox{lumi}) $, assuming a top mass of $m_t = 172.5$ GeV \cite{cdftotalxsection}. The predicted next-to-leading-order (NLO) SM cross-section at the value of the top mass we assumed in simulations, $m_t = 174.3$, is about $7.2$ pb \cite{Moch:2008ai}, whereas we find the LO SM cross-section is $5.6$ pb, implying a SM K-factor of about $7.2 / 5.6 \approx 1.3$. Of course, the NLO corrections to the new physics have not been calculated, so any comparison between the observed cross-section and the $t\bar{t}$ production cross-section is subject to some uncertainty. We do choose to show in Fig.~(\ref{modelScatterPlot}), however, the central value of the combined CDF $t\bar{t}$ production cross-section (7.5 pb) divided by the SM K-factor when comparing to the leading-order (LO) $t\bar{t}$ production cross-section of SM plus new physics against the observed production cross-section. From the figure it is clear that, in general, excepting the $Z'_H$ and axigluon models, models with couplings that are small enough to be in accord with the observed cross-section do not produce a large enough asymmetry in the high $t\bar{t}$ invariant mass bin.
\begin{figure}
\begin{center}
\includegraphics[width=0.32\textwidth]{pics/smcontours.png}\\
\includegraphics[width=0.95\textwidth]{pics/contours1.png}\\
\includegraphics[width=0.95\textwidth]{pics/contours2.png}
\caption[Event distributions densities for benchmark models.]{Event distribution densities for benchmark models at the parton level. The vertical axis is top quark pseudo-rapidity, and the horizonal axis it $t \bar{t}$ invariant mass.}\label{contours}
\end{center}
\end{figure}
This statement requires a strong qualification, however, which we investigate in detail below after carrying out a reconstruction of the top samples in these models. The qualification is that the efficiency for a $t\bar{t}$ event to pass cuts (the same as those used in the CDF analysis and our detector level analysis below, see Sec.~(\ref{Sec:reconstructed})) is strongly model dependent for cases where there is a large asymmetry. This is shown in Fig.~(\ref{efficiencies}), where we see that the efficiencies to pass cuts (after showering and jet clustering but no detector simulation) is suppressed by more than a factor of two relative to the Standard Model for the 400 GeV $Z'_H$ model shown and by about a factor of $1.5$ for the 800 GeV $Z'_H$ and the $W'$ model. The reason for this becomes clear after examining the distribution of events in top pseudorapidity, $\eta_t$, and and $t \bar{t}$ invariant mass, $m_{t \bar{t}}$, as shown in Fig.~(\ref{contours}). In order to generate a large asymmetry, the $\eta_t$ distribution must be skewed significantly with respect to the SM distribution. To generate a very large asymmetry in the high invariant mass bin, the distribution must be more skewed at high invariant mass. The distributions for the $Z'_H$ and $W'$ models are so skewed at high invariant mass that the peak of the distribution lies close to the $\eta_t = 1$ line. Thus a cut on lepton pseudorapidity of $|\eta| \leq 1$ and jet rapidity of $|\eta| \leq 2$ ends up cutting out a significantly greater fraction of events at high invariant mass than in the SM case.
Importantly, in unfolding the differential $t \bar{t}$ cross-section, assumptions about event selection efficiencies must be made; the assumption is that actual event selection efficiencies do not differ substantially from the Standard Model efficiencies \cite{Aaltonen:2009iz,CDF:thesis}. This assumption clearly breaks down in the case of the $Z'_H$ and $W'$. We investigate this effect in the reconstructed sample in the next section. For now, we compare the parton level asymmetries and cross-sections, and note caveats where this effect is important.
\begin{table}
\begin{tabular}{|c || c | c || c | c ||| c | c || c | c |}
\hline
& \multicolumn{4}{| c |||}{$m_{t \bar{t}} < 450 $ GeV } & \multicolumn{4}{| c |}{$m_{t \bar{t}} > 450 $ GeV } \\
\hline
& \multicolumn{2}{| c ||}{$ \Delta y < 0 $} & \multicolumn{2}{| c |||}{$\Delta y > 0$} & \multicolumn{2}{| c ||}{$\Delta y < 0 $} & \multicolumn{2}{| c |}{$\Delta y > 0 $} \\
\hline
Model & \emph{eff.} & $r$ & \emph{eff.} & $r$ & \emph{eff.} & $r$ & \emph{eff.} & $r$ \\
\hline
SM &0.079& 0.31&0.078& 0.3 &0.092& 0.2&0.089& 0.19\\
${Z_H'}^1$&0.078& 0.21&0.076& 0.22 &0.088& 0.15&0.063& 0.42\\
$W'$&0.079& 0.23&0.077& 0.27 &0.095& 0.16&0.075& 0.34\\
Triplet&0.084& 0.18&0.083& 0.23 &0.103& 0.2&0.095& 0.39\\
Sextet&0.075& 0.26&0.073& 0.28 &0.087& 0.19&0.08& 0.27\\
Axigluon&0.079& 0.26&0.077& 0.31 &0.096& 0.14&0.086& 0.28\\
${Z_H'}^2$&0.074& 0.18&0.072& 0.19 &0.089& 0.16&0.069& 0.47\\
\hline
\end{tabular}
\caption{Parton level efficiencies, \emph{eff.}, and bin fractions, $r$, for the Standard Model (SM) and for benchmark models ${Z_H'}^1$ ( $400$ GeV, $g_R = 1.75$), ${Z_H'}^2$ ( $800$ GeV, $g_R = 3.4$), $W'$ (400 GeV, $g_R = 2.55$), Triplet ($600$ GeV, $g = 4.4$), Sextet ($1.4$ TeV, $g = 4.0$), and Axigluon ($2$ TeV, $g_A^q = - g_A^t = 2.4$). Here $\text{\emph{eff.}} \equiv { \text{\# events in bin after cuts} \over \text{\# events in bin before cuts}}$ and $r \equiv { \text{\# events in bin after cuts} \over \text{total \# events after cuts}}$. }\label{coarse bin efficiencies}
\end{table}
Efficiencies can also affect the forward-backward asymmetry. If, for example, the efficiency for an event to pass cuts is lower for events with $\Delta y > 0$ than for events with $\Delta y < 0$, then cuts will wash out the asymmetry. We show the efficiency to pass cuts (after showering and jet clustering but no detector simulation) in four $m_{t \bar{t}}$, $\Delta y$ bins for the SM and for several benchmark models in Table~\ref {coarse bin efficiencies}. The difference between SM efficiencies and new physics model efficiencies is not as great for the coarse $m_{t \bar{t}}$ binning, implying that the effect on the coarse binned asymmetries reported by CDF will not be as great as for the more finely binned invariant mass distribution. However, we also see from the table that the efficiencies for $\Delta y > 0$ are smaller for some new models, such as the $Z'_H$ and $W'$ than for the SM and the other new models, implying more washout of the asymmetry for the $Z'_H$ and $W'$, which should be compensated. These effects will be seen when we compare our parton level asymmetries to the reconstructed asymmetries. The overall point here is that the efficiencies have some effect on unfolding the parton level asymmetries, but they are not as large an effect as on the invariant mass distribution.
\begin{table}
\begin{tabular}{| l | c | c | c| c| c | c| }
\hline
Model & \multicolumn{6}{|c|}{ Mass(GeV), coupling, cross-section(pb) } \\
\hline
FV $W'$ &{200, 1.4, 7.1}&{300, 1.8, 6.5}&{400, 2.4, 6.9}&{400, 2.6, 7.7}&{600, 3.4, 6.9}&{600, 3.6, 7.4}\\
\hline
FV $Z_H'$ &{300, 1.4, 6.}&{400, 1.6, 5.1}&{400, 1.8, 6.2}&{600, 2.4, 5.6}&{800, 3.2, 6.}&{800, 3.4, 6.8} \\
\hline
Triplet &{400, 3., 7.9}&{400, 3.2, 9.5}&{600, 3.6, 6.7}&{600, 3.8, 7.4}&{600, 4., 8.4} & \\
\hline
Sextet &{600, 2., 8.1}&{800, 2.4, 7.8}&{1000, 3., 8.}&{1200, 3., 7.1}&{1200, 3.4, 7.7}&{1400, 4., 7.7} \\
\hline
Axigluon &{2000, 2., 5.7}&{2000, 2.4, 5.8}&{2000, 3.2, 6.}&{2200, 3.2, 5.8}& & \\
\hline
\end{tabular}
\caption{Summary of benchmark models.
The coupling shown for axigluons is $g_A^q = -g_A^t$ (supposing $g_V^q = g_V^t = 0$) while the coupling shown for $Z_H'$ and $W'$ models is $g_R$ (assuming $g_L = 0$) and the quoted couplings for triplets and sextets is $g = \sqrt{(g_L^2 + g_R^2)/2}$.
For each model we consider, we include mass, coupling and total leading order matched $t\bar{t} + 0~\text{or}~1~\text{jet}$ production cross-section as calculated with {\tt MadGraph/MadEvent/Pythia}. The cross-sections above can be compared to the cross-section we obtained for the Standard Model using the same cuts and SM input parameters: 5.6 pb. No K-factors have been included in these quoted cross-sections. Note that care must be taken when comparing the new physics cross-sections against the Standard Model cross-section, as the selection efficiencies for the new physics models can be lower. This is discussed in more detail in the main text.}
\label{BenchmarkModels}
\end{table}
\subsection{Parton Level Asymmetries}
We now show the parton level asymmetries for each of the benchmark models. We choose several benchmark masses/couplings for $Z_H'$, $W'$, triplet, sextet, and axigluon models that give rise to large forward-backward asymmetries without generating too large a total cross-section. These benchmark models are listed in Table~\ref{BenchmarkModels}.
We show in Fig.~(\ref{ZpWpMttDely}) the forward-backward asymmetry in $M_{t\bar{t}}$ and $\Delta y$, comparing the reconstructed parton level asymmetry of the Tevatron to our simulated matched sample. The hatched regions correspond to the $1\sigma$ errors based on the limited statistics of the sample.
We can see that the horizontal $Z'$ and $W'$ can give good fits in both the high and low invariant mass bins, and in low and high $\Delta y$.
At least for the two bin case, the models also appear to give the correct shape as a function of $M_{t\bar{t}}$ and $\Delta y$.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{pics/ZhBestDMmatch.pdf}
\includegraphics[width=0.45\textwidth]{pics/ZhBestDYmatch.pdf}\\
\includegraphics[width=0.45\textwidth]{pics/WpBestDMmatch.pdf}
\includegraphics[width=0.45\textwidth]{pics/WpBestDYmatch.pdf}
\end{center}
\caption[Parton level $A_{FB}$ for $Z'_H,~W'$ models.]{$A_{FB}$ for $Z'_H,~W'$ models with couplings as indicated in the legend (with $g=g_R, g_L=0$), with the hatched regions corresponding to $1\sigma$ errors based on the limited statistics of the sample.
The contribution to $A_{FB}$ includes both $t$-channel $Z'_H,~W'$ exchange and single $Z'_H,~W'$ production.
The red bars are the CDF observation with $1\sigma$ errors, while the blue bars indicate the NLO SM contribution from \cite{Aaltonen:2011kc}, which has not been included in the LO contribution calculated via {\tt MadGraph} and {\tt PYTHIA}. The last bin includes all events with $m_{t \bar{t}} > 450$ GeV and $|\Delta y| > 1$, respectively.}
\label{ZpWpMttDely}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{pics/TripBestDMmatch.pdf}
\includegraphics[width=0.45\textwidth]{pics/TripBestDYmatch.pdf}\\
\includegraphics[width=0.45\textwidth]{pics/SixBestDMmatch.pdf}
\includegraphics[width=0.45\textwidth]{pics/SixBestDYmatch.pdf}\\
\includegraphics[width=0.45\textwidth]{pics/AxiDM.pdf}
\includegraphics[width=0.45\textwidth]{pics/AxiDY.pdf}
\end{center}
\caption[Parton-level $A_{FB}$ for triplet, sextet and axigluon models.]{$A_{FB}$ for triplet, sextet and axigluon models with couplings as indicated in the legend, with the hatched regions corresponding to $1\sigma$ errors based on the limited statistics of the sample. The red bars are the CDF observation with $1\sigma$ errors, while the blue bars indicate the NLO SM contribution, which has not been included in the LO contribution calculated via {\tt MadGraph} and {\tt PYTHIA}. The last bin includes all events with $m_{t \bar{t}} > 450$ GeV and $|\Delta y| > 1$, respectively.}
\label{TripMttDely}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{pics/ZpDsigDm.pdf}
\includegraphics[width=0.45\textwidth]{pics/WpDsigDm.pdf}\\
\includegraphics[width=0.45\textwidth]{pics/TripDsigDm.pdf}
\includegraphics[width=0.45\textwidth]{pics/SixDsigDm.pdf}\\
\includegraphics[width=0.45\textwidth]{pics/AxiDsigDm.pdf}
\end{center}
\caption[Parton-level $d \sigma \over d m_{t \bar{t}}$ for the benchmark models.]{$d \sigma \over d m_{t \bar{t}}$ for the benchmark models appearing in Figs.~(\ref{ZpWpMttDely},~\ref{TripMttDely}). The Tevatron measured cross section (red crosses, from \cite{Aaltonen:2009iz}), and LO SM cross-section with the same SM parameters and fixed renormalization scale used to generate benchmark model events are also shown. No K-factors are applied.}
\label{InvariantMassDistribution}
\end{figure}
A similar analysis is carried out for triplets, sextets and axigluons in Fig.~(\ref{TripMttDely}). While triplet and sextet models can marginally reproduce the asymmetries at LO, the rise between the low and high invariant mass bins and low and high rapidity bins is not as pronounced for the triplets and sextets as for the $W'$ and $Z'_H$. The reason for this in the sextet and triplet cases can, for example, be easily extracted from the analytical expressions, Eqs.~(\ref{Z'int}),~(\ref{Z'sq}),~(\ref{tripintsq}). At high invariant mass, the scattering amplitude is dominated by the squared term. There $\hat{u}_t \simeq \hat{u}_\phi$ in Eq.~(\ref{tripintsq}), and the effect of the ${\cal A}_{sq}$ on the asymmetry vanishes. By contrast, the $Z'_H,~W'$ are dominated by $\hat{u}_t^2/\hat{t}_M^2 \sim (1+c_\theta)^2/(1-c_\theta)^2$ which retains a contribution to the asymmetry at high invariant mass.
The axigluon models tend to significantly underproduce the asymmetry in the high invariant mass window. Choosing a larger coupling does not give rise to a larger asymmetry in the high invariant mass bin because of width effects, and the axigluon mass cannot be lowered in order to compensate on account of dijet constraints \cite{Bai:2011ed}. Thus we see that axigluon models have greater difficulty than $Z'_H$ and $W'$ for reproducing the observations. Some of these constraints can be relaxed somewhat by moving away from the point $g_A^q = - g_A^t$ \cite{Bai:2011ed}. In addition, on account of the large couplings present in these models, NLO corrections to the new physics must be considered in order to draw firm conclusions.
Before moving on to the fully reconstructed sample, we compare the invariant mass distributions of the LO {\tt PYTHIA} results against the observations in Fig.~(\ref{InvariantMassDistribution}). No K-factors for NLO corrections have been applied. All the models overproduce the extracted invariant mass spectrum in the high mass bins. Here again, however, the caveat must be applied that the $Z'_H$ and $W'$ models have lower selection efficiencies in the high invariant mass bins, so that one expects the discrepancy in the bins above $m_{t\bar{t}} \approx 500$ GeV to be greatly reduced for these models. On the other hand, the sextets and triplets severely overproduce the observed number of events and, based on Fig.~(\ref{contours}), are not helped by having lower selection efficiency in the high invariant mass bins. The axigluon models do not have as severe an overproduction problem, but also do not generate a large asymmetry, as can be seen in Fig.~(\ref{TripMttDely}). As commented earlier, it is of course possible that NLO corrections from the new physics will lead to significant changes in these distributions on account of the large couplings.
We now turn to the reconstructed sample from which it will be possible to make more quantitative statements about the observed versus model-dependent predicted invariant mass spectra and asymmetries.
\section{Fully Reconstructed Asymmetry and Invariant Mass Distributions}
\label{Sec:reconstructed}
To reconstruct the invariant mass spectrum and check the model dependence of the $t\bar{t}$ parton-level asymmetry extracted in \cite{Aaltonen:2011kc} for the class of models discussed here, we can send the showered $t\bar{t}$ events through {\tt PGS}. To select the $t\bar{t}$ signal, we take the same requirements as CDF in their analysis:
\begin{itemize}
\item Exactly one electron or muon with $p_T > 20$ GeV and $|\eta| < 1.0$.
\item Photon and $\tau$ veto.
\item At least four jets with $p_T > 20$ GeV and $|\eta| < 2.0$, with at least one of the jets having a $b$-tag.
\item $E_T^{miss} > 20$ GeV.
\end{itemize}
We must then fully reconstruct the decayed tops. We do a likelihood analysis on the lepton and jet kinematics to the $t \bar{t}$ hypothesis, using the algorithm described in our previous paper \cite{Gresham:2011dg}. The top is reconstructed out of the four hardest jets in the event. In order to gain enough statistics in the high invariant mass bin to reliably compare our results against the reconstructed CDF asymmetry, we generate 5 million $t\bar{t}$ events per model. Approximately 2\% of these events survive the cuts. In contrast to our previous analysis \cite{Gresham:2011dg}, but in accordance with the CDF analysis, we place no $\chi^2$ cut on the reconstruction of the tops. We explore later the effect of the $\chi^2$ cut on the size of the asymmetry.
Because of the large numbers of simulated and reconstructed events required, we consider only a representative subset of the models analyzed at the parton level in the previous section. We choose a 400 GeV $Z'_H$ with $g_R = 1.75$, a 400 GeV $W'$ with $g_R = 2.55$, a 600 GeV triplet with $g = 4.0$, a 1.4 TeV sextet with $g = 4.0$ and a 2.0 TeV axigluon with $g_A^q = -g_A^t = 2.4$.\footnote{Note that the $W'$ and $Z'_H$ models will require a triplet or higher Higgs representation in order to evade dijet constraints for the flavor-conserving $Z'$s, which exist in these models. On the other hand, we also compare a heavier 800 GeV $Z'_H$ to the lighter $W'$ and $Z'_H$, and find that it overproduces the high invariant mass spectrum.} As we will see, the CDF extraction of the parton level asymmetries and invariant mass spectra is somewhat model dependent, on account of the model dependent efficiencies shown in Fig.~(\ref{efficiencies}) and Table~(\ref{coarse bin efficiencies}), as well as detector effects.
We begin by comparing our reconstructed results against the CDF results for several models in Fig.~(\ref{ReconstructedWpZp}). The $A_{FB}^{t\bar{t}}$ shown there is as defined in Eq.~(\ref{AFBdef}), with the top and anti-top identified by the sign of the lepton. We see that models that reproduce the parton level asymmetry match well against the fully reconstructed CDF asymmetry. The $Z'_H$ and $W'$ models, however, receive a larger upward correction upon unfolding to the parton level than would be expected using SM efficiencies. The reason for this is clear from Table~(\ref{coarse bin efficiencies}): the efficiencies in the high invariant mass bin with $\Delta y > 0$ for $Z'_H$ and $W'$ models are lower than for the SM, leading to a greater washout of the asymmetry at the detector level. We also observe that the axigluon, while appearing to reproduce the reconstructed asymmetry marginally, tends to underproduce the unfolded parton level asymmetry as seen in Fig.~(\ref{TripMttDely}). We can also look at the partitioned asymmetry defined by
\begin{equation}
A^{t\bar{t}}(q, M_{t\bar{t},i}) = \frac{N((y_\ell - y_h)>0, q, M_{t\bar{t},i})-N((y_\ell - y_h) <0, q, M_{t\bar{t},i})}{N((y_\ell - y_h)>0, q, M_{t\bar{t},i})+N((y_\ell - y_h) <0, q, M_{t\bar{t},i})},
\label{AFBdeflep}
\end{equation}
where $y_\ell$ is the rapidity of the leptonic top, $y_h$ is the rapidity of the hadronic top, and $q$ is the charge of the lepton. The asymmetries obtained in this way are shown in Fig.~(\ref{ReconstructedWpZpTripSixPartitioned}).
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{pics/recoWpZpAFB.pdf}
\includegraphics[width=0.48\textwidth]{pics/recoTripSixAxiAFB.pdf}\end{center}
\caption[Reconstructed $A_{FB}^{t \bar{t}}(M_{t \bar{t},i})$ for benchmark models.]{$A_{FB}^{t \bar{t}}(M_{t \bar{t},i})$ for $W'$, $Z'_H$, triplet, sextet and axigluon models. Red crosses are the CDF values reconstructed from data. Blue crosses are the MC@NLO expectation. The last bin includes all events with $m_{t \bar{t}} > 700$ GeV.}
\label{ReconstructedWpZp}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{pics/recoWpZpAFBpartitioned.pdf}
\includegraphics[width=0.45\textwidth]{pics/recoTripSixAFBpartitioned.pdf} \\
\includegraphics[width=0.45\textwidth]{pics/recoAxiAFBpartitioned.pdf}
\end{center}
\caption[Reconstructed $ A_{FB}^{t \bar{t}}(q, M_{t \bar{t}})$ for benchmark models.]{$ A_{FB}^{t \bar{t}}(q, M_{t \bar{t}})$ as defined in Eq.\eqref{AFBdeflep} for $W'$, $Z_H'$, triplet, sextet, and axigluon models. Here, the data is divided according to the charge, $q$, of the lepton in the event. Black ($q >0$) and red ($q < 0$) crosses are the CDF values reconstructed from data (with background subtracted).}
\label{ReconstructedWpZpTripSixPartitioned}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.8\textwidth]{pics/recoEventCounts.pdf}
\end{center}
\caption[Number of expected reconstructed events versus $m_{t \bar{t}}$.]{Number of expected $t \bar{t}$ events with $5.3$ fb$^{-1}$ at the Tevatron, distributed over $m_{t \bar{t}}$. Events were passed through {\tt PGS} and then tops were reconstructed using the algorithm detailed in \cite {Gresham:2011dg}. The red bars indicate CDF's measurement, with expected background (as estimated by CDF) subtracted. The green histogram is our SM sample and the purple histograms represent model samples. For SM and model samples, a fixed number of events were generated, and then event counts were scaled appropriately for $5.3$ fb$^{-1}$ integrated luminosity.}
\label{ReconstructedEventCounts}
\end{figure}
We also compare the reconstructed invariant mass spectrum against that reported in \cite{Aaltonen:2011kc}. Since the efficiencies are lower for the $Z'_H$ and $W'$ than for the Standard Model, we may expect these models to agree better with the observations than suggested by the parton level invariant mass spectra shown in Fig.~(\ref{InvariantMassDistribution}).
We compare in Fig.~(\ref{ReconstructedEventCounts}) the simulated reconstructed invariant mass spectrum against that reported in \cite{Aaltonen:2011kc}. First we note the discrepancy at low invariant mass between all models (including the SM) and the observations, which we attribute to NLO corrections and to a difference between the {\tt PGS} detector simulation and the CDF simulation. However, we can see the effects of the efficiencies in the high invariant mass bins noted in Fig.~(\ref{efficiencies}). For example, we can see the efficiency correction does seem to bring the $W'$ model into agreement with the SM. On the other hand, the triplet model largely and almost uniformly overproduces the invariant mass distribution in all bins. While the sextet model appears in better agreement at high invariant mass, it underproduces the observed asymmetry as shown in Fig.~(\ref{ReconstructedWpZp}). In general, triplet and sextet models have greater difficulty producing the observed asymmetry while remaining consistent with the total cross-section and invariant mass distribution, as emphasized by Figs.~(\ref{modelScatterPlot},~\ref{efficiencies}).
There are two other comparisons that we are able to do with our fully reconstructed asymmetry. We are able to compare the center-of-mass versus lab frame asymmetries, which is shown in Fig.~(\ref{ReconstructedWpZpTripSixAxiLabCMafb}). While the models give rise to some difference between the CM and lab frames, the difference is less pronounced than what CDF observes.
We can also compare the asymmetries in the four and five jet samples, as shown in Fig.~(\ref{ReconstructedWpZpTripSixJetMult}). Here we see some washout of the asymmetry in the 5 jet sample, an effect that is observed in the CDF data.
\begin{figure}
\begin{center}
\includegraphics[width=0.30\textwidth]{pics/recoWpZpAFBlabVcm.pdf}
\includegraphics[width=0.30\textwidth]{pics/recoTripSixAFBlabVcm.pdf}
\includegraphics[width=0.30\textwidth]{pics/recoAxiAFBlabVcm.pdf}\end{center}
\caption[Reconstructed CM and lab frame $A_{FB}^{t \bar{t}}$.]{$A_{FB}^{t \bar{t}}$ and $A_{FB}^{p \bar{p}}$ in low ($m_{t \bar{t}} < 450$ GeV) and high ($m_{t \bar{t}} \geq 450$ GeV) $t \bar{t}$ invariant mass bins for $Z_H'$, $W'$, triplet, sextet, and axigluon models. Red / Blue crosses are the CDF values reconstructed from data in the lab / CM frames. Purple / Green crosses indicate the SM NLO predictions.}
\label{ReconstructedWpZpTripSixAxiLabCMafb}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=0.30\textwidth]{pics/recoWpZpAFBbyJets.pdf}
\includegraphics[width=0.30\textwidth]{pics/recoTripSixAFBbyJets.pdf}
\includegraphics[width=0.28\textwidth]{pics/recoAxiAFBbyJets.pdf}
\end{center}
\caption[Reconstructed 2-bin $ A_{FB}^{t \bar{t}}(m_{t \bar{t}})$ for 4- and 5-jet samples.]{$ A_{FB}^{t \bar{t}}$ in $t \bar{t}$ low ($m_{t \bar{t}} < 450$ GeV) and high ($m_{t \bar{t}} \geq 450$ GeV) invariant mass bins for benchmark models. The simulated data samples were partitioned according to whether the event had more than five jets with $p_T > 20$ GeV and $|\eta| < 2$. The CDF measured lab frame asymmetry for 4-jet and 5+-jet samples is shown as red crosses and black crosses, respectively. }
\label{ReconstructedWpZpTripSixJetMult}
\end{figure}
Lastly, though this is not considered in detail in the CDF analysis (a value for the raw asymmetry after a $\chi^2$ cut of $3$ is presented in Table XIV of \cite{Aaltonen:2011kc}), it is interesting to observe the effect of a $\chi^2$ cut in the top reconstruction on the size of the asymmetry.\footnote{Recall that we reconstruct tops by doing a $\chi^2$ fit on the lepton and jet kinematics to the $t \bar{t}$ hypothesis. The fit has three degrees of freedom.} We can see in Fig.~(\ref{ReconstructedZpTripChiSquaredCuts}) that a moderate $\chi^2$ cut increases the asymmetry especially in high invariant mass bins.
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{pics/ZhChiCuts.pdf}\end{center}
\caption[$A_{FB}^{t \bar{t}}(M_{t \bar{t},i})$ with various cuts on the $t \bar{t}$ reconstruction $\chi^2$.]{$A_{FB}^{t \bar{t}}(M_{t \bar{t},i})$ for $Z_H'$ model with various cuts on the $t \bar{t}$ reconstruction $\chi^2$. The last bin includes all events with $m_{t \bar{t}} > 700$ GeV. Red crosses are the CDF values reconstructed from data. CDF used a likelihood algorithm for top reconstruction, but made no $\chi^2$ cut.}
\label{ReconstructedZpTripChiSquaredCuts}
\end{figure}
\section{Lepton Asymmetry}
In order to avoid potential issues with the top reconstruction, one can also look at the asymmetries in di-leptons, where both tops decay leptonically. The CDF collaboration recently reported results from an analysis of di-leptonic $t \bar{t}$ events. In addition to reporting asymmetries obtained after reconstructing tops in events, they report the raw lepton asymmetry \cite{CDFLeptons}. We compare the raw lepton asymmetry in benchmark models at the showered parton level to the CDF measured value in Fig.~(\ref{LepAFB}a). To better compare with the CDF measurement, we show the asymmetry for only events that have electron rapidities in the range $|\eta| < 1.1$ or $1.2 < |\eta| < 2.8$ and muons with rapidities in the range $|\eta| < 1$, corresponding to the rapidity cuts placed on the leptons in their analysis. We also point out that examining the lepton asymmetry as a function of lepton-lepton invariant mass could be instructive. We have shown the lepton forward-backward asymmetry as a function of $m_{\ell^+ \ell^-}$ in Fig.~(\ref{LepAFB}b).
\begin{figure}
\begin{center}
\subfigure[]{\includegraphics[width=0.45\textwidth]{pics/lepsAFBDY.pdf} }
\subfigure[]{\includegraphics[width=0.45\textwidth]{pics/lepsAFB.pdf}}
\end{center}
\caption[Parton level lepton forward-backward asymmetry.]{Parton level lepton forward-backward asymmetry for $t \bar{t}$ events in which both tops decay leptonically, and in which both leptons pass the rapidity cuts corresponding to those used in the recent CDF analysis \cite{CDFLeptons}. Red points are extracted from the results in \cite{CDFLeptons}. {\bf (a): } $A_{FB (\Delta y_{\ell^+ \ell^-})} $ as a function of $|\Delta y| = |y_{\ell^+} - y_{\ell^-} | $. {\bf (b):} $A_{FB (\Delta y_{\ell^+ \ell^-})}$ $ = {N(y_{\ell^+} - y_{\ell^-} > 0) - N(y_{\ell^+} - y_{\ell^-} < 0) \over N(y_{\ell^+} - y_{\ell^-}> 0) + N(y_{\ell^+} - y_{\ell^-} < 0) } $ as a function of $\ell^+ \ell^-$ invariant mass. }
\label{LepAFB}
\end{figure}
\section{Early LHC Reach}
We now consider the feasibility of discovering such models in the early run of the LHC at 7 TeV. To this end, we make use of our previous results \cite{Gresham:2011dg}. According to these results, a 200 GeV $W'$ with coupling of 1 should be discoverable at $3 \sigma$ with 1 fb$^{-1}$ of data. For a 400 GeV $W'$, the coupling must be larger than 1.3, and for a 600 GeV $W'$, the requirement is a coupling of 1.8. Thus we see all of the $W'$ models giving rise to large asymmetries should be observable with an fb$^{-1}$ of data.
Likewise, for the $Z'_H$ model, a coupling larger than 0.7 is required to discover a 200 GeV state at 3 $\sigma$ with 1 fb$^{-1}$, while a coupling of 0.8 is required for a 400 GeV state, and a coupling of 1.2 for a 600 GeV state.
For the triplets, a coupling of $\sim 0.9$ is required for a 3 $\sigma$ discovery with 1 fb$^{-1}$ for 400 GeV or lower masses; for a 600 GeV triplet, the requirement strengthens to requiring a coupling of 1.3. Similar types of constraints can be obtained for the sextet models.
The broad conclusion here is that all of the $t$-channel models that we considered here to fit the Tevatron top forward-backward asymmetry should give rise to $3\sigma$ excesses with 1 fb$^{-1}$ at the LHC in the context of a top-jet resonance search.
According to the analysis in \cite{Bai:2011ed}, the axigluon benchmark models presented in this paper will be rapidly discoverable at the LHC through dijet events.
\section{Conclusions}
We examined models of new physics that could generate the top forward-backward asymmetry. We considered $W'$, $Z'_H$ triplet and sextet diquarks, as well as axigluon models. We compared the asymmetries produced by these models to those observed at the Tevatron, and concluded that of the models that generate a large enough asymmetry in the invariant mass bin $M_{t\bar{t}} > 450 \mbox{ GeV}$, the $Z'_H,~W'$ and axigluon models are the only ones that do not hugely overproduce the total $t\bar{t}$ production cross-section. To bring the $W'$ models into agreement with the total $t\bar{t}$ production cross-section extracted at the Tevatron, we noted an important effect: the efficiency to select $t\bar{t}$ events from $W'$ models is significantly lower than for the Standard Model. This same effect is also helpful in improving the agreement between the invariant mass spectra of the $W'$ and $Z'_H$ models with the Standard Model predictions (which agree with observations). Our result also differs from earlier studies on the diquark models which found that they could adequately produce the asymmetry without producing unduly large cross-sections. In order to further investigate the model-dependence in the extracted asymmetry and invariant mass spectra, we then proceeded to decay the top quarks and simulate detector effects, reconstructing the tops via a likelihood based algorithm. This allowed us to compare our results against the raw CDF results in the asymmetry as well as invariant mass spectra. We found that when this was done, some $W'$ and $Z'_H$ models adequately reproduced the invariant mass spectra. It also allowed us to compare our results against the CDF results for lab versus center-of-mass frames, as well as the 4 jet versus 5 jet asymmetries. We conclude that while the models reproduce the observed decrease in the asymmetry in the 5 jet sample, no appreciable difference occurs between the lab and center-of-mass frames. Lastly, we note that an LHC search at 7 TeV for top jet resonances could exclude at the $> 3 \sigma$ level any $Z'_H$, $W'$ or diquark model that produces the Tevatron asymmetry.
{\em Acknowledgments}: We thank Dan Amidei for many helpful conversations.
{\em Note added:} While this work was being finalized, \cite{Ligeti:2011vt} appeared, which explores sextet and triplet models. While our results agree quantitatively with theirs, our conclusions on the viability of these models for explaining the asymmetry are more pessimistic, on account of the total $t\bar{t}$ cross-section and invariant mass distribution.
|
\section{Introduction}
Accelerated expansion seems to play an important role in the dynamical history of the universe. There is a firm belief, at present time, that universe passed through inflationary phase at early times and there are growing evidences that it is accelerating at present. The study of large scale structures indicates that the universe is almost spatially flat and that dark energy accounts for about $70$ percent of the total energy content \cite{sal1,sal2,sal3,sal4}. Moreover, dark energy is believed to be the responsible of the acceleration of our expanding universe. This sort of fluid violates the strong version of the energy conditions. Besides, when the null version is also violated, the fluid is called phantom and then the universe may present future singularities at finite time \cite{stephane,shin,robert}. In the phantom phase, the energy density grows whereas it decreases in a non-phantom one. However, we know little about the nature of dark energy in general and of phantom fluid in particular, except for their negative pressure. Therefore, a large effort has been spent in recent years to explain this mystery \cite{saulo1,shinichi,salvatore,mano1,mano2,mano3}. Specially, this transition phenomenon known as quintom scenario is proposed by B. Feng and collaborators in \cite{feng}, and they found that it gives rise to the equation of state larger than $-1$ in the past and less than $-1$ today, satisfying current observations. For understanding the possible connections among the dark energy models, it is useful to study the cosmic duality. Then, it was studied the duality in two-field quintom models of dark energy and it has been found that an expanding universe dominated by quintom-A field is dual to a contracting universe with quintom-B field \cite{yifu2}. Recently, Yi-Fu. Cai and collaborators wrote a paper which introduced the experimental developments on finding the transition between quintessence and phantom phases and various theoretical realizations of such a scenario, see \cite{yifu}. On the other hand, the knowledge of some properties of the universe in the phantom phase, such as the singularities, motivated various investigations with the aim of dealing with them. A. B. Batista and collaborators \cite{brasil} investigated the effects of particle production when a massless minimally coupled scalar field is present in spacetimes where $\omega$ is a constant. To do so they used a state for which Bunch and Davies \cite{bunch} had previously computed the stress-energy tensor. They found that the energy density of the created particles never dominates over the phantom energy density. In the same way, quantum effects near the big rip is studied in \cite{flavio} where they used the n-wave regularization for calculating the energy density of particle creation and found that, in this case, it tends to infinity when the big rip is approached and becomes the dominant component of the universe. This means that the big rip can be avoided by a scalar massless field. Pavlov \cite{pavlov} computed both the number density of created particles and the stress-energy tensor for a conformally coupled massive scalar field for the case in which $\omega = -5/3$. It was found that quantum effects are not important for masses much smaller than the Planck mass and times which are early enough that the time until the Big Rip occurs is greater than the Planck time. J. D. Bates and P. R. Anderson \cite{anderson} used a background field approach in which the energy densities of the quantized fields are computed in the background spacetime which contains the Big Rip singularity. They found that for fields in realistic states for which the energy density of the quantized fields is small compared to that of the phantom energy density at early times, and for spacetimes with realistic values of $\omega$, there is no evidence that quantum effects become large enough to significantly affect the expansion of the spacetime until the spacetime curvature is of the order of the Planck scale or larger, at which point the semi classical approximation breaks down. Also in order to deal with singularity problem, it has been considered by Yi-Fu Cai and collaborators \cite{yifu3} the cosmology of the Higgs sector of the Lee-Wick Standard Model, an alternative to supersymmetry to solving the hierarchy problem. They found that homogeneous and isotropic solutions are non-singular and then, the Lee-Wick model can provide a possible solution of the cosmological singularity problem.
\par
One of the phenomenological ways to explain the dark energy problem is assuming a variable cosmological constant. The cosmological constant $\Lambda$ is pretty compatible with observation and effort is currently devoted to the investigation of the theoretical foundations of a variable cosmological constant and its model properties \cite{strominger,dymnikova1,dymnikova2,khlopov}. In this respect, various ansatz have been used. For example in \cite{saulo2}, S. Carneiro and collaborators considered a cosmological constant $\Lambda \propto H$, and studied a possible way to distinguish the validity of this scenario from the standard one. Note that a model $\Lambda \propto a^{-2}$ has earlier been proposed in \cite{ozer}, requiring that the cosmic density $\rho$ equal to the Einstein-de Sitter critical density $\rho_c$, leading to a close universe without singularity, horizon, entropy and monopole problems. Furthermore, a large number of phenomenological $\Lambda$ models have been constructed in order to describe the dynamics of the universe \cite{overduin}. The case which attracts our attention in this paper is $\Lambda\propto H^{2}$ and has been studied in several other works for other purposes, but always with the aim of clarifying some grey areas in cosmology \cite{freese,carvalho,lima1,lima2}. In this paper we propose to use $\Lambda\propto H^2$, for analysing the evolution of the universe, being in the phantom or non-phantom phase. The phase transition of the universe will be studied considering a model in which dark energy is described by some rather complicated ideal fluid with an unusual equation of state (EoS) which will be chosen to be an inhomogeneous one. It is important to emphasize that this inhomogeneous EoS corresponds to a pure dark energy models. This kind of models may reproduce late-time acceleration, but it is not easy to construct a model that keeps untouched the radiation and matter dominated epochs. However, it is easy to introduce into our considerations ordinary matter and radiation but in that case they only appear suddenly at some point.
Note that this kind of study has been also considered in \cite{brevik2} where the cosmological constant is a linear function of time. Here, we use the ansatz $\Lambda \propto H^2$ with the inhomogeneous EoS. We find that, depending one the choice of the input parameters of the EoS considered, the universe may transit from a non-phantom to a phantom phase leading to finite time singularities. Another interesting point of our result is that, in contrast to the model without variable cosmological constant in which the phantom universe ends with the singularity of type I (big rip), the phantom universe in this case ends with the singularity of type III.
\par
Another point we address here is the Cardy-Verlinde (CV) formula coming from inhomogeneous EoS. Verlinde \cite{verlinde} made an interesting proposal that Cardy formula \cite{cardy} in two-dimensional conformal field theory can be generalized to arbitrary spacetime dimensions. Verlinde further proposed that a closed universe has subextensive (Casimir) contribution to its energy and entropy with the Casimir energy conjectured to be bounded from above
by the Bekenstein-Hawking energy and as consequence, one obtains a very deep relation between gravity and thermodynamics \cite{youm}. Within the context of the radiation dominated universe, such bound on the Casimir energy is shown to lead to the Hubble and the Bekenstein entropy bounds respectively for the strongly and the weakly self-gravitating universes. The generalized entropy formula, called the CV formula, is further shown to coincide with the total entropy of the universe coming from the Friedmann equations. These results were later generalized \cite{group1,group2,group3,group4,group5,group6,group7,group8}. Our goal here is to analyse the equivalence between the CV formula and the total entropy coming from Friedmann equations assuming that the universe is conformally invariant. With this, we find that for the inhomogeneous EoS, the generalized entropy of the universe reduces to the CV formula with a special choice of the input parameter $m$ and this does not correspond to a radiative universe as in the case of homogeneous EoS. Note also that this equivalence occurs exclusively at the presente time .\par
The paper presents two sections, the first showing the inhomogeneous EoS with which the solution for the Hubble parameter is found and some discussions on the input constants are put forward, allowing to a whole analysis of the transition phenomenon. The second section shows a brief concept on the CV formula with homogeneous EoS and latter, a complete analysis of the equivalence between the CV formula and the Friedmann equations with inhomogeneous EoS. Finally, we present our conclusions and perspectives.
\section{Solving the inhomogeneous equation of state}
Les us consider the universe driven by an ideal fluid (dark energy) with the inhomogeneous equation of state \cite{brevik1}
\begin{equation}\label{e1}
p=\omega(t)\rho+\Lambda(t)\quad,
\end{equation}
where $\omega(t)$ and $\Lambda(t)$ depend on the time and $\rho$ and $p$ are respectively the energy density and the pressure of the fluid. This equation, for the case $\Lambda(t)=0$ and $\omega(t)$ as affine function of time, has been studied in \cite{brevik1,nojiri1}. Moreover, the case $\Lambda(t)\neq 0$ as a affine function of time has been examined in \cite{brevik2}.\par
The equation of energy conservation and the Friedmann equations are respectively
\begin{eqnarray}
\dot{\rho}+3H\left(\rho+p\right)=0\label{e2}\quad,\\
\frac{3}{\kappa^2}H^2=\rho\label{e3}\quad,\\
\frac{1}{\kappa^2}\left(2\dot H+3H^2\right)=-p\label{pressure}\quad,
\end{eqnarray}
where $\kappa^2=8\pi G$, with $G$ the gravitational constant and $H=\frac{\dot a}{a}$, the Hubble parameter where $a(t)$ is the scale factor. \par
Using (\ref{e1}) and (\ref{e3}), equation (\ref{e2}) is rewritten as
\begin{eqnarray}
\dot{\rho}+\sqrt{3}\kappa\left[1+\omega(t)\right]\rho^{3/2}+\sqrt{3}\kappa\rho^{1/2}\Lambda(t)=0\label{e4}\quad.
\end{eqnarray}
From now on, we suppose that $\omega(t)$ depends linearly on time and the cosmological constant $\Lambda(t)$ is proportional to the square of the Hubble parameter, that is
\begin{eqnarray}
\omega(t)&=& \alpha t+\beta \label{e5}\quad,\\
\Lambda(t)&=&\gamma H^2(t)\nonumber\\
&=&\frac{\gamma\kappa^2}{3}\rho(t)\label{e6}\quad.
\end{eqnarray}
Taking into account (\ref{e5}) and (\ref{e6}), equation (\ref{e4}) becomes
\begin{eqnarray}\label{e7}
\dot{\rho}+\left(At+B\right)\rho^{3/2}=0\quad,\quad A=\alpha\kappa\sqrt{3}\,\,\,,\quad B=\kappa\sqrt{3}\left(1+\beta+\frac{\gamma\kappa^2}{3}\right)\,\,\,.
\end{eqnarray}
The solution of this equation is
\begin{equation}\label{e8}
\rho(t)=\frac{16}{\left(At^2+2Bt-2C\right)^2}\,\,\,,
\end{equation}
and the Hubble parameter and its rate behave as
\begin{equation}\label{e9}
H(t)=-\frac{4\kappa}{\sqrt{3}\left(At^2+2Bt-2C\right)}\,\,\,,
\end{equation}
\begin{eqnarray}\label{e10}
\dot{H}(t)=\frac{8\kappa\left(At+B\right)}{\sqrt{3}\left(At^2+2Bt-2C\right)^2}\,\,,
\end{eqnarray}
where $C$ is an integration constant.\par
As we are dealing with an expanding universe, we need an increasing scale factor, that is $\dot{a}>0$. We know that $\dot{a}= Ha$, then for an expanding universe the Hubble parameter has also to be positive. Solving $\dot{a}= Ha$, one obtains
\begin{eqnarray}
a(t)&=&\exp{\left(\int H(t)dt\right)}\nonumber\\
&=&\exp{\left(\frac{4\kappa}{\sqrt{3}}\frac{g(t)}{\sqrt{B^2+2AC}}\right)}\,\,,\quad g(t)=\arctan{\left(\frac{At+B}{\sqrt{B^2+2AC}}\right)}\label{e11}\,\,.
\end{eqnarray}
The positivity of $\dot{a}$ depends on the sign of $H(t)$. Note that the expression $At^2+2Bt-2C$ vanishes for $t_{1,2}=-\left(B\pm\sqrt{B^2+2AC}\right)/A$. The universe expands when the Hubble parameter is positive and one has a phantom fluid when the weak version of the energy conditions is violated, that is when $\rho+p<0$. Combining equations (\ref{e3}) and (\ref{pressure}), one obtains
\begin{eqnarray}
\rho+p=-\frac{2}{\kappa^2}\dot{H}\label{e12}\,\,\,,
\end{eqnarray}
and it turns out that the phantom phase (respectively the non-phantom one) is obtained when $\dot{H}>0$ ($\dot{H}<0$). Two situations are important for a whole analysis: when the constant $A$ is positive or negative.\par
$\bullet$ {\bf Analysis for the case $A>0\, (\alpha>0)$}\par
In this case, a simple study of the Hubble parameter sign shows that one has an expanding universe for $t_1<t<t_2$ and a contracting one when $t<t_1$ and $t>t_2$. On the other hand, it is easy to see that $\dot{H}$ vanishes for $t_3=-B/A$. Then, the universe is in the phantom phase ($\dot{H}>0$) when $t>t_3$ and the energy density grows; for the non-phantom phase ($\dot{H}<0$) $t<t_3$, the energy density decreases. Consequently, the accelerated expanding universe begins with a non-phantom phase and enters in the phantom one at the transition time $t_{tr}=t_3$, the time at which the Hubble parameter and the energy are
\begin{eqnarray}
H_{tr}= \frac{4\kappa A}{\sqrt{3}\left(B^2+2AC\right)},\quad \rho_{tr}=\frac{16A^2}{\left(B^2+2AC\right)^2}\,\,\,.
\end{eqnarray}
In the non-phantom case, the energy density decreases and tends to $\rho_{tr}$ as the time goes to $t_{tr}$.
The simultaneous divergence of $\rho(t)$ and $H(t)$ appears at $t_1$ and $t_2$. However, $a(t_2)=\exp{\left( \kappa\pi/\sqrt{3(B^2+2AC)} \right)}$, which is finite. In fact, the phantom universe ends with a future singularity, in this case, the singularity is of type III (for a classification of future singularities see \cite{stephane}) since the energy density and the pressure at this time are divergent. The graph of $H(t)$ is shown in the left panel of Fig. $1$.\par
$\bullet$ {\bf Analysis for the case $A<0 \,(\alpha<0)$}\par
For $t<t_1$ and $t>t_2$, the universe expands whereas it contracts for $t_1<t<t_2$. The first derivative of the Hubble parameter is positive (respectively negative) for $t<t_3$ $(t>t_3)$. The expanding universe begins with a phantom phase which ends at $t_1$ and enters in a non-phantom phase at $t_2$. Here the transition is not instantaneous, it is the contracting phase of the universe. At $t_1$, the energy density $\rho(t_1)$ and the pressure $p(t_1)$ diverge. However, at $t_1$, the scale factor is finite, $a(t_1)=\exp{\left( - \kappa\pi/\sqrt{3(B^2+2AC)} \right)}$. Then, the phantom phase ends with the singularity of type III. In the non-phantom phase, the energy density decreases and goes to zero as $t\longrightarrow \infty$. The graph of the Hubble parameter, $H(t)$, versus time $t$ is shown in the right panel of Fig. $1$. \par
\par
\section{ CV formula from inhomogeneous EoS fluid}
This section is devoted to the application of CV formula to ideal fluids. In a first step, let us tell briefly introduce CV formula for homogeneous EoS. We consider a (n + 1)-dimensional space time described by the FRW metric, written in comoving coordinates as
\begin{eqnarray}\label{cv1}
ds^2=dt^2-\frac{a^2(t)dr^2}{1-kr^2}-r^2d\Omega^{2}_{n-1}\,\,,
\end{eqnarray}
where $k = -1, 0, +1 $ for an open, flat, or closed spatial Universe respectively, and $d\Omega^{2}_{n-1}$ is the metric of an $n-1$
sphere. Then, by inserting the metric (\ref{cv1}) in the Einstein equations the Friedmann equations are derived,
\begin{eqnarray}\label{cv2}
H^2=\frac{16\pi G}{n(n-1)}\rho-\frac{k}{a^2}\,\,\,, \quad\quad \dot{H}= -\frac{}{}\left(\rho+p\right)+\frac{k}{a^2}\,\,.
\end{eqnarray}
The total energy $E$ of the universe is $E=\rho V$, with $V$ its total volume. Since the Casimir energy may be include in the total energy, we consider a closed universe, $k=1$. For the homogeneous EoS, $p=\omega \rho$, with $\omega$ a constant. Then the conservation law for energy has the form
\begin{eqnarray}\label{cv3}
\dot{\rho}+nH\left(1+\omega\right)\rho=0\quad,
\end{eqnarray}
which reduces to (\ref{e2}) for $n=3$ and whose solution depends on the scale factor as
\begin{eqnarray}\label{cv4}
\rho \propto a^{-n(1+\omega)}\quad.
\end{eqnarray}
The total energy of the universe can be written as the sum of an extensive part $E_E$ and a subextensive part $E_C$, called the Casimir energy, and it takes the form:
\begin{eqnarray}\label{cv5}
E(S,V)= E_E(S,V)+\frac{1}{2}E_C(S,V)\,\,\,.
\end{eqnarray}
Under the transformations $S\rightarrow \xi S$ and $V\rightarrow \xi V$ with a constant $ \xi $, the extensive and
the subextensive parts of the total energy respectively scale as \cite{sergei1,cai}
\begin{eqnarray}\label{cv6}
E_E(\xi S,\xi V)= \xi E_E(S,V),\quad E_C(\xi S,\xi V)= \xi^{1-\frac{2}{n}}E_C(S,V)\quad,
\end{eqnarray}
and therefore, we have for the total energy
\begin{eqnarray}\label{cv7}
E(\xi S, \xi V)= \xi E_E(S,V)+\frac{1}{2}\, \xi^{1-\frac{2}{n}}E_C(S,V).
\end{eqnarray}
Taking the derivative of (\ref{cv7}) with respect to $\xi$ and letting $\xi=1$, one obtains
\begin{eqnarray}\label{cv8}
S\left(\frac{\partial E}{\partial S}\right)_V+V\left( \frac{\partial E}{\partial V}\right)_S = E_E+\left(\frac{1}{2}-\frac{1}{n}\right)E_C\,\,.
\end{eqnarray}
Assuming that the universe satisfies the first law of thermodynamics $dE = TdS-pdV$, we have the thermodynamics relations $\left(\frac{\partial E}{\partial V}\right)_S= - p$ and $\left(\frac{\partial E}{\partial S}\right)_V = T$. Using these thermodynamics relations and Eq. (\ref{cv5}), one can put Eq. (\ref{cv8}) into the following form for the Casimir energy, as the violation of the Euler identity
\begin{eqnarray}\label{cv9}
E_C= n \left( E+pV- TS\right)\,\,.
\end{eqnarray}
Since the total energy behaves as $E \sim a^{-n\omega}$ and by Eq. (\ref{cv5}), the Casimir energy also goes as $E_C \sim a^{-n\omega}$. The FRW Universe expands adiabatically ($dS = 0$) so the products $E_C a^{n\omega}$ and $E_E a^{n\omega}$ should be independent of the
volume V , and be just a function of the entropy. Then, by the rescaling properties (\ref{cv6}), the extensive and subextensive
parts of the total energy can be written as functions of the entropy only \cite{youm},
\begin{eqnarray}\label{cv10}
E_E=\frac{\mu}{4\pi a^{n\omega}}S^{\omega+1}\,\,,\quad\quad E_C=\frac{\nu}{2\pi a^{n\omega}}S^{\omega+1-2/n}\,,
\end{eqnarray}
where $\mu$ and $\nu$ are undetermined constants and $4\pi$ and $2\pi$ are used for convenience. From these expressions for $E_E$ and $E_C$, one obtains the following expression for the entropy of the universe:
\begin{eqnarray}\label{cv11}
S= \left(\frac{2\pi a^{n\omega}}{\sqrt{\mu\nu}}\sqrt{E_C(2E-E_C)}\right)^{\frac{n}{n(\omega+1)-1}}\,.
\end{eqnarray}
This result, obtained in \cite{youm}, reduces to the CV formula when the universe is radiation dominated, $\omega = 1/n$, that is
\begin{eqnarray}\label{cv12}
S= \frac{2\pi a}{\sqrt{\mu\nu}}\sqrt{E_C(2E-E_C)}\,.
\end{eqnarray}
Let us now look to the case of the inhomogeneous EoS that we used in the precedent section and analyse the relationship between the CV formula and the entropy of the universe. As has been done in \cite{brevik12}, we assume an EoS expressed as a function of the scale factor and described by
\begin{eqnarray}\label{cv13}
p=\omega(a)\rho+j(a)\,\,.
\end{eqnarray}
Introducing (\ref{cv13}) in the energy conservation equation (\ref{cv3}), one obtains
\begin{eqnarray}\label{cv14}
\rho^{\,\prime}(a)+\frac{n(1+\omega(a))}{a}\rho(a)= - n\frac{j(a)}{a}\,\,,
\end{eqnarray}
where the prime denotes the derivative with respect to the scale factor and we took $t=t(a)$. The general solution of (\ref{cv14}) is
\begin{eqnarray}\label{cv15}
\rho(a)= e^{-F(a)}\left( Q-n\int e^{F(a)}\frac{j(a)}{a}da\right)\,\,, \mbox{with}\quad F(a)= n\int^{a}\frac{1+\omega(a^{\prime})}{a^{\prime}}da^{\prime}\,\,,
\end{eqnarray}
where $Q$ is an integration constant. In this analysis, making use of (\ref{e6}) and taking into account the derivative with respect to the scale factor, (\ref{e1}) can be written as
\begin{eqnarray}\label{cv16}
\rho^{\prime}(a)+\frac{n(1+\bar{\omega}(a))}{a}\rho(a)=0\,\,,\quad\quad \bar{\omega}(a)= \frac{\gamma\kappa^2}{3}+\omega(a)\,\,.
\end{eqnarray}
Identifying (\ref{cv16}) with (\ref{cv14}), one gets
\begin{eqnarray}\label{cv17}
j(a)=0\,,\quad F(a)= n\int^{\,a}\frac{1+\bar{\omega}({a^{\prime}})}{a^{\prime}}da^{\prime}\,\,,
\end{eqnarray}
from which one obtains the energy density as
\begin{eqnarray}\label{cv18}
\rho(a)= Q e^{-F(a)}\,\,.
\end{eqnarray}
On the other hand, using (\ref{e11}), one can write
\begin{eqnarray}\label{cv19}
\bar{\omega}(a)=\frac{\sqrt{B^2+2AC}}{\kappa\sqrt{3}}\tan{\left[ \ln{\left(a^{\sqrt{3(B^2+2AC)}/(4\kappa)}\right)} \right]}\,\,.
\end{eqnarray}
Making use of (\ref{cv16}), (\ref{cv17}), (\ref{cv19}) and (\ref{cv18}), the energy density is written as
\begin{eqnarray}\label{cv20}
\rho(a)=Q\,\left[ \cos{\left(\frac{\sqrt{3(B^2+2AC)}}{4\kappa} \ln{(a)} \right)}\right]^{\frac{4n}{3}}\,\,.
\end{eqnarray}
Note here that only for some special conditions of the functions $\omega(a)$ and $j(a)$ CV formula (\ref{cv12}) can be recovered . Let us assume that the present time is $t_0=0$ and then analyse the equivalence between CV formula and the total entropy of the universe at this moment. Note that as $t\rightarrow t_0$, $\bar{\omega}(a)\rightarrow m-1$, with $m=1+\beta+\frac{\gamma\kappa^2}{3}$. Then, $F(a)\rightarrow nm\ln{(a)}$ and $\rho\propto a^{-nm}$. Hence, the total energy in the volume $V=a^n$ behaves as $E=\rho V\propto a^{n(1-m)}$, which is the same behaviour for the extensive and subextensive energy through (\ref{cv9}) and (\ref{cv5}). If we assume the conformal invariance, the products $E_E a^{n(m-1)}$ and $E_C a^{n(m-1)}$ do not depend on the volume and are only functions of entropy. Then, we have for the extensive and subextensive energy,
\begin{equation} \label{cv21}
E_E= \frac{\mu}{4\pi a^{n(m-1)}}S^{m}\,\,, \quad E_C= \frac{\nu}{2\pi a^{n(m-1)}}S^{m-\frac{2}{n}}\,\,,
\end{equation}
from which we determine the entropy as
\begin{eqnarray}\label{cv22}
S=\left[\frac{2\pi n a^{n(m-1)}}{\sqrt{\mu\nu}}\sqrt{E_C(2E-E_C)}\right]^{\frac{n}{nm-1}}\,\,.
\end{eqnarray}
Then, for $m=\frac{n+1}{n}$, the CV formula is recovered. However, for any $m\neq \frac{n+1}{n}$ CV formula can not be reproduced. Note in this case that the universe does not correspond to a radiative one.
\section{Conclusion}
We studied the transition of the universe between a phantom and a non-phantom phases. Note that in the non-phantom phase, the energy density decreases while in the phantom one, it grows leading to singularities. We focused our attention on the variable cosmological constant which has been introduced in the EoS, which becomes inhomogeneous. Then, we solved the equation of motion which led to the explicit expression of the energy density and consequently to that of the Hubble parameter. The first derivative of the Hubble parameter played a crucial role in this analysis since it allowed us to know which time interval corresponds to the phantom or non-phantom universe. The input parameter $A$ also appeared to be an important one.\par
In the first part, with the EoS considered and the ansatz $\Lambda\propto H^2(t)$, two important cases have been found. For $A>0$, we saw that the universe evolves from a non-phantom phase, $t_1<t<t_{tr}$, to a phantom one $t_{tr}<t<t_2$. In the non-phantom phase, the energy density descreses and goes to the energy density at the transition time while in the phantom phase, the energy density and the pressure grow and go to infinity at the finite time $t_2$. At the same time the scale factor remains finite and we conclude that the universe ends with the singularity of type III. For $A<0$, we saw that the universe begins with a phantom phase which ends with the singularity of type III at $t_1$, enters in the non-phantom at $t_2$ where the energy density descreses end goes to zero as $t\longrightarrow\infty$. Here, the transition from the phantom phase to the non-phantom one is not instantaneous; it is a time interval corresponding to the contracting phase of the universe.\par
However, in the case in which the phantom phase precedes the contracting phase of the universe, it would be interesting to study the possible avoidance of this type of singularity introducing either the viscosity term in the cosmic fluid or taking into account quantum effects. On the other hand, the same analysis can be done with the ansatz that the variable cosmological constant is proportional to the Hubble parameter, $\Lambda(t)\propto H(t)$. We will address these considerations in a future work.
\par
We also analyse the equivalence between the CV formula and the Friedmann equations with inhomogeneous EoS. Note that this has been done in several works with the homogeneous EoS and the Friedmann equations coincide with CV formula only in radiative universe. In this work, we use the inhomogeneous EoS including the variable cosmological constant proportional to the square of the Hubble parameter. We find that the equivalence between the Friedmann equations occurs only at the present time in a special case which is not the radiative universe as for the homogeneous EoS. \par
\vspace{1cm}
{\bf Acknowledgement:} The author thanks professors S. D. Odintsov, S. Carneiro and O. Piattella for criticism and comments, and CNPq (Brazil) for partial
financial support.
\begin{figure}[htt!!!]
\begin{minipage}[t]{0.55\linewidth}
\includegraphics[width=\linewidth]{courbehubblepostdoc11.eps}
\end{minipage} \hfill
\begin{minipage}[t]{0.55\linewidth}
\includegraphics[width=\linewidth]{courbehubblepostdoc1.eps}
\end{minipage} \hfill
\caption{{\protect\footnotesize The Hubble parameter as function of the cosmic time with $B=1$, from the left to right, $A=C=1$ and $A=C=-1$ respectively.}}
\label{}
\end{figure}
|
\section{Emergent Gravity}
Consider the $k-$essence scalar field $\phi$ minimallly coupled to the gravitational
field $g_{\mu\nu}$. Then the $k-$essence action is
+$$
S_{k}[\phi,g_{\mu\nu}]= \int d^{4}x {\sqrt -g} L(X,\phi)
\eqno(1)$$
where $X={1\over 2}g^{\mu\nu}\nabla_{\mu}\phi\nabla_{\nu}\phi$ and $\nabla_{\mu}$ means the
covariant derivative asociated with the metric $g_{\mu\nu}$. The total action describing the
dynamics of $k-$essence and gravity is
$$
S[\phi,g_{\mu\nu}]= \int d^{4}x {\sqrt -g}[ -{1\over 2} M^{2}_{\mathrm Pl}R + L(X,\phi)]
\eqno(2)$$
where $R$ is the Ricci scalar and $M_{\mathrm Pl}$ the reduced Planck mass.
The energy momentum tensor for the $k-$essence field is (with $L_{\mathrm X}= {dL\over dX},~~ L_{\mathrm XX}= {d^{2}L\over dX^{2}},
L_{\mathrm\phi}={d\phi\over dX}$)
$$
T_{\mu\nu}= {2\over \sqrt {-g}}{\delta S_{k}\over \delta g^{\mu\nu}}= L_{X}\nabla_{\mu}\phi\nabla_{\nu}\phi - g_{\mu\nu}L
\eqno(3)$$
and the equation of motion for the $k-$essence field is
$$-{1\over \sqrt {-g}}{\delta S_{k}\over \delta \phi}= \tilde G^{\mu\nu}\nabla_{\mu}\nabla_{\nu}\phi +2XL_{X\phi}-L_{\phi}=0
\eqno(4a)$$
where the effective metric $\tilde G^{\mu\nu}$ is
$$\tilde G^{\mu\nu}= L_{X} g^{\mu\nu} + L_{XX} \nabla^{\mu}\phi\nabla^{\nu}\phi
\eqno(5a)$$
and is physically meaningful only when
$$1+ {2X L_{XX}\over L_{X}} > 0$$
i.e the sound speed $c_{\mathrm s}=(1+ {2X L_{XX}\over L_{X}})^{-1/2}$ is
a real quantity.
When this condition holds everywhere the effective metric $\tilde G^{\mu\nu}$ determines the
characteristics for $k-$essence \cite{armen4,gib3,gib4,ren} .For the non-trivial configurations
of the $k-$ essence field $\partial_{\mu}\phi\neq 0$ and $\tilde G^{\mu\nu}$ is not conformally
equivalent to $g^{\mu\nu}$. So the characteristics are different from canonical
scalar fields whose lagrangians are linear in $X$. The characteristics determine the
local causal structure of the spacetime at every point of the manifold. So the local causal
structure for the $k-$essence field is different from those ones defined
by $g^{\mu\nu}$.
Making a conformal transfornmation
$G^{\mu\nu}\equiv {c_{\mathrm s}\over L_{X}^{2}}\tilde G^{\mu\nu}$ and
using the expression for $T_{\mu\nu}$ from equation $(3)$ one can
write the inverse of the metric $G^{\mu\nu}$ as
$$G_{\mu\nu}= {L_{X}\over c_{\mathrm s}}g_{\mu\nu}
-c_{\mathrm s}L_{XX}\nabla_{\mu}\phi\nabla_{\nu}\phi\eqno(5b)$$
We will be using this expression for the effective metric in all
that follows. Also note that after this conformal transformation,
if we further assume that $L$ is not an explicit function of $\phi$
then the equation of motion $(4a)$ is replaced by ;
$$-{1\over \sqrt {-g}}{\delta S_{k}\over \delta \phi}
= {L_{X}^{2}\over c_{\mathrm s}}G^{\mu\nu}\nabla_{\mu}\nabla_{\nu}\phi=0
\eqno(4b)$$
\section{The Schwarzschild solution}
The Schwarzschild metric is given by ($r_{s}= 2GM/c^{2}\equiv 2GM $, taking $c=1$)
$$
ds^{2}=(1- {r_{\mathrm s}\over r}) dt^{2}
- (1- {r_{\mathrm s}\over r})^{-1} dr^{2}\nonumber\\
- r^{2} (d\theta^{2} + sin^{2}\theta d\Phi^{2})
\eqno(6)$$
and the emergent metric components $ G_{\mu\nu}$ are related to the
Schwarschild metric components
$g_{\mu\nu}$ by $(5b)$. Therefore for $L= X^{2}$,and
assuming the $k-$essence field to be spherically symmetric
i.e.$\phi\equiv\phi(r,t)$ one has
$$G_{00}=\bigl(1- {r_{\mathrm s}\over r}\bigr) 2{\sqrt 3}X
-{2\over {\sqrt 3}}\bigl({\partial\phi\over\partial t}\bigr)^{2}$$
$$G_{11}=-\bigl(1- {r_{\mathrm s}\over r}\bigr)^{-1} 2{\sqrt 3}X
- {2\over{\sqrt 3}}\bigl({\partial\phi\over\partial r}\bigr)^{2}$$
$$G_{22}=2{\sqrt 3}Xg_{22}=-2{\sqrt 3}Xr^{2}$$
$$G_{33}=2{\sqrt 3}Xg_{33}=-2{\sqrt 3}Xr^{2}sin^{2}\theta\eqno(7)$$
$$G_{01}= G_{10}=-{2\over {\sqrt 3}}{\partial\phi\over\partial t}{\partial\phi\over\partial r}\eqno(8)$$
All the other $G_{\mu\nu}$ are zero.
Assume $\phi(r,t)$ to be of the form
$\phi_{\mathrm s}(r,t) =\phi_{\mathrm 1s}(r) + \phi_{\mathrm 2s}(t)$.
Then
$$G_{00}=\bigl(1- {r_{\mathrm s}\over r}\bigr) 2{\sqrt 3}X
- {2\over{\sqrt 3}} \bigl({d\phi_{\mathrm 2s}\over dt}\bigr)^{2}$$
$$G_{11}=-\bigl(1- {r_{\mathrm s}\over r}\bigr)^{-1} 2{\sqrt 3}X
- {2\over {\sqrt 3}}\bigl({d\phi_{\mathrm 1s}\over dr}\bigr)^{2}
\eqno(9)$$
$$G_{01}=G_{10}
=-{2\over{\sqrt 3}}\dot\phi_{\mathrm 2s}\phi_{\mathrm 1s}'
\eqno(10)$$
where "dot" denotes differentiation with respect to time and the "prime" is
differentiation with respect to $r$.
As we are concerned only with the singularity structure of the metrics we are not discussing
the $G_{22}$ and $G_{33}$ components as $g_{22}$ and $g_{33}$ are well behaved
for $r\rightarrow 0$.
We assume that
$L_{\mathrm X}~~;
~~L_{\mathrm XX}~~;
~~L_{XX}\bigl({\partial\phi\over\partial t}\bigr)^{2}~~;
~~L_{XX}\bigl({\partial\phi\over\partial r}\bigr)^{2}$
are all well behaved quantities for $r\rightarrow 0$.
All these conditions hold true in the above equations if we
also assume that $({\partial\phi_{\mathrm 1s}\over\partial r})$ is well behaved for
$r\rightarrow 0$.
We shall consider only physical singularities.
Here these occur at $r=0$.
The singularities at $r=r_{\mathrm s}$ are coordinate
singularities and these can always be removed by some
coordinate transformations and we are not considering them.
Note that at $r=0$, the second terms on the
{\it r.h.s.} of $(9)$ and $(10)$ are well behaved
as per our assumptions. Therefore good behaviour
of $G_{00}$ and $G_{11}$ at $r=0$
is guaranteed if there exist two functions
$f_{1}(r), f_{2}(r)$ such that both these functions
are well behaved at $r=0$ and
$$ (1- {r_{\mathrm s}\over r}) 2{\sqrt 3}X = f_{\mathrm 1}(r)~~;
~~(1- {r_{\mathrm s}\over r})^{-1} 2{\sqrt 3}X = f_{\mathrm 2}(r)
\eqno(11)$$
These equations imply that
$f_{1}(r)=f_{2}(r)(1- {r_{\mathrm s}\over r})^{2}$.
It is readily seen that for $X$ to
be well behaved as $r\rightarrow0$ and for the two equations
in $(11)$ to be consistent one possibility is
$f_{1}(r)= constant=1$ and
$f_{2}(r)= (1- {r_{\mathrm s}\over r})^{-2}$.
Then $X$ is well behaved at $r=0$. So
$$G_{00}=1 - {2\over{\sqrt 3}}(\dot\phi_{\mathrm 2s})^{2}~~;~~
G_{11}= -(1- {r_{\mathrm s}\over r})^{-2} -{2\over{\sqrt 3}} (\phi_{\mathrm 1s}')^{2}
\eqno(12)$$
At $r=0$ both $G_{00}~~,~~ G_{11}$ are well behaved and
$$X= {1\over 2{\sqrt 3}}{1\over (1- {r_{\mathrm s}\over r})}
\eqno(13)$$
With our assumption regarding the form of $\phi(r,t)$ , this leads to
$$(\dot\phi_{\mathrm 2s}(t))^{2}
= {1\over{\sqrt 3}} + (1- {r_{\mathrm s}\over r})^{2})(\phi_{\mathrm 1s}'(r))^{2} = k
\eqno(14) $$
where $k$ is a constant. Note that $(12)$ and $(14)$ imply that if
the sign of (temporal component) $G_{00}$ has to remain positive
w.r.t. (spatial components) $G_{11},G_{22},G_{33}$, then
$G_{00} > 0$. This means $k < {3\over 4}=0.75$. Only these
values of $k$ are allowed.
We now discuss possible solutions to this equation. Note that
{\bf Case 1, $k=0$}
We rule out taking $k=0$ because then
$\dot\phi_{\mathrm 2s}(t)= 0$ which means that the $k-$essence scalar
field does not have any kinetic energy. This violates the basic premise of $k-$essence
where the kinetic energy drives the accelerated expansion.
{\bf Case 2, $k={1\over{\sqrt 3}}=0.5773<0.75$}
Now we have ,
$\phi_{\mathrm 1s}'= 0$ and $\dot\phi_{\mathrm 2s}=(3)^{-1/4}$ so that
$$\phi_{\mathrm 1s}(r)= c_{1}~~;~~\phi_{\mathrm 2s}(t) = (3)^{-1/4}t + c_{2}
\eqno(15)$$
where $c_{1}, c_{2}$ are constants. Now
$X= {1\over 2{\sqrt 3}}{1\over (1- {r_{\mathrm s}\over r})}$ and
$G_{00}= 1- {2\over{\sqrt 3}}{1\over{\sqrt 3}} = {1\over 3}\neq g_{00}~;~
G_{11}= -{1\over (1- {r_{\mathrm s}\over r})^{2}}\neq g_{11}~;~
G_{22}=2{\sqrt 3}Xg_{22}~;~ G_{33}=2{\sqrt 3}Xg_{33}~;~
G_{01}= G_{10}=0$. All the other off-diagonal components are also
zero. So the emergent metric without any singularity at $r=0$ is
$$ G_{\mu\nu} = \left(\begin{array}{cccc}
{1\over 3} & 0 & 0 & 0\\
0 & {-1\over (1- {r_{\mathrm s}\over r})^{2}}
& 0 & 0\\
0 & 0 & {-r^{2}\over (1- {r_{\mathrm s}\over r})} & 0\\
0 & 0 & 0 & {-r^{2} sin^{2}\theta\over (1- {r_{\mathrm s}\over r})} \\
\end{array}\right) \eqno(16)$$
It is straightforward to see from eqs.$(16)$ that
$G_{\mu\nu}$ and $g_{\mu\nu}$ are not conformally equivalent.
Therefore there exist homogeneous (i.e. independent of $r$)
$k-$essence scalar field configurations, {\it viz.} $\phi (r,t)= c_{1} + (3)^{-1/4}t + c_{2}$,
that can give rise to an emergent gravity metric where the singularity in the
gravitational metric $g_{\mu\nu}$ is masked for observers riding on the scalar
field perturbations.These configurations also satisfy the emergent gravity
equations of motion $(4b)$ as is easily seen:
${L_{X}^{2}\over c_{\mathrm s}}[G^{00}\partial_{0}^{2}\phi_{\mathrm 2s}
+ G^{11}(\partial_{1}^{2}\phi_{\mathrm 1s}
-\Gamma_{11}^{1}\partial_{1}\phi_{\mathrm 1s})
+G^{01}\nabla_{0}\nabla_{1}\phi
+G^{10}\nabla_{1}\nabla_{0}\phi]= 0$. The first two terms within third brackets
vanish because $\phi_{\mathrm 2s}$ is linear in $t$ and $\phi_{\mathrm 1s}$ is
a constant. The last two terms vanish because $G^{01}\nabla_{0}\nabla_{1}\phi
+G^{10}\nabla_{1}\nabla_{0}\phi=G_{01}\nabla^{0}\nabla^{1}\phi
+G_{10}\nabla^{1}\nabla^{0}\phi$ and $G_{01}= G_{10}=0$.
\section{The Reissner-Nordstrom black hole}
For a static charged black hole with charge $Q$ the metric is the Reissner-Nordstrom metric:
$$ds^{2}=(1- {r_{\mathrm s}\over r} + {r_{\mathrm Q}^{2}\over r^{2}}) dt^{2}
- (1- {r_{\mathrm s}\over r} + {r_{\mathrm Q}^{2}\over r^{2}})^{-1} dr^{2}\nonumber\\
- r^{2} (d\theta^{2} + sin^{2}\theta d\Phi^{2})
\eqno(17)$$
with $r_{\mathrm Q}^{2}= GQ^{2}/4\pi\epsilon_{0}c^{4}\equiv GQ^{2}/4\pi\epsilon_{0}$
taking $c=1$.
We now carry out an exactly similar analysis
as before for the same lagrangian $L= X^{2}$ and assume the solutions for
$\phi$ to be of the form
$\phi_{\mathrm n}(r,t) =\phi_{\mathrm 1n}(r) + \phi_{\mathrm 2n}(t)$.
Then
$$G_{00}=\biggl(1- {r_{\mathrm s}\over r}+ {r_{\mathrm Q}^{2}\over r^{2}}\biggr)2{\sqrt 3}X
- {2\over {\sqrt 3}}\biggl({d\phi_{\mathrm 2n}\over\partial t}\biggr)^{2}$$
$$G_{11}=-\biggl(1- {r_{\mathrm s}\over r}+{r_{\mathrm Q}^{2}\over r^{2}}\biggr)^{-1} 2{\sqrt 3}X
-{2\over{\sqrt 3}}\biggl({d\phi_{\mathrm 1n}\over\partial r}\biggr)^{2}
\eqno(18a)$$
$$G_{01}=G_{10}
=-{2\over {\sqrt 3}}\dot\phi_{\mathrm 2n}\phi_{\mathrm 1n}';~
G_{22}=2{\sqrt 3}X g_{22}~;~G_{33}=2{\sqrt 3}X g_{33}
\eqno(18b)$$
All the other $G_{\mu\nu}$ are zero.
As before at $r=0$, the second terms on the
{\it r.h.s.} of $(18a)$ are well behaved
as per our assumptions. So good behaviour
of $G_{00}$ and $G_{11}$ at $r=0$
is guaranteed if there exist two functions
$g_{1}(r), g_{2}(r)$ such that both these functions
are well behaved at $r=0$ and
$$ (1- {r_{\mathrm s}\over r}+{r_{\mathrm Q}^{2}\over r^{2}})2{\sqrt 3}X = g_{\mathrm 1}(r)~~;
~~(1- {r_{\mathrm s}\over r}+{r_{\mathrm Q}^{2}\over r^{2}})^{-1}2{\sqrt 3}X = g_{\mathrm 2}(r)
\eqno(19)$$
These equations imply that
$g_{1}(r)=g_{2}(r)(1- {r_{\mathrm s}\over r}+{r_{\mathrm Q}^{2}\over r^{2}})^{2}$.
For $X$ to be well behaved as $r\rightarrow0$ and for consistency
one possibility is
$g_{1}(r)= constant=1$ and
$g_{2}(r)= \biggl(1- {r_{\mathrm s}\over r}+ {r_{\mathrm Q}^{2}\over r^{2}}\biggr)^{-2}$.
Then $X$ is well behaved at $r=0$. So
$$G_{00}=1 -{2\over{\sqrt 3}}(\dot\phi_{\mathrm 2n})^{2}~~;~~
G_{11}= -(1- {r_{\mathrm s}\over r}+{r_{\mathrm Q}^{2}\over r^{2}})^{-2} -{2\over{\sqrt 3}}
(\phi_{\mathrm 1n}')^{2}
\eqno(20)$$
At $r=0$ both $\tilde G_{00}~~,~~\tilde G_{11}$ are well behaved and
$$X= {1\over 2{\sqrt 3}}{1\over (1- {r_{\mathrm s}\over r}+{r_{\mathrm Q}^{2}\over r^{2}})}
\eqno(21)$$
With our assumption regarding the form of $\phi(r,t)$ , this leads to
$$(\dot\phi_{\mathrm 2n}(t))^{2}
={1\over{\sqrt 3}} + (1- {r_{\mathrm s}\over r}+{r_{\mathrm Q}^{2}\over r^{2}})^{2}(\phi_{\mathrm 1n}'(r))^{2} = k
\eqno(22)$$
where $k$ is a constant. For $k={1\over{\sqrt 3}}$, again $\phi_{\mathrm 1n}'(r)=0$
and we can write the solution for the $k-$essence field as a homogeneous field
$\phi (r,t)= d_{1} + (3)^{-1/4}t + d_{2}$, where $d_{1,2}$ are constants.
These configurations again satisfy the emergent gravity
equations of motion $(4b)$ as before:
${L_{X}^{2}\over c_{\mathrm s}}[G^{00}\partial_{0}^{2}\phi_{\mathrm 2n}
+ G^{11}(\partial_{1}^{2}\phi_{\mathrm 1n}
-\Gamma_{11}^{1}\partial_{1}\phi_{\mathrm 1n})
+G^{01}\nabla_{0}\nabla_{1}\phi
+G^{10}\nabla_{1}\nabla_{0}\phi]= 0$. The first two terms within third brackets
vanish because $\phi_{\mathrm 2n}$ is linear in $t$ and $\phi_{\mathrm 1n}$ is
a constant. The last two terms vanish because $G^{01}\nabla_{0}\nabla_{1}\phi
+G^{10}\nabla_{1}\nabla_{0}\phi=G_{01}\nabla^{0}\nabla^{1}\phi
+G_{10}\nabla^{1}\nabla^{0}\phi$ and $G_{01}= G_{10}=0$.
The emergent metric in the case of Reisner-Nordstrom background is then :
$$G_{\mu\nu} = \left(\begin{array}{cccc}
{1\over 3} & 0 & 0 & 0\\
0 & {-1\over (1- {r_{\mathrm s}\over r} + {r_{\mathrm Q}^{2}\over r^{2}})^{2}}
& 0 & 0\\
0 & 0 & -r^{2}\over (1- {r_{\mathrm s}\over r}+ {r_{\mathrm Q}^{2}\over r^{2}}) & 0\\
0 & 0 & 0 & -r^{2} sin^{2}\theta\over (1- {r_{\mathrm s}\over r}+ {r_{\mathrm Q}^{2}\over r^{2}})\\
\end{array}\right) \eqno(23)$$
\section{ "Masking" }
We now briefly discuss the "masking" of the singularity
(we ignore the singularity at $r=r_{\mathrm s}$ which is a
coordinate singularity as already mentioned in the beginning).
In our treatment, the scalar field is a homogeneous scalar field ,
i.e., depends only on the time coordinate. As the spatial part of the scalar field
is always a constant in $r$ and the temporal part is linear in the temporal coordinate,
our scalar field is well behaved at the central (physical) singularity, {\it viz.}, $r=0$.
The perturbations of the scalar field travels in $G_{\mu\nu}$.
This metric $G_{\mu\nu}$ is perfectly well behaved at $r=0$ as can
be easily seen. Hence in this metric one can never "see"
the physical singularity. This is what we mean by "masking" of the
singularity at $r=0$.
Does this mean that we have done away with the singularity. The answer is
obviously {\it no} as we now show.
For illustrative purposes we shall confine ourselves to the
Schwarzschild case. Let us define $\delta g_{\mu\nu}= G_{\mu\nu} - g_{\mu\nu}$.
Then it is easy to see that
$$\delta g_{00}= {r_{s}\over r}-{2\over 3}~;~ \delta g_{11}={-r_{s}r\over (r-r_{s})^{2}}$$
$$\delta g_{22}={-r_{s}r^{2}\over (r-r_{s})}~;~\delta g_{33}={-r_{s}r^{2}sin^{2}\theta\over (r-r_{s})}\eqno(24)$$
For the Reisner-Nordstrom background the above equations take the form:
$$\delta g_{00}= {r_{s}\over r}-{r_{\mathrm Q}^{2}\over r^{2}}-{2\over 3}~;~
\delta g_{11}={-r_{s}r^{3}+r_{\mathrm Q}^{2}r^{2}\over (r^{2}-rr_{s}+r_{\mathrm Q}^{2})^{2}}$$
$$\delta g_{22}={-r_{s}r^{3}+r_{\mathrm Q}^{2}r^{2}\over (r^{2}-rr_{s}+r_{\mathrm Q}^{2})}~;~
\delta g_{33}={(-r_{s}r^{3}+r_{\mathrm Q}^{2}r^{2})sin^{2}\theta\over (r^{2}-r_{s}r+r_{\mathrm Q}^{2})}\eqno(25)$$
Note that in both the above examples the change in the original metric $\delta g_{\mu\nu}$
still carries the same singularity structure at $r=0$ as $g_{\mu\nu}$. This is as it should be.
Therefore the singularity is still there but it is impossible to be aware of it
if we use $G_{\mu\nu}$. This is what we call "masking".
\section{Conclusion}
In this work we have shown that for observers whose world line is in an
emergent gravity metric $G_{\mu\nu}$,
(a)The physical singularity at $r=0$ in the gravitational
metric $g_{\mu\nu}$ can remain masked for certain configurations of the
$k-$essence field $\phi$ and observers travelling with the perturbations of such $k-$essence
fields will never be aware of the physical singularity of
the gravitational metric as this is not conformally equivalent to the emergent gravity metric.
(b)These configurations are homogeneous (i.e. functions of time $t$ only ) and satisfy the
equations of motion in the emergent gravity metric.
(c)The above have been shown here for the Schwarzschild and the Reissner-Nordstrom metrics.
Back reaction effects and inhomogeneous field configurations will be discussed in future communications.
|
\section{Introduction}
The precise mathematical definition of spontaneous symmetry
breaking (SSB) in quantum theory is somewhat up for grabs. But
all hands agree that, in the case of infinitely many degrees of
freedom, unitarily inequivalent representations are needed.
In physics more generally, SSB occurs when a ground state is not
invariant under a symmetry of the laws. This means that a
symmetry transformation will take a ground state to another
(mathematically distinct) ground state. But in quantum field
theory, an (irreducible) Hilbert space representation of the
commutation relations can include only a single vacuum state. So
a spontaneously broken symmetry must map between different
unitarily inequivalent representations.
Thus for SSB to occur in infinite quantum theory, the broken
symmetry must not be implemented by a unitary operator. This is
generally agreed to be a necessary condition \citep{GEandCLssb}
and is sometimes taken to be both necessary and sufficient
\citep{JEroughssb,FSsymbreaking}.
Put this way, it can be difficult to see how a symmetry can
possibly be spontaneously broken. The difficulty arises from an
apparent conflict with Wigner's unitary-antiunitary theorem, a
foundational result that applies to all quantum theories. Since a
symmetry ought to preserve all the empirical predictions of a
quantum state, it must not change the transition probabilities
between pure states, which are represented by the inner products
between vectors in a Hilbert space representation. Wigner's
theorem shows that any mapping that preserves these probabilities
for all vector states in a Hilbert space must be a unitary
mapping.
All symmetries preserve transition probabilities in this way. Besides
being physically intuitive, this can be proven rigorously even in
paradigm cases of SSB. But this seems to lead to paradox. We know
SSB is possible in infinite quantum theory --- there are well-known
examples. But we also know that any symmetry, even a broken one, must
satisfy the premises of Wigner's theorem. Since SSB requires unitary
inequivalence, these two facts appear inconsistent.
This inconsistency must be only apparent. Our task is to explain
why. We'll begin by explaining some general features that apply
in all cases of quantum SSB.
We will show that in such cases, Wigner's theorem applies. The
seeming paradox therefore threatens. To resolve it, we'll show
that the existence of a unitary symmetry in Wigner's sense does
not entail the unitary equivalence of the Hilbert space
representations it connects. There remains a sense in which the
symmetry is not (strictly speaking) unitarily implementable.
\section{Quantum SSB}
We begin by recalling some general properties of quantum theories on
the algebraic approach. At the broadest level of generality, a
quantum theory is described by a $C^*$-algebra $\mathfrak{A}$ obeying the
canonical commutation or anticommutation relations (CCRs or CARs
respectively) in their bounded form. This is either an algebra of
observables or (as in the present case) a field algebra. The
self-adjoint operators in $\mathfrak{A}$ stand for physical quantities and are
often called observables. The states of the algebra are the possible
assignments of expectation values to the operators in $\mathfrak{A}$. These
are given by normed linear functionals $\omega: \mathfrak{A} \rightarrow
\mathbb{C}.$ The expectation value of $A$ in state $\omega$ is written
$\omega(A)$.
A clear connection exists between this algebraic formalism and the
better-known Hilbert space formalism. The abstract algebraic states
and observables can be concretely realized by a Hilbert space
representation of $\mathfrak{A}$ (also called a representation of the
CCRs/CARs). Such a representation is a mapping $\pi$ from $\mathfrak{A}$ into
the algebra of bounded operators $B(\mathcal{H})$ on a Hilbert space $\mathcal{H}$.
The representation map is not usually a bijection; Hilbert space
representations will include more operators, and in particular more
observables, than the $C^*$-algebra. Some of the states $\omega$ of
$\mathfrak{A}$ will be representable by density operators on $\mathcal{H}$ that agree
with their expectation values for all $A$ in $\mathfrak{A}$. Collectively
these states are called the \emph{folium} of the representation $\pi$.
Every algebraic state $\omega$ has a unique ``home''
representation in which it is given by a cyclic vector. This is
established by the
\begin{gns} For each state $\omega$ of $\mathfrak{A}$, there is a representation
$\pi$ of $\mathfrak{A}$ on a Hilbert space $\mathcal{H}$, and a vector $\Omega \in
\mathcal{H}$ such that $\omega (A)=\langle \Omega ,\pi (A) \Omega \rangle$, for all
$A\in \mathfrak{A}$, and the vectors $\{ \pi (A)\Omega : A\in \mathfrak{A}\}$ are dense
in $\mathcal{H}$. (Call any representation meeting these criteria a \emph{GNS representation.}) The GNS representation is unique in the sense that for any other representation $(\3H ',\pi ',\Omega ')$ satisfying the previous
two conditions, there is a unique unitary operator $U:\mathcal{H} \to \mathcal{H}
'$ such that $U\Omega =\Omega '$ and $U\pi (A)=\pi '(A)U$, for all $A$ in
$\mathfrak{A}$ \emph{\citep[see][278--279]{kr}}.\end{gns}The definition of
``same representation'' presumed in this statement of uniqueness is called \emph{unitary
equivalence}. We call representations $\pi$ and $\pi'$
\emph{unitarily inequivalent}, and treat them as
distinct,\footnote{Although we always treat inequivalent
representations as formally or mathematically distinct, note that
they may not always be physically inequivalent. As we shall see,
they are sometimes related by symmetries, which are normally assumed
to preserve all the physical facts.} if there is no unitary operator
$U$ between their Hilbert spaces which relates the representations by
\begin{equation}\label{EQuequiv}
U\pi(A)=\pi' (A)U.
\end{equation}
When equation (\ref{EQuequiv}) does hold, we say that the unitary $U$ \emph{intertwines} the representations $\pi$ and $\pi'$.
A useful source on the representation of symmetry in this
framework is \citet{JRandGRbasicaqft}. They posit (very
reasonably) that any symmetry of a quantum system must at a
minimum consist of two bijections, $\alpha$ from the algebra of
physical quantities $\mathfrak{A}$ onto itself and $\alpha'$ from the space of
states of $\mathfrak{A}$ onto itself. These must preserve all expectation
values, so that
\begin{equation}\label{EQsym}
\alpha'(\omega(\alpha(A)))=\omega(A).
\end{equation}
They then show that any such $\alpha$ is a $*$-automorphism of $\mathfrak{A}$, a
bijection $\alpha:\mathfrak{A} \to \mathfrak{A}$ which preserves its algebraic structure
and commutes with the adjoint mapping $(\cdot)^*$. Furthermore,
$\alpha'$ can be defined in terms of this $*$-automorphism as it acts
on states. Clearly $\alpha'(\omega)$ is given by $\omega \circ \alpha^{-1}=\omega(\alpha^{-1}(A))$. We
therefore have a justification, from physical principles, of the
oft-cited fact that symmetries in quantum theory are given by
$*$-automorphisms.
Clearly if $\alpha$ is a symmetry and $\pi(A)$ is a representation of $\mathfrak{A}$ on a Hilbert space $\mathcal{H}$, $\pi \circ \alpha(A)=\pi(\alpha(A))$ is also a representation of $\mathfrak{A}$ on $\mathcal{H}$. In this case $\alpha$ will act as a bijective mapping from $\pi$ to $\pi \circ \alpha$. We call $\alpha$ \emph{unitarily implementable in the representation $\pi$} when there is a unitary mapping $U:\3H\to\3H$ such that
\begin{equation}\label{EQimplem}
\pi'(A)=\pi(\alpha(A))= U \pi(A) U^* .
\end{equation}
This means the symmetry $\alpha$ is unitarily implementable in $\pi$ iff $\pi$ and
$\pi \circ \alpha$ are unitarily equivalent representations of $\mathfrak{A}$.
This is where spontaneous symmetry breaking comes in. In general, a state $\omega$ breaks the symmetry $\alpha$ only if $\alpha$ is not unitarily implementable in $\omega$'s GNS representation.\footnote{On the approach shared by Strocchi and Earman, this is both a necessary and sufficient condition.} When this occurs, $\omega$'s GNS representation and the GNS representation of the symmetry-transformed state $\alpha'(\omega)=\omega \circ \alpha^{-1}$ will be unitarily inequivalent.
In one example, the CAR algebra (so-called because it obeys the
canonical anticommutation relations) is the field algebra for a system
of interacting spin-1/2 systems. We may use the infinite version of
the algebra to represent an infinitely long chain of spins confined to
a one-dimensional lattice, as in the Heisenberg model of a
ferromagnet. This infinite CAR algebra possesses a non-unitarily implementable
automorphism which represents a symmetry of the ferromagnet: namely, a
180-degree rotation which flips all of the spins in the chain. The
rotation is therefore a spontaneously broken symmetry. See
\citet{LRferro} for a detailed study of this case; we present only a
few general features it shares in common with other examples of SSB.
The lowest-energy states available to an infinite spin chain are
ones in which all of the spins align in the same direction. The
Heisenberg ferromagnet has two such ground states: $\omega$, in which
all of the spins point along $+x$ (where $x$ is the axis of the
one-dimensional chain) and $\omega'$, in which they all point along
$-x$. These states each define a GNS representation ($\pi,\pi'$
respectively) on Hilbert spaces $\mathcal{H}$ and $\mathcal{H}'$. If $\alpha$
is the automorphism of the CAR algebra representing a 180-degree
rotation along an axis perpendicular to $x$, $\omega'=\alpha(\omega)$.
But since $\alpha$ is spontaneously broken, $\pi$ must be
unitarily inequivalent to $\pi'$. This is where the seeming
paradox comes in.
\section{Wigner's theorem and the paradox}
The expectation values of observables aren't the only important
quantities in quantum physics. We should also expect a symmetry
to preserve the transition probabilities between (pure) states of
any quantum theory. In the Hilbert space formalism these are
given by inner products: $\langle \psi,\psi' \rangle$ represents the
likelihood of a spontaneous transition from vector state
$\psi$ to $\psi'$.\footnote{Only on collapse
interpretations do such transitions actually occur, of course. But
we should nevertheless expect other interpretations to retain the
statistical predictions codified in transition probabilities.}
It seems obvious that no symmetry worth its salt will alter any of
the transition probabilities. This led Wigner to conclude that
any symmetry worth its salt is given by a unitary
operator.\footnote{Or by an antiunitary operator; for present
purposes we ignore the difference.} For the following can be
proven \citep[see][]{VBwignersthm}:
\begin{wt} Any bijection from the unit rays (vectors states) of a Hilbert space $\mathcal{H}$
to the unit rays of $\mathcal{H}'$ which preserves the inner product is
given by a unitary mapping $W:\mathcal{H}\to\mathcal{H}'$.
\end{wt}
This is puzzling. A spontaneously broken symmetry should still map
bijectively between the vector states of the two representations, and
it would be very strange if it failed to preserve transition
probabilities. But we also know that broken symmetries are not
unitarily implementable. This seems impossible, but since well-known
examples of SSB exist, paradox threatens.
We are not the first to notice this problem. Considering the
$C^{*}$-algebra $B(\mathcal{H})$ of all bounded operators on a Hilbert space,
Earman presents the problem as follows:
\begin{quote}
Obviously, if $U$ is unitary then [the map $A \rightarrow UAU^*$]
is an automorphism of $B(\mathcal{H})$. Conversely, any automorphism of
$B(\mathcal{H})$ takes this form for a unitary $U$ -- ... $T$ is an inner
automorphism\footnote{An automorphism $\alpha$ is inner just in
case there is a unitary $V\in \mathfrak{A}$ such that $\alpha
(A)=VAV^*$ for all $A\in \mathfrak{A}$}
of the $C^*$-algebra $B(\mathcal{H})$. How then can a symmetry
in the guise of an automorphism of a $C^*$-algebra $\mathfrak{A}$ fail to be
an unbroken or Wigner symmetry [one which preserves transition
probabilities]?
\citep[341]{JEroughssb}
\end{quote}
Earman provides an answer, but not a satisfactory one. He notes
correctly that in cases of SSB, the automorphism that represents
the broken symmetry is not an automorphism of all of $B(\mathcal{H})$.
Instead it is defined only on the field algebra or algebra of
observables (the CAR algebra in the ferromagnet example). In a
Hilbert space representation this distinguished algebra is given
by a much smaller subalgebra of $B(\mathcal{H})$. So since the symmetry
is not an automorphism of $B(\mathcal{H})$, there is no reason to assume
it should be unitary in the first place.
This only appears to resolve the problem because Earman has presented
it in a limited form. He considers the symmetry solely as it acts on
operators, but the paradox can only be properly motivated by
considering the symmetry's action on states. Whether $\alpha$ is an
automorphism of $B(\mathcal{H})$ or not, $\alpha'$ is still a bijection between
the states of representations $\mathcal{H}$ and $\mathcal{H}'$. If it's not
unitary, it must fail to meet one of the premises of Wigner's theorem.
The problem is that the premises of Wigner's theorem are satisfied by
\emph{all} symmetries. Consider $\alpha'$ as it maps between the GNS
representations of two ground states. Since a symmetry always takes
pure states to pure states, and the vector states of a GNS
representation are all and only its pure states, $\alpha'$ is indeed a
bijection of vectors states (unit rays). Furthermore, as
\citet[335]{JRandGRbasicaqft} prove, $\alpha'$ must preserve all
transition probabilities. Wigner's theorem therefore applies.
Earman is not alone in struggling with this seeming paradox. Other authors have suggested (contrary to the theorem of Roberts and Roepstorff) that perhaps spontaneously broken symmetries do fail to preserve transition probabilities. For example, \citet[302 fn 111]{AAfields} claims that, because ``the preservation of probabilities in the Hilbert space formulation implies the existence of a unitary (or antiunitary) operator,'' no mapping between the folia of inequivalent representations -- presumably not even a symmetry transformation -- can preserve transition probabilities. Even \citet[483]{LRferro} makes a rare slip on this point:
\begin{quote} A state $\omega$ breaks a dynamical symmetry $\alpha$ iff $\alpha$
is not unitarily implementable on $\omega$'s GNS representation.
Unlike rotations on ground states of the finite spin chain,
\emph{dynamical symmetries do not conserve transition probabilities}
in the GNS representation of states that break them. (Emphasis
ours) \end{quote} But to the contrary, by the very definition of a symmetry, all
symmetries --- whether broken or not --- preserve transition
probabilities. Therefore, broken symmetries satisfy the premises of
Wigner's theorem.
So the answer to Earman's question --- how can a symmetry fail to be a
Wigner-type unitary symmetry? --- is that it cannot. And yet, there
are unitarily non-implementable symmetries. Paradox threatens.
\section{The resolution}
Since Wigner's theorem applies to all symmetries, a spontaneously
broken symmetry must in some sense give a unitary mapping between the
states of unitarily inequivalent representations. So there must be
some wiggle room in the definition of unitary equivalence that makes
this possible. To resolve the paradox, we must look again at the
definition and find the wiggle room.
In effect, we have two data points to work with. First, as Roberts and Roepstorff prove, Wigner's theorem applies to all algebraic symmetries. This means that any symmetry $\alpha'$, as it acts on states, must be implemented by some unitary operator. Since the existence of this operator is guaranteed by Wigner's theorem, we'll call it the \emph{Wigner unitary} $W$. Since for any state $\omega$, $\alpha'(\omega)=\omega \circ \alpha^{-1}$, Wigner's theorem is telling us that $W$ must take the state vector that represents $\omega$ to a vector that represents $\omega \circ \alpha^{-1}$.
Our second data point is the fact that spontaneous symmetry breaking is possible. This implies that some symmetries are not unitarily implementable, in the sense that they fail to satisfy Eq. (\ref{EQimplem}) for any unitary $U$. When Eq. (\ref{EQimplem}) is not satisfied, the symmetry's action on operators (which takes each operator $A$ to $\alpha(A)$) cannot be implemented by a unitary operator. Combining our two data points, this means that spontaneously broken symmetries are unitarily implementable as they act on states, but not as they act on operators.
This means that the unitary $W$ must map the vector states of $\mathcal{H}$ to those of $\mathcal{H}'$ without satisfying Eq.~\ref{EQuequiv} for $U=W$. In such a case, the representations are unitarily inequivalent even though their Hilbert spaces are related by a unitary operator. There is nothing contradictory about this, since the existence of a unitary operator implies unitary equivalence only if the operator intertwines the representations. Their respective representations $\pi,\pi'$ of the $C^*$-algebra may nonetheless be preserved by $W$, as long as $W$ does not map the representation $\pi$ pointwise to the representation $\pi'$. So it may still be that
\begin{equation}
W \pi(\mathfrak{A}) = \pi' (\mathfrak{A}) W \label{foobar}
\end{equation}
although there are individual operators $A \in \mathfrak{A}$ for which (\ref{EQuequiv}) does not hold.
In fact, an operator meeting these criteria exists whenever the states of two GNS representations are connected by a symmetry. We will show in steps that in every such case a unitary $W$ exists which satisfies Eq. (\ref{foobar}), and which implements the symmetry as it acts on states without implementing it as it acts on operators. First, we establish its existence:
\begin{gwt} Let $\langle \8H,\pi,\Omega\rangle$ be a GNS representation
for $\omega$, and let $\langle \8H',\pi',\Omega '\rangle$ be a GNS
representation for $\omega\circ\alpha^{-1}$. Then there is a unique
unitary operator $W=W_{\pi ,\pi'}:\8H\to \8H'$ such that $W\Omega =\Omega
'$, and $W\pi (\alpha ^{-1}(A))= \pi'(A)W$ for all $A\in \2A$. \emph{(Proof in Appendix 1.)}
\end{gwt}
Since $\alpha^{-1} (\mathfrak{A})=\mathfrak{A}$ and $W$ intertwines $\pi \circ \alpha^{-1}$ and $\pi'$, Eq. (\ref{foobar}) follows. This means that when $\pi$ and $\pi'$ are not unitarily equivalent, $W$ maps between these two representations without mapping $\pi$ pointwise to $\pi'$, which is what we expected. In other words, $W$ acts as a bijection between the operators of these representations but does not, in general, implement the symmetry as it acts on operators.
We have established that a unitary mapping preserving transition probabilities can exist even between unitarily inequivalent representations. Indeed, such a mapping always exists in cases of SSB. This is not yet enough to ensure that the conclusion of Wigner's theorem is true (as it must be). Wigner's theorem ensures, not just that such a mapping exists, but that every mapping which preserves transition probabilities must be unitary. This includes every symmetry (whether spontaneously broken or not) as it acts on states.
So we must also show that a spontaneously broken symmetry can be given by a unitary operator in the sense just discussed, without being unitarily implemented --- that is, without satisfying Eq.~\ref{EQimplem}. To establish this, we will show that $W$ itself implements the symmetry as it acts on states.
Keep in mind that any representation $\pi :\2A \to B(\8H )$ of a $C^*$-algebra gives rise to a map $\pi^*$ of unit vectors of $\8H$ into the state space of $\2A$. In particular, $$\pi ^*(x)(A) = \langle x,\pi (A)x\rangle ,\qquad (A\in \2A) .$$ We now use this map to show that $W$ implements the symmetry $\alpha'$ as it acts between the states of representations $\pi$ and $\pi'$.
\begin{cor} Let $(\8H ,\pi ,\Omega)$ be a GNS representation for $\omega$.
Then the Wigner unitary $W$ for $\alpha$ implements the action of $\alpha$
on vectors in $\8H$. That is, $(\pi ')^*(Wx)=\pi ^*(x)\circ \alpha
^{-1}$ for any unit vector $x$ in $\8H$. \emph{(Proof in Appendix 1.)}\end{cor}
In other words, when we apply $W$ to the state vector $x \in \mathcal{H}$ which represents the algebraic state $\pi^*(x)$ in the GNS representation $\pi$, the result is the vector $Wx \in \mathcal{H}'$ which represents the state $\alpha'(\pi^*(x))=\pi^*(x) \circ \alpha^{-1}$ in the representation $\pi'$. This is just what it means for $W$ to implement the symmetry as it acts on states.
Finally, we confirm that $W$ does not in general implement the symmetry $\alpha$ as it acts on operators. This is just to say that it does not in general intertwine the representations $\pi$ and $\pi'$. In fact, we can show that $W$ intertwines these representations only if it is trivial:
\begin{cor2} If the Wigner unitary $W:\8H\to\8H'$ also induces a
unitary equivalence between $\pi$ and $\pi'$, then $\alpha '\circ \pi
^*=\pi ^*$. \label{trivialize} \emph{(Proof in Appendix 1.)} \end{cor2}
That is, every vector state in $\pi$'s Hilbert space is invariant under the symmetry $\alpha'$, and hence left unchanged by $W$. This means $W$ must be the identity.
The astute reader may be puzzled by some of the properties we ascribe to the Wigner unitary. In particular, we've shown that the Wigner unitary, which implements a symmetry as it acts on states, never implements that same symmetry as it acts on operators unless it is the trivial identity operator. How, then, can a non-trivial symmetry be unitarily implemented on both states and operators (and hence unbroken)? This sort of puzzle is best resolved by looking at concrete examples of Wigner unitaries in the case of both broken and unbroken symmetries, which examples we provide in Appendix 2.
We've shown that whenever two states are related by a symmetry, a unitary mapping exists between the Hilbert spaces of their GNS representations and has the properties we would expect. This is the Wigner unitary. Its existence vindicates Wigner's theorem, in that it shows how the theorem can be true even when spontaneous symmetry breaking prevents a symmetry from being unitarily implemented as it acts on operators. The seeming paradox is no paradox at all.
\section{Foundational significance}
Besides the dissolution of a confusing paradox, are there foundational implications of this result? We believe so. To underscore the foundational importance of our resolution of the paradox, let's briefly explore how it bears on one vexed question in the philosophy of quantum field theory. What are the necessary conditions for physical equivalence between field-theoretic states? In AQFT, the representations of the field algebra separate the states into natural ``families:'' the folia of states given by density operators in each representation. We may therefore ask what conditions must be met for two such families of states -- two folia -- to represent the same set of physical possibilities.
For two folia to be physically equivalent, they must at least be empirically equivalent. The paradoxical line of reasoning suggests that unitary equivalence is necessary if we want to preserve transition probabilities. Since a quantum theory's transition probabilities are part of its empirical content, it would seem to follow that the folia of unitarily inequivalent representations cannot predict the same empirical consequences -- making them physically inequivalent by the above reasoning.
This is why Arageorgis, while attempting ``to clarify the
connection between `intertranslatability' [a necessary condition for physical equivalence] and `unitary equivalence,'" writes in his seminal dissertation,
\begin{quote}
Intertranslatability requires a mapping between theoretical
descriptions that preserves the reports of empirical findings.
These are couched in terms of probabilities in quantum theory. And
as Wigner has taught us, the preservation of probabilities in the
Hilbert space formulation implies the existence of a unitary (or
antiunitary)\footnote{Recall that for purposes of this paper we
ignore the distinction between unitary and antiunitary.} operator.
\citep[302 fn 111]{AAfields}
\end{quote}
He takes this point to establish that folia must belong to unitarily equivalent representations if they are to count as physically equivalent. But his argument includes a false premise: the assumption that the existence of a unitary operator connecting the folia of two representations implies a unitary equivalence between those representations. As we have shown, though, there is no such implication if the unitary operator is what we've called a Weyl unitary. Arageorgis's argument is unsound.
This means that at least one significant part of the empirical content of a quantum theory -- its transition probabilities -- can be preserved by a mapping between the folia of two inequivalent representations. If we further assume (as conventional wisdom dictates) that a quantum theory's symmetries preserve all empirical content, then the folia of at least some pairs of inequivalent representations must be empirically equivalent if spontaneous symmetry breaking is possible. The notion that unitary equivalence is a necessary condition for physical equivalence should now appear quite suspect. Insofar as the so-called ``Hilbert space conservative'' interpretation of quantum field theory identifies physical equivalence with unitary equivalence \citep[see][]{LRinterpreting}, that interpretation must come into question as well.
\section*{Appendix 1: Representation Wigner Theorem}
Here we prove the results mentioned in the main text.
\begin{defn} Let $\omega$ be a state on a $C^*$-algebra $\2A$, let $\8H$
be a Hilbert space with $\Omega$ a unit vector in $\8H$, and $\pi
:\2A\to B(\8H )$ a representation of $\2A$. We say that the triple
$\langle \8H,\pi ,\Omega \rangle$ is a \emph{GNS representation} for
$\omega$ just in case:\begin{enumerate}
\item $\langle \Omega ,\pi (A)\Omega\rangle = \omega (A)$, for all $A\in
\2A$, and
\item $\{ \pi (A)\Omega :A\in \2A \}$ is dense in the Hilbert space
$\8H$. \end{enumerate} \end{defn}
The GNS theorem shows that for each state $\omega$, there is a GNS
representation; and that any two GNS representations of $\omega$ are
unitarily equivalent.
Let $\2A$ be a $C^*$-algebra, let $\omega$ be a state of $\2A$, and let
$\alpha$ be a $*$-automorphism of $\3A$. Let $(\3H,\pi ,\Omega)$ be the GNS
triple of $\2A$ induced by $\omega$, and let $(\3H ',\pi ',\Omega ')$ be the
GNS triple of $\2A$ induced by $\omega\circ\alpha^{-1}$. For brevity, we
sometimes just use $\pi$ and $\pi '$ to denote the corresponding
triples.
\begin{gwt} Let $\langle \8H,\pi,\Omega\rangle$ be a GNS representation
for $\omega$, and let $\langle \8H',\pi',\Omega '\rangle$ be a GNS
representation for $\omega\circ\alpha^{-1}$. Then there is a unique
unitary operator $W=W_{\pi ,\pi'}:\8H\to \8H'$ such that $W\Omega =\Omega
'$, and $W\pi (\alpha ^{-1}(A))= \pi'(A)W$ for all $A\in \2A$.
\end{gwt}
\begin{proof} Let $(\8H ,\pi ,\Omega)$ be a GNS representation of $\2A$
for the state $\omega$, and let $(\8H ',\pi ',\Omega ')$ be a GNS
representation of $\2A$ for the state $\omega\circ \alpha ^{-1}$. Define
$W:\8H\to\8H '$ by setting
$$ W\pi (A)\Omega = \pi '(\alpha (A))\Omega ' ,\qquad \forall A\in
\2A.$$ Since $\alpha (I)=I$, it follows that $W\Omega =\Omega '$. Since
$$ \norm{\pi '(\alpha (A))\Omega '}^2 = \langle \Omega ',\pi '(\alpha (A^*A))\Omega
'\rangle = \omega (\alpha ^{-1}(\alpha (A^*A)))=\omega (A^*A)=\norm{\pi (A)\Omega
}^2,$$ it follows that $W$ is well defined and extends uniquely to a
unitary operator from $\8H$ to $\8H '$. Note that since $\pi (A)\Omega
=W^*W\pi (A)\Omega =W^*\pi '(\alpha (A))\Omega '$, it follows that $W^*\pi
'(B)\Omega '=\pi (\alpha ^{-1}(B))\Omega$ for all $B\in \2A$. Therefore,
$$ W^*\pi '(A)W\pi (B)\Omega = W^*\pi '(A\alpha (B))\Omega '= \pi (\alpha
^{-1}(A)B)\Omega = \pi (\alpha ^{-1}(A))\pi (B)\Omega ,$$ for all $A,B\in \2A$.
Since the vectors $\pi (B)\Omega$, for $B\in \2A$, are dense in $\8H$, it
follows that $W^*\pi '(A)W=\pi (\alpha ^{-1}(A))$ for all $A\in \2A$.
That is, $W$ implements a unitary equivalence from $\pi \circ \alpha
^{-1}$ to $\pi '$.
To show the uniqueness of $W$, it suffices to note that $\Omega$ is a
cyclic vector for $\pi\circ\alpha$, $\Omega '$ is a cyclic vector for $\pi
'$, and $W\Omega =\Omega '$. Thus, there is at most one unitary intertwiner
from $\pi\circ\alpha ^{-1}$ to $\pi'$ that maps $\Omega$ to
$\Omega'$. \end{proof}
Note that if $\alpha =\iota$ is the identity automorphism, and if we
take $\langle \8H',\pi',\Omega '\rangle = \langle \8H,\pi,\Omega \rangle$,
then $I$ satisfies the conditions of the theorem, hence by uniqueness
$W_{\pi,\pi'}=I$.
For the following corollary, recall that any representation $\pi :\2A
\to B(\8H )$ of a $C^*$-algebra gives rise to a map $\pi ^*$ of unit
vectors of $\8H$ into the state space of $\2A$. In particular,
$$\pi ^*(x)(A) = \langle x,\pi (A)x\rangle ,\qquad (A\in \2A) .$$
\begin{cor} Let $(\8H ,\pi ,\Omega)$ be a GNS representation for $\omega$.
Then the Wigner unitary $W$ for $\alpha$ implements the action of $\alpha$
on vectors in $\8H$. That is, $(\pi ')^*(Wx)=\pi ^*(x)\circ \alpha
^{-1}$ for any unit vector $x$ in $\8H$. \end{cor}
\begin{proof} By the Theorem, a Wigner unitary $W$ intertwines $\pi\circ
\alpha ^{-1}$ and $\pi'$, that is
$$ \pi (\alpha ^{-1}(A))= W^*\pi '(A)W ,$$
for all $A\in \2A$. Hence $$ (\pi ')^*(Wx)(A) \:=\: \langle Wx,\pi '
(A)Wx \rangle \: =\:\langle x,W^*\pi '(A)Wx\rangle \:=\: \langle
x,\pi (\alpha ^{-1}(A))x\rangle\:=\: \pi ^*(x)(\alpha ^{-1}(A)) ,$$ for all
$A\in \2A$.
\end{proof}
The preceding corollary can be conveniently pictured via a commuting diagram:
$$\bfig
\square/>`<-`<-`>/[S(\2A)`S(\2A)`\8H`\8H ';\alpha '`\pi ^*`(\pi ')^*`W]
\efig $$ where $S(\2A )$ is the state space of $\2A$, and $\alpha
':S(\2A)\to S(\2A )$ is the symmetry $\omega\mapsto \omega \circ \alpha ^{-1}$.
\begin{cor2} If the Wigner unitary $W:\8H\to\8H'$ also induces a
unitary equivalence between $\pi$ and $\pi'$, then $\alpha '\circ \pi
^*=\pi ^*$. \label{trivialize} \end{cor2}
Recall that $\alpha ':S(\2A )\to S(\2A )$ is defined by $\alpha '(\omega )=\omega
\circ \alpha ^{-1}$.
\begin{proof} If $W$ induces a unitary equivalence between $\pi$ and
$\pi '$ then
$$ \pi (A)W^* = W^*\pi '(A) = \pi (\alpha ^{-1}(A))W^* ,$$
for all $A\in \2A$. Canceling the unitary operator $W^*$ on the right
gives $\pi (A)=\pi (\alpha ^{-1}(A))$ for all $A\in \2A$, that is $\pi =
\pi \circ \alpha ^{-1}$. From the latter equation it clearly follows
that $\pi ^*=\alpha '\circ \pi ^*$. \end{proof}
\section*{Appendix 2: Properties of the Wigner unitary operator}
We now give a special case of the Representation Wigner Theorem which will illustrate some properties of the Wigner unitary. But first we need a lemma.
\begin{lemma} If $\langle \8H,\pi ,\Omega \rangle$ is a GNS
representation for the state $\omega$ then $\langle \8H,\pi \circ
\alpha ^{-1},\Omega \rangle$ is a GNS representation for the state $\omega
\circ \alpha ^{-1}$.
\end{lemma}
\begin{proof} Since $\langle \Omega ,\pi (\alpha ^{-1}(A))\Omega \rangle = \omega
(\alpha ^{-1}(A))$ and since $\Omega$ is cyclic under $\{ \pi (\alpha
^{-1}(A)) :A\in \mathfrak{A}\}$, it follows that $\langle \8H,\pi
\circ \alpha ^{-1},\Omega \rangle$ is a GNS representation for $\omega \circ
\alpha ^{-1}$. \end{proof}
We can now apply the Representation Wigner Theorem to the representations
$\langle \8H,\pi,\Omega\rangle$ and $\langle \8H,\pi\circ \alpha
^{-1},\Omega\rangle$
\begin{swt} If $W:\8H\to \8H$ is the Wigner operator for the
representations $\langle \8H,\pi,\Omega \rangle$ and $\langle
\8H,\pi\circ \alpha ^{-1},\Omega \rangle$, then $W=I$.
\end{swt}
\begin{proof} By RWT, $W\Omega =\Omega$ and $W\pi (\alpha ^{-1}(A))=\pi (\alpha
^{-1}(A))W$ for all $A\in \2A$. Since $\alpha$ is an automorphism,
$W\pi (A)=\pi (A)W$ for all $A\in \2A$. Hence
$$ W\pi (A)\Omega = \pi (A)W\Omega = \pi (A)\Omega ,$$
for all $A\in \2A$. Since the set $\{ \pi (A)\Omega :A\in
\mathfrak{A}\}$ of vectors is dense in $\8H$, it follows that $Wx=x$
for every vector in $\8H$; that is, $W=I$. \end{proof}
We now apply SRWT to the case of symmetries in elementary quantum mechanics. Let $\8H$ be a finite-dimensional Hilbert space, and let $U:\8H\to \8H$ be
a unitary operator that induces a symmetry. Then we have the
following transformations:
\[ \begin{array}{ll}
\varphi \longmapsto U\varphi & \text{transformed state} \\
A\longmapsto UAU^* & \text{transformed observable} \end{array} \] Of
course, $B(\8H )$ is a $C^*$-algebra, and each (pure) state of $B(\8H
)$ is represented uniquely by a ray in $\8H$. The unitary $U$ induces
the automorphism $\alpha (A)=UAU^*$ of $B(\8H )$, as well as the
corresponding state mapping.
In order to apply SRWT, we need to find representations. The first
representation is $\langle \8H,\iota ,\varphi\rangle$, where $\iota
:B(\8H )\to B(\8H )$ is the identity, and $\varphi$ is an arbitrarily
chosen unit vector. The second representation is $\langle \8H ,\iota
\circ \alpha ^{-1},\varphi\rangle$. By GWT, there is a Wigner unitary
$W:\8H\to \8H$, and by SRWT, $W=I$.
So, in what sense does $W$ induce the symmetry $U$ on states? Should
we not have $W=U$? No, because a vector $\varphi$ in $\8H$ names
different states on $B(\8H )$ according to which representation we
consider, either $\iota$ or $\alpha ^{-1}$. Relative to the first,
$\varphi$ represents the state $A\mapsto \langle \varphi
,A\varphi\rangle$, and relative to the second, $\varphi$ represents
the state $A\mapsto \langle \varphi ,\alpha ^{-1}(A)\varphi \rangle$.
What $W$ does is to map a vector representing some state $\omega$
relative to $\pi$ to a vector representing the state $\omega \circ \alpha
^{-1}$ relative to $\pi\circ \alpha ^{-1}$. In the way we have set
things up, $W=1_H$, which just means that if $\varphi$ represents
$\omega$ relative to $\pi$, then $\varphi$ represents $\omega\circ\alpha ^{-1}$
relative to $\pi'=\pi\circ \alpha ^{-1}$. So, indeed, the identity map
implements the symmetry $\omega \mapsto \omega \circ \alpha^{-1}$ of states!
Let us look now, more generally, at the case of an
\emph{unbroken} symmetry. By hypothesis, the symmetry $\alpha$ is unbroken
just in case the representations $(\8H ,\pi ,\Omega )$ and $(\8H
,\pi\circ \alpha ^{-1},\Omega )$ are unitarily equivalent. That is, there
is a unitary operator $V:H\to H$ such that $V\pi (\alpha ^{-1}(A)) = \pi
(A)V$. In fact, in the most interesting case where $\omega$ is a pure
state, $V$ can be chosen such that $V=\pi (U)$ for some unitary
operator $U\in \2A$, hence
$$ \pi (\alpha (A)) \: = \: V\pi (A)V^* \:=\: \pi (UAU^*) ,$$
for all $A\in \2A$. (To verify the existence of such a $U\in \2A$,
see \citet[730]{kr}.)
Of course, we are still guaranteed the existence of the Wigner Unitary
$W:\8H\to \8H$. (In fact, we know that $W=I$; but ignore that fact
for now.) Which operator, $W$ or $V$, implements the symmetry
$\alpha$ on states? The answer is that they \emph{both} do, but in
different senses.
Compare the following two diagrams:
$$\bfig
\square/>`<-`<-`>/[S(\2A)`S(\2A)`\8H`\8H ;\alpha '`\pi ^*`(\pi ')^*`W]
\square(1000,0)/>`<-`<-`>/[S(\2A)`S(\2A)`\8H`\8H;\alpha '`\pi ^*`\pi
^*`V] \efig $$
The square on the left shows the action of the Wigner
unitary $W$ for the special case of the GNS representation $(\8H
,\pi\circ \alpha ^{-1},\Omega)$ for $\omega\circ \alpha ^{-1}$. The square on the
right shows that action of the unitary $V$ that implements the
equivalence between $\pi$ and $\pi\circ \alpha ^{-1}$. The key
difference, of course, is that $V$ implements the symmetry in such a
way that the correspondence between vectors and states can be held
invariant (the vertical arrows are the same), whereas $W$'s
implementation requires a change of correspondence ($\pi ^*$ versus
$(\pi ')^*$). But a state is a way to map observables to numbers, so
changing the correspondence between vectors and states is equivalent
to leaving this correspondence fixed and instead changing the labels
of observables. In equation form:
$$ (\pi ')^* = \alpha '\circ \pi ^* ,$$
i.e.\ the correspondence $(\pi ')^*$ matches vectors with an
observable $A$ in exactly the way that the correspondence $\pi ^*$
matches vectors with the observable $\alpha ^{-1}(A)$.
\subsection*{Acknowledgements}
Thanks to Wayne Myrvold and Giovanni Valente for detailed comments on
an earlier draft. DJB would like to thank Laura Ruetsche for several
valuable discussions of the seeming paradox and John Earman for
helpful correspondence. HPH thanks Jeremy Butterfield, Joe Rachiele,
and Noel Swanson for helping him clarify the contrast between Wigner
and intertwining unitaries.
|
\section{{Introduction\label{firstsection}}}
{It is well-known (see \cite{GiTr}) that if }$A${\ is \emph{elliptic}, and }$%
A${\ and }$b${\ are smooth functions of their arguments, then quasilinear
operators in divergence form
\begin{equation*}
\mathcal{Q}w=\mathop{\rm div}A\left( x,w,\nabla w\right) +b\left( x,w,\nabla
w\right)
\end{equation*}%
are hypoelliptic: any weak solution $w$ of $\mathcal{Q}w=0$ is smooth. When $%
\mathcal{Q}$ is \emph{subelliptic} - i.e., when ellipticity fails only to
finite order - then hypoellipticity still holds if $\mathcal{Q}$ is \emph{%
linear} (see e.g. \cite{Treves}). When $\mathcal{Q}$ is linear but fails to
be subelliptic, the situation is more delicate. For example, Fedi\u{\i}
showed in \cite{Fe} that the two-dimensional operator
\begin{equation}
\partial _{x}^{2}+k\left( x\right) \partial _{y}^{2} \label{fedi}
\end{equation}%
is hypoelliptic if $k$ is smooth and positive for all $x\neq 0$. In this
case $k$ is allowed to vanish at any rate at $x=0$. However, by \cite{KuStr}%
, $\partial _{x}^{2}+k\left( x\right) \partial _{y}^{2}+\partial _{z}^{2}$
is hypoelliptic in $\mathbb{R}^{3}$ if and only if $\lim_{x\rightarrow
0}x\log k\left( x\right) =0$. A quasilinear version of operators of the form
(\ref{fedi}) arises when one considers two-dimensional Monge-Amp\`{e}re
equations
\begin{equation}
u_{ss}u_{tt}-u_{st}^{2}=k\left( s,t\right) ,\qquad \left( s,t\right) \in
\widetilde{\Omega }\subset \mathbb{R}^{2}, \label{MA2D}
\end{equation}%
together with the classical partial Legendre transformation $\left(
x,y\right) =T\left( s,t\right) $ given by
\begin{equation}
\left\{
\begin{array}{lll}
x & = & s \\
y & = & u_{t}%
\end{array}%
\right. . \label{PLT2D}
\end{equation}%
Indeed, assuming that $T$ is invertible, (\ref{MA2D}) and (\ref{PLT2D}) lead
to the two-dimensional quasilinear equation
\begin{equation}
\partial _{x}^{2}w+\partial _{y}\left\{ k\left( x,w\left( x,y\right) \right)
\partial _{y}w\right\} =0,\;\;\;\;\;\left( x,y\right) \in \Omega =T(%
\widetilde{\Omega }), \label{qfedi}
\end{equation}%
satisfied in the weak sense by $w\left( x,y\right) =t$. In \cite{SaW} and
\cite{SaW2}, two of the authors extended Fedi\u{\i}'s two-dimensional
regularity result for linear equations to certain solutions $w$ of (\ref%
{qfedi}) obtained through the transformation (\ref{PLT2D}) from a solution
of the Monge-Amp\`{e}re equation (\ref{MA2D}). The coefficient $k$
considered there is assumed to satisfy
\begin{equation}
\left\vert k_{t}\left( s,t\right) \right\vert \leq C\,k\left( s,t\right) ^{%
\frac{3}{2}},\qquad \left( s,t\right) \in \widetilde{\Omega }. \label{thw}
\end{equation}%
Thus $k$ is required to become more independent of its second variable as $k$
degenerates. Notice that the coefficient $k\,$ in (\ref{fedi}) is
independent of the second variable, and then (\ref{thw}) is automatically
true. The main result in \cite{SaW2} establishes that degenerate
two-dimensional Monge-Amp\`{e}re equations (\ref{MA2D}) with smooth
right-hand side $k$ satisfying (\ref{thw}) are hypoelliptic. This was the
first known hypoellipticity result for infinitely degenerate Monge-Amp\`{e}%
re equations. More general equations than (\ref{qfedi}) are also treated in
the papers above, including the equation for prescribed Gaussian curvature. }
\section{Description of the results}
In the present work, we improve the two-dimensional results above by
lowering the exponent $\frac{3}{2}$ in (\ref{thw}) to the optimal exponent $%
1 $; this optimallity is shown in Section 3 below. We also extend the theory
of regularity for degenerate quasilinear equations of the type treated in
\cite{SaW} and \cite{SaW2} to any dimension $n\geq 2$ and to more general
quasilinear problems. In this process we have to deal with several
fundamental difficulties associated with higher dimensions and the more
general structure of the equations. {We consider quasilinear equations of
the divergence form
\begin{equation}
\mathcal{Q}w=\mathop{\rm div}\mathcal{A}\left( x,w\right) \nabla w+\vec{%
\gamma}\left( x,w\right) \cdot \nabla w+f\left( x,w\right) =0\qquad \text{in
}\Omega , \label{equation}
\end{equation}%
where $\Omega $ is an open bounded connected subset of $\mathbb{R}^{n}$ and
the matrix $\mathcal{A}$, vector function }${\vec{\gamma}}$ {and scalar
function $f$ are smooth functions of their arguments. }Here we adopt the
vector notation $\vec{u}=\left( u^{1},u^{2},\dots ,u^{n}\right) $; $\nabla w$
denotes the gradient $\nabla w=\left( \partial _{1}w,\partial _{2}w,\dots
,\partial _{n}w\right) ^{\prime }$ where $\partial _{i}=\frac{\partial }{%
\partial x_{i}}$ is the $i^{th}$ partial derivative; and the divergence
operator applied to a vector function $\vec{u}$ is given by $\mathop{\rm div}%
\vec{u}=\partial _{1}u^{1}+\partial _{2}u^{2}+\dots +\partial _{n}u^{n}$. {\
In our applications, (\ref{equation}) will sometimes be satisfied in the
strong sense, i.e. in the pointwise sense for }$\mathcal{C}^{2}$ functions $%
w $, and other times in the {weak sense; see Section \ref%
{Subsubsection-weak-solution} in the Appendix for a precise definition. Note
that in the case $n=2$, equation (\ref{qfedi}) is included among equations
of the form $\mathcal{Q}w\left( x_{1},x_{2}\right) =0$ by choosing }${\vec{%
\gamma}}${$=\vec{0}$, $f=0$ and $\mathcal{A}\left( x_{1},x_{2},z\right) $ to
be the diagonal matrix $\mathop{\rm diag}\left( 1,k\left(
x_{1},x_{2},z\right) \right) $ with $k$ independent of $x_{2}$. However,
equation (\ref{equation}) does not include systems obtained from the
Monge-Amp\`{e}re equation by the partial Legendre transform introduced in
\cite{RSaW} for higher dimensions, and the treatment of such systems when
ellipticity fails to infinite order remains a challenging open problem in
dimensions bigger than two.}
{In Section \ref{hypo}, under structural restrictions on $\mathcal{A}$ and }$%
{\vec{\gamma}}${\ which are similar to (\ref{thw}) (although weaker, see
Conditions \ref{hyp1} and \ref{hyp2}), we first obtain local a priori bounds
for the Lipschitz norm of smooth solutions in terms of their $L^{\infty }$
norm and the parameters inherent to the equation. This result together with
the main theorem in \cite{RSaW1} (see Theorem \ref{quasi} below) provides a
priori control of \emph{all} derivatives of a smooth solution in terms of
the supremum norm of the solution. }
{In Section \ref{section-Proof}, we apply the a priori estimates together
with an approximation scheme to prove that continuous weak solutions are
smooth, and to establish existence, uniqueness, and regularity of solutions
of the Dirichlet problem. To do so, we use a class of }custom-built {\
barrier functions, a maximum principle and a comparison principle adapted to
our class of equations (see Sections \ref{barriersect}, \ref{maxpsection}
and \ref{compsect} in the Appendix).}
The method used to construct the barriers in Lemma \ref{barrier} takes into
account only the modulus of continuity of solutions on the boundary. This
generalizes most known barrier constructions, which usually require higher
regularity of solutions on the boundary.
{As already mentioned, one of our main results, Theorem \ref{application}
below, states that under certain hypotheses on the coefficients $\mathcal{A}$
, every \emph{continuous} weak solution $w$ of (\ref{equation}) is
infinitely differentiable, and all of its derivatives are locally controlled
by $\left\Vert w\right\Vert _{L^{\infty }}$. The conditions imposed on the
coefficients allow them to vanish to infinite order, so the quasilinear
operator $\mathcal{Q}$ in (\ref{equation}) is not in general uniformly
elliptic or even subelliptic. This is the first known hypoellipticity result
for infinitely degenerate quasilinear equations in $n$ dimensions.}
We now state special cases of the main Theorems \ref{DP} and \ref%
{application}.\ We include these simpler versions to illustrate the
principal features of our results without the technical assumptions of the
general case. {A \emph{domain} will always mean an open connected set. }
\begin{theorem}[Dirichlet problem]
Let $\Omega $ {be a strongly convex domain in $\mathbb{R}^{n}$ containing
the origin. }Let $k^{i}\left( x,z\right) $, $i=2,\dots ,n,$ be smooth
nonnegative functions in $\Omega \times \mathbb{R}$ such that
\begin{equation*}
k^{i}\left( x,z\right) >0\qquad \text{if}\quad x_{j}\neq 0\quad \text{for
some }j\neq i
\end{equation*}%
(this means that $k^{i}\left( x,z\right) $ may vanish only for those $\left(
x,z\right) $ so that $x$ lies on the $i^{th}$-coordinate axis), and such
that for some $B>0,$
\begin{equation*}
\left\vert \frac{\partial }{\partial z}k^{i}\left( x,z\right) \right\vert
\leq B~k^{\ast }\left( x,z\right) \qquad \text{for all\quad }\left(
x,z\right) \in \Omega \times \mathbb{R},
\end{equation*}%
where $k^{\ast }=\min_{i=2,\dots ,n}k^{i}$. Then, {for any continuous
function }$\varphi $ on $\partial \Omega $, there exists a unique continuous
strong solution $w$ to the Dirichlet problem%
\begin{equation*}
\left\{
\begin{array}{rcll}
\frac{\partial ^{2}}{\partial x_{1}^{2}}w\left( x\right) +\sum_{i=2}^{n}%
\frac{\partial }{\partial x_{i}}k^{i}\left( x,w\left( x\right) \right) \frac{%
\partial }{\partial x_{i}}w\left( x\right) & = & 0\quad & \text{in }\Omega ,
\\
w & = & \varphi & \text{on }\partial \Omega ,%
\end{array}%
\right.
\end{equation*}%
i.e., there exists a unique $w$ that is both a strong solution of the
differential equation in $\Omega $ and continuous in $\overline{\Omega }$
with boundary values ${\varphi }$. Moreover, this solution $w\in \mathcal{C}%
^{0}\left( \overline{\Omega }\right) \bigcap \mathcal{C}^{\infty }\left(
\Omega \right) $.
\end{theorem}
\begin{theorem}[Regularity of solutions]
{\ Let $\Omega \subset \mathbb{R}^{n}$ be a bounded domain containing the
origin, and suppose that $k^{i}\left( x,z\right) $, }$i=2,\dots ,n,${\ are
as in Theorem 2.1. Then any }continuous weak solution{\ $w$ of}%
\begin{equation*}
\frac{\partial ^{2}}{\partial x_{1}^{2}}w\left( x\right) +\sum_{i=2}^{n}%
\frac{\partial }{\partial x_{i}}k^{i}\left( x,w\left( x\right) \right) \frac{%
\partial }{\partial x_{i}}w\left( x\right) =0{\quad }\text{in }\Omega {\ }
\end{equation*}%
is also a strong solution, and {satisfies $w\in \mathcal{C}^{\infty }\left(
\Omega \right) $. }
\end{theorem}
\subsection{A priori estimates\label{statements}}
For $\tilde{x}\in \mathbb{R}^{n}$ and $\vec{r}\in \mathbb{R}_{+}^{n},$ we
denote by $\mathcal{R}\left( \tilde{x},\vec{r}\right) $ the box centered at $%
\tilde{x}$ with edges parallel to the coordinate axes and half-edgelengths
given by $\vec{r}$, i.e.,
\begin{equation}
\mathcal{R}\left( \tilde{x},\vec{r}\right) =\left[ \tilde{x}_{1}-r^{1},%
\tilde{x}_{1}+r^{1}\right] \times \cdots \times \left[ \tilde{x}_{n}-r^{n},%
\tilde{x}_{n}+r^{n}\right] . \label{rectangulo}
\end{equation}%
When $\tilde{x}$ is the origin we will just write $\mathcal{R}\left( \vec{r}%
\right) $ for $\mathcal{R}\left( 0,\vec{r}\right) $ and we also adopt the
summation notation $\mathcal{R}\left( \tilde{x},\vec{r}\right) =\tilde{x}+%
\mathcal{R}\left( \vec{r}\right) $. When $\vec{r}$ and $\tilde{x}$ are fixed
or clear from context, we will omit them and simply write $\mathcal{R}$ for $%
\mathcal{R}\left( \tilde{x},\vec{r}\right) $. For any positive constant $%
\gamma $, $\gamma \mathcal{R}\left( \tilde{x},\vec{r}\right) $ will denote
the box centered at $\tilde{x}$ with half-edgelengths given by $\gamma \vec{r%
}$, i.e.,%
\begin{equation}
\gamma \mathcal{R}\left( \tilde{x},\vec{r}\right) =\mathcal{R}\left( \tilde{x%
},\gamma \vec{r}\right) =\tilde{x}+\mathcal{R}\left( \gamma \vec{r}\right) .
\label{rectanguloG}
\end{equation}%
We define the \emph{i-wrap} $\mathcal{T}_{i}\left( \mathcal{R}\right) $ of
the box $\mathcal{R}$ as the set of faces of $\partial \mathcal{R}$
containing the direction $\vec{e}_{i}=\left( \delta _{ij}\right) _{j=1,\dots
,n}$:
\begin{equation*}
\mathcal{T}_{i}\left( \mathcal{R}\right) =\overline{\partial \mathcal{R}%
\backslash \left\{ y:\left\vert y_{i}-\tilde{x}_{i}\right\vert
=r^{i}\right\} }.
\end{equation*}%
The following figure illustrates $i$-wraps in $\mathbb{R}^{3}$ :
\begin{center}
\includegraphics[width=5.7in]{wraps}
\end{center}
We will use the notation $\Gamma =\Omega \times \mathbb{R}$ unless specified
to the contrary.
\begin{axiom}[Nondegeneracy condition]
\label{hyp1}We say that $\vec{k}=\left( k^{1},\dots ,k^{n}\right) $
satisfies Condition \ref{hyp1} in $\Gamma$ if $\vec{k}$ has continuous
second order derivatives in $\Gamma $, $\vec{k}\geq 0 $ in $\Gamma $, and $%
\vec{k}$ satisfies the following:
\begin{enumerate}
\item \label{ax1c2}For every subdomain $\Omega ^{\prime }$ with $\overline{%
\Omega ^{\prime }}\subset \Omega $, there exists $c>0$ such that if $\Gamma
^{\prime }=\Omega ^{\prime }\times \mathbb{R}$, then
\begin{equation*}
\inf_{\left( x,z\right) \in \Gamma ^{\prime }}\max_{1\leq i\leq
n}k^{i}(x,z)\geq c>0.
\end{equation*}
\item \label{ax1c3}For all $x\in \Omega $ and $0<\varepsilon <n^{-\frac{1}{ 2%
}}\mathop{\rm dist}\left( x,\partial \Omega \right) $, there exists $\vec{r}
=\vec{r}\left( x,\varepsilon ,\vec{k}\right) =\left( r^{1},\dots
,r^{n}\right) $, $0<r^{1},\dots ,r^{n}<\varepsilon $, and a box $\mathcal{R}%
= \mathcal{R}\left( \tilde{x},\vec{r}\right)$ so that $x\in \frac{1}{3}
\mathcal{R}$ and
\begin{equation*}
k^{i}\left( y,z\right) >0\text{ whenever }y\in \mathcal{T}_{i}\left(
\mathcal{R}\right) ,~z\in \mathbb{R},~i=1,\dots ,n.
\end{equation*}
The restriction on the size of $\epsilon$ ensures that $\mathcal{R} \subset
\Omega$.
\end{enumerate}
\end{axiom}
Note that if $k^{1}\equiv 1$, then property (\ref{ax1c2}) in Condition \ref%
{hyp1} is trivially satisfied. If each $k^{i}$ is nonnegative and has only
\emph{isolated zeros}, then property (\ref{ax1c3}) in Condition \ref{hyp1}
holds.
\begin{remark}
\label{spans}Property (\ref{ax1c3}) in Condition \ref{hyp1} holds for $\vec{%
k }$ in $\Gamma =\Omega \times \mathbb{R}$ if and only if for every $x\in
\Omega $ and $\varepsilon >0,$ there exists a box $\mathcal{R}$ in $\Omega $
with $x\in \frac{1}{3}\mathcal{R}$ and $0<r^{1},\dots ,r^{n}<\varepsilon $
such that for all $y\in \partial \mathcal{R}$ and $z\in \mathbb{R}$, the set
of vectors
\begin{equation*}
\mathcal{S}\left( y,z\right) =\left\{ k^{1}\left( y,z\right) \vec{e}
_{1},k^{2}\left( y,z\right) \vec{e}_{2},\dots ,k^{n}\left( y,z\right) \vec{e}
_{n}\right\} ,\quad \vec{e}_{i}=\left( \delta _{ij}\right) _{j=1,\dots ,n},
\end{equation*}
spans the tangent space to $\partial \mathcal{R}$ at $y$.
\end{remark}
{A structural condition that we impose on the matrix $\mathcal{A}$ in (\ref%
{equation}) is that }it is equivalent to a diagonal matrix in the following
sense:
\begin{axiom}[Diagonal condition]
\label{hellipp} For $\vec{k}$ as in Condition \ref{hyp1}, we assume that for
some $\Lambda \geq 1$, the matrix $\mathcal{A}$ satisfies
\begin{equation}
\sum_{i=1}^{n}k^{i}\left( x,z\right) \xi _{i}^{2}\leq \xi ^{t}\mathcal{A}
\left( x,z\right) \xi \leq \Lambda \sum_{i=1}^{n}k^{i}\left( x,z\right) \xi
_{i}^{2} \label{hellip}
\end{equation}
for all $\xi \in \mathbb{R}^{n}$ and $\left( x,z\right) \in \Gamma =\Omega
\times \mathbb{R}$.
\end{axiom}
\begin{remark}
Because of Condition \ref{hellipp} or Remark \ref{spans}, we can state
property (\ref{ax1c3}) in the nondegeneracy Condition \ref{hyp1} in terms of
the matrix $\mathcal{A}$ as follows: For every $x\in \Omega $ and $%
\varepsilon $ with $0<\varepsilon <n^{-\frac{1}{2}}\mathop{\rm dist}\left(
x,\partial \Omega \right) $, there exist $\vec{r}=\left( r^{1},\dots
,r^{n}\right) $ with $0<r^{1},\dots ,r^{n}<\varepsilon $ and a box $\mathcal{%
\ R}=\mathcal{R}\left( \tilde{x},\vec{r}\right) $ such that $x\in \frac{1}{3}
\mathcal{R}$ and for every $y\in \partial \mathcal{R}$ and nonzero $\vec{v}
\left( y\right) $ \emph{tangent} to $\partial \mathcal{R}$ at $y$,
\begin{equation}
\vec{v}\left( y\right) \cdot \mathcal{A}\left( y,z\right) \vec{v}\left(
y\right) >0\qquad \text{for all }z\in \mathbb{R}\text{.} \label{lidAk}
\end{equation}
\end{remark}
{Recall that a domain }$\Omega ${\ is an open connected set. $\Omega
^{\prime }$ will always denote a domain with compact closure in $\Omega$;
this will be abbreviated }$\Omega ^{\prime }\Subset \Omega ${. }
\begin{definition}[Subunit type]
{\label{subunitt}}We say that a vector field $G=\sum_{i=1}^{n}\gamma
^{i}\left( x,z\right) \frac{\partial }{\partial x_{i}}$ with bounded
coefficients $\gamma ^{i}$ is of \emph{subunit type} with respect to $%
\mathcal{A}$ {in $\Gamma =\Omega \times \mathbb{R}$ if for every $\Omega
^{\prime }\Subset \Omega $ and $M_{0}\geq 1,$ there is a constant $B_{\gamma
}=B_{\gamma }\left( \Omega ^{\prime },M_{0}\right) >0$ such that
\begin{equation*}
\left( \sum_{i=1}^{n}\gamma ^{i}\left( x,z\right) \xi _{i}\right) ^{2}\leq
B_{\gamma }^{2}~\xi ^{t}\mathcal{A}\left( x,z\right) \xi \quad \text{for all
}\left( x,z\right) \in \Gamma _{M_{0}}^{\prime }=\Omega ^{\prime }\times %
\left[ -M_{0},M_{0}\right], \xi \in \mathbb{R}^n.
\end{equation*}
\ }
\end{definition}
We will impose further conditions on $\mathcal{A}$. To motivate them, we
recall the classical inequality of Wirtinger type on a domain $\Phi \subset
\mathbb{R}^{n+1}:$ if $k$ is nonnegative with bounded second derivatives on $%
\Phi$, then
\begin{equation}
\left\vert \nabla _{\!\!x,z}k\left( x,z\right) \right\vert \leq C\left\{
\left\Vert \nabla _{\!\!x,z}^{2}k\right\Vert _{L^{\infty }\left( \Phi
\right) }^{\frac{1}{2}}+\left( \mathop{\rm dist}\left( \left( x,z\right)
,\partial \Phi \right) \right) ^{-\frac{1}{2}}\right\} \sqrt{k\left(
x,z\right) } \label{genWirt}
\end{equation}
for all $\left( x,z\right) \in \Phi $ (see e.g. the appendix in \cite{SaW}).
{Inequality (\ref{genWirt}) is crucial in our calculations, and although it
has an analogue for nonnegative diagonal matrices, it does not extend to
general matrix functions.}
\begin{definition}[Subordinate matrix]
\label{submat}{We say that $\mathcal{A}$ is \emph{subordinate} in $\Gamma
=\Omega \times \mathbb{R}$ if for every $\Omega ^{\prime }\Subset \Omega $
and $M_{0}\geq 1$, there exists $B_{\mathcal{A}}=B_{\mathcal{A}}\left(
\Omega ^{\prime },M_{0}\right) >0$ such that
\begin{equation}
\sum_{i=1}^{n}\left\vert \partial_i\mathcal{A}\left( x,z\right) \xi
\right\vert ^{2}+\left\vert \partial_z\mathcal{A}\left( x,z\right) \xi
\right\vert ^{2}\leq B_{\mathcal{A}}^{2}~\xi ^{t}\mathcal{A}\left(
x,z\right) \xi \label{wirtm}
\end{equation}
}for all $\xi \in \mathbb{R}^{n}$, $\left( x,z\right) \in \Gamma
_{M_{0}}^{\prime }$, {where $\Gamma_{M_0}^{\prime }=\Omega ^{\prime }\times %
\left[ -M_{0},M_{0}\right] $. }
\end{definition}
{We always consider locally bounded solutions $w$, i.e. $\left\Vert
w\right\Vert _{L^{\infty }\left( \Omega ^{\prime }\right) }<\infty $ for all
subdomains $\Omega ^{\prime }\Subset \Omega $. Thus, we deal only with a
solution $w$ whose graph on $\Omega ^{\prime }$ is contained in a \emph{%
bounded} connected set
\begin{equation}
\Gamma _{M_{0}}^{\prime }=\Omega ^{\prime }\times \left[ -M_{0},M_{0}\right]
\Subset \Gamma =\Omega \times \mathbb{R} \label{domain}
\end{equation}%
for some $M_{0}=M_{0}\left( w,\Omega ^{\prime }\right) <\infty $. For
convenience we also assume $M_{0}\geq 1$.\ }
{To obtain our main results, we will use the following a priori estimates
obtained in \cite{RSaW1} for the class of equations (\ref{equation} ). \ }
\begin{theorem}[Theorem 1.8 in~%
\cite{RSaW1}%
]
{\ \label{quasi}Let $\Omega $ be a bounded domain in $\mathbb{R}^{n}$ and $%
\Gamma =\Omega \times \mathbb{R}$. Let $\vec{k}\left( x,z\right) \in
\mathcal{C}^{2}\left( \Gamma \right) $ and $\mathcal{A}\left( x,z\right) $, $%
f\left( x,z\right) $ and $\vec{\gamma}\left( x,z\right) \in \mathcal{C}
^{\infty }\left( \Gamma \right) $. Suppose that}
\begin{enumerate}
\item {$\mathcal{A\,}$satisfies (\ref{hellip}), where $\vec{k}\left(
x,z\right) $ satisfies the nondegeneracy }Condition{\ \ref{hyp1} in $\Gamma $
,}
\item {$\mathcal{A\,}$is subordinate in $\Gamma $ (Definition \ref{submat}),}
\item {\ $\vec{\gamma}$ is of subunit type with respect to }$\mathcal{A}${\
in $\Gamma $ (Definition \ref{subunitt}). }
\end{enumerate}
{Then for every smooth solution $w$ of (\ref{equation}) in $\Omega $,
integer $N\geq 0$ and subdomains $\Omega ^{\prime \prime }\Subset \Omega
^{\prime }\Subset \Omega ,$ there exists a constant }$\mathcal{C}
_{\left\Vert w\right\Vert _{L^{\infty }\left( \Omega ^{\prime }\right)
},\left\Vert \nabla w\right\Vert _{L^{\infty }\left( \Omega ^{\prime
}\right) },N}=${\
\begin{equation*}
\mathcal{C}_{\left\Vert w\right\Vert _{L^{\infty }\left( \Omega ^{\prime
}\right) },\left\Vert \nabla w\right\Vert _{L^{\infty }\left( \Omega
^{\prime }\right) },N}\left( n,B,\vec{k},\Lambda ,\left\Vert \mathcal{A}
\right\Vert _{\mathcal{C}^{N+2}\left( \tilde{\Gamma}\right) },\left\Vert
f\right\Vert _{\mathcal{C}^{N+1}\left( \tilde{\Gamma}\right) },\left\Vert
\vec{\gamma}\right\Vert _{\mathcal{C}^{N+1}\left( \tilde{\Gamma}\right)
},\Omega ,\Omega ^{\prime },\Omega ^{\prime \prime }\right)
\end{equation*}
such that
\begin{equation}
\sum_{\left\vert \vec{\alpha}\right\vert \leq N}\left\Vert D^{\alpha
}w\right\Vert _{L^{\infty }\left( \Omega ^{\prime \prime }\right) }\leq
\mathcal{C}_{\left\Vert w\right\Vert _{L^{\infty }\left( \Omega ^{\prime
}\right) },\left\Vert \nabla w\right\Vert _{L^{\infty }\left( \Omega
^{\prime }\right) },N}. \label{control}
\end{equation}
Here $\tilde{\Gamma}=\Omega ^{\prime }\times \left[ -2\left\Vert
w\right\Vert _{L^{\infty }\left( \Omega ^{\prime }\right) },2\left\Vert
w\right\Vert _{L^{\infty }\left( \Omega ^{\prime }\right) }\right] $, and $B$
denotes $B_\gamma$, $B_{\mathcal{A}}$. }
\end{theorem}
{Our main application of these a priori estimates is the hypoellipticity
result stated in Theorem \ref{application} for (infinitely degenerate)
quasilinear equations of the form (\ref{equation}). In it, as in the special
two-dimensional case contained in \cite{SaW}, we will assume extra
conditions on the coefficients, namely, that the \emph{nonlinear} and the
\emph{\ infinitely degenerate} characters do not occur simultaneously in the
sense described below. We denote}%
\begin{equation}
{k^{\ast }\left( x,z\right) =\min_{i=1,\dots ,n}\left\{ k^{i}\left(
x,z\right) \right\} .} \label{kstar}
\end{equation}
\begin{axiom}[Super Subordination condition]
{\label{hyp2}We say that $\mathcal{A}$ satisfies the super subordination
condition in $\Gamma =\Omega \times \mathbb{R}$} if {for every $\Omega
^{\prime }\Subset \Omega $ and $M_{0}\geq 1,$ there exist constants $B_{
\mathcal{A}}=B_{\mathcal{A}}\left( \Omega ^{\prime },M_{0}\right) $ and }$%
B_{ \mathcal{A}}^{\prime }=B_{\mathcal{A}}^{\prime }\left( \Omega ^{\prime
},M_{0}\right) ${\ such that if }$\left( x,z\right) \in \Gamma
_{M_{0}}^{\prime }$ and $~\xi \in \mathbb{R}^{n},$ then{\
\begin{eqnarray}
\left\vert \partial _{z}\mathcal{A}\left( x,z\right) \xi \right\vert ^{2}
&\leq &B_{\mathcal{A}}^{2}~k^{\ast }\left( x,z\right) \,\xi ^{t}\mathcal{A}
\left( x,z\right) \xi , \label{xtra} \\
\sum_{i=1}^{n}\left\vert \partial _{i}\partial _{z}\mathcal{A}\left(
x,z\right) \xi \right\vert ^{2}+\left\vert \partial _{z}^{2}\mathcal{A}
\left( x,z\right) \xi \right\vert ^{2} &\leq &\left( B_{\mathcal{A}}^{\prime
}\right) ^{2}\,\xi ^{t}\mathcal{A}\left( x,z\right) \xi . \label{xxtra}
\end{eqnarray}
\indent%
\noindent If $\mathcal{A}$ is diagonal, then condition (\ref{xxtra}) follows
from (\ref{xtra}) by Wirtinger's inequality: see Remark \ref{diagcase}. }
We say $\vec{\gamma}\left( x,z\right) $ satisfies the super subordination
condition if {for all }$\left( x,z\right) \in \Gamma _{M_{0}}^{\prime }$ and
$\xi \in \mathbb{R}^{n},$ {\
\begin{equation}
\left\vert \partial _{z}\vec{\gamma}\left( x,z\right) \cdot \xi \right\vert
^{2}\leq B_{\gamma }^{2}~k^{\ast }\left( x,z\right) \xi ^{t}\mathcal{A}%
\left( x,z\right) \xi , \label{gxtra}
\end{equation}%
for some $B_{\gamma }=B_{\gamma }\left( \Omega ^{\prime },M_{0}\right) $.}
\end{axiom}
{The extra vanishing }condition{\ (\ref{xtra}) on $\partial _{z}\mathcal{A}$
is a stronger form of the part of (\ref{wirtm}) involving $\partial _{z}
\mathcal{A}$. In the two-dimensional diagonal case $\mathcal{A}=
\mathop{\rm
diag}\left( 1,k\right) $, inequality (\ref{wirtm}) always holds for any $%
\mathcal{C}^{2}$ nonnegative $k\left( x,z\right) $, and it takes the form (%
\ref{genWirt}), while the more restrictive (\ref{xtra}) with $k^{\ast }=k$
takes the form
\begin{equation}
\left\vert \partial _{z}k\left( x,z\right) \right\vert \leq B~k\left(
x,z\right) \label{thw2}
\end{equation}
(compare (\ref{thw})), which does not hold in general for nonnegative $%
k\left( x,z\right) $. On the other hand, if $f\left( x\right) $ is any
smooth nonnegative function in $\mathbb{R}^{n}$ and $h\left( x,z\right) $ is
a nonnegative Lipschitz function in $\mathbb{R}^{n+1}$, then
\begin{equation*}
k\left( x,z\right) =f\left( x\right) \left[ 1+h\left( x,z\right) \right]
\end{equation*}
satisfies (\ref{thw2}). Indeed, we have the following lemma; see Section 6.4
in the appendix of \cite{SaW} for details. }
\begin{lemma}[\protect\cite{SaW}]
{\ \label{chaz}Let $k(x,z)$ be a smooth nonnegative function in a bounded
region $T \subset \mathbb{R}^{n}\times \mathbb{R}$, and assume that for some
$\gamma, B \geq 1$,
\begin{equation}
\left\vert \partial _{z}k(x,z)\right\vert \leq B\, k\left( x,z\right)
^{\gamma }. \label{xwirt}
\end{equation}%
Then, for every $\left( x_{0},z_{0}\right) \in T $, there exists a smooth
function $f(x)\geq 0$ and a Lipschitz function $h(x,z)$, with Lipschitz
constant depending only on $B$, $\left\Vert k\right\Vert _{L^{\infty}(T)}$
and $T$, such that
\begin{equation}
k\left( x,z\right) =f\left( x\right) \left( 1+f\left( x\right) ^{\gamma
-1}h\left( x,z\right) \right) \label{charat}
\end{equation}%
for all $\left( x,z\right) $ in a neighborhood of $\left( x_{0},z_{0}\right)
$. Moreover, $h\left( x_{0},z_{0}\right) =0$. In particular,
\begin{equation*}
C^{-1}k\left( x,z^{\prime }\right) \leq k\left( x,z\right) \leq Ck\left(
x,z^{\prime }\right) ,\qquad \left( x,z\right) ,\,\left( x,z^{\prime
}\right) \in T ,
\end{equation*}
where $C=C\left( B,\mathop{\rm diam}T \right) $. Conversely, if $h\left(
x,z\right) $ is smooth and $f\left( x\right) \,$is a nonnegative smooth
function such that $f\left( x\right) ^{\gamma }$ is smooth, then $k\left(
x,z\right) $ given by (\ref{charat}) is smooth and satisfies (\ref{xwirt})
for some $B=B\left( h,f,\mathop{\rm
diam}T \right) $.}
\end{lemma}
\begin{remark}
\label{diagcase} As noted earlier, if $\mathcal{A}$ is diagonal then the
second extra vanishing condition (\ref{xxtra}) follows from the first one (%
\ref{xtra}) by (\ref{genWirt}). Indeed, it is enough to prove this for a
scalar function $k\left( x,z\right) $. If (\ref{thw2}) holds, then
\begin{equation*}
\tilde{k}\left( x,z\right) : =\partial _{z}k\left( x,z\right) +B~k\left(
x,z\right) \geq 0,
\end{equation*}
so by (\ref{genWirt}), {for all }$\left( x,z\right) \in \Gamma
_{M_{0}}^{\prime }$,
\begin{equation*}
\left\vert \nabla _{\!\!x,z}\tilde{k}\left( x,z\right) \right\vert \leq
C\left\{ \left\Vert \nabla _{\!\!x,z}^{2}\tilde{k}\right\Vert _{L^{\infty
}\left( \Gamma _{2M_{0}}^{\prime \prime }\right) }^{\frac{1}{2}}+\left( %
\mathop{\rm dist}\left( \left( x,z\right) ,\partial \Gamma _{2M_{0}}^{\prime
\prime }\right) \right) ^{-\frac{1}{2}}\right\} \sqrt{\tilde{k}\left(
x,z\right), }
\end{equation*}
where $\Gamma _{2M_{0}}^{\prime \prime }=\Omega ^{\prime \prime }\times %
\left[ -2M_{0},2M_{0}\right] $ with $\Omega ^{\prime }\Subset \Omega
^{\prime \prime }\Subset \Omega $. Hence from (\ref{thw2}) we obtain
\begin{equation*}
\sum_{i=1}^{n}\left\vert \partial _{i}\partial _{z}k\right\vert
^{2}+\left\vert \partial _{z}^{2}k\right\vert ^{2}\leq \tilde{B}%
^{2}\,k\left( x,z\right)
\end{equation*}
for a suitable constant $\tilde{B}$.
\end{remark}
{Under the extra assumptions in }the super subordination Condition{\ \ref%
{hyp2}, we will obtain interior a priori control of all derivatives
(including first order ones) of smooth solutions in terms of the supremum
norm of the solutions: }
\begin{theorem}[A priori estimate]
{\ \label{apriorial}Let $\Omega \subset \mathbb{R}^{n}$ be a bounded domain.
Suppose that $\Gamma $, $\mathcal{A}\left( x,z\right) $, $f\left( x,z\right)
$, $\vec{\gamma}\left( x,z\right) $ and $\vec{k}\left( x,z\right) $ are as
in Theorem \ref{quasi}, and also that $\mathcal{A}$ and $\vec{\gamma}$
satisfy }the super subordination Condition{\ \ref{hyp2} in $\Gamma $. If $w$
is a smooth solution of (\ref{equation}) in $\Omega $, then for any
nonnegative integer $N$ and subdomain $\Omega ^{\prime }\Subset \Omega $,
there exists a constant }$\mathcal{C}_{\left\Vert w\right\Vert _{L^{\infty
}\left( \Omega ^{\prime }\right) },N}=${\
\begin{equation*}
\mathcal{C}_{\left\Vert w\right\Vert _{L^{\infty }\left( \Omega ^{\prime
}\right) },N}\left( n,B,\vec{k},\Lambda ,\left\Vert \mathcal{A}\right\Vert _{%
\mathcal{C}^{N+3}\left( \tilde{\Gamma}\right) },\left\Vert f\right\Vert _{%
\mathcal{C}^{N+3}\left( \tilde{\Gamma}\right) },\left\Vert \vec{\gamma}%
\right\Vert _{\mathcal{C}^{N+3}\left( \tilde{\Gamma}\right) },\Omega ,\Omega
^{\prime }\right)
\end{equation*}%
such that
\begin{equation*}
\sum_{\left\vert \vec{\alpha}\right\vert \leq N}\left\Vert D^{\vec{\alpha}%
}w\right\Vert _{L^{\infty }\left( \Omega ^{\prime }\right) }\leq \mathcal{C}%
_{\left\Vert w\right\Vert _{L^{\infty }\left( \Omega ^{\prime }\right) },N}.
\end{equation*}%
Here $\tilde{\Gamma}=\Omega ^{\prime }\times \left[ -2\left\Vert
w\right\Vert _{L^{\infty }\left( \Omega ^{\prime }\right) },2\left\Vert
w\right\Vert _{L^{\infty }\left( \Omega ^{\prime }\right) }\right] $ and $B$
denotes $B_{\gamma },B_{\mathcal{A}},B_{\mathcal{A}^{\prime }}$. }
\end{theorem}
\begin{remark}
{\ A special case of Theorem \ref{apriorial} is established in Theorem 2.4
in \cite{SaW}, namely when $n=2$, $\vec{k}\left( x_{1},x_{2},z\right)
=\left( 1,k\left( x_{1},z\right) \right) $ is independent of $x_{2}$ with
\begin{equation}
\left\vert \partial _{z}k\left( x_{1},z\right) \right\vert \leq Ck\left(
x_{1},z\right) ^{\frac{3}{2}}, \label{thww}
\end{equation}
and $f=0$, $\vec{\gamma}=0$. Also, much more is required there of the
solution $w$, namely $w$ must satisfy (see (2.22) in \cite{SaW})
\begin{equation*}
\begin{array}{rrl}
1+\left( \partial _{x_{1}}w\left( x_{1},x_{2}\right) \right) ^{2} & \leq &
C~\partial _{x_{2}}w\left( x_{1},x_{2}\right) , \\
k\left( x_{1},w\left( x_{1},x_{2}\right) \right) ~\partial _{x_{2}}w\left(
x_{1},x_{2}\right) & \leq & C.%
\end{array}%
\end{equation*}
These restrictions are removed in Theorem \ref{apriorial}, in which we also
generalize the result to higher dimensions and allow lower order terms. }%
\newline
\indent%
{An important improvement found in Theorem \ref{apriorial} relative to
Theorem 2.4 in \cite{SaW} is the reduction of the power $3/2$ in (\ref{thww}%
) to the sharp power $1$. See Section \ref{section-sharp}.}
\end{remark}
The next lemma shows that if only the first part of Condition \ref{hyp2}
holds, then the bilinear form induced by $\mathcal{A}\left( x,z\right) $ is
equivalent to one which is independent of $z$ in any set on which $z$ is
bounded.
\begin{lemma}
\label{admisall}{Let $\mathcal{A}=\left( a_{ij}\left( x,z\right) \right)
_{i,j=1,\dots ,n}$ be a smooth symmetric matrix satisfying (\ref{hellip}) in
$\Gamma $ and such that for every $\Omega ^{\prime }\Subset \Omega $ and $%
M_{0}\geq 1,$ there exists $\tilde{B}=\tilde{B}_{\mathcal{A}}\left( M_{0}, %
\mathop{\rm diam}\Omega ,\mathop{\rm dist}\left( \Omega ^{\prime },\partial
\Omega \right) \right) $ such that
\begin{equation}
\left\vert \partial _{z}\mathcal{A}\left( x,z\right) \xi \right\vert
^{2}\leq \tilde{B}^{2}k^{\ast }\left( x,z\right) \,\xi ^{t}\mathcal{A}\left(
x,z\right) \xi ,\qquad \left( x,z\right) \in \Gamma _{M_{0}}^{\prime },~\xi
\in \mathbb{R}^{n}. \label{extratwo}
\end{equation}
Then there exists $\mathcal{C}=\mathcal{C}\left( \tilde{B},M_{0}\right) $
such that for all }$\left( x,z\right) ,\left( x,\tilde{z}\right) \in \Gamma
_{M_{0}}^{\prime }$ and $\,\xi \in \mathbb{R}^{n},$
\begin{equation*}
\mathcal{C}^{-1}\,\xi ^{t}\mathcal{A}\left( x,z\right) \xi \leq \xi ^{t}%
\mathcal{A}\left( x,\tilde{z}\right) \xi \leq \mathcal{C}\,\xi ^{t}\mathcal{%
A }\left( x,z\right) \xi .
\end{equation*}
Moreover, for all $i=1,\cdots ,n$ and $\left( x,z\right) ,\left( x,\tilde{z}
\right) \in \Gamma _{M_{0}}^{\prime },$
\begin{equation}
\mathcal{\tilde{C}}^{-1}k^{i}\left( x,z\right) \leq k^{i}\left( x,\tilde{z}
\right) \leq \mathcal{\tilde{C}}\,k^{i}\left( x,z\right), \label{chaak}
\end{equation}
where $\mathcal{\tilde{C}}=\mathcal{\tilde{C}}\left( \mathcal{C},\Lambda
\right) $.
\end{lemma}
\proof%
{By (\ref{extratwo}), for all $\xi \in \mathbb{R}^{n}$, the function $%
h\left( x,z,\xi \right) =\xi ^{t}\mathcal{A}\left( x,z\right) \,\xi $
satisfies
\begin{eqnarray*}
\left\vert \partial _{z}h\left( x,z,\xi \right) \right\vert ^{2}
&=&\left\vert \xi ^{t}\mathcal{A}_{z}\left( x,z\right) \,\xi \right\vert ^{2}
\\
&\leq &\left\vert \xi \right\vert ^{2}\left\vert \mathcal{A}_{z}\left(
x,z\right) \,\xi \right\vert ^{2} \\
&\leq &\left\vert \xi \right\vert ^{2}\tilde{B}^{2}k^{\ast }\left(
x,z\right) \,\xi ^{t}\mathcal{A}\left( x,z\right) \xi \\
&=&\left\vert \xi \right\vert ^{2}\tilde{B}^{2}k^{\ast }\left( x,z\right)
h\left( x,z,\xi \right) ,
\end{eqnarray*}
for all $\left( x,z\right) \in \Gamma _{M_{0}}^{\prime }=\Omega ^{\prime
}\times \left[ -M_{0},M_{0}\right] $ and$~\xi \in \mathbb{R}^{n}$. Since
from (\ref{hellip}) we have
\begin{equation*}
k^{\ast }\left( x,z\right) \left\vert \xi \right\vert
^{2}=\sum_{i=1}^{n}k^{\ast }\left( x,z\right) \xi _{i}^{2}\leq
\sum_{i=1}^{n}k^{i}\left( x,z\right) \xi _{i}^{2}\leq \xi ^{t}\mathcal{A}
\left( x,z\right) \xi =h\left( x,z,\xi \right) ,
\end{equation*}
we obtain
\begin{equation*}
\left\vert \partial _{z}h\left( x,z,\xi \right) \right\vert \leq \tilde{B}
\,h\left( x,z,\xi \right) .
\end{equation*}
By Gronwall's inequality it follows that for some $C=C\left( \tilde{B}
,M_{0}\right) ,$
\begin{equation*}
C^{-1}\,\xi ^{t}\mathcal{A}\left( x,z\right) \xi \leq \xi ^{t}\mathcal{A}
\left( x,\tilde{z}\right) \xi \leq C\,\xi ^{t}\mathcal{A}\left( x,z\right)
\xi ,\qquad \left( x,z\right) ,\left( x,\tilde{z}\right) \in \Gamma
_{M_{0}}^{\prime },\,\xi \in \mathbb{R}^{n}.
\end{equation*}
Inequality (\ref{chaak}) then follows immediately from (\ref{hellip}). }
\endproof%
\subsection{Hypoellipticity Main Results}
{Our main results Theorems \ref{DP} and \ref{application} are obtained as
applications of the a priori estimates above. Theorem \ref{application}
establishes smoothness of weak solutions, and Theorem \ref{DP} deals with
existence, uniqueness, and regularity of strong solutions to the Dirichlet
problem. The concept of weak solutions of our infinitely degenerate
operators is similar to that of classical weak solutions for accretive
operators, defined via the associated Hilbert space. We denote by }$H_{%
\mathcal{X}}^{1,2}\left( \Omega \right) ${\ the Hilbert space on }$\Omega ${%
\ induced by the quadratic form }$\mathcal{X}\left( x,\xi \right)
=\sum_{i=1}^{n}k^{i}\left( x,0\right) \xi _{i}^{2}$ (see Appendix, Section %
\ref{Section-Weak-Sol}). {We say that }$w$ is a weak solution of (\ref%
{equation}) in $\Omega $ if $w\in H_{\mathcal{X}}^{1,2}\left( \Omega \right)
\bigcap L^{\infty }\left( \Omega \right) $ and {\ }
\begin{equation*}
\int \mathcal{A}\left( x,w\right) \nabla w\cdot \nabla \varphi ~dx-\int
\varphi \vec{\gamma}\left( x,w\right) \cdot \nabla w~dx-\int f\left(
x,w\right) \varphi ~dx=0
\end{equation*}%
for all $\varphi \in Lip_{0}(\Omega )$. See Section \ref{deg-sob-section} of
the Appendix for a detailed discussion of the degenerate Sobolev spaces $H_{%
\mathcal{X}}^{1,2}\left( \Omega \right) $ and the meaning of $\nabla w$ if $%
w\in H_{\mathcal{X}}^{1,2}\left( \Omega \right) $.
Some extra conditions are required in order to make our approximation scheme
work.
\begin{definition}[Strongly convex domain]
{\ \label{strcv} A convex domain $\Phi \subset \mathbb{R}^{n}$ with $%
\partial \Phi \in \mathcal{C}^{2}$ is called \emph{\ strongly convex} (with
convex character $\lambda_0 $) if there exists $\lambda_0 >0$ such that
\begin{equation*}
\inf_{p\in \partial \Omega }\min_{1\leq i\leq n-1}\lambda ^{i}\left(
p\right) =\lambda_0 >0,
\end{equation*}
where $\lambda ^{1}\left( p\right) ,\dots ,\lambda ^{n-1}\left( p\right) $
denote the principal curvatures (see \cite{GiTr} p.354) of $\partial \Phi $
at a point $p\in \partial \Phi $. }
\end{definition}
{Theorem \ref{application} is obtained from the following existence and
uniqueness theorem for the Dirichlet problem.}
\begin{theorem}[Dirichlet problem]
\bigskip \label{DP}{Let $\widetilde{\Omega }\subset \mathbb{R}^{n}$ be a
bounded open set and let }$\Omega \Subset \widetilde{\Omega }${\ be a
strongly convex domain. For }$\Gamma =\widetilde{\Omega }\times \mathbb{R},$
suppose{\ $\vec{k}\in \mathcal{C}^{2}\left( \Gamma \right) $ and $\mathcal{A}%
\left( x,z\right) $, $f\left( x,z\right) $, $\vec{\gamma}\left( x,z\right)
\in \mathcal{C}^{\infty }\left( \Gamma \right) $ are such that }
\begin{enumerate}
\item {\ \label{appc01}$\vec{k}\left( x,z\right) $ satisfies }the
nondegeneracy Condition{\ \ref{hyp1} in $\Gamma $, }
\item {\ \label{appc02}$\mathcal{A\,}$satisfies the diagonal }Condition{\ %
\ref{hellipp} in $\Gamma $, }
\item {\ \label{appc07}$f\left( x,z\right) \mathop{\rm sign}z\leq 0\text{
and }f_{z}\left( x,z\right) \leq 0\text{ in }\Gamma $, }
\item {\ \label{appc04}$\vec{\gamma}$ }is of subunit type with respect to $%
\mathcal{A}${\ in $\Gamma $, }
\item {\ }\label{appc06}$\vec{\gamma}$ has compact support in $\Omega $ in
the $x$ variable, locally in the $z$ variable, i.e., for all $M_{0}>0$ there
exists open $\Omega ^{\prime }\Subset \Omega $ such that $\vec{\gamma}
\left( x,z\right) =0$ if $\left( x,z\right) \in \left( \Omega \backslash
\overline{ \Omega ^{\prime }}\right) \times \left[ -M_{0},M_{0}\right] $,
\item {\ }\label{appc03}{$\mathcal{A}$ satisfies }the super subordination
Condition{\ \ref{hyp2} in $\Gamma $}$,${\ }
\item {\ \label{appc05}$\vec{\gamma}$ satisfies }the super subordination
Condition{\ \ref{hyp2} in $\Gamma $}$.${\ }
\end{enumerate}
\end{theorem}
\noindent
\it%
{Then given any continuous function }$\varphi $ on $\partial \Omega$, there
exists a strong solution {$w$ to the Dirichlet problem}
\begin{equation}
\left\{
\begin{array}{rcll}
\mathop{\rm div}\mathcal{A}\left( x,w\right) \nabla w+\vec{\gamma}\left(
x,w\right) \cdot \nabla w+f\left( x,w\right) & = & 0\quad & \text{in }\Omega
\\
w & = & \varphi & \text{on }\partial \Omega ,%
\end{array}%
\right. \label{dirchlet}
\end{equation}%
{i.e., there exists} $w$ which is continuous in $\overline{\Omega }$, equal
to $\varphi $ on $\partial \Omega $, and a strong solution of the
differential equation in $\Omega $. {Moreover, $w\in \mathcal{C}^{0}\left(
\overline{\Omega }\right) $}$\bigcap \mathcal{C}^{\infty }\left( \Omega
\right) $, and {for any nonnegative integer $N$ and subdomain $\Omega
^{\prime }\Subset $}${\Omega }${, there exists a constant $\mathcal{C}_{N}=$
\begin{equation*}
\mathcal{C}_{N}\left( n,B,\vec{k},\Lambda ,\left\Vert \varphi \right\Vert
_{L^{\infty }\left( \partial \Omega \right) },\left\Vert \mathcal{A}%
\right\Vert _{\mathcal{C}^{N+2}\left( \tilde{\Gamma}\right) },\left\Vert
f\right\Vert _{\mathcal{C}^{N+1}\left( \tilde{\Gamma}\right) },\left\Vert
\vec{\gamma}\right\Vert _{\mathcal{C}^{N+1}\left( \tilde{\Gamma}\right)
},\lambda _{0},\Omega ,\Omega ^{\prime }\right)
\end{equation*}%
such that }$\sum_{\left\vert \vec{\alpha}\right\vert \leq N}\left\Vert D^{%
\vec{\alpha}}w\right\Vert _{L^{\infty }\left( \Omega ^{\prime }\right) }\leq
\mathcal{C}_{N}$. H{ere $\tilde{\Gamma}=\Omega \times \left[ -2\left\Vert
\varphi \right\Vert _{L^{\infty }\left( \partial \Omega \right)
},2\left\Vert \varphi \right\Vert _{L^{\infty }\left( \partial \Omega
\right) }\right] $, $B$ denotes the various constants in Condition \ref{hyp2}%
, and }${\lambda _{0}}${$=\lambda _{0}\left( \mathcal{A},\Omega \right) $ is
the convex character of }$\partial \Omega ${. }\newline
Moreover, {if }$\vec{\gamma}\equiv 0$, then the solution $w$ is unique.
\rm%
An important consequence of Theorem \ref{DP} in the case $\vec{\gamma}\equiv
0$ is the following interior regularity result:
\begin{theorem}[Regularity of solutions]
{\ \label{application}Let $\Omega \subset \mathbb{R}^{n}$ be a bounded
domain and $\Gamma =\Omega \times \mathbb{R}$. Suppose that $\vec{k}\in
\mathcal{C}^{2}\left( \Gamma \right) $, that $\mathcal{A}\left( x,z\right) $%
, $f\left( x,z\right) \in \mathcal{C}^{\infty }\left( \Gamma \right) $ and
satisfy (\ref{appc01}),(\ref{appc02}), (\ref{appc07}), (\ref{appc03}) from
Theorem \ref{DP}, and that }$\mathcal{A}$ satisfies the super subordination
Condition{\ \ref{hyp2}} in $\Gamma ${. Then any weak solution $w$ of}
\begin{equation*}
\mathop{\rm div}\mathcal{A}\left( x,w\right) \nabla w+f\left( x,w\right)
=0\qquad \text{in }\Omega
\end{equation*}%
{which is \emph{continuous} in $\Omega $ is also a strong solution and
satisfies $w\in \mathcal{C}^{\infty }\left( \Omega \right) $. Moreover, for
any nonnegative integer $N$ and subdomain $\Omega ^{\prime \prime }\Subset
\Omega ^{\prime }\Subset \Omega $, there exists a constant
\begin{equation*}
\mathcal{C}_{N}=\mathcal{C}_{N}\left( \left\Vert w\right\Vert _{L^{\infty
}\left( \Omega ^{\prime }\right) },n,B,\vec{k},\Lambda ,\left\Vert \mathcal{A%
}\right\Vert _{\mathcal{C}^{N+2}\left( \tilde{\Gamma}\right) },\left\Vert
f\right\Vert _{\mathcal{C}^{N+1}\left( \tilde{\Gamma}\right) },\Omega
,\Omega ^{\prime },\Omega ^{\prime \prime }\right)
\end{equation*}%
such that }$\sum_{\left\vert \vec{\alpha}\right\vert \leq N}\left\Vert D^{%
\vec{\alpha}}w\right\Vert _{L^{\infty }\left( \Omega ^{\prime \prime
}\right) }\leq \mathcal{C}_{N}${. Here $\tilde{\Gamma}=\Omega ^{\prime
}\times \left[ -2\left\Vert w\right\Vert _{L^{\infty }\left( \Omega ^{\prime
}\right) },2\left\Vert w\right\Vert _{L^{\infty }\left( \Omega ^{\prime
}\right) }\right] $ and $B$ denotes the relevant constants in Condition \ref%
{hyp2}.}
\end{theorem}
{Note that in case $n=2,$ }the super subordination Condition{\ \ref{hyp2}
reduces to (\ref{thw2}) if $\mathcal{A}=\mathop{\rm diag}\left( 1,k\right) $
and $\vec{\gamma}=0$. Moreover, in this case, whether $\vec{\gamma}=0$ or
not, }the nondegeneracy Condition{\ \ref{hyp1} means that given $\left(
x_{1},x_{2}\right) \in \Omega $ and $\varepsilon >0$, there exist $%
0<r^{1},r^{2}<\varepsilon $ such that $k\left( x_{1}\pm r^{1},\xi
_{2},z\right) >0$ if $\left\vert \xi _{2}-x_{2}\right\vert \leq r^{2}$. \ }
{As an example in case $n\geq 2$, we consider a diagonal matrix
\begin{equation*}
\mathcal{A}\left( x,z\right) =\left(
\begin{array}{cccc}
1 & 0 & \cdots & 0 \\
0 & k^{2}\left( x,z\right) & \cdots & 0 \\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \cdots & k^{n}\left( x,z\right)%
\end{array}%
\right) ,
\end{equation*}%
where $k^{i}$ are smooth nonnegative functions satisfying }the nondegeneracy
Condition{\ \ref{hyp1}, and such that
\begin{equation*}
\left\vert k_{z}^{i}\left( x,z\right) \right\vert \leq B\,k^{\ast }\left(
x,z\right) ,\qquad i=1,\dots ,n,\quad \left( x,z\right) \in \Gamma ,
\end{equation*}%
with $k^{\ast }=\min \left( k^{1},\dots ,k^{n}\right) $. Then $\mathcal{A}$
satisfies the hypotheses of Theorem \ref{application}. In particular, if $%
k\left( x,z\right) $ is nonnegative and satisfies $\left\vert
k_{z}\right\vert =O\left( k\right) $ as $k\rightarrow 0$, then $\mathcal{A}=%
\mathop{\rm diag}\left( 1,k,\dots ,k\right) $ is an admissible matrix for
Theorem \ref{application} provided property (\ref{ax1c3}) of }the
nondegeneracy Condition{\ \ref{hyp1} holds. }
\begin{remark}
{\ As a consequence of the previous observations in the case $\mathcal{A}= %
\mathop{\rm diag}\left( 1,k^{2},\dots ,k^{n}\right) $, we \emph{partially}
recover Fedi\u{\i}'s two-dimensional result \cite{Fe} that $\partial
_{x_{1}}^{2}+k\left( x_{1}\right) \partial _{x_{2}}^{2}$ is hypoelliptic if $%
k$ is smooth and positive for all $x_{1}\neq 0$. Indeed, since $k\left(
x_{1}\right) $ is independent of the $z$ variable it automatically satisfies
$\left\vert k_{z}\right\vert =O\left( k\right) $ as $k\rightarrow 0$. We
only partially recover Fedi\u{\i}'s result because our theorem applies only
to continuous weak solutions. }
\end{remark}
We {also obtain a partial extension (namely, for \emph{continuous} weak
solutions) to higher dimensions of Fedi\u{\i}'s result:}
\begin{theorem}
\label{fediihigh} {Let $k^{i}\left( x_{1},\dots ,x_{n}\right) $, $i=2,\dots
,n$,$\,$be smooth functions in $\mathbb{R}^{n}$ such that $k^{i}$ is
independent of the $i^{\text{th}}$ variable, i.e.,
\begin{equation*}
k^{i}\left( x_{1},\dots ,x_{n}\right) =k^{i}\left( \hat{x}^{i}\right) ,\quad
\text{with }\hat{x}^{i}=\left( x_{1},\dots ,x_{i-1},x_{i+1},\dots
,x_{n}\right) ,
\end{equation*}%
and $k^{i}\left( x\right) >0$ if $\hat{x}^{i}\neq 0$. Then any continuous
weak solution of
\begin{equation*}
\left\{ \partial _{x_{1}}^{2}+k^{2}\left( x\right) \partial
_{x_{2}}^{2}+\dots +k^{n}\left( x\right) \partial _{x_{n}}^{2}\right\} w=0
\end{equation*}%
in $\mathbb{R}^{n}\,$is smooth everywhere. }
\end{theorem}
\begin{remark}
{\ It is shown in \cite{KuStr} that if $k\left( x_{1}\right) $ is smooth and
positive for all $x_{1}\neq 0$, then $\partial _{x_{1}}^{2}+k\left(
x_{1}\right) \partial _{x_{2}}^{2}+\partial _{x_{3}}^{2}$ is hypoelliptic in
$\mathbb{R}^{3}$ if and only if $\lim_{x_{1}\rightarrow 0}x_{1}\log k\left(
x_{1}\right) =0$. We are not able to recover this result since $k\left(
x_{1}\right) $ vanishes identically at all points of the form $(x_1,x_2,x_3)
= (0, x_2,x_3)$, and so the hypothesis of our theorem is not met. Also, our
solutions are required to be continuous. On the other hand, from Theorem \ref%
{fediihigh}, we see that if $k\left( x_{1},x_{3}\right) $ is smooth and
positive for all $\left( x_{1},x_{3}\right) \neq \left( 0,0\right) $, then
every weak solution of $\left\{ \partial _{x_{1}}^{2}+k\left(
x_{1},x_{3}\right) \partial _{x_{2}}^{2}+\partial _{x_{3}}^{2}\right\} w=0$
is smooth in the interior of the domain of continuity of $w$ in $\mathbb{R}%
^{3}$. }
\end{remark}
{The following is a striking consequence of Theorem \ref{application} in $%
\mathbb{R}^{2}$:}
\begin{theorem}
\label{striking} \label{RSaW3}{\ If $k\left( x_{1},x_{2},z\right) \,$is
smooth, nonnegative, satisfies
\begin{equation}
\left\vert \partial _{z}k\right\vert =O\left( k\right) , \label{thw1}
\end{equation}%
and $k\left( \cdot ,\cdot ,0\right) $ \emph{does not vanish identically on
any horizontal line segment} in $\Omega $, then any continuous weak solution
$w$ of
\begin{equation}
\partial _{x_{1}}^{2}w+\partial _{x_{2}}k\left( x_{1},x_{2},w\left(
x_{1},x_{2}\right) \right) \partial _{x_{2}}w=0,\;\;\;\;\;\left(
x_{1},x_{2}\right) \in \Omega , \label{maeasy}
\end{equation}%
is smooth in $\Omega $.}
\end{theorem}
{%
\proof%
We will prove that the hypotheses of Theorem \ref{application} are satisfied
by $\vec{k}=\left( 1,k\left( x_{1},x_{2},z\right) \right) $ and $\mathcal{A}%
=\left(
\begin{array}{ll}
1 & 0 \\
0 & k%
\end{array}%
\right) $. Since $k^{1}\equiv 1$, property (\ref{ax1c2}) in }the
nondegeneracy Condition{\ \ref{hyp1} is trivially satisfied. Since $k\left(
\cdot ,\cdot ,0\right) $ is nonnegative, continuous and does not vanish
identically on any horizontal segment, given $\varepsilon >0$ and $\left(
x_{1},x_{2}\right) \in \Omega $, there exist $\bar{x}_{1}<x_{1}<\bar{x}%
_{1}^{\prime }$ with $\left\vert x_{1}-\bar{x}_{1}\right\vert =\left\vert
x_{1}-\bar{x}_{1}^{\prime }\right\vert =r^{1}<\varepsilon $, such that $%
k\left( \bar{x}_{1},x_{2},0\right) >0$ and $k\left( \bar{x}_{1}^{\prime
},x_{2},0\right) >0$. From (\ref{thw1}) and Lemma \ref{chaz} it follows that
$k\left( x_{1},x_{2},z\right) $ is either identically zero as a function of $%
z$ or strictly positive in $z$, and hence $k\left( \bar{x}%
_{1},x_{2},z\right) >0$ and $k\left( \bar{x}_{1}^{\prime },x_{2},z^{\prime
}\right) >0$. Then property (\ref{ax1c3}) in }the nondegeneracy Condition{\ %
\ref{hyp1} follows from the continuity of $k$ with respect to the second
variable and therefore $\vec{k}$ satisfies hypothesis (\ref{appc01}) in
Theorem \ref{application}. Since (\ref{thw1}) holds, $\mathcal{A}$ satisfies
(\ref{xtra}). Since $\mathcal{A}$ is diagonal, (\ref{xxtra}) then follows
from Remark \ref{diagcase}. Thus $\mathcal{A}$ is super subordinate{, so
hypotheses (\ref{appc02}) and (\ref{appc04}) in Theorem \ref{application}
are satisfied. Since $f=0$, hypothesis (\ref{appc07}) of Theorem \ref%
{application} is trivially satisfied.
\endproof%
} }
{From Theorem \ref{striking} and the techniques used in \cite{SaW}, we can
derive an extension of Theorem 2.1 in \cite{SaW}, which is a regularity
result for convex solutions to Monge-Amp\`{e}re equations. Consider a
smooth, bounded, strongly convex domain $\Phi \subset \mathbb{R}^{2}$. Given
a convex function $u\in \mathcal{C}^{1}\left( \Phi \right) $, following \cite%
{SaW} we set
\begin{equation*}
\omega _{-}\left( s\right) =\omega _{-}\left( s,\Phi ,u\right)
=\inf_{t:\left( s,t\right) \in \Phi }u_{t}\left( s,t\right) \quad \text{and}%
\quad \omega _{+}\left( s\right) =\omega _{+}\left( s,\Phi ,u\right)
=\sup_{t:\left( s,t\right) \in \Phi }u_{t}\left( s,t\right)
\end{equation*}%
for any $s$ lying in the projection of $\Phi $ onto the $s$ -axis. Let
\begin{equation*}
\mathtt{I}_{s}=\left\{ \omega :\omega _{-}\left( s\right) <\omega <\omega
_{+}\left( s\right) \right\} \quad \text{if $\omega _{-}\left( s\right)
<\omega _{+}\left( s\right) $ and $\mathtt{I}_{s}=\emptyset $ otherwise, and}
\end{equation*}%
\begin{equation*}
\Phi _{u}=\left\{ \left( s,t\right) \in \Phi :u_{t}(s,t)\in \mathtt{I}%
_{s}\right\} .
\end{equation*}%
Note that}
\begin{itemize}
\item $u\left( s,t \right) $ is affine in the $t$ variable if and only if $%
\mathtt{I}_{s}=\emptyset $.
\item If $u\in \mathcal{C}^{1}\left( \overline{\Phi }\right) $ is not affine
in the $t$ variable for any fixed $s$, {then }$\Phi _{u}$ is an open
connected set. Indeed, to see that $\Phi _{u}$ is open, let $\left(
s,t\right) \in \Phi _{u}$. Then $\omega _{-}(s)<u_{t}\left( s,t\right)
<\omega _{+}(s)$, and since $u\in \mathcal{C}^{1}(\overline{\Phi })$, the
functions $\omega _{-},\omega _{+}$ and $u_{t}$ are continuous. Hence, for $%
(s^{\prime },t^{\prime })$ near $(s,t)$, $\omega _{-}(s^{\prime
})<u_{t}\left( s^{\prime },t^{\prime }\right) <\omega _{+}(s^{\prime })$.
Therefore $\Phi _{u}$ is open. To see that $\Phi _{u}$ is connected it
suffices to show it is pathwise connected. This follows from the fact that
the midpoint $(\omega _{-}(s)+\omega _{+}(s))/2$ of I$_{s}$ is a continuous
function of $s$. Note the arguments above only use the continuity of $u_{t}$%
. See \cite{SaW} for further discussion about when $\Phi _{u}$ is connected.
\item Even if $u$ is not affine in the $t$ variable for any fixed $s$, the
set
\begin{equation*}
\Pi _{u}=\left\{ \left( s,t\right) \in \Phi _{u}:u_{t}\left( s,t\right)
=\omega _{-}\left( s\right) ~\text{or}~u_{t}\left( s,t\right) =\omega
_{+}\left( s\right) \right\}
\end{equation*}%
may be non-empty. This exceptional set $\Pi _{u}$ is composed by what are
called \textquotedblleft Pogorelov segments\textquotedblright\ in \cite{SaW}.
\end{itemize}
\begin{theorem}
\label{RSaW2} Suppose $k\left( s,t\right) \,$is smooth, nonnegative,
satisfies $\left\vert \partial _{t}k\right\vert =O\left( k\right) $, and $%
k\left( \cdot ,0\right) $ \emph{does not vanish identically on any
horizontal line segment} in $\Phi $, where $\Phi $ is as above. If $u\in
\mathcal{C}^{1}\left( \overline{\Phi }\right) $ is a convex solution of the
Monge-Amp\`{e}re boundary value problem
\begin{eqnarray*}
\det \left(
\begin{array}{cc}
u_{ss} & u_{st} \\
u_{ts} & u_{tt}%
\end{array}
\right) &=&k\left( s,t\right) ,\;\;\;\;\;\left( s,t\right) \in \Phi , \\
u &=&\phi \left( s,t\right) ,\qquad \left( s,t\right) \in \partial \Phi ,
\end{eqnarray*}
where $\phi $ is smooth on $\partial \Phi $, then $u$ is smooth in $\Phi
_{u}.$
\end{theorem}
{To obtain our main results, we follow the approach in \cite{SaW} for
two-dimensional equations, although our objectives are more general. We
consider equations in any dimension at least two, our equations may include
a first order drift term and a zero order term, and our notion of solution
is more general since we only require continuity instead of Lipschitz
continuity. To prove our main hypoellipticity result, Theorem \ref%
{application}, we use an approximation argument based on the a priori
estimates in Theorem \ref{quasi} and on the construction in Lemma \ref%
{barrier} of new custom-built barriers.} One of the core ingredients needed
to derive Theorem \ref{quasi} is the interpolation inequality given in Lemma %
\ref{lemma-interpolation}, proved in \cite{RSaW1}.
\section{Sharpness\label{section-sharp}}
Our results are sharp in the sense that the power $1$ in the super
subordination Condition \ref{hyp2} cannot be decreased. Indeed, we will now
show that for any $\varepsilon >0,$ there exists a nonnegative smooth
function $k=k\left( x,y,z\right) $ in $\mathbb{R}^{2}\times \mathbb{R}$
which is not identically zero on any horizontal segment in $\mathbb{R}^{2}$
(moreover $k>0$ unless $x=z=0$) and satisfies
\begin{equation*}
\left\vert \partial _{z}k\left( x,y,z\right) \right\vert \leq C~\left[
k\left( x,y,z\right) \right] ^{1-\varepsilon }\qquad \text{in }\Omega \times
\mathbb{R},
\end{equation*}%
and that there is a continuous weak solution $w$ of {%
\begin{equation}
\partial _{x}^{2}w+\partial _{y}k\left( x,y,w\left( x,y\right) \right)
\partial _{y}w=0 \label{equation-x}
\end{equation}%
in any neighborhood $\Omega $ of the origin, \emph{but $w$ is not smooth} in
}$\Omega $. This example is derived by applying the partial Legendre
transform to non-smooth solutions of the Monge-Amp\'{e}re equation which are
suitable powers of the distance function to the origin.
Given $\varepsilon >0,$ let $m$ be a positive integer such that
\begin{equation}
\frac{1}{4m-2}<\varepsilon . \label{meps}
\end{equation}%
Consider the function $w=w\left( x,y\right) $ defined implicitly by the
equation
\begin{equation}
F(x,y,w)=0\quad \text{where $F(x,y,z)=z\left( |x|^{2}+|z|^{2}\right) ^{m-%
\frac{1}{2}}+y$.} \label{eq-implicit}
\end{equation}
\begin{center}
\includegraphics[width=3.5in]{sh-example-bw}
\end{center}
Since $F(0,0,0)=0$ and $F_{z}(x,y,z)=(|x|^{2}+|z|^{2})^{m-\frac{3}{2}%
}(x^{2}+2mz^{2})$, it follows from the implicit function theorem that $%
w=w(x,y)$ is well-defined by (\ref{eq-implicit}) and smooth in $\mathbb{R}%
^{2}\setminus \{(0,0)\}$. Also, $w$ extends continuously to $(x,y)=(0,0)$
with $w(0,0)=0$, and thus for any neighborhood $\Omega $ of the origin, $%
w\in \mathcal{C}^{0}\left( \overline{\Omega }\right) \bigcap \mathcal{C}%
^{\infty }\left( \Omega \backslash \left\{ \left( 0,0\right) \right\}
\right) $. Moreover, $w(x,y)=0$ if and only if $y=0$.
Now let
\begin{equation*}
f\left( x,y\right) =x\left( |x|^{2}+|w(x,y)|^{2}\right) ^{m-\frac{1}{2}%
}\qquad \left( x,y\right) \in \Omega .
\end{equation*}%
Then $f(0,y)=0$, $f(x,0)=x|x|^{2m-1}$ and $f(x,y)=-\frac{xy}{w(x,y)}$ if $%
w(x,y)\neq 0$. By direct computation, if $(x,y)\neq (0,0)$,
\begin{equation}
\left\{
\begin{array}{rcl}
f_{x}\left( x,y\right) & = & -2m\left( x^{2}+w^{2}\right) ^{2m-1}w_{y}\left(
x,y\right) \\
f_{y}\left( x,y\right) & = & w_{x}\left( x,y\right) .%
\end{array}%
\right. \label{eq-CR}
\end{equation}%
Thus, if $k\left( x,y,z\right) =k\left( x,z\right) $ is the smooth function
in $\mathbb{R}^{2}\times \mathbb{R}$ defined by
\begin{equation*}
k\left( x,z\right) =2m\left( x^{2}+z^{2}\right) ^{2m-1},
\end{equation*}%
then by using the formulas $F_{x}=(2m-1)xz(|x|^{2}+|z|^{2})^{m-\frac{3}{2}}$
and $F_{y}=1$, we have
\begin{eqnarray*}
\left\vert w_{x}\right\vert ^{2} &=&\left\vert -\frac{\left( 2m-1\right) xw}{%
x^{2}+2mw^{2}}\right\vert ^{2}\leq \left\vert \frac{\left( 2m-1\right) }{2}%
\frac{x^{2}+w^{2}}{x^{2}+2mw^{2}}\right\vert ^{2}\leq m^{2}, \\
k\left\vert w_{y}\right\vert ^{2} &=&\frac{2m\left( x^{2}+w^{2}\right)
^{2m-1}}{\left( x^{2}+w^{2}\right) ^{2m-3}\left( x^{2}+2mw^{2}\right) ^{2}}
\\
&=&2m\frac{\left( x^{2}+w^{2}\right) ^{2}}{\left( x^{2}+2mw^{2}\right) ^{2}}%
\leq 2m.
\end{eqnarray*}%
In particular, this implies that%
\begin{equation*}
\int_{\Omega }(\left\vert w_{x}\right\vert ^{2}+k\left\vert w_{y}\right\vert
^{2})~dxdy\leq 3m^{2}\left\vert \Omega \right\vert .
\end{equation*}%
that is, $w\in H_{\mathcal{X}}^{1,2}\left( \Omega \right) \bigcap L^{\infty
}\left( \Omega \right) $, where $\mathcal{X}\left( x,y,w,\xi _{1},\xi
_{2}\right) =\xi _{1}^{2}+k\xi _{2}^{2}$. From (\ref{eq-CR}) it follows that
$w\left( x,y\right) $ is a continuous weak solution of the quasilinear
equation (\ref{equation-x}). Moreover, as a function of $(x,y)$, $k\left(
x,y,z\right) $ does not vanish on any horizontal line segment, and it
satisfies
\begin{eqnarray}
\left\vert \partial _{z}k\left( x,z\right) \right\vert &=&4m\left(
2m-1\right) \left\vert z\right\vert \left( x^{2}+z^{2}\right) ^{2m-2} \notag
\\
&\leq &C_{m}\left( |x|^{2}+|z|^{2}\right) ^{2m-\frac{3}{2}}=C_{m}k\left(
x,z\right) ^{1-\frac{1}{4m-2}}\leq C_{m}k^{1-\varepsilon }, \label{xtra-k}
\end{eqnarray}%
where we used the inequality $\left\vert z\right\vert \leq \sqrt{x^{2}+z^{2}}
$ and the bound (\ref{meps}) for $m$.
On the other hand, from (\ref{eq-implicit}), noting that $y$ and $w(x,y)$
have the same sign, we have
\begin{equation*}
w\left( 0,y\right) =\left( \mathop{\rm sign}y\right) \left\vert y\right\vert
^{\frac{1}{2m}}.
\end{equation*}
Hence $w\left( x,y\right) \not\in \mathcal{C}^{\frac{1}{2m}+ \delta
}(\Omega) $ for any $\delta >0$. In particular, $w$ is not smooth in $\Omega
$. {%
\endproof%
}
\section{Preliminaries}
\subsection{{Notation\label{notation}}}
{\ Throughout the paper, $C$ will denote a constant that may change from
line to line but that is independent of any significant quantities. In
general, $C$ may depend on the dimension $n$, $\vec{k},$ and the fixed
cutoff functions defined below. We will use calligraphic $\mathcal{C}$ to
denote a function of one or more variables, increasing in each variable
separately, that may also change from line to line, but that is independent
of any significant parameters except its variables. }
{We denote by $\left\Vert f\right\Vert _{L^{\infty }\left( \Omega \right) }$
the essential supremum of $|f|$ in $\Omega $\ and by $\left\Vert
f\right\Vert _{L^{p}\left( \Omega \right) }$ the $L^{p}$-norm of $f\,$in $%
\Omega $:
\begin{equation*}
\left\Vert f\right\Vert _{L^{p}\left( \Omega \right) }=\left( \int_{\Omega
}\left\vert f\right\vert ^{p}dx\right) ^{\frac{1}{p}},\quad 1\le p <\infty.
\end{equation*}
When $\Omega =\mathbb{R}^{n}$ we omit mentioning the set. The collection of
real-valued functions in $\Omega $ with $m\,$continuous (but not necessarily
bounded) derivatives in $\Omega $ will be denoted $\mathcal{C}^{m}\left(
\Omega \right) $, and for $f\in \mathcal{C}^{m}\left( \Omega \right) $ we
let
\begin{equation}
\left\Vert f\right\Vert _{\mathcal{C}^{m}\left( \Omega \right)
}=\sum_{i=0}^{m}\sum_{\left\vert \vec{\alpha}\right\vert =i}\left\Vert
\partial ^{\vec{\alpha}}f\right\Vert _{L^{\infty }\left( \Omega \right) },
\label{cmnorm}
\end{equation}
where $\partial^{\vec{\alpha}}=\left(\partial _{1}\right) ^{\alpha
_{1}}\cdots \left( \partial _{n}\right) ^{\alpha _{n}}$} for $\vec{\alpha}%
=\left( \alpha _{1},\dots ,\alpha _{n}\right) $.
{Let $\Gamma _{M_{0}}^{\prime }\subset \Gamma $ be the domains in $\mathbb{R}%
^{n}\times \mathbb{R}$ given in (\ref{domain}) and let $k^{i}\left(
x,z\right) \in \mathcal{C}^{2}\left( \Gamma \right) $ be nonnegative, $%
i=1,\dots ,n$. From now on, let $\vec{k}(x,z)=\left( k^{1}(x,z),\dots
,k^{n}(x,z)\right) $ satisfy the nondegeneracy Condition~ \ref{hyp1} with
\begin{equation}
k^{1}\left( x,z\right) =1\qquad \text{for all }\left( x,z\right) \in \Gamma ,
\label{unifbdk1}
\end{equation}%
which clearly implies property (\ref{ax1c2}) in }Condition{\ \ref{hyp1}.
Since our theorems are local, this assumption causes no loss of generality. }
\subsection{Boxes around points}
{Given $x\in \Omega $, we will consider rectangular neighborhoods }$\mathcal{%
\ R}${\ of $x$ of the form described in Section \ref{statements}, with $x\in
\frac{1}{3}\mathcal{R}$. The maximum sidelength }$R${\ of }$\mathcal{R}${\
will be chosen so that $2\mathcal{R}\subset \Omega $ and possibly even
smaller to allow the absorption of various terms involving $R$ as a factor.
We will always assume that }$\mathcal{R}${\ satisfies property (\ref{ax1c3})
of }the nondegeneracy Condition{\ \ref{hyp1}. }Since{\ $\vec{k}$ is
continuous, given such $\mathcal{R}$ and $M_{0}>0,$ there exist positive
numbers $\delta ^{i}=\delta ^{i}\left( \vec{k},\mathcal{R},M_{0}\right) $
such that $\delta ^{i}<\frac{1}{2}r^{i}$, where $r^{i}$ denotes the $i^{th}$
half-sidelength of $\mathcal{R}$, $i=1,\dots ,n$, and }
\begin{eqnarray}
k^{i}\left( y,z\right) &>&0\qquad \text{whenever }z\in \left[ -M_{0},M_{0}%
\right] \text{ and} \label{lidrfat} \\
y &\in &\mathcal{T}_{i}\left( \mathcal{R},\delta ^{i}\right) =\left\{ Y\in
\mathbb{R}^{n}:\mathop{\rm dist}\left( Y,\mathcal{T}_{i}\left( \mathcal{R}%
\right) \right) \leq \delta ^{i}\right\} . \notag
\end{eqnarray}
\begin{remark}
\label{noM0} Under the hypotheses of Theorem \ref{application}, or more
precisely, if $\mathcal{A}$ satisfies (\ref{xtra}) and (\ref{hellip}), the
parameters $\delta ^{i}$ above can be taken independent of $M_{0}$. Indeed,
( \ref{xtra}) and Lemma \ref{admisall} imply that for fixed $x,\xi \in
\mathbb{R} ^{n}$, the function $\xi ^{t}\mathcal{A}\left( x,z \right) \xi $
is either identically zero in $z$ or strictly positive in $z$. Therefore, (%
\ref{hellip}) implies a similar property for each $k^{i}\left( x,z\right) $.
\end{remark}
\subsection{A class of adapted cutoff functions}
A \emph{cutoff function} is any nonnegative smooth function with compact
support, i.e., $\varphi \,$is a cutoff function if $\varphi \in \mathcal{C}
_{0}^{\infty }\left( \mathbb{R}^{n}\right) $ and $\varphi \geq 0$.
We now define a special class of cutoff functions around $x$ which are
adapted to our operator as in \cite{SaW2} (see also \cite{Fe}). The main
property of these functions is that they are supported in a (small enough)
neighborhood of $x$, while their derivatives are supported \emph{away} from $%
x$, essentially where $\vec{k}$ has positive components.
\begin{definition}[Supporting relation]
\label{rhost}Given two cutoff functions $\zeta $, $\xi \in \mathcal{C}
_{0}^{\infty }\left( \Omega \right) $, we say that $\zeta $ \emph{supports} $%
\xi $ and denote it $\xi \succeq \zeta $ or $\zeta \preceq \xi $ if $\xi =1 $
on a neighborhood of $\mathop{\rm support}\left( \zeta \right) $. Note in
particular that if $\xi \succeq \zeta $ then $\zeta \,\xi =\zeta $ and $%
\left\Vert \zeta \right\Vert _{L^{\infty }}\xi \geq \zeta $.
\end{definition}
\begin{definition}[Special cutoff functions]
\label{rhos}Let $x\in \Omega $, $M_{0}\geq 1$, $\mathcal{R}=\tilde{x}+ \big(%
\left[-r^{1},r^{1}\right] \times ...\times \left[ -r^{n},r^{n}\right]\big) $
be a rectangular box with $x\in \frac{1}{3}\mathcal{R}$, and $\delta
^{i}=\delta^{i}( \vec{k},\mathcal{R},M_{0}) >0$ be as in (\ref{lidrfat}).
Let
\begin{equation*}
\eta _{i},\phi _{i},\zeta _{i},\theta _{i}\in \mathcal{C}_{0}^{\infty
}\left( \tilde{x}_{i}+\left( -2r^{i},2r^{i}\right) \right) ,\qquad 1\leq
i\leq n,
\end{equation*}
be functions of one variable which satisfy the following:
\begin{enumerate}
\item \label{one}$0\leq \eta _{i},$ $\phi _{i},~\zeta _{i},~\theta _{i}\leq
1 $
\item \label{intervals}$\eta _{i}$, $\phi _{i}$ and$~\zeta _{i}$ are equal
to $1$ in $\tilde{x}_{i}+\left[ -\left( r^{i}-\delta ^{i}\right)
,r^{i}-\delta ^{i}\right] $ and vanish \newline
outside $\tilde{x}_{i}+\left[ -\left( r^{i}+\delta ^{i}\right) ,r^{i}+\delta
^{i}\right] $; and for all integers $m\geq 1$ and some universal constants $%
c_{m}\geq 1$,
\begin{equation}
\left( \frac{1}{c_{m}\delta ^{i}}\right) ^{m}\leq \left\Vert \frac{d^{m}}{
dt^{m}}\eta _{i}\right\Vert _{L^{\infty }\left( \mathbb{R}\right)
}+\left\Vert \frac{d^{m}}{dt^{m}}\phi _{i}\right\Vert _{L^{\infty }\left(
\mathbb{R}\right) }+\left\Vert \frac{d^{m}}{dt^{m}}\zeta _{i}\right\Vert
_{L^{\infty }\left( \mathbb{R}\right) }\leq \left( \frac{c_{m}}{\delta ^{i}}
\right) ^{m}. \label{intervalsconstants}
\end{equation}
\item $\eta _{i}\preceq \phi _{i}\preceq \zeta _{i}$ and $\eta _{i}\leq \phi
_{i}\leq \zeta _{i}$
\item \label{includd}Let $J^{i}\,$be the smallest set of the form $J^{i}=
\tilde{x}_{i}+ \left(\left[ -b^{i},-a^{i}\right] \bigcup \left[ a^{i},b^{i}%
\right] \right) \subset \mathbb{R}$ ,with $0<a^{i}<b^{i}$, such that
\begin{equation*}
\mathop{\rm support}\left( \eta _{i}^{\prime }\right) \bigcup \mathop{\rm
support}\left( \phi _{i}^{\prime }\right) \bigcup \mathop{\rm support}\left(
\zeta _{i}^{\prime }\right) \subset J^{i}.
\end{equation*}
Note by (\ref{intervals}) that $J^{i}\subset \tilde{x}_{i}+ \left(\left[
-r^{i}-\delta ^{i},-r^{i}+\delta ^{i}\right] \bigcup \left[
r^{i}-\delta^{i},r^{i}+\delta ^{i}\right]\right)$. Let $\theta _{i}\equiv 1$
on $J^{i}$, so that $\theta _{i}\equiv 1$ on the supports of $\eta
_{i}^{\prime }$, $\phi _{i}^{\prime }$ and $\zeta _{i}^{\prime }$, i.e., $%
\eta _{i}^{\prime }\preceq \theta _{i},\,\phi _{i}^{\prime }\preceq \theta
_{i},$ and $\zeta _{i}^{\prime }\preceq \theta _{i}$.
\item {\ \label{five}$\mathop{\rm support}\left( \theta _{i}\right) \subset
\tilde{x}_{i}+\left(\left[ -r^{i}-2\delta ^{i},-r^{i}+2\delta ^{i}\right]
\bigcup \left[ r^{i}-2\delta ^{i},r^{i}+2\delta ^{i}\right] \right)$; in
particular, $\tilde{x}_{i}\notin \mathop{\rm
support}\left( \theta _{i}\right) $. Moreover, we assume that }for all
integers $m\geq 1${\ and for some universal constants $C_{m}>0,$} {\
\begin{equation}
\left( \frac{1}{C_{m}\delta ^{i}}\right) ^{m}\leq \left\Vert \frac{d^{m}}{%
dt^{m}}\theta _{i}\right\Vert _{L^{\infty }\left( \mathbb{R}\right) }\leq
\left( \frac{C_{m}}{\delta ^{i}}\right) ^{m}. \label{fiveconstants}
\end{equation}
}
\item \label{six}$\left\vert \eta _{i}^{\prime }\right\vert $, $\left\vert
\phi _{i}^{\prime }\right\vert $ and $\left\vert \zeta _{i}^{\prime
}\right\vert $ are smooth functions (see the discussion following the figure
below). \newline
\end{enumerate}
\end{definition}
Finally, let
\begin{equation*}
\begin{array}{ll}
\eta \left( x\right) =\prod_{i=1}^{n}\eta _{i}\left( x_{i}\right) ,\qquad &
\phi \left( x\right) =\prod_{i=1}^{n}\phi _{i}\left( x_{i}\right) , \\
\zeta \left( x\right) =\prod_{i=1}^{n}\zeta _{i}\left( x_{i}\right) , &
\varrho _{i}\left( x\right) =\theta _{i}\left( x_{i}\right) \prod_{j\neq
i}\zeta _{j}\left( x_{j}\right) ,%
\end{array}%
\end{equation*}%
and let $\xi ,\chi \in \mathcal{C}_{0}^{\infty }\left( 2\mathcal{R}\right) $
satisfy
\begin{eqnarray*}
\xi &=&\chi =1\quad \text{in }\mathcal{R} \\
\zeta &\preceq &\xi \preceq \chi ,~\quad \zeta \leq \xi \leq \chi ,\quad
\text{and} \\
\varrho _{i} &\preceq &\xi ,\quad i=1,\dots ,n.
\end{eqnarray*}
The figure below will serve as a reminder of the order between some of the
cutoff functions:
\begin{center}
\includegraphics[width=4.04in]{phies}
\end{center}
Property (\ref{six}) above is easily satisfied by assuming (in addition to
properties (\ref{one}) to (\ref{includd})) that $\eta _{i}$, $\phi _{i}$ and
$\zeta _{i}$ are smooth, non-decreasing in $\left( -\infty ,\tilde{x}%
_{i}\right) $ and non-increasing in $\left( \tilde{x}_{i},\infty \right) $.
Indeed, under such conditions and since these functions are constant on a
neighborhood of $\tilde{x}_{i}$, their derivatives are of the form $\eta
_{i}^{\prime }=\left( \eta _{i}^{\prime }\right) ^{+}-\left( \eta
_{i}^{\prime }\right) ^{-}$, $\phi _{i}^{\prime }=\left( \phi _{i}^{\prime
}\right) ^{+}-\left( \phi _{i}^{\prime }\right) ^{-}$ and $\zeta
_{i}^{\prime }=\left( \zeta _{i}^{\prime }\right) ^{+}-\left( \zeta
_{i}^{\prime }\right) ^{-}$ where all these functions are compactly
supported, $\left( \eta _{i}^{\prime }\right) ^{+}$, $\left( \phi
_{i}^{\prime }\right) ^{+}$ and $\left( \zeta _{i}^{\prime }\right) ^{+}$
are smooth, supported in $\left( -\infty ,\tilde{x}_{i}\right) $ and
nonnegative, while $\left( \eta _{i}^{\prime }\right) ^{-}$, $\left( \phi
_{i}^{\prime }\right) ^{-}$ and $\left( \zeta _{i}^{\prime }\right) ^{-}$
are smooth, supported in $\left( \tilde{x}_{i},\infty \right) $ and
nonnegative. It follows that $\left\vert \eta _{i}^{\prime }\right\vert
=\left( \eta _{i}^{\prime }\right) ^{+}+\left( \eta _{i}^{\prime }\right)
^{-}$, $\left\vert \phi _{i}^{\prime }\right\vert =\left( \phi _{i}^{\prime
}\right) ^{+}+\left( \phi _{i}^{\prime }\right) ^{-}$ and $\left\vert \zeta
_{i}^{\prime }\right\vert =\left( \zeta _{i}^{\prime }\right) ^{+}+\left(
\zeta _{i}^{\prime }\right) ^{-}$ are smooth functions.
It will be convenient to set
\begin{eqnarray}
\qquad A^{6} &=&1+\left\Vert \nabla \eta \right\Vert _{L^{\infty
}}^{6}+\left\Vert \nabla \phi \right\Vert _{L^{\infty }}^{6}+\left\Vert
\nabla \zeta \right\Vert _{L^{\infty }}^{6}+\left\Vert \nabla \varrho
_{1}\right\Vert _{L^{\infty }}^{6}+\dots +\left\Vert \nabla \varrho
_{n}\right\Vert _{L^{\infty }}^{6} \label{defA} \\
&&+\left\Vert \nabla ^{2}\eta \right\Vert _{L^{\infty }}^{3}+\left\Vert
\nabla ^{2}\phi \right\Vert _{L^{\infty }}^{3}+\left\Vert \nabla ^{2}\zeta
\right\Vert _{L^{\infty }}^{3} \notag \\
&&+\left\Vert \nabla ^{3}\eta \right\Vert _{L^{\infty }}^{2}+\left\Vert
\nabla ^{3}\phi \right\Vert _{L^{\infty }}^{2}+\left\Vert \nabla ^{3}\zeta
\right\Vert _{L^{\infty }}^{2} \notag
\end{eqnarray}
in order to collect constants in front of the lower order terms in what
follows.
\begin{remark}
\label{AdependsonR}Note that $A$ depends only on $\vec{k}$, $\mathcal{R}$
and $M_{0}$. Let $\delta ^{\ast }=\delta ^{\ast }\left( \vec{k}, \mathcal{R}%
,M_{0}\right) =\min_{i=1,\dots ,n}\delta ^{i}$ where $\delta ^{i}=\delta
^{i}\left( \vec{k},\mathcal{R},M_{0}\right) $ are as in (\ref{lidrfat}).
Then from (\ref{defA}),(\ref{intervalsconstants}) and (\ref{fiveconstants}),
\begin{equation*}
A\approx \left( \delta ^{\ast }\right) ^{-1}.
\end{equation*}
\end{remark}
The main property of the cutoff functions $\eta $, $\phi $ and $\zeta $
above is that their $i^{th}$ partial derivative, $i=1,\dots ,n$, is
supported in the set $\mathcal{K}_{i}=\bigcap_{\ell \neq i,~\left\vert
z\right\vert \leq M_{0}}\left\{ x:k^{\ell }\left( x,z\right) >0\right\} $.
Indeed, let $\sigma \left( x\right) =\prod_{\ell =1}^{n}\sigma _{\ell
}\left( x_{\ell }\right) $ with $\eta _{\ell }\leq \sigma _{\ell }\leq \zeta
_{\ell }$, and fix $z\in \left[ -M_{0},M_{0}\right] $. It follows from
Definition \ref{rhos} (\ref{includd}) and (\ref{five}) that $\sigma _{\ell
}^{\prime }\preceq \theta _{\ell }$, i.e., $\mathop{\rm
supp}\left( \sigma _{\ell }^{\prime }\right) \subset \left\{ \theta _{\ell
}=1\right\} $. Hence,
\begin{equation*}
\mathop{\rm supp}\left( \sigma _{i}^{\prime }\right) \subset \mathop{\rm
supp}\left( \theta _{i}\right) \subset \tilde{x}_{i}+\left( \left[
-r^{i}-2\delta ^{i},-r^{i}+2\delta ^{i}\right] \bigcup \left[ r^{i}-2\delta
^{i},r^{i}+2\delta ^{i}\right] \right) .
\end{equation*}%
Hence, since $\mathop{\rm supp}(\sigma _{\ell })\subset \tilde{x}_{\ell }+%
\left[ -r^{\ell }-\delta ^{\ell },r^{\ell }+\delta ^{\ell }\right] $, we
have $\mathop{\rm supp}\left( \partial _{i}\sigma \right) \subset %
\mathop{\rm supp}\left( \varrho _{i}\right) $. Now set $\mathrm{I}_{\ell }=%
\left[ -r^{\ell }-\delta ^{\ell },r^{\ell }+\delta ^{\ell }\right] $ and $%
\mathcal{R}_{i}=\mathcal{R}_{i}^{+}\bigcup \mathcal{R}_{i}^{-}$, where
\begin{equation}
\begin{array}{l}
\mathcal{R}_{i}^{+}=\tilde{x}+\left\{ \left( \prod_{\ell <i}\mathrm{I}_{\ell
}\right) \times \left[ r^{i}-2\delta ^{i},r^{i}+2\delta ^{i}\right] \times
\left( \prod_{\ell >i}\mathrm{I}_{\ell }\right) \right\} , \\
\mathcal{R}_{i}^{-}=\tilde{x}+\left\{ \left( \prod_{\ell <i}\mathrm{I}_{\ell
}\right) \times \left[ -(r^{i}+2\delta ^{i}),-(r^{i}-2\delta ^{i})\right]
\times \left( \prod_{\ell >i}\mathrm{I}_{\ell }\right) \right\} .%
\end{array}
\label{rss}
\end{equation}%
From (\ref{lidrfat}) it follows that
\begin{equation*}
\mathop{\rm supp}\left( \partial _{i}\sigma \right) \subset \mathop{\rm supp}%
\left( \varrho _{i}\right) \subset \mathcal{R}_{i}\subset \mathcal{K}_{i},
\end{equation*}%
as wanted. The figure below represents the sets $\mathcal{R}_{i}$ when $n=3$.
\begin{center}
\includegraphics[width=5.56in]{rhoes}
\end{center}
Since $\mathop{\rm supp}\left( \varrho _{i}\right) $ is compact and $%
\mathcal{K}_{i}$ is open, it follows from (\ref{chaak}) that there exists $%
\tilde{C}_{1}=\tilde{C}_{1}\left( \vec{k},\mathcal{R},M_0\right) >0$ such
that
\begin{equation*}
\varrho _{i}\left( x\right) \min_{\ell \neq i} k^{\ell }(x,z)\geq \frac{1 }{%
\tilde{C}_{1}}\varrho _{i}\left( x\right),\qquad i=1,\dots ,n,\quad \hspace{%
-0.0943pc}x\in \Omega \quad\hspace{-0.0943pc} z\in \left[ -M_{0},M_{0}\right]%
.
\end{equation*}
Since $\vec{k}$ is bounded in any compact set, it follows from Condition \ref%
{hyp1} that there exists $C_{1}=C_{1}\left( \vec{k},\mathcal{R},M_{0}\right)
$ such that for $1\leq i\leq n,$
\begin{equation}
\varrho _{i}\left( x\right) \,k^{i}(x,z)\,\leq C_{1}\,\varrho _{i}\left(
x\right) \,\min_{1\leq j\leq n}k^{j}(x,z),\quad \hspace{-0.0939pc}\left(
x,z\right) \in \Omega \times \left[ -M_{0},M_{0}\right] . \label{mink}
\end{equation}
We will often want to show that a certain term is small by applying the
one-dimensional Sobolev inequality in the $x_{1}$-variable, i.e., by
applying the estimate
\begin{equation}
\left\Vert \varphi \right\Vert _{L^{2}\left( \mathbb{R}^{n}\right) }\,\leq
Cr^{1}\left\Vert \partial _{1}\varphi \right\Vert _{L^{2}\left( \mathbb{R}%
^{n}\right) }\,, \label{spoinc0}
\end{equation}%
where $\varphi $ is a function with compact support in $2\mathcal{R}$ and $C$
is a universal constant. Then by (\ref{hellip}), (\ref{unifbdk1}), and the
definition in (\ref{gradk}) below, we have
\begin{equation}
\left\Vert \varphi \right\Vert _{L^{2}\left( \mathbb{R}^{n}\right) }\,\leq
Cr^{1}\left\Vert \nabla _{\!\!\mathcal{A},w}\,\varphi \right\Vert
_{L^{2}\left( \mathbb{R}^{n}\right) }\qquad\text{if support$(\varphi)
\subset 2\mathcal{R}.$} \label{spoinc}
\end{equation}%
The constant factor $r^{1}$ which appears here will often be chosen small to
help in absorption arguments, but it is important to observe that since $%
A\geq \left( \delta ^{1}\right) ^{-1}\geq 3\left( r^{1}\right) ^{-1}$, we
must ensure that a term to be shown small because it contains an $r^{1}$
factor \emph{does not also include a factor of positive powers of }$A$.
For simplicity, we will often restrict our calculations to the case when the
center $\tilde{x}$ of $\mathcal{R}$ is the origin.
\subsection{{Auxiliary Results\label{prelressection}}}
Given a weak solution $w\in H_{\mathcal{X}}^{1,2}\left(\Omega\right) \bigcap
L_\infty \left(\Omega \right) $ to (\ref{equation}) , we denote $\mathbf{A}%
\left( x\right) =\mathcal{A}\left( x,w\left( x\right) \right) $. It is
convenient to define the \emph{linear} operator
\begin{equation}
\mathcal{L}_{w}=\mathop{\rm div}\mathbf{A}\left( x\right) \nabla = %
\mathop{\rm div}\mathcal{A}\left( x,w\left( x\right) \right) \nabla
,\;\;\;\;\;x\in \Omega . \label{operatorw}
\end{equation}
Given $\varphi \in H_{\mathcal{X}}^{1,2}\left( \Omega \right) $, we{\ denote
by }$\nabla _{\!\!\sqrt{\mathcal{A}},w}\varphi \,${\ the }$\sqrt{\mathcal{A}}
${-gradient of }$\varphi, $ formally defined by
\begin{equation}
\nabla _{\!\!\sqrt{\mathcal{A}},w}\varphi =\sqrt{\mathcal{A}\left( x,w\left(
x\right) \right) }\nabla {\varphi .} \label{gradk}
\end{equation}
See the Appendix for a discussion of the meaning of $\nabla\varphi$ in case $%
\varphi$ is not smooth.
{We now list four useful lemmas obtained in \cite{RSaW1}.}
\begin{lemma}
{\ \label{Moser}Let $u\in \mathcal{C}^{\infty }\left( 2\mathcal{R}\right) $,
$\psi $ be a nonnegative cutoff function supported in $2\mathcal{R},$ and $%
\beta \in \mathbb{N}$. Then
\begin{align*}
\int_{2\mathcal{R}}\left\vert \psi \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}\,u^{\beta }\right\vert ^{2}& \leq \frac{2\beta ^{2}}{2\beta -1}%
\left\vert \int_{2\mathcal{R}}\left( \psi \mathcal{L}_{w}u\right) \left(
\psi u^{2\beta -1}\right) \right\vert \\
& +\left( \frac{4\beta }{2\beta -1}\right) ^{2}\int_{2\mathcal{R}}\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}\,\psi \right\vert ^{2}\left\vert u^{\beta
}\right\vert ^{2},
\end{align*}%
where $\mathcal{L}_{w}$ is the linear operator (\ref{operatorw}), and $%
\nabla _{\!\!\sqrt{\mathcal{A}},w}$ is given by (\ref{gradk}). }
\end{lemma}
\begin{lemma}
\label{minkgrad} Let ${\vec{k}}$ satisfy Condition \ref{hyp1} and $\mathcal{R%
}$ be a box satisfying property (\ref{ax1c3}) in Condition \ref{hyp1}. For
any smooth function $\varphi $ and smooth cutoff function $\psi $ of the
form $\psi =\prod_{i=1}^{n}\psi _{i}\left( x_{i}\right) $, where $\eta
_{i}\leq \psi _{i}\leq \zeta _{i}$ for all $i$ (see Definition (\ref{rhos}%
)), we have
\begin{equation*}
\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\psi \right\vert ^{2}\left\vert
\nabla \varphi \right\vert ^{2}\leq C_{1}\Lambda \left\vert \nabla \psi
\right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\varphi
\right\vert ^{2},
\end{equation*}
with $\Lambda,$ $C_{1}=C_{1}\left( \vec{k},\mathcal{R},M_{0}\right) $ and $%
\nabla _{\!\!\sqrt{\mathcal{A}},w}$ as in (\ref{hellip}), (\ref{mink}) and (%
\ref{gradk}).
\end{lemma}
\begin{lemma}
\label{lemma-interpolation} Suppose that $\zeta $, $\chi $ are cutoff
functions as in Definition \ref{rhos}, and $\mathcal{R}$ is a rectangle with
$2\mathcal{R} \subset \Omega $ which satisfies property (\ref{ax1c2}) of
Condition \ref{hyp1}. Then for each $q>n$, there exists $1<p<2$ such that
for all $u\in \mathcal{C}_{0}^{1}\left( 2\mathcal{R}\right) $ satisfying $%
u\preceq \chi $ (see Definition \ref{rhost}), all $\beta \in \mathbb{N}$,
and all $0<\epsilon \le 1$,
\begin{equation*}
\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\, \zeta
\right\vert ^{2}\left\vert u^{\beta }\right\vert ^{2}\leq \varepsilon ^{-1}%
\mathcal{C} \left( n,A,q,\mathcal{K},\mathcal{R}, C_{1},\Lambda \right)
\left\Vert u^{\beta }\right\Vert _{L^{p}}^{2}+\varepsilon \int_{2\mathcal{R}%
}\left\vert \zeta \nabla _{\!\!\sqrt{\mathcal{A}},w}\,u^{\beta }\right\vert
^{2}\,,
\end{equation*}
where $\mathcal{K}=\left\Vert \nabla \chi \mathbf{A}\right\Vert _{L^{q}}$,
and $C_1$ is as in Lemma \ref{minkgrad}.
\end{lemma}
\begin{lemma}
{\ \label{secondcacc}} Suppose that $w$ is a smooth solution of (\ref%
{equation}) in $2\mathcal{R}\subset \Omega ^{\prime }$. Then for $\beta \in
\mathbb{N}$ and any $0<\alpha \leq 1$,
\begin{equation*}
\sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}\left( \mathcal{L}
_{w}w_{ij}\right) w_{ij}^{2\beta -1}\right\vert
\end{equation*}
\begin{eqnarray*}
&\leq &2\sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\left( \nabla
w_{j}\right) \cdot \mathbf{A}_{i}\nabla \zeta ^{2}w_{ij}^{2\beta
-1}\right\vert +2\sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}w_{i}\left(
\nabla w_{j}\right) \cdot \mathbf{A}_{z}\nabla \zeta ^{2}w_{ij}^{2\beta
-1}\right\vert \\
&&+\sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}w_{ij}\left( \nabla
w\right) \cdot \mathbf{A}_{z}\nabla \zeta ^{2}w_{ij}^{2\beta -1}\right\vert
+ \frac{\mathcal{C}_{0}}{\alpha }\int_{2\mathcal{R}}\zeta ^{2}\left\vert
\nabla ^{2}w\right\vert ^{2\beta } \\
&&+\alpha \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \vec{\gamma%
}\cdot \nabla w_{ij}^{\beta }\right\vert ^{2}+\alpha \int_{2\mathcal{R}%
}\zeta ^{2}\left\vert \vec{\gamma}_{z}\cdot \nabla w\right\vert
^{2}\left\vert \nabla ^{2}w\right\vert ^{2\beta } \\
&&+2\sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}\left( \left(
\partial _{j}\vec{\gamma}\right) \cdot \nabla w_{i}\right) w_{ij}^{2\beta
-1}\right\vert +2\sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}w_{i}\left(
\nabla w\right) \cdot \mathbf{A}_{jz}\nabla \zeta ^{2}w_{ij}^{2\beta
-1}\right\vert \\
&&+\sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}w_{i}w_{j}\left( \nabla
w\right) \cdot \mathbf{A}_{zz}\nabla \zeta ^{2}w_{ij}^{2\beta -1}\right\vert
+\mathcal{C}_{0}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla w\right\vert
^{6\beta }+\mathcal{C}_{0},
\end{eqnarray*}
{where $\vec{\gamma}=\vec{\gamma}\left( x,w\right) $, $f=f\left( x,w\right)$%
, ${\mathbf{A}}=\mathcal{A}\left( x,w\right) $, $\mathbf{A}_{i}=\mathcal{A}%
_{i}\left( x,w\right) $, etc., and
\begin{equation*}
\mathcal{C}_{0}=C\left\{ \left\Vert \mathcal{A}\right\Vert _{\mathcal{C}%
^{3}\left( \tilde{\Gamma}\right) }+\left\Vert f\right\Vert _{\mathcal{C}
^{2}\left( \tilde{\Gamma}\right) }+\left\Vert \vec{\gamma}\right\Vert _{%
\mathcal{C}^{2}\left( \tilde{\Gamma}\right) }+1\right\} .
\end{equation*}
}
\end{lemma}
\begin{remark}
Lemma \ref{secondcacc} is a slightly different version of Lemma 5.6 in \cite%
{RSaW1} and readily follows from its proof. Indeed the fourth term on the
right side of the conclusion of the original lemma is replaced, via
straightforward changes in the proof, by the sixth and seventh terms on the
right side above.
\end{remark}
The following result is used in the proof of Theorem \ref{DP}.
\begin{lemma}
\label{concave-majorant}Given $u$ continuous on $\left[ 0,1\right] $, there
exists $w\in \mathcal{C}^0\left( \left[ 0,1\right] \right) \bigcap \mathcal{C%
}^{2}\left( \left( 0,1\right] \right) $ such that $w$ is concave, strictly
increasing, $w\left( x\right) \geq u\left( x\right) $ for all $x\in \left[
0,1\right] $, and $w\left( 0\right) =u\left( 0\right) $.
\end{lemma}
\begin{proof}
Let $\tilde{u}\left( x\right) =\max_{t\in \left[ 0,x\right] } u\left(
t\right) $ for $x\in \left[ 0,1\right] $, and $\tilde{u}\left( x\right)
=\max_{t\in \left[ 0,1\right] }u\left( t\right) $ for $x>1$. Since $%
\max_{t\in \left[ 0,x\right] }u\left( t\right) $ is nondecreasing, $\tilde{u}%
\left( x\right) $ is nondecreasing in $\left[ 0,\infty \right) $, $\tilde{u}%
\left( x\right) \geq u\left( x\right) $ in $\left[ 0,1\right] $, and $\tilde{%
u}\left( 0\right) =u\left( 0\right) $.
Next, for a smooth nonnegative function $\eta $ with compact support in $%
\left[ -1,1\right] $ and $\int \eta =1$, set $\eta _{\varepsilon }\left(
x\right) =\frac{1}{\varepsilon }\eta \left( \frac{x}{\varepsilon }\right) $
and let $v\left( x\right) =\tilde{u}\ast \eta _{\frac{x}{2}}\left( 2x\right)
$ for $x>0$, and $v\left( 0\right) =u\left( 0\right) $. Then $v\left(
x\right) \geq \tilde{u}\left( x\right) $ for all $x,$ and $v\in \mathcal{C}%
^0\left( \left[ 0,\infty \right) \right) \bigcap \mathcal{C}^{\infty }\left(
\left( 0,\infty \right) \right) $.
Taking now $\tilde{v}\left( x\right) =\max_{t\in \left[ 0,x\right] }v\left(
t\right) +x$, we have that $\tilde{v}\in \mathcal{C}^0\left( \left[ 0,\infty
\right) \right) \bigcap \mathcal{C}^{\infty }\left( \left( 0,\infty \right)
\right) $, $\tilde{v}\left( x\right) \geq v\left( x\right) \geq \tilde{u}%
\left( x\right) $ for all $x$, $\tilde{v}\left( 0\right) =u\left( 0\right) $%
, and $\tilde{v}$ is strictly increasing.
By the fundamental theorem of calculus,%
\begin{equation*}
\tilde{v}\left( x\right) =\tilde{v}\left( 1\right) -\int_{x}^{1}\tilde{v}%
^{\prime }\left( t\right) ~dt=\tilde{v}\left( 1\right) -\left( 1-x\right)
\tilde{v}^{\prime }\left( 1\right) +\int_{x}^{1}\int_{t}^{1}\tilde{v}%
^{\prime \prime }\left( s\right) ~ds~dt
\end{equation*}%
for all $x\in \left( 0,\infty \right) $. Let $\left[ \tilde{v}^{\prime
\prime }\left( s\right) \right] ^{+}=\max \left\{ \tilde{v}^{\prime \prime
}\left( s\right) ,0\right\} $ and $\left[ \tilde{v}^{\prime \prime }\left(
s\right) \right] ^{-}=\max \left\{ -\tilde{v}^{\prime \prime }\left(
s\right) ,0\right\} $. Then $\left[ \tilde{v}^{\prime \prime }\left(
s\right) \right] ^{\pm }$ are continuous in $\left( 0,\infty \right) $ and $%
\tilde{v}^{\prime \prime }\left( s\right) =\left[ \tilde{v}^{\prime \prime
}\left( s\right) \right] ^{+}-\left[ \tilde{v}^{\prime \prime }\left(
s\right) \right] ^{-}$. Since%
\begin{equation*}
\max_{x\in \left[ 0,1\right] }\left\vert \int_{x}^{1}\int_{t}^{1}\tilde{v}%
^{\prime \prime }\left( s\right) ~ds~dt\right\vert =\max_{x\in \left[ 0,1%
\right] }\left\vert \tilde{v}\left( x\right) -\tilde{v}\left( 1\right)
+\left( 1-x\right) \tilde{v}^{\prime }\left( 1\right) \right\vert <\infty ,
\end{equation*}%
it follows that the functions $w_{+}$ and $w_{-}$ defined by
\begin{equation*}
w_{+}\left( x\right) =\int_{x}^{1}\int_{t}^{1}\left[ \tilde{v}^{\prime
\prime }\left( s\right) \right] ^{+}~ds~dt\qquad \text{and}\qquad
w_{-}\left( x\right) =\int_{x}^{1}\int_{t}^{1}\left[ \tilde{v}^{\prime
\prime }\left( s\right) \right] ^{-}~ds~dt
\end{equation*}%
are finite and belong to $\mathcal{C}^{2}\left( 0,1\right] $. Also note that
$\tilde{v}\left( x\right) =\tilde{v}\left( 1\right) -\left( 1-x\right)
\tilde{v}^{\prime }\left( 1\right) +w_{+}\left( x\right) -w_{-}\left(
x\right) $ . Moreover, since%
\begin{equation*}
\left( w_{\pm }\left( x\right) \right) ^{\prime \prime }=\left[ \tilde{v}%
^{\prime \prime }\left( x\right) \right] ^{\pm }\geq 0,
\end{equation*}%
it follows that $w_{\pm }$ are convex in $\left[ 0,1\right] $. In
particular,
\begin{equation*}
w_{+}\left( x\right) \leq \left( 1-x\right) ~w_{+}\left( 0\right)
+x~w_{+}\left( 1\right) ,\qquad x\in \left[ 0,1\right] .
\end{equation*}%
We claim that the function
\begin{equation*}
w\left( x\right) =\tilde{v}\left( 1\right) -\left( 1-x\right) \tilde{v}%
^{\prime }\left( 1\right) +\left( 1-x\right) ~w_{+}\left( 0\right)
+x~w_{+}\left( 1\right) -w_{-}\left( x\right)
\end{equation*}%
satisfies all the properties stated in the lemma. Indeed, it is clear that $%
w $ is continuous in $\left[ 0,1\right] $, $\mathcal{C}^{2}$ in $\left( 0,1%
\right] $, and $w\left( 0\right) =u\left( 0\right)$ since $w\left( 0\right)
= \tilde{v}(1) -\tilde{v}(0) + w_{+}(0) - w_{-}(0) = \tilde{v}(0) = u(0)$.
From the last inequality,%
\begin{equation*}
w\left( x\right) \geq \tilde{v}\left( 1\right) -\left( 1-x\right) \tilde{v}%
^{\prime }\left( 1\right) +w_{+}\left( x\right) -w_{-}\left( x\right) =%
\tilde{v}\left( x\right) \geq \tilde{u}\left( x\right) \geq u\left( x\right)
.
\end{equation*}%
Finally, since%
\begin{equation*}
w^{\prime \prime }\left( x\right) =-\left( w_{-}\left( x\right) \right)
^{\prime \prime }\leq 0
\end{equation*}%
we have that $w$ is concave, as required.
\end{proof}
\section{Proof of the {a priori estimates\label{hypo}}}
\subsection{{$L^{p}$ estimates for the gradient\label{apriori}}}
{In this section we prove higher integrability properties of $\nabla w$
(Theorem \ref{extraintegrability}) by using the extra hypotheses in
Condition \ref{hyp2}.}
\begin{lemma}
{\label{extraprev}Under the hypotheses of Theorem \ref{quasi}, for all
integers }$\beta \geq 1$, {every smooth solution $w$ of (\ref{equation}) in $%
\Omega $ satisfies }
\begin{eqnarray*}
&&\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}
,w}\zeta \right\vert ^{2}\left\vert w_{j}\right\vert ^{2\beta } \\
&\leq &CC_{1}\Lambda \left( \left( A^{4}+A^2B^{2}\right) M_{0}^{2}+
A^2\left\Vert f\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right)
}\right) \sum_{j=1}^{n}\int_{2\mathcal{R}}\xi ^{2} w_{j}^{2\beta -2} \\
&&+\mathcal{C}\left( C_{1},\Lambda ,M_{0}\right) \sum_{j=1}^{n}\int_{2
\mathcal{R}}\left\vert \nabla \zeta \right\vert ^{2}\left\vert \nabla _{\!\!
\sqrt{\mathcal{A}},w}w_{j}^{\beta -1}\right\vert ^{2}.
\end{eqnarray*}
{Here $C_1$ as in Lemma \ref{minkgrad}, $\mathcal{R}$ is any box such that }$%
3\mathcal{R} \subset \Omega $ and $\mathcal{R}$ satisfies property (\ref%
{ax1c3}) in the nondegeneracy Condition{\ \ref{hyp1}, }$M_{0}=\left\Vert
w\right\Vert _{L^{\infty }\left( 2 \mathcal{R}\right) }${, }$\tilde{\Gamma}=2%
\mathcal{R}\times \left[ -M_{0},M_{0}\right] $, {and $B=B_{\gamma }\left( 2%
\mathcal{R},M_{0}\right) $ is as in Definition \ref{subunitt}. }
\end{lemma}
{\ \vspace*{-0.2in}%
\proof%
By the product rule,
\begin{eqnarray*}
\left\vert w_{j}\right\vert ^{2\beta } &=&\left( \partial _{j}\big(%
ww_{j}^{\beta -1}\big)-w\partial _{j}\big(w_{j}^{\beta -1}\big)\right) ^{2}
\\
&\leq &2\left( \partial _{j}\big(ww_{j}^{\beta -1}\big)\right)
^{2}+2w^{2}\left( \partial _{j}\big(w_{j}^{\beta -1}\big)\right) ^{2}.
\end{eqnarray*}%
Then, using that $w$ is bounded by }$M_{0}${\ and applying Lemma \ref%
{minkgrad} gives
\begin{eqnarray}
&&\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}\zeta \right\vert ^{2}\left\vert w_{j}\right\vert ^{2\beta } \notag \\
&\leq &2\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}\zeta \right\vert ^{2}\left\vert \partial _{j}\big(%
ww_{j}^{\beta -1}\big)\right\vert ^{2}+2M_{0}^{2}\sum_{j=1}^{n}\int_{2%
\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\zeta \right\vert
^{2}\left\vert \partial _{j}\big(w_{j}^{\beta -1}\big)\right\vert ^{2}
\notag \\
&\leq &CC_{1}\Lambda \sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla
\zeta \right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\big(%
ww_{j}^{\beta -1}\big)\right\vert ^{2} \notag \\
&&+CC_{1}\Lambda M_{0}^{2}\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla
\zeta \right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}w_{j}^{\beta -1}\right\vert ^{2} \notag \\
&\leq &CC_{1}\Lambda \sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla
\zeta \right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}w\right\vert ^{2}w_{j}^{2\beta -2} \label{xint003} \\
&&+\mathcal{C}\left( C_{1},\Lambda ,M_{0}\right) \sum_{j=1}^{n}\int_{2%
\mathcal{R}}\left\vert \nabla \zeta \right\vert ^{2}\left\vert \nabla _{\!\!%
\sqrt{\mathcal{A}},w}w_{j}^{\beta -1}\right\vert ^{2}. \notag
\end{eqnarray}%
Integrating by parts in the first term on the right, and using the fact that
$w\,$is a solution of (\ref{equation}), we obtain}
\begin{equation*}
\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla \zeta \right\vert
^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert
^{2}w_{j}^{2\beta -2}=-\sum_{j=1}^{n}\int_{2\mathcal{R}}w\mathop{\rm div}%
\left[ \left\vert \nabla \zeta \right\vert ^{2}w_{j}^{2\beta -2}\mathcal{A}%
(x,w)\nabla w\right]
\end{equation*}%
\begin{equation*}
=-\sum_{j=1}^{n}\int_{2\mathcal{R}}ww_{j}^{2\beta -2}\left( \nabla
\left\vert \nabla \zeta \right\vert ^{2}\right) \cdot \mathcal{A}\left(
x,w\right) \nabla w
\end{equation*}%
\begin{equation*}
-\sum_{j=1}^{n}\int_{2\mathcal{R}}w\left\vert \nabla \zeta \right\vert
^{2}\left( \nabla \big(w_{j}^{2\beta -2}\big)\right) \cdot \mathcal{A}\left(
x,w\right) \nabla w
\end{equation*}%
\begin{equation*}
+\sum_{j=1}^{n}\int_{2\mathcal{R}}w\left\vert \nabla \zeta \right\vert
^{2}w_{j}^{2\beta -2}\big(\vec{\gamma}\left( x,w\right) \cdot \nabla
w+f\left( x,w\right) \big)
\end{equation*}%
\begin{equation}
=I+II+III. \label{xint004}
\end{equation}%
In $I$, using the estimate
\begin{equation*}
\left\vert \left( \nabla \left\vert \nabla \zeta \right\vert ^{2}\right)
\cdot \mathcal{A}\left( x,w\right) \nabla w\right\vert \leq \left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}\left( \left\vert \nabla \zeta \right\vert
^{2}\right) \right\vert \,\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}w\right\vert \leq CA^{2}\left\vert \nabla \zeta \right\vert \left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert ,
\end{equation*}%
we have
\begin{eqnarray}
\left\vert I\right\vert &\leq &M_{0}\sum_{j=1}^{n}\int_{2\mathcal{R}%
}\left\vert w_{j}\right\vert ^{2\beta -2}\left\vert \left( \nabla \left\vert
\nabla \zeta \right\vert ^{2}\right) \cdot \mathcal{A}\left( x,w\right)
\nabla w\right\vert \notag \\
&\leq &\dfrac{1}{4}\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert
w_{j}\right\vert ^{2\beta -2}\left\vert \nabla \zeta \right\vert
^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert
^{2}+CA^{4}M_{0}^{2}\sum_{j=1}^{n}\int_{2\mathcal{R}}\xi ^{2}\left\vert
w_{j}\right\vert ^{2\beta -2}. \label{xint005}
\end{eqnarray}%
Using the identity $\nabla w_{j}^{2\beta -2}=2w_{j}^{\beta -1}\nabla
w_{j}^{\beta -1}$ in $II$, we obtain
\begin{eqnarray}
\left\vert II\right\vert &\leq &2M_{0}\sum_{j=1}^{n}\int_{2\mathcal{R}%
}\left\vert \nabla \zeta \right\vert ^{2}\left\vert w_{j}\right\vert ^{\beta
-1}\left\vert \left( \nabla w_{j}^{\beta -1}\right) \cdot \mathcal{A}\left(
x,w\right) \nabla w\right\vert \notag \\
&\leq &\frac{1}{4}\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla \zeta
\right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert
^{2}w_{j}^{2\beta -2}+CM_{0}^{2}\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert
\nabla \zeta \right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}w_{j}^{\beta -1}\right\vert ^{2}. \label{xint0055}
\end{eqnarray}%
{Finally, since }$\vec{\gamma}$ is subunit with respect to $\mathcal{A}$ in $%
\Gamma =\Omega \times \mathbb{R}$, we have $\left\vert \vec{\gamma}\left(
x,w\right) \cdot \nabla w\right\vert \leq B\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w\right\vert $ for all $x\in 2\mathcal{R}$, where {$B=B\left(
2\mathcal{R},M_{0}\right) $. Then}
\begin{eqnarray}
\left\vert III\right\vert &\leq &M_{0}\sum_{j=1}^{n}\int_{2\mathcal{R}%
}\left\vert \nabla \zeta \right\vert ^{2}w_{j}^{2\beta -2}\big(\left\vert
\vec{\gamma}\left( x,w\right) \cdot \nabla w\right\vert +\left\vert f\left(
x,w\right) \right\vert \big) \notag \\
&\leq &\dfrac{1}{4}\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla \zeta
\right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert
^{2}w_{j}^{2\beta -2} \label{xint006} \\
&&+\left( CB^{2}M_{0}^{2}+\left\Vert f\right\Vert _{L^{\infty }\left( \tilde{%
\Gamma}\right) }\right) \sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla
\zeta \right\vert ^{2}w_{j}^{2\beta -2}. \notag
\end{eqnarray}%
{\ \ Combining (\ref{xint004}), (\ref{xint005}), (\ref{xint0055}) and ( \ref%
{xint006}), and absorbing into the left yields}
\begin{eqnarray*}
&&\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla \zeta \right\vert
^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert
^{2}w_{j}^{2\beta -2} \\
&\leq &CM_{0}^{2}\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla \zeta
\right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w_{j}^{\beta
-1}\right\vert ^{2} \\
&&+C\left( \left( A^{4}+A^{2}B^{2}\right) M_{0}^{2}+A^{2}\left\Vert
f\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) }\right)
\sum_{j=1}^{n}\int_{2\mathcal{R}}\xi ^{2}w_{j}^{2\beta -2}.
\end{eqnarray*}%
Using this estimate in the first term on the right of ({\ref{xint003}}){\
finishes the proof of Lemma \ref{extraprev}.
\endproof%
}
\begin{lemma}
{\ \label{extraprev2}Under the hypothesis of Theorem \ref{quasi}, if $w$ is
a smooth solution of (\ref{equation}) in $\Omega $, then for any $\beta \in
\mathbb{N}$ and any box $\mathcal{R}\subset \Omega $ satisfying property (%
\ref{ax1c3}) in }the nondegeneracy Condition{\ \ref{hyp1},
\begin{eqnarray*}
&&\sum_{j=1}^n \int_{2\mathcal{R}}\zeta ^{2}\left\vert w_{j}\right\vert
^{2\beta }\left\vert \nabla _{\!\!\mathcal{A},w}w\right\vert ^{2} \\
&\leq &CB^{2}\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
_{\!\!\mathcal{A},w}w_{j}^{\beta }\right\vert ^{2}+CB^{2} \int_{2\mathcal{R}%
}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}, w}\zeta \right\vert
^{2}\left\vert \nabla w\right\vert ^{2\beta }.
\end{eqnarray*}
Here $B=B_{\gamma }\left( 2\mathcal{R},M_{0}\right) $ is as in Definition %
\ref{subunitt} and $C$ depends on $M_0$ and $||f||_{L^\infty(\tilde{\Gamma}%
)} $. }
\end{lemma}
{\vspace*{-0.2in}%
\proof%
Integrating by parts, we have
\begin{eqnarray}
&&\int_{2\mathcal{R}}\zeta ^{2}w_{j}^{2\beta }\left\vert \nabla _{\!\!%
\mathcal{A},w}w\right\vert ^{2} \notag \\
&=&-\int_{2\mathcal{R}}w\mathop{\rm div}\left[ \zeta ^{2}w_{j}^{2\beta }%
\mathcal{A}\left( x,w\right) \nabla w\right] \notag \\
&=&-2\int_{2\mathcal{R}}w\zeta w_{j}^{2\beta }\left( \nabla \zeta \right)
\cdot \mathcal{A}\left( x,w\right) \nabla w \notag \\
&&-2\int_{2\mathcal{R}}w\zeta ^{2}w_{j}^{\beta }\left( \nabla w_{j}^{\beta
}\right) \cdot \mathcal{A}\left( x,w\right) \nabla w \notag \\
&&+\int_{2\mathcal{R}}w\zeta ^{2}w_{j}^{2\beta }\big(\vec{\gamma}(x,w)\cdot
\nabla w+f(x,w)\big) \notag \\
&=&I+II+III. \label{xxtp200}
\end{eqnarray}%
By Schwarz's' inequality and since }$\vec{\gamma}${\ }is of subunit type
with respect to $\mathcal{A},${\
\begin{eqnarray*}
\left\vert I\right\vert &\leq &CM_{0}^{2}\int_{2\mathcal{R}}\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2}w_{j}^{2\beta }+%
\frac{1}{6}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w\right\vert ^{2}w_{j}^{2\beta } \\
\left\vert II\right\vert &\leq &CM_{0}^{2}\int_{2\mathcal{R}}\zeta
^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w_{j}^{\beta }\right\vert
^{2}+\frac{1}{6}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w\right\vert ^{2}w_{j}^{2\beta } \\
\left\vert III\right\vert &\leq &CB^{2}M_{0}^{2}\left( 1+||f||_{L^{\infty }(%
\tilde{\Gamma})}^{2}\right) \int_{2\mathcal{R}}\zeta ^{2}w_{j}^{2\beta }+%
\frac{1}{6}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w\right\vert ^{2}w_{j}^{2\beta }.
\end{eqnarray*}%
Applying these estimates to (\ref{xxtp200}) and absorbing into the left
gives
\begin{eqnarray*}
&&\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}w_{j}^{2\beta }\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert ^{2} \\
&\leq &C\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!%
\mathcal{A},w}w_{j}^{\beta }\right\vert ^{2}+C\sum_{j=1}^{n}\int_{2\mathcal{R%
}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\zeta \right\vert
^{2}w_{j}^{2\beta }+CB^{2}\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta
^{2}w_{j}^{2\beta }
\end{eqnarray*}%
with $C$ depending on $M_{0}$ and $||f||_{L^{\infty }(\tilde{\Gamma})}$. To
obtain the conclusion of the lemma, we apply the Sobolev inequality (\ref%
{spoinc}) to the last term on the right and note that
\begin{equation*}
CB^{2}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\big(%
\zeta w_{j}^{\beta }\big)\right\vert ^{2}
\end{equation*}%
is bounded by the sum of the first two terms on the right.
\endproof%
}
\begin{theorem}
{\ \label{extraintegrability}Under the hypothesis of Theorem \ref{apriorial}
, if $w$ is a smooth solution of (\ref{equation}) in $\Omega $, then for all
integers $\beta \geq 1$ and every open }$\Omega ^{\prime }$ with $\Omega
^{\prime }\Subset \Omega ${, there exists a positive constant
\begin{equation*}
\mathcal{C}_{\beta }=\mathcal{C}_{\beta }\left( M_0,n,B,\Lambda ,\vec{k}
,\left\Vert f\right\Vert _{\mathcal{C}^{1}\left( \Gamma ^{\prime }\right)
},\left\Vert \vec{\gamma}\right\Vert _{\mathcal{C}^{1}\left( \Gamma ^{\prime
}\right) }, \Omega,\mathop{\rm dist}\left( \Omega ^{\prime },\partial \Omega
\right) \right)
\end{equation*}
such that%
\begin{equation*}
\sum_{j=1}^{n}\int_{\Omega ^{\prime }} w_{j}^{2\beta
}+\sum_{j=1}^{n}\int_{\Omega ^{\prime }}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}} ,w}w_{j}^{\beta }\right\vert ^{2}\leq \mathcal{C}_{\beta }.
\end{equation*}
Here }$M_{0}=\left\Vert w\right\Vert _{L^{\infty }\left( \Omega ^{\prime
}\right) }$, $B$ denotes the constants $B_{\mathcal{A}}\left( \Omega
^{\prime },M_{0}\right) ,B_{\gamma }\left( \Omega ^{\prime },M_{0}\right)$
in (\ref{xtra}) and (\ref{gxtra}), and $\Gamma ^{\prime }=\Omega ^{\prime
}\times \left[ -M_{0},M_{0}\right]$.
\end{theorem}
{\vspace*{-0.2in}%
\proof%
Let $\mathcal{R}$ be a box satisfying property (\ref{ax1c3}) in }the
nondegeneracy Condition{\ \ref{hyp1} and such that $2\mathcal{R}\Subset
\Omega $. From Lemma \ref{Moser},
\begin{eqnarray}
&&\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \zeta \nabla _{\!\!\sqrt{%
\mathcal{A}},w}\,w_{j}^{\beta }\right\vert ^{2} \notag \\
&\leq &C\beta \sum_{j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}\left(
\mathcal{L}_{w}w_{j}\right) \left( w_{j}^{2\beta -1}\right) \right\vert
+C\sum_{j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}\,\zeta \right\vert ^{2}w_{j}^{2\beta }. \label{mmoo}
\end{eqnarray}%
Differentiating (\ref{equation}) with respect to the $j^{th}\,$variable, we
obtain
\begin{equation}
-\mathcal{L}_{w}w_{j}=\mathop{\rm div}\left\{ \partial _{j}\mathbf{A}\left(
x\right) \right\} \nabla w+\left( \partial _{j}\vec{\gamma}\right) \cdot
\nabla w+\vec{\gamma}\cdot \nabla w_{j}+\partial _{j}f \label{nonhom}
\end{equation}%
for all $j$, where $\mathcal{L}_{w}$ is the linear operator (\ref{operatorw}%
), $\vec{\gamma}=\vec{\gamma}\left( x,w\right) $ and $\partial
_{j}f=\partial \lbrack f(x,w)]$. Hence
\begin{eqnarray*}
&&\sum_{j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}\left( \mathcal{L}%
_{w}w_{j}\right) \left( w_{j}^{2\beta -1}\right) \right\vert \\
&=&\sum_{j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}%
{\Big (}%
\mathop{\rm div}\left\{ \partial _{j}\mathbf{A}\left( x\right) \right\}
\nabla w+\left( \partial _{j}\vec{\gamma}\right) \cdot \nabla w+\vec{\gamma}%
\cdot \nabla w_{j}+\left( \partial _{j}f\right)
{\Big )}%
\left( w_{j}^{2\beta -1}\right) \right\vert \\
&\leq &\sum_{j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}%
{\Big (}%
\mathop{\rm div}\left\{ \partial _{j}\mathbf{A}\left( x\right) \right\}
\nabla w%
{\Big )}%
\left( w_{j}^{2\beta -1}\right) \right\vert \\
&&+\sum_{j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}%
{\Big (}%
\left( \partial _{j}\vec{\gamma}\right) \cdot \nabla w+\left( \partial
_{j}f\right)
{\Big )}%
\left( w_{j}^{2\beta -1}\right) \right\vert +\sum_{j=1}^{n}\left\vert
\sum_{i=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\gamma ^{i}w_{ij}w_{j}^{2\beta
-1}\right\vert
\end{eqnarray*}%
\begin{equation}
=I+II+III. \label{extint00}
\end{equation}%
From (\ref{wirtm}), (\ref{xtra}) and the inequality
\begin{equation*}
\sqrt{k^{\ast }\left( x,w\right) }\left\vert \partial _{j}u\right\vert \leq
\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}u\right\vert ,\qquad 1\leq
j\leq n,
\end{equation*}%
where $u$ is any smooth function, we get (with $B_{\mathcal{A}}$ as in (\ref%
{xtra}))
\begin{eqnarray}
\left\vert \left\{ \partial _{j}\mathbf{A}\left( x\right) \right\} \nabla
u\right\vert &=&\left\vert \left\{ \mathcal{A}_{j}+w_{j}\mathcal{A}%
_{z}\right\} \nabla u\right\vert \notag \\
&\leq &CB_{\mathcal{A}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}u\right\vert +CB_{\mathcal{A}}\sqrt{k^{\ast }\left( x,w\right) }%
\left\vert w_{j}\right\vert \left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}u\right\vert \notag \\
&\leq &CB_{\mathcal{A}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}u\right\vert +CB_{\mathcal{A}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}w\right\vert \left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}u\right\vert .
\label{wirtX}
\end{eqnarray}%
Writing
\begin{equation*}
\nabla \left( \zeta ^{2}w_{j}^{2\beta -1}\right) =\frac{2\beta -1}{\beta }%
\zeta ^{2}w_{j}^{\beta -1}\nabla w_{j}^{\beta }+w_{j}^{2\beta -1}\nabla
\zeta ^{2},
\end{equation*}%
integrating by parts, and using $\frac{2\beta -1}{\beta }\leq 2$ and (\ref%
{wirtX} ), we obtain
\begin{eqnarray}
I &=&\sum_{j=1}^{n}\left\vert \int_{2\mathcal{R}}\left( \nabla w\right)
\cdot \left\{ \partial _{j}\mathbf{A}\left( x\right) \right\} \nabla \left(
\zeta ^{2}w_{j}^{2\beta -1}\right) \right\vert \notag \\
&\leq &C\sum_{j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}w_{j}^{\beta
-1}\left( \nabla w\right) \cdot \left\{ \partial _{j}\mathbf{A}\left(
x\right) \right\} \nabla w_{j}^{\beta }\right\vert \notag \\
&&+\sum_{j=1}^{n}\left\vert \int_{2\mathcal{R}}w_{j}^{2\beta -1}\left(
\nabla w\right) \cdot \left\{ \partial _{j}\mathbf{A}\left( x\right)
\right\} \nabla \zeta ^{2}\right\vert \notag \\
&\leq &CB_{\mathcal{A}}\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert
w_{j}\right\vert ^{\beta -1}\left\vert \nabla w\right\vert \left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}w_{j}^{\beta }\right\vert \left\{ 1+\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert \right\} \notag \\
&&+CB_{\mathcal{A}}\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta \left\vert
w_{j}\right\vert ^{2\beta -1}\left\vert \nabla w\right\vert \left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}\zeta \right\vert \left\{ 1+\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert \right\} \notag \\
&\leq &\alpha \sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}w_{j}^{\beta }\right\vert ^{2}+C\int_{2\mathcal{R}%
}\left\vert \nabla w\right\vert ^{2\beta }\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}\zeta \right\vert ^{2} \label{xxiiI} \\
&&+\frac{CB_{\mathcal{A}}^{2}}{\alpha }\int_{2\mathcal{R}}\zeta
^{2}\left\vert \nabla w\right\vert ^{2\beta }+\frac{CB_{\mathcal{A}}^{2}}{%
\alpha }\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla w\right\vert ^{2\beta
}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert ^{2}. \notag
\end{eqnarray}%
}
From the chain rule, the super subordination Condition (\ref{gxtra}) for $%
\gamma ${, and Young's inequality,
\begin{eqnarray}
II &\leq &\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left( \left\vert
f_{j}\right\vert +\left\vert w_{j}\right\vert \left\vert f_{z}\right\vert
+\sum_{i=1}^{n}\left\vert \gamma _{j}^{i}\right\vert \left\vert
w_{i}\right\vert \right) \left\vert w_{j}\right\vert ^{2\beta -1} \notag \\
&&+\sum_{j=1}^{n}\left\vert \int_{2\mathcal{R}}\sum_{i=1}^{n}\zeta
^{2}w_{i}\left( \gamma _{z}^{i}\right) w_{j}^{2\beta }\right\vert \notag \\
&\leq &\mathcal{C}_{a}\int_{2\mathcal{R}}\zeta ^{2}\left( \left\vert \nabla
w\right\vert ^{2\beta -1}+\left\vert \nabla w\right\vert ^{2\beta }\right)
+C\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}w_{j}^{2\beta }\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert ^{2} \notag \\
&\leq &\mathcal{C}_{a}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
w\right\vert ^{2\beta }+C\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
w\right\vert ^{2\beta }\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}w\right\vert ^{2}+\mathcal{C}_{a}, \label{xxiiII}
\end{eqnarray}%
where } $\mathcal{C}_{a}=\mathcal{C}_{a}\left( B_{\gamma },\Lambda
,\left\Vert f\right\Vert _{\mathcal{C}^{1}\left( \tilde{\Gamma}\right)
,}\left\Vert \vec{\gamma}\right\Vert _{\mathcal{C}^{1}\left( \tilde{\Gamma}%
\right) }\right) $ with $B_{\gamma }$ as in (\ref{gxtra}).
Since $\vec{\gamma}$ {\ }is of subunit type with respect to $\mathcal{A}$,
\begin{eqnarray}
III &=&\frac{1}{\beta }\sum_{j=1}^{n}\left\vert \sum_{i=1}^{n}\int_{2%
\mathcal{R}}\left( \zeta \gamma ^{i}\partial _{i}\left( w_{j}^{\beta
}\right) \right) \left( \zeta w_{j}^{\beta }\right) \right\vert \notag \\
&\leq &\alpha \sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}w_{j}^{\beta }\right\vert ^{2}+\frac{B_{\gamma
}^{2}}{\alpha }\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}w_{j}^{2\beta }.
\label{xxiiIII}
\end{eqnarray}%
Using (\ref{xxiiI}), (\ref{xxiiII}) and (\ref{xxiiIII}) in (\ref{extint00})
yields
\begin{eqnarray*}
&&\sum_{j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}\left( \mathcal{L}%
_{w}w_{j}\right) w_{j}^{2\beta -1}\right\vert \\
&\leq &C\alpha \sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}w_{j}^{\beta }\right\vert ^{2}+C\int_{2\mathcal{R}%
}\left\vert \nabla w\right\vert ^{2\beta }\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}\zeta \right\vert ^{2} \\
&&+\frac{\mathcal{C}_{a}B^{2}}{\alpha }\int_{2\mathcal{R}}\zeta
^{2}\left\vert \nabla w\right\vert ^{2\beta }+\frac{CB_{\mathcal{A}}^{2}}{%
\alpha }\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla w\right\vert ^{2\beta
}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert ^{2}+\mathcal{C}%
_{a}.
\end{eqnarray*}%
Substituting on the right of (\ref{mmoo}) and absorbing into the left, we
get
\begin{eqnarray}
\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w_{j}^{\beta }\right\vert ^{2} &\leq &C(\beta +1)\int_{2%
\mathcal{R}}\left\vert \nabla w\right\vert ^{2\beta }\left\vert \nabla _{\!\!%
\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2} \label{xint02} \\
&&+\,\mathcal{C}_{a}\beta B^{2}\int_{2\mathcal{R}}\zeta ^{2}\left\vert
\nabla w\right\vert ^{2\beta } \notag \\
&&+\,C\beta B_{\mathcal{A}}^{2}\int_{2\mathcal{R}}\zeta ^{2}\left\vert
\nabla w\right\vert ^{2\beta }\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}w\right\vert ^{2}+\beta \mathcal{C}_{a}. \notag
\end{eqnarray}%
Now we will further assume that for fixed constants $C$ and $\beta ,$ the
constant $B_{\mathcal{A}}$ is small enough so that
\begin{equation}
C\beta B_{\mathcal{A}}^{2}B^{2}\leq \frac{1}{2}. \label{double-xtra}
\end{equation}%
We will show later that this assumption causes no loss of generality.
Applying Lemma \ref{extraprev2} to the third term on the right of (\ref%
{xint02}), and absorbing into the left using (\ref{double-xtra}), we obtain
\begin{eqnarray*}
\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w_{j}^{\beta }\right\vert ^{2} &\leq &C(\beta +1)\int_{2%
\mathcal{R}}\left\vert \nabla w\right\vert ^{2\beta }\left\vert \nabla _{\!\!%
\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2}+\mathcal{C}_{a}\beta B^{2}\int_{2%
\mathcal{R}}\zeta ^{2}\left\vert \nabla w\right\vert ^{2\beta } \\
&&+C\beta B_{\mathcal{A}}^{2}B^{2}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!%
\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2}\left\vert \nabla w\right\vert
^{2\beta }+\beta \mathcal{C}_{a}
\end{eqnarray*}%
where $C$ depends on $M_{0}$ and $||f||_{L^{\infty }(\tilde{\Gamma})}$. In
turn, {applying the Sobolev inequality (\ref{spoinc}) to each $\zeta
w_{j}^{\beta }$ in the second term on the right, and applying Lemma \ref%
{extraprev} to the first and third terms on the right (and the similar term
arising from the Sobolev inequality), }we obtain
\begin{eqnarray}
&&\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w_{j}^{\beta }\right\vert ^{2} \notag \\
&\leq &\beta \mathcal{C}_{a}+\mathcal{C}_{b}\sum_{j=1}^{n}\int_{2\mathcal{R}%
}\left\vert \nabla \zeta \right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w_{j}^{\beta -1}\right\vert ^{2}+\mathcal{C}%
_{b}\sum_{j=1}^{n}\int_{2\mathcal{R}}\xi ^{2}w_{j}^{2\beta -2},
\label{xint-induct}
\end{eqnarray}%
after absorbing into the left, by taking $r^{1}$ small enough depending on $%
\beta $, $B$, $\Lambda $, $\left\Vert f\right\Vert _{\mathcal{C}^{1}\left(
\tilde{\Gamma}\right) }$ and $\left\Vert \vec{\gamma}\right\Vert _{\mathcal{C%
}^{1}\left( \tilde{\Gamma}\right) }$; here it is important to note that the
constants multiplying $r^{1}$ do not depend on the size of $\mathcal{R}$.
Also,
\begin{equation*}
\mathcal{C}_{b}=\mathcal{C}\left( M_{0},\vec{k},\mathcal{R},\beta
,n,B,\Lambda ,\left\Vert f\right\Vert _{\mathcal{C}^{1}\left( \tilde{\Gamma}%
\right) },\left\Vert \vec{\gamma}\right\Vert _{\mathcal{C}^{1}\left( \tilde{%
\Gamma}\right) }\right) .
\end{equation*}
The conclusion of the theorem will now follow by induction from (\ref%
{xint-induct}) and the Sobolev inequality. Indeed, if $\beta =1,$
\begin{equation*}
\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w_{j}\right\vert ^{2}\leq \beta\mathcal{C}_{a}+\mathcal{C}%
_{b} \int_{2\mathcal{R}}\xi ^{2}\leq \mathcal{C}_{b}.
\end{equation*}
By the Sobolev inequality (\ref{spoinc}) {and Lemma \ref{extraprev}},
\begin{equation}
\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2} w_{j}^{2}+ \sum_{j=1}^{n}\int_{2%
\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{\mathcal{A}}%
,w}w_{j}\right\vert ^{2}\leq \mathcal{C}_{b}. \label{xint-beta-0}
\end{equation}
Since $\Omega ^{\prime }\Subset \Omega $, there are a finite number of
rectangles $\left\{ \mathcal{R}_{i}\right\}$ satisfying (\ref{xint-beta-0})
and such that $\Omega ^{\prime }\subset \bigcup \mathcal{R} _{i}$. Choosing $%
\mathcal{C}_1$ as in Theorem \ref{extraintegrability} in case $\beta =1$, it
follows that{\
\begin{equation*}
\sum_{j=1}^{n}\int_{\Omega ^{\prime }} w_{j}^{2}+\sum_{j=1}^{n} \int_{\Omega
^{\prime }}\left\vert \nabla _{\!\!\sqrt{\mathcal{A} },w}w_{j}\right\vert
^{2}\leq \mathcal{C}_{1}.
\end{equation*}
}Suppose now that we have shown that {\
\begin{equation}
\sum_{j=1}^{n}\int_{\Omega ^{\prime }} w_{j}^{2\beta}+
\sum_{j=1}^{n}\int_{\Omega ^{\prime }}\left\vert \nabla _{\!\!\sqrt{\mathcal{%
A}},w}w_{j}^{\beta }\right\vert ^{2}\leq \mathcal{C}_{\beta }.
\label{xint-induct-M}
\end{equation}
for }$\beta =1,2,\dots ,M$. Given $x\in \Omega ^{\prime }$, let $\mathcal{R}$
be a{\ box satisfying property (\ref{ax1c3}) in }Condition{\ \ref{hyp1} and
such that $x\in \frac{1}{3}\mathcal{R}\subset 2\mathcal{R}\subset \Omega
^{\prime }\Subset \Omega $. }Then, from (\ref{xint-induct}),
\begin{eqnarray*}
\sum_{j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w_{j}^{M+1}\right\vert ^{2} \leq (M+1)\mathcal{C}_{a} &+&
\mathcal{C} _{b}\sum_{j=1}^{n} \int_{2\mathcal{R}}\left\vert \nabla \zeta
\right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w_{j}^{M}
\right\vert ^{2} \\
&+& \mathcal{C}_{b}\int_{2\mathcal{R}}\xi ^{2}\left\vert \nabla w\right\vert
^{2M}
\end{eqnarray*}
\begin{equation*}
\leq (M+1)\mathcal{C}_{a}+A^{2}\mathcal{C}_{b}\sum_{j=1}^{n}\int_{\Omega
^{\prime }}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w_{j}^{M}
\right\vert ^{2}+ \mathcal{C}_{b}\int_{\Omega ^{\prime }}\left\vert \nabla
w\right\vert ^{2M} \leq \mathcal{C}_{M+1}.
\end{equation*}
An application of the Sobolev inequality,{\ Lemma \ref{extraprev}}, and a
covering argument complete the induction.
It remains to show that the extra assumption (\ref{double-xtra}) imposes no
loss of generality. For a constant $M\geq 1,$ let $u\left( x\right)
=Mw\left( x\right) $ and
\begin{eqnarray*}
\widetilde{\mathcal{A}}\left( x,q\right) &=&\mathcal{A}\left( x,q/M\right)
,\qquad \qquad \widetilde{k^{\ast }}\left( x,q\right) =k^{\ast }\left(
x,q/M\right) , \\
\widetilde{\vec{\gamma}}\left( x,q\right) &=&\vec{\gamma}\left( x,q/M\right)
,\qquad \qquad \widetilde{f}\left( x,q\right) =Mf\left( x,q/M\right) .
\end{eqnarray*}%
Then $u$ satisfies
\begin{eqnarray*}
&&\mathop{\rm div}\widetilde{\mathcal{A}}\left( x,u\right) \nabla u+%
\widetilde{\vec{\gamma}}\left( x,u\right) \cdot \nabla u+\widetilde{f}\left(
x,u\right) \\
&=&\mathop{\rm div}\widetilde{\mathcal{A}}\left( x,Mw\left( x\right) \right)
\nabla Mw\left( x\right) +\widetilde{\vec{\gamma}}\left( x,Mw\left( x\right)
\right) \cdot \nabla Mw\left( x\right) +\widetilde{f}\left( x,Mw\left(
x\right) \right) \\
&=&M\left[ \mathop{\rm div}\mathcal{A}\left( x,w\left( x\right) \right)
\nabla w\left( x\right) +\vec{\gamma}\left( x,w\left( x\right) \right) \cdot
\nabla w\left( x\right) +f\left( x,w\left( x\right) \right) \right] =0.
\end{eqnarray*}%
On the other hand, by (\ref{xtra}),
\begin{eqnarray*}
\left\vert \partial _{q}\widetilde{\mathcal{A}}\left( x,q\right) \xi
\right\vert ^{2} &=&\left\vert \partial _{q}\mathcal{A}\left( x,q/M\right)
\xi \right\vert ^{2}=M^{-2}\left\vert \mathcal{A}_{z}\left( x,q/M\right) \xi
\right\vert ^{2} \\
&\leq & M^{-2}B_{\mathcal{A}}^{2}\,k^{\ast }\left( x,q/M\right) \,\xi ^{t}%
\mathcal{A}\left( x,q/M\right) \xi \\
&=&B_{\mathcal{A}^{2}}M^{-2} \widetilde{k^{\ast }}\left( x,q\right) \,\xi
^{t}\widetilde{\mathcal{A}}\left( x,q\right) \xi \\
&=&\left( \widetilde{B}_{\widetilde{\mathcal{A}}}\right) ^{2}\,\widetilde{%
k^{\ast }}\left( x,q\right) \,\xi ^{t}\widetilde{\mathcal{A}}\left(
x,q\right) \xi
\end{eqnarray*}%
with $\widetilde{B}_{\widetilde{\mathcal{A}}}=B_{\mathcal{A}}M^{-1}$. Since $%
\gamma $ is of subunit type (Definition \ref{subunitt}),
\begin{equation*}
\left( \sum_{i=1}^{n}\widetilde{\gamma }^{i}\left( x,q\right) \xi
_{i}\right) ^{2}=\left( \sum_{i=1}^{n}\gamma ^{i}\left( x,q/M\right) \xi
_{i}\right) ^{2}\leq B_{\gamma }^{2}~\xi ^{t}\mathcal{A}\left( x,q/M\right)
\xi = B_{\gamma }^{2}~\xi ^{t}\widetilde{\mathcal{A}}\left( x,q\right)\xi.
\end{equation*}%
Also, by (\ref{gxtra}),%
\begin{eqnarray*}
\left\vert \partial _{q}\widetilde{\vec{\gamma}}\left( x,q\right) \cdot \xi
\right\vert ^{2} &=&M^{-2}\left\vert \vec{\gamma}_{z}\left( x,q/M\right)
\cdot \xi \right\vert ^{2}\leq M^{-2}B_{\gamma }^{2}~k^{\ast }\left(
x,q/M\right) \xi ^{t}\mathcal{A}\left( x,q/M\right) \xi \\
&=&\left( B_{\gamma }M^{-1}\right) ^{2}~\widetilde{k^{\ast }}\left(
x,q\right) \xi ^{t}\widetilde{\mathcal{A}}\left( x,q\right) \xi.
\end{eqnarray*}%
Hence $u\left( x\right) $ satisfies the equation
\begin{equation*}
\mathop{\rm div}\widetilde{\mathcal{A}}\left( x,u\right) \nabla u+\widetilde{%
\vec{\gamma}}\left( x,u\right) \cdot \nabla u+\widetilde{f}\left( x,u\right)
=0
\end{equation*}%
with the corresponding constants in Condition (\ref{hyp2}), namely
\begin{equation*}
\widetilde{B}_{\widetilde{\mathcal{A}}}= B_{\mathcal{A}}M^{-1}\qquad \text{%
and}\qquad \widetilde{B}_{\gamma }=B_{\gamma }.
\end{equation*}%
Hence, taking $M$ large enough and letting $\widetilde{B}=\max \{B_{\mathcal{%
A}}M^{-1},B_{\gamma }\}$, we have that
\begin{equation*}
C\beta \left( \widetilde{B}_{\widetilde{\mathcal{A}}}\right) ^{2}\widetilde{B%
}^{2}=C\beta \left( \frac{B_{\mathcal{A}}}{M}\right) ^{2}\max \left\{ \left(
\frac{B_{\mathcal{A}}}{M}\right) ^{2},B_{\gamma }^{2}\right\} \leq \frac{1}{2%
},
\end{equation*}%
so the extra assumption (\ref{double-xtra}) holds for the operator $%
\mathop{\rm div}\widetilde{\mathcal{A}}\left( x,\cdot \right) \nabla +%
\widetilde{\vec{\gamma}}\left( x,\cdot \right) \cdot \nabla +\widetilde{f}%
\left( x,\cdot \right) $. By the previous calculations, there is a constant
\begin{equation*}
\widetilde{\mathcal{C}}_{\beta }=\widetilde{\mathcal{C}}_{\beta }\left(
\widetilde{M_{0}},n,\widetilde{B},\Lambda ,\widetilde{\vec{k}},\left\Vert
\widetilde{f}\right\Vert _{\mathcal{C}^{1}\left( \widetilde{\Gamma }^{\prime
}\right) },\left\Vert \widetilde{\vec{\gamma}}\right\Vert _{\mathcal{C}%
^{1}\left( \widetilde{\Gamma }^{\prime }\right) },\Omega ,\mathop{\rm dist}%
\left( \Omega ^{\prime },\partial \Omega \right) \right)
\end{equation*}%
such that
\begin{equation*}
\sum_{j=1}^{n}\int_{\Omega ^{\prime }}u_{j}^{2\beta
}+\sum_{j=1}^{n}\int_{\Omega ^{\prime }}\left\vert \nabla _{\!\!\sqrt{%
\widetilde{\mathcal{A}}},u}u_{j}^{\beta }\right\vert ^{2}\leq \widetilde{%
\mathcal{C}}_{\beta };
\end{equation*}%
here $\widetilde{M_{0}}=\left\Vert u\right\Vert _{L^{\infty }\left( \Omega
^{\prime }\right) }=M\left\Vert w\right\Vert _{L^{\infty }\left( \Omega
^{\prime }\right) }$ {\ and} $\widetilde{\Gamma }^{\prime }=\Omega ^{\prime
}\times \left[ -\widetilde{M_{0}},\widetilde{M_{0}}\right] $. The general
result for $w$ follows from the identity $w=Mu$ and the definitions of $%
\widetilde{f}$ and $\widetilde{\vec{\gamma}}$ .{%
\endproof%
}
\subsection{Proof of Theorem \protect\ref{apriorial}}
{In this section we prove the a priori estimate in Theorem \ref{apriorial}
as a consequence of the higher integrability of $\nabla w$ established in
the previous section. Theorem \ref{apriorial} will be the main tool in the
proof of Theorem \ref{application}. }
{By Theorem \ref{quasi}, we only need to show that $\nabla w$ is locally
bounded in terms of appropriate parameters, i.e., we only need to show that
for every box $\mathcal{R}\subset 4\mathcal{R}\subset \Omega ^{\prime
}\Subset \Omega $,
\begin{equation*}
\left\Vert \nabla w\right\Vert _{L^{\infty }\left( \mathcal{R}\right) }\leq
\mathcal{C}\left( \left\Vert w\right\Vert _{L^{\infty }\left( \Omega
^{\prime }\right) },\mathcal{R},\vec{k},\mathcal{A},f,\vec{\gamma}\right) ;
\end{equation*}%
the dependence of }$\mathcal{C}${\ on its arguments will be made more
explicit as we proceed. By the Sobolev imbedding theorem, it is enough to
show that for some $\beta >n$,
\begin{equation}
\left\Vert w_{ij}\right\Vert _{L^{\beta }\left( 2\mathcal{R}\right) }\leq
\mathcal{C}_{\beta }\left( \left\Vert w\right\Vert _{L^{\infty }\left(
\Omega ^{\prime }\right) },\mathcal{R},\vec{k},\mathcal{A},f,\vec{\gamma}%
\right) ,\qquad 1\leq i,j\leq n. \label{secondkbeta}
\end{equation}%
}
\vspace*{-0.2in}{Let $\mathcal{R}\subset {4\mathcal{R}\subset \Omega }%
^{\prime }$ be a box satisfying property (\ref{ax1c3}) in the nondegeneracy
Condition\ \ref{hyp1}. Since such boxes cover }$\Omega ^{\prime }$, there is
no loss of generality in adding this extra condition. {Applying the
Caccioppoli inequality for the ${\mathcal{A}}$-gradient, Lemma \ref{Moser},
to the smooth functions }$w_{ij}${$=\partial _{i}\partial _{j}w$, we get for
$\beta \in \mathbb{N}$ that
\begin{eqnarray*}
&&\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}\,w_{ij}^{\beta }\right\vert ^{2} \\
&\leq &C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}\left(
\mathcal{L}_{w}w_{ij}\right) w_{ij}^{2\beta -1}\right\vert
+C\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A%
}},w}\,\zeta \right\vert ^{2}w_{ij}^{2\beta }.
\end{eqnarray*}%
We estimate the first term on the right by Lemma \ref{secondcacc}, obtaining}
\begin{eqnarray*}
&&\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}\,w_{ij}^{\beta }\right\vert ^{2} \\
&\leq &C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\left( \nabla
w_{j}\right) \cdot \mathbf{A}_{i}\nabla \zeta ^{2}w_{ij}^{2\beta
-1}\right\vert +C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}%
}w_{i}\left( \nabla w_{j}\right) \cdot \mathbf{A}_{z}\nabla \zeta
^{2}w_{ij}^{2\beta -1}\right\vert \\
&&+C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}w_{ij}\left( \nabla
w\right) \cdot \mathbf{A}_{z}\nabla \zeta ^{2}w_{ij}^{2\beta -1}\right\vert +%
\frac{\mathcal{C}_{0}\beta }{\alpha }\int_{2\mathcal{R}}\zeta ^{2}\left\vert
\nabla ^{2}w\right\vert ^{2\beta } \\
&&+C\beta \alpha \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert
\vec{\gamma}\cdot \nabla w_{ij}^{\beta }\right\vert ^{2}+C\beta \alpha
\int_{2\mathcal{R}}\zeta ^{2}\left\vert \vec{\gamma}_{z}\cdot \nabla
w\right\vert ^{2}\left\vert \nabla ^{2}w\right\vert ^{2\beta } \\
&&+C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}\left(
\left( \partial _{j}\vec{\gamma}\right) \cdot \nabla w_{i}\right)
w_{ij}^{2\beta -1}\right\vert +C\beta \sum_{i,j=1}^{n}\left\vert \int_{2%
\mathcal{R}}w_{i}\left( \nabla w\right) \cdot \mathbf{A}_{jz}\nabla \zeta
^{2}w_{ij}^{2\beta -1}\right\vert \\
&&+C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}w_{i}w_{j}\left(
\nabla w\right) \cdot \mathbf{A}_{zz}\nabla \zeta ^{2}w_{ij}^{2\beta
-1}\right\vert +\mathcal{C}_{0}\beta \int_{2\mathcal{R}}\zeta ^{2}\left\vert
\nabla w\right\vert ^{6\beta } \\
&&+C\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}\,\zeta \right\vert ^{2}w_{ij}^{2\beta }+\mathcal{C}_{0}\beta
\end{eqnarray*}%
\vspace*{-0.2in}%
\begin{eqnarray}
&=&I+II+\cdots +VIII+IX+\mathcal{C}_{0}\beta \int_{2\mathcal{R}}\zeta
^{2}\left\vert \nabla w\right\vert ^{6\beta } \label{snd-0} \\
&&+C\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}\,\zeta \right\vert ^{2}w_{ij}^{2\beta }+\mathcal{C}_{0}\beta
, \notag
\end{eqnarray}%
where $0<\alpha <1$ is arbitrary, $\vec{\gamma}=\vec{\gamma}\left(
x,w\right) $, $f=f\left( x,w\right) $, ${\mathbf{A}}=\mathcal{A}\left(
x,w\right) $, $\mathbf{A}_{i}=\mathcal{A}_{i}\left( x,w\right) $, etc., and
\begin{equation*}
\mathcal{C}_{0}=\mathcal{C}_{0}\left( \left\Vert \mathcal{A}\right\Vert _{%
\mathcal{C}^{3}\left( \Gamma ^{\prime }\right) },\left\Vert f\right\Vert _{%
\mathcal{C}^{2}\left( \Gamma ^{\prime }\right) },\left\Vert \vec{\gamma}%
\right\Vert _{\mathcal{C}^{2}\left( \Gamma ^{\prime }\right) }\right)
\end{equation*}%
with $\Gamma ^{\prime }=\Omega ^{\prime }\times \left[ -M_{0},M_{0}\right] $
and $M_{0}=\left\Vert w\right\Vert _{L^{\infty }\left( \Omega ^{\prime
}\right) }$.
Note that
\begin{equation*}
\partial _{i}\mathbf{A}\left( x\right) =\partial _{i}\mathcal{A}\left(
x,w\right) =\mathcal{A}_{i}\left( x,w\right) +w_{i}\mathcal{A}_{z}\left(
x,w\right) =\mathbf{A}_{i}+w_{i}\mathbf{A}_{z}.
\end{equation*}%
{If $u$ is any smooth function, from (\ref{wirtm}), (\ref{xtra}), (\ref%
{xxtra}), (\ref{gxtra}) and since }$\vec{\gamma}${\ }is of subunit type with
respect to $\mathcal{A},$\ it follows that
\begin{eqnarray}
\left\vert \mathbf{A}_{i} \nabla u\right\vert ^{2} &\leq &B_{\mathcal{A}%
}^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}, w}u\right\vert ^{2},
\label{snd-1} \\
\left\vert w_{i}\mathbf{A}_{z} \nabla u\right\vert ^{2} &\leq &B_{\mathcal{A}%
}^{2}\left\vert w_{i}\right\vert ^{2}k^{\ast }\left( x,w\right)
\,\sum_{j=1}^{n}k^{j}\left( x,w\right) \,\left( \partial _{j}u\right) ^{2}
\notag \\
&\leq &B_{\mathcal{A}}^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},
w}w\right\vert ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},
w}u\right\vert ^{2}, \label{snd-2} \\
\sum_{i=1}^{n}\left\vert \mathbf{A}_{iz}\nabla u\right\vert ^{2}+\left\vert
\mathbf{A}_{zz}\nabla u\right\vert ^{2} &\leq &\left( B_{\mathcal{A}%
}^{\prime }\right) ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},
w}u\right\vert ^{2}, \label{snd-3} \\
\left\vert \vec{\gamma}\cdot \nabla u\right\vert ^{2} &\leq &B_{\gamma
}^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}u\right\vert ^{2}\qquad
\text{and} \label{snd-4} \\
\sum_{i=1}^{n}\left\vert \vec{\gamma}_{i}\cdot \nabla u\right\vert
^{2}+\left\vert \vec{\gamma}_{z}\cdot \nabla u\right\vert ^{2} &\leq
&B_{\gamma }^{2}k^{\ast }\left( x,w\right) \left\vert \nabla _{\!\!\sqrt{%
\mathcal{A} },w}u\right\vert ^{2}. \label{snd-5}
\end{eqnarray}
We will now apply {these estimates together with the Schwarz and triangle
inequalities to treat each term of (\ref{snd-0}). } We will incorporate the
constants $B_{\mathcal{A}}, B_\gamma,$ etc. in our generic constant $C$.
By definition of $I$ and ({\ref{snd-1}}),
\begin{eqnarray*}
I &\leq &C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta
w_{ij}^{2\beta -1}\left( \nabla w_{j}\right) \cdot \mathbf{A}_{i}\left(
\nabla \zeta \right) \right\vert +C\beta \sum_{i,j=1}^{n}\left\vert \int_{2
\mathcal{R}}\zeta ^{2}w_{ij}^{\beta -1}\left( \nabla w_{j}\right) \cdot
\mathbf{A}_{i}\nabla w_{ij}^{\beta }\right\vert \\
&\leq &\alpha \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}w_{ij}^{\beta }\right\vert ^{2}+C\beta
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}
,w}\zeta \right\vert ^{2} w_{ij}^{2\beta }+\frac{C\beta ^{2}}{\alpha }%
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2} w_{ij}^{2\beta }.
\end{eqnarray*}
Similarly, using (\ref{snd-2}),
\begin{eqnarray*}
II &\leq &C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta
w_{i}w_{ij}^{2\beta -1}\left( \nabla w_{j}\right) \cdot \mathbf{A}%
_{z}\left(\nabla \zeta \right) \right\vert +C\beta
\sum_{i,j=1}^{n}\left\vert \int_{2 \mathcal{R}}\zeta ^{2}w_{ij}^{\beta
-1}w_{i}\left( \nabla w_{j}\right) \cdot \mathbf{A}_{z}\nabla w_{ij}^{\beta
}\right\vert \\
&\leq &\alpha \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}w_{ij}^{\beta }\right\vert ^{2}+C\beta
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}
,w}\zeta \right\vert ^{2} w_{ij}^{2\beta } \\
&&+\frac{C\beta ^{2}}{\alpha }\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta
^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert ^{2}
w_{ij}^{2\beta }.
\end{eqnarray*}
Treating $III$ analogously, we get
\begin{eqnarray}
&&I+II+III+IV \notag \\
&\leq &\alpha \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}w_{ij}^{\beta }\right\vert ^{2}+C\beta
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}
,w}\zeta \right\vert ^{2}\left\vert w_{ij}\right\vert ^{2\beta }
\label{snd-I-IV} \\
&&+\frac{C\beta ^{2}}{\alpha }\sum_{i,j=1}^{n}\int_{2\mathcal{R}} \zeta ^{2}
w_{ij}^{2\beta }+\frac{C\beta ^{2}}{\alpha } \sum_{i,j=1}^{n}\int_{2\mathcal{%
R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{\mathcal{A}}S,w}w\right\vert
^{2} w_{ij}^{2\beta }. \notag
\end{eqnarray}
By (\ref{snd-4}) and (\ref{snd-5}),
\begin{equation}
V+VI\leq C\beta \alpha \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta
^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w_{ij}^{\beta }\right\vert
^{2}+C\beta \alpha \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert ^{2} w_{ij}^{2\beta }.
\label{snd-V-VI}
\end{equation}
Now, using the identity $\partial _{j}\vec{\gamma} =\vec{\gamma}_{j}+w_{j}%
\vec{\gamma}_{z}$,
\begin{eqnarray*}
VII &\leq &C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta
^{2}\left( \vec{\gamma}_{j}\cdot \nabla w_{i}\right) w_{ij}^{2\beta
-1}\right\vert +C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta
^{2}w_{j}\left( \vec{\gamma}_{z}\cdot \nabla w_{i}\right) w_{ij}^{2\beta
-1}\right\vert \\
&=&VII_{1}+VII_{2}.
\end{eqnarray*}
We have
\begin{eqnarray*}
VII_{1} &\leq &C\beta \sum_{i,j=1}^{n} \int_{2\mathcal{R}} \zeta^{2}
\left\Vert \vec{\gamma} \right\Vert_{C^1(2\mathcal{R})} w_{ij}^{2\beta} \\
&\leq &\mathcal{C}_0 \beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}
w_{ij}^{2\beta }.
\end{eqnarray*}
Now, integrating by parts,
\begin{eqnarray*}
VII_{2} &=&C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\nabla
w\cdot \partial _{i}\left( \zeta ^{2}\vec{\gamma}_{z}w_{j}w_{ij}^{2\beta
-1}\right) \right\vert \\
&\leq &C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta w_{j}\zeta
_{i}w_{ij}^{2\beta -1}\left( \nabla w\right) \cdot \vec{\gamma}%
_{z}\right\vert +C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta
^{2}w_{j}w_{ij}^{2\beta -1}\left( \nabla w\right) \cdot \left( \partial _{i}%
\vec{\gamma}_{z}\right) \right\vert \\
&&+C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}\left(
\nabla w\right) \cdot \vec{\gamma}_{z}w_{ij}^{2\beta }\right\vert +C\beta
\sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}w_{j}w_{ij}^{\beta
-1}\left( \nabla w\right) \cdot \vec{\gamma}_{z}\left( \partial
_{i}w_{ij}^{\beta }\right) \right\vert \\
&=&VII_{2,1}+VII_{2,2}+VII_{2,3}+VII_{2,4}.
\end{eqnarray*}
Then
\begin{eqnarray}
VII_{2,1} &=&C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta
w_{j}w_{ij}^{2\beta -1}\zeta _{i}\vec{\gamma}_{z}\cdot \left( \nabla
w\right) \right\vert \notag \\
&\leq &\frac{C\beta }{\alpha }\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}
w_{ij}^{2\beta }+C\alpha \beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}%
}w_{j}^{2}\zeta _{i}^{2} w_{ij}^{2\beta -2}\left\vert \vec{\gamma}_{z}\cdot
\nabla w\right\vert ^{2}. \label{VII-2-1-0}
\end{eqnarray}
Assume for the moment that $\beta>1$. Then by using the super subordination
Condition \ref{hyp2} for $\gamma $, and Young's inequality in the form
\begin{equation*}
\int \left\vert fg\right\vert \leq \frac{\beta -1}{\beta }\int \left\vert
f\right\vert ^{\frac{2\beta }{2\beta -2}}+\frac{1}{\beta}\int \left\vert
g\right\vert^{\beta },
\end{equation*}
the second term on the right of (\ref{VII-2-1-0}) is bounded by
\begin{eqnarray*}
&& C\alpha \beta \sum_{i,j=1}^{n}\int_{2 \mathcal{R}}w_{j}^{2}
w_{ij}^{2\beta -2}\zeta _{i}^{2}k^{\ast }\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w \right\vert ^{2} \\
&\leq &C\alpha \beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}} w_{ij}^{2\beta
-2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w} \zeta \right\vert
^{2}\left\vert \nabla w\right\vert ^{4} \\
&\leq &C\alpha \beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}} w_{ij}^{2\beta
}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2}+C\alpha
\beta A^{2} \sum_{i,j=1}^{n}\int_{2 \mathcal{R}}\xi ^{2}\left\vert \nabla
w\right\vert ^{4\beta },
\end{eqnarray*}
where we used that $\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}, w}\zeta
\right\vert ^{2}\leq C A^{2}\xi ^{2}$. If on the other hand $\beta =1$, the
estimation of the second term in (\ref{VII-2-1-0}) is simpler; without using
Young's inequality, we can estimate it by just the second term above. Thus,
\begin{eqnarray*}
VII_{2,1} &\leq &\frac{C\beta }{\alpha }\sum_{i,j=1}^{n}\int_{2\mathcal{R}
}\zeta ^{2} w_{ij}^{2\beta }+C\alpha \beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}%
} w_{ij}^{2\beta} \left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\zeta
\right\vert ^{2} \\
&&+C\alpha \beta A^{2} \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\xi ^{2}\left\vert
\nabla w\right\vert ^{4\beta }.
\end{eqnarray*}
Similarly,
\begin{eqnarray*}
VII_{2,2} &\leq &\frac{C\beta }{\alpha }\sum_{i,j=1}^{n}\int_{2\mathcal{R}
}\zeta ^{2} w_{ij}^{2\beta }+ C\alpha \beta \sum_{i,j=1}^{n}\int_{2\mathcal{R%
}}\zeta ^{2} w_{j}^{2\beta-2}\left\vert \nabla w\right\vert ^4 \\
&\leq &\frac{C\beta }{\alpha }\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}
w_{ij}^{2\beta } +\mathcal{C}_0 \alpha \beta \int_{2\mathcal{R}}\zeta
^{2}\left\vert \nabla w\right\vert ^{4\beta },
\end{eqnarray*}
and
\begin{equation*}
VII_{2,3}\leq \frac{C\beta }{\alpha }\sum_{i,j=1}^{n}\int_{2 \mathcal{R}%
}\zeta ^{2} w_{ij}^{2\beta }+ C\alpha \beta \sum_{i,j=1}^{n}\int_{2\mathcal{R%
}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{\mathcal{A}},w}w\right\vert ^{2}
w_{ij}^{2\beta }.
\end{equation*}
Now, from (\ref{snd-5}),
\begin{eqnarray*}
\left\vert \left( \nabla w\right) \cdot \vec{\gamma}_{z}\left( \partial
_{i}w_{ij}^{\beta }\right) \right\vert &\leq &B_{\gamma }\left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}w\right\vert \sqrt{k^{\ast }\left( x,w\right) }
\left\vert \partial _{i}w_{ij}^{\beta }\right\vert \\
&\leq &B_{\gamma }\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}, w}w\right\vert
\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w} w_{ij}^{\beta }\right\vert .
\end{eqnarray*}
Thus,
\begin{eqnarray*}
VII_{2,4} &\leq &\frac{C\beta }{\alpha }\sum_{i,j=1}^{n} \int_{2\mathcal{R}%
}\zeta ^{2} w_{j}^{2} w_{ij}^{2\beta -2} \left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}} ,w}w\right\vert ^{2}+C\alpha \beta \sum_{i,j=1}^{n}\int_{2%
\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}w_{ij}^{\beta } \right\vert ^{2} \\
&\leq &\frac{C\beta }{\alpha }\sum_{i,j=1}^{n}\int_{2\mathcal{\ R}}\zeta
^{2} w_{ij}^{2\beta } +\frac{\mathcal{C}_0\beta }{\alpha }\int_{2\mathcal{R}%
}\zeta ^{2}\left\vert \nabla w\right\vert ^{4\beta }+C\alpha \beta
\sum_{i,j=1}^{n}\int_{2\mathcal{\ R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w} w_{ij}^{\beta }\right\vert ^{2}.
\end{eqnarray*}
Assembling all these estimates, we obtain
\begin{eqnarray}
VII &\leq &\frac{\mathcal{C}_0\beta }{\alpha } \sum_{i,j=1}^{n} \int_{2%
\mathcal{R}} \zeta ^{2} w_{ij}^{2\beta }+ C \alpha\beta
\sum_{i,j=1}^{n}\int_{2\mathcal{R}} w_{ij}^{2\beta } \left\vert \nabla _{\!\!%
\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2} \label{snd-VII} \\
&&+C\alpha \beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}w\right\vert ^{2} w_{ij}^{2\beta } \notag
\\
&&+\frac{\mathcal{C}_0 A^{2}\beta }{\alpha }\int_{2\mathcal{R} }\xi
^{2}\left\vert \nabla w\right\vert ^{4\beta }+C\alpha \beta
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{%
\mathcal{A}},w}w_{ij}^{\beta }\right\vert ^{2} . \notag
\end{eqnarray}
By (\ref{snd-3}),
\begin{eqnarray}
VIII &\leq &C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta
w_{i}w_{ij}^{2\beta -1}\left( \nabla w\right) \cdot \mathbf{A}_{jz}\left(
\nabla \zeta \right) \right\vert \notag \\
&&+C\beta \sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta
^{2}w_{i}\left( \nabla w\right) w_{ij}^{\beta -1}\cdot \mathbf{A}_{jz}\nabla
w_{ij}^{\beta }\right\vert \notag \\
&\leq &C\beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}} w_{ij}^{2\beta }\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2}+C\alpha \beta
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}},w}w_{ij}^{\beta }\right\vert ^{2} \notag \\
&&+\frac{C\left( B_{\mathcal{A}}^{\prime }\right) ^{2}\beta }{\alpha }
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}w_{i}^{2} w_{ij}^{2\beta
-2}\left\vert \nabla w\right\vert ^{2} \notag \\
&\leq &C\beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}} w_{ij}^{2\beta }\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2}+\frac{C\left( B_{%
\mathcal{A}}^{\prime }\right) ^{2}\beta }{\alpha }\sum_{i,j=1}^{n}\int_{2%
\mathcal{R}}\zeta ^{2} w_{ij}^{2\beta } \label{snd-VIII} \\
&&+\frac{\mathcal{C}_0\left( B_{\mathcal{A}}^{\prime }\right) ^{2}\beta }{%
\alpha }\int_{2 \mathcal{R}}\zeta ^{2}\left\vert \nabla w\right\vert
^{4\beta }+C\alpha \beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta
^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w_{ij}^{\beta }\right\vert
^{2}. \notag
\end{eqnarray}
Similarly,
\begin{eqnarray}
IX &\leq &C\beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}} w_{ij}^{2\beta
}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2}+\frac{%
\mathcal{C}_0\left( B_{\mathcal{A}}^{\prime }\right) ^{2}\beta }{\alpha }%
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2} w_{ij}^{2\beta }
\label{snd-IX} \\
&&+\frac{C\left( B_{\mathcal{A}}^{\prime }\right) ^{2}\beta }{\alpha }%
\int_{2 \mathcal{R}}\zeta ^{2}\left\vert \nabla w\right\vert ^{6\beta
}+C\alpha \beta \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}w_{ij}^{\beta }\right\vert ^{2}. \notag
\end{eqnarray}
Using estimates (\ref{snd-I-IV}), (\ref{snd-V-VI}), (\ref{snd-VII}), (\ref%
{snd-VIII}) and (\ref{snd-IX}) in (\ref{snd-0}), and absorbing into the
left, we obtain
\begin{eqnarray}
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{%
\mathcal{A}},w}\,w_{ij}^{\beta }\right\vert ^{2} &\leq &C\beta
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}
,w}\zeta \right\vert ^{2} w_{ij}^{2\beta } \label{tta} \\
&&+\mathcal{C}_0 \beta ^{2}\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}
w_{ij}^{2\beta } \notag \\
&&+C\beta ^{2}\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}w\right\vert ^{2} w_{ij}^{2\beta } \notag \\
&&+\mathcal{C}_0 \beta ^{2}\int_{2\mathcal{R}}\xi ^{2}\left\vert \nabla
w\right\vert ^{6\beta }+\mathcal{C}_0 A^{2}\beta . \notag
\end{eqnarray}
{We now estimate the third term on the right of (\ref{tta}) proceeding as in
the proof of Lemma \ref{extraprev2}. Integrating by parts,
\begin{eqnarray}
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{%
\mathcal{A}},w}w\right\vert ^{2} w_{ij}^{2\beta } &=&-\sum_{i,j=1}^{n}\int_{2%
\mathcal{R}}w\mathop{\rm div} \left\{\zeta ^{2}w_{ij}^{2\beta }\mathcal{A}%
\left( x,w\right) \nabla w\right\} \notag \\
&\leq &M_{0}\sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}} w_{ij}^{2\beta
}\left( \nabla \zeta ^{2}\right) \cdot \mathcal{A} \left( x,w\right) \nabla
w\right\vert \notag \\
&&+M_{0}\sum_{i,j=1}^{n}\left\vert \int_{2\mathcal{R}}\zeta ^{2}\left(
\nabla w_{ij}^{2\beta }\right) \cdot \mathcal{A}\left( x,w\right) \nabla
w\right\vert \notag \\
&&+M_{0}\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2} w_{ij}^{2\beta
}\left\vert \vec{\gamma}\cdot \nabla w+f\right\vert . \notag
\end{eqnarray}
By Schwarz's inequality, the identity }$\nabla w_{ij}^{2\beta
}=2w_{ij}^{\beta }\nabla w_{ij}^{\beta }${, and (\ref{snd-4}), we obtain
after absorbing into the left,
\begin{eqnarray*}
&&\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!
\sqrt{\mathcal{A}},w}w\right\vert ^{2} w_{ij}^{2\beta }\leq C(M_{0}^{2}+1)
\left( \left\Vert f\right\Vert _{L^{\infty }\left( \Gamma ^{\prime }\right)
}+B_{\gamma}^{2}\right) \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}
w_{ij}^{2\beta } \\
&&+CM_{0}^{2}\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!
\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2} w_{ij}^{2\beta}+
CM_{0}^{2}\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}\,w_{ij}^{\beta }\right\vert ^{2}.
\end{eqnarray*}
Using this on the right of (\ref{tta}) gives}
\begin{eqnarray}
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{%
\mathcal{A}},w}\,w_{ij}^{\beta }\right\vert ^{2} &\leq &C\beta ^{2}\left(
M_{0}^{2}+1\right) \sum_{i,j=1}^{n}\int_{2\mathcal{R}} \left\vert \nabla
_{\!\!\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2} w_{ij}^{2\beta }
\label{tta1} \\
&&+ \mathcal{C}_1 \sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}
w_{ij}^{2\beta } \notag \\
&&+C\beta ^{2}M_{0}^{2}\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta
^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\,w_{ij}^{\beta }
\right\vert ^{2} \notag \\
&&+ \mathcal{C}_0 B^2\beta ^{2}\int_{2\mathcal{R}}\xi ^{2}\left\vert \nabla
w\right\vert ^{6\beta }+ \mathcal{C}_0 A^{2}\beta , \notag
\end{eqnarray}
with
\begin{equation*}
\mathcal{C}_{1}=\mathcal{C}_{1}\left( M_0, B,\beta,\left\Vert \mathcal{A}
\right\Vert _{\mathcal{C}^{3}\left( \Gamma ^{\prime }\right) },\left\Vert
f\right\Vert _{\mathcal{C}^{2}\left( \Gamma ^{\prime }\right) },\left\Vert
\vec{\gamma}\right\Vert _{\mathcal{C}^{2}\left( \Gamma ^{\prime }\right)
}\right) .
\end{equation*}
{By the Sobolev inequality (\ref{spoinc}) and the product rule,
\begin{equation*}
\int_{2\mathcal{R}}\zeta ^{2} w_{ij}^{2\beta }\leq C\left( r^{1}\right)
^{2}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}w_{ij}^{\beta }\right\vert ^{2}+C\left( r^{1}\right) ^{2}\int_{2\mathcal{R%
}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}\zeta \right\vert ^{2}\,
w_{ij}^{2\beta }.
\end{equation*}
Applying this to the second term on the right of (\ref{tta1}), and taking }$%
r^1$ small enough depending on $M_0, \beta $, $B$, $\left\Vert \mathcal{A}%
\right\Vert _{\mathcal{C}^{3}\left( \Gamma ^{\prime }\right) }$, $\left\Vert
f\right\Vert _{\mathcal{C}^{2}\left( \Gamma ^{\prime }\right) }$ and $%
\left\Vert \vec{\gamma}\right\Vert _{\mathcal{C}^{2}\left( \Gamma ^{\prime
}\right) }$ (in order to absorb the term resulting from the first term on
the right of the last estimate), we get
\begin{equation*}
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{%
\mathcal{A}},w}\,w_{ij}^{\beta }\right\vert ^{2} \leq \mathcal{C}
_{1}\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A} } ,w}\zeta \right\vert ^{2} w_{ij}^{2\beta }
\end{equation*}
\begin{equation*}
+C\beta ^{2}M_{0}^{2}\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w}\,w_{ij}^{\beta }\right\vert ^{2} +%
\mathcal{C}_{0}B^{2}\beta ^{2}\int_{2\mathcal{R}}\xi ^{2}\left\vert \nabla
w\right\vert ^{6\beta }+\mathcal{C}_{0}A^{2}\beta .
\end{equation*}
{Now restrict }$1\leq \beta \leq n+1$, and assume that $M_{0}$ is small
enough so that
\begin{equation}
C\beta ^{2}M_{0}^{2}\leq C\left( n+1\right) ^{2}M_{0}^{2}\leq 1/2.
\label{triple-xtra}
\end{equation}
Then the second term on the right above may be absorbed into the left side
to obtain
\begin{eqnarray}
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{%
\mathcal{A}},w}\,w_{ij}^{\beta }\right\vert ^{2} &\leq &\mathcal{C}
_{1}\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{%
\mathcal{A}} ,w}\zeta \right\vert ^{2} w_{ij}^{2\beta } \label{tta2} \\
&&+\mathcal{C}_{0}B^{2}\beta ^{2}\int_{2\mathcal{R}}\xi ^{2}\left\vert
\nabla w\right\vert ^{6\beta }+\mathcal{C}_{0}A^{2}\beta . \notag
\end{eqnarray}
{For any $q\geq 1,$
\begin{equation*}
\left\Vert \chi \nabla \mathbf{A}\right\Vert _{L^{q}}\leq C\left( \int_{2
\mathcal{R}}\sum_{i=1}^n \left\vert \mathbf{A}_{i}\right\vert
^{q}+\left\vert \nabla w\right\vert ^{q}\left\vert \mathbf{A}_{z}\right\vert
^{q}\right) ^{\frac{1}{ q}}\leq C\left\Vert \mathcal{A}\right\Vert _{%
\mathcal{C}^{1}\left( \Gamma ^{\prime }\right) }\left( 1+\left\Vert \nabla
w\right\Vert _{L^{q}\left( 2\mathcal{R}\right) }\right).
\end{equation*}
Then from Lemma \ref{lemma-interpolation} applied to the function $u = \xi
w_ij$, choosing $q=n+1$, there exists $1<p=p\left( n\right) <2$ such that
\begin{eqnarray*}
&&\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A%
}},w}\,\zeta \right\vert ^{2} w_{ij}^{2\beta } = \sum_{i,j=1}^{n} \int_{2%
\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}} ,w}\,\zeta
\right\vert ^{2} \left(\xi w_{ij}\right)^{2\beta } \\
&\leq& \varepsilon ^{-1}\mathcal{C}\left( n,A,\left\Vert \nabla w\right\Vert
_{L^{n+1}\left( 2\mathcal{R}\right) }\right) \sum_{i,j=1}^{n}\left\Vert \xi
w_{ij}^{\beta }\right\Vert _{L^{p}}^{2} +C\varepsilon \sum_{i,j=1}^{n}\int_{2%
\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w}\,w_{ij}^{\beta }\right\vert ^{2}
\end{eqnarray*}
for all }$1\leq \beta \leq n+1$. {On the other hand, by Theorem \ref%
{extraintegrability},
\begin{equation}
\left\Vert \nabla w\right\Vert _{L^{n+1}\left( 2\mathcal{R}\right)
}+\left\Vert \nabla w\right\Vert _{L^{6\beta }\left( 2\mathcal{R}\right)
}\leq \mathcal{C}_{2} \label{ttw}
\end{equation}
where }$\mathcal{C}_{2}=\mathcal{C}_{2}\left( M_0,n,B,\Lambda ,\vec{k}
,\left\Vert f\right\Vert _{\mathcal{C}^{1}\left( \Gamma ^{\prime }\right)
},\left\Vert \vec{\gamma}\right\Vert _{\mathcal{C}^{1}\left( \Gamma ^{\prime
}\right) }, \Omega ,\mathop{\rm dist}\left( \Omega ^{\prime },\partial
\Omega \right) \right) $. {Using these estimates in (\ref{tta2}) and
absorbing into the left gives}
\begin{equation}
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{%
\mathcal{A}},w}\,w_{ij}^{\beta }\right\vert ^{2}\leq \mathcal{C}
_{3}\sum_{i,j=1}^{n}\left\Vert \xi w_{ij}^{\beta }\right\Vert _{L^{p}}^{2}+
\mathcal{C}_{3}. \label{tttc}
\end{equation}
{\ with
\begin{equation*}
\mathcal{C}_{3}=\mathcal{C}_{3}\left( M_0, n,B,\Lambda ,\vec{k},\mathcal{R},
\left\Vert \mathcal{A}\right \Vert _{\mathcal{C}^{3}\left( \Gamma ^{\prime
}\right) },\left\Vert f\right\Vert _{\mathcal{C}^{2}\left( \Gamma ^{\prime
}\right) },\left\Vert \vec{\gamma}\right\Vert _{\mathcal{C} ^{2}\left(
\Gamma ^{\prime }\right) },\Omega ,\mathop{\rm dist}\left( \Omega ^{\prime
},\partial \Omega \right) \right) ,
\end{equation*}%
where we have used Remark \ref{AdependsonR} to substitute the dependence on }%
$A$ by dependence on $\vec{k}$, $\mathcal{R}$, $M_{0}$.
Choosing {$\beta =1$ in (\ref{tta2}) and applying Lemma \ref{minkgrad} and (%
\ref{ttw}) gives}
\begin{eqnarray*}
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{%
\mathcal{A}},w}\,w_{ij}\right\vert ^{2} &\leq &\mathcal{C}_{1}
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}}
,w}\zeta \right\vert ^{2} w_{ij}^{2}+\mathcal{C}_{0}B^{2} \int_{2\mathcal{R}%
}\xi ^{2}\left\vert \nabla w\right\vert ^{6}+\mathcal{C}_{0}A^{2} \\
&\leq &\mathcal{C}_{1}C_{1}A^{2}\Lambda \sum_{i=1}^{n}\int_{2\mathcal{R}}\xi
^{2}\left\vert \nabla _{\!\!\sqrt{\mathcal{A}},w}w_{i}\right\vert ^{2} +%
\mathcal{C}_{3}.
\end{eqnarray*}
{Estimating the first term on the right by Theorem \ref{extraintegrability},
we then get}
\begin{equation}
\sum_{i,j=1}^{n}\int_{2\mathcal{R}}\zeta ^{2}\left\vert \nabla _{\!\! \sqrt{%
\mathcal{A}},w}\,w_{ij}\right\vert ^{2}\leq \mathcal{C}_{3}. \label{tttd}
\end{equation}
{To finish the proof, we iterate (\ref{tttc}) in a similar fashion as in the
proof of Theorem \ref{extraintegrability}, using (\ref{tttd}) to start the
iteration. We omit the details. As a result, we obtain
\begin{equation*}
\sum_{i,j=1}^{n}\int_{\Omega ^{\prime }}\left\vert \,w_{ij}\right\vert
^{n+1}\leq \mathcal{C}^{\ast }\left( M_0, n,B,\Lambda ,\vec{k},\left\Vert
\mathcal{A}\right\Vert _{\mathcal{C}^{3}\left( \Gamma ^{\prime }\right)
},\left\Vert f\right\Vert _{\mathcal{C}^{2}\left( \Gamma ^{\prime }\right)
},\left\Vert \vec{\gamma}\right\Vert _{\mathcal{C}^{2}\left( \Gamma ^{\prime
}\right) }, \mathop{\rm dist}\left( \Omega ^{\prime },\partial \Omega
\right) ,\Omega \right).
\end{equation*}
From the Sobolev embedding theorem it follows that
\begin{equation*}
\left\Vert \nabla w\right\Vert _{L^{\infty }\left( \Omega^{\prime} \right)
}\,\leq \mathcal{C}^{\ast },
\end{equation*}
as desired.}
It remains to prove that assumption (\ref{triple-xtra}) does not result in a
loss of generality.{\ This will be accomplished by a change of variables as
at the end of the proof of Theorem \ref{extraintegrability}. In fact,
letting }$u\left( x\right) =w\left( x\right)/N$, $u$ satisfies the equation
\begin{equation}
\mathop{\rm div}\widetilde{\mathcal{A}}\left( x,u\right) \nabla u+\widetilde{%
\vec{\gamma}}\left( x,u\right) \cdot \nabla u+\widetilde{f}\left( x,u\right)
=0 \label{eq2}
\end{equation}
in $\Omega $, where
\begin{equation*}
\widetilde{\mathcal{A}}\left( x,q\right) =\mathcal{A}\left( x,Nq\right)
,\qquad \widetilde{\vec{\gamma}}\left( x,q\right) =\vec{\gamma}\left(
x,Nq\right) ,\qquad \widetilde{f}\left( x,q\right) =\frac{1}{N}f\left(
x,Nq\right) .
\end{equation*}
Moreover,
\begin{equation*}
\widetilde{M}_{0}=\left\Vert u\right\Vert _{L^{\infty }\left( \Omega
^{\prime }\right) }=\frac{M_{0}}{N}.
\end{equation*}
Hence, for $N$ big enough, the analogue of (\ref{triple-xtra}) holds for $u$%
. The result for $w$ then follows from the identity $w=Nu$.{\
\endproof%
}
\section{{\label{section-Proof}Proof of the Hypoellipticity Theorem}}
{In this section we prove our main results Theorems \ref{application} and %
\ref{DP}. To do so, we will apply Theorem 15.19 of \cite{GiTr}. For easy
reference we now state a special version of this theorem suitable for our
needs. }
\begin{theorem}[Theorem 15.19 in \protect\cite{GiTr}]
{\ \label{th15.19} Let $\Omega $ be a bounded domain in $\mathbb{R}^{n}$
satisfying an exterior sphere condition at each point of $\partial \Omega $.
Let $\mathcal{M}$ be a divergence structure operator,
\begin{equation*}
\mathcal{M}\left( w\right) =\mathop{\rm div}\mathcal{S}\left( x,w,\nabla
w\right) +\mathcal{T}\left( x,w,\nabla w\right) ,
\end{equation*}
where $\mathcal{S}=\left( \mathcal{S}^{1},\mathcal{S}^{2},\dots ,\mathcal{S}
^{n}\right) $, $\mathcal{S}^{i}\in \mathcal{C}^{1+\delta }\left( \Omega
\times \mathbb{R}\times \mathbb{R}^{n}\right) $ for all $i$, $\mathcal{T} $ $%
\in \mathcal{C}^{\delta }\left( \Omega \times \mathbb{R}\times \mathbb{R}
^{n}\right) $ for some $0<\delta <1$, and where there exist positive
constants $a_{0}$, $b_{0}$, $b_{1}$, $c_{0}$ and $d_{0}$ such that for all $%
\xi \in \mathbb{R}^{n}$ and all $\left( x,z,p\right) \in \Omega \times
\mathbb{R}\times \mathbb{R}^{n}$, }
\begin{eqnarray}
\xi ^{t}\left\{ \nabla _{p}\mathcal{S}\left( x,z,p\right) \right\} \xi &\geq
&c_{0}\left\vert \xi \right\vert ^{2} \label{15.59-1} \\
\left\vert \nabla _{p}\mathcal{S}\left( x,z,p\right) \right\vert &\leq &d_{0}
\label{15.64-1} \\
\left( 1+\left\vert p\right\vert \right) \left\vert \partial _{z}\mathcal{S}
\right\vert +\left\vert \nabla _{x}\mathcal{S}\right\vert +\left\vert
\mathcal{T}\right\vert &\leq &d_{0}\left( 1+\left\vert p\right\vert \right)
^{2} \label{15.66-1} \\
p\cdot \mathcal{S}\left( x,z,p\right) &\geq &a_{0}\left\vert p\right\vert
^{2} \label{10.23-a} \\
\mathcal{T}\left( x,z,p\right) \mathop{\rm sign}z &\leq &b_{0}\left\vert
p\right\vert +b_{1}. \label{10.23-b}
\end{eqnarray}
{Then for any function $\varphi \in \mathcal{C}^{0}\left( \partial \Omega
\right) $, there exists a solution $w\in \mathcal{C}^{2+\delta }\left(
\Omega \right) \bigcap \mathcal{C}^{0}\left( \overline{\Omega}\right) $ of
the Dirichlet problem
\begin{eqnarray*}
\mathcal{M}\left( w\right) &=&0\qquad \text{in }\Omega \\
w &=&\varphi \qquad \text{on }\partial \Omega .
\end{eqnarray*}
}
\end{theorem}
\begin{remark}
{\ Theorem 15.19 in \cite{GiTr} is established only for $a_{0}=1$. Its
generalization to any positive constant $a_{0}$ is straightforward. See the
proof of Theorem 10.9 in \cite{GiTr} for details. }
\end{remark}
To prove Theorem \ref{DP}, we will apply Theorem \ref{th15.19} to a family
of truncations of the operator $\mathcal{Q}$ defined in (\ref{equation}).
For each $M>0$, let $\chi _{M}\in \mathcal{C}^{\infty }\left( \mathbb{R}%
\right) $ satisfy
\begin{equation}
\chi _{M}\left( z\right) =\left\{
\begin{array}{ll}
z & \qquad \text{if }\left\vert z\right\vert \leq M \\
& \\
\frac{3}{2}M & \qquad \text{if }z\geq 2M \\
& \\
-\frac{3}{2}M & \qquad \text{if }z<2M%
\end{array}%
\right. ,\qquad \left\vert \frac{d}{dz}\chi _{M}\left( z\right) \right\vert
\leq 1. \label{chi_M}
\end{equation}%
Define
\begin{equation*}
\mathcal{A}^{M}\left( x,z\right) =\mathcal{A}\left( x,\chi _{M}\left(
z\right) \right) ,\quad \vec{\gamma}^{M}\left( x,z\right) =\vec{\gamma}%
\left( x,\chi _{M}\left( z\right) \right) ,\quad f^{M}\left( x,z\right)
=f\left( x,\chi _{M}\left( z\right) \right) ,
\end{equation*}%
and set
\begin{equation*}
\mathcal{Q}^{M}w\left( x\right) =\mathop{\rm div}\mathcal{A}^{M}\left(
x,w\left( x\right) \right) \nabla w\left( x\right) +\vec{\gamma}^{M}\left(
x,w\left( x\right) \right) \cdot \nabla w\left( x\right) +f^{M}\left(
x,w\left( x\right) \right) .
\end{equation*}%
Note that if $\left\vert w\left( x\right) \right\vert \leq M$ then $\mathcal{%
Q}^{M}w\left( x\right) =\mathcal{Q}w\left( x\right) $.
\begin{proposition}
\label{proptouse} For each $\varepsilon ,M>0,${\ the} operators $\mathcal{Q}%
_{\varepsilon }^{M}=\mathcal{Q}^{M}+\varepsilon \mathbf{\Delta }$ satisfy
the hypothesis of Theorem \ref{th15.19} with{\
\begin{eqnarray*}
a_{0} &=&c_{0}=\varepsilon \\
b_{0} &=&\left\Vert \vec{\gamma}\right\Vert _{L^{\infty }\left( \tilde{\Gamma%
}\right) },\qquad b_{1}=\left\Vert f\right\Vert _{L^{\infty }\left( \tilde{%
\Gamma}\right) } \\
d_{0} &=&\left\Vert \nabla _{\left( x,z\right) }\mathcal{A}\right\Vert
_{L^{\infty }\left( \tilde{\Gamma}\right) }+\left\Vert \vec{\gamma}%
\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) }+\left\Vert
f\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) }+\varepsilon ,
\end{eqnarray*}%
where $\tilde{\Gamma}=\Omega \times \left[ -\frac{3}{2}M,\frac{3}{2}M\right]
.$ Moreover, since $\mathcal{A}^{M}$, $\vec{\gamma}^{M}$ and $f^{M}$ are
smooth functions, the value of ${\delta }$ in Theorem \ref{th15.19} can be
any value $0<\delta <1$}.
\end{proposition}
\begin{proof}
With the notation of Theorem \ref{th15.19}, and $\mathcal{A}\left(
x,z\right) =\left\{ a^{ij}\left( x,z\right) \right\} _{i,j=1}^{n}$, we have
\begin{equation*}
\mathcal{Q}_{\varepsilon}^{M}\left( w\right) = \mathcal{M}\left( w\right) =%
\mathop{\rm div}\mathcal{S}\left( x,w,\nabla w\right) +\mathcal{T}\left(
x,w,\nabla w\right)
\end{equation*}
with
\begin{eqnarray}
\mathcal{S}^{i}\left( x,z,p\right) &=&\sum_{j=1}^{n}\left( a^{ij}\left(
x,\chi _{M}\left( z\right) \right) +\varepsilon \delta _{ij}\right)
p_{j},\qquad i=1,\dots ,n, \label{Sform} \\
\mathcal{T}\left( x,z,p\right) &=&\vec{\gamma}\left( x,\chi _{M}\left(
z\right) \right) \cdot p+f\left( x,\chi _{M}\left( z\right) \right) .
\label{Tform}
\end{eqnarray}
Here $\delta _{ij}=1$ if $i=j$ and $\delta _{ij}=0$ otherwise. Let us now
verify (\ref{15.64-1})--(\ref{10.23-b}). \newline
\textbf{(\ref{15.59-1}).} By (\ref{Sform}),
\begin{eqnarray*}
\xi ^{t}\left\{ \nabla _{p}\mathcal{S}\left( x,z,p\right) \right\} \xi
&=&\sum_{i,\ell =1}^{n}\partial _{p_{\ell }}\left( \sum_{j=1}^{n}\left(
a^{ij}\left( x,\chi _{M}\left( z\right) \right) + \varepsilon \delta _{ij}
\right) p_{j}\right) \xi _{\ell }\xi _{i} \\
&=&\sum_{i,\ell =1}^{n}\left( a^{i\ell }\left( x,\chi _{M}\left( z\right)
\right) + \varepsilon \delta _{i\ell }\right) \xi _{\ell }\xi _{i} \\
&=&\xi ^{t}\mathcal{A}\left( x,\chi _{M}\left( z\right) \right) \xi
+\varepsilon \left\vert \xi \right\vert ^{2} \\
&\geq &\varepsilon \left\vert \xi \right\vert ^{2}.
\end{eqnarray*}
Hence, (\ref{15.59-1}) holds with $c_{0}=\varepsilon $ . \newline
\textbf{(\ref{15.64-1}).} From (\ref{Sform}),
\begin{eqnarray*}
\left\vert \partial _{p_{j}}\mathcal{S}^{i}\left( x,z,p\right) \right\vert
&=&\left\vert a^{ij}\left( x,\chi _{M}\left( z\right) \right) +\varepsilon
\delta _{ij}\right\vert \\
&\leq &\left\vert a^{ij}\left( x,\chi _{M}\left( z\right) \right)
\right\vert +\varepsilon \\
&\leq &\left\Vert \mathcal{A}\right\Vert _{L^{\infty }\left( \tilde{\Gamma}
\right) }+\varepsilon ,
\end{eqnarray*}
where $\tilde{\Gamma}=\Omega \times \left[ -\frac{3}{2}M,\frac{3}{2}M \right]
$. Thus, (\ref{15.64-1}) holds with $d_{0}=\left\Vert \mathcal{A}
\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) }+\varepsilon $.
\newline
\textbf{(\ref{15.66-1}).} From (\ref{Sform}) and (\ref{Tform}),
\begin{eqnarray*}
&&\left( 1+\left\vert p\right\vert \right) \left\vert \partial _{z}\mathcal{%
S }\right\vert +\left\vert \nabla _{x}\mathcal{S}\right\vert +\left\vert
\mathcal{T}\right\vert \\
&=&\left( 1+\left\vert p\right\vert \right) \left\vert \tfrac{d}{dz}\chi
_{M}\left( z\right) \right\vert \left\vert \mathcal{A}_{z}\left( x,\chi
_{M}\left( z\right) \right) p\right\vert +\left\vert \nabla _{x}\mathcal{A}
\left( x,\chi _{M}\left( z\right) \right)p \right\vert \\
&&+\left\vert \vec{\gamma}\left( x,\chi _{M}\left( z\right) \right) \cdot
p+f\left( x,\chi _{M}\left( z\right) \right) \right\vert \\
&\leq &\left( \left\Vert \nabla _{\left( x,z\right) }\mathcal{A}\right\Vert
_{L^{\infty }\left( \tilde{\Gamma}\right) }+\left\Vert \vec{\gamma}
\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) }+\left\Vert
f\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) }\right) \left(
1+\left\vert p\right\vert \right) ^{2},
\end{eqnarray*}
where we used that $\left\vert \frac{d}{dz}\chi _{M}\left( z\right)
\right\vert \leq 1$. Thus (\ref{15.66-1}) holds with
\begin{equation*}
d_{0}=\left\Vert \nabla _{\left( x,z\right) }\mathcal{A}\right\Vert
_{L^{\infty }\left( \tilde{\Gamma}\right) }+\left\Vert \vec{\gamma}
\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) }+\left\Vert
f\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) }.
\end{equation*}
\textbf{(\ref{10.23-a}).} By (\ref{Sform}) it follows that
\begin{equation*}
p\cdot \mathcal{S}\left( x,z,p\right) =p\cdot \left( \mathcal{A}\left(
x,\chi _{M}\left( z\right) \right) +\mathbf{I}\varepsilon \right) p\geq
\varepsilon \left\vert p\right\vert ^{2}.
\end{equation*}
Thus (\ref{10.23-a}) holds with $a_{0}=\varepsilon $.\newline
\textbf{(\ref{10.23-b}).} By (\ref{Tform}),
\begin{eqnarray*}
\mathcal{T}\left( x,z,p\right) \,\mathop{\rm sign}z &=&\left( \vec{ \gamma}%
\left( x,\chi _{M}\left( z\right) \right) \cdot p\,+f\left( x,\chi
_{M}\left( z\right) \right) \right) \,\mathop{\rm sign}z \\
&\leq &\left\Vert \vec{\gamma}\right\Vert _{L^{\infty }\left( \tilde{\Gamma}
\right) }~\left\vert p\right\vert +\left\Vert f\right\Vert _{L^{\infty
}\left( \tilde{\Gamma}\right) }.
\end{eqnarray*}
\newline
Hence (\ref{10.23-b}) holds with $b_{0}=\left\Vert \vec{\gamma}\right\Vert
_{L^{\infty }\left( \tilde{\Gamma}\right) }$ and $b_{1}=\left\Vert
f\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) }$.
\end{proof}
\subsection{Proof of Theorem \protect\ref{DP}}
Let $\Omega \Subset \widetilde{\Omega }$ be a strongly convex domain as in
the hypotheses of Theorem \ref{DP}. Then $\Omega $ satisfies an exterior {%
sphere condition at each point of $\partial \Omega $}. Given a continuous
function $\varphi $ on $\partial \Omega $, let $M_{0}=\sup_{\partial \Omega
}\left\vert \varphi \right\vert $. {By Theorem \ref{th15.19} and Proposition %
\ref{proptouse} with $M=M_{0}$, for all $\varepsilon >0$ there exists $%
w^{\varepsilon }\in \mathcal{C}^{2+\delta }\left( \Omega \right) \bigcap
\mathcal{C}^{0}\left( \overline{{\Omega }}\right) $ , $0<\delta <1$, such
that $w^{\varepsilon }$ is a solution of the Dirichlet problem}
\begin{equation*}
\left\{
\begin{array}{rcll}
\mathcal{Q}_{\varepsilon }^{M_{0}}w & = & 0\quad & \text{in }\Omega \\
w & = & \varphi & \text{on }\partial \Omega .%
\end{array}%
\right.
\end{equation*}%
{\ The solution $w^{\varepsilon }$ also depends on $M_{0}$, which is fixed.
The smoothness assumptions on $\mathcal{A}$, $\vec{\gamma}$ and $f$ and a
standard bootstrapping argument (see, e.g., Theorems 6.2 and 6.3 in \cite{Ev}%
) imply that }$w^{\varepsilon }${$\in \mathcal{\ C}^{\infty }\left( \Omega
\right) \bigcap \mathcal{C}^{0}\left( \overline{\Omega }\right) $. Since the
operators $\mathcal{Q}_{\varepsilon }^{M_{0}}$ satisfy the hypotheses of
Theorem \ref{maxp} in the Appendix (the maximum principle), we have
\begin{equation}
\left\Vert w^{\varepsilon }\right\Vert _{L^{\infty }\left( \Omega \right)
}\leq \left\Vert w^{\varepsilon }\right\Vert _{L^{\infty }\left( \partial
\Omega \right) }=\left\Vert \varphi \right\Vert _{L^{\infty }\left( \partial
\Omega \right) }=M_{0}. \label{mp000}
\end{equation}%
Thus the functions }$w^{\varepsilon }${\ are uniformly bounded by $M_{0}$ in
}$\overline{{\Omega }}${\ for all $\varepsilon >0$. From the definition of }$%
\mathcal{Q}_{\varepsilon }^{M_{0}}$ it follows that{\ } $\mathcal{Q}%
w^{\varepsilon }+\varepsilon \triangle w^{\varepsilon }=\mathcal{Q}%
_{\varepsilon }^{M_{0}}w^{\varepsilon }=0$ in $\Omega $.
{By (\ref{hellip}), the coefficients of $\mathcal{Q}+\varepsilon \mathbf{I}$%
, namely the entries of $\mathcal{A}_{\varepsilon }=\mathcal{A+\varepsilon }%
\mathbf{I}$, satisfy
\begin{equation*}
\sum_{i=1}^{n}\left( k^{i}\left( x,z\right) +\varepsilon \right) \xi
_{i}^{2}\leq \xi ^{t}\mathcal{A}_{\varepsilon }\left( x,z\right) \xi \leq
\Lambda \sum_{i=1}^{n}\left( k^{i}\left( x,z\right) +\varepsilon \right) \xi
_{i}^{2}
\end{equation*}%
for all $\xi \in \mathbb{R}^{n}$ and $\left( x,z\right) \in \Gamma =\Omega
\times \mathbb{R}$. That is, $\mathcal{A}_{\varepsilon }$} satisfies the
diagonal condition for diagonal entries $k^{i}+\varepsilon $. Next, by (\ref%
{wirtm}), {\
\begin{equation*}
\sum_{i=1}^{n}\left\vert \partial _{i}\mathcal{A}_{\varepsilon }\left(
x,z\right) \xi \right\vert ^{2}+\left\vert \partial _{z}\mathcal{A}%
_{\varepsilon }\left( x,z\right) \xi \right\vert ^{2}\leq B_{\mathcal{A}%
}^{2}~\xi ^{t}\mathcal{A}\xi \leq B_{\mathcal{A}}^{2}~\xi ^{t}\mathcal{A}%
_{\varepsilon }\xi
\end{equation*}%
for all $\xi \in \mathbb{R}^{n}$, $\left( x,z\right) \in \Gamma
_{M_{0}}^{\prime }$. Hence }$\mathcal{A}_{\varepsilon }$ is subordinate,
with the same constant\ $B_{\mathcal{A}}$ as for $\mathcal{A}$ in $\Gamma
_{M_{0}}^{\prime }$. We also have by (\ref{xtra}) and (\ref{xxtra}) that {\
\begin{eqnarray*}
\left\vert \partial _{z}\mathcal{A}_{\varepsilon }\left( x,z\right) \xi
\right\vert ^{2} &=&\left\vert \partial _{z}\mathcal{A}\left( x,z\right) \xi
\right\vert ^{2}\leq B_{\mathcal{A}}^{2}~k^{\ast }\left( x,z\right) \,\xi
^{t}\mathcal{A}\left( x,z\right) \xi \\
&\leq &CB_{\mathcal{A}}^{2}~\left( k^{\ast }\left( x,z\right) +\varepsilon
\right) \,\xi ^{t}\mathcal{A}_{\varepsilon }\left( x,z\right) \xi
\end{eqnarray*}%
and}
\begin{eqnarray*}
\sum_{i=1}^{n}\left\vert \partial _{i}\partial _{z}\mathcal{A}_{\varepsilon
}\left( x,z\right) \xi \right\vert ^{2}+\left\vert \partial _{z}^{2}\mathcal{%
\ A}_{\varepsilon }\left( x,z\right) \xi \right\vert ^{2} &\leq &\left( B_{%
\mathcal{A}}^{\prime }\right) ^{2}\,\xi ^{t}\mathcal{A}\left( x,z\right) \xi
\\
&\leq &\left( B_{\mathcal{A}}^{\prime }\right) ^{2}\,\xi ^{t}\mathcal{A}%
_{\varepsilon }\left( x,z\right) \xi
\end{eqnarray*}%
{for all $\xi \in \mathbb{R}^{n}$, $\left( x,z\right) \in \Gamma
_{M_{0}}^{\prime }$.}
Thus $\mathcal{A}_{\varepsilon }$ satisfies the super subordination
condition with the same constants as $\mathcal{A}$. Hence $\mathcal{Q}%
+\varepsilon $ satisfies the hypotheses of Theorem \ref{apriorial} uniformly
in $\varepsilon $, $0\leq \varepsilon \leq 1${. Applying Theorem \ref%
{apriorial} to the solutions $w^{\varepsilon }$, it follows that for any
multi-index $\vec{\alpha}$ of nonnegative integers, the family $\left\{ D^{%
\vec{\alpha}}w^{\varepsilon },0<\varepsilon \leq 1\right\} $ is
equicontinuous and uniformly bounded in any subdomain }${\Omega }${$^{\prime
}\Subset $}${\Omega }${. By the Arzela-Ascoli theorem, there is a
subsequence $\left\{ w^{\varepsilon _{i}}\right\} $ with $\varepsilon
_{i}\rightarrow 0$ which converges in $\mathcal{C}^{\infty }\left( \Omega
\right) $ to a function $w^{0}\in \mathcal{C}^{\infty }\left( \Omega \right)
$, i.e., $D^{\vec{\alpha}}w^{\varepsilon _{i}}\,$converges to $D^{\vec{\alpha%
}}w^{0}\,$uniformly on compact subsets of $\Omega $, for all multi-indexes $%
\vec{\alpha}$. We will show that $w^{0}$ is a solution of the Dirichlet
problem (\ref{dirchlet}). }
{Since $w^{0}$ and all its derivatives are uniform limits of $w^{\varepsilon
_{i}}$ and their corresponding derivatives in compact subsets of $\Omega $,
then $\left\vert \bigtriangleup w^{0}\right\vert <\infty $ in $\Omega $, and
for all $x\in \Omega $,
\begin{eqnarray*}
\mathcal{Q}w^{0}\left( x\right) &=&\mathop{\rm div}\mathcal{A}\left(
x,w^{0}\right) \nabla w^{0}+\vec{\gamma}\left( x,w^{0}\right) \cdot \nabla
w^{0}+f\left( x,w^{0}\right) \\
&=&\lim_{i\rightarrow \infty }\mathop{\rm div}\mathcal{A}\left(
x,w^{\varepsilon _{i}}\right) \nabla w^{\varepsilon _{i}}+\vec{\gamma}\left(
x,w^{\varepsilon _{i}}\right) \cdot \nabla w^{\varepsilon _{i}}+f\left(
x,w^{\varepsilon _{i}}\right) \\
&=&\lim_{i\rightarrow \infty }\mathcal{Q}_{\varepsilon
^{i}}^{M_{0}}w^{\varepsilon _{i}}\left( x\right) -\varepsilon
_{i}\bigtriangleup w^{\varepsilon _{i}}\left( x\right) \\
&=&-\varepsilon _{i}\lim_{i\rightarrow \infty }\bigtriangleup w^{\varepsilon
_{i}}\left( x\right) =0.
\end{eqnarray*}%
Therefore }$w^{0}\in \mathcal{C}^{\infty }\left( \Omega \right) $ is a
strong solution of the differential equation in the Dirichlet problem (\ref%
{dirchlet}). Define $w^{0}=\varphi $ on $\partial \Omega $ and recall that $%
w^{\varepsilon }=\varphi $ on $\partial \Omega $ if $\varepsilon >0$. To
finish the proof of the theorem we must check that $w^{0}\in \mathcal{C}^{0}(%
\overline{\Omega })$.
Let $\omega \left( r\right) $ be the modulus of continuity of ${\varphi }$
on $\partial \Omega $:
\begin{equation*}
\omega \left( r\right) =\sup_{x,y\in \partial \Omega ,\left\vert
x-y\right\vert \leq r}\left\vert {\varphi }\left( x\right) -{\varphi }\left(
y\right) \right\vert .
\end{equation*}%
Then $\omega $ is continuous and $\omega \left( 0\right) =0$. By Lemma \ref%
{concave-majorant}, taking a bigger $\omega $ if necessary, we may assume
that $\omega $ is also concave and strictly increasing in $\left[ 0,%
\mathop{\rm diam}\Omega \right] $, and $\mathcal{C}^{2}$ in $\left( 0,%
\mathop{\rm diam}\Omega \right] $.
By our hypothesis on $\gamma $, there exists $\eta _{0}>0$ such that $\gamma
\left( x,z\right) =0$ if $x\in \Omega $, $\mathop{\rm dist}\left( x,\partial
\Omega \right) <\eta _{0}$ and $\left\vert z\right\vert \leq M_{0}$. {Let $%
x_{0}$ be an arbitrary point on $\partial \Omega $, and let $h\left(
x\right) $ be the barrier function for $\omega $ at $x_{0}$ provided by
Lemma \ref{barrier} in the Appendix, with $\Phi $ and $\Omega $ there chosen
to be $\Omega $ and $\tilde{\Omega}$ respectively, and with $\nu =2M_{0}$, }$%
m_{0}=2M_{0}$, $\eta =\eta _{0}$ and $K=\left\Vert f\right\Vert _{L^{\infty
}\left( \tilde{\Gamma}\right) }$, where ${\tilde{\Gamma}}${$=\overline{%
\Omega }\times \left[ -2M_{0},2M_{0}\right] $. {Thus } there is a
neighborhood $\mathcal{N}$ of $x_{0}$ with $\mathcal{N}\subset
\{|x-x_{0}|<\eta _{0}\}$ and a function $h\in \mathcal{C}^{\infty }\left(
\mathcal{N}\right) \bigcap \mathcal{C}^{0}\left( \overline{\mathcal{N}}%
\right) $ such that\
\begin{eqnarray}
h\left( x\right) &\leq &-\omega \left( \left\vert x-x_{0}\right\vert \right)
, \label{mp001} \\
\mathop{\rm div}\mathcal{A}\left( x,h\left( x\right) +m\right) \nabla h
&\geq &\left\Vert f\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) },
\label{mp002} \\
\bigtriangleup h &=&\sum_{i=1}^{n}\partial _{i}^{2}h>0 \label{mp003}
\end{eqnarray}%
for all $x\in \Omega \bigcap \mathcal{N}$ and }$\left\vert {m}\right\vert
\leq 2M_{0}${. Moreover,
\begin{eqnarray}
h\left( x\right) &\leq &-2M_{0}\text{ \qquad if }x\in \partial \mathcal{N}%
\bigcap \Omega , \label{mp004} \\
h\left( x_{0}\right) &=&0. \label{mp005}
\end{eqnarray}%
Now, by (\ref{mp001}) and the continuity of $h$ on $\overline{\mathcal{N}}$,
\begin{equation}
h\left( x\right) \leq -\omega \left( \left\vert x-x_{0}\right\vert \right)
\leq {\varphi }\left( x\right) -{\varphi }\left( x_{0}\right)
=w^{\varepsilon }\left( x\right) -{\varphi }\left( x_{0}\right) \qquad \text{%
if }x\in \overline{\mathcal{N}}\bigcap \partial \Omega ,\varepsilon >0.
\label{out}
\end{equation}%
By (\ref{mp004}) and (\ref{mp000}),
\begin{equation*}
h\left( x\right) \leq -2M_{0}\leq w^{\varepsilon }\left( x\right) -{\varphi }%
\left( x_{0}\right) ,\qquad \text{if }x\in \partial \mathcal{N}\bigcap
\Omega ,\varepsilon >0.
\end{equation*}%
Therefore,
\begin{equation}
h\left( x\right) +\varphi \left( x_{0}\right) \leq w^{\varepsilon }\left(
x\right) \qquad \text{if }x\in \partial \mathcal{N},\varepsilon >0.
\label{bdryh}
\end{equation}%
\ On the other hand, since $w^{\varepsilon }$ is a solution of $\mathcal{Q}%
_{\varepsilon }^{M_{0}}w^{\varepsilon }=0$ and }$\mathcal{N}\subset \left\{
\left\vert x-x_{0}\right\vert <\eta _{0}\right\} $, {\
\begin{equation*}
\mathop{\rm div}\mathcal{A}^{\varepsilon }\left( x,w^{\varepsilon }\right)
\nabla w^{\varepsilon }=-f\left( x,w^{\varepsilon }\right) \leq \left\Vert
f\right\Vert _{L^{\infty }\left( \tilde{\Gamma}\right) }\qquad \text{in }%
\mathcal{N}\bigcap \Omega ,
\end{equation*}%
where the last inequality follows from (\ref{mp000}). Thus, letting $%
\mathcal{L}_{\varepsilon }$ be the quasilinear operator }$\mathcal{L}%
_{\varepsilon }=\mathop{\rm div}\mathcal{A}^{\varepsilon }\left( x,\cdot
\right) \nabla $, we have by (\ref{mp002}) and (\ref{bdryh}) that{\
\begin{eqnarray*}
\mathcal{L}_{\varepsilon }\left( h+\varphi \left( x_{0}\right) \right) &\geq
&\mathcal{L}_{\varepsilon }w^{\varepsilon }\qquad \text{in }\mathcal{N}%
\bigcap \Omega , \\
h\left( x\right) +\varphi \left( x_{0}\right) &<&w^{\varepsilon }\left(
x\right) \qquad \text{if }x\in \partial \left( \mathcal{N}\bigcap \Omega
\right) .
\end{eqnarray*}%
From the comparison principle Lemma \ref{comparison} applied to $\mathcal{L}%
_{\varepsilon }$ and the functions $w^{\varepsilon }$, $h+\varphi \left(
x_{0}\right) $ in $\mathcal{N}\bigcap \Omega $, we obtain
\begin{equation}
h\left( x\right) \leq w^{\varepsilon }\left( x\right) -\varphi \left(
x_{0}\right) \qquad \text{if }x\in \mathcal{N}\bigcap \Omega .
\label{mpalleps}
\end{equation}%
Since $h$ is continuous in $\overline{\mathcal{N}}$ and $h\left(
x_{0}\right) =0$, given any $\sigma >0$, there exists $\delta _{0}>0$
independent of $\varepsilon $ such that
\begin{equation*}
-\sigma <w^{\varepsilon }\left( x\right) -\varphi \left( x_{0}\right) \qquad
\text{if }x\in \mathcal{N}\bigcap \left\{ \left\vert x-x_{0}\right\vert
<\delta _{0}\right\} \bigcap \Omega ,\quad \varepsilon >0.
\end{equation*}%
Proceeding in a similar fashion for the function $\varphi \left(
x_{0}\right) -w^{\varepsilon }\left( x\right) $, we obtain
\begin{equation}
\left\vert w^{\varepsilon }\left( x\right) -\varphi \left( x_{0}\right)
\right\vert <\sigma \qquad \text{if }x\in \mathcal{N}\bigcap \left\{
\left\vert x-x_{0}\right\vert <\delta _{0}\right\} \bigcap \Omega ,\quad
\varepsilon >0. \label{close}
\end{equation}%
}
Let us now show that $w^{0}$ is continuous on $\overline{\Omega }$. Suppose
not, and let $x_{0}$ be a point of discontinuity of $w^{0}$ in $\overline{%
\Omega }$. Since $w^{0}$ is smooth in $\Omega $, $x_{0}$ must lie on $%
\partial \Omega $. Since $w^{0}=\varphi $ on $\partial \Omega $ and $\varphi
$ is continuous by hypothesis, there exist points $\{x_{k}\}_{k=1}^{\infty
}\subset \Omega $ and $\tilde{\sigma}>0$ such that $x_{k}\rightarrow x_{0}$
and $|w^{0}(x_{k})-\varphi (x_{0})|\geq \tilde{\sigma}$ for all $k$. By (\ref%
{close}) with $\sigma =\tilde{\sigma}/2$, there exists $\delta >0$
independent of $\varepsilon $ such that $|w^{\varepsilon }(x)-\varphi
(x_{0})|\leq \tilde{\sigma}/2$ if $x\in \Omega $ and $|x-x_{0}|<\delta $.
Choose $x=x_{k_{0}}$ for $k_{0}$ so large that $|x_{k_{0}}-x_{0}|<\delta $.
Then $|w^{\varepsilon }(x_{k_{0}})-\varphi (x_{0})|\leq \tilde{\sigma}/2$
for all $\varepsilon $. However, $w^{\varepsilon }(x_{k_{0}})\rightarrow
w^{0}(x_{k_{0}})$ as $\varepsilon \rightarrow 0$, so $|w^{0}(x_{k_{0}})-%
\varphi (x_{0})|\leq \tilde{\sigma}/2$, which is a contradiction. Hence $%
w^{0}\in \mathcal{C}^{\infty }\left( \Omega \right) \bigcap \mathcal{C}%
^{0}\left( \overline{\Omega }\right) $ and $w^{0}$ is a solution to the
Dirichlet problem (\ref{dirchlet}).
When $\gamma \equiv 0$, uniqueness follows by Lemma \ref{comparison} in the
Appendix. This finishes the proof of Theorem \ref{DP}. \endproof
\subsection{Proof of Theorem \protect\ref{application}}
Under the hypotheses of Theorem \ref{application}, let $w$ be a continuous
weak solution of
\begin{equation*}
\mathop{\rm div}\mathcal{A}\left( x,w\right) \nabla w+f\left( x,w\right)
=0\qquad \text{in }\Omega .
\end{equation*}%
Given $\bar{x}\in \Omega $, let $\Phi $ be the ball centered at $\bar{x}$
with radius $r=\frac{1}{2}\mathop{\rm dist}\left( \bar{x},\partial \Omega
\right) $. Then $\Phi $ is strongly convex. By Theorem \ref{DP}, there is a
continuous strong solution $u$ of the Dirichlet problem
\begin{equation*}
\left\{
\begin{array}{rcll}
\mathcal{Q}u & = & 0\quad & \text{in }\Phi \\
u & = & w & \text{on }\partial \Phi .%
\end{array}%
\right.
\end{equation*}%
{Moreover, }$u\in \mathcal{C}^{0}\left( \overline{\Phi }\right) \bigcap
\mathcal{C}^{\infty }\left( \Phi \right) $. By restricting $\bar{x}$ to a
compact set $\Omega ^{\prime }\subset \Omega $, the convex character $%
\lambda _{0}$ of $\Phi $ is bounded below away from zero, the bound
depending on $\mathop{\rm dist}(\Omega ^{\prime },\partial \Omega )$, and
hence the constants $\mathcal{C}_{N}$ controlling the derivatives are
independent of $\lambda _{0}$. By the uniqueness part of the comparison
principle, Lemma \ref{comparison}, it follows that $u=w$ in $\overline{\Phi }
$ and therefore $w$ is smooth inside $\Phi $ with control on all its
derivatives in compact subsets of $\Phi $ . {\ This finishes the proof of
Theorem \ref{application}.\endproof}
\section{Appendix}
This Appendix is divided into four subsections in which we give some
technical details about facts that we used earlier: degenerate Sobolev
spaces and weak solutions, a maximum principle, a comparison principle, and
barriers for the Dirichlet problem.
\label{deg-sob-section}
\subsection{Degenerate Sobolev Spaces and Weak Solutions}
\subsubsection{The weak degenerate Sobolev space $H_{\mathcal{X}%
}^{1,2}(\Omega) \label{Section-Weak-Sol}$}
We first describe the degenerate Sobolev spaces used in the paper, beginning
with a standard definition.
\begin{definition}[Weak $X$ derivative]
Let $X$ be a locally Lipschitz vector field on $\Omega $, i.e., $X=\mathbf{v}%
\cdot \nabla $ with $\mathbf{v}\in \mathop{\rm Lip_{loc}}\left( \Omega
\right) $, the class of locally Lipschitz continuous $\mathbb{R}^{n}-$valued
functions on $\Omega $. $X$ is initially defined on real-valued functions $%
w\in \mathop{\rm Lip_{loc}}(\Omega )$ by $Xw=\mathbf{v}\cdot \nabla w$. We
say that a locally integrable function $g$ is the weak derivative $Xw$ of a
locally integrable function $w$ if
\begin{equation}
\int_{\Omega }g\varphi =-\int_{\Omega }wX^{\prime }\varphi =-\int_{\Omega
}w\nabla \cdot \left( \mathbf{v}\varphi \right) \quad \text{for all $\varphi
\in \mathop{\rm Lip_{0}}\left( \Omega \right) $.} \label{weakdef}
\end{equation}
\end{definition}
The weak derivative $Xw$ is clearly unique if it exists, and $Xw$ exists and
coincides with $\mathbf{v}\cdot \nabla w$ if $w\in \mathop{\rm Lip_{loc}}%
\left( \Omega \right) $.
\begin{definition}[Weak degenerate Sobolev space]
Let $\mathcal{X}=\left\{ X_{j}\right\} _{j=1}^{m}$ where $X_{j}=\mathbf{v}_j$
are $\mathop{\rm Lip_{loc}}(\Omega)$ vector fields on $\Omega \subset
\mathbb{R}^{n}.$ The degenerate Sobolev space $H_{\mathcal{X}}^{1,2}\left(
\Omega \right) $ is defined as the inner product space consisting of all $%
w\in L^{2}\left( \Omega \right) $ whose weak derivatives $X_{j}w$ are also
in $L^{2}\left( \Omega \right) $. The inner product in $H_{\mathcal{X}%
}^{1,2}\left( \Omega \right) $ is defined by
\begin{equation}
\left\langle w,v\right\rangle _{\mathcal{X}}=\int_{\Omega
}wv~dx+\int_{\Omega }\mathcal{X}w\cdot \mathcal{X}v~dx, \label{innpdt}
\end{equation}
where we denote $\mathcal{X}w= \left(X_1w,\dots,X_mw\right)$, and the norm
is
\begin{equation*}
\left\Vert w\right\Vert _{H_{\mathcal{X}}^{1,2}\left( \Omega \right)
}=\left\langle w,w\right\rangle _{\mathcal{X}}^{1/2} = \left(\left\Vert
w\right\Vert _{L^{2}\left( \Omega \right) }^{2}+\left\Vert \mathcal{X}%
w\right\Vert _{L^{2}\left( \Omega \right) }^{2}\right)^{1/2}.
\end{equation*}
\end{definition}
We now define what we mean by $\nabla w$ if $w \in H_{\mathcal{X}%
}^{1,2}\left( \Omega \right) $ for a collection $\mathcal{X} =
\{X_j\}_{j=1}^m = \{\mathbf{v}_j\cdot \nabla\}_{j=1}^m$ of $%
\mathop{\rm
Lip_{loc}}(\Omega)$ vector fields. For such $w$ and $\mathcal{X}$, there is
a sequence $\{w_k\}_{k=1}^\infty$ of $\mathop{\rm Lip}(\Omega)$ functions
and a vector $\vec{W}(x) \in \mathbb{R}^n$ satisfying $\mathbf{v}_j\cdot
\vec{W} \in L^2(\Omega)$ for all $j$ and
\begin{equation} \label{density}
||w_k -w||_{L^2(\Omega)} + \sum_j ||X_jw_k -\mathbf{v}_j\cdot \vec{W}
||_{L^2(\Omega)} \rightarrow 0\quad \text{as $k \to \infty$}.
\end{equation}
This is proved in \cite{SaW3} (see also \cite{FSS}, \cite{GN}) in case all $%
\mathbf{v}_j \in \mathop{\rm Lip}(\Omega)$ but remains true if all $\mathbf{v%
}_j \in \mathop{\rm Lip_{loc}}(\Omega)$ by examining the proof in \cite{SaW3}%
.
Then
\begin{equation}
\mathcal{X}w=\left( X_{1}w,\dots ,X_{m}w\right) =\left( \mathbf{v}_{1}\cdot
\vec{W},\dots \mathbf{v}_{m}\cdot \vec{W}\right) \label{Xform}
\end{equation}%
since for all $\varphi \in \mathop{\rm Lip_{0}}\left( \Omega \right) $,
\begin{eqnarray*}
\int_{\Omega }w\,\nabla \cdot (\mathbf{v}_{j}\varphi ) &=&\lim_{k\rightarrow
\infty }\int_{\Omega }w_{k}\nabla \cdot (\mathbf{v}_{j}\varphi ) \\
&=&\lim_{k\rightarrow \infty }\int_{\Omega }(\mathbf{v}_{j}\cdot \nabla
w_{k})\varphi =\lim_{k\rightarrow \infty }\int_{\Omega }(X_{j}w_{k})\varphi
\\
&=&-\int_{\Omega }\left( \mathbf{v}_{j}\cdot \vec{W}\right) \varphi .
\end{eqnarray*}%
Moreover, if $\{w_{k}^{\prime }\}$ and $\vec{W^{\prime }}$ are another such
sequence and vector for the same $w$, it follows similarly that $%
\int_{\Omega }w\,\nabla \cdot (\mathbf{v}_{j}\varphi )=-\int_{\Omega }\left(
\mathbf{v}_{j}\cdot \vec{W^{\prime }}\right) \varphi $. Hence
\begin{equation*}
\int_{\Omega }\left( \mathbf{v}_{j}\cdot \vec{W}\right) \varphi
=\int_{\Omega }\left( \mathbf{v}_{j}\cdot \vec{W^{\prime }}\right) \varphi
\end{equation*}%
for all $\varphi \in \mathop{\rm Lip_{0}}\left( \Omega \right) $, so that
\begin{equation}
\mathbf{v}_{j}\cdot \vec{W}=\mathbf{v}_{j}\cdot \vec{W^{\prime }}\quad \text{%
for all $j$}. \label{uniquegrad}
\end{equation}%
In this sense, $\vec{W}$ is unique, i.e., $\vec{W}$ is uniquely determined
by $w$ up to its dot product with each vector $\mathbf{v}_{j}$. We will
often abuse notation by writing $\vec{W}=\nabla w$. Any particular $\vec{W}$
as above with be called a \emph{representative} of $\nabla w$. Then $X_{j}w=%
\mathbf{v}_{j}\cdot \nabla w,j=1,\dots ,m,$ for all $w\in H_{\mathcal{X}%
}^{1,2}\left( \Omega \right) .$ Furthermore, by (\ref{density}), the
sequence $\{w_{k}\}$ above satisfies
\begin{equation*}
||w_{k}-w||_{H_{\mathcal{X}}^{1,2}\left( \Omega \right) }\rightarrow 0\quad
\text{as $k\rightarrow \infty $.}
\end{equation*}
Suppose that $\Omega ^{\prime }\subset \Omega $ and $M>0$, and let $\mathcal{%
X}=\{X_{j}\}=\{\mathbf{v}_{j}\cdot \nabla \}$ and $H_{\mathcal{X}%
,0}^{1,2}\left( \Omega \right) $ be as above. We claim that if $\mathcal{A}%
(x,z)$ and $\vec{\gamma}(x,z)$ satisfy
\begin{equation}
\xi \cdot \mathcal{A}(x,z)\xi \leq c\sum_{j}(\mathbf{v}_{j}(x)\cdot \xi
)^{2}\quad \text{and}\quad (\vec{\gamma}(x,z)\cdot \xi )^{2}\leq c\sum_{j}(%
\mathbf{v}_{j}(x)\cdot \xi )^{2} \label{subu}
\end{equation}%
for all $(x,z,\xi )\in \Omega ^{\prime }\times (-M,M)\times \mathbb{R}^{n}$,
then $\sqrt{\mathcal{A}(x,z)}\nabla w$ and $\vec{\gamma}(x,z)\cdot \nabla w$
are well-defined for any $(x,z)\in \Omega ^{\prime }\times (-M,M)$ and any $%
w\in H_{\mathcal{X},0}^{1,2}\left( \Omega \right) ,$ i.e., that if $\vec{W}$
and $\vec{W^{\prime }}$ are any two representatives of $\nabla w$, then $%
\sqrt{\mathcal{A}(x,z)}\vec{W}(x)=\sqrt{\mathcal{A}(x,z)}\vec{W^{\prime }}%
(x) $ and $\vec{\gamma}(x,z)\cdot \vec{W}(x)=\vec{\gamma}(x,z)\cdot \vec{%
W^{\prime }}(x)$. In fact, since $\mathbf{v}_{j}\cdot (\vec{W}-\vec{%
W^{\prime }})=\mathbf{v}_{j}\cdot \vec{W}-\mathbf{v}_{j}\cdot \vec{W^{\prime
}}=0$ for all $j$ by (\ref{uniquegrad}), this follows immediately from (\ref%
{subu}) by choosing $\xi =\vec{W}(x)-\vec{W^{\prime }}(x)$. \newline
Let {$\vec{k}\left( x,z\right) =$}$\left( k^{i}\left( x,z\right) \right)
_{i=1,\cdots ,n}$ and $\mathcal{A}\left( x,z\right) $ be as in Theorem \ref%
{DP}, that is, with $\Gamma =\widetilde{\Omega} \times \mathbb{R}$,
\begin{enumerate}
\item {$\vec{k}\left( x,z\right) \in \mathcal{C}^2(\Gamma)$ and satisfies }%
the nondegeneracy Condition{\ \ref{hyp1} in $\Gamma$, }
\item {$\mathcal{A} \in \mathcal{C}^\infty(\Gamma)$ and satisfies the
diagonal }Condition{\ \ref{hellipp} in $\Gamma $, }
\item {$\mathcal{A}$ satisfies }the super subordination Condition{\ \ref%
{hyp2} in $\Gamma $.}
\end{enumerate}
The particular vector fields that we will use are
\begin{equation}
\partial _{1},\sqrt{k^{2}(x,0)}\,\partial _{2},\dots ,\sqrt{k^{n}(x,0)}%
\,\partial _{n}. \label{ourvf}
\end{equation}%
We claim that since $k^{2}(x,0),\dots k^{n}(x,0)\in \mathcal{C}^{2}(\Omega )$
and are nonnegative, the Wirtinger inequality (\ref{genWirt}) implies that
the vector fields (\ref{ourvf}) belong to $\mathop{\rm Lip_{loc}}(\Omega )$.
To see why, fix $i$ and denote $k^{i}(x,0)=k(x)$. For a Euclidean ball $%
D\Subset \Omega $, $\varepsilon >0$ and all $x_{1},x_{2}\in D$, we have
\begin{eqnarray*}
\left\vert \sqrt{k(x_{1})+\varepsilon }-\sqrt{k(x_{2})+\varepsilon }%
\right\vert &\leq &\left\Vert \nabla \sqrt{k+\varepsilon }\right\Vert
_{L^{\infty }(D)}|x_{1}-x_{2}| \\
&=&\left\Vert \frac{\nabla k}{\sqrt{k+\varepsilon }}\right\Vert _{L^{\infty
}(D)}|x_{1}-x_{2}| \\
&\leq &C_{D}\left\Vert \frac{\sqrt{k}}{\sqrt{k+\varepsilon }}\right\Vert
_{L^{\infty }(D)}|x_{1}-x_{2}|\quad \mbox{by (\ref{genWirt}),}
\end{eqnarray*}%
where $C_{D}$ depends on $k$ and dist$(D,\partial \Omega )$. Hence
\begin{equation*}
\left\vert \sqrt{k(x_{1})+\varepsilon }-\sqrt{k(x_{2})+\varepsilon }%
\right\vert \leq C_{D}|x_{1}-x_{2}|,\quad x_{1},x_{2}\in D,
\end{equation*}%
uniformly in $\varepsilon $. Letting $\varepsilon \rightarrow 0$ and using
the Heine-Borel theorem to cover any compact subset of $\Omega $ by a finite
number of balls proves our claim.
\subsubsection{$\mathcal{X}$-weak solutions of quasilinear equations\label%
{Subsubsection-weak-solution}}
Here we make precise the notion of a ``weak solution'' of the quasilinear
equation (\ref{equation}). For $k^{i}\left( x,z\right) $ as in the
hypotheses of our main results Theorems \ref{application} and \ref{DP}, we
let $\mathcal{X} =\left\{ X_{j}\right\} _{j=1}^{n}$ with $X_{j}=\sqrt{%
k^{i}\left( x,0\right) }\frac{\partial }{\partial x_{j}}$. An analogous
definition can be given for any collection of locally Lipschitz vector
fields.
\begin{definition}
\label{Xweak}A function $w\in H_{\mathcal{X}}^{1,2}\left( \Omega \right)
\bigcap L_{\mathop{\rm loc}}^{\infty }\left( \Omega \right) $ is a weak
solution of
\begin{equation}
\mathcal{Q}w=\mathop{\rm div}\mathcal{A}\left( x,w\right) \nabla w+\vec{
\gamma}\left( x,w\right) \cdot \nabla w+f\left( x,w\right) =0\qquad \text{in
}\Omega \label{equation-weak}
\end{equation}
if for all $u\in \mathop{\rm Lip}_0(\Omega)$,
\begin{equation}
\int_{\Omega }\left( \nabla u\right) ^{t}\mathcal{A}\left( x,w\right) \nabla
w~dx=\int_{\Omega }u~\vec{\gamma}\left( x,w\right) \cdot \nabla
w~dx+\int_{\Omega }u~f\left( x,w\right) ~dx. \label{weak-eq}
\end{equation}
Given $w_{0}$, $w_{1}\in H_{\mathcal{X}}^{1,2}\left( \Omega \right) \bigcap
L_{\mathop{\rm loc}}^{\infty }\left( \Omega \right),$ we say that $\mathcal{Q%
}w_{1}\geq \mathcal{Q}w_{0}$ in $\Omega$ if
\begin{eqnarray*}
&&\int_{\Omega }\left( \nabla u\right) ^{t}\mathcal{A}\left( x,w_{1}\right)
\nabla w_{1}~dx-\int_{\Omega }u~\vec{\gamma}\left( x,w_{1}\right) \cdot
\nabla w_{1}~dx-\int_{\Omega }u~f\left( x,w_{1}\right) ~dx \\
&\leq &\int_{\Omega }\left( \nabla u\right) ^{t}\mathcal{A}\left(
x,w_{0}\right) \nabla w_{0}~dx-\int_{\Omega }u~\vec{\gamma}\left(
x,w_{0}\right) \cdot \nabla w_{0}~dx-\int_{\Omega }u~f\left( x,w_{0}\right)
~dx
\end{eqnarray*}
for all $u\in \mathop{\rm Lip}_0(\Omega)$.
\end{definition}
To show that the integrals in (\ref{weak-eq}) converge absolutely, note that
if $w\in $ $L_{\mathop{\rm loc}}^{\infty }\left( \Omega \right) $, then by
Lemma \ref{admisall} and the diagonal condition (\ref{hellip}), we have that
for all $x\in \Omega ^{\prime }\Subset \Omega,$
\begin{equation*}
\xi ^{t}\mathcal{A}\left( x,w\left( x\right) \right) \xi \approx \xi ^{t}
\mathcal{A}\left( x,0\right) \xi \approx \sum_{i=1}^{n}k^{i}\left(
x,0\right) \xi _{i}^{2},
\end{equation*}
with constants which depend on $\mathcal{A}$, $\Omega ^{\prime }$ and $%
M_{0}=\left\Vert w\right\Vert _{L^{\infty }\left( \Omega ^{\prime }\right) }$%
. Hence, since $w\in H_{\mathcal{X}}^{1,2}\left( \Omega \right) \bigcap L_{%
\mathop{\rm loc}}^{\infty }\left( \Omega \right) $, $u\in \mathop{\rm Lip}%
_0(\Omega)$, and $f\left(x,z\right) $ is continuous, it follows that
\begin{equation*}
\nabla _{\!\!\sqrt{\mathcal{A}},w}w,~\nabla _{\!\!\sqrt{\mathcal{A}},w}u,~%
\vec{\gamma }\left( x,w\left( x\right) \right) \cdot \nabla w,~-f\left(
x,w\left( x\right) \right) \in L_{\mathop{\rm loc}}^{2}\left( \Omega \right)
.
\end{equation*}
Consequently, each integral in (\ref{weak-eq}) converges absolutely, and the
same is true for the integrals in Definition \ref{Xweak}.
Alternately, we can make sense of weak solutions $w\in H_{\mathcal{X}
}^{1,2}\left( \Omega \right) $ which are not necessarily locally bounded by
assuming that the coefficient matrix is bounded in $z$ locally in $x$, that $%
\vec{\gamma}(x,z)$ is of subunit type globally in $z$ locally in $x$, and
that $\sup_z |f(x,z)|$ is locally integrable.
\subsection{{\label{maxpsection}A Maximum Principle}}
We will now prove the following result.
\begin{theorem}
{\ \label{maxp}Let }$\mathcal{A}$ satisfy (\ref{hellip}), $\vec{\gamma}${\ }
be of subunit type with respect to $\mathcal{A}${\ in $\Omega \times \mathbb{%
R}$, } and $f\,$ be a continuous function on $\Omega \times \mathbb{R}^{n}$
which satisfies $f(x,0)=0$ for $x\in \Omega $ and
\begin{eqnarray}
f\left( x,z\right) \mathop{\rm sign}z &\leq &0\qquad \text{in }\Omega \times
\mathbb{R}, \label{eff} \\
f\left( x,z_{1}\right) -f\left( x,z_{2}\right) &\leq &0\qquad \text{if }x\in
\Omega \text{ and }z_{1}\geq z_{2}. \label{fnc}
\end{eqnarray}%
{If $w$ is a smooth function in }$\Omega $ which is continuous on $\overline{%
\Omega }$ and satisfies
\begin{equation}
\mathop{\rm div}\mathcal{A}\left( x,w\right) \nabla w+\vec{\gamma}\left(
x,w\right) \cdot \nabla w+f\left( x,w\right) \geq 0\qquad \text{in }\Omega
\label{subso}
\end{equation}%
in the weak sense, i.e., satisfies
\begin{equation}
\int \nabla {\varphi }\cdot \mathcal{A}\left( x,w\right) \nabla w\leq \int {%
\varphi }\,\vec{\gamma}\left( x,w\right) \cdot \nabla w+\int {\varphi }%
\,f\left( x,w\right) \label{subsso}
\end{equation}%
for all nonnegative ${\varphi \in }\mathop{\rm Lip_{0}}\left( \Omega \right)
$, then
\begin{equation*}
\sup_{\Omega }w\leq \sup_{\partial \Omega }w^{+}.
\end{equation*}%
On the other hand, if the opposite inequality holds in (\ref{subso}), i.e.,
if
\begin{equation}
\mathop{\rm div}\mathcal{A}\left( x,w\right) \nabla w+\vec{\gamma}\left(
x,w\right) \cdot \nabla w+f\left( x,w\right) \leq 0\qquad \text{in }\Omega
\label{superso}
\end{equation}%
in the weak sense, then
\begin{equation}
\inf_{\Omega }w\geq \inf_{\partial \Omega }-(w^{-})\quad \left(
\,=-\sup_{\partial \Omega }w^{-}\right) , \label{lowerbound}
\end{equation}%
where $w^{-}:=-w$ if $w\leq 0$ and $w^{-}:=0$ if $w>0$. In particular, if $w$
is a weak solution of $\mathop{\rm
div}\mathcal{A}\left( x,w\right) \nabla w+\vec{\gamma}\left( x,w\right)
\cdot \nabla w+f\left( x,w\right) =0$ in $\Omega $, then $\sup_{\Omega
}|w|\leq \sup_{\partial \Omega }|w|$.
\end{theorem}
{\
\proof%
Assume first that $w$ satisfies (\ref{subso}), and recall that $w$ is smooth
by assumption. Let $\omega _{0}=\sup_{\partial \Omega }w^{+}$ and
\begin{equation*}
v_{\tau }\left( x\right) =\left( w(x)-\omega _{0}-\tau \right) ^{+},\qquad
\tau >0,\,x\in \Omega .
\end{equation*}%
Then $v_{\tau }$ is nonnegative and Lipschitz continuous in $\Omega $, and $%
v_{\tau }$ has compact support in $\Omega $ since $w$ is continuous on $%
\overline{\Omega }$ by hypothesis and $\tau >0$. Also, for any $x$, $v_{\tau
}(x)>0$ if and only if $w(x)>\omega _{0}+\tau $. Let $\Phi _{\tau }=\{x\in
\Omega :v_{\tau }(x)>0\}.$ Then $v_{\tau }=\chi _{\Phi _{\tau }}(w-\omega
_{0}-\tau )$ on $\Omega $ and $\nabla v_{\tau }=\chi _{\Phi _{\tau }}\nabla
w $ a.e. on $\Omega $. By choosing ${\varphi =}v_{\tau }$ in (\ref{subsso}),
we obtain
\begin{equation*}
\int \nabla v_{\tau }\cdot \mathcal{A}\left( x,w\right) \nabla w\leq \int
v_{\tau }\vec{\gamma}\left( x,w\right) \cdot \nabla w+\int v_{\tau }f\left(
x,w\right) .
\end{equation*}%
In this inequality, the right-hand side satisfies
\begin{equation*}
\int_{\Phi _{\tau }}v_{\tau }\vec{\gamma}\left( x,w\right) \cdot \nabla
w+\int_{\Phi _{\tau }}v_{\tau }f\left( x,w\right) =\int_{\Phi _{\tau
}}v_{\tau }\vec{\gamma}\left( x,w\right) \cdot \nabla v_{\tau }+\int_{\Phi
_{\tau }}v_{\tau }f\left( x,w\right) ,
\end{equation*}%
while the left-hand side is $\int_{\Phi _{\tau }}\nabla v_{\tau }\cdot
\mathcal{A}\left( x,w\right) \nabla v_{\tau }.$ Hence
\begin{equation}
\int_{\Phi _{\tau }}\nabla v_{\tau }\cdot \mathcal{A}\left( x,w\right)
\nabla v_{\tau }\leq \int_{\Phi _{\tau }}v_{\tau }\vec{\gamma}\left(
x,w\right) \cdot \nabla v_{\tau }+\int_{\Phi _{\tau }}v_{\tau }f\left(
x,w\right) =I+II. \label{7.12}
\end{equation}%
We have
\begin{equation*}
II=\int_{\Phi _{\tau }}v_{\tau }f\left( x,v_{\tau }\right) +\int_{\Phi
_{\tau }}v_{\tau }\big[f\left( x,w\right) -f\left( x,v_{\tau }\right) \big]%
\leq 0\,+\,0=0
\end{equation*}%
by (\ref{eff}) and (\ref{fnc}) since $v_{\tau }>0$ and $w>v_{\tau }$ on $%
\Phi _{\tau }$ (note that $\omega \geq 0$). Recalling that $w$ is assumed to
be continuous on $\overline{\Omega }$ and so is bounded there, we may choose
$M$ with $M>w$ on $\Omega $. Since $\vec{\gamma}$ is of subunit type with
respect to $\mathcal{A}$, Schwarz's inequality implies that
\begin{equation*}
I\leq \frac{1}{4}\int_{\Phi _{\tau }}\nabla v_{\tau }\cdot \mathcal{A}\left(
x,w\right) \nabla v_{\tau }+4B_{\gamma }^{2}\int_{\Phi _{\tau }}v_{\tau
}^{2},
\end{equation*}%
where $B_{\gamma }=B_{\gamma }\left( \Omega _{\tau },M\right) $. From (\ref%
{7.12}) and the estimates for $I$ and $II$, we obtain
\begin{equation}
\int_{\Phi _{\tau }}\nabla v_{\tau }\cdot \mathcal{A}\left( x,w\right)
\nabla v_{\tau }\leq C^{2}B_{\gamma }^{2}\int_{\Phi _{\tau }}v_{\tau }^{2}.
\label{7.13}
\end{equation}%
By the one-dimensional Sobolev estimate,
\begin{equation*}
\int_{\Phi _{\tau }}v_{\tau }^{2}\leq C^{2}R^{2}\int_{\Phi _{\tau
}}\left\vert \partial _{1}v_{\tau }\right\vert ^{2},\quad R=\mathop{\rm diam}%
(\Omega ).
\end{equation*}%
Combining this with (\ref{7.13}) gives
\begin{equation*}
\int_{\Phi _{\tau }}v_{\tau }^{2}\leq C^{2}R^{2}B_{\gamma }^{2}\int_{\Phi
_{\tau }}v_{\tau }^{2}.
\end{equation*}%
Thus, assuming that $R<(CB_{\gamma })^{-1}$, we obtain $\int_{\Phi _{\tau
}}v_{\tau }^{2}=0$. Then $\Phi _{\tau }$ is empty, i.e., $v_{\tau }=0$ on $%
\Omega $ and therefore $w\leq \omega _{0}+\tau $ on $\Omega $. }
To drop the restriction that $R<\big(CB_{\gamma }(\Omega _{\tau },M)\big)%
^{-1}$, let $N=CB_{\gamma }\mathop{\rm diam}(\Omega )+1$ and
\begin{equation*}
\widetilde{\Omega }=\frac{\Omega }{N}=\left\{ \frac{x}{N}:x\in \Omega
\right\} .
\end{equation*}%
Also, for $x\in \widetilde{\Omega }$, let
\begin{eqnarray*}
\widetilde{w}\left( x\right) &=&w\left( Nx\right) ,\qquad \widetilde{%
\mathcal{A}}\left( x,z\right) =N^{-2}\mathcal{A}\left( Nx,z\right) , \\
\widetilde{\vec{\gamma}}\left( x,z\right) &=&N^{-1}\vec{\gamma}\left(
Nx,z\right) ,\qquad \widetilde{f}\left( x,z\right) =f\left( Nx,z\right) .
\end{eqnarray*}%
Then if $x\in \widetilde{\Omega }$,
\begin{eqnarray*}
&&\mathop{\rm div}\widetilde{\mathcal{A}}\left( x,\widetilde{w}\right)
\nabla \widetilde{w}+\widetilde{\vec{\gamma}}\left( x,\widetilde{w}\right)
\cdot \nabla \widetilde{w}+\widetilde{f}\left( x,\widetilde{w}\right) \\
&=&N^{-2}\mathop{\rm div}\big[\mathcal{A}\left( Nx,w\left( Nx\right) \right)
\nabla \big(w\left( Nx\right) \big)\big]+N^{-1}\vec{\gamma}\left( Nx,w\left(
Nx\right) \right) \cdot \nabla \big(w\left( Nx\right) \big) \\
&&+f\left( Nx,w\left( Nx\right) \right) \\
&=&\mathop{\rm div}\mathcal{A}\left( y,w\left( y\right) \right) \nabla
w\left( y\right) +\vec{\gamma}\left( y,w\left( y\right) \right) \cdot \nabla
w\left( y\right) +f\left( y,w\left( y\right) \right) \geq 0
\end{eqnarray*}%
by (\ref{subso}), where $y=Nx\in \Omega $. Thus $\widetilde{w}$ satisfies an
analogue of (\ref{subso}) in $\tilde{\Omega}$. Moreover, since $\vec{\gamma}$
is of subunit type with respect to $\mathcal{A}$ in $\Omega \times \mathbb{R}
$ (Definition \ref{subunitt}), it follows that $\widetilde{\vec{\gamma}}$ is
of subunit type with respect to $\widetilde{\mathcal{A}}$ in $\tilde{\Omega}%
\times \mathbb{R}$ with constant $\widetilde{B_{\gamma }}=B_{\gamma }$:
indeed,
\begin{eqnarray*}
\left\vert \widetilde{\vec{\gamma}}\left( x,z\right) \cdot \xi \right\vert
^{2} &=&\left\vert N^{-1}\vec{\gamma}\left( Nx,z\right) \cdot \xi
\right\vert ^{2}\leq N^{-2}B_{\gamma }^{2}\xi ^{t}\mathcal{A}\left(
Nx,z\right) \xi \\
&=&B_{\gamma }^{2}\xi ^{t}\widetilde{\mathcal{A}}\left( x,z\right) \xi .
\end{eqnarray*}%
Also,
\begin{eqnarray*}
\widetilde{f}\left( x,z\right) \mathop{\rm sign}z &=&f\left( Nx,z\right) %
\mathop{\rm sign}z\leq 0\qquad \text{in }\widetilde{\Omega }\times \mathbb{R}%
, \\
\widetilde{f}\left( x,z_{1}\right) -\widetilde{f}\left( x,z_{2}\right)
&=&f\left( Nx,z_{1}\right) -f\left( Nx,z_{2}\right) \leq 0\quad \text{ if }%
z_{1}\geq z_{2},\,x\in \widetilde{\Omega }
\end{eqnarray*}%
and
\begin{equation*}
\mathop{\rm diam}\widetilde{\Omega }=\dfrac{\mathop{\rm diam}\Omega }{N}=%
\dfrac{\mathop{\rm diam}\Omega }{CB_{\gamma }\mathop{\rm diam}\Omega +1}<%
\dfrac{1}{CB_{\gamma }}=\dfrac{1}{C\widetilde{B_{\gamma }}}.
\end{equation*}%
Then $\widetilde{w}\leq \sup_{\partial \widetilde{\Omega }}(\widetilde{w}%
)^{+}+\tau $ in $\widetilde{\Omega }$ for all $\tau >0$, i.e., $w\leq
\sup_{\partial \Omega }w^{+}+\tau $ in $\Omega $ for all $\tau >0$. Letting $%
\tau \rightarrow 0$, we obtain $w\leq \sup_{\partial \Omega }w^{+}$ as
desired.
To prove (\ref{lowerbound}), let (\ref{superso}) hold and define $\widetilde{%
\mathcal{A}}(x,z)=\mathcal{A}(x,-z)$, $\widetilde{\vec{\gamma}}(x,z)=\vec{%
\gamma}(x,-z)$ and $\widetilde{f}(x,z)=-f(x,z)$ for $x\in \Omega $. Since $%
\mathcal{A}$ satisfies (\ref{hellip}) and $\vec{\gamma}$ is subunit with
respect to $\mathcal{A}$ in $\Omega \times \mathbb{R}$, it follows that $%
\widetilde{\mathcal{A}}$ satisfies (\ref{hellip}) and $\widetilde{\vec{\gamma%
}}$ is subunit with respect to $\widetilde{\mathcal{A}}$ in $\Omega \times
\mathbb{R}$. From (\ref{eff}) and (\ref{fnc}) for $f$, we obtain (\ref{eff})
and (\ref{fnc}) for $\widetilde{f}$:
\begin{eqnarray*}
\widetilde{f}\left( x,z\right) \mathop{\rm sign}z &=&-f\left( x,-z\right) %
\mathop{\rm sign}z \\
&=&f(x,-z)\mathop{\rm sign}(-z)\leq 0\qquad \text{in }\Omega \times \mathbb{R%
}, \\
\widetilde{f}\left( x,z_{1}\right) -\widetilde{f}\left( x,z_{2}\right)
&=&f\left( x,-z_{2}\right) -f\left( x,z_{1}\right) \leq 0\quad \text{ if }%
z_{1}\geq z_{2},\,x\in \Omega .
\end{eqnarray*}%
Now let $\widetilde{w}(x)=-w(x)$ and note that
\begin{equation*}
\mathop{\rm div}\widetilde{\mathcal{A}}\left( x,\widetilde{w}(x)\right)
\nabla \widetilde{w}(x)+\widetilde{\vec{\gamma}}\left( x,\widetilde{w}%
(x)\right) \cdot \nabla \widetilde{w}(x)+\widetilde{f}\left( x,\widetilde{w}%
(x)\right) \,=
\end{equation*}%
\begin{equation*}
-\left[ \mathop{\rm div}\mathcal{A}\left( x,w(x)\right) \nabla w(x)+\vec{%
\gamma}\left( x,w(x)\right) \cdot \nabla w(x)+f\left( x,w(x)\right) \right]
\,\geq 0\quad \text{in $\Omega $.}
\end{equation*}%
By the previous case, $\sup_{\Omega }\widetilde{w}\leq \sup_{\partial \Omega
}\widetilde{w}^{+}$. Equivalently,
\begin{equation*}
\sup_{x\in \Omega }(-w(x))\leq \sup_{x\in \partial \Omega
}(-w(x))^{+},\qquad \text{or}\qquad \inf_{x\in \Omega }w(x)\geq \inf_{x\in
\partial \Omega }-(w(x)^{-}),
\end{equation*}
which completes the proof of Theorem \ref{maxp}.{%
\endproof%
}
\subsection{{A Comparison Principle\label{compsect}}}
\begin{lemma}
{\ \label{comparison}Suppose that }$\mathcal{A}\left( x,z\right) $ satisfies
(\ref{hellip}) and (\ref{xtra}), and that $f$ is nonincreasing in $z$, i.e.,
\begin{equation}
f_{z}\left( x,z\right) \leq 0,\qquad \left( x,z\right) \in \Gamma .
\label{finc}
\end{equation}%
{Let $w_{0}$, $w_{1}\in $}$\left( H_{\mathcal{X}}^{1,2}\left( \Omega \right)
\bigcup \mathcal{C}^{\infty }\left( \Omega \right) \right) \bigcap \,${$%
\mathcal{C}^{0}\left( \overline{\Omega }\right) $\ satisfy $w_{0}+\kappa
\geq w_{1}$ on $\partial \Omega $ for some constant }$\kappa \geq 0${\ and
\begin{equation}
\mathcal{P}\left( w_{1}\right) \geq \mathcal{P}\left( w_{0}\right) \,\text{
in }\Omega \label{O}
\end{equation}%
(in the sense of Definition \ref{Xweak}), where
\begin{equation*}
\mathcal{P}\left( w\right) =\mathop{\rm div}\mathcal{A}\left( x,w\right)
\nabla w+f\left( x,w\right) .
\end{equation*}%
Then $w_{0}+\kappa \geq w_{1}$ in $\Omega $. In particular, if $\mathcal{P}%
w_{0}=\mathcal{P}w_{1}\,$ in $\Omega $ and $w_{0}=w_{1}\,$on $\partial
\Omega $, then $w_{0}=w_{1}$ in $\overline{\Omega }$. }
\end{lemma}
{\
\proof%
First we will assume that $w_{0}$, $w_{1}\in $}$H_{\mathcal{X}}^{1,2}\left(
\Omega \right) \bigcap \,$$\mathcal{C}^{0}\left( \overline{\Omega }\right) $%
. {Given $\tau >0,$ {let }$u_{\tau }=w_{1}-w_{0}-\kappa -\tau $ and {\ }$%
u_{\tau }^{+}=\max \left\{ u_{\tau },0\right\} $. Then $u_{\tau }^{+}$ is a
nonnegative continuous function compactly supported in $\Omega $. Denote $K=%
\mathop{\rm supp}\left( u_{\tau }^{+}\right) $ and $\delta _{0}=%
\mathop{\rm
dist}\left( K,\partial \Omega \right) $. Let $\psi _{\delta }$ be a smooth
approximation of the identity with $\delta >0$, i.e., $\psi _{\delta }\left(
x\right) =\delta ^{-n}\psi \left( x/\delta \right) $ where $\psi \in
\mathcal{C}_{0}^{\infty }\left( B_{1}\right) $ and $\int \psi dx=1$; here $%
B_{1}$ denotes the unit ball in $\mathbb{R}^{n}$. For $0<\varepsilon <1$ and
$0<\delta <\delta _{0}/2$, set
\begin{equation*}
{\varphi }_{\tau ,\varepsilon ,\delta }=\frac{u_{\tau }^{+}}{u_{\tau
}^{+}+\varepsilon }\ast \psi _{\delta }.
\end{equation*}%
Then $\varphi _{\tau ,\varepsilon ,\delta }$ is nonnegative, smooth and
compactly supported in $\Omega $. {From (\ref{O}),
\begin{equation*}
\int_{\Omega }\left[ \mathcal{A}\left( x,w_{1}\right) \nabla w_{1}-\mathcal{A%
}\left( x,w_{0}\right) \nabla w_{0}\right] \,\nabla {\varphi }_{\tau
,\varepsilon ,\delta }-\int_{\Omega }\left[ f\left( x,w_{1}\right) -f\left(
x,w_{0}\right) \right] \,{\varphi }_{\tau ,\varepsilon ,\delta }\leq 0.
\end{equation*}%
}Since $w_{0},w_{1},\varphi _{\tau ,\varepsilon ,\delta }\in H_{\mathcal{X}%
}^{1,2}\left( \Omega \right) $ and $\varphi _{\tau ,\varepsilon ,\delta }$
is continuous and has compact support, all the integrals above are
absolutely convergent. By Proposition 1.2.2 in \cite{FSS}, ${\varphi }_{\tau
,\varepsilon ,\delta }\rightarrow u_{\tau }^{+}/$}$\left( {u_{\tau
}^{+}+\varepsilon }\right) ${\ in $H_{\mathcal{X}}^{1,2}\left( \Omega
^{\prime \prime }\right) $ as $\delta \rightarrow 0^{+}$ for any open $%
\Omega ^{\prime \prime }\Subset \Omega ^{\prime }$. Letting $\delta
\rightarrow 0^{+}$, we obtain{{\
\begin{eqnarray}
&&\varepsilon \int_{\Omega }\left[ \mathcal{A}\left( x,w_{1}\right) \nabla
w_{1}-\mathcal{A}\left( x,w_{0}\right) \nabla w_{0}\right] \,\frac{\nabla
u_{\tau }^{+}}{\left( u_{\tau }^{+}+\varepsilon \right) ^{2}}
\label{subsol2} \\
&&-\int_{\Omega }\left[ f\left( x,w_{1}\right) -f\left( x,w_{0}\right) %
\right] \,\frac{u_{\tau }^{+}}{u_{\tau }^{+}+\varepsilon }\leq 0, \notag
\end{eqnarray}%
where we used that }}$\nabla \left( u_{\tau }^{+}/(u_{\tau }^{+}+\varepsilon
)\right) =\varepsilon \nabla u_{\tau }^{+}/$}$\left( u_{\tau
}^{+}+\varepsilon \right) ^{2}${. }
{Set {$u=w_{1}-w_{0}-\kappa $ and {$w_{t}=tw_{1}+\left( 1-t\right) w_{0}$, $%
0\leq t\leq 1$.} Then $\partial _{t}w_{t}=$}$u+\kappa $, and} {by (\ref{finc}%
),
\begin{eqnarray}
\left[ f\left( x,w_{1}\right) -f\left( x,w_{0}\right) \right] \frac{u_{\tau
}^{+}}{u_{\tau }^{+}+\varepsilon } &=&\frac{u_{\tau }^{+}}{u_{\tau
}^{+}+\varepsilon }\int_{0}^{1}\partial _{t}\left[ f\left( x,w_{t}\right) %
\right] \,dt \notag \\
&=&\frac{u_{\tau }^{+}\left( u+\kappa \right) }{u_{\tau }^{+}+\varepsilon }%
\int_{0}^{1}f_{z}\left( x,w_{t}\right) \,dt\leq 0, \label{subsll022}
\end{eqnarray}%
where we used that }$u+\kappa \geq 0$ on the support of $u_{\tau }^{+}$;
recall that $\kappa \ge 0$ by hypothesis. Also,{{\
\begin{eqnarray}
&&\mathcal{A}\left( x,w_{1}\right) \nabla w_{1}-\mathcal{A}\left(
x,w_{0}\right) \nabla w_{0}=\int_{0}^{1}\partial _{t}\left\{ \mathcal{A}%
\left( x,w_{t}\right) \nabla w_{t}\right\} \,dt \notag \\
&=&\left( u+\kappa \right) \,\left\{ \int_{0}^{1}\mathcal{A}_{z}\left(
x,w_{t}\right) \nabla w_{t}\,\,dt\right\} +\left\{ \int_{0}^{1}\mathcal{A}%
\left( x,w_{t}\right) \,dt\right\} \nabla u \notag \\
&=&\left( u+\kappa \right) \,\vec{a}\left( x\right) +\mathbf{\mathbf{\tilde{A%
}}}\left( x\right) \nabla u,\qquad \text{where} \label{subsll020}
\end{eqnarray}%
\begin{equation*}
\vec{a}\left( x\right) =\int_{0}^{1}\mathcal{A}_{z}\left( x,w_{t}\right)
\nabla w_{t}\,\,dt\qquad \text{and}\qquad \mathbf{\mathbf{\tilde{A}}}\left(
x\right) =\int_{0}^{1}\mathcal{A}\left( x,w_{t}\right) \,dt.
\end{equation*}%
Using (\ref{subsll020}) in (\ref{subsol2}),{\ using (\ref{subsll022}) to
omit the term in (\ref{subsol2}) which involves the difference of $f$
values, }and using the facts that $u=u_{\tau }^{+}+\tau $ and $\nabla
u=\nabla u_{\tau }^{+}$ on the support of $u_{\tau }^{+}$ yields}}
\begin{equation}
\int_{\Omega }\mathbf{\mathbf{\tilde{A}}}\left( x\right) \nabla u_{\tau
}^{+}\cdot \frac{\nabla u_{\tau }^{+}}{\left( u_{\tau }^{+}+\varepsilon
\right) ^{2}}\leq -\int_{\Omega }\left[ \left( {{u_{\tau }^{+}+\tau }}%
\right) \,\vec{a}\left( x\right) \right] \,\frac{\nabla u_{\tau }^{+}}{%
\left( u_{\tau }^{+}+\varepsilon \right) ^{2}}. \label{subsol05}
\end{equation}%
{{{\ } Now, by Schwarz's inequality and the definition of }}$\,\vec{a}\left(
x\right) ,$
\begin{eqnarray*}
&&\left\vert \int_{\Omega }\left[ \left( {{u_{\tau }^{+}+\tau }}\right) \,%
\vec{a}\left( x\right) \right] \,\frac{\nabla u_{\tau }^{+}}{\left( u_{\tau
}^{+}+\varepsilon \right) ^{2}}\right\vert \\
&=&\left\vert \int_{\Omega }\left[ \left( {{u_{\tau }^{+}+\tau }}\right)
\,\left\{ \int_{0}^{1}\nabla u_{\tau }^{+}\cdot \mathcal{A}_{z}\left(
x,w_{t}\right) \nabla w_{t}\,\,dt\right\} \right] \,\frac{1}{\left( u_{\tau
}^{+}+\varepsilon \right) ^{2}}\right\vert \\
&\leq &\alpha \int_{\Omega }\left\{ \int_{0}^{1}\dfrac{\left\vert \mathcal{A}%
_{z}\left( x,w_{t}\right) \nabla u_{\tau }^{+}\right\vert ^{2}}{k^{\ast
}\left( x,w_{t}\right) }\,dt\right\} \,\frac{1}{\left( u_{\tau
}^{+}+\varepsilon \right) ^{2}} \\
&&+\dfrac{C}{\alpha }\int_{\Omega }\,\left\{ \int_{0}^{1}k^{\ast }\left(
x,w_{t}\right) \left\vert \nabla w_{t}\right\vert ^{2}\,dt\right\} \,\frac{%
\left( {{u_{\tau }^{+}+\tau }}\right) ^{2}}{\left( u_{\tau }^{+}+\varepsilon
\right) ^{2}}.
\end{eqnarray*}%
{{In fact, in the last four integrations as well as those below, the domain
of integration can be restricted to the compact subset supp $u_{\tau }^{+}$
of $\Omega $. Then, since $k^{\ast }\left( x,w_{t}\right) \left\vert \nabla
w_{t}\right\vert ^{2}\leq \left\vert \nabla _{\!\!\sqrt{\mathcal{A}}%
,w_{t}}w_{t}\right\vert ^{2}$ due to (\ref{hellip}), by assuming that }}$%
\tau \leq \varepsilon ${\ and{\ applying{\ (\ref{xtra})} to the first term
on the right, we obtain }}
\begin{eqnarray*}
&&\left\vert \int_{\Omega }\left[ \left( {{u_{\tau }^{+}+\tau }}\right) \,%
\vec{a}\left( x\right) \right] \,\frac{\nabla u_{\tau }^{+}}{\left( u_{\tau
}^{+}+\varepsilon \right) ^{2}}\right\vert \\
&\leq &\alpha B_{\mathcal{A}}^{2}\int_{\Omega }\dfrac{\left\{
\int_{0}^{1}\nabla u_{\tau }^{+}\cdot \mathcal{A}\left( x,w_{t}\right)
\nabla u_{\tau }^{+}\,\,dt\right\} }{\left( u_{\tau }^{+}+\varepsilon
\right) ^{2}}+\dfrac{C}{\alpha }\int_{\Omega }\left\{ \int_{0}^{1}\left\vert
\nabla _{\!\!\sqrt{\mathcal{A}},w_{t}}w_{t}\right\vert ^{2}\,\,dt\right\} \\
&\leq &\alpha B_{\mathcal{A}}^{2}\int_{\Omega }\dfrac{\nabla u_{\tau
}^{+}\cdot \mathbf{\mathbf{\tilde{A}}}\left( x\right) \nabla u_{\tau }^{+}}{%
\left( u_{\tau }^{+}+\varepsilon \right) ^{2}}+\frac{C}{\alpha },
\end{eqnarray*}%
{{where we used that $w_{t}\in $}}$H_{\mathcal{X}}^{1,2}\left( \Omega
\right) ${\ for $0\leq t\leq 1$, and hence the constant }$C$ is independent
of $\tau $ and $\varepsilon ${{. Taking $\alpha =1/2B_{\mathcal{A}^{2}}$,
combining with (\ref{subsol05}) and absorbing into the left gives
\begin{equation}
\int_{\Omega }\mathbf{\mathbf{\tilde{A}}}\left( x\right) \nabla u_{\tau
}^{+}\cdot \frac{\nabla u_{\tau }^{+}}{\left( u_{\tau }^{+}+\varepsilon
\right) ^{2}}\leq CB_{\mathcal{A}}^{2},\qquad 0<\tau \leq \varepsilon .
\label{subsll23}
\end{equation}%
}}
From (\ref{hellip}) and the fact that $k^{1}(x,z)=1$, we have
\begin{equation*}
\mathbf{\mathbf{\tilde{A}}}\left( x\right) \nabla u_{\tau }^{+}\cdot \nabla
u_{\tau }^{+}\geq \left( \partial _{1}u_{\tau }^{+}\right) ^{2}.
\end{equation*}%
Then, by (\ref{subsll23}) and the identity
\begin{equation*}
\frac{\partial _{1}\left( u_{\tau }^{+}\right) }{u_{\tau }^{+}+\varepsilon }=%
\frac{\partial _{1}\left( \frac{u_{\tau }^{+}}{\varepsilon }+1\right) }{%
\frac{u_{\tau }^{+}}{\varepsilon }+1}=\partial _{1}\ln \left( \frac{u_{\tau
}^{+}}{\varepsilon }+1\right) ,
\end{equation*}%
it follows that
\begin{equation*}
\int_{\Omega }\left\vert \partial _{1}\ln \left( \frac{u_{\tau }^{+}}{%
\varepsilon }+1\right) \right\vert ^{2}dx\leq CB_{\mathcal{A}}^{2},\qquad
0<\tau \leq \varepsilon .
\end{equation*}%
Applying the one-dimensional Sobolev inequality and letting $\tau
\rightarrow 0$ gives
\begin{equation*}
\dfrac{1}{C\mathop{\rm diam}\Omega }\int_{\Omega }\left\vert \ln \left(
\frac{u^{+}}{\varepsilon }+1\right) \right\vert ^{2}dx\leq \int_{\Omega
}\left\vert \partial _{1}\ln \left( \frac{u^{+}}{\varepsilon }+1\right)
\right\vert ^{2}dx\leq CB_{\mathcal{A}}^{2}
\end{equation*}%
{{uniformly in $\varepsilon >0$. Since $u$ is continuous in $\overline{%
\Omega }$, it follows that $u^{+}=0$ in $\Omega $, so that $u\leq 0$ in $%
\Omega $. Hence $w_{1}\leq w_{0}+\kappa $ in $\Omega $ as desired.}}
Now, for the general case, assume that {$w_{0}$, $w_{1}\in $}$\left( H_{%
\mathcal{X}}^{1,2}\left( \Omega \right) \bigcup \mathcal{C}^{\infty }\left(
\Omega \right) \right) \bigcap \,${$\mathcal{C}^{0}\left( \overline{\Omega }%
\right) $ and satisfy the hypotheses of the lemma. Consider a family} $%
\left\{ \Omega _{\varepsilon }\right\} _{\varepsilon >0}$ of open sets such
that $\Omega _{\varepsilon }\nearrow \Omega $ and%
\begin{equation*}
0<\inf \left\{ \mathop{\rm dist}\left( x,\partial \Omega \right) :x\in
\partial \Omega _{\varepsilon }\right\} <\varepsilon .
\end{equation*}%
Then $\Omega _{\varepsilon }\Subset \Omega $ for all $\varepsilon >0$, and
the function $\mu \left( \varepsilon \right) $ defined for $\varepsilon >0$
by $\mu \left( \varepsilon \right) =\max_{\partial \Omega _{\varepsilon
}}\left( w_{1}-w_{0}-\kappa \right) $ satisfies $\lim_{\epsilon \rightarrow
0}\mu \left( \epsilon \right) \leq 0$.
Since {$w_{0}$, $w_{1}\in $}$\left( H_{\mathcal{X}}^{1,2}\left( \Omega
\right) \bigcup \mathcal{C}^{\infty }\left( \Omega \right) \right) $ and $%
\Omega _{\varepsilon }\Subset \Omega $, it follows that {$w_{0}$, $w_{1}\in $%
}$H_{\mathcal{X}}^{1,2}\left( \Omega _{\varepsilon }\right) $ for each $%
\varepsilon >0$. Moreover, {$w_{0}+\kappa +\mu \left( \varepsilon \right)
\geq w_{1}$ on $\partial \Omega _{\varepsilon }$. By the previous case,
\begin{equation*}
{w_{0}+\kappa +\mu \left( \varepsilon \right) \geq w_{1}\qquad }\text{in }%
\Omega _{\varepsilon }.
\end{equation*}%
The lemma now follows by letting $\varepsilon \rightarrow 0^{+}$.%
\endproof%
}
\subsection{Barriers for the Dirichlet problem\label{barriersect}}
In this section, we construct barrier functions for continuous weak
solutions of the Dirichlet problem in a smooth, strictly convex domain. An
interesting aspect of these barriers is that even though they are
specialized to a particular solution $w$, they depend only on the modulus
continuity of $w$.
\begin{lemma}
\label{barrier} Let $\Phi \Subset \Omega $ be a strongly convex domain. Let $%
\bar{r}=\mathop{\rm diam}\Phi $ and $\omega $ be a concave, strictly
increasing function with $\omega \in \mathcal{C}^{0}\left( \left[ 0,\bar{r}%
\right] \right) \bigcap \mathcal{C}^{2}\left( \left( 0,\bar{r}\right]
\right) $ and $\omega \left( 0\right) =0$. Suppose $\mathcal{A}$ satisfies (%
\ref{xtra}) and $\gamma $ is of subunit type with respect to $\mathcal{A}$.
For $m\in \mathbb{R}$, define a differential operator $\mathcal{L}_{m}$ by
\begin{equation*}
\mathcal{L}_{m}h=\mathop{\rm div}\mathcal{A}\left( x,h\left( x\right)
+m\right) \nabla h.
\end{equation*}%
Then for every $x_{0}\in \partial \Phi $ and all positive constants $\eta $,
$\nu $, $m_{0},K$, there exists a neighborhood $\mathcal{N}$of $x_{0}$ and a
function $h\in \mathcal{C}^{0}\left( \overline{\mathcal{N}}\right) \bigcap
\mathcal{C}^{\infty }\left( \mathcal{N}\right) $ such that $\mathcal{N}%
\subset \left\{ \left\vert x-x_{0}\right\vert <\eta \right\} $ and
\begin{equation*}
\left\{
\begin{array}{ll}
h\left( x_{0}\right) =0, & \\
h\left( x\right) \leq -\omega \left( \left\vert x-x_{0}\right\vert \right) ,
& x\in \Phi \bigcap \mathcal{N}, \\
h\left( x\right) \leq -\nu , & x\in \Phi \bigcap \partial \mathcal{N}, \\
\mathcal{L}_{m}h\geq K, & x\in \Phi \bigcap \mathcal{N},\quad \left\vert
m\right\vert \leq m_{0}, \\
\bigtriangleup h=\sum_{i=1}^{n}\partial _{i}^{2}h>0, & x\in \Phi \bigcap
\mathcal{N}.%
\end{array}%
\right.
\end{equation*}
\end{lemma}
\noindent {\bf Proof }%
Let $x_{0}\in \partial \Phi $. By translation, we may assume that $x_{0}=0.$
By using a rotation $\Theta =\left( \theta _{ij}\right) _{i,j=1}^{n}$, we
may represent $\partial \Phi $ locally as $y=\Theta x$ so that for $%
y^{\prime }=\left( y_{1},\dots ,y_{n-1}\right) $ and positive $\kappa _{0}$,
$r_{0}$ depending on $\partial \Phi $, we have
\begin{equation}
\kappa _{0}\left\vert y^{\prime }\right\vert ^{2}\leq y_{n},\qquad \left(
y^{\prime },y_{n}\right) \in \partial \Phi ,\qquad \left\vert y^{\prime
}\right\vert <r_{0}. \label{bdry}
\end{equation}%
By hypothesis, $\omega \,$satisfies
\begin{eqnarray}
\omega \left( r\right) &\geq &a_{0}\,r\qquad \text{if $r\in \left[ 0,\bar{r}%
\right] $, where $a_{0}=\frac{\omega \left( \bar{r}\right) }{\bar{r}}$},
\label{tang} \\
\liminf_{r\rightarrow 0^{+}}\omega ^{\prime }\left( r\right) &\geq &a_{0}>0,
\label{concav} \\
\omega ^{\prime \prime }\left( r\right) &\leq &0\qquad \text{if $r\in \left(
0,\bar{r}\right] .$} \label{conse}
\end{eqnarray}%
For fixed $\alpha _{0}\in (0,1]$, set
\begin{equation}
r_{0}=\left\{
\begin{array}{ll}
\omega ^{-1}\left( \alpha _{0}\right) & \qquad \text{if }\omega \left( \bar{r%
}\right) >\alpha _{0} \\
\bar{r} & \qquad \text{otherwise,}%
\end{array}%
\right. \label{barrr0}
\end{equation}%
i.e., $r_{0}$ is the largest $r\in (0,\bar{r}]$ with $\omega \left( r\right)
\leq \alpha _{0}$. Letting $\psi \left( r\right) =\sqrt{\omega \left(
r\right) }$ for $0<r\leq r_{0}$, we have by (\ref{tang})--(\ref{conse}) and
since $\lim_{r\rightarrow 0^{+}}\omega \left( r\right) =0$ that
\begin{eqnarray}
1 &\geq &\sqrt{\alpha _{0}}\geq \psi \left( r\right) \geq \omega \left(
r\right) , \label{psione} \\
-\psi ^{\prime \prime } &=&-\dfrac{2\omega ^{\prime \prime }\omega -\left(
\omega ^{\prime }\right) ^{2}}{4\omega ^{3/2}}\geq \frac{\left( \omega
^{\prime }\right) ^{2}}{4\omega ^{3/2}}=\frac{\left( \psi ^{\prime }\right)
^{2}}{\psi }>0, \label{psign} \\
\psi \left( r\right) &\geq &a_{1}\sqrt{r},\qquad \text{where $a_{1}=\sqrt{%
a_{0}}$,} \label{psiline} \\
\lim_{r\rightarrow 0+}\psi ^{\prime }\left( r\right) &=&\lim_{r\rightarrow
0+}\frac{\omega ^{\prime }\left( r\right) }{2\sqrt{\omega \left( r\right) }}%
=+\infty \text{.} \label{blow}
\end{eqnarray}%
For $t>0$, let $\mathcal{N}_{t}=\left\{ y_{n}<t\right\} \bigcap \left\{
|y|<r_{0}\right\} $. Because of (\ref{hellip}) (recall that $k^{1}=1$) and
continuity of $\mathcal{A}$, there exist $1\leq \ell \leq n$ and $c_{1}>0$
such that for all small $t$,
\begin{equation}
\vec{\theta}_{\ell }\mathcal{A}\left( y,z\right) \left( \vec{\theta}_{\ell
}\right) ^{\prime }\geq c_{1}>0,\qquad y\in \partial \Phi \bigcap \mathcal{N}%
_{t},\quad \quad \left\vert z\right\vert \leq 2m_{0}. \label{ne}
\end{equation}%
Here $\vec{\theta}_{\ell }^{\prime }$ is the $\ell ^{\text{th}}$ column of $%
\Theta $. For $m_{1}>0$ and $0<t_{1}\leq 1$ to be determined, define
\begin{equation}
h\left( y\right) =-2\psi \left( \sqrt{\rho y_{n}}\right) +m_{1}\frac{y_{\ell
}^{2}}{2}+\frac{1}{\ln y_{n}},\qquad y\in \mathcal{N}_{t_{1}}, \label{barrh}
\end{equation}%
where $\rho =\left( \kappa _{0}^{-\frac{1}{2}}+1\right) ^{2}$. For $t_{1}$
small enough $h$ is well-defined in $\mathcal{N}_{t_{1}}$ and extends
continuously to $\overline{\mathcal{N}_{t_{1}}}$ with
\begin{equation}
h\in \mathcal{C}^{\infty }\left( \mathcal{N}_{t_{1}}\right) \bigcap \mathcal{%
\ C}^{0}\left( \overline{\mathcal{N}_{t_{1}}}\right) ,\qquad h\left(
0\right) =0. \label{barr99}
\end{equation}%
From (\ref{bdry}),
\begin{eqnarray}
\sqrt{\rho y_{n}} &=&\left( \kappa _{0}^{-\frac{1}{2}}+1\right) \sqrt{y_{n}}=%
\sqrt{\frac{y_{n}}{\kappa _{0}}}+\sqrt{y_{n}} \notag \\
&\geq &\left\vert y^{\prime }\right\vert +y_{n}\geq \left\vert y\right\vert ,
\label{rho-y}
\end{eqnarray}%
where we used that $y_{n}\leq 1$ in $\mathcal{N}_{t_{1}}$.
Also, by taking $t_{1}$ small enough depending on $\kappa _{0}$, $a_{1}$, $%
m_{1} $, we obtain from (\ref{bdry}) and (\ref{psiline}) that
\begin{eqnarray*}
m_{1}\frac{y_{\ell }^{2}}{2}-\psi \left( \sqrt{\rho y_{n}}\right) &\leq
&m_{1}\frac{y_{n}}{2\kappa _{0}}-a_{1}\sqrt[4]{\rho y_{n}} \\
&\leq &m_{1}\frac{\sqrt[4]{y_{n}}}{2\kappa _{0}}\left( t_{1}^{3/4}-\frac{
2\kappa _{0}a_{1}\sqrt{\kappa _{0}^{-\frac{1}{2}}+1}}{m_{1}}\right) \leq 0.
\end{eqnarray*}
Then, using (\ref{rho-y}), the fact that $\psi $ is increasing, the
definition of $\psi$ and (\ref{psione}), we get
\begin{eqnarray}
h\left( y\right) +\omega \left( \left\vert y\right\vert \right) &=&-2\psi
\left( \sqrt{\rho y_{n}}\right) +m_{1}\frac{y_{\ell }^{2}}{2}+\frac{1}{\ln
y_{n}}+\omega \left( \left\vert y\right\vert \right) \notag \\
&\leq &\omega \left( \left\vert y\right\vert \right) -\psi \left( \sqrt{\rho
y_{n}}\right) \notag \\
&\leq &\psi ^{2}\left( \left\vert y\right\vert \right) -\psi \left(
\left\vert y\right\vert \right) <0. \label{bigger}
\end{eqnarray}
Now set
\begin{equation*}
G_{n}\left( y\right) =\left( \frac{\rho }{y_{n}}\right) ^{\frac{1}{2}}\psi
^{\prime }\left( \sqrt{\rho y_{n}}\right) +\frac{1}{\left( \ln y_{n}\right)
^{2}y_{n}}>0,
\end{equation*}
and write $h_{j}=\frac{\partial h}{\partial y_{i}}$, $h_{ij}=\frac{ \partial
^{2}h}{\partial y_{i}\partial y_{j}}$. If $\delta _{ij}$ denotes the
Kronecker delta, we have
\begin{eqnarray}
h_{n}\left( y\right) &=&-G_{n}+\delta _{\ell n}m_{1}y_n,\qquad h_{\ell
}\left( y\right) =m_{1}y_{\ell }-\delta _{\ell n}G_{n} \label{derh} \\
h_{i}\left( y\right) &=&0\qquad \text{for }i\neq n,\ell \label{derhz} \\
h_{\ell \ell }\left( y\right) &=& m_{1}+ \delta_{\ell n} \dfrac{1}{2y_{n}}%
\left[ G_{n}\left( y\right) -\rho \psi ^{\prime \prime }\left( \sqrt{\rho
y_{n}}\right) \right] \\
&& +\delta_{\ell n}\frac{1}{\left( \ln y_{n}\right) ^{2}y_{n}^{2}}\left(
\dfrac{2}{\ln y_{n}}+\dfrac{1}{2}\right) \label{h11} \\
h_{nn}\left( y\right) &=&\dfrac{1}{2y_{n}}\left[ G_{n}\left( y\right) -\rho
\psi ^{\prime \prime }\left( \sqrt{\rho y_{n}}\right) \right] \label{hnn} \\
&&+\frac{1}{\left( \ln y_{n}\right) ^{2}y_{n}^{2}}\left( \dfrac{2}{\ln y_{n}}
+\dfrac{1}{2}\right) +\delta _{\ell n}m_{1} \notag \\
h_{ii}\left( y\right) &=&0 \quad \text{for }i\neq n,\ell ,\qquad
h_{ij}\left( y\right) =0 \quad \text{for }i\neq j. \label{derhnn}
\end{eqnarray}
In particular, for $t_1$ small enough,
\begin{equation}
\begin{array}{lll}
\bigtriangleup h\left( y\right) & = & \dfrac{1}{2y_{n}}\left[ G_{n}\left(
y\right) -\rho \psi ^{\prime \prime }\left( \sqrt{\rho y_{n}}\right) \right]
+\dfrac{1}{\left( \ln y_{n}\right) ^{2}y_{n}^{2}}\left( \dfrac{2}{\ln y_{n}}%
+ \dfrac{1}{2}\right) +m_{1}>0%
\end{array}
\label{lapp}
\end{equation}
in $\mathcal{N}_{t_{1}}$ because of (\ref{psign}). Moreover, by (\ref{psign}%
) and the formulas for derivatives of $h$,
\begin{eqnarray}
\left\vert h_{\ell }\left( y\right) \right\vert ^{2} &\leq
&2m_{1}^{2}y_{\ell }^{2}+2\delta _{\ell n}G_{n}^{2} \notag \\
&\leq & c\max \left\{ m_{1}y_{\ell }^{2}, \, \delta _{\ell n}\psi \left(
\sqrt{ \rho y_{n}}\right) ,\frac{\delta _{\ell n}}{\left( \ln y_{n}\right)
^{2}} \right\} h_{\ell \ell }\left( y\right), \label{hl-s}
\end{eqnarray}
\begin{eqnarray}
\left\vert h_{n}\left( y\right) \right\vert ^{2} &\leq &2G_{n}^{2}+2\delta
_{\ell n}m_{1}^{2}y_n^{2} \notag \\
&\leq & c\max \left\{ \psi \left( \sqrt{\rho y_{n}}\right) ,\frac{1}{\left(
\ln y_{n}\right) ^{2}}, \delta _{\ell n}m_{1}y_n^{2}\right\} h_{nn}\left(
y\right) . \label{hn-s}
\end{eqnarray}
Let $H(x)= h(\Theta x)$ and $\tilde{H}(x) = H(x) +m = h(\Theta x) +m$ for
fixed $m\in [-m_0,m_0]$. Using $\Theta^t$ to denote the transpose of $\Theta$%
, we then have $\nabla H(x)= \Theta^t (\nabla h)(\Theta x)$, $H_j(x) =
\sum_k \theta_{jk}h_k(\Theta x)$ and
\begin{equation*}
H_{ij}(x) = \sum_k \theta_{jk} \sum_\lambda \theta_{i\lambda}
h_{k\lambda}(\Theta x) = \sum_k \theta_{jk} \theta_{ik} h_{kk}(\Theta x)
\end{equation*}
\begin{equation*}
= \theta_{j\ell}\theta_{i\ell} h_{\ell \ell}(\Theta x) +
\theta_{jn}\theta_{in} h_{nn}(\Theta x),
\end{equation*}
where only one of the last two terms appears in case $\ell =n$.
Setting $\mathcal{A}(x,\tilde{H}(x))=(a_{ij}(x))$ and letting $\left[ %
\mathop{\rm div}\mathcal{A}(x,\tilde{H}(x))\right] =\left[ \mathop{\rm div}%
(a_{ij}(x))\right] $ denote the vector whose components are the divergence
of the columns of $\big(a_{ij}(x)\big)$, we obtain
\begin{eqnarray*}
\mathcal{L}_{m}h &=&\mathop{\rm div}\left[ \mathcal{A}\left( x,\tilde{H}%
(x)\right) \nabla H(x)\right] =\mathop{\rm div}\left(
\sum_{i}a_{ij}(x)H_{i}(x)\right) _{j=1,\dots ,n} \\
&=&\left[ \mathop{\rm div}(a_{ij}(x))\right] \cdot \nabla
H(x)+\sum_{i,j}a_{ij}(x)H_{ij}(x)
\end{eqnarray*}%
\begin{equation*}
=\left[ \mathop{\rm div}(a_{ij}(x))\right] \cdot \Theta ^{t}(\nabla
h)(\Theta x)+h_{\ell \ell }(\Theta x)\sum_{i,j}a_{ij}(x)\theta _{j\ell
}\theta _{i\ell }+h_{nn}(\Theta x)\sum_{i,j}a_{ij}(x)\theta _{jn}\theta _{in}
\end{equation*}%
\begin{equation*}
=\left[ \mathop{\rm div}(a_{ij}(x))\right] \cdot \Theta ^{t}(\nabla
h)(\Theta x)+h_{\ell \ell }(\Theta x)\vec{\theta}_{\ell }\mathcal{A}(x,%
\tilde{H}(x))\vec{\theta}_{\ell }^{~t}+h_{nn}(\Theta x)\vec{\theta}_{n}%
\mathcal{A}(x,\tilde{H}(x))\vec{\theta}_{n}^{~t}.
\end{equation*}%
Here $\vec{\theta}_{j}^{t}$ denotes the $j^{\text{th}}$ column of $\Theta $
and $\vec{\theta}_{j}$ denotes the $j^{\text{th}}$ column of $\Theta $
rewritten as a row vector, and in the last two equalities, only one of the
second and third terms on the right appears in case $\ell =n$.
Now recall that $h_{i}=0$ if $i\neq \ell ,n$. Then the first term on the
right of the last equality equals
\begin{equation*}
h_{\ell }(\Theta x)\sum_{k}\left( \sum_{i}\frac{\partial }{\partial x_{i}}%
a_{ik}(x)\right) \theta _{k\ell }+h_{n}(\Theta x)\sum_{k}\left( \sum_{i}%
\frac{\partial }{\partial x_{i}}a_{ik}(x)\right) \theta _{kn}
\end{equation*}%
\begin{equation*}
=h_{\ell }(\Theta x)\left[ \mathop{\rm div}\mathcal{A}(x,\tilde{H}(x))\right]
\cdot \vec{\theta}_{\ell }^{~t}+h_{n}(\Theta x)\left[ \mathop{\rm div}%
\mathcal{A}(x,\tilde{H}(x))\right] \cdot \vec{\theta}_{n}^{~t},
\end{equation*}%
where as usual only one of the two terms on the right appears in case $\ell
=n$.
Altogether,
\begin{eqnarray*}
\mathcal{L}_{m}H(x) &=&h_{\ell }(\Theta x)\left[ \mathop{\rm div}\mathcal{A}%
\left( x,\tilde{H}(x)\right) \right] \cdot \vec{\theta}_{\ell
}^{~t}+h_{n}(\Theta x)\left[ \mathop{\rm div}\mathcal{A}\left( x,\tilde{H}%
(x)\right) \right] \cdot \vec{\theta}_{n}^{~t} \\
&&+h_{\ell \ell }(\Theta x)\vec{\theta}_{\ell }\mathcal{A}\left( x,\tilde{H}%
(x)\right) \vec{\theta}_{\ell }^{~t}+h_{nn}(\Theta x)\vec{\theta}_{n}%
\mathcal{A}\left( x,\tilde{H}(x)\right) \vec{\theta}_{n}^{~t},
\end{eqnarray*}%
where on the right side, only one of the first two terms and one of the
second two terms appears in case $\ell =n$.
By direct computation, we have
\begin{eqnarray*}
\left[ \mathop{\rm div}\mathcal{A}(x,\tilde{H}(x))\right] \cdot \vec{\theta}%
_{r}^{~t} &=&\sum_{i=1}^{n}\mathcal{A}_{i}^{i}(x,\tilde{H}(x))\cdot \vec{%
\theta}_{r}^{~t}+h_{\ell }(\Theta x)\vec{\theta}_{\ell }\mathcal{A}_{z}(x,%
\tilde{H}(x))\vec{\theta}_{r}^{~t} \\
&+&h_{n}(\Theta x)\vec{\theta}_{n}\mathcal{A}_{z}(x,\tilde{H}(x))\vec{\theta}%
_{r}^{~t},
\end{eqnarray*}%
where $\mathcal{A}^{i}$ denotes the $i^{th}$-column of $\mathcal{A}$ and $%
\mathcal{A}_{i}^{i}=\partial _{i}\mathcal{A}^{i}$, and where only one of the
last two terms on the right side appears in case $\ell =n$. Substituting
this in the formula above for $\mathcal{L}_{m}H$ gives
\begin{eqnarray*}
\mathcal{L}_{m}H(x) &=&\mathcal{L}_{m}[h(\Theta x)]=h_{\ell }(\Theta
x)\sum_{i=1}^{n}\mathcal{A}_{i}^{i}(x,\tilde{H}(x))\cdot \vec{\theta}_{\ell
}^{~t}+h_{\ell }(\Theta x)^{2}\vec{\theta}_{\ell }\mathcal{A}_{z}(x,\tilde{H}%
(x))\vec{\theta}_{\ell }^{~t} \\
&&+h_{\ell }(\Theta x)h_{n}(\Theta x)\vec{\theta}_{n}\mathcal{A}_{z}(x,%
\tilde{H}(x))\vec{\theta}_{\ell }^{~t}+h_{\ell }(\Theta x)h_{n}(\Theta x)%
\vec{\theta}_{\ell }\mathcal{A}_{z}(x,\tilde{H}(x))\vec{\theta}_{n}^{~t} \\
&&+h_{n}(\Theta x)\sum_{i=1}^{n}\mathcal{A}_{i}^{i}(x,\tilde{H}(x))\cdot
\vec{\theta}_{n}^{~t}+h_{n}(\Theta x)^{2}\vec{\theta}_{n}\mathcal{A}_{z}(x,%
\tilde{H}(x))\vec{\theta}_{n}^{~t} \\
&&+h_{\ell \ell }(\Theta x)\vec{\theta}_{\ell }\mathcal{A}\left( x,\tilde{H}%
(x)\right) \vec{\theta}_{\ell }^{~t}+h_{nn}(\Theta x)\vec{\theta}_{n}%
\mathcal{A}\left( x,\tilde{H}(x)\right) \vec{\theta}_{n}^{~t},
\end{eqnarray*}%
and in case $\ell =n$, only the first, second and seventh terms on the right
appear. By (\ref{wirtm}), the sum of the first and fifth terms on the right
is at most
\begin{equation*}
cB_{\mathcal{A}}^{2}+\dfrac{h_{\ell }(\Theta x)^{2}}{4}\vec{\theta}_{\ell }%
\mathcal{A}(x,\tilde{H}(x))\vec{\theta}_{\ell }^{~t}+\dfrac{h_{n}(\Theta
x)^{2}}{4}\vec{\theta}_{n}\mathcal{A}(x,\tilde{H}(x))\vec{\theta}_{n}^{~t},
\end{equation*}%
and the sum of the second, third, fourth and sixth terms is bounded by
\begin{equation*}
cB_{\mathcal{A}}\left( h_{\ell }(\Theta x)^{2}\vec{\theta}_{\ell }\mathcal{A}%
(x,\tilde{H}(x))\vec{\theta}_{\ell }^{~t}+h_{\ell }(\Theta x)^{2}\vec{\theta}%
_{n}\mathcal{A}(x,\tilde{H}(x))\vec{\theta}_{n}^{~t}\right) ,
\end{equation*}%
where {$B_{\mathcal{A}}=B_{\mathcal{A}}\left( \Phi ,m_{0}\right) $. }Using
these estimates, we obtain
\begin{eqnarray*}
\mathcal{L}_{m}H(x) &\geq &-cB_{\mathcal{A}}^{2}+\left[ h_{\ell \ell
}(\Theta x)-\left( cB_{\mathcal{A}}+\dfrac{1}{4}\right) h_{\ell }(\Theta
x)^{2}\right] ~\vec{\theta}_{\ell }\mathcal{A}\left( x,\tilde{H}(x)\right)
\vec{\theta}_{\ell }^{~t} \\
&&+\left[ h_{nn}(\Theta x)-\left( cB_{\mathcal{A}}+\dfrac{1}{4}\right)
h_{n}(\Theta x)^{2}\right] ~\vec{\theta}_{n}\mathcal{A}\left( x,\tilde{H}%
(x)\right) \vec{\theta}_{n}^{~t}.
\end{eqnarray*}%
From (\ref{hl-s}) and (\ref{hn-s}), by taking $t_{1}$ small enough
(depending on $\kappa _{0}$, $m_{1}$, and $B_{\mathcal{A}}$), we obtain
\begin{equation*}
\mathcal{L}_{m}H(x)\geq -cB_{\mathcal{A}}^{2}+\dfrac{h_{\ell \ell }(\Theta x)%
}{2}~\vec{\theta}_{\ell }\mathcal{A}\left( x,\tilde{H}(x)\right) \vec{\theta}%
_{\ell }^{~t}.
\end{equation*}%
Then from (\ref{ne}) and the fact that $h_{\ell \ell }\geq m_{1},$
\begin{equation}
\mathcal{L}_{m}h\geq -cB_{\mathcal{A}}^{2}+\dfrac{c_{1}m_{1}}{2}>K
\label{PK}
\end{equation}%
by taking $m_{1}$ large enough.
Finally, note that $y_{\ell }^{2}\leq y_{n}/\kappa _{0}$ by (\ref{bdry}).
Then if $x\in \Phi \bigcap \partial \mathcal{N} _{t_{1}} $,
\begin{eqnarray}
h\left( y\right) &=&-2\psi \left( \sqrt{\rho y_{n}}\right) +m_{1}\frac{
y_{\ell }^{2}}{2}+\frac{1}{\ln y_{n}} \notag \\
&\leq &\frac{m_{1}}{2\kappa _{0}}t_{1}^{2}+\frac{1}{\ln t_{1}}\leq -\nu
\label{Pnu}
\end{eqnarray}
by taking $t_{1}$ small enough. The condition $\mathcal{N}\subset \left\{
\left\vert x-x_{0}\right\vert <\eta \right\} $ can then be met by taking $%
t_{1}$ even smaller. Lemma \ref{barrier} now follows from (\ref{barr99}), (%
\ref{bigger}), (\ref{lapp}), (\ref{PK}) and (\ref{Pnu}).{\
\endproof%
}
|
\section{Introduction}
Threshold neutral pion photo- and electroproduction off the nucleon is one
of the finest reactions to test the chiral QCD dynamics, see
\cite{Bernard:2007zu}
for a recent review. While elementary proton targets are accessible
directly in experiment,
pion production off neutrons requires the use of nuclear targets
like the deuteron or three-nucleon bound states like $^3$H (triton) or $^3$He.
For a recent review on revealing the neutron structure from electron or photon
scattering off light nuclei, see \cite{Phillips:2009af}.
Of particular interest is to test the counterintuitive chiral perturbation
theory prediction (CHPT) that the elementary neutron S-wave multipole
$E_{0+}^{\pi^0
n}$ is larger in magnitude than the corresponding one of the proton,
$E_{0+}^{\pi^0 p}$ \cite{Bernard:1994gm,Bernard:2001gz}. This prediction
was already successfully tested in neutral pion photo- \cite{Beane:1997iv}
and electroproduction off the deuteron \cite{Krebs:2004ir}.
However, given the
scarcity and precision of the corresponding data, it is mandatory to study
also pion production off three-nucleon bound states, that can be calculated
nowadays to high precision based on chiral nuclear effective field theory
(EFT), that extends CHPT to nuclear physics (for a recent review, see
\cite{Epelbaum:2008ga}). $^3$He appears to be a particularly promising target
to extract the information about the neutron amplitude. Its wave function is
strongly dominated by the principal ``s-state'' component which suggests that
the spin of $^3$He is largely driven by the one of the neutron.
Consequently, in this letter we calculate
thres\-hold pion photo- and electroproduction based on chiral 3N wave
functions at next-to-leading order in the chiral expansion. Experimentally,
neutral pion photoproduction off light nuclei has so far only been studied
at Saclay~\cite{Argan:1980zz,Argan:1987dm} and at Saskatoon
\cite{Bergstrom:1998zz,Barnett:2008zz}.
In general, one has three different topologies for pion production off a
three-nucleon bound state as shown in Fig.~\ref{fig:123N}.
While the single-nucleon contribution (a) features the
elementary neutron and proton production amplitudes, the
nuclear corrections are given by two-body (b) and three-body
(c) terms. Based on the power counting developed in \cite{Beane:1995cb},
at next-to-leading order (NLO), only the to\-po\-lo\-gies (a) and (b)
contribute. Here, we will specifically consider threshold photo- and
electroproduction parameterized in terms of the electric $E_{0+}$
and longitudinal $L_{0+}$ S-wave multipoles. In particular, we
will study the sensitivity of the three-body S-wave multipoles
to the elementary $E_{0+}^{\pi^0 n}$ multipole, taking the proton
amplitude $E_{0+}^{\pi^0 p}$ from CHPT (as this value is consistent
with the data \cite{Schmidt:2001vg} and a recent study based on a
chiral unitary approach \cite{Gasparyan:2010xz}).
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.7\linewidth]{piprod3N.eps}
\end{center}
\vspace{-0.2cm}
\caption{Different topologies contributing to pion
production off the three-nucleon bound state (triangle).
(a), (b) and (c) represent the single-, two- and three-nucleon
contributions, respectively. Solid, dashed and wiggly lines
denote nucleons, pions and photons, in order. Topology (c)
does not contribute to the order considered here (NLO).}
\label{fig:123N}
\end{figure}
\section{Anatomy of the Calculation}
To analyze the process under consideration, we calculate
the the nuclear matrix element of the given
transition operator $\hat{O}$ as:
\begin{equation}\label{eq:OME}
\langle M_J^\prime|\hat{O}|M_J\rangle_{\psi}
:=\langle \psi M_J^\prime \vec{P}^\prime_{3N}\vec{q}_\pi|\hat{O}|\psi
M_J \vec{P}_{3N}\vec{k}_\gamma \rangle\, ,
\end{equation}
where $\psi$ refers to the three-nucleon wave function and
$\vec{k}_\gamma$, $\vec{q}_\pi$, $\vec{P}_{3N}$ and $\vec{P}^\prime_{3N}$
denote the
momentum of the exchanged (virtual) photon, produced pion and the initial
and final momentum of the 3N nucleus, respectively.
The $3N$ bound state has total nuclear angular momentum $J=1/2$ with magnetic
quantum numbers $M_J$ for the initial and $M_J^\prime$ for the final nuclear
state. $J$ can be decomposed in total spin $S=1/2,3/2$ and total orbital
angular momentum $L=0,1,2$. The total isospin is a mixture of two components,
$T=1/2$ and $3/2$. While the $T=1/2$ component is large, the small $T=3/2$
component emerges due to isospin breaking and is neglected in
our calculation.
Here, we consider neutral pion production by real or virtual photons
of a spin-1/2 particle - either the nucleon or the $^3$H and $^3$He nuclei.
At threshold, the corresponding transition matrix takes the form
\begin{equation}
\label{eq:multidef}
\mathcal{M}_\lambda =
2i\, E_{0+} \, (\vec{\epsilon}_{\lambda, \text{T}}\cdot\vec{S} )
+ 2i \, L_{0+}\, (\vec{\epsilon}_{\lambda, \text{L}}\cdot\vec{S}) \,,
\end{equation}
with $\vec{\epsilon}_{\lambda, \text{T}} = \vec{\epsilon}_\lambda
-(\vec{\epsilon}_\lambda\cdot\hat{k}_\gamma)\hat{k}_\gamma$ and $\vec{\epsilon}_{\lambda,
\text{L}} = (\vec{\epsilon}_\lambda\cdot\hat{k}_\gamma)\hat{k}_\gamma$ the transverse and
longitudinal photon polarization vectors. The transverse and
longitudinal S-wave multipoles are denoted by $ E_{0+}$ and $L_{0+}$,
respective\-ly. Note that $L_{0+}$ contributes only for virtual photons.
As explained before, the matrix element Eq.~(\ref{eq:OME}) receives
contributions from one- and two-nucleon operators at the order we are working.
Consider first the single
nucleon contribution, given in terms of the 1-body transition operator
$\hat{O}^{\rm 1N}$. After some algebra, one finds
\begin{align}\label{eq:O1fin}
&\langle M_J^\prime|\hat{O}^{\rm 1N}|M_J\rangle_{\psi} \nonumber\\
& =
i \vec{\epsilon}_{\lambda,\text{T}}\cdot\vec{S}_{M_J^\prime
M_J}\Big( E_{0+}^{\pi^0p} F_T^{S+V} + E_{0+}^{\pi^0n} F_T^{S-V}\Big)\nonumber\\
&+ i \vec{\epsilon}_{\lambda,\text{L}}\cdot\vec{S}_{M_J^\prime
M_J}\Big( L_{0+}^{\pi^0p}
F_L^{S+V}+ L_{0+}^{\pi^0n} F_L^{S-V}\Big)~.
\end{align}
where $F_{T/L}^{S\pm V}\equiv F_{T/L}^S\pm F_{T/L}^V$ and
$F_{T/L}^{S,V}$ denote the corresponding form factors of the
3N bound state,
\begin{align}
\label{eq:F1N}
F_{T/L}^S \, \vec{\epsilon}_{\lambda,\text{T/L}}\cdot\vec{S}_{M_J^\prime
M_J}
&= \frac{3}{2} \, \langle M_J^\prime
| \vec{\epsilon}_{\lambda,\text{T/L}}\cdot\vec{\sigma}_{1}
|
M_J\rangle_{\psi}~, \\
F_{T/L}^V \, \vec{\epsilon}_{\lambda,\text{T/L}}\cdot\vec{S}_{M_J^\prime
M_J}
&= \frac{3}{2} \, \langle M_J^\prime
|\vec{\epsilon}_{\lambda,\text{T/L}}\cdot\vec{\sigma}_{1}\tau_{1}^z|
M_J\rangle_{\psi}~, \nonumber
\end{align}
which parametrize the overall normalization of the
response of the
composite system to the excitation by photons in spin-isospin space.
In the above equation, $\vec \sigma_{i}$ ($\vec \tau_i$) denote the spin
(isospin) Pauli matrices corresponding to the nucleon $i$. Furthermore, $z$
refers to the isospin quantization axis.
Using the 3N wave functions from chiral nuclear EFT at the appropriate order,
the pertinent matrix elements in Eq.~(\ref{eq:F1N}) can be evaluated.
Here, we use chiral 3N wave functions obtained from the N$^2$LO
interaction in the Weinberg power counting
\cite{Epelbaum:2003gr,Epelbaum:2003xx}.\footnote{The consistency of the
Weinberg counting for short-range operators and
the non-perturbative renormalization of chiral EFT
are currently under discussion, see the
review~\cite{Epelbaum:2008ga} for more details.
A real alternative to the Weinberg approach for practical
calculations, however, is not available.}
In order to estimate the
error from higher order corrections, we use wave functions
for five different combinations of the
cutoff $\tilde{\Lambda}$ in the spectral function regularization
of the two-pion exchan\-ge and the cutoff $\Lambda$ used to regularize the
Lipp\-mann-Schwinger equation for the two-body T-ma\-trix.
The wave functions are taken from Ref.~\cite{Liebig:2010ki,Noggaprivate}
and the corresponding cutoff combinations in units of MeV are
$(\tilde{\Lambda},\Lambda)$ = (450,500), (600,500),
(550,600), (450,700), (600,700).
All five sets describe the binding energies of the
$^3$He and $^3$H nuclei equally well.
The one-body contributions to the 3N multipoles are given by
\begin{align}\label{eq:mult1N}
E_{0+}^{\rm 1N}
& = \frac{K_{\rm 1N}}{2} \left( E_{0+}^{\pi^0 p} \, F_T^{S+V} + E_{0+}^{\pi^0 n}
\, F_T^{S-V}\right)~,\nonumber\\
L_{0+}^{\rm 1N}
& = \frac{K_{\rm 1N}}{2} \left( L_{0+}^{\pi^0p} \, F_L^{S+V} + L_{0+}^{\pi^0n}
\, F_L^{S-V} \right)~.
\end{align}
Here, $K_{1N}$ is the kinematical factor to account for the change in phase
space from the 1N to the 3N system,
\begin{equation}\label{eq:kine1N}
K_{\rm 1N}
= \frac{m_N+M_\pi}{m_{\rm 3N}+M_\pi}\frac{m_{\rm 3N}}{m_N}
\approx 1.092\,,
\end{equation}
with $m_N$ being the nucleon mass and
$m_{\rm 3N}$ the mass of the three-nucleon bound state.
We evaluate the matrix elements for the one-body contribution
in Eq.~(\ref{eq:F1N}) numerically
with Monte Carlo integration using the VEGAS algorithm~\cite{Vegas}.
The results for the form factors $F_{T/L}^{S\pm V}$ are given
in Table~\ref{tab:FTL}.
\renewcommand{\arraystretch}{1.3}
\begin{table}[t]
\begin{center}
\begin{tabular}{|c||c|c|}
\hline
nucleus & $^3$He & $^3$H \\
\hline
$F_{T}^{S+V}$ & $0.017(13)(3)$ & $1.493(25)(3)$ \\
$F_{T}^{S-V}$ & $1.480(26)(3)$ & $0.012(13)(3)$ \\
$F_{L}^{S+V}$ & $-0.079(14)(8)$ & $1.487(27)(8)$ \\
$F_{L}^{S-V}$ & $1.479(26)(8)$ & $-0.083(14)(8)$ \\
\hline
\end{tabular}
\caption{Numerical results for the form factors $F_{T/L}^{S\pm V}$.
The first error is an estimate of the theory error
from higher orders in chiral EFT while the second error
is the statistical error from the Monte Carlo integration.}
\label{tab:FTL}
\end{center}
\end{table}
The first error represents the theoretical uncertainty estimated from the
cutoff variation
in the wave functions. We take the central value defined by the five
different cutoff sets as our prediction and estimate the theory error
from higher-order corrections
from the spread of the calculated values. Strictly speaking, this procedure
gives a lower bound on the error, but in practice it generates a
reasonable estimate.
We stress that we follow the nuclear EFT formulation of Lepage,
in which the whole effective potential is iterated to all orders when
solving the Schr\"odinger equation for the nuclear states.
As discussed in Ref.~\cite{Lepage:1997cs}, the cutoff should be kept
of the order of the breakdown scale or below
in order to avoid unatural scaling
of the coefficients of higher order terms. Indeed, using larger cutoffs
can lead to a violation of certain low-energy theorems as demonstrated in
Ref.~\cite{Epelbaum:2009sd} for an exactly solvable model.
The error related to the expansion of the
production operator is difficult to estimate given that the convergence in the
expansion for the single nucleon S-wave multipoles is known to be slow, see
Ref.~\cite{Bernard:1994gm} for an extended discussion.
We therefore give here only a rough estimate of this uncertainty. The
extractions of the proton S-wave photoproduction amplitude based
on CHPT using various approximations \cite{FernandezRamirez:2009jb}
lead to an uncertainty $\Delta E_{0+}^{\pi^0 p} \approx \pm 0.05\times
10^{-3}/M_{\pi^+}$, which is about 5\%. Similarly, we estimate the
uncertainty of the neutron S-wave threshold amplitude to be the same.
Consequently, our estimate of the error on the single nucleon amplitude
is 5\%.
The statistical error from the evaluation of the
integrals is typically one order of magnitude
smaller than the estimated theory error and can be neglected.
We now switch to the two-nucleon contribution.
In Coulomb gauge, only the two Feynman diagrams shown in
Fig.~\ref{fig:abterm} contribute at threshold to the order
we are working~\cite{Beane:1997iv}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.8\linewidth]{abterm.eps}
\end{center}
\vspace{-0.2cm}
\caption{Leading two-nucleon contributions to the
nuclear pion production matrix element at threshold.
Solid, dashed and wiggly lines
denote nucleons, pions and photons, in order.}
\label{fig:abterm}
\end{figure}
Their contribution to the multipoles can be written as
\begin{eqnarray}\label{eq:mult2N}
E_{0+}^{\rm 2N} &=& K_{\rm 2N} \, (F_T^{(a)} -F_T^{(b)})~,\nonumber\\
L_{0+}^{\rm 2N} &=& K_{\rm 2N} \, (F_L^{(a)} -F_L^{(b)})~,
\end{eqnarray}
with the prefactor
\begin{eqnarray}
K_{\rm 2N} &=& \frac{M_\pi e g_A m_{\rm 3N}}{16\pi(m_{\rm 3N}+M_\pi)
(2\pi)^3 F_\pi^3}\nonumber\\
&\approx& 0.135\mbox{ fm }\times 10^{-3}/M_{\pi^+}~.
\end{eqnarray}
The numerical value for $K_{\rm 2N}$
was obtained using $g_A=1.26$ for the axial
coupling constant, $F_\pi=93$ MeV for the pion decay constant, and the
neutral pion mass $M_\pi=135$ MeV.
The transverse and longitudinal form factors $F_{T/L}^{(a)}$
and $F_{T/L}^{(b)}$ corresponding to diagrams (a) and (b), respectively,
are
\begin{align}
&F_{T/L}^{(a)} \, \vec{\epsilon}_{\lambda,\text{T/L}}\cdot\vec{S}_{M_J^\prime
M_J} \nonumber\\
&= \frac{3}{2}\,
\langle M_J'|\frac{\vec{\epsilon}_{\lambda,\text{T/L}}\cdot (\vec \sigma_1
+ \vec \sigma_2)
(\vec{\tau}_1\cdot\vec{\tau}_2-
\tau_1^z\tau_2^z)}{\left(\vec{p}_{12}-\vec{p}_{12}^{\;\prime}+
\vec{k}_\gamma/2\right)^2}|
M_J\rangle_\psi~, \nonumber\\
\label{eq:valF2Na}
\end{align}
and
\begin{align}
&F_{T/L}^{(b)} \, \vec{\epsilon}_{\lambda,\text{T/L}}\cdot\vec{S}_{M_J^\prime
M_J} =
3 \, \langle M_J' |\, (\vec{\tau}_1\cdot\vec{\tau}_2-\tau_1^z\tau_2^z)
\nonumber\\
&\times \frac{\big[(\vec{p}_{12}-\vec{p}_{12}^{\;\prime}-
\vec{k}_\gamma/2)\cdot (\vec \sigma_1 + \vec \sigma_2 ) \big] }
{\big[(\vec{p}_{12}-\vec{p}_{12}^{\;\prime}- \vec{k}_\gamma/2 )^2+M_\pi^2 \big]
\big[\vec{p}_{12}-\vec{p}_{12}^{\;\prime}+\vec{k}_\gamma/2 \big]^2} \nonumber \\
&\times \big[
\vec{\epsilon}_{\lambda,\text{T/L}}
\cdot (\vec{p}_{12}-\vec{p}_{12}^{\;\prime}) \big]\,
|M_J\rangle_\psi~,
\label{eq:valF2Nb}
\end{align}
where $\vec{p}_{12} =(\vec{k}_1-\vec{k}_2)/2$ and
$\vec{p}_{12}^\prime=(\vec{k}_1^\prime-\vec{k}_2^\prime)/2$
are the initial and final Jacobi momenta
of nucleons 1 and 2, respectively. The integral for the
form factors $F_{T/L}^{(a)}$ contains an integrable singularity which can
be removed by an appropriate variable transformation. Then,
the form factors can
be evaluated using Monte Carlo integration in the same way as the
form factors for the single-nucleon contribution.
Our results for $F_{T/L}^{(a)}-F_{T/L}^{(b)}$ are given
in Table~\ref{tab:Fab}.
\renewcommand{\arraystretch}{1.3}
\begin{table}[t]
\begin{center}
\begin{tabular}{|c||c|c|}
\hline
nucleus & $^3$He & $^3$H \\
\hline
$F_{T}^{(a)} -F_{T}^{(b)}$ [fm$^{-1}$]
& $-29.3(2)(1)$ & $-29.7(2)(1)$ \\
$F_{L}^{(a)} -F_{L}^{(b)}$ [fm$^{-1}$]
& $-22.9(2)(1)$ & $-23.2(1)(1)$ \\
\hline
\end{tabular}
\caption{Numerical results for the form factors
$F_{T/L}^{(a)} -F_{T/L}^{(b)}$ parametrizing two-body contributions in units of fm$^{-1}$.
The first error is an estimate of the theory error
from higher orders in chiral EFT while the second error
is the statistical error from the Monte Carlo integration.}
\label{tab:Fab}
\end{center}
\end{table}
The first error is again the theory error estimated from the cutoff variation
in the chiral interaction as described above.
The second error is the statistical error from the Monte Carlo integration
which is about half the size of the theory error.
\section{Results and Discussion}
We are now in the position to evaluate the
nuclear S-wave multipoles. They are given as the
sum of the one- and the two-nucleon contributions given
in Eqs.~(\ref{eq:mult1N}, \ref{eq:mult2N}) in the previous section,
\begin{eqnarray}
E_{0+} &=& E_{0+}^{\rm 1N}+E_{0+}^{\rm 2N}~,\nonumber\\
L_{0+} &=& L_{0+}^{\rm 1N}+L_{0+}^{\rm 2N}~.
\end{eqnarray}
Using the values for the one- and two-body form factors
in Tables \ref{tab:FTL} and \ref{tab:Fab} together with
the subleading chiral perturbation theory results for the
single-nucleon multipoles at ${\cal O}(p^4)$
\cite{Bernard:1994gm,Bernard:2001gz}
\begin{eqnarray}
E_{0+}^{\pi^0 p} &=& -1.16 \times 10^{-3}/M_{\pi^+}~,\nonumber\\
E_{0+}^{\pi^0 n} &=& +2.13 \times 10^{-3}/M_{\pi^+}~,\nonumber\\
L_{0+}^{\pi^0 p} &=& -1.35 \times 10^{-3}/M_{\pi^+}~,\nonumber\\
L_{0+}^{\pi^0 n} &=& -2.41 \times 10^{-3}/M_{\pi^+}~,
\label{eq:1Nmultipoles}
\end{eqnarray}
we obtain for the threshold multipoles on $^3$He and on $^3$H
the values in Table~\ref{tab:multip}.
\renewcommand{\arraystretch}{1.3}
\begin{table*}[t]
\begin{center}
\begin{tabular}{|c|c||c|c|c|}
\hline
nucleus & multipole & 1N & 2N & total \\
\hline
$^3$He & $E_{0+}$ [$10^{-3}/M_{\pi^+}$]
& +1.71(4)(9) & -3.95(3) & -2.24(11)\\
& $L_{0+}$ [$10^{-3}/M_{\pi^+}$]
& -1.89(4)(9) & -3.09(2) & -4.98(12)\\
\hline
$^3$H & $E_{0+}$ [$10^{-3}/M_{\pi^+}$]
& -0.93(3)(5) & -4.01(3) & -4.94(7)\\
& $L_{0+}$ [$10^{-3}/M_{\pi^+}$]
& -0.99(4)(5) & -3.13(1) & -4.12(7)\\
\hline
\end{tabular}
\caption{Numerical results for the threshold multipoles $E_{0+}$
and $L_{0+}$ on $^3$He and $^3$H. The error estimates are explained in the
text.}
\label{tab:multip}
\end{center}
\end{table*}
For the 1N contribution, the first error is the theory error
from higher orders in chiral EFT estimated from the cutoff variation
as explained above,
while the second error is from the 5\% uncertainty of the one-nucleon
amplitudes. In the case of the 2N contribution, only the
theory error is given. The total error is obtained by adding the theory
error and the uncertainty of the one-nucleon amplitudes in quadrature.
As noted before, the total error is
dominated by the uncertainty in the single nucleon amplitudes.
We stress that our estimate of the theory error is only a lower bound.
One observes that the multipoles get a large contribution from the
two-body terms.
This behavior is similar to the deuteron case \cite{Beane:1997iv}.
For example, in the case of $E_{0+}$ for $^3$He, the proton contribution is
$-0.01$,
the neutron one is 1.72 while the two-body contribution is $-3.95$
in the canonical units of $ 10^{-3}/M_{\pi^+}$.
There is, however, still a large sensitivity to the
single-neutron contribution.
The corresponding threshold S-wave cross section
for pion photoproduction $a_0$ is given by
\begin{equation}
a_0 =\left.\frac{|\vec{k}_\gamma|}{|\vec{q}_\pi|}
\frac{d\sigma}{d\Omega}\right|_{\vec{q}_\pi=0}=\left|E_{0+}\right|^2~.
\end{equation}
The longitudinal multipole $L_{0+}$ contributes only in
electro-production.
The corresponding threshold cross
section contains an extra term $\sim \left|L_{0+}\right|^2$.
From here on, we will, however, concentrate on photoproduction.
In Fig.~\ref{fig:E0psq}, we illustrate the sensitivity of
$a_0$ to the single-neutron multipole $E_{0+}^{\pi^0 n}$.
\begin{figure}[ht]
\begin{center}
\includegraphics*[width=0.95\linewidth]{E0psensiSQ3.eps}
\end{center}
\vspace{-0.2cm}
\caption{Sensitivity of $a_0$ for $^3$He in units of
$ 10^{-6}/M^2_{\pi^+}$
to the single-neutron multipole $E_{0+}^{\pi^0 n}$
in units of $ 10^{-3}/M_{\pi^+}$. The vertical dashed line
gives the CHPT prediction for $E_{0+}^{\pi^0 n}$ and the
vertical dotted lines indicate the 5\% error in the prediction.
The inner shaded band indicates the theory error estimated
from the cutoff variation as described in the text.
The outer shaded band corresponds to
a 10\% uncertainty in the 2N contribution.}
\label{fig:E0psq}
\end{figure}
The inner shaded band indicates the theory error estimated
from the cutoff variation as described above. The outer shaded band illustrates the
effect of a 10\% uncertainty in the 2N contribution. This corresponds to the
size of the correction to the 2N contribution at the next order in the
deuteron case~\cite{Beane:1997iv}.
The vertical dashed line indicates the CHPT prediction
$E_{0+}^{\pi^0 n}= 2.13 \times 10^{-3}/M_{\pi^+}$.
Changing this value by $\pm20\%$ leads to changes in $a_0$
of about $\pm30\%$.
Thus, the $^3$He nucleus appears to be a very promising target to test
the CHPT prediction for $E_{0+}^{\pi^0 n}$.
On the contrary, neutral pion production on $^3$H is rather
insensitive to $E_{0+}^{\pi^0 n}$:
a variation of $E_{0+}^{\pi^0 n}$ from 0 to 3 changes $a_0$
only by $1\%$.
Next we compare our predictions with the available data.
The consistency of the CHPT prediction for the single-neutron
multipole with the measured S-wave threshold amplitude on the
deuteron from Saclay and Saskatoon
is well established, see Refs.~\cite{Beane:1997iv,Bernard:2007zu}.
The reanalyzed measurement of the S-wave amplitude for $^3$He
at Saclay gives
$E_3 = (-3.5\pm 0.3)\times 10^{-3}/M_{\pi^+}$ \cite{Argan:1980zz,Argan:1987dm},
which is related to $a_0$ according to
\begin{equation}
\left|E_{0+}\right|^2 = |E_3|^2 \left|\frac{F_T^{S-V}}{2} \right|^2 \left(
\frac{1+M_\pi/m_N}{1+M_\pi/3m_N}\right)^2~.
\end{equation}
Here, we have approximated the $A=3$ body form factor ${\cal F}_A$
of Argan et al.~\cite{Argan:1987dm} by the numerically dominant form factor
$F_{T}^{S-V}$
for $^3$He, cf.~Tab.~\ref{tab:FTL}. This results in
\begin{equation}
E_{0+} = (-2.8 \pm 0.2) \times 10^{-3}/M_{\pi^+}~,
\end{equation}
assuming the same sign as for our $^3$He prediction in
Table~\ref{tab:multip}.
In magnitude,
the extracted value is about 25\% above
the predicted one. Given the
model-de\-pen\-dence that is inherent to the analysis of Ref.~\cite{Argan:1987dm},
it is obvious that a more precise measurement using CW beams and modern
detectors is very much called for.
\section{Summary and Outlook}
In this letter, we have presented a calculation of neutral pion
production off $^3$H and $^3$He at threshold to leading one-loop order
for the production operator in the framework of chiral nuclear effective field theory.
We used the chiral wave functions of
Refs.~\cite{Epelbaum:2003gr,Epelbaum:2003xx}
which are consistent with the
pion production operator to calculate the S-wave 3N multipoles $E_{0+}$ and $L_{0+}$.
To this order, the production operator gets both
one- and two-body contributions. Our calculation shows that the two-body
contributions are of the same order of magnitude as the one-body
contributions. A similar behavior was observed in
the deuteron case~\cite{Beane:1997iv}.
The theoretical uncertainty resulting from the cutoff variation in the
employed wave functions appears to be small (of the order of
$3\%$). The dominant theoretical error at this order stems
from the threshold pion production amplitude off the proton and the
neutron, which is estimated to be about 5\%.
We explored the sensitivity
of neutral pion photoproduction on $^3$He to the elementary neutron multipole
$E_{0+}^{\pi^0 n}$ and found a large sensitivity. This makes $^3$He
a promising target to test the counterintuitive CHPT prediction
for $E_{0+}^{\pi^0 n}$~\cite{Bernard:1994gm,Bernard:2001gz}.
The cutoff variation estimate leads to a very small error for the 2N
contribution. If the error of this contribution is artificially
enlarged by a factor of 10, the extraction of the neutron multipole
is still feasible experimentally.
We have shown that our prediction for the $^3$He S-wave multipole
$E_{0+}$ is roughly consistent with the value deduced from the old
Saclay measurement of the threshold cross section \cite{Argan:1987dm}.
A new measurement using modern technology and better
methods to deal with few-body dynamics is urgent\-ly called for.
There are many natural extensions of this work. They include investigating
higher orders, pion production above threshold, the extension to virtual photons
and pion electroproduction, production of charged pions, and
considering heavier
nuclear targets such as $^4$He.
Further work in these directions is in progress.
\section*{Acknowledgements}
We thank Andreas Nogga for providing us with the chiral 3N
wave functions and Hermann Krebs for discussions.
Financial support by the Deutsche Forschungsgemeinschaft
(SFB/TR 16, ``Subnuclear
Structure of Matter''),
by the European Community Research Infrastructure Integrating Activity
``Study of Strongly Interacting Matter''
(acronym HadronPhysics2, Grant A\-gree\-ment n.~227431) under the 7th
Framework Programme of the EU and by
the European Research Council (acronym NuclearEFT, ERC-2010-StG 259218)
is gratefully
acknowled\-ged.
\medskip
|
\section{Introduction}
Fuzzy set theory was first introduce by Zadeh\cite{zadeh} in 1965 to describe
the situation in which data are imprecise or vague or uncertain.
Thereafter the concept of fuzzy set was generalized as intuitionistic
fuzzy set by K. Atanassov\cite{Atanasov,Atanasov1} in 1984. It has a wide
range of application in the field of population dynamics , chaos control ,
computer programming , medicine , etc.\\
The concept of fuzzy metric was first introduced by Kramosil and
Michalek\cite{Kramosil} but using the idea of intuitionistic fuzzy set,
Park\cite{Park} introduced the notion of intuitionistic fuzzy metric spaces
with the help of continuous t-norms and continuous t-conorms, which is a
generalization of fuzzy metric space due to George and Veeramani\cite{Veeramani}. \\
Introducing the contraction mapping with the help of the membership
function for fuzzy metric, several
authors\cite{Abdul,Grab,Veeramani2} established the Banach fixed
point theorem in fuzzy metric space. In the paper\cite{Abdul}, to
prove the Banach Fixed Point theorem in intuitionistic fuzzy metric
space, Mohamad\cite{Abdul} also introduced one concept of
contractive mapping, which is not so natural. There he proved that
every iterative sequence is a contractive sequence and then assumed
that every contractive sequences are Cauchy. But all these
contraction mappings, which they have considered to establish
different type fixed point theorem, do not bear the intension of the
contraction mapping with respect to a fuzzy metric, when a fuzzy
metric gives the degree of nearness of two points with respect to a
parameter $t$. Considering this meaning of fuzzy metric, in our
paper\cite{T.Samanta}, we have redefined the notion of contraction
mapping in a intuitionistic fuzzy metric space and then directly, it
has been proved that the every iterative sequence is a Cauchy
sequence, that is, we don't need to assume that every contractive
sequences are Cauchy sequences. Thereafter we have established the
Banach Fixed Point theorem there. In this paper, first we have
established two sets of sufficient conditions for a TS-IF
contractive mapping to have unique fixed point in a intuitionistic
fuzzy metric space. Then we have defined \,$(\,\varepsilon \,,\,
\lambda\,)$\, IF-uniformly locally contractive mapping and
\,$\eta\,-$\,chainable space, where it has been proved that the
\,$(\,\varepsilon \,,\, \lambda\,)$\, IF-uniformly locally
contractive mapping possesses a fixed point.
\section{Preliminaries}
We quote some definitions and statements of a few theorems which will be
needed in the sequel.
\begin{Definition} \cite{Schweizer}.
A binary operation \, $\ast \; : \; [\,0 \; , \; 1\,] \; \times \;
[\,0 \; , \; 1\,] \;\, \longrightarrow \;\, [\,0 \; , \; 1\,]$ \, is
continuous \, $t$ - norm if \,$\ast$\, satisfies the
following conditions \, $:$ \\
$(\,i\,)$ \hspace{0.2cm} $\ast$ \, is commutative and associative ,
\\ $(\,ii\,)$ \hspace{0.1cm} $\ast$ \, is continuous , \\
$(\,iii\,)$ \hspace{0.01cm} $a \;\ast\;1 \;\,=\;\, a \hspace{1.2cm}
\forall \;\; a \;\; \varepsilon \;\; [\,0 \;,\; 1\,]$ , \\
$(\,iv\,)$ \hspace{0.1cm} $a \;\ast\; b \;\, \leq \;\, c \;\ast\; d$
\, whenever \, $a \;\leq\; c$ , $b \;\leq\; d$ and $a \, , \, b
\, , \, c \, , \, d \;\, \varepsilon \;\;[\,0 \;,\; 1\,]$.
\end{Definition}
\begin{Definition}
\cite{Schweizer}. A binary operation \, $\diamond \; : \; [\,0 \; ,
\; 1\,] \; \times \; [\,0 \; , \; 1\,] \;\, \longrightarrow \;\,
[\,0 \; , \; 1\,]$ \, is continuous \, $t$-conorm if \,$\diamond$\,
satisfies the
following conditions \, $:$ \\
$(\,i\,)\;\;$ \hspace{0.01cm} $\diamond$ \, is commutative and
associative ,
\\ $(\,ii\,)\;$ \hspace{0.01cm} $\diamond$ \, is continuous , \\
$(\,iii\,)$ \hspace{0.01cm} $a \;\diamond\;0 \;\,=\;\, a
\hspace{1.2cm}
\forall \;\; a \;\; \in\;\; [\,0 \;,\; 1\,]$ , \\
$(\,iv\,)$ \hspace{0.1cm} $a \;\diamond\; b \;\, \leq \;\, c
\;\diamond\; d$ \, whenever \, $a \;\leq\; c$ , $b \;\leq\; d$
and $a \, , \, b \, , \, c \, , \, d \;\; \in\;\;[\,0
\;,\; 1\,]$.
\end{Definition}
\begin{Result}
\cite{klement}. $(\,a\,)\;$ For any \, $r_{\,1} \; , \;
r_{\,2} \;\; \in\;\; (\,0 \;,\; 1\,)$ \, with \, $r_{\,1} \;>\;
r_{\,2}$, there exist $\;r_{\,3} \; , \; r_{\,4} \;\; \in \;\; (\,0
\;,\; 1\,)$ \, such that \, $r_{\,1} \;\ast\; r_{\;3} \;>\; r_{\,2}$
\, and \, $r_{\,1} \;>\; r_{\,4} \;\diamond\; r_{\,2}.$
\\ $(\,b\,)$ \, For any \, $r_{\,5} \;\,
\in\;\, (\,0 \;,\; 1\,)$ , there exist \, $r_{\,6} \; , \; r_{\,7}
\;\, \in\;\, (\,0 \;,\; 1\,)$ \, such that \, $r_{\,6} \;\ast\;
r_{\,6} \;\geq\; r_{\,5}$ \,and\, $r_{\,7} \;\diamond\; r_{\,7}
\;\leq\; r_{\,5}.$
\end{Result}
\begin{Definition}
\cite{Park}
Let \,$\ast$\, be a continuous \,$t$-norm ,
\,$\diamond$\, be a continuous \,$t$-conorm and $X$ be any non-empty set.
An \textbf{intuitionistic fuzzy metric} or in short $\textit{\textbf{IFM}}$ on
\,$X$\, is an object of the form \\ $A \;\,=\;\, \{\; (\,(\,x \;,\,y \;,\;
t\,) \;,\; \mu\,(\,x \;,\,y\;,\; t\,) \;,\; \nu\,(\,x \;,\,y \;,\; t\,) \;) \;\,
: \;\, (\,x \;,\,y \;,\; t\,) \;\,\in\;\, X^{2}\times(0\,,\,\infty)\}$\,
where $\mu \,,\, \nu\;$ are fuzzy sets on \, $X^{2}\times(0\,,\,\infty)$ ,
\,$\mu$\, denotes the degree of nearness and
\,$\nu$\, denotes the degree of non$-$nearness of $x$ and $y$ relative
to $t$ satisfying the following conditions $:$ \, for all $x , y , z
\, \in \,X , \, s , t \, > \, 0$ \\
$(\,i\,)$ \hspace{0.10cm} $\mu\,(\,x \;,y\;, t\,) \;+\; \nu\,(\,x
\;,y\;, t\,) \;\,\leq\;\, 1 \hspace{0.5cm} \forall \;\; (\,x \;,\,y \;,\;
t\,)
\;\,\in\;\,X^{2}\times(0,\infty) ;$
\\$(\,ii\,)$ \hspace{0.10cm}$\mu\,(\,x \;,\,y \;,\; t\,) \;\,>\;\, 0 \, ;$ \\
$(\,iii\,)$ $\mu\,(\,x \;,\,y\;,\; t\,) \;\,=\;\, 1$ \, if
and only if \, $x \;=\;y \,$\,
\\$(\,iv\,)$ $\mu\,(\,x \;,\,y\;,\; t\,) \;\,=\;\,\mu\,(\,y \;,\,x\;,\; t\,);$
\\ $(\,v\,)$
\hspace{0.01cm} $\mu\,(\,x \;,\,y \;,\; s\,) \;\ast\; \mu\,(\,y \;,\,z \;,\; t\,)
\;\,\leq\;\, \mu\,(\,x \;,\,z \;,\; s\;+\;t\,
\,) \, ;$
\\$(\,vi\,)$ $\mu\,(\,x \,,\,y\,,\,\cdot\,) :(0 \,,\;\infty\,)\,\rightarrow
\,(0 \,,\;1]$ \, is continuous;
\\$(\,vii\,)$ \hspace{0.10cm}$\nu\,(\,x \;,\,y \;,\; t\,) \;\,>\;\, 0 \, ;$
\\$(\,viii\,)$ $\nu\,(\,x \;,\,y\;,\; t\,) \;\,=\;\, 0$ \, if
and only if \, $x \;=\;y \,;$\,
\\$(\,ix\,)$ $\nu\,(\,x \;,\,y\;,\; t\,) \;\,=\;\,\nu\,(\,y \;,\,x\;,\; t\,);$
\\ $(\,x\,)$\hspace{0.10cm} $\nu\,(\,x \;,\,y \;,\; s\,) \;\diamond\;
\mu\,(\,y \;,\,z \;,\; t\,)
\;\,\geq\;\, \nu\,(\,x \;,\,z \;,\; s\;+\;t\,
\,) \, ;$
\\$(\,xi\,)$ $\nu\,(\,x \,,\,y\,,\,\cdot\,) :(0\,,\;\infty\,)\,\rightarrow
\,(0\,,\;1]$ \, is continuous.
\\\\ If \,$A$\, is a $\textit{\textbf{IFM}}$ on
\,$X$ , the pair \,$(\,X \,,\, A\,)$\, will be called a \textbf{intuitionistic
fuzzy metric space} or in short $\textit{\textbf{IFMS}}$.\\\\ We further assume
that \,$(\,X \,,\, A\,)$\, is a $\textit{\textbf{IFMS}}$ with the property \\
$(\,xii\,)\;$ For all \,$a\,\in\,(0 \,,\, 1)$,\, $a \,\ast\, a \,=\, a$\,
and \,$a \,\diamond\, a \,=\, a$
\end{Definition}
\begin{Remark}
\cite{Park}
In intuitionistic fuzzy metric space $X$, \, $\mu\,(\,x \,,\,y \,,\,\cdot\,)$ \,
is non-decreasing and \,$\nu\,(\,x \,,\,y \,,\,\cdot\,)$\, is non-increasing
for all \,$x\,,\,y\,\in\,X$.
\end{Remark}
\begin{Definition}
\cite{Veeramani}
A sequence $\{\,x_{\,n}\,\}_{\,n}$ in a intuitionistic fuzzy metric space
is said to be a \textit{\textbf{Cauchy sequence}} if and only if for each
\,$r\,\in\,(\,0\,,\,1\,)$\, and \,$t>0$\, there exists \,$n _{\,0}\,\in\,N$\,
such that \,$\mu\,(\,x_{n} \,,\,x_{m} \,,\, t\,) \;>\;1\,-\,r $\, and
\,$\nu\,(\,x_{n} \,,\,x_{m} \,,\, t\,) \;<\;r$\, for all
$\;n\,,\,m\;\geq\;n_{\,0}.$\, \\A sequence $\{\,x_{\,n}\,\}$ in
a intuitionistic fuzzy metric space is said to converge to \,$x\,\in\,X$\,
if and only if for each \,$r\,\in\,(\,0\,,\,1\,)$\, and \,$t>0$\,
there exists \,$n _{\,0}\,\in\,N$\, such that
\,$\mu\,(\,x_{n} \,,\,x \,,\, t\,) \;>\;1\,-\,r $\, and
\,$\nu\,(\,x_{n} \,,\,x \,,\, t\,) \;<\;r$\, for all $\;n\,,\,m\;\geq\;n_{\,0}.$\,
\end{Definition}
\begin{Note}
\cite{Samanta}
A sequence $\{\,x_{\,n}\,\}_{\,n}$ in an intuitionistic fuzzy metric space
is a Cauchy sequence if and only if \[\mathop {\lim }\limits_{n\,\, \to
\,\,\infty } \,\mu\,(\,x_{n} \,,\,x_{n\,+\,p} \,,\, t\,)\;=\;1\;\; and
\;\; \mathop {\lim }\limits_{n\,\, \to \,\,\infty } \,\nu\,(\,x_{n} \,,
\,x_{n\,+\,p} \,,\, t\,)\;=\;0\;\]
A sequence $\{\,x_{\,n}\,\}_{\,n}$ in an intuitionistic fuzzy metric space converges to
\,$x\,\in\,X$\, if and only if \[\mathop {\lim }\limits_{n\,\, \to \,\,\infty }
\,\mu\,(\,x_{n} \,,\,x \,,\, t\,)\;=\;1\;\; and \;\; \mathop {\lim }\limits_{n\,\,
\to \,\,\infty } \,\nu\,(\,x_{n} \,,\,x \,,\, t\,)\;=\;0\;\]
\end{Note}
\begin{Definition}
\cite{Abdul}
Let $(\,X\,,\,A\,)$ be a intuitionistic fuzzy metric space. We will say
the mapping \,$f : X \rightarrow X$\, is \textbf{t-uniformly continuous}
if for each \,$\varepsilon,$\, with \,$0\, < \,\varepsilon \,< \,1,$\,
there exists \,$0 \,<\, r \,<\, 1,$\, such that
\,$\mu\,(\,x \,,\, y \,,\, t\,) \,\geq\,1\,-\,r\, $\, and
\,$\nu\,(\,x \,,\, y \,,\, t\,) \,\leq\,r\, $ implies
$\mu\,\left(\,f(x) \,,\, f(y) \,,\, t\,\right) \,\geq\,1\,-\,\varepsilon \, $
and \,$\nu\,\left(\,f(x) \,,\, f(y) \,,\, t\,\right) \,\leq\,\varepsilon \, $
for each \,$x , y\,\in\,X$\, and \,$t \,> \,0.$
\end{Definition}
\begin{Definition}
\cite{T.Samanta}
Let \,$(\,X \,,\, A\,)$\, be \textit{\textbf{IFMS}} and $\;T:X\rightarrow X$.
T is said to be TS-IF contractive mapping if there exists \,$k\;\in\;(0 \,,\, 1)$
such that \[k\;\mu\,\left(\,T(x) \;,\; T(y) \;,\; t\,\right)\;\geq\;
\mu\,(\,x \;,\; y \;,\; t\,)\;\] \[ and \hspace{0.8cm}\frac{1}{k}\;
\nu\,\left(\,T(x) \;,\; T(y) \;,\; t\,\right)\;\leq\;\nu\,(\,x \;,\;
y \;,\; t\,) \;\;\forall \;\;t \,>\, 0.\]
\end{Definition}
\begin{Proposition}\label{P1}
Let \,$(\,X \,,\, A\,)$\, be a intuitionistic fuzzy metric space.
If \,$f:X\rightarrow\;X$ is TS-IF contractive then f is t-uniformly continuous.
\end{Proposition}
{\bf Proof.} Obvious
\begin{Theorem}
\cite{T.Samanta}
Let \,$(\,X \,,\, A\,)$\, be a complete \textit{\textbf{IFMS}} and
$\,T : X \rightarrow X$ be TS-IF contractive mapping with k its
contraction constant. Then T has a unique fixed point.
\end{Theorem}
\section{ Fixed-point theorems}
\begin{Definition}
Let \,$(\,X \,,\, A\,)$\, be an \textit{\textbf{IFMS}} , $ x\,\in\,X,$\,
$r\,\in\,(0 \,,\, 1)$ , \,$t \,>\, 0$,
\\${\hspace{2.0cm}}B(\,x \,,\, r \,,\, t\,)\,=\,\left\{\,y\,\in \,X\,/\,
\mu(\,x \,,\, y \,,\, t\,) \,>\, 1\,-\,r \,,\, \nu(\,x \,,\, y \,,\, t\,)
\,<\, r\,\right\}$.
\\ Then \,$B(\,x \,,\, r \,,\, t\,)$\, is called an \textbf{open ball}
centered at x of radius r w.r.t. t.
\end{Definition}
\begin{Definition}
Let \,$(\,X \,,\, A\,)$\, be an \textit{\textbf{IFMS}} and \,$P\,\subseteq \,X$\,.
P is said to be a \textbf{closed set} in \,$(\,X \,,\, A\,)$\, if and only if
any sequence \,$\{\,x_{n}\}$\, in P converges to \,$x\,\in P$\, i.e, iff.
$\mathop {\lim }\limits_{n\; \to \;\infty } \,\mu\,(\,x_{n} \,,\, x \,,\, t\,)\,=\,1$\,
and \,$\mathop {\lim }\limits_{n\; \to \;\infty } \,\nu\,(\,x_{n}
\,,\,x \,,\, t\,)\,=\,0\;$ $\Rightarrow$ $x\in\,P$.
\end{Definition}
\begin{Definition}\label{D1}
Let \,$(\,X \,,\, A\,)$\, be an \textit{\textbf{IFMS,}} $ x\,\in\,X,$\,
$r\,\in\,(0 \,,\, 1)$ , \,$t \,>\, 0$,
\\${\hspace{2.0cm}}S(\,x \,,\, r \,,\, t\,)\,=\,\left\{\,y\,\in \,X\,/\,
\mu(\,x \,,\, y \,,\, t\,) \,>\, 1\,-\,r \,,\, \nu(\,x \,,\, y \,,\, t\,)
\,<\, r\,\right\}$.
\\Hence \,$S(\,x \,,\, r \,,\, t\,)$\, is said to be a \textbf{closed ball}
centered at x of radius r w.r.t. t iff. any sequence $\{\,x_{n}\}$ in
\,$S(\,x \,,\, r \,,\, t\,)$\, converges to y then $y\in\,\,S(\,x \,,\, r \,,\, t\,)$.
\end{Definition}
\begin{Theorem}
$($Contraction on a closed ball$)$ :- Suppose \,$(\,X \,,\, A\,)$\,
is a complete IFMS. Let $T:X\,\rightarrow\,X$ be TS-IF contractive
mapping on \,$S(\,x \,,\, r \,,\, t\,)$ \,with contraction constant k.
Moreover, assume that
\[k\,{\mu\,\left(\,x \,,\,T(x) \,,\, t\,\right)}
\,>\, (1 \,-\, r)\;\; and \;\; \frac{1}{k}
\;{\nu\,\left(\,x \,,\,T(x) \,,\, t\,\right)}\,<\,r\]
Then T has unique fixed point in \,$S(\,x \,,\, r \,,\, t\,)$.
\end{Theorem}
{\bf Proof.} Let $x_{1}\,=\,T(x) \,,\, x_{2}\,=\,T(x_{1})\,=\,T^{\,2}(x)
\,,\; \cdots \;,\, x_{n}\,=\,T(x_{n-1})$ \\i.e, \,$\,x_{n}\,=\,T^{\,n}(x)$\,
for all $n\,\in\,N.\;$ Now \[k\,{\mu\,\left(\,x \,,\,T(x) \,,\, t\,\right)}
\,>\, (1 \,-\, r){\hspace{2.6cm}}\]
\[\Rightarrow\;\;{\mu\,\left(\,x \,,\,T(x) \,,\, t\,\right)} \,>\,
\frac{(1 \,-\, r)}{k} \,>\, (1 \,-\, r)\hspace{1.5cm} \]
\[\Rightarrow\;\;{\mu\,\left(\,x \,,\,x_1 \,,\, t\,\right)} \,>\,
(1 \,-\, r) \hspace{1.5cm} \cdots \hspace{1.5cm}(\,i\,)\]
Again,
\[\frac{1}{k}\; {\nu\,\left(\,x \,,\,T(x) \,,\, t\,\right)}\,<\,
r {\hspace{2.5cm}}\]
\[\Rightarrow\;\;{\nu\,\left(\,x \,,\,T(x) \,,\, t\,\right)} \,<\,
r\,k\,<\,r\hspace{3.2cm} \]
\[\Rightarrow\;\;{\nu\,(\,x \,,\, x_1 \,,\, t\,)}\,<\,r \hspace{1.0cm}
\cdots \hspace{2.5cm} (\,ii\,) {\hspace{0.1cm}}\]
$(\,i\,)$ and $(\,ii\,)$ $\;\Rightarrow\; x_{1}\,\in\,S(\,x \,,\, r \,,\, t\,).$
\\\\Assume that \,$x_{1} \,,\, x_{2} \,,\; \cdots \;,\, x_{n \,-\, 1}
\,\in\,S(\,x \,,\, r \,,\, t\,)$.
We show that $x_{n}\,\in\,S(\,x \,,\, r \,,\, t\,).$
\[k\;\mu\,(\,x_{1} \;,\; x_{2}
\;,\; t\,)\;=\;k\;\mu\,\left(\,T(x) \;,\; T(x_{1}) \;,\; t\,\right) {\hspace{3.5cm}}\]
\[\;\geq\;\mu\,(\,x \;,\; x_{1}
\;,\; t\,) {\hspace{2.2cm}}\]
\[\Rightarrow\;\mu\,(\,x_{1} \;,\; x_{2}
\;,\; t\,)\,>\,\frac{(1-r)}{k}\,>\,(1-r)\hspace{4.8cm}\]
\[k\;\mu\,(\,x_{2} \;,\; x_{3}
\;,\; t\,)\;=\;k\;\mu\,\left(\,T(x_{1}) \;,\; T(x_2) \;,\; t\,\right) {\hspace{3.5cm}}\]
\[\;\geq\;\mu\,(\,x_{1} \;,\; x_{2}
\;,\; t\,) {\hspace{2.2cm}}\]
\[\Rightarrow\;\mu\,(\,x_{2} \;,\; x_{3}
\;,\; t\,)\,\geq\,\frac{1}{k}\,\mu\,(\,x_{1} \;,\; x_{2}
\;,\; t\,)\hspace{5.2cm}\]
\[>\;\,\frac{1-r}{k}\,>\,(1-r)\hspace{1.7cm}\]
Again,
\[\frac{1}{k}\;\nu\,(\,x_{1} \;,\; x_{2}
\;,\; t\,)\;=\;\frac{1}{k}\;\nu\,\left(\,T(x) \;,\; T(x_{1}) \;,\; t\,\right)
{\hspace{3.5cm}}\]
\[\;\leq\;\nu\,(\,x \;,\; x_{1}
\;,\; t\,) {\hspace{2.2cm}}\]
\[\Rightarrow\;\nu\,(\,x_{1} \;,\; x_{2}
\;,\; t\,)\,\leq\,k\,\nu\,(\,x \;,\; x_{1}
\;,\; t\,)\,<\,k\,r\,<\,r{\hspace{3.2cm}}\]
\[\frac{1}{k}\;\nu\,(\,x_{2} \;,\; x_{3}
\;,\; t\,)\;=\;\frac{1}{k}\;\nu\,\left(\,T(x_{1}) \;,\; T(x_{2}) \;,\;
t\,\right) {\hspace{3.5cm}}\]
\[\;\leq\;\nu\,(\,x_{1} \;,\; x_{2}
\;,\; t\,) {\hspace{2.2cm}}\]
\[\Rightarrow\;\nu\,(\,x_{2} \;,\; x_{3}
\;,\; t\,)\,\leq\,k\,\nu\,(\,x_{1} \;,\; x_{2}
\;,\; t\,)\,<\,k\,r\,<\,r{\hspace{3.2cm}}\]
Similarly it can be shown that ,\\
$\mu\,(\,x_{3} \,,\, x_{4} \,,\, t\,) \;>\; 1 \,-\, r \;,\;
\nu\,(\,x_{3} \,,\, x_{4} \,,\, t\,) \;<\; r$ \,,\,
$\; \cdots\; ,\, \mu\,(\,x_{n \,-\, 1} \,,\, x_{n} \,,\, t\,) \;>\; 1 \,-\, r$ and
$\;\nu\,(\,x_{n \,-\, 1} \,,\, x_{n} \,,\, t\,) \;<\; r.$ \\Thus, we see that ,
\[\mu\,(\,x \,,\, x_{n} \,,\, t\,)\;\geq\;\mu\,\left(\,x \,,\, x_{1}
\,,\, \frac{t}{n}\,\right)\;\ast\;\mu\,\left(\,x_{1} \,,\, x_{2} \,,\,
\frac{t}{n}\,\right)\;\ast\;\cdots\;\ast\;\mu\,\left(x_{n \,-\, 1}
\,,\, x_{n} \,,\, \frac{t}{n}\,\right)\]
\[>\; (1 \,-\, r)\;\ast\;(1 \,-\, r)\;\ast\;\cdots\;\ast\;(1 \,-\, r)
\;=\; 1 \,-\, r\;\;\]
\[i.e. \,, \hspace{0.5cm}\mu\,(\,x \,,\, x_{n} \,,\, t\,) \;>\; 1
\,-\, r {\hspace{9.5cm}}\]
\[\nu\,(\,x \,,\, x_{n} \,,\, t\,)\;\leq\;\nu\,\left(\,x \,,\, x_{1}
\,,\, \frac{t}{n}\,\right)\;\diamond\;\nu\,\left(\,x_{1} \,,\, x_{2} \,,\,
\frac{t}{n}\,\right)\;\diamond\;\cdots\;\diamond\;\nu\,\left(x_{n \,-\, 1}
\,,\, x_{n} \,,\, \frac{t}{n}\,\right)\]
\[\;<\;r\;\diamond\;r\;\diamond\;\cdots\;\diamond\;r\;=\;r {\hspace{4.6cm}}\]
Thus , $\;\mu\,(\,x \,,\, x_{n} \,,\, t\,) \;>\; 1 \,-\, r \;and\;
\nu\,(\,x \,,\, x_{n} \,,\, t\,) \;<\; r$
\\$\Rightarrow\;x_{n}\;\in\;S(\,x \,,\, r \,,\, t\,)$
\\Hence, by the theorem 3.10\cite{T.Samanta} and the definition\ref{D1},
T has unique fixed point in \,$S(\,x \,,\, r \,,\, t\,)$.
\begin{Note}
It follows from the proof of Theorem 3.10\cite{T.Samanta} that for
any $\;x\in\;X$ the sequence of iterates \,$\{\,T^{\,n}(x)\}$\,
converges to the fixed point of T.
\end{Note}
\begin{Lemma}\label{L2}
Let \,$(\,X \,,\, A\,)$\, be \textit{\textbf{IFMS}} and $T:\,X\rightarrow\,X$
be t-uniformly continuous on X. If $\;x_{n}\;\rightarrow\;x\; $ as
$\; n\;\rightarrow\infty\;$\, in \,$(\,X \,,\, A\,)$\, then
$\;T(x_{n})\;\rightarrow\;T(x)\;$ as $\; n\;\rightarrow\infty\;$
in \,$(\,X \,,\, A\,)$ .
\end{Lemma}
{\bf Proof.} \, Proof directly follows from the definitions of
t-uniformly continuity and convergence of a sequence in a \textit{\textbf{IFMS}}.
\begin{Lemma}\label{L1}
Let \,$(\,X \,,\, A\,)$\, be \textit{\textbf{IFMS}}.
If $\;x_{n}\;\rightarrow\;x$ and $\;y_{n}\;\rightarrow\;y$ in
\,$(\,X \,,\, A\,)$\, then $\mu\,(\,x_{n} \,,\, y_{n} \,,\, t\,)
\;\rightarrow\;\mu\,(\,x \,,\, y \,,\, t\,) \;and\; \nu\,(\,x_{n}
\,,\, y_{n} \,,\, t\,)\;\rightarrow\;\nu\,(\,x \,,\, y \,,\, t\,)
\;as \;n\;\rightarrow\infty \;\; for\; all\;\; t \,>\, 0\; \;in \;\;R\;.$
\end{Lemma}
{\bf Proof.} \,\,We have,
\[ {\hspace{1.0cm}}\mathop {\lim }\limits_{n\;\, \to \;\,\infty }
\;\mu\,(\,x_{n} \;,\,x \;,\; t\,)\;=\;1\; , \;\mathop {\lim }\limits_{n\;\,
\to \;\,\infty } \;\nu\,(\,x_{n} \;,\,x \;,\; t\,)\;=\;0\;\]
\[ and \;\;\mathop {\lim }\limits_{n\;\, \to \;\,\infty } \;
\mu\,(\,y_{n} \;,\,y \;,\; t\,)\;=\;1\; , \;\mathop {\lim }\limits_{n\;\,
\to \;\,\infty } \;\nu\,(\,y_{n} \;,\,y \;,\; t\,)\;=\;0\;\]
\[ \mu\,(\,x_{n} \,,\, y_{n} \,,\, t\,)\;\geq\;\mu\,\left(\,x_{n} \,,
\, x \,,\, \frac{t}{2}\,\right)\;\ast\;\mu\,\left(\,x \,,\,y_{n} \,,\,
\frac{t}{2}\,\right)\]
\[{\hspace{4.8cm}}\;\geq\;\mu\,\left(\,x_{n} \,,\, x \,,\, \frac{t}{2}\,
\right)\;\ast\;\mu\,\left(\,x \,,\, y \,,\, \frac{t}{4}\,\right)\;\ast\;\mu\,
\left(\,y \,,\, y_{n} \,,\, \frac{t}{4}\,\right)\]
\\${\hspace{0.5cm}}\Rightarrow\;\mathop {\lim }\limits_{n\;\, \to \;\,\infty }
\;\mu\,(\,x_{n} \;,\,y_{n} \;,\; t\,)\;\geq\;\mu\,(\,x\;,\;y\;,\;t\,)\;$
\[ \mu\,(\,x \,,\, y \,,\, t\,)\;\geq\;\mu\,\left(\,x \,,\, x_{n} \,,\,
\frac{t}{2}\,\right)\;\ast\;\mu\,\left(\,x_{n} \,,\, y \,,\, \frac{t}{2}\,\right)
{\hspace{2.5cm}}\]
\[ {\hspace{2.7cm}}\;\geq\;\mu\,\left(\,x \,,\, x_{n} \,,\, \frac{t}{2}\,\right)\;
\ast\;\mu\,\left(\,x_{n} \,,\, y_{n} \,,\, \frac{t}{4}\,\right)\;\ast\;\mu\,
\left(\,y_{n} \,,\, y \,,\, \frac{t}{4}\,\right)\]
${\hspace{0.5cm}}\Rightarrow\;\mu\,(\,x \;,\,y \;,\; t\,)\;\geq\;\;
\mathop {\lim }\limits_{n\;\, \to \;\,\infty } \;\mu\,(\,x_{n}\;,\;y_{n}\;,\;t\,)\;$
$\;\forall\;t\,\,>\,0.$
\\Then,\\ ${\hspace{0.9cm}}\;\mathop {\lim }\limits_{n\;\, \to \;\,\infty }
\;\mu\,(\,x_{n} \;,\;y_{n} \;,\; t\,)\;=\;\mu\,(\,x\;,\;y\;,\;t\,)\;$
for all $\;t\;>\,0$,
\\\\Similarly, $\;\mathop {\lim }\limits_{n\;\, \to \;\,\infty } \;
\nu\,(\,x_{n} \;,\;y_{n} \;,\; t\,)\;=\;\nu\,(\,x\;,\;y\;,\;t\,)\;$
for all $\;t\;>\,0$.
\begin{Theorem}
Let \,$(\,X \,,\, A\,)$\, be a complete \textit{\textbf{IFMS}} and
\,$T:X\rightarrow X$\, be a t-uniformly continuous on X. If for same
positive integer m, \,$T^{\,m}\,$ is a TS-IF contractive mapping with
k its contractive constant then $T$ has a unique fixed point in \,$X$.
\end{Theorem}
{\bf Proof.} \,Let $B\;=\;T^{\,m}$,\, $n$ \,be an arbitrary but fixed
positive integer and \,$x\;\in\;X$. \\we now show that $\;B^{\,n}\,T(x)
\;\rightarrow\; B^{\,n}\,(x)$ in \,$(\,X \,,\, A\,)$\,.
\\Now,
\[k\;\mu\,\left(\,B^{\,n}T(x) \;,\; B^{\,n}(x) \;,\; t\,\right)\;=\;
k\;\mu\,\left(\,B(B^{\,n \,-\, 1}\,T(x)) \;,\; B(B^{\,n \,-\, 1}(x))
\;,\; t\,\right) {\hspace{3.5cm}}\]
\[\;{\hspace{1.5cm}}\geq\;{\mu\,(\,B^{n \,-\, 1}\,T(x) \;,\; B^{n \,-
\, 1}(x) \;,\; t\,)}\]
\[i.e,\;\mu\,\left(\,B^{\,n}T(x) \;,\; B^{\,n}(x) \;,\; t\,\right)\;\geq\;
\,\frac{1}{k}\,\;{\mu\,(\,B^{n \,-\, 1}\,T(x) \;,\; B^{n \,-\, 1}(x) \;,\; t\,)}
{\hspace{2.8cm}} \]
\[{\hspace{4.5cm}}\;=\;\frac{1}{k}\,\;\mu\,\left(\,B(B^{\,n \,-\, 2}\,T(x))
\;,\; B(B^{\,n \,-\, 2}(x)) \;,\; t\,\right) \]
\[{\hspace{5.2cm}}\;\geq\;\frac{1}{k^{2}}\,\;{\mu\,(\,B^{n \,-\, 2}\,T(x)
\;,\; B^{n \,-\, 2}(x) \;,\; t\,)} {\hspace{2.2cm}}\]
\[\;\geq\;{\hspace{1.0cm}}\cdots{\hspace{1.1cm}}\]
\[{\hspace{0.6cm}}\;\geq\;\frac{1}{k^{n}}\,\;\mu\,(\,T(x) \;,\; x \;,\; t\,)\]
\[ \Rightarrow\;\;\mathop {\lim }\limits_{n\; \to \;\infty}
\mu\,\left(\,B^{\,n}T(x) \;,\; B^{\,n}(x) \;,\; t\,\right)\;
\geq\;\mathop {\lim }\limits_{n\; \to \;\infty }\;\frac{1}{k^{n}}\;
\mu\,(\,T(x) \;,\; x \;,\; t\,)\]
\[\Rightarrow \;\;\mathop {\lim }\limits_{n\; \to \;\infty} \mu\,
\left(\,B^{\,n}T(x) \;,\; B^{\,n}(x) \;,\; t\,\right)\;=\;1 {\hspace{4.4cm}}\]
\\Similarly,
${\hspace{0.5cm}}\mathop {\lim }\limits_{n\;\, \to \;\,\infty }
\nu\,\left(\,B^{\,n}\,T(x) \;,\; B^{\,n}(x) \;,\; t\,\right)\;=\;0,\;$
for all $\; t\;>\;0$. \\
Thus, $\;B^{\,n}\,T(x) \;\rightarrow\; B^{\,n}\,(x)$ in \,$(\,X \,,\, A\,)$\,.
\\Again, by the theorem 3.10\cite{T.Samanta}, we see that B has a unique
fixed point \,$y$(say), and from the note [3.5], it follows that
$\;B^{\,n}(x)\,\rightarrow \,y$ as $\;n\;\rightarrow\;\infty\;$
in \,$(\,X \,,\, A\,)$\,.
\\\\Since T is t-uniformly continuous on X, it follows from the above
lemma[3.6] that $\,B^{\,n}\,T(x)\;=\;T\;B^{\,n}(x)\;\rightarrow\;T(y)$
as $n\;\rightarrow\;\infty$ in \,$(\,X \,,\, A\,)$\,.
\\\\ Again since $\mathop {\lim }\limits_{n\;\, \to \;\,\infty }\mu\,
\left(\,B^{\,n}\,T(x) \;,\; B^{\,n}(x) \;,\; t\,\right)\;=\;1\;$ and
$\;\mathop {\lim }\limits_{n\;\, \to \;\,\infty }\nu\,\left(\,B^{\,n}\,T(x)
\;,\; B^{\,n}(x) \;,\; t\,\right)\\=\;0$, \,we have by the lemma [\ref{L1}]
\\$\mathop {\lim }\limits_{n\;\, \to \;\,\infty }\mu\,\left(\,T(y) \;,
\; y \;,\; t\,\right)\;=\;1$ and $\mathop {\lim }\limits_{n\;\, \to
\;\,\infty }\nu\,\left(\,T(y) \;,\; y \;,\; t\,\right)\;=\;0,\;$ for
all $\;t\;>\;0$,
\\$i.e. \,, \;\;\mu\,\left(\,T(y) \;,\; y \;,\; t \,\right)\;=\;1\;$
and $\;\nu\,\left(\,T(y) \;,\; y \;,\; t\,\right)\;=\;0,\;$ for all $\;t\;>\;0$.
\\$\Rightarrow\;T(y)\;=\;y\;$
$\;\Rightarrow\;y$ is a fixed point of T. \\If $y^{'}$ is a fixed point of T,
$\,i.e.\,, \;T(y^{'})\;=\;y^{'}$, then $\;T^{\,m}(y^{'})\;=\;T^{\,m \,-
\, 1}(T(y^{'}))\;=\;T^{\,m \,-\, 1}(y^{'})\;=\; \cdots \;=\;y^{'}$
$\Rightarrow\;B(y^{'})\;=\;y^{'}\;\Rightarrow\;y^{'}\,$ is a fixed point of B.
\\But $y$ is the unique fixed point of B, therefore $y\;=\;y^{'}$ which
implies that $y$ is the unique fixed point of T. This completes the proof.
\begin{Definition}
Let \,$(\,X \,,\, A\,)$\, be a \textit{\textbf{IFMS}} and \,$T:X\rightarrow X$.
For \,$\varepsilon\;>\;0$\, and \,$0 \;<\; \lambda \;<\; 1$\, , $T$\,
is said to be \,$(\,\varepsilon \,,\, \lambda\,)$\,
\textit{\textbf{IF-uniformly locally contractive}}
if \[\mu\,(\,x\;,\;y\;,\;t\,) \;>\; \varepsilon \;\Longrightarrow\;
\lambda\;\mu\,(\,Tx\;,\;Ty\;,\;t\,) \;>\; \mu\,(\,x\;,\;y\;,\;t\,)\]
\[\nu\,(\,x\;,\;y\;,\;t\,) \;<\; 1 \;-\; \varepsilon \;\Longrightarrow\;
\frac{1}{\lambda}\;\nu\,(\,Tx\;,\;Ty\;,\;t\,) \;<\; \nu\,(\,x\;,\;y\;,\;t\,)
\hspace{0.9cm}\]
\end{Definition}
\begin{Definition}
Let \,$0 \;<\; \eta \;<\; 1$\, and \,$(\,X \,,\, A\,)$\, be a \textit{\textbf{IFMS}}.
Then \,$(\,X \,,\, A\,)$\, is said to be \textit{\textbf{IF}}\,
\,$\eta\,-$\,\textit{\textbf{chainable}} space if for every
\,$a \,,\, b \; \in \;X$\, there exist a finite set of points
\,$a \,=\, x_{\,0} \;,\; x_{\,1} \;,\; \cdots \;,\; x_{\,n} \,=\, b$\, such that
\[\mu\,(\,x_{\,i \,-\, 1}\;,\;x_{\,i}\;,\;t\,) \;>\; \eta \hspace{0.3cm}
and \hspace{0.3cm} \nu\,(\,x_{\,i \,-\, 1}\;,\;x_{\,i}\;,\;t\,) \;<\; 1
\;-\; \eta \; \; ,\hspace{0.3cm} i \,=\, 1 \,,\, 2 \,,\; \cdots \;,\, n\]
\end{Definition}
\begin{Theorem}
Let \,$(\,X \,,\, A\,)$\, be a complete \textit{\textbf{IFMS}} and
IF\,$\varepsilon\,-$\,chainable space. If \,$T:X\rightarrow X$\, is
\,$(\,\varepsilon \,,\, \lambda\,)$\, IF-uniformly locally contractive
then $T$ has a fixed point in \,$X$.
\end{Theorem}
{\bf Proof.} \,Let \,$x$\, be an arbitrary but fixed point of \,$X$.
If \,$T x \,=\, x$\, then a fixed point is assured. We assume therefore
that \,$T x \;\neq\; x$. Since \,$X$\, is IF\,$\varepsilon\,-$\,chainable
space, there exists a finite set of points \,$x \,=\, x_{\,0} \;,\; x_{\,1}
\;,\; \cdots \;,\; x_{\,n} \,=\, T x$\, such that
\[\mu\,(\,x_{\,i \,-\, 1}\;,\;x_{\,i}\;,\;t\,) \;>\; \varepsilon \hspace{0.3cm}
and \hspace{0.3cm} \nu\,(\,x_{\,i \,-\, 1}\;,\;x_{\,i}\;,\;t\,) \;<\; 1 \;-\;
\varepsilon \; \; ,\hspace{0.3cm} i \,=\, 1 \,,\, 2 \,,\; \cdots \;,\, n\]
Again, since \,$T$\, is \,$(\,\varepsilon \,,\, \lambda\,)$\, IF-uniformly
locally contractive, we have
\[ \mu\,(\,x_{\,i \,-\, 1}\;,\;x_{\,i}\;,\;t\,) \;>\; \varepsilon \;
\Longrightarrow \; \lambda\;\mu\,(\,T x_{\,i \,-\, 1}\;,\;T x_{\,i}\;,\;t\,)
\;>\; \mu\,(\,x_{\,i \,-\, 1}\;,\;x_{\,i}\;,\;t\,) \;>\; \varepsilon\]
\[i.e. , \hspace{0.5cm} \mu\,(\,T x_{\,i \,-\, 1}\;,\;T x_{\,i}\;,\;t\,)
\;>\; \frac{\varepsilon}{\lambda} \;>\; \varepsilon \;;\hspace{6.5cm}\]
\[ \nu\,(\,x_{\,i \,-\, 1}\;,\;x_{\,i}\;,\;t\,) \;<\; 1 \;-\; \varepsilon \;
\Longrightarrow \; \frac{1}{\lambda}\;\nu\,(\,T x_{\,i \,-\, 1}\;,
\;T x_{\,i}\;,\;t\,) \;<\; \nu\,(\,x_{\,i \,-\, 1}\;,\;x_{\,i}\;,\;t\,)
\;<\; 1 \;-\; \varepsilon\]
\[i.e. , \hspace{0.5cm} \nu\,(\,T x_{\,i \,-\, 1}\;,\;T x_{\,i}\;,\;t\,)
\;<\; \lambda\,(\,1 \;-\; \varepsilon\,) \;<\; 1 \;-\; \varepsilon \hspace{6.5cm}\]
and therefore, \[ \lambda^{\,2}\;\mu\,(\,T^{\,2} x_{\,i \,-\, 1}\;,
\;T^{\,2} x_{\,i}\;,\;t\,) \;=\; \lambda\;\left(\, \lambda\;
\mu\,(\,T\,(\,T x_{\,i \,-\, 1}\,)\;,\; T\,(\,T x_{\,i}\,)\;,\;t\,) \,\right)\]
\[\hspace{3.7cm} >\; \lambda\;\mu\,(\,T x_{\,i \,-\, 1}\;,\;T x_{\,i}\;,\;t\,)
\;>\; \lambda\;\varepsilon\]
\[\Longrightarrow\;\; \mu\,(\,T^{\,2} x_{\,i \,-\, 1}\;,\;T^{\,2} x_{\,i}\;,\;t\,)
\;>\; \varepsilon \;;\hspace{6.5cm}\]
\[ \frac{1}{\lambda^{\,2}}\;\nu\,(\,T^{\,2} x_{\,i \,-\, 1}\;,
\;T^{\,2} x_{\,i}\;,\;t\,) \;=\; \frac{1}{\lambda}\;\left(\, \frac{1}{\lambda}\;
\nu\,(\,T\,(\,T x_{\,i \,-\, 1}\,)\;,\; T\,(\,T x_{\,i}\,)\;,\;t\,) \,\right)\]
\[\hspace{5.0cm} <\; \frac{1}{\lambda}\;\nu\,(\,T x_{\,i \,-\, 1}\;,\;T x_{\,i}\;,
\;t\,) \;<\; \frac{1}{\lambda}\;(\,1 \;-\; \varepsilon\,)\]
\[{\hspace{0.2cm}}\Longrightarrow\;\; \nu\,(\,T^{\,2} x_{\,i \,-\, 1}\;,
\;T^{\,2} x_{\,i}\;,\;t\,) \;<\; 1 \;-\; \varepsilon \hspace{6.5cm}\]
In the similar way we have,
\[ \lambda^{\,3}\;\mu\,(\,T^{\,3} x_{\,i \,-\, 1}\;,\;T^{\,3} x_{\,i}\;,
\;t\,) \;=\; \lambda^{\,2}\;\left(\, \lambda\;\mu\,(\,T\,(\,T^{\,2}
x_{\,i \,-\, 1}\,)\;,\; T\,(\,T^{\,2} x_{\,i}\,)\;,\;t\,) \,\right)\]
\[\hspace{3.7cm} >\; \lambda^{\,2}\;\mu\,(\,T^{\,2} x_{\,i \,-\, 1}\;,
\;T^{\,2} x_{\,i}\;,\;t\,) \;>\; \lambda^{\,2}\;\varepsilon\]
\[\Longrightarrow\;\; \mu\,(\,T^{\,3} x_{\,i \,-\, 1}\;,\;T^{\,3}
x_{\,i}\;,\;t\,) \;>\; \varepsilon \hspace{7.5cm}\]
\[\cdots\]
\[ \lambda^{\,m}\;\mu\,(\,T^{\,m} x_{\,i \,-\, 1}\;,\;T^{\,m} x_{\,i}\;,\;t\,)
\;=\; \lambda^{\,m \,-\, 1}\;\left(\, \lambda\;\mu\,(\,T\,(\,T^{\,m \,-\, 1}
x_{\,i \,-\, 1}\,)\;,\; T\,(\,T^{\,m \,-\, 1} x_{\,i}\,)\;,\;t\,) \,\right)\]
\[\hspace{5.2cm} >\; \lambda^{\,m \,-\, 1}\;\mu\,(\,T^{\,m \,-\, 1}
x_{\,i \,-\, 1}\;,\;T^{\,m \,-\, 1} x_{\,i}\;,\;t\,) \;>\; \lambda^{\,m \,-\, 1}
\;\varepsilon\]
\[\Longrightarrow\;\; \mu\,(\,T^{\,m} x_{\,i \,-\, 1}\;,\;T^{\,m} x_{\,i}\;,\;t\,)
\;>\; \varepsilon \;;\hspace{7.5cm}\]
\[ \frac{1}{\lambda^{\,3}}\;\nu\,(\,T^{\,3} x_{\,i \,-\, 1}\;,\;T^{\,3}
x_{\,i}\;,\;t\,) \;=\; \frac{1}{\lambda^{\,2}}\;\left(\, \frac{1}{\lambda}
\;\nu\,(\,T\,(\,T^{\,2} x_{\,i \,-\, 1}\,)\;,\; T\,(\,T^{\,2} x_{\,i}\,)\;,
\;t\,) \,\right)\]
\[\hspace{5.2cm} <\; \frac{1}{\lambda^{\,2}}\;\nu\,(\,T^{\,2} x_{\,i \,-\, 1}
\;,\;T^{\,2} x_{\,i}\;,\;t\,) \;<\; \frac{1}{\lambda^{\,2}}\;(\,1 \;-\;
\varepsilon\,)\]
\[\Longrightarrow\;\; \nu\,(\,T^{\,3} x_{\,i \,-\, 1}\;,\;T^{\,3} x_{\,i}\;,
\;t\,) \;<\; (\,1 \;-\; \varepsilon\,) \hspace{7.5cm}\]
\[\cdots\]
\[ \frac{1}{\lambda^{\,m}}\;\nu\,(\,T^{\,m} x_{\,i \,-\, 1}\;,\;T^{\,m}
x_{\,i}\;,\;t\,) \;=\; \frac{1}{\lambda^{\,m \,-\, 1}}\;
\left(\, \frac{1}{\lambda}\;\nu\,(\,T\,(\,T^{\,m \,-\, 1} x_{\,i \,-\, 1}\,)
\;,\; T\,(\,T^{\,m \,-\, 1} x_{\,i}\,)\;,\;t\,) \,\right)\]
\[\hspace{3.7cm} <\; \frac{1}{\lambda^{\,m \,-\, 1}}\;\nu\,(\,T^{\,m \,-\, 1}
x_{\,i \,-\, 1}\;,\;T^{\,m \,-\, 1} x_{\,i}\;,\;t\,) \;<\;
\frac{1}{\lambda^{\,m \,-\, 1}}\;(\,1 \;-\; \varepsilon\,)\]
\[\Longrightarrow\;\; \nu\,(\,T^{\,m} x_{\,i \,-\, 1}\;,\;T^{\,m} x_{\,i}\;,
\;t\,) \;<\; 1 \;-\; \varepsilon \hspace{7.5cm}\]
Now,\[ \mu\,(\,T^{\,m}\,x \;,\;T^{\,m \,+\, 1}\,x \;,\; t\,) \;=\;
\mu\,(\,T^{\,m}\,x_{\,0} \;,\;T^{\,m}\,x_{\,n} \;,\; t\,) \hspace{3.5cm}\]
\[ \hspace{5.1cm} \geq \; \;\left(\, \mu\,\left(\,T^{\,m}\,x_{\,0} \;,
\;T^{\,m}\,x_{\,1} \;,\; \frac{t}{n}\,\right) \;\ast\;
\mu\,\left(\,T^{\,m}\,x_{\,1} \;,\;T^{\,m}\,x_{\,2} \;,\;
\frac{t}{n}\,\right)\,\right. \]
\[\left. \hspace{6.9cm} \ast \hspace{0.5cm} \cdots \hspace{0.5cm} \ast \;
\mu\,\left(\,T^{\,m}\,x_{\,n \,-\, 1} \;,\;T^{\,m}\,x_{\,n} \;,\;
\frac{t}{n}\,\right)\,\right)\] \[ >\; \varepsilon \hspace{3.0cm}\]
\[ i.e. \hspace{0.5cm} \mu\,(\,T^{\,m}\,x \;,\;T^{\,m \,+\, 1}\,x \;,\; t\,)
\;>\; \varepsilon \hspace{0.5cm} for \;\; all \;\; t \;>\; 0 \; \; and \;\; for
\;\; all \;\; m\;\in\;\mathbf{N}\;;\hspace{5.5cm} \]
and,\[ \nu\,(\,T^{\,m}\,x \;,\;T^{\,m \,+\, 1}\,x \;,\; t\,) \;=\;
\nu\,(\,T^{\,m}\,x_{\,0} \;,\;T^{\,m}\,x_{\,n} \;,\; t\,) \hspace{3.5cm}\]
\[ \hspace{5.1cm} \leq \; \;\left(\, \nu\,\left(\,T^{\,m}\,x_{\,0}
\;,\;T^{\,m}\,x_{\,1} \;,\; \frac{t}{n}\,\right) \;\diamond\;
\nu\,\left(\,T^{\,m}\,x_{\,1} \;,\;T^{\,m}\,x_{\,2} \;,\;
\frac{t}{n}\,\right)\,\right. \]
\[\left. \hspace{6.9cm} \diamond \hspace{0.5cm} \cdots \hspace{0.5cm}
\diamond \; \nu\,\left(\,T^{\,m}\,x_{\,n \,-\, 1} \;,\;T^{\,m}\,x_{\,n}
\;,\; \frac{t}{n}\,\right)\,\right)\] \[ <\; 1 \;-\; \varepsilon \hspace{2.0cm}\]
\[ i.e. \hspace{0.5cm} \nu\,(\,T^{\,m}\,x \;,\;T^{\,m \,+\, 1}\,x
\;,\; t\,) \;<\; 1 \;-\; \varepsilon \hspace{0.5cm} for \;\; all \;\;
t \;>\; 0 \; \; and \;\; for \;\; all \;\; m\;\in\;\mathbf{N}.\hspace{5.5cm} \]
Now, for all \,$t \;>\; 0$\, and \,$j \;<\; k$\, we have,
\[ \mu\,(\,T^{\,j}\,x \;,\;T^{\,k}\,x \;,\; t\,) \;\geq\;
\mu\,\left(\,T^{\,j}\,x \;,\;T^{\,j \,+\, 1}\,x \;,\; \frac{t}{k \,-\, j}\,\right)
\;\;\ast\;\; \mu\,\left(\,T^{\,j \,+\, 1}\,x \;,\;T^{\,j \,+\, 2}\,x
\;,\; \frac{t}{k \,-\, j}\,\right) \]
\[\hspace{4.5cm} \ast \hspace{0.5cm} \cdots \hspace{0.5cm} \ast \;\;
\mu\,\left(\,T^{\,k \,-\, 1}\,x \;,\;T^{\,k}\,x \;,\; \frac{t}{k \,-\, j}\,\right)
\;\;>\;\;\varepsilon \;;\]
\[ \nu\,(\,T^{\,j}\,x \;,\;T^{\,k}\,x \;,\; t\,) \;\leq\;
\nu\,\left(\,T^{\,j}\,x \;,\;T^{\,j \,+\, 1}\,x \;,\; \frac{t}{k \,-\, j}\,\right)
\;\;\diamond\;\; \nu\,\left(\,T^{\,j \,+\, 1}\,x \;,\;T^{\,j \,+\, 2}\,x
\;,\; \frac{t}{k \,-\, j}\,\right) \]
\[\hspace{4.5cm} \diamond \hspace{0.5cm} \cdots \hspace{0.5cm} \diamond
\;\; \nu\,\left(\,T^{\,k \,-\, 1}\,x \;,\;T^{\,k}\,x \;,\;
\frac{t}{k \,-\, j}\,\right) \;\;<\;\;1 \;-\; \varepsilon.\]
$\Longrightarrow \;\; \left\{\,T^{\,j}\,x\,\right\}$\, is a Cauchy
sequence in \,$(\,X \,,\, A\,)$. Since \,$(\,X \,,\, A\,)$\, is
complete, there exists \,$\xi \;\in\; X$\, such that \,$T^{\,i}\,x
\;\longrightarrow\; \xi$\, as \,$i \,\longrightarrow\, \infty$\, in
\,$(\,X \,,\, A\,)$. Again, since \,$T$\, is IF-uniformly locally
contractive, it follows that \,$T$\, is t-uniformly continuous on X
and hence by the lemma(\ref{L2}), we get \[ T\,\xi \;=\; \mathop
{\lim }\limits_{i\;\, \to \;\,\infty }T\,T^{\,i}x \;=\; \mathop
{\lim }\limits_{i\;\, \to \;\, \infty }T^{\,i \;+\; 1}x \;=\; \xi\]
which shows that \,$\xi$\, is a fixed point of \,$T$.
|
\section{Introduction}\label{intro}
The X-ray binary PSR~B1259$-$63 is a long-period system, formed by a Be star and a young and powerful pulsar with spin-down luminosity $L_{\rm sd}\approx 8\times 10^{35}$~erg~s$^{-1}$
(\cite{man95}). Non-thermal emission in radio (e.g. \cite{joh05}), X-rays (e.g. \cite{com94,uch09,che09}) and gamma rays (e.g. \cite{aha05}) has been detected from this source. This
emission is thought to originate in the region where the star and pulsar winds collide (\cite{tav97}). Therefore, PSR~B1259$-$63 is not an accretion powered system, like many Be X-ray
binaries, since the pulsar wind ram pressure keeps the stellar wind beyond the neutron star gravitational capture radius (\cite{bon44}). The two-winds collision leads to two shocks, one in
the stellar wind, and another in the pulsar wind. The contact discontinuity between the two flows is located where the wind ram pressures are equal, at a minimum distance from the pulsar
$R_{\rm off}=\sqrt{\eta}\,R_{\rm orb}/(1+\sqrt{\eta})$, where $R_{\rm orb}$ is the star-pulsar separation distance, $\eta=L_{\rm dsd}/\dot{M}_*\,v_{\rm w}\,c$, and $\dot{M}_*$ and $v_{\rm
w}$ are the stellar mass loss rate and wind velocity, respectively (\cite{bog08}). Other three high-mass binaries in the Galaxy may harbor a non-accreting pulsar: LS~I~+61~303, LS~5039 and
HESS~J0632$+$057 (e.g. \cite{mar81,mar05,dub06,dha06,che06,sie08,hin09}), although their accreting nature cannot be discarded yet (e.g. see the discussion in \cite{bos09}; see also
\cite{rom07,mas09}).
In binaries hosting a non-accreting pulsar, the X- and gamma rays are produced close or within the binary system, where the shocks are the strongest. X-rays are of likely synchrotron
origin, and gamma rays, of inverse Compton (IC) nature (e.g. \cite{kir99,kha07,sier08,tak09,ker11}; see however \cite{ner07}), and the cooling timescales of the electrons and positrons
emitting at these energies are probably very short. These particles are likely accelerated in the pulsar wind shock, more suitable for particle acceleration than the slower and less
energetic stellar wind shock. Accelerated particles are expected to follow a power-law of index $p$ ($\propto E^{-p}$) between a minimum ($E_{\rm min}$) and a maximum energy ($E_{\rm
max}$), cooling down while advected away in the shocked flow. At the scales of the shock, $\sim R_{\rm off}$, synchrotron and IC cooling dominate, but at farther distances energy losses are
dominated by adiabatic cooling (i.e. work). The velocity at which the flow leaves the shock region, and eventually the system, determines how much radiation is produced at places where
adiabatic cooling is dominant. This is so because the non-thermal power scales as $\propto 1/t_{\rm ad}\sim R/v_{\rm esc}$, where $R$ and $v_{\rm esc}$ are the characteristic size of the
flow and the shock escape velocity, respectively. The emission coming from the outskirts of the binary and beyond can be better studied in radio, and provides information on the material
flowing away from the system. An important factor affecting the colliding wind region is the development of instabilities, which can mix the shocked pulsar wind with the much denser and
colder stellar wind.
High-resolution VLBI studies can be of very much help characterizing the shocked flows. Remarkably, extended radio emission has been recently detected from PSR~B1259$-$63 (\cite{mol11}; see
also \cite{dha06} and \cite{rib08} regarding the candidates LS~I~+61~303 and LS~5039). To fully profit from these observations, however, detailed radiation modeling is required. From the
radio emission from binaries hosting a young pulsar, one can derive important information of the shocked medium, like its velocity and magnetic field, both to be affected by mixing. The
emitting particles themselves can also be studied, since the value of $E_{\rm min}$ affects strongly the number of the radio emitting particles. It is useful to compute maps of the radio
emitter, to compare with observations. A study of the radio emission appearance was already done in \cite{dub06}, although the radio emitter was treated there as 1-dimensional, with
point-like injection. In this work, we present calculations done adopting a prescription for the shocked star-pulsar wind structure that accounts for the 3-dimensional extension of the
particle injection and the radio emitter. The results are preliminary, but they can already shed light on the properties of the radio emitting flow in binaries hosting a non-accreting
pulsar. In Figure~1 we show a sketch of the considered problem.
\begin{figure}[]
\begin{center}
\includegraphics[width=0.5\textwidth]{sk.eps}
\caption{Sketch of the scenario and the model presented here. The pulsar and the stellar wind collide forming two shocks
separated by a contact discontinuity. The shocked material flows away and gets spiral-shaped by the orbital motion of the
system. Electrons and positrons are injected in the interaction surface, in the form of different rings,
at different distances from the
pulsar. These particles flow with the shocked material producing radio to X-rays through synchrotron, and X-rays to gamma
rays through IC scattering. Here we concentrate in the radio emission. Overlaid on the system and interaction region, the
contours of the source radio image are shown.}
\label{p1}
\end{center}
\end{figure}
\section{The model}
We model the radio emitter injecting particles, electrons and positrons, as rings. These rings are distributed along a paraboloid that is connected to a cone beyond a certain distance. The
flux of energy at the shock, going to non-thermal particles, is computed from $\sin(\theta)\,L_{\rm sd}/4\pi\,l_{\rm p0}^2c$, where $\theta$ is the angle between the shock surface and the
pulsar radial direction, and $l_{\rm p0}$ is the distance from a particular injection point to the pulsar (or characteristic injection region size). The paraboloid is initially defined as
$r=R_{\rm orb}\sqrt{y-d_{\rm off}}$, where $r$ is the radius of the paraboloid, $y$ the distance to the star from the paraboloid axis, and $d_{\rm off}=R_{\rm orb}-R_{\rm off}$. The final
opening angle of the cone is: $\approx 0.5\,(4-\eta^{2/5})\,\eta^{1/3}$, which corresponds in reality to the opening angle of the contact discontinuity between the two winds \cite{bog08},
simplified here as the whole shocked structure. Particles are advected away along the parabolic/conical surface in the direction opposite to the star. In fact, these rings of particles are
injected, with time, in different locations following the pulsar orbital motion around the companion. Once injected, rings are assumed to follow ballistic motion, and thus the shape of the
whole structure at large scales tends to form a spiral. Note that ballistic motion is just a rough approximation; the shocked stellar wind exerts a Coriolis force on the pulsar wind,
bending it even farther in the direction opposite to the pulsar orbital motion.
We do not adopt here the Kennel \& Coroniti solution (\cite{ken84a}) for the postshock flow, as it was done in \cite{dub06}. The assumption of a flow being like the one treated in
\cite{ken84a}, of spherical nature, is not realistic here since the flow evolution is strongly affected by the inhomogeneity and anisotropy of the pulsar environment (i.e. the stellar
wind). This is clearly seen in the results of \cite{bog08}. However, instabilities were not considered in that work, and simulations including instability development have not been
performed yet. For this reason, some parameters are treated here phenomenologically. The flow velocity is derived defining a velocity of the shocked flow along its trajectory, parabolic
first, straight later on. The velocity is taken to be $c/3$ at the injection locations, although this is only strictly true at the regions in which the pulsar shock is roughly
perpendicular. At distances bigger than the injection region ($>l_{\rm p0}$), we have assumed that the flow speed decreases exponentially down to some intermediate velocity $v_{\rm
w}<v_{\rm esc}<c/3$ (typically $v_{\rm w}\sim (1-2)\times 10^8$~cm~s$^{-1}$). This decrease in the bulk flow velocity relates to mixing due to hydrodynamical instabilities\footnote{As shown
in \cite{bog08}, a hydrodynamical approximation for the flow seems appropriate.} in the contact discontinuity, most likely of Kelvin-Helmholtz nature. Kelvin-Helmholtz instabilities would
start to grow at scales of the order of those of the shock region, thus mainly affecting the radio emitting region. Eventually, the final turbulent and mixed flow velocity should be close
to the one obtained assuming momentum conservation, due to significant kinetic to internal/turbulent energy conversion. In this work, $v_{\rm esc}$ is fixed to $10^9$~cm~s$^{-1}$ for
simplicity. Adiabatic losses have been computed as given in Sect.~\ref{intro}. To compute synchrotron losses, we have adopted a magnetic field $B_0\approx 3\sqrt{\sigma\,2\,L_{\rm
sd}/l_{\rm lp0}^2\,c}$ (\cite{ken84b}), and $B$ decreases farther in the shocked flow as $1/l_{\rm p}$. Turbulent mixing should affect $B$, but we have not accounted for it at this stage.
The value of $E_{\rm min}$ has been fixed to $\sim 0.1\Gamma_{\rm p}\,m_{\rm e}\,c^2$, where $\Gamma_{\rm p}$ is the Lorentz factor of the particles in the pulsar wind (\cite{ken84b}),
fixed by us to $10^5$. The value of $E_{\rm max}$ has been set to 1~TeV. The spin-down luminosity of the pulsar has been taken $L_{\rm sd}=10^{36}$~erg~s$^{-1}$. Regarding the system
properties, the orbit is circular, with $R_{\rm orb}=3\times 10^{12}$~cm and an inclination angle $i=45^\circ$. The inferior conjunction of the compact object corresponds to phase 0.5. The
star luminosity has been fixed to $5\times 10^{38}$~erg~s$^{-1}$, and the temperature to $3\times 10^4$~K. The stellar mass loss rate is $\dot{M}_*=10^{-6}\,M_{\odot}$/yr, and $v_{\rm
w}=2\times 10^8$~cm. The IC losses are derived from the stellar photon energy density (see \cite{bos09} and references therein). Densities are low enough to neglect ionization/coulombian
losses and relativistic Bremsstrahlung.
The radio emission has been calculated from the evolved particle populations after propagating from their injection point, at $l_{\rm p0}$. Synchrotron self-absorption, and free-free
absorption in the stellar wind (ionization fraction of 0.1), have been taken into account. The impact of these absorption processes is rather small provided that most of the radio emission
comes from a fairly large region, well outside the binary.
\section{Results and discussion}
\begin{figure}[]
\begin{center}
\includegraphics[width=0.7\textwidth]{psr.ps}
\includegraphics[width=0.7\textwidth]{psr2.ps}
\caption{Top: Simulated radio map at 5~GHz, smoothed with a gaussian of FWHM $\approx 1$~mas,
for the generic case studied here and phases 0 and 0.5. Intensity units are mJy per beam. Axis units are in mas.
Bottom: The same as in the top of the figure, but without gaussian smoothening.}
\label{p1}
\end{center}
\end{figure}
In Figure~2, top, 5~GHz radio maps are computed for a generic binary system and two different orbital phases. The maps have been obtained smoothening the computed flux-per-beam values with
a 1~milliarcsecond (mas) FWHM Gaussian, reproducing the effect of a radio interferometer with a 1~mas beam. In Figure~2, bottom, the same maps without Gaussian smoothening are also shown.
From Fig.~2, it is clear that the overall picture differs strongly from a point-like source. The emitter spiral shape is hinted in Fig.~2, top, and much more clear in Fig.~2, bottom. The
center of gravity is displaced by few mas between phases 0 and 0.5, and the total flux in both phases is $\approx 13$~mJy. Despite the flow prescriptions are quite distinct, and that we
have accounted for the 3-dimensional extension of particle injection and the emitter, our results and those of \cite{dub06} are qualitatively similar.
The adopted value for $\sigma$, 0.1, may appear rather high. In the case of Crab, the parameter $\sigma$ is much lower (\cite{ken84a}). A higher $\sigma$ in our case is however plausible if
the transfer of energy, from the Pointing flux to matter in the pulsar wind, is a process extended in space. The conversion may even take place in the termination shock itself
(\cite{pet07}). Since $R_{\rm off}$ in a close binary system is much smaller than in Crab, the $\sigma$-value may well be significantly higher in the former case. Another possibility would
be an effective increase (or decrease) of $B$ through the mixing of the shocked stellar and pulsar winds, as the radio emitting plasma gets entangled with the stellar wind magnetic field.
Concerning $v_{\rm esc}$, simulations predict values close to $c$ due to the acceleration of the shocked flow because of strong pressure gradients \cite{bog08}. Under such conditions, as
shown in \cite{kha08}, adiabatic cooling timescales can be indeed very short. However, instabilities and mixing at scales significantly larger than $R_{\rm off}$ should slow down the flow.
This has been included here and allow radio fluxes to reach significant values. Otherwise, the emission would be lower by more than one order of magnitude, i.e. radio fluxes $<1$~mJy.
Nowadays, radio observations start to allow detailed comparisons with models. In this paper, we present preliminary results of such a comparisons in the context of a semi-phenomenological
work, based on current state of our knowledge on the flow structure. A more detailed study of the properties of the radio emitter, as well as applications to specific objects, will be
presented elsewhere. We remark that, to better characterize the shocked wind evolution, farther numerical magnetohydrodynamics simulations are required, with a detailed study of the
instability development. Pulsar wind physics is also a very important but open issue, central for wind propagation, particle acceleration at the wind termination, and postshock region
conditions.
\acknowledgments
V.B-R. want to thank Dmitry Khangulyan for very helpful discussions on the topic treated here.
The research leading to these results has received funding from the European Union
Seventh Framework Program (FP7/2007-2013) under grant agreement
PIEF-GA-2009-252463.
V.B.-R. acknowledges support by the Spanish Ministerio de Ciencia e
Innovaci\'on (MICINN) under grants AYA2010-21782-C03-01 and
FPA2010-22056-C06-02.
|
\section{Introduction}
Let $T$ be a connected abelian Lie group, i.e. a torus.
Let $(M,\omega)$ be a closed $2n$-dimensional Hamiltonian $T$-manifold with the moment map $\mu \colon M \rightarrow \textbf{t}^*$, where $\textbf{t}$ is the Lie algebra of $T$. A Liouville measure $m_L$ on $M$ is defined by
$$m_L(U) := \int_U \frac{\omega^n}{n!} $$ for any open set $U \subset M$. Then the push-forward measure $m_{\mathrm{DH}} := \mu_* m_L$, called
the \textit{Duistermaat-Heckman measure}, can be regarded as a measure on $\mathfrak{t^*}$ such that for any Borel subset $B
\subset \mathfrak{t^*}$, $m_{\mathrm{DH}}(B) = \int_{\mu^{-1}(B)}
\frac{\omega^n}{n!}$ tells us that how many states of our system have momenta in $B.$
Due to \cite{DH}, the Duistermaat-Heckman measure is absolutely continuous with respect to the Leabesque measure on $\mathbf{t}^*$ and the corresponding density function, denoted by $\mathrm{DH}$ and called the $\textit{Duistermaat-Heckman function}$, is a polynomial on any regular open set.
More precisely, for any regular value $\xi \in \textbf{t}^*$, $$\mathrm{DH}(\xi) = \int_{M_{\xi}} \frac{1}{(n-l)!}\omega_{\xi}^{n-l}$$ where $l$ is a dimension of $T$, $M_{\xi}$ is the symplectic reduction at $\xi$, and $\omega_{\xi}$ is the corresponding reduced symplectic form on $M_{\xi}$.
In 1996, W.Graham \cite{Gr} proved that if the Hamiltonian $T$-action on a K\"{a}hler manifold $(M,\omega,J)$ is holomorphic, then the Duistermaat-Heckman measure is log-concave, i.e. $\log{\mathrm{DH}}$ is a concave function on the image of the moment map.
\begin{example}\label{ex1}
Consider a compact symplectic toric manifold $(M,\omega,T)$ with the moment map $\mu \colon M \rightarrow \textbf{t}^*$. Since any symplectic
toric manifold can be obtained by the K\"{a}hler reduction, $M$ is a smooth toric variety and the given $T$-action is holomorphic with
some $\omega$-compatible complex structure on $M$.
By Atiyah-Guillemin-Sternberg convexity theorem, the image of $\mu$ is a rational convex polytope satisfying the non-singularity condition.
Choose any circle subgroup $S^1 \subset T$. Then the moment map of the circle action on $M$ is a composition map $\mu_{S^1} \colon M \stackrel{\mu}\longrightarrow \textbf{t}^* \stackrel{\pi}\longrightarrow \textbf{s}^*$ where $\pi$ is a dual map of $i \colon \textbf{s} \rightarrow \textbf{t}$, the inclusion of Lie algebra induced by $S^1 \hookrightarrow T$. The corresponding Duistermaat-Heckman function $\mathrm{DH}(t)$ is just $C \cdot \mathrm{vol}(\mu(M) \cap \pi^{-1}(t))$ with respect to the Lebesque measure on $\textbf{t}^*$, which is log-concave by Graham's theorem \cite{Gr}. Here, $C$ is a global constant depending on the embedding $S^1 \hookrightarrow T$. Therefore, for any given Delzant polytope $\triangle$ and any height function $h(t) \colon \triangle \rightarrow \R$, the corresponding slice volume function $g : h(\triangle) \rightarrow \R$ defined by $g(t) := \mathrm{vol}(h^{-1}(t))$ is log-concave.
The following figure is the case when $M = S^2 \times S^2 \times S^2$, $\omega = 2\omega_{FS} \oplus 3\omega_{FS} \oplus 6\omega_{FS}$, and a Hamiltonian circle action on $M$ is given by
$t \cdot (z_1,z_2,z_3) = (tz_1, tz_2, tz_3)$ for any $t \in S^1$.
\begin{figure}[ht]
\centering
\psset{unit=20pt}
\begin{pspicture}(0,-9)(27,8)\footnotesize
\pspolygon[fillstyle=solid, linewidth=1.5pt](2,0)(5,3)(5,4.5)(4,5.5)(1,2.5)(1,1) \psline(9,0)(9,5.5)
\psline(2,0)(2,1.5) \psline(2,1.5)(5,4.5) \psline(2,1.5)(1,2.5)
\psdot(4,4)\psdot(2,0)\psdot(1,1)\psdot(5,3)\psdot(5,4.5)\psdot(2,1.5)\psdot(1,2.5)\psdot(4,5.5)
\psdot(9,0)\psdot(9,1)\psdot(9,1.5)\psdot(9,2.5)\psdot(9,3)\psdot(9,4)\psdot(9,4.5)\psdot(9,5.5)
\psline[linestyle=dashed,linecolor=black](1,1)(4,4)
\psline[linestyle=dashed,linecolor=black](4,4)(4,5.5)
\psline[linestyle=dashed,linecolor=black](4,4)(5,3)
\rput(2,-0.5){\tiny{$(S,S,S)$}}
\rput(0.2,1){\tiny{$(N,S,S)$}}
\rput(3,1.7){\tiny{$(S,N,S)$}}
\rput(0.2,2.5){\tiny{$(N,N,S)$}}
\rput(6,3.2){\tiny{$(S,S,N)$}}
\rput(6,4.7){\tiny{$(S,N,N)$}}
\rput(3.4,4.2){\tiny{$(N,S,N)$}}
\rput(4,6){\tiny{$(N,N,N)$}}
\rput(9.4,0){\tiny{$0$}}
\rput(9.4,1){\tiny{$2$}}
\rput(9.4,1.5){\tiny{$3$}}
\rput(9.4,2.5){\tiny{$5$}}
\rput(9.4,3){\tiny{$6$}}
\rput(9.4,4){\tiny{$8$}}
\rput(9.4,4.5){\tiny{$9$}}
\rput(9.4,5.5){\tiny{$11$}}
\rput(12,3.2){\tiny{DH}}
\rput(7.5,2.4){$\mu$}
\rput(16.2,-0.2){\tiny{$0$}}
\rput(17.2,-0.2){\tiny{$2$}}
\rput(17.8,-0.2){\tiny{$3$}}
\rput(19,-0.2){\tiny{$5$}}
\rput(19.6,-0.2){\tiny{$6$}}
\rput(20.8,-0.2){\tiny{$8$}}
\rput(21.4,-0.2){\tiny{$9$}}
\rput(22.6,-0.2){\tiny{$11$}}
\rput(23.9,-0.2){$\mu(M)$}
\rput(15,6){$\mathrm{DH}(t)$}
\psdot(16,0)\psdot(17.2,0)\psdot(17.8,0)\psdot(19,0)\psdot(19.6,0)\psdot(20.8,0)\psdot(21.4,0)\psdot(22.6,0)
\pscurve(16,0)(16.15,0.03125)(16.3,0.125)(16.45,0.28125)(16.6,0.5)(16.75,0.78125)(16.9,1.125)(17.05,1.53125)(17.2,2)
\psdot(17.2,2)
\psline(17.2,2)(17.8,4)
\psdot(17.8,4)
\pscurve(17.8,4)(18.1,4.875)(18.4,5.5)(18.7,5.875)(19,6)
\psdot(19,6)
\psline(19,6)(19.6,6)
\psdot(19.6,6)
\pscurve(19.6,6)(19.9,5.875)(20.2,5.5)(20.5,4.875)(20.8,4)
\psdot(20.8,4)
\psline(20.8,4)(21.4,2)
\psdot(21.4,2)
\pscurve(21.4,2)(21.7,1.125)(22,0.5)(22.3,0.125)(22.6,0)
\psline(8,-5)(16,-5)
\psline(9,-9)(9,-2)
\rput(7.8,-2){$\log\mathrm{DH}(t)$}
\rput(17,-5.2){$\mu(M)$}
\psdot(9,-5)\psdot(10.2,-5)\psdot(10.8,-5)\psdot(12,-5)\psdot(12.6,-5)\psdot(13.8,-5)\psdot(14.4,-5)\psdot(15.6,-5)
\rput(9.2,-5.2){\tiny{$0$}}
\rput(10.2,-5.2){\tiny{$2$}}
\rput(10.8,-5.2){\tiny{$3$}}
\rput(12,-5.2){\tiny{$5$}}
\rput(12.6,-5.2){\tiny{$6$}}
\rput(13.8,-5.2){\tiny{$8$}}
\rput(14.4,-5.2){\tiny{$9$}}
\rput(15.6,-5.2){\tiny{$11$}}
\pscurve(9.15,-8.46573)(9.3,-7.07944)(9.45,-6.26851)(9.6,-5.69314)(9.75,-5.24686)(9.9,-4.88222)(10.05,-4.57392)(10.2,-4.30686)
\psdot(10.2,-4.30686)
\pscurve(10.2,-4.30686)(10.5,-3.88)(10.8,-3.6137)
\psdot(10.8,-3.6137)
\pscurve(10.8,-3.6137)(11.1,-3.41588)(11.4,-3.29525)(11.7,-3.22929)(12,-3.20824)
\psdot(12,-3.20824)
\psline(12,-3.20824)(12.6,-3.20824)
\psdot(12.6,-3.20824)
\pscurve(12.6,-3.20824)(12.9,-3.22929)(13.2,-3.29525)(13.5,-3.41588)(13.8,-3.6137)
\psdot(13.8,-3.6137)
\pscurve(13.8,-3.6137)(14.1,-3.88)(14.4,-4.30686)
\psdot(14.4,-4.30686)
\pscurve(14.4,-4.30686)(14.7,-4.88222)(15,-5.69314)(15.3,-7.07944)(15.45,-8.46573)
\psline(15,0)(23,0)
\psline(16,0)(16,6)
\psline{->}(7,2.8)(8,2.8)
\psline{->}(11,2.8)(13,2.8)
\psline[linestyle=dashed,linecolor=red,linewidth=0.2pt](2,0)(9,0)
\psline[linestyle=dashed,linecolor=red,linewidth=0.2pt](1,1)(9,1)
\psline[linestyle=dashed,linecolor=red,linewidth=0.2pt](2,1.5)(9,1.5)
\psline[linestyle=dashed,linecolor=red,linewidth=0.2pt](5,3)(9,3)
\psline[linestyle=dashed,linecolor=red,linewidth=0.2pt](5,4.5)(9,4.5)
\psline[linestyle=dashed,linecolor=red,linewidth=0.2pt](1,2.5)(9,2.5)
\psline[linestyle=dashed,linecolor=red,linewidth=0.2pt](4,5.5)(9,5.5)
\psline[linestyle=dashed,linecolor=red,linewidth=0.2pt](4,4)(9,4)
\end{pspicture}
\caption{Example \ref{ex1}} \label{figure: illustrate}
\end{figure}
\end{example}
Note that any holomorphic Hamiltonian $S^1$-manifold $(M,\omega,J)$ with an $S^1$-invariant K\"{a}hler structure $\omega$ satisfies the followings.
\begin{enumerate}
\item $\omega$ satisfies the Hard Lefschetz property,
\item any reduced symplectic form $\omega_t$ is K\"{a}hler, and hence it satisfies the Hard Lefschetz property, and
\item the Duistermaat-Heckman measure is log-concave.
\end{enumerate}
Hence it is natural to ask that when a Hamiltonian $S^1$-manifold $(M,\omega)$ and the symplectic reduction $(M_t, \omega_t)$ satisfy the Hard Lefschetz property, and when the corresponding Duistermaat-Heckman measure is log-concave.
In this paper, we focus on the case when $(M,\omega)$ is a closed $2n$-dimensional symplectic manifold with a semifree Hamiltonian $S^1$-action whose fixed point set $M^{S^1}$ consists of isolated points. Let $\mu \colon M \rightarrow \R$ be a corresponding moment map which satisfies $i_X\omega = -d\mu$.
As noted above, the Duistermaat-Heckman function $\mathrm{DH} \colon \mu(M) \rightarrow \R$ is defined by $$\mathrm{DH}(t) = \int_{M_t} \frac{1}{(n-1)!} \omega_t^{n-1}$$ where $M_t$ is the reduced space $\mu^{-1}(t)/S^1$ and $\omega_t$ is the reduced symplectic form on $M_t$.
Now we state our main results.
\begin{theorem}\label{first}
Let $(M,\omega)$ be a closed symplectic manifold with a semifree Hamiltonian $S^1$-action whose fixed point set $M^{S^1}$ consists of isolated points. Then the Duistermaat-Heckman measure is log-concave.
\end{theorem}
\begin{theorem}\label{second}
Let $(M,\omega)$ be a closed semifree Hamiltonian $S^1$-manifold whose fixed points are all isolated, and let $\mu$ be the moment map. Then $\omega$ satisfies the Hard Lefschetz property. Moreover, the reduced symplectic form $\omega_t$ satisfies the Hard Lefschetz property for every regular value $t$.
\end{theorem}
In Section 2, we briefly review the Tolman and Weitsman's work \cite{TW1} which is very powerful to analyze the equivariant cohomology of the Hamiltonian $S^1$-manifold with isolated fixed points. Especially we use the Tolman-Weitsman's basis of the equivariant cohomology $H^*_{S^1}(M)$ which is constructed by using the equivariant version of the Morse theory \cite{TW2}.
In Section 3, we express the Duistermaat-Heckman function explicitly in terms of the integration of some cohomology class on the reduced space. And then we compute the integration by using the Jeffrey-Kirwan residue formula \cite{JK}. Consequently, we will show that the log-concavity of the Duistermaat-Heckman measure is completely determined by
the set of pairs $\{(\mu(F),m_F)_F | F \in M^{S^1}\}$, where $\mu(F)$ is the image of the moment map of $F$ and $m_F$ is the product of all weights of the $S^1$-representation on $T_F M$. In Sections 4, we will prove Theorem \ref{first}. And we will prove Theorem \ref{second} in Section 5.
\begin{remark}\label{ree}
As noted in \cite{O2}, \cite{K} and \cite{Lin}, Ginzburg and Knudsen conjectured that for any closed Hamiltonian $T$-manifold, the corresponding Duistermaat-Heckman measure is log-concave. Note that the log-concavity conjecture is motivated by a concavity property of the Boltzmann' entropy in statistical mechanics. (See \cite{O3} for the detail.)
The log-concavity problem of the Duistermaat-Heckman measure is proved by A.Okounkov \cite{O1} when $M$ is a co-adjoint orbit of the classical Lie groups of type $A_n$, $B_n$, or $C_n$ with the maximal torus $T$-action, by using the log-concavity of the multiplicities of the irreducible representation of each Lie groups.
But the counter-example was found by Y.Karshon \cite{K}. By using the Lerman's symplectic cutting method, she constructed a closed 6-dimensional semifree Hamiltonian $S^1$-manifold with two fixed components such that the Duistermaat-Heckman measure is not log-concave. Later, Y.Lin \cite{Lin} generalized the construction of 6-dimensional Hamiltonian $S^1$-manifolds not satisfying the log-concavity of the Duistermaat-Heckman measure, and also proved the log-concavity conjecture for any Hamiltonian $T^{n-2}$-manifold $(M^{2n},\omega)$ whose reduced space has $b_2^+ = 1$.
\end{remark}
\section{Tolman-Weitsman basis of the equivariant cohomology $H^*_{S^1}(M;\Z)$}
In this section, we breifly review Tolman and Weitsman's results in \cite{TW1}.
Throughout this section, we assume that $(M^{2n},\omega)$ is a closed semifree Hamiltonian $S^1$-manifold whose fixed points are isolated.
Note that for each fixed point $p \in M^{S^1}$, \textit{the index of $p$} is the Morse index of the moment map at $p$ which is the same as the twice of the number of negative weights of the tangential $S^1$-representation at $p$.
\begin{proposition}\cite{TW1}
Let $N_k$ be the number of fixed points of index $2k$. Then $N_k = {n \choose k}$. Hence $N_k$ is the same as the one of the standard diagonal circle action on $(S^2 \times \cdots \times S^2, \omega_1 \oplus \cdots \oplus \omega_n)$, where $\omega_i$ is the Fubini-Study form on $S^2$ of $i$-th factor.
\end{proposition}
\begin{theorem}\cite{TW1}\label{basis}
Let $2^{[n]}$ be the power set of $\{1,\cdots,n\}$. Then there exist a bijection $\phi : M^{S^1} \rightarrow 2^{[n]}$ such that
\begin{enumerate}
\item For each index-$2k$ fixed point $x \in M^{S^1}$, $|\phi(x)| = k$.
\item Let $u$ be the generator of $H^*(BS^1,\Z)$. For each index-$2k$ fixed point $x \in M^{S^1}$, there exists a unique cohomology class $\alpha_x \in H^{2k}_{S^1}(M;\Z)$ such that for any $x' \in M^{S^1}$, \\
$\alpha_x|_{x'} = u^k$ if $\phi(x) \subset \phi(x')$. \\
$\alpha_x|_{x'} = 0$ otherwise.
\end{enumerate}
Moreover $\{\alpha_x | x \in M^{S^1} \}$ forms a basis of $H^*_{S^1}(M,\Z)$.
\end{theorem}
Applying Theorem \ref{basis} to $(S^2 \times \cdots \times S^2, \omega_1 \oplus \cdots \oplus \omega_n)$ with the diagonal semifree Hamiltonian circle action, we have a bijection
$\psi : (S^2 \times \cdots \times S^2)^{S^1} \rightarrow 2^{[n]}$ and there is a basis $\{\beta_y| y \in (S^2 \times \cdots \times S^2)^{S^1} \}$
of $H^*_{S^1}(S^2 \times \cdots \times S^2;\Z)$ satisfies the conditions in Theorem \ref{basis}. Hence we have an identification map
$$\psi^{-1} \circ \phi : M^{S^1} \rightarrow (S^2 \times \cdots \times S^2)^{S^1}$$ and $\psi^{-1} \circ \phi$ preserves the indices of the fixed points.
Note that $\psi^{-1} \circ \phi$ gives an identification between $H^*_{S^1}(M;\Z)$ and $H^*_{S^1}(S^2 \times \cdots \times S^2;\Z)$ as follow.
Let $a_i = \alpha_{\phi^{-1}\{i\}} \in H^2_{S^1}(M;\Z)$ and $b_i = \beta_{\psi^{-1}\{i\}} \in H^2_{S^1}(S^2 \times \cdots \times S^2;\Z)$.
\begin{lemma}\cite{TW1}\label{tol}
For each $x \in M^{S^1}$, $\alpha_x = \prod_{j \in \phi(x)} a_j$. Similarly, we have $\beta_y = \prod_{j \in \psi(y)} b_j$ for each $y \in
(S^2 \times \cdots \times S^2)^{S^1}$.
\end{lemma}
\begin{proof}
For an inclusion $i \colon M^{S^1} \hookrightarrow M$, we have a natural ring homomorphism
$i^* \colon H^*_{S^1}(M) \rightarrow H^*_{S^1}(M^{S^1}) \cong H^*(M^{S^1}) \otimes H^*(BS^1)$. Kirwan injectivity implies that
$i^*$ is injective. Hence it is enough to show that $\alpha_x|_z = (\prod_{j \in \phi(x)} a_j)|_z$ for all $x,z \in M^{S^1}$.
Here, the operation $``|_z"$ is the restriction $H^*_{S^1}(M^{S^1}) \rightarrow H^*_{S^1}(z) \cong H^*(BS^1)=\R[u]$ induced by the inclusion
$z \hookrightarrow M^{S^1}$. For any $x,z \in M^{S^1}$ with $\mathrm{Ind}(x) = 2k$,
\begin{itemize}
\item $\alpha_x|_{z} = u^k$ if $\phi(x) \subset \phi(z)$.
\item $\alpha_x|_{z} = 0$ otherwise.
\end{itemize}
On the other hand, $(\prod_{j \in \phi(x)} a_j)|_z = \prod_{j \in \phi(x)} a_j|_z$. Since $a_j|_z = u$ if and only if $j \in \phi(z)$, we have
\begin{itemize}
\item $(\prod_{j \in \phi(x)} a_j)|_z = u^k$ if $\phi(x) \subset \phi(z)$.
\item $(\prod_{j \in \phi(x)} a_j)|_z = 0$ otherwise.
\end{itemize}
Therefore, $\alpha_x = \prod_{j \in \phi(x)} a_j$ by the Kirwan injectivity theorem. The proof of the second statement is similar.
\end{proof}
Hence the $H^*(BS^1)$-module isomorphism $f : H^*_{S^1}(M;\Z) \rightarrow H^*_{S^1}(S^2 \times \cdots \times S^2;\Z)$ which sends $\alpha_x$ to $\beta_{\psi^{-1} \circ \phi(x)}$ for each $x \in M^{S^1}$ is in fact a ring isomorphism by the lemma \ref{tol}.
To sum up, we have the following corollary.
\begin{corollary}\label{iso}\cite{TW1}
There is a ring isomorphism $f : H^*_{S^1}(M;\Z) \rightarrow H^*_{S^1}(S^2 \times \cdots \times S^2;\Z)$ which sends $\alpha_x$ to $\beta_{\psi^{-1} \circ \phi(x)}$. Moreover, for any $\alpha \in H^*_{S^1}(M;\Z)$ and any fixed point $x \in M^{S^1}$,
we have $\alpha|_x = f(\alpha)|_{\psi^{-1} \circ \phi(x)}$.
\end{corollary}
\section{The Duistermaat-Heckman function and the residue formula}
Let $(M,\omega)$ be a $2n$-dimensional closed Hamiltonian $S^1$-manifold with the moment map $\mu \colon M \rightarrow \R$. We may assume that $0$ is a regular value of $\mu$ such that $\mu^{-1}(0)$ is non-empty. Choose two consecutive critical values $c_1$ and $c_2$ of $\mu$ so that the open interval $(c_1,c_2)$ consists of regular values of $\mu$ and contains $0$. By the Duistermaat-Heckman's theorem \cite{DH}, $[\omega_t] = [\omega_0] - et$ where $e$ is the Euler class of $S^1$-fibration $\mu^{-1}(0) \rightarrow M_0$, where $M_0$ is the symplectic reduction at 0 with the induced symplectic form $\omega_0$. Hence we have
\begin{equation}\label{Duist}
\mathrm{DH}(t) = \int_{M_0} \frac{1}{(n-1)!}([\omega_0] - et)^{n-1}
\end{equation}
on $(c_1,c_2) \subset \textrm{Im} \mu$.
Note that a continuous function on an open interval $g : (a,b) \rightarrow \R$ is \textit{concave} if $g(tc + (1-t)d) \geq tg(c) + (1-t)g(d)$ for any $c,d \in (a,b)$ and for any $t \in (0,1)$. We remark the basic property of a concave function as follow.
\begin{remark}\label{logconcave}
Let $g$ be a continuous, piecewise smooth function on a connected interval $I \subset \R$. Then g is concave on $I$ if and only if
the derivative of $g$ is decreasing, i.e. $g''(t) \leq 0$ for every smooth point $t \in I$ and $g'_+(c) - g'_-(c) < 0$ for every singular point $c \in I$, where $g'_+(c) = \lim_{t\rightarrow c, t > c} g'(t)$ and $g'_-(c) = \lim_{t\rightarrow c, t < c} g'(t)$.
\end{remark}
Note that Duistermaat and Heckman proved that $\mathrm{DH}$ is a polynomial on a connected regular open interval $U \subset \mu(M)$.
The following formula due to Guillemin, Lerman and Sternberg describes the behavior of $\mathrm{DH}$ near the critical value of $\mu$.
In particular, it implies that $\mathrm{DH}$ is $k$-times differentiable at a critical value $c \in \mu(M)$ if and only if
$\mu^{-1}(c)$ does not contain a fixed component whose codimension is less than $4+2k$.
\begin{theorem}\cite{GLS}\label{GLS}
Assume that $c$ is a critical value which corresponds to the fixed components $C_{i}$'s. Then the jump of $DH(t)$ at $c$ is given by $$DH_+ - DH_- = \sum_{i} \frac{\textrm{vol}(C_i)}{(d-1)!\prod_ja_j} (t-c)^{d-1} + O((t-c)^d)$$
where the sum is over the components $C_i$ of $M^{S^1} \bigcap \mu^{-1}(c)$, $d$ is half the real codimension of $C_i$ in $M$, and the $a_j$'s are the weights of the $S^1$-representation on the normal bundle of $C_i$.
\end{theorem}
If $c$ is a critical value which is not an extremum, then the codimension of the fixed point set in $\mu^{-1}(c)$ is at least 4. Therefore
the theorem \ref{GLS} implies that $\mathrm{DH}(t)$ is continuous at non-extremal critical values and $\mathrm{DH}'(t)$ jumps at $c$ when $d$ equals 2. In the case
when $d=2$, the two nonzero weights must have opposite signs, so the jump in the derivative is negative, i.e. $\mathrm{DH}'(t)$ decreases when it passes through the critical value with $d=2$. Since $\mathrm{DH}$ is continuous, the jump in $\frac{d}{dt}\ln{\mathrm{DH}(t)} = \frac{\mathrm{DH}'(t)}{\mathrm{DH}(t)}$ is negative at $c$. Combining with the lemma \ref{logconcave}, we have the following corollary.
\begin{corollary}
Let $(M,\omega)$ be a closed Hamiltonian $S^1$-manifold with the moment map $\mu : M \rightarrow \R$. Then the corresponding Duistermaat-Heckman function $\mathrm{DH}$ is log-concave on $\mu(M)$ if $(\log{\mathrm{DH}(t)})'' \leq 0$ for every regular value $t \in \mu(M)$.
\end{corollary}
Note that $(\log{\mathrm{DH}(t)})'' = \frac{\mathrm{DH}(t) \cdot \mathrm{DH}''(t) - \mathrm{DH}'(t)^2}{\mathrm{DH}(t)^2}$. Therefore
$(\log{\mathrm{DH}(t)})'' \leq 0$ is equivalent to $\mathrm{DH}(t) \cdot \mathrm{DH}''(t) - \mathrm{DH}'(t)^2 \leq 0$.
The equation (\ref{Duist}) implies that
\begin{equation}\label{Duist2}
\mathrm{DH}(t) \cdot \mathrm{DH}''(t) = (n-1)(n-2)\int_{M_0}e^2[\omega_t]^{n-3} \cdot \int_{M_0}[\omega_t]^{n-1}
\end{equation}
\begin{equation}\label{Duist3}
\mathrm{DH}'(t)^2 = (n-1)^2\left(\int_{M_0}e[\omega_t]^{n-2}\right)^2
\end{equation}
To compute the integrals appeared in the equations (\ref{Duist2}) and (\ref{Duist3}), we need the following procedure :
For an inclusion $\iota : \mu^{-1}(0) \hookrightarrow M$, we have a ring homomorphism $\kappa : H^*_{S^1}(M;\R) \rightarrow H^*_{S^1}(\mu^{-1}(0);\R) \cong H^*(M_0;\R)$ which is called the Kirwan map. Due to the Kirwan surjectivity \cite{Ki}, $\kappa$ is a ring surjection.
Consider a 2-form $\widetilde{\omega} := \omega - d(\mu\theta)$ on $ES^1 \times M$ where $\theta$ is the connection form on $ES^1$. We denote by $x = \pi^*u \in H^2_{S^1}(M;\Z)$ where $\pi : M \times_{S^1} ES^1 \rightarrow BS^1$ and $u$ is a generator of $H^*(BS^1;\Z)$ such that
the Euler class of the Hopf bundle $ES^1 \rightarrow BS^1$ is $-u$.
Some part of the following two lemmas are given in \cite{Au}, but we give the complete proofs here.
\begin{lemma}\label{equivariantform}
$\widetilde{\omega}$ is $S^1$-invariant and closed, and $i_X \widetilde{\omega} = 0$ so that $\widetilde{\omega}$ represents a cohomology class in $H^*_{S^1}(M;\R)$. Moreover, for any fixed component $F \in M^{S^1}$, we have $\kappa([\widetilde{\omega}]) = [\omega_0]$ and $[\widetilde{\omega}]|_{F} = [\omega]|_F + \mu(F)u$. In particular, if $F$ is isolated, then $[\widetilde{\omega}]|_{F} = \mu(F)u$.
\end{lemma}
\begin{proof}
For the first statement, it is enough to show that $i_X \widetilde{\omega}$ and $L_X \widetilde{\omega}$ vanish. Note that $i_X \widetilde{\omega} =
i_X \omega - i_X d(\mu\theta) = -d\mu + di_X(\mu\theta) - L_X(\mu\theta)$ by Cartan's formula. Since $i_X(\mu\theta) = \mu$ and $\mu\theta$ is invariant under the circle action, we have $i_X \widetilde{\omega} = -d\mu + d\mu = 0$. Moreover, it is obvious that $\widetilde{\omega}$ is closed by definition. Hence $L_X \widetilde{\omega} = 0$ by Cartan's formula again.
To prove the second statement, consider the following diagram.
\begin{displaymath}
\begin{array}{ccc}
\mu^{-1}(0) \times ES^1 & \hookrightarrow & M \times ES^1\\
\downarrow & & \downarrow \\
\mu^{-1}(0) \times_{S^1} ES^1 & \hookrightarrow & M \times_{S^1} ES^1 \\
\downarrow_{\cong} & & \\
\mu^{-1}(0)/S^1 \cong M_{red} & & \\
\end{array}
\end{displaymath}
Since $d\mu$ is zero on the tangent bundle $\mu^{-1}(0) \times ES^1$, the pull-back of $\widetilde{\omega} = \omega - d\mu \wedge \theta - \mu d\theta$ to $\mu^{-1}(0) \times ES^1$ is the restriction $\omega|_{\mu^{-1}(0) \times ES^1}$. And the lift-down of $\omega|_{\mu^{-1}(0) \times ES^1}$ to $\mu^{-1}(0)/S^1$ is just a reduced symplectic form at the level $0$. Hence $\kappa([\widetilde{\omega}]) = [\omega_0]$.
To show the last statement, consider $[\widetilde{\omega}]|_{F} = [\omega - d(\mu\theta)]|_F = [\omega - d\mu\theta - \mu d\theta]|_F$. Since the restriction $d\mu|_{F \times ES^1}$ vanishes, we have $[\widetilde{\omega}]|_{F} = [\omega]|_{F} - \mu(F) \cdot [d\theta]|_{ES^1} = [\omega]|_{F} + \mu(F)u$. If $F$ is isolated, then we have $[\widetilde{\omega}]|_{F} = \mu(F)u$.
\end{proof}
\begin{lemma}\label{euler}
Consider a 2-form $d\theta$ on $ES^1 \times M$ where $\theta$ is a connection 1-form on $ES^1$. Then we can lift $d\theta$ down to $ES^1 \times_{S^1} M$ so that $d\theta$ represents a cohomology class in $H^*_{S^1}(M;\R)$. Moreover, $[d\theta] = -x$ and $\kappa([d\theta]) = -\kappa(x) = e$ where $e$ is the Euler class of the $S^1$-fibration $\mu^{-1}(0) \rightarrow M_{red}$.
\end{lemma}
\begin{proof}
Note that $i_X d\theta = L_X\theta - di_X\theta = 0$. Hence we can lift $d\theta$ down to $ES^1 \times_{S^1} M$. For any fixed point $x \in
M^{S^1}$, the restriction $[d\theta]|_x$ is the Euler class of $ES^1 \times x \rightarrow BS^1$. Hence $[d\theta] = -u \cdot 1 = -x$. Here, the multiplication $\cdot $ comes from the $H^*(BS^1)$-module structure on $H^*_{S^1}(M)$. By the diagram in the proof of the Lemma \ref{equivariantform}, $\kappa([d\theta])$ is just the Euler class of the $S^1$-fibration $\mu^{-1}(0) \rightarrow \mu^{-1}(0)/S^1$. Therefore $\kappa([d\theta]) = -\kappa(x) = e$.
\end{proof}
Combining the equations (\ref{Duist2}), (\ref{Duist3}), Lemma \ref{equivariantform}, and Lemma \ref{euler}, we have the following corollary.
\begin{corollary}\label{log2}
$\mathrm{DH}(0)\cdot \mathrm{DH}''(0) - \mathrm{DH}'(0)^2 \leq 0$ if and only if $$(n-2)\int_{M_0} \kappa([d\theta]^2 [\widetilde{\omega}]^{n-3}) \cdot \int_{M_0}\kappa([\widetilde{\omega}]^{n-1}) - (n-1)(\int_{M_0}\kappa([d\theta] [\widetilde{\omega}]^{n-2}))^2 \leq 0.$$
\end{corollary}
To compute the above integrals $\int_{M_0} \kappa([d\theta]^2 [\widetilde{\omega}]^{n-3}), \int_{M_0}\kappa([\widetilde{\omega}]^{n-1}),$ and $\int_{M_0}\kappa([d\theta] [\widetilde{\omega}]^{n-2})$, we need the residue formula due to Jeffrey and Kirwan. (See \cite{JK} and \cite{J}).
\begin{theorem}\cite{JK}\label{JK}
Let $\nu \in H^*_{S^1}(M;\R)$. Then $$\int_{M_0} \kappa(\nu) = \sum_{F \in M^{S^1}, ~\mu(F)>0} Res (\frac{\nu|_F}{e_F}).$$
Here $e_F$ is the equivariant Euler class of the normal bundle to F, and $Res(f)$ is a residue of $f$.
\end{theorem}
Now, let's compute $\int_{M_0} \kappa([d\theta]^2 [\widetilde{\omega}]^{n-3})$. By Theorem \ref{JK},
$$\int_{M_0} \kappa([d\theta]^2 [\widetilde{\omega}]^{n-3}) = \sum_{F \in M^{S^1}, ~\mu(F)>0} Res (\frac{[d\theta]^2 [\widetilde{\omega}]^{n-3}|_F}{e_F})$$
Since $[\widetilde{\omega}]|_z = \mu(z)u$ and $[d\theta]|_z = -u$ by lemma \ref{equivariantform} and \ref{euler}, we have
\begin{displaymath}
\begin{array}{cl}
\int_{M_0} \kappa([d\theta]^2 [\widetilde{\omega}]^{n-3}) & = \sum_{F \in M^{S^1}, ~\mu(F)>0} Res (\frac{\mu(F)^{n-3}u^{n-1}}{e_F})\\
& \\
& =\sum_{F \in M^{S^1}, ~\mu(F)>0} Res(\frac{\mu(F)^{n-3}u^{n-1}}{m_Fu^n})\\
& \\
& =\sum_{F \in M^{S^1}, ~\mu(F)>0} \frac{1}{m_F}\mu(F)^{n-3}\\
\end{array}
\end{displaymath}
where $m_F$ is the product of all weights of tangential $S^1$-representation at $F$.
Similarly, if $\xi \in \R$ is a regular value of $\mu$, then we put $\widetilde{\mu} = \mu - \xi$ be the new moment map. By the same argument,
we have the following lemma
\begin{lemma}\label{eq}
For a regular value $\xi$ of the moment map $\mu$,
\begin{enumerate}
\item
\begin{displaymath}
\int_{M_{\xi}} \kappa([d\theta]^2 [\widetilde{\omega}]^{n-3}) = \sum_{\substack{F \in M^{S^1},\\ \mu(F)>\xi}} \frac{1}{m_F}(\mu(F)-\xi)^{n-3}.
\end{displaymath}
\item
\begin{displaymath}
\int_{M_{\xi}} \kappa([d\theta] [\widetilde{\omega}]^{n-2}) = \sum_{\substack{F \in M^{S^1}, \\\mu(F)>\xi}} \frac{-1}{m_F}(\mu(F)-\xi)^{n-2}.
\end{displaymath}
\item
\begin{displaymath}
\int_{M_{\xi}} \kappa([\widetilde{\omega}]^{n-1}) = \sum_{\substack{F \in M^{S^1}, ~\\ \mu(F)>\xi}} \frac{1}{m_F}(\mu(F)-\xi)^{n-1}.
\end{displaymath}
\end{enumerate}
\end{lemma}
Combining the corollary \ref{log2} and the lemma \ref{eq}, we have the following proposition.
\begin{proposition}\label{det}
Let $(M,\omega)$ be a closed Hamiltonian $S^1$-manifold with the moment map $\mu : M \rightarrow \R$. Assume that $M^{S^1}$ consists of isolated fixed points. Then a density function of the Duistermaat-Heckman measure with respect to $\mu$ is log-concave if and only if
\begin{displaymath}
\sum_{\substack{F \in M^{S^1}, \\ \mu(F)>\xi}} \frac{1}{m_F}(\mu(F)-\xi)^{n-3} \cdot \sum_{\substack{F \in M^{S^1}, \\ \mu(F)>\xi}} \frac{1}{m_F}(\mu(F)-\xi)^{n-1}
\\ - \left(\sum_{\substack{F \in M^{S^1}, \\ \mu(F)>\xi}} \frac{1}{m_F}(\mu(F)-\xi)^{n-2}\right)^2 \leq 0
\end{displaymath}
for all regular value $\xi \in \mu(M)$, where $m_F$ is the product of all weights of the $S^1$-representation on $T_F M$. In particular, the log-concavity of the Duistermaat-Heckman measure is completely determined by
the set $\{(\mu(F), m_F)_F | F \in M^{S^1} \}$.
\end{proposition}
\begin{corollary}\label{det2}
Let $(M^{2n},\omega)$ and $(N^{2n}, \sigma)$ be two closed Hamiltonian $S^1$-manifold with the moment map $\mu_1$ and $\mu_2$ respectively.
Assume there exist a bijection $\phi : M^{S^1} \rightarrow N^{S^1}$ which satisfies
\begin{enumerate}
\item for each $F \in M^{S^1}$, $m_F = m_{\phi(x)}$, and
\item for each $F \in M^{S^1}$, $\mu_1(F) = \mu_2(\phi(F))$.
\end{enumerate}
where $m_F$ is the product of all weights of the tangential $S^1$-representation at $F$.
If $N$ satisfies the log-concavity of the Duistermaat-Heckman measure with respect to $\mu_2$, then so does $M$ with respect to $\mu_1$.
\end{corollary}
\begin{remark}
The integration formulae (1) and (3) in the lemma \ref{eq} are proved by Wu by using the stationary phase method. See Theorem 5.2
in \cite{Wu} for the detail.
\end{remark}
\section{proof of the Theorem \ref{first}}
As noted in the introduction, if a Hamiltonian $S^1$-action on the K\"{a}hler manifold is holomorphic, then the corresponding Duistermaat-Heckman function is log-concave by \cite{Gr}. Let $(M^{2n},\omega)$ be a closed semifree Hamiltonian circle action with the moment map $\mu$. Assume that all fixed points are isolated. Let $\mathrm{DH}$ be the corresponding Duistermaat-Heckman function with respect to $\mu$. We will show that there is a K\"{a}hler form $\omega_1 \oplus \cdots \oplus \omega_n$ on $S^2 \times \cdots \times S^2$ with the standard diagonal holomorphic semifree circle action such that a bijection $\psi^{-1} \circ \phi : M^{S^1} \rightarrow (S^2 \times \cdots \times S^2)^{S^1} $ given in Section 2 satisfies the conditions in the corollary \ref{det2}, which implies the log-concavity of $\mathrm{DH}$. Now we start with the lemma below.
\begin{lemma}\label{determine}
Let $(M^{2n},\omega)$ be a closed semifree Hamiltonian circle action with the moment map $\mu$. Assume that all fixed points are isolated. Then $\{(\mu(F), m_F)_F | F \in M^{S^1} \}$ is completely determined by $\mu(p_0^1), \mu(p_1^1), \cdots, \mu(p_1^n)$, where $p_k^j$'s are the fixed points of index $2k$ for $j=1,\cdots,{n \choose k}$.
\end{lemma}
\begin{proof}
Consider an equivariant symplectic 2-form $\widetilde{\omega}$ on $ES^1 \times_{S^1} M$ which is given in Section 3. Because
$x, a_1, \cdots, a_n$ form a basis of $H^2_{S^1}(M;\Z)$, we may let $[\widetilde{\omega}] = m_0x + m_1a_1 + \cdots + m_na_n$ for some real numbers $m_i$'s. (See Section 2 : the definition of $x, a_1, \cdots, a_n$.) By lemma \ref{equivariantform}, we have $[\widetilde{\omega}]|_{p_0^1} = \mu(p_0^1)u$. On the other hand, the right hand side is $(m_0x + m_1a_1 + \cdots + m_na_n)|_{p_0^1} = m_0u$, since every $a_i$ vanishes on $p_0^1$. Hence $m_0 = \mu(p_0^1)$. Similarly, $[\widetilde{\omega}]|_{p_1^i} = \mu(p_1^i)u$ and $(m_0x + m_1a_1 + \cdots + m_na_n)|_{p_1^i} = m_0u + m_iu$. Hence we have $m_i = \mu(p_1^i) - m_0 = \mu(p_1^i) - \mu(p_0^1)$ for each $i=1 , \cdots, n$. Therefore $\mu(p_0^1), \mu(p_1^1), \cdots, \mu(p_1^n)$ determine the coefficients $m_i$'s of $[\widetilde{\omega}]$.
For $p_k^j$, $[\widetilde{\omega}]|_{p_k^j} = \mu(p_k^j)u$ and $(m_0x + m_1a_1 + \cdots + m_na_n)|_{p_k^j} = m_0u + \sum_{i \in \phi(p_k^j)} m_iu$. Hence for fixed $k$, the set $\{(\mu(p_k^j),m_{p_k^j})_j | j=1,\cdots, {n \choose k} \}$ is just a $\{(m_0 + m_{i_1} + \cdots, + m_{i_k},(-1)^k)_{\{i_1, \cdots, i_k\}} | \{i_1, \cdots, i_k\} \subset \{1,2,\cdots,n\}\}$. Hence $\{(\mu(F),m_F)_F | F \in M^{S^1}\} = \cup_{k=0}^{k=n} \{(\mu(p_k^j),m_{p_k^j})_j | j=1,\cdots, {n \choose k} \}$ is completely determined by $\mu(p_0^1)$, $\mu(p_1^1)$, $\cdots, \mu(p_1^n)$.
\end{proof}
Now we are ready to prove the theorem \ref{first}.
\begin{proof}[\textit{Proof of Theorem \ref{first}}]
For $\psi : (S^2 \times \cdots \times S^2)^{S^1} \rightarrow 2^{[n]}$ defined in Section 2, Let $\omega_i$ be the Fubini-Study form on $S^2$ such that the symplectic volume is $\mu(p_1^i) - \mu(p_0^1)$. Let $\mu' : S^2 \times \cdots \times S^2 \rightarrow \R$ be the moment map whose minimum is
$\mu(p_0^1)$. Then we have $\{(\mu(F), m_F)_F | F \in M^{S^1} \} = \{(\mu(F), m_F)_F | F \in (S^2 \times \cdots \times S^2)^{S^1} \}$ and $\psi^{-1} \circ \phi : M^{S^1} \rightarrow (S^2 \times \cdots \times S^2)^{S^1}$ satisfies the condition in the corollary \ref{det2}.
Therefore the Duistermaat-Heckman measure is log-concave on $\mu(M)$.
\end{proof}
\section{The hard Lefschetz property of the reduced symplectic forms}
For a K\"{a}hler manifold $(N,\sigma)$ with a holomorphic circle action preserving $\sigma$, its symplectic reduction is again K\"{a}hler, and hence
the reduced symplectic form $\sigma_t$ is K\"{a}hler. Therefore $\sigma_t$ satisfies the Hard Lefschetz property for every regular value $t$.
In this section, we show that the same thing happens when $(M,\omega)$ is a closed semifree Hamiltonian $S^1$-space whose fixed points are all isolated. The following theorem is due to Tolman and Weitsman.
\begin{theorem}\cite{TW2}\label{Kir}
Let $S^1$ act on a compact symplectic manifold $(M,\omega)$ with moment map $\mu : M \rightarrow \R$. Assume that all fixed points are isolated and $0$ is a regular value. Let $M^{S^1}$ denote the set of fixed points.
Define $K_+ := \{\alpha \in H^*_{S^1}(M;\Z) | \alpha_{F_+} = 0\}$ where $F_+ := M^{S^1} \bigcap \mu^{-1}(0,\infty)$ and $K_- := \{\alpha \in H^*_{S^1}(M;\Z) | \alpha_{F_-} = 0\}$ where $F_- := M^{S^1} \bigcap \mu^{-1}(-\infty,0)$.
Then there is a short exact sequence:
$$ 0 \longrightarrow K \longrightarrow H^*_{S^1}(M;\Z) \stackrel{\kappa}\longrightarrow (M_{red};\Z) \longrightarrow 0$$ where $\kappa : H^*_{S^1}(M;\Z) \rightarrow H^*(M_{red};\Z)$ is the Kirwan map.
\end{theorem}
\begin{proof}[\textit{Proof of Theorem \ref {second}}]
Let $\kappa_M : H^*_{S^1}(M;\R) \rightarrow H^*(M_{red};\R)$ be the Kirwan map for $(M,\omega)$ and let $\kappa$ be the one for $(S^2 \times \cdots \times S^2, \sigma)$, where $\sigma := \omega_1 \oplus \cdots \oplus \omega_n$ is chosen in the proof of Theorem \ref{first} in Section 4.
As in the proof of Theorem \ref{first}, we proved that there exists a semifree holomorphic Hamiltonian $S^1$-manifold $(S^2 \times \cdots \times S^2, \sigma)$ with the moment map $\mu'$ such that $\psi^{-1} \circ \phi : M^{S^1} \rightarrow (S^2 \times \cdots \times S^2)^{S^1}$ preserves their indices, weights, and the values of the moment map. Hence $\psi^{-1} \circ \phi$ identifies $K^M_+$ with $K^{S^2 \times \cdots \times S^2}_+$ and $K^M_-$ with $K^{S^2 \times \cdots \times S^2}_-$. The ring isomorphism $f : H^*_{S^1}(M;\Z) \rightarrow H^*_{S^1}(S^2 \times \cdots \times S^2;\Z)$ given in Corollary \ref{iso} satisfies $\alpha|_x = f(\alpha)|_{\psi^{-1} \circ \phi(x)}$ for any $\alpha \in H^*_{S^1}(M;\Z)$ and any fixed point $x \in M^{S^1}$.
Hence if $\alpha \in K^M_+$, then $f(\alpha) \in K^{S^2 \times \cdots \times S^2}_+$. Similarly for any $\alpha \in K^M_-$, we have $f(\alpha) \in K^{S^2 \times \cdots \times S^2}_-$. Therefore $f$ preserves the kernel of the Kirwan map $\kappa$ by Theorem \ref{Kir}.
Note that $\kappa(f([\widetilde{\omega}]))$ is the reduced symplectic class of $S^2 \times \cdots \times S^2$ at 0. Since the K\"{a}hler quotient of the holomorphic action is again K\"{a}hler, $\kappa(f([\widetilde{\omega}]))$ satisfies the hard Lefschetz property. Now, assume that $\omega_0$ does not satisfy the hard Lefschetz property. Then there exists a positive integer $k(<n)$ and some nonzero $\alpha \in H^k(M_{red};\R)$ such that $\alpha \cdot [\omega_0]^{n-k} = 0$ in $H^{2n-k}(M_{red})$. By the Kirwan surjectivity, we can find $\widetilde{\alpha} \in H^*_{S^1}(M;\R)$
with $\kappa(\widetilde{\alpha}) = \alpha$. Then $\widetilde{\alpha} \cdot [\widetilde{\omega}]^{n-k}$ is in $\ker{\kappa}$ and hence the image
$f(\widetilde{\alpha} \cdot [\widetilde{\omega}]^{n-k})$ is in $\ker{\kappa}$, since $f$ maps $\ker{\kappa}$ of $M$ to $\ker{\kappa}$ of $S^2 \times \cdots \times S^2$. It implies that $f(\widetilde{\alpha}) = 0$ by the hard Lefschetz condition for $f([\widetilde{\omega}])$. It contradicts that $f$ is an isomorphism.
It remains to show that $(M,\omega)$ satisfies the Hard Lefschetz property. Remind that $\psi^{-1} \circ \phi : M^{S^1} \rightarrow (S^2 \times \cdots \times S^2)^{S^1}$ induces an isomorphism $$f : H^*_{S^1}(M;\R) \rightarrow H^*_{S^1}(S^2 \times \cdots \times S^2;\R),$$ which sends the equivariant symplectic class $[\widetilde{\omega}]$ to $[\widetilde{\sigma}]$ as we have seen in Section 4. Here, $\widetilde{\sigma}$ is an equivariant symplectic form induced by $\sigma - d(\mu'\theta)$ in $H^*_{S^1}(S^2 \times \cdots \times S^2;\R)$.
Since $f$ is a $H^*(BS^1;\R)$-isomorphism, it induces a ring isomorphism
\begin{displaymath}
f_u : \frac{H^*_{S^1}(M;\R)}{u \cdot H^*_{S^1}(M;\R)} \rightarrow \frac{H^*_{S^1}(S^2 \times \cdots \times S^2;\R)}{u \cdot H^*_{S^1}(S^2 \times \cdots \times S^2;\R)}
\end{displaymath}
Moreover, the quotient map $\pi_M : H^*_{S^1}(M;\R) \rightarrow \frac{H^*_{S^1}(M;\R)}{u \cdot H^*_{S^1}(M;\R)} \cong H^*(M;\R)$ ($\pi_{S^2 \times \cdots \times S^2}$ respectively) is just a ring homomorphism which comes from an inclusion $M \hookrightarrow M \times_{S^1} ES^1$ as a fiber. Therefore $\pi_M([\widetilde{\omega}]) = [\omega]$ and $\pi_{S^2 \times \cdots \times S^2}([\widetilde{\sigma}]) = [\sigma]$. It means that
$f_u : H^*(M;\R) \rightarrow H^*(S^2 \times \cdots \times S^2 ;\R)$ is a ring isomorphism which sends $[\omega]$ to $[\sigma]$. Since $\sigma$ is a K\"{a}hler form, it satisfies the Hard Lefschetz property. Hence so $\omega$ does.
\end{proof}
\bigskip
\bibliographystyle{amsalpha}
|
\subsection{What broke the symmetry?}
\label{quarks}
The classical QCD Lagrangean has a symmetry under a rotation of the
underlying quark fields
\begin{equation}
\matrix{
\psi \rightarrow e^{i\phi\gamma_5/2}\psi\cr
\overline\psi \rightarrow \overline\psi e^{i\phi\gamma_5/2}.\cr
}
\end{equation}
This corresponds directly to the transformation of the composite
fields given in Eq.~\ref{rotate}. This symmetry is ``anomalous'' in
that any regulator must break it with a remnant surviving in the
as the regulator is removed \cite{Adler:1969gk,Adler:1969er,Bell:1969ts}.
With the lattice this concerns the continuum limit.
The specifics of how the anomaly works depend on the details of the
regulator. Here we will follow Fujikawa \cite{Fujikawa:1979ay} and
consider the fermionic measure in the path integral. If we make the
above rotation on the field $\psi$, the measure changes by the
determinant of the rotation matrix
\begin{equation}
\label{fujikawa}
d\psi\rightarrow |e^{-i\phi\gamma_5/2}| d\psi
=e^{-i\phi {\rm Tr}\gamma_5/2} d\psi.
\end{equation}
Here is where the subtlety of the regulator comes in. Naively
$\gamma_5$ is a simple four by four traceless matrix. If it is indeed
traceless, then the measure would be invariant. However, in the
regulated theory this is not the case. This is intimately tied with
the index theorem for the Dirac operator in topologically non-trivial
gauge fields.
A typical Dirac action takes the form $\overline\psi (D+m)\psi$ with
the kinetic term $D$ a function of the gauge fields. In the naive
continuum theory $D$ is anti-Hermitean, $D^\dagger=-D$, and
anti-commutes with $\gamma_5$, i.e. $[D,\gamma_5]_+=0$. What
complicates the issue with fermions is the index theorem discussed
earlier and reviewed in Ref.~\cite{Coleman:1978ae}. If a background
gauge field has winding $\nu$, then there must be at least $\nu$ exact
zero eigenvalues of $D$. Furthermore, on the space spanned by the
corresponding eigenvectors, $\gamma_5$ can be simultaneously
diagonalized with $D$. The net winding number equals the number of
positive eigenvalues of $\gamma_5$ minus the number of negative
eigenvalues. In this subspace the trace of $\gamma_5$ does not
vanish, but equals $\nu$.
What about the higher eigenvalues of $D$? We discussed these earlier
when we formulated the index theorem. Because $[D,\gamma_5]_+=0$,
non-vanishing eigenvalues appear in opposite sign pairs; i.e. if
$D|\psi\rangle =\lambda|\psi\rangle$ then $D\gamma_5|\psi\rangle
=-\lambda\gamma_5|\psi\rangle$. For an anti-Hermitean $D$, these
modes are orthogonal with $\langle\psi|\gamma_5|\psi\rangle=0$. As a
consequence, $\gamma_5$ is traceless on the subspace spanned by each
pair of eigenvectors.
So what happened to the opposite chirality states to the zero modes?
In a regulated theory they are in some sense ``above the cutoff.'' In
a simple continuum discussion they have been ``lost at infinity.''
With a lattice regulator there is no ``infinity''; so, something more
subtle must happen. With the
overlap\cite{Ginsparg:1981bj,Neuberger:1997fp} or
Wilson\cite{Wilson:1975id} fermions, discussed in more detail later,
one gives up the anti-Hermitean character of $D$. Most eigenvalues
still occur in conjugate pairs and do not contribute to the trace of
$\gamma_5$. However, in addition to the small real eigenvalues
representing the zero modes, there are additional modes where the
eigenvalues are also real but large. With Wilson fermions these
appear as massive doubler states. With the overlap, the eigenvalues
are constrained to lie on a circle. In this case, for every exact
zero mode there is another mode with the opposite chirality lying on
the opposite side of the circle. These modes are effectively massive
and break chiral symmetry. The necessary involvement of both small
and large eigenvalues warns of the implicit danger in attempts to
separate infrared from ultraviolet effects. When the anomaly is
concerned, going to short distances is not sufficient for ignoring
non-perturbative effects related to topology.
So with the regulator in place, the trace of $\gamma_5$ does not
vanish on gauge configurations of non-trivial topology. The change of
variables indicated in Eq.~\ref{fujikawa} introduces into the path
integral a modification of the weighting by a factor
\begin{equation}
\label{nu}
e^{-i\phi {\rm Tr}\gamma_5}=e^{-i\phi N_f\nu}.
\end{equation}
Here we have applied the rotation to all flavors equally, thus the
factor of $N_f$ in the exponent. The conclusion is that gauge
configurations that have non-trivial topology receive a complex weight
after the anomalous rotation.
Although not the topic of discussion here, note that this factor
introduces a sign problem if one wishes to study this physics via
Monte Carlo simulations. Here we treat all $N_f$ flavors
equivalently; this corresponds to dividing the conventionally defined
CP violation angle, to be discussed later, equally among the flavors,
i.e. effectively $\phi=\Theta/N_f$.
\subsection{A discrete chiral symmetry}
\label{znf}
We now return to the effective Lagrangean language of before. For
the massless theory, the symmetry under $\sigma\leftrightarrow
-\sigma$ indicates that we expect at least two minima for the
effective potential considered in the $\sigma,\eta^\prime$ plane.
These are located as sketched in Fig.~\ref{potential0}. Do we know
anything about the potential elsewhere in this plane? The answer is
yes, indeed there are actually $N_f$ equivalent minima.
\begin{figure}
\centering
\includegraphics[width=2.5in]{potential0.eps}
\caption{\label{potential0} We have two minima in the
$\sigma,\eta^\prime$ plane located at $\sigma=\pm v$ and
$\eta^\prime=0$. The circles represent that the fields will
fluctuate in a small region about these minima. Can we find any
other minima? }
\end{figure}
It is convenient to separate the left- and right-hand parts of the
fermion field
\begin{eqnarray}
&&\psi_{L,R}={1\over 2} (1\pm\gamma_5) \psi\cr
&&\overline\psi_{L,R}=\overline\psi {1\over 2} (1\mp\gamma_5).
\end{eqnarray}
The mass term is thus
\begin{equation}
m\overline\psi \psi=m(\overline\psi_L\psi_R
+\overline\psi_R \psi_L)
\end{equation}
and mixes the left and right components.
Using this notation, due to the anomaly the singlet rotation
\begin{equation}
\psi_L\rightarrow e^{i\phi}\psi_L
\label{singlet}
\end{equation}
is not a valid symmetry of the theory for generic values of the angle
$\phi$. On the other hand, flavored chiral symmetries should
survive, and in particular
\begin{equation}
\psi_L\rightarrow
g_L\psi_L = e^{i\phi_\alpha \lambda_\alpha}\psi_L
\label{flavored}
\end{equation}
is expected to be a valid symmetry for any set of angles
$\phi_\alpha$. The point of this subsection is that, for special
special discrete values of the angles, the rotations in
Eq.~\ref{singlet} and Eq.~\ref{flavored} can coincide. At such values
the singlet rotation becomes a valid symmetry. In particular, note
that
\begin{equation}
g=e^{2\pi i\phi/N_f} \in Z_{N_f} \subset SU(N_f).
\end{equation}
Thus a valid discrete symmetry involving only $\sigma$ and
$\eta^\prime$ is
\begin{equation}
\matrix{
\sigma\rightarrow \phantom{+} \sigma\cos(2\pi/N_f)
+\eta^\prime \sin(2\pi/N_f)\cr
\eta^\prime\rightarrow -\sigma \sin(2\pi/N_f)+\eta^\prime
\cos(2\pi/N_f).\cr
}
\end{equation}
The potential $V(\sigma,\eta^\prime)$ has a $Z_{N_f}$ symmetry
manifested in $N_f$ equivalent minima in the $(\sigma,\eta^\prime)$
plane. For four flavors this structure is sketched in
Fig.~\ref{potential1}.
\begin{figure}
\centering
\includegraphics[width=2.5in]{potential1.eps}
\caption{\label{potential1}
For four massless flavors we have four equivalent
minima in the $\sigma,\eta^\prime$ plane. This generalizes to $N_f$
minima with $N_f$ flavors.
}
\end{figure}
This discrete flavor singlet symmetry arises from the trivial fact
that $Z_N$ is a subgroup of both $SU(N)$ and $U(1)$. At the quark
level the symmetry is easily understood since the quark measure
receives an additional phase proportional to the winding number from
every flavor. With $N_F$ flavors, these multiply together making
\begin{equation}
\psi_L\rightarrow e^{2\pi i/N_f} \psi_L
\end{equation}
a valid symmetry even though rotations by smaller angles are not.
The role of the $Z_N$ center of $SU(N)$ is illustrated graphically in
Fig.~\ref{su3circle}, taken from Ref.~\cite{Creutz:1995wf}. Here we
plot the real and the imaginary parts of the traces of 10,000 $SU(3)$
matrices drawn randomly with the invariant group measure. The region
of support only touches the $U(1)$ circle at the elements of the
center. All elements lie on or within the curve mapped out by
elements of form $\exp(i\phi\lambda_8)$. Figure \ref{su4circle} is a
similar plot for the group $SU(4)$.
\begin{figure}
\centering
\includegraphics[width=2.5in]{su3circle.eps}
\caption{\label{su3circle} The real and imaginary parts for the traces
of 10,000 randomly chosen $SU(3)$ matrices. All points lie within the
boundary representing matrices of the form $\exp(i\phi\lambda_8)$.
The tips of the three points represent the center of the group. The
outer curve represents the boundary that would be found if the group
was the full $U(1)$. Taken from Ref.~\cite{Creutz:1995wf}. }
\end{figure}
\begin{figure}
\centering
\includegraphics[width=2.5in]{su4circle.eps}
\caption{\label{su4circle} The generalization of Fig.~\ref{su3circle}
to $SU(4)$. The real and imaginary parts for the traces of 10,000
randomly chosen $SU(4)$ matrices. Taken from
Ref.~\cite{Creutz:1995wf}. }
\end{figure}
\subsection{Coupling constant renormalization}
At the level of tree Feynman diagrams, relativistic quantum field
theory is well defined and requires no renormalization. However as
soon as loop corrections are encountered, divergences appear and must
be removed by a regularization scheme. In general the theory then
depends on some cutoff, which is to be removed with a simultaneous
adjustment of the bare parameters while keeping physical quantities
finite.
For example, consider a lattice cutoff with spacing $a$. The proton
mass $m_p$ is a finite physical quantity, and on the lattice it will
be some, a priori unknown, function of the cutoff $a$, the bare gauge
coupling $g$ and the bare quark masses. For the quark-less theory we
could use the lightest glueball mass for this purpose. The basic idea
is to hold enough physical properties constant to determine how the
coupling and quark masses behave as the lattice spacing is reduced.
As the quark masses go to zero the proton mass is expected to remain
finite; thus, to simplify the discussion, temporarily ignore the quark
masses. Thus consider the proton mass as a function of the gauge
coupling and the cutoff, $m_p(g,a)$. Holding this constant as the
cutoff varies determines how $g$ depends on $a$. This is the basic
renormalization group equation
\begin{equation}
a{d\over da} m_p(g(a),a)=0=a{\partial\over \partial a}m_p(g,a)
+ a\left({ dg\over da}\right)\ {\partial\over\partial g} m_p(g,a).
\end{equation}
By dimensional analysis, the proton mass should scale as $a^{-1}$
at fixed bare coupling. Thus we know that
\begin{equation}
a{\partial\over \partial a}m_p(g,a)=-m_p(g,a).
\end{equation}
The ``renormalization group function''
\begin{equation}
\beta(g)=a{dg\over da}={m_p(g,a)
\over {\partial\over\partial g} m_p(g,a)}
\end{equation}
characterizes how the bare coupling is to be varied for the continuum
limit. Note that this particular definition is independent of
perturbation theory or any gauge fixing.
As renormalization is not needed until quantum loops are encountered,
$\beta(g)$ vanishes as $g^3$ when the coupling goes to zero.
Define perturbative coefficients from the asymptotic series
\begin{equation}
\beta(g)=\beta_0 g^3+\beta_1 g^5 +\ldots
\label{AF}
\end{equation}
Politzer \cite{Politzer:1973fx} and Gross and Wilczek \cite
{Gross:1973id,Gross:1973ju} first calculated the coefficient $\beta_0$
for non-Abelian gauge theories, with the result
\begin{equation}
\label{betazero}
\beta_0={1\over 16\pi^2}(11N/3-2N_f/3)
\end{equation}
where the gauge group is $SU(N)$ and $N_f$ denotes the number of
fermionic species. As long as $N_f < 11 N/2$ this coefficient is
positive. Assuming we can reach a region where this first term
dominates, decreasing the cutoff corresponds to decreasing the
coupling. This is the heart of asymptotic freedom, which tells us
that the continuum limit of vanishing lattice spacing requires taking
a limit towards vanishing coupling. The two loop contribution to
Eq.~(\ref{AF}) is also known \cite{Caswell:1974gg,Jones:1974mm}
\begin{equation}
\label{betaone}
\beta_1=\left({1\over 16\pi^2}\right)^2
\left(34N^2/3-10NN_f/3-N_f(N^2-1)/N\right).
\end{equation}
In general the function $\beta(g)$ depends on the regularization
scheme in use. For example it might depend on what physical property
is held fixed as well as details of how the cutoff is imposed.
Remarkably, however, these first two coefficients are universal.
Consider two different schemes each defining a bare coupling as a
function of the cutoff, say $g(a)$ and $g^\prime(a)$. The expansion
for one in terms of the other will involve all odd powers of the
coupling. In the weak coupling limit each formulation should reduce
to the classical Yang-Mills theory, and thus to lowest order they
should agree
\begin{equation}
g^\prime=g+cg^3+O(g^5).
\label{gprime}
\end{equation}
We can now calculate the new renormalization group function
\begin{eqnarray}
\beta^\prime(g^\prime)&=&a{dg^\prime\over da}={\partial
g^\prime\over\partial g} \beta(g)\cr
&=&(1+3cg^3)(\beta_0 g^2+\beta_1 g^3)+O(g^5)\cr
&=&\beta_0 {g^\prime}^3+\beta_1 {g^\prime}^3+O({g^\prime}^5).
\label{universal}
\end{eqnarray}
Through order ${g^\prime}^3$ the dependence on the parameter $c$
cancels. This, however, does not continue to higher orders, where
alternate definitions of the beta function generally differ. We will
later comment further on this non-uniqueness.
The renormalization group function determines how rapidly the
coupling decreases with cutoff. Separating variables
\begin{equation}
d(\log(a))=
{dg\over \beta_0 g^3+\beta_1 g^5+O(g^7)}
\end{equation}
allows integration to obtain
\begin{equation}
\log(a\Lambda)=-{1\over 2 \beta_0 g^2}+
{\beta_1\over \beta_0^2}\log(g)+O(g^2)
\end{equation}
where $\Lambda$ is an integration constant.
This immediately shows that the lattice spacing decreases
exponentially in the inverse coupling
\begin{equation}
\label{rgfinal}
a={1\over \Lambda} e^{-1/2\beta_0 g^2} g^{-\beta_1/\beta_0^2}
(1+O(g^2)).
\end{equation}
Remarkably, although the discussion began with the beta function
obtained in perturbation theory, the right hand side of
Eq.~(\ref{rgfinal}) has an essential singularity at vanishing
coupling. The renormalization group provides non-perturbative
information from a perturbative result.
Dropping the logarithmic corrections, the coupling as a function of
the cutoff reduces to
\begin{equation}
g^2\sim {1\over 2\beta_0 \log(1/\Lambda a)}
\end{equation}
showing the asymptotic freedom result that the bare coupling goes to
zero logarithmically with the lattice spacing in the continuum limit.
The integration constant $\Lambda$ is defined from the bare charge and
in a particular cutoff scheme. Its precise numerical value will
depend on details, but once the scheme is chosen, it is fixed relative
to the scale of the quantity used define the physical scale. In the
above discussion this was the proton mass. The existence of a scheme
dependence can be seen by considering two different bare couplings as
related in Eq.~(\ref{gprime}). The relation between the integration
constants is
\begin{equation}
\log(\Lambda^\prime/\Lambda)={c\over 2\beta_0}.
\end{equation}
The mass $m$ of a physical particle, perhaps the proton used above, is
connected to an inverse correlation length in the statistical analogue
of the theory. Measuring this correlation length in lattice units, we
can consider the dimensionless combination $\xi=1/am$. For the
continuum limit, we want this correlation length to diverge.
Multiplying Eq.~(\ref{rgfinal}) by the mass tells us how this
divergence depends on the lattice coupling
\begin{equation}
ma=\xi^{-1}
={m\over \Lambda} e^{-1/2\beta_0 g^2} g^{-\beta_1/\beta_0^2}
(1+O(g^2)).
\label{correlation}
\end{equation}
Conversely, if we know how a correlation length $\xi$ of the
statistical system diverges as the coupling goes to zero, we can read
off the particle mass in units of $\Lambda$ as the coefficient of the
behavior in this equation. This exemplifies the close connection
between diverging correlation lengths in a statistical system and the
continuum limit of the corresponding quantum field theory.
We emphasize again the exponential dependence on the inverse coupling
appearing in Eq.~(\ref{correlation}). This is a function that is
highly non-analytic at the origin. This demonstrates quite
dramatically that QCD cannot be fully described by perturbation theory.
\subsection{A parameter free theory}
This discussion brings us to the remarkable conclusion that, ignoring
the quark masses, the strong interactions have no free parameters.
The cutoff is absorbed into $g(a)$, which in turn is absorbed into
the renormalization group dependence. The only remaining dimensional
parameter $\Lambda$ serves to set the scale for all other masses. In
the theory considered in isolation, one may select units such that
$\Lambda$ is unity. After such a choice, all physical mass ratios are
determined. Coleman and Weinberg \cite{Coleman:1973jx} have given
this process, wherein a dimensionless parameter $g$ and a
dimensionful one $a$ manage to ``eat'' each other, the marvelous name
``dimensional transmutation.''
In the theory including quarks, their masses represent further
parameters. Indeed, these are the only parameters in the theory of
the strong interactions. In the limit where the quark masses vanish,
referred to as the chiral limit, we return to a zero parameter theory.
In this approximation to the physical world, the pion mass is expected
to vanish and all dimensionless observables should be uniquely
determined. This applies not only to mass ratios, such as of the rho
mass to the proton, but as well to quantities such as the pion-nucleon
coupling constant, once regarded as a parameter for a perturbative
expansion. As the chiral approximation has been rather successful in
the predictions of current algebra, we expect an expansion in the
small quark masses to be a fairly accurate description of hadronic
physics. Given a qualitative agreement, a fine tuning of the small
quark masses should give the pion its mass and complete the theory.
The exciting idea of a parameter-free theory is sadly lacking from
most treatments of the other interactions such as electromagnetism or
the weak force. There the coupling $\alpha\sim 1/137$ is treated as a
parameter. One might optimistically hope that the inclusion of the
appropriate non-perturbative ideas into a unified scheme would
ultimately render $\alpha$ and the quark and lepton masses calculable.
\subsection{Including quark masses}
Above we concentrated on the flow of the bare coupling as one takes
the continuum limit. Of course with massive quarks in the theory, the
bare quark mass is also renormalized. Here we extend the above
discussion to see how the two bare parameters flow together to zero in
a well defined way.
Including the mass flow, the renormalization group equations become
\begin{eqnarray}
&&a{dg\over da}=\beta(g)=\beta_0 g^3+\beta_1 g^5 +\ldots
+{\rm non{\hbox{-}}perturbative}\cr
&&a{dm\over da}=m\gamma(g)=m(\gamma_0 g^2+\gamma_1 g^4 +\ldots)
+{\rm non{\hbox{-}}perturbative.}
\end{eqnarray}
Now we have three perturbative coefficients
$\beta_0,\ \beta_1,\ \gamma_0$ which are scheme independent and known
\cite{Politzer:1973fx,Gross:1973id,Gross:1973ju,
Caswell:1974gg,Jones:1974mm, Vermaseren:1997fq,Chetyrkin:1997fm }.
For $SU(3)$ we have
\begin{equation}
\matrix{
&\beta_0=&{11-2N_f/3 \over (4\pi)^2}
&\qquad=.0654365977\qquad (N_f=1)\hfill\cr
&\beta_1=&{102-12N_f\over (4 \pi)^4}
&\qquad = .0036091343\qquad (N_f=1)\hfill\cr
&\gamma_0=&{8\over (4 \pi)^2}&\qquad=.0506605918\hfill\cr
}
\end{equation}
For simplicity we work with $N_f$ degenerate quarks, although this is
easily generalized to the non-degenerate case. It is important to
recognize that the ``non-perturbative'' parts fall faster than any
power of $g$ as $g\rightarrow 0$. As we will discuss later, unlike
the perturbative pieces, the non-perturbative contributions to
$\gamma$ in general need not be proportional to the quark mass.
As with the pure gauge theory discussed earlier, these equations are
easily solved to show
\begin{eqnarray}
&&a={1\over \Lambda} e^{-1/2\beta_0 g^2} g^{-\beta_1/\beta_0^2}
(1+O(g^2))\cr
&&m=Mg^{\gamma_0/\beta_0}
(1+O(g^2)).
\end{eqnarray}
The quantities $\Lambda$ and $M$ are ``integration constants'' for the
renormalization group equations. Rewriting these relations gives the
coupling and mass flow in the continuum limit $a\rightarrow 0$
\begin{eqnarray}
&&g^2\sim {1\over \log(1/\Lambda a)}\rightarrow 0\qquad\qquad
\hbox{``asymptotic freedom''}\cr
&&m\sim M\ \left({1\over \log(1/\Lambda
a)}\right)^{\gamma_0/\beta_0}\rightarrow 0.
\end{eqnarray}
Here $\Lambda$ is usually regarded as the ``QCD scale'' and $M$ as the
``renormalized quark mass.'' The resulting flow is sketched in
Fig.~\ref{rgflow}.
\begin{figure*}
\centering
\includegraphics[width=2.5in]{rgflow.eps}
\caption{ In the continuum limit both the bare coupling and bare mass
for QCD flow to zero.}
\label{rgflow}
\end{figure*}
The rate of this flow to the origin is tied to the renormalization
group constants, which can be obtained from the inverted equations
\begin{eqnarray}
&&\Lambda=\lim_{a\rightarrow 0}
\ { e^{-1/2\beta_0 g^2} g^{-\beta_1/\beta_0^2}\over a}\\
&&M=\lim_{a\rightarrow 0}\ m g^{-\gamma_0/\beta_0}.
\end{eqnarray}
Of course, as discussed for $\Lambda$ above, the specific numerical
values of these parameters depend on the detailed renormalization
scheme.
Defining $\beta(g)$ and $\gamma(g)$ is most naturally done by fixing some
physical quantities and adjusting the bare parameters as the cutoff is
removed. Because of confinement we can't use the quark mass itself, but we
can select several physical particle masses $m_i(g,m,a)$ to hold
fixed. This leads to the constraint
\begin{equation}
a{dm_i(g,m,a)\over da}=0={\partial m_i \over \partial g}\beta(g)
+{\partial m_i \over \partial m} m\gamma(g)+a{\partial m_i\over
\partial a}.
\end{equation}
For simplicity, continue to work with degenerate quarks of mass $m$.
Then we have two bare parameters $(g,m)$, and we need to fix two
quantities.
\footnote{Actually there is a third parameter related to CP
conservation. Here we assume CP is a good symmetry and ignore this
complication. This issue will be further discussed in later
sections.} Natural candidates are the lightest baryon mass, denoted
here as $m_p$, and the lightest boson mass, $m_\pi$. Then we can
explicitly rearrange these relations to obtain a somewhat formal but
explicit expression for the renormalization group functions
\begin{eqnarray}
&\beta(g)=
\left(
a{\partial m_\pi \over \partial a}
{\partial m_p \over \partial m}
-a {\partial m_p \over \partial a}
{\partial m_\pi \over \partial m}
\right)
\bigg /
\left(
{\partial m_p \over \partial g}
{\partial m_\pi \over \partial m}
-{\partial m_\pi \over \partial g}
{\partial m_p \over \partial m}
\right)
\cr
&\cr
&\gamma(g)=
\left(
a{\partial m_\pi \over \partial a}
{\partial m_p \over \partial g}
-a {\partial m_p \over \partial a}
{\partial m_\pi \over \partial g}
\right)
\bigg /
\left(
{\partial m_p \over \partial m}
{\partial m_\pi \over \partial g}
-{\partial m_\pi \over \partial m}
{\partial m_p \over \partial g}
\right).
\end{eqnarray}
Note that this particular definition includes all perturbative and
non-perturbative effects. In addition, this approach avoids any need
for gauge fixing.
Once given $m_p$, $m_\pi$, and a renormalization scheme, then the
dependence of the bare parameters on the cutoff is completely fixed.
The physical masses are mapped onto the integration constants
\begin{eqnarray}
&&\Lambda=\Lambda(m_p,m_\pi)\\
&&M=M(m_p,m_\pi).
\end{eqnarray}
Formally these relations can be inverted to express the masses as
functions of the integration constants,
$m_i=m_i(\Lambda,M)$. Straightforward
dimensional analysis tells us that the masses must take the form
\begin{equation}
m_i=\Lambda f_i(M/\Lambda).
\end{equation}
As we will discuss in more detail in later sections, for the
multi-flavor theory we expect the pions to be Goldstone bosons with
$m_\pi^2 \sim m_q$. This tells us that the above function for the
pion should exhibit a square root singularity {$f_\pi(x)\sim
x^{1/2}$}. This relation removes any additive ambiguity in defining
the renormalized quark mass $M$. As will be discussed in more detail
later, this conclusion does not persist if the lightest quark becomes
non-degenerate.
\subsection{Which beta function?}
Thus far our discussion of the renormalization group has been in terms
of the bare charge with a cutoff in place. This is the natural
procedure in lattice gauge theory; however, there are alternative
approaches to the renormalization group that are frequently used in
the continuum theory. We now make some comments on connection between
the lattice and the continuum approaches.
An important issue is that there are many different ways to define a
renormalized coupling; it should first of all be an observable that
remains finite in the continuum limit
\begin{equation}
\lim_{a\rightarrow 0} g_r(\mu,a,g(a)) = g_r(\mu).
\end{equation}
Here $\mu$ is a dimensionful energy scale introduced to define the
renormalized coupling. The subscript $r$ is added to distinguish this
coupling from the bare one. For perturbative purposes one might use a
renormalized three-gluon vertex in a particular gauge and with all
legs at a given scale of momentum proportional to $\mu$. But many
alternatives are possible; for example, one might use as an observable
the force between two quarks at separation $1/\mu$.
Secondly, to be properly called a renormalization of the classical
coupling, $g_r$ should be normalized such that it reduces to the bare
coupling in lowest order perturbation theory for the cutoff theory
\begin{equation}
\label{renormalizedg}
g_r(\mu,a,g) = g+O(g^3).
\end{equation}
Beyond this, the definition of $g_r$ is totally arbitrary. In
particular, given any physical observable $H$ defined at scale $\mu$ and
satisfying a perturbative expansion
\begin{equation}
H(\mu,a,g)=h_0+h_1 g^2+O(g^4)
\label{obsa}
\end{equation}
we can define a corresponding renormalized coupling
\begin{equation}
g_H^2(\mu)=(H(\mu)-h_0)/h_1.
\end{equation}
As the energy scale goes to infinity, this renormalized charge should
go to zero. But with a different observable, we will generally obtain
a different functional behavior for this flow. From this flow of the
renormalized charge we can define a renormalized beta function
\begin{equation}
\beta_r=-\mu{\partial g_r(\mu)\over \partial \mu}.
\end{equation}
We now draw a remarkable connection between the renormalized
renormalization group function $\beta_r(g_r)$ and the function
$\beta(g)$ defined earlier for the bare coupling. When the cutoff
is still in place, the renormalized coupling is a function of the
scale $\mu$ of the observable, the cutoff $a$, and the bare coupling
$g$.
Since we are working with dimensionless couplings, $g_r$ can depend
directly on $\mu$ and $a$ only through their product. This simple
application of dimensional analysis implies
\begin{equation}
a\ \left.{\partial g_r\over \partial a}\right |_{g}
=\mu\ \left.{\partial g_r\over \partial \mu}\right |_{g}
=-\beta_r.
\end{equation}
Now, in the continuum limit as we take $a$ to zero and adjust $g$
appropriately, $g_r$ should become a function of the physical scale
$\mu$ alone. Indeed, we could use $g_r(\mu)$ itself as the physical
quantity to hold fixed for the continuum limit. Then we obtain
\begin{equation}
a{\partial g_r\over \partial a}+{\partial g_r\over \partial g}
a{\partial g\over \partial a}=0.
\end{equation}
Using this in an analysis similar to that in Eq.~(\ref{universal}), we
find
\begin{equation}
\beta_r(g_r)=\beta_0 g_r^3+\beta_1 g_r^5 +O(g_r^7).
\label{renormalizedgamma}
\end{equation}
Where $\beta_0$ and $\beta_1$ are the same coefficients that appear in
Eq.~(\ref{AF}). Both the renormalized and the bare $\beta$ functions
have the same first two coefficients in their perturbative expansions.
Indeed, it was through consideration of the renormalized coupling that
$\beta_0$ and $\beta_1$ were first calculated.
It is important to reiterate the considerable arbitrariness in
defining both the bare and the renormalized couplings. Far from the
continuum there need be no simple relationship between different
formulations. Once one leaves the perturbative region, even such
things as zeros in the $\beta$ functions are not universal. For one
extreme example, it is allowed to force the beta function to consist
of only the first two terms. In this case, as long as $N_f$ is small
enough that $\beta_1>0$, there is explicitly no other zero of the beta
function except at $g=0$. On the other hand, one might think it
natural to define the coupling from the force between two quarks.
When dynamical quarks are present, at large distances this falls
exponentially with the pion mass at large distances. In this case the
beta function must have another zero in the vicinity of where the
screening sets in. Thus, even the existence of zeros in the beta
function is scheme dependent. The only exception to this is if a zero
occurs in a region of small enough coupling that perturbation theory
can be trusted. This has been conjectured to happen for a sufficient
number of flavors \cite{Banks:1981nn}.
The perturbative expansion of $\beta_r$ has important experimental
consequences. If, as expected, the continuum limit is taken at
vanishing bare coupling and the renormalized coupling is small enough
that the first terms in Eq.~(\ref{renormalizedgamma}) dominate, then
the renormalized coupling will be driven to zero logarithmically as
its defining scale $\mu$ goes to infinity. Not only does the bare
coupling vanish, but the effective renormalized coupling becomes
arbitrarily weak at short distances. This is the physical implication
of asymptotic freedom; phenomena involving only short-distance effects
may be accurately described with a perturbative expansion. Indeed,
asymptotically free gauge theories were first invoked for the strong
interactions as an explanation of the apparently free parton behavior
manifested in the structure functions associated with deeply inelastic
scattering of leptons from hadrons.
The dependence of the integration constant $\Lambda$ on the details of
the renormalization scheme carries over to the continuum
renormalization group as well. Given a particular definition of the
renormalized coupling $g_r(\mu)$, its behavior as $r$ goes to zero
will involve a scale $\Lambda_r$ in analogy to the scale in the bare
coupling. Hasenfratz and Hasenfratz
\cite{Hasenfratz:1980kn,Hasenfratz:1981tw} were the first to perform
the necessary one loop calculations to relate $\Lambda$ from the
Wilson lattice gauge theory with $\Lambda_r$ defined from the
three-gluon vertex in the Feynman gauge and with all legs carrying
momentum $\mu^2$. They found
\begin{equation}
{\Lambda_r\over \Lambda}=\pmatrix{
57.5 & SU(2)\cr
83.5 & SU(3)\cr
}
\end{equation}
for the pure gauge theory. Note that not only is $\Lambda$ scheme
dependent, but that different definitions can vary by rather large
factors. The original calculation of these numbers was rather
tedious. They have been verified with calculationally more efficient
techniques based on quantum fluctuations around a slowly varying
classical background field \cite{Dashen:1980vm}.
\subsection{Effective potentials}
We begin with an elementary review of the concept of effective
potentials in quantum field theory. In generic continuum field theory
language, consider the path integral for a scalar field
\begin{equation}
Z=\int d\phi e^{-S(\phi)}.
\end{equation}
After adding in some external sources
\begin{equation}
Z(J)=\int d\phi e^{-S(\phi)+J\phi},
\end{equation}
general correlation functions can be found by differentiating with
respect to $J$. Here we use a shorthand notation that suppresses the
space dependence; {\it i.e.} $J\phi=\int dx J(x)\phi(x)$ in the
continuum, or $J\phi=\sum_i J_i\phi_i$ on the lattice.
One can think of $J$ as an external force pulling on the field. Such
a force will tend to drive the field to have an expectation value
\begin{equation}
\label{fieldforce}
\langle \phi \rangle_J=-{\partial F\over \partial J}
\end{equation}
where the free energy in the presence of the source is defined as
$F(J)=-\log(Z(J))$.
Now imagine inverting Eq.~(\ref{fieldforce}) to determine what value
of the force $J$ would be needed to give some desired expectation
value $\Phi$; {\it i.e.} we want to solve
\begin{equation}
\Phi(J)=\langle \phi \rangle_{J(\Phi)}=-{\partial F\over \partial J}
\end{equation}
for $J(\Phi)$. In terms of this formal solution, construct the
``Legendre transform''
\begin{equation}
V(\Phi)=F(J(\Phi))+\Phi J(\Phi)
\end{equation}
and look at
\begin{equation}
{\partial V \over \partial \Phi}
=-\Phi {\partial J \over \partial \Phi}+J+\Phi{\partial J \over \partial \Phi}
=J.
\end{equation}
If we now turn off the sources, this derivative vanishes. Thus the
expectation value of the field in the absence of sources occurs at an
extremum of $V(\Phi)$. This quantity $V$ is referred to as the
``effective potential.''
An interesting formal property of this construction follows from
looking at the second derivative of $V$
\begin{equation}
{\partial^2 V \over \partial \Phi^2}={\partial J\over \partial \Phi}.
\end{equation}
Actually, it is easier to look at the inverse
\begin{equation}
{\partial \Phi \over \partial J}=
-{\partial^2 F \over \partial J^2}
=\langle \phi^2 \rangle - \langle \phi \rangle ^2
=\langle (\phi-\langle\phi\rangle)^2 \rangle \ge 0.
\end{equation}
Thus this second derivative is never negative! This first of all
shows we are actually looking for a minimum and not a maximum of $V$,
but it also implies that $V(\Phi)$ can only have ONE minimum!
This convexity property is usually ignored in conventional
discussions, where phase transitions are signaled by jumps between
distinct minima of the potential.. So what is going on? Are phase
transitions impossible? Physically, the more you pull on the field,
the larger the expectation of $\Phi$ will become. It won't go back.
The proper interpretation is that we must do Maxwell's construction.
If we force the expectation of $\phi$ to lie between two distinct
stable phases, the system will phase separate into a heterogeneous
mixture. In this region the effective potential is flat. Note that
there is no large volume limit required in the above discussion.
However other definitions of $V$ can allow a small barrier at finite
volume due to surface tension effects. A mixed phase must contain
interfaces, and their energy represents a small barrier.
\subsection{Goldstone Bosons}
Now we turn to a brief discussion on some formal aspects of Goldstone
Bosons. Suppose we have a field theory containing a conserved current
\begin{equation} \partial_\mu j_\mu=0
\end{equation}
so the corresponding charge $Q=\int d^3x j_0(x)$ is a constant
\begin{equation} {dQ\over
dt}=-i[H,Q]=0.
\end{equation}
Here $H$ is the Hamiltonian for the system under consideration.
Suppose, however, that for some reason the vacuum is not a singlet
under this charge
\begin{equation} Q|0\rangle \ne 0.
\end{equation}
Then there must exist a massless particle in the theory. Consider the
state
\begin{equation}
\exp(i\theta\int d^3x j_0(x) e^{-\epsilon x^2}) |0\rangle
\end{equation}
where $\epsilon$ is a convenient cutoff and $\theta$ some parameter.
As epsilon goes to zero this state by assumption is not the vacuum,
but since the Hamiltonian commutes with $Q$, the expectation value of
the Hamiltonian goes to zero (normalize so the ground state energy is
zero). We can thus find a state that is not the vacuum but with
arbitrarily small energy. The theory has no mass gap. This situation
of having a symmetry under which the vacuum is not invariant is
referred to as ``spontaneous symmetry breaking.'' The low energy
states represent massless particles called Goldstone
bosons \cite{Goldstone:1962es}.
Free massless field theory is a marvelous example where everything can
be worked out. The massless equation of motion
\begin{equation} \partial_\mu\partial_\mu \phi=0
\end{equation}
can be written in the form
\begin{equation}
\partial_\mu j_\mu=0
\end{equation}
where
\begin{equation} j_\mu=\partial_\mu \phi.
\end{equation}
The broken symmetry is the invariance of the Lagrangean $L=\int
d^4x(\partial_\mu\phi)^2/2$ under constant shifts of the field
\begin{equation}
\phi\rightarrow\phi+c.
\end{equation}
Note that $j_0=\partial_0\phi=\pi$ is the conjugate variable to
$\phi$. Since it is a free theory, one could work out explicitly
\begin{equation} \langle 0 | \exp(i\theta\int d^3x j_0(x)
e^{-\epsilon x^2/2}) |0\rangle.
\end{equation}
We can, however, save ourselves the work using dimensional analysis.
The field $\phi$ has dimensions of inverse length, while $j_0$ goes as
inverse length squared. Thus $\theta$ above has units of inverse
length. These are the same dimensions as $\epsilon^2$. Now for a
free theory, by Wick's theorem, the answer must be Gaussian in
$\theta$. We conclude that the above overlap must go as
\begin{equation}
\exp(-C\theta^2/\epsilon^4)
\end{equation}
where $C$ is some non-vanishing dimensionless number. This expression
rapidly goes to zero as epsilon becomes small, showing that the vacuum
is indeed not invariant under the symmetry. In the limit of
$\epsilon$ going to zero, we obtain a new vacuum that is not even in
the same Hilbert space. The overlap of this new state with any local
polynomial of fields on the original vacuum vanishes.
It is perhaps interesting to note that the canonical commutation
relations $[\pi(x),\phi(y)]=i\delta(x-y)$ imply for the currents
\begin{equation}
[j_0(x),j_i(y)]=-i{d\over dx}\delta(x-y).
\end{equation}
In a Hamiltonian formulation, equal time commutators of different
current components must involve derivatives of delta functions. This
is a generic property and does not depend on the symmetry being
spontaneously broken \cite{Schwinger:1959xd}.
\subsection{Surface terms}
A remarkable feature of this formalism is that the combination ${\rm
Tr}\ F\tilde F$ is a total derivative. To see this first construct
\begin{eqnarray}
&&F\tilde F={1\over 2}\epsilon_{\mu\nu\rho\sigma}
(2\partial_\mu A_\nu+igA_\mu A_\nu)
(2\partial_\rho A_\sigma+igA_\rho A_\sigma)\cr
&&={1\over 2}\epsilon_{\mu\nu\rho\sigma}\left(
4\partial_\mu A_\nu \partial_\rho A_\sigma
+4 ig(\partial_\mu A_\nu) A_\rho A_\sigma
-g^2 A_\mu A_\nu A_\rho A_\sigma
\right).
\end{eqnarray}
If we take a trace of this quantity, the last term will drop out due
to cyclicity. Thus
\begin{equation}
{\rm Tr}F\tilde F=2\partial_\mu \epsilon_{\mu\nu\rho\sigma}
{\rm Tr}\left(A_\nu \partial_\rho A_\sigma
+ig A_\nu A_\rho A_\sigma\right)=2\partial_\mu K_\mu
\end{equation}
where we define
\begin{equation}
K_\mu\equiv \epsilon_{\mu\nu\rho\sigma}
{\rm Tr}\left(A_\nu \partial_\rho A_\sigma
+2ig A_\nu A_\rho A_\sigma\right).
\end{equation}
Note that although ${\rm Tr}\ F\tilde F$ is gauge invariant, this is
not true for $K_\mu$.
Being a total derivative, the integral of this quantity
\begin{equation}
\int d^4x\ {1\over 2}{\rm Tr}F\tilde F
\end{equation}
would vanish if we ignore surface terms. What is remarkable is that
there exist finite action gauge configurations for which this does not
vanish even though the field strengths all go to zero rapidly at
infinity. This is because the gauge fields $A_\mu$ that appear
explicitly in the current $K_\mu$ need not necessarily vanish as
rapidly as $F_{\mu\nu}$.
These surface terms are closely tied to the topology of the gauge
potential at large distances. As we want the field strengths to go to
zero at infinity, the potential should approach a pure gauge form
$A_\mu\rightarrow {-i\over g} h^\dagger \partial_\mu h$. In this case
\begin{equation}
K_\mu \rightarrow -{1\over g^2}\epsilon_{\mu \nu \rho \sigma}
{\rm Tr}
(h^\dagger \partial_\nu h)
(h^\dagger \partial_\rho h)
(h^\dagger \partial_\sigma h).
\end{equation}
Note the similarity of this form to that for the group measure in
Eq.~(\ref{measure}). Indeed, it is invariant if we take $h\rightarrow
h^\prime h$ with $h^\prime$ being a constant group element. The
surface at infinity is topologically a three dimensional sphere $S_3$.
If we concentrate on $SU(2)$, this is the same as the topology of the
group space. For larger groups we can restrict $h$ to an $SU(2)$
subgroup and proceed similarly. Thus the integral of $K_\mu$ over the
surface reduces to the integral of $h$ over a sphere with the
invariant group measure. This can give a non-vanishing contribution
if the mapping of $h$ onto the sphere at infinity covers the entire
group in a non-trivial manner. Mathematically, one can map the $S_3$
of infinite space onto the $S_3$ of group space an integer number of
times, i.e. $\Pi_3(SU(2))=Z$. Thus we have
\begin{equation}
\int d^4x\ {1\over 2} {\rm Tr}F\tilde F \propto \nu
\end{equation}
where $\nu$ is an integer describing the number of times $h(x)$ wraps
around the group as $x$ covers the sphere at infinity. The
normalization involves the surface area of a three dimensional sphere
and can be worked out with the result
\begin{equation}
\label{windingone}
\int d^4x\ {1\over 2}{\rm Tr}F\tilde F ={8\pi^2 \nu\over g^2}.
\end{equation}
For groups larger than $SU(2)$ one can deform $h$ to lie in an $SU(2)$
subgroup, and thus this quantization of the surface term applies to
any $SU(N)$.
If we were to place such a configuration into the path integral for
the quantum theory, we might expect a suppression of these effects by
a factor of $\exp(-8\pi^2/g^2)$. This is clearly non-perturbative,
however this factor strongly underestimates the importance of
topological effects. The problem is that we only need to excite
non-trivial fields over the quantum mechanical vacuum, not the
classical one. The correct suppression is indeed exponential in the
inverse coupling squared, but the coefficient in the exponent can be
determined from asymptotic freedom and dimensional transmutation. We
will return to this point in Subsection \ref{mixing}.
The combination ${\rm Tr}F\tilde F$ is formally a dimension four
operator, the same as the basic gauge theory action density ${\rm Tr}F
F$. This naturally leads to the question of what would happen if we
consider a new action which also includes this parity odd term.
Classically it does nothing since it reduces to the surface term
described above. However quantum mechanically this is no longer the
case. As we will discuss extensively later, the physics of QCD
depends quite non-trivially on such a term. An interesting feature of
this new term follows from the quantization of the resulting surface
term. Because of the above quantization and an imaginary factor in
the path integral, physics is periodic in the coefficient of ${1\over
2}{\rm Tr}F\tilde F$. Although discussing the consequences directly
with such a term in the action is traditional, we will follow a
somewhat different path in later sections and introduce this physics
through its effects on fermions.
\subsection{An explicit solution}
To demonstrate that non-trivial winding solutions indeed exist,
specialize to $SU(2)$ and find an explicit example. To start,
consider the positivity of the norm of $F\pm\tilde F$
\begin{equation}
0\le \int d^4x\ (F\pm \tilde F)^2= 2\int d^4x\ F^2 \pm 2\int d^4x\
F\tilde F.
\end{equation}
This means that the action is bounded below by $
\int d^4x\ {1\over 2}{\rm Tr}F\tilde F
$ and this bound is reached only if $F=\pm\tilde F$. As mentioned
earlier, reaching this is sufficient to guarantee a solution to the
equations of motion. We will now explicitly construct such a self
dual configuration.
Start with a gauge transformation function which is singular at the
origin but maps around the group at a constant radius
\begin{equation}
h(x_\mu)=
{t+i\vec\tau\cdot \vec x \over \sqrt{x^2}}={T_\mu x_\mu \over |x|}.
\end{equation}
Here we define the four component object $T_\mu=\{1,i\vec\tau\}$.
Considering space with the origin removed, construct the
pure gauge field
\begin{equation}
B_\mu={-i\over g}h^\dagger \partial_\mu h
={-i\over g}h^\dagger (T_\mu x^2 - x_\mu T\cdot x)/|x|^3.
\end{equation}
Because this is nothing but a gauge transformation of a vanishing
gauge field, the corresponding field strength
automatically vanishes
\begin{equation}
\partial_\mu B_\nu -\partial _\nu B_\mu +ig[B_\mu,B_\nu]=0.
\end{equation}
This construction gives a unit winding at infinity. However this
gauge field is singular at the origin where the winding unwraps.
If we smooth this singularity at $x=0$ with a field of form
\begin{equation}
A_\mu=f(x^2)B_\mu.
\end{equation}
where $f(0)=0$ and $f(\infty)=1$, this will remove the unwrapping at
the origin and automatically leave a field configuration with
non-trivial winding. The idea is to find a particular $f(x^2)$
such that $A$ also gives a self dual field strength and thereby is a
solution to the equations of motion.
We have set things up symmetrically under space-time rotations about
the origin. This connection with $O(4)$ is convenient in that we only
need to verify the self duality along a single direction. Consider
this to be the time axis, along which self duality requires
\begin{equation}
F_{01}(\vec x=0, t)=\pm F_{23}(\vec x=0, t).
\end{equation}
A little algebra gives
\begin{equation}
F_{\mu\nu}=
(f-f^{2})(\partial_\mu B_\nu -\partial _\nu B_\mu)
+2 f^\prime (x_\mu B_\nu-x_\nu B_\mu).
\end{equation}
So along the time axis we have
\begin{equation}
F_{01}\rightarrow
2f^\prime t {\tau_1\over gt}
\qquad \qquad
F_{23}\rightarrow (f-f^2) {2\tau_1\over g t^2}.
\end{equation}
Thus the self duality condition reduces to a simple first order
differential equation
\begin{equation}
z f^\prime(z)=\pm (f-f^2).
\end{equation}
This is easily solved to give
\begin{equation}
f(z)={1\over 1+\rho^2 z^{\pm 1}}
\end{equation}
where $\rho$ is an arbitrary constant of integration. To have the
function vanish at the origin we take the minus solution. The
resulting form for the gauge field
\begin{equation}
\label{instantonfield}
A_\mu={-i x^2\over g (x^2+\rho^2)} h^\dagger \partial_\mu h
\end{equation}
is the self dual instanton. The parameter $\rho$ controls the size
of the configuration. Its arbitrary value is a consequence of the
conformal invariance of the classical theory. Switching $h$ and
$h^\dagger$ gives a solution with the opposite winding.
\subsection {Zero modes and the Dirac operator}
A particularly important and intriguing aspect of the above field
configuration is that it supports an exact zero mode for the classical
Dirac operator. We will later discuss the rigorous connection between
the gauge field winding and the zero modes of the Dirac operator.
Here, however, we will verify this connection explicitly for the above
solution. Thus we look for a spinor field $\psi(x)$ satisfying
\begin{equation}
\gamma_\mu D_\mu \psi(x)=
\gamma_\mu\left(\partial_\mu+
i g A_\mu\right)
\psi(x)=0
\end{equation}
where we insert the gauge field from Eq.~(\ref{instantonfield}). The
wave function $\psi$ is a spinor in Dirac space and a doublet in
$SU(2)$ space; i.e. it has 8 components. Similarly, $\gamma_\mu
A_\mu$ is an 8 by 8 matrix, with a factor of four from spinor space
and a factor of two from the internal gauge symmetry. The solution
entangles all of these indices in a non-trival manner.
Since we don't want a singularity in $\psi$ at the origin, it is
natural to look for a solution of form
\begin{equation}
\psi(x)= p(|x|) V
\end{equation}
where $p$ is a scalar function of the four dimensional radius and $V$
is a constant vector in spinor and color space. As before, it is
convenient to look for the solution along the time axis. There $A_0$
vanishes and we have
\begin{equation}
\vec A=
{1\over g}
\ { t\over t^2+\rho^2}
\vec \tau.
\end{equation}
Then the equation of interest reduces to
\begin{equation}
\gamma_0 {d\over dt} \psi(t)=-{ t\over t^2+\rho^2}
\ \vec \tau \cdot \vec \gamma\ \psi(t).
\end{equation}
The 8 by 8 matrix $\vec \tau \cdot \vec \gamma$ is readily
diagonalized giving the eigenvalues $\{-3,-1,-1,-1,1,1,1,3\}$. Only the
$+3$ eigenvalue gives a normalizable solution
\begin{equation}
\psi(t)=\psi(0) \exp\left(-3\int_0^t {t\ dt\over t^2+\rho^2}\right).
\end{equation}
For general $x_\mu$ this becomes
\begin{equation}
\psi(x)
=\psi(0)\left(
{\rho^2 \over x^2+\rho^2}
\right)^{3/2}.
\end{equation}
At large $x$ this goes at $x^{-3}$ so its square is normalizable.
None of the other eigenvalues of $\vec\tau\cdot\vec\gamma$ go to zero
fast enough for normalization; thus, the solution is unique.
We see the appearance of a direct product of two $SU(2)$'s, one from
spin and one from isospin. As we rotate around the origin, for the
large eigenvalue these rotate together as an overall singlet. The
other positive eigenvalues of $\vec\tau\cdot\vec\gamma$ represent the
triplet combination while the negative eigenvalues come from
antiparticle states.
This zero eigenvalue of $D$ is robust under smooth deformations of the
gauge field. This is because the anti-commutation of $D$ with
$\gamma_5$ says that all non-zero eigenvalues of $D$ occur in
conjugate pairs. Without bringing in another eigenvalue, the isolated
one at zero cannot move. In the next subsection we will demonstrate
the general result that for arbitrary smooth gauge fields the number
of zero modes of the Dirac operator is directly given by the
topological winding number.
\subsection{Topology and eigenvalue flow}
There is a close connection between the zero modes of the Dirac
operator in the Euclidean path integral and a flow of eigenvalues of
the fermion Hamiltonian in Minkowski space. To see how this works it
is convenient to work in the temporal gauge with $A_0=0$ and separate
out the space-like part of the Dirac operator
\begin{equation}
D=\gamma_0 \partial_0 + \gamma_0 H(\vec A(t)).
\end{equation}
Consider $\vec A$ as some time dependent gauge field through which the
fermions propagate. Assume that at large positive or negative times
this background field reduces to a constant. Without a mass term, the
continuum theory Hamiltonian $H$ commutes with $\gamma_5$ and
anti-commutes with $\gamma_0$. Therefore its eigenvalues appear in
pairs of opposite energy and opposite chirality; {\it i.e.} if we have
\begin{eqnarray}
H\phi=E\phi\cr
\gamma_5\phi=\pm\phi
\end{eqnarray}
then
\begin{eqnarray}
H\gamma_0\phi=-E\gamma_0\phi\cr
\gamma_5\gamma_0\phi=\mp\gamma_0\phi.
\end{eqnarray}
\begin{figure}
\centering
\includegraphics[width=2.5in]{eflow.eps}
\caption{An energy eigenvalue that changes in sign between the distant
past and future.}
\label{eflow}
\end{figure}
Now suppose we diagonalize $H$
at some given time
\begin{equation}
H(\vec A(t))\phi_i(t) = E_i(t) \phi_i(t)
\end{equation}
where the wave function $\phi(t)$ implicitly depends on space, spinor,
and color indices. Suppose further that we can find some eigenvalue
that changes adiabatically from negative to positive in going from
large negative to large positive time, as sketched in
Fig.~\ref{eflow}. From this particular eigenstate construct the four
dimensional field
\begin{equation}
\psi(t)=e^{-\int_0^t E(t^\prime)\ dt^\prime} \phi(t).
\label{eigenflow}
\end{equation}
Because of the change in sign of the energy, the exponential factor
function goes to zero at both positive and negative large times, as
sketched in Fig.~\ref{zeromode}.
If we now consider the four dimensional Dirac operator applied to this
function we obtain
\begin{equation}
(\gamma_0 \partial_0 + \gamma_0 H(\vec A(t))) \psi(t)
=O(\partial_0 \psi(t)).
\end{equation}
If the evolution is adiabatic, the last term is small and we have an
approximate zero mode. The assumption of adiabaticity is unnecessary
in the chiral limit of zero mass. Then the eigenvalues of $D$ are
either real or occur in complex conjugate pairs. Any unaccompanied
eigenvalue of $D$ occurs robustly at zero. This is another
manifestation of the index theorem; we can count Euclidean-space zero
modes by studying the zero crossings appearing in the eigenvalues of
the Minkowski-space Hamiltonian.
In the above construction, the evolving eigenmode of $H$ is
accompanied by another of opposite energy and chirality. Inserted
into Eq.~(\ref{eigenflow}), this will give a non-normalizable form for
the four dimensional field. Thus we obtain only a single normalizable
zero mode for the Euclidean Dirac operator. Note that if a small mass
term is included, the up going and down going Hamiltonian eigenstates
will mix and the crossing is forbidden.
This eigenvalue flow provides an intuitive picture of the anomaly
\cite{Ambjorn:1983hp}. Start at early times with a filled Dirac sea
and all negative-energy eigenstates filled and then slowly evolve
through one of the above crossings. In the process one of the filled
states moves to positive energy, leaving a non-empty positive energy
state. At the same time the opposite chirality state moves from
positive to negative energy. As long as the process is adiabatic, we
wind up at large time with one filled positive-energy state and one
empty negative-energy state. As these are of opposite chirality,
effectively chirality is not conserved. The result is particularly
dramatic in the weak interactions, where anomalies are canceled
between quarks and leptons. This flow from negative to positive
energy states results in baryon non-conservation, although at an
unobservably small rate \cite{'tHooft:1976fv}.
\begin{figure}
\centering
\includegraphics[width=3in]{zeromode.eps}
\caption{The adiabatic evolution gives rise to a normalizable zero
mode of the four dimensional Dirac operator.}
\label{zeromode}
\end{figure}
\subsection{Hopping and doublers}
The essence of lattice doubling already appears in the quantum
mechanics of the simplest fermion Hamiltonian in one space dimension
\begin{equation}
H=iK\sum_j a_{j+1}^\dagger a_j - a_j^\dagger a_{j+1}.
\end{equation}
Here $j$ is an integer labeling the sites of an infinite
chain and the $a_j$ are fermion annihilation operators satisfying
standard anti-commutation relations
\begin{equation}
\left[ a_j, a_k^\dagger\right]_+\equiv a_j a_k^\dagger+ a_k^\dagger a_j
=\delta_{j,k}.
\end{equation}
The fermions hop from site to neighboring site with amplitude $K$;
thus, we refer to $K$ as the ``hopping parameter'' and by convention
take it to be positive. The bare vacuum $\vert 0 \rangle$ satisfies
$a_j \vert 0 \rangle=0.$ This vacuum is not the physical one, which
requires constructing a filled Dirac sea. Energy eigenstates in the
single fermion sector
\begin{equation}
\vert \chi\rangle=\sum_j \chi_j a_j^\dagger \vert 0 \rangle
\end{equation}
can be easily found in momentum space
\begin{equation}
\chi_j(q)= e^{iqj} \chi_0
\end{equation}
where we can restrict $-\pi < q \le \pi$. The resulting energy is
\begin{equation}
E(q)=2K \sin(q).
\end{equation}
The physical vacuum fills all the negative energy states, {\it i.e.}
those with $-\pi<q < 0.$
On this vacuum, consider constructing a fermionic wave packet by
exciting a few modes of small momentum $q$. This packet will have a
group velocity $dE/dq\sim 2K$ that is positive. Thus it moves to the
right and represents a right-moving fermion. On the other hand, a
wave packet of low energy can also be produced by exciting momenta in
the vicinity of $q\sim \pi$. This packet will have group velocity
$\left . {dE\over dq}\right|_{q=\pi}\sim -2K$ and therefore be left moving.
The essence of the Nielsen Ninomiya theorem \cite{Nielsen:1980rz} is
that we must have both types of excitation. We will go into this in
more detail later, but for this one dimensional case the periodicity
in $q$ requires the dispersion relation to have an equal number of
zeros with positive and negative slopes. If we now consider a two
component spinor to describe the fermion, we will have independent
states corresponding to each component. This is the so called
``doubling'' issue.
As a preliminary to later discussion, here we concentrate on a
Hamiltonian version of the Wilson approach to remove the doublers.
Continuing to work in one dimension, consider a two component spinor
\begin{equation}
\psi=\pmatrix{a\cr b\cr}.
\end{equation}
where $a$ and $b$ are distinct fermion annihilation operators on the
lattice sites. The so called ``naive'' lattice Hamiltonian begins
with the simple hopping case of above and adds in the lower components
and a mass term that mixes the upper and lower components
\begin{equation}
H=iK\sum_j
a_{j+1}^\dagger a_j - a_j^\dagger a_{j+1}
-b_{j+1}^\dagger b_j + b_j^\dagger b_{j+1}
+M\sum_j a^\dagger_j b_j + b^\dagger_j a_j.
\end{equation}
Introducing two by two Dirac matrices
\begin{equation}
\gamma_0=\sigma_1=\pmatrix{0&1\cr
1&0\cr},
\ \ \gamma_1=\sigma_2=\pmatrix{0&-i\cr
i&0\cr},
\ \ \gamma_5=\sigma_3=\pmatrix{1&0\cr
0& -1\cr},
\end{equation}
and defining
\def \overline\psi{\overline\psi}
\def \overline\chi{\overline\chi}
$\overline\psi=\psi^\dagger\gamma_0$,
we write the Hamiltonian compactly as
\begin{equation}
H=\sum_j K(\overline\psi_{j+1} \gamma_1 \psi_j-\overline\psi_j \gamma_1 \psi_{j+1})
+M\sum_j\overline\psi_j \psi_j.
\end{equation}
This looks very much like the continuum Dirac Hamiltonian with the
derivative term represented on the lattice by a nearest neighbor
difference. Chiral symmetry is manifest in the possibility of
independent rotations of the $a$ and $b$ type particles when the mass
term is absent. The latter mixes these components and opens a gap in
the spectrum.
As before, the single particle states are found by Fourier
transformation and satisfy
\begin{equation}
E^2=4K^2 \sin^2(q)+M^2.
\end{equation}
At each momentum there is one positive and one negative energy state.
Again, we are to fill the negative energy sea to form the physical
vacuum. The doubling issue is that there are low energy excitations
that satisfy the Dirac equation appearing both at $q\sim 0$ and $q\sim
\pi$. The physical momenta $k$ of the latter excitations appear in
expanding about pi, i.e $k=q-\pi$. These states have a smooth spatial
dependence in a redefined field $\chi_j=(-1)^j \psi_j$. The doublers
at $q\sim\pi$ are still with us.
\subsection{Wilson fermions}
One way to remove the degeneracy of the doublers is to make the mixing
of the upper and lower components momentum dependent. A simple way of
doing this was proposed by Wilson \cite{Wilson:1975id}. To our
Hamiltonian model, we add one more term
\begin{eqnarray}
H&=&iK\sum_j
a_{j+1}^\dagger a_j - a_j^\dagger a_{j+1}
-b_{j+1}^\dagger b_j + b_j^\dagger b_{j+1}+\cr
&& M\sum_j a^\dagger_j b_j + b^\dagger_j a_j
-rK \sum_j a_j^\dagger b_{j+1}+b_j^\dagger a_{j+1}
+b_{j+1}^\dagger a_j+a_{j+1}^\dagger b_j \cr
&=&\sum_j K(\overline\psi_{j+1} (\gamma_1-r) \psi_j-\overline\psi_j (\gamma_1+r)
\psi_{j+1})+\sum_jM\overline\psi_j \psi_j.
\end{eqnarray}
Now the spectrum satisfies
\begin{equation}
E^2=4K^2 \sin^2(q)+(M-2rK\cos(q))^2.
\end{equation}
The doublers at $q\sim \pi$ are increased in energy relative
to the states at $q\sim 0$. The physical particle mass is now $
m=M-2rK $ while that of the doubler is at $M+2rK$.
When $r$ becomes large, the dip in the spectrum of near $q=\pi$
actually becomes a maximum. This is irrelevant for our discussion,
although we note that the case $r=1$ is somewhat special. For this
value, the matrices $(\gamma_1\pm 1)/2$, which determine how the
fermions hop along the lattice, are projection operators. In a sense,
the doubler is removed because only one component can hop. This
choice $r=1$ has been the most popular in practice.
The hopping parameter has a critical value at
\begin{equation}
K_c={M \over 2r}.
\end{equation}
At this point the gap in the spectrum closes and one species of
fermion becomes massless. The Wilson term, proportional to $r$, still
mixes the $a$ and $b$ type particles; so, there is no exact chiral
symmetry. Nevertheless, in the continuum limit this represents a
candidate for a chirally symmetric theory. Before the limit, chiral
symmetry does not provide a good order parameter.
Now we generalize this approach to the Euclidean path integral
formulation in four space-time dimensions. In the continuum, one
usually writes for the free fermion action density
\begin{equation}
\overline\psi D \psi=\overline\psi (\slashchar\partial +m) \psi
\end{equation}
or in momentum space
\begin{equation}
\overline\psi (i\slashchar p +m) \psi.
\end{equation}
By convention we use Hermitean gamma matrices. Note that $D$ is the
sum of Hermitean and anti-Hermitean parts. In the continuum the
former is just a constant, the mass. A Hermitean operator appears in
the combination $\gamma_5 D$, but we don't need that just now.
A matrix can be diagonalized when it commutes with its adjoint; then
it is called ``normal.'' For the naive continuum operator this is the
case, and we see that all eigenvalues of $D$ lie along a line parallel
to the imaginary axis intersecting the real axis at $m$. This simple
structure will be lost on the lattice.
As discussed earlier, a simple transcription of derivatives onto the
lattice replaces factors of $p$ with trigonometric functions. Thus
the naive lattice action becomes
\begin{equation}
\overline\psi\left ({i\over a}\sum_\mu \gamma_\mu\sin(p_\mu a) +m\right) \psi
\end{equation}
where we have explicitly included the lattice spacing $a$. For small
momentum this reduces to the continuum result $\overline\psi(i\gamma_\mu
p_\mu+m)\psi$. Now let one component of $p$ get large and be near
$\pi/a$. Then we again have small eigenvalues and a nearby pole in
the propagator. As any of the four components of momentum can be near
$0$ or $\pi$ there are a total of 16 places in momentum space that
give rise to a Dirac like behavior. The naive fermion action gives
rise to 16 doublers.
As in the earlier example, the Wilson solution adds a momentum
dependent mass. Wishing to maintain only nearest neighbor terms, it
also involves trigonometric functions. To maintain hyper-cubic
symmetry, we put in the Wilson term symmetrically for all space-time
directions. For simplicity we set the Wilson parameter $r$ from
before to unity. Explicitly for free fields we consider the momentum
space form
\begin{equation}
\overline\psi D_W \psi = \overline\psi \left(
{1\over a} \sum_\mu(i\gamma_\mu\sin(p_\mu a)+1-\cos(p_\mu a))
+m\right) \psi.
\end{equation}
Now for a momentum component near $\pi$ the eigenvalues are of order
$1/a$ in size. Note that the lattice artifacts in the propagator
start at order $p^2 a$, rather than $O(a^2)$ as for naive fermions.
The eigenvalue structure of $D_W$ is rather interesting. The
eigenvalues for the free Wilson theory occur at
\begin{equation}
\label{freewilson}
\lambda=\pm {i\over a} \sqrt{\sum_\mu \sin^2(p_\mu a)} +
{1\over a}\sum_\mu 1-\cos(p_\mu a)
+m.
\end{equation}
The eigenvalues of this free operator lie on a set of ``nested
circles,'' as sketched in Fig.~\ref{eigen1}. Note that
$m\leftrightarrow -m$ is not a symmetry. Naively it would be in the
continuum, but as we discussed earlier, it cannot be so in the quantum
theory when one has an odd number of flavors.
Note that to obtain real eigenvalues in Eq.~(\ref{freewilson}), each
component of the momentum must be an integer multiple of $\pi$. There
are actually several critical values that can give rise to massless
fermions. For $m=0,-2,-4,-6,-8$ we have $1,4,6,4,1$ massless species.
When interactions are present these values of the mass will also be
renormalized.\footnote{Actually the 6 flavor case at m=-4 does have a
discrete symmetry that will protect against additive mass
renormalization.} Whether a continuum limit at these alternative
points is useful has not been investigated.
Rescaling to lattice units and restoring the hopping parameter, the
Wilson fermion action with the site indices explicit becomes
\begin{equation}
{D_W}_{ij}=\delta_{i,j}+K\sum_\mu
(1-\gamma_\mu )\delta_{i,j+e_\mu}
+(1+\gamma_\mu )\delta_{i,j-e_\mu}.
\end{equation}
By taking the coefficient $r$ of the Wilson term as unity we have
projection operators in the hoppings. The physical fermion mass
is read off from the small momentum behavior as $m={1\over
2a}(1/K-8)$. This vanishes at at $K=K_c=1/8$.
Here we consider that the gauge fields are formulated as usual with
group valued matrices on the lattice links. These are to be inserted
into the above hopping terms. One could use the simple Wilson gauge
action as a sum over plaquettes
\begin{equation}
S_g={\beta\over 3}\sum_p {\rm Re\ Tr\ } U_p
\end{equation}
although this specific form is not essential the qualitative nature of
the phase diagram. When the gauge fields are turned on, the dynamics
will move the fermion eigenvalues around, partially filling the holes
in eigenvalue pattern of Fig.~\ref{eigen1}. Some eigenvalues can
become real and are related to gauge field topology
\cite{Creutz:2006ts}.
For the free theory the Hermitean and anti-Hermitean parts of the
action commute. This ceases to be true in the interacting case
since both terms contain gauge matrices that themselves do not
commute. Thus the left eigenvalues are generally different from right
ones. Nevertheless, it is still true that the eigenvalues
either appear in complex conjugate pairs or they are real. This
follows from $\gamma_5$ Hermiticity, $D^\dagger=\gamma_5 D \gamma_5$.
Since $\gamma_5$ has unit determinant, $|D-\lambda|=0$ implies
$|D^\dagger-\lambda|=|D-\lambda^*|^*=0$.
A technical difficulty with this approach is that gauge interactions
will renormalize the parameters. To obtain massless pions one must
finely tune $K$ to $K_{c}$, an {\it a priori} unknown function of
the gauge coupling. Despite the awkwardness of such tuning, this is
how numerical simulations with Wilson quarks generally proceed. The
hopping parameter is adjusted to get the pion mass right, and one
assumes that the remaining predictions of current algebra reappear in
the continuum limit.
\begin{figure}
\centering
\includegraphics[width=3in]{eigen1.eps}
\caption{ The eigenvalue spectrum of the free Wilson fermion operator
is a set of nested circles. On turning on the gauge fields, some
eigenvalues drift into the open regions. Some complex pairs can
collide and become real. These are connected to gauge field topology.
Figure taken from Ref.~\cite{Creutz:2007fe}.}
\label{eigen1}
\end{figure}
Note that the basic lattice theory has two parameters, $\beta$ and
$K$. These are related to bare coupling, $\beta\sim 6/g_0^2$, and
quark mass, $(1/K-1/K_c)\sim m_q$. We will now turn to a discussion
of this relation in more detail.
\subsection{Lattice versus continuum parameters}
\label{parameters}
As emphasized earlier, QCD is a remarkably economical theory in terms
of the number of adjustable parameters. First of these is the overall
strong interaction scale, $\Lambda_{qcd}$. This is scheme dependent,
but once a renormalization procedure has been selected, it is well
defined. It is not independent of the coupling constant, the
connection being fixed by asymptotic freedom. In addition, the theory
depends on the renormalized quark masses $m_i$, or more precisely the
dimensionless ratios $m_i/\Lambda_{qcd}$. As with the overall scale,
the definition of $m_i$ is scheme dependent. The two flavor theory
with degenerate quarks and $\Theta=0$ has one such mass parameter. As
we wish to formulate the theory with a lattice cutoff in place, there
is a scale for this cutoff. As with everything else, it is convenient
to measure this in units of the overall scale; so, a third parameter
for the cutoff theory is $a\Lambda_{qcd}$, where one can regard $a$ as
the lattice spacing.
How the bare parameters behave as the continuum limit is taken was
discussed rather abstractly in Section \ref{asymptoticfreedom}. The
goal here is to explore some of the the lattice artifacts that arise
with Wilson fermions \cite{Wilson:1975id}. On the lattice it is
generally easier to work directly with lattice parameters. One of
these is the plaquette coupling $\beta$, which, with the usual
conventions, is related to the bare coupling $\beta=6/g_0^2$. For the
quarks, the natural lattice quantity is the ``hopping parameter'' $K$.
And finally, the connection with physical scales appears via the
lattice spacing $a$.
The set of physical parameters and the set of lattice parameters are,
of course, equivalent, and there is a well understood non-linear
mapping between them
\begin{equation}
\left\{
a\Lambda_{qcd},
{m\over\Lambda_{qcd}}
\right\}
\longleftrightarrow \{\beta,K\}.
\end{equation}
Of course, to extract physical predictions we are interested in the
continuum limit $a\Lambda_{qcd}\rightarrow 0$. For this, asymptotic
freedom tells us we must take $\beta\rightarrow \infty$ at a rate tied
to $\Lambda_{qcd}$. Simultaneously we must take the hopping parameter
to a critical value. With normal conventions, this takes
$K\rightarrow K_c\rightarrow 1/8$ at a rate tied to desired quark mass
$m$. Figure \ref{kbeta1} sketches how the continuum limit is taken in
the $\beta,K$ plane. Here we wish to further explore this phase
diagram with particular attention to hopping parameters larger than
$K_c$. This discussion is adapted from Ref.~\cite{Creutz:2007fe} and
adds the possible twisted mass term to the exposition from
Ref.~\cite{Creutz:1996bg}.
\begin{figure}
\centering
\includegraphics[width=3in]{kbeta1.eps}
\caption{ The continuum limit of lattice gauge theory with Wilson
fermions occurs at $\beta\rightarrow\infty$ and $K\rightarrow 1/8$.
Coming in from this point to finite beta is the curve $K_{c}(\beta)$,
representing the lowest phase transition in $K$ for fixed beta. The
nature of this phase transition is a delicate matter, discussed in the
text. Figure taken from Ref.~\cite{Creutz:2007fe}.}
\label{kbeta1}
\end{figure}
\subsection{Artifacts and the Aoki phase}
We previously made extensive use of an effective field theory to
describe the interactions of the pseudo-scalar mesons. Here we will
begin with the simplest form for the two flavor theory and then add
terms to mimic possible lattice artifacts. The language is framed as
before in terms of the isovector pion field $\vec\pi\sim
i\overline\psi \gamma_5 \vec \tau \psi$ and the scalar sigma
$\sigma\sim \overline\psi \psi$. The starting point for this
discussion is the canonical ``Mexican hat'' potential
\begin{equation}
V_0=\lambda (\sigma^2+\vec\pi^2-v^2)^2
\end{equation}
schematically sketched earlier in Fig.~\ref{v0}. The potential has a
symmetry under $O(4)$ rotations amongst the pion and sigma fields
expressed as the four vector $\Sigma=(\sigma,\vec\pi)$. This
represents the axial symmetry of the underlying quark theory.
As discussed before, the massless theory is expected to spontaneously
break chiral symmetry with the minimum for the potential occurring at
a non-vanishing value for the fields. As usual, we take the vacuum to
lie in the sigma direction with $\langle\sigma\rangle > 0$. The pions
are then Goldstone bosons, being massless because the potential
provides no barrier to oscillations of the fields in the pion
directions. Also as discussed before, we include a quark mass by
adding a constant times the sigma field
\begin{equation}
V_1=-m\sigma.
\end{equation}
This explicitly breaks the chiral symmetry by ``tilting'' the
potential as sketched in Fig.~\ref{v3}. That selects a unique
vacuum which, for $m>0$, gives a positive expectation for sigma. In
the process the pions gain a mass, with $m_\pi^2\sim m$.
Because of the symmetry of $V_0$, it does not matter physically in
which direction we tilt the vacuum. In particular, a mass term of
form
\begin{equation}
m\sigma\rightarrow m\cos(\theta) \sigma + m\sin(\theta)\pi_3
\label{mrot}
\end{equation}
should give equivalent physics for any $\theta$. In the earlier
continuum discussion we used this freedom to rotate the second term
away. However, as we will see, lattice artifacts can break this
symmetry, introducing the possibility of physics at finite lattice
spacing depending on this angle. As mentioned before, the second term
in this equation is what is usually called a ``twisted mass.''
The Wilson term inherently breaks chiral symmetry. This will give
rise to various modifications of the effective potential. The first
correction is expected to be an additive contribution to the quark
mass, i.e. an additional tilt to the potential. This means that the
critical kappa, defined as the smallest kappa where a singularity is
found in the $\beta,K$ plane, will move away from the limiting value
of $1/8$. Thus we introduce the function $K_c(\beta)$ and imagine
that the mass term is modeled with the form
\begin{equation}
m\rightarrow c_1(1/K-1/K_c(\beta)).
\label{c1}
\end{equation}
In general the lattice modification of the effective potential will
have further corrections of higher order in the effective fields. A
natural way to include such is as an expansion in the chiral fields.
With this motivation we include a term in the potential of form $c_2
\sigma^2$. Including these ideas in the effective model, we are led
to
\begin{equation}
V(\vec\pi,\sigma)=
\lambda(\sigma^2+\vec\pi^2-v^2)^2-c_1(1/K-1/K_c(\beta))\sigma
+c_2\sigma^2.
\end{equation}
Such a term was considered in
Refs.~\cite{Creutz:1996bg,Sharpe:1998xm}. The predicted phase
structure depends qualitatively on the sign of $c_2$, but a priori we
have no information on this.\footnote{Ref.~\cite{Damgaard:2010cz} has
argued that $c_2$ should be positive. We will return to this argument
a bit later in this section.} Indeed, as it is a lattice artifact, it
is expected that this sign might depend on the choice of gauge action.
Note that we could have added a term like $\vec\pi^2$, but this is
essentially equivalent since
$\vec\pi^2=(\sigma^2+\vec\pi^2)-\sigma^2$, and the first term here can
be absorbed, up to an irrelevant constant, into the starting Mexican
hat potential.
First consider the case when $c_2$ is less than zero, thus lowering
the potential energy when the field points in the positive or negative
sigma direction. This quadratic warping helps to stabilize the sigma
direction, as sketched in Fig.~\ref{cltzero}, and the pions cease to be
true Goldstone bosons when the quark mass vanishes. Instead, as the
mass passes through zero, we have a first order transition as the
expectation of $\sigma$ jumps from positive to negative. This jump
occurs without any physical particles becoming massless.
\begin{figure}
\centering
\includegraphics[width=3in]{c2lt0.eps}
\caption{
Lattice artifacts could quadratically warp the effective potential.
If this warping is downward in the sigma direction, the chiral
transition becomes first order without the pions becoming massless.
Figure taken from Ref.~\cite{Creutz:2007fe}.}
\label{cltzero}
\end{figure}
Things get a bit more complicated if $c_2>0$, as sketched in
Fig.~\ref{cgtzero}. In that case the chiral transition splits into
two second order transitions separated by a phase with an expectation
for the pion field, i.e. $\langle \vec\pi \rangle \ne 0$. The
behavior is directly analogous to that shown in
Fig.~\ref{condensate1}, the main difference being that now the two
quarks are degenerate. Since the pion field has odd parity and charge
conjugation as well as carries isospin, all of these symmetries are
spontaneously broken in the intermediate phase. As isospin is a
continuous group, this phase will exhibit Goldstone bosons. The
number of these is two, representing the two flavor generators
orthogonal to the direction of the expectation value. If higher order
terms do not change the order of the transitions, there will be a
third massless particle exactly at the transition endpoints. In this
way the theory reveals three massless pions exactly at the
transitions, as discussed by Aoki
\cite{Aoki:1983qi}. The intermediate phase is usually referred to as
the ``Aoki phase.'' Assuming this $c_2>0$ case, Fig.~\ref{kbeta3}
shows the qualitative phase diagram expected.
Note the similarity of this discussion to that leading to the phase
diagram in Fig.~\ref{phasediagram}. Indeed, lattice artifacts can
lead to the spontaneously broken CP region found there for the
$(m_1,m_2)$ plane to open up and remain present for degenerate quarks.
The Aoki phase is closely related to the possibility of CP violation
at $\Theta=\pi$ for unequal mass quarks. Note also that this
connection with the earlier continuum discussion shows that with an
odd number of flavors, the spontaneous breaking of parity is the
normal expectation whenever the hopping parameter exceeds its critical
value. Indeed, in this case the Aoki phase is less a lattice artifact
than a direct consequence of the CP violation expected in the
continuum theory at $\theta=\pi$.
\begin{figure}
\centering
\includegraphics[width=3in]{c2gt0.eps}
\caption{
If the lattice artifacts warping the potential upward in the sigma
direction, the chiral transition splits into two second order
transitions separated by a phase where the pion field has an
expectation value.
Figure taken from Ref.~\cite{Creutz:2007fe}.}
\label{cgtzero}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=3in]{kbeta3.eps}
\caption{The qualitative structure of the $\beta,K$ plane including
the possibility of an Aoki phase. }
\label{kbeta3}
\end{figure}
\subsection{Twisted mass}
\label{twisted}
The $c_2$ term breaks the equivalence of different chiral directions.
This means that physics will indeed depend on the angle $\theta$ if
one takes a mass term of the form in Eq.~(\ref{mrot}).
Consider complexifying the fermion mass in the usual way
\begin{equation}
m\overline\psi\psi\rightarrow m\overline\psi_L\psi_R
+m^*\overline\psi_R\psi_L
={\rm Re}\ m\ \overline\psi\psi+i{\rm Im}\ m\ \overline\psi\gamma_5\psi.
\end{equation}
The rotation of Eq.~(\ref{mrot}) is equivalent to giving the up and
down quark masses opposite phases
\begin{eqnarray}
m_u\rightarrow e^{+i\theta}m_u\cr
m_d\rightarrow e^{-i\theta}m_d.
\end{eqnarray}
Thus motivated, we can consider adding a new mass term to the lattice theory
\begin{equation}
\mu\ i\overline\psi \tau_3\gamma_5\psi \sim \mu\pi_3.
\end{equation}
This extends our effective potential to
\begin{equation}
V(\vec\pi,\sigma)=\lambda(\sigma^2+\vec\pi^2-v^2)^2-c_1(1/K-1/K_c(\beta))\sigma
+c_2\sigma^2-\mu \pi_3.
\end{equation}
The twisted mass represents the addition of a ``magnetic field''
conjugate to the order parameter for the Aoki phase.
There are a variety of motivations for adding such a term to our
lattice action \cite{Frezzotti:2003ni,Munster:2004wt}. Primary among
them is that $O(a)$ lattice artifacts can be arranged to cancel. With
two flavors of conventional Wilson fermions, these effects change sign
on going from positive to negative mass, and if we put all the mass
into the twisted term we are half way between. It should be noted
that this cancellation only occurs when all the mass comes from the
twisted term; for other combinations with a traditional mass term,
some lattice artifacts of $O(a)$ will survive. Also, although it
looks like we are putting phases into the quark masses, these cancel
between the two flavors. The resulting fermion determinant remains
positive and we are working at $\Theta=0$. Furthermore, the algorithm
is considerably simpler and faster than either overlap
\cite{Narayanan:1994gw,Neuberger:1997fp} or domain wall
\cite{Kaplan:1992bt,Furman:1994ky} fermions while avoiding the
diseases of staggered quarks \cite{Creutz:2007yg}. Another nice
feature of adding a twisted mass term is that it allows a better
understanding of the Aoki phase and shows how to continue around it.
Figures\ref{tm1} and \ref{tm2} show how this works for the case
$c_2>0$ and $c_2<0$, respectively.
\begin{figure}
\centering
\includegraphics[width=3in]{tm1.eps}
\caption{
Continuing around the Aoki phase with twisted mass. This sketch
considers the case $c_2>0$ where the parity broken phase extends over
a region along the kappa axis.
Figure taken from Ref.~\cite{Creutz:2007fe}.}
\label{tm1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=3in]{tm2.eps}
\caption{
As in Fig.~\ref{tm1}, but for the case $c_2<0$ so the chiral
transition on the kappa axis becomes first order.
Figure taken from Ref.~\cite{Creutz:2007fe}.}
\label{tm2}
\end{figure}
Of course some difficulties come along with these advantages. First,
one needs to know $K_c$. Indeed, with the Aoki phase present, the
definition of this quantity is a bit fuzzy.
And second, the mass needs to be larger than the
$c_2$ artifacts. Indeed, as Figs.~\ref{tm1} and \ref{tm2} suggest, if
it is not, then one is really studying the physics of the Aoki phase,
not the correct continuum limit. This also has implications for how
close to the continuum one must be to study this structure; in
particular, one must have $\beta$ large enough so the Aoki phase does
not extend into the doubler region.
The question of the sign of $c_2$ remains open. Simulations suggest
that the usual Aoki phase with $c_2>0$ is the situation with the
Wilson gauge action. Recently Ref.~\cite{Damgaard:2010cz} has pointed
out that with the twisted mass term present, all eigenvalues of the
product of gamma five with the Dirac operator will have non-zero
imaginary parts. Thus to have $c_2<0$, the phase transition at
non-vanishing twisted mass must occur where the fermion determinant
does not vanish on any configuration. This contrasts with the $c_2>0$
case where small eigenvalues of $D$ are expected in the vicinity of
the critical hopping. This at first sight makes $c_2<0$ seem somewhat
unnatural; however, this is not a proof since we saw in
Subsection \ref{phasesection} that phase transitions without small
eigenvalues of the Dirac operator do occur in the continuum theory for
two flavors with non-degenerate quarks.
This picture of the artifacts associated with Wilson fermions raises
some interesting questions. One concerns the three flavor theory. As
discussed previously, in this case a parity broken phase becomes
physical with negative mass. Indeed, three degenerate quarks of
negative mass represent QCD with a strong CP angle $\theta=\pi$, for
which spontaneous breaking of CP is expected. In some sense the Aoki
phase becomes physical. Also, with three flavors the twisting process
is not unique, with possible twists in the $\lambda_3$ or $\lambda_8$
directions. For example, using only $\lambda_3$ would suggest a
possible twisted mass of form $m_u\sim e^{2\pi i/3}$, $m_d\sim
e^{-2\pi i/3}$, $m_s\sim 1$. Whether there is an optimum twisting
procedure for three flavors is unclear.
Another special case is one flavor QCD \cite{Creutz:2006ts}. In this
situation the anomaly removes all chiral symmetry, and the quark
condensate loses meaning as an order parameter. The critical value of
kappa where the mass gap disappears is decoupled from the point of
zero physical quark mass. There is a parity broken phase, but it
occurs only at sufficiently negative mass. And from the point of view
of twisting the mass, without chiral symmetry there is nothing to
twist other than turning on the physical parameter $\Theta$.
\newpage
\Section{Lattice actions preserving chiral symmetry}
\subsection{The Nielsen-Ninomiya theorem}
As discussed some time ago \cite{Nielsen:1980rz}, the doubling issue
is closely tied to topology in momentum space. To see how this works,
let us first establish a gamma matrix convention
\begin{eqnarray}
&&\vec\gamma=\sigma_1\otimes \vec\sigma=
\pmatrix{0 & \vec\sigma\cr \vec\sigma & 0}\\
&&\gamma_0=\sigma_2\otimes I=
\pmatrix{0 & -i\cr i & 0}\\
&&\gamma_5=\sigma_3\otimes I=
\pmatrix{1 & 0\cr 0 & -1}.
\end{eqnarray}
Now suppose we have an anti-Hermitean Dirac operator $D$ that anti-commutes
with $\gamma_5$
\begin{equation}
D=-D^\dagger=-\gamma_5 D \gamma_5.
\end{equation}
Considering this quantity in momentum space, its most general form is
\begin{equation}
D(p)=\pmatrix{0&z(p) \cr -z^*(p)&0\cr}
\end{equation}
where $z(p)$ is a quaternion
\begin{equation}
z(p)=z_0(p)+i\vec\sigma\cdot\vec z(p).
\end{equation}
Thus we see that any chirally symmetric Dirac operator maps momentum
space onto the space of quaternions.
The Dirac equation is obtained by expanding the momentum space
operator around a zero, i.e. $D(p)\simeq i\slashchar p =
i\gamma_\mu p_\mu$. Now consider a three dimensional sphere embedded
in four-dimensional momentum-space and surrounding the zero with a
constant $D^2\sim p^2$. The above quaternion wraps non-trivially
about the the origin as we cover this sphere. Here is where the
topology comes in \cite{Nielsen:1980rz,Creutz:2007af}. Momentum space
is periodic over Brillouin zones. We must have $z(p)=z(p+2\pi n)$
where $n$ is an arbitrary four vector with integer components.
Because of this, we can restrict the momentum components to lie in the
range $-\pi<p_\mu\le\pi$, and we cannot have any non-trivial topology
on the surface of this zone. Any mapping associated with a zero in
$z(p)$ must unwrap somewhere else before we get to the surface.
Assuming $D(p)$ remains finite, any zero must be accompanied by
another wrapping in the opposite sense. Because of doubling, the 16
species with naive fermions split up into 8 zeros of each sense.
The above argument only tells us that a chiral lattice theory must
have an even number of species. The case of a minimal doubling with
only two species is in fact possible, although all methods presented
so far \cite{Misumi:2010ea} appear to involve a breaking of
hyper-cubic symmetry. This breaking is associated with the direction
between the zeros; this makes one direction special, although it might
be possible to avoid it by having the zeros form a symmetric lattice
using the periodicity of momentum space. This has not yet been
demonstrated.
In earlier sections we discussed how an odd number of flavors raised
some interesting issues; in particular the sign of the mass becomes
relevant. In spite of this, there seems to be no contradiction with
having, say, three light flavors in the continuum with a well defined
chiral limit. The above lattice argument, however, seems to indicate
troubles with maintaining an exact chiral symmetry with an odd number
of flavors. Whether this apparent conflict is serious is unclear.
One could always start with a multiple fermion theory and then, with
something like a Wilson term, give a few species masses while leaving
behind an odd number of massless fermions. This will involve some
parameter tuning, but presumably can give a reasonable chiral limit
for odd $N_f>2$. This does not obviate the fact emphasized earlier
that with only one flavor there must not be any remaining chiral
symmetry even in the continuum.
\subsection{Minimal doubling}
Several chiral lattice actions satisfying the minimal condition of
$N_f=2$ flavors are known. Some time ago Karsten
\cite{Karsten:1981gd} presented a simple form by inserting a factor of
$i\gamma_4$ into a Wilson like term for space-like hoppings. A slight
variation appeared in a discussion by Wilczek \cite{Wilczek:1987kw} a
few years later. More recently, a new four-dimensional action was
motivated by the analogy with two dimensional graphene
\cite{Creutz:2007af}. Since then numerous variations have been
presented \cite{Borici:2007kz, Bedaque:2008jm,Kimura:2009di,
Creutz:2010cz,Misumi:2010ea}.
The main potential advantage with these approaches lies in their
ultra-locality. They all involve only nearby neighbor hoppings for
the fermions. Thus they should be extremely fast in simulations while
still protecting masses from additive renormalization and helping
control mixing of operators with different chirality. The approach
also avoids the uncontrolled errors associated with the rooting
approximation discussed later
\cite{Creutz:2007yg,Creutz:2007rk,Creutz:2009kx,Creutz:2009zq}. On
the other hand, all minimally-doubled actions presented so far have
the above mentioned disadvantage of breaking lattice hyper-cubic
symmetry. With interactions, this will lead to the necessity of
renormalization counter-terms that also violate this symmetry
\cite{Capitani:2010nn}. The extent to which this will complicate
practical simulations remains to be investigated.
Minimally-doubled chiral fermions have the unusual property of a
single local field creating two distinct fermionic species. Here we
discuss a point-splitting method for separating the effects of the two
flavors which can be created by a single fermion field. For this we
will work with one specific form for the fermion action, but the method
should be easily extended to other minimally-doubled formulations.
We concentrate on a minimally-doubled fermion action which
is a slight generalization of those presented by Karsten
\cite{Karsten:1981gd} and Wilczek \cite{Wilczek:1987kw}. The fermion
term in the lattice action takes the form $\overline\psi D\psi$. For
free fermions we start in momentum space with
\begin{equation}
\label{pspace}
D(p)=
i\sum_{i=1}^3 \gamma_i \sin(p_i)
+{i\gamma_4\over \sin(\alpha)}\left(
\cos(\alpha)+3
-\sum_{\mu=1}^4\cos(p_\mu)
\right).
\end{equation}
This includes a Wilson like term for the space-like hoppings but
containing an extra factor of $i\gamma_4$. As a function of the
momentum $p_\mu$, the propagator $D^{-1}(p)$ has two poles, located at
$\vec p = 0$, $p_4=\pm \alpha$. Relative to the naive fermion action,
the other doublers have been given a large ``imaginary chemical
potential'' by the Wilson like term. The parameter $\alpha$ allows
adjusting the relative positions of the poles. The original
Karsten/Wilczek actions correspond to $\alpha=\pi/2$.
This action maintains one exact chiral symmetry, manifested in the
anti-commutation relation $[D,\gamma_5]_+=0$. The two species,
however, are not equivalent, but have opposite chirality. To see
this, expand the propagator around the two poles and observe that one
species, that corresponding to $p_4=+\alpha$, uses the usual gamma
matrices, but the second pole gives a proper Dirac behavior using
another set of matrices $\gamma_\mu^\prime=\Gamma^{-1}\gamma_\mu
\Gamma$. The Karsten/Wilczek formulation uses
$\Gamma=i\gamma_4\gamma_5$, although other minimally-doubled actions
may involve a different transformation. After this transformation
$\gamma_5^\prime=-\gamma_5$, showing that the two species rotate
oppositely under the exact chiral symmetry, and this symmetry should
be regarded as ``flavored.'' One can think of the physical chiral
symmetry as that generated in the continuum theory by
$\tau_3\gamma_5$.
It is straightforward to transform the momentum space action in
Eq.~(\ref{pspace}) to position space and insert gauge fields
$U_{ij}=U_{ji}^\dagger$ on the links connecting lattice sites.
Explicitly indicating the site indices, the Dirac operator becomes
\begin{eqnarray}
D_{ij}&=&
U_{ij}\sum_{\mu=1}^3 \gamma_i{\delta_{i,j+e_\mu}
-\delta_{i,j-e_\mu}\over
2}\cr
&+&
{i\gamma_4\over \sin(\alpha)}
\bigg(
(\cos(\alpha)+3)\delta_{ij}
-U_{ij} \sum_{\mu=1}^4{\delta_{i,j+e_\mu}
+\delta_{i,j-e_\mu}\over 2}
\bigg).
\end{eqnarray}
Again we see analogy with Wilson fermions \cite{Wilson:1975id} for the
space directions but augmented with an {$i\gamma_4$} inserted in the
Wilson term.
Perturbative calculations \cite{Capitani:2010nn} have shown that
interactions with the gauge fields can shift the relative positions of
the poles along the direction between them. In other words, the
parameter $\alpha$ receives an additive renormalization. Furthermore,
the form of the action treats the fourth direction differently than
the spatial coordinates, this is the breaking of hyper-cubic symmetry
mentioned above. There arise three possible new counter-terms for the
renormalization of the theory. First there is a possible
renormalization of the on-site contribution to the action proportional
to $i\overline\psi\gamma_4\psi$. This provides a handle on the shift
of the parameter $\alpha$. Secondly, the breaking of the hyper-cubic
symmetry indicates one may need to adjust the fermion ``speed of
light.'' This involves a combination of the above on-site term and
the strength of temporal hopping proportional to ${\delta_{i,j+e_4}
+\delta_{i,j-e_4}}$. Finally, the breaking of hyper-cubic symmetry
can feed back into the gluonic sector, suggesting a possible
counter-term of form $F_{4\mu}F_{4\mu}$ to maintain the gluon ``speed
of light.'' In lattice language, this corresponds to adjusting the
strength of time-like plaquettes relative to space-like ones.
Of these counter-terms, $i\overline\psi\gamma_4\psi$ is of dimension 3
and is probably the most essential. Quantum corrections induce the
dimension 4 terms, suggesting they may be small and could partially be
absorbed by accepting a lattice asymmetry. How difficult these
counter-terms are to control awaits simulations.
Note that all other dimension 3 counter-terms are forbidden by basic
symmetries. For example, chiral symmetry forbids $\overline\psi\psi$
and $i\overline\psi\gamma_5\psi$ terms, and spatial cubic symmetry
removes $\overline\psi\gamma_i\psi$,
$\overline\psi\gamma_i\gamma_5\psi$, and
$\overline\psi\sigma_{ij}\psi$ terms. Finally, commutation with
$\gamma_4$ plus space inversion eliminates $\overline\psi
\gamma_4\gamma_5\psi$.
The fundamental field $\psi$ can create either of the two species.
For a quantity that creates only one of them, it is natural to combine
fields on nearby sites in such a way as to cancel the other. In other
words, one can point split the fields to separate the poles which
occur at distinct ``bare momenta.'' For the free theory, one
construction that accomplishes this is to consider
\begin{eqnarray}
&&u(q)={1\over 2}\left(1+{\sin(q_4+\alpha)\over \sin(\alpha)}
\right)\psi(q+\alpha e_4)\cr
&&d(q)={1\over 2}\ {\Gamma}\left(1-{\sin(q_4-\alpha)\over \sin(\alpha)}
\right)\psi(q-\alpha e_4)
\end{eqnarray}
where $\Gamma=i\gamma_4\gamma_5$ for the Karsten/Wilczek formulation.
Here we have inserted factors containing zeros cancelling the undesired
pole. This construction is not unique, and specific details will
depend on the particular minimally-doubled action in use. The factor
of $\Gamma$ inserted in the $d$ quark field accounts for the fact that
the two species use different gamma matrices. This is required since
the chiral symmetry is flavored, corresponding to an effective minus
sign in $\gamma_5$ for one of the species.
It is now straightforward to proceed to position space and insert
gauge field factors to keep gauge transformation properties simple
\begin{eqnarray}
&&u_x
=
{1\over 2}{e^{i\alpha x_4}}\left(\psi_x+i\
{U_{x,x-e_4}\psi_{x-e_4}
-U_{x,x+e_4}\psi_{x+e_4}\over 2 \sin(\alpha)}\right)\cr
&&d_x
=
{1\over 2}{\Gamma e^{-i\alpha x_4}}\left(\psi_x-i\ {
U_{x,x-e_4}\psi_{x-e_4}-U_{x,x+e_4}\psi_{x+e_4}\over 2 \sin(\alpha)}\right).
\end{eqnarray}
The various additional phase factors serve to remove the oscillations
associated with the bare fields having their poles at non-zero
momentum.
Given the basic fields for the individual quarks, one can go on to
construct mesonic fields, which then also involve point splitting. To
keep the equations simpler, we now consider the case $\alpha=\pi/2$.
For example, the neutral pion field becomes
\begin{eqnarray}
&\pi_0(x)
={i\over 2}(\overline u_x\gamma_5 u_x-\overline d_x \gamma_5 d_x)=\cr
&{i\over 16}\bigg(4\overline\psi_x\gamma_5\psi_x
+\overline\psi_{x-e_4}\gamma_5\psi_{x-e_4}
+\overline\psi_{x+e_4}\gamma_5\psi_{x+e_4}
\cr&
-\overline\psi_{x+e_4}UU\gamma_5\psi_{x-e_4}
-\overline\psi_{x-e_4}UU\gamma_5\psi_{x+e_4}\bigg).
\end{eqnarray}
Note that this involves combinations of fields at sites separated
by either 0 or 2 lattice spacings. In contrast, the $\eta^\prime$ takes
the form
\begin{eqnarray}
&\eta^\prime(x)
={i\over 2}(\overline u_x \gamma_5 u_x+\overline d_x \gamma_5 d_x)=
\cr
&{1\over 8}\bigg(
\overline\psi_{x-e_4}U\gamma_5\psi_x
-\overline\psi_xU\gamma_5\psi_{x-e_4}
-\overline\psi_{x+e_4}U\gamma_5\psi_x
+\overline\psi_xU\gamma_5\psi_{x+e_4}\bigg)
\end{eqnarray}
where all terms connect even with odd parity sites. In a recent
paper, Tiburzi \cite {Tiburzi:2010bm} has discussed how the anomaly,
which gives the $\eta^\prime$ a mass of order $\Lambda_{qcd}$, can be
understood in terms of the necessary point splitting.
\subsection{Link variables}
Lattice gauge theory is closely tied to two of the above concepts; it
is a theory of phases and it exhibits an exact local symmetry. Indeed
it is directly defined in terms of group elements representing the
phases acquired by quarks as they hop around the lattice. The basic
variables are phases associated with each link of a four dimensional
space time lattice. For non-Abelian case, these variables become an
elements of the gauge group, i.e. $U_{ij}\in SU(3)$ for the strong
interactions. Here $i$ and $j$ denote the sites being conneted by the
link in question. We suppress the group indices to keep the notation
under control. These are three by three unitary matrices satisfying
\begin{equation}
U_{ij}=U_{ji}^{-1}= (U_{ji})^\dagger.
\end{equation}
The analogy with continuum vector fields $A_\mu$ is
\begin{equation}
U_{i,i+e_\mu} = e^{i a g_0 A_\mu}.
\end{equation}
Here $a$ represents the lattice spacing and $g_0$ is the bare
coupling considered at the scale of the cutoff.
In the continuum, a non-trivial gauge field arises when the curl (in a
four dimensional sense) of the potential is non zero. This in turn
means the phase factor around a small closed loop is not unity. The
smallest closed path in the lattice is a ``plaquette,'' or elementary
square. Consider the phase corresponding to one such
\begin{equation}
U_P=U_{12}U_{23}U_{34}U_{41}
\end{equation}
where sites 1 through 4 run around the square in question. In an
intuitive sense this measures the flux through this plaquette
$U_P\sim \exp(i a^2 g_0 F_{\mu,\nu})$. This motivates using this
quantity to define an action. For this,
look at the real part of the trace of $U_P$
\begin{equation}
{\rm Re Tr} U_P =
N - a^4 g_0^2\ {\rm Tr}\ F_{\mu\nu}F_{\mu\nu} + O(a^6).
\end{equation}
The overall added constant $N$ is physically irrelevant. This leads
directly to the Wilson gauge action
\begin{equation}
S(U)=-\sum_P {\rm Re Tr} U_P.
\end{equation}
Now we have our gauge variables and an action. To proceed we turn to
a path integral as an integral over all fields of the exponentiated
action. For a Lie group, there is a natural measure that we will
discuss shortly. Using this measure, the path integral is
\begin{equation}
Z=\int (dU) e^{-\beta S}
\end{equation}
where $(dU)$ denotes integration over all link variables. This leads
to the conventional continuum expression ${1\over 2}\int d^4x\
{\rm Tr}\ F_{\mu\nu}F_{\mu\nu}$ if we choose $\beta=2N/g_0^2$ for group $SU(N)$
and use the conventionally normalized bare coupling $g_0$.
Physical correlation functions are obtained from the path integral as
expectation values. Given an operator $B(U)$ which depends on the
link variables, we have
\begin{equation}
\langle B \rangle = {1\over Z}\int (dU) B(U) e^{-\beta S(U)}.
\end{equation}
Because of the gauge symmetry, discussed further later, this only
makes physical sense if $B$ is invariant under gauge transformations.
\subsection {Group Integration}
The above path integral involves integration over variables which are
elements of the gauge group. For this we use a natural measure with a
variety of nice properties. Given any function $f(g)$ of the group
elements $g\in G$, the Haar measure is constructed so as to be
invariant under ``translation'' by an arbitrary fixed element $g_1$ of
the group
\begin{equation}
\label{groupinvariance}
\int dg\ f(g)=\int dg\ f(g_1g).
\end{equation}
For a compact group, as for the $SU(N)$ relevant to QCD, this is
conventionally normalized so that $\int dg\ 1=1$. These simple
properties are enough for the measure to be uniquely determined.
An explicit representation for this integration measure is almost
never needed, but fairly straightforward to write down formally.
Suppose a general group element is parameterized by some variables
$\alpha_1, ... \alpha_n$. Considering here the case $SU(N)$, there
are $n=N^2-1$ such parameters. Then assume we know some region $R$ in
this parameter space that covers the group exactly once. Define the
$n$ dimensional fully antisymmetric tensor $\epsilon_{i_1,
\ldots i_n}$ such that, say, $\epsilon_{1,2,...n}=1$.
Now look at the integral
\begin{equation}
\label{measure}
I=A\int_R \{d\alpha\}\ f(g(\vec\alpha))\
\epsilon_{i_1, ...i_n} {\rm Tr}\left( (g^{-1}\partial_{i_1} g)
... (g^{-1}\partial_{i_n} g)\right).
\end{equation}
This has the required invariance properties of
Eq.~(\ref{groupinvariance}). The properties of a group imply there
should be a set of parameters $\alpha^\prime$ depending on $\alpha$
such that $g_1 g(\vec\alpha) = g(\vec\alpha^\prime)$. If we change
the integration variables from $\alpha$ to $\alpha^\prime$, then the
epsilon factor generates exactly the Jacobian needed for this variable
change. The normalization factor $A$ is fixed by the above condition
$\int dg\ 1 = 1.$ Once this is done, we have the invariant measure.
The above form for the measure will appear again when we discuss
topological issues for gauge fields in Section \ref{classical}.
Several interesting properties of the Haar measure are easily found.
If the group is compact, the left and right measures are equal
\begin{equation}
\int d_R g\ f(g)=\int d_R g\ f(gg_1)
=\int d_L g_1\ \int d_Rg_2\ f(g_2g_1)
=\int d_L g\ f(g).
\end{equation}
This also shows the measure is unique since any left invariant measure
could be used. (For a non-compact group the normalization can
differ.) A similar argument shows
\begin{equation}
\int dg\ f(g)=\int dg\ f(g^{-1}).
\end{equation}
For a discrete group, $\int dg$ is simply a sum over the elements.
For $U(1)=\{e^{i\theta}|0\le\theta<2\pi\}$ the measure is simply an
integral over the circle
\begin{equation}
\int dg\ f(g)=\int_0^{2\pi} {d\theta\over 2\pi} f(e^{i\theta}).
\end{equation}
For $SU(2)$, group elements take the form
\begin{equation}
g=\{a_0+i\vec a\cdot\vec\sigma | a_0^2+\vec a^2=1\}
\end{equation}
and the measure is
\begin{equation}
\int dg\ f(g)={1\over \pi^2}\int d^4a\ f(g)\delta(a^2-1).
\end{equation}
In particular, $SU(2)$ is a 3-sphere.
Some integrals are easily evaluated if we realize that group
integration picks out the ``singlet'' part of a function. Thus
\begin{equation}
\int dg R_{ab}(g) = 0
\end{equation}
where $R(g)$ is any irreducible matrix representation other than the
trivial one, $R=1$. For the group $SU(3)$ one can write
\begin{eqnarray}
&&\int dg\ {\rm Tr}g\ {\rm Tr}g^\dagger = 1\\
&&\int dg\ ({\rm Tr}g)^3 = 1
\end{eqnarray}
from the well known formulae $3\otimes\overline 3= 1\oplus 8$ and
$3\otimes 3\otimes 3= 1\oplus 8\oplus 8\oplus 10$.
A simple integral useful for the strong coupling expansion is
\begin{equation}
\int dg\ g_{ij}\ (g^\dagger)_{kl}=I_{ijkl}.
\end{equation}
The group invariance says we can multiply the indices arbitrarily by a
group element on the left or right. There is only one combination of
the indices that can survive for $SU(N)$
\begin{equation}
I_{ijkl}= \delta_{il} \delta_{jk}/N.
\end{equation}
The normalization here is fixed since tracing over $jk$
should give the identity matrix.
Another integral that has a fairly simple form is
\begin{equation}
\int dg\ g_{i_1j_1}\ g_{i_2j_2} \ldots g_{i_Nj_N}
={1\over N!}\epsilon_{i_1\ldots i_N}\epsilon_{j_1\ldots j_N}.
\end{equation}
This is useful for studying baryons in the strong coupling regime.
\subsection {Gauge invariance}
The action of lattice gauge theory has an exact local symmetry. If we
associate an arbitrary group element $g_i$ with each site $i$ of the
lattice, the action is unchanged if we replace
\begin{equation}
U_{ij}\rightarrow g_i^{-1} U_{ij} g_j.
\end{equation}
One consequence is that no link can have a vacuum expectation
value \cite{Elitzur:1975im}.
\begin{equation}
\langle U_{ij} \rangle= g_i^{-1} \langle U_{ij}\rangle g_j=0.
\end{equation}
Generalizing this, unless one does some sort of gauge fixing, the
correlation between any two separated $U$ matrices is zero. Indeed
many things familiar from perturbation theory often vanish without
gauge fixing, including such fundamental objects as quark and gluon
propagators!
An interesting consequence of gauge invariance is that we can forget
to integrate over a tree of links in calculating any gauge invariant
observable \cite{Creutz:1999zy}. An axial gauge represents fixing all
links pointing in a given direction.\footnote{Using a tree with small
highly-serrated leaves might be called a ``light comb gauge.''} Note
that this sort of gauge fixing allows the reduction of two dimensional
gauge theories to one dimensional spin models. To see this, pick the
tree to be a non-intersecting spiral of links starting at the origin
and extending out to the boundary. Links transverse to this spiral
interact exactly as a one dimensional system. This also shows that
two dimensional gauge theories are exactly solvable. Construct the
transfer matrix along this one dimensional system. The partition
function is the sum of the eigenvalues of this matrix each raised to
the power of the volume of the system.
The trace of any product of link variables around a closed loop is the
famous Wilson loop. These quantities are by construction gauge
invariant and are the natural observables in the lattice theory. The well
known criterion for confinement is whether the expectation of the
Wilson loop decreases exponentially in the loop area.
More general gauges can be introduced using an analogue of the
Fadeev-Popov factor. If $B(U)$ is gauge invariant, then
\begin{equation}
\langle B \rangle={1\over Z} \int d(U) e^{-S} B(U)
= {1\over Z} \int d(U) e^{-S} B(U) f(U)/\phi(U)
\end{equation}
where $f(U)$ is an arbitrary gauge fixing function and
\begin{equation}
\phi(U) = \int (dg) f(g_i^{-1} U_{ij} g_j)
\end{equation}
is the integral of the gauge fixing function $f$ over all gauges. A
possible gauge fixing scheme might be to ask that some function $h$ of
the links vanishes. In this case we could take $f=\delta(h)$ and then
$\phi=\int (dg) \delta(h)$. The integral of a delta function of
another function is generically a determinant $\phi=\det (\partial
g/\partial h)$. A determinant can generally be written as an integral
over a set of auxiliary ``ghost'' fields. Pursuing this yields the
usual Fadeev-Popov picture \cite{Faddeev:1967fc}.
Gauge fixing in the continuum raises several subtle issues if one
wishes to go beyond perturbation theory. Given some gauge fixing
condition $h=0$ and the corresponding $f=\delta(h)$, it is desirable
that this function vanish only once on any gauge orbit. Otherwise one
should correct for the over counting due to what are known of as
``Gribov copies'' \cite{Singer:1978dk}. This turns out to be
non-trivial with most perturbative gauges in practice, such as the
Coulomb or Landau gauge. One of the great virtues of the lattice
approach is that by not fixing the gauge, these issues are
sidestepped.
On the lattice gauge fixing is unnecessary and usually not done if one
only cares about measuring gauge invariant quantities such as Wilson
loops. But this does have the consequence that the basic lattice
fields are far from continuous. The correlation between link
variables at different locations vanishes. The locality of the gauge
symmetry literally means that there is an independent symmetry at each
space time point. If we consider a quark-antiquark pair located at
different positions, they transform under unrelated symmetries. Thus
concepts such as separating the potential between quarks into singlet and
octet parts are meaningless unless some gauge fixing is imposed.
\subsection {Numerical simulation}
Monte Carlo simulations of lattice gauge theory have come to dominate
the subject. We will introduce some of the basic algorithms in
Section \ref{montecarlo}. The idea is to use the analogy to
statistical mechanics to generate in a computer memory sets of gauge
configurations weighted by the exponentiated action of the path
integral. This is accomplished via a Markov chain of small weighted
changes to a stored system. Various extrapolations are required to
obtain continuum results; the lattice spacing needs to be taken to
zero and the lattice size to infinity. Also, such simulations become
increasingly difficult as the quark masses become small; thus,
extrapolations in the quark mass are generally necessary. It is not
the purpose of this review to cover these techniques; indeed, the
several books mentioned at the beginning of this section are readily
available. In addition, the proceedings of the annual Symposium on
Lattice Field Theory are available on-line for the latest results.
While confinement is natural in the strong coupling limit of the
lattice theory, we will shortly see that this is not the region of
direct physical interest. For this a continuum limit is necessary.
The coupling constant on the lattice represents a bare coupling
defined at a length scale given by the lattice spacing. Non-Abelian
gauge theories possess the property of asymptotic freedom, which means
that in the short distance limit the effective coupling goes to zero.
This remarkable phenomenon allows predictions for the observed scaling
behavior in deeply inelastic processes. The way quarks expose
themselves in high energy collisions was one of the original
motivations for a non-Abelian gauge theory of the strong interactions.
In addition to enabling perturbative calculations at high energies,
the consequences of asymptotic freedom are crucial for numerical
studies via the lattice approach. As the lattice spacing goes to
zero, the bare coupling must be taken to zero in a well determined
way. Because of asymptotic freedom, we know precisely how to adjust
our simulation parameters to take take the continuum limit!
In terms of the statistical analogy, the decreasing coupling takes us
away from high temperature and towards the low temperature regime.
Along the way a general statistical system might undergo dramatic
changes in structure if phase transitions are present. Such
qualitative shifts in the physical characteristics of a system can
only hamper the task of demonstrating confinement in the non-Abelian
theory. Early Monte Carlo studies of lattice gauge theory have
provided strong evidence that such troublesome transitions are avoided
in the standard four dimensional $SU(3)$ gauge theory of the nuclear
force \cite{Creutz:1980zw}.
Although the ultimate goal of lattice simulations is to provide a
quantitative understanding of continuum hadronic physics, along the
way many interesting phenomena arise which are peculiar to the
lattice. Non-trivial phase structure does occur in a variety of
models, some of which do not correspond to any continuum field theory.
We should remember that when the cutoff is still in place, the lattice
formulation is highly non-unique. One can always add additional terms
that vanish in the continuum limit. In this way spurious transitions
might be alternatively introduced or removed. Physical results
require going to the continuum limit.
\subsection {Order parameters}
Formally lattice gauge theory is like a classical statistical
mechanical spin system. The spins $U_{ij}$ are elements of a gauge
group $G$. They are located on the bonds of our lattice. Can this
system become ``ferromagnetic''? Indeed, as mentioned above, this is
impossible since $\langle U\rangle=0$ follows from the links
themselves not being gauge invariant \cite{Elitzur:1975im}.
But we do expect some sort of ordering to occur in the $U(1)$ theory.
If this is to describe physical photons, there should be a phase with
massless particles. Strong coupling expansions show that for large
coupling this theory has a mass gap \cite{Wilson:1974sk}. Thus a phase
transition is expected, and has been observed in numerical simulations
\cite{Creutz:1979zg}. Exactly how this ordering occurs remains
somewhat mysterious; indeed, although people often look for a
``mechanism for confinement,'' it might be interesting to rephrase
this question to ``how does a theory such as electromagnetism avoid
confinement.''
The standard order parameter for gauge theories and confinement
involves the Wilson loop mentioned above. This is the trace of the
product of link variables multiplied around a closed loop in
space-time. If the expectation of such a loop decreases exponentially
with the area of the loop, we say the theory obeys an area law and is
confining. On the other hand, a decrease only as the perimeter
indicates an unconfined theory. This order parameter by nature is
non-local; it cannot be measured without involving arbitrarily long
distance correlations. The lattice approach is well known to give the
area law in the strong coupling limit of the pure gauge theory.
Unfortunately, with dynamical quarks this ceases to be a useful
measure of confinement. As a loop becomes large, it will be screened
dynamically by quarks ``popping'' out of the vacuum. Thus we always
will have a perimeter law.
Another approach to understanding the confinement phase is to use the
mass gap. As long as the quarks themselves are massive, a confining
theory should contain no physical massless particles. All mesons,
glueballs, and nucleons are expected to gain masses through the
dimensional transmutation phenomenon discussed later. As with the
area law, the presence of a mass gap is easily demonstrated for the
strong coupling limit of the pure glue theory.
If the quarks are massless, this definition also becomes a bit tricky.
In this case we expect spontaneous breaking of chiral symmetry, also
discussed extensively later. This gives rise to pions as massless
Goldstone bosons. To distinguish this situation from the unconfined
theory, one could consider the number of massless particles in the
spectrum by looking at how the ``vacuum'' energy depends on
temperature using the Stefan-Boltzmann law. With $N_f$ flavors we
have $N_f^2-1$ massless scalar Goldstone bosons. On the other hand,
were the gauge group $SU(N)$ not to confine, we would expect $N^2-1$
massless vector gauge bosons plus $N_f$ massless quarks, all of which
have two degrees of freedom.
\subsection{Domain wall and overlap fermions}
The overlap fermion was originally developed \cite{Narayanan:1993sk}
as a limit of a fermion formulation using four dimensional surface
modes on a five dimensional lattice. This effectively amounts to
using Shockley surface states as the basis for a theory maintaining
chiral symmetry \cite{Kaplan:1992bt}. For a review see
Ref.~\cite{Jansen:1994ym}. The idea is to set up a theory in one
extra dimension so that surface modes exist, and our observed world is
an interface with our quarks and leptons being these surface modes.
Particle hole symmetry naturally gives the basic fermions zero mass.
In the continuum limit the extra dimension becomes unobservable due to
states in the interior requiring a large energy to create. In this
picture, opposing surfaces carry states of opposite helicity, and the
anomalies are due to a tunnelling through the extra dimension.
Ref. \cite{Creutz:1994ny} discussed the general conditions for surface
modes to exist. Normalized solutions are bound to any interface
separating a region with supercritical from sub-critical hopping.
Kaplan's original paper \cite{Kaplan:1992bt} considered not a surface,
but an interface with $M=M_{cr}+m \epsilon(x)$, where $M_{cr}$ is the
critical value for the mass parameter where the five dimensional
fermions would be massless. Shamir \cite{Shamir:1993zy} presented a
somewhat simpler picture where the hopping vanishes on one side, which
then drops out of the problem and we have a surface.
To couple gluon fields to this theory without adding unneeded degrees
of freedom, the gauge fields are taken to lie in the four physical
space-time directions and be independent of the fifth coordinate. In
this approach, the extra dimension is perhaps best thought of as a
flavor space \cite{Narayanan:1992wx}. With a finite lattice this
procedure gives equal couplings of the gauge field to the fermion
modes on opposing walls in the extra dimension. Since the left and
right handed modes are separated by the extra dimension, they only
couple through the gauge field. The result is an effective light
Dirac fermion. In the case of the strong interactions, this provides
an elegant scheme for a natural chiral symmetry without the tuning
inherent in the Wilson approach. The breaking of chiral symmetry
arises only through finiteness of the extra dimension.\footnote{The
anomaly, however, shows that some communication between the surfaces
survives even as the extra dimension becomes infinite. This is
possible since the same gauge fields are on each surface.}
The name ``overlap operator'' comes from the overlap of eigenstates of
the different five dimensional transfer matrices on each side of the
interface. Although originally derived from the infinite limit of the
five dimensional formalism, one can formulate the overlap operator
directly in four dimensions. We begin with the fermionic part of some
generic action as a quadratic form
$
S_f= \overline\psi D \psi.
$
The usual ``continuum'' Dirac operator $D=\sum\gamma_\mu D_\mu$
naively anti-commutes with $\gamma_5$, i.e. $[\gamma_5, D]_+=0$. Then
the change of variables $\psi \rightarrow e^{i\theta\gamma_5} \psi$
and $\overline\psi \rightarrow \overline\psi e^{i\theta\gamma_5}$
would be a symmetry of the action. This, however, is inconsistent
with the chiral anomalies. The conventional continuum discussion
presented earlier maps this phenomenon into the fermionic measure
\cite{Fujikawa:1979ay}.
On the lattice we work with a finite number of degrees of freedom;
thus, the above variable change is automatically a symmetry of the
measure. To parallel the continuum discussion, it is necessary to
modify the symmetry transformation on the action so that the measure
is no longer invariant. Remarkably, it is possible to construct a
modified symmetry under which corresponding actions are exactly
invariant.
To be specific, one particular variation \cite{Neuberger:1997fp,
Neuberger:1998my,Neuberger:1998wv,Chiu:1998gp,Chandrasekharan:1998wg}
modifies the change of variables to
\begin{eqnarray}
&&\psi \longrightarrow
e^{i\theta\gamma_5}
\psi\cr
&&\overline\psi \longrightarrow \overline\psi
e^{i\theta(1-aD)\gamma_5}
\end{eqnarray}
where $a$ represents the lattice spacing. Note the asymmetric way in
which the independent Grassmann variables $\psi$ and $\overline\psi$
are treated. Requiring the action to be unchanged gives the relation
\cite{Ginsparg:1981bj,Hasenfratz:1998ri,Hasenfratz:1998jp}.
\begin{equation}
\label{relation}
D \gamma_5 = -\gamma_5 D+a D\gamma_5 D=
-\hat\gamma_5 D
\end{equation}
with $\hat\gamma_5=(1-aD)\gamma_5$. To proceed, we also assume the
Hermeticity condition $\gamma_5 D \gamma_5 = D^\dagger$. We see that
the naive anticommutation relation receives a correction of order the
lattice spacing. The above ``Ginsparg-Wilson relation'' along with
the Hermeticity condition is equivalent to the unitarity of the
combination $V=1-aD$.
Neuberger \cite{Neuberger:1998my,Neuberger:1998wv} and Chiu and Zenkin
\cite{Chiu:1998gp} presented an explicit operator with the above
properties. They first construct $V$ via a unitarization of an
undoubled chiral violating Dirac operator, such as the Wilson operator
$D_w$. This operator should also satisfy the above Hermeticity
condition $\gamma_5 D_w \gamma_5 = D_w^\dagger$. Specifically, they
consider
\begin{equation}
\label{projection}
V=-D_w(D_w^\dagger D_w)^{-1/2}.
\end{equation}
The combination $(D_w^\dagger D_w)^{-1/2}$ is formally defined by
finding a unitary operator to diagonalize the Hermitean combination
$D_w^\dagger D_w$, taking the square root of the eigenvalues, and then
undoing the unitary transformation.
Directly from $V$ we construct the overlap operator as
\begin{equation}
D=(1-V)/a.
\end{equation}
The Ginsparg-Wilson relation of Eq.~(\ref{relation}) is most
succinctly written as the unitarity of $V$ coupled with its $\gamma_5$
Hermeticity
\begin{equation}
\gamma_5 V\gamma_5 V=1.
\end{equation}
The basic projection process is illustrated in Fig.~\ref{eigen3}.
\begin{figure}
\centering
\includegraphics[width=4in]{eigen3.eps}
\caption{
\label{eigen3}
The overlap operator is constructed by projecting the eigenvalues of
the Wilson operator onto a circle. Figure taken from
Ref.~\cite{Creutz:2002qa}. }
\end{figure}
The overlap operator has several nice properties. Being constructed
from a unitary operator, the normality of $D$ is guaranteed. But,
most important, it exhibits a lattice version of an exact chiral
symmetry.\cite{Luscher:1998pqa} The fermionic action $\overline\psi
D\psi$ is invariant under the transformation
\begin{eqnarray}
&\psi\rightarrow e^{i\theta\gamma_5}\psi\cr
&\overline\psi\rightarrow \overline\psi e^{i\theta\hat\gamma_5}
\label{symmetry}
\end{eqnarray}
where
\begin{equation}
\hat\gamma_5 =V\gamma_5.
\end{equation}
As with $\gamma_5$, this quantity which appeared in
Eq.~(\ref{relation}) is Hermitean and its square is unity.
Thus its eigenvalues are all plus or minus unity. The trace
defines an index
\begin{equation}
\label{tracehat}
\nu={1\over 2}{\rm Tr}\hat\gamma_5
\end{equation}
which plays exactly the role of the index in the continuum. If the
gauge fields are smooth, this counts the topology of the gauge
configuration. The factor of $1/2$ in Eq.~(\ref{tracehat}) appears
because the exact zero modes of the overlap operator have partners on
the opposite side of the unitarity circle that also contribute to the
trace.
At this point the hopping parameter in $D_w$ is a parameter. To have
the desired single light fermion per flavor of the theory, the hopping
parameter should be appropriately adjusted to lie above the critical
value where $D_w$ describes a massless flavor, but not so large that
additional doublers come into play \cite{Neuberger:1999jw}. There are
actually two parameters to play with, the hopping parameter of $D_w$,
and the lattice spacing. When the latter is finite and gauge fields
are present, the location of the critical hopping parameter in $D_w$
is expected to shift from that of the free fermion theory. As we saw
when discussing the Aoki phase, there is potentially a rather complex
phase structure in the plane of these two parameters, with various
numbers of doublers becoming massless as the hopping is varied. The
Ginsparg-Wilson relation in and of itself does not in general
determine the number of physical massless fermions.
Although the Wilson operator entering this construction is local and
quite sparse, the resulting overlap action is not. Because of the
inversion in Eq.~(\ref{projection}), it involves direct couplings
between arbitrarily separated sites
\cite{Hernandez:1998et,Horvath:1998cm,Horvath:1999bk}. How rapidly
these couplings fall with distance depends on the gauge fields and is
not fully understood. The five dimensional domain-wall theory is
local in the most naive sense; all terms in the action only couple
nearest neighbor sites. However, were one to integrate out the heavy
modes, the resulting low energy effective theory would also involve
couplings with arbitrary range. Despite these non-localities,
encouraging studies
\cite{Neuberger:1998wv,Edwards:1998wx,Borici:1999ws,
Dong:2000mr,Gattringer:2000js} show that it may indeed be practical
to implement the required inversion in large scale numerical
simulations. The overlap operator should have memory advantages over
the domain wall approach since a large number of fields corresponding
to the extra dimension do not need to be stored.
The overlap approach hides the infinite sea of heavy fermion states in
the extra dimension of the domain wall approach. This tends to
obscure the possible presence of singularities in the required
inversion of the Wilson kernel.
Detailed analysis
\cite{Luscher:2000zd,Kikukawa:1998pd} shows that this operator is
particularly well behaved order by order in perturbation theory. This
has led to hopes that this may eventually lead to a rigorous
formulation of chiral models, such as the standard model.
Despite being the most elegant known way to have an exact remnant of
chiral symmetry on the lattice, the overlap operator raises several
issues. These complications probably become insignificant as the
continuum limit is approached, but should be kept in mind given the
high computational cost of this approach. To begin with, the overlap
is highly non-unique. It explicitly depends on the kernel being
projected onto the unitary circle. Even after choosing the Wilson
kernel, there is a dependence on the input mass parameter. One might
want to define topology in terms of the number of exact zero modes of
the overlap operator. However the non-uniqueness leaves open the
question of whether the winding number of a gauge configuration might
depend on this choice. Later we will return to the question of
possible ambiguities in defining topological susceptibility in the
continuum limit.
In this connection, it is possible to make a bad choice for the mass
parameter. In particular, if it is chosen below the continuum kappa
critical value of $1/8$, no low modes will survive. This is true
despite the fact that the corresponding operator will still satisfy
the Ginsparg-Wilson condition. This explicitly shows that just
satisfying the Ginsparg-Wilson condition is not a sufficient condition
for a chiral theory. Conversely, if one chooses the mass parameter
too far in the supercritical region, additional low modes will be
produced from the doublers. As mentioned earlier, the Ginsparg-Wilson
condition does not immediately determine the number of flavors in the
theory.
Another issue concerns the one flavor case, discussed earlier.
Because of the anomaly, this theory is not supposed to show any chiral
symmetry and has no Goldstone bosons. Nevertheless, one can construct
the overlap operator and it will satisfy the Ginsparg-Wilson
condition. This shows that the consequences of this condition are
weaker than for the usual continuum chiral symmetry. With a
conventional chiral symmetry, the spectrum cannot show a gap. Either
we have the Goldstone bosons of spontaneous chiral breaking or we have
massless fermions \cite{'tHooft:1979bh}.
It should also be noted that the overlap behaves peculiarly for
fermions in higher representations than the fundamental. As we
discussed earlier, the number of zero modes associated with a
non-trivial topology in the continuum theory depends on the fermion
representation being considered. It has been observed in numerical
simulations that the appropriate multiplicity is not always seen for
the overlap operator constructed on rough gauge configurations
\cite{Edwards:1998dj}.
As a final comment, note that these actions preserving a chiral
symmetry all involve some amount of non-locality. With minimal
doubling this has a finite range, but is crucial for allowing the
anomaly to work out properly. An important consequence is that the
operator product expansion, a standard perturbative tool, must involve
operators with a similar non-locality. The ambiguities in defining
non-degenerate quark masses lie in these details.
\subsection{Fermionic transfer matrices}
The concept of continuity is lost with Grassman variables. There is
no meaning to saying that fermion fields at nearby sites are near
each other. This is closely tied to the doubling issues that we will
discuss later. But is also raises interesting complications in
relating Hamiltonian quantum mechanics with the Euclidian formulation
involving path integrals. Here we will go into how this connection is
made with an extremely simple zero space-dimensional model.
Anti-commutation is at the heart of fermionic behavior. This is true
in both the Hamiltonian operator formalism and the Lagrangean path
integral, but in rather complementary ways. Starting with a
Hamiltonian approach, if an operator $a^\dagger$ creates a fermion in
some normalized state on the lattice or the continuum, it satisfies
the basic relation
\begin{equation}
[a,a^\dagger]_+\equiv a a^\dagger + a^\dagger a =1.
\end{equation}
This contrasts sharply with the fields in a path integral, which all
anti-commute
\begin{equation}
[\chi,\chi^\dagger]_+=0.
\end{equation}
The connection between the Hilbert space approach and the path
integral appears through the transfer matrix formalism. For bosonic
fields this is straightforward \cite{Creutz:1976ch}, but for fermions
certain subtleties arise related to the doubling issue \cite
{Creutz:1999zy}.
To be more precise, consider a single fermion state created by the
operator $a^\dagger$, and an antiparticle state created by another
operator $b^\dagger$. For an extremely simple model, consider the
Hamiltonian
\begin{equation}
H=m (a^\dagger a+b^\dagger b)
\end{equation}
Here $m$ can be thought of as a ``mass'' for the particle.
What we want is an exact path
integral expression for the partition function
\begin{equation}
Z={\rm Tr} e^{-\beta H}.
\end{equation}
Of course, since the Hilbert space generated by $a$ and $b$ has only
four states, this is trivial to work out: $
Z=1+2e^{-\beta m}+e^{-2\beta m} $. However, we want this in a form
that easily generalizes to many variables.
The path integral for fermions uses Grassmann variables. We introduce
a pair of such, $\chi$ and $\chi^\dagger$, which will be connected to
the operator pair $a$ and $a^\dagger$, and another pair, $\xi$ and
$\xi^\dagger$, for $b$, $b^\dagger$. All the Grassmann variables
anti-commute. Integration over any of them is determined by the
simple formulas mentioned earlier
\begin{equation}
\int d\chi \ 1 = 0\ ; \qquad \int d\chi\ \chi = 1.
\end{equation}
For notational simplicity combine the individual Grassmann variables
into spinors
\begin{equation}\matrix{
\psi = \pmatrix{\chi\cr\xi^\dagger\cr};&
\psi^\dagger = \pmatrix{\chi^\dagger&\xi\cr}.\cr
}
\end{equation}
To make things appear still more familiar, introduce a ``Dirac
matrix''
\begin{equation}
\gamma_0=\pmatrix{1&0\cr 0 & -1\cr}
\end{equation}
and the usual
\begin{equation}
\overline\psi=\psi^\dagger\gamma_0.
\end{equation}
Then we have
\begin{equation}
\overline \psi \psi = \chi^\dagger \chi + \xi^\dagger \xi.
\end{equation}
where the minus sign from using $\xi^\dagger$ rather than $\xi$ in
defining $\psi$ is removed by the $\gamma_0$ factor. The temporal
projection operators
\begin{equation}
P_\pm={1\over 2}(1\pm\gamma_0)
\end{equation}
arise when one considers the fields at
two different locations
\begin{equation}
\chi_i^\dagger \chi_j +\xi_i^\dagger \xi_j =
\overline \psi_i P_+\psi_j+\overline \psi_j P_-\psi_i.
\end{equation}
The indices $i$ and $j$ will soon label the ends of a temporal hopping
term; this formula is the basic transfer matrix justification for the
Wilson projection operator formalism that we will return to in later
sections.
\subsection{Normal ordering and path integrals}
For a moment ignore the antiparticles and consider some general
operator $f(a,a^\dagger)$ in the Hilbert space. How is this
related to an integration in Grassmann space? To proceed we need a
convention for ordering the operators in $f$. We adopt the usual
normal ordering definition with the notation $:f(a,a^\dagger):$
meaning that creation operators are placed to the left of destruction
operators, with a minus sign inserted for each exchange. In this case
a rather simple formula gives the trace of the operator as a Grassmann
integration
\begin{equation}
\label{dchi}
{\rm Tr}\ :f(a,a^\dagger):\ = \int d\chi d\chi^\dagger
e^{2\chi^\dagger\chi} f(\chi,\chi^\dagger).
\end{equation}
To verify, just check that all elements of the complete set of
operators $\{1,a,a^\dagger,a^\dagger a\}$ work. However, this formula
is actually much more general; given a set of many Grassmann variables
with one pair associated with each of several fermion states, this
immediately generalizes to the trace of any normal ordered operator
acting in a many fermion Hilbert space.
What about a product of several normal ordered operators? This leads
to the introduction of multiple sets of Grassmann variables and the
general formula
\begin{eqnarray}
&{\rm Tr}\ &
\left(:f_1(a^\dagger,a):\ :f_2(a^\dagger,a): \ldots :f_n(a^\dagger,a):
\right)\cr
&=&\int d\chi_1\ d\chi_1^*\ldots d\chi_n\ d\chi_n^*
\ e^{\chi_1^*(\chi_1+\chi_n)} e^{\chi_2^*(\chi_2-\chi_1)}
\ldots e^{\chi_n^*(\chi_n-\chi_{n-1})} \cr
&&\qquad \times f_1(\chi_1^*,\chi_1)f_2(\chi_2^*,\chi_2)
\ldots f_n(\chi_n^*,\chi_n).
\end{eqnarray}
The positive sign on $\chi_n$ in the first exponential factor
indicates the natural occurrence of anti-periodic boundary conditions;
{\i.e.} we can define $x_0=-x_n$. With just one factor, this formula
reduces to Eq.~(\ref{dchi}). Note how the ``time derivative'' terms
are ``one sided;'' this is how doubling is eluded.
This exact relationship provides the starting place for converting our
partition function into a path integral. The simplicity of our
example Hamiltonian allows this to be done exactly at every stage.
First we break ``time'' into a number $N$ of ``slices''
\begin{equation}
Z={\rm Tr} \left( e^{-\beta H/N}\right)^N.
\end{equation}
Now we need normal ordered factors for the above formula. For this we
use
\begin{equation}
e^{\alpha a^\dagger a} = 1+(e^\alpha-1) a^\dagger a
=\ :e^{(e^\alpha-1) a^\dagger a }:\ ,
\end{equation}
which is true for arbitrary $\alpha$.\footnote{The definition of
normal ordering gives
$:(a^\dagger a)^2:=0.$}
This is all the machinery we need to write
\begin{equation}
Z=\int (d\psi d\overline\psi) e^{S}
\end{equation}
where
\begin{equation}
S=\sum_{i=1}^n
\overline\psi_n
(e^{-\beta m/N}-1)
\psi_n
+\overline\psi_n P_+ \psi_{n-1}
+\overline\psi_{n-1} P_ - \psi_{n}.
\end{equation}
Note how the projection factors of $P_\pm$ automatically appear for
handling the reverse convention of $\chi$ versus $\xi$ in our field
$\psi$. Expanding the first term gives the $-\beta m/N$ factor
appearing in the Hamiltonian form for the partition function.
It is important to realize that if we consider the action as a
generalized matrix connecting fermionic variables
\begin{equation}
S=\overline\psi M \psi,
\end{equation}
the matrix $M$ is not symmetric. The upper components propagate
forward in time, and the lower components backward. Even though our
Hamiltonian was Hermitean, the matrix appearing in the corresponding
action is not. With further interactions, such as gauge field
effects, the intermediate fermion contributions to a general path
integral may not be positive, or even real. Of course the final
partition function, being a trace of a positive definite operator, is
positive. Keeping the symmetry between particles and antiparticles
results in a real fermion determinant, which in turn is positive for
an even number of flavors. We will later see that some rather
interesting things can happen with an odd number of flavors.
For our simple Hamiltonian, this discussion has been exact. The
discretization of time adds no approximations since we could do the
normal ordering by hand. In general with spatial hopping or more
complex interactions, the normal ordering can produce extra terms
going as $O(1/N^2)$. In this case exact results require a limit of a
large number of time slices, but this is a limit we need anyway to
reach continuum physics.
\subsection{General masses in two-flavor QCD}
Given the confusion over the meaning of quark masses, it is
interesting to explore how two-flavor QCD behaves as these quantities
are varied, including the possibility of explicit CP violation through
the Theta parameter. The full theory has a rather rich phase diagram,
including first and second order phase transitions, some occuring when
none of the quark masses vanish.
We consider the quark fields $\psi$ as carrying implicit isospin,
color, and flavor indices. Assume as usual that the theory in the
massless limit maintains the $SU(2)$ flavored chiral symmetry under
\begin{eqnarray}
\psi \longrightarrow e^{i\gamma_5 \tau_\alpha\phi_\alpha/2}\psi\cr
\overline \psi \longrightarrow \overline\psi
e^{i\gamma_5 \tau_\alpha\phi_\alpha/2}.
\end{eqnarray}
Here $\tau_\alpha$ represents the Pauli matrices generating isospin
rotations. The angles $\phi_\alpha$ are arbitrary rotation
parameters.
We wish to construct the most general two-flavor mass term to add to
the massless Lagrangean. Such should be a dimension 3 quadratic form
in the fermion fields and should transform as a singlet under Lorentz
transformations. For simplicity, only consider quantities that are
charge neutral as well. This leaves four candidate fields, giving the
general form for consideration
\begin{equation}
m_1\overline\psi\psi+
m_2\overline\psi\tau_3\psi+
im_3\overline\psi\gamma_5\psi+
im_4\overline\psi\gamma_5\tau_3\psi.
\label{genmass}
\end{equation}
The first two terms are naturally interpreted as the average quark
mass and the quark mass difference, respectively. The remaining two
are less conventional. The $m_3$ term is connected with the CP
violating parameter of the theory. The final $m_4$ term has been used
in conjunction with the Wilson discretization of lattice fermions,
where it is referred to as a ``twisted mass''
\cite{Frezzotti:2000nk,Boucaud:2007uk}. Its utility in this context is
the ability to reduce lattice discretization errors. We will return
to this term later when we discuss the effect of lattice artifacts on
chiral symmetry.
These four terms are not independent. Indeed, consider the above
flavored chiral rotation in the $\tau_3$ direction, $\psi\rightarrow
e^{i\theta\tau_3\gamma_5/2}\psi$. Under this the composite fields
transform as
\begin{eqnarray}
\overline\psi\psi\ &\longrightarrow\
\cos(\theta)\overline\psi\psi
+\sin(\theta)i\overline\psi\gamma_5\tau_3\psi\cr
\overline\psi\tau_3\psi\ &\longrightarrow\
\cos(\theta)\overline\psi\tau_3\psi
+\sin(\theta)i\overline\psi\gamma_5\psi\cr
i\overline\psi\tau_3\gamma_5\psi\ &\longrightarrow\
\cos(\theta)i\overline\psi\tau_3\gamma_5\psi
-\sin(\theta)\overline\psi\psi\cr
i\overline\psi\gamma_5\psi\ &\longrightarrow\
\cos(\theta)i\overline\psi\gamma_5\psi
-\sin(\theta)\overline\psi\tau_3\psi.
\end{eqnarray}
This rotation mixes $m_1$ with $m_4$ and $m_2$ with $m_3$. Using this
freedom, we can select any one of the $m_i$ to vanish and a second to
be positive.
The most common choice is to set $m_4=0$ and use $m_1$ as controlling
the average quark mass. Then $m_2$ gives the quark mass difference,
while CP violation appears in $m_3$. This, however, is only a
convention. The alternative ``twisted mass'' scheme
\cite{Frezzotti:2000nk,Boucaud:2007uk}, makes the choice $m_1=0$.
This uses {$m_4>0$} for the average quark mass and {$m_3$} becomes the
up-down mass difference. In this case $m_2$ becomes the CP violating
term. It is amusing to note that an up down quark mass difference in
such a formulation involves the naively CP odd
$i\overline\psi\gamma_5\psi$. The strong CP problem has been rotated
into the smallness of the $\overline\psi\tau_3\psi$ term, which with
the usual conventions is the mass difference. But because of the
flavored chiral symmetry, both sets of conventions are physically
equivalent.
For the following, take the arbitrary choice $m_4=0$, although one
should remember that this is only a convention and we could have
chosen any of the four parameters in Eq.~(\ref{genmass}) to vanish.
With this choice, two-flavor QCD, after scale setting, depends on
three mass parameters
\begin{equation}
m_1\overline\psi\psi+
m_2\overline\psi\tau_3\psi+
im_3\overline\psi\gamma_5\psi.
\end{equation}
It is the possible presence of $m_3$ that represents the strong CP
problem. As all the parameters are independent and transform
differently under the symmetries of the problem, there is no
connection between the strong CP problem and $m_1$ or $m_2$.
As discussed extensively above, the chiral anomaly is responsible for
the iso-singlet rotation
\begin{eqnarray}
\psi \longrightarrow e^{i\gamma_5 \phi/2}\psi\cr
\overline \psi \longrightarrow \overline\psi
e^{i\gamma_5 \phi/2}
\end{eqnarray}
not being a valid symmetry, despite the fact that $\gamma_5$ naively
anti-commutes with the massless Dirac operator.
Subection \ref{quarks} showed this anomaly is nicely summarized via
Fujikawa's \cite{Fujikawa:1979ay} approach where the fermion measure
in the path integral picks up a non-trivial factor. In any given gauge
configuration only the zero eigenmodes of $\slashchar D$ contribute,
and by the index theorem they are connected to the winding number of
the gauge configuration. The conclusion is that the above rotation
changes the fermion measure by an amount depending non-trivially on
the gauge field configuration.
Note that this anomalous rotation allows one to remove any topological
term from the gauge part of the action. Naively this would have been
yet another parameter for the theory, but by including all three mass
terms for the fermions, this can be absorbed. For the following we
consider that any topological term has thus been rotated away. After
this one is left with the three mass parameters above, all of which
are independent and relevant to physics.
These parameters are a complete set for two-flavor QCD; however, this
choice differs somewhat from what is often discussed. Formally we can
define the more conventional variables as
\begin{eqnarray}
&&m_u=m_1+m_2+im_3\cr
&&m_d=m_1-m_2+im_3\cr
&&e^{i\Theta}={m_1^2-m_2^2-m_3^2+2im_1m_3\over
\sqrt{m_1^4+m_2^4+m_3^4+2m_1^2m_3^2+2m_2^2m_3^2-2m_1^2m_2^2}.
}
\end{eqnarray}
Particularly for $\Theta$, this is a rather complicated change of
variables. For non-degenerate quarks in the context of the phase
diagram discussed below, the variables $\{m_1,m_2,m_3\}$ are more
natural.
\subsection {The strong CP problem and the up quark mass}
Strong interactions preserve CP to high accuracy \cite{Baluni:1978rf}.
Thus only two of the three possible mass parameters seem to be needed.
With the above conventions, it is natural to ask why is $m_3$ so
small? It is the concept of unification that brings this question to
the fore. We know that the weak interactions violate CP. Thus, if
the electroweak and the strong interactions separate at some high
scale, shouldn't some remnant of this breaking survive? How is CP
recovered for the strong force?
One possible solution is that there is no unification and one should
just consider the weak interactions as a small perturbation. Another
approach involves adding a new dynamical ``axion'' field that couples
to the quarks through a coupling to $i\overline\psi\gamma_5\psi$.
Shifts in this field make $m_3$ essentially dynamical, and potentially
the theory could relax to $m_3=0$.
There is a third proposed solution, being criticized here, that the up
quark mass might vanish. This would naively allow a flavored chiral
rotation to remove any phases from the quark mass matrix. Why is a
vanishing up quark mass not a sensible approach? From the above, one
can define the up quark mass as a complex number
\begin{equation}
m_u\equiv m_1+m_2+im_3.
\end{equation}
But the quantities $m_1$, $m_2$, and $m_3$ are independent parameters
with different symmetry properties. With our conventions, $m_1$
represents an iso-singlet mass contribution, $m_2$ is isovector in
nature, and $m_3$ is CP violating. And, as discussed earlier, the
combination $m_1+m_2=0$ is scale and scheme dependent. The strong CP
problem only requires small $m_3$. So while it may be true formally
that
\begin{equation}
m_1+m_2+im_3=0\ \Rightarrow\ m_3=0,
\end{equation}
this would depend on scale and might well be regarded as ``not even
wrong.''
\subsection {The two-flavor phase diagram}
\label{phasesection}
\begin{figure}
\centering
{\includegraphics[width=2.5in] {warp.eps}}
\caption{ The $m_2$ and $m_3$ terms warp the Mexican hat potential
into two separate minima. The direction of the warping is
determined by the relative size of these parameters. Figure taken
from Ref.~\cite{Creutz:2010ts}.}
\label{warping}
\end{figure}
As a function of the three mass parameters, QCD has a rather intricate
phase diagram that we now discuss. Using simple chiral Lagrangean
arguments, this can be qualitatively mapped out. To begin we consider
the composite fields similar to those used in the earlier discussion
of pions as Goldstone bosons
\begin{eqnarray}
&\sigma\sim \overline\psi\psi \qquad
&\eta\sim i\overline\psi\gamma_5\psi\cr
&\vec\pi\sim i\overline\psi\gamma_5\vec\tau\psi \qquad
&{\vec a_0}\sim \overline\psi\vec\tau\psi.
\end{eqnarray}
In terms of these, a natural model for a starting effective potential
is
\begin{eqnarray}
V=&\lambda(\sigma^2+\vec\pi^2-v^2)^2-{m_1}\sigma-{m_2}{a_0}_3
-{m_3}\eta\cr
&+\alpha ({\eta}^2+{\vec a_0}^2)
-\beta (\eta\sigma+ {\vec a_0} \cdot \vec \pi)^2.
\end{eqnarray}
Here $\alpha$ and $\beta$ are ``low energy constants'' that bring in a
chirally symmetric coupling of $(\sigma,\vec\pi)$ with $(\eta,\vec
a_0)$. As discussed in Ref.~\cite{Creutz:1995wf}, the sign of the
$\beta$ term is suggested so that $m_{\eta} < m_{a_0}$.
This potential augments the famous ``Mexican hat'' or ``wine bottle''
potential discussed earlier, in which the Goldstone pions are
associated with the flat directions running around at constant
$\sigma^2+\vec\pi^2=v^2$. The $m_2$ and $m_3$ terms do not directly
affect the $\sigma$ and $\pi$ fields, but induce an expectation value
for ${a_0}_3$ and $\eta$, respectively. This in turn results in the
$\alpha$ and $\beta$ terms inducing a warping of the Mexican hat into
two separate minima, as sketched in Fig.~\ref{warping}. The direction
of this warping is determined by the relative size of $m_2$ and $m_3$;
$m_2$ ($m_3$) warps downward in $\pi_0$ ($\sigma$) direction. If we
now turn on $m_1$, this will select one of the two minimum as
favored. This gives rise to a generic first order transition at
$m_1=0$.
There is additional structure in the $m_1,m_2$ plane when $m_3$
vanishes. In this situation the quadratic warping is downward in the
$\sigma$ direction. For large $|m_1|$ only $\sigma$ will have an
expectation, with sign determined by the sign of $m_1$. The pion will
be massive, but with $m_2$ reducing the neutral pion mass below that
of the charged pions. If now $m_1$ is decreased in magnitude at fixed
$m_2$, eventually the neutral pion becomes massless and condenses.
How this occurs is sketched in Fig.~\ref{ising}. An order parameter
for the transition is the expectation value of the $\pi_0$ field, with
the transition being Ising-like.
\begin{figure}
\centering
\includegraphics[width=3.5in]{ising.eps}
\caption{In the $m_1,m_2$ plane, $m_{\pi_0}^2$ can pass through zero,
giving rise to pion condensation at an Ising-like transition. Figure
taken from \cite{Creutz:1995wf}. }
\label{ising}
\end{figure}
In this simple model the ratio of the neutral to charged pion masses
can be estimated from a quadratic expansion about the minimum of the
potential. For $m_3=0$ and $m_1$ above the transition line, this
gives
\begin{equation}
{m_{\pi_0}^2\over m_{\pi_\pm}^2}
=1-{\beta v m_2^2\over 2\alpha^2 m_1}+O(m^2).
\end{equation}
The second order transition is located where this vanishes, and thus
occurs for $m_1$ proportional to $m_2^2$. Note that this equation
verifies the important result that a constant quark mass ratio does
not correspond to a constant meson mass ratio and vice versa. This is
the ambiguity discussed at the beginning of this section.
This structure can also be observed in the expectation values for the
pion and sigma fields as functions of the average quark mass while
holding the quark mass difference fixed. This is sketched in
Fig.~\ref{condensate1}. The jump in $\sigma$ as we go from large
positive to large negative masses is split into two transitions with
the pion field acquiring an expectation value in the intermediate
region.
This second order transition occurs when both $m_u$ and $m_d$ are
non-vanishing but of opposite sign, i.e. $|m_1|<|m_2|$. This is
required to avoid the Vafa-Witten theorem \cite{Vafa:1984xg}, which
says that no parity breaking phase transition can occur if the fermion
determinant is positive definite. At the transition the correlation
length diverges. This shows that it is possible to have significant
long distance physics without the presence of small Dirac eigenvalues.
In contrast, we see that there is no transition at the point where
only one of the quark masses vanishes. In this situation there is no
long distance physics despite the possible existence of small Dirac
eigenvalues.
\begin{figure}
\centering
\includegraphics[width=3in]{condensate1.eps}
\caption{ With a constant up-down quark mass difference, the jump in
the chiral condensate splits into two second order transitions. The
order parameter distinguishing the intermediate phase is the
expectation value of the neutral pion field. Figure taken from
Ref.~\cite{Creutz:2005gb}. }
\label{condensate1}
\end{figure}
Putting this all together, we obtain the phase diagram sketched in
Fig.~\ref{phasediagram}. There are two intersecting first order
transition surfaces, one at $(m_1=0$, $m_3\ne 0)$ and the second at
$(m_1<m_2$,\ $m_3=0)$. These each occur where $\Theta=\pi$. However,
note that with non-degenerate quarks there is also a $\Theta=\pi$
region at $m_2=m_1+\epsilon$ for small but non-vanishing $\epsilon$
where there is no transition. The absence of a physical singularity
at $m_u=0$ when $m_d\ne 0$ lies at the heart of the problem in
defining a vanishing up quark mass.
In the next section we will see that the structure in the $m_1,m_2$
plane is closely related to an interesting lattice artifact in the
degenerate quark limit. Aoki \cite{Aoki:1983qi} discussed a possible
phase with spontaneous parity violation with the Wilson fermion
formulation. Indeed, lattice artifacts can modify the effective
potential in a similar way to the $m_2$ term and allow the CP
violating phase at finite cutoff to include part of the $m_1$ axis as
well.
\begin{figure}
\centering
{\includegraphics[width=3.5in] {phasediagram.eps}}
\caption{The full phase diagram for two-flavor QCD as a function of
the three mass parameters. It consists of two intersecting first
order surfaces with a second order edge along curves satisfying
$m_3=0$, $|m_1|<|m_2|$. There is no structure along the $m_u=0$
line except when both quark masses vanish. Figure from
Ref.~\cite{Creutz:2010ts}}
\label{phasediagram}
\end{figure}
\subsection{Bosonic fields}
A generic path integral
\begin{equation}
Z=\int (dU) e^{-S}
\end{equation}
on a finite lattice is a finite dimensional integral. One might try
to evaluate it numerically. But it is a many dimensional integral.
With $SU(3)$ on $10^4$ lattice we have $4*10^4$ links, each parametrized by
8 numbers. Thus it is a $320,000$ dimensional integral. Taking two
sample points for each direction, this already gives
\begin{equation}
2^{320,000}=3.8\times 10^{96,329}\qquad {\rm terms.}
\end{equation}
The age of the universe is only $\sim 10^{27}$ nanoseconds, so adding
one term at a time will take a while.
Such big numbers suggest a statistical approach. The goal of a Monte
Carlo simulation is to
find a few ``typical'' equilibrium configurations with probability distribution
\begin{equation}
p(C)\sim e^{-\beta S(C)}.
\end{equation}
On these one can measure observables of choice along with their
statistical fluctuations.
The basic procedure is a Markov process
\begin{equation}
C\rightarrow C^\prime \rightarrow \ldots
\end{equation}
generating a chain of configurations that eventually should approach
the above distribution. In general we take a configuration $C$ to a
new one with some given probability $P(C\rightarrow C^\prime)$. As a
probability, this satisfies $0\le P\le 1$ and
$\sum_{C^\prime}P(C\rightarrow C^\prime)=1$.\footnote{For continuous
groups the sum really means integrals.} For a Markov process, $P$
should depend only on the current configuration and have no dependence
on the history.
The process should bring us closer to ``equilibrium'' in a sense
shortly to be defined. This requires at least two things. First,
equilibrium should be stable; {\it i.e.} equilibrium is an
``eigen-distribution'' of the Markov chain
\begin{equation}
\sum_{C^\prime}P(C^\prime\rightarrow C) e^{-S(C^\prime)}= e^{-S(C)}.
\end{equation}
Second, we should have ergodicity; {\rm i.e.} all possible states must
in principle be reachable.
A remarkable result is that these conditions are sufficient for an
algorithm to approach equilibrium, although without any guarantee of
efficiency. Suppose we start with an ensemble of states, E,
characterized by the probability distribution $p(C)$. A distance
between ensembles is easily defined
\begin{equation}
D(E,E^\prime)\equiv \sum_C |p(C)-p^\prime(C)|.
\end{equation}
This is positive and vanishes only if the ensembles are equivalent.
A step of our Markov process takes ensemble $E$ into another
$E^\prime$ with
\begin{equation}
p^\prime(C)=\sum_{C^\prime}P(C^\prime\rightarrow C) p(C^\prime).
\end{equation}
Now assume that $P$ is chosen so that the equilibrium distribution
$p_{eq}(C)=e^{-S(C)}/Z$ is an eigenvector of eigenvalue 1. Compare
the new distance from equilibrium with the old
\begin{equation}
D(E^\prime,E_{eq})=\sum_C |p^\prime(C)-p_{eq}(C)|
=\sum_C\left |\sum_{C^\prime} P(C\rightarrow C^\prime)(p(C)-p_{eq}(C))\right|.
\end{equation}
Now the absolute value of a sum is always less than the sum of the
absolute values, so we have
\begin{equation}
D(E^\prime,E_{eq})\le
\sum_C \sum_{C^\prime} P(C\rightarrow C^\prime)|(p(C)-p_{eq}(C))|.
\end{equation}
Since each $C$ must go somewhere, the sum over $C^\prime$ gives
unity and we have
\begin{equation}
D(E^\prime,E_{eq})\le
\sum_C |(p(C)-p_{eq}(C))|=D(E,E_{eq}).
\end{equation}
Thus the algorithm automatically brings one closer to equilibrium.
How can one be sure that equilibrium is an eigen-ensemble? The usual
way in practice invokes a principle of detailed balance, a sufficient
but not necessary condition. This states that the forward and
backward rates between two states are equal when one is in equilibrium
\begin{equation}
p_{eq}(C)P(C\rightarrow C^\prime) = p_{eq}(C^\prime)P(C^\prime\rightarrow C).
\end{equation}
Summing this over $C^\prime$ immediately gives the fact that the
equilibrium distribution is an eigen-ensemble.
The famous Metropolis et al.~approach \cite{Metropolis:1953am} is an
elegant and simple way to construct an algorithm satisfying detailed
balance. This begins with a trial change on the configuration,
specified by a trial probability $P_T(C\rightarrow C^\prime)$. This
is required to be constructed in a symmetric way, so that
\begin{equation}
P_T(C\rightarrow C^\prime)=P_T(C^\prime\rightarrow C).
\end{equation}
This by itself would just tend to randomize the system. To restore
the detailed balance, the trial change is conditionally accepted with
probability
\begin{equation}
A(C,C^\prime)={\rm min}(1, p_{eq}(C^\prime)/p_{eq}(C)).
\end{equation}
In other words, if the Boltzmann weight gets larger, make the change;
otherwise, accept it with probability proportional to the ratio of the
Boltzmann weights. An explicit expression for the final transition
probability is
\begin{equation}
P(C\rightarrow C^\prime)=P_T(C\rightarrow C^\prime)A(C,C^\prime)
+\delta(C,C^\prime)\left(1-\sum_{C^{\prime\prime}}
P_T(C\rightarrow C^{\prime\prime})A(C,C^{\prime\prime})\right).
\end{equation}
The delta function accounts for the possibility that the change is
rejected.
For lattice gauge theory with its $U$ variables in a group, the trial
change can be most easily set up via a table of group elements $T=\{
g_1, ... g_n\}$. The trial change consists of picking an element
randomly from this table and using $U_T=gU$. These can be chosen
arbitrarily with two conditions: (1) multiplying them together in
various combinations should generate the whole group and (2) for each
element in the table, its inverse must also be present, i.e. $g\in T
\Rightarrow g^{-1} \in T$. The second condition is essential for
having the forward and reverse trial probabilities equal. An
interesting feature of this approach is that the measure of the group
is not used in any explicit way; indeed, it is generated
automatically.
Generally the group table should be weighted towards the identity.
Otherwise the acceptance gets small and you never go anywhere. But
this weighting should not be too extreme, because then the motion
through configuraton space becomes slow. Usually the width of the
table is adjusted to give an acceptance of order 50\%. For free field
theory the optimum can be worked out, it is a bit less. In general a
big change with a small acceptance can sometimes be better than small
changes; this appears to be the case with simulating self avoiding
random walks\cite{Madras:1988ei}.
The acceptance criterion involves the ratio
${p_{eq}(C^\prime)\over p_{eq}(C)}$.
An interesting quantity is
the expectation of this ratio in equilibrium. This is
\begin{equation}
\left\langle {p_{eq}(C^\prime)\over p_{eq}(C)} \right\rangle=
\sum_C p_{eq}(C) \sum_{C^\prime} P_T(C\rightarrow
C^\prime)p_{eq}(C^\prime)/p_{eq}(C)= 1
\end{equation}
since
\begin{equation}
\sum_C P_T(C\rightarrow C^\prime)=
\sum_C P_T(C^\prime \rightarrow C)=1
\end{equation}
and $\sum_{C^\prime}p_{eq}(C^\prime)=1.$ Of course the average
acceptance is not unity since it is expectation of the minimum of this
ratio and 1. However monitoring this expectation provides a simple
way to follow the approach to equilibrium.
A full Monte Carlo program consists of looping over all the lattice
links while considering such tentative changes. To improve
performance there are many tricks that have been developed over the
years. For example, in a lattice gauge calculation the calculation of
the ``staples'' interacting with a given link takes a fair amount of
time. This makes it advantageous to apply several Monte Carlo
``hits'' to the given link before moving on.
\subsection{Fermions}
The numerical difficulties with fermionic fields stem from their being
anti-commuting quantities. Thus it is not immediately straightforward
to place them on a computer, which is designed to manipulate numbers.
Indeed, the Boltzmann factor with fermions is formally an operator in
Grassmann space, and cannot be directly interpreted as a probability.
All algorithms in current use eliminate the fermions at the outset by
a formal analytic integration. This is possible because most actions
in practice are, or can easily be made, quadratic in the fermionic
fields. The fermion integrals are then over generalized gaussians.
Unfortunately, the resulting expressions involve the determinant of a
large, albeit sparse, matrix. This determinant introduces non-local
couplings between the bosonic degrees of freedom, making the path
integrals over the remaining fields rather time consuming.
For this brief overview we will be quite generic and assume we are
interested in a path integral of form
\begin{equation}
Z=\int (dA)(d\psi)(d\overline\psi)\ \exp(-S_G(A)-\overline\psi D(A) \psi).
\end{equation}
Here the gauge fields are formally denoted $A$ and fermionic fields
$\psi$ and $\overline\psi$. Concentrating on fermionic details, in
this section we ignore the technicality that the gauge fields are
actually group elements. All details of the fermionic formulation are
hidden in the matrix $D(A)$. While we call $A$ a gauge field, the
algorithms are general, and have potential applications in other field
theories and condensed matter physics.
In the section on Grassmann integration we found the basic formula
for a fermionic Gaussian integral
\begin{equation}
\int (d\psi d\overline\psi) \ e^{-\overline\psi D\psi} = \vert D
\vert
\end{equation}
where $(d\psi d\overline\psi)=d\psi_1\ d\overline\psi_1 \ldots d\psi_n
\ d\overline\psi_n$. Using this, we can explicitly integrate out the
fermions to convert the path integral to
\begin{equation}
Z=\int (dA)\ \vert D \vert\ e^{-S_G}
=\int (dA) \exp(-S_g+{\rm Tr}\ \log(D)).
\end{equation}
This is now an integral over ordinary numbers and therefore in
principle amenable to Monte Carlo attack.
For now we assume that the fermions have been formulated such that
$\vert D\vert$ is positive and thus the integrand can be regarded as
proportional to a probability measure. This is true for several of
the fermion actions discussed later. However, if $\vert D\vert$ is
not positive, one can always double the number of fermionic species,
replacing $D$ by $D^\dagger D$. We will see in later sections that
the case where $D$ is not positive can be rather interesting, but how
to include such situations in numerical simulations is not yet well
understood.
Direct Monte Carlo study of the partition function in this form is
still not practical because of the large size of the matrix $D$. In
our compact notation, this is a square matrix of dimension equal to
the number of lattice sites times the number of Dirac components times
the number of internal symmetry degrees of freedom. Thus, it is
typically a hundreds of thousands by hundreds of thousands matrix,
precluding any direct attempt to calculate its determinant. It is,
however, generally an extremely sparse matrix because most popular
actions do not directly couple distant sites. All the Monte Carlo
algorithms used in practice for fermions make essential use of this
fact.
Some time ago Weingarten and Petcher \cite{Weingarten:1980hx}
presented a simple ``exact'' algorithm. By introducing
``pseudofermions'' \cite{Fucito:1980fh,Scalapino:1981qs}, an auxiliary
set of complex scalar fields $\phi$, one can rewrite the path integral
in the form
\begin{equation}
Z= \int (dA) (d\phi^*\ d\phi) \exp(-S_G-\phi^* D^{-1}\phi).
\end{equation}
Thus a successful fermionic simulation would be possible if one could
obtain configurations of fields $\phi$ and $A$ with probability
distribution
\begin{equation}
P(A,\phi) \propto \exp(-S_G-\phi^* D^{-1}\phi).
\end{equation}
To proceed we again assume that $D$ is a positive matrix so this
distribution is well defined.
For an even number of species, generating an independent set of $\phi$
fields is actually quite easy. If we consider a field $\chi$ that is
gaussianly randomly selected, i.e. $P(\chi)\sim e^{-\chi^2}$, then
the field $\phi=D\chi$ is distributed as desired for two flavors
$P(\phi)\sim e^{-(D^{-1}\phi)^2}$. The hard part of the algorithm is
the updating of the $A$ fields, which requires knowledge of how
$\phi^* D^{-1}\phi$ changes under trial changes in $A$.
\subsection{The conjugate-gradient algorithm}
While $D^{-1}$ is the inverse of an enormous matrix, one really only
needs $\phi^* D^{-1}\phi$, which is just one matrix element of this
inverse. Furthermore, with a local fermionic action the matrix $D$ is
extremely sparse, the non-vanishing matrix elements only connecting
nearby sites. In this case there exist quite efficient iterative
schemes for finding the inverse of a large sparse matrix applied to a
single vector. Here we describe one particularly simple approach.
The conjugate gradient method to find $\xi=D^{-1} \phi$ works by
finding the minimum over $\xi$ of the function $\vert
D\xi-\phi\vert^2$. The solution is iterative; starting with some
$\xi_0$, a sequence of vectors is obtained by moving to the minimum of
this function along successive directions $d_i$. The clever trick of
the algorithm is to choose the $d_i$ to be orthogonal in a sense
defined by the matrix $D$ itself; in particular $(Dd_i, D d_j)=0$
whenever $i\ne j$. This last condition serves to eliminate useless
oscillations in undesirable directions, and guarantees convergence to
the minimum in a number of steps equal to the dimension of the
matrix. There are close connections between the conjugate gradient
inversion procedure and the Lanczos algorithm for tridiagonalizing
sparse matrices.
The procedure is a simple recursion. Select some arbitrary
initial pair of non-vanishing vectors $g_0=d_0$. For the inversion
problem, convergence will be improved if these are a good guess to
$D^{-1}\phi$. Then generate a sequence of further vectors by
iterating
\begin{eqnarray}
&& g_{i+1}=(Dg_i,Dd_i)g_i-(g_i,g_i)D^\dagger Dd_i \cr
&& d_{i+1}=(Dd_i,Dd_i)g_{i+1}-(Dd_,Dg_{i+1})d_i.
\end{eqnarray}
This construction assures that $g_i$ is orthogonal to $g_{i+1}$
and $(Dd_i,Dd_{i+1})=0$. It should also be clear that the three sets
of vectors $\{d_0,...d_k\}$,
$\{g_0,...g_k\}$, and $\{d_0,...(D^\dagger D)^kd_0\}$ all span the
same space.
The remarkable core of the algorithm, easily proved by induction, is
that the set of $g_i$ are all mutually orthogonal, as are $Dd_i$.
For an $N$ dimensional matrix, there can be no more than $N$
independent orthogonal vectors. Thus, ignoring round-off errors, the
recursion in Eq.~(15) must terminate in $N$ or less steps with the
vectors $g$ and $d$ vanishing from then on. Furthermore, as the
above sets of vectors all span the same space, in a basis defined by
the $g_i$ the matrix $D^\dagger D$ is in fact tri-diagonal, with
$(Dg_i,Dg_j)$ vanishing unless $i=j\pm 1$.
To solve $\phi=D\xi$ for $\xi$, simply expand in the $d_i$
\begin{equation}
\xi=\sum_i \alpha_i d_i.
\end{equation}
The coefficients are immediately found from the orthogonality
conditions
\begin{equation}
\alpha_i=(Dd_i,\phi)/(Dd_i,Dd_i).
\end{equation}
Note that if we start with the solution $d_0=D^{-1}\phi$, then we have
$\alpha_i=\delta_{i0}$.
This discussion applies for a general matrix $D$. If $D$ is
Hermitean, then one can work with better conditioned matrices by
replacing the orthogonality condition for the $d_i$ with $(d_i,Dd_j)$
vanishing for $i\ne j$.
In practice, at least when the correlation length is not too large,
this procedure adequately converges in a number of iterations which
does not grow severely with the lattice size. As each step involves
vector sums with length proportional to the lattice volume, each
conjugate gradient step takes a time which grows with the volume of
the system. Thus the overall algorithm including the sweep over
lattice variables is expected to require computer time which grows as
the square of the volume of the lattice. Such a severe growth has
precluded use of this algorithm on any but the smallest lattices.
Nevertheless, it does show the existence of an exact algorithm with
considerably less computational complexity than would be required for
a repeated direct evaluation of the determinant of the fermionic
matrix.
Here and below when we discuss volume dependences, we ignore additional
factors from critical slowing down when the correlation length is
also allowed to grow with the lattice size. The assumption is that
such factors are common for the local algorithms treated here. In
addition, such slowing occurs in bosonic simulations, and we are
primarily concerned here with the extra problems presented by the
fermions.
\subsection{Hybrid Monte Carlo}
One could imagine making trial changes of all lattice variables
simultaneously, and then accepting or rejecting the entire new
configuration using the exact action. The problem with this approach
is that a global random change in the gauge fields will generally
increase the action by an amount proportional to the lattice volume,
and thus the final acceptance rate will fall exponentially with the
volume. The acceptance rate could in principle be increased by
decreasing the step size of the trial changes, but then the step size
would have to decrease with the volume. Exploration of a reasonable
region of phase space would thus require a number of steps growing as
the lattice volume. The net result is an exact algorithm which
still requires computer time growing as volume squared.
So far this discussion has assumed that the trial changes are made in
a random manner. If, however, one can properly bias these variations,
it might be possible to reduce the volume squared behavior. The
``hybrid Monte Carlo'' scheme \cite{Duane:1987de} does this with a
global accept/reject step on the entire lattice after a microcanonical
trajectory.
The trick here is to add yet further auxiliary variables in the
form of ``momentum variables'' $p$ conjugate to the gauge fields $A$.
Then we look for a coupled distribution
\begin{equation}
P(p,A,\phi)=e^{-H(p,A,\phi)}
\end{equation}
with
\begin{equation}
H=p^2/2+V(A)
\end{equation}
and
\begin{equation}
V(A)=-S_g(A)- \phi^* D^{-1}\phi.
\end{equation}
The basic observation is that this is a simple classical Hamiltonian
for the conjugate variables $A$ and $p$, and evolution using Newton's
laws will conserve energy. For the gauge fields one sets up a
``trajectory'' in a fictitious ``Monte Carlo'' time variable $\tau$
and consider the classical evolution
\begin{eqnarray}
&&{dA_i\over d\tau}=p_i\cr
&&{dp_i\over d\tau}= F_i(A)=-{\partial V(A)\over \partial A_i}.
\end{eqnarray}
Under such evolution an equilibrium ensemble will remain in equilibrium.
An approximately energy conserving algorithm is given by a
``leapfrog'' discretization of Newton's law. With a microcanonical
time discretization of size $\delta$, this involves two half steps in
momentum sandwiching a full step in the coordinate $A$
\begin{eqnarray}
&&p_{1\over 2} = p + \delta\ F(A)/2\cr
&&A^\prime = A+\delta\ p_{1\over 2}\cr
&&p^\prime = p_{1\over 2}+\delta\ F(A^\prime)/2
\end{eqnarray}
or combined
\begin{eqnarray}
&& A^{\prime}=A+\delta\ p +\delta^2\ F(A)/2 \cr
&& p^{\prime}=p+\delta\ (F(A)+F(A^\prime))/2.
\end{eqnarray}
Even for finite step size $\delta$, this is an area preserving map of
the $(A,p)$ plane onto itself. The scheme iterates this mapping
several times before making a final Metropolis accept/reject decision.
This iterated map also remains reversible and area preserving. The
computationally most demanding part of this process is calculating the
force term. The conjugate gradient algorithm mentioned above can
accomplish this.
The important point is that after each step the momentum remains
exactly the negative of that which would be required to reverse the
entire trajectory and return to the initial variables. If at some
point on the trajectory we were to reverse all the momenta, the system
would exactly reverse itself and return to the same set of states from
whence it came. Thus a final acceptance with the appropriate
probability still makes the overall procedure exact. After each
accept/reject step, the momenta $p$ can be refreshed, their values
being replaced by new Gaussian random numbers. The pseudofermion
fields $\phi$ could also be refreshed at this time. The goal of the
procedure is to use the micro-canonical evolution as a way to restrict
changes in the action so that the final acceptance will remain high
for reasonable step sizes.
This procedure contains several parameters which can be
adjusted for optimization. First is $N_{mic}$, the number of
micro-canonical iterations taken before the global accept/reject
step and refreshing of the momenta $p$. Then there is the step
size $\delta$, which presumably should be set to give a
reasonable acceptance. Finally, one can also vary the frequency
with which the auxiliary scalar fields $\phi$ are updated.
The goal of this approach is to speed flow through phase space by
replacing a random walk of the $A$ field with a coherent motion in the
dynamical direction determined by the conjugate momenta. A simple
estimate \cite{Creutz:1988wv} suggests a net volume dependence
proportional to $V^{5/4}$ rather the naive volume squared without
these improvements.
As mentioned above, using pseudofermions is simplest if the fermion
matrix is a square, requiring an even number of species. Users of the
hybrid algorithm without the global accept-reject step have argued for
adjusting the number of fermion species by inserting a factor
proportional to the number of flavors in front of the pseudofermionic
term when the gauge fields are updated. This modification is simple
to make, but raises some theoretical issues that will be discussed
later. In particular, it is crucial that the underlying fermion
operator break any anomalous symmetries associated with the reduced
theory.
Despite the successes of these fermion algorithms, the overall
procedure still seems somewhat awkward, particularly when compared
with the ease of a pure bosonic simulation. This appears to be tied
to the non-local actions resulting from integrating out the fermions.
Indeed, had one integrated out a set of bosons coupled quadratically
to the gauge field, one would again have a non-local effective action,
indicating that this analytic integration was not a good idea.
Perhaps we should step back and explore algorithms before integrating
out the fermions.
An unsolved problem is to find a practical simulation approach to
fermionic systems where the corresponding determinant is not always
positive. This situation is of considerable interest because it
arises in the study of quark-gluon thermodynamics when a chemical
potential is present. All known approaches to this problem are
extremely demanding on computer resources. One can move the phase of
the determinant into the observables, but then one must divide out the
average value of this sign. This is a number which is expected to go
to zero exponentially with the lattice volume; thus, such an algorithm
will require computer time growing exponentially with the system size.
Another approach is to do an expansion about zero baryon density, but
again to get to large chemical potential will require rapidly growing
resources. New techniques are badly needed to avoid this growth;
hopefully this will be a particularly fertile area for future
algorithm development.
\section{#1} \setcounter{equation}{0}}
\long \def \blockcomment #1\endcomment{}
\def\slashchar#1{\setbox0=\hbox{$#1$}
\dimen0=\wd0
\setbox1=\hbox{/} \dimen1=\wd1
\ifdim\dimen0>\dimen1
\rlap{\hbox to \dimen0{\hfil/\hfil}}
#1
\else
\rlap{\hbox to \dimen1{\hfil$#1$\hfil}}
/
\fi} %
\begin{document}
\pagerange{1}{126}
\title{Confinement, chiral symmetry, and the lattice}
\author{Michael Creutz
\email{<EMAIL>}}
{Brookhaven National Laboratory\\
Upton, NY 11973
}
\abstract{Two crucial properties of QCD, confinement and chiral
symmetry breaking, cannot be understood within the context of
conventional Feynman perturbation theory. Non-perturbative
phenomena enter the theory in a fundamental way at both the
classical and quantum levels. Over the years a coherent qualitative
picture of the interplay between chiral symmetry, quantum mechanical
anomalies, and the lattice has emerged and is reviewed here.
}
\pacs{11.30.Rd, 12.39.Fe, 11.15.Ha, 11.10.Gh}
\tableofcontents
\newpage
\Section{QCD}
\input{quarksandgluons}
\subsection{Perturbation theory is not enough}
\input{whynonpert}
\newpage
\Section{Path integrals and statistical mechanics}
\label{pathintegrals}
\input{pathintegrals}
\newpage
\Section{Quark fields and Grassmann integration}
\input{grassmann}
\newpage
\Section{Lattice gauge theory}
\input{gaugefields}
\newpage
\Section{Monte Carlo simulation}
\label{montecarlo}
\input{montecarlo}
\newpage
\Section{Renormalization and the continuum limit}
\label{asymptoticfreedom}
\input{asymptoticfreedom}
\input{rgflow}
\newpage
\Section{Classical gauge fields and topology}
\label{classical}
\input{classical}
\subsection{The index theorem}
\input{winding}
\input{eigenflow}
\newpage
\Section{Chiral symmetry}
\label{chiral}
\input{chiral}
\input {effective}
\input{sigma}
\newpage
\Section{The chiral anomaly}
\input{anomaly}
\subsection{The 't Hooft vertex}
\input{vertex}
\newpage
\Section{Massive quarks and the Theta parameter}
\label{mass}
\input{theta}
\subsection{Quark scattering and mass mixing}
\label{mixing}
\input{masses}
\newpage
\Section{Lattice fermions}
\input{fermions}
\input{ginspargwilson}
\input{staggered}
\subsection{The rooting issue}
\input{rooting}
\newpage
\Section{Other issues}
\subsection{Quantum fluctuations and topology}
\input{fluctuations.tex}
\subsection{The standard model}
\input{standardmodel}
\newpage
\Section{Final remarks}
We have seen how many features of QCD are influenced by
non-perturbative physics. This is particularly important to various
aspects of chiral symmetry breaking. Taken as a whole, these fit
together into a rather elegant and coherent picture. In particular,
chiral symmetry is broken in three rather different ways. We have
concentrated on the interplay of these mechanisms.
The primary and most important effect is the dynamical symmetry
breaking that leads to the pions being light pseudo-Goldstone bosons.
Their dynamics represents the most important physics for QCD at low
energies. The popular and useful chiral expansion is a natural
expansion in the momenta and masses of these particles.
In addition to the basic dynamical breaking is the anomaly, which
eliminates the flavor-singlet axial $U(1)$ symmetry of the classical
theory. Thus the $\eta^\prime$ meson is not a Goldstone boson and
acquires a mass of order $\Lambda_{qcd}$. Understanding this breaking
requires non-perturbative physics associated with the zero modes of
the Dirac operator.
Finally, we have the explicit symmetry breaking from the quark masses.
This is responsible for the pseudo-scalar mesons not being exactly
massless. Using the freedom to redefine fields using chiral rotations,
the number of independent mass parameters is $N_f+1$ where $N_f$ is
the number fermion species under consideration. This includes the
possibility of CP violation coming from the interplay of the mass term
with the anomaly.
Throughout we have used only a few widely accepted assumptions, such
as the existence of QCD as a field theory and standard ideas about
chiral symmetry. Thus it is perhaps somewhat surprising that several
of the conclusions remain controversial. The first of these is that
chiral symmetry is lost in a theory with only one light quark. The
resulting additive non-perturbative renormalization of the mass
precludes using a massless up quark to solve the strong CP problem.
Tied to this is the issue of whether topological susceptibility is
well defined when non-differentiable fields dominate the path
integral. Finally, probably the most bitter controversies revolve
about the symmetries inherent in the staggered formulation and how
these invalidate the use of rooting to remove unwanted degeneracies.
As simple as the overall picture is, it requires understanding effects
that go well beyond perturbation theory. We need aspects of the Dirac
spectrum that rely on gauge fields of non-trivial topology. Such
appear already in the classical theory, although their true importance
only appears in the context of the anomaly. Including this physics
properly in a lattice formulation is a rich and sometimes
controversial topic of active research.
\begin{ack}
The author is grateful to Ivan Horvath for suggesting this article.
He is also indebted to Ivan as well as to Stefano Capitani and
Nobuyoshi Ohta for finding numerous typos in the original version.
The Alexander von Humboldt Foundation provided valuable support for
visits to the University of Mainz where a portion of this research was
carried out. This manuscript has been authored by employees of
Brookhaven Science Associates, LLC under Contract
No. DE-AC02-98CH10886 with the U.S. Department of Energy. The
publisher by accepting the manuscript for publication acknowledges
that the United States Government retains a non-exclusive, paid-up,
irrevocable, world-wide license to publish or reproduce the published
form of this manuscript, or allow others to do so, for United States
Government purposes.
\end{ack}
\input{nonpertrefs}
\end{document}
\subsection{Discretizing time}
We begin with the Lagrangean for a free
particle of mass $m$ moving in potential $V(x)$
\begin{equation}
L(x,\dot x)=K(\dot x) + V(x)
\label{nrl}
\end{equation}
where $K(\dot x)={1\over 2} m\dot x^2$ and $\dot x$ is the time
derivative of the coordinate $x$. Note the unconventional relative
positive sign between the two terms in Eq.~(\ref{nrl}). This is
because we formulate the path integral directly in imaginary time.
This improves mathematical convergence, yet we are left with the usual
Hamiltonian for diagonalization.
For a particle traversing a trajectory $x(t)$, we have the action
\begin{equation}
S=\int dt\ L(\dot x(t),x(t)).
\end{equation}
This appears in the path integral
\begin{equation}
Z=\int(dx) e^{-S}.
\label{pathintegral}
\end{equation}
Here the integral is over all possible trajectories $x(t)$. As it
stands, Eq.~(\ref{pathintegral}) is rather poorly defined. To
characterize the possible trajectories we introduce a cutoff in the
form of a time lattice. Putting our system into a temporal box of
total length $T$, we divide this interval into $N={T\over a}$ discrete
time slices, where $a$ is the timelike lattice spacing. Associated
with the $i$'th such slice is a coordinate $x_i$. This construction
is sketched in Figure \ref{slicing}. Replacing the time derivative of
$x$ with a nearest-neighbor difference, we reduce the action to a sum
\begin{equation}
S=a\sum_i\left[{1\over 2}m\left({x_{i+1}-x_i\over a}\right)^2+V(x_i)\right].
\end{equation}
The integral in Eq.~(\ref{pathintegral}) is now defined as an ordinary
integral over all the coordinates
\begin{equation}
Z=\int\left(\prod_i dx_i\right)\ e^{-S}.
\end{equation}
\begin{figure}
\centering
\includegraphics[width=3in]{slicing.eps}
\caption {Dividing time into a lattice of $N$ slices of timestep $a$.}
\label{slicing}
\end{figure}
This form for the path integral is precisely in the form of a
partition function for a statistical system. We have a one
dimensional polymer of coordinates $x_i$. The action represents the
inverse temperature times the Hamiltonian of the thermal analog. This
is a special case of a deep result, a $D$ space-dimensional quantum
field theory is equivalent to the classical thermodynamics of a $D+1$
dimensional system. In this example, we have one degree of freedom
and $D$ is zero; for the lattice gauge theory of quarks and gluons,
$D$ is three and we work with the classical statistical mechanics of a
four dimensional system.
We will now show that the evaluation of this partition function is
equivalent to diagonalizing a quantum mechanical Hamiltonian obtained
from the action via canonical methods. This is done with the use of
the transfer matrix.
\subsection{The transfer matrix}
The key to the transfer-matrix analysis is to note that the local
nature of the action permits us to write the partition function as a
matrix product
\begin{equation}
Z=\int\prod_i dx_i\ T_{x_{i+1},x_i}
\end{equation}
where the transfer-matrix elements are
\begin{equation}
T_{x^\prime,x}=\exp\left[-{m\over 2a}(x^\prime-x)
-{a\over 2}(V(x^\prime)+V(x))\right].
\label{transfer}
\end{equation}
The transfer matrix itself is an operator in the Hilbert space of
square integrable functions with the standard inner product
\begin{equation}
\langle\psi^\prime\vert\psi\rangle=\int dx {\psi^\prime}^*(x)\psi(x).
\end{equation}
We introduce the non-normalizable basis states $\vert x \rangle$ such
that
\begin{eqnarray}
&&\vert\psi\rangle=\int dx\ \psi(x)\ \vert x \rangle\\
&&\langle x^\prime \vert x \rangle=\delta(x^\prime-x)\\
&&1=\int dx\ \vert x\rangle\langle x\vert.
\end{eqnarray}
Acting on the Hilbert space are the canonically conjugate operators
$\hat p$ and $\hat x$ that satisfy
\begin{eqnarray}
&&\hat x\vert x\rangle=x\vert x\rangle\cr
&&[{\hat p,\hat x}]=-i\cr
&&e^{-i\hat py}\vert x\rangle=\vert x+y\rangle.
\end{eqnarray}
The operator $T$ is defined via its matrix elements
\begin{equation}
\langle x^\prime\vert T\vert x\rangle
=T_{x^\prime,x}
\end{equation}
where $T_{x^\prime,x}$ is given in Eq.~(\ref{transfer}). With periodic
boundary conditions on our lattice of $N$ sites, the path integral is
compactly expressed as as a trace over the Hilbert space
\begin{equation}
Z={\rm Tr}\ T^N.
\end{equation}
Expressing $T$ in terms of the basic operators $\hat p,\hat x$ gives
\begin{equation}
T=\int dy\ e^{-y^2/(2a)}\ e^{-aV(\hat x)/2}\ e^{-i\hat py}
\ e^{-aV(\hat x)/2}.
\end{equation}
To prove this, check that the right hand side has the appropriate
matrix elements. The integral over $y$ is Gaussian and gives
\begin{equation}
T=\left({2\pi a\over m}\right)^{1/2} \ e^{-aV(\hat x)/2}
e^{-a\hat p^2/(2m)}
e^{-aV(\hat x)/2}.
\end{equation}
The connection with the usual quantum mechanical Hamiltonian appears
in the small lattice spacing limit. When $a$ is small, the exponents
in the above equation combine to give
\begin{equation}
T=\left({2\pi a\over m}\right)^{1/2}\ e^{-aH+O(a^2)}
\end{equation}
with
\begin{equation}
H={\hat p^2\over 2m}+V(\hat x).
\end{equation}
This is just the canonical Hamiltonian operator following from our
starting Lagrangean.
The procedure for going from a path-integral to a Hilbert-space
formulation of quantum mechanics consists of three steps. First
define the path integral with a discrete time lattice. Then construct
the transfer matrix and the Hilbert space on which it operates.
Finally, take the logarithm of the transfer matrix and identify the
negative of the coefficient of the linear term in the lattice spacing
as the Hamiltonian. Physically, the transfer matrix propagates the
system from one time slice to the next. Such time translations are
generated by the Hamiltonian.
The eigenvalues of the transfer matrix are related to the energy
levels of the quantum system. Denoting the $i$'th eigenvalue of $T$
by $\lambda_i$, the path integral or partition function
becomes
\begin{equation}
Z=\sum_i\lambda_i^N.
\end{equation}
As the number of time slices goes to infinity, this expression is
dominated by the largest eigenvalue
$\lambda_0$ \begin{equation}
Z=\lambda_0^N\times[1+O(\exp[-N\log(\lambda_0/\lambda_1)])].
\end{equation}
In statistical mechanics the thermodynamic properties of a system
follow from this largest eigenvalue. In ordinary quantum mechanics
the corresponding eigenvector is the lowest eigenstate of the
Hamiltonian. This is the ground state or, in field theory, the vacuum
$\vert 0\rangle$. Note that in this discussion, the connection
between imaginary and real time is trivial. Whether the generator of
time translations is $H$ or $iH$, we have the same operator to
diagonalize.
In statistical mechanics one is often interested in correlation
functions between the statistical variables at different points. This
corresponds to a study of the Green's functions of the corresponding
field theory. These are obtained upon insertion of polynomials of the
fundamental variables into the path integral.
An important feature of the path integral is
that a typical path is
non-differentiable \cite{feynmanhibbs,Creutz:1980gp}. Consider the
discretization of the time derivative
\begin{equation}
\dot x \sim {x_{i+1}- x_i \over a}.
\end{equation}
The kinetic term in the path integral controls how close the fields
are on adjacent sites. Since this appears as simple Gaussian factor
$\exp(-(x_{i+1}- x_i)^2m/a)$ we see that
\begin{equation}
{1\over 2} m \langle \dot x^2\rangle=O(1/ma).
\end{equation}
This diverges as the lattice spacing goes to zero.
One can obtain the average kinetic energy in other ways, for example
through the use of the virial theorem or by point splitting. However,
the fact that the typical path is not differentiable means that one
should be cautious about generalizing properties of classical fields
to typical configurations in a numerical simulation. We will see that
such questions naturally arise when considering the topological
properties of gauge fields.
\subsection{Why quarks}
Although an isolated quark has not been seen, we have a variety of
reasons to believe in the reality of quarks as the basis for this next
layer of matter. First, quarks provide a rather elegant explanation
of certain regularities in low energy hadronic spectroscopy. Indeed,
it was the successes of the eightfold way
\cite{GellMann:1964nj} which originally motivated the quark model.
Two ``flavors' of low mass quarks lie at the heart of isospin symmetry
in nuclear physics. Adding the somewhat heavier ``strange'' quark
gives the celebrated multiplet structure described by representations
of the group $SU(3)$.
Second, the large cross sections observed in deeply inelastic
lepton-hadron scattering point to structure within the proton at
distance scales of less than $10^{-16}$ centimeters, whereas the
overall proton electromagnetic radius is of order $10^{-13}$
centimeters \cite{Mishra:1989jc}. Furthermore, the angular
dependences observed in these experiments indicate that any underlying
charged constituents carry half-integer spin.
Yet a further piece of evidence for compositeness lies in the excitations of
the low-lying hadrons. Particles differing in angular momentum fall
neatly into place along the famous ``Regge trajectories''
\cite{Collins:1977jy}. Families of states group together as orbital
excitations of an underlying extended system. The sustained rising of
these trajectories with increasing angular momentum points toward
strong long-range forces between the constituents.
Finally, the idea of quarks became incontrovertible with the discovery
of heavier quark species beyond the first three. The intricate
spectroscopy of the charmonium and upsilon families is admirably
explained via potential models for non-relativistic bound states.
These systems represent what are sometimes thought of as the
``hydrogen atoms'' of elementary particle physics. The fine details
of their structure provides a major testing ground for quantitative
predictions from lattice techniques.
\subsection{Gluons and confinement}
Despite its successes, the quark picture raises a variety of puzzles.
For the model to work so well, the constituents should not interact so
strongly that they loose their identity. Indeed, the question arises
whether it is possible to have objects display point-like behavior in
a strongly interacting theory. The phenomenon of asymptotic freedom,
discussed in more detail later, turns out to be crucial to realizing
this picture.
Perhaps the most peculiar aspect of the theory relates to the fact
that an isolated quark has never been observed. These basic
constituents of matter do not copiously appear as free particles
emerging from high energy collisions. This is in marked contrast to
the empirical observation in hadronic physics that anything which can
be created will be. Only phenomena prevented by known symmetries are
prevented. The difficulty in producing quarks has led to the concept
of a principle of exact confinement. Indeed, it may be simpler to
have a constituent which can never be produced than an approximate
imprisonment relying on an unnaturally small suppression factor. This
is particularly true in a theory like the strong interactions, which
is devoid of any large dimensionless parameters.
But how can one ascribe any reality to an object which cannot be
produced? Is this just some sort of mathematical trick? Remarkably,
gauge theories potentially possess a simple physical mechanism for
giving constituents infinite energy when in isolation. In this
picture a quark-antiquark pair experiences an attractive force which
remains non-vanishing even for asymptotically large separations. This
linearly rising long distance potential energy forms the basis of
essentially all models of quark confinement.
For a qualitative description of the mechanism, consider coupling the
quarks to a conserved ``gluo-electric'' flux. In usual
electromagnetism the electric field lines thus produced spread and
give rise to the inverse square law Coulombic field. If one can
somehow eliminate massless fields, then a Coulombic spreading will no
longer be a solution to the field equations. If in removing the
massless fields we do not destroy the Gauss law constraint that the
quarks are the sources of electric fields, the electric lines must
form into tubes of conserved flux, schematically illustrated in
Fig.~\ref{fig:flux}. These tubes begin and end on the quarks and
their antiparticles. The flux tube is meant to be a real physical
object carrying a finite energy per unit length. This is the storage
medium for the linearly rising inter-quark potential. In some sense
the reason we cannot have an isolated quark is the same as the reason
that we cannot have a piece of string with only one end. In this
picture a baryon would require a string with three ends. It lies in
the group theory of non-Abelian gauge fields that this peculiar state
of affairs is allowed.
Of course a length of real string can break into two, but then each
piece has itself two ends. In the QCD case a similar phenomenon
occurs when there is sufficient energy in the flux tube to create a
quark-antiquark pair from the vacuum. This is qualitatively what
happens when a rho meson decays into two pions.
\begin{figure}[t]
\begin{center}
\includegraphics[width=3in]{fluxtube.eps}
\caption {A tube of gluonic flux connects quarks and anti-quarks. The
strength of this string is 14 tons.}
\label{fig:flux}
\end{center}
\end{figure}
One model for this phenomenon is a type II superconductor containing
magnetic monopole impurities. Because of the Meissner
effect \cite{meissner}, a superconductor does not admit magnetic
fields. However, if we force a hypothetical magnetic monopole into
the system, its lines of magnetic flux must go somewhere. Here the
role of the ``gluo-electric'' flux is played by the magnetic field,
which will bore a tube of normal material through the superconductor
until it either ends on an anti-monopole or it leaves the boundary of
the system \cite{Abrikosov:1956sx}. Such flux tubes have been
experimentally observed in real superconductors \cite{Hess:1989zz}.
Another example of this mechanism occurs in the bag model
\cite{Chodos:1974je}. Here the gluonic fields are unrestricted in the
bag-like interior of a hadron, but are forbidden by {\it ad hoc}
boundary conditions from extending outside. In attempting to extract
a single quark from a proton, one would draw out a long skinny bag
carrying the gluo-electric flux of the quark back to the remaining
constituents.
The above models may be interesting phenomenologically, but they are
too arbitrary to be considered as the basis for a fundamental theory.
In their search for a more elegant approach, theorists have been drawn
to non-Abelian gauge fields \cite{Yang:1954ek}. This dynamical system
of coupled gluons begins in analogy with electrodynamics with a set of
massless gauge fields interacting with the quarks. Using the freedom
of an internal symmetry, the action also includes self-couplings of
the gluons. The bare massless fields are all charged with respect to
each other. The confinement conjecture is that this input theory of
massless charged particles is unstable to a condensation of the vacuum
into a state in which only massive excitations can propagate. In such
a medium the gluonic flux around the quarks should form into the flux
tubes needed for linear confinement. While this has never been proven
analytically, strong evidence from lattice gauge calculations
indicates that this is indeed a property of these theories.
The confinement phenomenon makes the theory of the strong interactions
qualitatively rather different from the theories of the
electromagnetic and weak forces. The fundamental fields of the
Lagrangean do not manifest themselves in the free particle spectrum.
Physical particles are all gauge singlet bound states of the
underlying constituents. In particular, an expansion about the free
field limit is inherently crippled at the outset. This is perhaps the
prime motivation for the lattice approach.
In the quark picture, baryons are bound states of three quarks. Thus
the gauge group should permit singlets to be formed from three objects
in the fundamental representation. This motivates the use of $SU(3)$
as the underlying group of the strong interactions. This internal
symmetry must not be confused with the broken $SU(3)$ represented in
the multiplets of the eightfold way. Ironically, one of the original
motivations for quarks has now become an accidental symmetry, arising
only because three of the quarks are fairly light. The gauge symmetry
of importance to us now is hidden behind the confinement mechanism,
which only permits observation of singlet states.
The presentation here assumes, perhaps too naively, that the nuclear
interactions can be considered in isolation from the much weaker
effects of electromagnetism, weak interactions, and gravitation. This
does not preclude the possible application of the techniques to the
other interactions. Indeed, unification may be crucial for a
consistent theory of the world. To describe physics at normal
laboratory energies, however, it is only for the strong interactions
that we are forced to go beyond well-established perturbative methods.
Thus we frame the discussion around quarks and gluons.
\subsection{Flows and irrelevant operators}
We now briefly discuss another way of looking at the renormalization
group as relating theories with different lattice spacings. Given one
lattice theory, one could imagine generating another with a larger
lattice spacing by integrating over all links except those on some
subset of the original lattice, thus generating an equivalent theory
with, say, a larger lattice spacing. While this is conceptually
possible, to do it exactly in more than one dimension will generate an
infinite number of couplings. If we could keep track of such, the
procedure would be ``exact,'' but in reality we usually need some
truncation. Continuing to integrate out degrees of freedom, the
couplings flow and might reach some ``fixed point'' in this infinite
space. With multiple couplings, there can be an attractive ``sheet''
towards which couplings flow, and then they might continue to flow
towards a fixed point, as sketched in Fig.~\ref{flow}. If the fixed
point has only one attractive direction, then two different models
that flow towards that same fixed point will have the same physics in
the large distance limit. This is the concept of universality; {\it
i.e.} exponents are the same for all models with the same
attractor.
\begin{figure*}
\centering
\includegraphics[width=2.5in]{flow.eps}
\caption{\label{flow}
A generic renormalization group flow. In general this occurs in an
infinite dimensional space.
}
\end{figure*}
Some hints on this process come from dimensional analysis, although,
in ignoring non-perturbative effects that might occur at strong
coupling, the following arguments are not rigorous. In $d$ dimensions
a conventional scalar field has dimensions of $M^{d-2\over 2}$. Thus
the coupling constant $\lambda$ in an interaction of form $\int d^dx\
\lambda \phi^n$ has dimensions of $M^{d-n{d-2\over2}}$. On a lattice
of spacing $a$, the natural unit of dimension is the inverse lattice
spacing. Thus without any special tuning, the renormalized coupling
at some fixed physical scale would naturally run as $\lambda\sim
a^{n{d-2\over2}-d}$. As long as the exponent in this expression is
positive, i.e.
$$
n\ge {2d\over d-2}
$$
we expect the coupling to become ``irrelevant'' in the continuum
limit. The fixed point is driven towards zero in the corresponding
direction. If $d$ exceeds four, this is the case for all
interactions. (We ignore $\phi^3$ in 6 dimensions because of
stability problems.) This suggests that four dimensions is a critical
case, with mean field theory giving the right qualitative critical
behavior for all larger dimensions. In four dimensions we have
several possible ``renormalizable'' couplings which are dimensionless,
suggesting logarithmic corrections to the simple dimensional
arguments. Indeed, four-dimensional non-Abelian gauge theories
display exactly such a logarithmic flow; this is asymptotic freedom.
This simple dimensional argument applied to the mass term suggests it
would flow towards infinity in all dimensions. For a conventional
phase transition, something must be tuned to a critical point. In
statistical mechanics this is the temperature. In field theory
language we usually remap this onto a tuning of the bare mass term,
saying that the transition occurs as bare masses go through zero. For
a scalar theory this tuning for a continuum limit seems unnatural and
is one of the unsatisfying features of the standard model, driving
particle physicists to try to unravel how the Higg's mechanism really
works.
In non-Abelian gauge theories with multiple massless fermions, chiral
symmetry protects the mass from renormalization, avoiding any special
tuning. Indeed, as we have discussed, because of dimensional
transmutation, all dimensionless parameters in the continuum limit are
completely determined by the basic structure of the initial
Lagrangean, without any continuous parameters to tune. In the limit
of vanishing pion mass, the rho to nucleon mass ratio should be
determined from first principles; it is the goal of lattice gauge
theory to calculate just such numbers.
As we go below four dimensions, this dimensional argument suggests
that several couplings can become ``relevant,'' requiring the
renormalization group picture of flow towards a non-trivial fixed
point. Above two dimensions the finite number of renormalizable
couplings corresponds to the renormalization group argument for a
finite number of ``universality classes,'' corresponding to different
basic symmetries.
One might imagine dimensionality as being a continuously variable
parameter. Then just below four dimensions a renormalizable coupling
becomes ``super-renormalizable'' and a new non-trivial fixed point
breaks away from vanishing coupling. Near four dimensions this point
is at small coupling, forming the basis for an expansion in $4-d$.
This has become a major industry, making remarkably accurate
predictions for critical exponents in three dimensional
systems \cite{Wilson:1973jj}.
An important consequence of this discussion is that a lattice action
is in general highly non-unique. One can always add irrelevant
operators and expect to obtain the same continuum limit.
Alternatively, one might hope to improve the approach the continuum
limit by a judicious choice of the lattice action.
The renormalization group is indeed a rich subject. We have only
touched on a few issues that are particularly valuable for the lattice
theory. Perhaps the most remarkable result of this section is how a
perturbative analysis of the renormalization-group equation gives rise
to information on the non-perturbative behavior in the particle
masses, as exhibited in Eq.~(\ref{correlation}).
\subsection{The Sigma model}
Much of the structure of low energy QCD is nicely summarized in terms
of an effective chiral Lagrangean formulated in terms of a field which
is an element of the underlying flavor group. In this section we
review this model for the strong interactions with three quarks,
namely up, down, and strange. The theory has an approximate SU(3)
symmetry, broken by unequal masses for the quarks. We work with the
familiar octet of light pseudoscalar mesons $\pi_\alpha$ with
$\alpha=1\ldots 8$ and consider an SU(3) valued field
\begin{equation}
\label{sigma}
\Sigma=\exp(i\pi_\alpha \lambda_\alpha/f_\pi)\in SU(3).
\end{equation}
Here the $\lambda_\alpha$ are the usual Gell-Mann matrices which
generate the flavor group and $f_\pi$ is a dimensional constant with a
phenomenological value of about 93 MeV. We follow the normalization
convention that ${\rm Tr} \lambda_\alpha \lambda_\beta =
2\delta_{\alpha\beta}$. The neutral pion and the eta meson will play
a special role later in this review; they are the coefficients of
the commuting generators
\begin{equation}
\lambda_3={1\over \sqrt 3}\pmatrix{1&0&0\cr
0&-1&0\cr 0&0&0\cr}
\end{equation}
and
\begin{equation}
\lambda_8={1\over \sqrt 3}\pmatrix{1&0&0\cr
0&1&0\cr 0&0&-2\cr},
\end{equation}
respectively. In the chiral limit of vanishing quark masses, the
interactions of the eight massless Goldstone bosons are modeled with
the effective Lagrangean density
\begin{equation}
\label{kinetic}
L_0={f_\pi^2\over 4}{\rm Tr}(\partial_\mu \Sigma^\dagger \partial_\mu \Sigma).
\end{equation}
The non-linear constraint of $\Sigma$ onto the group SU(3) makes this
theory non-renormalizable. It is to be understood only as the
starting point for an expansion of particle interactions in powers of
their momenta. Expanding Eq.~(\ref{kinetic}) to second order in the
meson fields gives the conventional kinetic terms for our eight
mesons.
This theory is invariant under parity and charge conjugation. These
operators are represented by simple transformations
\begin{equation}
\matrix{
&P:\ \Sigma\ \rightarrow\ \Sigma^{-1}\cr
&CP:\ \Sigma\ \rightarrow\ \Sigma^{*}\cr
}
\end{equation}
where the operation $*$ refers to complex conjugation. The eight
meson fields are pseudoscalars. The neutral pion and the eta meson
are both even under charge conjugation.
With massless quarks, the underlying quark-gluon theory has a chiral
symmetry under
\begin{equation}
\matrix{
&\psi_L\rightarrow g_L\psi_L \cr
&\psi_R\rightarrow g_R\psi_R. \cr
}
\end{equation}
Here $(g_L,g_L)$ is in $SU(3)\otimes SU(3)$ and $\psi_{L,R}$
represent the chiral components of the quark fields, with flavor
indices understood. This symmetry is expected to be broken
spontaneously to a vector SU(3) via a vacuum expectation value for
$\overline \psi_L \psi_R$. This motivates the sigma model through the
identification
\begin{equation}
\langle 0 | \overline \psi_L \psi_R | 0 \rangle \leftrightarrow v \Sigma.
\label{vec}
\end{equation}
The quantity $v$, of dimension mass cubed, characterizes the strength
of the spontaneous breaking of this symmetry. Thus the effective
field transforms under a chiral symmetry of form
\begin{equation}
\Sigma\rightarrow g_L \Sigma g_R^\dagger.
\end{equation}
The Lagrangean density in Eq.~(\ref{kinetic}) is the simplest
non-trivial expression invariant under this symmetry.
The quark masses break the chiral symmetry explicitly. From the
analogy in Eq.~(\ref{vec}), these are introduced through a 3 by 3 mass
matrix $M$ appearing in a potential term added to the Lagrangean
density
\begin{equation}
L= L_0 - v {\rm Re\ Tr}(\Sigma M).
\end{equation}
Here $v$ is the same dimensionful factor appearing in Eq.~(\ref{vec}).
The chiral symmetry of our starting theory shows the physical
equivalence of a given mass matrix $M$ with a rotated matrix
$g_R^\dagger M g_L$. Using this freedom we can put the mass matrix
into a standard form. We will assume it is diagonal with increasing
eigenvalues
\begin{equation}
M=\pmatrix{
m_u & 0 & 0 \cr
0 & m_d & 0 \cr
0 & 0 & m_s \cr
}
\end{equation}
representing the up, down, and strange quark masses. Note that this
matrix has both singlet and octet parts under flavor symmetry
\begin{equation}
M={m_u+m_d+m_s\over 3}
+{m_u-m_d\over 2}\ \lambda_3+{m_u+m_d-2 m_s\over 2\sqrt 3}\ \lambda_8.
\end{equation}
In general the mass matrix can still be complex. The chiral symmetry
allows us to move phases between the masses, but the determinant of
$M$ is invariant and physically meaningful. Under charge conjugation
the mass term would only be invariant if $M=M^*$. If $|M|$ is not
real, then its phase is the famous CP violating parameter that we will
extensively discuss later. For the moment, however, we take all quark
masses as real.
To lowest order the pseudoscalar meson masses appear on expanding the
mass term quadratically in the meson fields. This generates an
effective mass matrix for the eight mesons
\begin{equation}
{\cal M}_{\alpha\beta}\ \propto\ {\rm Re\ Tr}\ \lambda_\alpha\lambda_\beta M.
\end{equation}
The isospin breaking up-down mass difference gives this matrix an off
diagonal piece mixing the $\pi_0$ and the $\eta$
\begin{equation}
{\cal M}_{3,8}\ \propto\ m_u-m_d.
\end{equation}
The eigenvalues of this matrix give the standard mass relations
\begin{eqnarray}
\label{mesons}
&& m_{\pi_0}^2 \propto\ {2\over 3} \bigg(m_u+m_d+m_s
-\sqrt{m_u^2+m_d^2+m_s^2-m_um_d-m_um_s-m_dm_s}\bigg)\cr
&&m_{\eta}^2 \propto \ {2\over 3} \bigg(m_u+m_d+m_s
+\sqrt{m_u^2+m_d^2+m_s^2-m_um_d-m_um_s-m_dm_s}\bigg)\cr
&&m_{\pi_+}^2= \ m_{\pi_-}^2\propto m_u+m_d\cr
&&m_{K_+}^2= \ m_{K_-}^2\propto m_u+m_s\cr
&&m_{K_0}^2= \ m_{\overline K_0}^2\propto m_d+m_s.
\end{eqnarray}
Here we label the mesons with their conventional names.
Redundancies in these relations test the validity of the model. For
example, comparing two expressions for the sum of the three quark
masses
\begin{equation}
{2(m_{\pi_+}^2+m_{K_+}^2+m_{K_0}^2)
\over 3(m_{\eta}^2+m_{\pi_0}^2)} \sim 1.07
\end{equation}
suggests the symmetry should be good to a few percent. Further ratios
of meson masses then give estimates for the ratios of the quark masses
\cite{Kaplan:1986ru,Weinberg:1977hb,Leutwyler:1996qg}. For one such
combination, look at
\begin{equation}
{m_u\over m_d}=
{
m_{\pi^+}^2+m_{K_+}^2-m_{K_0}^2\over
m_{\pi^+}^2-m_{K_+}^2+m_{K_0}^2}\sim 0.66
\end{equation}
This particular combination is polluted by electromagnetic effects;
another combination that partially cancels such while ignoring small
$m_um_d/m_s$ corrections in expanding the square root in
Eq.~(\ref{mesons}) is
\begin{equation}
{m_u\over
m_d}= { 2m_{\pi^0}^2-m_{\pi^0}^2+m_{K_+}^2-m_{K_0}^2\over
m_{\pi^+}^2-m_{K_+}^2+m_{K_0}^2}\sim 0.55
\label{updown}
\end{equation}
In a moment we will comment on a third combination for this ratio.
For the strange quark, one can take
\begin{equation}
{2 m_s\over m_u+m_d}=
{
m_{K_+}^2+m_{K_0}^2-m_{\pi^+}^2\over
m_{\pi^+}^2
}\sim 26.
\end{equation}
Of course as discussed earlier the quark masses are scale dependent.
While their ratios are more stable, we will see later how these ratios
also acquire some scale dependence. Nevertheless, from mass
differences such as $m_n-m_p\sim 1.3{\rm MeV} $ and
$m_{K_0}-m_{K_+}\sim 4.0{\rm MeV}$ we conclude that the up and down
quark masses in these effective models are typically of order a few
MeV, while the strange quark mass is of order 100 MeV. These are what
are known as ``current'' quark masses, related to chiral symmetries
and current algebra. In contrast, since the proton is made of three
quarks, some simple quark models consider ``constituent'' quark masses
of a few hundred Mev; these are substantially larger because they
include the energy contained in the gluon fields.
While phenomenology, i.e. Eq.~(\ref{updown}), seems to suggest that
the up quark is not massless, there remains a lot of freedom in
extracting that ratio from the pseudoscalar meson masses. From
Eq.~(\ref{mesons}), the sum of the $\eta$ and $\pi_0$ masses squared
should be proportional to the sum of the three quark masses.
Subtracting off the neutral kaon mass should leave just the up quark.
Thus motivated, look at
\begin{equation}
{m_u\over m_d}=
{
3(m_{\eta}^2+m_{\pi_0}^2)/2-2m_{K_0}^2
\over
m_{\pi^+}^2-m_{K_+}^2+m_{K_0}^2
}
\sim-0.8
\end{equation}
This strange result is probably a consequence of $SU(3)$ breaking
inducing eta and eta-prime mixing, thus lowering the eta mass. But
one might worry that depending on what combination of mesons one uses,
even the sign of the up quark mass is ambiguous. Attempts to extend
the naive quark mass ratio estimates to higher orders in the chiral
expansion have shown that there are fundamental ambiguities in the
definition of the quark masses \cite{Kaplan:1986ru}. An important
message of later sections is that this ambiguity is an inherent
property of QCD.
Note that in Eq.~(\ref{mesons}) the neutral pion mass squared can
become negative if
\begin{equation}
m_u< {-m_dm_s\over m_d+m_s}.
\end{equation}
This unphysical situation will result in a condensation of the pion
field and a spontaneous breaking of CP symmetry \cite{Creutz:2003xu}.
This is closely tied to the possibility of a CP violating term in QCD
that we will discuss in later sections.
\subsection{Staggered fermions}
Another fermion formulation that has an exact chiral symmetry is the
so called ``staggered'' approach. To derive this it is convenient to
begin with the ``naive'' discretization of the Dirac equation from
before. This considers fermions hopping between nearest neighbor
lattice sites while picking up a factor of $\pm i\gamma_\mu$ for a hop
in direction $\pm \mu$. Going to momentum space, the discretization
replaces powers of momentum with trigonometric functions, for example
\begin{equation}
\gamma_\mu p_\mu\rightarrow \gamma_\mu {\sin(p_\mu)}.
\end{equation}
Here we work in lattice units and thus drop factors of $a$. As
discussed before, this formulation reveals the famous ``doubling''
issue, arising because the fermion propagator has poles not only for
small momentum, but also whenever any component of the momentum is at
$\pi$. The theory represents not one fermion, but sixteen. And the
various doublers have differing chiral properties. This arises from
the simple observation that
\begin{equation}
{d\over dp}\sin(p)|_{p=\pi}=-{d\over dp}\sin(p)|_{p=0}.
\end{equation}
The consequence is that the helicity projectors $(1\pm\gamma_5)/2$ for
a travelling particle depend on which doubler one is observing.
Now consider a fermion traversing a closed loop on the lattice. As
illustrated in Fig.~\ref{loop}, the corresponding gamma matrix factors
will always involve an even number of any particular $\gamma_\mu$.
Thus the resulting product is proportional to the identity. If a
fermion starts off with a particular spinor component, it will wind up
in the same component after circumnavigating the loop. This means
that the fermion determinant exactly factorizes into four equivalent
pieces. The naive theory has an exact $U(4)$ symmetry, as pointed out
some time ago by Karsten and Smit \cite{Karsten:1980wd}. Indeed, for
massless fermions this is actually a $U(4)\otimes U(4)$ chiral
symmetry. This symmetry does not contradict any anomalies since it is
not the full naive $U(16)\otimes U(16)$ of 16 species. The chiral
symmetry generated by $\gamma_5$ remains exact, but is allowed because
it is actually a flavored symmetry. As mentioned above, the helicity
projectors for the various doubler species use different signs for
$\gamma_5$.
\begin{figure*}
\centering
\includegraphics[width=1.6in]{loop.eps}
\caption{When a fermion circumnavigates a loop in the naive
formulation, it picks up a factor that always
involves an even power of any particular gamma matrix.
Figure from Ref.~\cite{Creutz:2007rk}.}
\label{loop}
\end{figure*}
The basic idea of staggered fermions is to divide out this $U(4)$
symmetry \cite{Kogut:1974ag,Susskind:1976jm,Sharatchandra:1981si} by
projecting out a single component of the fermion spinor on each site.
Taking $\psi\rightarrow P\psi$, one projector that accomplishes this
is
\begin{equation}
P={1\over 4} \bigg(1+i\gamma_1\gamma_2 (-1)^{x_1+x_2}
+i\gamma_3\gamma_4 (-1)^{x_3+x_4}
+\gamma_5(-1)^{x_1+x_2+x_3+x_4}\bigg)
\end{equation}
where the $x_i$ are the integer coordinates of the respective lattice
sites. This immediately reduces the doubling from a factor of sixteen
to four. It is the various oscillating sign factors in this formula
that give staggered fermions their name.
At this stage the naive $U(1)$ axial symmetry remains. Indeed, the
projector used above commutes with $\gamma_5$. This symmetry is
allowed since four species, often called ``tastes,'' remain. Among
them the symmetry is a taste non-singlet; under a chiral rotation, two
rotate one way and two the other.
The next step taken by most of the groups using staggered fermions is
the rooting trick. In the hope of reducing the multiplicity down from
four, the determinant is replaced with its fourth root,
$|D|\rightarrow |D|^{1/4}$. With several physical flavors this trick
is applied separately to each. In simple perturbation theory this is
correct since each fermion loop gets multiplied by one quarter,
cancelling the extra factor from the four ``tastes.''
At this point one should be extremely uneasy: the exact chiral
symmetry is waving a huge red flag. Symmetries of the determinant
survive rooting, and thus the exact $U(1)$ axial symmetry for the
massless theory remains. For the unrooted theory this was a flavored
chiral symmetry. But, having reduced the theory to one flavor, how
can there be a flavored symmetry without multiple flavors? We will
now show why this rooting trick fails non-perturbatively when applied
to the staggered quark operator.
\subsection{Where is the parity violation?}
The standard model of elementary particle interactions is based on the
product of three gauge groups, $SU(3)\otimes SU(2) \otimes U(1)_{em}$.
Here the $SU(3)$ represents the strong interactions of quarks and
gluons, the $U(1)_{em}$ corresponds to electromagnetism, and the
$SU(2)$ gives rise to the weak interactions. We ignore here the
technical details of electroweak mixing. The full model is, of
course, parity violating, as necessary to describe observed helicities
in beta decay. This violation is normally considered to lie in the
$SU(2)$ of the weak interactions, with both the $SU(3)$ and
$U(1)_{em}$ being parity conserving. We will show here that this is
actually a convention, adopted primarily because the weak interactions
are small compared to the others. We show below that reassigning
degrees of freedom allows a reinterpretation where the $SU(2)$ gauge
interaction is vector-like. Since the full model is parity violating,
this process shifts the parity violation into the strong,
electromagnetic, and Higgs interactions. The resulting theory pairs
the left handed electron with a right handed anti-quark to form a
Dirac fermion. With a vector-like weak interaction, the chiral issues
which complicate lattice formulations now move to the other gauge
groups. Requiring gauge invariance for the re-expressed
electromagnetism then clarifies the mechanism behind one proposal for
a lattice regularization of the standard
model \cite{Creutz:1996xc,Creutz:1997fv}.
For simplicity we consider here only the first generation, which
involves four left handed doublets. These correspond to the
neutrino/electron lepton pair plus three colors for the up/down quarks
\begin{equation}
\pmatrix{\nu \cr e^-\cr}_L,
\ \pmatrix{{u^r} \cr {d^r}\cr}_L,
\ \pmatrix{{u^g} \cr {d^g}\cr}_L,
\ \pmatrix{{u^b} \cr {d^b}\cr}_L.
\end{equation}
Here the superscripts from the set $\{r,g,b\}$ represent the internal
$SU(3)$ index of the strong interactions, and the subscript $L$ indicates
left-handed helicities.
If we ignore the strong and electromagnetic interactions, leaving only
the weak $SU(2)$, each of these four doublets is
equivalent and independent. We now arbitrarily pick two of them
and do a charge conjugation operation, thus switching to their
antiparticles
\begin{equation}\matrix{
\pmatrix{{u^g} \cr {d^g}\cr}_L \longrightarrow
\pmatrix{\overline{{d^g}} \cr \overline{{u^g}}\cr}_R \cr
\cr
\pmatrix{{u^b} \cr {d^b}\cr}_L \longrightarrow
\pmatrix{\overline{{d^b}} \cr \overline{{u^b}}\cr}_R .\cr
}
\end{equation}
In four dimensions anti-fermions have the opposite helicity; so, we
label these new doublets with $R$ representing right handedness.
With two left and two right handed doublets, we can combine them into
two Dirac doublets
\begin{equation}
\pmatrix{
\pmatrix{\nu \cr e^-\cr}_L\cr
\pmatrix{\overline{{d^g}} \cr \overline{{u^g}}\cr}_R\cr
}
\qquad
\pmatrix{
\pmatrix{{u^r} \cr {d^r}\cr}_L\cr
\pmatrix{\overline{{d^b}} \cr \overline{{u^b}}\cr}_R \cr
}.
\end{equation}
\def {1\over 2} { {1\over 2} }
Formally in terms of the underlying fields, the construction takes
\begin{eqnarray}
&&\psi= {1\over 2} (1-\gamma_5)\psi_{(\nu,e^-)}+ {1\over 2} (1+\gamma_5)
\psi_{({\overline{d^g}},{\overline{u^g}})} \\
&&\chi= {1\over 2} (1-\gamma_5)\psi_{({u^r}, {d^r})}+ {1\over 2} (1+\gamma_5)
\psi_{({\overline{d^b}},{\overline{u^b}})}.
\end{eqnarray}
From the conventional point of view, these fields have rather peculiar
quantum numbers. For example, the left and right parts have different
electric charges. Electromagnetism now violates parity. The left and
right parts also have different strong quantum numbers; the strong
interactions violate parity as well. Finally, the components have
different masses; parity is violated in the Higgs mechanism.
The different helicities of these fields also have variant baryon
number. This is directly related to the known baryon violating
processes through weak ``instantons'' and axial
anomalies\cite{'tHooft:1976fv}. When a topologically non-trivial weak
field is present, the axial anomaly arises from a level flow out of
the Dirac sea \cite{Ambjorn:1983hp}. This generates a spin flip in the
fields, {\it i.e.} $e^-_L \rightarrow ({\overline{u^g}})_R$. Because
of the peculiar particle identification, this process does not
conserve charge, with $\Delta Q= -{2\over 3} +1={1\over 3}$. This
would be a disaster for electromagnetism were it not for the fact that
simultaneously the other Dirac doublet also flips {${d^r}_L
\rightarrow ({\overline{u^b}})_R$} with a compensating $\Delta Q =
-{1\over 3}$. This is anomaly cancellation, with the total $\Delta Q
= {1\over 3}-{1\over 3}=0$. Only when both doublets are considered
together is the $U(1)$ symmetry restored. In this anomalous process
baryon number is violated, with $L+Q\rightarrow \overline Q +
\overline Q$. This is the famous `` `t Hooft vertex'' \cite {'tHooft:1976fv}
discussed earlier in the context of the strong interactions.
\subsection {A lattice model}
The above discussion on twisting the gauge groups has been in the
continuum. Now we turn to the lattice and show how this picture leads
to a possible lattice model for the strong interactions, albeit with
an unusual added coupling that renders the treatment quite difficult
to make rigorous \cite{Creutz:1996xc,Creutz:1997fv}. Whether this
model is viable remains undecided, but it does incorporate many of the
required features.
For this we use the domain wall approach for the fermions
\cite{Kaplan:1992bt,Shamir:1993zy}.
As discussed earlier, in this picture, our four dimensional world is a
``4-brane'' embedded in 5-dimensions. The complete lattice is a five
dimensional box with open boundaries, and the parameters are chosen so
the physical quarks and leptons appear as surface zero modes. The
elegance of this scheme lies in the natural chirality of these modes
as the size of the extra dimension grows. With a finite fifth
dimension one doubling remains, coming from interfaces appearing as
surface/anti-surface pairs. It is natural to couple a four
dimensional gauge field equally to both surfaces, giving rise to a
vector-like theory.
We now insert the above pairing into this five dimensional scheme. In
particular, consider the left handed electron as a zero mode on one
wall and the right handed anti-green-up-quark as the partner zero mode
on the other wall, as sketched in Fig.~\ref{fig:1}. This provides a
lattice regularization for the $SU(2)$ of the weak interactions.
\begin{figure}
\centering
\includegraphics[width=2.5in]{{ehopping.eps}}
\caption{Pairing the electron with the anti-green-up-quark.
Figure taken from \cite{Creutz:1997fv}.}
\label{fig:1}
\end{figure}
However, since these two particles have different electric charge,
$U(1)_{EM}$ must be broken in the interior of the extra dimension. We
now proceed in analogy to the ``waveguide''
picture\cite{Golterman:1993th} and restrict this charge violation to
$\Delta Q$ to one layer at some interior position $x_5=i$. Using
Wilson fermions, the hopping term from $x_5=i$ to $i+1$
\begin{equation}
\overline\psi_{i}P\psi_{i+1}\qquad{(P=(\gamma_5+r)/2)}
\end{equation}
is a $Q=1/3$ operator. At this layer, electric charge is not
conserved. This is unacceptable and needs to be fixed.
To restore the $U(1)$ symmetry one must transfer the charge from
$\psi$ to the compensating doublet $\chi$. For this we replace
the sum of hoppings with a product on the offending layer
\begin{equation}
\overline\psi_{i}P\psi_{i+1}
{+}\overline\chi_{i}P\chi_{i+1}
{\longrightarrow}
\overline\psi_{i}P\psi_{i+1}
{\times}\overline\chi_{i}P\chi_{i+1}.
\end{equation}
This introduces an electrically neutral four-fermi operator. Note
that it is baryon violating, involving a ``lepto-quark/diquark''
exchange, as sketched in Fig.~\ref{fig:2}. One might think of the
operator as representing a ``filter'' at $x_5=i$ through which only
charge compensating pairs of fermions can pass.
\begin{figure}
\centering
\includegraphics[width=2.5in]{{transfer.eps}}
\caption {Transferring charge between the doublets.
Figure taken from \cite{Creutz:1997fv}.}
\label{fig:2}
\end{figure}
In five dimensions there is no chiral symmetry. Even for the free
theory, combinations like $\overline\psi_{i}P\psi_{i+1} $ have
non-vanishing vacuum expectation values. We use such as a
``tadpole,'' with $\chi$ generating an effective hopping for $\psi$
and {\it vice versa}.
Actually the above four fermion operator is not quite sufficient for
all chiral anomalies, which can also involve right handed
singlet fermions. To correct this we need explicitly include the right
handed sector, adding similar four fermion couplings
(also electrically neutral). The main difference is that this sector
does not couple to the weak bosons.
Having fixed the $U(1)$ of electromagnetism, we restore the strong
$SU(3)$ with an anti-symmetrization of the quark color indices in the
new operator, $ {Q^r}{Q^g}{Q^b}{\longrightarrow
\epsilon^{\alpha\beta\gamma}Q^\alpha Q^\beta Q^\gamma}$. Note that
similar left-right inter-sector couplings are needed to correctly
obtain the effects of topologically non-trivial strong gauge fields.
An alternative view of this picture folds the lattice about the
interior of the fifth dimension, placing all light modes on one wall
and having the multi-fermion operator on the other. This is the model
of Ref.~\cite {Creutz:1996xc}, with the additional inter-sector couplings
correcting a technical error \cite{Neuberger:1997cz}.
Unfortunately the scheme is still not completely rigorous. In
particular, the non-trivial four-fermion coupling represents a new
defect and we need to show that this does not give rise to unwanted
extra zero modes. Note, however, that the five dimensional mass is
the same on both sides of defect; thus there are no topological
reasons for such.
A second worry is that the four fermion coupling might induce an
unwanted spontaneous symmetry breaking of one of the gauge symmetries.
We need to remain in a strongly coupled paramagnetic phase without
spontaneous symmetry breaking. Ref.~\cite{Creutz:1996xc} showed that
strongly coupled zero modes do preserve the desired symmetries, but
the analysis ignored contributions from heavy modes in the fifth
dimension.
Assuming all works as desired, the model raises several other
interesting questions. As formulated, we needed a right handed
neutrino to provide all quarks with partners. Is there some variation
that avoids this particle, which decouples from all gauge fields in
the continuum limit? Another question concerns possible numerical
simulations; is the effective action positive? Finally, we have used
the details of the usual standard model, leaving open the question of
whether this model is somehow special. Can we always find an
appropriate multi-fermion coupling to eliminate undesired modes in
other chiral theories where anomalies are properly canceled?
\subsection{Twisted tilting}
Conventionally the mass tilts the potential downward in the positive
$\sigma$ direction. However, it is an interesting exercise to
consider tilts in other directions in the $\sigma,\eta^\prime$ plane.
This is accomplished with an anomalous rotation on the mass term
\begin{eqnarray}
-m\overline\psi\psi&\rightarrow&
-m\cos(\phi)\overline\psi\psi
-im\sin(\phi)\overline\psi\gamma_5\psi\cr
&\sim& -m\cos(\phi)\sigma+m\sin(\phi)\eta^\prime.
\end{eqnarray}
Were it not for the anomaly, this would just be a redefinition of
fields. However the same effect that gives the $\eta^\prime$ its mass
indicates that this new form for the mass term gives an inequivalent
theory. As $i\overline\psi\gamma_5\psi$ is odd under CP, this theory
is explicitly CP violating.
The conventional notation for this effect involves the angle
$\Theta=N_f\phi$. Then the $Z_{N_f}$ symmetry amounts to a $2\pi$
periodicity in $\Theta$. As Fig.~\ref{potential} indicates, at
special values of the twisting angle $\phi$, there will exist two
degenerate minima. This occurs, for example, at $\phi=\pi/N_f$ or
$\Theta=\pi$. As the twisting increases through this point, there
will be a first order transition as the true vacuum jumps from the
vicinity of one minimum to the next.
\begin{figure*}
\centering
\includegraphics[width=2.5in]{potential.eps}
\caption{\label{potential}
With massive quarks and a twisting angle of $\phi=\pi/N_f$, two of the
minima in the $\sigma,\eta^\prime$ plane become degenerate. This
corresponds to a first order transition at $\Theta=\pi$.
}
\end{figure*}
Because of the $Z_{N_f}$ symmetry of the massless theory, all the
$N_f$ separate minima are physically equivalent. This means that if
we apply our mass term in the direction of any of them, we obtain the
same theory. In particular, for four flavors the usual mass term
$m\overline \psi \psi$ is equivalent to using the alternative mass
term $i m\overline\psi\gamma_5\psi$. This result, however, is true if
and only if $N_f$ is a multiple of four.
\subsection{Odd $N_f$}
One interesting consequence of this picture concerns QCD with an odd
number of flavors. The group $SU(N_f)$ with odd $N_f$ does not
include the element $-1$. In particular, the $Z_{N_f}$ structure is
not symmetric under reflections about the $\eta^\prime$ axis.
Figure \ref{potential2} sketches the situation for $SU(3)$. One
immediate conclusion is that positive and negative mass are not
equivalent. Indeed, a negative mass with three degenerate flavors
corresponds to the $\Theta=\pi$ case and a spontaneous breaking of CP
is expected. In this case there is no symmetry under taking
$\sigma\sim \overline\psi \psi$ to its negative. The simple picture
sketched in Fig.~\ref{v1} no longer applies.
At $\Theta=\pi$ the theory lies on top of a first order phase
transition line. A simple order parameter for this transition is the
expectation value for the $\eta^\prime$ field. As this field is odd
under CP symmetry, this shows that negative mass QCD with an odd
number of flavors spontaneously breaks CP.\footnote{Dashen's
original paper \cite{Dashen:1970et} speculates that this might be
related to the parity breaking seen in nature. This presumably
requires a new ``beyond the standard model'' interaction rather than
QCD.} This does not contradict the Vafa-Witten theorem
\cite{Vafa:1984xg} because in this regime the fermion determinant is
not positive definite.
Note that the asymmetry in the sign of the quark mass is not easily
seen in perturbation theory. Any quark loop in a perturbative diagram
can have the sign of the quark mass flipped by a $\gamma_5$
transformation. It is only through the subtleties of regulating the
divergent triangle diagram
\cite{Adler:1969gk,Adler:1969er,Bell:1969ts} that the sign of the mass
enters.
A remarkable conclusion of these observations is that two physically
distinct theories can have identical perturbative expansions. For
example, with flavor $SU(3)$ the negative mass theory has spontaneous
$CP$ violation, while the positive mass theory does not. Yet both
cases have exactly the same perturbation theory. This dramatically
demonstrates what we already knew: non-perturbative effects are
essential to understanding QCD.
\begin{figure*}
\centering
\includegraphics[width=2.5in]{potential2.eps}
\caption{\label{potential2} For odd $N_f$, such as the three flavor
case sketched here, QCD is not symmetric under changing the sign of
the quark mass. Negative mass corresponds to taking $\Theta=\pi$.
}
\end{figure*}
A special case of an odd number of flavors is one-flavor QCD
\cite{Creutz:2006ts}. In this case the anomaly removes all chiral
symmetry and there is a unique minimum in the $\sigma,\eta^\prime$
plane, as sketched in Fig.~\ref{potential3}. This minimum does not
occur at the origin, being shifted to $\langle \overline\psi
\psi\rangle > 0$ by the 't Hooft vertex, which for one flavor is just
an additive mass shift \cite{Creutz:2007yr}. Unlike the case with
more flavors, the resulting expectation value for $\sigma$ is not from
a spontaneous symmetry breaking; indeed, there is no chiral symmetry
to break in one flavor QCD. Any regulator that preserves a remnant of
chiral symmetry must inevitably fail \cite{Creutz:2008nk}. Note also
that for one-flavor QCD there is no longer the necessity of a first
order phase transition at $\Theta=\pi$. It has been argued
\cite{Creutz:2006ts} that for finite quark mass such a transition
should still occur if the mass is sufficiently negative, but the region
around vanishing mass is not expected to show any singularity.
\begin{figure*}
\centering
\includegraphics[width=2.5in]{potential3.eps}
\caption{ The effective potential for one-flavor QCD with small quark
mass has a unique minimum in the $\sigma,\eta^\prime$ plane. The
minimum is shifted from zero due to the effect of the 't Hooft
vertex. }
\label{potential3}
\end{figure*}
An unusual feature of one-flavor QCD is that the renormalization of
the quark mass is not multiplicative when non-perturbative effects are
taken into account. The additive mass shift is generally scheme
dependent since the details of the instanton effects depend on scale.
This is the basic reason that a massless up quark is not a possible
solution to the strong CP problem \cite{Creutz:2003xc}. Later we will
discuss this in more detail in the context of the two flavor theory
with non-degenerate masses.
Because of this shift in the mass, the conventional parameters
$\Theta$ and $m$ are singular coordinates for the one-flavor theory.
A cleaner set of variables would be the coefficients of the two
possible mass terms $\overline\psi \psi$ and $i\overline\psi
\gamma_5\psi$ appearing in the Lagrangean. The ambiguity in the quark
mass is tied to rough gauge configurations with ambiguous winding
number. This applies even to the formally elegant overlap operator
that we will discuss later; when rough gauge fields are present, the
existence of a zero mode can depend on the detailed fermion operator
in use. Smoothness conditions imposed on the gauge fields to remove
this ambiguity appear to conflict with fundamental principles, such as
reflection positivity \cite{Creutz:2004ir}.
The $Z_{N_f}$ symmetry discussed here is a property of the fermion
determinant and is independent of the gauge field dynamics. In Monte
Carlo simulation language, this symmetry appears configuration by
configuration. With $N_f$ flavors, we always have $|D|=|e^{2\pi
i/N_f} D|$ for any gauge field. This discrete chiral symmetry is
inherently discontinuous in $N_f$. This non-continuity lies at the
heart of the issues with the rooted staggered quark approximation. We
will return to this topic in a later section.
\subsection{Fermions in higher representations}
\label{reps}
When the quarks are massless, the classical field theory corresponding
to the strong interactions has a $U(1)$ axial symmetry under the
transformation
\begin{equation}
\psi\rightarrow
e^{i\theta\gamma_5}\psi\qquad
\overline\psi\rightarrow \overline\psi e^{i\theta\gamma_5}.
\label{thetarot}
\end{equation}
It is the 't Hooft vertex that explains how this symmetry does not
survive quantization. In this subsection we discuss how when the
quarks are in non-fundamental representations of the gauge group,
discrete subgroups of this symmetry can remain because of additional
zeros in the Dirac operator.
While these considerations do not apply to the usual theory of the
strong interactions where the quarks are in the fundamental
representation, there are several reasons to study them anyway. At
higher energies, perhaps as being probed at the Large Hadron Collider,
one might well discover new strong interactions that play a
substantial role in the spontaneous breaking of the electroweak
theory. Also, many grand unified theories involve fermions in
non-fundamental representations. As one example, we will see that
massless fermions in the 10 representation of $SU(5)$ possess a $Z_3$
discrete chiral symmetry. Similarly the left handed 16 covering
representation of $SO(10)$ gives a chiral gauge theory with a
surviving discrete $Z_2$ chiral symmetry. Understanding these
symmetries may eventually play a role in a discretization of chiral
gauge theories on the lattice.
Here we are generalizing the index theorem relating gauge field
topology to zero modes of the Dirac operator. In particular, fermions
in higher representations can involve multiple zero modes for a
given winding. Being generic, consider representation $X$ of a gauge
group $G$. Denote by $N_X$ the number of zero modes that are required
per unit of winding number in the gauge fields. That is, suppose the
index theorem generalizes to
\begin{equation}
n_r-n_l=N_X\nu
\end{equation}
where $n_r$ and $n_l$ are the number of right and left handed zero
modes, respectively, and $\nu$ is the winding number of the associated
gauge field. The basic 't Hooft vertex receives contributions from
each zero mode, resulting in an effective operator which is a product
of $2N_X$ fermion fields. Schematically, the vertex is modified along
the lines $\overline\psi_L \psi_R \longrightarrow (\overline\psi_L
\psi_R)^{N_X}$. While this form still breaks the $U(1)$ axial
symmetry, it is invariant under $\psi_R\rightarrow e^{2\pi
i/N_X}\psi_R$. In other words, there is a $Z_{N_X}$ discrete axial
symmetry.
There are a variety of convenient tools for determining $N_X$.
Consider building up representations from lower ones. Take two
representations $X_1$ and $X_2$ and form the direct product
representation $X_1\otimes X_2$. Let the matrix dimensions for $X_1$
and $X_2$ be $D_1$ and $D_2$, respectively. Then for the product
representation we have
\begin{equation}
N_{X_1\otimes X_2}= N_{X_1} D_{X_2}+N_{X_2} D_{X_1}.
\end{equation}
To see this, start with $X_1$ and $X_2$ representing two independent
groups $G_1$ and $G_2$. With $G_1$ having winding, there will be a
zero mode for each of the dimensions of the matrix index associated
with $X_2$. Similarly there will be multiple modes for winding in
$G_2$. These modes are robust and all should remain if we now
constrain the groups to be the same.
As a first example, denote the fundamental representation of $SU(N)$
as $F$ and the adjoint representation as $A$. Then using $\overline F
\otimes F = A\oplus 1$ in the above gives $N_A=2N_F$, as noted some
time ago \cite{Witten:1982df}. With $SU(3)$, fermions in the
adjoint representation will have six-fold degenerate zero modes.
For another example, consider $SU(2)$ and build up towards arbitrary
spin $s\in\{0,{1\over 2}, 1, {3\over 2},\ldots\}$. Recursing the
above relation gives the result for arbitrary spin
\begin{equation}
N_s=s(2s+1)(2s+2)/3.
\end{equation}
Another technique for finding $N_X$ in more complicated groups begins
by rotating all topological structure into an $SU(2)$ subgroup and
then counting the corresponding $SU(2)$ representations making up the
larger representation of the whole group. An example to illustrate
this procedure is the antisymmetric two indexed representation of
$SU(N)$. This representation has been extensively used in
\cite{Corrigan:1979xf,Armoni:2003fb,Sannino:2003xe,Unsal:2006pj} for
an alternative approach to the large gauge group limit. The basic
$N(N-1)/2$ fermion fields take the form
\begin{equation}
\psi_{ab}=-\psi_{ba}, \qquad a,b\in 1,2,...N.
\end{equation}
Consider rotating all topology into the $SU(2)$ subgroup involving the
first two indices, i.e. 1 and 2. Because of the anti-symmetrization,
the field $\psi_{12}$ is a singlet in this subgroup. The field pairs
$(\psi_{1,j},\psi_{2,j})$ form a doublet for each $j\ge 3$. Finally,
the $(N-2)(N-3)/2$ remaining fields do not transform under this
subgroup and are singlets. Overall we have $N-2$ doublets under the
$SU(2)$ subgroup, each of which gives one zero mode per winding
number. We conclude that the 't Hooft vertex leaves behind a
$Z_{N-2}$ discrete chiral symmetry. Specializing to the 10
representation of $SU(5)$, this is the $Z_3$ mentioned earlier.
Another example is the group $SO(10)$ with fermions in the 16
dimensional covering group. This forms the basis of a rather
interesting grand unified theory, where one generation of fermions is
placed into a single left handed 16 multiplet \cite{Georgi:1979dq}.
This representation includes two quark species interacting with the
$SU(3)$ subgroup of the strong interactions, Rotating a topological
excitation into this subgroup, we see that the effective vertex will
be a four fermion operator and preserve a $Z_2$ discrete chiral
symmetry.
It is unclear whether these discrete symmetries are expected to be
spontaneously broken. Since they are discrete, such breaking is not
associated with Goldstone bosons. But the quark condensate does
provide an order parameter; so, when $N_X>1$, any such breaking would
be conceptually meaningful. This could be checked in numerical
simulations.
|
\section{Conformal generator for arbitrary $n>1$} \label{Appendix A}
In this appendix, we derive the conformal generator $K$ for Lagrangian in (\ref{eq1})
\beq \label{eqL}
L=\frac{1}{n}\dot{q}^nq^{n-2}
\eeq
for arbitrary $n>1$ and two different orderings of the Hamiltonian $H$ which have the same classical transformation law of $q$ with different Casimirs $\mathcal{C}$.
\subsection{$H=\frac{N}{N+1} p^{\frac{1}{2}\left(1+\frac{1}{N}\right)}q^{-1+\frac{1}{N}} p^{\frac{1}{2}\left(1+\frac{1}{N}\right)}$, $N \equiv n-1$}
We start from the classical Hamiltonian $H$:
\begin{subequations}
\beq \label{eqB0}
\label{eqB0a} H=&&\frac{N}{N+1} p^{1+\frac{1}{N}} q^{-1+\frac{1}{N}} , \\
\label{eqB0b} p=&& \dot{q}^N q^{N-1}.
\eeq
\end{subequations}
Then we write (\ref{eqB0}) and time dilatation operator $D$ associated with Lagrangian (\ref{eqL}) quantum mechanically with the ordering
\begin{subequations} \label{eqB1}
\beq
\label{eqB1a} H=&
\frac{N}{N+1} p^{\frac{1}{2}\left(1+\frac{1}{N}\right)} q^{-1+\frac{1}{N}} p^{\frac{1}{2}\left(1+\frac{1}{N}\right)} \\
\label{eqB1b} D=&& t H - \frac{1}{4}(pq+qp) \equiv t H+D_0,
\eeq
and using (\ref{eq11}) for conformal operator $K$:
\beq
\label{eqB1c}
=&&t^2 H+2tD_0 +K_0.
\eeq
\end{subequations}
After a straightforward calculation, the $K_0$ is:
\beq \label{eqB2}
K_0&& = D_0 H^{-1} D_0 + \frac{\mathcal{C}}{H} \nonumber \\
=&&\frac{(N+1)}{4N} \ p^{\frac{1}{2}\left(1-\frac{1}{N}\right)} q^{3-\frac{1}{N}} p^{\frac{1}{2}\left(1-\frac{1}{N}\right)}+\left( \frac{ \hbar^2 } {16} \left(\frac{4}{N}-\frac{1}{N^2}\right)+\mathcal{C}\right) \frac{1}{H}.
\eeq
By choosing Casimir $\mathcal{C}$
\beq
\label{eqB4} \mathcal{C}=&& - \frac{ \hbar^2 } {16} \left(\frac{4}{N}-\frac{1}{N^2}\right),
\eeq
we eliminate the $H^{-1}$ term in $K_0$:
\beq
\label{eqB3} K_0
&&\frac{(N+1)}{4N} \ p^{\frac{1}{2}\left(1-\frac{1}{N}\right)} q^{3-\frac{1}{N}} p^{\frac{1}{2}\left(1-\frac{1}{N}\right)}.
\eeq
Note from (\ref{eqB3}) that only $n=2$ $(N=1)$ and $n=4/3$ $(N=1/3)$ produce a primary operator.
The transformation law for $q$ can be reproduced by computing
\begin{subequations} \label{eqB5}
\beq \label{eqB5a}
\delta_T q =&&\frac{i}{\hbar} \left[ H, q \right
\nonumber \\
=&&\frac{1}{2} \left( p^{\frac{1}{2}\left(1+\frac{1}{N}\right)} q^{-1+\frac{1}{N}} p^{\frac{1}{2}\left(-1+\frac{1}{N}\right)}+p^{\frac{1}{2}\left(-1+\frac{1}{N}\right)} q^{-1+\frac{1}{N}} p^{\frac{1}{2}\left(1+\frac{1}{N}\right)}\right) \equiv \dot{q} \\
\label{eqB5b}
\delta_D q = && \frac{i}{\hbar} \left[ tH+D_0, q \right]
= t\dot{q}-\frac{1}{2}q \\
\label{eqB5c}
\delta_C q =&& \frac{i}{\hbar} \left[ t^2H+2tD_0+K_0, q \right] \nonumber \\
=&& t^2 \dot{q
-tq +\frac{N^2-1}{8N^2} p^{-\frac{1}{2}\left(1+\frac{1}{N}\right)}\left(p q^{3-\frac{1}{N}}+q^{3-\frac{1}{N}} p\right) p^{-\frac{1}{2}\left(1+\frac{1}{N}\right)}
\eeq
We then obtain the classical transformation law of $q$ by using (\ref{eqB0b}):
\beq
\label{eqB5d} \delta_C q = && t^2 \dot{q}-tq+ \frac{1}{4} \left(1-\frac{1}{N^2}\right) \dot{q}^{-1}q^2
\eeq
\end{subequations}
\subsection{$H=\frac{N}{N+1} q^{\frac{1}{2}\left(-1+\frac{1}{N}\right)} p^{1+\frac{1}{N}}q^{\frac{1}{2}\left(-1+\frac{1}{N}\right)}$ }
We change (\ref{eqB1a}) to
\beq \label{eqB7}
H
\frac{N}{(N+1)} q^{\frac{1}{2}\left(-1+\frac{1}{N}\right)} p^{1+\frac{1}{N}} q^{\frac{1}{2}\left(-1+\frac{1}{N}\right)} .
\eeq
The generator $K_0$ is:
\beq \label{eqB8}
K_0 && =D_0 \frac{1}{H} D_0+\frac{\mathcal{C}}{H} \nonumber \\
=&&\frac{N+1}{4N} \ q^{\frac{1}{2}\left(3-\frac{1}{N}\right)} p^{1-\frac{1}{N}} q^{\frac{1}{2}\left(3-\frac{1}{N}\right)} +\frac{\hbar^2}{16}\left( \left(4-\frac{1}{N^2}\right)+\mathcal{C}\right) \frac{1}{H} .
\eeq
The choice of Casimir $\mathcal{C}$
\beq \label{eqB9}
\mathcal{C}= - \frac{\hbar^2}{16}\left(4-\frac{1}{N^2}\right) ,
\eeq
leads to
\beq \label{eqB10}
K_0=\frac{N+1}{4N} \ q^{\frac{1}{2}\left(3-\frac{1}{N}\right)} p^{1-\frac{1}{N}} q^{\frac{1}{2}\left(3-\frac{1}{N}\right)}.
\eeq
Following the same procedure, we obtain the classical transformation law of $q$
\begin{subequations} \label{eqB11}
\beq
\label{eqB11a} \delta_T q =&& \dot{q} \\
\label{eqB11b} \delta_D q = && t\dot{q}-\frac{1}{2}q \\
\label{eqB11c} \delta_C q = && t^2 \dot{q}-tq+ \frac{1}{4} \left(1-\frac{1}{N^2}\right) \dot{q}^{-1}q^2
\eeq
\end{subequations}
which are the same as (\ref{eqB5}).
We can see the variation of Lagrangian (\ref{eqL}) under transformation (\ref{eqB11c})
\beq \label{eqB12}
\delta_C L=&&\frac{\partial L}{\partial \dot{q}} \delta_C \dot{q} + \frac{\partial L}{\partial q} \delta_C q \nonumber \\
=&&\frac{d}{dt}\left(t^2 L - \frac{(N+1)^2}{4N^2} \dot{q}^{N-1}q^{N+1} \right)
\eeq
is a total time-derivative. The constant of motion is
\beq \label{eqB13}
K=&&\frac{\partial L}{\partial \dot{q}} \delta_C q -\left(t^2 L - \frac{(N+1)^2}{4N^2} \dot{q}^{N-1}q^{N+1} \right) \nonumber \\
=&& \frac{t^2 N}{N+1} p^{1+\frac{1}{N}} q^{-1+\frac{1}{N}} -tpq + \frac{N+1}{4N} p^{1-\frac{1}{N}}q^{3-\frac{1}{N}}
\eeq
This is classical form of (\ref{eqB1c}) with $K_0$ given in (\ref{eqB3}) (and (\ref{eqB10})).
\section{Another ordering of Hamiltonian for $n=4/3$} \label{Appendix B}
In this appendix, we use for $H$ an expression, which is ordered differently from (\ref{eq17a}):
\beq \label{eqA1}
H=\frac{1}{4} q p^4 q.
\eeq
Using (\ref{eq13}), we obtain
\beq \label{eqA2}
K_0= p^{-2} + \left(\frac{-5 \hbar^2}{16}+\mathcal{C}\right)\frac{1}{H} .
\eeq
Hence, we choose Casimir to be
\beq \label{eqA4}
\mathcal{C}=\frac{5 \hbar^2}{16}.
\eeq
and have
\beq \label{eqA3}
K_0= p^{-2} .
\eeq
Comparing (\ref{eq19a}) with (\ref{eqA3}), we observe the two $K_0$'s are the same in the two orderings. The Casimirs differ because of reordering.
\begin{acknowledgements}
I am grateful to Professor Roman Jackiw for providing to me the main idea of this work
and continuous and patient guidance. This work is supported by the National Science Council of R.O.C. under Grant number:
NSC98-2917-I-564-122.
\end{acknowledgements}
|
\section{Introduction}
In a hierarchical process of galaxy formation \citep{2000MNRAS.319..168C},
within the currently favored $\Lambda$-CDM cosmological
framework, the present properties of galaxies are decided by their individual
histories of being assembled from smaller halos \citep{1993MNRAS.262..627L}.
Moreover, these galaxies evolve continuously under the influence of self-gravity, rotation,
accretion and feedback.
However, \citet{2008Natur.455.1082D} have recently reported surprising
correlations among the properties of galaxies, to the extent of forming a
one-parameter set lying on a single fundamental line \citep{2009MNRAS.394..340G}.
\citet{2008Natur.455.1049V} has argued that such simplicity
is hard to explain within the paradigm of hierarchical galaxy mergers.
One of the puzzling results
\citep{1983AJ.....88..881G,1984AJ.....89..758H,2001A&A...370..765V,2003ApJ...585..256R,2009MNRAS.394..340G},
is that the neutral hydrogen mass across these galaxies scales almost
linearly with surface area, implying that widely different galaxies
share very similar neutral hydrogen surface densities.
This nearly constant HI surface density has been pointed out in literature
to be an intriguing puzzle, demanding an explanation.
In this work we argue that correlation between the neutral hydrogen mass
and surface area may be preserved, despite the complex
merger histories of the galaxies, by self-regulated star formation,
driven by the competition between gravitational instabilities in the
rotating disk and mechanical feedback from supernovae.
When mergers drive a galaxy away from the fundamental line, self
regulation of the porosity of the ISM and gravitational instability of
the star forming disk, as proposed by \citet{1997ApJ...481..703S},
can bring the neutral hydrogen surface density
back to the value which is predicted in our simple model.
This can explain
the regulation of the neutral hydrogen surface density in galaxies,
explaining part of the surprising simplicity in the observed properties
of galaxies.
\section{Surface density of gas in galaxies}
Evidence of a nearly universal neutral hydrogen (HI) surface density
in galaxies has been accumulating over the past three decades.
The initial hints came from single-dish 21 cm radio observations
of nearby galaxies.
\citet{1983AJ.....88..881G} reported neutral hydrogen (HI) observations of 24 galaxies
in the Virgo cluster, using the Arecibo telescope.
The HI sizes and masses were found to be correlated in the same
manner, irrespective of whether they were HI rich or HI poor.
A similar correlation was soon found between the optical sizes and
HI masses of 288 isolated galaxies \citep{1984AJ.....89..758H}.
It was also found that the optical diameters of spiral disk are
better correlated with the HI mass than the morphological type
\citep{1984AJ.....89..758H}.
A deeper survey by \citet{2001A&A...370..765V} has revealed similar
correlations for galaxies in the Ursa Major Cluster.
The Arecibo Dual-Beam Survey \citep{2000ApJS..130..177R} (ADBS) has
found HI in 265 galaxies in a ``blind'' survey of $\sim430$deg$^2$ of sky.
While most of the ADBS galaxies were unresolved at the resolution
of the Arecibo, the Very Large Array (VLA) was used for interferometric
mapping of 84 galaxies and determine accurate sizes of 50 of them.
A comparison of HI masses and HI sizes of these along with 53 galaxies
with high resolution maps from literature revealed that they were
consistent with a nearly constant average HI surface density of the order
$\sim10^7M_\odot$ kpc$^{-2}$ \citep{2003ApJ...585..256R}.
The HI masses of individual ADBS galaxies, spanning 3 orders of
magnitude, deviate by only $\sim0.13$ dex ($1\sigma$) from those
expected from a constant HI surface density. This puts the evidence, for a
regulated HI surface density across galaxies, on a firm observational basis.
A recent study of HI selected galaxies (free from optical selection effects)
found using the Parkes radio telescope and identified with SDSS sources,
has shown that 6 observed parameters, namely the dynamical mass
($M_d$), HI mass ($M_{HI}$), luminosity, color, and two radii
containing $50\%$ and $90\%$ of the observed luminosity, have 5
independent correlations among themselves \citep{2008Natur.455.1082D}.
This implies that the galaxies form a single parameter family and are
not removed significantly from their fundamental
line by the diverse merger histories that they would have had in
the process of hierarchical galaxy formation.
Some of the correlations have already been widely
known and discussed in other forms, such as the
correlation \citep{1996A&A...312..397G} between luminosity and dynamical mass.
\citet{2008Natur.455.1082D,2009MNRAS.394..340G} report a tight linear
correlation in the already established relation between the HI mass and
the surface area. The nearly constant HI surface density is argued in
literature to be an intriguing puzzle which demands an
explanation \citep{2009MNRAS.394..340G}. We discuss below, a physically
motivated explanation for this important observation.
\section{Gravitational instability in disk galaxies}
Below a surface density threshold, azimuthally integrated star
formation in giant HII regions across a galaxy
ceases \citep{1989ApJ...344..685K}.
The existence of this threshold is traditionally stated in the
form of the Toomre parameter for gravitational instability
\citep{1968dms..book.....S}. Below a Toomre parameter $Q\lesssim1$,
rotation support and gas pressure cannot stabilize a thin self-gravitating
disk against gravitational instability
\citet{1960AnAp...23..979S,1964ApJ...139.1217T}.
Here $Q$ is defined as
\begin{equation}
Q\equiv \mu _{cr} / \mu _{gas} ,
\end{equation}
where the critical gas surface density is given by
$\mu _{cr} \equiv \Omega \sigma_g / \pi G$, in terms of the angular
velocity $\Omega$ and velocity dispersion $\sigma_g$ in the gas disk.
\citet{2004ApJ...609..667S} has demonstrated that results from detailed
considerations including not only self-gravity, but metals, dust and
UV radiation, coincide with this empirically derived surface density
threshold for star formation.
This condition depends on the local angular velocity, which
we may define the as $\mathbf{\Omega}=\nabla \times \mathbf{V}$.
Using cylindrical polar coordinates we can now write down the curl operator as
\begin{align}\nonumber
\nabla \times \mathbf{A} = & \left({1 \over \rho}{\partial A_z \over \partial \phi}
- {\partial A_\phi \over \partial z}\right) \boldsymbol{\hat \rho} +
\left({\partial A_\rho \over \partial z} - {\partial A_z \over \partial \rho}\right)
\boldsymbol{\hat \phi} \\
& + {1 \over \rho}\left({\partial \left( \rho A_\phi \right) \over \partial \rho}
- {\partial A_\rho \over \partial \phi}\right) \boldsymbol{\hat z} .
\end{align}
Assuming, rotation and translation symmetry along $\boldsymbol{\hat \phi}$
and $\boldsymbol{\hat z}$ respectively, if the bulk motion of the gas is only
in the $\boldsymbol{\hat \phi}$ direction (as would be the case if the motion
is purely Keplarian) we may write down the angular velocity as
\begin{equation}
\Omega=\left|\mathbf{\Omega}\right| = \left|\nabla \times \mathbf{V} \right|
= \left|{1 \over r}
\left({\partial \left( r V (r) \right) \over \partial r}\right)
\boldsymbol{\hat z}\right| ,
\end{equation}
where the radial coordinate $\rho$ has been replaced by $r$, to avoid
confusion with density. In nearly rigid rotors such as dwarf galaxies,
the velocity is given by $V (r)=\Omega_0 r$. In such a case,
$\left|\nabla \times \mathbf{V} \right|$ is given by $2\Omega_0$
\citep[problem 2.4.7]{2005mmp..book.....A}. For flat ($V(r)=V_0$) rotation curves
$\left|\nabla \times \mathbf{V} \right|=V_0/r$. Hence, for
all these cases, at boundary of the starforming disk, the local value
of the angular velocity is comparable to the global value. Hence, following
\citet{1997ApJ...481..703S} we shall use the global angular velocity
of the disk in the rest of this work.
The
global
angular velocity can now be expressed as
\begin{equation}
\Omega_0=\sqrt{\frac{G M_d}{R^3}}
\end{equation}
in terms of the enclosed dynamical mass within the gas disk. Replacing
$M_d=4 \pi \rho_d R^3 / 3$, where $\rho_d$ is the nearly universal
dynamical mass density \citep{2008Natur.455.1082D,2009MNRAS.394..340G},
we have $\Omega_0=\sqrt{4 \pi G \rho_d / 3}$. Substituting, we get
\begin{equation}\label{mucr}
\mu _{cr} = \frac{\sigma_g}{\pi} \sqrt{\frac{4 \pi \rho_d}{3 G}}.
\end{equation}
Now, we note that for the star forming disk to be long lived, it
may only be marginally unstable. Q is seen to be $\sim1$ throughout
the star forming regions in well observed disk
galaxies \citep{1989ApJ...344..685K}. This implies that throughout
the star forming disk the mean gas surface density is of order
$\mu _{gas} \sim \mu _{cr}$. Hence, the mean surface density of gas
is controlled primarily by the dynamical mass density $\rho_d$ and
the gas velocity dispersion $\sigma_g$.
One of the tight correlations seen by
\citet{2008Natur.455.1082D,2009MNRAS.394..340G} is
between the dynamical mass and cube of the optical radius, implying
a roughly constant average dynamical mass density (within the radii
of their star forming disks) for galaxies.
It is expected that Dark Matter contributes most of the gravitational mass
of the galaxies, hence a physical explanation of this observation would require
an understanding of the nature of Dark Matter.
Whether or not hierarchical halo build-up within a $\Lambda$-CDM cosmology
can explain this observation is still under investigation.
\citet{2010arXiv1011.6374L} have suggested that cold dark matter particles
interacting through a Yukawa potential could make halos beyond a certain
critical density evaporate over an Hubble time. This may set a characteristic
scale to the peak density of dark matter halos. Note however that the dark
matter halos of galaxies are much larger than the sizes of their star forming
disks. As a result, most of the dark matter may lie further out.
\citet{1987AJ.....93..816K} points out that the relative contributions of
the dark matter halo and stellar contents to the dynamical mass of a galaxy
vary significantly with its luminosity and morphological type. In the
rest of this work, we shall use the simple $M_d \propto R^3$ relation observed
by \citet{2008Natur.455.1082D,2009MNRAS.394..340G}, showing a
shared $\rho_d\sim10^7 M_\odot kpc^{-3}$ across galaxies.
This implies, that the mean surface density is controlled essentially
by the gas velocity dispersion $\sigma_g$ which is driven by energy
input from supernova explosions in the disk.
\section{Mechanical feedback from supernovae}
The distribution of gas in our galaxy has been variously described in
the past as being similar to that of Swiss Cheese \citep{1974ApJ...189L.105C},
as a Cosmic Bubble Bath \citep{1975A&A....38..363B} or as the Violent
Interstellar Medium \citep{1979ARA&A..17..213M}.
The basic idea is that supernova remnants
are full of coronal gas which can persist for long enough, as it radiates
very inefficiently by bremsstrahlung, so that a modest supernova rate may
produce an interconnected morphology of hot gas. Shells and supershells
\citep{1979ApJ...229..533H} have been found in the neutral hydrogen
distribution of galaxies. \citet{1987ApJ...317..190M} suggested that
stellar winds and repeated supernovae from an OB association may create
cavities of coronal gas in the interstellar medium leading to the formation
of supershells. \citet{2011ApJ...728...24C} have demonstrated
that this mechanism can explain the dynamics of a HI supershell found in
M101 driven by mechanical energy input from supernovae in a giant young
stellar association.
The fraction
of volume filled by supernova-driven hot gas is referred to as the porosity (P)
of the interstellar matter (ISM). The porosity is driven by the supply of hot gas
from supernova remnants (SNRs) and is related to the 4-volume of a SNR in the
cooling phase \citep{1988ApJ...334..252C} ($\nu_{SN}$) and the supernova rate
per unit volume ($r_{SN}$) as
\begin{equation}
P=\nu_{SN} \times r_{SN}.
\end{equation}
The supernova rate is related to the recent star formation rate per
unit volume ($\rho_{\ast}$) as
\begin{equation}
r_{SN}=\rho_{\ast}/m_{SN},
\end{equation}
where $m_{SN}$ is the total mass of star formation required on an average
to produce each core collapse supernova.
\citet{1997ApJ...481..703S} has argued that if P is too high,
it would suppress the efficiency of star formation.
Since the supernova rate follows the recent star formation rate,
the situation called ``blowout'' would throttle the supply of hot gas and bring
down the value of P. On the other hand if P was too low, it would allow
the cold gas phase to dominate and form new stars more efficiently, some
of which would soon explode as supernovae and drive up the value of P.
Hence, there would be a self regulation process that would control P.
In this manner, \citet{1997ApJ...481..703S} shows that supernova explosions
supply the energy input necessary to maintain the velocity dispersion in
the gas phase at
\begin{equation}
\sigma_g=6.90 P^{-0.58}n_g^{0.1}\, E_{51}^{0.2}\zeta^{0.008}\rm km\, s^{-1}.
\end{equation}
This expression depends only weakly on the gas density $n_g$, energy input
from individual supernovae $E_{51}\times10^{51}$ ergs, and the metallicity
$\zeta$ all of which have typical values of order unity.
The dependence on these parameters is dropped from our subsequent
equations. Assuming self-regulated star formation ($P\sim0.5$) the predicted
gas velocity dispersion has been shown \citep{1997ApJ...481..703S} to be
$\sigma_g\sim11$ km s$^{-1}$, close to the value of observed by
\citet{1989ApJ...339..763S} for the three-dimensional peculiar velocity
dispersion of interstellar molecular clouds within 3 kpc of the Sun.
\citet{2009ApJ...704..137J} have studied the effect of the supernova rate
on the interstellar turbulent pressure and confirmed a very weak dependence
of $\sigma_g$ on the star formation rate. They find simulated HI emission
lines widths of $10-18$ km s$^{-1}$ for models with SN rates that range from
1 to 512 times the Galactic SN rate.
Hence, the characteristic values for $\rho_d$ and $\sigma_g$ when plugged
into Equation \ref{mucr}, set the characteristic scale for the gas
surface density to a ballpark figure of $\sim10^7M_\odot$ kpc$^{-2}$, which
is close the observed value \citep{2003ApJ...585..256R} for the ADBS galaxies.
\begin{figure}
\includegraphics[width=0.95\columnwidth]{adbs_theory.eps}
\caption{\textbf{Surface area and HI masses:} ADBS galaxies
\citep{2003ApJ...585..256R} are plotted as black squares. The solid
line is the theoretical curve (Equation \ref{M_HI}) from this work
and the area between the dashed lines represents the $2\sigma$ variance
region. Note the simple relation in which the HI mass scales almost
linearly with the surface area, implying a regulated HI surface density
across galaxies. This hitherto unexplained feature is reproduced by our
simple model.}
\end{figure}
\section{Regulated surface gas density}
Integrating the star formation rate per unit volume over an entire galaxy
\citet{1997ApJ...481..703S} provides the total SFR as
\begin{equation}
\dot M_{\ast} \backsimeq 2.6 \times v_{rot,200}^{5/2} \, Q^{-3/2} \left( \frac{P}{0.5} \right)^{-0.87}\,\rm M_\odot\, yr^{-1},
\end{equation}
where $v_{rot,200}$ is the maximum rotational velocity of the galaxy,
normalized to 200 km s$^{-1}$. The exact expression \citep{1997ApJ...481..703S}
depends on well understood quantities like the total mass of stars formed to
yield one core collapse supernova, the mechanical energy output of an average
supernova, and depends only weakly on the metallicity.
As the disk becomes gravitationally unstable if $Q$ decreases, the SFR
increases and the resulting supernovae increase the porosity $P$. This
reduces the supply of cold gas and hence suppresses the star formation
rate. Hence \citet{1997ApJ...481..703S} points out that this expression
provides an explicit demonstration of disk self-regulation and
self-regulated star formation ensures $P\sim0.5$ and $Q\sim1$.
We use these values in the rest of this work.
Using the definition of the dynamical
mass ($M_d=(\Delta V)^2R_{90\%}/G)$) from \citet{1988gera.book...95K},
we can express $v_{rot,200}$ as
\begin{equation}
v_{rot,200} \backsimeq 1.04 \times \sqrt{\frac{M_{d,11}}{R_{10}}},
\end{equation}
where $M_{d,11}$ is the dynamical mass normalized to $10^{11} M_\odot$ and
$R_{10}$ is the 90\% optical containment radius ($R_{90\%}$) normalized to 10 kpc.
The dynamical mass has been found to be correlated with the cube of the
optical radius \citep{2009MNRAS.394..340G}. The implied roughly constant
global dynamical density is of the order of $\rho_d\sim10^7 M_\odot kpc^{-3}$.
Exploiting this observed relation we have
\begin{equation}
M_{d,11}\backsimeq0.42 \rho_7 R_{10}^3,
\end{equation}
where $\rho_7=\rho_d / 10^7 M_\odot kpc^{-3}$.
Substituting for $v_{rot,200}$ and then for $M_{d,11}$, we get
\begin{equation}
\dot M_{\ast} \backsimeq0.95 \times \rho_7^\frac{5}{4} R_{10}^\frac{5}{2} \,\rm M_\odot\, yr^{-1}.
\end{equation}
One may interpret the \citet{1996A&A...312..397G} relation between total
luminosity (a proxy for the stellar mass) and dynamical mass (proportional
to $R^3$ because of the shared $\rho_d$) as $M_{\ast}\propto R^3$.
This implies that the doubling time for the stellar mass, scales as
\begin{equation}
\tau=\frac{M_{\ast}}{\dot M_{\ast}} \propto R^{\frac{1}{2}}.
\end{equation}
Objects with larger doubling times have older stellar populations on
an average. This may explain why bigger galaxies are systematically redder.
This
star formation rate
allows us to estimate the mean SFR per unit area as
\begin{equation}
\Sigma_{SFR} \backsimeq3.0 \times 10^{-3} \times \rho_7^{1.25} R_{10}^{0.5} {\rm~M_\odot~yr^{-1}~kpc^{-2}}.
\end{equation}
However $\Sigma_{SFR}$ is related to the surface density of gas $\Sigma_{gas}$
by the Kennicutt-Schmidt Law \cite{1998ARA&A..36..189K} as
\begin{align}
\Sigma_{SFR} = & {{(2.5 \pm 0.7)} \times 10^{-4}}~
\\ \nonumber &
\times{\Big({\Sigma_{gas} \over
{1~M_\odot~{\rm pc}^{-2}}}\Big)^{1.4\pm0.15}}{\rm~M_\odot~yr^{-1}~kpc^{-2}}.
\end{align}
If the molecular gas fraction is $\ll1$, which is true for most disk galaxies,
the mean surface density of HI $\Sigma_{HI}$
is comparable to the total gas surface density $\Sigma_{gas}$.
This assumption will make a proportional error comparable to the
molecular gas fraction.
Hence assuming $\Sigma_{HI}\sim\Sigma_{gas}$,
eliminating the $\Sigma_{SFR}$ and integrating over the surface area,
we obtain the relation between the gas mass and the surface area as
\begin{align}
\nonumber
log\Big(\frac{M_{HI}}{1 M_\odot}\Big) \backsimeq & (6.91 \pm 0.08)
+ 1.18 \times log\Big(\frac{A}{1 kpc^2}\Big)
\\ &
+ 0.89 \times log\Big(\frac{\rho_d}{10^7 M_\odot kpc^{-3}}\Big),
\label{M_HI}
\end{align}
where $A\equiv\pi R^2$ is the cross sectional surface area presented
by the galaxy.
Given that dynamical mass density ($\rho_d$) is not seen to vary much across
galaxies, the almost linear relation between the total HI mass and the surface
area implies very similar HI surface densities across a range of
galaxies. The normalization matches the
observed \citep{2003ApJ...585..256R} relation for $\rho_7=0.26$ which lies within
the range of its observed \citep{2009MNRAS.394..340G} values. The HI masses,
of ADBS galaxies \citep{2003ApJ...585..256R}, spanning 3 orders of magnitude,
vary only by $\sim0.14$ dex ($1\sigma$) from the theoretical curve. However,
this scatter is larger than the scatter propagated from the Kennicutt-Schmidt
Law \citep{1998ARA&A..36..189K}. The excess scatter may be attributed to the
scatter in core properties of halos such as $\rho_7$. It has been pointed out
by \citet{2010arXiv1011.6374L} that numerical simulations are
required to determine the scatter in dark matter core densities as a function
of mass and redshift. This would be important for dark matter dominated halos.
\citet{1996A&A...312..397G} show a correlation between total luminosity and
the dynamical mass. If most of the dynamical mass is provided by the stellar
content, it would be important to study the scatter in this relationship.
\section{Discussions}
Our result shows that the present understanding of mechanical
feedback from supernovae, leading to self regulated star formation
\citep{1997ApJ...481..703S}, can account for the regulation of HI
surface density across galaxies to a characteristic value of around
$\sim10^7M_\odot$ kpc$^{-2}$ \citep{2003ApJ...585..256R}.
The predicted HI surface density depends on the mean density of the
dynamical mass, which is likely to be provided by a combination of
dark matter and stellar content.
As to why galaxies are observed \citep{2009MNRAS.394..340G} to share
similar dynamical mass densities is still an open case requiring further
investigation. Cold dark matter particles interacting through a Yukawa
potential \citep{2010arXiv1011.6374L} could provide a natural explanation
for a characteristic density in dark matter dominated halos.
The correlation \citep{1996A&A...312..397G} between luminosity
and dynamical mass may be important in halos dominated by the stellar content.
The model of self regulated star formation \citep{1997ApJ...481..703S}
has been shown in this work to provide a scaling between doubling time
and radius. This could explain why larger galaxies are systematically redder.
This relation should be used in conjunction with population synthesis models
such as Starburst99 \citep{1999ApJS..123....3L}
to predict colors as a function of galaxy size. This could provide an
interesting test of this model in future.
Even if mergers drive a galaxy away from the fundamental line, self
regulation of the porosity of the ISM and Toomre parameter of the star forming
disk can bring the HI surface density back to the value which is
predicted in our simple model and observed in a wide range of galaxies.
In a framework for hierarchical galaxy
mergers this can happen multiple times in a galaxy's history, until it
eventually runs out of neutral hydrogen. A detailed understanding of
the proposed mechanism would require self consistent galaxy simulations
taking into account the supply of hot gas into the ISM from
individual supernovae.
Our model relates the total neutral hydrogen mass of the galaxy with
its projected surface area. However in practice, the quantities observed
with radio telescopes are the redshift, integrated HI line fluxes and
the solid angles on the sky.
Fluxes and solid angles behave differently in different cosmological
scenarios, as they scale with the luminosity distance $d_L$ and angular
diameter distance $d_A$ respectively. This could facilitate their use in
testing the \citet{1933PMag...15..761E} relation, $d_L = (1 + z)^2 d_A$,
when future telescopes such as the Square Kilometer
Array start to detect red-shifted HI from very distant galaxies.
\acknowledgments
The author wishes to thank thank Alak Ray for guidance, a careful
reading of the manuscript and numerous suggestions. Abraham Loeb,
Swastik Bhattacharya and Satej Khedekar are thanked for discussions
about the work. An anonymous referee is thanked for suggestions.
\bibliographystyle{apj}
|
\section{Introduction}
\label{Intro}
Solar filament eruptions are energetic events occurring due to the
explosive release of magnetic energy. Understanding the driver and
trigger mechanisms of these eruptions is one of the most
challenging, ongoing research problems in solar physics. These
events manifest as prominence eruptions if seen at the limb, and as
filament eruptions if seen against the disk. They can be broadly
divided into two classes - (i) ejective eruptions which give CMEs
and long duration two-ribbon flares, and (ii) confined eruptions
which give short duration flares. However, there are also
filament/prominence eruptions that are not associated with any
flares. What defines the eruption to be ejective or confined event
remains an unanswered question (For detailed reviews on theories of
eruptions, refer to \inlinecite{priest2002}, \inlinecite{lin2003}).
In filament eruptions, one commonly observed feature in
chromospheric H$\alpha$ images is two bright flare ribbons
separating away as the flare progresses. This process is explained
as follows: Reconnection of magnetic fields is believed to be the
underlying mechanism for energy release as proposed by
\inlinecite{carmi1964}, \inlinecite{sturrock1966},
\inlinecite{hirayama1974}, \inlinecite{kopp1976}. As a coronal
magnetic flux rope loses equilibrium and travels upwards, an extreme
reconnection current sheet (RCS) is formed underneath. Reconnection
in this RCS releases most of the magnetic energy stored in the
magnetic field configuration \cite{forbes1984,lin2000}. Charged
particles can be effectively accelerated by the electric field in
the RCS \cite{martens1990,litv1995}. Some particles, energized
during a solar flare, gyrate around the field lines and propagate
toward the underlying foot points, precipitating at different layers
of the solar atmosphere to produce the two-ribbon flares. Separation
of these chromospheric flare ribbons is believed to provide a
signature of the reconnection process occurring progressively higher
up in the corona. This is the standard flare model scenario. All
eruption models lead to this standard model in the onset phase of
eruption.
Several ideas have been advanced for explaining the eruption onset
mechanism \cite{klim2001,forbes2006,moore2006}. In tether-cutting
model proposed by \inlinecite{moore1980}, and further elaborated by
\inlinecite{moore2001}, magnetic tension restraining the sheared
core field of a bipolar magnetic arcade is released by internal
reconnection above the PIL. Evidence for the tether-cutting model
can be found in several recent observational studies
\cite{liu2007,wang2006,yurc2006}. External tether cutting or
``breakout'' reconnection is similar to tether cutting in that it is
a tension release mechanism via reconnection. But here it occurs
between the arcade envelope of the erupting field and an
over-arching restraining, reversed field of quadrupolar magnetic
configuration \cite{antioc1998,antioc1999}. Flux cancelation
\cite{martin1985,martin1989}, emergence of twisted flux ropes from
below the surface \cite{leka1996}, and ideal MHD instability
\cite{kliem2006,fan2007} are some other mechanisms proposed to
explain the eruption process.
Although there exist several proposed mechanisms, it is difficult to
disentangle as to which particular mechanism is responsible for the
fast eruption in complex ARs. This difficulty arises due to the wide
variety of dynamic processes involved in such ARs. Therefore, we
have selected a filament eruption event that occurred in a
relatively simple solar AR NOAA 11093 on 7 August 2010 that led to
M1.0 class two-ribbon flare and a fast CME. Our aim in this study is
to understand the driver and trigger mechanisms of the eruption and
associated processes. We investigate various possible conditions of
eruption process in relation to flux emergence/cancellation at
selected locations. From a morphological analysis, we attempt to
find triggering mechanism of the flare. We derive the flare
energetics in order to quantify flare characteristics and then
compare it with previous such studies.
The essence of high cadence and high resolution multi-wavelength
observations has already been revealed by earlier studies
\cite{liu2007,liu2005,wang1999}. In the present study, we utilize
the unique opportunity of coordinated observations in
multi-wavelength channels corresponding to different atmospheric
heights provided by Atmospheric Imaging Assembly (AIA;
\inlinecite{lemen2011}) and Helioseismic and Magnetic Imager (HMI;
\inlinecite{schou2011}) on board Solar Dynamics Observatory(SDO).
The rest of the paper is organized as follows: The observational
data and reduction procedures are presented in Section~\ref{data}.
Results and discussions are described in Section~\ref{analys} while
Section~\ref{summ} gives the summary and conclusions.
\section{Observational Data and Reduction}
\label{data}
The eruption event in AR NOAA 11093 ($N12^\circ E31^\circ$) occurred
on 7 August 2010. It produced a GOES M1.0 class flare starting at
17:55 UT and peaking at 18:20 UT. This event was covered by SDO's
AIA and HMI, Reuven Ramaty High-Energy Solar Spectroscopic Imager
(RHESSI; \inlinecite{lin2002}), as well as, by the ground based GONG
H$\alpha$ network telescope at BBSO. The associated CME was detected
by the COR1 coronagraph \cite{thompson2003} on board both the Ahead
and Behind satellites of the Solar TErrestrial RElations Observatory
(STEREO) which were separated by about $150^\circ$.
AIA takes multi-wavelength images at pixel size of
$0^{\prime\prime}.6$ pixel$^{-1}$ and 12 s cadence. To study the
flaring plasma, we focused on the images obtained in $\mbox{94\AA}$~
(Fe XVIII; $\log T = 6.8$), $\mbox{171\AA}$~ (Fe IX; $\log T =
5.8$) corresponding to the upper transition region, and
$\mbox{304\AA}$~ (He II; $\log T = 4.8$) corresponding to the
chromosphere and lower transition region. Images were added to
enhance the signal to noise ratio, giving a cadence of 1 minute. We
have used preprocessed images (level 1.0) provided after
calibration, involving bad pixel correction, aligning, and scaling.
HMI makes measurements of line-of-sight magnetic field of the full
solar disk at $\mbox{6173\AA}$~ with a pixel size of
$0^{\prime\prime}.5$ and $45$ s cadence with a precision of 10G. We
rotated images (level 1.0) for solar north pointing up. Scaling of
data was done using the header information. We added every four
images for increasing the signal to noise ratio, giving a cadence of
3 minutes.
COR1 is an internally-occulted coronagraph and is one of the STEREO
SECCHI suite of remote sensing telescopes. It takes observations of
CME from 1.3 to 4 solar radii in three different polarizing angles
every five minutes. We have used \texttt{secchi\_prep.pro} and
\texttt{cor1\_quickpol.pro} routines in STEREO software package to
process the images and finally to get total polarization brightness
images. For observing ejected material within 1.2 $R_\odot$, we have
also examined EUVI observations on board STEREO-A.
We obtained H$\alpha$ 6563 \AA~ filtergrams at pixel size of
$1^{\prime\prime}$ and 1 minute cadence from the GONG telescope
operating at BBSO.
All full disk images obtained from different instruments were
aligned by differentially rotating to a reference image at 18:00 UT.
The offset was corrected after remapping and by overlaying magnetic
contours on images taken by different instruments. We have used
standard SolarSoftWare (SSW) library routines for our study.
\begin{figure}
\centering
\includegraphics[width=.9\textwidth, clip=,bb=43 27 416 228]{fig1_2}
\caption{HMI line of sight magnetogram (left panel), and GONG
H$\alpha$ filtergram (right panel) of NOAA 11093 overlaid with red
(blue) contours corresponding to positive (negative) magnetic
fluxes. North is directed upward in these and subsequent
maps.}\label{fig1}
\end{figure}
\section{Results and Discussions}
\label{analys}
\subsection{Filament Evolution Leading to the Two-Ribbon M1.0 Flare}
\subsubsection{Morphology}
\begin{figure}
\centering
\includegraphics[width=.75\textwidth,clip=,bb=6 14 425 490]{litcur}
\caption{Light curves of the M1.0 class flare of 7 August 2010 in
AR NOAA 11093: (top) GOES soft X-rays, (middle) AIA wavelengths
and GONG H$\alpha$, and (bottom) RHESSI hard X-rays. Gray shaded
region represents the impulsive phase of the flare as inferred from
GOES soft X-rays. Gaps in RHESSI light curves are due to the
spacecraft's passage through night-times and the South Atlantic
Anomaly.}\label{litcur}
\end{figure}
Figure~\ref{fig1}(left panel) shows an HMI magnetogram of AR NOAA
11093 taken on 7 August 2010 at 17:10 UT. The AR possessed a simple
magnetic configuration reported as $\beta-$ class by NOAA/USAF AR
Summary issued on 7 August 2010. The GONG H$\alpha$ filtergram of
the AR with contours of the line-of-sight magnetic field is shown in
the right panel. It consisted of a single main sunspot and an
inverse S-shaped filament with its one end connected to the sunspot,
as observed more than an hour before the eruption event ensued.
These images show the polarity of the dominant main sunspot and the
diffused fluxes of opposite polarities surrounding the filament.
The filament was oriented in nearly NE-SW direction along the PIL.
Concerning the observed filament, the AR was bipolar with two main
leading polarities located on either side of PIL. We have
concentrated our study mainly on the time interval 17:00--20:00 UT
for identifying the changes occurring in morphological structure as
well as the connectivity of field-lines in the AR leading to
filament eruption and the flare.
Light curves of the flare in different wavelengths are shown in
Figure~\ref{litcur}. The gray shaded region from the start to peak
time of GOES flux represents the impulsive phase of the flare. The
decay phase of GOES soft X-rays flux lasted over three hours from
the peak time, implying that this relatively small M1.0 class flare
was a long duration event (LDE). The AIA 304 \AA~ and H$\alpha$
profiles essentially followed the GOES profiles, peaking at around
18:20 UT. However, AIA 94 and 171 \AA~ profiles did not agree well
in the impulsive phase, and peaked much later at 18:45 UT. By
examining the corresponding images, we found that the delayed
peaking in these wavelengths corresponded with the post-flare loops
that appeared in the decay phase.
The RHESSI hard X-rays (HXR) profiles show two gaps owing to the
spacecraft's passage through the night-time and South Atlantic
Anomaly. However, the available data in the shaded region, i.e., the
impulsive phase of the flare, clearly shows a short duration
enhancement in the level of HXR emissions in 6-12 and 12-25 keV
channels. Notably, however, no enhancement was seen in the harder
emission in 25-50 keV (blue) channel.
\begin{figure}
\centering
\includegraphics[width=1.0\textwidth,clip=,bb=43 18 565 565]{preris}
\caption{(Top row) Images in AIA 304 \AA~ exhibit a rising motion of
flux rope lying along PIL (marked by arrows) with the brightening
region encircled. (Middle row) H$\alpha$ images showing
corresponding morphological changes in the chromosphere. White thick
curves drawn on these images represent the observed rising flux rope
as seen in 304 \AA~ images at respective times. In the frame (d),
overlaid red(blue) contours represent positive (negative) polarity
fluxes. (Bottom row) Flux rope rise observed in STEREO-A/EUVI 304
\AA~ at the east limb. }\label{preris}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=.98\textwidth,clip=,bb=24 14 420 220]{171_over_mag_schm}
\includegraphics[width=.98\textwidth,clip=,bb=0 0 456 218]{fil_eru_1and2}
\caption{Top: Field line topology before(a) and after(b) the onset
of eruption as seen in 94 and 171 \AA~ channels, respectively, with
superposed magnetic field contours. Bottom: Schematics of triggering
of flux rope/filament rise motion by tether weakening due to
interaction of side lobe field line with twisted overlying field
line from the sunspot. This rise motion enhanced the twist in the
inner core of the filament further, forming a current sheet beneath
it. Twisted inner core can be clearly seen in frame
(b).}\label{schem_fil_eru}
\end{figure}
>From movies made using the images in H$\alpha$ and AIA channels, we
observed that a flux rope having an average diameter of 6.9~Mm
($\sim9^{\prime\prime}.5$) started to rise from the inner core of
the filament at around 17:40 UT. Figure~\ref{preris}(a-c) shows, as
pointed by arrows, the rise of flux rope connecting the sunspot and
the other end of the filament elbow. The projection of this flux
rope on chromosphere is seen as a dark shadow moving toward east and
disappearing after 18:04 UT. At that time it had risen high, nearly
vertical to the disk plane. About 15 min after the start of its
rise, flare brightenings appeared; first the northern ribbon at the
hump part and then the southern one. The rise of filament was seen
as plasma jet ejection in EUVI images of STEREO A in 304 \AA~
(Figure~\ref{preris}(g-i)). STEREO A and B were separated by
150$^\circ$, and STEREO B was situated at 71$^\circ$ west from the
Earth, AR 11093 was 36$^\circ$ west of STEREO-B central meridian and
24$^\circ$ away from east limb of STEREO-A. Therefore, the flux rope
was not visible till 18:03 UT in STEREO-A. Furthermore, from
STEREO-B the event was not as clearly visible as in AIA 304 \AA~ due
to the lower spatial resolution (~1.6$^{\prime\prime}$/pixel).
We provide the schematics of eruption event constructed by a careful
examination of observed plasma tracers using animations of 171 and
94 \AA~ images in Figure~\ref{schem_fil_eru}. The EUV brightening in
the smaller encircled location of Figure~\ref{preris}(a) resulted
due to the interaction of overlying sheared field lines from the
sunspot with the side lobe field lines, as shown in
Figure~\ref{schem_fil_eru}(frame a and b). This interaction
substantially weakened the overlying field lines, allowing the flux
rope to rise upward. This rising motion formed current sheet in
cusped region of field lines connecting polarities on either side of
the filament leading to the onset of internal reconnection to
commence the flare. In 94 \AA~ image (frame b), the twisted bright
flux system can be seen in the inner core below the flux rope. It is
important to notice that the bright flux system appeared only in 94
\AA~ because soft X-ray emission began due to reconnection by the
above process. This scenario agrees well with the model of
\inlinecite{moore2001} that describes ``tether cutting'' as the
trigger mechanism. It is worth mentioning that the tether weakening
first occurred at the location of brightening and then induced
further in the core region beneath the filament. Tether cutting as a
trigger for the onset of eruption within the erupting system is
addressed by \inlinecite{yurc2006} in a quadrupolar configuration.
We will discuss the magnetic flux changes at the locations of these
brightenings in Section~\ref{FlxCha}
\begin{figure}
\centering
\includegraphics[width=1.0\textwidth,clip=,bb=35 10 560 560]{mosaic}
\caption{A mosaic of images showing the filament evolution during
various phases of the flare -- start(row 1), impulsive(rows 2 and
3), and decay (row 4) as observed in GONG H$\alpha$ and AIA
wavelengths corresponding to successively higher atmospheric
layers.}\label{mosaic}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1.0\textwidth,clip=,bb=35 14 488 198]{rheimg}
\caption{Contours of RHESSI HXR overlaid on AIA 171 \AA~ image at
18:09 UT in three energy bands: (left) 3-6 keV, (middle) 6-12 keV,
and (right) 12-25 keV, reconstructed by clean algorithm with one
minute integration time. Contour levels correspond to 50,70,80 and
90\% of respective peak flux.}\label{rheimg}
\end{figure}
Figure~\ref{mosaic} is a mosaic of multi-wavelength images
illustrating filament eruption and the flare in H$\alpha$ and
various AIA channels corresponding to successively higher
atmospheric layers (from left to right columns). Rows correspond to
the time of start (top), impulsive (second and third), and decay
(bottom) phases of the flare. First column shows flare ribbons in
chromospheric H$\alpha$. As time progressed, flare ribbons
brightened and separated away up to the peak time 18:25 UT, and
decayed there after. Images in the second column correspond to AIA
304 \AA, showing flare ribbons in the upper chromosphere along with
the overlying field lines filled with $6\times10^4$K plasma.
Post-flare loops connecting the flare ribbons are clearly seen in
the decay phase. These post flare loops gave rise to the second peak
of 304 \AA~ light curve in Figure~\ref{litcur}. In third column, AIA
171 \AA~ images show plasma loops more clearly as these are
sensitive to $T \sim 6\times10^5$K. Fourth column corresponds to AIA
94 \AA~ images showing plasma loops at even higher temperatures
($\approx6\times10^6$K). We can observe here an increased twist in
core flux system due to tether cutting reconnection with increased
flux rope height (18:06:02 UT frame). It is important to note that
this twisted flux system is not visible in 304 and 171 \AA~
corresponding to lower heights. Once the main reconnection phase
commenced, these twisted flux system below the flux rope, as seen at
18:06:02 UT in the inner core of filament, relaxed to arcades of
lower twist seen at 19:06:26 UT in the decay phase. This twisted, or
sigmoid to arcade evolution is an important mechanism of energy
release process as studied in many events \cite{liu2007,liu2005}. (A
movie, ``mosaic.mpeg'', is available on request.)
>From the RHESSI light curves, it is evident that HXR sources were
produced during the impulsive phase of the flare, at least in the
lower energies of 6-12 and 12-25 keV. Unfortunately no data was
available during 18:17-18:36 UT due to the spacecraft's night-time
and afterward when attenuators were out of the field of view. The
data with sufficient counts for imaging was available only during
18:00-18:10 UT in the impulsive phase of the flare.
We constructed the HXR images with ``clean'' algorithm from the
modulated data and looked for HXR sources in the flaring region. In
Figure~\ref{rheimg}, we have plotted the contours of these
reconstructed images in the energy bands of 3-6, 6-12 and 12-25 keV
at 18:09 UT and overlaid on AIA 171 \AA~ images. Integration time of
the images was taken as 1 min; adequate to detect the changes. Due
to the rather low count rates in this event, it was not possible to
reconstruct images in higher energy bands.
The HXR contours were found to be localized between the flare ribbon
kernels suggesting that reconnection took place at these locations,
where particles accelerated along the field lines and propagated
toward foot points anchored in the photosphere. In the 12-25 keV
image, a break in contours was observed on ribbon hump part. This
suggests that non-thermal HXR sources were located at foot points of
the flare ribbons. In the absence of additional HXR data, however,
it was not possible to follow up their further evolution.
\subsubsection{The Flare Energetics}
We can study the flare energetics by evaluating two main physical
parameters, viz., the rates of reconnection and energy release.
Reconnection rate is the electric-field strength in RCS and is
defined as the reconnected magnetic flux per unit time expressed as
\begin{equation}
\dot{\Phi} = B_{c} v_{in}.
\end{equation}
The magnetic energy release rate during a solar flare is the product
of Poynting flux and the area of RCS that is generated during
magnetic reconnection. On the basis of reconnection model, it has
been shown by \inlinecite{isobe2002} that energy release rate can be
written as,
\begin{equation}
\frac{dE}{dt}=S A_{r}f_{r}=\frac{1}{2\pi}B_{c}^{2}v_{in}f_{r},
\end{equation}
where S is the Poynting flux into the reconnection region, $B_{c}$,
$v_{in}$, $A_{r}$ and $f_{r}$ are coronal magnetic field strength,
inflow velocity, area of the reconnection region, and the filling
factor of reconnection inflow, respectively.
The inflow velocity can be determined from observations. Using the
magnetic flux conservation theorem, one can write
\begin{equation}
B_{c} v_{in}=B_{\rm chro} v_{\rm ribb}=B_{\rm phot} v_{\rm ribb},
\end{equation}
where $v_{\rm ribb}$ is the velocity of the H$\alpha$ ribbon
separation, and $B_{\rm phot}$, $B_{\rm chro}$ are the
photospheric and chromospheric magnetic field strengths
respectively.
\begin{figure}
\centering
\includegraphics[width=.6\textwidth,clip=,bb=33 33 369 388]{nutlin}
\caption{H$\alpha$ image (in negative) of the flare at
18:12:54\,UT, overlaid with magnetic field contours. Solid black
line is the simplified PIL. Arrowed lines L1, L2, and L3 are drawn
perpendicular to PIL at points marked by ``+'' to follow the
separation of flare ribbons.}\label{nutlin}
\end{figure}
Figure~\ref{nutlin} depicts the scenario of flare ribbon separation
motion along the arbitrarily selected lines L1, L2, and L3. We have
measured distances from points marked as ``+'' on both sides of PIL
shown in thick solid line, i.e toward the north and south
directions. Separation velocities are then determined after fitting
Boltzmann sigmoids through the measurements. Further, using these we
calculated the reconnection rates, and Poynting fluxes for a
quantitative study of flare energetics.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=.7\textwidth,clip=,bb=8 8 314 497]{fig09_mod}
\caption{(a) GOES X-ray flux and its derivative, (b) flare ribbons
separation distance and velocity profiles, (c) reconnection rate
and Poynting flux, and (d) filament/flux rope height as function of
time. Vertical dashed lines mark the start and peak times of flare
plotted for reference.}\label{flxrop_ht}
\end{center}
\end{figure}
For obtaining ribbons' separation velocity, we followed the
technique and assumptions as discussed in \inlinecite{maurya2010}
[and references therein] with $a = 0.2$ and $\epsilon=0.4$. We thus
obtained the reconnection rates in the range 0.5-3 Vcm$^{-1}$ and
Poynting fluxes in the range 0.01-3.8 G erg cm$^{-2}$s$^{-1}$ as
measured along the lines L1, L2, and L3. We have plotted one of the
temporal profiles in Figure~\ref{flxrop_ht} for comparing with the
filament/CME rise. This flare lasted for about 3000 seconds with
current sheet spread over an area $\approx 10^{19} cm^{2}$, with
average energy release rate of $10^{7}$ erg cm$^{-2}$s$^{-1}$. Using
these typical values, we estimate the total energy released during
the flare event to be a modest $10^{29} $ergs, in conformity with
the magnitude of the flare. Our results are consistent with that of
quiescent filament eruption reported by \inlinecite{wang2003}.
We notice here that kernel velocities were not uniform along the
selected lines. Further, electric field that is expected to release
energy through dissipation of electric currents in RCS, temporally
did not correlate well with impulsive hard X-ray emission in some
directions. To investigate the electron acceleration in the
impulsive phase of the flare, we require hard X-ray and microwave
observations. But, due to the data gap in hard X-rays during the
impulsive phase, we have used the time derivative of GOES soft X-ray
light curve as an alternative to the hard X-ray with the assumption
that Neupert effect is valid in this event (as in reference XXXX).
Our estimates for this M1.0 are lower as compared to the velocities
of 15-50 kms$^{-1}$ and reconnection rates of 2.7-11.8 Vcm$^{-1}$
for the M3.9 flare event reported by \inlinecite{miklenic2007}. Also, for
a rather short duration, small C9.0 flare, \inlinecite{qiu2002}
reported a significantly larger kernel velocity of 20-100 kms$^{-1}$
and peak electric field of 90 Vcm$^{-1}$. They found ribbon motions
both parallel and perpendicular to PIL and concluded that either the
2D-magnetic reconnection theory related to the H$\alpha$ kernel
motion was applicable only to a part of the flare region due to its
particular magnetic geometry, or the electron acceleration was
dominated by some other mechanisms depending on the electron energy.
We therefore suggest that the estimated electric fields and
reconnection rates depend not only on the magnitude of the flare,
but also on the flare kernel velocity and the magnetic field
geometry.
\subsection{The CME and its dynamics}
\label{erucme} The flux rope started rising at 17:40 UT as seen in
AIA 304\AA~ (cf., Figure~\ref{preris}). Further, with its rise, the
surrounding overlying loops also started rising as if they formed a
balloon like cavity expanding temporally. The jet like ejection of
plasma came out from the reconnection region with ejection velocity
presumably proportional to the reconnection rate as its height
increased. Eventually, it is observed as the CME event in COR1
coronagraph at about 18:20 UT.
We project the lateral displacements of filament in vertical
direction in AIA 304 \AA~ in pre-rise phase(see for details
\inlinecite{wang2003}) and corrected heights of flux rope as it is
seen exactly on the limb in EUVI/STEREO in impulsive phase. Then we
reconstructed the heights of CME front end from triangulation method
in decay phase to obtain the heights of filament/flux with its rise.
In order to reduce the measurement errors, we fit them with Boltmann
sigmoid function as used for flare ribbon motion \cite{maurya2010}.
This function suits well as a model to the data because the lower,
steep and upper parts of this function resemble the rise, impulsive
and decay phases of CME/flux rope dynamics and can be fitted by four
parameters in the least square sense.
Heights of filament and CME front is plotted in
Figure~\ref{flxrop_ht} for comparing with temporal evolution of
flare ribbon separation, filament eruption and flare energy release
in terms of reconnection rate \cite{wang2003}. The fast-rising stage
coincided with the flare impulsive phase, and the mass acceleration
increased rapidly along with the increase of magnetic reconnection
rate. As evident from the figure, flare ribbons appeared at 17:58 UT
and attained maximum velocity at the peak phase of flare. This
motion agreed with hard X-ray profile as electrons injected onto
photosphere in the impulsive phase with increasing reconnection
rate. From the figure, we can see that filament started to rise in
the first 15 minutes with velocity in the range 8-10 kms$^{-1}$
reaching 100 kms$^{-1}$ at an average acceleration of 60 ms$^{-2}$.
>From STEREO observations including COR1, we found that the CME
traveled with an average speed of 590 kms$^{-1}$ and reached peak
acceleration of 220 ms$^{-2}$ at the peak phase of the flare. It
then decelerated in the decay phase gradually. This is in
correspondence with magnetic reconnection rate and flux rope
acceleration obtained from a study of 13 well observed two-ribbon
flares \cite{jing2005}. The acceleration in the range 50-400
ms$^{-2}$ and peak reconnection rates in the range 0.2-5.0
Vcm$^{-1}$ as obtained by them are consistent with our results.
\subsection{Changes in the Photospheric Magnetic Field }
\label{FlxCha} Theoretical models suggest that evolution of magnetic
fields at or below the photosphere, in the form of flux emergence
and cancellation, could result in a loss of equilibrium of the
magnetic structures \cite{martin1985,martin1989,leka1996}. Changes
in the photospheric longitudinal magnetic field around the time of
eruption have been examined by many workers in the past
\cite{wang1999,mathew2000,green2003,ambastha2007,jiang2007,sterl2007}.
\inlinecite{zhang2008b} studied the relationship between flux
emergence and CME initiation inferring that $60\%$ of CME source
regions have increase and $40\%$ have decrease of magnetic flux.
Here, we look for regions, if any, of flux emergence/cancellation in
photospheric LOS magnetic field using high resolution SDO/HMI
magnetograms. We carried out registration of the images by
differentially rotating to a reference image at 18:00 UT. Effects of
telescope jitter and other pointing errors were corrected by using a
cross correlation method reducing the uncertainty within 1-2 arc
sec. Every four images were added to yield a cadence of 3min. As the
HMI precision is 10G, we neglected magnetic fields below this value.
\begin{figure}[htbp]
\centering
\includegraphics[width=.50\textwidth,clip=,bb=82 393 362 670]{magbox_1}
\includegraphics[width=.48\textwidth,clip=,bb=28 8 356 346]{magbox_2}
\caption{(Left) GONG H$\alpha$ image of NOAA11093 overlaid with HMI
LOS magnetogram contours. Boxes(1-7) mark the selected
regions-of-interest (ROI) for calculation of unsigned flux.
(Right):~Temporal profiles of unsigned flux in the
ROIs.}\label{magbx_prof}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=.95\textwidth,clip=,bb=12 8 358 274]{mag_box_mosaic}
\includegraphics[width=.95\textwidth,clip=,bb=17 48 358 234]{dif_img_171_304}
\caption{Enlarged box regions 3,5 and 7 of Figure~\ref{magbx_prof},
showing magnetic flux evolution with time (top panel). Difference
images in pre-eruption/flare phase of 17:45 - 17:40 UT in AIA 171
\AA~, 304 \AA~ wavelengths showing associated brightenings (bottom
panel).}\label{brig}
\end{center}
\end{figure}
A typical GONG H$\alpha$ image of NOAA 11093 overlaid with contours
of LOS magnetic fields is shown in Figure~\ref{magbx_prof} (left
frame). From a movie of the registered images, we identified sites
of flux emergence/cancellation, marked in the figure by boxes,
located around the filament and the sunspot in the AR. Time profiles
of unsigned magnetic fluxes corresponding to the selected boxes are
plotted during the period 16:00-20:00 UT. Sufficient care was taken
in selecting the box size of
$30^{\prime\prime}\times30^{\prime\prime}$. A very small box size
would not adequately cover the region of interest, while averaging
over too large a region would dilute the magnitude of changes.
We interpret temporal evolution of fluxes in each box and its
contribution to the stability of sunspot-filament magnetic system.
Boxes 1 and 2 are located around the leading sunspot. Box 1 selected
at the penumbral location of sunspot also covered regions of
opposite polarity fluxes around the filament. Total unsigned flux
increased there till the time of onset of the filament eruption, and
decreased thereafter. On the other hand, changes were oscillatory in
box 2 before the flare, where a small region of negative (red) flux
emerged along with pre-flare brightening.
Boxes 3, 4, 5 and 6 are located at either side of the PIL from where
field lines originated to tie the filament as ropes required for a
stable configuration. Changes in these boxes are important for
examining the stability of the filament. In box 3, a gradual
decrease of $0.2\times10^{20}$ Mx flux occurred at the flare onset
time from the time of start of the rope rise. On the other hand,
flux increased in boxes 5, 6, and 7 as a result of new flux
emergence. Of these, box 6 was located under the flux rope.
Evidently, there were sites of flux emergence/cancellation in and
around the filament influencing its stability. As the HMI
measurements do not suffer by the degrading effect of Earth's
atmosphere, one can be reasonably confident about the observed flux
changes.
To further corroborate these temporal changes, we enlarged some ROIs
as shown in Figure~\ref{brig} (top panel). Emerging flux
regions(EFR), marked by arrows, can be seen in these regions.
However, we did not find appreciable disappearance or cancelation in
box 3. We also looked for any brightening expected to be associated
with flux emergence/cancellation, by examining difference images of
pre-eruption/flare phase, i.e., 17:45 -17:40 UT, in AIA 171 \AA~ and
304 \AA~ (Figure~\ref{brig} (bottom panel)). This showed a
brightening (seen as darkening in the difference images) in box 3
where a gradual decrease of unsigned flux was observed during the
rise phase of the filament. Similar changes were also found in boxes
5 and 7 (seen as white dots in difference images).
\begin{figure}
\centering
\includegraphics[width=.5\textwidth,clip=,bb=9 1 280 352]{prot_wand_liu}
\caption{Time profiles of positive and negative flux within the
region covering the entire filament area (gray big box in
Figure~\ref{magbx_prof}(left pannel)). The start and peak times of
the flare are marked by the dashed vertical lines.}\label{flx_chan}
\end{figure}
In a recent work, \inlinecite{wang2010} found observational evidence
of back reaction on the solar surface fluxes with coronal magnetic
restructuring after reconnection in well observed X-class flares.
They suggested that such results may also be detectable for smaller
flares from the low threshold HMI data. To investigate it, we
computed positive and negative fluxes separately in the region shown
as the large box (drawn in gray color) covering the filament
(Figure~\ref{magbx_prof}(left panel)). Figure~\ref{flx_chan} shows
that positive/negative fluxes increased/decreased by similar amount
of $\approx 10^{20}$ Mx in the post flare phase. This gives an
indirect evidence of reconnection as a consequence of tether
cutting, mentioned in the previous sections.
Thus, we have found regions of flux emergence, cancelation and
associated EUV brightening in some locations of the AR (see
Figure~\ref{preris}(a)). The interaction of field lines at one of
these brightening regions gives an evidence of reconnection leading
to tether weakening. These processes could have destabilized the
filament to rise upward leading to tether cutting reconnection. The
observed photospheric flux changes found in the decay phase of the
event are in conformity with \inlinecite{wang2010}, i.e., evidence
of the back reaction on solar surface as a consequence of
reconnection.
\section{Summary and Conclusions}
\label{summ}
The current models assume that the stored energy lies in a low lying
magnetic flux system which is twisted or sheared. This system is the
so-called core flux or flux rope. Such twisted systems in the form
of inverse S (or forward S shape), called sigmoid, are usually seen
in association with eruptions as the present event.
\inlinecite{canf1999} had studied the nature of activity with
respect to morphology of eruptive or non-eruptive events with soft
X-ray observations and inferred that sigmoids are more likely to
erupt.
In this paper, we have presented multi-instrument multi-wavelength
observations and analysis of the eruption of an inverse S shaped
filament. This event exhibits a good example of standard solar flare
characterized by the filament eruption, two ribbon separation and
its association with a fast CME. We summarize below the main
findings of our analysis of this eruption event:
\begin{enumerate}
\item From morphological study of this event, we
inferred that the rising motion of the filament was triggered by
remote tether weakening at coronal brightening region,
which further induced tether cutting reconnection
underneath it to unleash the eruption process, leading to a fast CME
subsequent to the two-ribbon flare.
\item Flare ribbons or kernels separated out with modest velocities in the range
12-16~kms$^{-1}$. This was used to estimate various physical parameters for
evaluating the flare energetics using a 2-D model. Reconnection rates
and Poynting fluxes were estimated in the range of 0.5-3.0Vcm$^{-1}$ and
0.01-3.8 G erg cm$^{-2}$s$^{-1}$, respectively. These were sufficient
to release free energy of $10^{29}$ ergs of an M-class flare.
\item Filament/flux rope rising motion profile indeed showed
correspondence with various flare characteristics, viz., reconnection rate and
hard X-ray emission profiles. It started rising with velocity of 8-10
kms$^{-1}$, reaching a maximum of around 100 kms$^{-1}$ with
an average acceleration of 60 ms$^{-2}$ (estimated by projecting lateral
displacements of filament on to the vertical
direction observed in 304 \AA~ channel of AIA). Further, this flux rope accelerated
to the maximum velocity as the CME, observed at the peak phase of
the flare, followed by its deceleration to an average velocity of 590 kms$^{-1}$.
\item Flux variations in and around the filament were examined
before and after the eruption. We found some areas of flux changes,
co-temporal with the onset of
filament rise. For example, gradual reduction of positive flux was
found in box 3 (cf., Figure~\ref{magbx_prof}) at the onset time of filament rise.
From a careful study of 171 \AA~ images, it is interpreted as a consequence
of interaction of overlying field lines across the filament with side lobe field
lines resulting in tether weakening of the sigmoidal filament
system. In addition, flux emergence in box 5 located near the rising part of
the filament might have contributed to destabilize the system.
In summary, we infer that these flux changes
caused the loss of equilibrium leading to slow, upward rise of the
filament, and the onset of eruption by tether cutting reconnection.
In turn, changes occurred in photospheric fluxes in the decay
phase of the flare as a back reaction of this
reconnection (Figure~\ref{flx_chan}), in accordance with the
recent findings of \inlinecite{wang2010}.
\end{enumerate}
Destabilization of the filament system can occur due to either
ideal-MHD, or kink, instability \cite{kliem2006}. An increased
twist in the filament or flux rope system can become kinked by flux
emergence, because of which filament itself can rise. Another
possibility could be shear motions of photospheric fluxes; not
studied here. We suggest that the observed emergence/cancellation of
magnetic fluxes near the filament caused the flux rope to rise,
resulting in the tethers to cut and reconnection to take place
beneath the filament; in agreement with the tether cutting model. We
intend to pursue this study further by invoking observations of
vector magnetograms as boundary conditions for extrapolations to
look for changes in the coronal magnetic field and other associated
parameters.
\begin{acks}
The AIA(HMI) data used here are courtesy of SDO(NASA) and the
AIA(HMI) consortium. We thank AIA team for making available the
processed data. This work utilizes data obtained by the Global
Oscillation Network Group (GONG) Program, managed by the National
Solar Observatory, which is operated by AURA, Inc., under a
cooperative agreement with the National Science Foundation. We thank
the anonymous referee for carefully going through the manuscript and
making valuable comments which improves readability of manuscript
appreciably.
\end{acks}
|
\section{Introduction}
Hawking radiation is one of the most intriguing yet unobserved predictions of modern physics.
It is believed to be generated near the horizon surrounding a black hole and to be responsible for its ultimate decay
\cite{hawking1974,Unruh1976}. As noted by Unruh \cite{Unruh1981}, a phenomenon analog to Hawking radiation
should occur near a sonic horizon in a moving fluid,
i.e. near the interface separating regions of subsonic and supersonic current.
He argued that a thermal flux of phonons
should be spontaneously generated from the sonic horizon towards the subsonic region.
The effect finds its origin in the impossibility of defining a global quasiparticle vacuum which suits both incoming and outgoing states.
More recently, it has been proposed that flowing condensates of bosonic atoms could provide interesting
analogs of black hole physics \cite{Garay2000,Laughlin2003} and in particular of their Hawking radiation \cite{balbinot2008,carusotto2008,macher2009,finazzi2010,coutant2010}. The interface between subsonic and supersonic regions has also been shown to provide a scenario for the bosonic analog of Andreev reflection \cite{zapata2009}.
Except for black hole lasers (discussed in Refs. \cite{finazzi2010,coutant2010,leonhardt2008,corley1999}),
most proposals predict a spectrum (phonon current distribution per unit frequency) with a single peak at frequency $\omega=0$ falling as $1/\omega$ for small $\omega>0$, where $\hbar\omega$ is the quasiparticle energy measured with respect to the condensate chemical potential.
This means that the low-frequency peak is essentially thermal in character.
However, because of dispersive effects,
the effective temperature characterizing that zero-point radiation
is no longer universal. In a Bose-Einstein condensate, the group velocity of phonons increases at high frequency, unlike in the original model considered by Unruh \cite{Unruh1995}. As a result, this ``superluminal" transport dilutes the sonic horizon into a spatial interval of finite size \cite{finazzi2010,finazzi2011}. The blue-shifting effect accompanying with the sonic horizon implies that the dispersive properties of phonons are always involved. Nevertheless, when the dispersive length scale is much smaller than the horizon curvature scale, the temperature is determined by the local properties of condensate flow near the horizon (the gradient of the flow~\cite{Unruh1981})
in strict analogy with the standard Hawking radiation which is fixed by the surface gravity of the black hole.
This regime is found when the gradient is much smaller than one in units of the healing length~\cite{macher2009}.
Instead, when it is higher than one, the dispersion effects dominate and as a result
the effective temperature is fixed by the healing length and the
jump of the velocities across the sonic horizon~\cite{recati2009}.
In any case, the effective temperature will be smaller than the chemical potential \cite{lahav2010}.
As a consequence, a direct attempt to measure this radiation profile appears extremely difficult.
An alternative proposal to indirectly measure Hawking radiation relies on the squeezed character of the
state \cite{balbinot2008,carusotto2008} which could be observed at currently attainable temperatures by counting atoms on both sides of the horizon at coinciding times. However it should be pointed out that the main contribution to these correlations are due to the stimulated
amplification of pre-existing phonons \cite{carusotto2008,macher2009},
and not to the spontaneous amplification of vacuum fluctuations.
Here we study a new method which specifically aims at detecting the spontaneous contribution to
Hawking radiation. This approach relies on the strong frequency-dependence of resonant tunneling through a double barrier structure. Such a sonic black-hole analog behaves as a Fabry-Perot resonator for quasiparticles, with the peculiar feature that quasiparticles propagate linearly against a condensate background which is itself governed by a nonlinear equation. The Hawking emitted phonon spectrum shows peaks at frequencies different from zero. We find that, at currently achievable temperatures, thermal noise could be weak enough not to blur the characteristics of this resonant radiation, and a time-of-flight experiment could allow for its detection. Our proposal is quite similar to the black-hole laser setup \cite{finazzi2010,coutant2010,leonhardt2008,corley1999} in that
sharp peaks are found in both cases. However, although dynamical instabilities may appear occasionally, they are not a necessary feature of these type of setups (see Section \ref{QNM.sect}). We propose here to focus on situations that are dynamically stable, i.e. where all peaks are due to resonances. The dynamical stability of the flow is likely to be a valuable asset in actual experiments. A systematic study of resonances and instabilities is currently under investigation.
This paper is arranged as follows: In section II we present a mean-field study of the considered setup, discuss the main results and some general features. Section III is devoted to formulate the scattering problem of Bogoliubov quasiparticles propagating against the condensate background and to identify the essential features of Hawking radiation. In section IV we present and discuss the numerical results obtained for the Hawking radiation spectrum emitted from a double delta-barrier interface separating a subsonic and a supersonic region. Section V deals with the general distinction between quasinormal modes (or resonances) and dynamical instabilities, both of which are candidates to lie behind the sharp peaks in the radiation spectrum. Finally, Section VI is devoted to a summary and discussion of the main results. The main text is complemented by two appendices. Appendix A presents the analytical calculation of the mean-field model. Appendix B deals with the analytical resolution of the quasiparticle eigenvalue problem in the inhomogeneous region on the subsonic and supersonic sides near the double-barrier interface.
\section{Formulation of the model. Condensate wave function.}
We study an atom transport setup which is schematically depicted in Fig. \ref{figDDSamplePlot}. A quasi-1D bosonic condensate, which occupies the left region $x<0$, is allowed to leak to the right though two identical delta potential barriers. The leftmost barrier is conventionally placed at $x=0$ separated by a distance $d$ from the second barrier. We will see that this double-barrier setup behaves as a resonant structure. For convenience, we neglect quantum fluctuations of the condensate stemming from its one-dimensional character \cite{jackson1998,leboeuf2001}.
We may decompose the Heisenberg second-quantized field operator
\begin{equation}\label{totalfield.eq}
\widehat{\Psi}(x,t)=e^{-i \mu t/\hbar}\Psi_0(x)+\delta\widehat{ \Psi}(x,t)
\end{equation}
into a stationary condensate wave-function $e^{-i \mu t/\hbar}\Psi_0(x)$, with $\mu$ the chemical potential, and its fluctuations $\delta\widehat{\Psi}(x,t)$. In this section we focus on the condensate behavior. At low temperatures and densities, the mean-field equation which governs the stationary flow of a Bose-Einstein condensate is the time-independent Gross-Pitaevskii (GP) equation for the condensate wave function:
\begin{equation}\label{GPDim.eq}
\left[-\frac{\hbar^2}{2m} \frac{\partial^2}{\partial x^2} - \mu + V_{\rm ext}(x)+g_{\rm 1D}|\Psi_0(x)|^2\right] \Psi_0(x)=0 \; .
\end{equation}
We wish to study the effect of a potential consisting of two delta barriers of equal strength, although for comparison we will occasionally consider the single barrier case (see Ref. \cite{pavloffTBP}). Thus we will assume an external potential $V_{\rm ext}(x)$ which takes one of the two following forms:
\begin{equation}\label{potential.eq}
V_{\rm ext}(x)=\left\{
\begin{array}{c}
V_{1}(x)\equiv \hbar c_u z\delta(x) \\
V_{2}(x)\equiv \hbar c_u z\left[\delta(x)+\delta(x-d)\right]
\end{array}\right. \; ,
\end{equation}
where $c_u=\sqrt{g_u n_u / m}$, with $n_u \equiv \lim_{x \rightarrow -\infty} |\Psi_0(x)|^2$, is the speed of sound on the subsonic, upstream ($x<0$) side, and $z$ the dimensionless strength of each barrier. The healing length on the asymptotic upstream side is $\xi_u=\hbar/mc_u$.
In both the single- and double-barrier case, solutions can be found in which the condensate velocity is supersonic on the right of the second barrier, and subsonic from $-\infty$ to some point in the vicinity of the leftmost ($x=0$) barrier (see Appendix \ref{MFT.app} for a general discussion of these solutions). Those flow profiles must have one or more horizons, defined as points where the local condensate velocity equals the local speed of sound. This subsonic-supersonic scenario is the same which has been shown to display the bosonic analog of Andreev reflection, where the supersonic side plays the role of the normal fluid \cite{zapata2009}. Thus one may generally speak of Andreev-Hawking processes when dealing with scattering events undergone by elementary excitations of condensate flow through a subsonic-supersonic interface.
Such sonic analogs relying on condensate flow have a superluminal dispersion relation at higher frequencies and their horizons do not imply a strict causal disconnection among regions, but they are sufficient to produce Hawking radiation analogs (see however Ref. \cite{barcelo2006a} for a discussion of a scenario where horizons would not be needed). For the case of a single delta barrier the only horizon lies on the near left of the barrier. The double barrier case is richer: there appears one horizon in the vicinity of the $x=0$ barrier (which may lie on either side) and possibly one or more horizon pairs. The leftmost horizon, and the only one in the single delta barrier case, can be viewed as a black hole (BH) analog, because there the flow goes from the subsonic to the supersonic side. The possible additional pairs of horizons appearing in a double barrier structure are the analogs of white-hole/black-hole pairs, where a white hole (WH) is the time-reversed version of a BH.
White hole analogs seem to be extremely difficult to generate experimentally. There is a debate in the literature on whether white holes are stable at all, a question to which
linear stability analysis has not yet provided an answer \cite{macher2009,mayoral2011}.
To clearly identify what constitutes Hawking radiation, the experiment should ideally be done in such a way that the condensate wave-function is stationary and asymptotically flat: $\rho'(\pm\infty)=0$, where $\rho(x)\equiv|\Psi_0(x)|^2$ \cite{birrell1982}. In Appendix \ref{MFT.app} we show that, with those homogenous boundary conditions, there is one or more mean-field solutions for two delta potential barriers, and just one for a single delta barrier. A sketch of a typical density for two delta barriers is shown in Fig. \ref{figDDSamplePlot}.
In Fig. \ref{figDeltaStrengthPlot} we plot the region of the $(z,q)$ plane for which stationary solutions with homogeneous boundary conditions exist to the double-barrier problem. We recall that $z$ is the dimensionless parameter characterizing the strength of the two identical delta barriers, while $q$ is the condensate momentum on the subsonic (upstream) side. We note that, for a given value of $q$, there is a minimum and a maximum barrier strength between which a solution is guaranteed to exist for that particular value of $q$. This contrasts with the behavior of the single barrier case, for which only one solution exists for a given value of $q$. Interestingly, the resulting line $z(q)$ line for the single-barrier case coincides with the upper boundary of the shaded region in Fig. \ref{figDeltaStrengthPlot}.
Figure \ref{figPhaseDiagplot} shows a representative sample of solutions in the $(d,z)$ plane, where $d$ is the distance between barriers. Each curve represents a value of the condensate momentum $q$. Inspection of Fig. \ref{figPhaseDiagplot} reveals that, for a given $q$, there is a finite interval of allowed $z$ values. That range has been shown in Fig. \ref{figDeltaStrengthPlot}. Figure \ref{figPhaseDiagplot} reveals that, for given $z$ and $q$, a multiplicity of $d$ solutions exist. The two lowest $d$ values can be characterized by two different values of $\rho(0)$, which we call $\rho_{\rm max}$ and $\rho_{\rm min}$. The upper $d$ solutions come also in pairs and are regularly separated by a distance difference equal to the period of the nonlinear oscillations between the two barriers, which is the same for both $\rho_{\rm max}$ and $\rho_{\rm min}$.
An interesting feature is that, at small barrier separations, a single $q$ solution exists for a given $d$ and $z$. For higher $d$ (slightly above 2 for the sample of curves shown in Fig. \ref{figPhaseDiagplot}), two $q$ solutions exist for given $d$ and $z$. For still higher $d$, three $q$ solutions exist for given $d$ and $z$, and so forth. Since the barrier separation $d$ and the barrier strength $z$ are expected to be experimentally adjustable parameters, the clear trend is that, the larger the distance, the higher the number of allowed stationary solutions, a fact which could translate itself into instabilities.
\section{Bogoliubov analysis}\label{BdG.sect}
A general and detailed introduction to the Hawking radiation physics in Bose-Einstein condensates, within a Bogoliubov-deGennes description can be found in Refs. \cite{recati2009,macher2009,leonhardt2003,leonhardt2003A}. In this section we mainly introduce notation while refer the reader to those works for a more complete presentation.
The quantum fluctuation part introduced in Eq. (\ref{totalfield.eq}), $\delta\widehat{ \Psi}(x,t)$, can be subject to a canonical transformation resulting in the expansion
\begin{equation}\label{BogDecomp.eq}
\delta\widehat{ \Psi}(x,t)= e^{-i \mu t/\hbar} \sum_i \left[ u_i(x)e^{-i\omega_i t} \hat{\gamma}_i+v_i^*(x)e^{i \omega_i t} \hat{\gamma}^\dag_i \right] \; ,
\end{equation}
where $u_i(x)$ and $v_i^*(x)$ are the components of the wave function of the bosonic quasiparticle created by $\hat{\gamma}^\dag_i$. These $u,v$ components satisfy the bosonic Bogoliubov-deGennes (BdG) equations:
\begin{eqnarray}\label{BdGDim.eq}
\hbar \omega \left[
\begin{array}{c}
u(x) \\
v(x) \\
\end{array}
\right]
&=&
\left[
\begin{array}{cc}
\hat{H}& g_{\rm 1D} \Psi_0(x)^2 \\
-g_{\rm 1D} \Psi_0^*(x)^2 & -\hat{H} \\
\end{array}
\right]
\left[
\begin{array}{c}
u(x) \\
v(x) \\
\end{array}
\right]\\
\nonumber
\hat{H}&\equiv&-\frac{\hbar^2}{2m} \frac{\partial^2}{\partial x^2} - \mu + V_{\rm ext}(x)+2 g_{\rm 1D}|\Psi_0(x)|^2 \; ,
\end{eqnarray}
while the $\hat{\gamma},\hat{\gamma}^\dag$ operators satisfy, for real $\omega$,
\begin{equation}\label{commutator.eq}
\left[ \hat{\gamma}_i, \hat{\gamma}^\dag_j \right] = \int dx \left[ u_i^*(x) u_j(x) - v_i^*(x) v_j(x) \right]=\nu_i \delta_{ij} \; ,
\end{equation}
where the normalization $\nu_i$ can be set to $\pm 1$.
In Eq. (\ref{BdGDim.eq}) $g_{\rm 1D}=2\hbar^2 a_0 /(m a_\bot^2)$ is the effective 1D coupling constant, where $a_0$ is the s-wave scattering length and $a_\bot$ the transverse confinement harmonic oscillator length, and $V_{\rm ext}(x)$ is an externally imposed potential. We note that the sum over $i$ in the first equation may be interpreted as including both $\omega_i\gtrless0$, with $\nu_i>0$, or both $\nu_i\gtrless0$, with $\omega_i>0$.
The role of the horizon is appreciated when doing a WKB-type approximation for the excitations (see Refs. \cite{leonhardt2003,leonhardt2003A}). Close to the horizon are the turning points of the low energy classical trajectories where the WKB-solution is not appropriate. By matching the solutions on both sides and assuming some general scale separation (see also Ref. \cite{unruh2005}), one can show that in the case of just one horizon the profile of emitted Hawking radiation approaches $1/\omega$ at low frequencies. This important result guarantees the (approximately) thermal radiation profile. Thus at low frequencies, and in the absence of other scattering obstacles, the zero-point radiation has a $1/\omega$ behavior which makes it in principle difficult to distinguish from truly thermal behavior.
Due to the presence of delta barriers in our chosen configuration, WKB-type approximations cannot be used uncritically. However, thanks to a theorem on dark-soliton perturbation theory (see Ref. \cite{chen1998}) we know that we have at our disposal a complete set of solutions on the left of the $x=0$ barrier. In the flat supersonic region ($x>d$) the solutions are even simpler to work out. We refer the reader to Appendix \ref{BdG.app} for both cases, $x<0$ and $x>d$. The only non-analytically solvable problem lies between barriers ($0<x<d$), but this is a finite region where numerical integration of the BdG equations (\ref{BdGDim.eq}) requires only a moderate computational effort.
The next step is to identify the relevant scattering states. This has been done for Bose-Einstein condensates in studies of BH analogs \cite{recati2009,macher2009} and of bosonic Andreev reflection \cite{zapata2009}. In this article we focus on the scattering states and their connection to quasi-normal modes (or resonances) and dynamical instabilities. A systematic study of complex-energy eigenmodes is left for a future work \cite{zapataTBP}. In the asymptotic regions, propagating modes obey the Bogoliubov dispersion relation. Following Ref. \cite{recati2009} we label the asymptotic regions with indeces $u$, for upstream ($x\rightarrow-\infty$), and $d$ for downstream ($x\rightarrow\infty$). The upstream dispersion relation is
\begin{equation}\label{BogDR.eq}
\omega_u(k)=v_{u} k \pm c_{u}|k|\sqrt{1+(k\xi_{u})^2/4} \; ,
\end{equation}
where $v_u$ is the upstream flow velocity, and similarly for downstream $\omega_d(k)$.
A graph of this relation is shown in Fig. \ref{figDispRelation} for both the subsonic and supersonic side. The branches shown in blue/red correspond to the $+/-$ of Eq. (\ref{BogDR.eq}) and can be shown to lead to positive/negative normalization. Modes are named after the sign of their group velocity (in/out according to whether they approach/leave the scattering structure), location (u/d) and, in the supersonic case, 1-2 stands for modes with normal/anomalous (i.e. positive/negative) normalization. Importantly, the anomalous d2 modes exist only for frequencies $\omega < \omega_{\rm max}$.
Here we are mainly interested in the scattering state with frequency $\omega$ characterized by the incoming channel d2-in. Its wave function reads
\begin{equation}\label{d2-in.eq}
{u_{\rm{d2-in, \omega}}(x)\brack v_{\rm{d2-in, \omega}}(x)}=\left\{
\begin{array}
[c]{ll}
\psi_u(x), & x\rightarrow-\infty\\
\psi_d(x), & x\rightarrow\infty
\end{array}
\right.
\end{equation}
where
\begin{equation}\label{psid.eq}
\begin{split}
\psi_d(x)&={u_{d2}(k_{\rm{d2-in}}) e^{i q_d x} \brack v_{d2}(k_{\rm{d2-in}}) e^{-i q_d x}} \frac{e^{i k_{\rm{d2-in}} x}}{\sqrt{2\pi |w_{d2}(k_{\rm{d2-in}})|}}
\\ &+S_{\rm{d1d2}}(\omega) {u_{d1}(k_{\rm{d1-out}}) e^{i q_d x}\brack v_{d1}(k_{\rm{d1-out}}) e^{-i q_d x}} \frac{e^{i k_{\rm{d1-out}} x}}{\sqrt{2\pi |w_{d1}(k_{\rm{d1-out}})|}}\\
&+S_{\rm{d2d2}}(\omega){u_{d2}(k_{\rm{d2-out}}) e^{i q_d x}\brack v_{d2}(k_{\rm{d2-out}}) e^{-i q_d x}} \frac{e^{i k_{\rm{d2-out}} x}}{\sqrt{2\pi |w_{d2}(k_{\rm{d2-out}})|}},
\end{split}
\end{equation}
\begin{equation}\label{psiu.eq}
\psi_u(x)=S_{\rm{ud2}}(\omega){u_u(k_{\rm{u-out}}) e^{i q_u x}\brack v_u(k_{\rm{u-out}}) e^{-i q_u x}} \frac{e^{i k_{\rm{u-out}} x}}{\sqrt{2\pi |w_u(k_{\rm{u-out}})|}}
\end{equation}
where $q_u,q_d$ are the up- and down-stream condensate momenta, $w=d\omega/dk$ denote the relevant group velocities, and the $S$-matrix elements are shown. In Eqs. (\ref{d2-in.eq})-(\ref{psiu.eq}) $u_\alpha(k)$ and $v_\alpha(k)$ are the spinor components for a quasiparticle of the uniform Bose gas, in channel $\alpha$ at momentum $k$. A similar decomposition can be made for the other in-modes. We note that the $k$'s of the previous equation are solutions of Eq. (\ref{BogDR.eq}) for a given $\omega$ lying in the interval $0<\omega<\omega_{\rm max}$, each solution defining a particular mode. Not shown in the previous equation but necessary for the matching, an evanescent solution exists on the subsonic side. The same holds on the supersonic side for $\omega>\omega_{\rm max}$.
To avoid double counting one can either choose all the positive normalization modes (for both in and out), as is done in Ref. \cite{zapata2009}, and then one has to deal with negative frequency modes, or as is usual in this Hawking context, choose only positive frequency modes and then interchange $\hat{\gamma}_i \leftrightarrow -\hat{\gamma}^\dag_i$ for the negative normalization ones \cite{footnote1}. In this work we adopt the convention of $\omega_i>0$ and write:
\begin{eqnarray}
\delta\hat{ \Psi}(x)=\int_0^\infty d\omega \sum_{I={\rm u-in,d1-in}} [ u_{I, \omega}(x) \hat{\gamma}_{I, \omega}+v^*_{I, \omega}(x) \hat{\gamma}^{\dag}_{I, \omega} ] \\
\nonumber
+\int_0^{\omega_{\rm max}} d\omega [ u_{\rm{d2-in, \omega}}(x) \hat{\gamma}^{\dag}_{\rm{d2-in, \omega}}+v^*_{\rm{d2-in, \omega}}(x) \hat{\gamma}_{\rm{d2-in, \omega}} ] \\ \; .
\end{eqnarray}
The normalization chosen in Eqs. \ref{d2-in.eq}-\ref{psiu.eq} guarantees that modes are normalized to unit quasiparticle current and so $[\hat{\gamma}_{I, \omega},\hat{\gamma}^{\dag}_{I', \omega'}]=\delta_{II'} \delta(\omega-\omega')$. An identical expression may be written changing in $\rightarrow$ out. Standard scattering theory arguments show that the $S(\omega)$-matrix coefficients connect the in-modes to the out-modes:
\begin{equation}\label{S-matrix.eq}
(\hat{\gamma}^{\dag}_{\rm{u-out, \omega}}, \hat{\gamma}^{\dag}_{\rm{d1-out, \omega}}, \hat{\gamma}_{\rm{d2-out, \omega}})=(\hat{\gamma}^{\dag}_{\rm{u-in, \omega}}, \hat{\gamma}^{\dag}_{\rm{d1-in, \omega}}, \hat{\gamma}_{\rm{d2-in, \omega}}) S^{\dag}(\omega) \; .
\end{equation}
Due to the pseudo-Hermitian character of the BdG equations (\ref{BdGDim.eq}),
this $S(\omega)$-matrix obeys, for frequencies $\omega<\omega_{\rm max}$, a pseudo-unitary relation $S^{\dag}(\omega)\eta S(\omega)=\eta$ with $\eta={\rm diag}(1,1,-1)$. This enforces non-standard relations among the transmission and reflection coefficients, for example, $|S_{\rm{d2d2}}(\omega)|^2-|S_{\rm{ud2}}(\omega)|^2-|S_{\rm{d1d2}}(\omega)|^2=|S_{\rm{ud1}}(\omega)|^2+|S_{\rm{d1d1}}
(\omega)|^2-|S_{\rm{d2d1}}(\omega)|^2=1$ (see Refs. \cite{zapata2009,recati2009,macher2009}).
The case $\omega>\omega_{\rm max}$ is simpler because the $S(\omega)$ matrix is unitary and there are no anomalous reflections or transmissions. Thus all asymptotic states can be chosen with $\nu>0$. If we rewrite $R_{I}(\omega):=S_{II}(\omega), T_{IJ}(\omega):=S_{IJ}(\omega)$ with $I,J$ different and taking values $I,J = u,d1$, then the usual unitarity relation $|R_{I}(\omega)|^2+|T_{IJ}(\omega)|^2=1$ applies.
The upstream phonon flux spectrum can be computed from these considerations and for $\omega<\omega_{\rm max}$ shows the remarkable form \cite{recati2009,macher2009}:
\begin{equation}\label{HawkingFlux.eq}
\frac{dI_{\rm{u-out}}}{d\omega}=|S_{\rm{uu}}(\omega)|^2 \frac{dI_{\rm{u-in}}}{d\omega}+|S_{\rm{ud1}}(\omega)|^2 \frac{dI_{\rm{d1-in}}}{d\omega}+|S_{\rm{ud2}}(\omega)|^2 \left(\frac{dI_{\rm{u-in}}}{d\omega}+1\right).
\end{equation}
Assuming that we can populate the ingoing fluxes with comoving thermal populations, namely, $dI_{\rm{\alpha-in}}/d\omega=n_B(\Omega_{\rm{\alpha-in}})$, where $\alpha=u,d1,d2$, $n_B(\Omega)=(e^{\beta\hbar\Omega}-1)^{-1}$ is the Bose-Einstein occupation at the common temperature $\beta^{-1}:=k_B T$, and $\Omega_{\alpha-\rm{in}}$ is the comoving frequency of the mode $\alpha-\rm{in}$:
\begin{equation}\label{Doppler.eq}
\Omega_{\rm{\alpha-in}}(\omega)=\omega-v_{\alpha} k_{\rm{\alpha-in}}(\omega),
\end{equation}
where $k_{\rm{\alpha-in}}(\omega)$ is the solution to Eq. (\ref{BogDR.eq}) for given $\omega$, with $v_{\alpha}=v_u,v_d$ the flow velocities. Equation (\ref{HawkingFlux.eq}) reveals that, even at $T=0$, a nonzero upstream flow of energy (phononic Hawking radiation) must be expected. The spectrum for $\omega>\omega_{\rm max}$ is of the same form as Eq. (\ref{HawkingFlux.eq}) but with the last term removed, i.e. without any zero-point energy flux.
\section{Hawking radiation spectra}
Figure \ref{grHawkingPlot} shows some frequency spectra of the upstream phononic flow [see Eq. (\ref{HawkingFlux.eq})] for the setups and condensate solutions depicted previously in Fig. \ref{grDdPlot}, with a one-to-one correspondence between the graphs in the two figures. The top-left graph of Fig. \ref{grHawkingPlot} shows a structureless profile with a peak at $\omega=0$ followed by a $1/\omega$ tail, which is typical of Hawking radiation profiles in the absence of barriers, and very similar to what is obtained in the single delta barrier case. The message is that two nearby barriers behave similarly to a single barrier of double strength.
We may also note that the zero temperature contribution (thick blue line) is not easily distinguishable from the total contribution at nonzero temperature (thin red line), because at low frequencies they both follow the same thermal law $1/\omega$. This property seems to be common to all structures which do not permit one or more nonlinear oscillations of the condensate between the barriers. An exception to this trend occurs when the delta barrier strength $z$ is close to its upper limit. Then a bigger separation among barriers is possible and, as a consequence, a peak at nonzero $\omega$ may develop (see the upper-right graph at Fig. \ref{grHawkingPlot}). A general trend that can be clearly appreciated in the rest of the graphs is that the largest the separation the more peaks appear. In particular the two bottom graphs in Fig. \ref{grHawkingPlot} exhibit a double peak in the allowed frequency interval. The reader might wonder why panels (e) and (f) of Fig. \ref{grHawkingPlot} look somewhat different since the parameters hardly vary, the only difference being a relative change in the inter-barrier distance of approximately $0.01$. The reason is that the net amplification factor results from a rather complicated expression involving interferences between the scattering coefficients on both sides of the scattering region. We refer to Eq. (69) and Fig. 5, right panel of Ref. \cite{finazzi2010}, where a similar sensitive dependence has been found.
We note that spontaneous phonons are generated at the event horizons and, after multiple scattering by barriers and horizons, they are partly emitted into the subsonic side. Those scattering events include normal processes ($u$ and $v$ components do not mix) and Andreev (or anomalous) processes ($u$ converts into $v$ or viceversa, see. Ref. \cite{zapata2009}). The bigger the separation between the barriers, the more quasi-bound states can be accommodated between them and the more peaks appear in the Hawking radiation spectrum.
An important point which will be discussed in greater depth in the next section, is that peaks can be due to resonances (also called quasinormal modes) or instabilities, so more information is needed to classify them. An unequivocal experimental signature of Hawking radiation (associated to zero-point quantum fluctuations of an otherwise stationary state) would undoubtedly be favored by a transport regime where the classical flow is dynamically stable. We recall that the $S(\omega)$ matrix is pseudo-unitary in the Hawking sector $0<\omega<\omega_{\rm max}$. The general spectral theorem on these matrices \cite{mostafazadeh2004} guarantees that eigenvalues come in pairs of $s_i, 1/s_i^*$ or have unit modulus, $|s_j|=1$. A general trend we have observed is that in the region $\omega\alt\omega_{\rm max}/3$ most of scattering matrices show only one unit-modulus eigenvalue. By contrast, in the region $\omega\agt\omega_{\rm max}/2$, most of the $S$-matrices have three unimodular eigenvalues. Nevertheless, the theorem mentioned before guarantees that the determinant will be a pure phase. Then, as in a conventional phase shift analysis of quantum mechanical scattering, a phase jumping upwards when crossing through a peak from low to high $\omega$ can be interpreted as a resonance. On the contrary, a jump downwards will be reveal an instability. Hence, a most convenient way to detect resonances while distinguishing them from instabilities is to simultaneously plot the phase of the determinant of the $S(\omega)$ matrix. This corresponds to the thin green line appearing in all the graphs of Fig. \ref{grHawkingPlot}. All curves show a clear resonance behavior, with the exception of the middle-right phase curve, which reveals an instability, as indicated by the sudden drop of the green line when traversing the instability. Most of the cases which we have studied show QNM behavior, but the the occurrence of instabilities is not so rare that it can be ignored. From our analysis, one cannot rule out the existence of strong instabilities with a large imaginary part in the eigenfrequency, since these would generate not easily detectable structures in the frequency spectrum. However, from the systematic analysis of Refs. \cite{finazzi2010, coutant2010}, where such instabilities where not found, one can conjecture that also in the present case those strong instabilities will not be found. An analysis of the time-dependent Gross-Pitaevskii equation in real time could establish that this is indeed the case. These considerations underline the need for a systematic search of instabilities and QNMs as poles of the propagator in the complex energy plane \cite{zapataTBP}.
The examples shown in Figs. \ref{grDdPlot}-\ref{grHawkingPlot} were chosen because they clearly reveal general trends. Unfortunately, in a time of flight (TOF) expansion, which correlates the long time density distribution with the momentum distribution of the initial state, their peaks barely stand out above the background of the depletion cloud, and even less when thermal fluctuations are taken into account. We have included in Fig. \ref{grTOFPlot} a set of two more favorable setups which show large signals for zero-point Hawking radiation. The two bottom graphs show the momentum distribution of the initial trapped state and are computed in the approximation where only the subsonic flow is included and boundary effects are neglected.
When a TOF measurement is performed in such systems, the depletion contribution is negligible at momentum values which however reveal clear resonant peaks. Zero-point Hawking radiation nearly exhausts those peaks at low temperatures and still gives the main contribution to the area under the peak at temperatures as high as $0.9\mu$. This fact could allow for the unequivocal detection of Hawking radiation.
\section{Quasi-normal modes (resonances) and dynamical instabilities}\label{QNM.sect}
A discussion of instabilities in BEC black-hole analogs can be found in Refs. \cite{leonhardt2003,finazzi2010,leonhardt2003A,barcelo2006}. For a study of QNMs in a similar context, we refer to Ref. \cite{barcelo2007}. General studies of QNM in gravitational black holes and in optical analogs can be found in Refs. \cite{kokkotas1999} and \cite{ching1998,settimi2009}, respectively. A more complete study for the setup considered in this article will be presented \cite{zapataTBP}.
We have said that peaks in the spectrum of Hawking radiation may be due to resonances or instabilities. The former are characterized by poles of the analytical continuation of the $S(\omega)$ matrix in the complex $\omega$-plane with Im$(\omega)<0$ while the latter have Im$(\omega)>0$. Moreover, both types of complex modes may have $|\rm{Im}(\omega)|$ so large that they are not clearly appreciated in the radiation spectrum, yet they can hide important instabilities. While a systematic search for poles is left for a future work, we discuss here some general features of the behavior of QNMs and instabilities in the complex $k$-plane.
When a linear stability analysis is made of a stationary solution of the GP equation (\ref{GPDim.eq}), instabilities appear as solutions of the BdG equations (\ref{BdGDim.eq}) whose frequency $\omega$ has a positive imaginary part. Because complex frequency implies complex asymptotic wave numbers, these solutions must be localized in real space. For a given complex frequency, there are always four complex $k$'s (counting each different $k$ with its multiplicity), two with Im$(k)>0$ and two with Im$(k)<0$.
Like in the search for bound states in conventional quantum mechanics, a discrete, possibly empty set of modes is to be expected. The mode with the largest Im$(\omega)$ dominates the long time behavior and its inverse is a good measure of the decay time of the condensate due to that instability. There is a well defined procedure to quantize such unstable modes \cite{rossignoli2005}, which show up as essentially free particles (instead of oscillators). In open systems like the presently analyzed, there are however some subtleties in that quantization procedure which, to our knowledge, have not been properly addressed in the literature. Specifically, some spectral properties of the BdG operators are used which can only be guaranteed for finite-dimensional operators \cite{brink2001}.
At a given real frequency between 0 and $\omega_{\rm max}$ we may have have four complex $k$ solutions on each side of the interface. These are shown by circles in Fig. \ref{figComplexKPlot}. On the upstream side (left graph), they correspond to an incoming, outgoing, evanescent, and exploding solution. Downstream (right graph) we have two incoming and two outgoing solutions, corresponding to the normal (d1) and anomalous (d2) channels. For instance, a conventional retarded scattering state is characterized by one incoming channel matching all possible outgoing and evanescent solutions, with no amplitude for exploding or other incoming solutions.
An instability is a bound state made exclusively of spatially decaying solutions that happen to match at a particular complex frequency (with positive imaginary part). Localized solutions correspond to Im$(k)<0$ on the upstream side and to Im$(k)>0$ on the downstream side. Figure \ref{figComplexKPlot} shows how the various $k$ solutions evolve as $\omega$ varies from Im$(\omega)=0$ to Im$(\omega)>0$ while Re$(\omega)$ remains constant. Eventual instabilities can only be obtained by matching the evolved ``u-out'' and ``evanescent'' solutions upstream (Im$k<0$) and the evolved ``d1-out" and ``d2-out" downstream (Im$k>0$), with vanishing amplitudes for the other solutions. We emphasize that the wave function of these unstable modes involves exclusively analytical continuations of outgoing and evanescent solutions.
QNM modes, on the other hand, are obtained from the analytical continuation to the Im$(\omega)<0$ half-plane of the retarded Green's function of the time-dependent version of the Bogoliubov-deGennes equations (\ref{BdGDim.eq}) (see Refs. \cite{ching1998,settimi2009} for a discussion in an optical context but precisely translatable to the present one). More specifically, these QNMs can be obtained as an analytical continuation to Im$(\omega)<0$ of scattering states involving only outgoing and evanescent solutions, i.e. without the intervention of incoming waves. Like for instabilities, a discrete number of QNMs is to be expected. Therefore, in order to find resonances, we are only interested in the evolution of ``u-out'' and ``evanescent'' solutions on the upstream side, and ``d1-out" and ``d2-out" on the downstream side, ruling out the other solutions. Interestingly, this is exactly the same set of solutions which are relevant in the search for instabilities. We conclude that instabilities and resonances are obtained from the matching of the same set of solutions albeit in different regions of the complex energy plane.
Hence, QNM/instabilities correspond to poles of the $S(\omega)$ matrix when analytically continued to the lower/upper complex plane from the real energy line. This can be readily seen if one allows for a general coefficient in the in-wave component of the scattering solution Eq. (\ref{d2-in.eq}), for example. So both resonances and instabilities are candidates to explain peaks in the square of a given $S(\omega)$ matrix coefficient. The frequency of those possible underlying modes must of course have an imaginary part much smaller than their real part. Since the determinant of the $S(\omega)$ matrix is a pure phase (as noted in the previous section), it can be used to discriminate between both behaviors. Specifically, a jump upwards/downwards of the phase when traversing the peak with increasing frequency, implies a QNM/instability.
Figure \ref{grHawkingPlot}, together with other not shown data, reveals that peaks in the Hawking spectrum are mostly due to resonances, instabilities being more the exception than the norm. While this sampling is encouraging, a systematic search for QNMs and instabilities in various stationary flows is left for future work.
Finally we note that, both for QNMs and instabilities, the boundary conditions here described are different from those adopted in Ref. \cite{barcelo2006,barcelo2007}.
\section{Discussion and conclusion}
We have studied the flow of an atomic condensate through a double barrier interface separating regions of subsonic and supersonic flow. Such a setup provides a scenario where Hawking radiation into the subsonic side is enhanced at some frequencies due to multiple scattering of quasiparticles by the two barriers and the modulations of the condensate. The resulting highly non-thermal Hawking radiation presents peaks at frequencies which may lie well above the working temperature and thus can be unambiguously interpreted as stemming from quantum fluctuations of the quasiparticle vacuum. The non-thermal Hawking spectrum emitted by the double barrier interface represents and important advantage over the cases of single or zero barrier, where the low-frequency zero-point radiation has a thermal character which makes it more difficult to distinguish from genuinely thermal radiation.
Our calculation has been based on a model of stationary flow of condensate and quasiparticles through a double barrier structure with open boundary conditions, where the condensate density is asymptotically flat on both sides of the structure, while quasiparticle motion is described by scattering states characterized by incoming channels that are thermally populated. While a stationary scattering picture of transport has proved to hold predictive power in electron systems, it still represents an idealized scenario in cold atom contexts, where stationary circuits have not yet been developed and where transport of finite-sized condensates is mostly investigated within a time-dependent scheme \cite{Aspect2009}. As long as steady condensate transport is still an item for the future, it will be of interest to perform numerical simulations of time-dependent transport that may reveal those features of the Andreev-Hawking phenomena which we have explored here.
An important question is whether, in currently achievable setups, the resonance frequency can be tuned to lie well above the currently attainable temperatures that would characterize the incoming quasiparticle population. We have seen that, while in a time-of-flight experiment the contribution from Hawking radiation peaks may be easily overshadowed by the contribution of the depletion cloud, setups can be designed where the resonant Hawking peaks are sufficiently sharp to be clearly visible even at temperatures comparable to the chemical potential.
We thank A. Aspect, C. D\'{\i}az Guerra, L. Garay, P. Leboeuf, N. Pavloff, G. V. Shlyapnikov, and C. Westbrook for valuable discussions. This work has been supported by the joint France-Spain Acci\'on Integrada HF2008-0088 (PHC - Picasso Program). Support from MICINN (Spain) through grants FIS2007-65723 and FIS2010-21372, from Comunidad de Madrid through grant MICROSERES-CM (S2009/TIC-1476), and from the Swiss National Science Foundation, is also acknowledged.
|
\section{Introduction}
The thermodynamics of matter plays an important roll in the study of its fluid dynamics. The solution of hydrodynamic equations, both in the relativistic and non-relativistic formalism, require a knowledge about the equation of state (EOS) of the system.
An EOS is essentially a relation between the macroscopic quantities e.g. the pressure $p$, the total mass-energy density $\rho$, and rest mass energy density $\rho_0$. One could arrive into such a relation from the knowledge of the probability distribution of various microscopic quantities (e.g molecular velocity, energy). For the relativistic gas, such a study has been done using the formalism of statistical mechanics in a Lorentz covariant framework (See [\onlinecite{Synge1957}] and ref. therein, [\onlinecite{Chandrasekhar1939}], [\onlinecite{Tolman1934}], [\onlinecite{Taub1948}]). Distribution function of co-ordinate velocity ($\frac{dx}{dt}$) and the EOS can be derived using this formalism. Several authors tried to use relativistic EOS to solve the relativistic hydrodynamics (RHD) equation in various astrophysical problems ([\onlinecite{Shen1998}]). But a direct application ([\onlinecite{Falle1996}]) of the exact relativistic EOS turns out to be a less feasible method for numerical computing because of its complexity. Thus in the popular studies of numerical RHD in astrophysical problems (see review [\onlinecite{Wilson2003}]), people prefer to use various alternative models of EOS which can reproduce various features of exact EOS with a good approximation ([\onlinecite{Ryu2006}], [\onlinecite{Mignone2005}]). All these models ([\onlinecite{Sokolov2001}]) are proposed in an ad-hoc manner and are not derived consistently from the first principle.
The present status of RHD study motivates us to analyze this problem in a model independent way in particular, we wish to find a simpler form of EOS which, apart from being computer friendly, must also be derived from very first principle of kinetic theory. At this point, it would be interesting to see whether one could derive a probability distribution function of 4-velocity following the basic methods of probability theory and assumptions of isotropy as originally had been done by MB in the case of non-relativistic gas.This approach, to the best of our knowledge, is yet unexplored and one could examine it atleast for the sake of conceptual completeness. An EOS could be derived then using this distribution function.
In this article, we follow the original line of arguments of MB to find a distribution function of 4-velocity ($u^i=\frac{dx^i}{d\tau}$), instead of co-ordinate velocity $\frac{dx^i}{dt}$, for a relativistic ideal gas. We then use these Maxwell like 4-velocity distribution function to derive a form of EOS.
This paper is organized as follows. In \S II, we derive Maxwell like 4-velocity distribution function for relativistic gas.
In \S III, we apply it to find the EOS for relativistic gas. In \S IV, we investigate the behavior of some physical quantities (e.g. $\gamma$ and sound speed etc.). The extreme limits: ultra-relativistic and non-relativistic are discussed. Finally concluding remarks are drawn in \S V.
\section{Four velocity distribution of a relativistic ideal gas}
We consider a gas consisting of one species of non-interacting particles. The temperature of the gas is so high that the average kinetic energy of a constituent particle is comparable to its rest mass energy. In other words, the thermal energy of the gas is comparable with its rest mass energy.
The co-ordinate velocity and the 4-velocity of a gas particle are defined respectively as
$v^{i}\!=\!\frac{dx^{i}}{dt}$ and $u^{\mu}\!=\!c\frac{dx^{\mu}}{d\tau}$
where $d\tau^2\!=\!c^2dt^2\!-\!(d\vec x)^2$ is the length element in the Minkowski space-time.
\!$u^{\mu}$'s are related to the co-ordinate velocity as
\begin{eqnarray}
u^{\mu}=c\frac{dx^{\mu}}{d\tau}=v^{\mu}/\sqrt{1-{{v^2}\over{c^2}}},
\label{umu}
\end{eqnarray}
where $v^{\mu}=(1,\vec v)$.
The molecules of the gas interacts only through elastic collision
so that all of their energy is kinetic energy. In a Lorentz frame, where the center of mass of the container
is at rest, the distribution of molecular speed is isotropic.
The choice of such a frame is unique upto a three dimensional rotation.
In different frames in which the container is at rest, $u^0$ takes same value and $u^i$s are transformed
by a 3-D Euclidean rotation {\it i.e} they transform like a components of an ordinary three dimensional
vector in non-relativistic case. According to special relativity
\begin{eqnarray}
v_x^2+v_y^2+v_z^2\leq c^2.
\label{constraint}
\end{eqnarray}
where $v^i=\frac{d x^i}{dt}$, is the co-ordinate velocity.
In the non-relativistic case (where $v_x, v_y, v_z $ are not constraint by the above inequality),
we write the probability of finding the particle simultaneously in the velocity range
$v_x$ to $v_x+dv_x$, $v_y$ to $v_y+dv_y$, and, $v_z$ to $v_z+ dv_z$ as the product of the individual probabilities:
\begin{eqnarray}
p(v_x, v_y, v_z)dv_x dv_y dv_z= f(v_x)f(v_y)f(v_z)dv_xdv_ydv_z\ \ \ \ \
\label{cmbprbl}
\end{eqnarray}
where $f(v_i)dv_i$ represents the probability of finding the particle in the velocity component range $v_i$ to $v_i+dv_i$.
The functional form of `$f$' is same for all the components due to isotropy. This assumptions ($\ref{cmbprbl}$)
is no longer valid in relativistic case
as the co-ordinate velocity component are no longer independent due the above inequality and can only vary from $-c$ to $c$.
This problem can be circumvented if we seek a formula for the probability distribution of 4-velocity instead of coordinate velocity.
The 4-velocity components are related through the equation
\begin{eqnarray}
\sum_{i=1}^3(u^i)^2=c^2[(u^0)^2-1]=v^2/(1-\frac{v^2}{c^2}).
\end{eqnarray}
But, as $v\rightarrow c$, $u^{0}\rightarrow \infty$ and therefore, each of the three spacial components ($u^1\!, u^2\!, u^3$)
of the 4-velocity can take values from $-\infty$ to $\infty$ independently,
irrespective of the values of other two components. Therefore,
just as the 3-velocity components in the non-relativistic case,
the three spacial components of 4-velocities have no constraint among themselves.
In the chosen Lorentz frame, the center of mass of the gas system is
at rest. The average 4-momentum of the total gas system is zero in all directions. This implies that
the probability distribution function of the three spacial components of the 4-velocity is isotropic just as
the probability distribution function for velocities in non-relativistic case.
Let the probability of finding a gas molecule in between $u^i$ to $u^i+du^i$ be
\begin{eqnarray}
p_i(u^i)du^i=F_i(u^i)du^i, \ \ \ \ (i=1,2,3).
\label{4prbl1}
\end{eqnarray}
The assumption of isotropy demands that the functional from of $F_i$ is same for all values of $i$, {\it i.e.}
$F_1=F_2=F_3$. As already explained above the three spatial components of 4-velocity $u^1, \ u^2, \ u^3$ are not
constrained, hence the probability that a gas
particle is found simultaneously within the 4-velocity range $u^1$ to $u^1+du^1$, $u^2$ to $u^2+du^2$, and
$u^3$ to $u^3+du^3$ is
\begin{eqnarray}
p(u^1\!,u^2\!,u^3)du^1du^2du^3\!=\!F(u^1)F(u^2)F(u^3)du^1du^2du^3.\ \ \ \ \
\label{4prbl2}
\end{eqnarray}
But, on account of isotropy, $p(u^1, u^2, u^3)$ is invariant under a three dimensional rotation and
must be a function of $(u^1)^2+(u^2)^2+(u^3)^2=(u)^2$. This gives
\begin{eqnarray}
F(u^1)F(u^2)F(u^3)=\psi((u)^2).
\label{dife}
\end{eqnarray}
Differentiating Eq.{\ref{dife}} partially, w.r.t $u^1$,
$u^2$, and $u^3$, and dividing by \!$F(u^1)F(u^2)F(u^3)$, one arrives at the equation
\begin{eqnarray}
\frac{F^{\prime}(u^1)}{u^1F(u^1)}\!=\!\frac{F^{\prime}(u^2)}{u^2F(u^2)}\!=\!\frac{F^{\prime}(u^3)}{u^3F(u^3)}=\frac{\psi^{\prime}(u)}{u\psi}=-2 \lambda
\label{feqn}
\end{eqnarray}
where $\lambda$ is a constant which is positive because of the convergence requirement of $F$.
The Eq.($\ref{feqn}$) is solved to give:\
$F(u^i)=Ae^{-\lambda {(u^i)}^2} $\ and
therefore,
\begin{eqnarray}
F(u^1)F(u^2)F(u^3)={A}^3e^{-\lambda (u)^2}.
\end{eqnarray}
Integrating this probability over spherical shell of radius $u$ and $u+du$, one finds the total probability of finding
the particle between the speed range $u$ and $u+du$, as
\begin{eqnarray}
F(u)du=4\pi A^3 u^2 e^{-\lambda u^2}du.
\label{mxldstrbn}
\end{eqnarray}
The normalization condition: $\int_{0}^{\infty} F(u)du=1$ relates ${ A}$ with $\lambda$ as
${ A}=\sqrt{\frac{\lambda}{\pi}}$. The constant $\lambda$ is related to the average energy of the gas particle
as
\begin{eqnarray}
\hskip -0.3cm
<\!u^0\!>=\!\!\int_{0}^{\infty}\!\!\!\!u^0F(u)du
&=&4\pi{A}^3\!\int_{0}^{\infty}\!\!\!\!u^2\sqrt{(1+\frac{u^2}{c^2})}\ e^{-\lambda u^2}du.
\nonumber
\end{eqnarray}
As in the case of MB, the parameter $\lambda$ in this case, is also related to the average kinetic energy ($<\!E_k\!>$) of the gas molecule, however not in a simple manner. If $m$ is the rest mass of the gas molecule then
\begin{eqnarray}
{<\!E_k(\lambda)\!>=mc^2[<\!u^0\!>-1]\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\nonumber}\\
=\Bigl[4\pi mc^2{ A}^3\int_{0}^{\infty}u^2\sqrt{(1+\frac{u^2}{c^2})}\ e^{-\lambda u^2}du\Bigr]-mc^2.\ \ \ \ \ \ \
\end{eqnarray}
The above equation gives the physical meaning of the constant $\lambda$ as it relates this constant to the physically measurable
quantity $<\!E_k\!>$, i.e. \!$\lambda$ is a measure of the temperature $T$ of an ideal gas.
$<\!E_k\!>$, and $T$ are monotonically decreasing function of $\lambda$. Therefore, the non-relativistic limit
is achieved in large $\lambda$ limit.
The function $F(u)$ has non-negligible value only in the region where\ $\lambda u^2 \sim 1$\ or, $u^2\sim \frac{1}{\lambda}$.
For large $\lambda$,\ this implies $u^2={v^2}/({1-\frac{v^2}{c^2}}) \ll 1$,\ i.e.\ $\frac{v^2}{c^2} \ll 1$. This gives
$u^i \approx v^i$, therefore, $F(u)du=4\pi A^3 v^2 e^{-\lambda v^2}dv$, is the velocity distribution function in the
non-relativistic case.
\section{ The relativistic EOS }
In the general case of a relativistic gas, one can relate $p$, $\rho$, $\rho_0$ and
the adiabatic index $\gamma\ (=\frac{c_p}{c_v})$ as
\begin{eqnarray}
p=(\gamma -1)(\rho -\rho_0).
\label{gamaeqn}
\end{eqnarray}
It is easy to show that for a cool non-relativistic gas $\gamma$ becomes $\frac{5}{3}$, and for a extreme relativistic gas of
photon $\gamma=\frac{2}{3}$ [\onlinecite{Weinberg1972}].
Therefore, the value of $\gamma$ in general depends on the relativistic nature of the gas,
{\it i.e.} the dominance of the kinetic energy density over the rest mass density $\rho_0$.
To calculate the dependence of $\gamma$ on $\rho$, we first expressed $p$, and $\rho$, in terms of $\rho_0$.
This can be done by considering that the pressure $p$ is the momentum flux through unit area averaged over all the molecules of
the fluid-blob in the co-moving frame. We take a surface element of the fluid blob oriented to any one axis ($i^{th}$ axis say,)
the pressure $p$ is\ \
$p=n<\!p^iv^i\!>$.
Here $n$ is the number density of the molecule, $p^i$ is the $i^{th}$ component of the 4-momentum $p^{\mu}=mu^{\mu}$.
Using the relation $v^i=\frac{u^i}{u^0}$, and taking $<\!\!\frac{(u^1)^2}{u^0}\!\!>=<\!\!\frac{(u^2)^2}{u^0}\!\!>=<\!\!\frac{(u^3)^2}{u^0}\!\!>=<\!\!\frac{1}{3}\frac{(u)^2}{u^0}\!\!>$\ due to the isotropy, we write the pressure as
\begin{eqnarray}
p=\frac{1}{3}\rho_0 (<\!u^0\!>- <\!\frac{1}{u^0}\!> )=\rho_0 g(\lambda).
\label{pe}
\end{eqnarray}
Here and henceforth we work with $c=1$.
The total energy density $\rho$ is related to rest mass density $\rho_0$ as
\begin{eqnarray}
\rho=nm<\!u^{0}(\!\lambda\!)\!>=\rho_0 <\!u^0(\lambda)\!>=\rho_0 f(\lambda),
\label{rho}
\end{eqnarray}
where $f(\lambda)$ and $g(\lambda)$ are related to the modified Bessel functions of second kind $K_n(\lambda)$ through the
{\vskip -0.5cm
\begin{eqnarray}
{<\!\!u^{0}\!(\!\lambda\!)\!\!>\!=\!4\pi A^3\!\!\!\int_0^{\infty}\!\!\!\!\!\!\sqrt{1+u^2} u^2\!e^{-\lambda u^2}\!du\!=\!\frac{\pi A^3e^{\frac{\lambda}{2}}}{\lambda}K_1(\!\frac{\!\lambda\!}{2}\!),\ \ \ \ \ }
\label{au0}
\end{eqnarray}
}{\vskip -0.6cm
\begin{eqnarray}
{<\!\!\frac{1}{u^0(\!\lambda\!)}\!\!>=\!4\pi\!A^3\!\!\!\int_0^{\infty}\!\!\!\!\frac{\ e^{-\lambda u^2}\!u^2\!du}{\sqrt{1+u^2}}\!=\!\!\pi A^3\!e^{\frac{\lambda}{2}}\![K_1(\!\frac{\!\lambda\!}{2}\!)\!-\!K_0(\!\frac{\!\lambda\!}{2}\!)\!].\ \ \ \ \ }
\label{Ou0}
\end{eqnarray}
}
The properties of $K_n(\lambda)$ functions and its derivatives (require later) are well known in mathematics ([\onlinecite{Abramowitz1972}]). The value of $4\pi A^3$ is given before using the normalization condition.
\subsection{The Specific Heat Ratio and the Sound Speed}
Using the EOS (\ref{pe}, \ref{rho}), one could to find the variation of $\gamma$ w.r.t $\lambda$ for a
relativistic gas in many astrophysical applications.
Therefore, our next job is to express the thermodynamic variables in terms of $\lambda$ alone. Using thermodynamic
definitions, we obtain the specific internal energy
$\varepsilon=\frac{\rho-\rho_0}{\rho_0}=f-1$,\ and the specific enthalpy $h=\frac{p+\rho}{\rho_0}=f+g$ of the gas.
Variation of $\varepsilon,\ h$ are plotted in Fig.$\ref{IEeps}$ as a function of $\lambda$.
Further, simultaneous solutions of three equations: $p=(\gamma-1)(\rho-\rho_0)$,\ $\rho=\rho_0 f(\lambda)$,\ and $p=\rho_0 g(\lambda)$,
provide the specific heat ratio
\begin{eqnarray}
\gamma=(f-1)^{-1}(h-1),
\label{gma}
\end{eqnarray}
as function of $\lambda$ only.
To find the expression for the sound speed, one needs to
start from the definition $a_s^2=(\frac{\partial p}{\partial \rho})_s$, together with the
entropy equation $Tds=0$ and $p=p(\rho,\rho_0)$. Computing the quantities
$(\frac{\partial \rho_0}{\partial \rho})_s$,
$(\frac{\partial p}{\partial \lambda})_{\rho_0}$,\ $(\frac{\partial p}{\partial \rho_0})_{\lambda}$,
and simplifying one finally yields
\begin{eqnarray}
a_s^2\!=\!(\frac{\partial p}{\partial \rho})_{\rho_0}\!\!\left[1\!-\!(\frac{\partial \rho_0}{\partial \rho})_{\lambda}\!\right]\!+\!\frac{1}{h}\!(\frac{\partial p}{\partial \rho_0})_{\lambda}\!=\!(1\!+\!\frac{g^{\prime}}{f^{\prime}}\!)(\!\frac{g}{f\!+\!g}\!),\ \ \ \ \ \
\label{as2}
\end{eqnarray}
in terms of $\lambda$ completely. Here the {\it prime} (`$\prime$') denotes derivatives w.r.t $\lambda$.
To check the overall variation of $\gamma$ and $a_s$ w.r.t $\lambda$, we plot these functions in Fig.$\ref{aseps}$ and in Fig.$\ref{Gmaeps}$.
In Fig.$\ref{aseps}$, we see that there is no applicable variation of $\gamma$ throughout.
However there is a sharp decay of $\gamma$ values very close to zero.
This indicates that EOS is mostly non-relativistic in nature and becomes relativistic to ultra-relativistic at the end phase
of evolution. On the other hand, the deviation of the sound speed $a_s$ in Fig.$\ref{Gmaeps}$,
from non-relativistic value is prominent even in the relatively higher $\lambda$ values.
\begin{figure}
\includegraphics[height=6.0cm,width=8.0cm]{IE.eps}
\caption{\label{IEeps} The variation of specific internal energy ($\varepsilon$) and specific enthalpy ($h$)
are shown as a function of $\lambda$.}
\end{figure}
\section{Extreme limits: Ultra-relativistic and non-relativistic}
It is important to show that this expression for $p$ gives correct results for non-relativistic
and ultra-relativistic limit. In the {\it non-relativistic} limit $v<<c$.
Therefore, $u^0\approx (1+\frac{1}{2}\frac{v^2}{c^2})$ and $\frac{1}{u^0}\approx (1-\frac{1}{2}\frac{v^2}{c^2})$.
Therefore,\
$p=\frac{1}{3}\rho_0<\!v^2\!> $.
Now, the average kinetic energy density is\ \ $\rho_0<\!E_k\!>=\rho_0(<\!u^0\!>-1)=\frac{1}{2}\rho_0<\!v^2\!>$. Therefore,
$\rho -\rho_0=\frac{1}{2}\rho_0<\!v^2\!>$. Using above results, we find
\begin{eqnarray}
\rho -\rho_0=\frac{3}{2}p,
\end{eqnarray}
an exact EOS in the non-relativistic limit in which $\gamma\!=\!\frac{5}{3}$.\\
For an {\it ultra-relativistic} gas of photon
$v \rightarrow c$ and $m\rightarrow 0$. However, $c^2mu^0=E_{photon}$ remains finite. Therefore, for a photon gas\
$u^0 \rightarrow \infty$ and $\frac{1}{u^0}\rightarrow 0$. This gives\ $\rho_0=0$, $\rho=nE_{photon}$
and $p=\frac{1}{3}nE_{photon}$. Therefore, in the ultra-relativistic limit, EOS and $\gamma$ respectively becomes
\begin{eqnarray}
p=\frac{1}{3}\rho, \ \ \ \ \ \ \gamma=\frac{4}{3}.
\end{eqnarray}
{\vskip -0.1cm
For a gas of massive particle, the values of the $\gamma$ and $a_s$ in the extreme relativistic regime can be obtain
from Eq.($\ref{gma}$) and Eq.($\ref{as2}$) in the limit $\lambda \to 0$.}
The integral in Eq.($\ref{au0}$) \& ($\ref{Ou0}$) and their derivatives w.r.t $\lambda$,
take the value for $\lambda \to 0$ (by putting $\gamma u^2=y$)
\begin{eqnarray}
<\!\!u^0\!\!>=\!\frac{2\lambda^{-\frac{1}{2}}}{\sqrt{\pi}}, <\!\!u^0\!\!>^{\prime}=\!-\!\frac{\lambda^{-\frac{3}{2}}}{\sqrt{\pi}}, <\!\!\frac{1}{u^0}\!\!>=\!\frac{2\lambda^{\frac{1}{2}}}{\sqrt{\pi}}, <\!\!\frac{1}{u^0}\!\!>^{\prime}=\!\frac{\lambda^{-\frac{1}{2}}}{\sqrt{\pi}}.\nonumber
\end{eqnarray}
Substituting above values in the Eq.(\ref{gma}), \& (\ref{as2}), we find
\begin{eqnarray}
\lim_{\lambda \rightarrow 0}\gamma=\frac{4}{3} \ \ \ \mbox{and} \ \ \lim_{\lambda \rightarrow 0}a_s^2=\frac{1}{3}.
\end{eqnarray}
{\vskip -0.1cm
Similarly, it can be shown that in the large $\lambda$ limit, $\gamma=\frac{5}{3}$.}
We emphasis these points in Fig.$\ref{aseps}$, where y-axis are set at $y=\frac{4}{3}$ and $\frac{5}{3}$.
We see, $\frac{4}{3} \le \gamma \le \frac{5}{3}$ perfectly holds in our case. Also,
in the Fig.$\ref{Gmaeps}$, we find the sound speed $a_s \le \frac{1}{\sqrt{3}}$ satisfies.
\begin{figure}
\includegraphics[height=6.0cm,width=8.0cm]{as.eps}
\caption{\label{aseps} A plot of the sound speed $a_s$ vs. $\lambda$ is shown in figure.
The inequality, $a_s \le {1}/{\sqrt{3}}$ satisfies when $\lambda \to 0 $. }
\end{figure}
\begin{figure}
\includegraphics[height=6.0cm,width=8.0cm]{Gma.eps}
\caption{\label{Gmaeps} The adiabatic index $\gamma$ ($\frac{4}{3}\le\gamma\le \frac{5}{3}$) is plotted w.r.t $\lambda$.}
\end{figure}
\section{Conclusion}
In this communication we express the relativistic EOS in a new form through two parametric equations:
$\rho=\rho_0 f(\lambda)$,\ and $p=\rho_0 g(\lambda)$, where $\lambda$\ is a parameter related to kinetic energy of the gas.
The EOS is obtained by using a 4-velocity distribution function which we derived by applying the basic principles of probability
theory with the assumptions of isotropy.
In the ultra-relativistic regime,
these new equations perfectly produces well-known results, namely, the EOS:\ $\rho=3p$,\
the value of $\gamma=\frac{4}{3}$, and sound speed $\le \frac{1}{\sqrt{3}}$ have\\
\\
an exact match,whereas, in the non-relativistic regime, EOS correctly reduces to $\rho-\rho_0=\frac{3}{2}p$, which
implies $\gamma = \frac{5}{3}$.
The results we obtain are also in well agreement with the known results.
Our theoretical investigation indicates that EOS evolves abruptly at the end phase of ultra-relativistic regime.
Following the prescription of new EOS, it is easy to express the thermodynamic and hydrodynamic
variables in terms of $\lambda$ alone. In the next stage, one just need to estimate the variation of $\lambda$ w.r.t
the radial distance as to find an interesting solutions in different physical systems in interest.
We have studied one such solutions in the accretion around black hole which we wish communicated soon.\\
\begin{acknowledgments}
Author PB acknowledges SNBNCBS PDF support. Author SM acknowledges KASI PDF support and thanks to RKMRC to grant study leave under UGC-scheme.
\end{acknowledgments}
|
\section{Introduction}
Stellar variability studies provide critical access to a number of
astronomically significant properties including rotation rates,
eclipsing binaries, and pulsations. Aside from offering insight into
the nature of the stars themselves, statistical studies conducted on
large samples of field and cluster stars broaden our understanding of
stellar evolution and can assist in the search for exoplanets.
Variability surveys that endeavor to find variations with intermediate
periods of days to weeks studying objects
such as short period eclipsing binaries and planetary transits,
benefit from dedicated long-term, high cadence, dedicated observations
\citep[][and references therein]{howell2008,vonbraun2009b}.
NASA's \textsl{Kepler } mission \citep{keplerreview}, which was launched in
March 2009, is conducting a transit search in Cygnus with the goal of
finding Earth-like planets orbiting in the habitable zones of sun-like stars. The
signatures of transits of this nature would have relatively small
depths, making the inherent variability of the host star even more
relevant. The success of this mission partially depends on an accurate
characterization of the stellar variability in the field. To that
end, the Burrell-Optical-Kepler-Survey (BOKS) was designed to
determine the level and type of stellar variability in a small ($\approx$ 1\%)
subsection of the \textsl{Kepler } field. As an added goal, we can assess the frequency
of close-in Jovian-type planets (the so-called ``Hot Jupiters'') in the same field,
and allow for a comparison of ground-based and the \textsl{Kepler } based transit surveys.
A number of other Hot Jupiters have been discovered in the \textsl{Kepler } field prior to launch
from ground-based surveys \citep{odonovan2006,pai2008,bakos2010} and \textsl{Kepler }
itself has already discovered many additional exoplanets
\citep{kepler4,kepler5,kepler6,kepler7,kepler8,keplerbig,steffen2010}. Further comparisons
of ground-based and space-based transit candidates would be extremely beneficial due to the
high quality lightcurves that \textsl{Kepler } can provide. In particular, since the \textsl{Kepler }
mission must make numerous selection cuts in order to achieve their mission objectives
\citep{batalha2010}, there is a tendency to avoid fainter target stars that may have detectable
Hot Jupiters, but not be suitable for Earth-sized transit searches. By identifying additional
Hot Jupiter candidates, and then having \textsl{Kepler } undertake follow-up observations, we may be able to
study detailed properties of these systems and characterize them in exquisite detail. This
has already been done for one pre-launch Hot Jupiter candidate, HAT-P-7, and there are indications
that the extrasolar planet in this system is gravitationally distorting the host star \citep{welsh2010}.
This paper introduces the properties of BOKS, and gives a summary
of the data reduction and analysis of the survey. In \S\ref{sec:obs} we present
our observing strategy and a summary of our observations. We outline our data
reduction techniques and discuss the observational window function of our survey in \S\ref{sec:datared}.
We present the object detection, photometry and astrometry in \S\ref{sec:objphot}. Finally, in \S\ref{sec:analyze}
we discuss the initial results of our variability survey and of the search for exoplanets in the
BOKS field, specifically those of the ``Hot Jupiter'' variety. We review our conclusions
in \S\ref{sec:conclusions}. There are a number of other scientific
projects planned for the BOKS data, such as the comparison of stellar field to cluster variability,
cataloging the many variable stars found in this survey, and searching for any moving objects.
These results will be discussed in future papers. We plan to submit all of the BOKS
data to the NASA/IPAC/NExScI Star and Exoplanet Database\footnote{located at: http://nsted.ipac.caltech.edu/}
\citep{vonbraun2009a}, where it can be of service to the entire astronomical community
\section{ \label{sec:obs} Observations}
In order to maximize the scientific benefits from this survey, we chose our scientific
field under a number of constraints. First, besides being located within the \textsl{Kepler } field
our field ideally should have a large number of stars, as this will improve our chances
to find extrasolar planets and find other interesting objects. Second, given that our
ground-based imager has relatively large pixels in angular size, the field must not be so
close to the Galactic plane that photometric crowding would be a major factor. Third, in
order to compare our data against numerous stellar cluster variability surveys, such as
UStAPS\citep{hood2005}, EXPLORE/OC \citep{vonbraun2005}, PISCES \citep{moch2006}, and STEPSS
\citep{burke2006}, we decided on a field that had both field stars and an open cluster within it, so
that we could compare the variability properties of both stellar populations simultaneously.
In order to find the optimal field, we first pre-imaged a number of candidate fields on
2006 April 24. After visual inspection of all of the pre-images, we made a determination
of the field that best matched our conflicting criteria.
Our final selected target field of view covered 1.39 deg$^{2}$ in the constellation of
Cygnus and was centered on the open cluster NGC 6811 (RA$=19^{h}37^{m}17^{s}$,
Dec=$+46^{d}23^{m}18^{s}$; WEBDA\footnote{The WEBDA database, developed by J.-C. Mermilliod,
can be found at http://www.univie.ac.at/webda/}). Our field is completely contained
within the \textsl{Kepler } field, and the BOKS field is located on channels 63, 47, 23, and 39
of the \textsl{Kepler } imager in the Spring, Summer, Fall, and Winter seasons respectively
\footnote{Full Frame Images (FFIs) of the Kepler field for each observing season can be found at
http://archive.stsci.edu/kepler/ffi\_display.php}.
Our photometric survey began on 1 September 2006 and ended on 10 October 2006,
consisting of 40 nights in all. We observed with the Case Western Reserve University 0.61m Burrell
Schmidt telescope (hereafter the Burrell), located at Kitt Peak National Observatory. One advantage of a
dedicated observatory for this survey is that we could observe for a large number of consecutive
nights. Many researchers
\citep{pont2006,beatty2008,vonbraun2009b} have shown that the total duration of a photometric survey
is crucial in maximizing transit detection efficiency in the presence of statistically correlated (``red'')
noise.
The imager used for this survey was a SITe back-illuminated 2k$\times$4k CCD
with 15 micron (1.45 arcsecond) pixels, run by a Leach version 2 controller \citep{leach1998}, and
two output amplifiers. The long axis of the CCD was oriented East/West. We observed primarily in
the SDSS $r$-band filter, but we also obtained occasional Johnson $V$-band images of the field of
view in order to obtain two-color information and to allow for
cross-comparison between our data and other photometric catalogs.
In particular, we compared our photometry to that found
from the Kepler Input Catalog (KIC)\footnote{Version 10 of the KIC is available at:
http://archive.stsci.edu/kepler/kic10/search.php In some cases, we used the 7th and 8th version of the KIC for
steps in our analysis. These will be referred to as KIC78 in the text.}
Observations were ongoing during any weather conditions where stars were
visible on the sky, and it was safe to operate the telescope. As a result, the BOKS data
have large variations in seeing, transparency, and night sky brightness. Of the 40 nights of
observing, thirteen were completely lost to weather, leaving 27 nights of potential
data. Nearly all of our $r$-band and $V$ integrations were 180 seconds in duration, with the
exception of 31 images taken on night 14 that had exposure times of 300 seconds. The
CCD readout time was 45 seconds in length. A total of 1,924 $r$ images and 10 $V$ images were taken
over the entire run. The SITe CCD gain was fixed at 2 electrons per ADU, each pixel had
a full well capacity of at least 100,000 electrons, and the read noise was 12 electrons.
We note here that while the CCD pixel scale is fairly large compared to typical CCD
imagers, studies (Feldmeier et~al. 2011, in preparation) have shown that
milli-magnitude relative photometry is possible, even with such large pixels. Table~\ref{table:obslog}
provides a summary log of our observations and Figure~\ref{fig:hist} shows a graphical representation of
the number of exposures throughout the survey. An $r$-band exposure of BOKS, created from
co-adding the first 24 images from our survey is shown in Figure~\ref{fig:field}.
With the help of the American Association of Variable Star Observers (AAVSO), we also arranged to have
bright variable stars photometrically monitored in the field at the same time as BOKS was underway. The
preliminary results from this independent photometric survey are discussed in \citet{henden2006}, and a more
careful comparison will be discussed in a future paper\footnote{The AAVSO NGC~6811 campaign information
can be found at\\http://www.aavso.org/news/ngc6811.shtml}.
\section{ \label{sec:datared} Data Reduction }
Although the data reduction of BOKS is relatively straightforward, the large
number of images and the need to ensure highly precise relative photometry demands
some careful attention. We therefore began our
CCD reductions as follows. Since our imaging observations were obtained using the dual
amplifier mode with the SITe CCD, we first combined the two amplifier readouts into single
images using IRAF's\footnote{IRAF is distributed by the National Optical Astronomy
Observatory, which is operated by the Association of Universities for Research
in Astronomy, Inc, under cooperative agreement with the National Science Foundation.}
\textsf{mscred.mkmsc} task. The resulting images were then merged, trimmed, and overscan
subtracted using the \textsf{ccdred.ccdproc} task.
We obtained approximately ten bias and five twilight flatfield frames per night, which
were used on the respective night's images using IRAF's
\textsf{ccdred} package after they were checked for unwanted features
such as bright stars in the flatfields or amplifier noise in the bias
frames. If any such features were present, the corresponding bias or
flatfield frames were discarded. We created nightly master bias frames, but
due to the presence of dust grains on the dewar window and/or filter, that changed
positions between nights, we did not create a master flat for the entire run.
After the data had been processed, we inspected the entire data set for quality
issues. We determined basic parameters of each image such as the median sky level
and the median seeing in order to verify that our data was suitable for use.
To determine the median sky level, we used the IRAF task \textsf{imstatistics} in iterative mode.
The median seeing of each image was calculated by applying the \textsf{imexam} task on
150 bright, unsaturated stars on each image. Each star was fit using a Gaussian function, and
the median of all of the derived FWHM values was adopted for the median seeing for each image.
After applying the photometric zero point (\S 4.3), we present the median sky brightness
and median seeing for BOKS in Figure~\ref{fig:skysee}. As would be expected from a
telescope run of this length, these properties varied substantially as differing lunar phases
and weather patterns occurred. After this analysis, we removed night 37's data due to very
high sky levels by the nearly full moon. Next, we visually inspected each image for
quality issues. From this
process, we found that night 12 suffered from condensation on the CCD dewar window, leaving a
circular distortion feature on approximately 17\% of the area of each image. Night 19 suffered from
CCD electronics issues on one of the two CCD amplifiers. In both of these cases, many of the stars
remain unaffected on each image. However, to be conservative in this initial study, we have
removed the entire night's data from further consideration. After our initial light curve
analysis (discussed in \S\ref{sec:analyze}), we found that two additional nights, nights 13 and 23, had extremely
variable clouds (with such a large angular field, even the ensemble photometry method discussed
in \S\ref{sec:objphot} can fail if the clouds are variable enough; Feldmeier et~al. 2011, in preparation).
Although some of the exposures on these nights should be acceptable for our scientific goals, we again chose
to be conservative and removed the entire night's data from consideration. This left us with imaging
data from 22 different nights over the span of the survey and a total of 1,565 $r$-band images that could be
used for potential variability studies.
For a planetary transit survey, understanding the observational window function gives crucial insight
into the survey completeness and sensitivity to periodic variability. Normally, the window
function is defined as the probability that a planetary transit is detected in a given data set, as
a function of planetary orbital period \citep{vonbraun2009b}. For BOKS, we calculated the
approximate observable window function in the following manner: we simulated planetary transits
from periods ranging from 0.5 to 30 days
and divided each period into 10,000 phases. Given the starting and ending times of the 22 nights of
observations, we then determined the probability of observing a transit with that period over all phases.
The results of this analysis are plotted in Figure~\ref{fig:window}. As with all time-limited photometric surveys,
we are most sensitive to short period transits, and our ability to detect transits decreases with transit period.
Since our survey is ground-based, the characteristic aliasing period of integer days is also strongly present
in our data. We should note that this window function deals with temporal sampling only
and likely to be an overestimate: it does not take into account the effects of differing transparency,
seeing, and sky brightness on the detectability of transits. It also does not take into account the effects
of differing stellar and planetary radii on the detectability of transits. Finally, the effects of
statistically correlated ``red noise,'' which are likely to be significant, are not included in this
calculation. We plan to perform extensive Monte Carlo simulations on these effects, which will be presented
in a future paper. However, we note that BOKS has a significant advantage over many other ground-based
transit surveys: if even one transit appears in our survey, we could, in principle, verify it with \textsl{Kepler }
follow-up observations.
\section{\label{sec:objphot}Object Detection, Photometry, and Astrometry}
With our good quality data set finalized, we next focussed our attention on finding all stellar
sources in the BOKS field, and determining their magnitudes and positions throughout the survey period.
\subsection{Object Detection and Coordinate Transformations}
We next created a master image by combining six
individual images from night 25 using the \textsf{imcombine}
task. This master image was used to ensure that we detect all of the stellar sources in the frame and
remove the possibility
of radiation events contaminating our source catalog. We next created the master list of stars (point sources)
by running the \textsf{daofind} task within IRAF on the master image. We chose a \textsf{threshold} value
for point source detection of five times
the standard deviation of the sky background. Given the large pixel scale of our data, we adjusted
the \textsf{sharphi} value to 0.9 rather than the 1.2 that is normally adopted. All other data-independent
parameters were left at their default values. We found a total of 56,354 sources that matched the \textsf{daofind}
criteria we adopted. This initial list was then manually inspected to ensure that non-stellar objects were not
included, such as diffraction spikes, radiation events, or objects on the extreme edges of the frame. This left
a total of 54,687 objects for further study.
With the master list of coordinates determined, we then needed to re-identify each source on every frame
of our survey. Rather than shift each image, which would lead to unacceptable uncertainties in the magnitudes due
to interpolation, we instead re-determined the source coordinates for each individual image. To do this, we used a high
quality image from the middle of night 25 as our positional reference image. To match stars in individual images to
those found on the reference image, we split each image into eight rectangular subsections (512$\times$512 pixels).
For each subsection of each image, we summed a group of rows and group of columns to create a pair of one dimensional
arrays that contained the peak counts at the row or column location of each star. We then used the Fourier transform
between pairs of these arrays taken from different images to locate translational shifts between each subsection of
the images. These shifts were applied to the master coordinate list to locate each star on individual images.
The Burrell tracked fairly well overall but in periods of cloudy conditions, when the autoguider was
unable to hold the tracking steady, the spatial shifts could be up to 1--2 arcminutes. By dividing the image into
subsections, we were able to accommodate most of the magnification and rotational differences between images
in our final astrometric coordinate solution.
\subsection{Aperture Photometry and Ensemble Correction}
We performed aperture photometry using the \textsf{phot}
task within IRAF's \textsf{noao.digiphot.daophot} package \citep{stetson1987}. We experimented with
five seperate photometry apertures, from very small radii (1 pixel) to large radii (5 pixels), to
span the range of stellar brightness and seeing changes over the observed field of view. After inspection
of the output light curves, we selected two distinct aperture values for our
work (3 and 4 pixels) based on their small level of scatter around the median magnitude of a representative
light curve sample. These two aperture values, which correspond to 4.35 arcseconds and 5.8 arcseconds, with a
sky annulus of 11.6--33 arcseconds, gave good results and we used the smaller aperture for our final light curve set.
Once the raw photometry files were created, we then needed to account for the effects of seeing,
transparency, and airmassfor each star. To do this, we adopted a local ensemble
photometry technique, where we used a local set of bright stars that
are photometrically constant to determine changes in these parameters. The algorithms are discussed in detail
by \citet{everett2002} and \citet{everett2001}; we will briefly outline them here.
First, we divided each image up into 8 $\times$ 4 square regions with a size of 500 $\times$ 500 pixels (corresponding
to 725 arcseconds on the side). This size was chosen to allow a sufficient number of ensemble stars to be present in
the individual regions and at the same time, to optimize our sampling of positional dependence of photometric
effects such as variable point spread functions, color terms, or focus gradients. For a star to be an
acceptable ensemble star, it must fulfill certain quality criteria. Specifically: 1) the star must be present in
all frames, 2) it must have a photometrically constant light curve ($\chi^{2} < 3.$), 3) it must be bright enough that
photon noise is negligible (average flux must be greater than 50,000 ADU, corresponding to a S/N of 316 or better),
and 4) have no close-by stellar companions that would interfere with the light curve in poor seeing conditions (no stars
within 5 pixels that are within 5 magnitudes of the ensemble candidate's magnitude). We
created an initial list of ensemble stars by conducting an automated search through the stars in each region.
After this automated preselection, the ensemble stars were inspected by eye to eliminate stars that appeared to have
signs of residual variability compared to the remainder of the ensemble stars. A total of 688 stars with $r$ magnitudes
varying between 14.3 and 16.6 were used in the final ensembles.
After the final list of ensemble stars was created, the relative photometry procedure was rerun using
only the cleaned sample of stars. If a region had fewer than ten ensemble stars, we combined it with
a neighboring one to create a larger region.
The exact calculation of relative photometric offsets for the individual regions
was performed by a custom written routine based on \cite{everett2002}. An example of these offsets for a single
region is plotted in Figure~\ref{fig:offsets}. Due to the effects of airmass, seeing, and transparency, these
offset values vary significantly over the survey length.
Due to the differing positions of each individual image, various stars have differing numbers of observations, with
objects at the edges of the fields having fewer observations than object near the center. To ensure a high quality
set of light curves for study, we focus exclusively on light curves that have at least 1,000 photometric measurements.
This left 32,806 sources for further analysis.
An unfortunate issue we found during the data reduction was the fact that the start time (HJD) for
any given exposure listed in the image header was not sufficiently accurate for our purposes.
The internal clock on the data recording computer, which was a Microsoft Windows PC running the Voodoo
image acquisition software\footnote{located at http://www.astro-cam.com/} at the telescope was found to be
imprecise and could not be used alone for exact timing purposes as we found that it drifted by up to
several minutes over the course of a single night. As a result, we used the file creation times
(recorded by a different clock on an internet time controlled Linux machine) as HJD ``start time"
information in our image headers. From some experimentation, we found that the header start time recorded is
approximately 93 seconds after the true mid-exposure time in most cases. From comparing consecutive exposures
throughout the survey, the uncertainty in this correction is approximately one second.
Consequently, the times we recorded in our light curves will be precise relative
to each other, but our time zero point is not tied to UT or any other absolute time system to within several
minutes. In the future, we plan to correct this by correlating our observations against the AAVSO
and \textsl{Kepler } observations, which should allow us to reduce any time offsets.
\subsection{Photometric Zero Point}
The transformation from differential instrumental magnitudes to SDSS $r$ magnitudes for our stars was
performed by comparing the
magnitudes of the ensemble stars in each subregion to the corresponding $r$ stellar magnitudes listed
in the Kepler Input Catalog (KIC). The scatter in this comparison was approximately 0.1 magnitudes, which we
take as our systematic magnitude uncertainty for the BOKS survey. Given that we have minimal color
information in our observations and our survey is primarily interested in searching for stellar variability,
this amount of uncertainty is acceptable for our needs. For any individual object, however,
one can derive a more accurate magnitude by directly comparing the object to the KIC values, which have a
mean $r$ systematic uncertainty of less than 0.015 magnitudes\footnote{photometric uncertainties for the KIC
can be found at: http://www.cfa.harvard.edu/kepler/kic/kicindex.html}.
\subsection{\label{sec:ast}Astrometry}
We performed astrometry using the USNO-B1.0 catalog \citep{monet2003} and
\textsf{wcstools}\footnote{\url{http://tdc-www.harvard.edu/software/wcstools/}}. In order to eliminate the
effects of field rotation and distortion, we performed astrometry on the individual regions as described
above, rather than on the field as a whole. Our internal uncertainties on our coordinates were approximately
$\pm$ 0.5\arcsec right ascension and $\pm$ 0.3\arcsec declination. As an independent check on our astrometry,
we compared our final astrometric catalog to that found from the 2MASS Point Source Catalog \citep{2mass}.
Figure \ref{fig:match} shows the astrometric comparison between our BOKS astrometric solutions and those
listed for common stars in the 2MASS catalog. The one-sigma coordinate offset is near 0.4\arcsec both
RA and DEC with a small, but clear, asymmetry across the field. Given that the pixel size of the
Burrell is 1.45\arcsec and that the wide field of the Burrell produces differential focus and refraction
effects, this agreement is quite reasonable.
\section{ \label{sec:analyze} Light Curve Analysis }
With the light curves established for each star, we then began the search for photometric variability.
To characterize the possible variable nature of our light curve sample and to search for possible
exoplanet transits, we utilized OPTICSTAT, a custom-written statistical
analysis package created by M. Everett (discussed in Howell et~al. 2005). OPTICSTAT
returns several statistics related to stellar variability, including the reduced ${\chi}^2$
(${\chi}^2_{\nu}$), the standard deviation, the probability that the star varies periodically, and the
most likely period. Prior to our statistical light curve analysis, we removed the
effects of obvious cosmic rays which will artificially increase the apparent flux as follows.
If there were one or two
consecutive points in the light curve that deviated from the mean
magnitude by more than 3.5 times the standard deviation, those
points were rejected during statistical tests. This might remove some signs of
ultra-short variability, but since the exposures are 180 seconds in length, and the
readout time of the CCD is 45 seconds, the total scientific impact should be minimal.
To test each light curve for general variability, we fitted the
light curve with its mean flux and then calculated the probability that the
reduced $\chi_{\nu}^{2}$ statistic shows the data to be inconsistent with
this mean flux. This test can easily fail, however, in the presence
of systematic errors or uncertainties in the calculated magnitude
errors we assign to the individual data points. To compensate, we adopted a more
conservative threshold for the $\chi_{\nu}^{2}$ probability than our
formal errors would dictate. When applied to the full light curve,
we have adopted the criteria that point sources that
have $\chi_{\nu}^{2}\geq5$ are variable sources.
\subsection{ \label{sec:errors} Photometric Uncertainties}
In order to make a proper determination of variability for each star, we must determine
the random and systematic uncertainties for our target stars. By splitting the entire
BOKS field of view into 32 smaller sub-sections we
were able to remove the vast majority of systematic variations in our light
curves. However, some dispersion remains due to photometric uncertainties,
small differences in color and properties between the ensemble stars and the
target objects, and the statistically correlated (``red'' noise) that is
present in all ground-based transit surveys \citep{pont2006,beatty2008,vonbraun2009b}.
To estimate our remaining photometric uncertainties for every light curve, we
calculated a weighted average magnitude and the standard deviation of the entire light curve
about this average. Figure \ref{fig:sigma} shows the standard deviation of our BOKS light
curves as a function of their $r$ magnitude. As can be clearly seen, our brightest
stars have a 4 milli-magnitude dispersion about the average magnitude, which sets our noise floor
for this survey. For stars brighter than $r\simeq14$, saturation becomes an issue with seeing
changes making the exact magnitude of saturation somewhat imprecise.
As the apparent magnitude increases, the photometric errors also increase. Plotted on Figure~\ref{fig:sigma}
is an estimate of the photometric errors as a function of target flux \citep{everett2001}:
\begin{equation}
\sigma_{*} = 1.0857 \frac{\sqrt{N_{*} \times g + n_{pix} [ 1 + (n_{pix}/n_{sky})](N_{sky} \times g + R^{2})}}{N_{*} \times g}
\end{equation}
where $N_{*}$ is the number of ADUs from the star in the aperture, $g$ is the gain of the CCD in electrons
(2 e$^{-}$ per ADU), $n_{pix}$ is the number of pixels in the aperture, $n_{sky}$ is the number of
pixels in the annulus around
the aperture used to measure the sky flux, $N_{sky}$ is the flux in ADUs per pixel from the sky, and $R$ is
the rms read noise of the CCD in electrons (12 e$^{-1}$). For this comparison, we assumed the faintest
value of the night sky brightness in our survey, which should give a lower bound to the true photometric
uncertainties. As can be seen, this function is in good agreement with the lower edge of our
error distribution. At the faint
end of our photometry, near $r=19$, our uncertainties are $\sim$50 milli-magnitudes per observation,
substantially larger than the 10-30 milli-magnitude precisions needed to find transiting
extrasolar planets.
For completeness, we have also calculated the variability level expected from atmospheric
scintillation alone (Young 1967; see also Young 1993a; Young 1993b; Badiali et al. 1994
for some important comments) in order to rule it out as a significant contributor of our highest
precision photometric results. We find that the scintillation
at unit airmass is approximately 0.3 milli-magnitudes, about
a factor of 15 lower than the photon noise from the brightest
stars in our sample. This is an approximate value: Young (1993b) notes
that the value can vary by up to a factor of two on a timescale of a
few minutes. Nevertheless, given our measured photometric uncertainties, scintillation is not a major
contributor to our error.
\subsection{Point Source Variability Statistics}
\label{sec:vari}
An important goal of BOKS is to determine the variability properties of stellar sources in general
within one portion of the \textsl{Kepler } field.
As mentioned previously, DAOFIND identified 56,354 point sources in at least one BOKS survey image.
Of these, 32,806 point sources were observed at least 1,000 times within our survey. From this subset,
25,861 point sources are located within 2.5 arcseconds of a source in the KIC78, and therefore we have
additional photometric information. However, since the KIC78 data is non-contemporary with BOKS, any
values derived from this comparison should be treated carefully. For statistical purposes, we
excluded any source from the KIC78 that did not have a valid $r$ magnitude, even if it had measured
magnitudes in other bands. This left 22,340 sources for further analysis.
We then determined a magnitude cut-off value for our statistical analysis of variability.
Progressively fainter stars have larger photometric errors, are more likely to have poor quality light curves
due to contamination from nearby bright stars, and in conditions of poor seeing, and may have incorrect recentering
by the aperture photometry algorithm, which can cause the aperture to recenter on a nearby brighter star.
For these reasons, we restricted our variability statistics to the 13,786 stars brighter than $r=18.5$,
which also lie within 2.5\arcsec of a source within the KIC78 and which have over 1000 data points. We
hereafter refer to this subsample as the BOKS-KIC sample.
After applying the statistical tests from OPTICSTAT, we found 2,457 stars
with $\chi_{\nu}^{2}\geq5$ and $r<18.5$ in the BOKS-KIC sample. We note that this number is approximate,
as the reduced $\chi^{2}$ is strongly affected by a number of systematic uncertainties, such as residual
variability in our ensemble stars, color mismatches between the ensemble and the target stars, spatial
structure in the extinction correction, and other effects. Additionally, photometric uncertainties
and stellar variability applied to the KIC78 will move stars above and below the magnitude cutoff, creating
an additional systematic uncertainty.
The variability fraction found ($\approx$ 18\%) can be compared to other variability surveys
using identical search techniques. In an earlier study, \citet{everett2002} found a variability fraction
of 17\% over a five day survey period using a similar telescope and sampling rate, but using a less
strict variability criteria ($\chi_{\nu}^{2}\geq3$). In contrast, the variability study of NGC~2301
\citep{tonry2005,howell2005,sukhbold2009} found a much larger variability fraction (56\%) using an orthogonal
transfer CCD observing mode over a 12 night run, and with substantially better photometric precision
($\approx$ 1.6~mmag).
\citet{tonry2005} and \citet{howell2008} discussed a relation between the percentage of variable
sources that will be found in any given photometric survey and that
survey's photometric uncertainty. From the NGC~2301 results, the cumulative fraction of variable
stars found, as a function of quartile variability, $x$, (the quartile variability is $\approx$ 1.5 times
smaller than the standard deviation) is:\footnote{Again, we note that in this paper we use $\chi^{2} > 5$ while the
previous studies used $\chi^{2} >3$, which
would allow more low amplitude variables into the samples.}
\begin{equation}
f(<x) = 1 - \frac{1.6~mmag}{x}
\end{equation}
For a best precision of 4 mmag, as we have in the BOKS survey, the percentage of variable sources would
be expected to be $\sim$40\%, substantially larger than what was actually detected. Some of this difference
may be due to differing stellar populations: NGC~2301 is a young open cluster
\citep[250 Myr;][]{kim2001}, in contrast to NGC~6811 \citep[575 Myr;][]{luo2009}, and the level of stellar
activity on the main sequence may be significantly different between the two clusters. The amount of
background and foreground contamination may vary significantly between the two surveys as well. Given
the position of the BOKS field, it is likely that many of the stars are field stars and therefore have
lower amounts of variability.
Of the 2,457 variable sources, 776 (32\% of all the variables or $\approx$ 6\% of the total BOKS-KIC sample)
were found to be periodic variable candidates (see \S\ref{sec:pervar}). This is significantly larger than
the results of \citet{everett2002}, who found a 14\% periodic fraction in their survey and \citet{howell2005}, who found
a 13\% periodic variable fraction in NGC~2301. We believe this is due to the substantially longer
time coverage of BOKS, which should make it substantially easier for our search algorithm \S\ref{sec:pervar}
to detect periodicity. The remaining variable stars appear to be non-periodic within the limits of
our time sampling, photometric precision, and observational window. These non-periodic variable
stars are presented in Table~2, with the corresponding photometric information from the KIC given. We
caution that the standard deviations given for these light curves are likely to be an overestimate:
any photometric residuals in the ensemble stars and the effect of radiation events can increase this
value significantly.
There have been three other recent variability surveys of the regions of the BOKS field, all centered
on the NGC~6811 open cluster \citep{vanc2005,rose2007,luo2009}. Unfortunately, the majority of the stars
found to be variable by these surveys are saturated on the BOKS images. The three exceptions,
stars V8 and V9 found by \citet{vanc2005}, and star V17 found by \citet{luo2009}, were also
detected by BOKS, and all three were found to be periodic variable stars.
\subsubsection{Non-variable Stars}
Our variability analysis found 10,881 stars brighter than
$r=18.0$ with no detected variability (${\chi}^2_{\nu}<$5).
Figure~\ref{fig:const_grr_cmd} shows a color-magnitude diagram
(CMD) of these stars. This CMD is typical of stellar fields
within the galactic disk, showing a plume around a color
of $g-r = 0.4-0.6$, corresponding to the main sequence turn-off
for an old ($>$ 10 Gyr) stellar population. The BOKS survey combined with
the KIC does not go photometrically deep enough to detect the very faint and red low-mass
stars in the disk population, which typically appear at
$g-r \approx 1.4$ and begin at $r \approx 18.2$. Similar CMDs are found in
the work of \citet[][see their Figures 2 \& 3]{dejong2010} with data taken
from the Sloan Extension for Galactic Understanding and Exploration (SEGUE).
Figure~\ref{fig:const_n_vs_r} shows a histogram of the number of
photometrically constant stars as a function of $r$ magnitude. The number of photometrically
constant stars increases roughly linearly up to the final bin ($r=17.9$). This
reflects both the increase in the number of faint stars and our
decreased sensitivity to variability for fainter stars.
Some example light curves of stars that show little signs of variability over the
timescale of our survey are given in Figure~\ref{fig:constar}. Note that in such long
surveys, it is highly likely for a star to suffer at least one, and possibly
more, hits from radiation events, such as cosmic rays.
\subsubsection{Periodic Variable Stars}
\label{sec:pervar}
In order to test each variable star light curve for periodicity we applied the Lomb-Scargle
method \citep{lomb1976,scargle1982}, as described by \citet{press1992}.
The algorithm produces a periodogram giving probabilities for sinusoidal signals
in the light curves over a range of periods from our minimum sampled period of
$1/60$ days ($0.4$ hours), set by the spacing of consecutive exposures, up to 40 days,
the maximum possible period spanned by the entire data set. We identified stars
as possible periodic variables if: 1) OPTICSTAT returned a false probability of
periodic variability of less than $1\times 10^{-4}$, 2) the amplitude was less
than 2.5 magnitudes, and 3) the average magnitude was brighter than $r$=18.5.
This yielded a large but manageable list of candidates. We then visually inspected the light
curves of these candidates after phasing them to the best-fit period.
Only those light curves that had clear periodic signals, and whose light curves
showed no sign of systematic effects were accepted as potential periodic variable stars.
In order to reduce the effects of aliasing, we also removed from our sample any sources that
had periods of 1 $\pm$ 0.025 days, though some of these objects may be genuine variable stars.
We found that 776 stars from our variable sub-sample had periodic signals that
ranged from $\sim$0.2 day to $\sim$33 days. The coordinates, mean $r$ magnitudes,
determined period and amplitude and the corresponding KIC information are presented in Table~3.
Additionally, we have classified the periodic variables into approximate types, which are
also presented in Table~3. Of these objects, 78 ($\approx$ 10\%)
show variability amplitudes larger than
$\sim$0.1 magnitudes. Another 235 objects ($\approx$ 30\%) have periodic amplitudes of 1-3\%
and periods of 1-3 weeks, which is consistent with rotational modulation due
to star spots. A significant number ($N=93$; 12\%) of the periodic variables
remaining have periods less than 2~days and photometric amplitudes less
than 0$^m$.05. These short period low amplitude variables are likely to be
pulsational variables such as $\delta$~Scuti stars \citep{breger2000}.
We next compared the properties of the periodic variable stars as a function
of period, stellar color, and magnitude, as these distributions give insight
on stellar properties in general. It is well known that there are systematic
changes in the fraction and amount of stellar variability across the H-R
diagram \citep[][and references therein]{eyer2008,ciardi2011}, but the precise distributions
are still under debate. Second, studies of these distributions allow
us to compare our results with other high precision variability surveys and provide
confidence that the survey is valid. Figure~\ref{fig:logp} shows the overall distribution of
periods in our sample of periodic variables. The most notable features of this plot are the
large number of stars with periods longer than 10 days and the peak
in the period distribution near periods of $\sim$1~day. Despite our attempts to
remove them, this peak strongly suggests that a fraction of these variable stars have derived
periods that are reflective of aliasing due to our observational window function,
rather than their actual period. Additionally,
the stars found to be periodic with the longest periods only have one or two measured
periods, and the period-finding algorithm may have mistakenly flagged these objects, when in
fact, they may not be periodic. More observations of these particular stars will be required to
fully address this issue.
In contrast to the number of photometrically constant stars, a histogram of the number
of periodic variables vs $r$ magnitude, which is plotted in Figure~\ref{fig:var_n_vs_r},
shows a steep rise in the number of periodic variables between $r\sim14$ and
$r\sim16$, which is followed by a steep decline in the number of periodic stars
fainter than $r\sim17$. This decline is due to the rapid loss of photometric
sensitivity to low amplitude variations for the fainter stars.
Figure~\ref{fig:var_n_vs_gr} shows a histogram of the number of
periodic variables versus $g-r$ color. The overall shape of this
distribution is very similar to the color-magnitude distribution for constant stars
(see Figure~\ref{fig:const_grr_cmd}). There may be a small excess of
the bluest stars ($g-r<0.3$) and the reddest stars ($g-r>1.0$), but it is unclear
whether this is statistically significant.
In contrast, Figure~\ref{fig:var_p_vs_gr} plots the period of periodic stars versus
their $g-r$ color. Most variables have colors that are similar to
those of constant stars regardless of period. There seems to be a
small excess of stars with $0\leq g-r \leq0.5$ among the periodic
variables with the shortest periods (P$\lesssim$1 day).
Figure~\ref{fig:amplitude} presents the distribution of amplitude for the periodic
variables within the BOKS survey. As has been previously seen \citep{howell2005},
the number of variable stars increases as the amplitude of variability decreases.
The dashed line indicates the power law model of variability found by \citet{tonry2005},
which has a slope of -2. The fit is in good agreement, giving further evidence to
applicability of this model.
Although a full accounting of the BOKS periodic variable stars would be soporific, we
briefly discuss some of the more interesting stars here, and we discuss two extreme
examples in \S~\ref{sec:othervar}. Among the periodic stars with derived amplitudes larger
than 10\%, we find at least ten clear pulsational variables. An example light curve
of this type of variable star is plotted in Figure~\ref{fig:pulsvar}. We have also detected two
probable RR~Lyrae stars within the field, with approximate periods of 0.53 and 0.56 days,
respectively. A light curve of one of these objects is plotted in Figure \ref{fig:rrlyrae}.
There are at least 32 eclipsing/contact binaries within the BOKS survey field,
with periods varying from 0.13 to 6.10 days. Figures \ref{fig:eb1} and \ref{fig:eb2}
give examples of two of these systems. From visual inspection of the light curves,
the majority of these systems are of the W Ursae Majoris subtype, as would be expected
\citep{hoffman2009}. However, we have also found at least eight Algol-type binary systems.
\subsection{Exoplanet Transit Candidates}
\label{sec:occult}
A primary goal of BOKS is to search for any signs of transiting extrasolar planets in
the data set. To search the light curves for transits, we used a simple test to find
and flag light curves containing at least one occurrence of a
diminution in the relative flux with parameters specified below. The
algorithm searched a light curve for time intervals when the mean
magnitude is statistically significantly fainter than in the preceding
and following data points, in other words, an inverse ``top-hat'' light
curve. The algorithm stepped
through time in each light curve testing a grid of possible interval
starting and ending times and durations and reported back the most
significant possible transit found for each
light curve. A probability of significance is determined for all
light curves for each possible transit by calculating a Student's
t-test \citep{press1992} statistic comparing the mean magnitude during
transit to the mean magnitude of combined pre- and post-transit data
points. Those light curves with the most significant probabilities,
specifically those that have a formal false probability less than
$1 \times 10^{-16}$, were then subjected to further inspection. In order
to avoid detecting too many spurious light curve fluctuations as transits,
we next placed further restrictions on the events reported by the transit-finding
algorithm. First, we searched for transit durations only between one and 12 hours.
Second, we placed limits on the depth of transit that merited further study.
Large planets orbiting all but the smallest M-dwarf stars will result in transit
depths of 0.5 magnitudes or less. We therefore removed all possible transits
that had depths larger than 0.5 magnitudes.
For the purpose of completeness, we decided to use all of the stars in this
analysis. It is extremely unlikely that faint stars will show a genuine
transit event, but including these objects in the search allows us to test for
other systematic effects in the algorithms. The entire BOKS sample
contains 54,687 light curves that were all analyzed using OPTICSTAT.
Using our detection limit described above, OPTICSTAT identified approximately
1,445 light curves with evidence of transit events.
Each of these 1,445 light curves were then inspected by eye with careful attention
given to additional criteria. We required at least two data points during
transit and at least two data points in both the pre- and post-transit light
curve, and we also required that the pre- and post-transit data points are
separated by no more than 24 hours from the time of ingress or egress.
The transit search algorithm used by OPTICSTAT searched for an
inverted ``top hat" shape in the light curve. However, there are a
number of types of variability that can lead to a ``top hat" shape
that must be eliminated through a visual inspection of the light
curve. Extrasolar planets will cause transits that have a
flat-bottomed appearance, so any sharp-bottomed transits were rejected from further
consideration. It is not possible with our data set to observe
secondary transits resulting from the planet passing behind its star,
so any light curve showing secondary transits was also removed from further study.
Finally, in order to do follow-up observations we required that the light curve
have at least two transit events. We require this to confirm the
first transit-event and to obtain an accurate determination of the
planet period before any follow-up observations are planned.
The light curve shown in Figure \ref{fig:eb2} was incorrectly identified by OPTICSTAT
as a transit candidate, but was easily removed by our manual inspection criteria: the transits in this light curve
are too deep ($\approx$ 0.6 magnitudes), sharp-bottomed, and there is an obvious secondary
transit. Nearly all of the transit candidates detected by OPTICSTAT were rejected using the
simple requirements we have outlined.
At the end of our analysis we were left with three exoplanet candidates:
BOKS-45069, BOKS-40959, and BOKS-52481. Some basic properties of these candidates
are given in Table~\ref{table:bokscand}, and a plot of their lightcurves is given
in Figure~\ref{fig:candidates}. For each of the candidates, we then determined an approximate
period by phasing the light curve to the best-fitting phase. In the case of BOKS-52481,
we detected one full transit and only a portion of another transit, so the measured period
is substantially less certain than the other two candidates. We also determined an approximate
transit depth by averaging the closest 40 light curve points immediately before and after the
transit to obtain a baseline. We present an example of each transit in Figure~\ref{fig:candphase}.
The properties derived are similar to other ground-based transit detections, but a detailed analysis,
including follow-up spectroscopic and imaging observations of these candidates is presented in
Howell et~al. (2010).
\subsection{Two Specific Cases of Stellar Variability}
\label{sec:othervar}
During any large area photometric survey, there is the potential of
discovering unusual to rare objects. In the case of BOKS, we detail here two
unusual variable stars that we have found in the survey.
\subsubsection{BOKS-45906}
The first object, whose location is displayed in Figure~\ref{fig:dwarfnfc},
is known as BOKS-45906 (KIC 9778689). On the first two clear nights of
our survey (MJD 3980 \& 3988), the star had a mean magnitude of $r \approx 20$,
though there were clear signs of variability of up to a magnitude in amplitude.
The star continued to vary on both nightly and intra-nightly timescales. Then,
between MJD 4004 and 4005, the star had an eruption, reaching a maximum of
$r=16.6 \pm 0.01$ on MJD 4006. It then declined in flux, returning to the approximate quiescent
flux level on MJD 4009. The overall light curve is plotted in Figure~\ref{fig:dwarfn}.
From the light curve, it is likely that this star is a newly discovered
cataclysmic variable of the dwarf nova subtype. There is scientific interest in the light
curves of similar objects, both before and after eruption
\citep{rob1975,collazzi2009,schaefer2010}: therefore additional photometric monitoring
of this source may be helpful.
\subsubsection{BOKS-53856}
The second interesting variable we have found is a blue star in the
BOKS survey (a finding chart is displayed in Figure \ref{fig:bluefc}), known as
BOKS-53856. From comparison to the KIC78, it has a measured color of $g-r=-0.46$, making
it the bluest stellar source in our field. Analysis of its light curve indicates
periodic variability, with a period of 0.255 days, though of an unusual nature. The
phased light curve is presented in Figure \ref{fig:bluelc}.
We obtained a 900 second spectrum of BOKS-53856 using the Kitt Peak 2.1-m telescope
and the GoldCam spectrograph on UT 26 June 2008. We used the 300 l/mm grating (\#32) with a
one arcsecond slit to provide a mean spectral resolution of 2.4\AA~across the full wavelength
range. The spectra were reduced in the normal manner with observations
of calibration lamps and spectrophotometric standard stars obtained
before and after each sequence and bias and flat frames collected in
the afternoon. The final reduced spectrum is displayed in Figure
\ref{fig:bluespec}.
The most obvious features in the spectrum are the blue continuum and strong
Balmer lines, indicating a DA white dwarf type spectrum. In general,
blue variables are either of low amplitude and consist of pulsations, or as
in the case here, show larger variations and may be some sort of interacting
binary. Holberg \& Howell (2011; in preparation) present a
further study of this star.
\section{\label{sec:conclusions} Conclusions }
The goal of the BOKS survey was to constrain the amount and nature of
variability in a subsection of NASA's \textsl{Kepler } mission field of view.
The dedicated observations we conducted of the BOKS field allowed us
to observe variability on various timescales from a few minutes to
many days. The long-term observations also allowed us a reasonable
opportunity to search for hot Jupiter exoplanet transits.
Through our preliminary analysis of the variability in the BOKS field
we have identified $\sim$2,457 candidate variable stars with 776
candidate periodic variables. Most of this periodic variation can be attributed
to rotating, low-mass stars with magnetic activity star/spots. We have also found
over 90 $\delta$~Scuti stars, over 32 eclipsing binaries and contact
binaries, as well as tens of large amplitude pulsators, such as RR
Lyrae stars. Within the BOKS field of view, we have also identified
at least three exoplanet candidates, all of which are undergoing
observations by the \textsl{Kepler } mission. The comparison of ground-based
and space-based transit observations should be beneficial to many future
surveys.
\acknowledgements
We would like to thank the entire staff of Case Western Reserve University
Warner \& Swassey observatory, including Heather L. Morrison, Charles Knox,
and Colin Wallace for their invaluable assistance with the Burrell Schmidt.
We are also extremely grateful to the observers of the AAVSO who performed many
observations of our field concurrently. We thank Richard Wade for some useful
discussions. We also thank the anonymous referee for several suggestions that
improved the quality of this paper. This research has made use of the WEBDA database,
operated at the Institute for Astronomy of the University of Vienna. This publication
makes use of data products from the Two Micron All Sky Survey, which is a joint project
of the University of Massachusetts and the Infrared Processing and Analysis
Center/California Institute of Technology, funded by the National Aeronautics
and Space Administration and the National Science Foundation.
{\it Facilities:} \facility{KPNO:2.1m (Goldcam),} \facility{CWRU:Schmidt}
\pagebreak
|
\section{Introduction}\label{I}
In this note we are concerned with several questions related to probabilistic and analytic potential theory of generalized Dirichlet forms.
A particular aim is to find definitive analytic conditions for non-sectorial Dirichlet forms that ensure the existence of an associated Hunt process.
The question whether the associated process is a Hunt process is crucial for localizing purposes (see e.g. introduction of Ref. \cite{Tr5}).\\
A fundamental consequence of Theorem 2 in Ref. \cite{Tr5} and Theorem 3.2(ii) in Ref. \cite{RoTr} is that
any {\it transient} Hunt process $\mathbb{M}$ on a metrizable and
separable state space is strictly properly associated in the resolvent sense with a strictly quasi-regular generalized Dirichlet form.
This is relevant because we can then apply all the fine results from the potential theory of generalized Dirichlet forms w.r.t.
the strict capacity (see Ref. \cite{Tr5} for some strict potential theory,
and Remark 3.3(iv) in Ref. \cite{RoTr} which applies also to strictly quasi-regular
generalized Dirichlet forms and Hunt processes).
Moreover, if the state space is only slightly less general, namely (for tightness reasons) a metrizable Lusin space, then by Theorem 2.1 in Ref. \cite{mrs}
the Hunt process can be approximated by multivariate Poisson processes and the
approximation works for all $P_x$, i.e. for all $x$ in the state space. The canonical approximation of the Hunt
process by Markov chains is useful as it provides an additional tool for its analysis and for the analysis of the underlying generalized Dirichlet form.
Note that the just mentioned line of arguments is not valid for sectorial Dirichlet forms, which underlines a strength of generalized Dirichlet form theory.
In fact, for a given arbitrary Hunt process we first do not know how to check whether it is associated to a sectorial Dirichlet form, and second this is clearly is not true in general. \\
Here, we establish the \lq\lq quasi converse\rq\rq\ of the above with nearly no restriction on the state space. We consider
two problems, which, due to the method, are in fact solved simultaneously.
The first problem is to establish the existence of an associated Hunt process to a strictly quasi-regular generalized Dirichlet form on a general state space,
and the second is the approximation of this Hunt process in a canonical way through Markov chains.
The second problem goes back to an original idea of S. Ethier and T. Kurtz.
In fact, it is shown in Chapter 4.2 in Ref. \cite{ek} that for nice state spaces such
as locally compact and separable state spaces and nice transition semigroups like Feller ones, the Yosida approximation via multivariate
Poisson processes converges for all starting points to a Markov process with the given semigroup. This was generalized in Ref. \cite{mrz}
where it is shown that the Yosida approximation of the generator together with some tightness arguments that result from the strict quasi-regularity
leads to the approximation via multivariate Poisson processes of any Hunt process that is associated with a strictly quasi-regular sectorial Dirichlet form.
This also led to a new proof for the existence of an associated Hunt process.
However, the price for the increased generality is that the approximation only works for strictly quasi-every starting point $x$ of the state space.
Since we use the same method we have to pay the same price, and even more we have to assume the additional structural condition {\bf D3} that is however trivially
satisfied for any sectorial Dirichlet form (see Proposition \ref{sd3}). Nonetheless, since the class of generalized Dirichlet forms is much larger than the
class of sectorial Dirichlet forms our results represent a considerable generalization. In particular time-dependent processes and
processes corresponding to far-reaching perturbations of symmetric (or even sectorial) forms, are covered.\\
Besides the canonical approximation scheme through Markov chains we want to emphazise that our main result
Theorem \ref{hunt2} is a definitive improvement of Theorem 3 in Ref. \cite{Tr5}.
Applying here a quite different method than in Ref. \cite{Tr5} we were able to relax the algebra structure
condition {\bf SD3} of Ref. \cite{Tr5} to the quite weaker linear structure condition {\bf D3}.
Therefore our general analytic conditions
for non-sectorial Dirichlet forms to ensure the existence of an associated Hunt process are just {\bf D3} and the strict quasi-regularity.
The state space is only assumed to be a Hausdorff topological space such that its Borel $\sigma$-algebra is
generated by the set of continuous functions on the state space. Our result is hence the counterpart of IV. Theorem 2.2 in
Ref. \cite{St1} for Hunt processes.\\
Finally let us briefly summarize the main contents of this paper. Section \ref{2} contains some preliminaries and the fundamental results.
In particular, our way of defining the strict capacity (cf. Definition \ref{cap}) is more explicit than in V.2 of Ref. \cite{mr} and Section 2 of Ref. \cite{mrz},
but still equivalent (see Remark \ref{all}).
The strict capacity is defined w.r.t. some reference function $\varphi$, but it turns out to be like the $\varphi$-capacity independent of that function (see Remark \ref{all}(ii)).
Proposition \ref{a} provides a useful new estimate for the strict capacity.
A crucial result is the construction of the modified functions $\widehat{e}_n$ in Lemma \ref{gf} in comparison to the functions $e_n$ of Lemma 3.5 in Ref. \cite{mrz}.
This makes the difference and allows to handle the non-sectorial case (see also Remark \ref{enr} for some related explanations).
Lemma \ref{gf} allows to get the important tightness result of Lemma \ref{liq2}. Note that
we also correct an inaccuracy that appears in the proofs of the statements corresponding to Lemma \ref{liq2} in both
papers Ref. \cite{mrz} and Ref. \cite{mrs} (see Remark \ref{mis}) and that we partially improve results from Ref. \cite{mrz} (see e.g. paragraph in front of Proposition \ref{sp}).
Having developed the fundamental results of potential theory in Section 2, most of the results of Sections \ref{3} and \ref{4}
follow by \lq\lq routine\rq\rq\ arguments from Ref. \cite{mr}, Ref. \cite{ek}, and Ref. \cite{get},
similarly to the line of arguments used in Ref. \cite{mrz}. For the sake of completeness and convenience of the reader we summarize these results.
\section{Strict quasi-regularity, strict capacity, and the construction of $R_{\alpha}$}\label{2}
For notations and notions that might not be defined here we refer to Ref. \cite{Tr5} and references therein. Throughout the paper let $E$ be a Hausdorff space such that its Borel $\sigma$-algebra ${\cal{B}}(E)$ is
generated by the set ${\cal{C}}(E)$ of all continuous functions on $E$, and let $m$ be a
$\sigma$-finite measure on $(E,{\cal{B}}(E))$. Let $({\cal{E}},{\cal{F}})$ be a generalized Dirichlet form with sectorial part $({\cal A},{\cal V})$ on ${\cal{H}}=L^2(E,m)$.
Let $(G_\alpha)_{\alpha>0}$ be $L^2(E,m)$-{\it resolvent associated with} ${\cal{E}}$, and $(\widehat G_\alpha)_{\alpha >0}$ be
the adjoint of $(G_\alpha)_{\alpha>0}$ in ${\cal{H}}$.
\subsection{Quasi-regular generalized Dirichlet forms and the conditions {\bf D3} and {\bf SD3}}\label{2.1}
Given $\varphi\in L^2(E,m)$, $\varphi>0$, an increasing sequence of closed subsets $(F_k)_{k\ge 1}$ of $E$
is an ${\cal{E}}$-nest, if
$$
\mbox{Cap}_{\varphi}(F_k^c)=\int_E (G_1 \varphi)_{F_k^c}\varphi\,dm\longrightarrow 0 \ \mbox{ as } k \to \infty.
$$
By IV. Proposition 2.10 in Ref. \cite{St1} the notion of ${\cal{E}}$-nest is independent of the special choice of $\varphi$.
Accordingly to $\mbox{Cap}_{\varphi}$, ${\cal{E}}$-exceptional sets, ${\cal{E}}$-quasi-continuity, etc., and the quasi-regularity are defined (see Ref. \cite{St1}). \\
In contrast to the theory of sectorial Dirichlet forms in Ref. \cite{mr} and Ref. \cite{semi_df} it is not known whether
quasi-regularity is sufficient for the existence of an associated standard process in case of a generalized Dirichlet form.
Therefore the following condition
\begin{itemize}
\item[$\mathbf{D3}$]There exists a linear subspace $\mathcal{Y}\subset{\cal{H}}\cap L^{\infty}(E,m)$ such that $\mathcal{Y}\cap{\cal{F}}$ is dense in ${\cal{F}}$,
$\lim_{\alpha\rightarrow\infty}e_{\alpha G_{\alpha} u-u}=0$ in ${\cal{H}}$ for all $u\in\mathcal{Y}$ and for the closure $\overline{\mathcal{Y}}$ of $\mathcal{Y}$
in $L^{\infty}(E;m)$ it follows that $u\wedge\alpha\in \overline{\mathcal{Y}}$ for $u\in \overline{\mathcal{Y}}$ and $\alpha \geq 0.$
\end{itemize}
is introduced in IV.2 in Ref. \cite{St1} and it is shown in IV. Theorem 2.2 of Ref. \cite{St1} that a {\it quasi-regular generalized Dirichlet
form satisfying} {\bf D3}
is associated with an $m$-{\it tight special standard process}.\\
By an algebra of functions we understand a {\it linear} space that is closed under multiplication. The following condition
\begin{itemize}
\item[$\mathbf{SD3}$] There exists an algebra of functions $\mathcal{G}\subset {\cal{H}} \cap L^{\infty}(E,m)$ such that $\mathcal{G}\cap {\cal{F}}$ is dense in ${\cal{F}}$ and
$\lim_{\alpha\rightarrow \infty}e_{\alpha G_{\alpha} u-u}$ in ${\cal{H}}$ for every $u\in\mathcal{G}.$
\end{itemize}
was introduced in Ref. \cite{Tr5}. We have the following:
\begin{prop}\label{sd3}
It holds:
\begin{itemize}
\item[(i)] {\bf SD3} implies {\bf D3}.
\item[(ii)] {\bf SD3} holds for any (sectorial semi-)Dirichlet form $({\cal{E}},D({\cal{E}}))$ on $L^2(E;m)$ with $\mathcal{G}=D({\cal{E}})\cap L^{\infty}(E,m)$.
\end{itemize}
\end{prop}
\begin{proof}
(i) The proof is standard; cf. e.g. proof of IV. Proposition 2.1 in Ref. \cite{St1}.\\
(ii) Follows from I.Corollary 4.15 and Proposition 4.17(ii) in Ref. \cite{mr}.
\end{proof}
\subsection{Strict capacities and strictly quasi-regular generalized Dirichlet forms}
We fix $\varphi\in L^1(E,m)\cap {\cal{B}}$ with $0<\varphi(z)\leq 1$ for every $z\in E$. The
following definition is a notational simplification of Definition 1 of Ref. \cite{Tr5}.
\begin{defn}\label{cap}
For $U\subset E$, $U$ open, set
$$
\mbox{Cap}_{1,\widehat G_1 \varphi}(U):=\int_E e_{U} \varphi\, dm
$$
where $e_U:=\lim_{k\rightarrow\infty}(1\wedge G_1(k\varphi))_U$ exists as a bounded and increasing limit in $L^{\infty}(E,m)$. If
$A\subset E$ arbitrary then
$\mbox{Cap}_{1,\widehat G_1 \varphi}(A):=\inf\{\mbox{Cap}_{1,\widehat G_1 \varphi}(U)| \,U\supset A, U \mbox{open}\}$.
\end{defn}
By Theorem 1 of Ref. \cite{Tr5} $\mbox{Cap}_{1,\widehat G_1 \varphi}$ is a finite Choquet capacity. A priori the function $e_U$ depends on the chosen $\varphi$ but in the next lemma we will see that this is actually not the case.
\begin{lem}\label{capeq}
$U\subset E$, $U$ open.
\begin{itemize}
\item[(i)] We have
$$
\mbox{Cap}_{1,\widehat G_1 \varphi}(U)=\sup_{u\in {\cal{P}}_{\cal{F}},u\le 1}{\cal{E}}_1(u_U,\widehat G_1 \varphi),
$$
where ${\cal{P}}$ denotes the $1$-excessive elements of ${\cal{V}}$, and ${\cal{P}}_{\cal{F}}=\{u\in {\cal{P}}|\exists f\in{\cal{F}}, u\le f\}$.
\item[(ii)] It holds
$$
e_U=\esssup\{u_U\,|\,u\in {\cal{P}}_{\cal{F}},u\le 1\}.
$$
\end{itemize}
\end{lem}
\begin{proof}
(i) Clearly \lq\lq$\le$\rq\rq\ holds in the statement.
For $f\in L^2(E;m)$ it is not difficult to see that $G_1 f>0$ $m$-a.e. whenever $f>0$ $m$-a.e. Indeed, since $m$ is $\sigma$-finite and $(\widehat{G}_{\alpha})_{\alpha>0}$ is
positivity preserving by I. Remark 4.2 of Ref. \cite{St1}, we can easily show by I. Proposition 3.4 in Ref. \cite{St1} that
if $A\in {\cal B}(E)$ and $\int_A G_1 f \,dm=0$, then $m(A)=0$. Thus $\sup_{k\ge 1}1\wedge G_1(k \varphi)=1$ $m$-a.e.
But then we have by (9) of Ref. \cite{Tr2}
\begin{eqnarray*}
\sup_{u\in {\cal{P}}_{\cal{F}},u\le 1}{\cal{E}}_1(u_U,\widehat G_1 \varphi) & = & \sup_{u\in {\cal{P}}_{\cal{F}},u\le 1}\sup_{\alpha\ge 1}{\cal{E}}_1(u,(\widehat G_1 \varphi)_U^{\alpha})\\
& \le & \sup_{\alpha\ge 1}\sup_{k\ge 1}{\cal{E}}_1(1\wedge G_1(k \varphi),(\widehat G_1 \varphi)_U^{\alpha})=\mbox{Cap}_{1,\widehat G_1 \varphi}(U).\\
\end{eqnarray*}
(ii) Define ${\cal{P}}_{\cal{F}}^{1}:=\{u\in {\cal{P}}_{\cal{F}}\,|\,u\le 1\}$. Since $e_U=\esssup\{\big (1\wedge G_1(k\varphi)\big )_U\,|\, k\ge 1\}$, it is enough to show that $e_U\ge u_U$ for every $u\in {\cal{P}}_{\cal{F}}^{1}$. Suppose this is wrong. Then there is some $f\in {\cal{P}}_{\cal{F}}^{1}$ such that
\begin{equation}\label{zwei}
\int_E \left (f_U\vee e_{U}\right ) \,\varphi\, dm>\int_E e_{U} \varphi\, dm=\mbox{Cap}_{1,\widehat G_1 \varphi}(U).
\end{equation}
On the other hand since obviously $f_U\vee \big (1\wedge G_1(k\varphi)\big )_U\le \big (f\vee \big (1\wedge G_1(k\varphi)\big )\big )_U\in {\cal{P}}_{\cal{F}}^{1}$ for any $k$ we obtain by (i)
\begin{eqnarray*}
\int_E \left (f_U\vee e_{U}\right ) \,\varphi\, dm & = &
\lim_{k\rightarrow\infty}\int_E \big (f_U\vee \big (1\wedge G_1(k\varphi)\big )_U \big ) \,\varphi\, dm\\
& \le & \limsup_{k\rightarrow\infty} \int_E \big (f\vee \big (1\wedge G_1(k\varphi)\big )\big )_U\,\varphi\, dm\le \mbox{Cap}_{1,\widehat G_1 \varphi}(U),\\
\end{eqnarray*}
which contradics (\ref{zwei}).
\end{proof}
\begin{rem}\label{all}
\begin{itemize}
\item[(i)] Let $\overline{\mbox{Cap}}_{1,\widehat G_1 \varphi}$ be the strict capacity
as defined in V.2 of Ref. \cite{mr}. Then in the sectorial case, i.e. when $({\cal E}, {\cal F})$ is a Dirichet form, we have ${\cal{F}}={\cal{V}}$ and ${\cal{P}}_{\cal{F}}={\cal{P}}$.
Thus $\mbox{Cap}_{1,\widehat G_1 \varphi}=\overline{\mbox{Cap}}_{1,\widehat G_1 \varphi}$ by Lemma \ref{capeq} and Definition V.2.1 in \cite{mr}.
Moreover, the function $e_U$ defined in Definition \ref{cap} is an explicit realization of the functon $e_U$ of Lemma 2.2 in Ref. \cite{mr}.
\item[(ii)] It follows immediately from Lemma \ref{capeq} that the strict ${\cal{E}}$-nests, and hence the strict notions, do not depend on the special choice of $\varphi$.
\end{itemize}
\end{rem}
Adjoining the cemetery $\Delta$ to $E$ we let $E_{\Delta}:=E\cup\{\Delta\}$ and ${\cal{B}}(E_{\Delta})={\cal{B}}(E)\cup\{B\cup \{\Delta\}|B\in{\cal{B}}(E)\}$.
We will consider different topologies on
$E_{\Delta}$.
If $E$ is a locally compact separable metric space but not compact,
$E_{\Delta}$ will be the one point compactification of $E$, i.e. the open sets of $E_{\Delta}$ are the open sets of $E$
together with the sets of the form $E_{\Delta}\setminus K$, $K\subset E$, $K$ compact in $E$. Otherwise
we adjoin the cemetery $\Delta$ to $E$ as an isolated point.
We extend
$m$ to $(E_{\Delta},{\cal{B}}(E_{\Delta}))$ by setting $m(\{\Delta\})=0$.
Any real-valued function u on $E$ is extended to $E_{\Delta}$ by setting $u(\Delta)=0$.
\\
Given an increasing sequence $(F_k)_{k\ge 1}$ of closed subsets of $E$, we define
$$
C_{l,\infty}(\{F_k\})=\{f:A\rightarrow \mathbb{R}\mid \bigcup_{k\ge 1}F_k\subset A\subset E,\,
f_{\mid F_k\cup \{\Delta\}} \mbox{ is lower semicontinuous } \forall k\},
$$
and $C(\{F_k\})$, $C_{\infty}(\{F_k\})$ as on page 360 of Ref. \cite{Tr5}.\\
Accordingly to $\mbox{Cap}_{1,\widehat G_1 \varphi}$, the notions strictly ${\cal{E}}$-exceptional (s.${\cal{E}}$-exceptional), strict ${\cal{E}}$-nest (s.${\cal{E}}$-nest),
strictly ${\cal{E}}$-quasi-everywhere (s.${\cal{E}}$-q.e.), strictly ${\cal{E}}$-quasi-continuous (s.${\cal{E}}$-q.c.), and strictly ${\cal{E}}$-quasi-lower-semicontinuous (s.${\cal{E}}$-q.l.s.c.), are defined (see Ref. \cite{Tr5}).
We observe that Proposition 2(i) in Ref. \cite{Tr5} can be generalized as follows:
\begin{prop}\label{a}
Let $u\in {\cal{H}}$ with s.${\cal{E}}$-q.l.s.c. $m$-version $\widetilde{u}$ and suppose further that $e_u$ exist.
Then for any $\varepsilon>0$
\begin{eqnarray*}
\mbox{Cap}_{1,\widehat G_1 \varphi}(\{\widetilde{u} >\varepsilon\})\le \varepsilon^{-1}
\int_E e_u \,\varphi\,dm.
\end{eqnarray*}
\end{prop}
\begin{proof} We have
\begin{eqnarray*}
\mbox{Cap}_{1,\widehat G_1 \varphi}(\{\widetilde{u} >\varepsilon\}) &\le& \lim_{k\to \infty}\lim_{\alpha\to \infty}{\cal{E}}_1(1\wedge G_1(k\varphi), (\widehat G_1 \varphi)^{\alpha}_{\{\widetilde{u} >\varepsilon\}}).
\end{eqnarray*}
Since $1\wedge G_1(k\varphi)\le 1\le \varepsilon^{-1} e_u$ $m$-a.e. on $\{\widetilde{u} >\varepsilon\}$ for any $k\in \mathbb{N}$, and
$(\widehat G_1 \varphi)^{\alpha}_{\{\widetilde{u} >\varepsilon\}}$ is $1$-coexcessive, and since $e_u$ is $1$-excessive and
$(\widehat G_1 \varphi)^{\alpha}_{\{\widetilde{u} >\varepsilon\}}\le \widehat G_1 \varphi$ $m$-a.e. for any $\alpha>0$ the result
now easily follows.
\end{proof}\\
{\bf From now on} we fix a {\it generalized Dirichlet form $({\cal{E}},{\cal{F}})$ that is strictly quasi-regular}
(see Definition 2 in Ref. \cite{Tr5}).
Using Proposition \ref{a} it is not difficult to see that strict versions of statements in Ref. \cite{St1} hold as stated in the following Lemma \ref{l1}. However, we remark that the strict quasi-regularity in Lemma \ref{l1} is only used to ensure the existence of a strict ${\cal{E}}$-nest of compact metrizable sets for the proof of (ii) (cf. III. Proposition 3.2 in Ref. \cite{St1}).
\begin{lem}\label{l1}
\begin{itemize}
\item[(i)] Let $S$ be a countable family of s.${\cal{E}}$-q.c. functions (resp. s.${\cal{E}}$-q.l.s.c. functions). Then there exists
a s.${\cal{E}}$-nest $(F_k)_{k\geq 1}$ such that $S\subset C_{\infty}(\{F_k\})$ (resp. $S\subset C_{l,\infty}(\{F_k\})$ ).
\item[(ii)] If $f$ is s.${\cal{E}}$-q.s.l.c. and $f\leq 0$ $m$-a.e. on an open set $U\subset E$, then $f\leq 0$ s.${\cal{E}}$-q.e. on $U.$
If $f,g$ are s.${\cal{E}}$-q.c. and $f=g$ $m$-a.e. on an open set $U\subset E$, then $f=g$ s.${\cal{E}}$-q.e. on $U.$
\item[(iii)] Let $u_n\in {\cal{H}}$ with s.${\cal{E}}$-q.c. $m$-version $\widetilde{u}_n$, $n\ge 1$, such that $e_{u_n-u}+e_{u-u_n}\to 0$ in ${\cal{H}}$ as $n\to \infty$
for some $u\in{\cal{H}}$. Then there is a subsequence $(\widetilde{u}_{n_k})_{k\ge 1}$ and a s.${\cal{E}}$-q.c. $m$-version $\widetilde{u}$ of $u$ such that $\lim_{k\ge 1}\widetilde{u}_{n_k}=\widetilde{u}$
s.${\cal{E}}$-quasi-uniformly.
\item[(iv)] Let $u_n\in {\cal{F}}$ with s.${\cal{E}}$-q.c. $m$-version $\widetilde{u}_n$, $n\ge 1$, and $u_n\to u$ in ${\cal{F}}$. Then
there is a subsequence $(\widetilde{u}_{n_k})_{k\ge 1}$ and a s.${\cal{E}}$-q.c. $m$-version $\widetilde{u}$ of $u$ such that $\lim_{k\ge 1}\widetilde{u}_{n_k}=\widetilde{u}$
s.${\cal{E}}$-quasi-uniformly.
\end{itemize}
\end{lem}
By strict quasi-regularity one can find a strict ${\cal{E}}$-nest of compact metrizable sets $(E_k)_{k\ge 1}$ as in IV. Lemma 1.10 in Ref. \cite{St1} (see also Ref. \cite{mr}). Let
$$
Y_1:=\bigcup_{k\in \mathbb{N}} E_k.
$$
Then $Y_1$ is a Lusin space. Since
$E\setminus Y_1$ is strictly ${\cal{E}}$-exceptional it is ${\cal{E}}$-exceptional, hence $m(E\setminus Y_1)=0$ and
we may identify $L^2(E;m)$ with $L^2(Y_1,m)$ canonically. \\
By Lemma 2 in Ref. \cite{Tr5} we know that for any $\alpha>0$ there exists a kernel $\widetilde{R}_{\alpha}$
from $(E,{\cal{B}}(E))$ to $(Y_1,{\cal{B}}(Y_1))$ such that
\begin{enumerate}
\item[(R1)] $\alpha\widetilde{R}_{\alpha}(z,Y_1)\le 1$ for all $z\in E$.
\item[(R2)] $\widetilde{R}_{\alpha}f$ is a s.${\cal{E}}$-q.c. $m$-version of $G_{\alpha}f$ for all (measurable) $f\in {\cal{H}}$.
\end{enumerate}
Moreover, the kernel $\widetilde{R}_{\alpha}$ is unique in the sense that, if $K$ is another kernel
from $(E,{\cal{B}}(E))$ to $(Y_1,{\cal{B}}(Y_1))$ satisfying $(R1)$ and $(R2)$, it follows that
$K(z,\cdot)=\widetilde{R}_{\alpha}(z,\cdot)$ s.${\cal{E}}$-q.e. \\
The next Proposition \ref{sp}(ii) is even when applied to the sectorial case an improvement over Lemma 3.4 in Ref. \cite{mrz} since we can choose
the function $h$ of Lemma 3.4 in Ref. \cite{mrz} to be in the domain of the infinitesimal generator. Note also that Proposition \ref{sp}(ii) is a statement about existence and not stated for any $\varphi$ with the given properties.
\begin{prop}\label{sp}
\begin{itemize}
\item[(i)] Let $(u_n)_{n\ge 1}\subset {\cal{H}}$ densely. Then $\{\widetilde{R}_{1}u_n^+, \widetilde{R}_{1}u_n^-; n\ge 1\}$
separates the points of $E_{\Delta}\setminus N$, where $N$ is some s.${\cal{E}}$-exceptional set.
\item[(ii)] There is some $\varphi\in L^1(E,m)\cap {\cal{B}}$ such that $0<\varphi(z)\le 1$ for every $z\in E$, and such that $\widetilde{R}_{1}\varphi >0$ s.${\cal{E}}$-q.e.
\end{itemize}
\end{prop}
\begin{proof}
(i) Using Lemma \ref{l1}(iv) the proof is the same as in IV. Proposition 1.9 in Ref. \cite{St1}.\\
(ii) Choose $(u_n)_{n\ge 1}\subset L^1(E,m)\cap {\cal{B}}_b$ such that $(u_n)_{n\ge 1}\subset {\cal{H}}$ densely. Then by (i)
$\{\widetilde{R}_{1}u_n^+, \widetilde{R}_{1}u_n^-; n\ge 1\}$ separates the points of $E_{\Delta}\setminus N$, where $N$ is some s.${\cal{E}}$-exceptional set.
Define
$$
h(x):=\sum_{n\ge 1} c_n\widetilde{R}_{1}(u_n^+ + u_n^-)(x), \ \ \mbox{ with } \ c_n :=2^{-n}(1+\|u_n \|_{L^{1}(E,m)}+\|u_n \|_{L^{\infty}(E,m)}).
$$
Since $\{\widetilde{R}_{1}u_n^+, \widetilde{R}_{1}u_n^-; n\ge 1\}$ separates the points of $E_{\Delta}\setminus N$ we have
$h(x)>0$ for all $x\in E\setminus N$. Since $g_k:=\sum_{n = 1}^k c_n (u_n^+ + u_n^-)$ converges in $L^{1}(E,m)$ to some $g$ with $0\le g \le 1$ and $\widetilde{R}_{1}$ is
a kernel we obtain $h=\widetilde{R}_{1}g$. Now choose $\rho\in L^{1}(E,m)$ with $0< \rho\le 1$. Then $\varphi:=\rho \vee g$ is the desired function.
\end{proof}
{\bf From now on} we assume that the strictly quasi-regular generalized Dirichlet form ${\cal{E}}$ satisfies {\bf D3}.
Using Lemma \ref{l1}, Proposition \ref{sp}, and the strict version of IV. Proposition 2.8 in Ref. \cite{St1} we obtain the following:
\begin{lem}\label{J0}
There exists a countable family $J_0$ of bounded strictly ${\cal{E}}$-quasi-continuous 1-excessive functions and a Borel set $Y\subset Y_1$ satisfying:
\begin{enumerate}
\item[(i)] If $u,v\in J_0,\ \alpha,c_1,c_2\in\mathbb{Q}_+^*$, then $\widetilde{R}_{\alpha}u,\ u\wedge v,\ u\wedge 1,\ (u+1)\wedge v,\ c_1 u+c_2 v$ are all in $J_0$.
\item[(ii)] $N:=E\setminus Y$ is strictly ${\cal{E}}$-exceptional and $\widetilde{R}_{\alpha}(x,N)=0,$ for all $x\in Y,\ \alpha\in\mathbb{Q}_+^*$.
\item[(iii)] $J_0$ separates the points of $Y_{\Delta}$.
\item[(iv)] If $u\in J_0,\ x\in Y$, then $\beta \widetilde{R}_{\beta+1}u(x)\leq u(x)$ for all $\beta\in\mathbb{Q}_+^*$,\\
$\widetilde{R}_{\alpha}u(x)-\widetilde{R}_{\beta}u(x)=(\beta-\alpha)\widetilde{R}_{\alpha}\widetilde{R}_{\beta}u(x)$ for all $\alpha,\beta\in\mathbb{Q}_+^*$,\\
$\lim_{\mathbb{Q}_+^*\ni \alpha\rightarrow\infty}\alpha \widetilde{R}_{\alpha+1}u(x)=u(x).$\\
\end{enumerate}
\end{lem}
Next, we define for $\alpha\in\mathbb{Q}_+^*,\ A\in{\cal{B}}(Y_{\Delta}):={\cal{B}}(E_{\Delta})\cap Y_{\Delta}$
\begin{equation}\label{k1}
R_{\alpha}(x,A) :=
\begin{cases}
\widetilde{R}_{\alpha}(x,A\cap Y)+\left ( \frac{1}{\alpha} -\widetilde{R}_{\alpha}(x,Y)\right )1_A(\Delta),& if\ x\in Y \\
\frac{1}{\alpha} 1_A(\Delta),& if x=\Delta
\end{cases}
\end{equation}
and set
\begin{equation}\label{k2}
J:=\{u+c1_{Y_{\Delta}}\mid u\in J_0, c\in\mathbb{Q}_+\}.
\end{equation}
Since $J_0$ separates the points of $Y_{\Delta}$, so does $J$. The following lemma is also clear.
\begin{lem}\label{J}
Let $(R_{\alpha})_{\alpha\in\mathbb{Q}_+^*}$ and J be as in (\ref{k1}), (\ref{k2}). Then the statements of Lemma \ref{J0} remain true with
$J_0,\ Y$ and $\widetilde{R}_{\alpha}$ replaced by $J,\ Y_{\Delta}$ and $R_{\alpha}$ respectively.
\end{lem}
\subsection{The construction of nice excessive functions}\label{2.3}
Since strict quasi-regularity implies quasi-regularity by Proposition 2(ii) in Ref. \cite{Tr5}, and since {\bf D3} is in force, we obtain
by IV.Theorem 2.2 in Ref. \cite{St1}
that $({\cal{E}},{\cal{F}})$ is associated with some $m$-tight $m$-special standard process. We denote the process resolvent by
$$
V_{\alpha} f(z)=E_z\left [\int_0^{\infty}e^{-\alpha t}f(Y_t)dt\right ], \ \ \ \ \alpha>0,\ f\in{\cal{H}} \cap L^{\infty}(E,m).
$$
By Remark \ref{all}(ii) the strict capacity does not depend on the special choice of $\varphi$. We may and will hence {\bf from now on} assume that $\varphi$ is as in Proposition \ref{sp}(ii).
The following two lemmas are crucial for the later study of weak limits.
\begin{lem}\label{gf}
Let $U_n\subset E,\ n\geq 1$ be a decreasing sequence of open sets such that $\mbox{Cap}_{1,\widehat G_1 \varphi}(U_n)\rightarrow 0$, as $n\rightarrow\infty$.
Then we can find $m$-versions $e_n$ of $e_{U_n}$ such that:
\begin{enumerate}
\item[(i)] $e_n\geq 1$ ${\cal{E}}$-q.e. on $U_n,\ n\geq 1.$ In particular, there are ${\cal{E}}$-exceptional sets $N_n\in {\cal{B}}(E)$, $N_n\subset U_n$, such that
\begin{equation*}
\widehat{e}_n(x):=e_n(x)+ 1_{N_n}(x)\ge 1
\ \ \forall x\in U_n, \ n\ge 1.
\end{equation*}
\item[(ii)] $\alpha \widetilde{R}_{\alpha+1}e_n\leq e_n$ and $\alpha \widetilde{R}_{\alpha+1}\widehat{e}_n\leq \widehat{e}_n$ s.${\cal{E}}$-q.e. for any $\alpha\in\mathbb{Q}_+^*,\ n\geq 1.$
\item[(iii)] $e_n\searrow 0$ and $\widehat{e}_n\to 0$ s.${\cal{E}}$-q.e. as $n\rightarrow \infty$.
\end{enumerate}
\end{lem}
\begin{proof}
Define for $n\ge 1$
\begin{equation}\label{en}
e_n:=\sup_{\alpha\ge 1}\sup_{l\ge 1}\alpha \widetilde{R}_{\alpha +1}\big(\left (1\wedge V_1(l\varphi)\right )_{U_n}\big).
\end{equation}
where $\left (1\wedge V_1(l\varphi)\right )_{U_n}$ is some bounded measurable $m$-version of $\left (1\wedge G_1(l\varphi)\right )_{U_n}$.
Clearly $e_n$ is an $m$-version of $e_{U_n}$.
Since
$\left (1\wedge G_1(l\varphi)\right )_{U_n}$ is $1$-excessive, and $\widetilde{R}_{\alpha+1}f$ is s.${\cal{E}}$-q.c.
for any (measurable) $f\in {\cal{H}}$ by (R2),
by Lemma \ref{l1}(ii) it is clear that the first part of (ii) holds. The second part of (ii) similarly also holds once we have shown that
$N_n$ is ${\cal{E}}$-exceptional, hence in particular $m$-negligible. This is done at the end of the proof. \\
Obviously $e_n$ is s.${\cal{E}}$-q.l.s.c, s.${\cal{E}}$-q.e. decreasing in $n$, $\lim_{n\to \infty} e_n$ exists s.${\cal{E}}$-q.e. and $\lim_{n\to \infty} e_n\ge 0$ s.${\cal{E}}$-q.e.
We have
\begin{eqnarray*}
\mbox{Cap}_{1,\widehat G_1 \varphi}\left (\{\lim_{n\to \infty} e_n>0\}\right )
& \le & \sum_{k\ge 1}\mbox{Cap}_{1,\widehat G_1 \varphi}\left (\cap_{n\ge 1}\left \{e_n>k^{-1}\right \}\right ). \\
\end{eqnarray*}
Up to some ${\cal{E}}$-exceptional set $\left \{e_n>k^{-1}\right \}\subset \bigcup_{\alpha\ge 1}\bigcup_{l\ge 1}\left \{\alpha \widetilde{R}_{\alpha +1}\big (\left (1\wedge V_1(l\varphi)\right )_{U_n}\big )>k^{-1}\right \}$,
and $\left \{\alpha \widetilde{R}_{\alpha +1}\big (\left (1\wedge V_1(l\varphi)\right )_{U_n}\big )>k^{-1}\right \}$ is increasing in $\alpha$, $l$. Then, since $\mbox{Cap}_{1,\widehat G_1 \varphi}$ is a Choquet capacity we obtain with Proposition \ref{a}
\begin{eqnarray*}
\mbox{Cap}_{1,\widehat G_1 \varphi}\left (\left \{e_n>k^{-1}\right \}\right )
&\le& k\cdot \sup_{\alpha\ge 1}\sup_{l\ge 1} \int_E \alpha \widetilde{R}_{\alpha +1}\big (\left (1\wedge V_1(l\varphi)\right )_{U_n}\big )\varphi \,dm. \\
\end{eqnarray*}
Therefore
\begin{eqnarray*}
\mbox{Cap}_{1,\widehat G_1 \varphi}\left (\cap_{n\ge 1}\left \{e_n>k^{-1}\right \}\right )
& \le & k\cdot\inf_{n\ge 1} \int_E e_n \,\varphi \,dm=0. \\
\end{eqnarray*}
Thus the first part of (iii) holds. The second part of (iii) is clear since $\limsup_{n\ge 1}1_{N_n}\le 1_{\cap_{n\ge 1}U_n}=0$ s.${\cal{E}}$-q.e.\\
Since $\widetilde{R}_{\alpha+1}f$, $V_{\alpha+1}f$, $f\in {\cal{H}}\cap L^{\infty}(E,m)$, are ${\cal{E}}$-q.c. and coincide $m$-a.e. by (R2), it
follows by III.Corollary 3.4 in Ref. \cite{St1} that $V_{\alpha+1}f=\widetilde{R}_{\alpha+1}f$ ${\cal{E}}$-q.e. Using
$\left (1\wedge V_1(l\varphi)\right )_{U_n}=1\wedge V_1(l\varphi)$ $m$-a.e. on $U_n$, and $1\wedge V_1(l\varphi)\nearrow \mathbb{I}_E$ as
$l\to\infty$, it follows ${\cal{E}}$-q.e.
\begin{equation}\label{e2}
e_n \ge \limsup_{\alpha\ge 1}\alpha V_{\alpha +1} \mathbb{I}_{U_n}.
\end{equation}
By right-continuity and normality of the process $Y$ we obtain for all $z\in U_n$
$$
\lim_{\alpha\to\infty}\alpha V_{\alpha +1} \mathbb{I}_{U_n}(z)=\lim_{\alpha\to\infty}\frac{\alpha}{\alpha+1}E_z\left [\int_0^{\infty}e^{-t}\mathbb{I}_{U_n}(Y_{\frac{t}{\alpha+1}})dt\right ]=1.
$$
Hence the first part of (i) holds. For the second part of (i) we can find ${\cal{E}}$-exceptional sets $N_n\in {\cal{B}}(E)$, $N_n\subset U_n$, with $e_n \cdot 1_{U_n\setminus N_n}+1_{N_n}\ge 1$ pointwise on $U_n$.
But $e_n\ge 0$ everywhere since $\widetilde{R}_{\alpha +1}$ is a kernel and so we obtain $\widehat{e}_n\ge 1$ on $U_n$ as desired.
\end{proof}
\begin{rem}\label{enr}
In Lemma \ref{gf}(i) we were not able to show directly
\begin{equation}\label{enh}
e_n\ge 1 \mbox{ s.}{\cal{E}}\mbox{-q.e. on }\ U_n, \ n\ge 1.
\end{equation}
(Unfortunately, $\left (1\wedge V_1(l\varphi)\right )_{U_n}$ has only a s.${\cal{E}}$-q.l.s.c. $m$-version in general and the inequality in Lemma \ref{l1}(ii) is just the wrong way around.) (\ref{enh}) is used in Lemma 3.5, Lemma 3.6, and proof of Theorem 3.3 in Ref. \cite{mrz} in an essential way. Instead, we will use the functions $\widehat{e}_n$ defined in Lemma \ref{gf}(i) which is sufficient
(cf. Lemmas \ref{ss} and \ref{liq}, and Theorem \ref{key}). We remark that it is even sufficient to only know that $e_n\ge 1$ $m$-a.e. on $U_n$, so that the sets $N_n$ in Lemma \ref{gf}(i) are only $m$-negligible.\\
It will turn out a posteriori that (\ref{enh}) actually holds. In fact, by our main result Theorem \ref{hunt2} below it follows that the process resolvent
$V_{\alpha+1}f$, $f\in {\cal{H}}\cap L^{\infty}(E,m)$, is s.${\cal{E}}$-q.c. Thus applying Lemma \ref{l1}(ii) $V_{\alpha+1}f=\widetilde{R}_{\alpha+1}f$ s.${\cal{E}}$-q.e.
Therefore (\ref{e2}) holds s.${\cal{E}}$-q.e. and (\ref{enh}) follows.
\end{rem}
\begin{lem}\label{ss}
In the situation of Lemma \ref{gf} there exists $S\in{\cal{B}}(E),\ S\subset Y$ such that $E\setminus S$ is strictly ${\cal{E}}$-exceptional and the following holds:
\begin{enumerate}
\item[(i)] $\widetilde{R}_{\alpha}(x,Y\setminus S)=0\ \forall x\in S,\ \alpha\in\mathbb{Q}_+^*.$
\item[(ii)] $\widehat{e}_n(x)\geq 1\ for\ x\in U_n,\ n\geq 1$, and $\widetilde{R}_{\alpha} 1_{N_n}(x)=0$ $\forall x\in S,\ \alpha\in\mathbb{Q}_+^*, n\geq 1$.
\item[(iii)]$\alpha\widetilde{R}_{\alpha +1}\widehat{e}_n(x)\leq \widehat{e}_n(x) \ \forall x\in S,\ \alpha\in\mathbb{Q}_+^*,\ n\geq 1$.
\item[(iv)] $\widehat{e}_n(x)\to 0$ as $n\to \infty\ \ \forall x\in S$.
\end{enumerate}
\end{lem}
\begin{proof}
The first assertion of (ii) holds by definition in Lemma \ref{gf}(i).
By Lemma \ref{gf}, $(R2)$, and Lemma \ref{l1}(ii), the rest of the proof works as in IV.3.11 of Ref. \cite{mr}.
\end{proof}
\section{The approximating forms ${\cal{E}}^{\beta}$ and the approximating processes $X^{\beta}$}\label{3}
Let $J,\ Y_{\Delta}$ and $(R_{\alpha})_{\alpha\in\mathbb{Q}_+^*}$ be as in Lemma \ref{J}. First, we collect some results
of Chapter 4 section 2 of Ref. \cite{ek}. For a fixed $\beta\in\mathbb{Q}_+^*$, let $\{Y^{\beta}(k),\ k=0,1,\ldots\}$ be a Markov chain
in $Y_{\Delta}$ with initial distribution
$\nu$ and transition function $\beta R_{\beta}$. Let further $(\Pi^{\beta}_t)_{t\geq 0}$ be a Poisson process with parameter $\beta$
and independent of $\{Y^{\beta}(k),\ k=0,1,\ldots\}$. Then it is known that
$$
X^{\beta}_t:=Y^{\beta}(\Pi^{\beta}_t)
$$
is a strong Markov process in $Y_{\Delta}$ with transition semigroup
\begin{equation}\label{ak}
P^{\beta}_t f=e^{-\beta t}\sum_{k=0}^{\infty}\frac{(\beta t)^k}{k!}(\beta R_{\beta})^k f\quad \forall t\geq 0,\ f\in{\cal{B}}_b(Y_{\Delta}),
\end{equation}
i.e. we have for all $t,s\ge 0, f\in {\cal{B}}_b(Y_{\Delta})$
\begin{equation}\label{ts}
E[f(X^{\beta}_{t+s})\mid \sigma(X^{\beta}_u,u\le t)]=(P^{\beta}_s f)(X^{\beta}_t).
\end{equation}
Here (\ref{ts}) easily follows from (2.14) of Chapter 4 in Ref. \cite{ek}.
Furthermore from the formula (\ref{ak}) one can see that $(P_t^{\beta})_{t\geq 0}$ is a strongly continuous contraction semigroup on the Banach space $({\cal{B}}_b(Y_{\Delta}),\|\cdot\|_{\infty})$.
The corresponding generator is
\begin{equation}\label{g}
L^{\beta}f(x)=\frac{d}{dt}P^{\beta}_tf(x)\big|_{t=0} =\beta(\beta R_{\beta}f(x)-f(x)),\quad f\in{\cal{B}}_b(Y_{\Delta}).
\end{equation}
Define the forms ${\cal{E}}^{(\beta)},\ \beta >0,$ by
$$
{\cal{E}}^{(\beta)}(u,v):=\beta (u-\beta G_{\beta}u,v)_{{\cal{H}}},\quad u,v\in{\cal{H}},
$$
where we recall that $(G_{\beta})_{\beta >0}$ is the $L^2$-resolvent of ${\cal{E}}$.
It is known (see e.g. Chapter I in Ref. \cite{mr}), that the $C_0$-semigroup of submarkovian contractions on $L^2(E;m)$ that is associated to ${\cal{E}}^{(\beta)}$ is given by
\begin{equation}\label{as}
T_t^{\beta}f=e^{-\beta t}\sum_{j=0}^{\infty}\frac{(\beta t)^j}{j!}(\beta G_{\beta})^j f, \quad f\in {\cal{H}}.
\end{equation}
From (\ref{ak}), (\ref{ts}), and (\ref{as}) it follows that $(X^{\beta}_t)$ is associated with ${\cal{E}}^{(\beta)}$.
Since $R_{\beta}f$ is an $m$-version of $G_{\beta}f$ for any measurable $f\in {\cal{H}}$, by I. Examples 4.9(ii) in Ref. \cite{St1} we see that ${\cal{E}}^{(\beta)}$ is a generalized Dirichlet form and
\begin{eqnarray*}
{\cal{E}}^{(\beta)}(u,v)=(-L^{\beta}u,v)_{{\cal{H}}},\quad u,v\in{\cal{H}}.\\
\end{eqnarray*}
For an arbitrary subset $M\subset E_{\Delta}$ let $\Omega_{M}:=D_{M}[0,\infty)$
be the space of all c\`adl\`ag functions from $[0,\infty)$ to $M$.
Let $(X_t)_{t\geq 0}$ be the coordinate process on $\Omega_{E_{\Delta}}$, i.e. $X_t(\omega)=\omega(t)$ for $\omega\in\Omega_{E_{\Delta}}$.
$\Omega_{E_{\Delta}}$ is equipped with the Skorokhod topology (see Chapter 3 in Ref. \cite{ek}).
Let
$P_x^{\beta}$ be the law of $X^{\beta}$ on $\Omega_{E_{\Delta}}$ with initial distribution $\delta_x$ if $x\in Y_{\Delta}$,
and if $x\in E_{\Delta}\setminus Y_{\Delta}$ let $P^{\beta}_x$ be the Dirac measure on
$\Omega_{E_{\Delta}}$ such that $P_x^{\beta}[X_t=x\mbox{ for all }t\geq 0]=1$. Finally, let $({\cal{F}}^{\beta}_t)_{t\geq 0}$ be the completion w.r.t.
$(P_x^{\beta})_{x\in E_{\Delta}}$ of the natural filtration of $(X_t)_{t\geq 0}$.
\begin{prop}\label{ah}
$\mathbb{M}^{\beta}:=(\Omega_{E_{\Delta}},\ (X_t)_{t\geq 0},\ ({\cal{F}}^{\beta}_t)_{t\geq 0},\ (P^{\beta}_x)_{x\in E_{\Delta}})$ is a Hunt process associated with ${\cal{E}}^{(\beta)}$, i.e. for all $t\geq 0$ and any m-version of $u\in L^2(E;m)$, $x\mapsto \int u(X_t)\; dP^{\beta}_x$ is an m-version of $T^{\beta}_t u.$
\end{prop}
\begin{proof} By construction $\mathbb{M}^{\beta}$ is a right process that has left limits in $E_{\Delta}$. The quasi-left continuity up to $\infty$
can be shown by a routine argument following Ref. \cite{get} (cf. IV.3.21 in Ref. \cite{mr}).
\end{proof}
\centerline{}
Let $J=\{u_n\mid n\in\mathbb{N} \}$ and
$$
g_n:=R_1 u_n,\ n\in \mathbb{N}.
$$
Define for all $x,y\in Y_{\Delta}$
$$
\rho(x,y)=\sum_{n=1}^{\infty}\frac{1}{2^n}\lvert g_n(x)-g_n(y)\rvert\wedge 1.
$$
By Lemma \ref{J0}(ii) and Lemma \ref{J} $\{g_n\mid n\in\mathbb{N}\}$ separates the points of $Y_{\Delta}$ and
hence $\rho$ defines a metric on $Y_{\Delta}$.
We may assume that $Y_{\Delta}$ is a Lusin topological space (cf. IV.Remark 3.2(iii) in Ref. \cite{mr}). It follows by Lemma 18 on p.108 of Ref. \cite{schwartz}
that ${\cal{B}}(Y_{\Delta})=\sigma(g_n\mid n\in\mathbb{N})=(\rho\mbox{--}){\cal{B}}(Y_{\Delta})$.
Now define
$$
\overline{E}:=\overline{Y}_{\Delta}^{\rho}.
$$
$(\overline{E},\rho)$ is a compact metric space by Tychonoff's theorem.\\
We extend the kernel $(R_{\alpha})_{\alpha\in\mathbb{Q}_+^*}$ to the space $\overline{E}$ by setting for $\alpha \in\mathbb{Q}_+^*,\ A\in{\cal{B}}(\overline{E})$,
\begin{equation}\label{k1bis}
R_{\alpha}(x,A) := \begin{cases}R_{\alpha}(x,A\cap Y_{\Delta}),& x\in Y_{\Delta}\\ \frac{1}{\alpha}1_A(x),& x\in \overline{E}\setminus Y_{\Delta}.\end{cases}\\
\end{equation}
We may regard $(X^{\beta}_t)_{t\geq 0}$ as a c\`adl\`ag process with state space $\overline{E}$ and use the same notation as before:
$P_x^{\beta}$ denotes hence the law of $(X^{\beta}_t)_{t\geq 0}$ in $\Omega_{\overline{E}}$ with initial distribution $\delta_x$.
Each $g_n$ is $\rho$-uniformly continuous and extends therefore uniquely to a continuous function on $\overline{E}$ which we denote again by $g_n$.\\
For the convenience of the reader we include the proof of the following theorem, which as we feel is slightly more transparent than the corresponding proof of Theorem 3.2 in Ref. \cite{mrz}.
\begin{theo}\label{tt}
$\{P_x^{\beta}\mid \beta\in\mathbb{Q}_+^*\}$ is relatively compact for any $x\in \overline{E}$.
\end{theo}
\begin{proof}
We first show that assumptions of 9.4 Theorem of Chapter 3 in Ref. \cite{ek} are fulfilled with $C_a=C(\overline{E})$ (where $C_a$ is
as in the just mentioned theorem of Ref. \cite{ek}).
Since $g_n\in D(L^{\beta})$ it follows that
$$
\left (g_n(X^{\beta}_t)-\int_0^t L^{\beta}g_n(X^{\beta}_s)\; ds\right )_{t\geq 0}
$$
is an $(P_x^{\beta}, (\mathcal{F}_t^{\beta})_{t\ge 0})$-martingale for any $x\in \overline{E}$.
Since $L^{\beta}g_n=1_{Y_{\Delta}}\beta R_{\beta}(g_n-u_n)$
we have for all $n\in\mathbb{N}$
\begin{eqnarray*}
\sup_{\beta\in\mathbb{Q}_+^*} \|L^{\beta} g_n\|_{\infty} = \sup_{\beta\in\mathbb{Q}_+^*}\|1_{Y_{\Delta}}\beta R_{\beta}(g_n-u_n)\|_{\infty}\leq \|1_{Y_{\Delta}}(g_n-u_n)\|_{\infty} < +\infty.
\end{eqnarray*}
So, we proved that $R_1 J:=\{g_n\mid n\in\mathbb{N}\}\subset D$ where $D\subset C(\overline{E})$ is the linear space from the theorem in Ref. \cite{ek}. Since for any $u\in J$
$$
R_1 J \ni R_1\left (\alpha(u-\alpha R_{\alpha +1}u\right ))(x)=\alpha R_{\alpha +1}u(x)\nearrow u(x),\quad \mathbb{Q}_+^*\ni \alpha\rightarrow\infty, \forall x\in Y_{\Delta},
$$
we see by Dini's theorem that every $u\in J$ has a unique ($\rho$-uniformly) continuous extension to $\overline{E}$ that is again denoted by $u$.
Thus we may and do consider $J$ as a subset of $C(\overline{E})$. In particular,
if $\overline{A}^{\|\cdot\|_{\infty}}$ denotes the uniform closure of $A\subset C(\overline{E})$, we have
$$
\overline{J-J}^{\|\cdot\|_{\infty}}\subset \overline{R_1(J-J)}^{\|\cdot\|_{\infty}}\subset \overline{D}^{\|\cdot\|_{\infty}}\subset C(\overline{E}).
$$
Since $J-J$ contains the constant functions, is inf-stable and separates the points of $\overline{E}$ we obtain that $J-J$ is dense in $C(\overline{E})$ by the Stone-Weierstra\ss\ theorem.
Hence $\overline{D}^{\|\cdot\|_{\infty}}= C(\overline{E})$ and so by the theorem of Ref. \cite{ek} $\{f\circ X^{\beta}\mid \beta\in \mathbb{Q}_+^*\}$ is relatively compact for all $f\in C(\overline{E})$.
Since $\overline{E}$ is compact, the compact containment condition trivially holds and so by 9.1 Theorem of Chapter 3 in Ref. \cite{ek} $\{X^{\beta}\mid \beta\in \mathbb{Q}_+^*\}$ is relatively compact as desired.
\end{proof}
\section{Limiting process associated with the strictly quasi-regular generalized Dirichlet form}\label{4}
For a Borel subset $S\subset Y$, we write $S_{\Delta}:=S\cup\{\Delta\}$. On $S_{\Delta}$ we consider (except otherwise stated) the topology induced by the metric $\rho$. In particular the $\rho$-topology
and the original one generate the same Borel $\sigma$-algebra on $S_{\Delta}$.\\
Let $\mathbb{M}^{\beta}:=(\Omega_{E_{\Delta}},\ (X_t)_{t\geq 0},\ ({\cal{F}}_t^{\beta})_{t\geq 0},\ (P^{\beta}_x)_{x\in E_{\Delta}})$
be the Hunt process from Proposition \ref{ah}, $R_{\alpha}$ and $\overline{E}$ be as in Section \ref{3}. Let $S\in{\cal{B}}(E)$ and $\widehat{e}_n$ be as in Lemma \ref{ss}.
In view of Lemma \ref{ss}(i), exactly as in Lemma 3.7 of Ref. \cite{mrz} one can show that
\begin{equation}\label{iv}
P_x^{\beta}[X_t\in S_{\Delta},\ X_{t-}\in S_{\Delta}\ \forall t\geq 0]=1\ \ \forall x\in S_{\Delta}.
\end{equation}
Due to Lemma \ref{ss}(i), (iii) the proof of the next lemma is also the same as the proof of Lemma 3.8 in Ref. \cite{mrz}.
\begin{lem}\label{liq}
Let $\beta\in\mathbb{Q}_+^*,\ \beta \geq 2,\ n\geq 1.$ Then $\widehat{e}_n$ is a $(P^{\beta}_t)$-2-excessive function on $S_{\Delta}$, i.e.
$e^{-2t}P_t^{\beta}\widehat{e}_n(x)\leq \widehat{e}_n(x)$ and $\lim_{t\rightarrow 0}e^{-2t}P_t^{\beta}\widehat{e}_n(x)=\widehat{e}_n(x)\ \ \forall x\in S_{\Delta}$.
\end{lem}
\begin{rem}\label{mis}
The proof of Lemma 3.7 in Ref. \cite{mrz} (resp. Lemma 3.2 in Ref. \cite{mrs}) contains an inaccuracy, namely it is not as stated true that $X_{\tau_n}\in U_n$ $P^{\beta}_x$-a.s. on $\{\tau_n < \infty\}$, where
$$
\tau_n:=\inf\{t\geq 0\mid X_t\in U_n\},\ \ n\in \mathbb{N}.
$$
By right-continuity $X_{\tau_n}$ will be in general in the closure $\overline{U}_n$ of $U_n$. This leads to a wrong argument
so that the proof of Lemma 3.7 in Ref. \cite{mrz} (resp. Lemma 3.2 in Ref. \cite{mrs}) cannot be maintained.
However, these lemmas are quite important since they tell us that any s.${\cal{E}}$-nest is a pointwise strict ${\cal{E}}^{\beta}$-nest.
Following the proof of Lemma \ref{liq2} below one can see how this inaccuracy can be corrected. Therefore the statement of
Lemma 3.7 in Ref. \cite{mrz} (resp. Lemma 3.2 in Ref. \cite{mrs}) remains true and no results of Ref. \cite{mrz} (resp. Ref. \cite{mrs}) are affected.
\end{rem}
\begin{lem}\label{liq2}
Let $\beta\in \mathbb{Q}_+^*,\ \beta\geq 2$.
Then $E^{\beta}_x[e^{-2\tau_n}]\leq \widehat{e}_n(x)\ \ \forall x\in S_{\Delta}$.
\end{lem}
\begin{proof}
Since by (\ref{iv}) $S_{\Delta}$ is invariant set of $\mathbb{M}^{\beta}$, the restriction $\mathbb{M}^{\beta}_{S_{\Delta}}$ of $\mathbb{M}^{\beta}$
to $S_{\Delta}$ is still a Hunt process.
Since $\widehat{e}_n$ is a $(P^{\beta}_t)$-2-excessive function on $S_{\Delta}$ we have that $(e^{-2t}\widehat{e}_n(X_t))_{t\geq 0}$
is a positive right-continuous $(P_x^{\beta}, ({\cal{F}}_t^{\beta})_{t\ge 0})$-supermartingale for all $ x\in S_{\Delta}$
(use IV. Proposition 5.14 claim 1 of Ref. \cite{mr} and monotone classes as indicated in the proof of 5.8(iii) in Ref. \cite{get}).
By the optional sampling theorem and normality we have
$$
E^{\beta}_x[e^{-2\tau_n}\widehat{e}_n(X_{\tau_n})]\leq \widehat{e}_n(x),\ x\in S_{\Delta}.
$$
By Lemma \ref{ss}(ii) we have that $\widehat{e}_n(x)\geq 1$ for all $x\in U_n$.
Hence, by right-continuity for all $x\in S_{\Delta}$ we have $\widehat{e}_n(X_{\tau_n})=\lim_{\mathbb{Q}^*_+\ni t_n\downarrow \tau_n}\widehat{e}_n(X_{t_n})\geq 1\ P_x^{\beta}$-a.s. on $\{\tau_n < \infty\}$. (As usual we let $X_{\infty}:=\Delta$, and $f(\Delta):=0$ for any function $f$). It follows that for all $x\in S_{\Delta}$
$$
E_x^{\beta}[e^{-2\tau_n}]\leq E^{\beta}_x[e^{-2\tau_n}\widehat{e}_n(X_{\tau_n})] \leq \widehat{e}_n(x).
$$
\end{proof}\\
In view of Lemma \ref{liq2}, Lemma \ref{ss}, the proof of the following \lq\lq key theorem\rq\rq is the same as in Theorem 3.3 of Ref. \cite{mrz}.
\begin{theo}\label{key}
There exists a Borel subset $Z\subset Y$ and a Borel subset $\Omega\subset \Omega_{\overline{E}}$ with the following properties:
\begin{enumerate}
\item[(i)] $E\setminus Z$ is strictly ${\cal{E}}$-exceptional.
\item[(ii)] $R_{\alpha}(x, \overline{E}\setminus Z_{\Delta})=0\ \ \forall x\in Z_{\Delta},\ \alpha\in\mathbb{Q}_+^*.$
\item[(iii)] If $\omega\in\Omega$, then $\omega_t,\ \omega_{t-}\in Z_{\Delta}$ for all $t\geq 0$. Moreover, each $\omega\in\Omega$ is c\`adl\`ag in the
original topology of $Y_{\Delta}$ and $\omega_{t-}^0=\omega_{t-}$ for all $t>0$, where $\omega_{t-}^0$ denotes the left limit in the original topology.
\item[(iv)] If $x\in Z_{\Delta}$ and $P_x$ is a weak limit of some sequence $(P^{\beta_j}_x)_{j\in\mathbb{N}}$ with $\beta_j \in\mathbb{Q}_+^*,\ \beta_j\uparrow\infty$, then $P_x[\Omega]=1$.
\end{enumerate}
\end{theo}
For $\alpha, \beta\ \in\mathbb{Q}_+^*$ let
$R_{\alpha}^{\beta}f(x):=E_x^{\beta}\left[\int_0^{\infty}e^{-\alpha t}f(X_t)\; dt\right], f\in{\cal{B}}_b(\overline{E}),\ x\in\overline{E}$.
Since the identities (\ref{ts}) and (\ref{g}) carry over to $B_b(\overline{E})$ it is straight forward to check that the resolvent of the Yosida approximation has the following form
\begin{equation*}
R_{\alpha}^{\beta}f=\left(\frac{\beta}{\alpha +\beta}\right)^2 R_{\frac{\alpha \beta}{\alpha +\beta}}f+\frac{1}{\alpha +\beta}f.
\end{equation*}
For the explicit calculations we refer to Lemma 4.1 in Ref. \cite{mrz}. It is equally straight forward to check (see Lemma 4.2 in Ref. \cite{mrz}) that if $P_x$, $x\in\overline{E}$, is
a weak limit of a subsequence $(P_x^{\beta_j})_{j\geq 1}$ with $\beta_j\uparrow\infty,\ \beta_j\in\mathbb{Q}_+^*$, then the kernel $P_t f(x):=E_x[f(X_t)]:=\int_{\Omega} f(X_t(\omega))P_x(d\omega), f\in{\cal{B}}_b(\overline{E})$,
satisfies
\begin{equation}\label{atd}
\int_0^{\infty}e^{-\alpha t}P_t f(x)\; dt= R_{\alpha}f(x),\quad \forall f\in{\cal{B}}_b(\overline{E}),\ \alpha\in\mathbb{Q}_+^*.
\end{equation}
In particular, the kernels $P_t,\ t\geq 0$, are independent of the subsequence $(P_x^{\beta_j})_{j\geq 1}$.
Then for every $x\in Z_{\Delta}$ ($Z$ as in a Theorem \ref{key}) the relatively compact set $\{P_x^{\beta}\mid \beta\in\mathbb{Q}_+^*\}$
has a unique limit $P_x$ for $\beta\uparrow\infty$, and the process $(\Omega_{\overline{E}},(X_t)_{t\geq 0}, (P_x)_{x\in Z_{\Delta}})$ is a Markov process with the transition semigroup
$(P_t)_{t\geq 0}$ determined by (\ref{atd}). Moreover,
\begin{equation}\label{tmv}
P_x[X_t \in Z_{\Delta},X_{t-}\in Z_{\Delta}\ for\ all\ t\geq 0]=1
\end{equation}
for all $x\in Z_{\Delta}$. The proof of this is again the same as in Theorem 4.3 of Ref. \cite{mrz}.
\centerline{}
Up to this end let $(P_x)_{x\in Z_{\Delta}}$ be as in (\ref{tmv}), and $\Omega$ and $Z_{\Delta}$ be specified by Theorem \ref{key}. Since $P_x[\Omega]=1$ for all
$x\in Z_{\Delta}$, we may restrict $P_x$ and the coordinate process $(X_t)_{t\geq 0}$ to $\Omega$. Let $({\cal{F}}_t)_{t\geq 0}$ be the natural filtration of $(X_t)_{t\geq 0}$.
Then exactly as in Theorem 4.4. of Ref. \cite{mrz} one shows that
$\mathbb{M}^{Z}:=(\Omega, (X_t)_{t\geq 0}, ({\cal{F}}_t)_{t\geq 0},(P_x)_{x\in Z_{\Delta}})$ is a Hunt process with respect to both the $\rho$-topology and the original topology.
\begin{rem}\label{fin}
Let $\mathbb{M}$ be the trivial extension of $\mathbb{M}^Z$ to $E_{\Delta}$ (see IV. (3.48) in Ref. \cite{mr}, or IV. (2.18) in Ref. \cite{St1}). Then
$\mathbb{M}$ is again a Hunt process and strictly
properly associated in the resolvent sense with $({\cal{E}},{\cal{F}})$ by (R2) and (\ref{atd}). The Hunt process $\mathbb{M}$ is unique up to the equivalence
described in IV.6.3 of Ref. \cite{mr}. In this sense $\mathbb{M}$ is the same process as the one constructed in Theorem 3 of Ref. \cite{Tr5} under the condition {\bf SD3}.
\end{rem}
\begin{theo}\label{hunt2}
Let ${\cal{E}}$ be a strictly quasi-regular generalized Dirichlet form satisfying {\bf D3}. Then there exists a
strictly $m$-tight Hunt process which is strictly properly associated in the resolvent sense with ${\cal{E}}$.
\end{theo}
\begin{proof}
For the existence of the Hunt process which is strictly properly associated in the resolvent sense with ${\cal{E}}$ see Remark \ref{fin}.
The $m$-tightness is a direct consequence of the existence of a strict ${\cal{E}}$-nest like in Definition 2(i) of Ref. \cite{Tr5} and the representation of the capacity
from Lemma 1(i) in Ref. \cite{Tr5}.
\end{proof}
|
\section{Introduction}
The cooperon-fermion model (for a review, see Ref. \onlinecite{Chen-PR-05}) is a basic model for superconductivity that has widely been adopted to explain the BCS-BEC crossover in ultra-cold fermionic atomic gases \cite{Chen-PR-05, Holland-PRL-01, Ohashi-PRL-02} and high-$T_c$ superconductivity.\cite{Chen-PR-05, Minca-RMP-90, Friedberg-PRB-1989, Domanski-arxiv-03, Dima-97, Micnas-PRB-07, Ranning-PC-95,Neto-PRB-01, Perali-PRB-00, Altman-PRB-02, Tsai-PRB-06, Alexandrov-04}
The resonantly paired fermions, or cooperons, in this model can either be locally bound pairs of small polarons due to extremely strong electron-phonon coupling,\cite{Robaszkiewicz-PRB-87} or localized Cooper pairs due to strong local pairing as might be the case in high-$T_c$ superconductors \cite{Altman-PRB-02}, or molecular bosons in ultra-cold atoms.\cite{Chen-PR-05, Holland-PRL-01, Ohashi-PRL-02} The potential existence of finite energy cooperons with a local attraction has also been put forward a few years ago in a simple semiconductor system.\cite{Nozieres-EPJB-99}
One recent work based on the four-leg ladder Hubbard model \cite{KRT10} as well as earlier studies in the quasi-2D ladder Hubbard model \cite{LeHur-Review} reveals the important role of the cooperon excitations in the transition from insulating state to superconducting state.
In cuprates, the interplay between the finite energy cooperon excitations around the antinode and the truncated Fermi surface around the node has recently been proposed as a possible mechanism for the superconductivity in a two-gap scenario.\cite{YRZ, RYZ-Review}
The dominant underlying mechanism for driving superconductivity in this model is the scattering two electrons involving a virtual cooperon $c_{\uparrow, \boldsymbol{k}} + c_{\downarrow, -\boldsymbol{k}} \xrightarrow{b_{\boldsymbol{k}=0}} c_{\uparrow, \boldsymbol{k}'} + c_{\downarrow, -\boldsymbol{k}'}$.
Compared with the attractive Hubbard model, the cooperon-fermion model has much richer physics since it has the complete dynamical information of the interaction between two fermions and the delocalization of cooperons with decreasing temperature.
So far most work on this model is done at the mean field level, with either the T-matrix method or various other methods going beyond simple mean field theory by including more diagrams \cite{Chen-PR-05} but there have been few unbiased calculations.
A recent exact diagonalization study of this model has been limited by small sizes and special geometries. \cite{ED-BF} A direct quantum Monte Carlo simulation usually suffers from a sign problem,\cite{sign-problem} except for certain models and algorithms, such as determinant quantum Monte Carlo simulations of the attractive Hubbard model.
In this paper, we simulate the two-dimensional cooperon-fermion model using a continuous-time diagrammatic determinant quantum Monte Carlo method (DDQMC).
The Hamiltonian has the form:
\begin{equation}
H= \sum_{\sigma,\boldsymbol{k}} (c^{\dag}_{\boldsymbol{k},\sigma}\epsilon^{f}_{\boldsymbol{k}} c_{\boldsymbol{k},\sigma} +b^{\dag}_{\boldsymbol{k}}\epsilon^{b}_{\boldsymbol{k}} b_{\boldsymbol{k}})
+ U\sum_{i} ( c^{\dag}_{i, \uparrow}c^{\dag}_{i,\downarrow}b_{i} + h.c. )
\label{eq:H}
\end{equation}
where $c^{\dag}_{\sigma} (c_{\sigma})$ is the fermionic creation (annihilation) operator with spin $\sigma = \{\uparrow, \downarrow \}$ and $b^{\dag} (b)$ is the bosonic creation (annihilation) operator of a cooperon.
The interaction term $U$ leads to the $s$-wave pairing of fermions mediated by the originally localized cooperons with a band gap $\Delta$ and the delocalization of cooperons at low temperatures.
The bare dispersions are $\epsilon_{\boldsymbol{k}}^{f}$ = $2t_{f}[2- \cos(k_{x}) - \cos(k_{y}) ] + \mu$, and $\epsilon_{k}^{b} = 2t_{b}(2- \cos(k_{x}) - \cos(k_{y})) + \Delta$. It is a trivial generalization to include attractive interactions between the fermions.
By integrating out the bosonic degrees of freedom, an effective action for the fermionic part can be obtained, which has a form similar to the attractive Hubbard model but with full dynamic properties:
\begin{equation}
\begin{split}
&S^{\rm eff}_{f}(\bar{\psi}_{\sigma}, \psi_{\sigma}) = \int _{0}^{\beta} d\tau \sum_{\boldsymbol{k},\sigma} \bar{\psi}_{\boldsymbol{k}, \sigma} (\partial_{\tau} + \epsilon^{f}_{\boldsymbol{k}}) \psi_{\boldsymbol{k},\sigma} \\
&\!\!+ U^{2} \int_{0}^{\beta} \int_{0}^{\beta} d\tau d\tau ^{\prime} \sum_{i, i^{\prime}} \bar{\psi}_{i,\tau,\uparrow} \bar{\psi}_{i,\tau,\downarrow} G^{b,0}_{r_{i}-r_{i\prime},\tau-\tau^{\prime}} {\psi}_{i^{\prime},\tau^{\prime},\downarrow} {\psi}_{i^{\prime},\tau^{\prime},\uparrow}
\end{split}
\end{equation}
where $G^{b,0}_{r_{i} -r_{i\prime},\tau-\tau^{\prime}}$ is the bare bosonic Green's function.
The attractive Hubbard model can be obtained by approximating $U^{2} G^{b,0}_{r_{i}-r_{i\prime},\tau-\tau^{\prime}} =U_{H} \delta_{i, i\prime} \delta_{\tau, \tau\prime}$ with $U_{H} \sim -U^{2}/\Delta_{\rm eff}$ ($\Delta_{\rm eff}$ is the renormalized band gap of the cooperons) including the dominant contribution from the renormalized $\boldsymbol{k}=0$ cooperon.
We develop a continuous-time DDQMC algorithm, similar to the algorithms for an attractive Hubbard model\cite{Burovski-NJP-06, Burovski-PRL-08} for the cooperon-fermion model, and obtain the numerically exact solution to the model at the filling value $n=0.12$. The Kosterlitz-Thouless (KT) transition temperature is estimated from the finite-size scaling of the fermion pair correlation function and the cooperon Green's function. The renormalization of the cooperon band characterized by the effective gap $\Delta_{\rm eff}$ and the effective mass $m_{\rm eff}$ are examined carefully.
Applying these results to study the strong diamagnetism recently observed in cuprates, we find that the renormalization of the cooperon band will enhance the diamagnetism dramatically at low temperatures, which qualitatively agrees with the experimental data. \cite{Lu-epl}
\section{The algorithm}
\begin{figure}[tbf]
\centerline
{
\includegraphics[width = 6.0cm, height =5.0cm, angle= 0]
{configuration.eps}
}
\caption[]
{(Color online.) (a): elementary interaction events (vertices): two fermions are combined onto a cooperon (left) and vice versa (right). (b): a diagram contributing to the partition function $Z$.
(c): a diagram sampled in the DDQMC algorithm: open ends of the fermionic lines imply a sum over all the ways of connecting the vertices by the fermionic lines, which is represented by the corresponding determinant. }
\label{fig:event}
\end{figure}
Continuous-time DDQMC algorithms have been applied successfully in the past to obtain the critical temperature in the BCS-BEC crossover\cite{Burovski-NJP-06, Burovski-PRL-08}
The sign problem in these algorithms is avoided by collecting all Feynman diagrams with the same distribution of vertices as a single configuration which turns out to have a positive-definitive weight.
This approach can also be applied to the cooperon-fermion model eliminating the severe sign problem coming from the permutations of the fermionic lines.
In the interaction picture, the partition function for the model (\ref{eq:H}) can be expressed as
\begin{equation}
\begin{split}
Z &=
{\rm Tr} \Bigl[e^{-\beta (H_{f,\uparrow}^{0} + H_{f,\downarrow}^{0})} e^{-\beta H_{b}^{0}} \sum_{n} \Big[\int_{0}^{\beta} \int \Big]^{n} \\
&\cdot\frac{1}{n!} \prod_{i=1,n} dr_{i} d\tau_{i} \mathcal{T}_{\tau}( -U
c^{\dag}_{r_{i},\tau_{i},\uparrow}c^{\dag}_{r_{i},\tau_{i},\downarrow}b_{r_{i},\tau_{i}} + h.c.)\Bigr]
\end{split} \label{EQ:Z}
\end{equation}
where $H_{f, \sigma}^{0}$ in the Hamiltonian of free fermions with the spin $\sigma$ and $H_{b}^{0}$ is the Hamiltonian for free cooperons, the bilinear in $c^{\dag}_{\boldsymbol{k},\sigma}$, $c_{\boldsymbol{k},\sigma}$ and $b^{\dag}_{\boldsymbol{k}}$, $b_{\boldsymbol{k}}$ terms respectively in Eq.~\eqref{eq:H}. To describe the Feynman diagrams, we define two different kinds of events, shown in Fig.~\ref{fig:event}(a), representing the process of combining two fermions with opposite spins into one cooperon and the reverse process. For a typical Feynman diagram like the one shown in Fig.~\ref{fig:event}(b), the vertices are connected by the bare single particle propagators
\begin{align}
G^{f,0}_{r^{a}_{i}-r^{c}_{j}, \tau^{a}_{i}-\tau^{c}_{j}} &= - {\rm Tr}\Bigl[ e^{-\beta H_{f,\sigma}^{0}} \mathcal{T}_{\tau}c_{r^{a}_{i},\tau^{a}_{i},\sigma} c^{\dag}_{r^{c}_{j},\tau^{c}_{j},\sigma} \Bigr] \\
G^{b,0}_{r^{c}_{i}-r^{a}_{j}, \tau^{c}_{i}-\tau^{a}_{j}} &= - {\rm Tr}\Bigl[ e^{-\beta H_{b}^{0}} b_{r^{c}_{i},\tau^{c}_{i} } b^{\dag}_{r^{a}_{j},\tau^{a}_{j}} \Bigr]
\end{align}
The superscript $a/c$ on $r_{i,j}$ are for the events with annihilation and creation of a pair of fermions (in accompany of the creation and annihilation of a cooperon), respectively.
By applying the Wick's theorem, the partition function can be rewritten as
\begin{equation}
\begin{split}
Z =& \sum_{n} U^{2n} \Big[ \int _{0}^{\beta} \int \Big]^{2n}\frac{1}{n!n!} \Bigl[ \prod_{i=1,n} d^{D}r^{c}_{i}d^{D}r^{a}_{i} d\tau^{c}_{i}d\tau^{a}_{i} \Bigr] \\
& \cdot\det A_{{S}_{n}^{'},\uparrow} \det A_{{S}_{n}^{'},\downarrow} \text{Perm}(B_{{S}_{n}^{'}})
\end{split}
\end{equation}
where $S_{n}^{'}$ represents the configuration including all possible ways of connecting a specific distribution of vertices with the propagator lines. The matrix components $[A_{S_{n}^{'},\sigma}]^{i,j} = G^{f,0}_{r^{a}_{i}-r^{c}_{j}, \tau^{a}_{i}-\tau^{c}_{j}}$, and $[B_{S_{n}^{'}}]^{i,j} = - G^{b,0}_{r^{c}_{i}-r^{a}_{j}, \tau^{c}_{i}-\tau^{a}_{j}}$.
The determinant of the matrix $A$ comes from the anti-commutation relation of fermions, while the permanent Perm$(B_{S_{n}^{'}})$ originates in the commutation relation between cooperons. In contrast to determinants, which can be efficiently evaluated, the calculation of a permanent is an exponentially hard problem. Thus, we evaluate the permanent by individually sampling all permutations of the bosonic lines. A
typical diagram $\widetilde{S}_{n}$ encountered in the Monte Carlo sampling is shown in Fig.~\ref{fig:event}(c). The open ends of fermionic lines indicate that all possible connection ways of the fermion lines are summed up in the determinant, while the connection between the cooperon lines is fixed, indicating that the specific connections are sampled individually. Summing the fermion lines into a determinant completely removes the fermionic sign problem. However, sampling the permanent gives us a small remaining sign problem, which is tractable since the distribution of values in $B_{S_{n}^{'}}$ is dominantly positive.
The weight of a configuration $\widetilde{S}_{n}$ is
\begin{equation}
Z_{\widetilde{S}_{n}} =\frac{U^{2n} } {( L^{2} \beta )^{2n}} \det A_{\widetilde{S}_{n},\uparrow} \det A_{\widetilde{S}_{n},\downarrow} \prod_{i=1,n} B^{i,\mathcal{P}_{i}}_{\widetilde{S}_{n}},
\label{eq:Z}
\end{equation}
where $L^{2}$ is the spatial size of the system with periodic boundary conditions. Thermal averages are calculated by sampling all possible configurations $\widetilde{S}_{n}$.
\begin{figure}[tbh]
\centerline
{
\includegraphics[width = 7.5cm, height =4.0cm, angle= 0]
{update.eps}
}
\caption[]
{(Color online.) Two sets of complementary Monte Carlo updates: (a) adding or removing one pair of vertices, and (b) swapping the end points of cooperon lines.}
\label{fig:update}
\end{figure}
Our algorithm implements three different Monte Carlo updates: creating or deleting one pair of vertices which changes the order from $n$ to $n\pm1$, and swapping the connection of cooperon lines as shown in Fig.~\ref{fig:update}. In order to improve the efficiency of sampling we pick a pair of times with a probability proportional to the cooperon's bare Green's function $G^{b,0}_{r,\tau}$.
Adding a pair of vertices to go from configuration $\widetilde{S}_{n}$ to $ \widetilde{S}_{n+1}$ is accepted with an acceptance ratio of
\begin{align}
R_{\rm add} = \min \bigg( 1,&U^{2} \frac{ \left| \det A_{\widetilde{S}_{n+1},\uparrow} \right| ^{2}} {\left| \det A_{\widetilde{S}_{n},\uparrow} \right | ^{2}} \left| B_{\widetilde{S}_{n+1}}^{n+1,\mathcal{P}_{n+1}} \right| \nonumber \\
& \frac{\beta L^{2}} {n+1} \frac{\int_{0}^{\beta} d\tau \sum_{r} G^{b,0}_{r,\tau}} {G^{b,0}_{r,\tau}}
\bigg),
\label{eq:R}
\end{align}
and a corresponding equation for the removal.
For the self-complementary process of swapping the connection of cooperon Greens functions
the acceptance ratio is
\begin{equation}
R_{\rm flip} =\min \left(1, {\left| \prod_{i} B^{i,\mathcal{P}(i)}_{\widetilde{S}_{\mathcal{P}(n)}} \right|} \Big/ {\left| \prod_{i} B^{i,\mathcal{P'}(i)}_{\widetilde{S}_{\mathcal{P'}(n)}} \right|} \right)
\end{equation}
The fermionic and cooperon Green's functions can be measured as
\begin{align}
G^{f}_{R-R^{\prime}, \tau-\tau^{\prime},\sigma} &= \left\langle{\det \widetilde{A}_{\widetilde{S}_{n},\sigma}} / {\det {A}_{\widetilde{S}_{n},\sigma}}
\right\rangle _{\rm MC} \label{eq:Gf}\\
G^{b}_{R-R^{\prime}, \tau-\tau^{\prime}} &= \Bigl\langle G^{b,0}_{R-R^{\prime}, \tau-\tau^{\prime}} \label{eq:Gb} \\
&+\sum_{l} \frac{G^{b,0}_{R-r^{a}_{l}, \tau - \tau^{a}_{l}} G^{b,0}_{r^{c}_{l}-R^{\prime}, \tau^{c}_{l}-\tau^{\prime}} }{G^{b,0}_{r^{c}_{l}-r^{a}_{l},\tau^{c}_{l}-\tau^{c}_{l}}} \Bigr\rangle_{\rm MC} \notag
\end{align}
The matrix $\widetilde{A}_{\widetilde{S}_{n}, \sigma}$ is an $(n+1) \times (n+1)$ matrix extending $A$ by an extra column and row corresponding to the open vertices $ c_{(R^{\prime},\tau^{\prime}) }^{\dag}$ and $ c_{(R, \tau)}$. The notation $\langle .. \rangle_{\rm MC}$ denotes the Monte Carlo average. The particle-particle correlation function of fermions $\langle \mathcal{T}_{\tau} c_{R,\tau,\downarrow}c_{R,\tau,\uparrow} c^{\dag}_{R^{\prime}, \tau^{\prime}, \uparrow}c^{\dag}_{R^{\prime}, \tau^{\prime}, \downarrow} \rangle$ is measured as
\begin{equation}
\chi^{\rm pp}_{R-R^{\prime}, \tau-\tau^{\prime}} = \left\langle \left| \frac{\det \widetilde{A}_{\widetilde{S}_{n},\sigma}}{\det {A}_{\widetilde{S}_{n},\sigma}} \right|^{2} \right\rangle _{\rm MC}
\end{equation}
with the double occupancy characterizing the local pairing strength $\langle n_{\uparrow} n_{\downarrow} \rangle $
=$\chi^{\rm pp}_{R=(0,0), \tau=0^{-}}$.
The vertex correlation contribution to the particle-particle correlation of fermions, which indicates the formation of coherent cooper pairs, is defined as
\begin{equation}
\begin{split}
\chi^{\rm od}_{R-R^{\prime}, \tau-\tau^{\prime}} &=\left\langle \mathcal{T}_{\tau} c_{R,\tau,\downarrow}c_{R,\tau,\uparrow} c^{\dag}_{R^{\prime}, \tau^{\prime}, \uparrow}c^{\dag}_{R^{\prime}, \tau^{\prime}, \downarrow} \right \rangle \\
&- \left\langle \mathcal{T}_{\tau} c_{R,\tau,\uparrow} c^{\dag}_{R^{\prime}, \tau^{\prime}, \uparrow} \right\rangle \left \langle \mathcal{T}_{\tau}c_{R,\tau,\downarrow}c^{\dag}_{R^{\prime}, \tau^{\prime}, \downarrow} \right\rangle
\label{eq:chi-od}
\end{split}
\end{equation}
The cooperon Green's function and $\chi^{od}$ in the long wave-length and static limit for our finite system with periodic boundary conditions can be defined as $G^{b}_{\boldsymbol{k}=0,\omega =0} = \int_{0}^{\beta} d\tau \sum_{r} G^{b}_{r,\tau}$, and $\chi^{od}_{\boldsymbol{k}=0,\omega =0} = \int_{0}^{\beta} d\tau \sum_{r} \chi^{od}_{r,\tau}$ with $r$ being the distance confined by system size.
\begin{figure}[tb]
\centerline
{
\includegraphics[width = 7.5cm, height =2.0cm, angle= 0]
{RPA.eps}
}
\caption[]
{(Color online.) The Feynman diagrams comprising the fermionic (a) and cooperon (b) Green's function within the random phase approximation (RPA) .}
\label{fig:RPA}
\end{figure}
In addition to the DDQMC results, we also show results of random phase approximation (RPA) calculations which only take into account the contribution from a truncated set of Feynman diagrams, whose diagrammatic representations are shown in Fig.~\ref{fig:RPA}. The cooperon and fermionic Green's functions have the general form
\begin{align}
[G^{f/b}_{\boldsymbol{k},\omega}]^{-1}
= [G^{f/b,0}_{\boldsymbol{k},\omega}]^{-1}
- \Sigma^{f/b}_{\boldsymbol{k},\omega}
\end{align}
with the self energies $\Sigma^{f/b}_{\boldsymbol{k},\omega}$ estimated by only including ladder diagrams.
RPA works well at weak coupling and high temperature region, and is useful as a test for our numerical solution in that limit.
\section{Main results}
For our simulations we choose $U=1$ as the unit of energy and set $t^{f} = 1$, $t^{b} = 0.5$. The bare cooperon band lies above the bottom of the fermionic band with the offset $\Delta=0.75$ and the total charge density is fixed at $n^{f}_{\uparrow} + n^{f}_{\downarrow}+ 2n_{b}=0.12\pm0.002$.
In the parameter regime we are interested in, the chemical potential $\mu$ is around $0.22-0.45$, resulting in the effective renormalized gap $\Delta_{\rm eff} \le 0.3$ and the corresponding effective attractive Hubbard interaction $|U_{H}| \sim U^{2}/\Delta_{\rm eff} > 3$ at $\beta > 3$. For the finite-size scaling analysis, we use the set of linear system sizes $L = 11, 15, 21, 25$. The expectation values and the error bars are obtained from 96 independent sampling processes with different random number seeds. Each measurement is made after 2-3 autocorrelation times, {\it i.e.} around one measurement per 100$L^{2}$ Monte Carlo steps at high temperatures ($\beta \sim 3$), and 3000-5000 $L^{2}$ steps for low temperatures ($\beta \sim17$).
Fig.~\ref{fig:mu}(a) shows the temperature dependence of the chemical potential $\mu$. The solid curve is the RPA result with a mean-field critical temperature $T_{c}^{MF} \sim 0.09$ characterized by the closure of the effective gap $\Delta_{\rm eff}$.
Blue diamonds are the chemical potential at thermodynamic limit determined by DDQMC.
RPA generally overestimates the interplay between fermions and cooperons at low temperatures due to the logarithmic divergence leading to a larger value of the cooperon's self-energy from the simple particle-particle bubble diagram.
As a consequence, the RPA chemical potential is lower than the exact value in this regime.
However, in the high-temperature limit, the RPA calculation provides a consistency check for DDQMC and there we observe a perfect agreement.
Fig.~\ref{fig:mu}(b) shows the particle density $n_{\sigma}$ and $n_{b}$ at thermodynamic limit.
The cease of the suppression of $n_{b}$ as the temperature is decreased, i.e. the flattening out of $n_{b}$ v.s. $T$ at low $T(<0.1)$, indicates an approach to the KT transition.
\begin{figure}[tb]
\centerline
{
\includegraphics[width = 5.5cm, height =9.0cm, angle= 270]
{mu_n.eps}
}
\caption[]
{(Color online.) (a): the chemical potential versus temperature $T/U$.
(b): the particle density $n_{\sigma}$ of a single spin component and the cooperon density $n_{b}$ versus temperature $T/U$. DDQMC results in thermodynamic limit (blue points) are extrapolated to the infinite system size.}
\label{fig:mu}
\end{figure}
\begin{figure}[tbh]
\centerline
{
\includegraphics[width = 5.5cm, height =9.0cm, angle= 270]
{double_n.eps}
}
\caption[]
{(Color online.) The system size dependence of the onsite double occupancy $n_{d}$ for various values of $\beta$. }
\label{fig:double-n}
\end{figure}
As seen from Fig.~\ref{fig:double-n}, the double occupancy $n_{d} =\langle n_{i,\uparrow}n_{i, \downarrow} \rangle \gg \langle n_{i,\uparrow} \rangle \langle n_{i, \downarrow} \rangle$ reveals a strong on-site pairing mediated by cooperons.
The behavior of $n_{d}$ is determined by two aspects:
(i) the competition between the potential energy gain from pairing and the corresponding kinetic energy loss, and (ii) the balance between the particle number of fermions and cooperons. However, the latter is not expected to play an important role at low temperatures due to the plateau in $n_{b}(T)$.
The monotonous increase of $n_{d}$ with lowering the temperature is analogous to recent DMFT results for the attractive Hubbard model. \cite{Schollwock-prl-01}
We also note that a different low-$T$ behavior of $n_{d}$ for the attractive Hubbard model has been reported in early QMC studies. \cite{Singer-PRB-96}
\begin{figure}[b]
\centerline
{
\includegraphics[width = 5.5cm, height =9.0cm, angle= 270]
{chi_pp_scaling.eps}
}
\caption[]
{(Color online.) (a): system-size dependence of off-diagonal order $\chi^{\rm od}_{R_{x,{\rm max}},\omega =0}$ at distance $R_{x,{\rm max}}$. (b): the cooperon Green's function $G^{b}_{R_{x,{\rm max}}, \omega =0}$. $R_{x,{\rm max}} = (L-1)/2$ is the maximal distance along the $x$ direction in our system with periodic boundary conditions.
}
\label{fig:scaling}
\end{figure}
The 2D superconducting state is characterized by algebraically decaying off-diagonal order \cite{CNYang} in $\chi^{od}$. Figure ~\ref{fig:scaling}(a) shows $\chi^{\rm od}_{R_{x,{\rm max}},\omega =0}$ for different system sizes, where $R_{x,{\rm max}} = (L-1)/2$ is the maximum distance in the $x$ direction on the lattice with periodic boundary conditions. We also plot the size dependence of the cooperon Green's function $G^{b}_{R_{x,{\rm max}},\omega=0}$ in Fig.~\ref{fig:scaling}(b), which behaves very similarly to $\chi^{\rm od}_{R_{x,{\rm max}},\omega =0}$. Both grow substantially with decreasing the temperature, indicating that the Cooper pairs and cooperons become coherent.
\begin{figure}[tbh]
\centerline
{
\includegraphics[width = 5.5cm, height =9cm, angle= 270]
{GB_k0w0_L_scaling_RG.eps}
}
\caption[]
{(Color online.) Finite-size scaling of $\chi^{\rm od}_{\boldsymbol{k}=0, \omega=0}$ and $G^{b}_{\boldsymbol{k}=0, \omega=0}$ according to Eq.~(\ref{eq:scaling}); the error bars are smaller than the symbol size.
The uncertainties of $A=0.55\pm0.1$ and $T_{c}=0.03\pm0.01$ are estimated from a breakdown of the shown data collapse.}
\label{fig:GBk0w0-scale}
\end{figure}
We next perform a finite-size scaling analysis for the pair correlator $\chi^{\rm od}_{\boldsymbol{k}=0, \omega=0}$ and the cooperon Green's function $G^{b}_{\boldsymbol{k}=0, \omega=0}$. For $T_{c} > T > 0$, one expects $\chi^{\rm od}_{r,\omega=0}$ to exhibit a power-law decay with an exponent $\eta(T)$, such that $\eta(T_{c}) = 1/4$ and $\eta(T=0)=0$ indicating the emergence of the true log-range order at $T=0$. Above $T_{c}$, $\chi^{\rm od}_{r,\omega=0}$ shows an exponential decay. The pair correlator at $T>T_c$ should obey the scaling formula \cite{Scalapino-PRB-91}
\begin{equation}
\chi^{\rm od} = L^{2-\eta(T_{c})} f(L/\xi^{f}) \mbox{, for $L \gg 1, T \to T^{+}_{c}$}
\label{eq:scaling}
\end{equation}
with $\xi^{f} \sim e^{A/\sqrt{T-T_{c}}}$.
Since the critical behavior in both subsystems of cooperons and fermions is a manifestation of one and the same superfluid transition, $G^{b}_{\boldsymbol{k}=0, \omega =0}$ is also supposed to exhibit the scaling given by Eq.~\eqref{eq:scaling}. The parameters $A$ and $T_c$ are chosen so that measurements of
$\chi^{\rm od}_{\boldsymbol{k}=0, \omega =0}$ and $G^{b}_{\boldsymbol{k}=0, \omega =0}$ for different system sizes collapse in the vicinity of the phase transition as shown in Figs.~\ref{fig:GBk0w0-scale}, resulting in $A=0.55\pm0.1$ and $T_{c}=0.03\pm0.01$. The uncertainties in $A$ and $T_{c}$ are estimated from observing a noticeable distortion of the data from a single smooth curve as the parameters are varied beyond the claimed error bars.
The most important properties of cooperons are their effective band gap $\Delta_{\rm eff}$ and effective mass $m_{{\rm eff}}$ renormalized by interactions mediated by fermions. As we shall discuss in more details in the next section, these parameters will allow us to obtain an estimate of the diamagnetic susceptibility, which is expected to rise dramatically due to the quasi-condensation on approach to $T_c$.
Figure~\ref{fig:scale}(a) shows the dependence of $\Delta_{\rm eff}$, which is obtained from the cooperon Green's function according to $\Delta_{\rm eff} = - [G^{b}_{\boldsymbol{k}=0,\omega=0}]^{-1}$, on the linear system size.
The value $\Delta_{{\rm eff}, L\to \infty}$ in the thermodynamic limit, obtained from the extrapolation in the system size, is shown in Fig.~\ref{fig:scale}(b). At $T>0.1$, the gap $\Delta_{{\rm eff},L \to \infty}$ is quite close to the RPA estimate. Due to the logarithmic divergence of the bare particle-particle bubble, RPA leads to a substantially higher mean-field critical temperature $T^{MF}_{c} \sim 0.09$.
\begin{figure}[tbh]
\centerline
{
\includegraphics[width = 5.5cm, height =9cm, angle= 270]
{gap_scale.eps}
}
\caption[]
{(a) (Color online.) Finite size extrapolation of the renormalized cooperon gap $\Delta_{{\rm eff}}$. (b) Comparison of the extrapolated $\Delta_{{\rm eff}}$ at $L \to \infty$ with the RPA results. $T_{c} = 0.03$ obtained in Fig.~\ref{fig:GBk0w0-scale} is shown by the arrow. The error bars are smaller than symbol sizes.}
\label{fig:scale}
\end{figure}
The effective cooperon mass is calculated as
\begin{eqnarray}
m^{-1}_{{\rm eff}} &=& \frac{\sum_{r}G^{b}_{r, \omega =0} r^{2} } {[G^{b}_{\boldsymbol{k}=0, \omega=0}]^{2}},
\label{eq:mass}
\end{eqnarray}
where the factor of $r^{2}$ in the sum shows the importance of the long-range behaviour of the Green's function in determining $m_{{\rm eff}}$. Since the finite size of the system along with the periodic boundary conditions will enhance $G^{b}_{r, \tau}$ at large $r$, the straightforward evaluation of $m_{{\rm eff}}$ using Eq.~\eqref{eq:mass} is inadequate.
To get rid of the finite-size effects, we fit the measured $G^{b}_{r,\tau}$ according to
\begin{eqnarray}
G^{b}_{r,\tau} = \sum_{m,n} \mathcal{G}^{b}_{(r_{x} +mL, r_{y}+nL), \tau}
\label{eq:fit}
\end{eqnarray}
with $\mathcal{G}^{b}_{r,\tau} = a e^{-b |r|}$ for large $|r|$ (both $a$ and $b$ depend on $\tau$). The function $\mathcal{G}^{b}_{r,\tau}$ obtained thereby is then used instead of $G^{b}_{r,\tau}$ in Eq.~\eqref{eq:mass}.
The result obtained using the data for $L=11, 15$ is shown in Fig.~\ref{fig:dispersion}(b). The RPA calculation gives similar values at $T >0.15$. At lower temperatures, the effective mass continuously decreases and tends to a finite value. Fits to Eq.~\eqref{eq:mass} allow us to estimate the renormalized dispersion of the cooperon band $\epsilon^{b}_{\boldsymbol{k},eff}$ as shown in Fig.~\ref{fig:dispersion} using the Green's functions for the case $L=11$. For comparison, the RPA curve $\epsilon^{b}_{\boldsymbol{k},eff}$ is also shown in Fig.~\ref{fig:dispersion}, which agrees with the Monte Carlo results at high temperatures.
\begin{figure}[tbh]
\centerline
{
\includegraphics[width = 5.5cm, height =9.0cm, angle= 270]
{L11_dispersion.eps}
}
\caption[]
{
(Color online.) (a): The suppression of the renormalized cooperon mass with decreasing the temperature. Red curve corresponds to the RPA result. The blue (black) dots are for lattice size $L = 11, 15$. The same fitting process fails for larger lattice sizes $L = 21, 25$ due to the too small measured $G^{b}_{r,\tau}$ close to the lattice boundary and very large sampling error bars.
(b):
Comparison of the renormalized cooperon's dispersion $\epsilon^{b}_{\boldsymbol{k},eff}$ from RPA calculation and our simulation with $L=11$.
The fitting process (Eq.~\eqref{eq:fit}) is applied to obtain $\epsilon^{b}_{\boldsymbol{k},eff}$ in our simulation.
The error bars are smaller than the symbol.
}
\label{fig:dispersion}
\end{figure}
\section{Application to the strong diamagnetism in the cuprates}
Lots of anomalous properties of the pseudogap phase have been reported since the early stage of high-$T_{c}$ studies, for instance the existence of the partial gap, the linear resistivity, and the proportionality of the charge carrier density to doping concentration.\cite{RYZ-Review, pseudogap-review}
Strong superconducting fluctuations have been observed in a large temperature region in recent Nernst and torque magnetometry measurements.\cite{Lu, yayu} In contrast to conventional BCS superconductors, where the Gaussian (amplitude) fluctuations are dominant and the pairing length is quite long, in the cuprates the fluctuations of the phase rigidity are predominant while the cooper pairs are strongly bound with the energy scale around the spin-spin superexchange $J\sim 100$ meV and they localized in a small spatial area with $\xi \sim 3-4$ lattice constant (a = 3.8 \AA).
In momentum space the pseudogap phase is highly anisotropic. More and more evidence shows that the states at the antinode and node are intrinsically different \cite{Hufner-RPP-08}.
In the pseudogap phase, the single particle gap is partially opened only around the antinode, leaving either ``arcs" or hole-like Fermi pockets around the nodes.\cite{Norman-98, yang-nature-07,XJZhou-09} Recent angular-resolved photoemission spectroscopy (ARPES) experiments have observed the existence of such pocket and unmask the particle-hole symmetry of the spectrum around the antinode, and the particle-hole asymmetry around the node.\cite{yang-nature-07} Evidence from scanning tunneling microscopy (STM) reveals that the low-energy states around the nodes are homogenous and of long range correlation length, and quantum interference STM observed well-defined Fermi surface only inside AF reduce Brillouin zone \cite{Kohsaka-nature-08}. Meanwhile the high-energy states around the antinodes are inhomogeneous with short correlation length around four lattice constants.
Some evidence of particle-hole symmetry in ARPES \cite{yang-nature-07} around the antinodes is consistent with the functional renormalization group calculations,\cite{Honerkamp-RG} showing that the strong umklapp scattering enhances the cooperon channel.
Some possible scenarios leading to superconductivity by the finite energy cooperon excitations and fermion sea have been proposed from the Hubbard mode on a ladder \cite{KRT10, LeHur-Review} and semiconductors,\cite{Nozieres-EPJB-99} which may shed some light on the case for cuprates.\cite{RYZ-Review} So far the possible interplay between the states residing on the node and the antinode is still an open question. The superconducting gap on the Fermi surface around the node in SC state may be driven by this effect.
\begin{figure}[tbh]
\includegraphics[width=5.5cm,height=6cm, angle=0]
{cuprate.eps}
\caption{(Color online) Anisotropy in momentum space in the pseudogap phase: localized tightly bound cooperons are located at the antinodes and a hole-like Fermi sea resides on the nodes.
}
\label{fig:BF1}
\end{figure}
\begin{figure}[b]
\centering
\includegraphics[width=5.5cm,height=7.0cm, angle=0]
{diamagentism.eps}
\caption{(Color online.) The temperature dependence of the magnetization at $B=1$ T calculated using Eq.~\eqref{eq:chi_b}.The $x$-axis is normalized by $T_{c}\sim 0.03$ ($\sim80$ K with $U=250$ meV). Red dots are from the torque magnetization experiments with $T_c=50{\rm K}$.\cite{Lu} Note that there is no parametric fitting to the experimental data. While strong damping of cooperons at high temperature suppresses the magnetization at $T \gg T_{c}$, the exponential increase close to $T_{c}$ will not be affected.}
\label{fig:BF}
\end{figure}
It is reasonable to treat the states around the node as a free Fermi gas and the component around the antinode as tightly bound cooperons with a pairing gap of the order of $J\sim 100$ meV as sketched in Fig~\ref{fig:BF1}.
Our model \eqref{eq:H} can thus be a simplified picture for the cuprate pseudogap phase by ignoring the multi-patch structure and the $d$-wave phase of the momentum-space localized cooperons. There has been a lot of previous work attempting to use this model for the phenomenology of the pseudogap phase (for a review see Ref. \onlinecite{Chen-PR-05}). However most of these are based on mean field calculations, but an exact solution is still missing. Our numerical simulations can give useful quantitative insight. Here we will focus on the strong diamagnetism observed recently. From the recent phenomenological YRZ theory,\cite{YRZ} it is reasonable to further propose that the fermionic particles around the node are itinerant holes, and the cooperons around the antinode are tightly bound hole pairs. To describe the underdoped phase we assume the total charge carrier density is around 0.12, which is obtained by a value of $\mu \sim 100$meV (with $U \sim 250$meV), which is in a reasonable regime based on early analysis of the YRZ model \cite{yang-epl}. The bare hopping $t^{f}= 250$ meV is chosen comparable to the nearest neighbor hopping integral in cuprates.
In conventional BCS theory \cite{Tinkham} the contributions of the fermionic pairing fluctuation to diamagnetism are substantial only in a very narrow temperature region above $T_{c}$. In our model, however, the renormalized cooperon band contributes dominantly to the diamagnetism in a wide temperature range above $T_{c}$ in pseudogap phase. Since it is computationally extremely demanding to calculate the numeric value of the second order coefficient in $\boldsymbol{q}$ of the current-current correlation function $\mathcal{K}_{\boldsymbol{q},\omega=0}$, we approximate the diamagnetization of cooperons as that of free bosons with renormalized gap $\Delta_{\rm eff}$ in the limit $L\to \infty$ and the effective mass $m_{\rm eff}$ at $L=11$ (see Fig.\ref{fig:dispersion}). In
CGS units [$1/(4\pi)$] it has the form:\cite{May-PR-1959}
\begin{eqnarray}
\chi &=& \frac{(2e)^{2}} {c^{2}m_{{\rm eff}}d} \frac{n^{b}_{\boldsymbol{k}=0}} {6} \label{eq:chi_b},
\end{eqnarray}
where $d$=6$\AA$ is the interlayer distance for cuprates. The value $(2e)^{2}/(c^{2} 2m_{e}d)$ corresponds to $M$=$7.5 $ A/m at $H=1$ T.
The experimentally observed singular behavior $M(T,H) \sim -H^{1/\delta(T)}$ with $\delta \to 0$ at small $H$ and $T \to T_{c}^{+}$ might be related to the mesoscopic Meissner effect or the fragile Landon rigidity.\cite{Lu} Thermally excited vortices with exponentially increased inter-vortex length, however, are not sufficient to explain the experimental observation.
In Fig.~\ref{fig:BF} we show the diamagnetism at $H = 1$T with $M = \chi H$ and $\chi$ calculated from Eq.~\eqref{eq:chi_b}. Strong diamagnetism prevails in a very wide temperature region above $T_{c}$; this is in agreement with the experimental data (with $T_{c} = 50 K$). Note that even though the experimental data shows a much narrower temperature region, it is still orders of magnitude wider than that predicted by the conventional BCS theory. In the cooperon-fermion model, the strong damping of the cooperon at high temperature will suppress the diamagnetism greatly but will leave the exponential increase of diamagnetism at $T$ close to $T_{c}$ unchanged. We also note that an alternative explanation of the strong diamagnetism based on a vortex liquid picture has been proposed by Oganesyan \emph{et al.}. \cite{Oganesyan}
\section{Summary}
We have developed a continuous-time diagrammatic determinant quantum Monte Carlo algorithm for the cooperon-fermion model.
Our results for the fermionic part of the model show similar behavior to its twin model, the attractive Hubbard model, which is often used to describe the BCS-BEC crossover in the systems of ultra-cold atoms, where the cooperon-fermion model is the relevant model on the BEC side.
Besides the critical temperature we have calculated the renormalized band gap and mass of the cooperons. The decrease of the mass and the suppression of the renormalized gap have important effects on the thermodynamic properties of the cooperons. Applied to cuprate superconductors, the interplay between the cooperons at the antinode and the fermions at the node is expected to delocalize the cooperons and finally lead to a substantial enhancement of the diamagnetism in a wide temperature range. That could explain the strong diamagnetic signal observed recently in the underdoped state.
The numerical method developed here can be used to study the BCS-BEC crossover on lattices in the framework of the cooperon-fermion model, which gives direct access to the paring physics via the cooperon part. The universal results in terms of to the s-wave effective coupling between the fermions can in principle be obtained in the low-density limit, as was done, e.g., in Ref.~\cite{Burovski-NJP-06}.
We thank T. M. Rice, E. Burovski, M. Sigrist, B. Surer, E. Gull, P. N. Ma., L. Pollet, S. Pilati for discussions. We acknowledge financial support from the Swiss National Science Foundation and the NCCR MaNEP. Xin Wang acknowledge support from the Condensed Matter Theory Center of the University of Maryland.
We used the Brutus cluster at ETH Zurich for most of the simulations.
|
\section{Introduction}
The Tevatron and the LHC are becoming increasingly sensitive to
potential signals of a Standard Model (SM) Higgs boson. The highest
sensitivity is achieved for a Higgs boson of mass
\mbox{$m_h \sim 2 \, m_W$},
where the branching ratio of the Higgs decaying into a pair of $W$ bosons
is close to one.
The most significant production channel for these
searches is gluon fusion. This is a loop-mediated process, and in
the Standard Model the corresponding amplitudes are dominated by
top-quark loops. Gluon fusion may be sensitive to new coloured, heavy
particles
with a large coupling to the Higgs
boson. Limits on the Higgs production cross-section from ongoing
experimental searches are model-specific, and a dedicated theoretical
prediction to Standard Model extensions is often required.
A simple extension of the Standard Model may include a fourth family
of quarks and leptons. This model is theoretically constrained to not
very large fourth-generation quark masses, in order to preserve
perturbativity~\cite{Marciano:1989ns}.
To evade limits from precision electroweak
tests, the mass difference of the fourth-generation quarks is also
restricted to be small~\cite{Kribs:2007nz}. Finally, the fourth-generation neutrino
is required to have a mass greater than half the mass of the Z-boson
from the LEP limits on the
invisible Z-boson decay width.
One might find little theoretical appeal in a naive Standard Model with
four generations. As in the Standard Model, it suffers from the
hierarchy problem, and in addition it introduces a rather awkwardly
heavy neutrino. Nevertheless, its predictions are rather
spectacular in modifying the Higgs boson cross-section at hadron
colliders and can be tested easily with Tevatron and early LHC data.
The Tevatron has published limits on the Higgs boson
cross-section in this model, excluding a wide range of Higgs boson
masses~\cite{Aaltonen:2010sv}.
Recently the CMS collaboration carried out a similar study~\cite{Chatrchyan:2011tz}.
A precise calculation of the Higgs boson gluon fusion
cross-section has been made in Ref.~\cite{Anastasiou:2010bt},
where numerical results
have been given for the Tevatron. In this work, we compute the
cross-section at the LHC with center of mass energy $\sqrt{s} = 7\,
{\rm TeV}$.
\section{Calculation details}
In the Standard Model, gluon fusion Higgs production is mediated by massive
quarks and electroweak gauge bosons.
The dominant contribution is given by top-quark loops. Bottom-quark loops
and $W,Z$ loops yield a small contribution, and quarks lighter
than the bottom can be neglected.
The gluon fusion cross-section through massive quark loops
has been computed through NLO in perturbative QCD
both in the heavy quark effective theory~\cite{Dawson:1990zj,Djouadi:1991tka}
and retaining the full top-mass dependence~\cite{Graudenz:1992pv,Spira:1995rr}.
The NLO corrections are
large, motivating the calculation of the cross-section through NNLO
in perturbative QCD~\cite{Harlander:2002wh,Anastasiou:2002yz,Ravindran:2003um}.
The NNLO corrections have first been
computed in the heavy quark effective theory (HQET). Recently,
subleading corrections in the inverse top-quark mass
expansion have been calculated~\cite{Harlander:2009mq,Pak:2009dg},
validating the quality of the HQET
approximation. Other corrections include the exact two-loop
electroweak contributions~\cite{Actis:2008ug} and mixed QCD and electroweak
corrections~\cite{Anastasiou:2008tj} using an effective theory approach.
A precise prediction of the Higgs boson cross-section in fixed order
perturbation theory may be obtained by combining the above
contributions, with an estimated uncertainty of about
$\pm 10\%$ due to missing higher order effects. Calculations
resumming factorizable contributions of infrared nature are in
agreement with this error uncertainty~\cite{Catani:2003zt,deFlorian:2009hc}.
A similar uncertainty at the
LHC and slightly larger at the Tevatron is due to parton
distribution functions.
In this work, we compute the cross-section in a Standard Model with
four fermion generations.
The following points make this calculation
different from the calculation in the
Standard Model.
\begin{itemize}
\item The gluon fusion amplitudes receive contributions from the
up and down quarks of the fourth generation. We compute these
contributions through NLO retaining the exact mass
dependence. In the strict HQET limit, the cross-section is increased
by a factor of nine with respect to the SM cross-section.
\item Bottom-quark contributions are suppressed with respect to the
SM by roughly a factor of three. We compute bottom-quark
contributions exactly through NLO in perturbative QCD.
\item Through NLO, the Wilson coefficient of the HQET operator is
three times the Standard Model Wilson coefficient. At NNLO, the
Wilson coefficient depends on the masses of the the heavy quarks
and it also receives contributions from three-loop diagrams
containing two different heavy quarks~\cite{Anastasiou:2010bt}.
\item Two-loop electroweak corrections due to light quarks are
given by the same diagrams as in the Standard Model~\cite{Aglietti:2004nj}. They
predominate the
electroweak corrections for a light Higgs boson.
They are suppressed in the Standard Model with four
generations compared to the three-generation SM. In the
SM, additional electroweak corrections due
to the top quark are very small for a light Higgs boson
and have been computed in Ref.~\cite{Actis:2008ug}.
Here we include the full two-loop Standard Model electroweak
corrections of Ref.~\cite{Actis:2008ug}, but we neglect the two-loop electroweak
corrections with quarks of the fourth generation.
\end{itemize}
We compute the NNLO corrections in the effective theory
approximation. We normalize these corrections to the Born cross-section with the exact
mass dependence on the masses of the top-quark and the quarks of
the fourth generation. We do not include bottom-quark loops at NNLO.
\section{Predictions for the gluon fusion cross-section
for $\sqrt{s} = 7 \, {\rm TeV}$}
In this section we present our numerical results for the gluon
fusion cross-section in a Standard Model with four generations at
the current LHC energy of $7\, {\rm TeV}$ and for Higgs boson masses in
the range $m_h \in \left[ 110\,{\rm GeV}, 300\,{\rm GeV} \right]$.
The parameter space of the Standard Model with four generations is
quite limited. The model becomes non-perturbative for values of
the heavy quark masses roughly above $\sim 500 \, {\rm GeV}$~\cite{Marciano:1989ns}. A large mass
splitting of the quarks of the fourth generation induces large
corrections to oblique electroweak parameters~\cite{Kribs:2007nz}. To comply
with these constraints we present the cross-section for two
scenarios,
where the mass of the fourth-generation down quark is chosen as
\begin{equation}
\mbox{scenario 1: } m_{d_4} = 300 \, {\rm GeV} \; , \qquad
\mbox{scenario 2: } m_{d_4} = 400 \, {\rm GeV} \;.
\end{equation}
The mass of the fourth-generation up quark given by~\cite{Kribs:2007nz}
\begin{equation}
m_{u_4} - m_{d_4} = 50 \, {\rm GeV} + 10 \, {\rm GeV} \, \times \,
\log\left(
\frac{m_h}{115 \, {\rm GeV}}
\right) \, {\rm GeV} \;.
\end{equation}
For the top- and bottom-quark masses we take
\begin{equation}
m_t = 172 \, {\rm GeV} \; , \qquad m_b(m_b) = 4.2 \, {\rm GeV} \; .
\end{equation}
We use the MSTW08 NNLO parton densitities~\cite{Martin:2009iq} and quote the
``$\alpha_s$+pdf'' uncertainty at the $90\%$ confidence level (CL).
To estimate the uncertainty due to higher order perturbative effects,
we vary the renormalization and factorization scales in the range
$\frac{m_h}{4} \leq \mu_r=\mu_f \leq m_h$. The scale variation and pdf
uncertainty are very similar to the equivalent uncertainties in the
Standard Model. We present our results in Table~\ref{tab:cross-sections}.
\begin{table}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
{\rm GeV}
&
$
\begin{array}{c}
\sigma[pb] \\
\mbox{scenario 1}
\end{array}
$
&
$
\begin{array}{c}
\sigma[pb] \\
\mbox{scenario 2}
\end{array}
$
&
{\tiny
$
\begin{array}{c}
\delta ^{(+)} ({\rm pdf}+\alpha_s) \\
{\rm MSTW08} \\
90\% \, {\rm CL}
\end{array}
$
}\%
&
{\tiny
$
\begin{array}{c}
\delta ^{(-)} ({\rm pdf}+\alpha_s) \\
{\rm MSTW08} \\
90\% \, {\rm CL}
\end{array}
$
}\%
& $\delta ^{(+)} ({\mu})$\%
& $\delta ^{(-)} (\mu)$\% \\
\hline
105 & 202.33 & 201.39 & 7.9 & -7.6 & 9.2 & -9.7 \\
110 & 183.41 & 182.51 & 7.9 & -7.6 & 9.0 & -9.7 \\
115 & 166.85 & 165.97 & 7.9 & -7.6 & 8.9 & -9.6 \\
120 & 152.27 & 151.41 & 7.9 & -7.6 & 8.7 & -9.6 \\
125 & 139.38 & 138.54 & 7.9 & -7.6 & 8.6 & -9.6 \\
130 & 127.93 & 127.12 & 7.9 & -7.6 & 8.5 & -9.5 \\
135 & 117.72 & 116.93 & 7.9 & -7.6 & 8.4 & -9.5 \\
140 & 108.59 & 107.81 & 7.9 & -7.6 & 8.3 & -9.5 \\
145 & 100.39 & 99.628 & 7.9 & -7.6 & 8.2 & -9.4 \\
150 & 93.002 & 92.253 & 7.9 & -7.6 & 8.1 & -9.4 \\
155 & 86.298 & 85.563 & 7.9 & -7.6 & 8.0 & -9.4 \\
160 & 80.091 & 79.371 & 7.9 & -7.6 & 7.9 & -9.4 \\
165 & 74.221 & 73.516 & 7.9 & -7.7 & 7.8 & -9.4 \\
170 & 68.920 & 68.228 & 8.0 & -7.7 & 7.8 & -9.3 \\
175 & 64.249 & 63.570 & 8.0 & -7.7 & 7.7 & -9.3 \\
180 & 60.000 & 59.333 & 8.0 & -7.7 & 7.6 & -9.3 \\
185 & 56.080 & 55.424 & 8.0 & -7.8 & 7.6 & -9.3 \\
190 & 52.493 & 51.849 & 8.1 & -7.8 & 7.5 & -9.3 \\
195 & 49.246 & 48.612 & 8.1 & -7.8 & 7.4 & -9.3 \\
200 & 46.306 & 45.681 & 8.1 & -7.9 & 7.4 & -9.2 \\
205 & 43.620 & 43.005 & 8.1 & -7.9 & 7.3 & -9.2 \\
210 & 41.153 & 40.546 & 8.2 & -7.9 & 7.3 & -9.2 \\
215 & 38.878 & 38.279 & 8.2 & -8.0 & 7.2 & -9.2 \\
220 & 36.776 & 36.185 & 8.2 & -8.0 & 7.2 & -9.2 \\
225 & 34.832 & 34.249 & 8.3 & -8.0 & 7.2 & -9.2 \\
230 & 33.031 & 32.454 & 8.3 & -8.1 & 7.1 & -9.2 \\
235 & 31.357 & 30.788 & 8.3 & -8.1 & 7.1 & -9.2 \\
240 & 29.805 & 29.242 & 8.3 & -8.1 & 7.0 & -9.2 \\
245 & 28.357 & 27.800 & 8.4 & -8.2 & 7.0 & -9.1 \\
250 & 27.009 & 26.459 & 8.4 & -8.2 & 7.0 & -9.1 \\
255 & 25.751 & 25.206 & 8.5 & -8.3 & 6.9 & -9.1 \\
260 & 24.577 & 24.038 & 8.5 & -8.3 & 6.9 & -9.1 \\
265 & 23.480 & 22.945 & 8.5 & -8.3 & 6.9 & -9.1 \\
270 & 22.455 & 21.926 & 8.6 & -8.4 & 6.8 & -9.1 \\
275 & 21.495 & 20.970 & 8.6 & -8.4 & 6.8 & -9.1 \\
280 & 20.598 & 20.078 & 8.6 & -8.4 & 6.8 & -9.1 \\
285 & 19.756 & 19.241 & 8.7 & -8.5 & 6.7 & -9.1 \\
290 & 18.969 & 18.457 & 8.7 & -8.6 & 6.7 & -9.1 \\
295 & 18.232 & 17.725 & 8.8 & -8.6 & 6.7 & -9.1 \\
300 & 17.541 & 17.037 & 8.8 & -8.6 & 6.7 & -9.1 \\
\hline
\end{tabular}
\end{center}
\caption{Gluon fusion cross-section in a
Standard Model with four fermion generations.
The masses of the fourth generation quarks are
chosen according to ``scenario 1''
and ``scenario 2''. All cross-sections are computed with $0.1\%$ Monte-Carlo
integration error or better.}
\label{tab:cross-sections}
\end{table}
\section{Sensitivity to parton distributions}
Besides the MSTW08 pdf set, other two NNLO parton distribution sets are currently made available
by the GJR~\cite{JimenezDelgado:2009tv} and ABKM~\cite{Alekhin:2009ni} collaborations.
In Table~\ref{tab:pdfs} we present the central value of the cross-section for $\mu =
\frac{m_h}{2}$, the scale variation uncertainty and the
pdf uncertainty (at the $68\%$ CL) for $m_h=110,165,200, 300\, {\rm GeV}$
and the ABKM09, GJR and MSTW08 pdf sets.
\begin{table}[th]
\begin{center}
\begin{tabular}{|c||c|c|c|c|}
\hline
$\sigma [pb$] &
$\mbox{ABKM09}$& $\mbox{GJR}$ & $\left. \mbox{MSTW08}\right|_{68\%
{\rm CL}}$
& $\left. \mbox{MSTW08}\right|_{90\%
{\rm CL}}$
\\
\hline
$m_h= 110\, {\rm GeV}$ & $167.59 \pm 3.0\%_{\rm pdf}$
& $162.78 \pm 3.6\%_{\rm pdf}$
& $183.41 \begin{array}{c} +4.0
\\ - 3.1 \end{array}\%_{\rm
pdf}$
& $ \begin{array}{c} +7.9 \\ - 7.6 \end{array}\%_{\rm pdf}$ \\
$m_h= 165\, {\rm GeV}$ & $66.130 \pm 3.3\%_{\rm pdf}$
& $67.713 \pm 3.3\%_{\rm pdf}$
& $74.221 \begin{array}{c} +4.0
\\ - 3.3 \end{array}\%_{\rm
pdf}$
& $ \begin{array}{c} +7.9 \\ - 7.7 \end{array}\%_{\rm pdf}$ \\
$m_h= 200\, {\rm GeV}$ & $40.634 \pm 3.6\%_{\rm pdf}$
& $42.867 \pm 3.5\%_{\rm pdf}$
& $46.306 \begin{array}{c} +4.1 \\ - 3.4 \end{array}\%_{\rm pdf}$
& $ \begin{array}{c} +8.1 \\ - 7.9 \end{array}\%_{\rm pdf}$ \\
$m_h= 300\, {\rm GeV}$ & $14.768 \pm 4.7\%_{\rm pdf}$
& $16.786 \pm 5.0\%_{\rm pdf}$
& $17.541
\begin{array}{c} +4.3 \\ - 3.9 \end{array}
\%_{\rm pdf}$
& $
\begin{array}{c} +8.8 \\ - 8.6 \end{array}
\%_{\rm pdf}$
\\
\hline
\end{tabular}
\end{center}
\caption{A comparison for the gluon fusion cross-section in the
``scenario 1'' of the four-generation Standard Model with
the three available NNLO pdf sets: ABKM09, GJR and MSTW08.}
\label{tab:pdfs}
\end{table}
We find that the GJR and ABKM09 pfds give central values for the
cross-section which can be up to $12\%$ smaller than the one of
MSTW08. These differences are larger than what anticipated from the
quoted parton density uncertainties at the $68\%$ confidence level.
The MSTW08 provides uncertainties at the $90\%$ confidence level,
which yield about twice as large an uncertainty for the gluon fusion
cross-section. These overlap, albeit marginally for lower Higgs
boson masses, with the uncertainties of GJR and ABKM09.
All three groups provide consistent determinations of the parton
distributions with self-consistent choices and assumptions, and
there is no ``bullet-proof'' argument to choose one over the
others. Nevertheless, we prefer MSTW08 as our default pdf.
Our main reasons for this choice are that MSTW08 includes jet data directly sensitive
to the gluon density and that their central value of $\alpha_s(m_Z)$ is
in very good agreement with the world average \cite{Bethke:2009jm}
and determinations from jet data at $e^+e^-$ colliders using NNLO jet cross-sections~\cite{Dissertori:2009qa}.
We also find it prudent to estimate the pdf uncertainty of the cross-section
using parton density uncertainties at the $90\%$ CL.
The extraction of parton distribution functions will soon be assisted with early LHC data,
which we hope will help to resolve the discrepances among the various pdf sets.
\section{Branching ratios}
For a complete prediction of a Higgs signal cross-section at
colliders the branching ratios of the Higgs boson decays to
observable final states are needed. Branching ratios are significantly modified
with respect to the Standard Model when adding a fourth quark and
lepton generation.
In this model, the partial decay width of the
Higgs boson to gluons is enhanced by a large factor which reaches
nine in the HQET limit. This width dominates for a light Higgs boson.
The decay widths to $WW$ and $ZZ$ are significant for a range of
values of the Higgs boson mass above ~$140 \,{\rm GeV}$.
However, the corresponding branching ratios are smaller than in the Standard
Model. These decays are important, given that the Tevatron
and the LHC are quite sensitive and their experimental
study may lead to a Higgs boson discovery with a modest amount
of data. In the four-generation SM, novel decays of a heavy Higgs boson
to the leptons
of the fourth generation emerge and assume a significant width.
A systematic study of the Higgs boson decays has been made in
Ref.~\cite{Kribs:2007nz}, where the branching ratios have been
computed by modifying the program {\tt HDECAY}~\cite{Djouadi:1997yw}. Tabulated
results from Ref.~\cite{Kribs:2007nz} for the branching ratios of
the Higgs boson, corresponding to ``scenario 1'' and ``scenario 2''
of our study, can be found in Ref.~\cite{Aaltonen:2010sv}.
\section{Conclusions}
In this work we have studied the Higgs boson cross-section at the
LHC, in a Standard Model with a fourth generation. We have computed
the cross-section through NNLO in perturbative QCD, including finite
quark-mass effects and electroweak corrections through NLO. We have
provided an estimate of the uncertainty due to higher order perturbative
effects, and have studied the sensitivity of the cross-section on
various NNLO parameterizations of the parton densities and their
uncertainties. Our results are of direct relevance to the ongoing
studies for the discovery of the Higgs boson at the LHC.
\subsection*{Acknowledgments}
Research supported by the Swiss National Foundation under
contract SNF 200020-126632 and the DOE under Grant DE-AC02-98CH10886.
\section*{References}
|
\section{Introduction}
\label{sec:intro}
The masses of galaxies are revealed by the gravitational interaction
of their matter constituents, e.g. by stellar or gas kinematics or
gravitational lensing effects. The caveat is that these effects
rigorously constrain only the total amount of gravitating mass. The
decomposition into luminous and dark matter relies on further
assumptions. For example, in dynamical studies of early-type galaxies
it is commonly assumed that the stellar mass distribution follows the
light. Any radial increase in the mass-to-light ratio -- if needed to
explain the observed velocities of the stars -- is attributed to dark
matter.
The limitation of this approach becomes clear if one imagines a galaxy
dark matter halo that follows the light distribution
exactly. Discriminating between luminous and dark matter according to
their potentially different radial distributions would fail in this
case. In fact, there would be no direct way to dynamically unravel the
relative contributions of luminous and dark matter to the galaxy
mass. It is unlikely that a real galaxy halo follows the light
distribution exactly, yet it exemplifies the intrinsic degeneracies in
any mass decomposition. On top of this, even the most advanced
present-day dynamical modelling techniques rely on symmetry
assumptions and, often, also on the assumption of a steady-state
dynamical system. If the symmetry assumptions are strongly violated,
dynamical mass-to-light ratios can be biased by a factor of up to two
\citep{Tho07A}.
To crosscheck the validity of the assumptions in dynamical modelling
it is important to compare the resulting masses with other independent
methods. {\it Total} masses can be most directly compared to results
from gravitational lensing, which is the first goal of this paper.
An examination of the mass {\it decomposition} requires the
investigation of a galaxy's stellar population. Since the latter
provides an independent measure of the stellar mass-to-light ratio,
the comparison of dynamical and stellar-population masses yields
further constraints on the dark matter content. In the above mentioned
case, for example, the dynamical mass-to-light ratio $\Upsilon_{\ast,\mathrm{dyn}}$ would
exceed the corresponding stellar population value, indicating that
some fraction of the galaxy's mass is actually dark matter.
Unfortunately, stellar population models do not provide unique stellar
mass-to-light ratios. They suffer from an incomplete knowledge of the
initial stellar mass function (IMF), as well as age-metallicity
degeneracies. Observations in and around our own Galaxy indicate that
the IMF slope flattens below $0.5 \, \mathrm{M}_\odot$ \citep{Sca86,Kro01}.
Recent spectroscopic observations of massive ellipticals in the
near-infrared point towards the low-mass slope of the IMF in these
galaxies being steeper \citep{Dok10}. Yet, until now the IMF in
distant galaxies with unresolved stellar populations remains largely
uncertain. This translates into a significant indeterminacy of
population mass-to-light ratios: the steeper the slope at the low-mass
end, the higher the population mass-to-light ratio. Hence, the stellar
population approach is not directly conclusive as a probe of the mass
decomposition in dynamical models.
\begin{figure*}\centering
\begin{minipage}{166mm}
\includegraphics[width=144mm,angle=0]{figure1.eps}
\caption{Stellar population mass-to-light ratios $\Upsilon_\mathrm{Salp}$ (Salpeter
IMF) as function of radius along the major-axis (filled circles) and
minor-axis (open circles). Note that for some galaxies there
are systematic differences between the two sides of a slit which
are slightly larger than the statistical errors (e.g. along the
major-axis of GMP0144). The dotted line in each panel is the
light-weighted average of $\Upsilon_\mathrm{Salp}$ within $r<r_\mathrm{eff}$, the solid line
corresponds to the stellar mass-to-light ratio $\Upsilon_{\ast,\mathrm{dyn}}$ from
dynamical models. From top-left to bottom-right galaxies are plotted
in order of decreasing $\Upsilon_{\ast,\mathrm{dyn}}$.}
\label{fig:salpeter_data}
\end{minipage}
\end{figure*}
Conversely, if dynamical stellar mass determinations were free of
ambiguities with respect to a dark matter contamination then they
could serve as a measure for the slope of the IMF in distant
galaxies. The method would be to tweak the IMF in the stellar
population models until agreement with the dynamical stellar masses is
achieved.
The ambiguities in both, dynamical stellar masses as well as stellar
population models, make neither of the approaches directly
applicable. Nevertheless, comparing dynamical with stellar population
models is important to (1) narrow down the dynamically plausible range
of possible IMFs and to (2) delimit the range of dark matter fractions
compatible with observed stellar population properties. This is the
second goal of the present paper in which we compare dynamical and
stellar population masses in a sample of 16 Coma early-type galaxies.
Our dynamical models account for both the full variety of possible
orbit configurations in axisymmetric, flattened galaxies as well as
for dark matter. In this respect we extend previous
studies. \citet{Cap06}, for example, used a similar modelling
technique to compare dynamical and stellar population masses in the
SAURON sample, but they did not consider dark matter explicitly in
their dynamical models. The justification for this neglect was the
expected insignificance of dark matter in the central galaxy regions
probed by the SAURON observations ($r_\mathrm{obs,max} \la
r_\mathrm{eff}$). However, measuring only the sum of luminous and dark mass
makes the comparison with stellar population models potentially
uncertain. \citet{Nap10} analysed a large sample of early-type
galaxies taking into account the contribution from dark matter, but
their models do not account for galaxy flattening and rotation.
A subsample of the Coma galaxies was recently analysed by
\citet{Gri10}. While they used multi-band photometry to derive stellar
population properties, our approach is to measure ages, metallicities
and [$\alpha$/Fe] ratios from Lick indices to reduce potential biases
in population parameters. Likewise, the analysis of SLACS galaxies by
\citet{Tre10}, combining constraints from gravitational lensing and
stellar dynamics, relied on multi-band photometry for the stellar
population part.
Mass-to-light ratios of early-type galaxies are of particular interest
to understand the tilt of the fundamental plane. Virial relations
imply that the effective surface brightness $\langle I \rangle_\mathrm{eff}$, the effective
radius $r_\mathrm{eff}$ and the central velocity dispersion $\sigma_0$ in hot
stellar systems are not independent of each other. This is revealed
by the fundamental plane of early-type galaxies \citep{Djo87,Dre87}.
Yet, the observed fundamental plane is tilted with respect to the
simple case of a virialised, homologous family of dynamical
systems. This tilt can reflect (1) systematic variations of the
luminosity distribution (e.g. \citealt{Sag93,Tru04}), (2) systematic
variations of the orbital structure (e.g. \citealt{Cio96}) or (3)
systematic variations of the mass-to-light ratio, as a result of
varying stellar populations and/or dark matter distributions
(e.g. \citealt{Ren93}). Understanding these variations allows a
deeper insight into the formation process of early-type galaxies
\citep{Ben92}.
Most of the above mentioned effects can be factored out if additional
information about the stellar populations and/or the mass
distributions are available. Aperture spectroscopy is one way to
measure stellar population properties. By assuming simple scaling laws
it can also provide estimates for dynamical masses, such that the
contributions of stellar population and dynamical effects on the
fundamental plane tilt can be disentangled (e.g.
\citealt{Hyd09,Gra10}). More reliable constraints come from radially
resolved spectroscopy and detailed dynamical (or lensing) models of
galaxies (e.g. \citealt{Cap06,Bol07}). The third goal of this paper is
to follow the latter approach and to use the specific information
contained in our two-component dynamical models for further
investigations upon the origin of the fundamental plane tilt.
The paper is organised as follows. Sec.~\ref{data:review} reviews the
galaxy sample and models. in Sec.~\ref{sec:projmass} we compare
projected masses from dynamical models and from gravitational lensing
against each other. Sec.~\ref{sec:ml} deals with the comparison of
luminous dynamical and stellar population masses. The dark matter
distribution is analysed in Sec.~\ref{sec:dm} and implications for the
tilt of the fundamental plane are addressed in Sec.~\ref{sec:fp}. The
paper is summarised in Sec.~\ref{sec:sum}.
\section{Galaxy sample and modelling}
\label{data:review}
The sample analysed in this paper comprises 16 Coma early-type
galaxies in the luminosity range between $M_B = -19.88$ and $M_B = -
22.26$ (eight giant ellipticals, eight lenticular/intermediate type
galaxies). It is almost identical to the sample presented in
\citet{Tho09}. Only the galaxy GMP3958 has been omitted for its strong
gas emission, which hampers reliable stellar population modelling.
For the dynamical analysis of each galaxy a composite of ground-based
and HST photometry has been used. Stellar absorption line data for the
kinematics and Lick indices come from various long-slit spectra, at
least along the major and minor axes, but in many cases covering other
position angles as well. The spectra extend to $1-4\,r_\mathrm{eff}$. Details
about the photometric and spectroscopic data have been published in
\citet{Meh00}, \citet{Weg02}, \citet{Cor08}, and \citet{Tho09}.
\subsection{Dynamical modelling}
\label{subsec:dynmod}
To the kinematic and photometric data we applied our implementation of
Schwarzschild's orbit superposition technique \citep{S79} for
axisymmetric potentials \citep{Ric88,Geb03,Tho04,Tho05}. A detailed
description of the models is given in \citet{Tho07B}. Most important
for this paper are the assumptions about the mass structure. We will
consider two sets of models. For the first set it is assumed that all
the mass follows the light, i.e.
\begin{equation}
\label{eq:mfl}
\rho = \Upsilon \times \nu
\end{equation}
where $\nu$ is the three-dimensional luminosity density. By
construction, the mass-to-light ratio $\Upsilon$ here includes the
contribution of both the stellar and dark mass of a galaxy. It is not
known in advance and obtained by a $\chi^2$-minimisation with respect
to the kinematical observations. The best-fit $\Upsilon$ of
one-component (i.e. self consistent) models will be referred to as
$\Upsilon_{\ast,\mathrm{sc}}$ in the remainder of this paper and can be found in
Tab.~\ref{tab:dmfrac}.
For the second, and standard set of models we assume a mass density of
the form
\begin{equation}
\label{eq:massdecomp}
\rho = \Upsilon \times \nu + \rho_\mathrm{DM}.
\end{equation}
The first component again follows the light, while the second one,
$\rho_\mathrm{DM}$, accounts for dark matter. Eq.~\ref{eq:massdecomp} is
designed to separate the contributions of luminous and dark matter to
the total mass of a galaxy. Our basic assumption is that the best-fit
$\Upsilon$ of two-component models represents only the stellar mass of
a galaxy and we will refer to it as the dynamical stellar
mass-to-light ratio $\Upsilon_{\ast,\mathrm{dyn}}$ in the following. Strictly speaking,
$\Upsilon_{\ast,\mathrm{dyn}}$ measures all the mass which follows the light distribution,
be it stellar or be it dark matter. In this respect $\Upsilon_{\ast,\mathrm{dyn}}$ provides
only an upper limit for the stellar mass. As soon as there are other
mass components which -- for whatever reason -- follow the light, they
will contribute to $\Upsilon_{\ast,\mathrm{dyn}}$ as well and the actual galaxy stellar
mass-to-light ratio will be smaller than $\Upsilon_{\ast,\mathrm{dyn}}$. A more detailed
discussion upon this issue will be given in
Secs.~\ref{subsec:mldynerr} and \ref{subsec:lightdm}.
According to equation (\ref{eq:massdecomp}), the cumulative
(spherical) dark matter fraction $f_\mathrm{DM,dyn}$ of the models inside radius
$r$ is
\begin{equation}
\label{eq:dmfrac}
f_\mathrm{DM,dyn}(r) \equiv \frac{M_\mathrm{DM,dyn}(r)}{M_\mathrm{tot,dyn}(r)} = \frac{M_\mathrm{DM,dyn}(r)}{\Upsilon_{\ast,\mathrm{dyn}} \, L(r) + M_\mathrm{DM,dyn}(r)},
\end{equation}
where $M_\mathrm{tot,dyn}$, $M_\mathrm{DM,dyn}$ and $L$ stand for the cumulative total mass,
dark mass and light inside radius $r$. Tab.~\ref{tab:dmfrac} lists the
observationally derived dark matter fractions. Here we consider mostly
the region inside $r_\mathrm{eff}$, where the average dark matter fraction is
$\langle f_\mathrm{DM,dyn} \rangle = 23 \pm 17 \, \%$, but our data reach out to
radii where the stellar and the dark matter density become equal
\citep{Tho07B}.
\begin{table*}
\begin{center}
\begin{minipage}{170mm}
\begin{tabular}{lc||cccccc}
\multicolumn{2}{c}{galaxy} & $\Upsilon_{\ast,\mathrm{dyn}}$ & $\Upsilon_{\ast,\mathrm{sc}}$ & $f_\mathrm{DM,dyn}$ & $\sigma_\mathrm{eff}$ & $\Upsilon_\mathrm{Krou}$ & $\Upsilon_\mathrm{Salp}$\\ \\
GMP & NGC/IC & $[\mathrm{M}_\odot/\mathrm{L}_{R,\odot}]$ & $[\mathrm{M}_\odot/\mathrm{L}_{R,\odot}]$ & & $[\mathrm{km/s}]$ & $[\mathrm{M}_\odot/\mathrm{L}_{R,\odot}]$ & $[\mathrm{M}_\odot/\mathrm{L}_{R,\odot}]$\\ \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) \\
\hline
0144 & 4957 & $4.50 \pm 0.50$ & $7.00 $ & $0.30^{+0.10}_{-0.05} $ & $211.8 \pm 0.4$ & $3.03 \pm 0.17$ & $4.74 \pm 0.26$ \\
0282 & 4952 & $5.00 \pm 0.50$ & $6.50 $ & $0.25^{+0.13}_{-0.05} $ & $268.2 \pm 0.5$ & $3.57 \pm 0.22$ & $5.58 \pm 0.34$ \\
0756 & 4944 & $2.60 \pm 0.20$ & $3.00 $ & $0.11^{+0.05}_{-0.05} $ & $193.2 \pm 0.3$ & $1.40 \pm 0.06$ & $2.19 \pm 0.10$ \\
1176 & 4931 & $2.00 \pm 0.20$ & $2.50 $ & $0.31^{+0.05}_{-0.04} $ & $205.6 \pm 0.3$ & $1.54 \pm 0.04$ & $2.41 \pm 0.06$ \\
1750 & 4926 & $6.00 \pm 0.75$ & $7.00 $ & $0.15^{+0.47}_{-0.05} $ & $279.1 \pm 1.6$ & $3.41 \pm 0.50$ & $5.33 \pm 0.79$ \\
1990 & IC 843 & $10.00 \pm 1.00$ & $10.00 $ & $0.01^{+0.01}_{-0.01} $ & $281.3 \pm 1.1$ & $3.03 \pm 0.30$ & $4.74 \pm 0.47$ \\
2417 & 4908 & $8.00 \pm 0.75$ & $8.50 $ & $0.05^{+0.20}_{-0.04} $ & $211.0 \pm 1.7$ & $3.78 \pm 0.43$ & $5.91 \pm 0.67$ \\
2440 & IC 4045 & $6.50 \pm 0.50$ & $7.00 $ & $0.07^{+0.02}_{-0.02} $ & $225.9 \pm 0.9$ & $3.46 \pm 0.27$ & $5.41 \pm 0.42$ \\
3414 & 4871 & $4.00 \pm 0.62$ & $6.00 $ & $0.46^{+0.06}_{-0.27} $ & $169.6 \pm 1.3$ & $3.20 \pm 0.29$ & $4.99 \pm 0.45$ \\
3510 & 4869 & $5.50 \pm 0.50$ & $6.00 $ & $0.09^{+0.18}_{-0.05} $ & $177.7 \pm 1.7$ & $3.49 \pm 0.65$ & $5.45 \pm 1.02$ \\
3792 & 4860 & $8.00 \pm 1.00$ & $9.00 $ & $0.13^{+0.24}_{-0.05} $ & $284.0 \pm 1.9$ & $4.27 \pm 0.31$ & $6.66 \pm 0.48$ \\
4822 & 4841A & $5.50 \pm 1.00$ & $6.50 $ & $0.51^{+0.14}_{-0.27} $ & $272.1 \pm 2.7$ & $3.24 \pm 0.33$ & $5.07 \pm 0.52$ \\
4928 & 4839 & $8.50 \pm 2.00$ & $10.00 $ & $0.32^{+0.35}_{-0.21} $ & $314.8 \pm 2.9$ & $3.34 \pm 0.46$ & $5.22 \pm 0.72$ \\
5279 & 4827 & $6.50 \pm 0.50$ & $7.00 $ & $0.10^{+0.21}_{-0.07} $ & $244.1 \pm 1.2$ & $3.16 \pm 0.45$ & $4.93 \pm 0.71$ \\
5568 & 4816 & $6.00 \pm 1.00$ & $7.00 $ & $0.53^{+0.15}_{-0.25} $ & $233.4 \pm 1.7$ & $3.91 \pm 0.33$ & $6.11 \pm 0.52$ \\
5975 & 4807 & $3.00 \pm 0.50$ & $4.00 $ & $0.29^{+0.05}_{-0.01} $ & $195.9 \pm 0.8$ & $3.07 \pm 0.24$ & $4.80 \pm 0.37$ \\
\hline
\end{tabular}
\caption{Dynamical parameters of Coma galaxies. (1, 2) galaxy
identification (GMP from \citealt{GMP}). (3) best-fit dynamical
$\Upsilon_{\ast,\mathrm{dyn}}$ ($R$-band) in models that explicitly account for dark
matter (the quoted errors include all models that deviate by less
than $\Delta \chi^2 \le 1$ from the best-fit model. For a more
detailed discussion of the errors the reader is referred to
\citealt{Tho05}.) (4) best-fit dynamical $\Upsilon_{\ast,\mathrm{sc}}$ ($R$-band) assuming
that all the mass follows the light. (According to Fig.~2 in
\citealt{Tho07B} the formal errors on $\Upsilon_{\ast,\mathrm{sc}}$ are smaller than those
of $\Upsilon_{\ast,\mathrm{dyn}}$. However, since the assumption that mass-follows-light
does not yield satisfactory fits to the kinematics we don't give
errors for $\Upsilon_{\ast,\mathrm{sc}}$ here.) (5) dark matter fraction $f_\mathrm{DM,dyn}$ within
$r_\mathrm{eff}$. (6) galaxy velocity dispersion $\sigma_\mathrm{eff}$ (inside
$r_\mathrm{eff}$). (7, 8) stellar population mass-to-light ratios for the
Kroupa IMF ($\Upsilon_\mathrm{Krou}$; $R$-band) and Salpeter IMF ($\Upsilon_\mathrm{Salp}$;
$R$-band), respectively. The mass-to-light ratios are light-weighted
averages within $r_\mathrm{eff}$. Only radii with a stellar-population age
$\tau \le 14 \, \mathrm{Gyr}$ are considered.
\label{tab:dmfrac}}
\end{minipage}
\end{center}
\end{table*}
We probed two dark matter descriptions. Firstly, logarithmic halos
with a constant-density core of size $r_h$ and secondly NFW halos
\citep{Nav96}. The latter provide good fits to cosmological $N$-body
simulations and have a central logarithmic slope of $-1$. Neither of
the two profiles includes baryonic halo contraction explicitly
(cf. the discussion in Sec.~\ref{subsec:lightdm}). Coma galaxies are
in most cases better fit with logarithmic halos, but the significance
over NFW halo profiles is marginal. The steeper central slope of NFW
halos implies a higher central dark matter density. But even in cases
where NFW halos fit better, the innermost galaxy regions are still
dominated by the mass-component that follows the light (for details
see \citealt{Tho07B}). The assumptions about the halo-density profile
have therefore little influence on the best-fit stellar mass-to-light
ratio $\Upsilon_{\ast,\mathrm{dyn}}$ (typically, $\Upsilon_{\ast,\mathrm{dyn}}$ from logarithmic or NFW halos
differ by no more than $\Delta \Upsilon_{\ast,\mathrm{dyn}} \approx 0.5$; see also
Sec.~\ref{subsec:imfnorm}).
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure2.eps}
\caption{Observed versus predicted colours for the subsample of Coma
galaxies with SDSS photometry. Colours as indicated in each
panel. Model predictions are plotted along the vertical axis, SDSS
observations horizontally. All axes are in magnitudes. The mean
colour difference $\Delta$ and rms-scatter between models and
observations (in magnitudes) is quoted in each panel.}
\label{fig:farben}
\end{figure}
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure3.eps}
\caption{Stellar population $\Upsilon_\mathrm{Krou}$ (Kroupa IMF) from SDSS colours
(x-axis; cf. \citealt{Gri10}) and from Lick indices (y-axis;
cf. Tab.~\ref{tab:dmfrac}).}
\label{fig:grillo}
\end{figure}
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure4.eps}
\caption{Einstein radii $r_\mathrm{Ein}$ and effective velocity
dispersions $\sigma_\mathrm{eff}$ of SLACS galaxies from \citet{Aug09} are
shown by the crosses. Dotted open circles highlight lenses that
follow eq.~(\ref{def:subA}), shown by the dotted line, and form
subsample $\mathrm{SL}_{0.5}$; solid open circles are for subsample $\mathrm{SL}_{0.75}$
(i.e. galaxies that follow eq.~\ref{def:subB}, shown by the solid
line).}
\label{fig:fitrelation}
\end{figure}
The gravitational potential of the galaxies is assumed to be
axisymmetric. Contrasting other methods like using Jeans equations,
the Schwarzschild technique allows the exploration of all possible
orbit configurations. The Coma sample is unique in being the only
larger sample of axisymmetric dynamical models including dark
matter. Previous modelling attempts either assumed spherical symmetry
\citep{Ger01} or did not include dark matter explicitly \citep{Cap06}.
\subsection{Stellar population models}
\label{subsec:gradients}
Stellar ages, metallicities, [$\alpha$/Fe] ratios and $R$-band stellar
mass-to-light ratios are determined by fitting the single stellar
population models of \citet{Mar98,Mar05} with $\alpha$-elements
overabundance of \citet{DTho03} to the Lick indices H$\beta$, $\langle
\mathrm{Fe} \rangle$, $[\mathrm{MgFe}]$ and Mg$\, b$. Two
initial-stellar-mass functions are considered. Firstly, the Salpeter
IMF ($\Upsilon_\mathrm{Salp}$, with mass limits of $0.1 \, \mathrm{M}_\odot$ and $100 \, \mathrm{M}_\odot$)
and, secondly, the Kroupa IMF ($\Upsilon_\mathrm{Krou}$, with the same mass limits,
but a shallower slope for stars below $0.5 \, \mathrm{M}_\odot$). The Salpeter
IMF implies more low-mass stars and a higher mass-to-light ratio. In
the $R$-band the scaling between the two cases is $\Upsilon_\mathrm{Salp} \approx
1.56 \, \Upsilon_\mathrm{Krou}$ \citep{Sal55,Kro01}. When the IMF does not need to
be specified, we will refer to stellar-population mass-to-light ratios
as $\Upsilon_\mathrm{ssp}$.
Stellar population properties are calculated at each radius with
observations. Salpeter $R$-band mass-to-light ratios are shown in
Fig.~\ref{fig:salpeter_data}. The light-weighted averages within
$r_\mathrm{eff}$ are indicated by the dashed lines and form the basis for the
remainder of this paper (Tab.~\ref{tab:dmfrac}). For some galaxies
there are systematic differences between the two sides of a slit which
are slightly larger than the statistical errors (e.g. along the
major-axis of GMP0144). These systematic uncertainties are not
included in the errors quoted in Tab.~\ref{tab:dmfrac}. Irrespective
of the metalicity gradients present in early-type galaxies
\citep{Meh03}, mass-to-light ratio gradients in the R-band are
generally small for the Coma galaxies
(cf. Fig.~\ref{fig:salpeter_data}). The results presented here do not
depend significantly on the averaging radius. Its choice is driven by
the most massive sample objects. Their spectroscopic data reach only
out to $r_\mathrm{obs,max} \approx r_\mathrm{eff}$ and, for the purpose of
homogeneity, we restrict the averaging in other galaxies to the same
radius, even if the data extend further out.
For a subsample of the Coma galaxies studied here, multi-band
photometry from the Sloan Digital Sky Survey (SDSS; \citealt{Yor00})
is available. Fig.~\ref{fig:farben} compares observed SDSS colours
with the predictions of our best-fit SSP models. We have applied the
same colour corrections to the models as discussed in \citet{Sag10}
plus an additional $i-z = -0.05$ (Maraston, private communication), to
take into account the recent improvements in the calibration of the
Maraston SSP models \citep{Mar09}. The vertical error-bars indicate
the 68\% confidence region derived from the observational
errors. Model colours are averaged inside $r_\mathrm{eff}$. They fit well to
the SDSS colours with average differences smaller than $\la 0.01 \,
\mathrm{mag}$.
\citet{Gri10} used observed SDSS colours to derive photometric stellar
population parameters for some of our Coma galaxies. In contrast to
the analysis presented here, they (1) assume a solar metallicity for
all galaxies (but see the middle panel of Fig.~\ref{fig:ratml_pop})
(2) allow for an extended star-formation history and (3) use the
\citet{Mar05} models without colour corrections. In
Fig.~\ref{fig:grillo} we plot their photometric
$\Upsilon_\mathrm{Krou}^\mathrm{SDSS}$ (scaled to the Kroupa IMF) against our
$\Upsilon_\mathrm{Krou}$. On average, both approaches yield consistent results
($\langle \Upsilon_\mathrm{Krou}/\Upsilon_\mathrm{Krou}^\mathrm{SDSS} \rangle = 1.11$), though
the rms-scatter ($\pm 0.35$) is large.
\section{Projected masses from dynamics and lensing}
\label{sec:projmass}
As it has been stated in Sec.~\ref{sec:intro}, the dynamical models
rely on the assumption of axial symmetry and steady state dynamics. To
check how accurately these assumptions are fulfilled in real galaxies
we first compare our dynamical models against gravitational lensing
results. The latter constrain the total projected mass inside a
cylinder delimited by the Einstein radius $r_\mathrm{Ein}$ of the lens and are
less affected by symmetry assumptions \citep{Koc91}.
\subsection{Lens selection}
\label{subsec:lensselect}
The Einstein radius of a gravitational lens results from two
independent properties of a lensing configuration. Firstly, from the
physical deflection angle that the lensing galaxy gives rise to
according to its gravity. It depends only on the mass distribution of
the foreground galaxy. Secondly, from projection factors that depend
on the distances of the foreground lens and the background source,
respectively. For the Coma galaxies we only know their mass
distributions, but they are not part of real lenses. To compare them
with observed gravitational lenses, we need to define an appropriate
lensing configuration for each Coma galaxy. Here we do this implicitly
by seeking for lensing galaxies that accidentally happen to fall on a
linear relation $r_\mathrm{Ein}(\sigma_\mathrm{eff})$ between the Einstein radius and the
effective velocity dispersion. Defining a fiducial Einstein radius for
each Coma galaxy according to the same $r_\mathrm{Ein}(\sigma_\mathrm{eff})$ then ensures,
that there is at least a subsample of real lenses with similar lens
configurations at a given $\sigma_\mathrm{eff}$. The Coma galaxies can be compared
to the lenses in such a subsample, but not to the rest of the lensing
galaxies. The form of the selection function is arbitrary, any other
function would serve equally well. We choose a linear relation for
simplicity.
Fig.~\ref{fig:fitrelation} shows the observed Einstein radii $r_\mathrm{Ein}$
of SLACS lenses from \citet{Aug09} against their velocity
dispersions\footnote{Note that for Fig.~\ref{fig:fitrelation} we have
used the correction of \citet{Cap06} to transform the measured SDSS
aperture dispersions into $\sigma_\mathrm{eff}$.}. The large circles show two
different subsamples of lenses constructed as outlined above. The
lenses are selected to deviate by less than $0.2 \, \mathrm{kpc}$ from two
arbitrary linear selection functions $r_\mathrm{Ein}(\sigma_\mathrm{eff}) = 0.025 \,
\sigma_\mathrm{eff} - 3.0$ and $r_\mathrm{Ein}(\sigma_\mathrm{eff}) = 0.025 \, \sigma_\mathrm{eff} - 1.9$. After
having selected the lenses we fit a straight line to each subsample in
order to determine the actual best-fit $r_\mathrm{Ein}(\sigma_\mathrm{eff})$ that is used
to define fiducial Einstein radii for Coma galaxies. The fitted
relations differ only slightly from the original selection functions.
For subsample $\mathrm{SL}_{0.5}$ (dotted) we get
\begin{equation}
\label{def:subA}
\frac{r_\mathrm{Ein}}{\mathrm{kpc}} = 0.0238 \times \frac{\sigma_\mathrm{eff}}{\mathrm{km/s}} - 2.6709
\end{equation}
and for subsample $\mathrm{SL}_{0.75}$ (solid)
\begin{equation}
\label{def:subB}
\frac{r_\mathrm{Ein}}{\mathrm{kpc}} = 0.0249 \times \frac{\sigma_\mathrm{eff}}{\mathrm{km/s}} - 1.8564.
\end{equation}
The two subsamples are constructed as a compromise between (1) ending
up with a sizeable number of galaxies in each subsample and (2)
yielding sufficiently different subsamples to allow for a comparison
between Coma and lensing galaxies at different physical scales. For
subsample $\mathrm{SL}_{0.5}$ we get an average $\langle r_\mathrm{Ein}/r_\mathrm{eff} \rangle
\approx 0.5$ (11 lenses), while for subsample $\mathrm{SL}_{0.75}$ it is
$\langle r_\mathrm{Ein}/r_\mathrm{eff} \rangle \approx 0.75$ (17 lenses).
\begin{figure*}\centering
\begin{minipage}{166mm}
\includegraphics[width=144mm,angle=0]{figure5.eps}
\caption{The projected total (luminous+dark) mass $M_\mathrm{Ein}$ within a
fiducial Einstein radius $r_\mathrm{Ein}$. Coma galaxies are indicated by the
large symbols. Top row: two-component models with dark matter halos
($\rho = \Upsilon_{\ast,\mathrm{dyn}} \times \nu + \rho_\mathrm{DM,dyn}$ in equation
\ref{eq:meindef}); bottom row: dynamical mass under the assumption
that mass follows light ($\rho = \Upsilon_{\ast,\mathrm{sc}} \times \nu$). Small circles:
total projected masses of SLACS galaxies \citep{Aug09}. In the
left-hand panels we compare Coma galaxies with the SLACS subsample $\mathrm{SL}_{0.5}$
using fiducial Einstein radii calculated via eq.~(\ref{def:subA}). In
the right-hand panels we compare to subsample $\mathrm{SL}_{0.75}$ (using
eq.~\ref{def:subB}). Masses are plotted against the average velocity
dispersion $\sigma_\mathrm{eff}$ inside $r_\mathrm{eff}$.}
\label{fig:projmass_tot}
\end{minipage}
\end{figure*}
\begin{figure*}\centering
\begin{minipage}{166mm}
\includegraphics[width=144mm,angle=0]{figure6.eps}
\caption{As Fig.~\ref{fig:projmass_tot}, but for stellar population
masses. Top row: Salpeter IMF ($\rho = \Upsilon_\mathrm{Salp} \times \nu$); bottom
row: Kroupa IMF ($\rho = \Upsilon_\mathrm{Krou} \times \nu$).}
\label{fig:projmass_ssp}
\end{minipage}
\end{figure*}
\subsection{The total mass}
\label{subsec:totproj}
The large symbols in Fig.~\ref{fig:projmass_tot} show projected,
integrated masses of Coma galaxies and SLACS lenses as a function of
galaxy velocity dispersion $\sigma_\mathrm{eff}$. Coma masses are calculated from
the integral
\begin{equation}
\label{eq:meindef}
M_\mathrm{Ein} \equiv \int_{-10 \, r_\mathrm{eff}}^{10 \, r_\mathrm{eff}} \mathrm{d}z \int_0^{2\pi} \mathrm{d}\varphi \int_0^{r_\mathrm{Ein}(\sigma_\mathrm{eff})} \rho \, r \, \mathrm{d}r,
\end{equation}
where ($r,\varphi$) are polar coordinates on the sky and $z$ is the
direction of the line-of-sight. Formally, the integral in
eq.~(\ref{eq:meindef}) should be calculated over $-\infty \leq z \leq
+\infty$, but we limited it over $-10 \, r_{\rm eff} \leq z \leq +10 \,
r_{\rm eff}$. For an isothermal sphere this cut-off results in a
negligible underestimation of the integral ($\approx 4 \, \%$). Coma
galaxy velocity dispersions $\sigma_\mathrm{eff}$ are measured by (1)
reconstructing the line-of-sight velocity distributions (LOSVDs) from
the kinematic moments, (2) coadding all LOSVDs inside $r_\mathrm{eff}$, each
weighted by its projected light, and (3) fitting a Gaussian to the
resulting LOSVD. They are listed in Tab.~\ref{tab:dmfrac}. Because of
their higher signal-to-noise, only the major-axis data have been
considered. The upper integration limit $r_\mathrm{Ein}(\sigma_\mathrm{eff})$ in
eq.~(\ref{eq:meindef}) refers to eq.~(\ref{def:subA}) for the
comparison with SLACS subsample $\mathrm{SL}_{0.5}$ and to eq.~(\ref{def:subB}) for
the comparison with subsample $\mathrm{SL}_{0.75}$. In Fig.~\ref{fig:projmass_tot}
we only show those SLACS lenses that belong to either subsample $\mathrm{SL}_{0.5}$
(left-hand panels) or subsample $\mathrm{SL}_{0.75}$ (right-hand panels),
respectively.
The top row is for the total cylindrical mass of our standard
two-component models with dark matter halos (i.e. with the best-fit
density $\rho = \Upsilon_{\ast,\mathrm{dyn}} \times \nu + \rho_\mathrm{DM,dyn}$ in
eq.~\ref{eq:meindef}). The good agreement with the lensing results is
reassuring (the average mass offset is $0.05$ dex for subsample $\mathrm{SL}_{0.5}$
and $0.02$ dex for subsample $\mathrm{SL}_{0.75}$). It implies that the two
completely independent methods yield consistent results. Moreover, the
scatter in the dynamical masses is not larger than in the lensing
masses. Consequently, strong deviations from axisymmetry are unlikely
in the Coma galaxies. As shown in \citet{Tho07A}, strong triaxiality,
if not accounted for in the models, can bias dynamical masses by a
factor of up to two, depending on viewing angle. Assuming random
viewing angles, strongly triaxial mass distributions would therefore
likely cause a significant scatter in the dynamical masses,
which is however not observed (but see \citealt{Ven09}).
The bottom row of Fig.~\ref{fig:projmass_tot} is for projected masses
of self-consistent dynamical models in which all the mass is assumed
to follow the light (i.e. $\rho = \Upsilon_{\ast,\mathrm{sc}} \times \nu$). These models are
not consistent with the lensing masses. The discrepancy is larger for
the SLACS subsample $\mathrm{SL}_{0.5}$ than for subsample $\mathrm{SL}_{0.75}$ because the
smaller Einstein radii of subsample $\mathrm{SL}_{0.5}$ emphasise the central
regions in the comparison. As it has been shown earlier
(e.g. \citealt{Ger01}, \citealt{Tho07B}), the mass distribution in
early-type galaxies follows the light in the inner regions
($\rho_\mathrm{in} \approx \Upsilon_{\ast,\mathrm{dyn}} \times \nu_\mathrm{in}$), but has an
additional component in the outer parts ($\rho_\mathrm{out} \approx
\Upsilon_{\ast,\mathrm{dyn}} \times \nu_\mathrm{out} + \rho_\mathrm{DM}$). Then, assuming that all
the mass follows the light requires $\Upsilon_{\ast,\mathrm{sc}} > \Upsilon_{\ast,\mathrm{dyn}}$ to include the
outer dark matter ($\Upsilon_{\ast,\mathrm{sc}} \times \nu_\mathrm{out} \approx
\rho_\mathrm{out} \approx \Upsilon_{\ast,\mathrm{dyn}} \times \nu_\mathrm{out} + \rho_\mathrm{DM}$).
However, the central regions become proportionally more massive, too,
such that $\Upsilon_{\ast,\mathrm{sc}} \times \nu_\mathrm{in} > \Upsilon_{\ast,\mathrm{dyn}} \times
\nu_\mathrm{in} \approx \rho_\mathrm{in}$. This explains why the
offset in the lower-left panel of Fig.~\ref{fig:projmass_tot} is larger
than in the lower-right one.
\subsection{Stellar population masses}
In Fig.~\ref{fig:projmass_ssp} we show projected {\it stellar} masses
(i.e. using $\rho = \Upsilon_\mathrm{Salp} \times \nu$ and $\rho = \Upsilon_\mathrm{Krou} \times
\nu$, respectively, in eq.~\ref{eq:meindef}). The open dots in these
panels represent the projected {\it stellar} masses for either the
Salpeter IMF (top row) or the Kroupa IMF (bottom row). Kroupa stellar
masses are always below lensing masses. The mass difference between
the lenses and the Kroupa masses increases from subsample $\mathrm{SL}_{0.5}$ to
subsample $\mathrm{SL}_{0.75}$ for the same reason discussed at the end of
Sec.~\ref{subsec:totproj}. It could be due to dark matter. Salpeter
stellar masses are likewise consistent with the lensing results. In
low-dispersion galaxies the IMF cannot be much steeper than Salpeter
as otherwise the implied stellar masses would exceed the total
observed lens masses. In high-dispersion galaxies the Salpeter stellar
masses are however not enough to explain the total lensing
masses. Then, if all the lensing mass was stellar, the IMF would have
to change with galaxy velocity dispersion and in massive early-types
the stellar mass per stellar light would have to be larger than for a
Salpeter IMF (or any equivalent top-heavy IMF). If the IMF is
constant, then the top-left panel of Fig.~\ref{fig:projmass_ssp}
provides direct evidence for the presence of dark matter in
high-dispersion early-type galaxies.
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure7.eps}
\caption{Dark matter fractions inside the Einstein radius.
Top: projected dark matter fraction; bottom: deprojected dark matter
fraction. Red/open circles: Coma galaxies of this work; dark/filled
circles in the top panel: SLACS galaxies from \citet{Koo06}. Coma
galaxy dark matter fractions are plotted against the velocity
dispersion $\sigma_\mathrm{eff}$ inside $r_\mathrm{eff}$. SLACS galaxies are plotted
against the measured velocity dispersion inside the SDSS
aperture.}
\label{fig:projdmfrac}
\end{figure}
\subsection{Luminous and dark matter separated}
In our standard two-component models that take into account the
detailed, radially resolved stellar kinematics, the large projected
masses of high-dispersion galaxies do not entirely originate from
luminous mass. This is shown in the top panel of
Fig.~\ref{fig:projdmfrac} (projected dark matter fractions inside the
Einstein radius). Note that here we compare to the results of
\citet{Koo06}, who provided a combined lensing and dynamics analysis
of the first SLACS galaxies. We applied a similar lens selection as
described in Sec.~\ref{subsec:lensselect} (using $r_\mathrm{Ein} = 0.03244
\times \sigma_\mathrm{eff} - 4.6324$). Our dynamically derived dark matter
fractions are in good agreement with those from \citet{Koo06}. As
already stated above, they increase with $\sigma_\mathrm{eff}$. For comparison, we
have also plotted the corresponding deprojected dark matter fractions
(cf. eq.~\ref{eq:dmfrac}) inside the same three dimensional radius
$r_\mathrm{Ein}(\sigma_\mathrm{eff})$ in the bottom panel of
Fig.~\ref{fig:projdmfrac}. They are generally lower and do not vary
with $\sigma_\mathrm{eff}$. Projection effects therefore contribute to the trends
in Figs.~\ref{fig:projmass_tot}-\ref{fig:projdmfrac}. To get a better
understanding about the stellar IMF and dark matter distribution it is
necessary to analyse the intrinsic three dimensional properties of the
galaxies. This will be done in the following Secs.~\ref{sec:ml} and
\ref{sec:dm}.
\section{Dynamical luminous mass versus stellar population mass}
\label{sec:ml}
Fig.~\ref{fig:ml_sig} shows dynamical and stellar-population
mass-to-light ratios versus $\sigma_\mathrm{eff}$. The dynamical $\Upsilon_{\ast,\mathrm{dyn}}$ exhibit
a broad distribution ranging from $\Upsilon_{\ast,\mathrm{dyn}} \approx 2$ to $\Upsilon_{\ast,\mathrm{dyn}}
\approx 10$. In contrast, for the majority of Coma galaxies the
stellar population $\Upsilon_\mathrm{ssp}$ are almost constant with $\Upsilon_\mathrm{Salp} \approx
5-6$ for the Salpeter IMF and $\Upsilon_\mathrm{Krou} \approx 3-4$ for the Kroupa
IMF. In addition, while the dynamical $\Upsilon_{\ast,\mathrm{dyn}}$ clearly increase with
galaxy velocity dispersion, there is no similar correlation between
$\sigma_\mathrm{eff}$ and stellar-population $\Upsilon_\mathrm{ssp}$.
There are two galaxies (GMP0756 and GMP1176) with distinctly lower
stellar mass-to-light ratios than in the rest of the sample. As
Fig.~\ref{fig:ml_tau} shows, these two galaxies have overall young
stellar populations. In general, one can read off from
Figs.~\ref{fig:ml_sig} and \ref{fig:ml_tau} that dynamical $\Upsilon_{\ast,\mathrm{dyn}}$
depend on $\sigma_\mathrm{eff}$ but not primarily on the stellar population age,
while $\Upsilon_\mathrm{ssp}$ mostly reflect stellar ages and do not show any
dependency on $\sigma_\mathrm{eff}$.
\subsection{Variation in the stellar IMF?}
Since dynamical and stellar population masses scale differently with
galaxy velocity dispersion (cf. Fig.~\ref{fig:ml_sig}), the ratio
$\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{ssp}$ has to vary with galaxy $\sigma_\mathrm{eff}$ (for any fixed IMF).
This is explicitly shown by the large/red symbols in
Fig.~\ref{fig:comp}. In addition to our results, the figure also
combines work from other groups.
Pentagons represent the stellar-population analysis of a subsample of
our Coma galaxies by \citet{Gri10}
(cf. Sec.~\ref{subsec:gradients}). The pentagons differ from the large
filled symbols only in terms of $\Upsilon_\mathrm{Krou}$. The velocity dispersions
and dynamical $\Upsilon_{\ast,\mathrm{dyn}}$ are the same.
Triangles are for the SAURON survey \citep{Cap06}. In terms of both,
the stellar population analysis (based on spectral absorption line
indices) as well as the dynamical modelling (orbit-based), they can be
most directly compared to the Coma galaxies of this work. Note,
however, that \citet{Cap06} measured only the sum of luminous and dark
mass, assuming the latter to contribute only a small amount of mass in
the central galaxy regions observed with SAURON ($r_\mathrm{obs,max}
\la r_\mathrm{eff}$). Adopting equivalent modelling assumptions for the Coma
galaxies yields mass-to-light ratios typically $10-20 \, \%$ higher
than compared with stellar $\Upsilon_{\ast,\mathrm{dyn}}$ from models where dark matter is
accounted for explicitly (cf. Tab.~\ref{tab:dmfrac}). The overall
distributions of $\Upsilon_{\ast,\mathrm{dyn}}/\mlkroupa$ are nevertheless similar in both samples.
\begin{figure*}\centering
\begin{minipage}{166mm}
\includegraphics[width=144mm,angle=0]{figure8.eps}
\caption{Stellar mass-to-light ratios (R-band) against average
velocity dispersion $\sigma_\mathrm{eff}$ inside $r_\mathrm{eff}$. From left-to-right:
a) dynamical $\Upsilon_{\ast,\mathrm{dyn}}$, b) stellar-population $\Upsilon_\mathrm{Salp}$ for the
Salpeter IMF and c) stellar-population $\Upsilon_\mathrm{Krou}$ for the Kroupa
IMF. Circles: E/S0 and S0 galaxies; squares: ellipticals.}
\label{fig:ml_sig}
\end{minipage}
\end{figure*}
\begin{figure*}\centering
\begin{minipage}{166mm}
\includegraphics[width=144mm,angle=0]{figure9.eps}
\caption{As Fig.~\ref{fig:ml_sig}, but stellar mass-to-light ratios
are plotted against stellar population age $\tau$.}
\label{fig:ml_tau}
\end{minipage}
\end{figure*}
Finally, Fig.~\ref{fig:comp} also includes SLACS galaxies, analysed
with a combined dynamics and lensing approach \citep{Tre10}. For these
galaxies, the ratio of the total versus the stellar mass (the latter
derived from broad-band colours) inside the Einstein radius (typically
of the order of $r_\mathrm{eff}/2$) is plotted along the y-axis.
In all the samples included in Fig.~\ref{fig:comp} dynamical (or
lensing) stellar masses systematically exceed the required masses for
a Kroupa IMF. Above $\sigma_\mathrm{eff} \ga 150 \, \mathrm{km/s}$ the ratio
$\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Krou}$ tends to increase with velocity dispersion. For
lower mass galaxies the ratio becomes uncertain due to the more
frequent presence of multiple stellar populations
(e.g. \citealt{Cap06}). Moreover, for low-mass galaxies the assumption
that all the mass follows the light is in conflict with gravitational
lensing masses (cf. Sec.~\ref{sec:projmass}).
If the increase of $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Krou}$ with $\sigma_\mathrm{eff}$ was a pure
stellar population effect then we would have to assume that the IMF is
not universal. Interpreted in this way, Fig.~\ref{fig:comp} would
imply the IMF in high-dispersion galaxies to produce either more
low-mass stars than the Kroupa IMF (i.e. being Salpeter-like) or more
stellar remnants (i.e. being top-heavy). A higher fraction of low-mass
stars would reduce the number of SNe of type II (per stellar mass) and
would lead to an overall lower metallicity. Thus, if the IMF changed
from Kroupa towards Salpeter, then the increase in $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Krou}$
would be expected to come along with a decrease in [Z/H]. Instead, a
top-heavy IMF enhances the importance of type II SNe over type Ia
SNe. Accordingly, if the IMF changed from Kroupa towards being
top-heavy then one would expect higher [$\alpha$/Fe] in galaxies with
higher $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Krou}$ \citep{Tho99,Gra10}.
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure10.eps}
\caption{Ratio $\Upsilon_{\ast,\mathrm{dyn}}/\mlkroupa$ versus velocity dispersion $\sigma$. Large,
filled: spectroscopic $\Upsilon_\mathrm{Krou}$ and orbit-based dynamical models
accounting for dark matter (this work); triangles: spectroscopic
$\Upsilon_\mathrm{Krou}$ and orbit-based dynamical models neglecting dark matter
\citep{Cap06}; pentagons: photometric $\Upsilon_\mathrm{Krou}$ \citep{Gri10} and
orbit-based dynamical models accounting for dark matter; open
squares: photometric $\Upsilon_\mathrm{Krou}$ and combined lensing + dynamics
models \citep{Tre10}. For all but the SLACS galaxies the average
velocity dispersion $\sigma_\mathrm{eff}$ inside $r_\mathrm{eff}$ is plotted. The SLACS
dispersions are the average over the spectroscopic aperture of the
SDSS survey.}
\label{fig:comp}
\end{figure}
Fig.~\ref{fig:ratml_pop} shows $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Krou}$ against stellar
population age $\tau$, metallicity [Z/H] and [$\alpha$/Fe]
ratio. There is no correlation with any stellar population parameter.
Note, however, that stellar metallicities and [$\alpha$/Fe] ratios do
not only depend on the stellar IMF, but also on the duration of the
star-formation episode(s), the depth of the galaxy potential well and
on evolutionary processes related to the cluster environment. In this
respect, the lack of evidence for an IMF change cannot be taken as a
proof for a constant IMF. Alternatives to IMF variation are discussed
in Secs.~\ref{subsec:mldynerr} and \ref{subsec:lightdm}.
\subsection{The shape of the stellar IMF}
\label{subsec:imfnorm}
A direct comparison between dynamical and stellar-population
mass-to-light ratios is provided by Fig.~\ref{fig:ml_ml_krou}. As
already clear from the different scalings of $\Upsilon_{\ast,\mathrm{dyn}}$ on the one side
and $\Upsilon_\mathrm{ssp}$ on the other (cf. Figs.~\ref{fig:ml_sig} and
\ref{fig:ml_tau}), neither the Kroupa IMF nor the Salpeter IMF yields
a close match between $\Upsilon_{\ast,\mathrm{dyn}}$ and $\Upsilon_\mathrm{ssp}$.
Above the dotted line in Fig.~\ref{fig:ml_ml_krou} the luminous
dynamical mass is larger than the stellar mass required for the Kroupa
IMF, while below the line the dynamical $\Upsilon_{\ast,\mathrm{dyn}}$ is formally
insufficient for the Kroupa IMF. For the Salpeter IMF the
corresponding limit is shifted towards $\Upsilon_{\ast,\mathrm{dyn}}$ which are a factor
$1.6$ higher (dashed line). Accordingly, all Coma galaxies are
compatible with a Kroupa IMF. The majority of the galaxies is also
consistent with a Salpeter IMF, but there is at least one galaxy for
which the dynamical $\Upsilon_{\ast,\mathrm{dyn}}$ is significantly lower than $\Upsilon_\mathrm{Salp}$ (at
about the $3 \, \sigma$ level; GMP5975). Concerning the total sample,
however, the Salpeter IMF fits the dynamical masses better than the
Kroupa IMF. The corresponding sample averages are $\langle
\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Salp} \rangle = 1.15$ for the Salpeter IMF (with an rms
scatter of $0.35$) and $\langle \Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Krou} \rangle = 1.8$ for
the Kroupa IMF, respectively.
\begin{figure*}\centering
\begin{minipage}{166mm}
\includegraphics[width=144mm,angle=0]{figure11.eps}
\caption{Ratio $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Krou}$ against stellar population age
$\tau$ (panel a), metallicity [Z/H] (panel b) and abundance ratio
[$\alpha$/Fe] (panel c).}
\label{fig:ratml_pop}
\end{minipage}
\end{figure*}
These conclusions also hold for other galaxy samples. Since $\Upsilon_\mathrm{Salp}
\approx 1.6 \times \Upsilon_\mathrm{Krou}$, the one-to-one line for the Salpeter
IMF in Fig.~\ref{fig:comp} would occur at $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{ssp} \approx
1.6$. Then, the Salpeter IMF provides {\it on average} a better match
with dynamical/lensing masses. In line with this, recent near-infrared
spectroscopic observations point towards a bottom-heavy IMF in massive
early-type galaxies as well \citep{Dok10,Dok11}. However, our
dynamical models as well as previous lensing studies
\citep{Fer10,Tre10} indicate that around $\sigma_\mathrm{eff} \la 200 \, \mathrm{km/s}$ the
Salpeter stellar masses exceed the observed dynamical and/or lensing
limits. This rules out a Salpeter IMF for low-mass galaxies.
The results for the Coma galaxies are largely independent of the
parameterisation chosen for the dark matter halos. Fits with
logarithmic halos alone yield $\langle \Upsilon_{\ast,\mathrm{dyn}}^\mathrm{LOG}/\Upsilon_\mathrm{Salp}
\rangle = 1.16$, while NFW halos result in $\langle
\Upsilon_{\ast,\mathrm{dyn}}^\mathrm{NFW}/\Upsilon_\mathrm{Salp} \rangle = 1.06$. Both are consistent
within the rms scatter (about $\approx 0.32$).
\subsection{Uncertainties in population $\Upsilon_\mathrm{ssp}$}
Gas emission can refill the H$\beta$ line and lead to an overestimate
of stellar population ages and, then, of $\Upsilon_\mathrm{ssp}$. A young stellar
subpopulation (dominating in terms of light, but not in mass) can
likewise bias stellar ages and $\Upsilon_\mathrm{ssp}$, yet towards too low
values. In any case, $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{ssp}$ would systematically decrease
with stellar population age. Fig.~\ref{fig:ratml_pop}a shows however,
that this is not the case in the Coma galaxies, such that a strong
bias due to gas emission or young stellar subpopulations is unlikely.
The stellar population parameters of the Coma galaxies are derived
from spectral indices and, thus, represent averages along the
line-of-sight. In other words, at a given radius of observation
$r_\mathrm{obs}$, the SSP parameters combine the properties of stars
with $r>r_\mathrm{obs}$. In contrast, dynamical models are most
sensitive to the mass distribution inside $r_\mathrm{obs}$. Projection
effects can therefore introduce a systematic bias between dynamical
$\Upsilon_{\ast,\mathrm{dyn}}$ and stellar population $\Upsilon_\mathrm{ssp}$ if the stellar-population
changes with radius. A monotonic stellar population gradient is
enhanced after projection along the line-of-sight but diminished in
the cumulative mass-to-light ratio constrained by dynamical models
\citep{Tho06}. More specifically, a radial increase of $\Upsilon$
leads to an underestimation of $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{ssp}$, a radial decrease of
$\Upsilon$ to an overestimation of $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{ssp}$. For a rather steep
gradient of $\mathrm{d} \log \Upsilon/ \mathrm{d} \log r = \pm 0.23$
(a change by a factor of 1.7 per decade in radius), the expected
systematic difference between $\Upsilon_{\ast,\mathrm{dyn}}$ and $\Upsilon_\mathrm{ssp}$ would amount to
$\mp 30$ percent inside $r_\mathrm{eff}$ \citep{Tho06}. The observed gradients
in the Coma galaxies are however much smaller
(cf. Fig.~\ref{fig:salpeter_data}). Therefore systematics due to
projection effects seem negligible.
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure12.eps}
\caption{Dynamical $\Upsilon_{\ast,\mathrm{dyn}}$ versus stellar-population
mass-to-light ratio $\Upsilon_\mathrm{Krou}$. The dotted line shows the
one-to-one relation for the Kroupa IMF. Since $\Upsilon_\mathrm{Salp} = 1.56
\times \Upsilon_\mathrm{Krou}$, the corresponding one-to-one relation for the
Salpeter IMF occurs at larger $\Upsilon_{\ast,\mathrm{dyn}}$ (dashed line).}
\label{fig:ml_ml_krou}
\end{figure}
\subsection{Uncertainties and interpretation of dynamical $\Upsilon_{\ast,\mathrm{dyn}}$}
\label{subsec:mldynerr}
The dynamical mass-to-light ratios $\Upsilon_{\ast,\mathrm{dyn}}$ could be affected by
systematic biases in the modelling process, e.g. arising from false
symmetry assumptions. As it has been stated in Sec.~\ref{sec:intro}
the luminous $\Upsilon_{\ast,\mathrm{dyn}}$ can be biased by a factor of up to two, if the
studied galaxies deviate significantly from the symmetry assumed in
our models \citep{Tho07A}. However, there is no evidence for the Coma
galaxies to be strongly non-axisymmetric. Firstly, they do not show
significant isophotal twists as would be indicative for
triaxiality. Secondly, as already discussed in
Sec.~\ref{sec:projmass}, the good match between dynamical and
strong-lensing masses provides further evidence that the obtained
dynamical masses are unbiased.
Even if the luminous $\Upsilon_{\ast,\mathrm{dyn}}$ are accurate, they might not represent
the galaxy {\it stellar} mass in a one-to-one fashion. Ambiguities can
come from any non-stellar mass that follows the light and contributes
to $\Upsilon_{\ast,\mathrm{dyn}}$.
Such a mass component could be gas loss during stellar evolution. It
is not included in our $\Upsilon_\mathrm{ssp}$, which only encompass the baryonic
mass locked in stars or stellar remnants. For a 10 Gyr old population
the lost gas mass amounts to about 40 percent of the originally formed
stellar mass (e.g. \citealt{Mar05}). Provided that it remains in the
galaxies and provided it follows the light distribution, it would
contribute to the dynamical mass $\Upsilon_{\ast,\mathrm{dyn}}$ and the actual baryonic mass
in stars and stellar remnants would only be $(f_\ast \Upsilon_\mathrm{dyn}) \times L$ (with
$f_\ast \approx 0.6$). Correcting the dynamical $\Upsilon_{\ast,\mathrm{dyn}}$ for stellar
mass loss yields formally a good agreement with the Kroupa IMF:
$\langle (f_\ast \Upsilon_\mathrm{dyn})/\Upsilon_\mathrm{Krou} \rangle = 1.04 \pm 0.32$.
Theoretical arguments indicate that most of the lost gas is either
expelled from the galaxies or recycled into new stars. Only a few
percent of the original stellar mass is expected to remain in hot gas
halos around the galaxies (e.g. \citealt{Cio91,Dav91}). Fittingly,
observed gas mass fractions of hot X-ray halos around massive
early-type galaxies are typically less than a percent of the present
stellar mass (e.g. \citealt{Mat01}), such that stellar mass loss is
unlikely to explain the excess $\Upsilon_{\ast,\mathrm{dyn}} > \Upsilon_\mathrm{Krou}$.
A possible component of non-baryonic matter that follows the light is
discussed below in Sec.~\ref{subsec:lightdm}.
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure13.eps}
\caption{From top to bottom: total (luminous + dark) three dimensional
mass density $\rho$, dark matter density $\rho_\mathrm{DM,dyn}$ and
luminous mass density $\rho_\mathrm{\ast,dyn}$. All densities are
spherically averaged. Left-hand panels: galaxies with less massive
halos ($v_h \le 400 \, \mathrm{km/s}$); right-hand panels: galaxies with
massive halos ($v_h > 400 \, \mathrm{km/s}$. In the middle row the dotted
lines indicate the spatial region where $\rho_\mathrm{DM,dyn} <
\rho_\mathrm{\ast,dyn}/10$. }
\label{fig:densplot}
\end{figure}
\section{Dark matter}
\label{sec:dm}
\subsection{Mass that does not follow the light}
\begin{figure*}\centering
\begin{minipage}{166mm}
\includegraphics[width=144mm,angle=0]{figure14.eps}
\caption{Comparison of the local logarithmic density slope $\xi =
\mathrm{d} \ln \rho(r)/\mathrm{d} \ln(r)$ with the density ratio
$\rho_\mathrm{DM}(r)/\rho_\mathrm{\ast,dyn}(r)$ of dark to luminous
matter at the same radius. Left panel: logarithmic slope of the
luminous mass density $\rho_\mathrm{\ast,dyn} = \Upsilon_{\ast,\mathrm{dyn}} \times \nu$; right panel:
logarithmic slope of the total mass density $\rho = \Upsilon_{\ast,\mathrm{dyn}} \times
\nu + \rho_\mathrm{DM}$. Vertical dotted lines indicate where luminous and
dark matter equalise. To the left of these lines luminous mass
dominates (inner regions of the galaxies), to the right dark matter
dominates (outer regions of galaxies). Horizontal lines are for $\xi
= -2.8$, the luminous slope at which dark matter takes over luminous
matter.}
\label{fig:slopes}
\end{minipage}
\end{figure*}
Fig.~\ref{fig:densplot} shows the spherically averaged three
dimensional density distributions of luminous and dark matter, as well
as the sum of both. The sample is subdivided into galaxies with
LOG-halo circular velocities $v_h \le 400 \, \mathrm{km/s}$ and galaxies with
$v_h > 400 \, \mathrm{km/s}$. The reason is that the latter galaxies have very
uniform outer dark and total mass density profiles. The corresponding
dark matter fractions (inside $r_\mathrm{eff}$) scatter around a mean of
$\langle f_\mathrm{DM,dyn} \rangle = 23 \pm 17 \, \%$ and do not depend on
$\sigma_\mathrm{eff}$. Note that in the central galaxy regions our models
implicitly maximise the mass contribution from the light. Similar
maximum-bulge models for lensing galaxies yield dark matter fractions
around $25 \, \%$ as well \citep{Bar09}.
Fig.~\ref{fig:slopes} shows the logarithmic slope of the luminosity
density and the total mass density in the Coma galaxies. The slope is
plotted against the ratio of dark to luminous matter densities at the
same radii. In the inner regions the slope of the total mass density
follows the light, since dark matter is negligible
($\rho_\mathrm{\ast,dyn} \gg \rho_\mathrm{DM}$), in the outer regions the total
mass density profile is flatter than the light profile. Overall, the
total mass density is roughly isothermal: $\rho \sim r^{-2}$. This
reflects the nearly flat circular velocity curves of early-type
galaxies \citep{Ger01,Tho07B}. Similar slopes for the total mass
distribution have been seen in lensing galaxies
\citep{Koo06,Bar11}. Fig.~\ref{fig:slopes} illustrates that the actual need
for a dark matter component in our models comes from the fact that the
outer mass distribution does not follow the light in early-type
galaxies. The radius where the density of dark matter takes over
luminous matter is roughly where the slope of the luminosity density
falls below $\xi \approx -2.8$ (indicated by the horizontal dotted
lines in Fig.~\ref{fig:slopes}).
\subsection{A component of dark matter that follows the light?}
\label{subsec:lightdm}
A galaxy might have more dark matter than captured by $\rho_\mathrm{DM}$ if some
fraction of the halo mass follows the light so closely that it is
mapped onto $\Upsilon_{\ast,\mathrm{dyn}}$ rather than $\rho_\mathrm{DM}$. In particular,
Fig.~\ref{fig:slopes} leaves the possibility open that this could
happen in the inner galaxy regions where the slope of the luminosity
distribution is $-2 \la \xi \la -1$. If the fraction of dark matter
that follows the light is larger in galaxies with higher $\sigma_\mathrm{eff}$,
then this would be a possible explanation for the trend between
$\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Krou}$ and $\sigma_\mathrm{eff}$ seen in Fig.~\ref{fig:comp}.
\begin{figure*}\centering
\begin{minipage}{166mm}
\includegraphics[width=144mm,angle=0]{figure15.eps}
\caption{$\Upsilon_{\ast,\mathrm{dyn}}/\mlkroupa$ against dark matter fraction $f_\mathrm{DM,dyn}$ (panel a) and
against halo core radius $r_h$ (scaled by the effective radius
$r_\mathrm{eff}$; panel b). Shifting mass from the dark halo component into
the luminous one at constant {\it total} mass moves galaxies along
the direction indicated by the arrow in the left-hand panel
(a). Note that for the right-hand panel the galaxy GMP1990 has been
omitted, since its dark matter fraction is so low that the
determination of a halo core-radius becomes meaningless.}
\label{fig:ratml_dm}
\end{minipage}
\end{figure*}
Because the {\it total} (luminous + dark) mass $M_\mathrm{tot,dyn}$ is well
constrained by the dynamical models (inside the region with
kinematical data), a spurious increase in the luminous mass component
would be accompanied by a corresponding decrease in the nominal dark
matter fraction $f_\mathrm{DM,dyn}$ of the models. More specifically, under the
assumption $M_\mathrm{tot,dyn} = M_\mathrm{tot,gal}$, the model parameters ($\Upsilon_{\ast,\mathrm{dyn}},f_\mathrm{DM,dyn}$)
and the actual galaxy parameters ($\Upsilon_{\ast,\mathrm{gal}},f_\mathrm{DM,gal}$) would be
related via
\begin{equation}
\label{eq:mldegen}
\Upsilon_{\ast,\mathrm{dyn}} = \Upsilon_{\ast,\mathrm{gal}} + \left( f_\mathrm{DM,gal} - f_\mathrm{DM,dyn} \right) \, \Upsilon_{\mathrm{tot,dyn}},
\end{equation}
where $\Upsilon_{\mathrm{tot,dyn}} \equiv \mtot/L$ is the total mass-to-light
ratio (including dark matter).
A degeneracy in the mass decomposition (at fixed {\it total} mass)
would therefore correlate the offset
\begin{equation}
\Delta \Upsilon \equiv \frac{\Upsilon_{\ast,\mathrm{dyn}} - \Upsilon_{\ast,\mathrm{gal}}}{\Upsilon_\mathrm{Krou}}
\end{equation}
in stellar mass-to-light ratios with the offset
\begin{equation}
\Delta f_\mathrm{DM} \equiv f_\mathrm{DM,dyn} - f_\mathrm{DM,gal}
\end{equation}
in dark matter fractions as
\begin{equation}
\label{eq:mlshift}
\Delta \Upsilon = - \frac{\Upsilon_{\mathrm{tot,dyn}}}{\Upsilon_\mathrm{Krou}} \Delta f_\mathrm{DM}.
\end{equation}
Fig.~\ref{fig:ratml_dm}a shows $\Upsilon_{\ast,\mathrm{dyn}}/\mlkroupa$ against the models' dark matter
fractions $f_\mathrm{DM,dyn}$ (inside $r_\mathrm{eff}$). There is a slight trend for
$\Upsilon_{\ast,\mathrm{dyn}}/\mlkroupa$ to be particularly large whenever $f_\mathrm{DM,dyn}$ is low. Note
that an intrinsic dark matter variation from galaxy to galaxy at a
constant IMF scatters galaxies horizontally in
Fig.~\ref{fig:ratml_dm}a, while an IMF variation at constant dark
matter fraction scatters galaxies vertically. The luminous-dark matter
degeneracy discussed above scatters galaxies along the arrow shown in
Fig.~\ref{fig:ratml_dm}a. It marks the direction along which the fitted
($\Upsilon_{\ast,\mathrm{dyn}},f_\mathrm{DM,dyn}$) are expected to separate from the galaxies'
($\Upsilon_{\ast,\mathrm{gal}},f_\mathrm{DM,gal}$) according to equation (\ref{eq:mlshift}) and --
for each galaxy -- depends on the ratio $\Upsilon_{\mathrm{tot,dyn}}/\Upsilon_\mathrm{Krou}$. For the
arrow in Fig.~\ref{fig:ratml_dm}a we have used the average $\langle
\Upsilon_{\mathrm{tot,dyn}}/\Upsilon_\mathrm{Krou} \rangle = 2.39$ over the Coma sample. The
distribution of the majority of Coma galaxies roughly follows the
arrow, in particular below $f_\mathrm{DM,dyn} \la 0.3$.
Fig.~\ref{fig:ratml_dm}a suggests, though does not unambiguously prove,
that the scatter in $\Upsilon_{\ast,\mathrm{dyn}}/\mlkroupa$ could reflect a degeneracy in the
dynamical mass decomposition. A contamination of some $\Upsilon_{\ast,\mathrm{dyn}}$ with
dark matter would also explain the trend seen in Fig.~\ref{fig:ratml_dm}b,
where $\Upsilon_{\ast,\mathrm{dyn}}/\mlkroupa$ is plotted against the halo core-radius $r_h$ (in units
of $r_\mathrm{eff}$). The more the inner dark matter goes into $\Upsilon_{\ast,\mathrm{dyn}}$, the
less the model component $\rho_\mathrm{DM}$ traces the actual inner dark matter
of the galaxies. Instead, it only represents the outer parts of the
dark matter halos. Correspondingly, one would expect relatively larger
halo core-radii (with respect to $r_\mathrm{eff}$) whenever $\Upsilon_{\ast,\mathrm{dyn}}$ is large
compared to $\Upsilon_\mathrm{ssp}$ -- as seen in Fig.~\ref{fig:ratml_dm}b.
\citet{Nap10} analysed a large set of early-type galaxies with
two-component spherical Jeans models and find dynamical $\Upsilon_{\ast,\mathrm{dyn}}$ close
to the Salpeter IMF when using collisionless halos from cosmological
simulations without baryon contraction. However, their $\Upsilon_{\ast,\mathrm{dyn}}$ get
closer to the Kroupa IMF when baryonic contraction is taken into
account. Baryonic contraction might therefore be one way to make
luminous and dark matter distributions similar enough to explain the
difference between $\Upsilon_{\ast,\mathrm{dyn}}$ and $\Upsilon_\mathrm{Krou}$ (see
Sec.~\ref{subsec:degdis}).
In our modelling approach, the halo parameters are allowed to vary
freely, without being connected to results from cosmological
simulations. On the one hand this ensures that baryonic contraction is
implicitly taken into account: the best-fit models for more contracted
halos are simply expected to occur in a different region of parameter
space. On the other hand, the steepest halo density profiles that we
probed are those from cosmological simulations without baryon
contraction (NFW halos; cf. Sec.~\ref{subsec:dynmod}). If the actual
galaxy halo profiles are steeper, then the best-fit dynamical model
would still be obtained by shifting some fraction of the inner dark
mass into $\Upsilon_{\ast,\mathrm{dyn}}$.
More similar distributions of luminous and dark matter in some
galaxies than in others could also reflect differences in their
evolutionary histories. For example, the cosmological simulations of
\citet{Naa09} indicate a difference in the radial distribution of
in-situ formed stars relative to stars that were accreted during
mergers. In-situ formed stars have a more centrally concentrated
radial distribution than stars that were accreted in collisionless
mergers. The latter dominate the stellar mass density around
$r_\mathrm{eff}$. In any case, more detailed investigations of numerical
simulations are required to conclude about a possible degeneracy
between luminous and dark matter in the inner regions of galaxies.
\subsection{The distribution of dark matter in case of a universal Kroupa IMF}
If we adopt the point of view that the stellar IMF in early-type
galaxies is universal and Kroupa-like then this affects the
distribution of dark matter significantly. The reason is that in this
case our nominal halo component captures only a part of the galaxies'
dark matter, while a large fraction follows the light and is included
in $\Upsilon_{\ast,\mathrm{dyn}}$. In fact, a universal Kroupa IMF implies the dark matter
fractions $f_\mathrm{DM,Krou}$ to read
\begin{equation}
\label{eq:dmfrack}
f_\mathrm{DM,Krou}(r) \equiv f_\mathrm{DM,dyn}(r) + \frac{(\Upsilon_{\ast,\mathrm{dyn}}-\Upsilon_\mathrm{Krou}) \times L(r)}{M_\mathrm{tot,dyn}(r)}.
\end{equation}
These fractions are larger than the nominal $f_\mathrm{DM,dyn}$ derived from
$\rho_\mathrm{DM}$, have a smaller scatter and slightly increase with
galaxy $\sigma_\mathrm{eff}$ (cf. Fig.~\ref{fig:dmkroupa}).
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure16.eps}
\caption{Dark matter fractions (inside the effective radius $r_\mathrm{eff}$)
against velocity dispersion $\sigma_\mathrm{eff}$. Open circles: fraction of dark matter
$f_\mathrm{DM,dyn}$ that does not follow the light; filled circles: dark matter fraction
$f_\mathrm{DM,Krou}$ assuming i) a universal Kroupa IMF and ii) that the
excess mass $(\Upsilon_{\ast,\mathrm{dyn}}-\Upsilon_\mathrm{Krou}) \times L$ is a component of dark
matter that follows the light.}
\label{fig:dmkroupa}
\end{figure}
Fig.~\ref{fig:rhokroupa} shows the spherically averaged dark matter
density profiles
\begin{equation}
\label{eq:rhodmk}
\rho_\mathrm{DM,Krou} \equiv \rho_\mathrm{DM} + (\Upsilon_{\ast,\mathrm{dyn}}-\Upsilon_\mathrm{Krou}) \times \nu
\end{equation}
including the extra dark matter required for a Kroupa IMF
(cf. equation \ref{eq:massdecomp}). These density profiles are smooth
and close to a power-law with logarithmic slope slightly shallower
than -2. The slight wiggles in the profiles around $\approx 5 \, \mathrm{kpc}$
indicate the transition from the outer parts, which are dominated by
$\rho_\mathrm{DM}$, to the inner parts, which are dominated by the second term
on the right-hand side of equation (\ref{eq:rhodmk}).
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure17.eps}
\caption{Spherically averaged dark matter density $\rho_\mathrm{DM,Krou}$ against
radius. The excess mass $(\Upsilon_{\ast,\mathrm{dyn}}-\Upsilon_\mathrm{Krou}) \times L$ with respect
to a Kroupa IMF stellar population has been added to the dark matter
halo. The dashed line indicates a power-law density with
logarithmic slope of -2.}
\label{fig:rhokroupa}
\end{figure}
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure18.eps}
\caption{Star formation epoch (x-axis) against halo assembly epoch
(y-axis). The dotted line represents the one-to-one relation.}
\label{fig:zz}
\end{figure}
\subsection{Dark matter density and halo assembly epoch}
\label{subsec:degdis}
In \citet{Tho09} we estimated halo assembly epochs $z_\mathrm{form}$
based on the assumption that the average dark matter density $\langle
\rho_\mathrm{DM} \rangle$ inside $2 \, r_\mathrm{eff}$ scales with
$(1+z_\mathrm{form})^3$. From the overdensity of dark matter in
early-type relative to spiral galaxies one can then narrow down
elliptical galaxy assembly redshifts with an additional assumption
about the typical formation redshift of spiral galaxies
($z_\mathrm{form} \approx 1$). Adopting a universal Kroupa IMF
results in average dark matter densities a factor of $\approx 3$
larger than the nominal ones of the dynamical models. The strongest
contribution to this increase comes from the single galaxy GMP1990
which has a negligible $f_\mathrm{DM,dyn}$, but a large $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Krou}$.
Without this galaxy, dark matter densities are only larger by about a
factor of $1.6$, which lets the halo assembly redshifts increase from
$z_\mathrm{form} \approx 1 - 3 $ \citep{Tho09} to about
$z_\mathrm{form} \approx 1.5 - 3.5$.
As a second test, we have compared the $\rho_\mathrm{DM,Krou}$ from
Fig.~\ref{fig:rhokroupa} directly to the galaxy formation models of
\citet{deL07}. These models do not include the dynamical reaction of
dark matter on the baryon infall. Therefore, we first subtracted from
the observed dark matter profile $\rho_\mathrm{DM,Krou}$ the expected effect of
baryon contraction by assuming the adiabatic approximation (i.e. by
inverting the equations of \citealt{Blu86}). The relationship between
the average dark matter density and halo formation redshift in the
models of \citet{deL07} is well fitted by $\log \langle
\rho_\mathrm{DM} \rangle \approx -2.9 + \log (1+z_\mathrm{form})^3$.
We used this relation to calculate Coma galaxy assembly redshifts from
the decontracted Kroupa dark matter halo densities. Fig.~\ref{fig:zz}
shows that these halo assembly redshifts cover a similar range
($z_\mathrm{form} \approx 1 - 3$) as the ones obtained through the
comparison with spiral galaxy dark matter densities. Note that we here
computed star-formation redshifts from the average stellar ages inside
$r_\mathrm{eff}$, while in \citet{Tho09} we used the central stellar ages from
\citet{Meh03}.
For many of the Coma galaxies, star formation redshifts and dark halo
assembly redshifts are similar. Two galaxies appear far to the left of
the one-to-one relation. These are GMP0756 and GMP1176, which have
extended and relatively young ($\tau \approx 3 \, \mathrm{Gyr}$) stellar
disks. The galaxy on the bottom-right, where the stars seem
significantly older than the halo is GMP5568. As discussed in
\citet{Tho09}, dry mergers can move galaxies below the one-to-one
relation in Fig.~\ref{fig:zz}. Besides this possibility the effective
radius of GMP5568 is exceptionally large ($r_\mathrm{eff} \approx 27 \, \mathrm{kpc}$).
Therefore, the dark matter density refers to a spatial region
significantly larger than in any of our other Coma galaxies. If the
effective radius of this galaxy was overestimated, then the average
dark matter density would be underestimated and so would be the halo assembly
redshift. This could also explain the high dark matter fraction of GMP5568
(cf. Tab.~\ref{tab:dmfrac}), since it generally increases with the physical distance
from the galaxy centre.
\section{The tilt of the fundamental plane}
\label{sec:fp}
The effective radius $r_\mathrm{eff}$, mean surface brightness $\langle I \rangle_\mathrm{eff}$ inside
$r_\mathrm{eff}$ and the central velocity dispersion $\sigma_0$ of a galaxy are
connected via
\begin{equation}
\label{fp1}
\sigma_0^2 = c_M \frac{G \, M}{r_\mathrm{eff}}
\end{equation}
where $M$ is the total mass and
\begin{equation}
\label{fp2}
\langle I \rangle_\mathrm{eff} = \frac{L}{2 \pi \, r_\mathrm{eff}^2}
\end{equation}
where $L$ is the total luminosity. In virial equilibrium the structure
coefficient $c_M$ depends on the orbital structure and the radial
distributions of the mass and the tracer population. For a homologous
family of dynamical objects $c_M$ is a constant and with $\langle \Upsilon \rangle \equiv
M/L$ equations (\ref{fp1}) and (\ref{fp2}) lead to
\begin{equation}
\label{eq:FPvir}
\log \frac{r_\mathrm{eff}}{\mathrm{kpc}} = 2 \, \log \frac{\sigma_0}{\mathrm{km/s}}
- \log \frac{\langle I \rangle_\mathrm{eff}}{\mathrm{L}_\odot \mathrm{pc}^{-2}} + \gamma
\end{equation}
and
\begin{equation}
\label{fp3}
\gamma = - \log \frac{\langle \Upsilon \rangle}{\mathrm{M}_\odot/\mathrm{L}_{R,\odot}} - \log \left( 2 \pi \, \frac{G}{(\kms)^2 \, \kpc \, \mathrm{M}_\odot^{-1}} c_M \right).
\end{equation}
The actually observed fundamental plane (FP; \citealt{Djo87,Dre87}) of
early-type galaxies reads
\begin{equation}
\label{eq:FP}
\log \frac{r_\mathrm{eff}}{\mathrm{kpc}} = \alpha \, \log \frac{\sigma_0}{\mathrm{km/s}}
+ \beta \, \log \frac{\langle I \rangle_\mathrm{eff}}{\mathrm{L}_\odot \mathrm{pc}^{-2}} + \gamma
\end{equation}
with $\alpha \ne 2$ and $\beta \ne -1$. It is tilted with respect to
the virial plane of equation (\ref{eq:FPvir}).
\begin{figure*}\centering
\begin{minipage}{166mm}
\includegraphics[width=144mm,angle=0]{figure19.eps}
\caption{Fundamental plane and mass plane of 16 Coma galaxies. In each
panel the best-fit parameters of an orthogonal fit and the
rms-scatter in $\log \, r_\mathrm{eff}$ are quoted. a) Traditional FP; b)
fundamental mass plane from stellar populations, i.e. the effective
surface brightness is $\langle I \rangle_\mathrm{eff}$ is replaced by the mass-density
$\langle \Sigma \rangle_\mathrm{eff} \equiv \Upsilon_\mathrm{Krou} \times \langle I \rangle_\mathrm{eff}$; c) as b) but for the
dynamically derived stellar-population $\Upsilon_{\ast,\mathrm{dyn}}$: $\langle \Sigma \rangle_{\ast,\mathrm{dyn}} \equiv
\Upsilon_{\ast,\mathrm{dyn}} \times \langle I \rangle_\mathrm{eff}$; d) as c), but instead of the dynamically
derived {\it stellar} mass-to-light ratio the {\it total} $\mtot/L$
(including dark matter) is used: $\langle \Sigma \rangle_{\mathrm{tot,dyn}} \equiv \mtot/L \times
\langle I \rangle_\mathrm{eff}$ ($\mtot/L$ is taken at the effective radius). Solid lines
trace the one-to-one relation. }
\label{fig:FPall}
\end{minipage}
\end{figure*}
\begin{figure*}\centering
\begin{minipage}{166mm}
\includegraphics[width=144mm,angle=0]{figure20.eps}
\caption{As Fig.~\ref{fig:FPall}, but four galaxies with young stellar
cores are omitted.}
\label{fig:FPyoung}
\end{minipage}
\end{figure*}
Fig.~\ref{fig:FPall}a shows the FP\footnote{Note that in this section
we use the average velocity dispersion $\sigma_0$ inside the central
$2\arcsec$, derived in the same way as the effective $\sigma_\mathrm{eff}$
discussed in Sec.~\ref{sec:projmass}.} of Coma galaxies. The
best-fit parameters of an orthogonal fit are $\alpha = 1.26 \pm 0.58$
and $\beta = -0.72 \pm 0.10$ (bootstrap errors). Within the
statistical uncertainties, the fit is consistent with the $r$-band FP
of SDSS early-type galaxies ($\alpha = 1.49 \pm 0.05$, $\beta = -0.75
\pm 0.01$; \citealt{Ber03,Hyd09}). The larger errors result from the
smaller sample size. Most of the difference with respect to
\citet{Ber03} goes back to four galaxies (GMP0144, GMP0756, GMP1176,
and GMP5975) that are distinct from the rest of the sample in several
respects: (1) they have young central stellar cores ($\tau_0 < 7 \,
\mathrm{Gyr}$; cf. \citealt{Meh03}); (2) at least two of them have an extended
thin stellar disk; (3) they follow different dark halo scaling
relations \citep{Tho09}. Fig.~\ref{fig:FPyoung} is equivalent to
Fig.~\ref{fig:FPall} except that these four galaxies have been
removed. The corresponding FP matches well with the SDSS results.
\subsection{The fundamental mass plane}
The tilt in the FP can reflect non-homology (i.e. $c_M$ being a
function of galaxy mass) or variations in $\langle \Upsilon \rangle$ (or both). A
systematic variation of $\langle \Upsilon \rangle$, in turn, could reflect stellar
population effects or changes in the dark matter distribution. In any
case, the dynamical models allow to incorporate variations of $\langle \Upsilon \rangle$
into the FP. For this purpose let $\langle \Sigma
\rangle_\mathrm{eff}$ denote the average surface mass density inside
$r_\mathrm{eff}$. Then, equation (\ref{eq:FP}) can be written
\begin{equation}
\label{eq:MP}
\log \frac{r_\mathrm{eff}}{\mathrm{kpc}} = \alpha \, \log \frac{\sigma_0}{\mathrm{km/s}}
+ \beta \, \log \frac{\langle \Sigma \rangle_\mathrm{eff}}{\mathrm{M}_\odot \mathrm{pc}^{-2}}
+ \gamma,
\end{equation}
where
\begin{equation}
\label{fp4}
\gamma = - \log
\left( 2 \pi \, \frac{G}{(\kms)^2 \, \kpc \, \mathrm{M}_\odot^{-1}} c_M \right).
\end{equation}
Equation (\ref{eq:MP}) defines the so-called fundamental mass plane
\citep{Bol07}.
Figs.~\ref{fig:FPall}b - \ref{fig:FPall}d show Coma galaxy MPs for
different choices of $\langle \Sigma \rangle_\mathrm{eff}$. Using
either $\langle \Sigma \rangle_\mathrm{Krou} \equiv \Upsilon_\mathrm{Krou} \times
\langle I \rangle_\mathrm{eff}$ (panel b) or $\langle \Sigma \rangle_\mathrm{\ast,dyn} \equiv
\Upsilon_{\ast,\mathrm{dyn}} \times \langle I \rangle_\mathrm{eff}$ (panel c), the tilt is reduced, but does not
vanish ($\beta \ne -1$). The change in $\alpha$ from
Fig.~\ref{fig:FPall}a to Fig.~\ref{fig:FPall}b is consistent with the
SDSS analysis of \citet{Hyd09}. Fig.~\ref{fig:FPall}d shows the case
of $\langle \Sigma \rangle_\mathrm{eff}$ including all the projected
mass (luminous and dark) inside $r_\mathrm{eff}$, respectively. The mass plane
of Fig.~\ref{fig:FPall}d is tilt-free within the errors. The match to
the virial plane of equation (\ref{eq:FPvir}) improves further when
omitting the four galaxies harbouring young stellar cores
(cf. Fig.~\ref{fig:FPyoung}). With $\langle \Upsilon \rangle \equiv \mtot/L$ the MP has
the same scatter as the FP itself (cf. Figs.~\ref{fig:FPall}a and
d). For the subsample of old Coma galaxies the scatter slightly
increases (cf. Fig.~\ref{fig:FPyoung}).
\subsection{The tilt of the fundamental plane}
The tilt in the FP is reduced when the effective surface brightness is
replaced by an effective surface mass derived from
$\Upsilon_\mathrm{ssp}$\footnote{In Sec.~\ref{sec:fp} the surface mass was derived
from the Kroupa IMF, but the difference between the Kroupa and the
Salpeter IMF is only a constant scaling factor. The results for the
Salpeter IMF are therefore similar.}.
The amount of reduction is consistent with the SDSS results of \citet{Hyd09}.
In accordance with \citet{Gra10}, the tilt is further reduced if
$\Upsilon_\mathrm{ssp}$ is replaced by the dynamical $\Upsilon_{\ast,\mathrm{dyn}}$. As discussed above,
the different scalings of $\Upsilon_{\ast,\mathrm{dyn}}$ and
$\Upsilon_\mathrm{ssp}$ with $\sigma_\mathrm{eff}$ could reflect (1) changes in the IMF or (2) changes
in the distribution of dark matter. The absence of any correlation between $\Upsilon_{\ast,\mathrm{dyn}}/\mlkroupa$ and
stellar population parameters makes (2) more likely than (1).
Finally, the tilt vanishes completely (for a subsample of Coma
galaxies with uniformly old stellar populations), if the remaining dark mass inside
the effective radius $r_\mathrm{eff}$ is taken into account as well. This behaviour is also found
in lensing studies \citep{Bol07}.
In our FP analysis we have not tried to calculate $c_M$ directly from
the dynamical models. However, the density distribution (in particular
the galaxy flattening) as well as the orbital structure vary among the
Coma galaxies \citep{Tho09a}. Nevertheless, Fig.~\ref{fig:FPyoung}d
indicates that for old Coma early-types the tilt of the FP is
dominated by mass-to-light ratio effects rather than any possible
variation in $c_M$.
\subsection{Mass estimators}
\label{subsec:virest}
\cite{Wol10} provide an estimation for the mass
\begin{equation}
\label{eq:Wol1}
M_\mathrm{W10}(r_3) = \frac{3 \langle \sigma^2 \rangle r_3}{G}
\end{equation}
inside the radius $r_3$ where the logarithmic density slope of the
tracer population is $\xi=-3$. Here, $\langle \sigma^2 \rangle$ is the
average of the projected $\sigma^2$ over the whole
galaxy. \citet{Wol10} derived eq.~(\ref{eq:Wol1}) for non-rotating
spherical galaxies. For the Coma galaxies, we averaged the measured
$\sigma^2$ up to $r_\mathrm{eff}$, ignoring the galaxies' rotation velocities.
Fig.~\ref{fig:magic_wolf}a compares Coma galaxy masses $M_\mathrm{tot,dyn}(r_3)$
against the predictions of eq.~(\ref{eq:Wol1}). Averaged over all Coma
galaxies we find $\langle M_\mathrm{tot,dyn}(r_3)/M_\mathrm{W10}(r_3) \rangle =
1.03 \pm 0.27$ (rms scatter). Fig.~\ref{fig:magic_wolf}b is similar to
Fig.~\ref{fig:magic_wolf}a, except for the additional approximation
$r_3 \approx r_{1/2} \approx 4/3 \, r_\mathrm{eff}$, where $r_{1/2}$ is the
deprojected half-light radius. The agreement with the Coma galaxies is
still very good.
Fig.~\ref{fig:magic_cap} compares luminous dynamical masses with the
estimator
\begin{equation}
\label{virialest}
M_\mathrm{C06} = \frac{5 \sigma^2_\mathrm{eff} r_\mathrm{eff}}{G}.
\end{equation}
from \citet{Cap06}. When using luminous dynamical masses $M_{\ast,\mathrm{dyn}} =
\Upsilon_{\ast,\mathrm{dyn}} \times L$ from model fits that do have a separate dark matter
component, then we find $\langle M_{\ast,\mathrm{dyn}}/M_\mathrm{C06} \rangle = 0.86
\pm 0.35$ (cf. Fig.~\ref{fig:magic_cap}a). The small offset
disappears if we use model fits that do not have an additional dark
matter component (equivalent to the approximation made in
\citealt{Cap06}). The corresponding luminous dynamical masses $M_{\ast,\mathrm{sc}} =
\Upsilon_{\ast,\mathrm{sc}} \times L$ (cf. Tab.~\ref{tab:dmfrac}) are slightly larger and
the comparison with eq.~(\ref{virialest}) yields $\langle
M_{\ast,\mathrm{sc}}/M_\mathrm{C06} \rangle = 1.01 \pm 0.36$
(cf. Fig.~\ref{fig:magic_cap}). The rms scatter includes both the
measurement errors and anisotropy variations \citep{Tho07B}. Note,
however, that the assumption that all the mass follows the light is
inconsistent with lensing masses (cf. Fig.~\ref{fig:projmass_tot}).
Thus, the \cite{Wol10} formula gives good estimates for the {\it
total} dynamical mass inside a radius which is a bit larger than
$r_\mathrm{eff}$. The virial estimator of \citet{Cap06} captures the entire
dynamical mass that results under the assumption that {\it total mass
follows light}. Assuming $r_3 \approx r_{1/2}$,
eq.~(\ref{virialest}) implies
\begin{equation}
\label{virialest2}
M_\mathrm{C06}(r_3) \approx M_\mathrm{C06}(r_{1/2}) = \frac{M_\mathrm{C06}}{2} = \frac{2.5 \sigma^2_\mathrm{eff} r_\mathrm{eff}}{G},
\end{equation}
whereas from eq.~\ref{eq:Wol1} and $r_3 \approx 4/3 \, r_\mathrm{eff}$ it
follows
\begin{equation}
M_\mathrm{W10}(r_3) \approx \frac{4 \langle \sigma^2 \rangle r_\mathrm{eff}}{G}.
\end{equation}
Concerning the enclosed mass inside $4/3r_\mathrm{eff}$, the two mass
estimators would differ by a factor $4/2.5 = 1.6$, unless $\langle
\sigma^2 \rangle \ne \sigma^2_\mathrm{eff}$. Due to the different
treatment of rotation $\sigma^2_\mathrm{eff} \approx 1.3 \, \langle
\sigma^2 \rangle$ in the Coma sample, such that the actual difference
is only about 20 percent.
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure21.eps}
\caption{Comparison of Coma galaxies with the \citet{Wol10} mass
estimator (cf. eq.~\ref{eq:Wol1}): (a) total dynamical mass enclosed
inside the radius $r_3$, where the logarithmic slope of the
luminosity distribution equals $\xi = -3$; (b) is similar to (a) but
with the additional approximation $r_3 \approx 4/3 \, r_\mathrm{eff}$.}
\label{fig:magic_wolf}
\end{figure}
\begin{figure}\centering
\includegraphics[width=84mm,angle=0]{figure22.eps}
\caption{Comparison of Coma galaxies with the \citet{Cap06} virial
mass estimator (cf. eq.~\ref{virialest}): (a) luminous dynamical
mass $\Upsilon_{\ast,\mathrm{dyn}} \times L$ from fits explicitly allowing for a separate
component of dark matter; (b) luminous dynamical mass
$\Upsilon_{\ast,sc} \times L$ from fits assuming that all the mass
follows the light.}
\label{fig:magic_cap}
\end{figure}
\section{Summary}
\label{sec:sum}
We compared dynamically derived stellar mass-to-light ratios $\Upsilon_{\ast,\mathrm{dyn}}$
with completely independent results from simple stellar population
models. Our dynamical models are based on Schwarzschild's orbit
superposition technique and have two mass components. One follows the
light and its mass-to-light ratio $\Upsilon_{\ast,\mathrm{dyn}}$ is assumed to approximate
the stellar mass distribution. The other mass component explicitly
accounts for dark matter. This way, any potential degeneracy between
the stellar mass and the dark matter halo is minimised. The Coma
galaxy sample studied here is currently the largest with axisymmetric
Schwarzschild models including dark matter explicitly.
Intrinsic uncertainties in the modelling, in particular related to the
assumption of axial symmetry, are unlikely to bias our results
significantly. The main reason is that in projection, our dynamical
masses match well with completely independent results from strong
gravitational lensing.
Our main findings are:
\begin{enumerate}
\item For galaxies with low velocity dispersions ($\sigma_\mathrm{eff} \approx 200
\, \mathrm{km/s}$), the assumption that all the mass follows the light yields
projected masses larger than in comparable strong-gravitational lens
systems.
\item In high-velocity dispersion galaxies ($\sigma_\mathrm{eff} \approx 300 \,
\mathrm{km/s}$) the assumption that mass follows light is consistent with
strong lensing results.
\item Two-component dynamical models with an explicit dark halo
component yield total projected masses that are in good agreement
with results from strong gravitational lensing for all galaxies.
\item In two-component models, the mass-to-light ratio $\Upsilon_{\ast,\mathrm{dyn}}$ of the
component that follows the light increases with galaxy velocity
dispersion $\sigma_\mathrm{eff}$.
\item Stellar population $\Upsilon_\mathrm{ssp}$ (for any fixed IMF) are largely
independent of $\sigma_\mathrm{eff}$. As a result, the ratio $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{ssp}$ of
luminous dynamical mass over stellar population mass increases with
galaxy velocity dispersion.
\item The luminous dynamical $\Upsilon_{\ast,\mathrm{dyn}}$ is always larger than, or at
least equalises, the stellar-population mass-to-light ratio
$\Upsilon_\mathrm{Krou}$ for a Kroupa IMF.
\item There is no correlation between $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{ssp}$ and stellar
population age, metallicity or [$\alpha$/Fe] ratio.
\item Inside $r_\mathrm{eff}$, the average fraction of dark matter (that does
not follow the light) is $f_\mathrm{DM,dyn} = 28 \pm 17 \, \%$ in the Coma
galaxies.
\item The tilt of the FP reduces if the effective surface brightness
$\langle I \rangle_\mathrm{eff}$ is replaced by the stellar population surface-mass density
$\Upsilon_\mathrm{ssp} \times \langle I \rangle_\mathrm{eff}$, further reduced if $\langle I \rangle_\mathrm{eff}$ is replaced by the
dynamical stellar surface-mass density $\Upsilon_{\ast,\mathrm{dyn}} \times \langle I \rangle_\mathrm{eff}$ and,
for a subsample of galaxies with uniformly old stellar populations,
vanishes completely with the {\it total} dynamical surface-mass
density $\langle \Sigma_\mathrm{tot,dyn} \rangle_\mathrm{eff}$.
\item Commonly used mass estimators are accurate to the $20 - 30 \,
\%$ level.
\end{enumerate}
The implications of these findings are as follows:
\begin{enumerate}
\item That luminous dynamical masses increase more rapidly with galaxy
velocity dispersion than stellar-population masses for a fixed IMF
could be due to a change in the IMF or due to an increasing amount
of dark matter following a spatial distribution similar to that of
the light.
\item If the IMF changes, then massive early-types ($\sigma_\mathrm{eff} \approx
300 \, \mathrm{km/s}$) have up to two times more stellar mass per stellar
light than lower-mass galaxies ($\sigma_\mathrm{eff} \approx 200 \, \mathrm{km/s}$),
which are consistent with a Kroupa IMF. However, the lack of any
correlation between $\Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{ssp}$ and stellar population age,
metallicity or [$\alpha$/Fe] ratio is consistent with, though does
not prove, that the IMF is actually universal.
\item If the IMF is universal, then the increase in luminous dynamical
masses must primarily come from a component of dark matter that
follows the light very closely and is more important in more massive
galaxies. The IMF would be consistent with being Kroupa in all
early-types.
\item Independent of the actual slope of the stellar IMF, luminous
dynamical masses are {\it on average} more accurately predicted by
assuming a Salpeter IMF: $\langle \Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Salp} \rangle = 1.15$,
but these masses may not represent exclusively stars. The Kroupa IMF
yields $\langle \Upsilon_{\ast,\mathrm{dyn}}/\Upsilon_\mathrm{Krou} \rangle = 1.8$.
\item Adopting a Kroupa IMF and counting the excess mass
$(\Upsilon_{\ast,\mathrm{dyn}}-\Upsilon_\mathrm{Krou}) \times L$ as dark matter that follows the light
doubles the average dark matter fractions inside $r_\mathrm{eff}$ to about
$f_\mathrm{DM,Krou} \approx 55 \pm 12 \, \%$. Moreover, it yields a smooth
trend between the resulting $f_\mathrm{DM,Krou}$ and galaxy velocity
dispersion and, also, smooth dark matter halo profiles.
\item The FP tilt is not a pure stellar population effect. Further
inferences about the tilt depend on the interpretation of the
observed $\Upsilon_{\ast,\mathrm{dyn}}/\mlkroupa$. As above, that the tilt reduces when considering
the dynamical mass $\Upsilon_{\ast,\mathrm{dyn}} \times L$ that follows the light could be
due to (1) variations in the relative distribution of luminous and
dark matter or (2) IMF variability.
\end{enumerate}
\section*{Acknowledgements}
We thank the referee Glenn van de Ven for useful comments that
helped us to improve the presentation of the results. JT acknowledges
financial support by the DFG through SFB 375 ``Astro-Teilchenphysik''.
RPS acknowledges support by the DFG Cluster of Excellence ``Origin and
Structure of the Universe''. EMC is supported by the University of
Padua through grants 60A02-1283/10 and CPDA089220 and by the Ministry
of Education, University, and Research (MIUR) through grant EARA
2004-2006. SS acknowledges support by the TR33 "The Dark Universe"
and by the Cluster of Excellence ``Origin and Structure of the
Universe''. The new Coma galaxy HST observations were obtained
through the program HSTGO-10884.0-A which was provided by NASA through
a grant from the Space Telescope Science Institute, which is operated
by the Association of Universities for Research in Astronomy,
Inc. under NASA contract NAS5-26555. Funding for the SDSS and SDSS-II
has been provided by the Alfred P. Sloan Foundation, the Participating
Institutions, the National Science Foundation, the U.S. Department of
Energy, the National Aeronautics and Space Administration, the
Japanese Monbukagakusho, the Max Planck Society and the Higher
Education Funding Council for England. The SDSS Web Site is
http://www.sdss.org/. The SDSS is managed by the Astrophysical
Research Consortium for the Participating Institutions. The
Participating Institutions are the American Museum of Natural History,
Astrophysical Institute Potsdam, University of Basel, University of
Cambridge, Case Western Reserve University, University of Chicago,
Drexel University, Fermilab, the Institute for Advanced Study, the
Japan Participation Group, Johns Hopkins University, the Joint
Institute for Nuclear Astrophysics, the Kavli Institute for Particle
Astrophysics and Cosmology, the Korean Scientist Group, the Chinese
Academy of Sciences (LAMOST), Los Alamos National Laboratory, the
Max-Planck-Institute for Astronomy (MPIA), the Max-Planck-Institute
for Astrophysics (MPA), New Mexico State University, Ohio State
University, University of Pittsburgh, University of Portsmouth,
Princeton University, the United States Naval Observatory and the
University of Washington.
\bibliographystyle{mn2e}
|
\section{\textrm{Introduction}}
The paper studies the weak Euler approximation for solutions to SDEs driven
by L\'{e}vy processes with a nondegenerate main part. The goal is to
investigate the dependence of the convergence rate on the regularity of
coefficients and driving processes. We use the method developed in \cit
{mikz1} and \cite{mikz2}. For the sake of completeness we repeat some
arguments. A methodical novelty is that contrary to \cite{mikz1} and \cit
{mikz2} we do not use Fourier transform. Also, the whole H\"{o}lder-Zygmund
scale is covered.
\subsection{Nondegenerate SDEs Driven by L\'{e}vy Processes}
Let $(\Omega ,\mathcal{F},\mathbf{P})$ be a complete probability space with
a filtration $\mathbb{F}=\{\mathcal{F}_{t}\}_{t\in \lbrack 0,T]}$ of $\sigma
$-algebras satisfying the usual conditions and $\alpha \in (0,2]$ be fixed.
Consider the following model in $\mathbf{R}^{d}$:
\begin{equation}
X_{t}=X_{0}+\int_{0}^{t}a(X_{s})ds+\int_{0}^{t}b(X_{s-})dU_{s}^{\alpha
}+\int_{0}^{t}G(X_{s-})dZ_{s},t\in \lbrack 0,T], \label{one}
\end{equation
where $a(x)=(a^{i}(x))_{1\leq i\leq d}$, $b(x)=(b^{ij}(x))_{1\leq i,j\leq d}
, $G(x)=(G^{ij}(x))_{1\leq i\leq d,1\leq j\leq m}$, $x\in \mathbf{R}^{d}$
are measurable and bounded, with $a=0$ if $\alpha \in (0,1)$ and $b$ being
nondegenerate. The main part of the equation is driven by $U^{\alpha
}=\{U_{t}^{\alpha }\}_{t\in \lbrack 0,T]}$, a standard $d$-dimensional
spherically-symmetric $\alpha $-stable process
\begin{equation*}
U_{t}^{\alpha }=\int_{0}^{t}\int (1-\bar{\chi}_{\alpha
}(y))yp_{0}(ds,dy)+\int_{0}^{t}\int \bar{\chi}_{\alpha
}(y)yq_{0}(ds,dy),\alpha \in (0,2),
\end{equation*
where $\bar{\chi}_{\alpha }(y)=\mathbf{1}_{\{\alpha \in (1,2)\}}+\mathbf{1
_{\{\alpha =1\}}\chi _{\left\{ |y|\leq 1\right\} }$ and $~p_{0}(dt,dy)$ is a
Poisson point measure on $[0,\infty )\times \mathbf{R}_{0}^{d}$ ($\mathbf{R
_{0}^{d}=\mathbf{R}^{d}\backslash \left\{ 0\right\} $) with
\begin{equation*}
\mathbf{E}[p_{0}(dt,dy)]=\frac{dtdy}{|y|^{d+\alpha }},\quad
q_{0}(dt,dy)=p_{0}(dt,dy)-\frac{dtdy}{|y|^{d+\alpha }}.
\end{equation*
If $\alpha =2$, $U^{\alpha }$ is the standard Wiener process in $\mathbf{R
^{d}$. The last term is driven by $Z=\{Z_{t}\}_{t\in \lbrack 0,T]}$, an $m
-dimensional L\'{e}vy process whose characteristic function is $\exp \left\{
t\eta (\xi )\right\} $ with
\begin{equation*}
\eta (\xi ) = \int_{\mathbf{R}_{0}^{m}}\big[e^{i(\xi ,y)}-1-i(\xi ,y)\chi
_{\left\{ |y|\leq 1\right\} }\mathbf{1}_{\{\alpha \in (1,2]\}}\big]\pi (dy).
\end{equation*}
Hence,
\begin{equation}
Z_{t}=\int_{0}^{t}\int (1-\chi _{\alpha }(y))yp(ds,dy)+\int_{0}^{t}\int \chi
_{\alpha }(y)yq(ds,dy), \notag
\end{equation
where $\chi _{\alpha }(y)=\mathbf{1}_{\{\alpha \in (1,2]\}}\chi _{\left\{
|y|\leq 1\right\} }$, $p(dt,dy)$ is a Poisson point measure on $[0,\infty
)\times \mathbf{R}_{0}^{m}$ with $\mathbf{E}[p(dt,dy)]=\pi (dy)dt$, and
q(dt,dy)=p(dt,dy)-\pi (dy)dt$ is the centered Poisson measure. It is assumed
that
\begin{equation*}
\int (|y|^{\alpha }\wedge 1)\pi (dy)<\infty .
\end{equation*}
\subsection{Motivation}
The process defined in (\ref{one}) is used as a mathematical model for
random dynamic phenomena in applications arising from fields such as finance
and insurance, to capture continuous and discontinuous uncertainty. For many
applications, the practical computation of functionals of the type $F
\mathbf{E}[g(X_{T})]$ and $F=\mathbf{E}\big[\int_{0}^{T}f(X_{s})ds\big]$
plays an important role. For instance in finance, derivative prices can be
expressed by such functionals. However in reality, a stochastic differential
equation does not always have a closed-form solution. In such cases, in
order to evaluate $F$, an alternative option is to numerically approximate
the It\^{o} process $X$ by a discrete-time Monte-Carlo simulation, which has
been widely applied. The simplest and the most commonly-used scheme is the
weak Euler approximation.
Let the time discretization $\{\tau _{i},i=0,\ldots ,n_{T}\}$ of the
interval $[0,T]$ with maximum step size $\delta \in (0,1)$ be a partition of
$[0,T]$ such that $0=\tau _{0}<\tau _{1}<\dots <\tau _{n_{T}}=T$ and
\max_{i}(\tau _{i}-\tau _{i-1})\leq \delta .$ \ The Euler approximation of
X $ is an $\mathbb{F}$-adapted stochastic process $Y=\{Y_{t}\}_{t\in \lbrack
0,T]}$ defined by the stochastic equatio
\begin{equation}
Y_{t}=X_{0}+\int_{0}^{t}a(Y_{\tau _{i_{s}}})ds+\int_{0}^{t}b(Y_{\tau
_{i_{s}}})dU_{s}^{\alpha }+\int_{0}^{t}G(Y_{\tau _{i_{s}}})dZ_{s},t\in
\lbrack 0,T], \label{du}
\end{equation
where $\tau _{i_{s}}=\tau _{i}$\ if $s\in \lbrack \tau _{i},\tau
_{i+1}),i=0,\ldots ,n_{T}-1.$ Contrary to those in (\ref{one}), the
coefficients in (\ref{du}) are piecewise constants in each time interval of
[\tau _{i},\tau _{i+1}).$
The weak Euler approximation $Y$\ is said to converge with order $\kappa >0
\ if for each bounded smooth function $g$ with bounded derivatives, there
exists a constant $C$, depending only on $g$, such that
\begin{equation*}
|\mathbf{E}[g(Y_{T})] - \mathbf{E}[g(X_{T})]|\leq C\delta ^{\kappa },
\end{equation*
where $\delta >0$\ is the maximum step size of the time discretization.
In the literature, the weak Euler approximation of stochastic differential
equations with smooth coefficients has been consistently studied. For
diffusion processes ($\alpha =2)$, Milstein was one of the first to
investigate the order of weak convergence and derived $\kappa =1$~\cit
{Mil79, Mil86}. Talay considered a class of the second order approximations
for diffusion processes~\cite{Tal84, Tal86}. For It\^{o} processes with jump
components, Mikulevi\v{c}ius \& Platen showed the first-order convergence in
the case in which the coefficient functions possess fourth-order continuous
derivatives~\cite{MiP88}. Platen and Kloeden \& Platen studied not only
Euler but also higher order approximations~\cite{KlP00, Pla99}. Protter \&
Talay analyzed the weak Euler approximation for
\begin{equation}
X_{t}=X_{0}+\int_{0}^{t}G(X_{s-})dZ_{s},t\in \lbrack 0,T], \label{pt}
\end{equation
where $Z_{t}=(Z_{t}^{1},\ldots ,Z_{t}^{m})$ is a L\'{e}vy process and
G=(G^{ij})_{1\leq i\leq d,1\leq j\leq m}$ is a measurable and bounded
function~\cite{PrT97}. They showed the order of convergence $\kappa =1,$
provided that $G$ and $g$\ are smooth and the L\'{e}vy measure of $Z$\ has
finite moments of sufficiently high order. Because of this, the main
theorems in \cite{PrT97} do not apply to (\ref{one}). On the other hand,
\ref{one}) with a nondegenerate matrix $b$ does not cover (\ref{pt}), which
can degenerate completely.
In general, the coefficients and the test function $g$\ do not always have
the smoothness properties assumed in the papers cited above. Mikulevi\v{c
ius \& Platen proved that there still exists some order of convergence of
the weak Euler approximation for nondegenerate diffusion processes under
\"{o}lder conditions on the coefficients and $g$~\cite{MiP911}. Kubilius \&
Platen and Platen \& Bruti-Liberati considered a weak Euler approximation in
the case of a nondegenerate diffusion process with a finite number of jumps
in finite time intervals~\cite{KuP01, PlB10}.
In this paper, we investigate the dependence of the rate of convergence on
the H\"{o}lder regularity of coefficients and the driving processes. For a
driving process, the variation of the process can be regarded as a part of
its regularity. In this sense, Wiener process is the worse, most
\textquotedblleft chaotic\textquotedblright , among $\alpha $-stable
processes. Also, as pointed out in \cite{PrT97}, the tails of L\'{e}vy
processes influence the convergence rate as well.
\subsection{Examples}
For $\beta >0\,,$ denote $C^{\beta }(\mathbf{R}^{d})$ the H\"{o}lder-Zygmund
space, and $\tilde{C}^{\beta }(\mathbf{R}^{d})$ the Lipshitz space ($\tilde{
}^{\beta }(\mathbf{R}^{d})=C^{\beta }(\mathbf{R}^{d})$ if $\beta \notin
\mathbf{N}$, see Section~\ref{sec:operator} for definitions). Let us look at
two examples.
\begin{example}
\label{ex1}$($see Corollary \ref{co1}$)$ Assume $\beta \leq \alpha $, the
coefficients $a^{i},b^{ij}\in \tilde{C}^{\beta }(\mathbf{R}^{d})$,
G^{ij}\in \tilde{C}^{\frac{\beta }{\alpha \wedge 1}}(\mathbf{R}^{d})$,
\inf_{x}|\det b(x)|>0,$ and
\begin{equation*}
\int_{\mathbf{R}^{m}}|y|^{\alpha }\pi (dy)<\infty ,
\end{equation*
where $\pi $ is the L\'{e}vy measure of the driving process $Z$. Then it
holds tha
\begin{eqnarray*}
|\mathbf{E}[g(Y_{T})]-\mathbf{E}[g(X_{T})]| &\leq &C|g|_{\alpha +\beta
}r(\delta ,\alpha ,\beta ), \\
|\mathbf{E}\big[\int_{0}^{T}f(Y_{\tau _{i_{s}}})ds\big]-\mathbf{E}\big
\int_{0}^{T}f(X_{s})ds\big]| &\leq &C|f|_{\beta }r(\delta ,\alpha ,\beta ),
\end{eqnarray*
where
\begin{equation*}
r(\delta ,\alpha ,\beta )=\left\{
\begin{array}{cc}
\delta ^{\frac{\beta }{\alpha }} & \text{if }\beta <\alpha , \\
\delta (1+\left\vert \ln \delta \right\vert ) & \text{if }\beta =\alph
\end{array
\right.
\end{equation*}
\end{example}
\begin{example}
\label{ex2}$($see Corollary \ref{co2}$)$ Consider the jump-diffusion case
(\alpha =2)$
\begin{equation*}
X_{t}=X_{0}+\int_{0}^{t}a(X_{s})ds+\int_{0}^{t}b(X_{s})dW_{s}
\int_{0}^{t}G(X_{s-})dZ_{s},t\in \lbrack 0,T],
\end{equation*
where $W=\{W_{t}\}_{t\in \lbrack 0,T]}$ is a standard Wiener process. Assume
$a,b^{ij}\in \tilde{C}^{\beta }(\mathbf{R}^{d})$, $\inf_{x}|\det b(x)|>0$,
and there exists $\mu \in (0,3)$ such tha
\begin{equation*}
\int_{|y|\leq 1}|y|^{2}\pi (dy)+\int_{|y|>1}|y|^{\mu }\pi (dy)<\infty
\end{equation*
Let $G^{ij}\in \tilde{C}^{\frac{\beta }{\mu \wedge 1}}(\mathbf{R}^{d})$.
Then it holds tha
\begin{eqnarray*}
|\mathbf{E}[g(Y_{T})]-\mathbf{E}[g(X_{T})]| &\leq &C|g|_{\alpha +\beta
}r(\delta ), \\
|\mathbf{E}\big[\int_{0}^{T}f(Y_{\tau _{i_{s}}})ds\big]-\mathbf{E}\big
\int_{0}^{T}f(X_{s})ds\big]| &\leq &C|f|_{\beta }r(\delta ),
\end{eqnarray*
wher
\begin{equation*}
r(\delta )=\left\{
\begin{array}{cc}
\delta ^{\frac{\beta \wedge \mu }{2}} & \text{if }\mu <2, \\
\delta (1+|\ln \delta |) & \text{if }\mu =\beta =2, \\
\delta & \text{if }\mu >2,\beta >
\end{array
\right.
\end{equation*
The assumption $G^{ij}\in \tilde{C}^{\frac{\beta }{\mu \wedge 1}}(\mathbf{R
^{d})$ shows that if $\mu <1$, the heavy tail of $\pi $ can be balanced by a
higher regularity of $G^{ij}$.
\end{example}
As in \cite{MiP911}, this paper employs the idea of Talay (see \cite{Tal84})
and uses the solution to the backward Kolmogorov equation associated with
X_{t},$ It\^{o}'s formula, and one-step estimates (see Section~\re
{sec:outline} for the outline of the proof).
The paper is organized as follows. In Section 2, the main result is stated
and the proof is outlined. In Section 3, we present the essential technical
results, followed by the proof of the main theorem in Section 4.
\section{Notation and Main Result}
\subsection{Main Result and Notation}
The main result of this paper is the following statement.
\begin{theorem}
\label{thm:main}Let $\beta \in (0,3)$, $0<\beta \leq \mu <\alpha +\beta $
and
\begin{equation*}
\int_{|y|\leq 1}|y|^{\alpha }\pi (dy)+\int_{|y|>1}|y|^{\mu }\pi (dy)<\infty .
\end{equation*
Assume $\inf_{x}|\det b(x)|>0$ and $a^{i},b^{ij}\in \tilde{C}^{\beta }
\mathbf{R}^{d}),G^{ij}\in \tilde{C}^{\frac{\beta }{\mu \wedge 1}}(\mathbf{R
^{d})$. Then there exists a constant $C$ such that for all $g\in C^{\alpha
+\beta }(\mathbf{R}^{d})$, $f\in C^{\beta }(\mathbf{R}^{d})$
\begin{eqnarray*}
|\mathbf{E}[g(Y_{T})]-\mathbf{E}[g(X_{T})]| &\leq &C|g|_{\alpha +\beta
}r(\delta ,\alpha ,\beta ), \\
|\mathbf{E}\big[\int_{0}^{T}f(Y_{\tau _{i_{s}}})ds\big]-\mathbf{E}\big
\int_{0}^{T}f(X_{s})ds\big]| &\leq &C|f|_{\beta }r(\delta ,\alpha ,\beta ),
\end{eqnarray*
wher
\begin{equation*}
r(\delta ,\alpha ,\beta )=\left\{
\begin{array}{cl}
\delta ^{\frac{\beta }{\alpha }}, & \beta <\alpha , \\
\delta (1+\left\vert \ln \delta \right\vert ), & \beta =\alpha , \\
\delta , & \beta >\alpha
\end{array
\right.
\end{equation*}
\end{theorem}
Applying Theorem~\ref{thm:main} to the case $\alpha =\mu $ and the case of
heavier tails results in Corollary~\ref{co1} and Corollary~\ref{co2},
respectively.
\begin{corollary}
\label{co1}Let $\beta \in (0,3)$, $\beta \leq \alpha $, and
\begin{equation*}
\int |y|^{\alpha }\pi (dy)<\infty .
\end{equation*
Assume $a^{i},b^{ij}\in \tilde{C}^{\beta }(\mathbf{R}^{d})$, $G^{ij}\in
\tilde{C}^{\frac{\beta }{\alpha \wedge 1}}(\mathbf{R}^{d})$, and
\inf_{x}|\det b(x)|>0.$ Then there exists a constant $C$ such that for all
g\in C^{\alpha +\beta }(\mathbf{R}^{d})$, $f\in C^{\beta }(\mathbf{R}^{d})$
\begin{eqnarray*}
|\mathbf{E}[g(Y_{T})]-\mathbf{E}[g(X_{T})]| &\leq &C|g|_{\alpha +\beta
}r(\delta ,\alpha ,\beta ), \\
|\mathbf{E}\big[\int_{0}^{T}f(Y_{\tau _{i_{s}}})ds\big]-\mathbf{E}\big
\int_{0}^{T}f(X_{s})ds\big]| &\leq &C|f|_{\beta }r(\delta ,\alpha ,\beta ),
\end{eqnarray*
where
\begin{equation*}
r(\delta ,\alpha ,\beta )=\left\{
\begin{array}{cc}
\delta ^{\frac{\beta }{\alpha }} & \text{if }\beta <\alpha , \\
\delta (1+\left\vert \ln \delta \right\vert ) & \text{if }\beta =\alph
\end{array
\right.
\end{equation*}
\end{corollary}
\begin{corollary}
\label{co2}Let $\beta \in (0,3)$, $0<\beta \leq \mu <\alpha $, and
\begin{equation*}
\int_{|y|\leq 1}|y|^{\alpha }\pi (dy)+\int_{|y|>1}|y|^{\mu }\pi (dy)<\infty .
\end{equation*
Let $\inf_{x}|\det b(x)|>0,a^{i},b^{ij}\in \tilde{C}^{\beta }(\mathbf{R
^{d}) $ and $G^{ij}\in \tilde{C}^{\frac{\beta }{\mu \wedge 1}}(\mathbf{R
^{d})$. Then there exists a constant $C$ such that for all $g\in C^{\alpha
+\beta }(\mathbf{R}^{d})$, $f\in C^{\beta }(\mathbf{R}^{d})$
\begin{eqnarray*}
|\mathbf{E}[g(Y_{T})]-\mathbf{E}[g(X_{T})]| &\leq &C|g|_{\alpha +\beta
}\delta ^{\frac{\beta \wedge \mu }{\alpha }}, \\
|\mathbf{E}\big[\int_{0}^{T}f(Y_{\tau _{i_{s}}})ds\big]-\mathbf{E}\big
\int_{0}^{T}f(X_{s})ds\big]| &\leq &C|f|_{\beta }\delta ^{^{\frac{\beta
\wedge \mu }{\alpha }}}.
\end{eqnarray*}
\end{corollary}
Denote $H=[0,T]\times \mathbf{R}^{d}$, $\mathbf{N}=\{0,1,2,\ldots \}$,
\mathbf{R}_{0}^{d}=\mathbf{R}^{d}\backslash \{0\}$. For $x,y\in \mathbf{R
^{d}$, write $(x,y)=\sum_{i=1}^{d}x_{i}y_{i}$. For $(t,x)\in H,$ multiindex
\gamma \in \mathbf{N}^{d}$ with $D^{\gamma }=\frac{\partial ^{|\gamma |}}
\partial x_{1}^{\gamma _{1}}\ldots \partial x_{d}^{\gamma _{d}}}$, and
i,j=1,\ldots ,d$, denote
\begin{eqnarray*}
\partial _{t}u(t,x) &=&\frac{\partial }{\partial t}u(t,x), \ D^{k}u(t,x)
\big(D^{\gamma }u(t,x)\big)_{|\gamma |=k},k\in \mathbf{N}\text{,} \\
\partial _{i}u(t,x) &=&u_{x_{i}}(t,x)=\frac{\partial }{\partial x_{i}
u(t,x), \ \partial _{ij}^{2}u(t,x)=u_{x_{i}x_{j}}(t,x)=\frac{\partial ^{2}}
\partial x_{i}x_{j}}u(t,x), \\
\partial _{x}u(t,x) &=&\nabla u(t,x)=\big(\partial _{1}u(t,x),\dots
,\partial _{d}u(t,x)\big), \\
\partial ^{2}u(t,x) &=&\Delta u(t,x)=\sum_{i=1}^{d}\partial _{ii}^{2}u(t,x).
\end{eqnarray*}
$C=C(\cdot ,\ldots ,\cdot )$ denotes constants depending only on quantities
appearing in parentheses. In a given context the same letter is (generally)
used to denote different constants depending on the same set of arguments.
\subsection{Outline of Proof}
\label{sec:outline}
Due to the lack of regularity, standard techniques such as stochastic flows
cannot be applied to prove Theorem~\ref{thm:main}. Instead, as in \cit
{MiP911}, the solution to the backward Kolmogorov equation associated with
X_{t}$ is used. In the following, the operators of the Kolmogorov equation
associated with $X_{t}$ are first defined.
For $u\in C^{\alpha +\beta }(H)$, denote
\begin{eqnarray*}
A_{z}u(t,x) &=&\mathbf{1}_{\{\alpha =1\}}(a(z),\nabla _{x}u(t,x))+\mathbf{1
_{\{\alpha =2\}}\frac{1}{2}\sum_{i,j=1}^{d}D^{ij}(z)\partial _{ij}^{2}u(t,x)
\\
&&+\mathbf{1}_{\{\alpha \in (0,2)\}}\int [u(t,x+b(z)y)-u(t,x)-(\nabla
u(t,x),b(z)y)\chi _{\alpha }(y)]\frac{dy}{|y|^{d+\alpha }}, \\
Au(t,x) &=&A_{x}u(t,x)=A_{z}u(t,x)|_{z=x},
\end{eqnarray*
with $\chi _{\alpha }(y)=\mathbf{1}_{\{\alpha \in (1,2)\}}+\mathbf{1
_{\{\alpha =1\}}\chi _{\left\{ |y|\leq 1\right\} },D=b^{\ast }b,$ an
\begin{eqnarray*}
B_{z}u(t,x) &=&\mathbf{1}_{\{\alpha \in (1,2]\}}(a(z),\nabla
_{x}u(t,x))+\int_{\mathbf{R}_{0}^{m}}\big[u(t,x+G(z)y))-u(t,x) \\
&&-\mathbf{1}_{\{\alpha \in (1,2]\}}\mathbf{1}_{\left\{ |y|\leq 1\right\}
}(\nabla _{x}u(t,x),G(z)y))\big]\pi (dy), \\
Bu(t,x) &=&B_{x}u(t,x)=B_{z}u(t,x)|_{z=x}.
\end{eqnarray*}
Applying It\^{o}'s formula to $X_{t}$ and $u\in C_{0}^{\infty }(\mathbf{R
^{d})$, we find that
\begin{equation*}
u(X_{t})-\int_{0}^{t}Au(X_{s})ds-\int_{0}^{t}Bu(X_{s})ds,t\in \lbrack 0,T]
\end{equation*
is a martingale.
\begin{remark}
\label{lrenew1}More precisely, under assumptions of Theorem \ref{thm:main},
there exists a unique weak solution to equation \textup{(\ref{one})} and the
stochastic process
\begin{equation*}
u(X_{t})-\int_{0}^{t}(A+B)u(X_{s})ds,\forall u\in C^{\alpha +\beta }(\mathbf
R}^{d}) \label{ff25}
\end{equation*
is a martingale~\textup{\cite{MiP923}}. The operator $\mathcal{L}=A+B$ is
the generator of $X_{t}$ defined in \textup{(\ref{one})}; $A$ is the
principal part of $\mathcal{L}$ and $B$ is the lower order or subordinated
part of $\mathcal{L}$.
\end{remark}
If $v(t,x),(t,x)\in H$ satisfies the backward Kolmogorov equation
\begin{eqnarray*}
\big(\partial _{t}+A+B\big)v(t,x) &=&0,\quad 0\leq t\leq T, \\
v(T,x) &=&g(x),
\end{eqnarray*
then as interpreted in Section 4, by It\^{o}'s formula
\begin{equation*}
\mathbf{E}[g(Y_{T})]-\mathbf{E}[g(X_{T})]=\mathbf{E}[v(T,Y_{T})-v(0,Y_{0})]
\mathbf{E}\big[\int_{0}^{T}(\partial _{t}+\mathcal{L}_{Y_{\tau
_{i_{s}}}})v(s,Y_{s})ds\big].
\end{equation*
The regularity of $v$ determines the one-step estimate and the rate of
convergence of the approximation. For $\beta \in (0,1)$, the results for the
Kolmogorov equation in H\"{o}lder classes are available~\cite{MiP922, MiP09
. In a standard way the results can be extended to the case $\beta >1$. The
main difficulty is to derive the one-step estimates (see Lemma~\re
{lem:expect}).
\section{Backward Kolmogorov Equation}
In H\"{o}lder-Zygmund spaces, consider the backward Kolmogorov equation
associated with $X_{t}$:
\begin{eqnarray}
\big(\partial _{t}+A+B\big)v(t,x) &=&f(t,x), \label{eq1} \\
v(T,x) &=&0. \notag
\end{eqnarray
The regularity of its solution is essential for the one-step estimate which
determines the rate of convergence.
\begin{definition}
\label{def1}Let $f$ be a measurable and bounded function on $\mathbf{R}^{d}
. We say that $u\in C^{\alpha +\beta }(H)$ is a solution to $(\ref{eq1})$ if
\begin{equation}
u(t,x)=\int_{t}^{T}\big[\mathcal{L}u(s,x)-f(s,x)\big]ds,\forall (t,x)\in H.
\label{defs}
\end{equation}
\end{definition}
The following theorem is the main result of this section.
\begin{theorem}
\label{thm:StoCP}Let $\beta \in (0,3)$, $0<\beta \leq \mu <\alpha +\beta $,
and
\begin{equation*}
\int_{|y|\leq 1}|y|^{\alpha }\pi (dy)+\int_{|y|>1}|y|^{\mu }\pi (dy)<\infty .
\end{equation*
Assume $a^{i},b^{ij}\in \tilde{C}^{\beta }(\mathbf{R}^{d}),G^{ij}\in \tilde{
}^{\frac{\beta }{\mu \wedge 1}}(\mathbf{R}^{d}),\inf_{x}|\det b(x)|>0$. Then
for each $f\in C^{\beta }(\mathbf{R}^{d})$, there exists a unique solution
v\in C^{\alpha +\beta }(H)$ to \textup{(\ref{eq1})} and a constant $C$
independent of $f$ such that $|u|_{\alpha +\beta }\leq C|f|_{\beta }.$
\end{theorem}
An immediate consequence of Theorem~\ref{thm:StoCP} is the following
statement.
\begin{corollary}
\label{lcornew1}Let $\beta \in (0,3)$, $0<\beta \leq \mu <\alpha +\beta $,
and
\begin{equation*}
\int_{|y|\leq 1}|y|^{\alpha }\pi (dy)+\int_{|y|>1}|y|^{\mu }\pi (dy)<\infty .
\end{equation*
Let $a^{i},b^{ij}\in \tilde{C}^{\beta }(\mathbf{R}^{d}),G^{ij}\in \tilde{C}^
\frac{\beta }{\mu \wedge 1}}(\mathbf{R}^{d}),\inf_{x}|\det b(x)|>0$. Then
for each $f\in C^{\beta }(\mathbf{R}^{d})$ and $g\in C^{\alpha +\beta }
\mathbf{R}^{d})$, there exists a unique solution $v\in C^{\alpha +\beta }(H)$
to the Cauchy problem
\begin{eqnarray}
\big(\partial _{t}+A+B\big)v(t,x) &=&f(x), \label{maf8} \\
v(T,x) &=&g(x) \notag
\end{eqnarray
and $|v|_{\alpha +\beta }\leq C(|f|_{\beta }+|g|_{\alpha +\beta })$ with a
constant $C$ independent of $f$ and $g$.
\end{corollary}
To prove Theorem \ref{thm:StoCP} and Corollary \ref{lcornew1}, in a standard
way the equation with constant coefficients is first solved. Then variable
coefficients are handled by using partition of unity and deriving apriori
Schauder estimates in H\"{o}lder-Zygmund spaces. Finally, the continuation
by parameter method is applied to extend solvability of an equation with
constant coefficients to (\ref{eq1}).
\subsection{Kolmogorov Equation with Constant Coefficients}
It is convenient to rewrite the principal operator $A$ by changing the
variable of integration in the integral part
\begin{eqnarray*}
A_{z}u(t,x) &=&\mathbf{1}_{\{\alpha =1\}}(a(z),\nabla _{x}u(t,x))+\mathbf{1
_{\{\alpha =2\}}\sum_{i,j=1}^{d}D^{ij}(z)\partial _{ij}^{2}u(t,x) \\
&&+\mathbf{1}_{\{\alpha \in (0,2)\}}\int [u(t,x+y)-u(t,x)-(\nabla
u(t,x),y)\chi _{\alpha }(y)]m(z,y)\frac{dy}{|y|^{d+\alpha }},
\end{eqnarray*
where $D=b^{\ast }b,
\begin{equation}
m(z,y)=\frac{1}{|\det b(z)|}\frac{1}{|b(z)^{-1}\frac{y}{|y|}|^{d+\alpha }
,\alpha \in (0,2). \label{foo}
\end{equation
Obviously,
\begin{equation}
\int_{S^{d-1}}ym(\cdot ,y)\mu _{d-1}(dy)=0. \label{ff27}
\end{equation
Here $S^{d-1}$ is the unit sphere in $\mathbf{R}^{d}$ and $\mu _{d-1}$ is
the Lebesgue measure.
For $z_{0}\in \mathbf{R}^{d}$, denote $A^{0}u(x)=A_{z_{0}}u(x)$. Consider a
backward Kolmogorov equation with constant coefficients and $\lambda \geq 0,
\begin{eqnarray}
\big(\partial _{t}+A^{0}-\lambda \big)v(t,x) &=&f(x), \label{eq0} \\
v(T,x) &=&0. \notag
\end{eqnarray}
\begin{proposition}
\label{prop1}Let $\beta >0$ and $f\in C^{\beta }(\mathbf{R}^{d})$. Assume
there are constants $c_{1},K>0$ such that for all $z\in \mathbf{R}^{d},$
\begin{equation*}
|\det b(z)|\geq c_{1},\quad \mathbf{1}_{\{\alpha =1\}}|a(z)|+|b(z)|\leq K.
\end{equation*
Then there exists a unique solution $u\in C^{\alpha +\beta }(H)$ to (\re
{eq0}) and
\begin{equation}
|u|_{\alpha +\beta }\leqslant C|f|_{\beta }, \label{1}
\end{equation
where the constant $C$ depends only on $\alpha ,\ \beta ,\ T,\ d$, $c_{1},\
K.$ Moreover
\begin{equation}
|u|_{\beta }\leq C(\alpha ,d)(\lambda ^{-1}\wedge T)|f|_{\beta } \label{2}
\end{equation
and there exists a constant $C$ such that for all $s\leq t\leq T,$
\begin{equation}
|u(t,\cdot )-u(s,\cdot )|_{\frac{\alpha }{2}+\beta }\leq C(t-s)^{\frac{1}{2
}|f|_{\beta }. \label{3}
\end{equation}
\end{proposition}
To derive Proposition~\ref{prop1}, some auxiliary results are presented
first.
\subsubsection{Continuity of Operator $A^{0}$ in H\"{o}lder-Zygmund Spaces}
\label{sec:operator}
To show that operator $A^{0}$ is continuous in H\"{o}lder-Zygmund spaces
C^{\beta }(\mathbf{R}^{d})$, first recall their definition.
For $\beta =[\beta ]^{-}+\left\{ \beta \right\} ^{+}>0$, where $[\beta
]^{-}\in \mathbf{N}$ and $\left\{ \beta \right\} ^{+}\in (0,1]$, let
C^{\beta }(H)$ denote the space of measurable functions $u$ on $H$ such that
the norm
\begin{eqnarray*}
|u|_{\beta } &=&\sum_{|\gamma |\leq \lbrack \beta ]^{-}}\sup_{(t,x)\in
H}|D_{x}^{\gamma }u(t,x)|+\mathbf{1}_{\{\{\beta \}^{+}<1\}}\sup_{\substack{
|\gamma |=[\beta ]^{-}, \\ t,x\neq \tilde{x}}}\frac{|D_{x}^{\gamma
}u(t,x)-D_{x}^{\gamma }u(t,\tilde{x})|}{|x-\tilde{x}|^{\{\beta \}^{+}}} \\
&&+\mathbf{1}_{\{\{\beta \}^{+}=1\}}\sup_{\substack{ |\gamma |=[\beta ]^{-},
\\ t,x,h\neq 0}}\frac{|D_{x}^{\gamma }u(t,x+h)+D_{x}^{\gamma
}u(t,x-h)-2D_{x}^{\gamma }u(t,x)|}{|h|^{\{\beta \}^{+}}}
\end{eqnarray*
is finite. Accordingly, $C^{\beta }(\mathbf{R}^{d})$ denotes the
corresponding space of functions on $\mathbf{R}^{d}$.~The classes $C^{\beta
} $ coincide with H\"{o}lder spaces if $\beta \notin \mathbf{N}$ (see 1.2.2
of \cite{Tri92}).
For $v\in C^{\beta }(\mathbf{R}^{d})$ with $\beta \in (0,1]$, denot
\begin{eqnarray*}
|v|_{0} &=&\sup_{x}|v(x)|,\ \\
\left[ v\right] _{\beta } &=&\sup_{x\neq y}\frac{|v(x)-v(y)|}{|x-y|^{\beta }
\text{ if }\beta \in (0,1), \\
\left[ v\right] _{\beta } &=&\sup_{x,h\neq 0}\frac{|v(x+h)-2v(x)+v(x-h)|}
|x-y|^{\beta }}\text{ if }\beta =1.
\end{eqnarray*
Similarly we define the spaces $\tilde{C}^{\beta }(\mathbf{R}^{d})$: $\tilde
C}^{\beta }(\mathbf{R}^{d})=C^{\beta }(\mathbf{R}^{d})$ if $\beta >0,\beta
\notin \mathbf{N}$, and $\tilde{C}^{k}(\mathbf{R}^{d})$ is the space of all
functions $f$ on $\mathbf{R}^{d}$ having $k-1$ continuous bounded
derivatives and such that $D^{\gamma }f,|\gamma |=k-1,$ are Lipshitz. We
introduce the norms in $\tilde{C}^{\beta }(\mathbf{R}^{d})
\begin{eqnarray*}
||f||_{\beta } &=&|f|_{\beta }\text{ if }\beta >0,\beta \notin \mathbf{N}, \\
||f||_{k} &=&\sum_{|\gamma |\leq k-1}|D^{\gamma }f|_{0}+\sup_{x\neq
y,|\gamma |=k-1}\frac{|D^{\gamma }f(x)-D^{\gamma }f(y)|}{|x-y|}.
\end{eqnarray*}
For $\alpha \in (0,2)$, define for $v\in C^{\alpha +\beta }(\mathbf{R}^{d})$
the fractional Laplacia
\begin{equation}
\partial ^{\alpha }v(x)=\int [v(x+y)-v(x)-\left( \nabla v(x),y\right) \chi
_{\alpha }(y)]\frac{dy}{|y|^{d+\alpha }},x\in \mathbf{R}^{d}. \label{fo4}
\end{equation}
For various estimates, the following representation of the difference is
useful.
\begin{lemma}
\label{r1}$($Lemma 2.1 in \textup{\cite{Kom84}}$)$ For $\delta \in (0,1)$
and $u\in C_{0}^{\infty }(\mathbf{R}^{d})$,
\begin{equation*}
u\big(x+y\big)-u(x)=K\int k^{(\delta )}(y,z)\partial ^{\delta }u(x-z)dz,
\label{eqn:diff_bound}
\end{equation*
where $K=K(\delta ,d)$ is a constant,
\begin{equation*}
k^{(\delta )}(y,z)=|z+y|^{-d+\delta }-|z|^{-d+\delta },
\end{equation*
and there exists a constant $C$ such that
\begin{equation*}
\int |k^{(\delta )}(y,z)|dz\leq C|y|^{\delta },\forall y\in \mathbf{R}^{d}.
\end{equation*}
\end{lemma}
By taking pointwise limit ($\partial ^{\delta}$ is defined by (\ref{fo4}))
and applying the dominated convergence theorem, the statement can be
extended to $u\in C^{\delta }(\mathbf{R}^{d})$.
Let $m(y)$ be a measurable and bounded function on $\mathbf{R}^{d}.$ Define
\begin{equation*}
L^{m}u(x)=\int_{\mathbf{R}^{d}}\left[ u(x+y)-u(x)-(\nabla u(x), y)\chi
_{\alpha}(y)\right] m(y)\frac{dy}{|y|^{d+\alpha }},u\in C^{\alpha +\beta }.
\end{equation*}
The following statement is proved in~\cite{MiP11} for $\beta \in (0,1]$.
\begin{lemma}
\label{prop2}Let $\alpha \in (0,2)$, $\beta >0$, $u\in C^{\alpha +\beta }
\mathbf{R}^{d})$, and $|m|\leq K$. Assume if $\alpha =1$,
\begin{equation}
\int_{r<|y|\leq 1}ym(y)\frac{dy}{|y|^{d+1}}=0,\forall r\in (0,1). \label{8}
\end{equation
Then there exists a constant $C$ independent of $u$ such tha
\begin{equation*}
|L^{m}u|_{\beta }\leq CK|u|_{\alpha +\beta }.
\end{equation*}
\end{lemma}
\begin{proof}
The result holds for $\beta \in (0,1]$ according to Proposition 11 in \cit
{MiP11}. If $\beta >1$, and $u\in C^{\alpha +\beta }(\mathbf{R}^{d})$, then
for any multiindex $|\gamma |=[\beta ]$, $D^{\gamma }u\in C^{\alpha +\beta
-[\beta ]}$, an
\begin{equation*}
|D^{\gamma }\left( L^{m}u\right) |_{\beta -[\beta ]^{-}}=|L^{m}\left(
D^{\gamma }u\right) |_{\beta -[\beta ]^{-}}\leq CK|D^{\gamma }u|_{\alpha
+\beta -[\beta ]^{-}}.
\end{equation*
The statement follows.
\end{proof}
\subsubsection{Proof of Proposition \protect\ref{prop1}}
The statement is proved by induction. Given $\alpha \in (0,2]$ and $f\in
C^{\beta }(H)$, for $\beta \in (0,1]$, there exists a unique solution $u\in
C^{\alpha +\beta }(H)$ to the Kolmogorov equation (\ref{eq0}) such that (\re
{1})-(\ref{3}) hold~\cite{MiP11}.
Assume the result holds for $\beta \in \bigcup\limits_{l=0}^{n-1}(l,l+1]$,
n\in \mathbf{N}$. Let $\beta \in (n,n+1]$, $\tilde{\beta}=\beta -1$, and
f\in C^{\beta }(H)$. Then $\tilde{\beta}\in (n-1,n]$, $f\in C^{\tilde{\beta
}(H)$, and there exists a unique solution $v\in C^{\alpha +\tilde{\beta}}(H)
, $\alpha \in (0,2]$ to the Cauchy problem such that (\ref{1})-(\ref{3})
hold for $v$ with $\tilde{\beta}$. For $h\in \mathbf{R}$ and $k=1,\dots ,d$,
denote
\begin{equation*}
v_{k}^{h}(t,x)=\frac{v(t,x+he_{k})-v(t,x)}{h},
\end{equation*
where $\{e_{k},k=1,\ldots ,d\}$ is the canonical basis in $\mathbf{R}^{d}$.
Obviously, $v_{k}^{h}\in C^{\alpha +\tilde{\beta}}(H)$ and
\begin{eqnarray}
\big(\partial _{t}+A^{0}-\lambda \big)v_{k}^{h}(t,x) &=&f_{k}^{h}(x),x\in
\mathbf{R}^{d}, \label{eqn:cauchy} \\
v_{k}^{h}(T,x) &=&0. \notag
\end{eqnarray}
Since $f\in C^{\beta }(H)$ and
\begin{equation*}
f_{k}^{h}(t,x)=\int_{0}^{1}\partial _{k}f(t,x+he_{k}s)ds,\forall h\neq 0,
\end{equation*
then
\begin{equation}
|f_{k}^{h}|_{\tilde{\beta}}\leq C|\nabla f|_{\beta -1}\leq C|f|_{\beta }
\label{ff22}
\end{equation
with a constant $C$ independent of $h$. Since $v\in C^{\alpha +\tilde{\beta
}(H)$, then $v_{k}^{h}\in C^{\alpha +\tilde{\beta}}(H)$. By (\ref{ff22}) and
the induction assumption, the estimates (\ref{1})-(\ref{3}) hold for
v_{k}^{h}$ with a constant independent of $h$. Hence $v_{k}^{h}(t,x)$ are
equicontinuous in $(t,x)$. By the Arzel\`{a}-Ascoli theorem, for each
h_{n}\rightarrow 0$, there exist a subsequence $\{h_{n_{j}}\}$ and
continuous functions $v_{k}(t,x),(t,x)\in H,k=1,\ldots ,d$, such that
v_{k}^{h_{n_{j}}}(t,x)\rightarrow v_{k}(t,x)$ uniformly on compact subsets
of $H$ as $j\rightarrow \infty $. Therefore, $v_{k}\in C^{\alpha +\tilde
\beta}}$ and $|v_{k}|_{\alpha +\tilde{\beta}}\leq C|f|_{\beta },k=1,\ldots
,d.$
It then follows from passing to the limit in the integral form of (\re
{eqn:cauchy}) (see (\ref{defs})) and the dominated convergence theorem that
u_{k}$ is the unique solution to
\begin{eqnarray*}
\big(\partial _{t}+A^{0}-\lambda \big)v_{k}(t,x) &=&\partial _{k}f(t,x), \\
v_{k}(T,x) &=&0,k=1,\dots ,d
\end{eqnarray*
and so $v_{k}^{h_{n}}(t,x)\rightarrow v_{k}(t,x),\forall h_{n}\rightarrow 0.$
Hence,
\begin{equation*}
v_{k}(t,x)=\lim_{h\rightarrow 0}v_{k}^{h}(t,x)=\lim_{h\rightarrow 0}\frac
v(t,x+he_{k})-v(t,x)}{h}=\partial _{k}v(t,x),
\end{equation*
$\partial _{k}v\in C^{\alpha +\tilde{\beta}}(H),k=1,\dots ,d$, and $|\nabla
v|_{\alpha +\tilde{\beta}}\leq C|f|_{\beta }.$ Therefore, $v\in C^{\alpha
+\beta }(H)$ and the statement follows.
\subsection{Kolmogorov Equation with Variable Coefficients}
In this section, an estimate is derived to show that $Bu$ is a lower order
operator, which is essential in deriving Schauder estimates in the case of
variable coefficients. To prove Theorem \ref{thm:StoCP}, in a standard way
we use partition of unity and the estimates for constant coefficients, which
allow to obtain apriori estimates. Then the continuation by parameter method
is applied to transfer from constant to variable coefficients.
\subsubsection{Estimates of $Bf$, $f\in C^{\protect\alpha +\protect\beta}$}
\begin{proposition}
\label{b1}Let $\beta \in (0,3)$, $0<\beta \leq \mu <\alpha +\beta $, and
\begin{equation*}
\int_{|y|\leq 1}|y|^{\alpha }\pi (dy)+\int_{|y|>1}|y|^{\mu }\pi (dy)<\infty .
\end{equation*
Assume $a\in C^{\beta }(\mathbf{R}^{d})$,$G^{ij}\in \tilde{C}^{\frac{\beta }
\mu \wedge 1}}(\mathbf{R}^{d})$. Then for each $\varepsilon >0$, there
exists a constant $C_{\varepsilon }$ such tha
\begin{equation*}
|Bf|_{\beta }\leq \varepsilon |f|_{\alpha +\beta }+C_{\varepsilon
}|f|_{0},f\in C^{\alpha +\beta }(\mathbf{R}^{d}).
\end{equation*}
\end{proposition}
\begin{proof}
Since the estimates involving the term with $a(x)$ are obvious, in the
following estimates, assume $a=0$.
\textit{Case I}: $\beta \in (0,1]$. Split for $\delta \in (0,1),
\begin{eqnarray*}
B_{z}f(x) &=&\int \big[f(x+G(z)y)-f(x)-\mathbf{1}_{\{\alpha \in
(1,2]\}}(\nabla f(x),G(z)y)\chi _{\{|y|\leq 1\}}\big]\pi (dy) \\
&=&\int_{|y|\leq \delta }...+\int_{|y|>\delta
}...=B_{z}^{1}f(x)+B_{z}^{2}f(x)
\end{eqnarray*
and $B_{z}^{2}f(x)=B_{z}^{21}f(x)+B_{z}^{22}f(x)$ with
\begin{eqnarray*}
B_{z}^{21}f(x) &=&f(x)\int_{|y|>\delta }\pi (dy)+\mathbf{1}_{\{\alpha \in
(1,2]\}}\big(\nabla f(x),\int_{\delta <|y|\leq 1}G(z)y\pi (dy)\big), \\
B_{z}^{22}f(x) &=&\int_{|y|>\delta }f(x+G(z)y)\pi (dy).
\end{eqnarray*
It follows by the assumptions that there exists $\beta ^{\prime }$ such that
$\mu <\alpha +\beta ^{\prime }<\alpha +\beta $ and
\begin{eqnarray}
|B_{z}^{21}f\left( \cdot \right) |_{\beta }+|B_{z}^{22}f\left( \cdot \right)
|_{\beta } &\leq &C[|f|_{\beta }+\mathbf{1}_{\{\alpha \in (1,2]\}}|\nabla
f|_{\beta }],z\in \mathbf{R}^{d}, \label{fo001} \\
|B_{z+h}^{21}f\left( x\right) -B_{z}^{21}f(x)| &\leq &C\mathbf{1}_{\{\alpha
\in (1,2]\}}|\nabla f|_{0}||G||_{\beta }|h|^{\beta }, \notag \\
|B_{z+h}^{22}f(x)-B_{z}^{22}f(x)| &\leq &C\int_{|y|\geq \delta }|y|^{\mu
}\pi (dy)|f|_{\alpha +\beta ^{\prime }}|G|_{\frac{\beta }{\mu \wedge 1
},x\in R^{d}. \notag
\end{eqnarray}
Consider different scenarios on values of $\alpha$ to show that
\begin{equation}
|B_{z}^{1}f(\cdot )|_{\beta }\leq C|f|_{\alpha +\beta }\int_{|y|\leq \delta
}|y|^{\alpha }d\pi ,z\in \mathbf{R}^{d}. \label{fo1}
\end{equation}
For $\alpha \in (0,1]$, by Lemma \ref{r1}$,
\begin{eqnarray*}
B_{z}^{1}f(x) &=&\int_{|y|\leq \delta }\int \partial ^{\alpha
}f(x-z)k^{(\alpha )}(C(z)y,z)dz\pi (dy)\text{ if }\alpha <1, \\
B_{z}^{1}f(x) &=&\int_{|y|\leq \delta }\int_{0}^{1}(\nabla
f(x+sC(z)y),y)ds\pi (dy)\text{ if }\alpha =1.
\end{eqnarray*
Hence, (\ref{fo1}) follows.
For $\alpha =2$, (\ref{fo1}) follows since
\begin{equation}
B_{z}^{1}f(x)=\int_{|y|\leq \delta }\Big[\int_{0}^{1}\Big(
D^{2}f(x+sC(z)y)C(z)y,C(z)y\Big) (1-s)ds\Big]\text{ if }\alpha =2. \notag
\end{equation}
For $\alpha \in (1,2)$, (\ref{fo1}) follows since by Lemma \ref{r1}
\begin{eqnarray*}
B_{z}^{1}f(x) &=&\int_{|y|\leq \delta }\Big[\int_{0}^{1}\Big(\nabla
f(x+sC(z)y)-\nabla f(x),C(z)y\Big)ds\Big]d\pi \\
&=&\int_{|y|\leq \delta }\Big[\int_{0}^{1}\Big(\int \partial ^{\alpha
-1}\nabla f(x-t)k^{(\alpha -1)}(sC(z)y,t)dt,C(z)y\Big)ds\Big]d\pi .
\end{eqnarray*}
Similarly, to estimate $|B_{\cdot }^{1}f(x)|_{\beta }$, consider different
scenarios on values of $\alpha$.
For $\alpha \in (0,1)$
\begin{equation}
|B_{\cdot }^{1}f(x)|_{\beta }\leq |f|_{(\alpha +\beta )}\int_{|y|\leq \delta
}|y|^{\alpha }\pi (dy)||G||_{\frac{\beta }{\mu \wedge 1}},x\in \mathbf{R
^{d}. \label{fo03}
\end{equation}
For $\alpha \in \lbrack 1,2]$, let $\beta \leq \mu <\alpha +\beta ^{\prime
}<\alpha +\beta$.
If $\alpha \in (1,2]$, for $|y|\leq 1,z,\bar{z}\in \mathbf{R}^{d},
\begin{eqnarray*}
&&|[f(x+G(z)y)-f(x)-(\nabla f(x),G(z)y)] \\
&&-\left[ f(x+G(\bar{z})y)-f(x)-(\nabla f(x),G(\bar{z})y)\right] | \\
&=&|[f(x+G(z)y)-f(x+G(\bar{z})y)-(\nabla f(x),(G(z)-G(\bar{z}))y)]| \\
&\leq &\int_{0}^{1}|\left( \nabla f(x+(1-s)G(\bar{z})y+sG(z)y)-\nabla
f(x),G(z)y-G(\bar{z})y\right) ds| \\
&\leq &|f|_{\alpha }\left( |G|_{0}^{\alpha -1}|y|^{\alpha -1}+|G(\bar{z
)-G(z)|^{\alpha -1}|y|^{\alpha -1}\right) |G(z)-G(\bar{z})||y| \\
&\leq &C|f|_{\alpha }|G|_{0}^{\alpha -1}|G(\bar{z})-G(z)||y|^{\alpha }
\end{eqnarray*
and if $\alpha =1,
\begin{eqnarray*}
&&|[f(x+G(z)y)-f(x)]-\left[ f(x+G(\bar{z})y)-f(x)\right] | \\
&\leq &\int_{0}^{1}|\left( \nabla f(x+(1-s)G(\bar{z})y+sG(z)y),G(z)y-G(\bar{
})y\right) ds| \\
&\leq &|\nabla f|_{0}|G(z)-G(\bar{z})||y|.
\end{eqnarray*
It then follows:
\begin{equation}
|B_{z+h}^{1}f(x)-B_{z}^{1}f(x)|\leq C|h|^{\beta }|f|_{\alpha +\beta ^{\prime
}}||G||_{\beta }^{\alpha },x\in R^{d}. \label{fo3}
\end{equation}
By (\ref{fo001})-(\ref{fo3}), for each $\varepsilon >0$ there exists a
constant $C_{\varepsilon }$ such tha
\begin{equation}
|Bf|_{\beta }\leq \varepsilon |f|_{\alpha +\beta }+C_{\varepsilon }|f|_{0}.
\label{fos}
\end{equation}
\textit{Case II}: $\beta \in (1,2],$ $\beta \leq \mu <\alpha +\beta .$ Note
tha
\begin{equation*}
\partial _{j}(Bf(x))=\big(\frac{\partial }{\partial _{z_{j}}}B_{z}f(x)\big
|_{z=x}+B_{z}f_{x_{j}}|_{z=x}=\big(\frac{\partial }{\partial _{z_{j}}
B_{z}f(x)\big)|_{z=x}+Bf_{x_{j}}.
\end{equation*
For the second term, apply estimate (\ref{fos}) of \textit{Case I}:
f_{x_{j}}\in C^{\alpha +\beta -1},$ the tail moment is 1 and $\beta -1$
\leq \mu -1<\alpha +\beta -1$. Also note that $|G|_{\frac{\beta -1}{(\mu
-1)\wedge 1}}\leq C|G|_{\beta }<\infty $. Hence
\begin{equation*}
|Bf_{x_{j}}|_{\beta -1}\leq \varepsilon |f_{x_{j}}|_{\alpha +\beta
-1}+C_{\varepsilon }|f_{x_{j}}|_{0}.
\end{equation*
Only the first term needs to be estimated:
\begin{eqnarray}
B_{z}^{j}f(x) &=&\frac{\partial }{\partial z_{j}}B_{z}f(x) \label{fo7} \\
&=&\int \left[ \nabla f(x+G(z)y)G_{z_{j}}(z)y-\mathbf{1}_{\{\alpha \in
(1,2]\}}\nabla f(x)G_{z_{j}}(z)y\chi _{\left\{ |y|\leq 1\right\} }\right]
d\pi . \notag
\end{eqnarray}
Let $B^{j}f(x)=B_{z}^{j}f(x)|_{z=x}$, $x\in R^{d}$. Consider different
scenarios on values of $\alpha$ to show that for each $\varepsilon >0$ there
exists a constant $C_{\varepsilon }$ such tha
\begin{equation}
|B^{j}f|_{\beta -1}\leq \varepsilon |f|_{\alpha +\beta }+C_{\varepsilon
}|f|_{0},f\in C^{\alpha +\beta }(\mathbf{R}^{d}). \label{fo8}
\end{equation}
For $\alpha \in (0,1]$, since
\begin{equation*}
B_{z}^{j}f(x)=\int \nabla f(x+G(z)y)G_{z_{j}}(z)yd\pi ,
\end{equation*
then for $\mu <\alpha +\beta ^{\prime }<\alpha +\beta $,
\begin{eqnarray*}
|B_{z}^{j}f(\cdot )|_{\beta -1} &\leq &C|\nabla f|_{\beta
-1}|G_{z_{j}}|_{0},z\in \mathbf{R}^{d}, \\
|B_{z+h}^{j}f(x)-B_{z}^{j}f(x)| &\leq &C|h|^{\beta -1}[|\nabla
f|_{0}||G_{z_{j}}||_{\beta -1}+|f|_{\alpha +\beta ^{\prime
}}|G_{z_{j}}|_{0}||G||_{\frac{\beta }{\mu \wedge 1}}],x\in \mathbf{R}^{d}.
\end{eqnarray*}
For $\alpha \in (1,2]$, spli
\begin{equation*}
B_{z}^{j}f(x)=\int_{|y|\leq
1}...+\int_{|y|>1}...=B_{z}^{j,1}f(x)+B_{z}^{j,2}f(x).
\end{equation*
Since by Lemma \ref{r1}
\begin{equation*}
B_{z}^{j,1}f(x)=\int_{|y|\leq 1}\int \partial ^{\alpha -1}\nabla
f(x-t)k^{(\alpha -1)}(G(z)y,t)G_{z_{j}}(z)ydtd\pi ,
\end{equation*
then $\displaystyle|B_{z}^{j,1}f(\cdot )|_{\beta -1}\leq C|f|_{\alpha +\beta
^{\prime }}||G||_{\beta }^{\alpha },z\in \mathbf{R}^{d}$, for some $\beta
^{\prime }\in (0,\beta )$. \newline
For $|y|\leq 1$, $z,\bar{z}\in \mathbf{R}^{d},$
\begin{eqnarray*}
&&[\nabla f(x+G(z)y)-\nabla f(x)]G_{z_{j}}(z)y]-[\nabla f(x+G(\bar{z
)y)-\nabla f(x)]G_{z_{j}}(\bar{z})y] \\
&\leq &|\nabla f(x+G(z)y)-\nabla f(x+G(\bar{z})y)|~|G_{z_{j}}(z)y| \\
&&+|\nabla f(x+sG(\bar{z})y)-\nabla f(x)||G_{z_{j}}(z)-G_{z_{j}}(\bar{z})||y|
\\
&\leq &|D^{2}f|_{0}|y|^{2}[|G|_{0}|G_{z_{j}}(z)-G_{z_{j}}(\bar{z
)|+|G_{z_{j}}|_{0}|G(z)-G(\bar{z})|]
\end{eqnarray*
and
\begin{equation*}
|B_{z}^{j,1}f-B_{\bar{z}}^{j,1}f|\leq C|D^{2}f|_{0}||G||_{\beta }^{2}|z-\bar
z}|^{\beta -1}.
\end{equation*
Sinc
\begin{equation*}
B_{z}^{j,2}f(x)=\int_{|y|>1}\nabla f(x+G(z)y)G_{z_{j}}(z)yd\pi ,
\end{equation*
then
\begin{eqnarray*}
|B_{z}^{j,2}f(\cdot )|_{\beta -1} &\leq &C|\nabla f|_{\beta
-1}|G_{z_{j}}|_{0}, \\
|B_{z+h}^{j,2}f(x)-B_{z}^{j,2}f(x)| &\leq &C|h|^{\beta -1}[|\nabla
f|_{0}||G_{z_{j}}||_{\beta -1}+\int_{|y|>1}|y|^{\mu }d\pi
|D^{2}f|_{0}(1+||G||_{\beta }^{2})],
\end{eqnarray*
$z,x,h\in \mathbf{R}^{d}.$
It hence proves that (\ref{fo8}) holds.
\textit{Case III}: $\beta \in (2,3),$ $\beta \leq \mu <\alpha +\beta $. Sinc
\begin{equation*}
\partial _{j}(Bf(x))=\big(\frac{\partial }{\partial _{z_{j}}}B_{z}f(x)\big
|_{z=x}+B_{z}f_{x_{j}}|_{z=x},
\end{equation*
then
\begin{eqnarray}
\frac{\partial ^{2}}{\partial x_{i}\partial x_{j}}\left( Bf(x)\right)
&=&B_{z}f_{x_{i}x_{j}}(x)|_{z=x}+\frac{\partial }{\partial z_{i}}\left(
B_{z}f_{x_{j}}\right) |_{z=x} \label{fo00} \\
&&+\frac{\partial }{\partial _{z_{j}}}B_{z}f_{x_{i}}(x)|_{z=x}+\frac
\partial ^{2}}{\partial z_{i}\partial z_{j}}B_{z}f(x)|_{z=x} \notag \\
&=&B\partial _{ij}^{2}f+B^{i}\partial _{j}f+B^{j}\partial _{i}f+B^{ij}f.
\notag
\end{eqnarray
The estimate (\ref{fos}) of \textit{Case I} can be used for the first term (
\beta -2\leq \mu -2<\alpha +\beta -2$ with $\beta -2\in (0,1)$). For each
\varepsilon ^{\prime }$, there exists a constant $C_{\varepsilon ^{\prime }}$
such tha
\begin{equation*}
|Bf_{x_{i}x_{j}}|_{\beta -2}\leq \varepsilon ^{\prime }|D^{2}f|_{\alpha
+\beta -2}+C_{\varepsilon ^{\prime }}|D^{2}f|_{0}.
\end{equation*
For the second and third term in (\ref{fo00}), estimate (\ref{fo8}) of
\textit{Case II} is applied. Indeed, $f_{x_{j}}\in C^{\alpha +\beta -1}
\mathbf{R}^{d})$, $\beta -1\in (1,2)$, $\beta -1\leq \mu -1<\alpha +\beta
-1. $ Hence, for each $\varepsilon ^{\prime }$, there exists a constant
C_{\varepsilon ^{\prime }}$ such tha
\begin{equation*}
|B^{i}f_{x_{j}}|_{\beta -2}+|B^{j}f_{x_{i}}|_{\beta -2}\leq \varepsilon
^{\prime }|\nabla f|_{\alpha +\beta -1}+C_{\varepsilon ^{\prime }}|\nabla
f|_{0}.
\end{equation*}
Therefore, only the last term is new. By (\ref{fo7})
\begin{eqnarray*}
\frac{\partial ^{2}}{\partial z_{i}\partial z_{j}}B_{z}f(x) &=&\int
(D^{2}f(x+G(z)y)G_{z_{j}}(z)y,G_{z_{i}}(z)y)d\pi \\
&&+\mathbf{1}_{\{\alpha \in (0,1]\}}\int \nabla
f(x+G(z)y)G_{z_{i}z_{j}}(z)yd\pi \\
&&+\mathbf{1}_{\{\alpha \in (1,2]\}}\int \left( \nabla f(x+G(z)y)-\nabla
f(x),G_{z_{i}z_{j}}(z)y\right) d\pi \\
&=&B_{z}^{ij,1}f(x)+B_{z}^{ij,2}f(x)+B_{z}^{ij,3}f(x),
\end{eqnarray*
and for $\alpha \in (1,2]$
\begin{equation*}
B_{z}^{ij,3}f(x)=\int
\int_{0}^{1}(D^{2}f(x+sG(z)y)G(z)y,G_{z_{i}z_{j}}(z)y)dsd\pi .
\end{equation*
It then follows that for $z\in \mathbf{R}^{d}$,
\begin{eqnarray*}
|B_{z}^{ij,1}f(\cdot )|_{\beta -2} &\leq &|\partial ^{2}f|_{\beta -2}|\nabla
G|_{0}^{2},|B_{z}^{ij,2}f(\cdot )|_{\beta -2}\leq |\nabla f|_{\beta
-2}|\nabla G|_{0}|\partial ^{2}G|_{0}, \\
|B_{z}^{ij,3}f(\cdot )|_{\beta -2} &\leq &|D^{2}f|_{\beta
-2}|D^{2}G|_{0}|G|_{0}.
\end{eqnarray*
Let $\beta \leq \mu <\alpha +\beta ^{\prime }<\alpha +\beta $. Then for
x,z,h\in \mathbf{R}^{d}$,
\begin{eqnarray*}
|B_{z+h}^{ij,1}f(x)-B_{z}^{ij,1}f(x)|_{\beta -2} &\leq &|h|^{\beta
-2}(|D^{2}f|_{0}||\nabla G||_{\beta -2}^{2}+|D^{2}f|_{\alpha +\beta ^{\prime
}}|\nabla G|_{0}^{3}\int |y|^{\mu }d\pi ), \\
|B_{z+h}^{ij,2}f(x)-B_{z}^{ij,2}f(x)|_{\beta -2} &\leq &C|h|^{\beta
-2}(|D^{2}f|_{0}||G||_{\beta }^{2}+|\nabla f|_{0}||G||_{\beta }), \\
|B_{z+h}^{ij,3}f(x)-B_{z}^{ij,3}f(x)|_{\beta -2} &\leq &|h|^{\beta
-2}(|D^{3}f|_{0}||G||_{\beta }^{3}+|D^{2}f|_{0}||G||_{\beta }^{2}).
\end{eqnarray*}
Sinc
\begin{eqnarray*}
&&B_{x+h}f(x+h)-B_{x}f(x) \\
&=&B_{x+h}f(x+h)-B_{x+h}f(x) \\
&&+B_{x+h}f(x)-B_{x}f(x)
\end{eqnarray*
an
\begin{eqnarray*}
&&B_{x+h}f(x+h)-2B_{x}f(x)+B_{x-h}f(x-h) \\
&=&B_{x+h}f(x+h)-2B_{x+h}f(x)+B_{x+h}f(x-h) \\
&&+2[B_{x+h}f(x)-B_{x}f(x)+ \\
&&+[B_{x-h}f(x-h)-B_{x+h}f(x-h)],
\end{eqnarray*
the statement follows.
\end{proof}
\subsubsection{Proof of Theorem \protect\ref{thm:StoCP}}
The proof follows that of Theorem 5 in~\cite{MiP09}, with some simple
changes.
It is well known that for an arbitrary but fixed $\delta >0$, there exist a
family of cubes $D_{k}\subseteq \tilde{D}_{k}\subseteq \mathbf{R}^{d}$ and a
family of deterministic functions $\eta _{k}\in C_{0}^{\infty }(\mathbf{R
^{d})$ with the following properties:
\begin{enumerate}
\item For all $k\geq 1,D_{k}$ and $\tilde{D}_{k}$ have a common center
x_{k},$ diam $D_{k}\leq \delta$, dist$(D_{k},\mathbf{R}^{d}\backslash \tilde
D}_{k})\leq C\delta $ for a constant $C=C(d)>0$, $\cup _{k}D_{k}=\mathbf{R
^{d},$ and $1\leq \sum_{k}\mathbf{1}_{\tilde{D}_{k}}\leq 2^{d}.$
\item For all $k$, $0\leq \eta _{k}\leq 1,\eta _{k}=1$ in $D_{k},\eta _{k}=0$
outside of $\tilde{D}_{k}$ and for all multiindices $\gamma$ with $|\gamma
|\leq 3,
\begin{equation*}
|\partial ^{\gamma }\eta _{k}|\leq C(d)\delta ^{-|\gamma |}.
\end{equation*}
\end{enumerate}
For $\alpha \in (0,2)$, $\lambda \geq 0$, $k\geq 1,u\in C^{\alpha +\beta
}(H),$ denot
\begin{eqnarray*}
Au(t,x) &=&A_{x}u(t,x),\quad Bu(t,x)=B_{x}u(t,x),A_{k}u(t,x)=A_{x_{k}}u(t,x),
\\
E_{k}u(t,x) &=&\int [u(t,x+y)-u(t,x)][\eta _{k}(x+y)-\eta _{k}(x)]m(x_{k},y
\frac{dy}{|y|^{d+\alpha }}, \\
F_{k}u(t,x) &=&u(t,x)A_{k}\eta _{k}(x).
\end{eqnarray*}
It is readily checked that there is $\beta ^{\prime }\in (0,\beta )$ and a
constant $C$ such tha
\begin{equation*}
\sup_{k}|E_{k}u(t,\cdot )|_{\beta }\leq C|u|_{\alpha +\beta ^{\prime }}
\end{equation*
an
\begin{equation*}
\sup_{k}|F_{k}u(t,\cdot )|_{\beta }\leq C|u|_{\beta }.
\end{equation*}
Elementary calculation shows that for every $u\in C^{\beta }(H),
\begin{eqnarray*}
|u|_{0} &\leq &\sup_{k}\sup_{x}|\eta _{k}(x)u(x)|, \\
|u|_{\beta } &\leq &\sup_{k}|\eta _{k}u|_{\beta }+C|u|_{0}\leq
C\sup_{k}|\eta _{k}u|_{\beta }, \\
\sup_{k}|\eta _{k}u|_{\beta } &\leq &|u|_{\beta }+C|u|_{0}\leq C|u|_{\beta },
\end{eqnarray*
with $C=C(\alpha ,\delta ,d)$. In particular,
\begin{equation}
|u|_{\alpha +\beta }\leq C\sup_{k}|\eta _{k}u|_{\alpha +\beta }. \label{in1}
\end{equation}
Let $u\in C^{\alpha +\beta }(H)$ be a solution to (\ref{eq1}). Then $\eta
_{k}u$ satisfies the equatio
\begin{eqnarray}
\partial _{t}(\eta _{k}u) &=&A_{k}(\eta _{k}u)-\lambda (\eta _{k}u)+\eta
_{k}(Au-A_{k}u) \label{eqk} \\
&&+\eta _{k}Bu+\eta _{k}f-F_{k}u-E_{k}u, \notag
\end{eqnarray
and by Proposition \ref{prop1}
\begin{equation*}
|\eta _{k}u|_{\alpha +\beta }\leq C[|\eta _{k}(Au-A_{k}u)|_{\beta }+|\eta
_{k}Bu|_{\beta }+|\eta _{k}f|_{\beta }+|F_{k}u|_{\beta }+|E_{k}u|_{\beta }.
\end{equation*
Hence
\begin{equation}
|u|_{\alpha +\beta }\leq C[\sup_{k}|\eta _{k}f|_{\beta }+I], \label{form00}
\end{equation
wher
\begin{eqnarray*}
I &\leq &C_{1}\sup_{k}[|\eta _{k}(Au-A_{k}u)|_{\beta }+|\eta _{k}Bu|_{\beta }
\\
&&+|F_{k}u|_{\beta }+|E_{k}u|_{\beta }]+C_{2}|u|_{0}.
\end{eqnarray*
By Lemma \ref{prop2}, there exist $\beta ^{\prime }<\beta $, a constant $C$
not depending on $\delta $ and a constant $C=C(\delta )$
\begin{equation*}
|\eta _{k}(Au-A_{k}u)|_{\beta }\leq C[C(\delta )|u|_{\alpha +\beta ^{\prime
}}+\delta ^{\beta }|u|_{\alpha +\beta }].
\end{equation*
Therefore, for each $\varepsilon >0$, there is a constant $C=C(\varepsilon )
$ such tha
\begin{equation*}
|\eta _{k}(Au-A_{k}u)|_{\beta }\leq \varepsilon |u|_{\alpha +\beta
}+C(\varepsilon )|u|_{0}.
\end{equation*
By the estimates of Proposition \ref{b1}, it follows that for each
\varepsilon >0$, there exists a constant $C_{\varepsilon }$ such tha
\begin{equation*}
I\leq \varepsilon |u|_{\alpha +\beta }+C_{\varepsilon }|u|_{0}.
\end{equation*}
By (\ref{form00}),
\begin{equation}
|u|_{\alpha +\beta }\leq C[|f|_{\beta }+|u|_{0}]. \label{form01}
\end{equation}
On the other hand, (\ref{eqk}) holds and by Proposition \ref{prop1}
\begin{eqnarray*}
|u|_{0} &\leq &\sup_{k}|\eta _{k}u|_{\beta } \\
&\leq &\mu (\lambda )\sup_{k}[|f|_{\beta }+|\eta _{k}(Au-A_{k})|_{\beta
}+|\eta _{k}Bu|_{\beta }+|F_{k}u|_{\beta }+|E_{k}u|_{\beta }],
\end{eqnarray*
where $\mu (\lambda )\rightarrow 0$ as $\lambda \rightarrow \infty $. Thus
\begin{equation}
|u|_{0}\leq C\mu (\lambda )[|f|_{\beta }+|u|_{\alpha +\beta }].
\label{form02}
\end{equation
The inequalities (\ref{form01}) and (\ref{form02}) imply that there exist
\lambda _{0}>0$ and a constant $C$ independent of $u$ such that if $\lambda
\geq \lambda _{0}$,
\begin{equation}
|u|_{\alpha +\beta }\leq C|f|_{\beta }. \label{form03}
\end{equation
In a standard way (see ~\cite{MiP09}), it can be verified that (\ref{form03
) holds for all $\lambda \geq 0.$ Again by Proposition \ref{prop1} and (\re
{in1}), there exists a constant $C$ such that for all $s\leq t\leq T,$
\begin{eqnarray*}
|u(t,\cdot )-u(s,\cdot )|_{\frac{\alpha }{2}+\beta } &\leq &\sup_{k}|\eta
_{k}u(t,\cdot )-\eta _{k}u(s,\cdot )|_{\frac{\alpha }{2}+\beta } \\
&\leq &C(t-s)^{\frac{1}{2}}\left( |f|_{\beta }+|u|_{\alpha +\beta }\right) .
\end{eqnarray*
Therefore there exists a constant $C$ such that for all $s\leq t\leq T,
\begin{equation*}
|u(t,\cdot )-u(s,\cdot )|_{\frac{\alpha }{2}+\beta }\leq C(t-s)^{\frac{1}{2
}|f|_{\beta }.
\end{equation*}
To finish the proof, apply the continuation by parameter argument. Let $\tau
\in \left[ 0,1\right] $, $L_{\tau }u=\tau Lu+\left( 1-\tau \right) \partial
^{\alpha }u$ with $L=A+B$ and introduce the space $\hat{C}^{\alpha +\beta
}\left( H\right) $ of functions $u\in C^{\alpha +\beta }(H)$ such that for
each $\left( t,x\right) $, $u\left( t,x\right) =\int_{t}^{T}F\left(
s,x\right) ds$, where $F\in C^{\beta }\left( H\right) .$ It is a Banach
space with respect to the norm
\begin{equation*}
\left\vert u\right\vert _{\alpha ,\beta }=\left\vert u\right\vert _{\alpha
+\beta }+\left\vert F\right\vert _{\beta }.
\end{equation*
Consider the mappings $T_{\tau }:\hat{C}^{\alpha +\beta }\left( H\right)
\rightarrow C^{\beta }(H)$ defined by $u\left( t,x\right)
=-\int_{t}^{T}F\left( s,x\right) \,ds\longmapsto F+L_{\tau }u$. By Lemma \re
{prop2} and Proposition \ref{b1}, for some constant $C$ independent of $\tau
,$ $\left\vert T_{\tau }u\right\vert _{\beta }\leq C\left\vert u\right\vert
_{\alpha ,\beta }$. On the other hand, there exists a constant $C$
independent of $\tau $ such that for all $u\in \hat{C}^{\alpha +\beta
}\left( H\right) $
\begin{equation}
\left\vert u\right\vert _{\alpha ,\beta }\leq C\left\vert T_{\tau
}u\right\vert _{\beta }. \label{cp1}
\end{equation
Indeed,
\begin{equation*}
u\left( t,x\right) =-\int_{t}^{T}F\left( s,x\right) \,ds=\int_{t}^{T}\left(
L_{\tau }u-(F+L_{\tau }u)\right) \,ds.
\end{equation*
According to the estimate (\ref{form03}), there exists a constant $C$
independent oh $\tau $ such that
\begin{equation}
\left\vert u\right\vert _{\alpha +\beta }\leq C\left\vert T_{\tau
}u\right\vert _{\beta }=C\left\vert F+L_{\tau }u\right\vert _{\beta }.
\label{cp2}
\end{equation
Hence, by Lemma \ref{prop2}, Proposition \ref{b1} and (\ref{cp2}),
\begin{eqnarray*}
\left\vert u\right\vert _{\alpha ,\beta } &=&\left\vert u\right\vert
_{\alpha +\beta }+\left\vert F\right\vert _{\beta }\leq \left\vert
u\right\vert _{\alpha +\beta }+\left\vert F+L_{\tau }u\right\vert _{\beta
}+\left\vert L_{\tau }u\right\vert _{\beta } \\
&\leq &C(\left\vert u\right\vert _{\alpha +\beta }+\left\vert F+L_{\tau
}u\right\vert _{\beta })\leq C\left\vert F+L_{\tau }u\right\vert _{\beta
}=C\left\vert T_{\tau }u\right\vert _{\beta },
\end{eqnarray*
and (\ref{cp1}) follows. Since $T_{0}$ is an onto map, by Theorem 5.2 in
\cite{GiT83}, all the $T_{\tau }$ are onto maps and the statement follows.
\subsubsection{Proof of Corollary \protect\ref{lcornew1}}
By Lemma \ref{prop2} and Proposition \ref{b1}, for $g\in C^{\alpha +\beta }
\mathbf{R}^{d})$, $|Ag|_{\beta }\leq C|g|_{\alpha +\beta }$ and $|Bg|_{\beta
}\leq C|g|_{\alpha +\beta }$ with a constant $C$ independent of $f$ and $g$.
It then follows from (\ref{eq1}) that there exists a unique solution $\tilde
v}\in C^{\alpha +\beta }(H)$ to the Cauchy problem
\begin{eqnarray} \label{eqn:cauchy_v}
\big(\partial _{t}+A_{x}+B_{x}\big)\tilde{v}(t,x)
&=&f(t,x)-A_{x}g(x)-B_{x}g(x), \\
\tilde{v}(T,x) &=&0 \notag
\end{eqnarray
and $|\tilde{v}|_{\alpha +\beta }\leq C\big(|g|_{\alpha +\beta }+|f|_{\beta
\big)$ with $C$ independent of $f$ and $g$. Let $v(t,x)=\tilde{v}(t,x)+g(x)
, where $\tilde{v}$ is the solution to problem (\ref{eqn:cauchy_v}). Then $v$
is the unique solution to the Cauchy problem (\ref{maf8}) and $|v|_{\alpha
+\beta }\leq C(|g|_{\alpha +\beta }+|f|_{\beta })$.
\begin{remark}
\label{rlast}If the assumptions of Corollary \ref{lcornew1} hold and $v\in
C^{\alpha +\beta }(H)$ is the solution to $(\ref{maf8})$, then $\partial
_{t}v=f-A_{x}v-B_{x}v$, and by Lemma \ref{prop2} and Proposition \ref{b1},
|\partial _{t}v|_{\beta }\leq C(|g|_{\alpha +\beta }+|f|_{\beta }).$
\end{remark}
\section{One-Step Estimate and Proof of Main Result}
The following Lemma provides a one-step estimate of the conditional
expectation of an increment of the Euler approximation.
\begin{lemma}
\label{lem:expect}Let $\beta \in (0,3)$, $0<\beta \leq \mu <\alpha +\beta $,
and
\begin{equation*}
\int_{|y|\leq 1}|y|^{\alpha }\pi (dy)+\int_{|y|>1}|y|^{\mu }\pi (dy)<\infty .
\end{equation*
Assume $a^{i},b^{ij}\in \tilde{C}^{\beta }(\mathbf{R}^{d}),G^{ij}\in \tilde{
}^{\frac{\beta }{\mu \wedge 1}}(\mathbf{R}^{d})$. Then there exists a
constant $C$ such that for all $f\in C^{\beta }(\mathbf{R}^{d}),
\begin{equation*}
\big\vert\mathbf{E}\big[f(Y_{s})-f(Y_{\tau _{i_{s}}})|\widetilde{\mathcal{F}
_{\tau _{i_{s}}}\big]\big\vert\leq C|f|_{\beta }r(\delta ,\alpha ,\beta
),\forall s\in \lbrack 0,T],
\end{equation*
where $i_{s}=i$ if $\tau _{i}\leq s<\tau _{i+1}$ and $r(\delta ,\alpha
,\beta )$ is as defined in Theorem \ref{thm:main}.
\end{lemma}
The proof of Lemma~\ref{lem:expect} is based on applying It\^{o}'s formula
to $f(Y_{s})-f(Y_{\tau _{i_{s}}})$, $f\in C^{\beta }(\mathbf{R}^{d})$. If
\beta >\alpha $, by Remark \ref{lrenew1} and It\^{o}'s formula, the
inequality holds. If $\beta \leq \alpha $, $f$ is first smoothed by using
w\in C_{0}^{\infty }(\mathbf{R}^{d}),$ a nonnegative smooth function with
support on $\{|x|\leq 1\}$ such that $w(x)=w(|x|)$, $x\in \mathbf{R}^{d},$
and $\int w(x)dx=1$ (see (8.1) in \cite{Fol99})$.$ Note that, due to the
symmetry,
\begin{equation}
\int_{\mathbf{R}^{d}}x^{i}w(x)dx=0,i=1,\ldots ,d. \label{ff26}
\end{equation}
For $x\in \mathbf{R}^{d}$ and $\varepsilon \in (0,1)$, define
w^{\varepsilon }(x)=\varepsilon ^{-d}w\left( \frac{x}{\varepsilon }\right) $
and the convolution
\begin{equation}
f^{\varepsilon }(x)=\int f(y)w^{\varepsilon }(x-y)dy=\int
f(x-y)w^{\varepsilon }(y)dy,x\in \mathbf{R}^{d}. \label{maf7}
\end{equation}
\subsection{Some Auxiliary Estimates}
For the estimates of $A_{z}f^{\varepsilon }$, the following simple integral
estimates are needed. Recall that $m(z,y)$ in the definition of operator
A_{z}$ (see (\ref{foo})) is bounded, smooth, and 0-homogeneous and symmetric
in $y$.
\begin{lemma}
\label{rem2} Let $v\in C_{0}^{\infty }(\mathbf{R}^{d})$.
\begin{enumerate}
\item[\textup{(i)}] For $\alpha \in (0,2)$
\begin{equation*}
\int_{\mathbf{R}^{d}}\int_{\mathbf{R}_{0}^{d}}|v(y+y^{\prime })-v(y)-\chi
^{(\alpha )}(y^{\prime })(\nabla v(y),y^{\prime })|\frac{dydy^{\prime }}
|y^{\prime }|^{d+\alpha }}<\infty ,
\end{equation*
where $\chi ^{(\alpha )}(y)=\mathbf{1}_{\{|y|\leq 1\}}\mathbf{1}_{\{\alpha
=1\}}+\mathbf{1}_{\{\alpha \in (1,2)\};}$
\item[\textup{(ii)}] For $\beta \in (0,1]$, $\beta <\alpha ,z\in \mathbf{R
^{d},
\begin{equation*}
\sup_{z}\int_{\mathbf{R}^{d}}|(A_{z}v)(y)||y|^{\beta }dy<\infty ,
\end{equation*
and for $\beta \in (0,1]$, $\beta =\alpha ,z\in \mathbf{R}^{d},k>1,
\begin{equation*}
\sup_{z}\int_{\mathbf{R}^{d}}|(A_{z}v)(y)|(|y|^{\alpha }\wedge k)dy\leq
C(1+\ln k).
\end{equation*}
\item[\textup{(iii)}] For $1<\beta <\alpha <2,
\begin{equation*}
\int_{\mathbf{R}^{d}}\int_{\mathbf{R}_{0}^{d}}\int_{0}^{1}|v(y+sy^{\prime
})-v(y)|~|y|^{\beta -1}\frac{dsdydy^{\prime }}{|y^{\prime }|^{d+\alpha -1}
<\infty ,
\end{equation*
and for $1<\beta =\alpha <2,k>1,
\begin{eqnarray*}
&&\int_{\mathbf{R}^{d}}\int_{\mathbf{R}_{0}^{d}}\int_{0}^{1}|v(y+sy^{\prime
})-v(y)|~(|y|^{\beta -1}\wedge k)\frac{dsdydy^{\prime }}{|y^{\prime
}|^{d+\alpha -1}} \\
&\leq &C(1+\ln k).
\end{eqnarray*}
\end{enumerate}
\end{lemma}
\begin{proof}
(i) Indeed
\begin{eqnarray*}
&&|v(y+y^{\prime })-v(y)-\chi ^{(\alpha )}(y^{\prime })(\nabla
v(y),y^{\prime })| \\
&\leq &\mathbf{1}_{\{|y^{\prime }|\leq 1\}}\big\{\int_{0}^{1}[\max_{i,j}
\partial _{ij}^{2}v(y+sy^{\prime })|~|y^{\prime }|^{2}+\mathbf{1}_{\{\alpha
\in (0,1)\}}|\nabla v(y+sy^{\prime })||y^{\prime }|]ds\big\} \\
&&+\mathbf{1}_{\{|y^{\prime }|>1\}}\big\{|v(y+y^{\prime })|+|v(y)|+\mathbf{1
_{\{\alpha \in (1,2)\}}|\nabla v(y)|~|y^{\prime }|\big\},y,y^{\prime }\in
\mathbf{R}^{d}.
\end{eqnarray*
The claim follows.
(ii) For $\beta \in (0,1]$, $\beta <\alpha $, $z\in \mathbf{R}^{d}$,
\begin{eqnarray*}
\int_{\mathbf{R}^{d}}|(A_{z}v)(y)||y|^{\beta }dy &\leq &\int_{\mathbf{R
^{d}}\int_{|y^{\prime }|>1}|v(y+y^{\prime })|~|y|^{\beta }\frac{dydy^{\prime
}}{|y^{\prime }|^{d+\alpha }} \\
&&+\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|>1}|v(y)||y|^{\beta }\frac
dydy^{\prime }}{|y^{\prime }|^{d+\alpha }} \\
&&+\max_{i,j}\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|\leq
1}\int_{0}^{1}|\partial _{ij}^{2}v(y+sy^{\prime })|~|y^{\prime
}|^{2}|y|^{\beta }\frac{dsdy^{\prime }dy}{|y^{\prime }|^{d+\alpha }}
\end{eqnarray*
an
\begin{eqnarray*}
\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|>1}|v(y+y^{\prime })||y|^{\beta
\frac{dydy^{\prime }}{|y^{\prime }|^{d+\alpha }} &\leq &C\Big[\int_{\mathbf{
}^{d}}\int_{|y^{\prime }|>1}|v(y+y^{\prime })||y+y^{\prime }|^{\beta }\frac
dydy^{\prime }}{|y^{\prime }|^{d+\alpha }} \\
&&+\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|>1}|v(y+y^{\prime })||y^{\prime
}|^{\beta }\frac{dydy^{\prime }}{|y^{\prime }|^{d+\alpha }}\Big].
\end{eqnarray*
Let $\beta \in (0,1],\beta =\alpha .$ Assume $v(x)=0$ if $|x|>R$. We have
for $k>1$ with $A=(R+1)^{1/\alpha },
\begin{eqnarray*}
&&\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|>1}|v(y+y^{\prime })|~(|y|^{\alpha
}\wedge k)\frac{dydy^{\prime }}{|y^{\prime }|^{d+\alpha }} \\
&\leq &\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|>1}|v(y+y^{\prime
})|~(|y|^{\alpha }\wedge Ak)\frac{dydy^{\prime }}{|y^{\prime }|^{d+\alpha }}
\\
&=&\int_{|y|\leq (R+1)k^{1/\alpha }}\int_{|y^{\prime }|>1}|v(y+y^{\prime
})|~|y|^{\alpha }\frac{dydy^{\prime }}{|y^{\prime }|^{d+\alpha }} \\
&&+k\int_{|y|>(R+1)k^{1/\alpha }}\int_{\left\vert y^{\prime }\right\vert
>1}|v(y+y^{\prime })|\frac{dydy^{\prime }}{|y^{\prime }|^{d+\alpha }} \\
&=&A_{1}+A_{2.}
\end{eqnarray*
Then
\begin{eqnarray*}
|A_{1}| &\leq &\int \int_{1\leq |y^{\prime }|\leq (R+1)(1+k^{1/\alpha
})}|v(y+y^{\prime })|\frac{dy^{\prime }dy}{|y^{\prime }|^{d}} \\
&\leq &C(1+\ln k).
\end{eqnarray*
Since for $|y+y^{\prime }|\leq R,|y|>(R+1)k^{1/\alpha },$ we have
|y^{\prime }|\geq (R+1)k^{1/\alpha }-R\geq k^{1/\alpha }$, it follows
\begin{eqnarray*}
|A_{2}| &\leq &k\int \int_{|y^{\prime }|\geq k^{1/\alpha }}|v(y+y^{\prime })
\frac{dydy^{\prime }}{|y^{\prime }|^{d+\alpha }} \\
&\leq &Ckk^{-1}=C.
\end{eqnarray*
The
\begin{eqnarray*}
&&\int \int_{|y^{\prime }|\leq 1}|v(y+y^{\prime })-v(y)-1_{\alpha =1}(\nabla
v(y),y^{\prime })|~(|y|^{\alpha }\wedge k)\frac{dydy^{\prime }}{|y^{\prime
}|^{d+\alpha }} \\
&\leq &\int_{0}^{1}\int \int_{|y^{\prime }|\leq 1}|\nabla v(y+sy^{\prime
})-1_{\alpha =1}\nabla v(y)|~|y|^{\alpha }\frac{dydy^{\prime }}{|y^{\prime
}|^{d+\alpha -1}}<\infty .
\end{eqnarray*
Part (ii) follows.
(iii) For $1<\beta <\alpha <2,$
\begin{eqnarray*}
&&\int_{\mathbf{R}^{d}}\int_{\mathbf{R}_{0}^{d}}\int_{0}^{1}|v(y+sy^{\prime
})-v(y)|~|y|^{\beta -1}\frac{dydy^{\prime }ds}{|y^{\prime }|^{d+\alpha -1}}
\\
&\leq &\int_{\mathbf{R}^{d}}\int_{|y^{\prime
}|>1}\int_{0}^{1}|v(y+sy^{\prime })||y|^{\beta -1}\frac{dydy^{\prime }ds}
|y^{\prime }|^{d+\alpha -1}} \\
&&+\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|>1}\int_{0}^{1}|v(y)||y|^{\beta
-1}\frac{dydy^{\prime }ds}{|y^{\prime }|^{d+\alpha -1}} \\
&&+\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|\leq
1}\int_{0}^{1}\int_{0}^{1}|\nabla v(y+s\tau y^{\prime })||y|^{\beta -1}\frac
dsd\tau dydy^{\prime }}{|y^{\prime }|^{d+\alpha -2}}.
\end{eqnarray*
Sinc
\begin{eqnarray*}
&&\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|>1}\int_{0}^{1}|v(y+sy^{\prime
})||y|^{\beta -1}\frac{dydy^{\prime }ds}{|y^{\prime }|^{d+\alpha -1}} \\
&\leq &C\Big[\int_{\mathbf{R}^{d}}\int_{|y^{\prime
}|>1}\int_{0}^{1}|v(y+sy^{\prime })||y+sy^{\prime }|^{\beta -1}\frac
dydy^{\prime }ds}{|y^{\prime }|^{d+\alpha -1}} \\
&&+\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|>1}\int_{0}^{1}|v(y+sy^{\prime
})||y^{\prime }|^{\beta -1}\frac{dydy^{\prime }ds}{|y^{\prime }|^{d+\alpha
-1}}\Big]
\end{eqnarray*
and similarl
\begin{eqnarray*}
&&\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|\leq
1}\int_{0}^{1}\int_{0}^{1}|\nabla v(y+s\tau y^{\prime })||y|^{\beta -1}\frac
dsd\tau dydy^{\prime }}{|y^{\prime }|^{d+\alpha -2}} \\
&\leq &C\Big[\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|\leq
1}\int_{0}^{1}\int_{0}^{1}|\nabla v(y+s\tau y^{\prime })||y+s\tau y^{\prime
}|^{\beta -1}\frac{dsd\tau dydy^{\prime }}{|y^{\prime }|^{d+\alpha -2}} \\
&&+\int_{\mathbf{R}^{d}}\int_{|y^{\prime }|\leq
1}\int_{0}^{1}\int_{0}^{1}|\nabla v(y+s\tau y^{\prime })||y^{\prime
}|^{\beta -1}\frac{dsd\tau dydy^{\prime }}{|y^{\prime }|^{d+\alpha -2}}\Big]
\end{eqnarray*
are finite, the first estimate in (iii) follows.
If $1<\beta =\alpha <2$, we simply repeat the proof in part (ii). The
statement follows.
\end{proof}
\begin{remark}
The estimate in part (iii) implies that (ii) can be extended to all $0<\beta
\leq \alpha <2.$
\end{remark}
We will use the following modulus of continuity estimate of a function $f\in
C^{1}(\mathbf{R}^{d}).$
\begin{lemma}
\label{lemn1}(cf. Lemma 5.6 in \cite{Caf}, Lemma 2.2 in \cite{kimd}) Let
f\in C^{1}(\mathbf{R}^{d})$ and $[f]_{1}\leq K$. Then there is a constant $C$
such that for all $x,h\in \mathbf{R}^{d},h\neq 0,
\begin{equation*}
|f(x+h)-f(x)|\leq C|h|(1+\left\vert \ln |h|\right\vert )|f|_{1}.
\end{equation*}
\end{lemma}
\begin{proof}
We follow the steps of Lemma 5.6 in \cite{Caf}. Fix $x,h\in \mathbf{R
^{d},h\neq 0$ such that $0<|h|<1/2$, we find a positive integer $k$ so tha
\begin{equation*}
2^{-k-1}\leq |h|<2^{-k}
\end{equation*
or $\frac{1}{2}2^{-k}\leq |h|<2^{-k}$. Set $\tau _{0}=2^{k}h$ (\textbf{note:
}$\frac{1}{2}\leq \tau _{0}<1,2^{-k}\leq 2|h|),\ln |h|<-k\ln 2$ or $k<\frac
-\ln |h|}{\ln 2}$). Define for $\tau \in \mathbf{R}^{d},
\begin{equation*}
v(\tau )=f(x+\tau )-f(x).
\end{equation*
Note tha
\begin{eqnarray*}
|v(\tau )-2v(\tau /2)| &=&|f(x+\tau )-2f(x+\tau /2)+f(x)| \\
&\leq &[f]_{1}|\tau |/2.
\end{eqnarray*
Thus,
\begin{equation*}
|2^{j-1}v(\tau _{0}/2^{j-1})-2^{j}v(\tau _{0}/2^{j})|\leq \lbrack
f]_{1}2^{j-1}|\tau _{0}|/2^{j}=2^{-1}[f]_{1}|\tau _{0}|,
\end{equation*
which implies ($v(\tau _{0})-2^{k}v(h)=\sum_{j=1}^{k}\left( 2^{j-1}v(\tau
_{0}/2^{j-1}\right) -\left( 2^{j}v(\tau _{0}/2^{j}\right) )$
\begin{eqnarray*}
|v(\tau _{0})-2^{k}v(h)| &\leq &\sum_{j=1}^{k}|2^{j-1}v(\tau
_{0}/2^{j-1})-2^{j}v(\tau _{0}/2^{j})| \\
&\leq &k2^{-1}|\tau _{0}|[f]_{1}.
\end{eqnarray*
Since $|v(\tau _{0})|\leq 2|f|_{0}$ or $|v(\tau _{0})|\leq \lbrack f]_{1/2})
, we deriv
\begin{eqnarray*}
|v(h)| &\leq &2^{-k}|2^{k}v(h)-v(\tau _{0})|+2^{-k}|v(\tau _{0})| \\
&\leq &[f]_{1}k2^{-1-k}|\tau _{0}|+2\cdot 2^{-k}|f|_{0} \\
&\leq &[f]_{1}k|h|+4|h||f|_{0} \\
&\leq &C|f|_{1}|h|(1-\ln |h|).
\end{eqnarray*
The statement follows.
\end{proof}
Now we prove some estimates for $Af^{\varepsilon }$ and $Bf^{\varepsilon }$.
\begin{lemma}
\label{lnew2}Let $\varepsilon \in (0,1).$
(i) Let $\alpha \in (0,2)$. Then there exists a constant $C$ such that for
all $z,x\in \mathbf{R}^{d},$
\begin{equation}
|A_{z}f^{\varepsilon }(x)|\leq C\kappa (\varepsilon ,\alpha ,\beta
)|f|_{\beta }, \label{ff29}
\end{equation
where $\kappa (\varepsilon ,\alpha ,\beta )=\varepsilon ^{-\alpha +\beta }$
if $\beta <\alpha $ and $\kappa (\varepsilon ,\alpha ,\beta )=1-\ln
\varepsilon $ if $\beta =\alpha ;$ in particular, for all $f\in C^{\beta }
\mathbf{R}^{d}),z,x\in \mathbf{R}^{d}$,
\begin{equation}
|\partial ^{\alpha }f^{\varepsilon }(x)|\leq C\kappa (\varepsilon ,\alpha
,\beta )|f|_{\beta }. \label{maf5}
\end{equation}
(ii)For each $\beta \in (0,2]$ there exists a constant $C$ such that for all
$f\in C^{\beta }(\mathbf{R}^{d}),x\in \mathbf{R}^{d}$,
\begin{equation*}
|f^{\varepsilon }(x)-f(x)|\leq C\gamma (\varepsilon ,\beta )|f|_{\beta },
\end{equation*
where $\gamma (\varepsilon ,\beta )=\varepsilon ^{\beta }$ if $\beta <2$ and
$\gamma (\varepsilon ,2)=\varepsilon ^{2}(1-\ln \varepsilon ).$
(iii) Let $\beta \in (0,2]$. For $k,l=1,\ldots ,d,x\in \mathbf{R}^{d},
\begin{eqnarray}
|\partial _{k}f^{\varepsilon }(x)| &\leq &C\kappa (\varepsilon ,1,\beta
)|f|_{\beta },\mbox{ if }\beta \leq 1, \label{ff30} \\
|\partial _{kl}^{2}f^{\varepsilon }(x)| &\leq &C\kappa (\varepsilon ,2,\beta
)|f|_{\beta },\mbox{ if }\beta \leq 2, \notag \\
|f^{\varepsilon }|_{1} &\leq &C|f|_{1}, \notag
\end{eqnarray
an
\begin{eqnarray}
|f^{\varepsilon }|_{\alpha } &\leq &C\varepsilon ^{-\alpha +\beta
}|f|_{\beta },\mbox{ if }\beta \in (0,1],\alpha \in \lbrack 1,2),
\label{maf5'} \\
|\partial ^{\alpha -1}\nabla f^{\varepsilon }(x)| &\leq &C\kappa
(\varepsilon ,\alpha ,\beta )|f|_{\beta },\mbox{ if }\beta \in (0,\alpha
],\alpha \in (1,2),\beta \neq \alpha -1. \label{maf6}
\end{eqnarray}
\end{lemma}
\begin{proof}
(i) For $z,x\in \mathbf{R}^{d},$ by changing the variable of integration
with $\bar{y}=\frac{y}{\varepsilon }$ and using (\ref{ff27}) for $\alpha =1
,
\begin{eqnarray}
A_{z}w^{\varepsilon }(x) &=&\mathbf{1}_{\{\alpha =1\}}(a(z),\nabla
w^{\varepsilon }(x)) \notag \\
&&+\int [w^{\varepsilon }(x+y)-w^{\varepsilon }(x)-\bar{\chi}_{\alpha
}(y)(\nabla w^{\varepsilon }(x),y)]m(z,y)\frac{dy}{|y|^{d+\alpha }} \notag
\\
&=&\varepsilon ^{-\alpha }\varepsilon ^{-d}(A_{z}w)(\frac{x}{\varepsilon }),
\label{and}
\end{eqnarray
where $\bar{\chi}_{\alpha }(y)=\mathbf{1}_{\{|y|\leq 1\}}\mathbf{1
_{\{\alpha =1\}}+\mathbf{1}_{\{\alpha \in (1,2)\}},y\in \mathbf{R}^{d}.$ It
follows from Lemma \ref{rem2}(i), the Fubini theorem, and (\ref{and}),
changing the variable of integration with $\bar{y}=\frac{y}{\varepsilon }$
as well, tha
\begin{eqnarray*}
A_{z}f^{\varepsilon }(x) &=&\int_{\mathbf{R}^{d}}\varepsilon ^{-\alpha
}\varepsilon ^{-d}(A_{z}w)(\frac{x-y}{\varepsilon })f(y)dy \\
&=&\int \varepsilon ^{-\alpha }\varepsilon ^{-d}(A_{z}w)(\frac{y}
\varepsilon })f(x-y)dy \\
&=&\int \varepsilon ^{-\alpha }(A_{z}w)(y)f(x-\varepsilon y)dy,x,z\in
\mathbf{R}^{d}.
\end{eqnarray*}
By Lemma \ref{rem2}(i) and the Fubini theorem
\begin{equation*}
\int_{\mathbf{R}^{d}}A_{z}w(y)dy=0.
\end{equation*
Also, it is easy to see that $A_{z}w(y)=A_{z}w(-y),y\in \mathbf{R}^{d}$.
Hence, if $\beta \in (0,1]$, $\beta \leq \alpha $
\begin{eqnarray*}
A_{z}f^{\varepsilon }(x) &=&\int \varepsilon ^{-\alpha
}(A_{z}w)(y)f(x-\varepsilon y)dy \\
&=&\frac{1}{2}\int \varepsilon ^{-\alpha }(A_{z}w)(y)[f(x-\varepsilon
y)+f(x+\varepsilon y)-2f(x)]dy
\end{eqnarray*
and
\begin{equation*}
|A_{z}f^{\varepsilon }(x)|\leq C\varepsilon ^{-\alpha +\beta }|f|_{\beta
}\int_{\mathbf{R}^{d}}|(A_{z}w)(y)|~(|y|^{\beta }\wedge \varepsilon ^{-\beta
})dy.
\end{equation*
So, by Lemma \ref{rem2}, (\ref{ff29}) holds for $\beta \leq \alpha ,\beta
\in (0,1].$
Assume $1<\beta \leq \alpha <2$. By Theorem 2.27 in \cite{Fol99},
differentiation and integration can be can switched
\begin{eqnarray*}
A_{z}w(y) &=&\int [w(y+y^{\prime })-w(y)-(\nabla w(y),y^{\prime
})]m(z,y^{\prime })\frac{dy^{\prime }}{|y^{\prime }|^{d+\alpha }} \\
&=&\int \int_{0}^{1}\big(\nabla _{y}w(y+sy^{\prime })-\nabla
_{y}w(y),y^{\prime }\big)dsm(z,y^{\prime })\frac{dy^{\prime }}{|y^{\prime
}|^{d+\alpha }} \\
&=&\sum_{i=1}^{d}\frac{\partial }{\partial y_{i}}\int
\int_{0}^{1}[w(y+sy^{\prime })-w(y)]y_{i}^{\prime }dsm(z,y^{\prime })\frac
dy^{\prime }}{|y^{\prime }|^{d+\alpha }}.
\end{eqnarray*
By integrating by parts,
\begin{eqnarray}
A_{z}f^{\varepsilon }(x) &=&\int \varepsilon ^{-\alpha
}A_{z}w(y)f(x-\varepsilon y)dy \notag \\
&=&\varepsilon ^{-\alpha +1}\int \int_{0}^{1}[w(y+sy^{\prime })-w(y)]
\label{ff28} \\
&&\times (\nabla f(x-\varepsilon y),y^{\prime })m(z,y^{\prime })\frac
dsdydy^{\prime }}{|y^{\prime }|^{d+\alpha }},x\in \mathbf{R}^{d}. \notag
\end{eqnarray
Since
\begin{equation*}
\int_{\mathbf{R}_{0}^{d}}\int_{\mathbf{R}^{d}}\int_{0}^{1}|w(y+sy^{\prime
})-w(y)|~|y^{\prime }|\frac{dsdydy^{\prime }}{|y^{\prime }|^{d+\alpha }
<\infty ,
\end{equation*
the Fubini theorem applies, $\int [w(y+sy^{\prime })-w(y)]dy=0$ and we can
rewrite (\ref{ff28}) a
\begin{eqnarray*}
A_{z}f^{\varepsilon }(x) &=&\varepsilon ^{-\alpha +1}\int_{\mathbf{R
^{d}}\int_{\mathbf{R}_{0}^{d}}\int_{0}^{1}[w(y+sy^{\prime })-w(y)] \\
&&\times (\nabla f(x-\varepsilon y)-\nabla f(x),y^{\prime })m(z,y^{\prime }
\frac{dsdydy^{\prime }}{|y^{\prime }|^{d+\alpha }},x,z\in \mathbf{R}^{d}.
\end{eqnarray*
Hence,
\begin{eqnarray*}
|A_{z}f^{\varepsilon }(x)| &\leq &C\varepsilon ^{-\alpha +1}\varepsilon
^{\beta -1}|\nabla f|_{\beta -1}\int \int \int_{0}^{1}|w(y+sy^{\prime
})-w(y)| \\
&&\times (|y|^{\beta -1}\wedge \varepsilon ^{-(\beta -1)})\frac{dydy^{\prime
}}{|y^{\prime }|^{d+\alpha -1}},
\end{eqnarray*
and by Lemma \ref{rem2}(iii), (\ref{ff29}) is proved for $1<\beta \leq
\alpha <2$. By taking $m=1,$ (\ref{maf5}) follows.
(ii) For $\beta \in (1,2)$, by (\ref{ff26}),
\begin{eqnarray*}
f^{\varepsilon }(x)-f(x) &=&\int [f(x-y)-f(x)]w^{\varepsilon }(y)dy \\
&=&\int [f(x+y)-f(x)-(\nabla f(x),y)]w^{\varepsilon }(y)dy \\
&=&\int \int_{0}^{1}(\nabla f(x+sy)-\nabla f(x),y)dsw^{\varepsilon }(y)dy
\end{eqnarray*}
an
\begin{equation*}
|f^{\varepsilon }(x)-f(x)|\leq C|\nabla f|_{\beta -1}\int |y|^{1+(\beta
-1)}w^{\varepsilon }(y)dy\leq C|f|_{\beta }\varepsilon ^{\beta }.
\end{equation*
If $\beta =2$, then by Lemma \ref{lemn1}
\begin{eqnarray*}
|f^{\varepsilon }(x)-f(x)| &\leq &C|\nabla f|_{1}\int |y|^{2}(1+\left\vert
\ln |y|\right\vert )w^{\varepsilon }(y)dy \\
&\leq &C|\nabla f|_{1}\varepsilon ^{2}(1-\ln \varepsilon ).
\end{eqnarray*}
For $\beta \in (0,1],
\begin{eqnarray*}
f^{\varepsilon }(x)-f(x) &=&\int [f(x-y)-f(x)]w^{\varepsilon }(y)dy \\
&=&\int [f(x+y)-f(x)]w^{\varepsilon }(y)dy
\end{eqnarray*
an
\begin{equation*}
f^{\varepsilon }(x)-f(x)=\frac{1}{2}\int [f(x+y)+f(x-y)-2f(x)]w^{\varepsilon
}(y)dy.
\end{equation*
Hence, for $\beta \in (0,1],
\begin{equation*}
|f^{\varepsilon }(x)-f(x)|\leq C|f|_{\beta }\varepsilon ^{\beta }.
\end{equation*}
(iii) If $\beta <1$, by changing the variable of integration
\begin{eqnarray*}
\partial _{k}f^{\varepsilon }(x) &=&\varepsilon ^{-1}\int_{\mathbf{R
^{d}}\varepsilon ^{-d}\partial _{k}w(\frac{x-y}{\varepsilon })f(y)dy \\
&=&\varepsilon ^{-1}\int_{\mathbf{R}^{d}}\varepsilon ^{-d}\partial _{k}w
\frac{y}{\varepsilon })f(x-y)dy \\
&=&\varepsilon ^{-1}\int_{\mathbf{R}^{d}}\partial _{k}w(y)[f(x-\varepsilon
y)-f(x)]dy,
\end{eqnarray*
and the first inequality follows by Lemma \ref{lemn1}. Sinc
\begin{eqnarray*}
&&f^{\varepsilon }(x+h)+f^{\varepsilon }(x-h)-2f^{\varepsilon }(x) \\
&=&\frac{1}{2}\int w_{\varepsilon }(y)[f(x-y+h)+f(x-y-h)-2f(x-y)]dy,
\end{eqnarray*
we have $|f^{\varepsilon }|_{1}\leq |f|_{1}$. Also, since $\partial
_{kl}^{2}w(y)=\partial _{kl}^{2}w(-y),k,l=1,\ldots ,d,y\in \mathbf{R}^{d},$
\begin{eqnarray*}
\partial _{kl}^{2}f^{\varepsilon }(x) &=&\varepsilon ^{-2}\int_{\mathbf{R
^{d}}\varepsilon ^{-d}\partial _{kl}^{2}w(\frac{x-y}{\varepsilon })f(y)dy \\
&=&\varepsilon ^{-2}\int_{\mathbf{R}^{d}}\varepsilon ^{-d}\partial
_{kl}^{2}w(\frac{y}{\varepsilon })f(x-y)dy \\
&=&\varepsilon ^{-2}\int_{\mathbf{R}^{d}}\partial
_{kl}^{2}w(y)[f(x-\varepsilon y)-f(x)]dy \\
&=&\frac{1}{2}\varepsilon ^{-2}\int_{\mathbf{R}^{d}}\partial
_{kl}^{2}w(y)[f(x+\varepsilon y)+f(x-\varepsilon y)-2f(x)]dy.
\end{eqnarray*}
\ \ \ \ \ Thus, for all $x\in \mathbf{R}^{d}$,
\begin{equation*}
\ |\partial _{kl}^{2}f^{\varepsilon }(x)|\leq C\varepsilon ^{-2+\beta
}|f|_{\beta }\text{ if }\beta \in (0,1].
\end{equation*
Similarly, if $1<\beta \leq 2,
\begin{eqnarray*}
\partial _{k}f^{\varepsilon }(x) &=&\int \varepsilon ^{-d}w(\frac{y}
\varepsilon })\partial _{k}f(x-y)dy \\
&=&\int \varepsilon ^{-d}w(\frac{x-y}{\varepsilon })\partial _{k}f(y)dy
\end{eqnarray*
an
\begin{eqnarray*}
\partial _{kl}^{2}f^{\varepsilon }(x) &=&\varepsilon ^{-1}\int \varepsilon
^{-d}\partial _{l}w(\frac{y}{\varepsilon })\partial _{k}f(x-y)dy \\
&=&\varepsilon ^{-1}\int \partial _{l}w(y)[\partial _{k}f(x-\varepsilon
y)-\partial _{k}f(x)]dy.
\end{eqnarray*
Hence, by Lemma \ref{lemn1},
\begin{equation*}
|\partial _{kl}^{2}f^{\varepsilon }(x)|\leq C\kappa (\varepsilon ,2,\beta
)|f|_{\beta }.
\end{equation*}
To prove (\ref{maf5'}), apply (\ref{ff30}) and the interpolation theorem.
Let $\beta \in (0,1]$. Consider an operator on $C^{\beta }$ defined by
T^{\varepsilon }(f)=f^{\varepsilon }$. According to (\ref{ff30}),
T^{\varepsilon }:C^{\beta }(\mathbf{R}^{d})\rightarrow C^{k}(\mathbf{R
^{d}),k=1,2,$ is bounded
\begin{equation*}
|T^{\varepsilon }(f)|_{k}\leq C\varepsilon ^{-k+\beta }|f|_{\beta
},k=1,2,f\in C^{\beta }(\mathbf{R}^{d}).
\end{equation*
By Theorem 6.4.5 in \cite{BeL76}, $T^{\varepsilon }:C^{\beta }(\mathbf{R
^{d})\rightarrow C^{\alpha }(\mathbf{R}^{d})$ is bounded an
\begin{equation*}
|T^{\varepsilon }(f)|_{\alpha }\leq C\varepsilon ^{(-1+\beta )(2-\alpha
)}\varepsilon ^{(-2+\beta )(\alpha -1)}|f|_{\beta }=C\varepsilon ^{-\alpha
+\beta }|f|_{\beta },f\in C^{\beta }(\mathbf{R}^{d}).
\end{equation*
If $\beta \in (1,\alpha ],$ $\partial ^{\alpha -1}\nabla f^{\varepsilon
}=\partial ^{\alpha -1}\left( \nabla f\right) ^{\varepsilon }$ and by (\re
{maf5})
\begin{equation*}
|\partial ^{\alpha -1}\nabla f^{\varepsilon }(x)|=|\partial ^{\alpha
-1}\left( \nabla f\right) ^{\varepsilon }(x)|\leq C\kappa (\varepsilon
,\alpha -1,\beta -1)|\nabla f|_{\beta -1},
\end{equation*
and (\ref{maf6}) follows.
Let $\beta \in (0,1],\alpha \in (1,2),\beta <\alpha -1.$ The
\begin{eqnarray*}
\nabla f^{\varepsilon } &=&\varepsilon ^{-1}\varepsilon ^{-d}\int \nabla w
\frac{y}{\varepsilon })f(x-y)dy \\
&=&\varepsilon ^{-1}\varepsilon ^{-d}\int \nabla w(\frac{y}{\varepsilon
)f(x-y)dy.
\end{eqnarray*
an
\begin{equation*}
\partial ^{\alpha -1}\nabla f^{\varepsilon }=\varepsilon ^{-1-(\alpha
-1)}\int \partial ^{\alpha -1}(\nabla w)(y)f(x-\varepsilon y)dy
\end{equation*
and we derive as in part (i) (using Lemma \ref{rem2}) tha
\begin{equation*}
|\partial ^{\alpha -1}\nabla f^{\varepsilon }(x)|\leq C\varepsilon ^{-\alpha
+\beta }|f|_{\beta }.
\end{equation*
If $\beta \in (0,1],\alpha \in (1,2),\beta >\alpha -1$, the
\begin{eqnarray*}
\partial ^{\alpha -1}\nabla f^{\varepsilon } &=&\varepsilon ^{-1}\int \nabla
w(y)(\partial ^{\alpha -1})f(x-\varepsilon y)dy \\
&=&\varepsilon ^{-1}\int \nabla w(y)[(\partial ^{\alpha -1})f(x-\varepsilon
y)-(\partial ^{\alpha -1})f(x)]dy
\end{eqnarray*
an
\begin{equation*}
|\partial ^{\alpha -1}\nabla f^{\varepsilon }|\leq C\varepsilon ^{-1+\beta
-\alpha +1}|\partial ^{\alpha -1}f|_{\beta -\alpha +1}\leq C\varepsilon
^{-\alpha +\beta }|f|_{\beta }.
\end{equation*}
The statement is proved.
\end{proof}
\begin{corollary}
\label{coro3}Let $\alpha \in (0,2],\beta \leq \alpha $. Assume $\varepsilon
\in (0,1)$, $a(x)$ is bounded, and
\begin{equation*}
\int \big(|y|^{\alpha }\wedge 1\big)\pi (dy)<\infty .
\end{equation*
Then there exists a constant $C$ such that for all $z,x\in \mathbf{R}^{d}$,
f\in C^{\beta }(\mathbf{R}^{d}),
\begin{equation*}
|B_{z}f^{\varepsilon }(x)|\leq C\kappa (\varepsilon ,\alpha ,\beta
)|f|_{\beta }.
\end{equation*}
\end{corollary}
\begin{proof}
If $\beta \leq \alpha <1$, by Lemmas \ref{r1} and \ref{lnew2}
\begin{equation*}
f^{\varepsilon }(x+y)-f^{\varepsilon }(x)=\int k^{(\alpha )}(y,y^{\prime
})\partial ^{\alpha }f^{\varepsilon }(x-y^{\prime })dy^{\prime },
\end{equation*
and by (\ref{maf5}),
\begin{equation*}
|f^{\varepsilon }(x+y)-f^{\varepsilon }(x)|\leq C\kappa (\varepsilon ,\alpha
,\beta )|f|_{\beta }(|y|^{\alpha }\wedge 1),x,y\in \mathbf{R}^{d}.
\end{equation*}
So
\begin{eqnarray*}
|f^{\varepsilon }(x+G(x)y)-f^{\varepsilon }(x)| &\leq &C\kappa (\varepsilon
,\alpha ,\beta )|f|_{\beta }(|G(x)y|^{\alpha }\wedge 1) \\
&\leq &C\kappa (\varepsilon ,\alpha ,\beta )|f|_{\beta }\big[\mathbf{1
_{\left\{ |y|\leq 1\right\} }|G(x)y|^{\alpha } \\
&&\quad +\mathbf{1}_{\left\{ |y|>1\right\} }(|G(x)y|^{\alpha }\wedge 1)\big].
\end{eqnarray*}
If $\beta \leq \alpha =1$, by Lemma \ref{lnew2}, (\ref{ff30})
\begin{eqnarray*}
|f^{\varepsilon }(x+y)-f^{\varepsilon }(x)| &\leq &C\sup_{x}[f(x)|+|\nabla
f^{\varepsilon }(x)|](|y|\wedge 1) \\
&\leq &C\kappa (\varepsilon ,1,\beta )|f|_{\beta }(|y|\wedge 1),x,y\in
\mathbf{R}^{d}
\end{eqnarray*
an
\begin{eqnarray*}
|f^{\varepsilon }(x+G(x)y)-f^{\varepsilon }(x)| &\leq &C\kappa (\varepsilon
,1,\beta )|f|_{\beta }(|G(x)y|\wedge 1) \\
&\leq &C\kappa (\varepsilon ,1,\beta )|f|_{\beta }[\mathbf{1}_{|y|\leq
1}|G(x)y|+\mathbf{1}_{|y|>1}(|G(x)y|\wedge 1)].
\end{eqnarray*}
Assume $\alpha \in (1,2],\beta \leq \alpha .$ Then for $x,y^{\prime }\in
\mathbf{R}^{d},
\begin{equation}
f^{\varepsilon }(x+y^{\prime })-f^{\varepsilon }(x)-(\nabla f^{\varepsilon
}(x),y^{\prime })=\int_{0}^{1}\big(\nabla f^{\varepsilon }(x+sy^{\prime
})-\nabla f^{\varepsilon }(x),y^{\prime }\big)ds. \label{ff34}
\end{equation
We will show that for all $x,y^{\prime }\in \mathbf{R}^{d}
\begin{eqnarray}
&&|f^{\varepsilon }(x+y^{\prime })-f^{\varepsilon }(x)-(\nabla
f^{\varepsilon }(x),y^{\prime })| \label{ff350} \\
&\leq &C|y^{\prime }|^{\alpha }\kappa (\varepsilon ,\alpha ,\beta
)|f|_{\beta } \notag
\end{eqnarray
If $\beta \in (0,\alpha ],\alpha \in (1,2),\beta \neq \alpha -1$, then we
have (\ref{ff350}) by Lemmas \ref{r1}, \ref{lnew2}$.$
If $\beta \in (0,\alpha ],\alpha \in (1,2),\beta =\alpha -1$,then for any
x,y^{\prime }\in \mathbf{R}^{d}
\begin{eqnarray*}
&&\nabla f^{\varepsilon }(x+y^{\prime })-\nabla f^{\varepsilon }(x) \\
&=&\varepsilon ^{-1}\varepsilon ^{-d}\int \nabla w(y/\varepsilon
)[f(x+y^{\prime }-y)-f(x-y)]dy
\end{eqnarray*
an
\begin{equation*}
|\nabla f^{\varepsilon }(x+y^{\prime })-\nabla f^{\varepsilon }(x)|\leq
C\varepsilon ^{-1}|y^{\prime }|^{\beta }|f|_{\beta }=C\varepsilon ^{-\alpha
+\beta }|f|_{\beta }|y^{\prime }|^{\alpha -1}.
\end{equation*
So, (\ref{ff350}) holds in this case as well. If $\beta \leq \alpha =2$,
then by Lemma \ref{lnew2}
\begin{eqnarray*}
|\nabla f^{\varepsilon }(x+y^{\prime })-\nabla f^{\varepsilon }(x)| &\leq
&\sup_{x}|D^{2}f^{\varepsilon }(x)|~|y^{\prime }| \\
&\leq &C\kappa (\varepsilon ,2,\beta )|f|_{\beta }|y^{\prime }|
\end{eqnarray*
and (\ref{ff350}) follows. Hence, for $|y|\leq 1,$ by (\ref{ff350})
\begin{equation*}
|f^{\varepsilon }(x+G(x)y)-f^{\varepsilon }(x)-(\nabla f^{\varepsilon
}(x),G(x)y)|\leq C\kappa (\varepsilon ,\alpha ,\beta )|G(x)y|^{\alpha
}|f|_{\beta }.
\end{equation*
Also, for $|y|>1,
\begin{equation*}
|f^{\varepsilon }(x+G(x)y)-f^{\varepsilon }(x)|\leq 2|f|_{\beta }.
\end{equation*}
Therefore, the statement follows by the assumptions and Lemma \ref{lnew2}.
\end{proof}
\subsection{Proof of Lemma \protect\ref{lem:expect}}
If $\beta \leq \alpha \,,$ define $f^{\varepsilon }$ by (\ref{maf7}) for
\varepsilon \in (0,1)$ and apply It\^{o}'s formula (see Remark \ref{lrenew1
): for $s\in \lbrack 0,T]$,
\begin{equation*}
\mathbf{E}[f^{\varepsilon }(Y_{s})-f^{\varepsilon }(Y_{\tau _{i_{s}}})
\mathcal{F}_{\tau _{i_{s}}}]=\mathbf{E}\big[\int_{\tau _{i_{s}}}^{s}\big
A_{Y_{\tau _{i_{s}}}}f^{\varepsilon }(Y_{r})+B_{Y_{\tau
_{i_{s}}}}f^{\varepsilon }(Y_{r})\big)dr\big|\mathcal{F}_{\tau _{i_{s}}}\big
.
\end{equation*
Hence, by Lemma \ref{lnew2} and Corollary \ref{coro3}, for $\varepsilon \in
(0,1)$
\begin{eqnarray*}
|\mathbf{E}[f(Y_{s})-f(Y_{\tau _{i_{s}}})|\mathcal{F}_{\tau _{i_{s}}}]|
&\leq &|\mathbf{E}[(f-f^{\varepsilon })(Y_{s})-(f-f^{\varepsilon })(Y_{\tau
_{i_{s}}})|\mathcal{F}_{\tau _{i_{s}}}]| \\
&&+|\mathbf{E}[f^{\varepsilon }(Y_{s})-f^{\varepsilon }(Y_{\tau _{i_{s}}})
\mathcal{F}_{\tau _{i_{s}}}]| \\
&\leq &CF(\varepsilon ,\delta )|f|_{\beta },
\end{eqnarray*
with a constant $C$ independent of $\varepsilon ,f$ and
\begin{equation*}
F(\varepsilon ,\delta )=\left\{
\begin{array}{cc}
(\varepsilon ^{2}+\delta )(1-\ln \varepsilon ) & \text{if }\alpha =\beta =2,
\\
\varepsilon ^{\beta }+\delta \kappa (\varepsilon ,\alpha ,\beta ) & \text
otherwise.
\end{array
\right.
\end{equation*
Minimizing $F(\varepsilon ,\delta )$ in $\varepsilon \in (0,1)$, we obtai
\begin{equation*}
|\mathbf{E}[f(Y_{s})-f(Y_{\tau _{i_{s}}})|\mathcal{F}_{\tau _{i_{s}}}]|\leq
Cr(\delta ,\alpha ,\beta )|f|_{\beta }.
\end{equation*}
If $\beta >\alpha ,$ apply It\^{o}'s formula directly (see Remark \re
{lrenew1})
\begin{equation*}
\mathbf{E}[f(Y_{s})-f(Y_{\tau _{i_{s}}})|\mathcal{F}_{\tau _{i_{s}}}]
\mathbf{E}\big[\int_{\tau _{i_{s}}}^{s}\big(A_{Y_{\tau _{i_{s}}}}^{(\alpha
)}f(Y_{r})+B_{Y_{\tau _{i_{s}}}}^{(\alpha )}f(Y_{r})\big)dr \big|\mathcal{F
_{\tau _{i_{s}}}\big].
\end{equation*
Hence, by Lemmas \ref{prop2} and \ref{lnew2},
\begin{equation*}
|\mathbf{E}[f(Y_{s})-f(Y_{\tau _{i_{s}}})|\mathcal{F}_{\tau _{i_{s}}}]|\leq
C\delta |f|_{\beta }.
\end{equation*
The statement of Lemma \ref{lem:expect} follows.
\subsection{Proof of Theorem \protect\ref{thm:main}}
Let $v\in C^{\alpha +\beta }(H)$ be the unique solution to (\ref{maf8}) (see
Corollary \ref{lcornew1}). By It\^{o}'s formula (see Remark \ref{lrenew1})
and (\ref{maf8})
\begin{eqnarray*}
\mathbf{E}[v(0,X_{0})] &=&\mathbf{E}[v(T,X_{T})] - \mathbf{E}\big
\int_{0}^{T}\big(\partial
_{t}v(s,X_{s})+A_{X_{s}}v(s,X_{s})+B_{X_{s}}v(s,X_{s})\big)ds\big] \\
&=&\mathbf{E}\big[g(X_{T})-\int_{0}^{T}f(X_{s})ds\big]
\end{eqnarray*
and
\begin{equation}
\mathbf{E}[v(0,X_{0})] = \mathbf{E}[v(0,Y_{0})]. \label{eqn:expect_terminal}
\end{equation}
By Proposition \ref{b1}, Corollary~\ref{lcornew1}, Remark \ref{rlast}, and
Lemma \ref{prop2},
\begin{eqnarray}
|A_{z}v(s,\cdot )|_{\beta }+|B_{z}v(s,\cdot )|_{\beta } &\leq &C|v|_{\alpha
+\beta }\leq C|g|_{\alpha +\beta }, \label{maf9} \\
|\partial _{t}v(s,\cdot )|_{\beta } &\leq &C|g|_{\alpha +\beta },s\in
\lbrack 0,T]. \notag
\end{eqnarray}
Then, by It\^{o}'s formula (Remark \ref{lrenew1}) and Corollary \re
{lcornew1}, with (\ref{eqn:expect_terminal}) and (\ref{maf9}), it follows
that
\begin{eqnarray*}
&&\mathbf{E}[g(Y_{T})] - \mathbf{E}[g(X_{T})] - \mathbf{E}\big
\int_{0}^{T}f(Y_{\tau _{i_{s}}})ds\big] + \mathbf{E}\big
\int_{0}^{T}f(X_{s})ds\big] \\
&=&\mathbf{E}[v(T,Y_{T})] - \mathbf{E}[v(0,Y_{0})] - \mathbf{E}\big
\int_{0}^{T}f(Y_{_{\tau _{i_{s}}}})ds\big] + \mathbf{E}\big
\int_{0}^{T}f(X_{s})ds\big] \\
&=&\mathbf{E}\Big[\int_{0}^{T}\Big\{\big[\partial _{t}v(s,Y_{s})-\partial
_{t}v(s,Y_{\tau _{i_{s}}})\big] \\
&&+\big[A_{Y_{\tau _{i_{s}}}}v(s,Y_{s})-A_{Y_{\tau _{i_{s}}}}v(s,Y_{\tau
_{i_{s}}})\big] \\
&&+\big[B_{Y_{\tau _{i_{s}}}}v(s,Y_{s})-B_{Y_{\tau _{i_{s}}}}v(s,Y_{\tau
_{i_{s}}})\big]\Big\}ds\Big].
\end{eqnarray*}
Hence, by (\ref{maf9}) and Lemma~\ref{lem:expect}, there exists a constant
C $ independent of $g$ such that
\begin{equation*}
|\mathbf{E}[g(Y_{T})]-\mathbf{E}[g(X_{T})]|\leq Cr(\delta ,\alpha ,\beta
)|g|_{\alpha +\beta }.
\end{equation*
The statement of Theorem \ref{thm:main} follows.
\section{Conclusion}
The paper studies weak Euler approximation of SDEs driven by L\'{e}vy
processes. The dependence of the rate of convergence on the regularity of
coefficients and driving processes is investigated under assumption of
\beta $-H\"{o}lder continuity of the coefficients. It is assumed that the
main term of the SDE is driven by a spherically-symmetric $\alpha $-stable
process and the tail of the L\'{e}vy measure of the lower order term has a
\mu $-order finite moment ($\mu \in (0,3)).$ The resulting rate depends on
\beta ,\alpha $ and $\mu $. In order to estimate the rate of convergence,
the existence of a unique solution to the corresponding backward Kolmogorov
equation in H\"{o}lder space is first proved. The assumptions on the
regularity of coefficients and test functions are different than those in
the existing literature.
One possible improvement could be to consider the asymptotics of the tails
at infinity instead of the tail moment $\mu $. Besides this, the stochastic
differential equations considered so far are associated with nondegenerate
\'{e}vy operators. A further step could be to study the case with degenerate
operators. That is, consider equation (\ref{one}) without assuming $\det
b\neq 0$. For example, let $\alpha \in \lbrack 1,2]$ and $\beta \in (\alpha
,2\alpha ].$ Assume the coefficients are in $C^{\beta }$ and
\begin{equation*}
\int_{|y|\leq 1}|y|^{\alpha }d\pi +\int_{|y|>1}|y|^{2\alpha }d\pi <\infty .
\end{equation*
In this case, a plausible convergence rate is $r(\delta ,\alpha ,\beta
)=\delta ^{\frac{\beta }{\alpha }-1}$. With $\det b=0$ being allowed, a
higher regularity of coefficients and lighter tails of $\pi $ would be
required.
|
\section{Introduction}
{\bf Notations and definitions:} Throughout this paper, $\F_q$ will be a finite field of characteristic $p$ where $q=p^r$ for some $r\in \N^*$.
When $F_1,\ldots,F_n\in k[X_1,\ldots,X_n]$ ($k$ a field), then $F:=(F_1,\ldots,F_n)$ is a polynomial endomorphism over $k$.
If there exists a polynomial endomorphism $G$ such that $F(G)=G(F)=(X_1,\ldots,X_n)$, then $F$ is a polynomial automorphism (which is stronger than stating that $F$ induces a bijection on $k^n$). We will write $X$ for $X_1,\ldots,X_n$. The
polynomial automorphisms in $n$ variables over $k$ form a group, denoted $\GA_n(k)$ (compare the notation $\GL_n(k)$), while the set of polynomial endomorphisms is denoted by $\ME_n(k)$ (the monoid of endomorphisms).
If $F\in \GA_n(k)$ such that $\deg(F_i)=1$ for all $1\leq i \leq n$, then $F$ is called affine. The affine automorphisms form a subgroup of $\GA_n(k)$ denoted by $\Aff_n(k)$. In case $F\in \GA_n(k)$ such that $F_i\in k[X_i,\ldots,X_n]$ for each $1\leq i\leq n$, then $F$ is called triangular, or Jonqui\'ere. The triangular automorphisms form a subgroup of $\GA_n(k)$, denoted by $\textup{J}_n(k)$.
The subgroup of $\GA_n(k)$ generated by $\Aff_n(k)$ and $\textup{J}_n(k)$ is called the {\bf tame automorphism group}, denoted by $\TA_n(k)$. By $\deg(F)$ we will denote the maximum of $\deg(F_i)$. The set of polynomial maps of degree $d$ or less we denote by $\ME^d_n(k)$, and the endomorphisms whose affine part is the identity we denote by
$\overline{\ME}_n(k)$. The following notations are now natural: $\overline{\ME}^d_n(k):=\overline{\ME}_n(k)\cap \ME_n^d(k)$, $\GA_n^d(k):=\ME_n^d(k)\cap \GA_n(k)$, $\overline{\GA}_n(k):=\overline{\ME}_n(k)\cap \GA_n(k)$, and
$\overline{\GA}_n^d(k):=\overline{\GA}_n(k)\cap \GA_n^d(k)$.
If $F,G\in \ME_n(k)$, then $F$ and $G$ are called {\bf equivalent (tamely equivalent)} if there exist $N,M\in \GA_n(k)$ ($N,M\in \TA_n(k)$ such that $NFM=G$. If $F\in \ME_n(k)$ then we say that $(F,X_{n+1},\ldots,X_{n+m})\in \MA_{n+m}(k)$ is a stabilisation of $F$. We hence introduce the terms {\bf``stably equivalent'' and ``stably tamely equivalent''} meaning that a stabilisation of $F$ and $G$ are equivalent or tamely equivalent.\\
The automorphism group $\GA_n(k)$ is one of the basic objects in (affine) algebraic geometry, and the understanding of its structure a highly-sought after question. If $n=1$ then $\GA_1(k)=\Aff_1(k)$, and if $n=2$ then one has the Jung-van der Kulk theorem \cite{Jung, Kulk}, stating among others that $\GA_2(k)=\TA_2(k)$. However, in dimension 3 the structure of $\GA_3(k)$ is completely dark.
The only strong result is in fact a {\em negative} result by Umirbaev and Shestakov \cite{SU1,SU2}, stating that if $\kar{k}=0$, then
$\TA_3(k)\not = \GA_3(k)$.
It might be that all types of automorphisms known in $\GA_3(k)$ have already surfaced, but the possibility exists that there are some
strange automorphisms that have eluded common knowledge so far. But, in the case $k=\F_q$, we have an opportunity: one could simply check the finite set of endomorphisms up to a certain degree $d$, i.e. $\MA_n^d(\F_q)$, and determine which ones are automorphisms. Any ``new'' type of automorphisms {\em have} to surface in this way.
Unfortunately, the computations rapidly become unfeasible if the degree $d$, the number of variables $n$, or the size of the finite field $\F_q$, are too large. We didn't find any significant shortcuts except the ones mentioned in section \ref{S2}. In the end, for us scanning through lists of $2^{30}=8^{10}$
endomorphisms was feasible, but $3^{20}=9^{10}$ was barely out of reach.
In the case of $k=\F_q$, another interesting class that surfaces are the (what we define as) {\bf mock automorphisms} of $\F_q$:
endomorphisms which induce bijections of $\F_q^n$, and whose determinant of the Jacobian is a nonzero constant. Such maps are interesting for example for cryptography (being ``multivariate permutation polynomials'').
In this article, we do the (for us) feasible computations, and analyze the resulting lists. In particular, we compute (some of) the (mock) automorphisms for $n=3$, $d\leq 3$, and $q\leq 5$.
\section{Generalities on polynomial automorphisms}
\label{S2}
The following lemma explains why we only study polynomial maps having affine part identity:
\begin{lemma}
Let $F\in \GA_n^d(k)$. Then there exists a unique $\alpha, \beta \in \Aff_n(k)$ and $ F', F'' \in \overline{\GA}^d_n(k)$ such that
\[ F=\alpha F' = F''\beta. \]
\end{lemma}
\begin{proof}
For the first equality, take $\alpha$ to be the affine part of $F$, and define $F':=\alpha^{-1}F$.
For the second, do the first equality for $F^{-1}$, i.e. $F^{-1}=\gamma G$ for some $\gamma\in \Aff_n(k), G\in \overline{\GA}_n(k)$.
Then $F=G^{-1}\gamma^{-1}$ i.e. take $\beta=\gamma^{-1}$, $F'':=G^{-1}$. The fact that $F''\in \GA_n^d(k)$ is easy to check by comparing the highest degrees of $F''$ and $F$.
\end{proof}
A useful criterion is that if $F$ is invertible, then $\det(Jac(F))\in k^*$. The converse is a notorious problem in characteristic zero:\\
\noindent
{\bf Jacobian Conjecture:} (Short JC) If $\kar(k)=0$, $F\in \ME_n(k)$, and $\det(\Jac(F))\in k^*$, then $F$ is an automorphism.\\
The JC in $\kar(k)=p$ is not true in general, as already in one variable, $F(X_1):=X_1-X_1^p$ has Jacobian $1-pX_1^{p-1}=1$, but
$F(0)=F(1)$ and so $F$ is not a bijection. However, the following two (well-known) lemma's show that the Jacobian conjecture is true for the special case where $degree(F)=2$ and $
char(k)\geq 3$.
\begin{lemma}
Let $F:k^n\rightarrow k^n$ be a polynomial endomorphism of degree $2$ with $det(Jac(F))$ nowhere zero.
Assume that $\kar(k)=p\not= 2$, then $F$ is injective. In particular, if $k$ is a finite field, then $F$ is bijective.
\end{lemma}
\begin{proof}
Suppose $F$ is not injective, then there exist $a,b\in k^n$ such that $F(a)=F(b)$. We may assume that $a=0=F(a)$, as we may replace $F$ by $F(a-X)-F(a)$ if necessary. Now consider $F(tX)=F_0+tF_1(X)+t^2F_2(X)$, where $F_i(X)$ is the homogeneous part of $F(X)$ of degree $i$ and $t$ is a new variable.
Then on the one hand $\frac{d}{dt}F(tX) = F_1(X)+2tF_2(X)$, on the other hand $\frac{d}{dt}F(tX)=Jac(F)\arrowvert_{tX}X$.
Now if we substitute $t=\frac{1}{2}$, then on the one hand we get $\frac{d}{dt}F(tX)\arrowvert_{t=\frac{1}{2}}= F_1(X)+F_2(X)=F(X)$, (by assumption $F(0)=0$, so $F_0=0$). And on the other hand we get $\frac{d}{dt}F(tX)\arrowvert_{t=\frac{1}{2}}=Jac(F)\arrowvert_{\frac{1}{2}X}X$.
But this means that $0=F(a)=Jac(F)\arrowvert_{\frac{1}{2}a}a$ but since $det(Jac(F)\arrowvert_{\frac{1}{2}a})\not=0$ it follows that $Jac(F)\arrowvert_{\frac{1}{2}a}$ is invertible, and this implies that $a=0$, which contradicts our assumption that $a\not=0$.
So $F$ is injective. If $k$ is a finite field then $k^n$ is a finite set, so injective implies bijective.
\end{proof}
\begin{corollary}\label{Jac1aut}
Let $F:k^n\rightarrow k^n$ be a polynomial endomorphism of degree $2$ with $det(Jac(F))=1$.
Assume that $char(k)\not= 2$, then $F$ is an automorphism.
\end{corollary}
\begin{proof}
Let $K$ be the algebraic closure of $k$. And consider $F$ as a polynomial endomorphism of $K^n$. For every finite extension $L$ of $k$, we have $F:L^n\rightarrow L^n$ is a bijection, by the above lemma. Hence, $F:K^n\rightarrow K^n$ is a bijection (as $K$ is the infinite union of all finite extensions of $k$). But $K$ is algebraically closed so a bijection of $K^n$ is a polynomial automorphism, so it has an inverse $F^{-1}$. Now Lemma 1.1.8 in \cite{EFC} states that $F^{-1}$ has coefficients in $k$, which means that $F^{-1}$ is defined over $k$, which means that $F$ is a polynomial automorphism over $k$.
\end{proof}
One remark on the previous result about the difference between $det(Jac(F))=1$ and $det(Jac(F))$ is nowhere zero.
If $det(Jac(F))$ is nowhere zero over $k$ this does not imply that $det(Jac(F))$ over $K$ is nowhere zero, consider the following example (see the warning after Corollary 1.1.35 in \cite{EFC}):
\begin{example}
Let $F=(x,y+axz,z+bxy)\in k[x,y,z]$ with $k$ a finite field of characteristic $p$ and $a,b\not= 0\ \in k$, such that $ab$ is not a square. Then $det(Jac(F))=1-abx^2$ is nowhere zero, but obviously $F$ not invertible.
\end{example}
The following result is on subsets of groups that are invariant under a subgroup.
\begin{lemma}\label{FS}
Let $G$ be a group, $H$ a finite subgroup of $G$ and $V$ a finite subset of $G$ such that $HV\subseteq V$, then $\#H|\#V$.
\end{lemma}
\begin{proof}
Since $G$ acts transitively on $G$, $H$ acts transitively on $V$. Thus, for every $v\in V$, $Hv$ is an orbit set-isomorphic to $H$.
Also, $HV=V$ consists of disjoint orbits of the form $Hv$, so $\# H | \# V$.
\end{proof}
In this article we also consider so-called {\em locally finite polynomial automorphisms}. A motivation for studying these automorphisms is that they might generate the automorphism group in a natural way (see \cite{Fu-Mau} for a more elaborate motivation of studying these maps). The reason that we make computations and classifications on them in this article is to have some examples on hand to work with in the future, as there can be rather complicated locally finite polynomial automorphisms.
\begin{definition} Let $F\in \MA_n(k)$. Then $F$ is called locally finite (short LFPE) if $\deg(F^n)$ is bounded, or equivalently, there exists $n\in \N$ and $a_i\in k$ such that $F^n+a_{n-1}F^{n-1}+\ldots + a_1F+a_0 I=0$. We say that $T^n+a_{n-1}T^{n-1}+\ldots + a_1T+a_0$ is a vanishing polynomial for $F$. In \cite{Fu-Mau} theorem 1.1 it is proven that these vanishing polynomials form an ideal of $k[T]$, and that there exists a minimum polynomial $\m_F(T)$.
\end{definition}
When trying to classify LFPEs and their minimal polynomials (i.e using computer calculations) one can use the following lemmas to reduce computations:
\begin{lemma}\label{LFPEconjugacy}
Let $F\in \GA_n(k)$ and $L \in GL_n(k)$ then $F$ is locally finite iff $L^{-1}FL$ is locally finite. Furthermore if $\m(T)\in k[T]$ is the minimum polynomial for $F$, then $\m(T)$ is also the minimum polynomial of $L^{-1}FL$.
\end{lemma}
\begin{proof}
Suppose $F$ is locally finite with minimum polynomial $\m(T)=m_0+m_1T+\cdots +m_dT^d$, so $m_0I+m_1F+m_2F^2+\cdots + m_dF^d=0$, where $I$ is the identity mam, and $F^d$ is the commosition of $d$ $F$'s.
Now $0=L^{-1}(m_0I+m_1F+m_2F^2+\cdots + m_dF^d)L=m_0L^{-1}IL+m_1L^{-1}FL+m_2L^{-1}F^2L+\cdots + m_dL^{-1}F^dL=
m_0I+m_1L^{-1}FL+m_2(L^{-1}FL)^2+\cdots + m_d(L^{-1}FL)^d=\m(L^{-1}FL)$.
This shows that if $F$ is locally finite with minimum polynomial $\m(T)$, then so is $L^{-1}FL$.
\end{proof}
When classifying LFPEs, one cannot simply restrict to $\overline{\GA}_n(k)$, as it is very well possible that $F\in \overline{\GA}_n(k)$
is not an LFPE, but $\alpha F$ is where $\alpha\in \Aff_n(k)$. However, we can restrict to classes of affine parts under linear maps, by the following lemma:
\begin{lemma}\label{CompLocFin}
Let $\alpha\in Aff_n(k)$ and $F\in \overline{GA}^d_n(k)$. Suppose that $\beta=L^{-1}\alpha L$ , where $L\in GL_n(k)$, and that $\beta F$ is locally finite. Now there exists an automorphism $G\in \overline{GA}^d_n(k)$, such that $\beta F=L^{-1}\alpha G L$, i.e. $\beta F$ is in the conjugacy class of $\alpha G$. Furthermore, the minimum polynomials of $F$ and $G$ are the same.
\end{lemma}
\begin{proof}
Just take $G=LFL^{-1}$, then $L^{-1}\alpha GL=L^{-1}\alpha LFL^{-1}L=L^{-1}\alpha LF=\beta F$.
\end{proof}
So in order to classify the locally finite automorphisms (up to some degree $d$), it suffices to compute the conjugacy classes of $\Aff_n(k)$ under conjugacy by $\GL_n(k)$, and compose a representative of each class with all the elements of $\GA_n^d(k)$.
When considering LFPEs over finite fields, we have the additional following lemma:
\begin{lemma} \label{order} Let $F\in \GA_n(\F_q)$ be an LFPE. Then $F$ has finite order (as element of $\GA_n(\F_q)$).
\end{lemma}
\begin{proof} If $F$ is an LFPE, then there exists a minimum polynomial $\m(T)$ generating the ideal of vanishing polynomials for $F$.
There exists some $r\in \N$ such that $\m(T)~|~ T^{q^r}-T$, yielding the result.
\end{proof}
Another concept that surfaces, is the following:
\begin{definition} Let $F=I+H \in \overline{\MA}_n(k)$ where $H=(H_1,\ldots, H_n)$ is the non-linear part. Then $F$ is said to satisfy the {\em dependence criterion} if $(H_1,\ldots, H_n)$ are linearly dependent.
\end{definition}
Notice that $F\in \overline{\MA}_n(k)$ satisfying the dependence criterion is equivalent to being able to apply a linear conjugation to isolate one variable, i.e. $L^{-1}FL=(X_1,X_2+H_2, \ldots, X_n+H_n)$ for some linear map $L$.
\section{Computations on endomorphisms of low degree}
It is clear that $\#\Aff_n(k)|\#\GA_n^d(k)$.
Let $F:k^n\rightarrow k^n$ be an automorphism then we can consider $\alpha$ to be the affine part of $F$, which is obviously invertible and we can then look at $G=\alpha^{-1}F$, where $G$ now has affine part the identity. This means that to compute all automorphisms it suffices to compute all automorphisms having affine part the identity and compose each of them from the left with all affine automorphisms.
This suffices for our computations over $\F_2$ and $\F_3$, but for larger finite fields ($\F_4,\F_5$ and $\F_7$) we will add additionally the dependence criterion.
Finally recall that over $\F_q$ there are $(q^n-1)(q^n-q)\cdots (q^n-q^{n-1})$ linear automorphisms, and that there are $q^n$ times as many affine automorphisms as linear.
For the rest of the article, as we stay in 3 dimensions, we will rename our variables $x,y,z$.
\subsection{The finite field of two elements: $\mathbb{F}_2$}\label{F2}
As mentioned in the previous section to find all polynomial automorphisms it suffices to find all automorphisms having affine part equal to the identity, and there are $(2^3-1)(2^3-2)(2^3-2^2)=168$ linear automorphisms. There are $2^3*168=1344$ affine automorphisms.
\subsubsection{Degree 2 over $\F_2$}
We remind the reader that by a {\em mock automorphism} we mean an endomorphism $E\in \MA_n(k)$ such that $E$ induces a permutation of $k^n$ and $\det(\Jac(E))\in k^*$.
Over $\F_2$, Corollary \ref{Jac1aut} does not hold so there do exist mock automorphisms which are not automorphisms in $\MA_3^2(\F_2)$.
\begin{theorem}\label{T1} If $F\in \MA^2_3(\F_2)$ is a mock automorphism, then $F$ is in one of the following four classes:
\begin{itemize}
\item[1)] The 176 tame automorphisms, equivalent to $(x,y,z)$.
\item[2)] 48 endomorphisms tamely equivalent to $(x^4+x^2+x,y,z)$.
\item[3)] 56 endomorphisms tamely equivalent to $(x^8+x^2+x,y,z)$.
\item[4)] 56 endomorphisms tamely equivalent to $(x^8+x^4+x,y,z)$.
\end{itemize}
In particular, all automorphisms of this type are tame, i.e. $\GA_3^2(\F_2)=\TA_3^2(\F_2)$
Furthermore, the equivalence classes are all distinct, except possibly class (3) and (4) (see conjecture \ref{Qu1}). \\
There are in total $1344\cdot 176=236544$ automorphisms of $\F_2^3$ of degree less or equal to $2$.
\end{theorem}
\begin{proof} The classification is done by computer, see \cite{RoelThesis} chapter 5.
We can show how for example $(x^8+x^4+x, y,z)$ is tamely equivalent to a polynomial endomorphism of degree 2:
\[ \begin{array}{l}
(x+y^2,y+z^2,z)(x^8+x^4+x,y,z)(x,y+x^4+x^2,z+x^2)=\\
(x+y^2,y+x^2+z^2,z+x^2)
\end{array} \]
What is left is to show that the classes (1),(2), and (3)+(4) are different.
Class (1) consists of automorphisms while (2),(3),(4) are not.
Using the below lemma \ref{L1}, the endomorphisms of type (2) are all bijections of $\F_{2^m}^3$ if $3\not|m$, and the endomorphisms of type (3) and (4) are all bijections of $\F_{2^m}^3$ if $7\not|m$.
The last sentence follows since $\#\Aff_3(\F_2)=1344$.
\end{proof}
\begin{lemma}\label{L1}
$x^4+x^2+x$ is a bijection of $\F_{2^r}$ if $3\not |r$, and $x^8+x^4+x$ and $x^8+x^2+x$ are bijection of $\F_{2^r}$ if $7\not | r$.
\end{lemma}
\begin{proof}
Let us do $f(x):=x^8+x^4+x$, the other proofs go similarly. $f$ is a bijection if and only if $f$ is injective if and only if $f(x)=f(y)$ has
only $x=y$ as solutions. $f(x)=f(y)$ if and only if $(x-y)^8+(x-y)^4+(x-y)=0$. $x=y$ is a solution, another solution would be equivalent to finding a zero of $x^7+x^3+1$. Now it is an elementary exercise to see that if $\alpha\in \F_{2^r}$ is a zero of this polynomial, then
$7| r$.
\end{proof}
\begin{question}\label{Qu1}
(1) Are $F=(x^8+x^2+x,y,z)$ and $G=(x^8+x^4+x,y,z)$ (tamely) equivalent? \\
(2) More general: Are $x^8+x^2+x$ and $x^8+x^4+x$ stably (tamely) equivalent?\\
\end{question}
The above question is particular to characteristic $p$, for consider the following:
\begin{lemma} \label{equiv} Let $P,Q\in k[x]$. Assume that $F:=(P(x), y, z)$ is equivalent to $G:=(Q(x),y,z)$.
Then $P'$ and $Q'$ are equivalent, in particular $Q'(ax+b)=cP'$ for some $a,b,c\in k$, $ac\not =0$.
\end{lemma}
\begin{proof} Equivalent means there exist $S, T\in \GA_3(k)$ such that $SF=GT$. Write $\J$ for $\det(\Jac)$.
Now $\J(S)=\lambda, \J(T)=\mu$ for some $\lambda,\mu\in k^*$. Using the chain rule we have
\[ \begin{array}{ll}
&\J(SF)=\J(F)\cdot (\J(S)\circ (F))= \frac{\partial P}{\partial x} \cdot (\lambda\circ (F)) =\lambda \frac{\partial P}{\partial x}\\
=&\J(GT)=\J(T)\cdot (\J(G)\circ (T)) =\mu \cdot ( \frac{\partial Q}{\partial x} \circ T)
\end{array}
\]
so
\[
Q'(T)= \frac{\lambda}{\mu}P'\]
which means that $T=(T_1,T_2,T_3)$ and $T_1=ax+b$ where $a\in k^*, b\in k$, proving the lemma.
\end{proof}
\begin{corollary} Assume $\kar(k)=0$. Let $P,Q\in k[x]$. Assume that $F:=(P(x), y, z)$ is equivalent to $G:=(Q(x),y,z)$.
Then $P$ and $Q$ are equivalent.
\end{corollary}
\begin{proof} Lemma \ref{equiv} shows that $ P' (ax+b)= c Q'$ for some $a,b,c\in k$, $ac\not =0$. In
characteristic zero we can now integrate both sides and get $a^{-1}P(ax+b)=cQ$ proving the corollary.
\end{proof}
Note that in the ``integrate both sides'' part the characteristic zero is used, as $(x+x^2+x^8)'=(x+x^4+x^8)'$ in characteristic $2$.
Note that all the above one-variable polynomials $x^8+x^2+x, x^4+x^2+x$ have a stabilisation which is tamely equivalent to a polynomial endomorphism of degree 2. In this respect, note the following proposition, which is lemma 6.2.5 from \cite{EFC}:
\begin{proposition}
Let $F\in \ME_n(k)$ where $k$ is a field. Then there exists $m\in \N$, and $G,H\in \TA_{n+m}(k)$ such that (writing the stabilisation of $F$ as $\tilde{F}\in \ME_{n+m}(k)$) $G\tilde{F}H$ is of degree 3 or less.
\end{proposition}
\subsubsection{Locally finite in degree 2 over $\F_2$}
We now want to classify the locally finite automorphisms among the 236544 automorphisms over $\F_2$ of degree 2 (or less),
and we want to determine the minimum polynomial of each. Using lemma \ref{LFPEconjugacy}, we may classify up to conjugation by a linear map.
We found $262$ locally finite classes under linear conjugation, with the following minimum polynomials:\\
\ \\
\begin{tabular}{|l|c|c|}
\hline
Minimumpolynomial & $\#$ & $t$ \\ \hline \hline
$F^5+F^4+F+I$ & $16$ & $8$\\ \hline
$F^4+F^3+F^2+I$ & $8$ & $7$\\ \hline
$F^4+F^3+F+I$ & $26$ & $6$\\ \hline
$F^4+I$ & $12$ & $4$\\ \hline
$F^4+F^2+F+I$ & $8$ & $7$\\ \hline
$F^3+F^2+F+I$ & $139$ & $4$\\ \hline
$F^3+F^2+I$ & $2$ & $7$\\ \hline
$F^3+F+I$ & $2$ & $7$\\ \hline
$F^3+I$ & $14$ & $3$\\ \hline
$F^2+I$ & $34$ & $2$\\ \hline
$F+I$ & $1$ & $1$\\ \hline
\end{tabular}\\
\ \\
In the above tabular, $\#$ denotes the number of {\em conjugacy classes} (i.e. not elements) having this minimum polynomial, while $t$ denotes the order of the automorphism (see lemma \ref{order}). Furthermore, observe that $\#$ displayed is the number of conjugacy classes that satisfy this relation, not the total number of automorphisms.
\subsubsection{Degree 3 over $\F_2$}
We only comsidered the endomorphisms of the form $F=I+H$, where $H$ is homogeneous of degree $3$. (The automorphisms of degree 3 or less in general was just out of reach.)
The below tabular describes the set of $F\in \MA^3_3(\F_2)$ having the following criteria:
\begin{itemize}
\item $F$ is a mock automorphism,
\item $F=I+H$, $H$ homogeneous of degree 3.
\end{itemize}
We found
$1520$ endomorphisms satisfying the above requirements. The tabular lists them in 20 classes up to conjugation by linear maps:\\
\ \\
\begin{tabular}{|l|l|l|l|c|}
\hline
& Representant & Bijection over & $\#$\\ \hline \hline
{\bf 1.}& $\mathbf{(x,y,z)}$ \\ \hline
1a. &$(x,y,z)$ &all & 1\\ \hline
1b. &$(x,y,z+x^2y+xy^2)$ &all & 7 \\ \hline
1c. &$(x,y,z+x^3+x^2y+y^3)$ &all & 14 \\ \hline
1d. &$(x,y+x^3,z+x^3)$ &all & 21 \\ \hline
1e. &$(x,y,z+x^3+x^2y+xy^2)$ & all & 21\\ \hline
1f. &$(x,y,z+x^2y)$ & all & 42\\ \hline
1g. &$(x,y+x^3,z+xy^2)$ &all & 42\\ \hline
1h. &$(x,y+x^3,z+x^2y+xy^2)$ & all & 42 \\ \hline
1i. &$(x,y + z^3,z+x^2y)$ &all & 42 \\ \hline
1j. &$(x,y+x^3,z+ x^2y + y^3)$ & all & 84\\ \hline
1k. &$(x,y+x^3,z+y^3)$ & all & 84\\ \hline
{\bf 2.} & $\mathbf{(x,y, z+x^3z^4+xz^2)}$\\\hline
2&$(x,y+x^3+xz^2,z+xy^2+xz^2)$ & $\F_2,\F_4,\F_{16},\F_{32}$ & 56\\ \hline
{\bf 3.}&$\mathbf{ (x,y,z+x^3z^2+x^3z^4)}$ \\ \hline
3a. &$(x,y+ xz^2,z+x^2y + xy^2)$ & $\F_2,\F_4$ & 84 \\ \hline
3b. &$(x,y+xz^2,z+x^3+x^2y+xy^2)$ & $\F_2,\F_4$ & 84\\ \hline
{\bf 4.}& $\mathbf{(x,y,z+xz^2+xz^6)}$ \\ \hline
4a. &$(x,y+x^3+z^3,z+x^3+xy^2+xz^2)$ & $\F_2$ & 168\\ \hline
4b. &$(x,y+z^3,z+xy^2+xz^2)$ & $\F_2$& 168\\ \hline
{\bf 5.}&$\mathbf{ (x,y,z+x^3z^2+xy^2z^4+x^2yz^4+x^3z^6)}$ \\ \hline
5a. &$(x,y+xz^2,z+xy^2+y^3)$ & $\F_2$ & 168\\ \hline
5b. &$(x,y+xz^2,z+x^3+x^2y+y^3)$ & $\F_2$ & 168\\ \hline
{\bf 6.}& $\mathbf{(x,y,z+x^3z^2+xy^2z^2+x^2yz^4+x^3z^6)}$ \\ \hline
6. &$(x,y+xy^2+xz^2,z+x^3+x^2y)$ & $\F_2$ & 168\\ \hline
{\bf 7.}& $\mathbf{(x+y^2z,y+x^2z+y^2z,z+x^3+xy^2+y^3)}$ \\ \hline
7. &$(x+y^2z,y+x^2z+y^2z,z+x^3+xy^2+y^3)$ & $\F_2$ & 56\\ \hline
\end{tabular}\\
\ \\
The first column gives a representant up to linear conjugation, and the bold fonted one gives a representant under tamely equivalence for the classes listed beneath it.
The second column lists for which field extensions (from $\F_{2^r}$ where $1\leq r\leq 5$) the map is also a bijection of $\F_{2^r}^3$
Class 1 are the 400 automorphisms, all of them are tame and sastisfy the dependence conjecture. All classes
are tamely equivalent to a map of the form $(x,y,P(x,y,z))$, except the last class 7 - these maps do not satisfy the Dependence Criterium, which makes them very interesting!
The above tabular might make one think that any mock automorphism in $\MA_3(\F_2)$ of the form $F= (x,y+H_2, z+H_3)$ where $H_2,H_3$ are homgeneous of the same degree, then one can tamely change the map into one of the form $(x,y,z+K)$, but the below conjecture might give a counterexample:
\begin{conjecture} Let $F=(x, y+y^8z^2+y^2z^8, z+y^6z^4+y^4z^6)$. Then $F$ is not tamely equivalent to
a map of the form $(x,y,z+K)$ .
\end{conjecture}
Due to our lack of knowledge of the automorphism group $\TA_3(\F_2)$, this conjecture is a hard one unless one finds a good invariant
of maps of the form $(x,y, z+K)$.
\subsection{The finite field of three elements: $\mathbb{F}_3$}\label{F3}
\subsubsection{Degree 2 over $\F_3$}
Over $\F_3$, there are $(27-1)(27-3)(27-9)=11232$ linear automorphisms and $27*11232=303264$ affine automorphisms.\\
From corollary \ref{Jac1aut} it follows that if $\det(\Jac(F))=1$ and $\deg(F)\leq 2$, then $F$ is an automorphism - so we will not encounter any mock automorphisms which aren't an automorphism in this class.
There are $2835$ automorphisms of degree less or equal to $2$ having affine part identity, so there are $2835\cdot 303264=$
automorphisms of degree $2$ or less. They all turned out to be tame.
\subsubsection{Locally finite}\label{locFin32}
We computed all conjugacy classes under linear maps of locally finite automorphisms of $\F_3^3$ (see lemma
\ref{CompLocFin}). There are $80$ orbits of affine automorphisms, composing a representative of each class with all of the $2,835$ tame automorphisms, gives us $226,800$ representatives of ``conjugacy classes''. We checked for each of them whether it was locally finite or not.
It turns out that $25,872$ of these conjugacy classes are locally finite. And there are exactly a hundred different minimum polynomials that can appear.
Of the appearing minimal polynomials in this list, all polynomials of degree 3 appear in this list. The highest minimum polynomials are of degree 10. We list just a (sort of random, non-affine) ten minimum polynomials, their order (which is determined by the minimum polynomial), number of {\em conjugacy classes} with this minimum polynomial, and one example. The reader interested in the complete list we refer to chapter 6 of the Ph.-D. thesis of the second author \cite{RoelThesis}.\\
\ \\
{\tiny
\begin{tabular}{|l|c|c|l|}
\hline
Minimum polynomial & order & $\sharp$ & example\\ \hline \hline
$F^2+2I$ & 2 & 509 & $\left(\begin{array}{l} 2x^2+xy+xz+2x+y^2+z^2\\ 2x^2+xy+xz+y^2+2y+z^2\\ 2x^2+xy+xz+y^2+z^2+2z\end{array}\right)$ \\ \hline
$F^3+F^2+2F+2I$ & 6 & 5084 & $\left(\begin{array}{l} x^2+xy+2x+y^2\\ x^2+xy+y^2+2y\\ 2x^2+2y^2+2z\end{array}\right)$ \\ \hline
$F^4+2F^2+2F+2I$ & 24 & 2 & $\left(\begin{array}{l} x^2+xz+2x+y^2+2y+z^2\\ 2x+y+z\\ x^2+xz+x+y^2+y+z^2+z\end{array}\right)$ \\ \hline
$F^4+2F^3+2F+I$ & 9 & 3804 & $\left(\begin{array}{l} 2x^2+2xy+x+2y^2+1\\ 2x^2+2xy+2y^2+y+1\\ 2x^2+xy+2x+2y^2+z+1\end{array}\right)$\\ \hline
$F^4+F^3+F^2+2F+I$ & 8 & 38 & $\left(\begin{array}{l} x^2+2xy+xz+x+y^2+yz+z^2+2z+2\\ x^2+2xy+xz+y^2+yz+z^2+2z\\ 2x^2+xy+2xz+2x+2y^2+2yz+2y+2z^2+2\end{array}\right)$ \\ \hline
$F^5+2F^3+2F^2+F+2I$ & 8 & 8 & $\left(\begin{array}{l} 2x^2+xy+xz+y^2+2y+z^2\\ 2x^2+xy+xz+2x+y^2+2y+z^2+z\\ 2x^2+xy+xz+x+y^2+z^2+z\end{array}\right)$ \\ \hline
$F^6+F^5+2F^4+F^3+2I$ & 24 & 16 & $\left(\begin{array}{l} y^2+yz+2y+z^2\\ 2x+y^2+yz+2y+z^2+z\\ x+y^2+yz+z^2+z\end{array}\right)$ \\ \hline
$F^7+F^6+2F+2I $ & 18 & 396 & $\left(\begin{array}{l} 2x^2+2xz+2y^2+2y+2z^2+2z+1\\ x^2+xz+2y+z^2\\ 2x^2+2xz+2x+2y^2+2y+2z^2+2\end{array}\right)$ \\ \hline
$F^{10}+F^8+2F^5+F^2+2F+2I $ & 26 & 40 & $\left(\begin{array}{l} y+2z^2+z+1\\ x^2+2xz+x+z^2+1\\ x+z+1\end{array}\right)$ \\ \hline
$F^{10}+F^9+2F^8+F^7+F^6+F^5+2F^3+2F+I$ & 13 & 48 & $\left(\begin{array}{l} 2x^2+2xy+2xz+y^2+yz+y+2z^2+2z+1\\ 2x+y+z+2\\ 2x^2+2xy+2xz+2x+yz+y+2z^2+2z+1\end{array}\right)$ \\ \hline
\end{tabular}}
\subsubsection{Degree 3 over $\F_3$}
The amount of elements in $\overline{\ME}_3(\F_3)$ of the form $(x,y,z)+(0,H_2,H_3)$ (i.e. satisfying the dependency criterion)
where $H_2,H_3$ are homogeneous of degree 3 is too large: this set has $3^{20}$ elements which was too large for our system to scan through; however, we think that this case is feasible for someone having a stronger, dedicated system and a little more time.
\subsection{The finite fields $\mathbb{F}_4$ and $\mathbb{F}_5$}\label{F4F5F7}
In this section we will only restrict to degree 2, and to the maps which satisfy the dependency conjecture.
Thus, in this section we restrict to maps $F$ of the form $(x+H_1,y+H_2,z)$ where $H_1,H_2$ are of degree 2.
\subsection{The finite field $\F_4$}
There are $(64-1)(64-4)(64-16)=181,440$ linear automorphisms and $64*181,440=11,612,160$ affine automorphisms.
We considered the follwing maps:
\begin{itemize}
\item $F\in \overline{\ME}_3^2(\F_4)$,
\item $F$ is a mock automorphism,
\item $F$ is of the form $(x+H_1(x,y,z), y+H_2(x,y,z),z)$ (i.e. $F$ satisfies the dependency criterion).
\end{itemize}
and we counted $40,384$ such maps.
Under tame equivalence, we have the following classes:
\begin{itemize}
\item[1] $(x,y,z)$ (tame automorphisms)
\item[2] $(x+x^2+x^4, y, z)$
\end{itemize}
So, surprisingly, we only find a subset of the classes we found over $\F_2$. Well, not really surprising - the dependency criterion removes the classes 3 and 4 of theorem \ref{T1} from the list. We conjecture that the four classes of theorem \ref{T1} are the same for $\F_4$:
\begin{conjecture} (i) Suppose $F\in \ME_3^2(\F_4)$ is a mock automorphism of $\F_4$.
Then $F$ is tamely equivalent to $(P(x),y,z)$ where \[ P=x, P=x^4+x^2+x, P=x^8+x^4+x, \textup{~or~}P=x^8+x^2+x.\]
(ii) Suppose $F\in \ME_3^2(L)$ is a mock automorphism of $L$, where $[L:\F_2]<\infty$.
Then $F$ is tamely equivalent to $(P(x),y,z)$ where \[ P=x, P=x^4+x^2+x, P=x^8+x^4+x, \textup{~or~}P=x^8+x^2+x.\]
If $3|[L:\F_2]$ then one should remove the class of $P=x^4+x^2+x$, and if $7|[L:\F_2]$ then one should remove the classes of
$ P=x^8+x^4+x$ and $P=x^8+x^2+x$.
\end{conjecture}
It would be interesting to see a proof of this conjecture by theoretical means - or a counterexample of course.
\subsection{The finite field $\F_5$}
There are $(125-1)(125-5)(125-25)=1,488,000$ linear automorphisms and $125\cdot1,488,000=1,186,000,000$ affine automorphisms.
We consider maps of the following form:
There are $3,625$ mock automorphisms of $\F_5$ of degree at most 2. endomorphisms satisfying the following:
\begin{itemize}
\item $F\in \overline{\ME}_3^2(\F_5)$,
\item $F$ is a mock automorphism of $\F_5$,
\item $F$ satisfies the dependency criterion (i.e. $F=(x+H_1(x,y,z), y+H_2(x,y,z),z)$).
\end{itemize}
We counted $3,625$ such maps - and becaus of Corollary \ref{Jac1aut}, they are all automorphisms.
They all turned out to be tame maps.
\section{Conclusions}
We can gather some of the results in the below theorem:
\begin{theorem} Let $F\in \GA_3^d(\F_q)$. If one of the below conditions is met, then $F$ is tame:
\begin{itemize}
\item $d=3$, $q=2$,
\item $d=2$, $q=3$,
\item $d=2$, $q=4$ or $5$, and $F$ satisfies the Dependency criterion.
\end{itemize}
\end{theorem}
This gives rise to the following conjecture:
\begin{conjecture} If $F=I+H\in \GA_n(k)$ where $H$ is homogeneous of degree 2, then $F$ is tame.
\end{conjecture}
This natural conjecture might have been posed before, but we are unaware. This article proves this conjecture for $n=3$ and $k=\F_2,\F_3,\F_4,\F_5$. We expect that for $n=3$ and a generic field a solution is within reach.\\
Unfortunately, the computations did not allow us to go as far as finding some candidate non-tame automorphisms (though the Nagata automorphism is one, however it is of too high degree). However,
one of the interesting conclusions is that the set of {\em classes} (under tame automorphisms) of mock automorphisms seems to be much smaller than we originally expected: only 4 (perhaps 3) over $\F_2$ up to degree 2, and at most 7 over $\F_2$ of degree 3.
In particular, we are puzzled by the interesting question
whether the two endomorphisms over $\F_2$ described by $(x^8+x^4+x,y)$ and $(x^8+x^2+x,y)$ are not equivalent, as stated in question \ref{Qu1}.
{\bf Computations:} For computations we used the MAGMA computer algebra program. The reader interested in the routines we refer to chapter 6 of the thesis of the second author, \cite{RoelThesis}. Also, we posess databases usable in MAGMA, which we hope to share in the near future on a website. \\
{\bf Acknowledgements:} The second author would like to thank Joost Berson for some useful discussions.
|
\section{Introduction}
A puzzling feature of the hot Jupiters is that many of them have radii that are either larger or smaller than one would have guessed prior to the discovery of this class of objects. This ``radius anomaly problem'' has been present since the first transiting planet was discovered by \citet{2000ApJ...529L..45C} and \citet{2000ApJ...529L..41H}, and still has no universally acknowledged resolution. \citet{2010SSRv..152..423F} have reviewed many of the proposed solutions, and even in the short time since their review several other theories have been proposed \citep[see, e.g.,][]{2010ApJ...724..313P, 2010ApJ...714L.238B}. The small size of some planets can be explained as a consequence of heavy-metal enrichment, beyond the enrichment factors of Jupiter and Saturn and comparable to those of Uranus and Neptune. As for the larger radii, possible explanations include tidal friction, unexpected atmospheric properties, and resistive heating from electrical currents driven by star-planet interactions.
This paper presents follow-up observations of three exoplanets that were found to have anomalous radii. As in previous papers in this series, the purpose of the observations was to refine the system parameters (thereby checking on the magnitude of the radius problem) and to check for any transit timing anomalies that might be caused by additional gravitating bodies in the system.
Two of our targets are among the most ``bloated'' planets known. TrES-4 was discovered by \citet{2007ApJ...667L.195M}, and the two high-precision light curves that accompanied the discovery paper were reanalyzed by \citet{2008ApJ...677.1324T} and \citet{2009ApJ...691.1145S}. Here, we present 5 new light curves for this system. We also present 2 new light curves for WASP-12, a planet that was discovered by \citet{2009ApJ...693.1920H} and for which occultation photometry has been used to characterize the planet's atmosphere and orbit \citep{2010arXiv1003.2763C,2011Natur.469...64M}. Our third target, HAT-P-3 \citep{2007ApJ...666L.121T}, is in the opposite category of planets that are ``too small.'' \citet{2010MNRAS.401.1917G} have published 7 high-quality light curves of the system. We present six new light curves, and provide independent estimates of the planetary and stellar parameters.
\section{Observations and Data Reduction}
Almost all the observations were conducted at the Fred Lawrence Whipple Observatory (FLWO) located on Mt.\ Hopkins, Arizona, using the 1.2m telescope and KeplerCam detector. The KeplerCam is a $4096^2$ CCD with a field of view of $23\farcm1 \times 23\farcm1$. The pixels were binned $2 \times 2$ on the chip for faster readout. The binned pixels subtend $0\farcs68$ on a side. Observations were made through Sloan $i$ and $z$ filters. One of the WASP-12 transits was observed with the Nordic Optical Telescope (NOT) located in the Canary Islands, using the ALFOSC detector. The ALFOSC detector is a $2048^2$ CCD with a field of view of $6\farcm4 \times 6\farcm4$, corresponding to $0\farcs19$ per pixel. The observation was made through a Johnson $V$ filter. On each night we attempted to observe the entire transit, with at least an hour before ingress and an hour after egress, but the weather did not always cooperate.
\begin{figure*}[ht]
\begin{center}
\leavevmode
\hbox{
\epsfxsize=7.5in
\epsffile{light_curves_tres4_vertical.eps}}
\end{center}
\vspace{-0.25in}
\caption{Relative photometry of TrES-4 in the $i$-band. From top to bottom,
the observing dates are
2008~Jun~10, 2009~Apr~1, 2009~May~3, 2010~Apr~27 and 2010~May~4.
See Table 1 for the cadence and rms residual of each light curve.
The bottom plot is a composite light curve averaged into 3~min bins.}
\vspace{0.0in}
\label{fig:lc-tres4}
\end{figure*}
We performed overscan correction, trimming, bias subtraction and flat-field division with IRAF\footnote{IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy (AURA) under cooperative agreement with the National Science Foundation.}. To generate the light curves, we performed aperture photometry on the target star and all the comparison stars with similar brightnesses to the target star (within about a factor of two). We tried many different choices for the photometric aperture and found, unsurprisingly, that the best aperture diameter was approximately twice the full width at half maximum (FWHM) of the star. A comparison signal was formed from the weighted average of the flux histories of the comparison stars. The weights were chosen to minimize the out-of-transit (OOT) noise level. Some comparison stars that did not seem to provide a good correction were rejected. In general, 6-8 comparison stars were used to generate the final comparison signal. The target star's flux history was divided by the comparison signal and then multiplied by a constant to give a mean flux of unity outside of the transit.
Table 1 is a journal of all our observations, including those that were spoiled by bad weather. The light curves are displayed in Figures 1--4, after having been corrected for differential extinction. Section 3 explains how this correction was applied. The airmass-corrected data are given in electronic form in Tables 2-4.
\begin{figure*}[ht]
\begin{center}
\leavevmode
\hbox{
\epsfxsize=7.5in
\epsffile{light_curves_hat3_i_vertical.eps}}
\end{center}
\vspace{-0.25in}
\caption{Relative photometry of HAT-P-3 in the $i$-band.
From top to bottom,
the observing dates are
2008~Mar~8, 2008~Apr~6 and 2008~May~5.
See Table 1 for the cadence and rms residual of each light curve.
The bottom plot is a composite light curve averaged into 3~min bins.}
\label{fig:lc-hat3-i}
\end{figure*}
\begin{figure*}[ht]
\begin{center}
\leavevmode
\hbox{
\epsfxsize=7.5in
\epsffile{light_curves_hat3_z_vertical.eps}}
\end{center}
\vspace{-0.25in}
\caption{Relative photometry of HAT-P-3 in the $z$-band.
From top to bottom,
the observing dates are
2009~Mar~14, 2009~Mar~20 and 2009~Apr~15.
See Table 1 for the cadence and rms residual of each light curve.
The bottom plot is a composite light curve averaged into 3~min bins.}
\label{fig:lc-hat3-z}
\end{figure*}
\begin{figure*}[ht]
\begin{center}
\leavevmode
\hbox{
\epsfxsize=7.5in
\epsffile{light_curves_wasp12_vertical.eps}}
\end{center}
\vspace{-0.25in}
\caption{Relative photometry of WASP-12 in the $z$- and $V$-bands.
The top light curve is based on $z$-band observations on 2009~Jan~08.
See Table 1 for the cadence and rms residual of each light curve.
The bottom light curve is based on
$V$-band observations on 2009~Dec~06.}
\label{fig:lc-wasp12}
\end{figure*}
\section{Determination of System Parameters}
\label{sec:sys-params}
Our techniques for light-curve modeling and parameter estimation are similar to those
employed in previous papers in this series \citep[see, e.g.,][]{2006ApJ...652.1715H,2007AJ....133...11W}. The basis for the light-curve model was the formula of \citet{2002ApJ...580L.171M}, assuming
quadratic limb darkening and a circular orbit. The set of model parameters included the planet-to-star radius ratio ($R_p/R_\star$), the stellar radius in units of orbital distance ($R_\star/a$), the impact parameter ($b\equiv a\cos i/R_\star$, where $i$ is the orbital inclination), the time of conjunction for each individual transit ($T_c$), and the limb-darkening parameters $u_1$ and $u_2$.
We also fitted for two parameters ($\Delta m_0$, $k_z$) specifying a correction for differential extinction,
\begin{equation}
\label{eq:diff-ext}
\Delta m_{\rm cor} = \Delta m_{\rm obs} + \Delta m_0 + k_z z,
\end{equation}
where $z$ is the airmass, $\Delta m_{\rm obs}$ is the observed magnitude, and $\Delta m_{\rm cor}$ is the corrected magnitude that to be compared to the idealized transit model.
Since the data are not precise enough to determine both of the limb-darkening parameters, we followed the suggestion of \citet{2008MNRAS.390..281P} to form uncorrelated linear combinations of those parameters. We allowed the well-constrained combination to be a free parameter and held the poorly-constrained combination fixed at a tabulated value for a star of the appropriate type. In P{\'a}l's notation, the rotation angle $\phi$ was taken to be near $37^\circ$ in all cases. To determine the tabulated values we used a program kindly provided by J.\ Southworth to query and interpolate the tables of \citet{2004yCat..34281001C}.\footnote{The interpolated values $(u_1,u_2)$ for the cases of TrES-4 $i$, HAT-P-3 $i$, HAT-P-3 $z$, WASP-12 $z$, and WASP-12 $V$ are $(0.20,0.37)$, $(0.38,0.27)$, $(0.30,0.29)$, $(0.14,0.36)$, $(0.36,0.35)$, respectively.}
For parameter estimation, we used a Markov Chain Monte Carlo (MCMC) algorithm to sample from the posterior probability distribution, employing the
Metropolis-Hastings jump criterion and Gibbs sampling. Uniform priors were adopted for all parameters,
and the likelihood was taken to be $\exp(-\chi^2/2)$ with
\begin{equation}
\label{eqn:chisq}
\chi^2 = \sum_{i,j} \left( \frac{f_{{\rm obs},i,j} - f_{{\rm calc}, i,j}}{\sigma_{i,j}} \right)^2
\end{equation}
where $f_{{\rm obs}, i,j}$ is the $j$th data point from the $i$th light curve, $f_{{\rm calc},i,j}$ is the calculated light curve based on the current parameters, and $\sigma_{i,j}$ is the uncertainty associated with $f_{{\rm obs}, i,j}$. All the light curves for a given planet were fitted simultaneously. The uncertainties were determined in a two-step process. First, the standard deviation $\sigma_i$ of the OOT data was determined for each light curve. In a few cases, the pre-transit and post-transit noise levels were very different (due to different airmasses); in these cases the starting point was a function $\sigma_{i,j}$ that interpolated linearly between the two differing noise levels. Second, the preceding uncertainty estimates were multiplied by a correction factor $\beta\geq 1$ intended to account for time-correlated noise. The $\beta$ factor was determined with the ``time-averaging'' procedure
\citep{2006MNRAS.373..231P,2008ApJ...683.1076W,2009ApJ...704...51C}, using bin sizes bracketing the ingress/egress duration by a factor of 2. The values of $\beta$ are given in Table~\ref{tbl:obs}.
The starting point for each Markov chain was determined by minimizing $\chi^2$, and then perturbing those parameters by Gaussian random numbers with a standard deviation of $10\sigma$, where $\sigma$ is the rough uncertainty estimate returned by the least-squares fit. We ran several test chains to establish the appropriate jump sizes, giving acceptance rates near 40\%. Then we ran $4-5$ chains each with $10^6$ links, ignored the initial 20\% of each chain, and ensured convergence according to the Gelman-Rubin statistic \citep{GR1992}. The quoted value for each parameter is the median of the one-dimensional marginalized posterior, and the quoted uncertainty interval encloses 68.3\% of the probability (ranging from the 15.85\% to the 84.15\% levels of the cumulative probability distribution).
\section{Results}
The results for the model parameters, and various derived parameters of interest, are given in Tables~\ref{tbl:params-tres4},~\ref{tbl:params-hat3}~and~\ref{tbl:params-wasp12}. The next subsection explains how the stellar and plantetary dimensions were calculated from the combination of light-curve parameters and stellar-evolutionary models. This is followed by a subsection presenting an examination of the transit times.
\subsection{The stellar and planetary radii}
The stellar and planetary radii and masses cannot be determined from transit parameters alone. The route we followed to determining these dimensions was to set the mass scale by using an estimated stellar mass $M_\star$ and the observed semiamplitude $K_\star$ of the star's radial-velocity orbit. The stellar mass is itself estimated by using stellar-evolutionary models with inputs from the observed spectral parameters, as well as the mean density that is calculated from the light-curve parameters. For the relevant formulas and discussion see \citet{2007ApJ...664.1190S} or \citet{2010arXiv1001.2010W}. We used previously measured values of $K_\star$, documented in Tables~\ref{tbl:params-tres4},~\ref{tbl:params-hat3}~and~\ref{tbl:params-wasp12}. For the evolutionary models, we used the Yonsei-Yale ($\rm Y^2$) isochrones \citep{2001ApJS..136..417Y}.
The $\rm Y^2$ isochrones can be thought of as an algorithm that takes as input the age, metallicity, mass, and concentration of $\alpha$-elements of a star, and returns the star's temperature, mass, density, and other properties. We interpolated the $\rm Y^2$ isochrones in age from $0.1$ to $14$ Gyr in steps of $0.1$ Gyr, and in metallicity from $-0.20$ to $0.58$ dex in steps of $0.02$ dex. Then we used linear interpolation to create a $4\times$ finer mass sampling for each metallicity and age. We assumed the concentration of $\alpha$-elements to be solar.
To each model star in the resulting Y$^2$ grid, we assigned a likelihood
based on the measured metallicity $Z$ and effective temperature $T_{\rm eff}$ (taken
from the literature) as well as the
stellar mean density $\rho_\star$ determined solely from the transit parameters.
Following \citet{2009ApJ...696..241C}, the likelihood was taken to be proportional
to $n \exp( - \chi^2/2)$,
where
\begin{equation}
\chi^2 = \left( \frac{Z-Z_{\rm obs}}{\sigma_Z} \right)^2 +
\left( \frac{T_{\rm eff} - T_{\rm eff, obs}}{\sigma_{T_{\rm eff}}} \right)^2 +
\left( \frac{\rho_\star}{\rho_{\star, {\rm obs}}} \right)^2,
\end{equation}
and $n$ is the number density of stars as a function of mass, according to Salpeter's law with exponent $-1.35$. The effect of multiplying by $n$ is to set a prior so that extremely rare and short-lived stars are disfavored, even though they might provide a good fit. (The effect of this prior was generally small.) Finally, the ``best-fitting'' values and uncertainties were computed from the appropriate likelihood-weighted integrals in the space of model stars.
With the stellar mass thereby determined, it is possible to compute the other dimensions $R_p, R_\star,$ and $M_p$. The results are given in Tables~\ref{tbl:params-tres4},~\ref{tbl:params-hat3}~and~\ref{tbl:params-wasp12}. We note that this procedure does not take into account any uncertainty in the $\rm Y^2$ isochrones themselves and, therefore, is subject to systematic errors that probably amount to a few percent \citep[see, e.g.,][]{2008ApJ...677.1324T}.
\subsection{Transit times and revised ephemerides}
\label{subsec:ephemeris}
We analyzed the new transit times in conjunction with previously published midtransit times, to seek evidence for significant discrepancies from strict periodicity and refine the ephemerides to allow for accurate prediction of future events. The new transit times are given in Tables~\ref{tbl:midtransits-tres4},~\ref{tbl:midtransits-hat3}~and~\ref{tbl:midtransits-wasp12}, with uncertainties determined by the MCMC analysis described in Section 3. Previously published midtransit times were taken from \citet{2010MNRAS.401.1917G}, \citet{2009ApJ...693.1920H}, \citet{2009ApJ...691.1145S}, \citet{2007ApJ...666L.121T}, and \citet{2007ApJ...667L.195M}.
The transit times for each system were fitted with a linear function,
\begin{equation}
\label{eqn:linfit}
T_C = T_0 + E \cdot P
\end{equation}
where $E$ is an integer (the epoch), $P$ is the period, and $T_0$ is a particular reference time. The best-fitting values of $P$ and $T_0$ were determined by linear regression. Figures 5-7 show the residuals to the fits, and the captions specify the minimum $\chi^2$ and the number of degrees of freedom. In no case is there clear evidence of timing anomalies. For TrES-4, there is formally only a 6\% chance of obtaining such a large $\chi^2$ with random Gaussian errors, but we do not deem this significant enough to warrant special attention.
The case of WASP-12 required somewhat special treatment because \citet{2009ApJ...693.1920H} did not report individual midtransit times, but rather a consensus reference time based on observations of multiple transits. We included their quoted reference time as a single data point. \citet{2010arXiv1003.2763C} provided many transit times obtained over several years, but most of the data were from amateur observers and were not presented in detail or evaluated critically. For this reasons we did not include them; however, as a consequence, there are not many points in our fit.
The results for $P$ and $T_0$ are given in Tables 5-7. They are based on a fit in which all of the uncertainties of the transit times were rescaled by a common factor to give $\chi^2/N_{\rm dof} = 1$. Our intention is to provide conservative error estimates to allow for planning of future observations. The uncertainties on the individual transit times given in Tables 8-10 were {\it not} rescaled in this way, nor were the error bars that are plotted in Figures 5-7.
\begin{figure}[ht]
\epsscale{0.8}
\plotone{timing_residuals_tres4_unscaled.eps}
\caption{ Timing residuals for TrES-4. The data presented
in this paper are labeled with solid circles (complete transits) and solid
trianges (partial transits).
Open circles represent transits observed by \citet{2007ApJ...667L.195M}.
The best fit gives $\chi^2=12.1$ with 6 degrees of freedom.
The probability of obtaining a higher $\chi^2$ with random Gaussian
data points is about 6\%.
\label{fig:timing-residuals-tres4}}
\end{figure}
\begin{figure}[ht]
\epsscale{0.8}
\plotone{timing_residuals_hat3_unscaled.eps}
\caption{ Timing residuals for HAT-P-3.
The data presented
in this paper are labeled with solid circles (complete transits) and solid
trianges (partial transits).
Open circles represent transits observed by \citet{2010MNRAS.401.1917G, 2007ApJ...666L.121T}.
The best fit gives $\chi^2=10.4$ with 12 degrees of freedom.
The probability of obtaining a higher $\chi^2$ with random Gaussian
data points is about 60\%.
\label{fig:timing-residuals-hat3}}
\end{figure}
\begin{figure}[ht]
\epsscale{0.8}
\plotone{timing_residuals_wasp12_unscaled.eps}
\caption{ Timing residuals for WASP-12.
Solid circles are complete transits that we observed, and the open circle
is the reference transit time (derived from observations of multiple events) quoted
by \citet{2009ApJ...693.1920H}. The best fit gives $\chi^2=2.0$ with 1 degree of freedom.
\label{fig:timing-residuals-wasp12}}
\end{figure}
\section{Summary and Discussion}
We have presented new photometry and new analyses of the transiting exoplanets TrES-4b, HAT-P-3b, and WASP-12b. Whereas the discovery papers reporting TrES-4b and HAT-P-3b included only a few high-precision light curves, our analyses are based on 5-6 such datasets. Likewise, the WASP-12b discovery paper featured only one high-precision light curve, to which we have added two. We have applied consistent and conservative procedures for parameter estimation, including an accounting for uncertainties in the limb-darkening law and due to time-correlated noise, as well as linkage between the light curve parameters and stellar-evolutionary models, that were not always applied by previous authors. In the past these efforts have occasionally led to significant revisions of the planetary dimensions \citep[see, e.g.][]{2007AJ....134.1707W,2008ApJ...683.1076W}.
In the present case our results are in agreement with the previously reported results. All of the TrES-4 parameters agree to within 2$\sigma$ with the results reported by \citet{2008ApJ...677.1324T}. Two of the most important parameters, the mass and radius of the planet, agree to within 1$\sigma$. Our HAT-P-3 parameters agree to within 2$\sigma$ with those reported by \citet{2010MNRAS.401.1917G}, and the planetary mass and radius agree to within 1.3$\sigma$. Our WASP-12 parameters agree to within 2$\sigma$ with those reported by \citet{2009ApJ...693.1920H}, and the mass and radius of the planet agree to within 1$\sigma$.
Our transit ephemeris for TrES-4 agrees with that of \cite{2007ApJ...667L.195M}, and the refined orbital period is about 20 times more precise. Our ephemeris for HAT-P-3 agrees with that of \citet{2010MNRAS.401.1917G}, and is 2-3 times more precise. Our ephemeris for WASP-12 agrees with that of \citet{2009ApJ...693.1920H}, and is of comparable precision, despite being based on only a few well-documented data points.
The reason that TrES-4, HAT-P-3, and WASP-12 are of particular interest is because their measured dimensions do not agree with standard models of gas giant planets. TrES-4 and WASP-12 are heavily bloated, with radii too large for their masses, while HAT-P-3 is too small. With reference to the tables of \citet{2007ApJ...659.1661F}, TrES-4 and WASP-12 are incompatible with pure hydrogen-helium giant planets at the 10$\sigma$ and 5$\sigma$ levels, respectively. Enhancing these planets with metals to the degree of Jupiter or Saturn would only make the problem worse. HAT-P-3, in contrast, is compatible with the tabulated models if it is endowed with approximately 100~$M_\earth$ of heavy elements. Our results do not change these interpretations of the three systems we have studied. Rather, our results lend more confidence to the claims that the dimensions of the planets are anomalous, and merit attention by theoreticians who seek to solve the radius anomaly problem.
\acknowledgments We thank Gerald Nordley for checking some of the entries in Tables 5-7. We gratefully acknowledge support from the NASA Origins program through award NNX09AB33G and from the Research Science Institute, a program of the Center for Excellence in Education. Some of the data presented herein were obtained with the Nordic Optical Telescope (NOT), operated on the island of La Palma jointly by Denmark, Finland, Iceland, Norway, and Sweden, in the Spanish Observatorio del Roque de los Muchachos Instituto Astrofisica de Canarias, and ALFOSC, which is owned by the Instituto Astrofisica de Andalucia (IAA) and operated at the NOT under agreement between IAA and NBlfAFG of the Astronomical Observatory of Copenhagen.
\bibliographystyle{apj}
|
\section{Boundary conditions} \label{sec:BC}
In the moment methods, the boundary condition is always a complicated
issue when simulating microflows. As discussed in \cite{Grad,
Struchtrup2000,Thatcher,Torrilhon2008,Emerson} and the references
therein, delicate derivations and careful numerical techniques are
needed for a solid wall. In this section, a numerical way for dealing
with boundary conditions in the \NRxx method is introduced, which
appears to be uniform for all orders of moment systems.
\subsection{The kinetic boundary condition}
In the kinetic theory, the most extensively used boundary condition is
the one proposed by Maxwell in \cite{Maxwell}. According to the common
hyperbolic theory, for \eqref{eq:Shakhov}, the boundary condition is
only needed when $\bxi \cdot \bn < 0$, where $\bn$ is the outer normal
vector of the boundary. For a point $\bx$ on the wall, supposing the
velocity and temperature of the wall to be $\bu^W(t, \bx)$ and
$\theta^W(t, \bx)$ at time $t$, Maxwell proposed the following
boundary condition:
\begin{equation} \label{eq:bc}
f(t, \bx, \bxi) = \left\{ \begin{array}{ll}
\chi f_M^W(t,\bx,\bxi) + (1 - \chi) f(t, \bx, \bxi^*),
& \bC^W \cdot \bn < 0, \\[5pt]
f(t, \bx, \bxi), & \bC^W \cdot \bn \geqslant 0,
\end{array} \right.
\end{equation}
where $\chi \in [0,1]$ is a parameter for different gases and walls,
and
\begin{gather}
\bxi^* = \bxi - 2(\bC^W \cdot \bn) \bn,
\quad \bC^W = \bxi - \bu^W(t, \bx), \\
\label{eq:f_M}
f_M^W(t, \bx, \bxi) =
\frac{\rho^W(t, \bx)}{(2\pi \theta^W(t, \bx))^{3/2}}
\exp \left(
-\frac{|\bxi - \bu^W(t, \bx)|^2}{2\theta^W(t, \bx)}
\right).
\end{gather}
The functions $\bu^W(t, \bx)$ and $\theta^W(t, \bx)$ are prescribed
and stand for the wall velocity and temperature at time $t$ and
position $\bx$, and $\rho^W(t, \bx)$ ensures the conservation of the
mass at the wall, that is,
\begin{equation} \label{eq:mass_csv}
\begin{split}
& \int_{\bbR^3} (\bC^W \cdot \bn) f(t, \bx, \bxi) \dd \bxi \\
={} & \chi \left(
\int_{\bC^W \cdot \bn < 0}
(\bC^W \cdot \bn) f_M^W(t, \bx, \bxi) \dd \bxi +
\int_{\bC^W \cdot \bn \geqslant 0}
(\bC^W \cdot \bn) f(t, \bx, \bxi) \dd \bxi
\right) = 0.
\end{split}
\end{equation}
For this boundary condition, the normal velocity of gas on the
boundary is the same as the normal velocity of the wall. However, in
the case of shear flow, velocity slip and temperature jump will appear
on the boundary.
\subsection{The boundary conditions for the \bNRxx method}
The boundary condition can be derived by taking moments on both sides
on \eqref{eq:bc}. Before that, we define
\begin{equation}
C_{\theta,\alpha} = \frac{(2\pi)^{3/2} \theta^{|\alpha| + 3}}
{\alpha_1! \alpha_2! \alpha_3!},
\qquad \forall \theta > 0, \quad \alpha \in \bbN^3.
\end{equation}
This definition leads to
\begin{equation}
g_{\alpha} = C_{\theta,\alpha} \int_{\bbR^3}
g(\bxi) \mathcal{H}_{\theta,\alpha}(\bv) \exp(|\bv|^2 / 2)
\dd \bv,
\end{equation}
where $\bv = (\bxi - \bu) / \sqrt{\theta}$ and $g(\bxi)$ is a
distribution function expanded into Hermite series as $g(\bxi) =
\sum_{\alpha \in \bbN^3} g_{\alpha} \mathcal{H}_{\theta,\alpha}(\bv)$.
In order to simplify the calculation, we suppose $\bn = (0, 1, 0)^T$.
Thus, taking moments for \eqref{eq:bc} requires half-space integration
\begin{equation} \label{eq:hs_int}
C_{\theta,\alpha} \int_{\xi_2 \geqslant u_2^W}
g(\bxi) \mathcal{H}_{\theta,\alpha}(\bv) \exp(|\bv|^2/2) \dd \bv.
\end{equation}
Suppose an $M$-th order system is used in the \NRxx method; that is,
an $(M+1)$-st order approximation of the distribution can be obtained
through \eqref{eq:reg}. This approximation is directly used in \eqref
{eq:hs_int} so that the integral can be actually worked out.
Concretely speaking, \eqref{eq:hs_int} is approximated as
\begin{equation} \label{eq:sum}
\sum_{|\beta| \leqslant M+1} g_{\beta} C_{\theta,\alpha}
\int_{\xi_2 \geqslant u_2^W}
\mathcal{H}_{\theta,\alpha}(\bv)
\mathcal{H}_{\theta,\beta}(\bv) \exp(|\bv|^2 / 2)
\dd \bv.
\end{equation}
Since $u_2 = u_2^W$ on the boundary, the region of integration can be
written as $\{v_2 \geqslant 0\}$. Thus, we only need to calculate
\begin{equation} \label{eq:I_def}
I_{\alpha,\beta}(\theta) = C_{\theta,\alpha}
\int_{v_2 \geqslant 0}
\mathcal{H}_{\theta,\alpha}(\bv)
\mathcal{H}_{\theta,\beta}(\bv) \exp(|\bv|^2 / 2)
\dd \bv.
\end{equation}
The details can be found in Appendix \ref{sec:hs_int}, and the result
is
\begin{equation} \label{eq:I}
I_{\alpha,\beta}(\theta) =
S(\alpha_2,\beta_2) \theta^{\frac{\alpha_2 - \beta_2}{2}}
\cdot \delta_{\alpha_1 \beta_1} \delta_{\alpha_3 \beta_3}
\end{equation}
and
\begin{equation} \label{eq:S}
S(m,n) = \left\{ \begin{array}{ll}
1/2, & m = n = 0, \\
K(1,n-1), & m = 0 \text{ and } n \neq 0, \\
K(m,0), & m \neq 0 \text{ and } n = 0, \\
K(m,n) + S(m-1,n-1) \cdot n / m, & \text{otherwise},
\end{array} \right.
\end{equation}
where
\begin{equation} \label{eq:K}
K(m,n) = \frac{(2\pi)^{-1/2}}{m!} \He_{m-1}(0) \He_n(0).
\end{equation}
The above deduction leads to the following proposition:
\begin{proposition} \label{prop:hs}
Suppose $g(\bv)$ is a function defined on $\bbR^3$ which can be
denoted by a finite expansion of Hermite basis functions
\begin{equation}
g(\bv) = \sum_{|\alpha| \leqslant M + 1}
g_{\alpha} \mathcal{H}_{\theta,\alpha}(\bv).
\end{equation}
for some $\theta > 0$. Let $\tilde{g}(\bv)$ be a half-space cut-off of
$g(\bv)$ as
\begin{equation}
\tilde{g}(\bv) = \left\{ \begin{array}{ll}
g(\bv), & v_2 \geqslant 0, \\
0, & v_2 < 0. \\
\end{array} \right.
\end{equation}
Then $\tilde{g}$ can also be expanded into Hermite series as
\begin{equation} \label{eq:g_exp}
\tilde{g}(\bv) = \sum_{\alpha \in \bbN^3}
\sum_{|\beta| \leqslant M + 1} g_{\beta} I_{\alpha,\beta}(\theta)
\mathcal{H}_{\theta,\alpha}(\bv),
\end{equation}
where $I_{\alpha,\beta}(\theta)$ is defined in \eqref{eq:I}---\eqref
{eq:K}.
\end{proposition}
\begin{proof}
It is already known in \cite{NRxx} that $\{\mathcal{H}_{\theta,\alpha}
(\bv)\}_{\alpha \in \bbN^3}$ is an orthogonal basis of the weighted
$L^2$ space $L^2(\bbR^3; \exp(|\bv|^2 / 2) \dd \bv)$. Since
\begin{equation}
\begin{split}
& \int_{\bbR^3} |\tilde{g}(\bv)|^2 \exp(|\bv|^2/2) \dd \bv =
\int_{v_2 \geqslant 0} |g(\bv)|^2 \exp(|\bv|^2/2) \dd \bv \\
\leqslant {} & \int_{\bbR^3} |g(\bv)|^2 \exp(|\bv|^2/2) \dd \bv =
\sum_{|\alpha| \leqslant M + 1}
C_{\theta,\alpha}^{-1} |g_{\alpha}|^2 < +\infty,
\end{split}
\end{equation}
$\tilde{g}(\bv)$ also lies in $L^2(\bbR^3; \exp(|\bv|^2 / 2) \dd
\bv)$. Thus the validity of \eqref{eq:g_exp} can be naturally
obtained.
\end{proof}
The following proposition depicts the sparsity of $I_{\alpha,\beta}$.
\begin{proposition} \label{prop:sparsity}
If $I_{\alpha,\beta}(\theta)$ is nonzero, then (1) $\alpha_1 =
\beta_1$; (2) $\alpha_3 = \beta_3$; (3) $\alpha_2 - \beta_2$ is zero
or odd. When $\alpha = \beta$, $I_{\alpha,\beta}(\theta)$ is equal to
$1/2$.
\end{proposition}
\begin{proof}
If $I_{\alpha,\beta}(\theta)$ is nonzero, \eqref{eq:I} directly gives
$\alpha_1 = \beta_1$ and $\alpha_3 = \beta_3$. If $\alpha_2 - \beta_2$
is a nonzero even integer, $K(\alpha_2, \beta_2)$ is zero since $\He%
_n(0)$ is zero when $n$ is odd. In order to prove $I_{\alpha,\beta}%
(\theta) = 0$ in this case, according to \eqref{eq:I}, we only need to
prove $S(\alpha_2, \beta_2) = 0$. This can be done by induction:
(1) If $\alpha_2 = 0$ or $\beta_2 = 0$, $\alpha_2$ and $\beta_2$ must
be both even but one of them must be positive. Equation \eqref{eq:S}
shows $S(\alpha_2, \beta_2) = 0$ directly.
(2) Suppose $S(\alpha_2 - 1, \beta_2 - 1) = 0$. Then, according to the
last case in \eqref{eq:S}, $S(\alpha_2, \beta_2)$ is also zero.
Finally, when $\alpha = \beta$, $\eqref{eq:I}$ gives $I_{\alpha,\beta}
(\theta) = S(\alpha_2, \beta_2)$. The subsequent proof can also be
done by induction, since $S(0,0) = 1/2$ and $K(n,n) = 0$ for $n > 0$.
\end{proof}
According to Proposition \ref{prop:sparsity}, we find that only
$(\lceil \alpha_2 / 2 \rceil + 1)$ terms are nonzero in the summation
\eqref{eq:sum}. This greatly reduces the computational cost.
Now let us return to the boundary conditions. According to Grad's
theory \cite{Grad, Grad1958}, in order to ensure the continuity of
boundary conditions when $\chi \rightarrow 0$, only a subset of
moments $\{f_{\alpha} \mid |\alpha| \leqslant M + 1 \text{ and }
\alpha_2 \text{ is odd}\}$ should be used to formulate boundary
conditions. This will be completed in the following three subsections.
Later in this section, for conciseness, the variables $t$ and $\bx$
are omitted in our statement if not specified, and all spatially
dependent functions are considered to be on the boundary.
\subsubsection{Determination of $\rho^W$}
For simplicity, we factorize the right hand side of \eqref{eq:bc}
into three parts and consider each part independently. Define
\begin{equation}
\begin{gathered}
p(\bxi) = \left\{ \begin{array}{ll}
f_M^W(\bxi), & \xi_2 < u_2^W, \\
0, & \xi_2 \geqslant u_2^W,
\end{array} \right. \quad
q(\bxi) = \left\{ \begin{array}{ll}
f(\bxi), & \xi_2 \geqslant u_2^W, \\
0, & \xi_2 < u_2^W,
\end{array} \right. \quad
r(\bxi) = q(\bxi) + q(\bxi^*).
\end{gathered}
\end{equation}
Then \eqref{eq:bc} can be rewritten as
\begin{equation} \label{eq:factorization}
f(\bxi) = \chi p(\bxi) + \chi q(\bxi) + (1 - \chi) r(\bxi).
\end{equation}
Suppose the Hermite expansion of $f$ is
\begin{equation} \label{eq:f}
f(\bxi) = \sum_{|\alpha| \leqslant M + 1}
f_{\alpha} \mathcal{H}_{\theta,\alpha}
\left( \frac{\bxi - \bu}{\sqrt{\theta}} \right).
\end{equation}
Then $q(\bxi)$ can also be expanded into Hermite series according to
Proposition \ref{prop:hs} and \ref{prop:sparsity} as
\begin{equation} \label{eq:q}
q(\bxi) = \sum_{\alpha \in \bbN^3}
q_{\alpha} \mathcal{H}_{\theta,\alpha}
\left( \frac{\bxi - \bu}{\sqrt{\theta}} \right).
\end{equation}
Substituting \eqref{eq:f_M} and \eqref{eq:q} into \eqref{eq:mass_csv},
$\rho^W$ can be worked out as
\begin{equation} \label{eq:rho_W}
\rho^W = \sqrt{\frac{2\pi}{\theta^W}} q_{e_2}
= \sqrt{\frac{2\pi}{\theta^W}}
\sum_{k=0}^{\lceil M/2 \rceil} S(1,2k) \theta^{1/2-k} f_{2k e_2},
\end{equation}
where the expression of $q_{e_2}$ is derived from \eqref{eq:I}, \eqref
{eq:g_exp} and Proposition \ref{prop:sparsity}.
\subsubsection{The moments of $p$ and $r$}
Now the moments for $q(\bxi)$ have been calculated in \eqref{eq:q}, we
still need to get Hermite expansions of $p(\bxi)$ and $r(\bxi)$.
We suppose that $p(\bxi)$ can be expanded under the basis
$\big\{\mathcal{H}_{\theta, \alpha} \big( (\bxi - \bu) / \sqrt{\theta}
\big)\big\}_{\alpha \in \bbN^3}$ as
\begin{equation}
p(\bxi) = \sum_{\alpha \in \bbN^3}
p_{\alpha} \mathcal{H}_{\theta, \alpha} \left(
\frac{\bxi - \bu}{\sqrt{\theta}}
\right).
\end{equation}
Then, according to \eqref{eq:f_M}, the coefficients can be formulated
by
\begin{equation} \label{eq:p}
p_{\alpha} = C_{\theta,\alpha} \int_{v_2 < 0}
\frac{\rho^W}{(2\pi\theta^W)^{3/2}} \exp \left(
- \frac{|\bxi - \bu^W|^2}{2\theta^W}
\right) \mathcal{H}_{\theta,\alpha}(\bv) \exp \left(
\frac{|\bv|^2}{2}
\right) \dd \bv,
\end{equation}
where $\bxi = \sqrt{\theta} \bv + \bu$. Define
\begin{align}
\label{eq:J}
J_s(x) &= \frac{1}{s!} \theta^{\frac{s+1}{2}}
\int_{-\infty}^{+\infty} \frac{1}{\sqrt{2\pi \theta^W}}
\exp \left( -\frac{|\sqrt{\theta} y - x|^2}{2 \theta^W} \right)
\He_s(y) \dd y, \\
\label{eq:tilde_J}
\tilde{J}_s(x) &= \frac{1}{s!} \theta^{\frac{s+1}{2}}
\int_{-\infty}^0 \frac{1}{\sqrt{2\pi \theta^W}}
\exp \left( -\frac{|\sqrt{\theta} y - x|^2}{2 \theta^W} \right)
\He_s(y) \dd y.
\end{align}
Then $p_{\alpha}$ can be expressed by
\begin{equation} \label{eq:p_alpha}
p_{\alpha} = \rho^W
J_{\alpha_1}(u_1^W - u_1)
\tilde{J}_{\alpha_2}(u_2^W - u_2)
J_{\alpha_3}(u_3^W - u_3).
\end{equation}
$J_s(x)$ and $\tilde{J}_s(x)$ can be calculated recursively as
\begin{align}
\label{eq:J_s}
J_s(x) &= \frac{1}{s} \left[
(\theta^W - \theta) J_{s-2}(x) + x J_{s-1}(x)
\right], \quad s \geqslant 1; \\
\label{eq:tilde_J_s}
\tilde{J}_s(x) &= \frac{1}{s} \left[
(\theta^W-\theta) \tilde{J}_{s-2}(x) + x \tilde{J}_{s-1}(x)
\right] - H_s(x), \quad s \geqslant 1; \\
H_s(x) &= -\frac{s-2}{s(s-1)} \theta H_{s-2}(x),
\quad s \geqslant 2.
\end{align}
The starting values are
\begin{align}
\label{eq:J_start}
J_{-1}(x) = 0, & \qquad J_0(x) = 1, \\
\label{eq:tilde_J_start}
\tilde{J}_{-1}(x) = 0, & \qquad
\tilde{J}_0(x) = \frac{1}{2} \mathrm{erfc} \left(
\frac{x}{\sqrt{2 \theta^W}}
\right), \\
\label{eq:H_start}
H_0(x) = 0, & \qquad
H_1(x) = \sqrt{\frac{\theta^W}{2\pi}} \exp \left(
-\frac{x^2}{2 \theta^W}
\right),
\end{align}
The detailed derivation of \eqref{eq:J_s}-\eqref{eq:H_start} can be
found in the Appendix \ref{sec:half_Max}. Noting that $u_2 = u_2^W$,
\eqref{eq:p_alpha} can be further simplified as
\begin{equation} \label{eq:p_alpha_simplified}
p_{\alpha} = \rho^W J_{\alpha_1}(u_1^W - u_1)
\hat{J}_{\alpha_2} J_{\alpha_3}(u_3^W - u_3),
\end{equation}
where
\begin{equation} \label{eq:J_H}
\begin{gathered}
\hat{J}_s = \frac{1}{s}(\theta^W - \theta) \hat{J}_{s-2} - \hat{H}_s,
\quad s \geqslant 1, \qquad
\hat{H}_s = -\frac{s-2}{s(s-1)} \theta \hat{H}_{s-2},
\quad s \geqslant 2, \\
\hat{J}_{-1} = \hat{H}_0 = 0, \quad
\hat{J}_0 = 1/2, \quad \hat{H}_1 = \sqrt{\frac{\theta^W}{2\pi}}.
\end{gathered}
\end{equation}
Here we emphasize that due to equation \eqref{eq:rho_W}, all
$p_{\alpha}$'s are only related with $\{f_{2ke_2}\}_{0 \leqslant k
\leqslant \lceil M/2 \rceil}$ besides $\bu$, $\bu^W$, $\theta$ and
$\theta^W$.
Now we turn to the moments of $r(\bxi)$. Note that only the moments
with odd $\alpha_2$ are needed. However, $r(\bxi)$ is an even function
with respect to $C_2^W$, which causes all its moments with odd
$\alpha_2$ vanished. This indicates that $r(\bxi)$ can be simply
neglected when discussing the boundary conditions.
\subsubsection{Construction of boundary conditions}
Now we take moments with odd $\alpha_2$ on both sides of \eqref
{eq:factorization}. Making use of Proposition \ref{prop:sparsity},
we have
\begin{equation}
f_{\alpha} = \chi p_{\alpha} + \chi q_{\alpha}
= \chi p_{\alpha} + \frac{1}{2} \chi f_{\alpha} +
\chi \sum_{k=0}^{K_2(\alpha)}
S(\alpha_2, 2k) \theta^{\alpha_2/2 - k}
f_{\alpha+(2k-\alpha_2)e_2},
\end{equation}
where $K_2(\alpha) = \lceil (M - \alpha_1 - \alpha_3) / 2 \rceil$. A
simple rearrangement gives
\begin{equation} \label{eq:mnt_bc}
f_{\alpha} = \frac{2\chi}{2 - \chi} \left[
p_{\alpha} + \sum_{k=0}^{K_2(\alpha)}
S(\alpha_2, 2k) \theta^{\alpha_2/2 - k}
f_{\alpha+(2k-\alpha_2)e_2}
\right].
\end{equation}
Equations \eqref{eq:mnt_bc} with $|\alpha| \leqslant M+1$ and odd
$\alpha_2$, together with $u_2 = u_2^W$ form the boundary conditions
of the dynamic moment equations. Recalling
\begin{equation}
p_{\alpha} = p_{\alpha}(\bu, \bu^W, \theta, \theta^W,
f_0, f_{2e_2}, \cdots, f_{2 \lceil M/2 \rceil e_2}),
\end{equation}
one can find that the terms which appear on the left hand side of
\eqref{eq:mnt_bc} never appear on its right hand side. Thus, if an
arbitrary distribution function denoted as \eqref{eq:f} is given, we
can define a functional $F^b$ which maps \eqref{eq:f} to another
distribution $f^b(\bxi)$:
\begin{equation} \label{eq:f_b}
f^b(\bxi) = \sum_{|\alpha| \leqslant M + 1}
f_{\alpha}^b \mathcal{H}_{\theta^b, \alpha} \left(
\frac{\bxi - \bu^b}{\sqrt{\theta^b}}
\right),
\end{equation}
where $\bu^b = (u_1, u_2^W, u_3)$, $\theta^b = \theta$, and
\begin{equation} \label{eq:f_b_alpha}
f_{\alpha}^b = \left\{ \begin{array}{ll}
f_{\alpha}, & \text{if $\alpha_2$ is even}, \\
\text{the right hand side of \eqref{eq:mnt_bc}}, &
\text{if $\alpha_2$ is odd}.
\end{array} \right.
\end{equation}
Thus $f_{\alpha}^b$ satisfies the boundary condition. The mapping
$F^b$ will be used in the numerical implementation of boundary
conditions.
At the end of this section, we prove that $\bu^b$ and $\theta^b$ are
the corresponding velocity and temperature of the distribution
function $f^b(\bxi)$. This is equivalent to the following proposition:
\begin{proposition} \label{prop:conservation}
If a distribution $f(\bxi)$ with expression \eqref{eq:f} satisfies
\begin{equation}
f_{e_1} = f_{e_2} = f_{e_3} = \sum_{d=1}^3 f_{2e_d} = 0,
\end{equation}
then $f^b = F^b(f)$ with expression \eqref{eq:f_b} also satisfies
\begin{equation}
f_{e_1}^b = f_{e_2}^b = f_{e_3}^b = \sum_{d=1}^3 f_{2e_d}^b = 0.
\end{equation}
\end{proposition}
\begin{proof}
Equation \eqref{eq:f_b_alpha} gives
\begin{equation}
f_{e_1}^b = f_{e_1}, \qquad
f_{e_3}^b = f_{e_3}, \qquad
f_{2e_d}^b = f_{2e_d}, \quad d=1,2,3.
\end{equation}
Thus it only remains to prove $f_{e_2}^b = 0$. According to \eqref
{eq:J_start}, \eqref{eq:p_alpha_simplified} and \eqref{eq:J_H},
$p_{e_2}$ can actually be expressed by
\begin{equation}
p_{e_2} = \rho^W J_0(u_1^W - u_1) \hat{J}_1 J_0(u_3^W - u_3)
= \rho^W [(\theta^W - \theta) \hat{J}_{-1} - \hat{H}_1]
= -\rho^W \sqrt{\frac{\theta^W}{2\pi}}.
\end{equation}
Since $K_2(e_2) = \lceil M/2 \rceil$, the above equation together with
\eqref{eq:rho_W} and \eqref{eq:mnt_bc} immediately gives $f_{e_2}^b =
0$.
\end{proof}
\subsection{Numerical implementation of boundary conditions}
In a finite volume scheme, the boundary conditions is often applied by
ghost cell techniques. Suppose the distribution function of the cell
on the boundary is denoted as \eqref{eq:f}. The distribution function
of the ghost cell can be constructed as follows:
\begin{enumerate}
\setlength\itemsep{0cm}
\item Apply $F^b$ on $f(\bxi)$ and suppose the result is
\eqref{eq:f_b};
\item Construct the ghost cell distribution as
\begin{equation} \label{eq:ghost}
f^{\mathrm{ghost}}(\bxi) = \sum_{|\alpha| \leqslant M + 1}
(2f_{\alpha}^b - f_{\alpha}) \mathcal{H}_{\theta, \alpha} \left(
\frac{\bxi - (2\bu^b - \bu)}{\sqrt{\theta}}
\right).
\end{equation}
\end{enumerate}
Now we consider the time complexity of this operation. Suppose $N_M =
(M+2)(M+3)(M+4)/6$ is the number of moments involved in the boundary
condition. Obviously, \eqref{eq:ghost} requires $O(N_M)$ operations.
For the calculation of $F^b(f)$, we list the cost as follows:
\begin{enumerate}
\setlength\itemsep{0cm}
\item Half-space cut-off of $f$ \eqref{eq:q}: $O(M N_M)$ operations;
\item Calculation of $\rho^W$ \eqref{eq:rho_W}: $O(1)$ operations;
\item Calculation of $p_{\alpha}$ \eqref{eq:p_alpha_simplified}:
$O(N_M)$ operations;
\item Evaluation of \eqref{eq:f_b_alpha}: $O(N_M)$ operations.
\end{enumerate}
Thus, the total computational cost is $O(M N_M)$, while the time
complexity is $O(N_M)$ if no boundary condition is considered.
However, since this procedure only takes place on the boundary, it
produces little increment of the computational time in real
computation.
\begin{remark}
Proposition \ref{prop:conservation} indicates the conservation of
mass on the boundary when using the HLL numerical flux as in
\cite{Cai}. One can find that when $u_2^W = 0$, saying a special
reference coordinate system is used, the minimum and maximum signal
speeds in need of the HLL flux are opposite numbers. Together with
$\rho^{\mathrm{ghost}} = \rho$, $u_2^{\mathrm{ghost}} = -u_2$, the
mass conservation of the HLL scheme follows naturally.
\end{remark}
\section{The Boltzmann equation and the \bNRxx method}
\label{sec:NRxx}
The Boltzmann equation is the basic equation in the kinetic theory,
where a distribution function $f(t, \bx, \bxi)$ is introduced to
provide a statistical description for the motion of molecules. Here
$t\in \bbR^+$ is the time, and $\bx, \bxi \in \bbR^3$ are the position
and velocity of particles. The Boltzmann equation reads
\begin{equation}
\frac{\partial f}{\partial t} + \bxi \cdot \nabla_{\bx} f +
\bF \cdot \nabla_{\bxi} f = Q(f,f),
\end{equation}
where $\bF$ is the acceleration of particles caused by external
forces. The detailed expression of the collision term $Q(f,f)$ is not
presented here due to its complexity, but we stress that $Q(f,f)$
contains a five-dimensional integration which causes great difficulty
in the numerical simulation. Instead, simplified collision models such
as the BGK model \cite{BGK} and the Shakhov model \cite{Shakhov} are
adopted in this paper. These models read:
\begin{enumerate}
\item \it BGK model:
\begin{equation}
\frac{\partial f}{\partial t} + \bxi \cdot \nabla_{\bx} f +
\bF \cdot \nabla_{\bxi} f = \frac{1}{\tau}(f_M - f);
\end{equation}
\item \it Shakhov model:
\begin{equation} \label{eq:Shakhov}
\frac{\partial f}{\partial t} + \bxi \cdot \nabla_{\bx} f +
\bF \cdot \nabla_{\bxi} f =
\frac{1}{\tau} \left\{ \left[ 1 +
\frac{(1 - \mathrm{Pr})(\bxi - \bu) \cdot \bq}{5 \rho \theta^2}
\left( \frac{|\bxi - \bu|^2}{\theta} - 5 \right)
\right] f_M - f \right\}.
\end{equation}
\end{enumerate}
Here $\rho$, $\bu$, $\theta$ and $\bq$ denote the density, mean
velocity, temperature and heat flux respectively, and these
macroscopic variables are related with the distribution function $f$
by
\begin{equation}
\begin{gathered}
\rho = \int_{\bbR^3} f \dd\bxi,
\quad \bu = \frac{1}{\rho} \int_{\bbR^3} \bxi f \dd\bxi, \\
\theta = \frac{1}{3\rho} \int_{\bbR^3} |\bxi - \bu|^2 f \dd\bxi, \quad
\bq = \frac{1}{2} \int_{\bbR^3} |\bxi - \bu|^2 (\bxi - \bu) f \dd\bxi.
\end{gathered}
\end{equation}
Besides, $\tau$ is the relaxation time and $f_M$ is the local
Maxwellian which can be analytically formulated by
\begin{equation}
f_M = \frac{\rho}{(2\pi \theta)^{3/2}}
\exp \left( -\frac{|\bxi - \bu|^2}{2\theta} \right).
\end{equation}
In \eqref{eq:Shakhov}, $\mathrm{Pr}$ stands for the Prandtl number
which is a constant. One can easily observe that if $\mathrm{Pr} = 1$,
then the Shakhov model reduces to the BGK model, which agrees with the
common knowledge that the BGK model predicts an incorrect Prandtl
number $1$.
The \NRxx method is a numerical tool for solving large moment
equations. It originated in \cite{NRxx} and was simplified in \cite
{NRxx_new}. The basic idea is to expand the distribution function $f$
into the Hermite series:
\begin{equation} \label{eq:expansion}
f(t, \bx, \bxi) = \sum_{\alpha \in \bbN^3} f_{\alpha}(t, \bx)
\mathcal{H}_{\theta,\alpha} \left(
\frac{\bxi - \bu(t, \bx)}{\sqrt{\theta(t, \bx)}}
\right),
\end{equation}
where $\mathcal{H}_{\theta,\alpha}$ is the basis function defined as
\begin{equation} \label{eq:H}
\mathcal{H}_{\theta,\alpha}(\bv) =
\prod_{d=1}^3 \frac{1}{\sqrt{2\pi}}
\theta^{-\frac{\alpha_d + 1}{2}}
\He_{\alpha_d}(v_d) \exp\left( -\frac{v_d^2}{2} \right),
\quad \forall \alpha \in \bbN^3,
\end{equation}
and $\He_n$ is the Hermite polynomials
\begin{equation} \label{eq:He}
\He_n(x) = (-1)^n \exp \left( \frac{x^2}{2} \right)
\frac{\mathrm{d}^n}{\mathrm{d} x^n}
\exp \left( -\frac{x^2}{2} \right).
\end{equation}
For convenience, we let $\He_n(x) \equiv 0$ if $n < 0$. Thus $\mathcal
{H}_{\theta,\alpha}(\bv)$ is zero when any of the components of
$\alpha$ is negative.
With the expansion \eqref{eq:expansion}, the coefficients $f_{\alpha}$
can be considered as a set of infinite moments, and we have the
following relations:
\begin{equation} \label{eq:low_order_moments}
\begin{gathered}
f_0 = \rho, \quad f_{e_i} = 0, \quad \sum_{d=1}^3 f_{2e_d} = 0, \\
\sigma_{ij} = f_{e_i + e_j}, \quad \sigma_{ii} = 2 f_{2e_i}, \quad
q_i = 2 f_{3e_i} + \sum_{d=1}^3 f_{2e_d + e_i},
\end{gathered}
\end{equation}
where $i,j = 1,2,3$ and $i \neq j$, and $\sigma_{ij}$ is the stress
tensor or pressure deviators, which can be deduced from the
distribution function $f$ by
\begin{equation}
\sigma_{ij} = p_{ij} - \frac{1}{3} \delta_{ij} \sum_{d=1}^3 p_{dd},
\quad \text{with} \quad
p_{ij} = \int_{\bbR^3} (\xi_i - u_i)(\xi_j - u_j) f \dd \bxi,
\qquad i,j = 1,2,3.
\end{equation}
In order to implement \eqref{eq:expansion} numerically, a positive
integer $M \geqslant 3$ is chosen and only the coefficients $\{%
f_{\alpha}(t, \bx)\}_{|\alpha| \leqslant M}$ are stored. Due to the
absence of higher order moments, the resulting moment system is not
closed. According to \cite{NRxx_new}, the $(M+1)$-st order moments are
approximated by
\begin{equation} \label{eq:reg}
\begin{split}
f_{\alpha} &= \tau \Bigg\{
\frac{1}{\rho} \sum_{j=1}^D
\frac{\partial (\rho \theta)}{\partial x_j} f_{\alpha-e_j}
+ \frac{\theta}{D} \left(
\sum_{j=1}^D \frac{\partial u_j}{\partial x_j}
\right) \sum_{d=1}^D f_{\alpha-2e_d}
- \sum_{j=1}^D \Bigg[ \theta
\frac{\partial f_{\alpha-e_j}}{\partial x_j} \\
& \qquad {} + \sum_{d=1}^D \left(
\frac{\partial u_d}{\partial x_j} \theta f_{\alpha-e_d-e_j}
+ \frac{1}{2} \frac{\partial \theta}{\partial x_j}
(\theta f_{\alpha-2e_d-e_j} + (\alpha_j+1) f_{\alpha-2e_d+e_j})
\right)
\Bigg] \Bigg\}.
\end{split}
\end{equation}
Here $f_{\alpha}$ is taken as zero when any of $\alpha$'s components
is negative. The numerical scheme for the force-free BGK model has
been constructed in \cite{Cai} based on the finite volume scheme with
linear reconstruction and the fractional step method. Suppose the
problem is in 1D and the grid is uniform with cell size $\Delta x$. We
denote the cell centers as $x_j$, and then a full time step of the
scheme can be sketched as follows:
\begin{enumerate}
\setlength\itemsep{0cm}
\item Determine the time step size $\Delta t$.
\item Reconstruct the first $M$-th order moments for the distribution
functions on cell boundaries $x_{j \pm 1/2}$ with a conservative
linear reconstruction.
\item Get the $(M+1)$-st order moments for the distribution functions
on cell boundaries with a direct discretization of \eqref{eq:reg}.
\item \label{item:fvm} Apply the HLL scheme to solve the purely
advective equation $\partial_t f + \bxi \cdot \nabla_{\bx} f = 0$
over a time step of length $\Delta t$.
\item \label{item:BGK} Analytically solve the pure collision equation
of the BGK model $\partial_t f = (f_M - f) / \tau$ over a time step
of length $\Delta t$.
\end{enumerate}
We refer the readers to \cite{NRxx, NRxx_new, Cai} for details of the
algorithm. Here we only note that the Step \ref{item:fvm} is
nontrivial since two distributions cannot be added up directly, and in
Step \ref{item:BGK}, the reason why the collision-only equation can be
directly solved is that $f_M$ can be expressed in the Hermite series
$\{\mathcal{H}_{\theta,\alpha}\}$ trivially as $f_M = f_0 \mathcal{H}
_{\theta,0} \left( (\bxi - \bu) / \sqrt{\theta} \right)$.
\section{The \bNRxx method for Shakhov model with force terms}
\label{sec:Shakhov}
As is well known, the Prandtl number for monatomic gases is around
$2/3$, while the BGK model gives a Prandtl number $1$, which causes
incorrect prediction of the stress tensor $\sigma_{ij}$ or heat flux
$\bq$ for a dense gas. As a remedy, the Shakhov model was introduced
in \cite{Shakhov} as a generalization of the BGK model. The difference
between these two models has been investigated in \cite{Yang,
Kudryavtsev}. In this section, we extend the \NRxx method in
\cite{NRxx_new} to the Shakhov model, and the force terms in
\eqref{eq:Shakhov} is added.
\subsection{The governing equations} \label{sec:gov_eq}
The moment system for the Shakhov model \eqref{eq:Shakhov} with moment
set $\{f_{\alpha}(t,\bx)\}_{|\alpha| \leqslant M}$ will be deduced
here. As in \cite{NRxx_new}, the strategy is to expand
\eqref{eq:Shakhov} into Hermite series, and then match the
coefficients for the same basis functions. In order to simplify the
notation, we define
\begin{equation}
\begin{split}
A &= \frac{\partial f}{\partial t} + \bxi \cdot \nabla_{\bx} f, \\
B &= \bF \cdot \nabla_{\bxi} f, \\
C &= \frac{1}{\tau} \left\{ \left[
1 + \frac{(1 - \mathrm{Pr})(\bxi - \bu) \cdot \bq}{5 \rho \theta^2}
\left( \frac{|\bxi - \bu|^2}{\theta} - 5 \right)
\right] f_M - f \right\}.
\end{split}
\end{equation}
It has been deduced in \cite{NRxx_new} that the Hermite expansion of
$A$ is
\begin{equation} \label{eq:A}
\begin{split}
A &= \sum_{\alpha \in \bbN^3} \Bigg\{ \left(
\frac{\partial f_{\alpha}}{\partial t} +
\sum_{d=1}^3 \frac{\partial u_d}{\partial t} f_{\alpha-e_d} +
\frac{1}{2} \frac{\partial \theta}{\partial t}
\sum_{d=1}^3 f_{\alpha-2e_d}
\right) \\
& \qquad + \sum_{j=1}^3 \Bigg[ \left(
\theta \frac{\partial f_{\alpha - e_j}}{\partial x_j} +
u_j \frac{\partial f_{\alpha}}{\partial x_j} +
(\alpha_j + 1) \frac{\partial f_{\alpha+e_j}}{\partial x_j}
\right) \\
& \qquad \qquad + \sum_{d=1}^3 \frac{\partial u_d}{\partial x_j}
\left( \theta f_{\alpha-e_d-e_j} + u_j f_{\alpha-e_d}
+ (\alpha_j + 1) f_{\alpha-e_d+e_j} \right) \\
& \qquad \qquad + \frac{1}{2} \frac{\partial \theta}{\partial x_j}
\sum_{d=1}^3 \left(
\theta f_{\alpha-2e_d-e_j} + u_j f_{\alpha-2e_d}
+ (\alpha_j + 1) f_{\alpha-2e_d+e_j}
\right)
\Bigg] \Bigg\} \mathcal{H}_{\theta,\alpha} \left(
\frac{\bxi - \bu}{\sqrt{\theta}}
\right).
\end{split}
\end{equation}
Using the differential relation of the Hermite polynomials, we have
\begin{equation}
\frac{\partial}{\partial \xi_d} \mathcal{H}_{\theta,\alpha} \left(
\frac{\bxi - \bu}{\sqrt{\theta}}
\right) = -\mathcal{H}_{\theta,\alpha+e_d} \left(
\frac{\bxi - \bu}{\sqrt{\theta}}
\right).
\end{equation}
Thus the Hermite expansion of the force term $B$ can be easily deduced
as
\begin{equation} \label{eq:B}
B = -\sum_{\alpha \in \bbN^3} \sum_{d=1}^3 F_d f_{\alpha-e_d}
\mathcal{H}_{\theta,\alpha} \left(
\frac{\bxi - \bu}{\sqrt{\theta}}
\right).
\end{equation}
The expansion of the collision term $C$ can also be obtained by direct
calculation. The result is
\begin{equation} \label{eq:C}
C = \frac{1}{\tau} \left[
\frac{1 - \mathrm{Pr}}{5} \sum_{i=1}^3 \sum_{j=1}^3
q_i \mathcal{H}_{\theta, e_i + 2e_j} \left(
\frac{\bxi - \bu}{\sqrt{\theta}}
\right)
- \sum_{|\alpha| \geqslant 2} f_{\alpha}
\mathcal{H}_{\theta, \alpha} \left(
\frac{\bxi - \bu}{\sqrt{\theta}}
\right)
\right].
\end{equation}
Putting \eqref{eq:A}\eqref{eq:B} and \eqref{eq:C} into the
Boltzmann-Shakhov equation $A+B=C$ and extracting coefficients for all
basis functions, with a slight rearrangement, we get the following
general moment equations for Shakhov model:
\begin{equation} \label{eq:mnt_eq}
\begin{split}
& \frac{\partial f_{\alpha}}{\partial t} +
\sum_{d=1}^3 \left(
\frac{\partial u_d}{\partial t} +
\sum_{j=1}^3 u_j \frac{\partial u_d}{\partial x_j} - F_d
\right) f_{\alpha-e_d} + \frac{1}{2} \left(
\frac{\partial \theta}{\partial t} +
\sum_{j=1}^3 u_j \frac{\partial \theta}{\partial x_j}
\right) \sum_{d=1}^3 f_{\alpha-2e_d} \\
& \quad + \sum_{j,d=1}^3 \left[
\frac{\partial u_d}{\partial x_j} \left(
\theta f_{\alpha-e_d-e_j} + (\alpha_j + 1) f_{\alpha-e_d+e_j}
\right) + \frac{1}{2} \frac{\partial \theta}{\partial x_j} \left(
\theta f_{\alpha-2e_d-e_j} + (\alpha_j + 1) f_{\alpha-2e_d+e_j}
\right)
\right] \\
& \quad + \sum_{j=1}^3 \left(
\theta \frac{\partial f_{\alpha - e_j}}{\partial x_j} +
u_j \frac{\partial f_{\alpha}}{\partial x_j} +
(\alpha_j + 1) \frac{\partial f_{\alpha+e_j}}{\partial x_j}
\right) = \frac{1}{\tau} \left(
\frac{1 - \mathrm{Pr}}{5} \sum_{i,j=1}^3
\delta_{ij}(\alpha) q_i - \delta(\alpha) f_{\alpha}
\right),
\end{split}
\end{equation}
where $\delta_{ij}(\alpha)$ and $\delta(\alpha)$ are defined by
\begin{equation}
\delta_{ij}(\alpha) = \left\{ \begin{array}{ll}
1, & \text{if } \alpha = e_i + 2e_j, \\
0, & \text{otherwise,}
\end{array} \right. \qquad
\delta(\alpha) = \left\{ \begin{array}{ll}
1, & \text{if } |\alpha| \geq 2, \\
0, & \text{otherwise.}
\end{array} \right.
\end{equation}
Now we will explore something more from \eqref{eq:mnt_eq}. Noting
that $f_{e_j} = 0$, $\forall j = 1,2,3$, the following relation can be
obtained if we put $\alpha = 0$ into \eqref{eq:mnt_eq}:
\begin{equation}
\frac{\partial f_0}{\partial x_j} +
\sum_{j=1}^3 \left(
u_j \frac{\partial f_0}{\partial x_j} +
f_0 \frac{\partial u_j}{\partial x_j}
\right) = 0.
\end{equation}
This is the mass conservation law. If we set $\alpha = e_d$, $d =
1,2,3$, the equations are
\begin{equation}
f_0 \left(
\frac{\partial u_d}{\partial t} +
\sum_{j=1}^3 u_j \frac{\partial u_d}{\partial x_j} - F_d
\right) + f_0 \frac{\partial \theta}{\partial x_d}
+ \theta \frac{\partial f_0}{\partial x_d}
+ \sum_{j=1}^3 (\delta_{jd} + 1)
\frac{\partial f_{e_d + e_j}}{\partial x_j} = 0.
\end{equation}
This equation can be simplified as
\begin{equation} \label{eq:mtm}
f_0 \left(
\frac{\partial u_d}{\partial t} +
\sum_{j=1}^3 u_j \frac{\partial u_d}{\partial x_j} - F_d
\right) + \sum_{j=1}^3 \frac{\partial p_{jd}}{\partial x_j} = 0.
\end{equation}
Now we consider the case of $|\alpha| \geqslant 2$. Substituting
\eqref{eq:mtm} into \eqref{eq:mnt_eq}, the temporal differentiation of
$\bu$ can be eliminated. In order to eliminate the temporal
differentiation of $\theta$, we multiply \eqref{eq:Shakhov} by $|\bxi
- \bu|^2$ on both sides and then integrate on $\bbR^3$ with respect to
$\bxi$. The result is
\begin{equation} \label{eq:energy}
f_0 \left( \frac{\partial \theta}{\partial t}
+ \sum_{j=1}^3 u_j \frac{\partial \theta}{\partial x_j} \right)
+ \frac{2}{3} \sum_{j=1}^3 \left(
\frac{\partial q_j}{\partial x_j} +
\sum_{d=1}^3 p_{jd} \frac{\partial u_d}{\partial x_j}
\right) = 0.
\end{equation}
Note that the force term does not appear in this equation, since
\begin{equation}
\int_{\bbR^3} |\bxi-\bu|^2 \frac{\partial f}{\partial \xi_j}
\dd \bxi = -2 \int_{\bbR^3} (\xi_j - u_j) f \dd \bxi = 0.
\end{equation}
Thus, the final form of equations for $|\alpha| \geqslant 2$ reads
\begin{equation} \label{eq:mnt_system}
\begin{split}
& \frac{\partial f_{\alpha}}{\partial t} - \frac{1}{f_0}
\sum_{d=1}^3 \sum_{j=1}^3
\frac{\partial p_{jd}}{\partial x_j} f_{\alpha-e_d}
- \frac{1}{3f_0} \sum_{j=1}^3 \left(
\frac{\partial q_j}{\partial x_j} +
\sum_{d=1}^3 p_{jd} \frac{\partial u_d}{\partial x_j}
\right) \sum_{d=1}^3 f_{\alpha-2e_d} \\
& \quad + \sum_{j,d=1}^3 \left[
\frac{\partial u_d}{\partial x_j} \left(
\theta f_{\alpha-e_d-e_j} + (\alpha_j + 1) f_{\alpha-e_d+e_j}
\right) + \frac{1}{2} \frac{\partial \theta}{\partial x_j} \left(
\theta f_{\alpha-2e_d-e_j} + (\alpha_j + 1) f_{\alpha-2e_d+e_j}
\right)
\right] \\
& \quad + \sum_{j=1}^3 \left(
\theta \frac{\partial f_{\alpha - e_j}}{\partial x_j} +
u_j \frac{\partial f_{\alpha}}{\partial x_j} +
(\alpha_j + 1) \frac{\partial f_{\alpha+e_j}}{\partial x_j}
\right) = \frac{1}{\tau} \left(
\frac{1 - \mathrm{Pr}}{5} \sum_{i,j=1}^3
\delta_{ij}(\alpha) q_i - \delta(\alpha) f_{\alpha}
\right).
\end{split}
\end{equation}
In order to get a closed system, we collect \eqref{eq:reg},
\eqref{eq:mtm}, \eqref{eq:energy} and \eqref{eq:mnt_system} with $2
\leqslant |\alpha| \leqslant M$ together. Then the governing system
for the \NRxx method with Shakhov model and force terms is formed.
\begin{remark} \label{rem:reg}
In the Shakhov model, the equation \eqref{eq:reg}, the prediction of
$f_{\alpha}$ with $|\alpha| = M + 1$ derived for the BGK model, is
still available. In \cite{NRxx_new}, \eqref{eq:reg} is deduced in the
following two steps:
\begin{enumerate}
\item Determine the orders of magnitude for all $f_{\alpha}$ using
Maxwellian iteration.
\item \label{item:remove_hot} For $|\alpha| = M + 1$, remove all the
high order terms in the equations containing only $-f_{\alpha} /
\tau$ in their right hand sides.
\end{enumerate}
The Maxwellian iteration can also be applied to \eqref{eq:mnt_system},
and after the first iteration step, we immediately get
\begin{equation}
f_{\alpha} = O(\tau), \qquad |\alpha| = 2, \quad \text{or} \quad
\alpha = e_i + 2e_j, \quad i,j=1, 2, 3,
\end{equation}
and other moments with $|\alpha| \geqslant 2$ remain to be zero.
This result is the same as that we have derived in the BGK model.
We note that for $|\alpha| > 3$ and $\alpha = (1,1,1)$,
\eqref{eq:mnt_system} is just the corresponding equation for the BGK
model. Thus, in the view of order of magnitude, the subsequent
iterations are identical to the BGK case. Moreover, when $M \geqslant
3$, which we have assumed in the last section, step
\ref{item:remove_hot} is also identical for both models. Hence
\eqref{eq:reg} still applies for the Shakhov model.
\end{remark}
\subsection{The numerical approach}
The acceleration $\bF$ only appears in \eqref{eq:mtm} in the governing
system, thus a splitting method can be applied as follows:
\begin{enumerate}
\setlength\itemsep{0cm}
\item \emph{Transportation:} solve the force-free Shakhov equation
over a time step of length $\Delta t$.
\item \emph{Acceleration:} solve $\partial_t \bu = \bF$ over a time
step of length $\Delta t$.
\end{enumerate}
In order to solve the force-free Shakhov equation, another splitting
of the convection and collision part is needed. For the convection
part, the method is identical to that used in the BGK model. We refer
the readers to \cite{NRxx,Cai} for details. For the collision part,
since a new collision model is adopted, the procedure is slightly
different.
Now we consider the pure collision model, where $\rho, \bu$ and
$\theta$ are not changed while time evolves. Therefore, the collision
terms only exist in \eqref{eq:mnt_system} with $|\alpha| \geqslant 2$.
Two cases are considered below:
(1) $\alpha = e_i + 2e_j$, $i,j = 1,2,3$. In these cases, the pure
collision equations are written as
\begin{equation} \label{eq:coll}
\begin{split}
\frac{\partial f_{e_i + 2e_j}}{\partial t} &=
\frac{(1-\mathrm{Pr}) q_i - 5 f_{e_i + 2e_j}}{5 \tau} \\
&= \frac{1}{\tau} \left[
\frac{1-\mathrm{Pr}}{5} \left(
2f_{3e_i} + \displaystyle\sum_{j=1}^3 f_{e_i + 2e_j}
\right) - f_{e_i + 2e_j}
\right], \qquad i,j = 1,2,3.
\end{split}
\end{equation}
In the general case, $\tau$ only depends on $\rho$ and $\theta$. Thus
it is invariant in the collision-only system. This turns
\eqref{eq:coll} into a linear \emph{ordinary} differential system with
9 equations, which can be analytically integrated as
\begin{equation} \label{eq:sol1}
f_{e_i + 2e_j}(t) = \frac{1}{5} q_i(t_0) \exp \left(
-\frac{\mathrm{Pr} (t - t_0)}{\tau}
\right) - \left( \frac{1}{5} q_i(t_0) - f_{e_i+2e_j}(t_0) \right)
\exp \left( -\frac{t - t_0}{\tau} \right),
\end{equation}
where $t_0$ denotes the initial time.
(2) Other cases. For other $\alpha$'s, the collision-only equation is
the same as the BGK model:
\begin{equation}
\frac{\partial f_{\alpha}}{\partial t} = -\frac{1}{\tau} f_{\alpha}.
\end{equation}
The solution is
\begin{equation} \label{eq:sol2}
f_{\alpha}(t) = f_{\alpha}(t_0)
\exp \left( -\frac{t - t_0}{\tau} \right).
\end{equation}
When \eqref{eq:sol1} and \eqref{eq:sol2} are used in the numerical
scheme, we replace $t$ and $t_0$ with $t_{n+1}$ and $t_n$
respectively. Note that when $\tau$ is independent of $\bu$, the
acceleration and collision do not coupled with each other, thus the
splitting is applied only once rather than twice. This makes it more
efficient when the Strang splitting is employed.
\section{Introduction}
In the kinetic theory, the degree of rarefaction of a gas is often
characterized by the dimensionless Knudsen number $\Kn = \lambda / L$,
where $\lambda$ is the mean free path and $L$ is the relevant
characteristic length. The classic Navier-Stokes-Fourier (NSF)
equations are accurate only when $\Kn < 0.01$. However, the ongoing
miniaturization of technical devices requires modelling of gas in
microscopic channels, for which the characteristic length $L$ is so
small that even under normal density and temperature, the Knudsen
number is beyond the available region of NSF equations. Meanwhile, in
the transitional regime ($0.1 < \Kn < 10$), the traditional no-slip
wall boundary condition is no longer valid. In order to match the
physical experimentation, the interaction between wall and gas should
be carefully conducted. We refer the readers to \cite{Karniadakis} for
more details.
For microflows, it is known that the Boltzmann equation with Maxwell
boundary conditions \cite{Maxwell} is able to accurately describe the
flow state. However, on the computational perspective, the cost for
solving Boltzmann equation directly is unacceptable in the general
case. Grad \cite{Grad} did a pioneer work which extends Euler
equations to a thirteen-moment system, which opened a new way for
modelling rarefied gas flow called as moment method. However, it was
discovered by Grad himself in \cite{Grad1952} that this system fails
to give smooth shock profiles when the Mach number is larger than
$1.65$. To remedy this drawback, some authors tend to construct a
parabolic system similar to the NSF equations. In this field, some
methods such as Jin-Slemrod \cite{Jin}, COET \cite{Reitebuch} and R13
\cite{Grad1958, Struchtrup2003} were subsequently raised.
Concurrently, increasing attention is attracted to systems with more
than 13 moments (e.g. \cite{Shock, Struchtrup2002}). As a combination
of these two directions, R20 and R26 equations were respectively
studied in \cite{Mizzi} and \cite{Emerson}. In \cite{NRxx}, a
general method for numerically solving the regularized moment
equations of arbitrary order was proposed, and it was improved in
\cite{NRxx_new, Cai} and abbreviated as \NRxx method in \cite{Cai} for
convenience. On the other hand, the boundary condition for the moment
methods is a major obstacle for applications of moment methods in the
field of microflows. In Grad's paper \cite{Grad}, the basic idea for
the modelling of Maxwell boundary conditions in the framework of
moment method is raised. The idea was also used in
\cite{Torrilhon2008, Emerson} for R13 and R26 equations. However, for
general moment equations, the numerical method to process the boundary
conditions for \NRxx method is unavailable yet.
The major concern of this paper is to supply suitable boundary
conditions for \NRxx method. Before that, the \NRxx method is first
improved such that it is able to predict stress and heat flux
correctly in the dense case. This is achieved by replacing the BGK
collision model \cite{BGK} used in \cite{NRxx, NRxx_new, Cai} with the
Shakhov model \cite{Shakhov}. Recall that for the BGK model, collision
term can be analytically solved when using the \NRxx method.
Similarly, analytical solution for each moment can also be obtained
when using the Shakhov model. At the same time, the force term is also
applied to the \NRxx method, and one can find that this term only
affects the momentum equation that it is turned to be trivial when
splitting method is employed.
As to the wall boundary conditions, we follow the idea of Grad
\cite{Grad} and try to approximate Maxwell boundary condition using
moment method. The Maxwell boundary condition is a linear combination
of specular reflection and diffusive reflection. According to Grad's
theory, only the moments of odd order in the normal microscopic
velocity are controlled by boundary conditions. These moments for the
specularly reflective part vanish. For the diffusive reflection, the
incidence part and the emergence part are considered separately. For
the incidence part, one need to calculate the moments of a
distribution cut off by a half space. Since the distribution is
expressed by a finite expansion of Hermite series, the cut-off turns
out to be quite intricate. We eventually derive a simple recursive
formula to obtain these moments with careful investigation into the
detailed expressions. The obtained formula brings only slight
increment of the computational cost. For the emergence part, which is
a half Maxwellian, the moments are obtained by direct integration, and
the result is also given in a recursive form. The overall boundary
condition is the summation of both the specular part and diffusive
part, which is rearranged into a simple formulation. It is numerically
implemented by first constructing a set of moments satisfying the
boundary conditions, and then approximating the flow state in the
ghost cell with a first order extrapolation of each moment. Thus,
boundary conditions for the \NRxx method of all orders are collected
into a uniform framework, which avoids separate and involved
implementation for different systems with sophisticated expressions
\cite{Thatcher, Torrilhon2008, Emerson}.
A number of numerical examples are presented to show the validity of
the boundary conditions. Both steady and unsteady problems are
studied. Numerical simulations up to $455$-moment system are carried
out. The classic symmetric planar Couette flow and force-driven
Poiseuille flow are investigated as examples for steady problems. All
the numerical results exhibit the convergence of the \NRxx method as
the number of moments increases.
The layout of this paper is as follows: in Section \ref{sec:NRxx}, we
give a brief introduction to the Boltzmann equation and the \NRxx
method. In Section \ref{sec:Shakhov}, the Shakhov collision model and
the force-induced acceleration terms are coupled with the \NRxx
method. In Section \ref{sec:BC}, the derivation of boundary conditions
are carried out. Numerical examples are shown in Section
\ref{sec:example}, and some discussions on the validity and accuracy of
the \NRxx method are given in Section \ref{sec:discussion}. Finally,
we make some conclusion in Section \ref{sec:conclusion}.
\section*{Appendix}
\section{Concluding remarks} \label{sec:conclusion} A uniform
numerical scheme for coupling the \NRxx method and the wall boundary
conditions are developed in this paper, and the \NRxx method is
extended to apply the force term and predict correct Prandtl number by
using the Shakhov collision model. To validate the proposed method,
both steady and unsteady problems are simulated. We are currently
working on applying the \NRxx method to 2D problems.
\section*{Acknowledgements}
We thank Dr. Vladimir Titarev for some useful discussions on the
implementation of conservative discrete velocity method for Shakhov
collision model. The research of the second author was supported in
part by the National Basic Research Program of China (2011CB309704),
the National Science Foundation of China under grant 10731060 and NCET
in China. Z. Qiao is partially supported by the Hong Kong RGC grant
HKBU201710.
\section{Some discussions on the \bNRxx method} \label{sec:discussion}
\subsection{Order of accuracy}
For the macroscopic equations, a basic quantity describing its ability
is the \emph{order of accuracy} with respect the Knudsen number. The
definition of \emph{order of accuracy} can be found in textbook \cite
{Struchtrup}:
\begin{quote} \it%
A set of equations is said to be accurate of order $\lambda_0$, when
the pressure deviator $\sigma_{ij}$ and the heat flux $q_i$ are known
within the order $O(\varepsilon^{\lambda_0})$.
\end{quote}
Here $\varepsilon$ is a small parameter proportional to the relaxation
time $\tau$. As we have discussed in Remark \ref{rem:reg}, in the
view of order of magnitude, the process of Maxwellian iteration for
the Shakhov model is identical to the BGK model. Hence, for an
arbitrary $M \geqslant 3$, the leading order term of $f_{\alpha}$ with
$|\alpha| = M + 1$ is known from the corresponding moment equations
(see \cite{NRxx_new} for details). And it has been deduced in \cite
{NRxx_new} that $f_{\alpha} \sim O(\tau^{\lceil |\alpha| / 3 \rceil})$
for all $|\alpha| \geqslant 4$. Thus, from the analytical form of the
moment equations \eqref{eq:mnt_system}, we immediately have that
$f_{\alpha}$ with $|\alpha| = M$ is known up to $(\lceil (M + 1) / 3
\rceil + 1)$-th order. Subsequently, $f_{\alpha}$ with $|\alpha| = M -
1$ is know up to $(\lceil (M + 1) / 3 \rceil + 2)$-th order, and this
can be done until $|\alpha| = 2$. The general result is
\begin{proposition}
For the moment equations described in Section \ref{sec:gov_eq},
$f_{\alpha}$ has $(\lceil 4(M + 1) / 3 \rceil - |\alpha|)$-th order
accuracy if $2 \leqslant |\alpha| \leqslant M$.
\end{proposition}
\noindent Now, using \eqref{eq:low_order_moments} and the definition
of order of accuracy, we conclude that the \NRxx equations have the
order of accuracy $\lceil (4M - 5) / 3 \rceil$.
For boundary value problems, such discussion is only valid in the
bulk. In the Knudsen layer, which is known to be of width $O(\Kn)$,
we need to use $X = x / \Kn$ as the spatial variable while
investigating the accuracy of moment equations. In this case, if we
consider a steady state problem, the small parameter no longer
appears in the governing equations \eqref{eq:mnt_system}. This means
the order of magnitude for $f_{\alpha}$ does not increase as
$|\alpha|$ increases, as has been found in \cite{Struchtrup2007}.
\subsection{The validity of \bNRxx method for large Knudsen number and
in the Knudsen layer}
As we have discussed above, there are two cases when the order of
accuracy is not so meaningful for describing the accuracy of the \NRxx
method:
\begin{enumerate}
\item In the case of $\Kn \sim O(1)$, there are no ``small
parameters'' in our concept.
\item In the Knudsen layer, the orders of magnitude of moments do not
increase as they behave in the bulk.
\end{enumerate}
Nevertheless, we can still consider the \NRxx method as a solver for
Boltzmann equation with spectral expansion in the velocity space, and
the method should be valid when $f_{\alpha}$ decays sufficiently fast
as $|\alpha|$ increases. Now we follow \cite{Struchtrup2007} and give
the average absolute values of the moments with the same order for
different $\Kn$ and different $M$ in Figure \ref{fig:avg_abs_value}.
The result is based on the Couette flow problem in Section
\ref{sec:Couette}, and the \NRxx solution at $x = -0.5$ is used in the
these plots.
\begin{figure}[!ht]
\psfrag{M = 3}{\footnotesize{$M=3$}}
\psfrag{M = 6}{\footnotesize{$M=6$}}
\psfrag{M = 9}{\footnotesize{$M=9$}}
\centering
\subfigure[$\Kn=0.1$]{
\includegraphics[scale=.6]{Couette_abs_value_Kn=0.1.eps}
} \\
\subfigure[$\Kn=1.0$]{
\includegraphics[scale=.6]{Couette_abs_value_Kn=1.0.eps}
} \\
\subfigure[$\Kn=10.0$]{
\includegraphics[scale=.6]{Couette_abs_value_Kn=10.0.eps}
}
\caption{Average values of $\{|f_{\alpha}|\}_{|\alpha|=k}$ for $k=1$
to $9$. The results are based the \NRxx solutions of the Couette
flow.}
\label{fig:avg_abs_value}
\end{figure}
In these figures, we find that even when the Knudsen number is as
large as $10$, the magnitudes of moments still decay very fast. Thus,
the \NRxx method can still be considered to be valid and efficient.
Although the methodology of regularization which is based on a small
$\tau$ is not valid any more, the regularization term \eqref{eq:reg}
is simply a prediction of higher order moments. Such prediction
differs from Grad equations' guess $f_{\alpha} = 0$, but also has a
uniform expression for all Knudsen numbers. When $M$ goes to infinity,
the regularization term is expected to vanish since $f_{\alpha}$
decays. On the other hand, this term smooths the profiles of the
macroscopic variables, thus avoids the appearance of some unphysical
phenomena such as subshocks (see \cite{NRxx_new}). This indicates the
meaningfulness of regularization for practical use.
\section{Numerical examples} \label{sec:example} In this section,
three numerical examples are presented to validate our algorithm. In
all these examples, a hard sphere gas is assumed, for which the
relaxation time is defined as
\begin{equation} \label{eq:tau}
\tau = \frac{5}{16} \sqrt{\frac{2\pi}{\theta}} \frac{\Kn}{\rho}
\end{equation}
following \cite{Bird}, where $\Kn$ is the Knudsen number. The CFL
number is always $0.95$. And for all the tests, the wall is set to be
a fully diffusive one ($\chi = 1$) with $\theta^W = 1$. The POSIX
multithreading technique is utilized in our simulation, and at most 8
CPU cores are used.
\subsection{The beginning of a shock wave's formation} \label{sec:reflection}
The first example is a simulation of the interaction of a coming flow
with a diffusive wall. The computational domain is $[-5,0]$ and the
global Knudsen number $\Kn$ used in \eqref{eq:tau} is set to be $0.5$.
The left boundary is a free boundary, and the right is a stationary
diffusive wall parallel to the $xz$-plane. The initial condition is
given by
\begin{equation}
\rho_0(y) = 1.0, \quad \bu_0(y) = (0, 0.5, 0)^T,
\quad \theta_0(y) = 1.0, \qquad \forall y \in [-5, 0],
\end{equation}
and the gas is in equilibrium everywhere. A left-going shock wave will
form after a sufficiently long time. Here we stop the computation at
$t=1.0$ in order to check the validity of the boundary condition. For
a reference solution, we solve the Shakhov equation \eqref{eq:Shakhov}
directly using a Conservative Discrete Velocity Method (CDVM)
introduced in \cite{Titarev}. For the computation of both \NRxx method
and CDVM, a uniform mesh with $500$ grids are used to discretize the
domain. For CDVM, the computational velocity domain is $[-10, 10]
\times [-10, 10] \times [-10, 10]$ and discretized by $50 \times 100
\times 50$ grids.
Figure \ref{fig:Ref_rho_theta} and \ref{fig:Ref_sigma_q} are the
results for CDVM and \NRxx method for $M=3$ to $12$. Only the part $y
\in [-3, 0]$ is shown since all variables for the remaining part are
almost constant. Since a large Knudsen number is considered,
predictions from lower order moment equations give very large
deviations, so the necessity of high order moment theory is obvious.
As the number of moments increases, all profiles get closer and closer
to the results of CDVM. When $M$ reaches $11$, the density and
temperature plots agree with the CDVM results very well, and the errors
in $\sigma_{22}$ and $q_2$ are much smaller than the low order cases,
though it is still observable. It is reasonable that higher order
moments converge more slowly than lower order moments, which is also
observed in \cite{Emerson}.
\begin{figure}[!ht]
\centering
\psfrag{dashed, NRxx}{\footnotesize{$\rho$, \NRxx}}
\psfrag{dashed, CDVM}{\footnotesize{$\rho$, CDVM}}
\psfrag{solid, NRxx}{\footnotesize{$\theta$, \NRxx}}
\psfrag{solid, CDVM}{\footnotesize{$\theta$, CDVM}}
\subfigure[$M=3$]{
\includegraphics[width=.45\textwidth]{Ref_rho_theta_M=3.eps}
}
\subfigure[$M=4$]{
\includegraphics[width=.45\textwidth]{Ref_rho_theta_M=4.eps}
}
\subfigure[$M=5$]{
\includegraphics[width=.45\textwidth]{Ref_rho_theta_M=5.eps}
}
\subfigure[$M=6$]{
\includegraphics[width=.45\textwidth]{Ref_rho_theta_M=6.eps}
}
\subfigure[$M=7$]{
\includegraphics[width=.45\textwidth]{Ref_rho_theta_M=7.eps}
}
\subfigure[$M=8$]{
\includegraphics[width=.45\textwidth]{Ref_rho_theta_M=8.eps}
}
\caption{Density and temperature plots for the problem in section
\ref{sec:reflection}. The left axis is for the dashed lines, and the
right axis is for the solid lines (to be continued).}
\label{fig:Ref_rho_theta}
\end{figure}
\addtocounter{figure}{-1}
\begin{figure}[!ht]
\centering
\psfrag{dashed, NRxx}{\footnotesize{$\rho$, \NRxx}}
\psfrag{dashed, CDVM}{\footnotesize{$\rho$, CDVM}}
\psfrag{solid, NRxx}{\footnotesize{$\theta$, \NRxx}}
\psfrag{solid, CDVM}{\footnotesize{$\theta$, CDVM}}
\setcounter{subfigure}{6}
\subfigure[$M=9$]{
\includegraphics[width=.45\textwidth]{Ref_rho_theta_M=9.eps}
}
\subfigure[$M=10$]{
\includegraphics[width=.45\textwidth]{Ref_rho_theta_M=10.eps}
}
\subfigure[$M=11$]{
\includegraphics[width=.45\textwidth]{Ref_rho_theta_M=11.eps}
}
\subfigure[$M=12$]{
\includegraphics[width=.45\textwidth]{Ref_rho_theta_M=12.eps}
}
\caption{Density and temperature plots for the problem in section
\ref{sec:reflection}. The left axis is for the dashed lines, and the
right axis is for the solid lines.}
\end{figure}
\begin{figure}[!ht]
\centering
\psfrag{dashed, NRxx}{\scalebox{0.8}{$\sigma_{22}$, \NRxx}}
\psfrag{dashed, CDVM}{\scalebox{0.8}{$\sigma_{22}$, CDVM}}
\psfrag{solid, NRxx}{\scalebox{0.8}{$q_2$, \NRxx}}
\psfrag{solid, CDVM}{\scalebox{0.8}{$q_2$, CDVM}}
\subfigure[$M=3$]{
\includegraphics[width=.45\textwidth]{Ref_sigma_q_M=3.eps}
}
\subfigure[$M=4$]{
\includegraphics[width=.45\textwidth]{Ref_sigma_q_M=4.eps}
}
\subfigure[$M=5$]{
\includegraphics[width=.45\textwidth]{Ref_sigma_q_M=5.eps}
}
\subfigure[$M=6$]{
\includegraphics[width=.45\textwidth]{Ref_sigma_q_M=6.eps}
}
\subfigure[$M=7$]{
\includegraphics[width=.45\textwidth]{Ref_sigma_q_M=7.eps}
}
\subfigure[$M=8$]{
\includegraphics[width=.45\textwidth]{Ref_sigma_q_M=8.eps}
}
\caption{Stress and heat flux plots for the problem in section
\ref{sec:reflection}. The left axis is for the dashed lines, and the
right axis is for the solid lines (to be continued).}
\label{fig:Ref_sigma_q}
\end{figure}
\addtocounter{figure}{-1}
\begin{figure}[!ht]
\centering
\psfrag{dashed, NRxx}{\scalebox{0.8}{$\sigma_{22}$, \NRxx}}
\psfrag{dashed, CDVM}{\scalebox{0.8}{$\sigma_{22}$, CDVM}}
\psfrag{solid, NRxx}{\scalebox{0.8}{$q_2$, \NRxx}}
\psfrag{solid, CDVM}{\scalebox{0.8}{$q_2$, CDVM}}
\setcounter{subfigure}{6}
\subfigure[$M=9$]{
\includegraphics[width=.45\textwidth]{Ref_sigma_q_M=9.eps}
}
\subfigure[$M=10$]{
\includegraphics[width=.45\textwidth]{Ref_sigma_q_M=10.eps}
}
\subfigure[$M=11$]{
\includegraphics[width=.45\textwidth]{Ref_sigma_q_M=11.eps}
}
\subfigure[$M=12$]{
\includegraphics[width=.45\textwidth]{Ref_sigma_q_M=12.eps}
}
\caption{Stress and heat flux plots for the problem in section
\ref{sec:reflection}. The left axis is for the dashed lines, and the
right axis is for the solid lines.}
\end{figure}
\subsection{Planar Couette flow} \label{sec:Couette}
The planar Couette flow is a classic benchmark test in the field of
microflows. The moment method for this problem has been investigated
in a lot of papers such as \cite{Tallec, Reitebuch1999, Thatcher, Gu,
Torrilhon2008, Emerson}. Here we consider the symmetric Couette flow.
The gas lies between two plates parallel to the $xz$-plane. Two
plates move in the opposite direction with constant velocities within
their own planes. A steady state can be obtained for a fully developed
flow.
In this example, the computational domain is $[-0.5, 0.5]$. The
velocities of the left and right plates are
\begin{equation}
\bu_L^W = (-0.6296, 0, 0)^T, \quad \bu_R^W = (0.6296, 0, 0)^T.
\end{equation}
The initial state is a global equilibrium with
\begin{equation}
\rho_0(y) = 1, \quad \bu_0(y) = 0, \quad \theta_0(y) = 1,
\qquad \forall y \in [-0.5, 0.5].
\end{equation}
The steady state can be achieved if the computational time is
sufficiently long. Also, both the \NRxx method and CDVM are applied to
this problem. Three different Knudsen numbers, $Kn = 0.1, 0.5, 1.0$,
are investigated. For CDVM, the computational velocity domain is
chosen as $[-10, 10] \times [-10, 10] \times [-10, 10]$, and $50
\times 50 \times 50$ grids are used. Here we note that such
discretization may not produce numerical results accurate enough as
the reference solution, but the computation is already extremely
slow.
Numerical results for $\Kn = 0.1$ are shown in Figure
\ref{fig:rho_theta_Kn=0.1} and \ref{fig:sigma_sigma_yy_Kn=0.1}. In
this case, most lines agree with each other. The convergence in the
number of moments can be observed, however, due to the numerical error
from both \NRxx method and CDVM, small deviations between the CDVM
results and the possible limit of the \NRxx method can be found. One
can disclose that lower order \NRxx results deviate from the CDVM
results more than higher order ones. This correctly reflects the
behavior of \NRxx method under low Knudsen numbers, as is also found
by \cite{NRxx}.
Now a larger Knudsen number $\Kn = 0.5$ is considered, and the results
are given in Figure \ref{fig:rho_theta_Kn=0.5} and
\ref{fig:sigma_sigma_yy_Kn=0.5}. In this case, the results for odd and
even orders evidently break into two groups, and they approach closer
to the CDVM results separately. This can be also find in Figure \ref
{fig:Ref_rho_theta} and \ref{fig:Ref_sigma_q}. For $\rho$ and
$\theta$, the even group gives better results, while for $\sigma_{12}$
and $\sigma_{22}$, the odd group is more accurate. The reason remains
to be further explored. The two subfigures in Figure
\ref{fig:sigma_sigma_yy_Kn=0.5} clearly exhibit the convergence. In
\cite{Torrilhon2008, Emerson}, it was discovered that the normal
stress $\sigma_{22}$ is difficult to match by R13 and R26 equations.
Here one may find that when the number of moments is increasing, the
quality of the approximation to this quantity is improved
continuously. In case of $M=9$, the profile agrees with the CDVM
result quite well, and when $M=10$, the relative difference is below
$5\%$.
The severe case $\Kn=1.0$ is also studied. Similar results with the
case $\Kn=0.5$ are obtained in Figure \ref{fig:rho_theta_Kn=1.0} and
\ref{fig:sigma_sigma_yy_Kn=1.0}, while the magnitude of the difference
is much larger. For $\sigma_{22}$, now the relative difference for
$M=9$ is about $10\%$. But the rate of convergence is still
encouraging --- compared with the result with $M=4$, the error is
halved.
\psfrag{M=3}{\footnotesize{$M=3$}}
\psfrag{M=4}{\footnotesize{$M=4$}}
\psfrag{M=5}{\footnotesize{$M=5$}}
\psfrag{M=6}{\footnotesize{$M=6$}}
\psfrag{M=7}{\footnotesize{$M=7$}}
\psfrag{M=8}{\footnotesize{$M=8$}}
\psfrag{M=9}{\footnotesize{$M=9$}}
\psfrag{M=10}{\footnotesize{$M=10$}}
\psfrag{solid, CDVM}{\footnotesize{CDVM}}
\psfrag{circle, DSMC}{\footnotesize{DSMC}}
\begin{figure}[!ht]
\centering
\setlength\subfigcapskip{10pt}
\begin{tabular}{r}
\subfigure[Density, $\rho$]{
\includegraphics[scale=.75]{Couette_Kn=0.1_rho.eps}
} \\
\subfigure[Temperature, $\theta$]{
\includegraphics[scale=.75]{Couette_Kn=0.1_theta.eps}
}
\end{tabular}
\caption{Density and temperature plots for the planar Couette flow
with $\Kn=0.1$ (to be continued)}
\label{fig:rho_theta_Kn=0.1}
\end{figure}
\begin{figure}[!ht]
\centering
\setlength\subfigcapskip{10pt}
\begin{tabular}{r}
\subfigure[Shear stress, $\sigma_{12}$]{
\includegraphics[scale=.75]{Couette_Kn=0.1_sigma.eps}
} \\
\subfigure[Normal stress, $\sigma_{22}$]{
\includegraphics[scale=.75]{Couette_Kn=0.1_sigma_yy.eps}
}
\end{tabular}
\caption{Shear and normal stress plots for the planar Couette flow
with $\Kn=0.1$}
\label{fig:sigma_sigma_yy_Kn=0.1}
\end{figure}
\begin{figure}[!ht]
\centering
\setlength\subfigcapskip{10pt}
\begin{tabular}{r}
\subfigure[Density, $\rho$]{
\includegraphics[scale=.75]{Couette_Kn=0.5_rho.eps}
} \\
\subfigure[Temperature, $\theta$]{
\includegraphics[scale=.75]{Couette_Kn=0.5_theta.eps}
}
\end{tabular}
\caption{Density and temperature plots for the planar Couette flow
with $\Kn=0.5$}
\label{fig:rho_theta_Kn=0.5}
\end{figure}
\begin{figure}[!ht]
\centering
\setlength\subfigcapskip{10pt}
\begin{tabular}{r}
\subfigure[Shear stress, $\sigma_{12}$]{
\includegraphics[scale=.75]{Couette_Kn=0.5_sigma.eps}
} \\
\subfigure[Normal stress, $\sigma_{22}$]{
\includegraphics[scale=.75]{Couette_Kn=0.5_sigma_yy.eps}
}
\end{tabular}
\caption{Shear and normal stress plots for the planar Couette flow
with $\Kn=0.5$}
\label{fig:sigma_sigma_yy_Kn=0.5}
\end{figure}
\begin{figure}[!ht]
\centering
\setlength\subfigcapskip{10pt}
\begin{tabular}{r}
\subfigure[Density, $\rho$]{
\includegraphics[scale=.75]{Couette_Kn=1.0_rho.eps}
} \\
\subfigure[Temperature, $\theta$]{
\includegraphics[scale=.75]{Couette_Kn=1.0_theta.eps}
}
\end{tabular}
\caption{Density and temperature plots for the planar Couette flow
with $\Kn=1.0$}
\label{fig:rho_theta_Kn=1.0}
\end{figure}
\begin{figure}[!ht]
\centering
\setlength\subfigcapskip{10pt}
\begin{tabular}{r}
\subfigure[Shear stress, $\sigma_{12}$]{
\includegraphics[scale=.75]{Couette_Kn=1.0_sigma.eps}
} \\
\subfigure[Normal stress, $\sigma_{22}$]{
\includegraphics[scale=.75]{Couette_Kn=1.0_sigma_yy.eps}
}
\end{tabular}
\caption{Shear and normal stress plots for the planar Couette flow
with $\Kn=1.0$}
\label{fig:sigma_sigma_yy_Kn=1.0}
\end{figure}
\subsection{Force-driven Poiseuille flow}
This is another example which is frequently used to verify the
boundary conditions of moment methods \cite{Torrilhon2008, Emerson}.
Similar with the Couette flow, the gas also lies between two parallel
plates, but the plates are stationary and an external constant force
parallel to the plates causes the flow to reach a non-stationary
steady state. In our settings, the computational domain is again
$[-0.5, 0.5]$, and the Knudsen number is set to be $0.1$. The force
introduces an acceleration
\begin{equation}
\bF = (0.2555, 0, 0)^T.
\end{equation}
The initial condition is the same as the Couette flow. These settings
are the non-dimensional form of the test in \cite{Garcia}, where the
DSMC result is carried out, and this example is also considered in
\cite{Xu2007, Emerson}. Since it is quite difficult for us to exert
the force term in CDVM, we have to use the DSMC result in
\cite{Garcia} for comparison in spite of the difference in the
collision model.
The numerical results are presented in Figure \ref{fig:Poiseuille1}
and \ref{fig:Poiseuille2}. For all the profiles, the convergence in
the number of moments is legible, while the \NRxx results do not
converge to the results of DSMC. This may due to the difference
between the collision terms of Shakhov model and DSMC. Taking the
temperature plot (Figure \ref {fig:theta}) as an example, the result
of $M=3$ matches DSMC result best, since when $M=3$, the collision
term of Shakhov model is almost the same as that of DSMC. While when
the number of moments increases, the collision term deviates away from
DSMC's gradually. Here the accuracy of collision models is not the
topic of this paper. Even though, two results are very close
quantitatively, which indicates the correctness of the boundary
conditions and the Prandtl number of the \NRxx method.
\begin{figure}[p]
\centering
\setlength\subfigcapskip{10pt}
\begin{tabular}{r}
\subfigure[Density, $\rho$]{
\includegraphics[scale=.6]{Poiseuille_rho.eps}
} \\
\subfigure[Velocity, $u_2$]{
\includegraphics[scale=.6]{Poiseuille_u.eps}
} \\
\subfigure[Temperature, $\theta$]{
\label{fig:theta}
\includegraphics[scale=.6]{Poiseuille_theta.eps}
}
\end{tabular}
\caption{Density, velocity and temperature plots for the planar
Poiseuille flow}
\label{fig:Poiseuille1}
\end{figure}
\begin{figure}[p]
\centering
\setlength\subfigcapskip{10pt}
\begin{tabular}{r}
\subfigure[Shear stress, $\sigma_{12}$]{
\includegraphics[scale=.6]{Poiseuille_sigma.eps}
} \\
\subfigure[Normal stress, $\sigma_{22}$]{
\includegraphics[scale=.6]{Poiseuille_sigma_yy.eps}
} \\
\subfigure[Heat flux, $q_1$]{
\includegraphics[scale=.6]{Poiseuille_q_x.eps}
}
\end{tabular}
\caption{Stress and heat flux plots for the planar Poiseuille flow}
\label{fig:Poiseuille2}
\end{figure}
|
\section{Introduction}
This paper focusses on polytopes realizing flip graphs on certain geometric and combinatorial structures. Various examples of such polytopes are illustrated in \fref{fig:polytopalRealizations} and described along this introduction. The motivating example is the \defn{associahedron} whose vertices correspond to triangulations of a convex polygon~$\mathcal{P}$ and whose edges correspond to flips between them. The boundary complex of its polar is (isomorphic to) the simplicial complex of crossing-free sets of internal diagonals of~$\mathcal{P}$. The associahedron appears under various motivations ranging from geometric combinatorics to algebra, and several different constructions have been proposed (see~\cite{Lee,Loday,HohlwegLange,CeballosSantosZiegler}). We have represented two different realizations of the $3$-dimensional associahedron in \fref{fig:polytopalRealizations} (top). In fact, the associahedron is a specific case of a more general polytope: the \defn{secondary polytope}~\cite{gkz,bfs} of a $d$-dimensional set~$\mathcal{P}$ of $n$ points is a $(n-d-1)$-dimensional polytope whose vertices correspond to regular triangulations of~$\mathcal{P}$ and whose edges correspond to regular flips between them. Its boundary complex is (isomorphic to) the refinement poset of regular polyhedral subdivisions of~$\mathcal{P}$. See \fref{fig:polytopalRealizations}~(middle left). We refer to~\cite{lrs} for a reference on triangulations of point sets and of the structure of their flip graphs.
\begin{figure}[b]
\capstart
\centerline{\includegraphics[width=.95\textwidth]{geometricFlips}}
\caption{Flips in four geometric structures: a triangulation of a convex polygon, a triangulation of a general point set, a pseudotriangulation and a $2$-triangulation of a convex polygon.}
\label{fig:geometricFlips}
\end{figure}
\begin{figure}[p]
\capstart
\centerline{\includegraphics[width=1.08\textwidth]{transformationPolytopesAll}}
\caption{Polytopal realizations of various flip graphs: two constructions of the $3$-dimensional associahedron ---~the secondary polytope of the regular hexagon (top left) and Loday's construction (top right); the secondary polytope of a set of $6$ points (middle left); a $3$-dimensional pseudotriangulations polytope (middle right); the $3$-dimensional permutahedron (bottom left); the $3$-dimensional cyclohedron (bottom right).}
\label{fig:polytopalRealizations}
\end{figure}
Our work was motivated by two different generalizations of planar triangulations, whose combinatorial structures extend that of the associahedron~---~see \fref{fig:geometricFlips}:
\begin{enumerate}[(i)]
\item Let~$\mathcal{P}$ be a point set in general position in the Euclidean plane, with $i$ interior points and $b$ boundary points. A set of edges with vertices in~$\mathcal{P}$ is \defn{pointed} if the edges incident to any point of~$\mathcal{P}$ span a pointed cone. A \defn{pseudotriangle} on~$\mathcal{P}$ is a simple polygon with vertices in~$\mathcal{P}$, which has precisely three convex corners, joined by three concave chains. A (pointed) \defn{pseudotriangulation} of~$\mathcal{P}$ is a decomposition of its convex hull into~$i+b-2$ pseudotriangles on~$\mathcal{P}$~\cite{PocchiolaVegter,RoteSantosStreinu-survey}. Equivalently, it is a maximal pointed and crossing-free set of edges with vertices in~$\mathcal{P}$. The \defn{pseudotriangulations polytope}~\cite{RoteSantosStreinu-polytope} of the point set~$\mathcal{P}$ is a simple $(2i+b-3)$-dimensional polytope whose vertices correspond to pseudotriangulations of~$\mathcal{P}$ and whose edges correspond to flips between them. The boundary complex of its polar is (isomorphic to) the simplicial complex of pointed crossing-free sets of internal edges on~$\mathcal{P}$. See \fref{fig:polytopalRealizations} (middle right).
\item A \defn{$k$-triangulation} of a convex $n$-gon~$\mathcal{P}$ is a maximal set of diagonals with no $(k+1)$-crossing (no $k+1$ diagonals are mutually crossing)~\cite{CapoyleasPach, PilaudSantos, Pilaud}. We can forget the diagonals of the $n$-gon with less than $k$ vertices of~$\mathcal{P}$ on one side: they cannot appear in a $(k+1)$-crossing and thus they belong to all $k$-triangulations of~$\mathcal{P}$. The other edges are called \defn{$k$-relevant}. The simplicial complex~$\Delta_n^k$ of $(k+1)$-crossing-free sets of $k$-relevant diagonals of $\mathcal{P}$ is a topological sphere~\cite{Jonsson, Stump} whose facets are $k$-triangulations of $\mathcal{P}$ and whose ridges are flips between them. It remains open whether or not this simplicial complex is the boundary complex of a polytope.
\end{enumerate}
In~\cite{PilaudPocchiola}, Pilaud \& Pocchiola developed a general framework which generalizes both pseudotriangulations and multitriangulations. They study the graph of flips on \defn{pseudoline arrangements with contacts} supported by a given sorting network. The present paper is based on this framework. Definitions and basic properties are recalled in Section~\ref{sec:definitions}.
In this paper, we define and study the \defn{brick polytope} of a sorting network~$\mathcal{N}$, obtained as the convex hull of vectors associated to each pseudoline arrangement supported by~$\mathcal{N}$. Our main result is the characterization of the pseudoline arrangements which give rise to the vertices of the brick polytope, from which we derive a combinatorial description of the faces of the brick polytope. We furthermore provide a natural decomposition of the brick polytope into a Minkowski sum of matroid polytopes. These structural results are presented in Section~\ref{sec:structure}. We illustrate the results of this section with particular sorting networks whose brick polytopes are graphical zonotopes. Among them, the \defn{permutahedron} is a well-known simple $(n-1)$-dimensional polytope whose vertices correspond to permutations of~$[n]$ and whose edges correspond to pairs of permutations which differ by an adjacent transposition. Its boundary complex is (isomorphic to) the refinement poset of ordered partitions of~$[n]$. See \fref{fig:polytopalRealizations} (bottom left).
We obtain our most relevant examples of brick polytopes in Section~\ref{sec:associahedra}. We observe that for certain well-chosen sorting networks, our brick polytopes coincide (up to translation) with Hohlweg \& Lange's realizations of the associahedron~\cite{HohlwegLange}. We therefore provide a complementary point of view on their polytopes and we complete their combinatorial description. We obtain in particular a natural Minkowski sum decomposition of these polytopes into matroid polytopes.
Finally, Section~\ref{sec:multi} is devoted to our initial motivation for the construction of the brick polytope. We wanted to find a polytopal realization of the simplicial complex~$\Delta_n^k$ of $(k+1)$-crossing-free sets of $k$-relevant diagonals of the $n$-gon. Using Pilaud \& Pocchiola's correspondence between multitriangulations and pseudoline arrangement covering certain sorting networks~\cite{PilaudPocchiola}, we construct a point configuration in $\mathbb{R}^{n-2k}$ with one point associated to each $k$-triangulation of the $n$-gon. We had good reasons to believe that this point set could be a projection of the polar of a realization of~$\Delta_n^k$: the graph of the corresponding brick polytope (the convex hull of this point configuration) is a subgraph of flips, and all sets of $k$-triangulations whose corresponding points belong to a given face of this brick polytope are faces of~$\Delta_n^k$. However, we prove that our point configuration cannot be a projection of the polar of a realization of~$\Delta_n^k$.
After the completion of a preliminary version of this paper, Stump pointed out to us his paper~\cite{Stump} which connects the multitriangulations to the type~$A$ subword complexes of Knutson \& Miller~\cite{KnutsonMiller}. The latter can be visually interpreted as sorting networks (see Section~\ref{subsec:subwordComplex}). This opened the perspective of the generalization of brick polytopes to subword complexes on Coxeter groups. This generalization was achieved by Pilaud \& Stump in~\cite{PilaudStump}. This construction yields in particular the generalized associahedra of Hohlweg, Lange \& Thomas~\cite{HohlwegLangeThomas} for certain particular subword complexes described by Ceballos, Labb\'e \& Stump~\cite{CeballosLabbeStump}. In the present paper, we focus on the classical situation of type~$A$, which already reflects the essence of the construction. The only polytope of different type which appears here is the \defn{cyclohedron} via its standard embedding in the associahedron. The vertices of the cyclohedron correspond to centrally symmetric triangulations of a centrally symmetric convex $(2n)$-gon and its edges correspond to centrally symmetric flips between them (\ie either a flip of a centrally symmetric diagonal or a simultaneous flip of a pair of symmetric diagonals). The boundary complex of its polar is (isomorphic to) the refinement poset of centrally symmetric polygonal subdivisions of the $(2n)$-gon. See \fref{fig:polytopalRealizations} (bottom right).
We moreover refer to~\cite{PilaudStump} for further properties of the brick polytope which appeared when generalizing it to Coxeter groups of finite types. They relate in particular the graph of the brick polytope to a quotient of the weak order and the normal fan of the brick polytope to the Coxeter fan.
\section{The brick polytope of a sorting network}\label{sec:definitions}
\subsection{Pseudoline arrangements on sorting networks}\label{subsec:networks}
Consider a set of $n$ hori\-zontal lines (called \defn{levels}, and labeled from bottom to top), and place $m$ vertical segments (called \defn{commutators}, and labeled from left to right) joining two consecutive horizontal lines, such that no two commutators have a common endpoint~---~see \eg Figure~\ref{fig:network}. Throughout this paper, we fix such a configuration $\mathcal{N}$ that we call a \defn{network}. The \defn{bricks} of~$\mathcal{N}$ are its $m-n+1$~bounded cells. We say that a network is \defn{alternating} when the commutators adjacent to each intermediate level are alternatively located above and below it.
\begin{figure}[ht]
\capstart
\centerline{\includegraphics[width=\textwidth]{network}}
\caption{Three networks with $5$ levels, $14$ commutators and $10$ bricks. The first two are alternating.}
\label{fig:network}
\end{figure}
A \defn{pseudoline} is an abscissa monotone path on the network $\mathcal{N}$. A \defn{contact} between two pseudolines is a commutator whose endpoints are contained one in each pseudoline, and a \defn{crossing} between two pseudolines is a commutator traversed by both pseudolines. A \defn{pseudoline arrangement} (with contacts) is a set of $n$ pseudolines supported by $\mathcal{N}$ such that any two of them have precisely one crossing, some (perhaps zero) contacts, and no other intersection~---~see Figure~\ref{fig:flip}. Observe that in a pseudoline arrangement, the pseudoline which starts at level $\ell$ necessarily ends at level $n+1-\ell$ and goes up at $n-\ell$ crossings and down at $\ell-1$ crossings. Note also that a pseudoline arrangement supported by $\mathcal{N}$ is completely determined by its ${n \choose 2}$ crossings, or equivalently by its $m-{n \choose 2}$ contacts. Let~$\arr(\mathcal{N})$ denote the set of pseudoline arrangements supported by~$\mathcal{N}$. We say that a network is \defn{sorting} when it supports at least one pseudoline arrangement.
\begin{figure}
\capstart
\centerline{\includegraphics[width=\textwidth]{flip}}
\caption{Two pseudoline arrangements, both supported by the rightmost network~$\mathcal{N}$ of \fref{fig:network}, and related by a flip. The left one is the greedy pseudoline arrangement~$\Gamma(\mathcal{N})$, whose flips are all decreasing. It is obtained by sorting the permutation~$(5,4,3,2,1)$ according to the network~$\mathcal{N}$.}
\label{fig:flip}
\end{figure}
\subsection{The graph of flips}\label{subsec:flips}
There is a natural \defn{flip} operation which transforms a pseudoline arrangement supported by $\mathcal{N}$ into another one by exchanging the position of a contact. More precisely, if $V$ is the set of contacts of a pseudoline arrangement $\Lambda$ supported by~$\mathcal{N}$, and if $v\in V$ is a contact between two pseudolines of $\Lambda$ which cross at $w$, then $(V\smallsetminus\{v\}) \cup \{w\}$ is the set of contacts of another pseudoline arrangement supported by $\mathcal{N}$~---~see Figure~\ref{fig:flip}. The \defn{graph of flips}~$G(\mathcal{N})$ is the graph whose nodes are the pseudoline arrangements supported by~$\mathcal{N}$ and whose edges are the flips between them. This graph was studied in~\cite{PilaudPocchiola}, whose first statement is the following result:
\begin{theorem}[{\cite{PilaudPocchiola}}]\label{theo:connect}
The graph of flips $G(\mathcal{N})$ of a sorting network~$\mathcal{N}$ with $n$ levels and $m$ commutators is $\left(m-{n \choose 2}\right)$-regular and connected.
\end{theorem}
Regularity of the graph of flips is obvious since every contact induces a flip. For the connectivity, define a flip to be \defn{decreasing} if the added contact lies on the left of the removed contact. The oriented graph of decreasing flips is clearly acyclic and is proved to have a unique source in~\cite{PilaudPocchiola} (and thus, to be connected). This source is called the \defn{greedy pseudoline arrangement} supported by~$\mathcal{N}$ and is denoted by~$\Gamma(\mathcal{N})$. It is characterized by the property that any of its contacts is located to the right of its corresponding crossing. It can be computed by sorting the permutation~$(n,n-1,\dots,2,1)$ according to the sorting network~$\mathcal{N}$~---~see \fref{fig:flip}~(left). We will use this particular pseudoline arrangement later. We refer to~\cite{PilaudPocchiola} for further details.
For any given subset~$\gamma$ of the commutators of~$\mathcal{N}$, we denote by $\arr(\mathcal{N}|\gamma)$ the set of pseudoline arrangements supported by~$\mathcal{N}$ and whose set of contacts contains~$\gamma$. The arrangements of $\arr(\mathcal{N}|\gamma)$ are in obvious correspondence with that of $\arr(\mathcal{N}\smallsetminus\gamma)$, where~$\mathcal{N}\smallsetminus\gamma$ denotes the network obtained by erasing the commutators of~$\gamma$ in~$\mathcal{N}$. In particular, the subgraph of~$G(\mathcal{N})$ induced by~$\arr(\mathcal{N}|\gamma)$ is isomorphic to $G(\mathcal{N}\smallsetminus\gamma)$, and thus Theorem~\ref{theo:connect} ensures that this subgraph is connected for every~$\gamma$.
More generally, let~$\Delta(\mathcal{N})$ denote the simplicial complex of all sets of commutators of~$\mathcal{N}$ contained in the set of contacts of a pseudoline arrangement supported by~$\mathcal{N}$. In other words, a set~$\gamma$ of commutators of~$\mathcal{N}$ is a face of $\Delta(\mathcal{N})$ if and only if the network $\mathcal{N}\smallsetminus\gamma$ is still sorting. This complex is pure of dimension~$m-{n \choose 2}-1$, its maximal cells correspond to pseudoline arrangements supported by~$\mathcal{N}$ and its ridge graph is the graph of flips~$G(\mathcal{N})$. The previous connectedness properties ensure that~$\Delta(\mathcal{N})$ is an abstract polytope~\cite{Schulte}, and it is even a combinatorial sphere (see Corollary~\ref{coro:shellable} and the discussion in Section~\ref{subsec:multiassociahedron}). These properties motivate the following question:
\begin{question}\label{qu:polytope}
Is~$\Delta(\mathcal{N})$ the boundary complex of a $\left(m-{n \choose 2}\right)$-dimensional simplicial polytope?
\end{question}
In this article, we construct a polytope whose graph is a subgraph of~$G(\mathcal{N})$, and which combinatorially looks like ``a projection of'' the dual complex of $\Delta(\mathcal{N})$. More precisely, we associate a vector $\omega(\Lambda) \in \mathbb{R}^n$ to each~$\Lambda\in\arr(\mathcal{N})$, and we consider the convex hull~${\Omega(\mathcal{N}) \eqdef \conv\set{\omega(\Lambda)}{\Lambda\in\arr(\mathcal{N})} \subset\mathbb{R}^n}$ of all these vectors. The resulting polytope has the property that for every face~$F$ of~$\Omega(\mathcal{N})$ there is a set~$\gamma$ of commutators of~$\mathcal{N}$ such that $\arr(\mathcal{N}|\gamma) = \set{\Lambda\in\arr(\mathcal{N})}{\omega(\Lambda)\in F}$. In particular, when the dimension of~$\Omega(\mathcal{N})$ is $m-{n \choose 2}$, our construction answers Question~\ref{qu:polytope} in the affirmative. The relationship between our construction and Question~\ref{qu:polytope} is discussed in more details in Section~\ref{subsec:multiassociahedron}.
\subsection{Subword complexes on finite Coxeter groups}
\label{subsec:subwordComplex}
Before presenting our construction of the brick polytope of a sorting network, we make a little detour to connect the abovementioned simplicial complexes $\Delta(\mathcal{N})$ with the subword complexes of Knutson \& Miller~\cite{KnutsonMiller}.
Let $(W,S)$ be a finite Coxeter system, that is, $W$ is a finite reflection group and $S$ is a set of simple reflections minimally generating $W$. See for example~\cite{Humphreys} for background. Let $Q$ be a word on the alphabet $S$ and let $\rho$ be an element of~$W$. The \defn{subword complex} $\Delta(Q,\rho)$ is the pure simplicial complex of subwords of~$Q$ whose complements contain a reduced expression of $\rho$~\cite{KnutsonMiller}. The vertices of this simplicial complex are labeled by the positions in the word $Q$ (note that two positions are different even if the letters of $Q$ at these positions coincide), and its facets are the complements of the reduced expressions of $\rho$ in the word $Q$.
There is a straightforward combinatorial isomorphism between:
\begin{enumerate}[(i)]
\item the simplicial complex $\Delta(\mathcal{N})$ discussed in the previous section, where $\mathcal{N}$ is a sorting network on $n$ levels and $m$ commutators; and
\item the subword complex $\Delta(Q,w_\circ)$, where the underlying Coxeter group $W$ is the symmetric group $\mathfrak{S}_n$ on $n$ elements, the simple system $S$ is the set of adjacent transpositions $\tau_i \eqdef (i,i+1)$, the word $Q = \tau_{i_1}\tau_{i_2}\dots\tau_{i_m}$ is formed according to the positions of the commutators of~$\mathcal{N}$ --- the $j$\textsuperscript{th} leftmost commutator of $\mathcal{N}$ lies between the $i_j$\textsuperscript{th} and $(i_j+1)$\textsuperscript{th} levels of $\mathcal{N}$ ---, and the permutation ${w_\circ = [n,n-1,\dots,2,1]}$ is the longest element of $\mathfrak{S}_n$ --- its reduced expressions on $S$ all have the maximal length ${n \choose 2}$.
\end{enumerate}
Since Pilaud \& Pocchiola~\cite{PilaudPocchiola} were not aware of the definition of the subword complex, they studied the simplicial complexes $\Delta(\mathcal{N})$ independently and rediscovered some relevant properties which hold for any subword complex. The connection between multitriangulations and subword complexes was first done by Stump~\cite{Stump}, providing the powerful toolbox of Coxeter combinatorics to the playground.
In particular, Knutson \& Miller prove in~\cite{KnutsonMiller} that the subword complex $\Delta(Q,\rho)$ is either a combinatorial sphere or a combinatorial ball, depending on whether the Demazure product of $Q$ equals $\rho$ or not. The interested reader can refer to their article for details on this property and for other known properties on subword complexes. In the conclusion of their article, Question~6.4 asks in particular whether any spherical subword complex is the boundary complex of a convex simplicial polytope, which is a generalized version of Question~\ref{qu:polytope} stated above.
Throughout our article, we only consider subword complexes on the Coxeter system $(\mathfrak{S}_n, \set{\tau_i}{i\in[n-1]})$ and with $\rho=w_\circ$. However, we want to mention that Pilaud \& Stump~\cite{PilaudStump} extended the construction of this paper to any subword complex on any Coxeter system. This generalized construction yields in particular the generalized associahedra of Hohlweg, Lange \& Thomas~\cite{HohlwegLangeThomas} for certain particular subword complexes described by Ceballos, Labb\'e \& Stump~\cite{CeballosLabbeStump}.
\subsection{The brick polytope}
The subject of this paper is the following polytope:
\begin{definition}
Let $\mathcal{N}$ be a sorting network with $n$ levels. The \defn{brick vector} of a pseudoline arrangement $\Lambda$ supported by $\mathcal{N}$ is the vector $\omega(\Lambda) \in \mathbb{R}^n$ whose $i$\textsuperscript{th} coordinate is the number of bricks of~$\mathcal{N}$ located below the $i$\textsuperscript{th} pseudoline of $\Lambda$ (the one which starts at level $i$ and finishes at level $n+1-i$). The \defn{brick polytope} $\Omega(\mathcal{N}) \subset \mathbb{R}^n$ of the sorting network $\mathcal{N}$ is the convex hull of the brick vectors of all pseudoline arrangements supported by $\mathcal{N}$:
$$\Omega(\mathcal{N}) \eqdef \conv\set{\omega(\Lambda)}{\Lambda\in\arr(\mathcal{N})} \subset\mathbb{R}^n.$$
\end{definition}
This article aims to describe the combinatorial properties of this polytope in terms of the properties of the supporting network. In Section~\ref{sec:structure}, we provide a characterization of the pseudoline arrangements supported by~$\mathcal{N}$ whose brick vectors are vertices of the brick polytope~$\Omega(\mathcal{N})$, from which we derive a combinatorial description of the faces of the brick polytope. We also provide a natural decomposition of~$\Omega(\mathcal{N})$ into a Minkowski sum of simpler polytopes. In Section~\ref{sec:associahedra}, we recall the duality between the triangulations of a convex polygon and the pseudoline arrangements supported by certain networks~\cite{PilaudPocchiola}, whose brick polytopes coincide with Hohlweg \& Lange's realizations of the associahedron~\cite{HohlwegLange}. We finally discuss in Section~\ref{sec:multi} the properties of the brick polytopes of more general networks which support pseudoline arrangements corresponding to multitriangulations of convex polygons~\cite{PilaudSantos,PilaudPocchiola}.
We start by observing that the brick polytope is not full dimensional. Define the \defn{depth} of a brick of $\mathcal{N}$ to be the number of levels located above it, and let $D(\mathcal{N})$ be the sum of the depths of all the bricks of~$\mathcal{N}$. Since any pseudoline arrangement supported by $\mathcal{N}$ covers each brick as many times as its depth, all brick vectors are contained in the following hyperplane:
\begin{lemma}\label{lem:depth}
The brick polytope~$\Omega(\mathcal{N})\subset\mathbb{R}^n$ is contained in the hyperplane of equation $\sum_{i=1}^n x_i = D(\mathcal{N})$.
\end{lemma}
The dimension of~$\Omega(\mathcal{N})$ is thus at most $n-1$, but could be smaller. We obtain the dimension of~$\Omega(\mathcal{N})$ in Corollary~\ref{coro:dim}.
We can also describe immediately the action of the vertical and horizontal reflections of the network on the brick polytope.
The brick polytope of the network~$v(\mathcal{N})$ obtained by reflecting~$\mathcal{N}$ through the vertical axis is the image of~$\Omega(\mathcal{N})$ under the affine transformation $(x_1,\dots,x_n) \mapsto (x_n,\dots,x_1)$. Similarly, the brick polytope of the network~$h(\mathcal{N})$ obtained by reflecting~$\mathcal{N}$ through the horizontal axis is the image of~$\Omega(\mathcal{N})$ under the affine transformation $(x_1,\dots,x_n) \mapsto (m-n+1)1\!\!1 - (x_n,\dots,x_1)$.
\subsection{Examples}\label{subsec:examples}
Before going on, we present some examples which will illustrate our results throughout the paper. Further motivating examples will be studied in Sections~\ref{sec:associahedra} and~\ref{sec:multi}.
\begin{example}[Reduced networks]\label{exm:reduced}
A sorting network~$\mathcal{N}$ with~$n$ levels and~${m={n \choose 2}}$ commutators supports a unique pseudoline arrangement. Consequently, the graph of flips~$G(\mathcal{N})$, the simplicial complex~$\Delta(\mathcal{N})$ and the brick polytope~$\Omega(\mathcal{N})$ are all reduced to a single point. Such a network is said to be \defn{reduced}.
\end{example}
\begin{example}[$2$-level networks]
Consider the network $\mathcal{X}_m$ formed by two levels related by $m$ commutators. We obtain a pseudoline arrangement by choosing any of these commutators as the unique crossing between two pseudolines supported by~$\mathcal{X}_m$. Thus, the graph of flips~$G(\mathcal{X}_m)$ is the complete graph on $m$ vertices, and the simplicial complex~$\Delta(\mathcal{X}_m)$ is a $(m-1)$-dimensional simplex. The brick polytope $\Omega(\mathcal{X}_m)$ is, however, a segment.
\begin{figure}[ht]
\capstart
\centerline{\includegraphics[width=.8\textwidth]{2levels}}
\caption{The three pseudoline arrangements supported by the network~$\mathcal{X}_3$ with two levels and three commutators.}
\label{fig:2levels}
\end{figure}
\noindent
The brick vector of the pseudoline arrangement whose crossing is the $i$\textsuperscript{th} commutator of~$\mathcal{X}_m$ is the vector $(m-i,i-1)$. Thus, the brick polytope~$\Omega(\mathcal{X}_m)$ is the segment from $(m-1,0)$ to $(0,m-1)$. Its endpoints are the brick vectors of the pseudoline arrangements whose crossings are respectively the first and the last commutator of~$\mathcal{X}_m$. The former is the source (\ie the greedy pseudoline arrangement) and the latter is the sink in the oriented graph of decreasing flips. The polytope $\Omega(\mathcal{X}_m)$ is contained in the hyperplane of equation $x+y=m-1$.
\end{example}
\begin{example}[$3$-level alternating networks]\label{exm:3levels}
Let~$m \ge 3$ be an odd integer. Consider the network~$\mathcal{Y}_m$ formed by 3 levels related by $m$ alternating commutators. Any choice of~$3$ alternating crossings provides a pseudoline arrangement supported by~$\mathcal{Y}_m$. Consequently, $\mathcal{Y}_m$ supports precisely $\frac{1}{24}(m-1)m(m+1)$ pseudoline arrangements. The brick polytope~$\Omega(\mathcal{Y}_m)$ is a single point when $m=3$, a pentagon when $m=5$, and a hexagon for any $m \ge 7$. We have represented in \fref{fig:3levels} the projection of~$\Omega(\mathcal{Y}_{11})$ on the first and third coordinates plane, in such a way that the transformation $(x_1,x_3) \mapsto (x_3,x_1)$ appears as a vertical reflection.
\begin{figure}
\capstart
\centerline{\includegraphics[width=.85\textwidth]{3levels}}
\caption{The brick polytope~$\Omega(\mathcal{Y}_{11})$ of a $3$-level alternating network, projected on the first and third coordinates plane. Next to each vertex is drawn the corresponding pseudoline arrangement, and next to each edge is drawn the corresponding set of contacts.}
\label{fig:3levels}
\end{figure}
\end{example}
\begin{example}[Duplicated networks]\label{exm:duplicated1}
Consider a reduced network~$\mathcal{N}$ with $n$ levels and ${n \choose 2}$ commutators. For any distinct~$i,j\in[n]$, we labeled by $\{i,j\}$ the commutator of~$\mathcal{N}$ where the $i$\textsuperscript{th} and $j$\textsuperscript{th} pseudolines of the unique pseudoline arrangement supported by~$\mathcal{N}$ cross.
Let $\Gamma$ be a connected graph on~$[n]$. We define $\mathcal{Z}(\Gamma)$ to be the network with $n$ levels and $m={n \choose 2} + |\Gamma|$ commutators obtained from~$\mathcal{N}$ by duplicating the commutators labeled by the edges of~$\Gamma$~---~see \fref{fig:duplicated}. We say that~$\mathcal{Z}(\Gamma)$ is a \defn{duplicated network}. Observe that a pseudoline arrangement supported by~$\mathcal{Z}(\Gamma)$ has a crossing for each commutator which has not been duplicated, and a crossing and a contact among each pair of duplicated commutators. Thus, $\mathcal{Z}(\Gamma)$ supports precisely $2^{|\Gamma|}$ different pseudoline arrangements, the graph of flips~$G(\mathcal{Z}(\Gamma))$ is the graph of the $|\Gamma|$-dimensional cube, and more generally, the simplicial complex $\Delta(\mathcal{Z}(\Gamma))$ is the boundary complex of the $|\Gamma|$-dimensional cross-polytope. As an application of the results of Section~\ref{sec:structure}, we will see that the brick polytope of the duplicated network~$\mathcal{Z}(\Gamma)$ is a graphical zonotope~---~see Examples~\ref{exm:duplicated2},~\ref{exm:duplicated3},~\ref{exm:duplicated4}, and~\ref{exm:duplicated5}. In particular, when $\Gamma$ is complete we obtain the permutahedron, while when~$\Gamma$ is a tree, we obtain a cube.
\begin{figure}[ht]
\capstart
\centerline{\includegraphics[scale=.48]{duplicated}}
\caption{The graph of flips of the duplicated network $\mathcal{Z}(K_3)$.}
\label{fig:duplicated}
\end{figure}
\end{example}
\section{Combinatorial description of the brick polytope}\label{sec:structure}
In this section, we characterize the vertices and describe the faces of the brick polytope $\Omega(\mathcal{N})$. For this purpose, we study the cone of the brick polytope~$\Omega(\mathcal{N})$ at the brick vector of a given pseudoline arrangement supported by~$\mathcal{N}$. Our main tool is the incidence configuration of the contact graph of a pseudoline arrangement, which we define next.
\subsection{The contact graph of a pseudoline arrangement}
Let $\mathcal{N}$ be a sorting network with~$n$ levels and~$m$ commutators, and let~$\Lambda$ be a pseudoline arrangement supported by~$\mathcal{N}$.
\begin{definition}
The \defn{contact graph} of $\Lambda$ is the directed multigraph $\Lambda^\#$ with a node for each pseudoline of $\Lambda$ and an arc for each contact of $\Lambda$ oriented from the pseudoline passing above the contact to the pseudoline passing below it.
\end{definition}
\begin{figure}[h]
\capstart
\centerline{\includegraphics[width=\textwidth]{contact}}
\caption{The contact graphs of the pseudoline arrangements of \fref{fig:flip}. The connected components are preserved by the flip.}
\label{fig:contact}
\end{figure}
The nodes of the contact graph come naturally labeled by~$[n]$: we label by~$\ell$ the node corresponding to the pseudoline of~$\Lambda$ which starts at level~$\ell$ and finishes at level~$n+1-\ell$. With this additional labeling, the contact graph provides enough information to characterize its pseudoline arrangement:
\begin{lemma}\label{lem:contactgraph}
Let~$\mathcal{N}$ be a network and~$\Lambda^\#$ be a graph on~$[n]$. If $\Lambda^\#$ is the contact graph of a pseudoline arrangement~$\Lambda$ supported by~$\mathcal{N}$, then~$\Lambda$ can be reconstructed from~$\Lambda^\#$ and~$\mathcal{N}$.
\end{lemma}
\begin{proof}
To obtain a pseudoline arrangement from its contact graph~$\Lambda^\#$, we sort the permutation~$(n,n-1,\dots,2,1)$ on~$\mathcal{N}$ according to~$\Lambda^\#$. We sweep the network from left to right, and start to draw the $\ell$\textsuperscript{th} pseudoline at level~$\ell$. When we reach a commutator of~$\mathcal{N}$ with pseudoline~$i$ above and pseudoline~$j$ below,
\begin{itemize}
\item if there remains an arc $(i,j)$ in~$\Lambda^\#$, we insert a contact in place of our commutator and delete an arc~$(i,j)$ from~$\Lambda^\#$;
\item otherwise, we insert a crossing in place of our commutator: the indices $i$ and $j$ of the permutation get sorted at this crossing
\end{itemize}
This procedures is correct since the contacts between two pseudolines $i<j$ are all directed from $i$ to $j$ before their crossing and from $j$ to $i$ after their crossing.
\end{proof}
\begin{example}\label{exm:contactgreedy}
We already mentioned a relevant example of the sorting procedure of the previous proof: the greedy pseudoline arrangement~$\Gamma(\mathcal{N})$ is obtained by sorting the permutation~$(n,n-1,\dots,2,1)$ on~$\mathcal{N}$ and inserting the crossings as soon as possible. In other words any arc of the contact graph of the greedy pseudoline arrangement is sorted (\ie of the form~$(i,j)$ with~$i<j$)~---~see \fref{fig:contact} (left).
\end{example}
\begin{example}[Duplicated networks, continued]\label{exm:duplicated2}
For a connected graph $\Gamma$, consider the duplicated network $\mathcal{Z}(\Gamma)$ defined in Example~\ref{exm:duplicated1}. A pseudoline arrangement supported by~$\mathcal{Z}(\Gamma)$ has one contact among each pair of duplicated commutators, and thus its contact graph is an oriented copy of~$\Gamma$. Reciprocally, any orientation on~$\Gamma$ is the contact graph of a pseudoline arrangement: this follows from Lemma~\ref{lem:contactgraph} since~$\mathcal{Z}(\Gamma)$ supports~$2^{|\Gamma|}$ pseudoline arrangements, but one can also easily reconstruct the pseudoline arrangement whose contact graph is a given orientation on~$\Gamma$.
Note that the contact graphs of the pseudoline arrangements supported by a given network have in general distinct underlying undirected graphs (see \fref{fig:contact}).
\end{example}
\begin{remark}
In fact, any labeled directed multigraph arises as the contact graph of pseudoline arrangement on a certain sorting network. Indeed, consider a labeled directed multigraph~$G$ on $n$ vertices. Consider the unique pseudoline arrangement~$\Lambda$ supported by a reduced network~$\mathcal{N}$ with $n$ levels and ${n \choose 2}$ commutators. For any directed edge ${(i,j)\in G}$ with $i<j$ (resp.~with $i>j$), insert a new commutator immediately to the right (resp.~left) of the crossing between the $i$\textsuperscript{th} and $j$\textsuperscript{th} pseudolines of $\Lambda$. Then $G$ is precisely the contact graph $\Lambda^\#$ of $\Lambda$ (seen as a pseudoline arrangement supported by the resulting network).
\end{remark}
Let $\Lambda$ and $\Lambda'$ denote two pseudoline arrangements supported by~$\mathcal{N}$ and related by a flip involving their $i$\textsuperscript{th} and $j$\textsuperscript{th} pseudolines~---~see \fref{fig:contact} for an example. Then the directed multigraphs obtained by merging the vertices $i$ and $j$ in the contact graphs $\Lambda^\#$ and $\Lambda'^\#$ coincide. In particular, a flip preserves the connected components of the contact graph. Since the flip graph~$G(\mathcal{N})$ is connected (Theorem~\ref{theo:connect}), this implies the following result:
\begin{lemma}\label{lem:connectcomp}
The contact graphs of all pseudoline arrangements supported by~$\mathcal{N}$ have the same connected components.
\end{lemma}
We call a sorting network \defn{reducible} (resp.~\defn{irreducible}) when the contact graphs of the pseudoline arrangements it supports are disconnected (resp. connected).
Our next statement describes the structure of the simplicial complex~$\Delta(\mathcal{N})$ and of the brick polytope~$\Omega(\mathcal{N})$ associated to a reducible sorting network~$\mathcal{N}$. To formalize it, we need the following definition. If $\mathcal{N}$ is a network and $\Theta$ is a set of disjoint abscissa monotone curves supported by~$\mathcal{N}$, the \defn{restriction} of~$\mathcal{N}$ to~$\Theta$ is the network obtained by keeping only the curves of $\Theta$ and the commutators between them, and by stretching the curves of $\Theta$. In other words, it has $|\Theta|$ levels and a commutator between its $i$\textsuperscript{th} and $(i+1)$\textsuperscript{th} levels for each commutator of $\mathcal{N}$ joining the $i$\textsuperscript{th} and $(i+1)$\textsuperscript{th} curves of $\Theta$~---~see \fref{fig:restriction} (left).
\begin{figure}
\capstart
\centerline{\includegraphics[width=.9\textwidth]{restriction}}
\caption{A sorting network and a pseudoline arrangement covering it (top). Their restriction to the connected component $\{1,3,5\}$ of the contact graph (bottom).}
\label{fig:restriction}
\end{figure}
It makes sense to speak of the restriction~$\mathcal{N}(U)$ of~$\mathcal{N}$ to a connected component~$U$ of the contact graphs of the pseudoline arrangements supported by~$\mathcal{N}$. Indeed, if~$\Lambda$ is supported by~$\mathcal{N}$, the restriction of~$\mathcal{N}$ to the levels of the subarrangement formed by the pseudolines of~$\Lambda$ labeled by~$U$ does not depend on the choice of~$\Lambda$~---~see \fref{fig:restriction} (right). Furthermore, there is an obvious correspondence between the pseudoline arrangements supported by~$\mathcal{N}(U)$ and the subarrangements of the arrangements supported by~$\mathcal{N}$ formed by their pseudolines in~$U$. In particular,~$\mathcal{N}(U)$ is an irreducible sorting network; we say that it is an \defn{irreducible component} of~$\mathcal{N}$.
\begin{proposition}\label{prop:disconnected}
Let $\mathcal{N}$ be a sorting network whose irreducible components are $\mathcal{N}_1,\dots,\mathcal{N}_p$. Then the simplicial complex~$\Delta(\mathcal{N})$ is isomorphic to the join of the simplicial complexes~$\Delta(\mathcal{N}_1),\dots,\Delta(\mathcal{N}_p)$ and the brick polytope~$\Omega(\mathcal{N})$ is a translate of the product of the brick polytopes~$\Omega(\mathcal{N}_1),\dots,\Omega(\mathcal{N}_p)$.
\end{proposition}
\begin{proof}
The commutators of $\mathcal{N}$ can be partitioned into~$p+1$ sets: one set corresponding to each irreducible component of~$\mathcal{N}$, and the set~$X$ of commutators between two different connected components in the contact graphs. All pseudoline arrangements supported by~$\mathcal{N}$ have crossings at the commutators of~$X$, and are obtained by choosing independently their subarrangements on the irreducible components~$\mathcal{N}_1,\dots,\mathcal{N}_p$. The result immediately follows.
\end{proof}
In particular, the dimension of the brick polytope of a sorting network with $n$ levels and $p$ irreducible components is at most $n-p$. We will see in Corollary~\ref{coro:dim} that this is the exact dimension. For example, the brick polytope of a reduced network (see Example~\ref{exm:reduced}) has dimension $0$: it supports a unique pseudoline arrangement whose contact graph has no edge, and its brick polytope is a single point.
Proposition~\ref{prop:disconnected} enables us to only focus on irreducible sorting networks throughout this article. Among them, the following networks have the fewest commutators:
\begin{definition}\label{def:minimal}
An irreducible sorting network~$\mathcal{N}$ is \defn{minimal} if it satisfies the following equivalent conditions:
\begin{enumerate}[(i)]
\item $\mathcal{N}$ has $n$ levels and $m={n \choose 2}+n-1$ commutators.
\item The contact graph of a pseudoline arrangement supported by~$\mathcal{N}$ is a tree.
\item The contact graphs of all pseudoline arrangements supported by~$\mathcal{N}$ are trees.
\end{enumerate}
\end{definition}
For example, the networks of \fref{fig:network} all have $5$ levels and $14$ commutators. The rightmost is reducible, but the other two are minimal. To be convinced, draw the greedy pseudoline arrangement on these networks, and check that its contact graph is connected.
We come back to minimal irreducible sorting networks at the end of Section~\ref{subsec:fd} since their brick polytopes are of particular interest.
\subsection{The incidence cone of a directed multigraph}
In this section, we briefly recall classical properties of the vector configuration formed by the columns of the incidence matrix of a directed multigraph. We fix a directed multigraph~$G$ on~$n$ vertices, whose underlying undirected graph is connected.
Let~$(e_1,\dots,e_n)$ be the canonical basis of~$\mathbb{R}^n$ and $1\!\!1\eqdef\sum e_i$.
\begin{definition}
The \defn{incidence configuration} of the directed multigraph~$G$ is the vector configuration~$I(G) \eqdef \set{e_j-e_i}{(i,j)\in G} \subset\mathbb{R}^n$. The \defn{incidence cone} of~$G$ is the cone~$C(G)\subset\mathbb{R}^n$ generated by~$I(G)$, {\it i.e.} its positive span.
\end{definition}
In other words, the incidence configuration of a directed multigraph consists of the column vectors of its incidence matrix. Observe that the incidence cone is contained in the linear subspace of equation $\dotprod{1\!\!1}{x} = 0$. We will use the following relations between the graph properties of~$G$ and the orientation properties of~$I(G)$, which can be summed up by saying that the (oriented and unoriented) matroid of $G$ coincides with that of its incidence configuration $I(G)$. See~\cite{bvswz} for an introduction and reference on oriented matroids. In particular, Section 1.1 of that book explores the incidence configuration of a directed graph.
\begin{remark}\label{rem:incidenceconfiguration}
Consider a subgraph~$H$ of~$G$. Then the vectors of~$I(H)$:
\begin{enumerate}
\item are independent if and only if $H$~has no (not necessarily oriented) cycle, that is, if $H$ is a forest;
\item span the hyperplane~$\dotprod{1\!\!1}{x}= 0$ if and only if $H$~is connected and spanning;
\item form a basis of the hyperplane~$\dotprod{1\!\!1}{x}= 0$ if and only if $H$~is a spanning~tree;
\item form a circuit if and only if $H$~is a (not necessarily oriented) cycle; the positive and negative parts of the circuit correspond to the subsets of edges oriented in one or the other direction along this cycle; in particular,~$I(H)$ is a positive circuit if and only if $H$~is an oriented cycle;
\item form a cocircuit if and only if $H$~is a minimal (not necessarily oriented) cut; the positive and negative parts of the cocircuit correspond to the edges in one or the other direction in this cut; in particular,~$I(H)$ is a positive cocircuit if and only if $H$~is an oriented cut.
\end{enumerate}
\end{remark}
This remark on the incidence configuration translates into the following remark on the incidence cone:
\begin{remark}\label{rem:incidencecone}
Consider a subgraph~$H$ of~$G$. The incidence configuration $I(H)$ is the set of vectors of $I(G)$ contained in a $k$-face of $C(G)$ if and only if $H$ has $n-k$ connected components and the quotient graph $G/H$ is acyclic. In particular:
\begin{enumerate}
\item The cone~$C(G)$ has dimension~$n-1$ (since we assumed that the undirected graph underlying~$G$ is connected).
\item The cone~$C(G)$ is pointed if and only if~$G$ is an acyclic directed graph.
\item If~$G$ is acyclic, it induces a partial order on its set of nodes. The rays of~$C(G)$ correspond to the edges of the Hasse diagram of~$G$. The cone is simple if and only if the Hasse diagram of~$G$ is a tree.
\item The facets of~$C(G)$ correspond to the complements of the minimal directed cuts in~$G$. Given a minimal directed cut in~$G$, the characteristic vector of its sink is a normal vector of the corresponding facet.
\end{enumerate}
\end{remark}
\begin{example}
If~$G$ is the complete directed graph on~$n$ vertices, with one arc from any node to any other (and thus, two arcs between any pair of nodes, one in each direction), then its incidence configuration~$I(G) = \set{e_i-e_j}{i,j \in [n]}$ is the root system of type~$A$. See~\cite{Humphreys}.
If~$G$ is an acyclic orientation on the complete graph, then its incidence configuration~$I(G)$ is a positive system of roots~\cite{Humphreys}. The Hasse diagram of the order induced by~$G$ is a path, thus the incidence cone~$C(G)$ is simple. Its rays are spanned by the simple roots of the positive system~$I(G)$.
\end{example}
\subsection{The vertices of the brick polytope}\label{subsec:vc}
Let~$\mathcal{N}$ be an irreducible sorting network supporting a pseudoline arrangement~$\Lambda$. We use the contact graph~$\Lambda^\#$ to describe the cone of the brick polytope~$\Omega(\mathcal{N})$ at the brick vector~$\omega(\Lambda)$:
\begin{theorem}\label{theo:cones}
The cone of the brick polytope $\Omega(\mathcal{N})$ at the brick vector $\omega(\Lambda)$ is precisely the incidence cone $C(\Lambda^\#)$ of the contact graph~$\Lambda^\#$ of $\Lambda$:
$$\cone \set{\omega(\Lambda')-\omega(\Lambda)}{\Lambda' \in \arr(\mathcal{N})} = \cone \set{e_j-e_i}{(i,j) \in \Lambda^\#}.$$
\end{theorem}
\begin{proof}
Assume that $\Lambda'$ is obtained from $\Lambda$ by flipping a contact from its $i$\textsuperscript{th} pseudoline to its $j$\textsuperscript{th} pseudoline. Then the difference $\omega(\Lambda')-\omega(\Lambda)$ is a positive multiple of $e_j-e_i$. This immediately implies that the incidence cone $C(\Lambda^\#)$ is included in the cone of $\Omega(\mathcal{N})$ at $\omega(\Lambda)$.
Reciprocally, we have to prove that any facet $F$ of the cone $C(\Lambda^\#)$ is also a facet of the brick polytope $\Omega(\mathcal{N})$.
According to Remark \ref{rem:incidencecone}(4), there exists a minimal directed cut from a source set $U$ to a sink set $V$ (which partition the vertices of $\Lambda^\#$) such that $1\!\!1_V \eqdef \sum_{v \in V} e_v$ is a normal vector of $F$. We denote by $\gamma$ the commutators of $\mathcal{N}$ which correspond to the arcs of $\Lambda^\#$ between~$U$ and~$V$. We claim that for any pseudoline arrangement $\Lambda'$ supported by $\mathcal{N}$, the scalar product $\dotprod{1\!\!1_V}{\omega(\Lambda')}$ equals $\dotprod{1\!\!1_V}{\omega(\Lambda)}$ when $\gamma$ is a subset of the contacts of $\Lambda'$, and is strictly bigger than $\dotprod{1\!\!1_V}{\omega(\Lambda)}$ otherwise.
Remember first that the set of all pseudoline arrangements supported by $\mathcal{N}$ and whose set of contacts contains $\gamma$ is connected by flips. Since a flip between two such pseudoline arrangements necessarily involves either two pseudolines of $U$ or two pseudolines of $V$, the corresponding incidence vector is orthogonal to $1\!\!1_V$. Thus, the scalar product $\dotprod{1\!\!1_V}{\omega(\Lambda')}$ is constant on all pseudoline arrangements whose set of contacts contains $\gamma$.
Reciprocally, we consider a pseudoline arrangement $\Lambda'$ supported by $\mathcal{N}$ which minimizes the scalar product $\dotprod{1\!\!1_V}{\omega(\Lambda')}$. There is clearly no arc from $U$ to $V$ in~$\Lambda'^\#$, otherwise flipping the corresponding contact in $\Lambda'$ would decrease the value of $\dotprod{1\!\!1_V}{\omega(\Lambda')}$. We next prove that we can join $\Lambda$ to $\Lambda'$ by flips involving two pseudolines of $U$ or two pseudolines of $V$. As a first step, we show that we can transform $\Lambda$ and $\Lambda'$ into pseudoline arrangements $\hat\Lambda$ and $\hat \Lambda'$ in which the first pseudoline coincide, using only flips involving two pseudolines of $U$ or two pseudolines of $V$. We can then conclude by induction on the number of levels of~$\mathcal{N}$.
Assume first that the first pseudoline (the one which starts at level $1$ and ends at level $n$) of $\Lambda$ and $\Lambda'$ is in $U$. We sweep this pseudoline from left to right in $\Lambda$. If there is a contact above and incident to it, the above pseudoline must be in $U$. Otherwise we would have an arc between $V$ and $U$ in $\Lambda^\#$. Consequently, we are allowed to flip this contact. By doing this again and again we obtain a pseudoline arrangement $\hat\Lambda$ whose first pseudoline starts at the bottom leftmost point and goes up whenever possible until getting to the topmost level. Since this procedures only relies on the absence of arc from $V$ to $U$ in $\Lambda^\#$, we can proceed identically on $\Lambda'$ to get a pseudoline arrangement $\hat\Lambda'$ with the same first pseudoline. Finally, if the first pseudoline of $\Lambda$ and $\Lambda'$ is in~$V$, then we can argue similarly but sweeping the pseudoline from right to left.
\end{proof}
By Remark~\ref{rem:incidencecone}(1) and Proposition~\ref{prop:disconnected}, we obtain the dimension of~$\Omega(\mathcal{N})$:
\begin{corollary}\label{coro:dim}
The brick polytope of an irreducible sorting network with $n$ levels has dimension $n-1$. In general, the brick polytope of a sorting network with $n$ levels and $p$ irreducible components has dimension $n-p$.
\end{corollary}
According to Remark \ref{rem:incidencecone}(2), Theorem~\ref{theo:cones} also characterizes the pseudoline arrangements whose brick vector is a vertex of $\Omega(\mathcal{N})$:
\begin{corollary}\label{coro:vc}
The brick vector $\omega(\Lambda)$ is a vertex of the brick polytope $\Omega(\mathcal{N})$ if and only if the contact graph $\Lambda^\#$ of $\Lambda$ is acyclic.
\end{corollary}
For example, the brick vector of the greedy pseudoline arrangement~$\Gamma(\mathcal{N})$ is always a vertex of~$\Omega(\mathcal{N})$ since its contact graph is sorted (see Example~\ref{exm:contactgreedy}). Similarly, the brick vector of the sink of the oriented graph of decreasing flips is always a vertex of~$\Omega(\mathcal{N})$. These two greedy pseudoline arrangements can be the only vertices of the brick polytope, as happens for $2$-level networks:
\begin{example}[$2$-level networks, continued]
Let $\mathcal{X}_m$ be the sorting network formed by two levels related by $m$ commutators. The contact graph of the pseudoline arrangement whose unique crossing is the $i$\textsuperscript{th} commutator of~$\mathcal{X}_m$ is a multigraph with two vertices and $m-1$ edges, $m-i$ of them in one direction and $i-1$ in the other. Thus, only the first and last commutators give pseudoline arrangements with acyclic contact graphs.
\end{example}
In general, the map~$\omega:\arr(\mathcal{N})\to\mathbb{R}^n$ (which associates to a pseudoline arrangement its brick vector) is not injective on~$\arr(\mathcal{N})$. For example, many interior points appear several times in Examples~\ref{exm:3levels} and~\ref{exm:duplicated1}.
However, the vertices of the brick polytope have precisely one preimage by~$\omega$:
\begin{proposition}\label{prop:vertices}
The map~$\omega:\arr(\mathcal{N})\to\mathbb{R}^n$ restricts to a bijection between the pseudoline arrangements supported by~$\mathcal{N}$ whose contact graphs are acyclic and the vertices of the brick polytope~$\Omega(\mathcal{N})$.
\end{proposition}
\begin{proof}
According to Corollary~\ref{coro:vc}, the map~$\omega$ defines a surjection from the pseudoline arrangements supported by~$\mathcal{N}$ whose contact graphs are acyclic to the vertices of the brick polytope~$\Omega(\mathcal{N})$. To prove injectivity, we use an inductive argument based on the following claims:
\begin{enumerate}[(i)]
\item the greedy pseudoline arrangement~$\Gamma(\mathcal{N})$ is the unique preimage of~$\omega(\Gamma(\mathcal{N}))$;
\item if a vertex of $\Omega(\mathcal{N})$ has a unique preimage by~$\omega$, then so do its neighbors in the graph of~$\Omega(\mathcal{N})$.
\end{enumerate}
To prove~(i), consider a pseudoline arrangement~$\Lambda$ supported by~$\mathcal{N}$ such that $\omega(\Lambda)=\omega(\Gamma(\mathcal{N}))$. According to Theorem~\ref{theo:cones}, the contact graphs~$\Lambda^\#$ and~$\Gamma(\mathcal{N})^\#$ have the same incidence cone, which ensures that all arcs of~$\Lambda^\#$ are sorted. In other words, all flips in~$\Lambda$ are decreasing. Since this property characterizes the greedy pseudoline arrangement, we obtain that~$\Lambda=\Gamma(\mathcal{N})$.
To prove~(ii), consider two neighbors $v,v'$ in the graph of~$\Omega(\mathcal{N})$. Let ${i,j\in[n]}$ be such that $v'-v=\alpha(e_j-e_i)$ for some $\alpha>0$. Let $\Lambda$ be a pseudoline arrangement supported by~$\mathcal{N}$ such that $v=\omega(\Lambda)$. Let $\Lambda'$ denote the pseudoline arrangement obtained from $\Lambda$ by flipping the rightmost contact between its $i$\textsuperscript{th} and $j$\textsuperscript{th} pseudolines if $i<j$ and the leftmost one if $i>j$. Then $v'=\omega(\Lambda')$. In particular, if $v$ has two distinct preimages by $\omega$, then so does $v'$. This proves~(ii).
\end{proof}
\begin{corollary}\label{coro:graph}
The graph of the brick polytope is a subgraph of~$G(\mathcal{N})$ whose vertices are the pseudoline arrangements with acyclic contact graphs.
\end{corollary}
Example~\ref{exm:counterExample} shows that the graph of the brick polytope is not always the subgraph of~$G(\mathcal{N})$ induced by the pseudoline arrangements with acyclic~contact~graphs.
\begin{example}[Duplicated networks, continued]\label{exm:duplicated3}
For a connected graph~$\Gamma$, consider the duplicated network~$\mathcal{Z}(\Gamma)$ defined in Example~\ref{exm:duplicated1}. The contact graphs of the~$2^{|\Gamma|}$ pseudoline arrangements supported by~$\mathcal{Z}(\Gamma)$ are the $2^{|\Gamma|}$ orientations on~$\Gamma$, and two pseudoline arrangements are related by a flip if their contact graphs differ in the orientation of a single edge of~$\Gamma$. According to Proposition~\ref{prop:vertices}, the vertices of the brick polytope~$\Omega(\mathcal{Z}(\Gamma))$ correspond to the acyclic orientations on~$\Gamma$.
\begin{figure}[b]
\capstart
\centerline{\includegraphics[scale=.6]{permutahedron}}
\caption{The brick polytope~$\Omega(\mathcal{Z}(K_n))$ is (a translate of) the permutahedron~$\Pi_n$. For space reason, we have represented only certain arrangements. The rest of the drawing is left to the reader.}
\label{fig:permutahedron}
\end{figure}
When~$\Gamma=K_n$ is the complete graph on $n$ vertices, the contact graphs of the pseudoline arrangements supported by $\mathcal{Z}(K_n)$ are the tournaments on $[n]$, the vertices of $\Omega(\mathcal{Z}(K_n))$ correspond to the permutations of~$[n]$, and the graph of $\Omega(\mathcal{Z}(K_n))$ is a subgraph of that of the permutahedron~$\Pi_n=\conv\set{(\sigma(1),\dots,\sigma(n))^T}{\sigma\in\mathfrak{S}_n}$. Since $\Pi_n$ is simple and both $\Omega(\mathcal{Z}(K_n))$ and $\Pi_n$ have dimension~$n-1$, they must have in fact the same graph, and consequently the same combinatorial structure (by simplicity~\cite{BlindMani, Kalai}). In fact, $\Omega(\mathcal{Z}(K_n))$ is a translate of~$\Pi_n$. See \fref{fig:permutahedron}.
When $\Gamma$ is a tree, all possible orientations on $\Gamma$ are acyclic. The brick polytope $\Omega(\mathcal{Z}(\Gamma))$ is thus a cube.
\end{example}
\begin{remark}[Simple and non-simple vertices]
According to Remark~\ref{rem:incidencecone}(3), the brick vector $\omega(\Lambda)$ of a pseudoline arrangement~$\Lambda$ supported by~$\mathcal{N}$ is a simple vertex of~$\Omega(\mathcal{N})$ if and only if the contact graph~$\Lambda^\#$ is acyclic and its Hasse diagram is a tree. See \fref{fig:counterExample} for a brick polytope with non-simple vertices.
\end{remark}
\subsection{The faces of the brick polytope}\label{subsec:fd}
Let $\mathcal{N}$ be an irreducible sorting network. Theorem~\ref{theo:cones} and Remark~\ref{rem:incidencecone}(4) provide the facet description of the brick polytope~$\Omega(\mathcal{N})$:
\begin{corollary}\label{coro:fd}
The facet normal vectors of the brick polytope $\Omega(\mathcal{N})$ are precisely all facet normal vectors of the incidence cones of the contact graphs of the pseudoline arrangements supported by $\mathcal{N}$. Representatives for them are given by the characteristic vectors of the sinks of the minimal directed cuts of these~contact~graphs.
\end{corollary}
\begin{remark}
Since we know its vertices and its facet normal vectors, we obtain immediately the complete inequality description of the brick polytope. More precisely, for each given pseudoline arrangement~$\Lambda$ supported by~$\mathcal{N}$ with an acyclic contact graph~$\Lambda^\#$, and for each minimal directed cut in~$\Lambda^\#$ with source~$U$ and sink~$V$, the right-hand-side of the inequality of the facet with normal vector~$1\!\!1_V \eqdef \sum_{v \in V} e_v$ is given by the sum, over all pseudolines~$\ell$ of~$\Lambda$ in~$V$, of the number of bricks below~$\ell$.
\end{remark}
More generally, Theorem~\ref{theo:cones} implies a combinatorial description of the faces of the brick polytope. We need the following definition:
\begin{definition}
A set $\gamma$ of commutators of~$\mathcal{N}$ is \defn{$k$-admissible} if there exists a pseudoline arrangement $\Lambda\in\arr(\mathcal{N}|\gamma)$ such that $\Lambda^\#\smallsetminus\gamma^\#$ has $n-k$ connected components and $\Lambda^\#/(\Lambda^\#\smallsetminus\gamma^\#)$ is acyclic (where $\gamma^\#$ denotes the subgraph of~$\Lambda^\#$ corresponding to the commutators of~$\gamma$).
\end{definition}
Theorem~\ref{theo:cones}, Remark~\ref{rem:incidencecone}, and Proposition~\ref{prop:vertices} lead to our face description:
\begin{corollary}\label{coro:faces}
Let $\Phi$ be the map which associates to a subset $X$ of $\mathbb{R}^n$ the set of commutators of $\mathcal{N}$ which are contacts in all the pseudoline arrangements supported by~$\mathcal{N}$ whose brick vectors lie in~$X$. Let $\Psi$ be the map which associates to a set~$\gamma$ of commutators of~$\mathcal{N}$ the convex hull of $\set{\omega(\Lambda)}{\Lambda\in\arr(\mathcal{N}|\gamma)}$. Then the maps $\Phi$ and $\Psi$ define inverse bijections between the $k$-faces of~$\Omega(\mathcal{N})$ and the $k$-admissible sets of commutators of~$\mathcal{N}$.
\end{corollary}
\begin{example}[Duplicated networks, continued]\label{exm:duplicated4}
For a connected graph~$\Gamma$ on~$[n]$, consider the duplicated network~$\mathcal{Z}(\Gamma)$ defined in Example~\ref{exm:duplicated1}. According to Corollary~\ref{coro:faces}, the $k$-faces of $\Omega(\mathcal{Z}(\Gamma))$ are in bijection with the couples $(\pi,\sigma)$ where $\pi$ is a set of~$n-k$ connected induced subgraphs of~$\Gamma$ whose vertex sets partition~$[n]$, and $\sigma$ is an acyclic orientation on the quotient $\Gamma/\bigcup \pi$.
When $\Gamma=K_n$ is the complete graph on $n$ vertices, the $k$-faces of the brick polytope $\Omega(\mathcal{Z}(K_n))$ correspond to the ordered $(n-k)$-partitions of~$[n]$. This confirms that $\Omega(\mathcal{Z}(K_n))$ is combinatorially equivalent to the permutahedron $\Pi_n$
When~$\Gamma$ is a tree, we obtain a $k$-face by choosing $n-k$ connected subgraphs of~$\Gamma$ whose vertex sets partition~$[n]$ and an arbitrary orientation on the remaining edges. This clearly corresponds to the $k$-faces of the cube.
\end{example}
We now apply our results to minimal irreducible sorting networks (which support pseudoline arrangements whose contact graphs are trees~---~Definition~\ref{def:minimal}). For these networks, we obtain an affirmative answer to Question~\ref{qu:polytope}:
\begin{theorem}
For any minimal irreducible sorting network~$\mathcal{N}$, the simplicial complex~$\Delta(\mathcal{N})$ is the boundary complex of the polar of the brick polytope~$\Omega(\mathcal{N})$. In particular, the graph of $\Omega(\mathcal{N})$ is the flip graph $G(\mathcal{N})$.
\end{theorem}
\begin{proof}
Let~$\gamma$ be a set of $p$~commutators such that $\mathcal{N}\smallsetminus\gamma$ is still sorting, and let $\Lambda\in\arr(\mathcal{N}|\gamma)$. Since the contact graph of $\Lambda$ is an oriented tree, its subgraph $\Lambda^\#\smallsetminus\gamma^\#$ has $p+1$ connected components and its quotient $\Lambda^\#/(\Lambda^\#\smallsetminus\gamma^\#)$ is acyclic. Consequently, $\Omega(\mathcal{N})$ has a $(n-p-1)$-dimensional face corresponding to~$\gamma$.
\end{proof}
\begin{example}
When~$\Gamma$ is a tree, the polar of the cube~$\Omega(\mathcal{Z}(\Gamma))$ realizes~$\Delta(\mathcal{Z}(\Gamma))$.
\end{example}
Reciprocally, note that the dimension of the brick polytope~$\Omega(\mathcal{N})$ does not even match that of the simplicial complex~$\Delta(\mathcal{N})$ for an irreducible sorting network~$\mathcal{N}$ which is not minimal. Consequently, the minimal networks are precisely those irreducible networks for which the brick polytope provides a realization of $\Delta(\mathcal{N})$. We discuss in more details the relationship between the boundary complex of the brick polytope and the simplicial complex $\Delta(\mathcal{N})$ for certain non minimal networks in Section~\ref{sec:multi}.
\begin{example}\label{exm:counterExample}
We finish our combinatorial description of the face structure of the brick polytope with the example of the sorting network represented in \fref{fig:counterExample}. It illustrates that:
\begin{enumerate}
\item All pseudoline arrangements supported by~$\mathcal{N}$ can appear as vertices of~$\Omega(\mathcal{N})$ even for a non-minimal network~$\mathcal{N}$.
\item Even when all pseudoline arrangements supported by~$\mathcal{N}$ appear as vertices of~$\Omega(\mathcal{N})$, the graph can be a strict subgraph of $G(\mathcal{N})$.
\item The brick vector of a pseudoline arrangement~$\Lambda$ supported by~$\mathcal{N}$ is a simple vertex of~$\Omega(\mathcal{N})$ if and only if the contact graph~$\Lambda^\#$ is acyclic and its Hasse diagram is a tree.
\end{enumerate}
\begin{figure}
\capstart
\centerline{\includegraphics[width=\textwidth]{counterExample}}
\caption{The Schlegel diagram of the brick polytope~$\Omega(\mathcal{N})$ of a sorting network~$\mathcal{N}$ with $4$ levels and $10$ commutators. Since the graph of flips~$G(\mathcal{N})$ is that of the $4$-dimensional cube, $\Omega(\mathcal{N})$ has no missing vertex but $4$ missing edges (find them!).}
\label{fig:counterExample}
\end{figure}
\end{example}
\subsection{Brick polytopes as (positive) Minkowski sums}\label{subsec:minkowskiSum}
Let~$\mathcal{N}$ be a sorting network with $n$ levels and let~$b$ be a brick of~$\mathcal{N}$. For any pseudoline arrangement~$\Lambda$ supported by~$\mathcal{N}$, we denote by~$\omega(\Lambda,b)\in\mathbb{R}^n$ the characteristic vector of the pseudolines of~$\Lambda$ passing above~$b$. We associate to the brick~$b$ of ~$\mathcal{N}$ the polytope
$$\Omega(\mathcal{N},b) \eqdef \conv\set{\omega(\Lambda,b)}{\Lambda \in \arr(\mathcal{N})}\subset\mathbb{R}^n.$$
These polytopes provide a Minkowski sum decomposition of the brick polytope~$\Omega(\mathcal{N})$:
\begin{proposition}\label{prop:minkowskiSum}
The brick polytope~$\Omega(\mathcal{N})$ is the Minkowski sum of the polytopes~$\Omega(\mathcal{N},b)$ associated to all the bricks $b$ of~$\mathcal{N}$.
\end{proposition}
\begin{proof}
Since $\omega(\Lambda) = \sum \omega(\Lambda,b)$ for any pseudoline arrangement~$\Lambda$ supported by~$\mathcal{N}$, $\Omega(\mathcal{N})$ is included in the Minkowski sum~$\sum \Omega(\mathcal{N},b)$. To prove equality, we thus only have to prove that any vertex of~$\sum \Omega(\mathcal{N},b)$ is also a vertex of~$\Omega(\mathcal{N})$.
Let $f:\mathbb{R}^n\to\mathbb{R}$ be a linear function, and $\Lambda, \Lambda'$ be two pseudoline arrangements related by a flip involving their $i$\textsuperscript{th} and $j$\textsuperscript{th} pseudolines. If a brick~$b$ of~$\mathcal{N}$ is not located between the $i$\textsuperscript{th} pseudolines of~$\Lambda$ and~$\Lambda'$, then $f(\omega(\Lambda,b)) = f(\omega(\Lambda',b))$. Otherwise, the variation $f(\omega(\Lambda,b)) - f(\omega(\Lambda',b))$ has the same sign as the variation $f(\omega(\Lambda)) - f(\omega(\Lambda'))$. Consequently, the pseudoline arrangement~$\Lambda_f$ supported by~$\mathcal{N}$ which minimizes $f$ on $\Omega(\mathcal{N})$, also minimizes $f$ on $\Omega(\mathcal{N},b)$ for each brick~$b$ of~$\mathcal{N}$.
Now let $v$ be any vertex of~$\sum \Omega(\mathcal{N},b)$. Let $f:\mathbb{R}^n\to\mathbb{R}$ denote a linear function which is minimized by~$v$ on $\sum \Omega(\mathcal{N},b)$. Then $v$ is the sum of the vertices which minimize~$f$ in each summand $\Omega(\mathcal{N},b)$. Consequently, we obtain that $v=\sum\omega(\Lambda_f,b)=\omega(\Lambda_f)$ is a vertex of~$\Omega(\mathcal{N})$.
\end{proof}
\begin{remark}
This Minkowski sum decomposition comes from the fact that the summation and convex hull in the definition of the brick polytope commute:
$$\Omega(\mathcal{N}) \eqdef \conv_\Lambda \sum\nolimits_b \omega(\Lambda,b) \,=\, \sum\nolimits_b \conv_\Lambda \omega(\Lambda,b) \eqfed \sum\nolimits_b \Omega(\mathcal{N},b),$$
where the index~$b$ of the sums ranges over the bricks of~$\mathcal{N}$ and the index~$\Lambda$ of the convex hulls ranges over the pseudoline arrangements supported by~$\mathcal{N}$.
\end{remark}
Observe that the vertex set of~$\Omega(\mathcal{N},b)$ is contained in the vertex set of a hypersimplex, since the number of pseudolines above~$b$ always equals the depth of~$b$. Furthermore, since~$\Omega(\mathcal{N},b)$ is a summand in a Minkowski decomposition of~$\Omega(\mathcal{N})$, all the edges of the former are parallel to that of the latter, and thus to that of the standard simplex $\triangle_{[n]} \eqdef \conv\set{e_i}{i\in [n]}$. Consequently, the polytope~$\Omega(\mathcal{N},b)$ is a \defn{matroid polytope}, \ie the convex set of the characteristic vectors of the bases of a matroid (the ground set of this matroid is the set of pseudolines and its rank is the depth of~$b$).
\begin{example}[$2$-level networks, continued]
Let $\mathcal{X}_m$ be the sorting network formed by two levels and $m$ commutators, and let~$b$ be a brick of~$\mathcal{X}_m$. For every pseudoline arrangement~$\Lambda$ supported by~$\mathcal{X}_m$, we have $\omega(\Lambda,b)=(1,0)$ if the two pseudolines of~$\Lambda$ cross before~$b$ and $\omega(\Lambda,b)=(0,1)$ otherwise. Thus, the polytope $\Omega(\mathcal{X}_m,b)$ is the segment with endpoints $(1,0)$ and $(0,1)$, and the brick polytope~$\Omega(\mathcal{X}_m)$ is the Minkowski sum of $m-1$ such segments.
\end{example}
\begin{example}[Duplicated networks, continued]\label{exm:duplicated5}
For a connected graph~$\Gamma$, consider the duplicated network~$\mathcal{Z}(\Gamma)$ obtained by duplicating the commutators of a reduced network~$\mathcal{N}$ according to the edges of~$\Gamma$~---~see Example~\ref{exm:duplicated1}. This network has two kinds of bricks: those located between two adjacent commutators (which replace a commutator of~$\mathcal{N}$) and the other ones (which correspond to the bricks of~$\mathcal{N}$). For any brick~$b$ of the latter type, the polytope $\Omega(\mathcal{Z}(\Gamma),b)$ is still a single point. Now let~$b$ be a brick of~$\mathcal{Z}(\Gamma)$ located between a pair of adjacent commutators corresponding to the edge $\{i,j\}$ of~$\Gamma$. Then $\Omega(\mathcal{Z}(\Gamma),b)$ is (a translate of) the segment~$[e_i,e_j]$. Summing the contributions of all bricks, we obtain that the brick polytope~$\Omega(\mathcal{Z}(\Gamma))$ is the Minkowski sum of all segments $[e_i,e_j]$ for $\{i,j\}\in \Gamma$. Such a polytope is usually called a \defn{graphical zonotope}. When $\Gamma = K_n$, the permutahedron $\Pi_n = \Omega(\mathcal{Z}(K_n))$ is the Minkowski sum of the segments $[e_i,e_j]$ for all distinct $i,j\in[n]$. When $\Gamma$ is a tree, the cube $\Omega(\mathcal{Z}(\Gamma))$ is the Minkowski sum of linearly independent segments.
\end{example}
\subsection{Brick polytopes and generalized permutahedra}\label{subsec:generalizedPermutahedra}
Our brick polytopes are instances of a well-behaved class of polytopes studied in~\cite{Postnikov, ArdilaBenedettiDoker, PostnikovReinerWilliams}:
\begin{definition}[\cite{Postnikov}]
A \defn{generalized permutahedra} is a polytope whose inequality description is of the form:
$$\GP{Z}{z_I} \eqdef \set{\begin{pmatrix} x_1 \\ \vdots \\ x_n\end{pmatrix}\in\mathbb{R}^n}{\sum_{i=1}^n x_i = z_{[n]} \text{ and } \sum_{i\in I} x_i \ge z_I \text{ for } I\subset[n]}$$
for a family $\{z_I\}_{I\subset[n]}\in\mathbb{R}^{2^{[n]}}$ such that $z_I + z_J \le z_{I \cup J} + z_{I \cap J}$ for all $I,J \subset [n]$.
\end{definition}
In other words, a generalized permutahedron is obtained as a deformation of the classical permutahedron by moving its facets while keeping the direction of their normal vectors and staying in its deformation cone~\cite{PostnikovReinerWilliams}.
This family of polytopes contains many relevant families of combinatorial polytopes: permutahedra, associahedra, cyclohedra (and more generally, all graph-asso\-ciahedra~\cite{CarrDevadoss}), etc.
Since $\sGP{Z}{z_I} + \sGP{Z}{z'_I} = \sGP{Z}{z_I+z'_I}$, the class of generalized associahedra is closed by Minkowski sum and difference (a Minkowski difference $P-Q$ of two polytopes $P,Q\subset\mathbb{R}^n$ is defined only if there exists a polytope $R$ such that ${P = Q + R}$).
Consequently, for any $\{y_I\}_{I\subset[n]}\in\mathbb{R}^{2^{[n]}}$, the Minkowski sum and difference
$$\GP{Y}{y_I} \eqdef \sum_{I\subset[n]} y_I\triangle_I$$
of faces $\triangle_I \eqdef \conv\set{e_i}{i\in I}$ of the standard simplex $\triangle_{[n]}$ is a generalized permutahedron. Reciprocally, it turns out that any generalized permutahedron $\sGP{Z}{z_I}$ can be decomposed as such a Minkowski sum and difference $\sGP{Y}{y_I}$, and that $\{y_I\}$ is derived from $\{z_I\}$ by M\"obius inversion when all the inequalities defining $\sGP{Z}{z_I}$ are tight:
\begin{proposition}[\cite{Postnikov, ArdilaBenedettiDoker}]\label{prop:minkowskiSumDiff}
Every generalized permutahedron can be written uniquely as a Minkowski sum and difference of faces of the standard simplex:
$$\GP{Z}{z_I} = \GP{Y}{y_I}, \quad \text{where} \quad y_I = \sum_{J \subset I} (-1)^{|I\smallsetminus J|}z_J$$
if all inequalities $\sum_{i\in I} x_i\ge z_I$ are tight.
\end{proposition}
\begin{example}
The classical permutahedron can be written as
$$\Pi_n = \conv\set{(\sigma(1), \dots, \sigma(n))^T}{\sigma\in\mathfrak{S}_n} = \GP{Z}{\frac{|I|(|I|+1)}{2}} = \sum_{|I| = 2} \triangle_I.$$
\end{example}
The Minkowski decomposition of Proposition~\ref{prop:minkowskiSumDiff} is useful in particular to compute the volume of the generalized permutahedra~\cite{Postnikov}.
All our brick polytopes are generalized permutahedra (as Minkowski sums of matroid polytopes).
It raises three questions about our construction:
\begin{question}
Which generalized permutahedra are brick polytopes?
\end{question}
For example, we obtain all graphical zonotopes (Example~\ref{exm:duplicated1}) and all associahedra (Section~\ref{sec:associahedra}).
However, the only $2$-dimensional brick polytopes are the square, the pentagon and the hexagon: no brick polytope is a triangle.
\begin{question}\label{qu:coeffsMinkowskiSumDiff}
How to compute efficiently the coefficients $\{y_I\}$ in the Minkowski decomposition of a brick polytope~$\Omega(\mathcal{N})$ into dilates of faces of the standard simplex?
\end{question}
Lange studies this question in~\cite{Lange} for all associahedra of Hohlweg and Lange~\cite{HohlwegLange} (see also Remark~\ref{rem:minkowskiSumAssociahedra}). In general, observe that the Minkowski sum decomposition we obtain in Proposition~\ref{prop:minkowskiSum} has only positive coefficients, but its summands are matroid polytopes which are not necessarily simplices. In contrast, the Minkowski decomposition of Proposition~\ref{prop:minkowskiSumDiff} has nice summands but requires sums and differences. In general, the two decompositions of Propositions~\ref{prop:minkowskiSum} and~\ref{prop:minkowskiSumDiff} are therefore different. This last observation raises an additional question:
\begin{question}
Which generalized permutahedra can be written as a Minkowski sum of matroid polytopes?
\end{question}
\section{Hohlweg \& Lange's associahedra, revisited}\label{sec:associahedra}
In this section, we first recall the duality between triangulations of a convex polygon and pseudoline arrangements supported by the $1$-kernel of a reduced alternating sorting network~\cite{PilaudPocchiola}. Based on this duality, we observe that the brick polytopes of these particular networks specialize to Hohlweg \& Lange's realizations of the associahedron~\cite{HohlwegLange}.
\subsection{Duality}\label{subsec:duality}
Remember that we call reduced alternating sorting network any network with $n$ levels and ${n \choose 2}$ commutators and such that the commutators adjacent to each intermediate level are alternatively located above and below it. Such a network supports a unique pseudoline arrangement, whose first and last pseudolines both touch the top and the bottom level, and whose intermediate pseudolines all touch either the top or the bottom level.
To a word~$x\in\{a,b\}^{n-2}$, we associate two dual objects~---~see \fref{fig:alternating}:
\begin{enumerate}
\item $\mathcal{N}_x$ denotes the reduced alternating sorting network such that the $(i+1)$\textsuperscript{th} pseudoline touches its top level if $x_i=a$ and its bottom level if $x_i=b$, for all~${i\in[n-2]}$.
\item $\mathcal{P}_x$ denotes the $n$-gon obtained as the convex hull of $\set{p_i}{i\in[n]}$ where $p_1=(1,0)$, $p_n=(n,0)$ and $p_{i+1}$ is the point of the circle of diameter~$[p_1,p_n]$ with abscissa~$i+1$ and located above~$[p_1,p_n]$ if $x_i=a$ and below~$[p_1,p_n]$ if~$x_i=b$, for all $i\in[n-2]$.
\end{enumerate}
For any distinct~$i,j\in[n]$, we naturally both label by~$\{i,j\}$ the diagonal $[p_i,p_j]$ of~$\mathcal{P}_x$ and the commutator of~$\mathcal{N}_x$ where cross the $i$\textsuperscript{th} and $j$\textsuperscript{th} pseudolines of the unique pseudoline arrangement supported by~$\mathcal{N}_x$. Note that the commutators incident to the first and last level of~$\mathcal{N}_x$ correspond to the edges of the convex hull of~$\mathcal{P}_x$.
In \fref{fig:alternating}, we have represented~$\mathcal{N}_x$ and~$\mathcal{P}_x$ for $x\in\{bbb, aab, aba\}$. The five missing reduced alternating sorting networks with $5$ levels are obtained by reflection of these three with respect to the horizontal or vertical axis.
\begin{figure}[b]
\capstart
\centerline{\includegraphics[width=\textwidth]{alternating}}
\caption{The networks~$\mathcal{N}_x$ and the polygons~$\mathcal{P}_x$ for the words $bbb$, $aab$ and $aba$. The leftmost sorting network is the ``bubble sort'', while the rightmost is the ``even-odd transposition sort''.}
\label{fig:alternating}
\end{figure}
We call \defn{$1$-kernel} of a network~$\mathcal{N}$ the network~$\mathcal{N}^1$ obtained from~$\mathcal{N}$ by erasing its first and last horizontal lines, as well as all commutators incident to them. For a word $x\in\{a,b\}^{n-2}$, the network~$\mathcal{N}_x^1$ has~$n-2$ levels and~${n \choose 2}-n$ commutators. Since we erased the commutators between consecutive pseudolines on the top or bottom level of~$\mathcal{N}_x$, the remaining commutators are labeled by the internal diagonals of~$\mathcal{P}_x$. The pseudoline arrangements supported by~$\mathcal{N}_x^1$ are in bijection with the triangulations of~$\mathcal{P}_x$ through the following duality~---~see \fref{fig:dualite}:
\begin{proposition}[\cite{PilaudPocchiola}]\label{prop:duality}
Fix a word~$x\in\{a,b\}^{n-2}$. The set of commutators of~$\mathcal{N}_x$ labeled by the internal diagonals of a triangulation~$T$ of $\mathcal{P}_x$ is the set of contacts of a pseudoline arrangement~$T^*$ supported by $\mathcal{N}_x^1$. Reciprocally the internal diagonals of $\mathcal{P}_x$ which label the contacts of a pseudoline arrangement supported by~$\mathcal{N}_x^1$ form a triangulation of $\mathcal{P}_x$.
\end{proposition}
\begin{figure}
\capstart
\centerline{\includegraphics[scale=1]{dualite}}
\caption{A triangulation $T$ of $\mathcal{P}_{aab}$ and its dual pseudoline arrangement $T^*$ on the $1$-kernel of the sorting network~$\mathcal{N}_{aab}$.}
\label{fig:dualite}
\end{figure}
The dual pseudoline arrangement $T^*$ of a triangulation $T$ of $\mathcal{P}_x$ has one pseudoline $\Delta^*$ dual to each triangle $\Delta$ of $T$. A commutator is the crossing between two pseudolines $\Delta^*$ and $\Delta'^*$ of $T^*$ if it is labeled by the common bisector of the triangles $\Delta$ and $\Delta'$ (a \defn{bisector} of $\Delta$ is an edge incident to one vertex of $\Delta$ and which separates its remaining two vertices). A commutator is a contact between $\Delta^*$ and $\Delta'^*$ if it is labeled by a common edge of $\Delta$ and $\Delta'$. Consequently, this duality defines an isomorphism between the graph of flips on pseudoline arrangements supported by~$\mathcal{N}_x^1$ and the graph of flips on triangulations of~$\mathcal{P}_x$. Furthermore, we obtain the following interpretation of the contact graph of~$T^*$:
\begin{lemma}\label{lem:tree}
The contact graph $(T^*)^\#$ of the dual pseudoline arrangement $T^*$ of a triangulation~$T$ is precisely the dual tree of $T$, with some orientations on the edges.
\end{lemma}
\begin{remark}
This duality can be extended to any reduced sorting network. First, a reduced sorting network~$\mathcal{N}$ with $n$ levels can be seen as the dual arrangement of a set~$\mathcal{P}$ of $n$ points in a topological plane. Second, the pseudoline arrangements which cover the $1$-kernel of~$\mathcal{N}$ correspond to the pseudotriangulations of~$\mathcal{P}$. We refer to~\cite{PilaudPocchiola} or~\cite[Chapter~3]{Pilaud} for details. In Section~\ref{sec:multi}, we recall another similar duality between $k$-triangulations of the $n$-gon and pseudoline arrangements supported by the $k$-kernel of a reduced alternating sorting network with $n$ levels.
\end{remark}
\subsection{Associahedra}
Let $\mathcal{N}_x$ be a reduced alternating sorting network with $n$ levels. According to Proposition~\ref{prop:duality} and Lemma~\ref{lem:tree}, its $1$-kernel~$\mathcal{N}_x^1$ is a minimal network: the pseudoline arrangements it supports correspond to triangulations of~$\mathcal{P}_x$ and their contact graphs are the dual trees of these triangulations (with some additional orientations). Consequently, the brick polytope~$\Omega(\mathcal{N}_x^1)$ is a simple $(n-3)$-dimensional polytope whose graph is isomorphic to the graph of flips~$G(\mathcal{N}_x^1)$. Since this graph is isomorphic to the graph of flips on triangulations of the polygon~$\mathcal{P}_x$, we obtain many realizations of the $(n-3)$-dimensional associahedron with integer coordinates:
\begin{proposition}
For any word~$x\in\{a,b\}^{n-2}$, the simplicial complex of crossing-free sets of internal diagonals of the convex $n$-gon~$\mathcal{P}_x$ is (isomorphic to) the boundary complex of the polar of~$\Omega(\mathcal{N}_x^1)$. Thus, the brick polytope~$\Omega(\mathcal{N}_x^1)$ is a realization of the $(n-3)$-dimensional associahedron.
\end{proposition}
It turns out that these polytopes coincide up to translation with the associahedra of Hohlweg \& Lange~\cite{HohlwegLange}. For completeness, let us recall their construction (we adapt notations to fit our presentation). Consider the polygon~$\mathcal{P}_x$ associated to a word~$x\in\{a,b\}^{n-2}$, and let~$T$ be a triangulation of~$\mathcal{P}_x$. For~$j \in [n-2]$, there is a unique triangle~$\Delta_j(T)$ of~$T$ with vertices~$u < j+1 < v$. Let~$\pi_j(T)$ denote the product of the number of edges of~$\mathcal{P}_x$ between $u$ and~$j+1$ by the number of edges of~$\mathcal{P}_x$ between~$j+1$ and~$v$. Associate to the triangulation~$T$ the vector~$\omega(T)$ whose $j$\textsuperscript{th} coordinate is~$\pi_j(T)$ if $x_{j+1} = b$ and~$n+1-\pi_j(T)$ if $x_{j+1} = a$. The associahedra of Hohlweg \& Lange~\cite{HohlwegLange} is the convex hull of the vectors~$\omega(T)$ associated to the $\frac{1}{n-1}{2n-4 \choose n-2}$ triangulations of~$\mathcal{P}_x$. It is straightforward to check that our duality from~$T$ to~$T^*$ maps~$\Delta_j(T)$ to the $j$\textsuperscript{th} pseudoline of~$T^*$, and the vector~$\omega(T)$ to our brick vector~$\omega(T^*)$, up to a constant translation.
\begin{figure}[b]
\capstart
\centerline{\includegraphics[width=1\textwidth]{associahedra2B}}
\caption{The brick polytopes~$\Omega(\mathcal{N}_{b^4}^1)$ (left) and $\Omega(\mathcal{N}_{a^2b^2}^1)$ (right) provide two different realizations of the $3$-dimensional associahedron. The convex hull of the brick vectors of the centrally symmetric triangulations of~$\mathcal{P}_{a^2b^2}$ (colored in the picture) is a realization of the $2$-dimensional cyclohedron.}
\label{fig:associahedra2B}
\end{figure}
Observe that the associahedron~$\Omega(\mathcal{N}_x^1)$ does not depend on the first and last letters of~$x$ since we erase the first and last levels of~$\mathcal{N}_x$. Furthermore, a network~$\mathcal{N}_x$ and its reflection~$v(\mathcal{N}_x)$ (resp.~$h(\mathcal{N}_x)$) through the vertical (resp.~horizontal) axis give rise to affinely equivalent associahedra. Affine equivalence between these associahedra is studied in~\cite{bhlt}. Two non-affinely equivalent $3$-dimensional associahedra are presented in \fref{fig:associahedra2B}.
\begin{example}\label{exm:loday}
We obtain Loday's realization of the $(n-3)$-dimensional associahedron~\cite{Loday} as a translate of the brick polytope of the $1$-kernel of the ``bubble sort'' network~$\mathcal{B}_n \eqdef \mathcal{N}_{b^{n-2}}$ associated to the word $b^{n-2}$~---~see \fref{fig:associahedra2B} (left).
\end{example}
We now describe normal vectors for the facets of these associahedra. For any word ${x\in\{a,b\}^{n-2}}$, the facets of the brick polytope~$\Omega(\mathcal{N}_x^1)$ are in bijection with the commutators of~$\mathcal{N}_x^1$. The vertices of the facet corresponding to a commutator~$\gamma$ are the brick vectors of the pseudoline arrangements supported by~$\mathcal{N}_x^1$ and with a contact at~$\gamma$. We have already seen that a normal vector of this facet is given by the characteristic vector of the cut induced by~$\gamma$ in the contact graphs of the pseudoline arrangements supported by~$\mathcal{N}_x^1$ and with a contact at~$\gamma$. In the following lemma, we give an additional description of this characteristic vector:
\begin{lemma}\label{lem:facetsAssociahedra}
Let $\Lambda$ be a pseudoline arrangement supported by~$\mathcal{N}_x^1$ and let $\gamma$ be a contact of~$\Lambda$. The arc corresponding to~$\gamma$ is a cut of the contact graph~$\Lambda^\#$ which separates the pseudolines of~$\Lambda$ passing above~$\gamma$ from those passing below~$\gamma$.
\end{lemma}
\begin{proof}
Let $T$ be the triangulation of~$\mathcal{P}_x$ such that $T^*=\Lambda$. Let $\Delta$ and $\Delta'$ be two triangles of~$T$ whose dual pseudolines~$\Delta^*$ and~$\Delta'^*$ pass respectively above and below~$\gamma$. Then $\Delta$ and $\Delta'$ are located on opposite sides of the edge~$\gamma$ of~$T$, and thus, they cannot share an edge, except if it is~$\gamma$ itself. Consequently, their dual pseudolines~$\Delta^*$ and~$\Delta'^*$ cannot have a contact, except if it is~$\gamma$ itself. We obtain that~$\gamma$ is the only contact between the pseudolines of~$\Lambda$ passing above~$\gamma$ and those passing below~$\gamma$. This implies the lemma.
\end{proof}
\begin{remark}
The pseudolines passing below~$\gamma$ in any pseudoline arrangement supported by~$\mathcal{N}_x^1$ with a contact at~$\gamma$ are also the pseudolines of the greedy pseudoline arrangement~$\Gamma(\mathcal{N}_x^1)$ which pass below the cell immediately to the left of~$\gamma$. This provides a fast method to obtain a normal vector for each facet of $\Omega(\mathcal{N}_x^1)$~---~see \fref{fig:normalvectors} for illustration.
\begin{figure}[h]
\capstart
\centerline{\includegraphics[scale=1]{normalvectors}}
\caption{Normal vectors for the facets of the associahedra~$\Omega(\mathcal{N}_{b^4}^1)$ and~$\Omega(\mathcal{N}_{a^2b^2}^1)$, read on the greedy pseudoline arrangements~$\Gamma(\mathcal{N}_{b^4}^1)$ and~$\Gamma(\mathcal{N}_{a^2b^2}^1)$.}
\label{fig:normalvectors}
\end{figure}
\end{remark}
\begin{corollary}
For any~$x\in\{a,b\}^{n-2}$, the brick polytope~$\Omega(\mathcal{N}_x^1)$ has $n-3$ pairs of parallel facets. The diagonals of~$\mathcal{P}_x$ corresponding to two parallel facets of~$\Omega(\mathcal{N}_x^1)$ are crossing.
\end{corollary}
\begin{figure}
\capstart
\centerline{\includegraphics[width=1\textwidth]{associahedra2C}}
\caption{Pairs of parallel facets in the associahedra of \fref{fig:associahedra2B}. Since they are arranged differently (for example, there is a vertex adjacent to none of these facets in the rightmost realization), these two associahedra are not affinely equivalent.}
\label{fig:associahedra2C}
\end{figure}
\begin{proof}
For $i\in[n-3]$, the normal vector of the facet corresponding to the leftmost contact between the $i$\textsuperscript{th} and $(i+1)$\textsuperscript{th} level of~$\mathcal{N}_x^1$ is always the vector $f_i \eqdef \sum_{j=1}^i e_j$ while the normal vector of the facet corresponding to the rightmost contact between the $(n-i-2)$\textsuperscript{th} and $(n-i-1)$\textsuperscript{th} level of~$\mathcal{N}_x^1$ is always the vector $\sum_{j=i+1}^{n-2} e_j = 1\!\!1 - f_i$ ---~see \fref{fig:normalvectors}. Since~$\Omega(\mathcal{N}_x^1)$ is orthogonal to~$1\!\!1$, these two facets are thus parallel. We obtain $n-3$ pairs of parallel facets when $i$ varies from $1$ to $n-3$. Finally, since two parallel facets of~$\Omega(\mathcal{N}_x^1)$ have no vertex in common, the corresponding diagonals of~$\mathcal{P}_x$ are necessarily crossing each other.
\end{proof}
\begin{remark}\label{rem:minkowskiSumAssociahedra}
As discussed in Sections~\ref{subsec:minkowskiSum} and~\ref{subsec:generalizedPermutahedra}, the associahedron~$\Omega(\mathcal{N}_x^1)$ has two different Minkowski decompositions: as a positive Minkowski sum of the polytopes~$\Omega(\mathcal{N}_x^1,b)$ associated to each brick~$b$ of~$\mathcal{N}_x^1$, or as a Minkowski sum and difference of faces of the standard simplex~$\triangle_{[n-2]}$.
In Loday's associahedron (\ie when~$x=b^{n-2}$ and~$\mathcal{N}_x = \mathcal{B}_n$), these two decompositions coincide. Indeed, for any $i,j\in[n]$ with $j \ge i+3$, denote by $b(i,j)$ the brick of~$\mathcal{B}_n$ located immediately below the contact between the $i$\textsuperscript{th} and the $j$\textsuperscript{th} pseudoline of the unique pseudoline arrangement supported by~$\mathcal{B}_n$. Then the Minkowski summand~$\Omega(\mathcal{B}_n,b(i,j))$ is the face $\triangle_{\{i,\dots,j-2\}}$ of the standard simplex (up to a translation of vector~$1\!\!1_{\{1,\dots,i-1\}\cup\{j-1,\dots,n-2\}}$). This implies that
$$\Omega(\mathcal{B}_n^1) = \sum_{1\le i<j\le n-2} \triangle_{\{i,\dots,j\}}$$
up to translation, and by unicity, that the coefficient $y_I$ is $1$ if $I$ is an interval of $[n-2]$ and $0$ otherwise (singletons are irrelevant since they only involve translations).
For general~$x$, the Minkowski summands~$\Omega(\mathcal{N}_x^1,b)$ are not always simplices. In~\cite{Lange}, Lange computes the coefficients $\{y_I\}$ in the Minkowski decomposition of any associahedra~$\Omega(\mathcal{N}_x^1)$ into dilates of faces of the standard simplex, and therefore answers Question~\ref{qu:coeffsMinkowskiSumDiff} for those special networks.
\end{remark}
\begin{remark}\label{rem:cyclohedra}
To close this section, we want to mention that we can similarly present Hohlweg \& Lange's realizations of the cyclohedra~\cite{HohlwegLange}. Namely, consider an antisymmetric word $x\in\{a,b\}^{2n-2}$ (\ie which satisfies $\{x_i,x_{2n-1-i}\}=\{a,b\}$ for all $i$), such that the $(2n)$-gon~$\mathcal{P}_x$ is centrally symmetric. Then the convex hull of the brick vectors of the dual pseudoline arrangements of all centrally symmetric triangulations of~$\mathcal{P}_x$ is a realization of the $(n-1)$-dimensional cyclohedron. For example, the centrally symmetric triangulations of~$\mathcal{P}_{a^2b^2}$ are colored in the right associahedron of \fref{fig:associahedra2B}: the convex hull of the corresponding vertices is a realization of the $2$-dimensional cyclohedron.
\end{remark}
\section{Brick polytopes and multitriangulations}\label{sec:multi}
This last section is devoted to the initial motivation of this work. We start recalling some background on multitriangulations, in particular the duality between $k$-triangulations of a convex polygon and pseudoline arrangements supported by the $k$-kernel of a reduced alternating sorting network. We then describe normal vectors for the facets of the brick polytope of the $k$-kernel of the ``bubble sort network''. Finally, we discuss the relationship between our brick polytope construction and the realization of a hypothetical ``multiassociahedron''.
\subsection{Background on multitriangulations}\label{subsec:backgroundMulti}
We refer the reader to~\cite{PilaudSantos} and \mbox{\cite[Chapters 1--4]{Pilaud}} for detailled surveys on multitriangulations and only recall here the properties needed for this paper. We start with the definition:
\begin{definition}
A \defn{$k$-triangulation} of a convex polygon is a maximal set of its diagonals such that no $k+1$ of them are mutually crossing.
\end{definition}
Multitriangulations were introduced by Capoyleas \& Pach \cite{CapoyleasPach} in the context of extremal theory for geometric graphs: a $k$-triangulation of a convex polygon~$\mathcal{P}$ induces a maximal $(k+1)$-clique-free subgraph of the intersection graph of the diagonals of~$\mathcal{P}$. Fundamental combinatorial properties of triangulations (which arise when~$k=1$) extend to multitriangulations~\cite{Nakamigawa,dkm,Jonsson,PilaudSantos}. We restrict the following list to the properties we need later on:
\paragraph{Diagonals~\cite{Nakamigawa, dkm, PilaudSantos}}
Any $k$-triangulation of a convex $n$-gon~$\mathcal{P}$ has precisely ${k(2n-2k-1)}$ diagonals. A diagonal of~$\mathcal{P}$ is said to be \defn{\mbox{$k$-relevant}} (resp. \defn{$k$-boundary}, resp.~\defn{$k$-irrelevant}) if it has at least $k$ vertices of~$\mathcal{P}$ on each side (resp.~precisely $k-1$ vertices of~$\mathcal{P}$ on one side, resp.~less than $k-1$ vertices of~$\mathcal{P}$ on one side). All $k$-boundary and $k$-irrelevant diagonals are contained in all $k$-triangulations of~$\mathcal{P}$.
\begin{figure}[h]
\capstart
\centerline{\includegraphics[scale=1]{2triang}}
\caption{A $2$-triangulation of the octagon (left) and the \mbox{$2$-rele}\-vant, $2$-boundary and $2$-irrelevant edges of the octogon (right).}
\label{fig:2triang}
\end{figure}
\paragraph{Stars~\cite{PilaudSantos}}
A \defn{$k$-star} of~$\mathcal{P}$ is a star-polygon with $2k+1$ vertices $s_0,\dots,s_{2k}$ in convex position joined by the $2k+1$ edges $[s_i,s_{i+k}]$ (the indices have to be understood modulo $2k+1$). Stars in multitriangulations play the same role as triangles in triangulations. In particular, a $k$-triangulation of~$T$ is made up gluing $n-2k$ distinct $k$-stars: a $k$-relevant diagonal of~$T$ is contained in two $k$-stars of~$T$ (one on each side), while a $k$-boundary diagonal is contained in one $k$-star of~$T$~---~see \fref{fig:stars} for an illustration. Furthermore, any pair of $k$-stars of~$T$ has a unique \defn{common bisector}. This common bisector is not a diagonal of~$T$ and any diagonal of~$\mathcal{P}$ not in~$T$ is the common bisector of a unique pair of $k$-stars.
\begin{figure}[h]
\capstart
\centerline{\includegraphics[scale=1]{stars}}
\caption{The four $2$-stars in the $2$-triangulation of \fref{fig:2triang}.}
\label{fig:stars}
\end{figure}
\paragraph{Flip~\cite{Nakamigawa,Jonsson,PilaudSantos}}
Let $T$ be a $k$-triangulation of~$\mathcal{P}$, let $e$ be a \mbox{$k$-rele}\-vant edge of~$T$, and let $f$ denote the common bisector of the two $k$-stars of~$T$ containing~$e$. Then $T\triangle\{e,f\}$ is again a $k$-triangulation of~$\mathcal{P}$, and $T$ and $T\triangle\{e,f\}$ are the only $k$-triangulations of $\mathcal{P}$ containing $T\smallsetminus\{e\}$. We say that we obtain~$T\triangle\{e,f\}$ by the \defn{flip} of~$e$ in~$T$. The graph of flips is $k(n-2k-1)$-regular and connected.
\begin{figure}[h]
\capstart
\centerline{\includegraphics[scale=1]{multiFlip}}
\caption{A flip in the $2$-triangulation of \fref{fig:2triang}.}
\label{fig:multiFlip}
\end{figure}
\paragraph{Duality~\cite{PilaudPocchiola}}
Fix a word~$x\in\{a,b\}^{n-2}$ and consider the $n$-gon~$\mathcal{P}_x$ and the sorting network~$\mathcal{N}_x$ defined in Section~\ref{subsec:duality}. We denote by~$\mathcal{N}_x^k$ the \defn{$k$-kernel} of $\mathcal{N}_x$, \ie the network obtained from~$\mathcal{N}_x$ by erasing its first~$k$ and last~$k$ levels together with the commutators incident to them. The remaining commutators of~$\mathcal{N}_x^k$ are precisely labeled by the $k$-relevant diagonals of~$\mathcal{P}_x$, which provides a duality between $k$-triangulations of~$\mathcal{P}_x$ and pseudoline arrangements supported by~$\mathcal{N}_x^k$. Namely, the set of commutators of~$\mathcal{N}_x$ labeled by the internal diagonals of a $k$-triangulation~$T$ of $\mathcal{P}_x$ is the set of contacts of a pseudoline arrangement~$T^*$ supported by $\mathcal{N}_x^k$. Reciprocally the $k$-relevant diagonals of $\mathcal{P}_x$ which label the contacts of a pseudoline arrangement supported by~$\mathcal{N}_x^k$, together with all $k$-irrelevant and $k$-boundary diagonals of~$\mathcal{P}_x$, form a $k$-triangulation of $\mathcal{P}_x$.
\begin{figure}[h]
\capstart
\centerline{\includegraphics[width=\textwidth]{dualite2}}
\caption{A $2$-triangulation $T$ of $\mathcal{P}_{b^6}$ (left), a symmetric representation of it (center), and its dual pseudoline arrangement $T^*$ on the $2$-kernel of the sorting network~$\mathcal{N}_{b^6}$ (right).}
\label{fig:dualite2}
\end{figure}
The dual pseudoline arrangement $T^*$ of a $k$-triangulation $T$ of $\mathcal{P}_x$ has one pseudoline $S^*$ dual to each $k$-star $S$ of $T$. A commutator is the crossing (resp.~a contact) between two pseudolines $S^*$ and $R^*$ of $T^*$ if it is labeled by the common bisector (resp.~by a common edge) of the $k$-stars $S$ and~$R$. Consequently, this duality defines an isomorphism between the graph of flips on pseudoline arrangements supported by~$\mathcal{N}_x^k$ and the graph of flips on $k$-triangulations of~$\mathcal{P}_x$.
\subsection{The brick polytope of multitriangulations}
We consider the brick polytope~$\Omega(\mathcal{N}_x^k)$ of the $k$-kernel of a reduced alternating sorting network~$\mathcal{N}_x$ (for some word~$x\in\{a,b\}^{n-2}$). Since we erase the first and last $k$ levels, this polytope again does not depend on the first $k$ and last $k$ letters of~$x$. The vertices of this polytope correspond to $k$-triangulations of~$\mathcal{P}_x$ whose contact graph is acyclic. In contrast to the case of triangulations ($k=1$) discussed in Section~\ref{sec:associahedra},
\begin{enumerate}[(i)]
\item not all $k$-triangulations of~$\mathcal{P}_x$ appear as vertices of~$\Omega(\mathcal{N}_x^k)$ and the graph of~$\Omega(\mathcal{N}_x^k)$ is a proper subgraph of the graph of flips on $k$-triangulations of~$\mathcal{P}_x$;
\item the combinatorial structure of~$\Omega(\mathcal{N}_x^k)$ depends upon~$x$.
\end{enumerate}
In the rest of this section, we restrict our attention to the ``bubble sort network''. We denote by~$\mathcal{B}_n \eqdef \mathcal{N}_{b^{n-2}}$ and $\mathcal{P}_n \eqdef \mathcal{P}_{b^{n-2}}$. For this particular network, we can describe further the combinatorics of the brick polytope~$\Omega(\mathcal{B}_n^k)$.
\begin{example}
The $f$-vectors of the brick polytopes~$\Omega(\mathcal{B}_7^2)$, $\Omega(\mathcal{B}_8^2)$, $\Omega(\mathcal{B}_9^2)$ and $\Omega(\mathcal{B}_{10}^2)$ are $(6,6)$, $(22,33,13)$, $(92,185,118,25)$ and $(420,1062,945,346,45)$ respectively. We have represented~$\Omega(\mathcal{B}_8^2)$ and~$\Omega(\mathcal{B}_9^2)$ in Figures~\ref{fig:B82} and~\ref{fig:B92}. The polytopes~$\Omega(\mathcal{B}_7^2)$ and~$\Omega(\mathcal{B}_8^2)$ are simple while the polytope~$\Omega(\mathcal{B}_9^2)$ has two non-simple vertices (which are contained in the projection facet of the Schlegel diagram on the right of \fref{fig:B92}) and the polytope~$\Omega(\mathcal{B}_{10}^2)$ has~$24$ non-simple vertices.
\begin{figure}[p]
\capstart
\centerline{\includegraphics[width=1.1\textwidth]{B82}}
\caption{The $3$-dimensional polytope $\Omega(\mathcal{B}_8^2)$. Only $22$ of the $84$ $2$-triangulations of the octagon appear as vertices.}
\label{fig:B82}
\end{figure}
\begin{figure}[p]
\capstart
\centerline{\includegraphics[width=\textwidth]{B92}}
\caption{Two Schlegel diagrams of the $4$-dimensional polytope $\Omega(\mathcal{B}_9^2)$. On the second one, the two leftmost vertices of the projection facet are non-simple vertices.}
\label{fig:B92}
\end{figure}
\end{example}
As mentioned in Example~\ref{exm:loday}, the brick polytope~$\Omega(\mathcal{B}_n^1)$ coincides (up to translation) with Loday's realization of the $(n-3)$-dimensional associahedron~\cite{Loday}~---~see \fref{fig:associahedra2B} (left). The facet normal vectors of this polytope are all vectors of $\{0,1\}^{n-2}$ which are neither $0^{n-2}$ nor $1^{n-2}$ and whose $1$'s are consecutive. More precisely, the vector~$\sum_{\ell=i}^{j-2} e_\ell$ is a normal vector of the facet corresponding to the diagonal $[i,j]$, for any $i,j\in[n]$ with $j \ge i+2$.
For general~$k$, we can similarly provide a more explicit description of the facets of the brick polytope $\Omega(\mathcal{B}_n^k)$. Representatives for their normal vectors are given by the following sequences:
\begin{definition}
A sequence of~$\{0,1\}^q$ is \defn{$p$-valid} if it is neither $0^q$ nor $1^q$ and if it does not contain a subsequence $10^r1$ for $r\ge p$. In other words, all subsequences of $p$ consecutive zeros appear before the first $1$ or after the last $1$.
\end{definition}
\begin{remark}
Let $\textsc{vs}_{p,q}$ denote the number of $p$-valid sequences of~$\{0,1\}^q$. It is easy to see that $\textsc{vs}_{1,q}=\frac{1}{2}(q-1)(q+2)$ (the number of internal diagonals of a $(q+2)$-gon!) and that $\textsc{vs}_{2,q}=F_{q+4}-(q+4)$, where $F_n$ denotes the $n$\textsuperscript{th} Fibonacci number. To compute $\textsc{vs}_{p,q}$ in general, consider the non-ambiguous rational expression $0^*(1,10,100,\dots,10^{p-1})^*10^*$. The corresponding rational language consists of all $p$-valid sequences plus all non-empty sequences of~$1$. Thus, the generating function of $p$-valid sequences is:
$$\sum_{q\in\mathbb{N}} \textsc{vs}_{p,q}x^q=\frac{1}{1-x}\,\frac{1}{1-\sum_{i=1}^{p} x^i}\,x\,\frac{1}{1-x}-\frac{x}{1-x}=\frac{x^2(2-x^p)}{(1-2x+x^{p+1})(1-x)}.$$
\end{remark}
Let $\sigma$ be a sequence of $\{0,1\}^{n-2k}$. Let $|\sigma|_0$ denote the number of zeros in $\sigma$. For all $i\le |\sigma|_0+2k$, we denote by $\zeta_i(\sigma)$ the position of the $i$\textsuperscript{th} zero in the sequence $0^k\sigma0^k$, obtained from $\sigma$ by appending a prefix and a suffix of $k$ consecutive zeros. We associate to $\sigma$ the set of edges of~$\mathcal{P}_n$:
$$D(\sigma) \eqdef \set{[\zeta_i(\sigma),\zeta_{i+k}(\sigma)]}{i\in[|\sigma|_0+k]}.$$
Observe that $D(\sigma)$ can contain some $k$-boundary edges of~$\mathcal{P}_n$.
\begin{proposition}\label{prop:multi}
Each $k$-valid sequence $\sigma\in\{0,1\}^{n-2k}$ is a normal vector of a facet $F_\sigma$ of the brick polytope $\Omega(\mathcal{B}_n^k)$. The facet~$F_\sigma$ contains precisely the brick vectors of the dual pseudoline arrangements of the $k$-triangulations of~$\mathcal{P}_n$ containing $D(\sigma)$. Furthermore, every facet of $\Omega(\mathcal{B}_n^k)$ is of the form $F_\sigma$ for some $k$-valid sequence $\sigma\in\{0,1\}^{n-2k}$.
\end{proposition}
\begin{proof}[Proof (sketch)]
Consider a sequence $\sigma\in\{0,1\}^{n-2k}$ and let $(U,V)$ be a partition of $[n-2k]$ such that $\sigma=1\!\!1_U=1\!\!1-1\!\!1_V$. Since $D(\sigma)$ contains no $(k+1)$-crossing, there exists a $k$-triangulation~$T$ of~$\mathcal{P}_n$ containing $D(\sigma)$. We claim that the $k$-relevant edges of $D(\sigma)$ separate $U$ from $V$ in the contact graph~$(T^*)^\#$, and that this cut is minimal if and only if $\sigma$ is $k$-valid. This claim together with Corollary~\ref{coro:fd} prove our proposition. We refer the reader to \cite{Pilaud} for details.
\end{proof}
\subsection{Towards a construction of the multiassociahedron?}\label{subsec:multiassociahedron}
Let $\Delta_n^k$ denote the simplicial complex of $(k+1)$-crossing-free sets of $k$-relevant diagonals of a convex $n$-gon. Its maximal elements are $k$-triangulations of the $n$-gon and thus it is pure of dimension $k(n-2k-1)-1$. In an unpublished manuscript~\cite{Jonsson}, Jonsson proved that $\Delta_n^k$ is in fact a shellable sphere (see also the recent preprint~\cite{SerranoStump}). However, the question is still open to know whether the sphere~$\Delta_n^k$ can be realized as the boundary complex of a simplicial $k(n-2k-1)$-dimensional polytope. As a conclusion, we discuss some interactions between this question and our paper.
\paragraph{Universality}
According to Pilaud \& Pocchiola's duality presented in Section~\ref{subsec:backgroundMulti}, this question seems to be only a particular subcase of Question~\ref{qu:polytope}. However, we show in the next proposition that the simplicial complexes $\Delta_n^k$ $(n,k\in\mathbb{N})$ contain all simplicial complexes $\Delta(\mathcal{N})$ ($\mathcal{N}$ sorting network). If $X$ is a subset of the ground set of a simplicial complex~$\Delta$, remember that the \defn{link} of $X$ in $\Delta$ is the subcomplex~$\set{Y \smallsetminus X}{X \subset Y \in \Delta}$, while the \defn{star} of~$X$ is the join of $X$ with its~link.
\begin{proposition}\label{prop:universality}
For any sorting network~$\mathcal{N}$ with $n$ levels and $m$ commutators, the simplicial complex~$\Delta(\mathcal{N})$ is (isomorphic to) a link of~$\Delta_{n+2m-2}^{m-1}$.
\end{proposition}
\begin{proof}
Label the commutators of~$\mathcal{N}$ from left to right (and from top to bottom if some commutators lie on the same vertical line). For any~$i\in[m]$, let $i_\square$ be such that the $i$\textsuperscript{th} commutator of~$\mathcal{N}$ join the $i_\square$\textsuperscript{th} and~$(i_\square+1)$\textsuperscript{th} levels of~$\mathcal{N}$. Define the set~$W \eqdef \set{\{i,i+i_\square+m-1\}}{i\in[m]}$.
We claim that the simplicial complex~$\Delta(\mathcal{N})$ is isomorphic to the simplicial complex of all $m$-crossing-free sets of $(m-1)$-relevant diagonals of the \mbox{$(n+2m-2)$-gon} which contain the diagonals not labeled by~$W$. Indeed, by duality, the \mbox{$(m-1)$-trian}\-gulations of this complex correspond to the pseudoline arrangements supported by~$\mathcal{B}_{n+2m-2}^{m-1}$ whose sets of contacts contain the contacts of $\mathcal{B}_{n+2m-2}^{m-1}$ not labeled by~$W$. Thus, our claim follows from the observation that, by construction, the contacts labeled by~$W$ are positioned exactly as those of~$\mathcal{N}$~---~see \fref{fig:universality}.
\end{proof}
\begin{figure}[h]
\capstart
\centerline{\includegraphics[scale=1.2]{universality}}
\caption{Universality of the multitriangulations: any sorting network~$\mathcal{N}$ is a subnetwork of a bubble sort network~$\mathcal{B}_n^k$ for some parameters~$n$ and~$k$ depending on~$\mathcal{N}$.}
\label{fig:universality}
\end{figure}
Since links in shellable spheres are shellable spheres, this proposition extends Jonsson's result~\cite{Jonsson} to any sorting network:
\begin{corollary}\label{coro:shellable}
For any sorting network~$\mathcal{N}$, the simplicial complex~$\Delta(\mathcal{N})$ is a shellable sphere.
\end{corollary}
This corollary is also a consequence of Knutson \& Miller's results on the shellability of any subword complex~\cite{KnutsonMiller}.
Proposition~\ref{prop:universality} would also extend any proof of the polytopality of $\Delta_n^k$ to that of $\Delta(\mathcal{N})$:
\begin{corollary}
If the simplicial complex $\Delta_n^k$ is polytopal for any $n,k\in\mathbb{N}$, then the simplicial complex $\Delta(\mathcal{N})$ is polytopal for any sorting network~$\mathcal{N}$.
\end{corollary}
In particular, if one manages to construct a multiassociahedron realizing the simplicial complex~$\Delta_n^k$, it would automatically provide an alternative construction to the polytope of pseudotriangulations~\cite{RoteSantosStreinu-polytope}. We want to underline again that the brick polytope does not realizes the simplicial complex of pointed crossing-free sets of internal edges in a planar point set.
\paragraph{Projections and brick polytopes}
In this paper, we have associated to each \mbox{$k$-trian}gulation~$T$ of the $n$-gon the brick vector~$\omega(T^*) \in \mathbb{R}^{n-2k}$ of its dual pseudoline arrangement~$T^*$ on~$\mathcal{B}_n^k$. We have seen that the convex hull~$\Omega(\mathcal{B}_n^k)$ of the set $\set{\omega(T^*)}{T \; k\text{-triangulation of the } n\text{-gon}}$ satisfies the following two properties:
\begin{enumerate}[(i)]
\item The graph of~$\Omega(\mathcal{B}_n^k)$ is (isomorphic to) a subgraph of the graph of flips on $k$-triangulations of an $n$-gon (Corollary~\ref{coro:graph}).
\item The set of $k$-triangulations whose brick vector belongs to a given face of~$\Omega(\mathcal{B}_n^k)$ forms a star of~$\Delta_n^k$ (Proposition~\ref{prop:multi}).
\end{enumerate}
One could thus reasonably believe that our point set could be a projection of the polar of a hypothetical realization of~$\Delta_n^k$. The following observation kills this hope:
\begin{proposition}
Let $P$ be the polar of any realization of~$\Delta_7^2$. It is impossible to project~$P$ down to the plane such that the vertex of~$P$ labeled by each \mbox{$2$-triangulation}~$T$ of the heptagon is sent to the corresponding brick vector~$\omega(T^*)$.
\end{proposition}
\begin{proof}
The network~$\mathcal{B}_7^2$ is a $3$-level alternating network and thus has been addressed in Example~\ref{exm:3levels}. See \fref{fig:B72} (left) to visualize the brick polytope~$\Omega(\mathcal{B}_7^2)$. The contacts are labeled from left to right, and the brick vector of a pseudoline arrangement covering $\mathcal{B}_7^2$ is labeled by its four contacts.
\begin{figure}[h]
\capstart
\centerline{\includegraphics[scale=.9]{B72}}
\caption{The $2$-dimensional brick polytope $\Omega(\mathcal{B}_7^2)$.}
\label{fig:B72}
\end{figure}
Assume that there exists a polytope~$P$ whose polar realizes~$\Delta_7^2$ and which projects to the point configuration $\set{\omega(T^*)}{T \; 2\text{-triangulation of the heptagon}}$. Observe that the projections of two non-parallel edges of a $2$-dimensional face of~$P$ either are not parallel or lie on a common line (when the $2$-dimensional face is projected to a segment of this line). We will reach a contradiction by considering the facet of~$P$ labeled by $4$. We have represented in \fref{fig:B72bis} the projections of its \mbox{$2$-dimen}\-sional faces: there are two triangles $24$ and $46$ which project to a segment, two quadrilaterals $14$ and $47$ and two pentagons $34$ and $45$. Since they belong to the $2$-dimensional face~$45$ of~$P$, and since they project on two distinct parallel lines, the two edges $(3456,4567)$ and $(1245,1457)$ of~$P$ are parallel. Similarly, the edges $(1245,1457)$ and $(1234,1347)$ are parallel and the edges~$(1234,1347)$ and $(3456,3467)$ are parallel. By transitivity, we obtain that the edges $(3456,4567)$ and $(3456,3467)$ of~$P$ are parallel which is impossible since they belong to the triangular face~$46$.
\begin{figure}
\capstart
\centerline{\includegraphics[scale=.9]{B72bis}}
\caption{Projections of some $2$-dimensional faces of~$\Delta_7^2$ on $\Omega(\mathcal{B}_7^2)$.}
\label{fig:B72bis}
\end{figure}
\end{proof}
\section*{Acknowledgment}
We thank Christian Stump for helpful comments on a preliminary version of this paper. We are also grateful to Christophe Hohlweg and Carsten Lange for interesting discussion about this paper and related topics. Finally, we thank two anonymous referees for relevant suggestions.
\bibliographystyle{alpha}
|
\section{Introduction}
Several authors have been investigating real classes, characters
and blocks of finite groups, see e.g.
\cite{Brau,Bra,DolNavTie,Gow,G-M,HH,IMN,IN,Iwasaki,K-N,MN,M,Mur,NavSanTie, NavSanTie2}.
The aim of this note is to formulate real versions of Brauer's
${\rm k}(B)$ conjecture, see \cite{Bra56}, Olsson's conjecture, see \cite{Ol84},
and Eaton's conjecture, see \cite{Eat03}, for
$2$-blocks.
We
give special cases when we can prove the real versions of them.
The last part of the paper deals with fusion of elements
of defect groups.
\section{Notations and terminology}
In this note $G$ will always denote a finite group, $p$ a prime integer, which is $2$
except for the last section of the paper.
Let $(R,k,F)$ be a $p$-modular system, where $R$ is a complete discrete valuation
ring with quotient field $k$ of characteristic zero and residue class field $F$
of characteristic $p$. We assume that $k$ and $F$ are splitting fields of all the
subgroups of $G$. We may also assume that $k$ is a subfield of the
complex numbers.
Complex conjugation acts on ${\rm Irr}(G)$. A character is {\bf real} if it is conjugate
to itself, in other words if it is real valued.
An element of $G$ is {\bf real} if it is conjugate to its inverse.
An element $x$ of a subgroup $H$ of $G$ we call {\bf $H$-real}, if it can be conjugated to its inverse
by an element from the subgroup $H$.
We say that the {\bf conjugacy class $C$ is real } if it is equal to the class of
the inverse elements of the class. We use the notation ${\rm Cl}_{r}(G)$ for the set of these
classes.
A {\bf $p$-block $B$ }is called {\bf real } if it contains the complex conjugate of
an irreducible ordinary character (and hence of all irreducible characters)
in the block. It is known, see e.g. \cite[Thm.~3.33]{Nav}, that a real
$2$-block always contains real valued irreducible ordinary and Brauer characters, as well.
We use the notation ${\rm Irr}_{rv}(G)$ and ${\rm Irr}_{rv}(B)$ for the set of real valued irreducible ordinary
characters in $G$ and in $B$, respectively. Let ${\rm k}_{rv}(G)$ and ${\rm k}_{rv}(B)$
stand for the sizes of these sets.
We use the notation ${{\rm k}}_{{i},{rv}}(B)$ for the number of
real valued irreducible characters of height $i$ in the $p$-block $B$.
By Brauer's permutation lemma the number of real conjugacy classes of the group
$G$ is equal to $k_{rv}(G)$. We use the notation ${\mbox{\rm Bl}}(G|D)$ for the set of $p$-blocks
of
$G$ with defect group $D$, $D^{(n)}$ stands for the $n$-th derived subgroup
of $D$. For constructing examples we used the GAP system, see \cite{GAP}, and
we also describe these groups with their GAP notation.
\section {The real conjectures}
Unless otherwise stated, let $p=2$. Let $G$ be a finite group, $B$ a real
$2$-block of $G$ with defect group $D$.
\begin{conjecture}[Weak real version of Brauer's conjecture]\label{conj0}
We conjecture that
${\rm k}_{rv}(B)$
is bounded from above by the
number of $G$-real elements of $D$.
\end{conjecture}
\begin{conjecture}[Strong real version of Brauer's conjecture]\label{conj1}
We conjecture that ${\rm k}_{rv}(B)$
is bounded from above by the
number of $N_G(D)$-real elements of $D$.
\end{conjecture}
\begin{conjecture}[Real version of Olsson's conjecture]\label{conj2}
We conjecture that ${{\rm k}_{{0},{rv}}}(B)$
is bounded from above by
the number of $N_G(D)/D'$-real elements of $D/D'$.
\end{conjecture}
\begin{conjecture} [Real version of Eaton's conjecture]\label{conj3}
We conjecture that
$\sum_{i=0}^n {{\rm k}}_{{i},{rv}}(B)$
is bounded from above by the number of
$N_G(D)/D^{(n+1)}$-real elements
of $D/D^{(n+1)}$.
\end{conjecture}
\begin{remark} One could not replace in Conjecture \ref{conj1} the $N_G(D)$ by $D$.
The smallest example is a group of order $24$ which is the pullback of maps
$S_3\rightarrow C_2$ and $Q_8\rightarrow C_2$, (with GAP notations it is
$\texttt{SmallGroup}(24,4)$). In this group there are two $2$-blocks. The nonprincipal
block $B$ has a normal defect group $D\simeq C_4$, where there are just two
$D$-real elements, however ${\rm k}_{rv}(B)=4$. (In fact in this group all characters in ${\rm Irr}(G)$ are real).
However, we do not know any such example for the principal block,
or for blocks of maximal defect.
\end{remark}
\begin{remark}
If every irreducible character is real in the group $G$ then we get
stronger versions of the non-real conjectures, see Remark \ref{stronger}, namely
${\rm k}(B)$ (${{\rm k}_0}(B)$, $\sum_{i=0}^{n}{{\rm k}_i}(B))$ are bounded from above by the number of
elements of the defect group $D$ of $B$,($D/D'$ or $D/D^{(n+1)}$) that are
real inside $N_G(D)$ ($N_G(D)/D'$ or $N_G(D)/D^{(n+1)}$) respectively.
Of course Conjecture~\ref{conj3} implies Conjectures \ref{conj0}, \ref{conj1}
and \ref{conj2}.
\end{remark}
\begin{remark}\label{stronger}
If every irreducible character of a group $G$ is real, it does not follow that
the normalizer of its Sylow $2$-subgroup also has this property.
Let $G=\texttt{SmallGroup}(96,185)$. This group has selfnormalizing Sylow $2$-subgroups. In the principal block of $G$ all the $14$
irreducible characters are real valued, its Sylow $2$-subgroup has also $14$ irreducible characters, but only $12$ of them are real.
This example also shows that an element can
be real in one of the Sylow $2$-subgroups,
but not real in an other Sylow $2$-subgroup,
since one can find order $4$ elements in this with that property.
In this group all $32$ elements of the Sylow $2$-subgroup are real in $G$, but
only $28$ of them are real in $N_G(S)=S$.
\end{remark}
In the next remark we show that the $p$-analogue of Conjecture \ref{conj0}
is not true for $p>2$, and since the defect group is abelian the other Conjectures \ref{conj1}, \ref{conj2} and \ref{conj3} also cannot hold:
\begin{remark}\label{forp}
Let $G=M_{11}$, $p=11$ and let $B$ be the principal block.
Then $|D|=11$, ${\rm k}_{rv}(B)=3$, but in $D$ there is only one
$G$-real element. This group is also an example for the fact that the number of real
valued irreducible characters can be different in the Brauer correspondent blocks
with cyclic defect group if $p>2$.
Let $b\in Bl(N_G(D)|D)$ be the Brauer correspondent of $B$. Then ${\rm k}_{rv}(b)=1$.
If $p=2$ and the defect group is noncyclic then Brauer correspondent blocks might have
differrent number of real valued irreducible characters: let us take the same
group $G$, then the principal $2$-block has $6$, however
its Brauer correspondent block has $5$ real valued
irreducible characters.
\end{remark}
\section{Nilpotent groups, symmetric groups and blocks with central defect groups}
\begin{proposition}[The nilpotent groups]\label{nilp} A stronger form of Conjecture~\ref{conj3} (hence \ref{conj2}, \ref{conj1} and \ref{conj0})
holds for nilpotent groups. If $G$ is either a $2$-group or abelian, then in Conjecture
\ref{conj2} there is equality.
\end{proposition}
\noindent
Proof.
If $G$ is nilpotent then every $2$-block is of maximal defect, and by \cite{Gow}
the only real $2$-block of maximal defect is the principal block $B_0$. Then ${\rm Irr}(B_0)={\rm Irr}(G_2)$,
where $G_2\in Syl_2(G)$.
Characters of height $n$ of $G_2$ are those of degree $2^n$. This is at most the $n$th
character degree of $G_2$. By \cite[Lemma~5.12]{Is}, all irreducible characters
of height at most $n$ contain
${G_2}^{(n+1)}$ in their kernels, hence
$\sum_{i=0}^n {{\rm k}}_{{i},{rv}}(B_0)\leq
|{\rm Irr}_{rv}(G_2/{{G_2}^{(n+1)}})|$, which is at most the number of $G_2/G_2^{(n+1)}$-real elements
in $G_2/{{G_2}^{(n+1)}}$.
\begin{proposition}[The symmetric groups]\label{sym}
Conjectures \ref{conj1} and \ref{conj2} hold for the symmetric groups.
\end{proposition}
\noindent
Proof.
\begin{itemize}
\item[(a)]Since every irreducible character of the symmetric group is real valued and
since its Sylow $2$-subgroup also has this property by \cite[Thm.~4.4.8]{J-K},
Conjecture~\ref{conj1} for the principal $2$-blocks
reduces in this case to the non-real ${\rm k}(B)$ conjecture, which
holds by \cite{Ol84}.
In \cite{Ol76} it is proved that the defect group $D$ of each block $B$ of weight
$w$ of $S_n$ is isomorphic to the Sylow $p$-subgroup of $S_{pw}$ and there is a canonical
height preserving bijection between the irreducible characters of $B$ and that
of the principal block of $S_{pw}$. Thus if $p=2$ then in $D$ each element is also real,
and again by \cite{Ol84}, Conjecture~\ref{conj1} holds also for nonprincipal
$2$-blocks of $S_n$.
\item[(b)]
Olsson's conjecture also holds for $S_n$ by \cite{Ol84}. Thus by similar arguments as above, Conjecture~\ref{conj2}, also holds.
\end{itemize}
\begin{remark}
A positive answer to Eaton's conjecture for $S_n$, would imply a positive answer to
Conjecture \ref{conj3}.
\end{remark}
\begin{proposition}[Blocks with central defect groups]\label{central}
Conjecture \ref{conj3} (and hence, Conjectures \ref{conj0},\ref{conj1} and \ref{conj2}) holds for
central defect groups.
In fact we prove a slightly stronger statement:
the strong forms of the conjectures holds for $2$-blocks with defect group $D$, where $G=DC_G(D)$:
\end{proposition}
\noindent
Proof.
Let $B\in {\mbox{\rm Bl}}(G|D)$ be a $2$-block of $G$, where $G=DC_G(D)$.
By \cite[Thm.~9.12]{Nav} $|{\mbox{\rm IBr}}(B)|=1$ and there is a bijection between
${\rm Irr}(D)$ and ${\rm Irr}(B)$ mapping $\zeta $ to $\theta_{\zeta}$, where
$\theta_{\zeta}(g)=\zeta(g_2)\theta(g_{2'})$, if $g_2\in D$, otherwise it is zero.
Here $\theta $ is the unique
character in ${\rm Irr}(B)$ containing $D$ in its kernel, and ${\mbox{\rm IBr}}(B)=\{\theta^{0}\}$.
Moreover $ht(\theta_{\zeta})=n$ iff $\zeta(1)=2^n$.
If $B$ is a real $2$-block then $\theta $ is a real valued character and $\theta_{\zeta}$
is real valued if and only if $\zeta $ is real valued.
Thus ${\rm k}_{rv}(B)={\rm k}_{rv}(D)$ and $\sum_{i=0}^n k_{i,rv}(B)=\sum_{i=0}^n
k_{i,rv}(D)\leq |{\rm Irr}_{rv}(D/D^{(n+1)}|$ by Proposition \ref{nilp}, this is at most
the number of $D/D^{(n+1)}$-real elements of $D/D^{(n+1)}$.
\begin{remark}
It is easy to see that if the above conjectures are true for the direct factors of a group then they are also true for the
direct product: a tensor product of characters is real iff each component is
real, if we have defect classes $C_1\in {\rm Cl}(B_1,D_1)$ and $C_2\in {\rm Cl}(B_2,D_2)$,
then the pair $(c_1,c_2)\in C_1\times C_2$ is a defect class of $B_1\otimes B_2$.
The defect of the character $\chi_1\otimes \chi_2$ is the sum of the defects of $\chi_1$ and $\chi_2$. The height of the product of characters is the sum of the
heights. The number of real elements in $D_1\times D_2$ is just the product
of the numbers of real elements in the direct components.
\end{remark}
\section{Blocks with cyclic defect groups}
For a block $B\in{\mbox{\rm Bl}}(G)$ we consider the pairs $(x,\theta)$ with $x\in G$
a $p$-element $\theta\in{\mbox{\rm IBr}}(b)$, where $b\in {\mbox{\rm Bl}}(C_G(x))$ such that
$b^G=B$.
As in \cite{Mur}, we call the $G$-conjugacy classes of these pairs,
denoted by $(x,\theta)^G$, the {\bf columns of $B$}.
A {\bf column} $(x,\theta)^G$ is called ${\bf real}$ if $(x,\theta)^G=(x^{-1},\overline{\theta})^G$.
In \cite[Lemma 1.1]{Mur} it is proved that ${\rm k}_{rv}(B)$ is equal to the number of real
columns of $B$.
We will use Dade's description \cite[Thm 68.1]{Dor} of $p$-blocks with
cyclic defect
groups only for the special case $p=2$:
Let $B$ be a $2$-block with cyclic defect group $D=\langle x \rangle$ of order
$2^a$, $D_i=\langle x^{2^i}\rangle$, $C_i=C_G(D_i),N_i=N_G(D_i)$, for
$i=0,\cdots,a-1$. Let $B_0\in {\mbox{\rm Bl}}(N_G(D)|D)$ be the Brauer correspondent block of $B$. Let $b_0\in {\mbox{\rm Bl}}(C_G(D)|D)$ with $b_0^{N_0}=B_0$.
Such blocks are conjugate in $N_0$.
Similarly let $b_i={b_0}^{C_i}$, then every block of $C_i$ that induces $B$
is conjugate to $b_i$ in $N_i$.
Let $\theta^i$ be the unique irreducible Brauer character of $b_i$ for $i=0,\cdots,a-1$.
The inertia subgroup of $b_i$ in $N_i$ is $C_i$
for $i=0,\cdots,a-1$, and also $|{\mbox{\rm IBr}}(B)|=1$. Let ${\mbox{\rm IBr}}(B)=\{\phi\}$.
\bigskip
First we prove:
\begin{lemma}\label{cyclicdefect}
With the notation above, we have that
$(x,\theta^0)$, $(x^k,\theta^0)$ for $k$ odd,
$(x^2,\theta ^1)$, $(x^{2k}, \theta^1)$ for $k$ odd,
...
$(x^{2^{a-1}},\theta^{a-1})$,
$(x^{2^{(a-1)}k},\theta^{a-1})$ for $k$ odd,
and $(1, \phi)$ are representatives of the columns of $B$.
If $x^{2^i}$ is the smallest power of $x$ which is $G$-real then representatives of the real columns of $B$ are among
those columns whose first component is a power of $x^{2^i}$.
\end{lemma}
\noindent
Proof.
If the first components of two pairs generate different subgroups, then they cannot be conjugate.
Let us take the pair $(y,\psi)$, where $y$ generates $D_j$ and $j< a$. Then the block of $\psi$
is conjugate to $b_j$ in $N_j$, so $\psi$ is conjugate to $\theta^j$ in $N_j$.
The conjugation takes $y$ to another generator of $D_j$, i.e. to $x^{{2^j}k}$, where
$k$ is odd.
If the first component is $1$, then the second component must be $\phi $.
\begin{corollary}
Let $G$ be a finite group, let $B$ be a real $2$-block with cyclic defect group $D$.
Then Conjecture \ref{conj0} holds for $G$.
\end{corollary}
\noindent
Proof.
We use \cite[Lemma 1.1]{Mur}, Lemma \ref{cyclicdefect} and the notations above.
Then the number of $G$-real elements in $D$ is exactly $2^{a-i}$.
We have
that the representatives of real columns of $B$
are $(1,\phi)$ and some of those columns whose first component is an element of
$D_i$ and if it generates $D_j$ then
the second component is $\theta^j$. Their number is at most the number of elements of $D_i$, which is
$2^{a-i}$.
\begin{corollary} \label{normal}
Let $D$ be a cyclic normal $2$-subgroup of $G$.
Then Conjecture \ref{conj1} and hence Conjecture \ref{conj3} also
holds for blocks $B\in Bl(G|D)$.
\end{corollary}
\begin{remark}
Using similar arguments for the $p>2$ case, one gets for block with cyclic defect groups that ${\rm k}_{rv}(B)\leq {\rm l}(B)\cdot
|\{$ $G$-real elements in $D$ $\}|$.
This could be considered as some kind
of real analogue of the so called ``Trace inequality'' in \cite[Prop. 2, p. 272]{Ol80}.
\end{remark}
\bigskip
To prove Conjecture \ref{conj1} for $2$-blocks with cyclic defect group
we will need the following lemma (the $p$-analogue of it
for $p>2$ is not true, and if $p=2$, but the defect group is noncyclic
then the analogous result is not true either, see Example \ref{forp}):
\begin{lemma}\label{samenumber}
Let $G$ be a finite group, let $B\in {\mbox{\rm Bl}}(G|D)$ be a real $2$-block with cyclic
defect group $D$ and let $B_0\in {\mbox{\rm Bl}}(N_G(D)|D)$ be its Brauer correspondent block.
Then
${\rm k}_{rv}(B)={\rm k}_{rv}(B_0)$.
\end{lemma}
\noindent
Proof. We use the same notation as in the introduction to this section.
By \cite[Lemma 1.1]{Mur} and Lemma \ref{cyclicdefect} it is enough
to prove that if $(x^{2^{j}k},\theta^j)$ represents a real column of the block $B$,
(recall that $\theta^j\in {\mbox{\rm IBr}}(b_j)$ and $b_j\in {\mbox{\rm Bl}}(C_G(D_j)|D)$),
and $\tilde{b_j}\in {\mbox{\rm Bl}}(N_{C_G(D_j)}(D)|D)={\mbox{\rm Bl}}(C_{N_G(D)}(D_j)|D)$ is the
Brauer correspondent of $b_j$ containing the single irreducible Brauer character
$\tilde\theta^j$, then
$(x^{{2^j}k},\tilde \theta^j)$ belongs to a real column of $B_0$ and this correspondence defines a bijection
of real columns of $B_0$ and $B$.
Let $z\in G$ such that $(({x^{{2^j}k}})^z,{\theta^j}^z)=(({x^{{2^j}k}})^{-1},
\overline{\theta^j})$.
Then $\overline{\theta^j}\in {\mbox{\rm IBr}}(\overline{b_j})$. This
block's Brauer correspondent in $N_{C_G(D_j)}(D)$ is
$\overline{\tilde{b_j}}$, that contains the unique
irreducible Brauer character $\overline{\tilde{\theta^j}}$.
Since blocks of $C_{N_G(D)}(D_j)$ that induce $B_0$ are conjugate in $N_{N_G(D)}(D_j)=
N_G(D)$, there exists an element $z_1\in N_G(D)$ with ${\tilde{b_j}}^{z_1}=\overline
{\tilde{b_j}}$. Then ${b_j}^{z_1}=\overline {b_j}$
and ${\tilde {\theta^j}}^{z_1}=\overline{\tilde{\theta^j}}$.
But then ${b_j}^{zz_1}={b_j}$. But the inertia group of
$b_j$ in $N_j$ is $C_j$, thus $zz_1\in C_j$ and so $({x^{{2^j}k}})^{zz_1}=x^{{2^j}k}$, and
$({x^{{2^j}k}})^{-1}=({x^{{2^j}k}})^z=({x^{{2^j}k}})^{z_1}$, and hence
$(({x^{{2^j}k}})^{z_1},{\tilde {\theta^j}}^{z_1})=(({x^{2^{j}k}})^{-1},\overline{\tilde{\theta^j}})$. Thus it represents a real column of $B_0$.
The remaining column of $B$ containing $(1,\phi)$ is real and the corresponding column
containing
$(1,\tilde\phi)$ in $B_0$ is also real. So we are done.
Now we have:
\begin{theorem}
Let $G$ be a finite group, let $B\in {\mbox{\rm Bl}}(G|D)$ be a $2$-block with cyclic defect group
$D$. Then Conjecture \ref{conj1} and hence Conjecture \ref{conj3} also holds for $B$.
\end{theorem}
\noindent
Proof.
Using Lemma \ref{samenumber} and Corollary \ref{normal}, we have that
$|{\rm Irr}_{rv}(B)|=|{\rm Irr}_{rv}(B_0)|$ is bounded from above by the number of
the $N_G(D)$-real elements of $D$, thus
we are done.
\section{Groups with odd real core}
In \cite{HH} we defined the {\bf real core} $R(G)$ of $G$ as the subgroup
generated by the real elements of odd order.
\bigskip
Our main result is the following:
\begin{theorem}\label{theorem1}
If $|R(G)|$ is odd then Conjecture~\ref{conj3} holds (hence also Conjectures
\ref{conj2}, \ref{conj1}
and \ref{conj0}) for the principal $2$-block of
$G$.
In particular if any of the following cases occur Conjecture~\ref{conj3} holds for
the principal $2$-block of $G$.
\begin{itemize}
\item [(a)] The commutator subgroup $G'$ is $2$-nilpotent.
\item [(b)] $G=O_{2',2,2'}(G)$. (In fact this is equivalent to $|R(G)|$ being odd.)
\item [(c)] $G$ is solvable and its Sylow $2$-subgroup is abelian.
\end{itemize}
Moreover, we may replace in Conjecture~\ref{conj3} $N_G(D)$ by $D$.
\end{theorem}
\noindent
Proof. If $|R(G)|$ is odd then by \cite{HH} $G=O_{2',2,2'}(G)$, in particular
$G$ is solvable.
Let $B_0$ be the principal $2$-block of $G$.
Then ${\rm Irr}_{rv}(B_0)={\rm Irr}_{rv}(G/O_{2'}(G))$.
Let $\overline G=G/O_{2'}(G)$. Then $\overline S\in Syl_2(\overline G)$ is normal.
\noindent
Step 1 : For every real element $x$ there exists a $2$-element $g$
such that $x^g=x^{-1}$:
Let $g=g_2g_{2'}$ if $x^{g_2g_{2'}}=x^{-1}$, then an appropriate $2$-power
of $g$ is already a $2'$-element, which centralizes $x$. Thus $g_{2'}$ acts on $x$
trivially, and we are done.
\noindent
Step 2: $R(\overline G)=1$:
If $x\in \overline G$ is a real element of odd order, then by Step 1 there is
a $2$-element g inverting $x$. Since $\overline S\triangleleft \overline G$,
$[x,g]\in \overline S\cap \langle x \rangle =1$.
Thus $x^{-1}=x$, and so $x=1$.
\noindent
Step 3: Every real element in $\overline G$ is a $2$-element, hence it lies in
$\overline S\in Syl_2(\overline G)$:
Let $x=x_2x_{2'}$ be a real element in $\overline G$. Then ${x_2}^{-1}{x_{2'}}^{-1}=
x^{-1}=x^g={x_2}^g{x_{2'}}^g$, thus $x_2$ and $x_{2'}$ are both real. By Step 2,
$x_{2'}=1$.
Thus $|{\rm Irr}_{rv}(B_0)|=|{\rm Irr}_{rv}(\overline G)|=|{\rm Cl}_{r}(\overline G)|\leq |\{
x\in\overline G | x$ ${\rm real} \}|$$=|\{ x \in \overline S\in Syl_2(\overline G)|
x $ ${\rm real}$ ${\rm in}$ $\overline S \}$.
We prove Conjecture~\ref{conj3} by induction on $r$. Let $r=0$. An irreducible character $\chi \in {\rm Irr}(B_0)$ is of height zero iff its degree is odd.
We have that $\chi \in {\rm Irr}(\overline G)$, and $\chi_{\overline S}$ has only linear constituents, hence
${\overline{S}}'\leq \ker(\chi)$. Thus ${\rm k}_{0,rv}(B_0)\leq |{\rm Irr}_{rv}(\overline G/{\overline {S}}')|=|{\rm Cl}_r(\overline G/{\overline {S}}')|\leq |\{ x \in \overline G/{\overline {S}}'| x$ ${\rm real} $ ${\rm in}$ $\overline G/{\overline {S}}'$ $ \}|$.
If there would be a real $2'$-element in $G/{\overline {S}}'$ then by
Proposition 5.3 in \cite{HH} there would be a real $2'$ element in $\overline G\backslash
{\overline {S}}'$, which is not the case.
Thus there are also no real elements in $\overline G/{\overline {S}}'$ whose
$2'$-part is not $1$. Thus each real element belongs to
$\overline S/{\overline {S}}'$, and by Step 1 this element is
$\overline S /{\overline {S}}'$-real.
Thus we are done for $r=0$.
Let us suppose that Conjecture~\ref{conj3} is true for $r<n$.
If $\chi \in {\rm Irr}(B_0)$ is of height $n$, then its degree has $2$-part $2^n$.
Then all constituents of $\chi_{\overline S}$ have degree $2^n$. By
\cite[Thm.~5.12]{Is}, they contain in their kernels $\overline {S}^{(n+1)}$,
thus $\chi$ also contains it in its kernel. Similarly all irreducible characters of
$\overline G$ of smaller height also contain it in their kernels.
Hence $\sum_{i=0}^n k_{i,rv}(B_0)\leq |{\rm Irr}_{rv}(\overline G/{\overline S^{(n+1)})|=
|{\rm Cl}_r(\overline G/{\overline S^{(n+1)}}|\leq |\{x\in \overline G/{\overline S}^{(n+1)}}| x$ {\rm real}$\}|=
| \{x\in \overline S/{\overline S}^{(n+1)}| x $ {\rm real in\ }$
{\overline S}/{\overline S}^{(n+1)}\}|$. Hence
Conjecture~\ref{conj3} holds.
\begin{itemize}
\item[(a)] Since $R(G)\leq G'$ by \cite{HH}, if $G'$ is $2$-nilpotent, then $|R(G)|$ is odd.
\item[(b)] This is equivalent to $|R(G)|$ odd by \cite{HH}.
\item[(c)] By the Hall-Higman lemma, $\overline S$ is normal, thus we have case (b).
\end{itemize}
\noindent
\begin{corollary}
If $S\in Syl_2(G)$ is normal then Conjecture $\ref{conj3}$ holds for $G$, since
then each block is of maximal defect, and the only real $2$-block of maximal defect is the principal block, hence we can apply Theorem \ref{theorem1} (b).
\end{corollary}
\section{Computer results}
\bigskip
We have checked Conjecture~\ref{conj3} for the principal $2$-block
with \textsf{GAP} \cite{GAP} for the small groups library. We also checked
Conjecture~\ref{conj1} for the principal $2$-block
for the $26$ sporadic simple groups.
For these blocks the respective conjectures were true.
We also checked Conjecture \ref{conj1}
and Conjecture \ref{conj2} with the help of $GAP$ for all
$2$-blocks of groups up to order $1536$ except for groups of orders
$856,1048,1112,1192,1304,1352,1384,1432,1448$, where our computational methods did not work (Conway polynomials are not yet known).
We did not find any conterexamples for these conjectures among the investigated groups.
\section{$B$-classes and $G$-classes of $D$}
Let now $p$ be an arbitrary prime number.
First we prove the following
\begin{lemma}\label{Hiss}
Let $B$ be a $p$-block of $G$ with defect group $D$.
Then for every $x\in D$ there exists a
$\chi \in {\rm Irr}(B)$ with $\chi(x)\ne 0$.
\end{lemma}
\noindent
Proof. Let us suppose by contradiction that there exists
an element $x\in D$ with $\chi(x)=0$, for every $\chi\in {\rm Irr}(B)$.
If we can prove that there exists a trivial source $FG$-module
$M$ in this block with vertex $D$,
then by \cite[p.~175, Lemma~2.16]{Lan} this is
liftable to a trivial source $RG$-module $\tilde M$ and its character is nonzero on the elements of the vertex of $M$,
contradicting our assumption.
If $D$ is normal in G then by \cite[p.~247, Lemma~10.3]{Lan}
all simple modules in
$B$ are trivial source modules in $B$ with vertex $D$.
If $D$ is not normal then the Brauer correspondent
$b\in {\mbox{\rm Bl}}(N_G(D)|D)$ of $B$ has the property that every simple module
$S$ in it is a trivial source module with vertex $D$.
Let us lift a simple $FN_G(D)$-module $S$ in $b$ to an $RN_G(D)$-module
$\tilde S$. Then its Green correspondent, $f(\tilde S)$ is a
trivial source module
with
vertex $D$ and by \cite[p.~466, Thm.~59.9]{C-R}, $f(\tilde S)$ belongs
to the block $B$.
So we are done.
\noindent
\begin{definition} Let $B$ be a $p$-block of the finite group $G$ with defect group
$D$. We say that two elements $x,y\in D$ are in the same $B$-class, iff for every
irreducible character $\chi\in {\rm Irr}(B)$, $\chi (x)=\chi (y)$.
\end{definition}
We have the following result:
\begin{theorem}\label{bclass} Let $G$ be a finite group with $p$-block $B\in {\mbox{\rm Bl}}(G|D)$.
Then the $B$-classes of the defect group $D$ are exactly the $G$-classes ${\rm Cl}_G(D)$
of $D$ under conjugation.
\end{theorem}
\noindent
Proof. If two elements $x,y\in D$ are conjugate in $G$, then of course they are also in the same $B$-class.
Let us suppose now that $x,y\in D$ are in the same $B$-class, but they
are not conjugate in $G$. Then by the strong
block orthogonality relation, see \cite[p.~106, Cor.~5.11]{Nav} $\sum_{\chi\in {\rm Irr}(B)}\chi(x)\overline{\chi(y)}=0$.
Using that $x,y$ are in the same $B$-class this gives us
$\sum_{\chi(x)\in {\rm Irr}(B)}|\chi(x)|^2=0$.
Hence $\chi(x)=0$ for every $\chi\in {\rm Irr}(B)$.
By Lemma \ref{Hiss}, this is not possible.
\begin{definition} Let $B$ be a $p$-block of a finite group
$G$ with defect group $D$. We say that the element $x\in D$ is
$B$-real if $\chi(x)$ is real for every $\chi\in {\rm Irr}(B)$.
An element $x\in D$ is $B$-rational, if $\chi(x)$ is rational
for every
$\chi \in {\rm Irr}(B)$.
\end{definition}
\begin{corollary} Let $B\in {\mbox{\rm Bl}}(G|D)$ and let $x\in D$. Then $x$ is
$B$-real iff
it is real in $G$.
\end{corollary}
\begin{corollary} Let $B\in {\mbox{\rm Bl}}(G|D)$ and let $x\in D$. Then $x$ is $B$-rational
iff it is rational in $G$.
\end{corollary}
\begin{corollary} Let
Let $F$ be a field containing $Q$.
Let $B\in Bl(G|D)$ and let $x\in D$. Then $\chi (x)\in F$ for every $\chi \in {\rm Irr}(G)$
if and only if $\chi (x)\in F$ for every $\chi \in {\rm Irr}(B)$.
\end{corollary}
We have also the following
\begin{theorem}
Let $B\in {\mbox{\rm Bl}}(G|D)$. The restrictions $\chi_D$ of $\chi\in {\rm Irr}(B)$
to the defect group $D$ generate the vector space of the restrictions
of all complex $G$-class
functions to $D$.
\end{theorem}
\noindent
Proof.
Let us choose representatives $x_i\in C_i\cap D, i = 1, \ldots, t$
of $G$-conjugacy classes $C_i$ intersecting $D$.
We want to prove that if we restrict the character table of $G$ to
these columns
and to those rows which belong to the block $B$, then these columns
are independent. It implies that this
submatrix has rank $t$, hence, it has also
$t$ independent rows. But then any complex vector
of length $t$ can be expressed by these rows and we are done.
Let us suppose that the above mentioned $t$ columns are dependent.
Then
there are coefficients $\alpha_1,...,\alpha _t$, not all zero
with the property
that
$\sum_{i=1}^t\alpha_i\chi(x_i)=0$, for all $\chi \in {\rm Irr}(B)$.
By \cite[Lemma~4.6, Ch.~5]{N-T} the subsum, where $x_i$-s belong to
any
$p$-section is also zero. But the $x_i$-s all belong to different
$p$-sections,
thus $\alpha_i\chi(x_i)=0$ for every $i = 1, \ldots, t$ and
every $\chi \in {\rm Irr}(B)$.
By Lemma \ref{Hiss} we see that there is no $x_i$ where
every
$\chi \in {\rm Irr}(B)$ vanishes. Hence $\alpha _i=0$ for all
$i = 1, \ldots, t$. Thus
the columns of the above restricted matrix are independent and
we are done.
In this way we get another proof of the following:
\begin{corollary}\label{lowerbound} For a block $B\in {\mbox{\rm Bl}}(G|D)$, the number of $G$-conjugacy classes
$|{\rm Cl}_G(D)|$ of its defect group, is a lower bound for the number ${\rm k}(B)$.
\end{corollary}
\begin{example}
It is not true, however that $B$-classes (hence $G$-classes) of the defect group $D$
are the same as $b$-classes (hence $N_G(D)$-classes)
for the Brauer correspondent block $b\in {\mbox{\rm Bl}}(N_G(D)|D)$ even for
$2$-blocks $B\in {\mbox{\rm Bl}}(G|D)$ with cyclic defect group.
Let $G=\texttt{SmallGroup}(288,375)$. Then the third $2$-block has cyclic defect group of order $8$,
it contains four $G$-real (hence $B$-real) elements and only two $N_G(D)$-real
(hence $b$-real) elements.
\end{example}
{\rm ACKNOWLEDGEMENTS}
The authors are very grateful to Professor Gerhard Hiss for the
discussions about the topic of this paper and also
for his suggestions and comments, especially for his
contributions to the proof of Lemma \ref{Hiss}.
We also thank Thomas Breuer for his advices concering the GAP system.
Research supported by National Science Foundation Research Grant No. T049841
and K 77476.
\bibliographystyle{amsalpha}
|
\section{Introduction}
One of the most fascinating aspects of graphene is its band structure which can be fundamentally changed in several
different ways by modifying its lattice structure. This happens because the honeycomb lattice of monolayer graphene
has two independent sublattices and the electron, as it moves through the lattice, has to change its sublattice and
hence the character of its wavefunction.\cite{Revs} Thus, even small local changes in the lattice structure lead to the appearance
of gauge potentials which are associated with the phase of the electronic wavefunction in each sublattice. As a consequence,
there is a one-to-one correspondence between graphene's structure with the topological features of the electronic states.
Another amazing property of graphene is its honeycomb structure which yields an electronic low-energy effective theory
which is \textit{Lorentz-invariant} in two dimensions (2D), and thus corresponds to 2D Dirac fermions. This Lorentz
invariance is robust because the energy associated with sublattice coupling, that is, the inter-sublattice hopping energy
$t$ ($ \approx 3$ eV) is the dominant energy scale in the system. Lorentz invariance persists even when the lattice is
modified either by external forces (strain, shear, etc.),\cite{strain} by external fields, or by the addition of more
layers.\cite{Lorentz} For instance,
AB-stacked bilayer graphene is described, at low energies by two sets of massive Lorentz invariant Dirac particles per valley and spin.
In the simplest models where only nearest neighbor hoppings are taking into account, there is still
an accidental degeneracy that makes the particle-like band of one flavor to be degenerate with the anti-particle-like (hole-like)
band of the other flavor at the K (K') point in each valley. This degeneracy can be easily lifted by the application of a
perpendicular electric field that breaks the inversion symmetry in the system.\cite{Revs} Although the Lorentz invariance
is preserved, the wavefunction of the electrons at low energies is modified$-$whereas in monolayer graphene the Dirac fermions
carry a Berry phase $\pm \pi$, the Berry phase is $\pm 2\pi$ in AB-stacked bilayer graphene.\cite{MF06} At low energies, trigonal
warping splits this ``double'' Dirac point into three with a Berry phase $\pi$ and an additional one with $-\pi$,\cite{MF06}
a situation that persists for a translational mismatch between the layers.\cite{Son}
Twisted bilayer graphene, in which the two layers have a rotational mismatch described by an
angle $\theta$ as compared to the perfect AB stacking, is another example where lattice structure and wavefunction topology
are directly interconnected. In fact, from the experimental point of view, twisted graphene is more the rule than the exception.
It naturally occurs at the surface of graphite,\cite{Rong,Pong} in graphene grown in the surface of SiC,\cite{Hass} or
graphene grown by chemical vapor deposition on metal substrates.\cite{Eva} As compared to monolayer graphene, each Dirac flavor
is then split into two copies that are separated in reciprocal space by a wave vector $\Delta{\bf K}$, as a function of $\theta$.\cite{Lopes}
Inter-layer hopping results in a renormalization of the Fermi velocity\cite{Heer,Schmidt,Lopes,Shallcross} as well as
in a van Hove singularity at relatively low energy as compared to that in monolayer graphene.\cite{Lopes,Eva} Once again,
Lorentz invariance is preserved at low energies but the nature of the electronic wavefunctions is modified in a profound way.
In this paper, we investigate the topological aspects of the band structure of twisted graphene bilayers at low energies in the continuum
approximation, for small angle mismatches as compared to perfect AB stacking. We identify possible topological classes that describe band
inversion symmetry. These topological classes determine the relative Berry phase between the two copies of Dirac particles, which
have either the same or opposite Berry phase. If the two Dirac cones are related by time-reversal symmetry, such as the
$K$ and $K'$ points in monolayer graphene, the Berry phases are naturally opposite. In twisted graphene bilayers, however, the two Dirac points emanating from different layers are not time-related, and the symmetry of the inter-layer hopping term enforces
the Berry phases to be identical. We show that this feature yields a topologically
protected zero-energy Landau level, in contrast to the former case. The scenario may be tested in
quantum Hall measurements.
The paper is organized as follows. In Sec. \ref{sec:Mod4}, we discuss the model of twisted bilayer graphene and the symmetry
properties of their bands (Sec. \ref{sec:sym}) that are fixed by the form of the inter-layer hopping. Furthermore, we present an
effective two-band model (Sec. \ref{sec:Mod2}) that displays the same topological low-energy properties as the original four-band model.
Section \ref{sec:LL} is devoted to the discussion of the Landau-level spectrum in twisted bilayer graphene, in the perspective of
possible quantum-Hall measurements.
\begin{figure}[t]
{\label{bla}\includegraphics[width=0.18\textwidth]{FirstBZ_twisted.pdf}}
\caption{First Brillouin zone for twisted bilayer graphene. The 1BZ of the upper layer (dashed hexagon) is rotated by an
angle $\theta$ with respect to that of the lower layer (full hexagon). The corners, where Dirac points occur, are labeled by
$K_{\theta}^{(\prime)}$ and $K^{(\prime)}$, respectively.
}
\label{fig01}
\end{figure}
\section{Model of twisted bilayer graphene}
\label{sec:Mod4}
If we neglect, for the moment, hopping
between atoms in different layers, the electronic properties of twisted bilayer graphene are
described by two copies of the Hamiltonian for monolayer graphene (we use units with $\hbar=1$):
\begin{equation}\label{eq:ham0}
H_0({\bf k})\equiv v_F \left(\begin{array}{cc} 0 & k^*\\ k & 0
\end{array}\right),
\end{equation}
where $v_F$ is the Fermi velocity and $k=k_x+i k_y$ is the 2D wave-vector relative to the K (K') points of the rotated
layers (see Fig.\ref{fig01}). For a twist (rotation) angle $\theta\neq 0$, each of the two inequivalent Dirac points, which reside at the corners of the first BZ $K$ and $K'$,
are split into two, separated by a wave vector $\Delta {\bf K}={\bf K} -{\bf K}_{\theta}$, where ${\bf K}_{(\theta)}$
is the position of the $K$ point in the lower (upper) layer and $-{\bf K}_{(\theta)}$ the position of the
points $K'$ and $K_{\theta}'$, respectively.
Throughout this paper, we will work in the continuum limit around a single pair ($K$,$K_{\theta}$) and, therefore, neglect commensuration effects between the two layers which could
enlarge the unit cell in position space and thus fold it back in reciprocal space. This procedure is mostly justified because the coupling between the pairs of Dirac cones\cite{Mele,MacDonald} is negligible.\cite{Shallcross}
The Hamiltonian describing the electronic properties of twisted bilayer graphene reads
\begin{equation}\label{eq:ham1}
H({\bf k})=\left(\begin{array}{cc} H_0({\bf k}+\Delta {\bf K}/2) & H_{\perp}\\ H_{\perp}^{\dagger} & H_0({\bf k}-\Delta{\bf K}/2)
\end{array}\right),
\end{equation}
where $H_{\perp}$ is the hopping matrix between the two layers.
Equation (\ref{eq:ham1}) refers to an expansion around the ${\bf Q}$ point of Fig. \ref{fig01}.
The analysis of the Moir\'e pattern formed by the twisted bilayer shows that, for small angle $\theta$, the hopping matrix
$H_{\perp}$ may have three different forms corresponding to the three main Fourier components.\cite{Lopes,MacDonald}
This leads to three different types of inter-layer hopping terms,
\begin{equation}
\label{eq:hopping}
H_{\perp}^0 \equiv \tilde{t}_{\perp}
\left(
\begin{array}{cc}
1 & 1 \\
1 & 1 \\
\end{array}
\right),
\quad
H_{\perp}^{\pm} \equiv \tilde{t}_{\perp}
\left(
\begin{array}{cc}
e^{\mp i\phi} & 1 \\
e^{\pm i\phi} & e^{\mp i\phi} \\
\end{array}
\right),
\end{equation}
where $\phi=2\pi/3$ and $\tilde{t}_{\perp}$ is a hopping parameter which generally depends on $\theta$.\cite{Lopes,MacDonald,note}
\subsection{Symmetry of the bands and Berry phases}
\label{sec:sym}
In contrast to a lattice Hamiltonian that may be analyzed with the help of global symmetries, such as time reversal or lattice inversion,
$H({\bf k})$ in Eq. (\ref{eq:ham1}) is a continuum model in which the two Dirac points are no longer related by these symmetries.
However, $H({\bf k})$ and the inter-layer hopping term $H_{\perp}$ may be investigated via symmetries that involve directly
the energy bands, such as rotation, mirror, and inversion symmetry. Whereas for $H_{\perp}=0$ the rotation and inversion symmetries
are respected, the latter are broken for non-zero inter-layer hopping, and we therefore restrict the discussion to the inversion
symmetry $\mathcal{I}$ of the bands. We emphasize that this inversion symmetry is defined with respect to the bands in reciprocal space, in contrast to a previous analysis,\cite{Paco} in which the more common definition of inversion symmetry with respect
to the lattice was used.
In the absence of any inter-layer hopping, the Hamiltonian (\ref{eq:ham1}) would be
ambiguous since one could work equally well with $H_0^*$
in the second layer. As a consequence, there remain two possible representations of inversion (with different spinorial expressions) $\mathcal{I}_1$ and $\mathcal{I}_2$:
\begin{equation}
\label{eq:inv2}
\mathcal{I}_1:H(-{\bf k})=-H({\bf k}) \qquad {\rm or} \qquad \mathcal{I}_2: H(-{\bf k})=-H({\bf k})^*.
\end{equation}
The minus signs in both transformations maps positive to negative energy states
at opposite wave vectors and {\sl vice-versa}: $E({\bf k})\rightarrow -E(-{\bf k})$,
whereas the complex conjugation in $\mathcal{I}_2$ changes the relative phase between the spinor components and, hence,
the Berry phase of a cone.
The phases of the two Dirac points are thus different for the two representations.
For the $\mathcal{I}_1$ case, the Berry phases at fixed energy
are opposite such that a merging transition of the two Dirac points (at $\theta=0$) corresponds to a vanishing total Berry phase and consequently to the possible opening
of a band gap. This situation arises e.g. in the framework of the model discussed in Refs. \onlinecite{listDPmerg}, which describes
monolayer graphene under strong strain.\cite{strain}
In contrast to this rather well-known topological universality class, the transformation
$\mathcal{I}_2$ yields Berry phases that are the same at fixed energy $E$ and opposite to those at $-E$ and thus represents
a second universality class for Hamiltonians describing pairs of Dirac points.
The transformation $\mathcal{I}_2$ may be represented in terms of a tensor product of Pauli matrices,
$\mathcal{I}_2=\sigma_y^E\otimes \sigma_x^A$, where $\sigma_{x/y}^{A}$ and $\sigma_{x/y}^{E}$ describe the intra-layer
and the inter-layer spinorial spaces, respectively.
The phases are eventually fixed by the symmetry of the interlayer hopping $H_{\perp}$, and the particular forms (\ref{eq:hopping})
happen to be invariant under $\mathcal{I}_2$.
Notice, however, that as a consequence of the
inter-layer hopping terms (\ref{eq:hopping}), the two Dirac points are no longer exactly at the same energy, but we
neglect this energy shift here because it is associated with a very small energy scale $\sim 1$ meV. As a second-order perturbation, this indeed scales like $\tilde{t}_{\perp}^2/(v_F \Delta {\bf K})$ with $\tilde{t}_{\perp}$ of the order of $100$ meV and $v_F \Delta {\bf K} \sim 1$ eV.\cite{Lopes}
\subsection{Effective two-band model}
\label{sec:Mod2}
In the opposite limit, $\tilde{t}_{\perp}\gg v_F\Delta K$, the model (\ref{eq:ham1}) may be reduced to an
effective two-band model, similarly to the case of a perfectly AB-stacked ($\theta=0$) graphene bilayer.\cite{MF06}
This is done in two steps. First, one replaces $H_{\perp}$ in Eq. (\ref{eq:ham1})
by a simplified inter-layer hopping term,
\begin{equation}\label{eq2}
H_{\perp}^{\rm eff}=\tilde{t}_{\perp}\left(\begin{array}{cc}
0 & 0 \\
1 & 0
\end{array}\right),
\end{equation}
in the limit where
$\tilde{t}_{\perp}\gg v_F\Delta K$, i.e. for small tilt angles. The inter-layer hopping term (\ref{eq2})
is reminiscent of the Bernal bilayer case.
In spite of this simplification and the difference in the energy scales,
the resulting Hamiltonian may be viewed as a representative of the
topological universality class that also includes the original four-band model.
We consider the eigenvectors of Eq. (\ref{eq2}) in terms of the 4-spinor basis $\left\{ \psi_A, \psi_B, \psi_{A'}, \psi_{B'} \right\}$ where $(A,B)$ belongs to the two sublattices of the first layer and $(A',B')$ to that of the second one.
With the particular form of Eq. (\ref{eq2}), the zero-energy sector is spanned
$\psi_A$ and $\psi_{B'}$, whereas $\psi_{B}$ and $\psi_{A'}$ are strongly hybridized by the inter-layer hopping term.
Their symmetric and anti-symmetric combinations are the eigenstates at $-\tilde{t}_{\perp}$ and $\tilde{t}_{\perp}$, respectively.
In order to describe the electronic properties in the vicinity of $E=0$, one may therefore project the Hamiltonian onto the reduced $\left\{ \psi_A,\psi_{B'}\right\}$ basis and neglecting terms of the form $E \psi_{B/A'}$, which are a product of the energy
$E\sim 0$ and the small components $\psi_{B/A'}$. The eigenvalue problem reads
\begin{eqnarray}
\label{eq3a}
v_F (k+\Delta K/2)^*\psi_B & = & E\psi_A\\
\label{eq3b}
v_F (k+\Delta K/2)\psi_A + \tilde{t}_{\perp}\psi_{A'} & = & E \psi_B \simeq 0 \\
\label{eq3c}
\tilde{t}_{\perp}\psi_{B} + v_F (k-\Delta K/2)^*\psi_{B'} & = & E\psi_{A'} \simeq 0 \\
\label{eq3d}
v_F (k-\Delta K/2)\psi_{A'} & = & E\psi_{B'}.
\end{eqnarray}
By rewriting Eqs. (\ref{eq3c}) and (\ref{eq3d}),
\begin{eqnarray}
-v_F/\tilde{t}_{\perp} (k+\Delta K/2)\psi_A & =& \psi_{A'} \\
-v_F/\tilde{t}_{\perp} (k-\Delta K/2)^*\psi_{B'} & =& \psi_{B},
\end{eqnarray}
and substituting them into Eqs. (\ref{eq3a}) and (\ref{eq3d}), one obtains the Schr\"odinger equation
\begin{equation}
H^{\text{eff}}({\bf k})\left(
\begin{array}{c}
\psi_A\\ \psi_{B'}
\end{array}
\right)
= E
\left(
\begin{array}{c}
\psi_A\\ \psi_{B'}
\end{array}
\right),
\end{equation}
in terms of the effective two-band Hamiltonian
\begin{equation}
\label{eq:hamKK}
H^{\text{eff}}({\bf k})=-\frac{v_F^2}{\tilde{t}_{\perp}}
\left(
\begin{array}{cc}
0 & \left(k^*\right)^2 -\left(\Delta K^*\right/2)^2 \\
k^2-\left(\Delta K/2\right)^2 & 0 \\
\end{array}
\right),
\end{equation}
which is similar to that for the nematic transition of the interacting Bernal graphene bilayer.\cite{Vafek}
\begin{figure}[t]
\centering
{\label{bla1}\includegraphics[width=0.4\textwidth]{Group.pdf}}
\caption{(Color online) Generic band structure for the lowest four bands of the twisted bilayer graphene, around the ${\bf Q}$ point, with only $H_{\perp}^0$ considered. The inset of the figure pictures the effective bands of the model Hamiltonian (\ref{eq:hamKK}). Both band structures are inversion-invariant with respect to the ${\bf Q}$ point.}
\label{fig03}
\end{figure}
The generic form of the band structure obtained from the effective two-band model (\ref{eq:hamKK}) is depicted in the inset of Fig. \ref{fig03}.
One notices the two Dirac points originally situated at ${\bf K}$ and ${\bf K}_{\theta}$, separated by a wave vector $|\Delta {\bf K}|\sim \theta/a$
in the first BZ, in terms of the intra-layer distance $a=0.142$ nm between neighboring carbon atoms.
In order to see the linearity of the dispersion relation in the vicinity of these contact points, one can further expand the Hamiltonian
(5) around $\pm \Delta {\bf K}/2$, by defining ${\bf k} = {\bf q} \pm \Delta {\bf K}/2$, with $|{\bf q}| \ll |\Delta {\bf K}|/2$. This expansion yields two
Dirac Hamiltonians
\begin{equation}\label{eq:twoD}
\pm \frac{v_F^2 \Delta K}{\tilde{t}_{\perp}}
\left(
\begin{array}{cc}
0 & q^* \\
q & 0
\end{array}
\right),
\end{equation}
with identical chirality for the two contact points.
Furthermore,
the bands have saddle points at ${\bf k}=0$ between the two Dirac points. The effective
two-band Hamiltonian therefore captures the logarithmic van-Hove singularity
in agreement with previous theoretical\cite{Lopes} and experimental studies.\cite{Li10}
The most salient features of the Hamiltonian (\ref{eq:hamKK}) are its chiral properties.
In agreement with the above-mentioned general symmetry considerations, electrons at a fixed energy
at the ${\bf K}$-point have the same chirality as those at the ${\bf K}_{\theta}$ point, such that in both cases the electron
experiences the same Berry phase $\gamma=\pi$ (and $\gamma=-\pi$ at the points ${\bf K}'$ and ${\bf K}_{\theta}^{\prime}$)
on a closed orbit around one of the Dirac points. This is obvious in the low-energy expansion (\ref{eq:twoD}) around the
two Dirac points. As for AB-stacked bilayer graphene,
the Berry phase acquired on an orbit enclosing both Dirac points is then $2\gamma$, as one may also see from the limit
of vanishing twist angle ($\theta=0$) which reproduces the perfectly AB-stacked bilayer.\cite{MF06} Furthermore, it becomes apparent
from the form of the Hamiltonian (\ref{eq:hamKK}) that the merging of the Dirac points is not accompanied with a gap opening,
in contrast to the Hamiltonian discussed in Ref. \onlinecite{listDPmerg}, which desribes the same band structure in the semi-metallic
phase but which belongs to another topological class, described by the symmetry $\mathcal{I}_1$.
\section{Landau levels of twisted bilayer graphene}
\label{sec:LL}
One of the most prominent consequences of these topological properties is the presence of a (doubly-degenerate) zero mode
that emerges in the presence of a quantizing magnetic field, in which case the Hamiltonian (\ref{eq:hamKK}) may be
written in terms of the usual ladder operators $a$ and $a^{\dagger}$, with $[a,a^{\dagger}]=1$:
\begin{equation}\label{eq:hamKK_B}
H_B=\omega_C \left(
\begin{array}{cc} 0 & a^{2} - \alpha^{*2} \\ a^{\dagger 2} - \alpha^{2} & 0 \end{array}
\right),
\end{equation}
where $\omega_C= 2v_F^2eB/\tilde{t}_{\perp}$ is the cyclotron frequency and $\alpha\equiv \Delta K l_B/2\sqrt{2}$.
As compared to the perfectly AB-stacked bilayer ($\alpha=0$), where one readily obtains the Landau level (LL) spectrum,\cite{MF06}
one notices that the additional
terms couple states only of the same parity. One obtains thus two classes of eigenstates $\psi_{2n}$ and $\psi_{2n+1}$
that may be written in the usual
harmonic-oscillator basis $|m\rangle$, with $a^{\dagger}a|m\rangle = m|m\rangle$:
\begin{equation}\label{eq:eigen}
\Psi_{2n(+1)} = \sum_{m=0}^{\infty}\left(\begin{array}{c} \phi_{2m(+1)}^1 \\ \phi_{2m(+1)}^2
\end{array}\right) |2m(+1)\rangle,
\end{equation}
where the components $\phi_{m}^1$ and $\phi_{m}^2$ are to be determined recursively.
\subsection{Zero-energy levels}
\label{sec:0LL}
Before discussing the LL spectrum of Hamiltonian (\ref{eq:hamKK_B}), we investigate the zero-energy states, which may be obtained
analytically from the equation $H_B \Psi=0$.
As for the AB-stacked bilayer, one obtains two distinct solutions, one with even and one with
odd parity, such that the zero-energy level
is orbitally two-fold degenerate, in addition to its usual four-fold spin-valley degeneracy and the orbital degeneracy described
by the flux density $n_B=eB/h$. The two zero-energy states, which are reminiscent of coherent states,
are directly obtained from the secular equation:
\begin{eqnarray}\label{eq:states0}
\Psi_0 &=&
\mathcal{N}_0 \cosh(\alpha^* a^{\dagger})\left(\begin{array}{c} 0 \\ |n=0\rangle \end{array}\right),
\\
{\rm and} \qquad
\Psi_1 &=&
\mathcal{N}_1 \frac{\sinh(\alpha^* a^{\dagger})}{\alpha^*}\left(\begin{array}{c} 0 \\ |n=0\rangle \end{array}\right),
\end{eqnarray}
in terms of the normalization factors $\mathcal{N}_{0/1}$.
These states are generalizations of the zero-energy states $(0,|0\rangle)$ and $(0,|1\rangle)$ of the AB-stacked
bilayer.\cite{MF06}. We also notice that we neglect any other potential lifts due to Zeeman effect and/or interactions.\cite{goerbig10}
\begin{figure}[t]
\centering
{\label{bla2}\includegraphics[width=0.4\textwidth]{landau.pdf}}
\caption{(Color online) Landau levels of twisted bilayer graphene, obtained from the numerical solution of Eq. (\ref{eq:hamKK_B}). The characteristic
energy scale of the van-Hove singularity has been chosen as $|\alpha|^2\omega_C=(\Delta K)^2/8m=0.1$eV.
The red line indicates the zero-energy level, and the numbers correspond to the filling factors in the gaps.
}
\label{fig02}
\end{figure}
The existence of those zero-mode states is independent of the strength of the magnetic field. This is to be compared to the other topological class of Dirac cones with opposite Berry phases where the degeneracy of the zero-mode is lifted.\cite{listDPmerg} This protection is solely determined by the topology of the model Hamiltonian since the band-structures are otherwise identical. Notice further that the scaling of the Landau levels relative to $\Delta {\bf K}$ is different for the two topological classes.
\subsection{Full Landau-level spectrum}
\label{sec:nLL}
The full LL spectrum, which has been calculated numerically, is depicted in Fig. \ref{fig02}. In the regime of small magnetic fields,
the LLs with small index $n\ll n_C$, where $n_C$ denotes roughly the LL which crosses the van-Hove singularity,
$\sqrt{n_CB}\sim v_F|\Delta K|^2/4\sqrt{2}\tilde{t}_{\perp}$,
display the $\sqrt{Bn}$ behavior, which is the benchmark of massless Dirac fermions, as expected for the linear dispersion below
the van-Hove singularity. Because of the two Dirac points, these LLs, as well as the zero-energy
level discussed above, are twofold degenerate, in addition to the fourfold spin-valley degeneracy. The twofold degeneracy due to the
Dirac-point splitting in the twisted bilayer is lifted once the LLs approach and eventually cross the van-Hove singularity,
$n\gtrsim n_C$, above which the LLs scale as $\sim B(n+1/2)$, as one would expect from the parabolic dispersion relation revealed
by Hamiltonian (\ref{eq:hamKK}) at high energies. Notice, however, that as for AB-stacked bilayer graphene the effective two-band
approximation is no longer valid at higher energies (far beyond the van-Hove singularity) because of the presence of the remaining bands, which become visible there.
The LL spectrum in Fig. \ref{fig02} allows one to understand the main features of a quantum Hall effect (QHE) in twisted bilayer
graphene. The topologically protected zero-energy LL, with its altogether eightfold degeneracy, yields a QHE at filling
factors $\nu=\pm 4$ (taking into account spin degeneracy). This feature is independent of the interlayer hopping strength $t_{\perp}$ and of the van-Hove singularity,
which is triggered by the twist angle $\theta$. For other LLs, the position of the van-Hove singularity determines their
degeneracy. If the LLs remain well below the singularity, $n\ll n_C$, they maintain their eightfold degeneracy, and one would therefore
expect Hall plateaus at filling factors $\nu=\pm 4(2n+1)=\pm 4, \pm 12, \pm 20, ...$, but one would expect
additional plateaus at $\nu=\pm 8, \pm 16, ...$ for
LLs with $n\gtrsim n_C$.
The experiment\cite{Lee10} indicates, even at rather low magnetic fields of $B\sim 5$T, only
the zero-energy LL is eightfold degenerate, whereas a plateau at $\nu=\pm 8$ has been observed. This stipulates that
the energy of the van-Hove singularity is as small as the first excited LL.
\section{Conclusions}
In conclusion, we have investigated the topological band structure of twisted bilayer graphene, in the framework of a
symmetry analysis of the inter-layer hopping in the continuum limit. For small and moderate twist angles $\theta$, the two
copies of the Dirac point (that are not related by time-reversal symmetry) are described by the same Berry phase, due
to the the symmetry of the inter-layer hopping term.
Therefore, they
belong to a different topological class than the (usual) two Dirac points, which are related by time-reversal symmetry.
In the presence of a quantizing magnetic field,
these particular topological properties yield a protected zero-energy LL with an eight-fold degeneracy that may be evidenced in
quantum Hall transport measurements.
\acknowledgements
This work was supported by the ANR project NANOSIM GRAPHENE under Grant No. ANR-09-NANO-016 and by the Ecole Doctorale de Physique de
la R\'egion Parisienne (ED 107), AHCN acknowledges partial financial support from the U.S. DOE under grant DE-FG02-08ER46512.
FG is supported by by MICINN (Spain), grants FIS2008-00124 and CONSOLIDER CSD2007-00010.
We acknowledge fruitful discussions with Klaus von Klitzing, Nuno Peres, Jo\~ao Lopes dos Santos, and Jurgen Smet.
AHCN thanks the Laboratoire de Physique des Solides for hospitality.
|
\section{Introduction}
ALICE (A Large Ion Collider Experiment) is the dedicated heavy ion experiment at the CERN LHC, optimised to study matter under extreme conditions of temperature and pressure -- the Quark-Gluon Plasma (QGP) -- in collisions between heavy nuclei. With an energy up to 30 times higher than RHIC, the current energy frontier machine for heavy ion collisions at BNL, we expect both a very different type of QGP, e.g. in terms of initial temperature, lifetime and system volume, and an abundance of hard signals like jets and heavy quarks which serve as probes to study QGP properties. Data taking with pp (and later p-nucleus) is equally important, primarily to collect comparison data for the heavy ion program. In addition, a number of measurements concerning soft and semi-hard QCD processes in Minimum Bias (MB) and high multiplicity pp collisions are part of the initial physics program~\cite{Carminati:2004fp}, taking advantage of the specific and complementary capabilities of the detector. The large MB sample will also provide a detailed characterisation of global event properties over a range of LHC energies, which will be very useful for tuning Monte Carlo generators to better describe the QCD background underlying searches for new physics. With the high multiplicity pp sample, which is enriched with a dedicated trigger to reach particle densities comparable to nuclear collisions at lower energies, we can test if heavy ion like features become apparent as multiplicity increases or if they are dominated by hard multijet final states.
ALICE consists of a central part, which measures hadrons, electrons and photons, and a forward spectrometer to measure muons. The central part, which covers polar angles from $45^0$ to $135^0$ over the full azimuth, is embedded in the large L3 solenoidal magnet. It consists of an inner tracking system (ITS) of high-resolution silicon detectors, a cylindrical TPC, three particle identification arrays of Time-of-Flight (TOF), Cerenkov (HMPID) and Transition Radiation (TRD) counters, and two single-arm electromagnetic calorimeters (high resolution PHOS and large acceptance EMCAL). The forward muon arm ($2^0-9^0$) consists of a complex arrangement of absorbers, a large dipole magnet, and 14 stations of tracking and triggering chambers. Several smaller detectors for triggering and multiplicity measurements (ZDC, PMD, FMD, T0, V0) are located at small angles.
The main design features include a robust and redundant tracking over a limited region of pseudorapidity, designed to cope with the very high particle density of nuclear collisions, a minimum of material in the sensitive tracking volume (10\% radiation length between vertex and outer radius of the TPC) to reduce multiple scattering, and several detector systems dedicated to particle identification over a large range in momentum~\cite{JSPLHC2010,Evans:2009zz}.
The layout of the ALICE detector and its eighteen different subsystems are described in detail in~\cite{ALICEdet}. The experiment is fully installed, commissioned and operational, with the exception of two systems (TRD and EMCAL) which were added more recently and are only now nearing the end of construction. Both systems have currently about 40\% of their active area installed and will be completed during the winter shutdown in 2010/11 (EMCAL) and 2012/13 (TRD).
The LHC started operating in November 2009 with pp collisions at $\sqrt{s} = 900$ GeV and reached its current maximum energy of 7 TeV in March 2010. The primary ALICE goal for pp was to collect about $10^{9}$ minimum bias collisions under clean experimental conditions.
Therefore the experiment was running for much of 2010 in a special low luminosity mode, in which the LHC beams are separated by $ 3-5 \sigma$ in the ALICE interaction region, to keep event pile-up below a few percent per bunch crossing. The total pp data sample collected at 7 TeV (900 GeV) corresponds to some 700 M (7 M) MB triggers, 50 M muon triggers and about 20 M high multiplicity events. In addition, a very short test run was taken at 2.36 TeV with only a small subset of detectors being read out.
First heavy ion collisions at 2.76 TeV/nucleon started just prior to this conference in November 2010, after only a few days of switching over from the pp set-up. Like for pp, the LHC worked exceedingly well also for heavy ions, with a steep increase in luminosity, reaching about $2~10^{25} \rm cm^{-2} \rm s^{-1}$ towards the end of the run and yielding some 30 M nuclear MB interactions on tape.
The many years of preparation, analysis tuning with simulations, and detector commissioning with cosmics during much of 2008/9 paid off with most of the detector components working 'right out of the box' and rather close to performance specifications, with both pp and Pb--Pb collisions . Within days all experiments could show first qualitative (and some quantitative) results. By the time of writing these proceedings, many of the results presented here have been already submitted for publication; they are therefore only briefly summarised in these proceedings (see also~\cite{Schukraft:2010ru}). Further details and references can be found in the original publications and in the large number of ALICE contributions presented at this conference.
\section{Results from pp collisions}
The charged particle density ${\rm d}N_{\rm ch}/{\rm d}\eta$ as well as the multiplicity distributions were measured at 0.9, 2.36 and 7 TeV for inelastic and non-single diffractive collisions~\cite{:2009dt,Aamodt:2010pp,Aamodt:2010ft}. The energy dependence of the multiplicity is well described by a power law in energy, $s^{0.1}$, and increases significantly more strongly than predicted by most event generators. Most of this stronger increase happens in the tail of the multiplicity distribution, i.e. for events with much larger than average multiplicity. Likewise, neither the transverse momentum distribution at 900 GeV nor the dependence of average $p_T$ on $N_{\rm ch}$ is well described by various versions of generators~\cite{Aamodt:2010my}, in particular when including low momentum particles ($p_T < 500$ MeV). The shape of the $p_T$ spectrum as a function of multiplicity hardly changes below 0.8 GeV (which includes the large majority of all particles), whereas the power law tail increases rapidly with event multiplicity above about 1-2 GeV.
Bose Einstein (HBT) correlations between identical particles are used to measure the space-time evolution of the dense matter system created in heavy ion collisions; their interpretation in elementary reactions ($e^+e^-, pp$) is controversial. However, HBT measurements at LHC are important also in pp as a comparison for the heavy ion data, because non-HBT correlations, which must be subtracted, are expected to increase with energy (e.g. via increased jet and mini-jet activity in MB events). They should also clarify if systematic trends, seen e.g. as a function of multiplicity and pair momentum $k_T$, differ between pp and heavy ions. The experimental trends observed at both 900 GeV and 7 TeV~\cite{Aamodt:2010jj, Aamodt:2011kd} can be summarised as follows:
HBT radii, which are typically of order 1 fm in elementary collisions, increase smoothly with multiplicity as previously observed both at low (ISR, RHIC) and at high (FNAL) energies, a trend which is qualitatively similar to the one seen in heavy ion collisions.
However, the decrease of the radius with pair momentum appears to be much weaker in our MB data than seen at the Tevatron; it becomes increasingly significant only for high pp multiplicities.
However, the decrease of radius with pair momentum seems to be much weaker in our MB data than the one seen at the Tevatron by E735; it becomes increasingly significant only for high pp multiplicities. We observe, both in the data (with unlike-sign pairs) as well as in the Monte Carlo (both Pythia and Phojet), strong particle correlations at small relative pair momentum, i.e. in the region of the HBT effect, which had to be carefully studied and subtracted. If we evaluate the HBT radii without subtracting non-HBT correlations, as has been done traditionally in the past, we observe a much stronger momentum dependent decrease of the radius, as these minijet-like correlations become more pronounced at higher momentum, effectively simulating a decreasing radius. A detailed comparison in the various HBT components (side/long/out) between pp and nuclear reactions as a function of ${\rm d}N_{\rm ch}/{\rm d}\eta$ and $k_T$ reveals qualitatively similar, but quantitatively distinct trends. The question if high multiplicity pp collisions show 'collectivity' or 'expansion-like' features requires further digestion of these results and comparison with models.
At the LHC, by far the highest energy proton--proton collider, we have studied baryon transport over very large rapidity intervals
by measuring the antiproton-to-proton ratio at mid-rapidity~\cite{Aamodt:2010dx} in order to discriminate between various theoretical models of baryon stopping.
Baryon number transport over large gaps in rapidity is often described in terms of a nonlinear three gluon configuration called 'baryon string junction'. The dependence of this process on the size of the rapidity gap has been a longstanding issue (for large gaps, where it should be dominant), with advocates for both very weak and rather strong dependencies.
In either case, the $\bar{p}/p$ ratio at LHC is very close to 1.0, with the difference between various models only of the order of a few percent. So this ratio must be measured with high precision. While in the ratio many instrumental effects cancel, the very large difference between $p$ and $\bar{p}$ cross section for both elastic (track can get lost) and inelastic (particle can be absorbed) reactions with the detector material lead to corrections of order 10\%, even in the very thin central part of ALICE. As the corrections are much larger than the effect, a very precise knowledge of the detector material as well as of the relevant cross section values at low momentum is required. The former was measured with the data via photon conversions with a relative precision of $< 7\%$ (absolute precision of better than 7 per mil radiation length!); the latter had to be cross checked with experimental data and the Fluka transport code since all available versions of Geant overestimate the $\bar{p}$ cross sections by up to a factor of five. The $\bar{p}/p$ ratio was found to be compatible with 1.0 at 7 TeV and 4\% below 1.0 at 900 GeV, with an experimental uncertainty of about 1.5\%, dominated by the systematic error. This result favours models which predict a strong suppression of baryon transport over large gaps; they agree very well with standard event generators but not with those that have implemented enhanced proton stopping.
The yields and $p_T$ spectra of identified stable charged ($\pi$, K, p) and neutral strange particles
($\mbox{$\mathrm {K^0_S}$},\Lambda, \Xi, \phi$) have been measured for the small 900 GeV data set taken in 2009~\cite{Aamodt:2011zj,Collaboration:2010vf}. In most cases Phojet as well as several Pythia tunes are well below the data -- by factors of two to almost five -- and more so at high $p_T$ and for the heavier particles ($\Lambda, \Xi$). The ratio of $\Lambda$ to $\mbox{$\mathrm {K^0_S}$}$ agrees very well with the STAR data at 200 GeV but is significantly below the ratio measured by UA1 and CDF. This discrepancy merits further investigation; it could be due to differences between the experiments in the acceptance, triggers, or correction for feed-down from weak decays. Baryon to meson ratios are of particular interest, as they rise well above unity in nuclear collisions at RHIC. This 'meson-baryon' anomaly has been interpreted in the context of coalescence models as an indirect sign of the QGP, in which case it would not be obvious why similar ratios should be reached already in minimum bias pp interactions.
\section{Results from Pb--Pb collisions}
The main aim of ultra-relativistic heavy ion physics is to {\em search} for the QGP, to {\em measure} its properties, and along the way {\em discover} new aspects of QCD in the regime where the strong interaction is strong indeed. With the fantastic results from RHIC, I consider the {\em search} for the QGP essentially over and the {\em discovery} phase well on the way. However, {\em precision measurements} of QGP parameters are just starting (precision is meant here in the context of non-pertubative QCD, where a factor of two is already respectable). In this program, the LHC has some unique advantages and is complementary to RHIC in other aspects. We expect significant differences at the higher energy in terms of energy density, lifetime and volume of the state of matter created in the collisions, and rare 'hard probes' (jets, heavy flavours, quarkonia) will be abundantly produced. In this new environment, we can test and validate the 'heavy ion standard model' (HI-SM), which describes the QGP as a strongly interacting, very opaque, almost perfect fluid. This model has emerged over the last 10 years from RHIC, and it would be a non trivial test to see how it fares at much higher energy. Once we have verified that the global event characteristics (e.g. energy density, volume, lifetime) of matter at LHC are indeed rather different, but that evolution and intrinsic properties are still well described by the standard model, we can embark on the program of precision measurements of the QGP parameters (e.g. viscosity, equation-of-state, transport coefficients, Debye screening mass,..). And along the way, we may well discover some more surprises...
\begin{figure}[!t]
\centerline{\includegraphics[width=0.9\textwidth]{Fig1.pdf}}
\caption{Top left: Charged particle pseudorapidity density per participant versus centre of mass energy for pp and AA collisions. Top right: Charged particle density per participant versus number of participants $N_{\rm part}$ for LHC and RHIC data. Bottom left: Product of the three pion HBT radii at $k_T$ = 0.3 GeV/c compared to central gold and lead collisions at lower energies. Bottom right: The decoupling time $\tau_{f}$ compared to results from lower energies.
Figures taken from refs~\cite{Aamodt:2010pb, Collaboration:2010cz, Aamodt:2011mr}.}
\label{fig1}
\end{figure}
The first, and most anticipated, results concerned global event characteristics. When ALICE was conceived in the early '90s, the charged particle density predicted for central Pb--Pb collisions at LHC was extremely uncertain, varying between 2000 and $>4000$ charge particles per unit rapidity. With RHIC results, the uncertainties came down quite a bit, as did the central value, with most predictions clustering in the range 1000 - 1700~\cite{Abreu:2007kv}.
The value finally measured at LHC, ${\rm d}N_{\rm ch}/{\rm d}\eta \approx 1600$~\cite{Aamodt:2010pb}, was right in this range, if somewhat on the high side.
From the measured multiplicity in central collisions one can derive a rough estimate of the energy density, which gives at least a factor 3 above RHIC. The corresponding increase of the initial temperature is about 30\%, even with the conservative assumption that the formation time $\tau_0$ does not decrease from RHIC to LHC.
When combined with lower energy data, the charged particle production per participant, $N_{\rm part}$, rises stronger with energy than in pp, approximately with $s^{0.15}$ (Figure~\ref{fig1} top left). Even more surprising is the fact that the centrality dependence of ${\rm d}N_{\rm ch}/{\rm d}\eta$~\cite{Collaboration:2010cz} (Figure~\ref{fig1} top right)
is practically identical to Au--Au at RHIC (at least for $N_{\rm part} >50$), despite the fact that impact parameter dependent shadowing/saturation would be expected to be much stronger at LHC (much smaller Feynman-x). And indeed, models with strong shadowing (like the latest version of Hijing) and different saturation-type calculations do describe the impact parameter dependence best.
Does the similarity between the two energies then mean that saturation saturates already at RHIC? Or is it just that the nuclear geometry reigns supreme?�
\begin{figure}[!t]
\centerline{\includegraphics[width=0.90\textwidth]{Fig2.pdf}}
\caption{Left: Elliptic flow coefficient $v2$ versus $p_T$ compared to RHIC data. Right: Nuclear modification factor $R_{AA}$ for central Pb--Pb at LHC compared to RHIC data. Figures taken from refs~\cite{Aamodt:2010pa, Aamodt:2010jd}.}
\label{fig2}
\end{figure}
The freeze-out volume and total lifetime of the created system was measured with identical particle interferometry (HBT)~\cite{Aamodt:2011mr}. Compared to top RHIC energy, the 'volume of homogeneity' ((Figure~\ref{fig1} bottom left) increases by a factor 2 (to 300 $fm^3$) and the system lifetime (Figure~\ref{fig1} bottom right) increases by about 20\% (to 10 fm/c), pretty much in line with the predictions from hydrodynamics. Also a more detailed study of the out/long/side radii as a function of pair momentum shows very good agreement with hydro models, at least with those who do describe well also the RHIC data. After years of confusion and discrepancy with data (the dark ages of the 'HBT puzzle'), interferometry is again a predictive and consistent part of the HI-SM.
The most critical test of the heavy ion standard model however comes from the measurement of the elliptic flow at LHC, the pillar which supports the 'fluid' interpretation of the QGP (perfect or otherwise). Assuming only small or no changes in the fluid properties (EoS, viscosity) between RHIC and LHC, hydro predicts firmly that the elliptic flow coefficient $v2$, measured as a function of $p_T$, should not depend on beam energy. As shown in Figure~\ref{fig2} (left), this prediction was confirmed very quickly and with good precision~\cite{Aamodt:2010pa}. The $p_T$ integrated flow values however do increase compared to RHIC by some 30\% for mid-central collisions, because the average $p_T$ is significantly higher at LHC. While $<p_T>$ increases also in pp with energy, because hard and semi-hard processes become more important, hydro predicts in addition an increase in the radial flow velocity leading to a characteristic $p_T$ and mass dependence of the spectra. It will be very interesting to see, once identified particle spectra will be available, if also this prediction is borne out (as everyone assumes it is), and if the radial flow regime extends to even higher momentum than at RHIC (here predictions are less firm).
With the heavy ion standard model having passed its first tests (HBT, $v2$) with flying colours, the program of precision measurements is now starting at LHC. For example, a major advance in quantifying the shear viscosity, which at RHIC was measured to be within a factor 2-4 of the quantum limit of a perfect fluid, will require a better estimate of remaining non-flow contributions, as well as a better constraint on the initial conditions, i.e. the geometry of the collision zone (and its fluctuations) which drive the various flow components. Progress is already being made on both fronts, in particular by looking in more detail at the centrality dependence and at higher Fourier components ($v2, v3, v4,..$).
\begin{figure}[!t]
\centerline{\includegraphics[width=0.9\textwidth]{Fig3.pdf}}
\caption{ Top left: Reconstructed $\Omega$ decays. Top right and bottom left: Reconstructed hadronic charm decays. Bottom right: Dimuon invariant mass distribution around the $\Jpsi$ mass. }
\label{fig3}
\end{figure}
The energy advantage of LHC is most evident in the area of parton energy loss and jet quenching, where the kinematic reach vastly exceeds the one available at RHIC. With high $p_T$ jets easily sticking out of the soft background, jet quenching is qualitatively evident already by visual inspection of charged jets in the TPC where a striking jet energy imbalance develops for central collisions. A quantitative analysis will however take more time, in order to see how much energy is actually lost (i.e. measure the transport coefficient), how and where it is lost (multiple soft versus few hard scatterings), and if there is a response of the medium to this local energy deposit (shock waves, mach cones). Eventually, an answer to all these questions will require measuring the medium modification of the fragmentation function, down to low ($<$ 5 GeV) or even very low ($<$ 2 GeV) transverse momentum. In the meantime, we have measured the nuclear modification factor $R_{AA}$ of inclusive charged particle momentum distributions out to 20 GeV~\cite{Aamodt:2010jd}, where the spectra are dominated by leading jet fragments. As shown in Figure~\ref{fig2} (right), the $R_{AA}$ ratio has a minimum at around 6 GeV, where the suppression is modestly stronger than at RHIC, but then rises again smoothly towards higher momentum. This latter feature is not evident in published RHIC data, and while such a rise was qualitatively predicted by some models for LHC~\cite{Abreu:2007kv}, it looks stronger at first sight. However, initial state effects (shadowing/saturation), which presumably are very strong at LHC and which should depend on both impact parameter and momentum transfer, can complicate a straight forward interpretation of the data and the comparison between different beam energies. It will be interesting to see how this result will fit into the overall picture of jet quenching.
Finally, as an appetizer of results to come, Figure~\ref{fig3} shows some first performance plots on strange particle and heavy flavour production in Pb--Pb for a fraction of the collected statistics (for details and additional results see the other contributions of ALICE to this conference): the $\Omega$ (top left) and fully reconstructed hadronic charm decays in the central barrel region (top right, bottom left) and the $\Jpsi$ decaying into muons in the forward muon arm (bottom right). However, as in the case of $R_{AA}$, charm and quarkonia production and their distributions may be strongly modified by initial state effects which could mask the medium modification which is the prime interest. Therefore, like at RHIC, a pA comparison run at LHC may be required sooner rather than later.
\section{Summary}
After two decades of design, R\&D, construction, installation, commissioning and simulations, the ALICE experiment has 'hit the ground running' since LHC started its operation at the end of 2009. Most systems are fast approaching design performance, and physics analysis is well underway. With the first heavy ion collisions a mere four weeks ago, a new era has started for ultra-relativistic heavy ion physics and we can look forward to at least a decade of exciting (and hopefully revealing) new results.
|
\section{INTRODUCTION}
Highly energetic explosions in self-gravitating fluid distributions are common events in relativistic astrophysics (see for example \cite{1N,2N} and references therein). There are two main questions related to this issue:
\begin{itemize}
\item What is the energy source of such explosions?
\item How the system evolves after the explosion?
\end{itemize}
This work is devoted to the second of the above questions. Our approach is completely analytical, which presents both advantages and disadvantages. On the one hand the analytical approach allows one to carry out a rather model independent analysis, allowing in turn to obtain general conclusions about the pattern evolution of the system. On the other hand however, in order to handle analytically the involved expressions, one needs to introduce simplifying assumptions which might exclude important physical phenomena.
We are interested in analytical models which even if might be relatively simple to analyze, still contain some of the essential features of a realistic situation. It should be emphasized that we are not interested in the dynamics and the conditions of the creation of the cavity itself, but only in its evolution once it is already formed.
A similar study to the one presented here was recently published \cite{3N}, however the main difference between the two being the kinematical condition imposed on the evolution of the system. In \cite{3N} it was assumed that the proper radial distance between any two infinitesimally close fluid elements remains constant all along the evolution (the purely areal evolution condition). Here we assume that the expansion scalar vanishes for all fluid elements. As discussed in \cite {4N,5N,3N} either assumption creates conditions for the existence of a vacuum cavity surrounding the centre. The dynamical instability of expansion-free fluids is studied in \cite{Herrera}.
We have two hypersurfaces delimiting the fluid. The external one separating the fluid distribution from a Schwarzschild spacetime (for simplicity we assume the evolution to be adiabatic) and the internal one, delimiting the cavity within which we have Minkowski spacetime. It should be clear that for cavities with sizes of the order of 20 Mpc or smaller, the assumption of a spherically symmetric spacetime outside the cavity is quite reasonable, since the observed universe cannot be considered homogeneous on scales less than 150-300 Mpc. Nevertheless for larger cavities it should be more appropriate to consider their embedding in an expanding Lema\^{\i}tre-Friedmann-Robertson-Walker spacetime (for the specific case of void modeling in expanding universes see \cite{6N,7N} and references therein).
Thus, we have to consider junction conditions on both hypersurfaces. Depending on whether we impose Darmois conditions \cite{8N} or allow for the existence of thin shells \cite{9N}, different kinds of models are obtained. In this paper we are particularly interested in models satisfying Darmois conditions, however some models presenting thin shells will be also exhibited.
The first known solution satisfying the vanishing expansion scalar is due to Skripkin
\cite{13N,Mac}. It is worth noticing that Skripkin did not assume explicitly the expansion scalar to vanish, however the conditions he imposed (isotropic fluid with a constant energy-density) imply the vanishing of the expansion scalar. Unfortunately though, as it was shown in \cite{5N}, the Skripkin model does not satisfy Darmois junction conditions (in principle Skripkin models with thin shells satisfying Israel conditions can be obtained).
It is our purpose here to find exact solutions under the vanishing expansion scalar condition and satisfying Darmois junction conditions on both delimiting hypersurfaces.
For sake of generality we shall consider an anisotropic fluid. Indeed, as it has been shown in \cite{4N}, in the case of isotropic pressure, the expansion-free conditions imposes that the energy-density is independent on the time-like coordinate, which severely restricts the models (further arguments to justify such kind of fluid distributions may found in \cite{10N,11N,12N} and references therein). A detailed description of this kind of distribution, the Darmois junction conditions on both, the inner and the outer boundary surface, as well as definitions of kinematical variables, and basic equations, are given in section II. The general description of the expansion-free fluids is given in section III, whereas explicit solutions are found and described in sections IV and V. Finally results are summarized in the last section.
\section{FLUID DISTRIBUTION, KINEMATICAL VARIABLES AND BASIC EQUATIONS}
We consider a spherically symmetric distribution of collapsing
fluid, bounded by a spherical surface $\Sigma^{(e)}$. The fluid is
assumed to be locally anisotropic (principal stresses unequal).
Choosing comoving coordinates inside $\Sigma^{(e)}$, the general
interior metric can be written
\begin{equation}
ds^2_-=-A^2dt^2+B^2dr^2+R^2(d\theta^2+\sin^2\theta d\phi^2),
\label{1}
\end{equation}
where $A$, $B$ and $R$ are functions of $t$ and $r$ and are assumed
positive. We number the coordinates $x^0=t$, $x^1=r$, $x^2=\theta$
and $x^3=\phi$. Observe that $A$ and $B$ are dimensionless, whereas $R$ has the same dimension as $r$.
The matter energy-momentum $T_{\alpha\beta}^-$ inside $\Sigma^{(e)}$
has the form
\begin{equation}
T_{\alpha\beta}^-=(\mu +
P_{\perp})V_{\alpha}V_{\beta}+P_{\perp}g_{\alpha\beta}+\Pi \chi_{
\alpha}\chi_{\beta}, \label{3}
\end{equation}
where $\mu$ is the energy density, $\Pi\equiv P_r-P_\perp$, $P_r$ the radial pressure,
$P_{\perp}$ the tangential pressure, $V^{\alpha}$ the four velocity of the fluid,
and $\chi^{\alpha}$ a unit four vector along the radial direction. These quantities
satisfy
\begin{equation}
V^{\alpha}V_{\alpha}=-1, \;\; \chi^{\alpha}\chi_{\alpha}=1, \;\chi^{\alpha}V_{\alpha}=0,
\end{equation}
since we assumed the metric (\ref{1}) comoving then
\begin{equation}
V^{\alpha}=A^{-1}\delta_0^{\alpha}, \;\;
\chi^{\alpha}=B^{-1}\delta^{\alpha}_1.\label{5}
\end{equation}
\subsection{Einstein equations}
From (\ref{1}) and (\ref{3}), Einstein equations
\begin{equation}
G_{\alpha \beta} = 8 \pi T_{\alpha \beta}, \label{Eeq}
\end{equation}
read
\begin{eqnarray}
8\pi T_{00}=8\pi \mu A^2
=\left(2\frac{\dot{B}}{B}+\frac{\dot{R}}{R}\right)\frac{\dot{R}}{R}\nonumber\\
-\left(\frac{A}{B}\right)^2\left[2\frac{R^{\prime\prime}}{R}+\left(\frac{R^{\prime}}{R}\right)^2
-2\frac{B^{\prime}}{B}\frac{R^{\prime}}{R}-\left(\frac{B}{R}\right)^2\right],
\label{12} \\
8\pi T_{01}=0 =-2\left(\frac{{\dot R}^{\prime}}{R} -\frac{\dot
B}{B}\frac{R^{\prime}}{R}-\frac{\dot
R}{R}\frac{A^{\prime}}{A}\right),
\label{13} \\
8\pi T_{11}=8\pi P_rB^2 \nonumber\\
=-\left(\frac{B}{A}\right)^2\left[2\frac{\ddot{R}}{R}
-\left(2\frac{\dot A}{A}-\frac{\dot{R}}{R}\right)\frac{\dot R}{R}\right]\nonumber\\
+\left(2\frac{A^{\prime}}{A}+\frac{R^{\prime}}{R}\right)\frac{R^{\prime}}{R}-\left(\frac{B}{R}\right)^2,
\label{14} \\
8\pi T_{22}=\frac{8\pi}{\sin^2\theta}T_{33}=8\pi P_{\perp}R^2\nonumber \\
=-\left(\frac{R}{A}\right)^2\left[\frac{\ddot{B}}{B}+\frac{\ddot{R}}{R}
-\frac{\dot{A}}{A}\left(\frac{\dot{B}}{B}+\frac{\dot{R}}{R}\right)
+\frac{\dot{B}}{B}\frac{\dot{R}}{R}\right]\nonumber\\
+\left(\frac{R}{B}\right)^2\left[\frac{A^{\prime\prime}}{A}
+\frac{R^{\prime\prime}}{R}-\frac{A^{\prime}}{A}\frac{B^{\prime}}{B}
+\left(\frac{A^{\prime}}{A}-\frac{B^{\prime}}{B}\right)\frac{R^{\prime}}{R}\right],\label{15}
\end{eqnarray}
where the prime stands for $r$
differentiation and the dot stands for differentiation with respect to $t$.
\subsection{Kinematical variables and the mass function}
The four acceleration $a_{\alpha}$ is given by
\begin{equation}
a_{\alpha}=V_{\alpha ;\beta}V^{\beta}, \label{4b}
\end{equation}
from (\ref{4b}) and (\ref{5}) we obtain
\begin{equation}
a_1=\frac{A^{\prime}}{A}, \;\;
a^2=a^{\alpha}a_{\alpha}=\left(\frac{A^{\prime}}{AB}\right)^2
\label{5c}
\end{equation}
with $a^\alpha= a \chi^\alpha$. The expansion $\Theta$ is given by
\begin{equation}
\Theta={V^{\alpha}}_{;\alpha}=\frac{1}{A}\left(\frac{\dot{B}}{B}+2\frac{\dot{R}}{R}\right).\label{th}
\end{equation}
and for the shear we have
\begin{equation}
\sigma_{\alpha\beta}=V_{(\alpha
;\beta)}+a_{(\alpha}V_{\beta)}-\frac{1}{3}\Theta h_{\alpha\beta}.
\label{4a}
\end{equation}
where $h_{\alpha \beta} = g_{\alpha \beta} + V_\alpha V_\beta $.
Using (\ref{5}) we obtain the non-vanishing components of
(\ref{4a})
\begin{equation}
\sigma_{11}=\frac{2}{3}B^2\sigma, \;\;
\sigma_{22}=\frac{\sigma_{33}}{\sin^2\theta}=-\frac{1}{3}R^2\sigma,
\label{5a}
\end{equation}
with
\begin{equation}
\sigma^{\alpha\beta}\sigma_{\alpha\beta}=\frac{2}{3}\sigma^2,
\label{5b}
\end{equation}
being
\begin{equation}
\sigma=\frac{1}{A}\left(\frac{\dot{B}}{B}-\frac{\dot{R}}{R}\right).\label{5b1}
\end{equation}
$\sigma_{\alpha\beta}$ may also be written as
\begin{equation}
\sigma_{\alpha \beta}= \sigma \left(\chi_\alpha \chi_\beta -
\frac{1}{3} h_{\alpha \beta}\right). \label{sh}
\end{equation}
Next, the mass function $m(t,r)$ introduced by Misner and Sharp \cite{14N} (see also \cite{15N}) is
given by
\begin{equation}
m=\frac{R^3}{2}{R_{23}}^{23} =\frac{R}{2}
\left[\left(\frac{\dot{R}}{A}\right)^2
-\left(\frac{R^{\prime}}{B}\right)^2+1\right].
\label{18}
\end{equation}
We can define the velocity $U$ of the collapsing
fluid as the variation of the areal radius $R$ with respect to proper time, i.e.
\begin{equation}
U=\frac{\dot R}{A}. \label{19}
\end{equation}
Then (\ref{18}) can be rewritten as
\begin{equation}
E \equiv \frac{R^{\prime}}{B}=\left[1+U^2-2\frac{m(t,r)}{R}\right]^{1/2}.
\label{20}
\end{equation}
With the above we can express (\ref{13}) as
\begin{equation}
0=\frac{1}{3 R^{\prime}}(\Theta-\sigma)^{\prime}
-\frac{\sigma}{R}.\label{21a}
\end{equation}
While from (\ref{18}) we have
\begin{eqnarray}
\dot m=-4\pi P_r \dot RR^2,
\label{22}
\end{eqnarray}
and
\begin{eqnarray}
m^{\prime}=4\pi\mu R^{\prime}R^2.
\label{27}
\end{eqnarray}
Combining (\ref{22}) and (\ref{27}) we obtain
\begin{equation}
\dot \mu R^\prime +P^\prime _r \dot R+(P_r+\mu)\left(\dot R^\prime
+2R^\prime \frac{\dot R}{R}\right)=0.\label{j}
\end{equation}
The integration of (\ref{27}) produces
\begin{equation}
m=\int^{r}_{0}4\pi\mu R^2R^{\prime}dr\label{27int}
\end{equation}
where a regular centre to the distribution, $m(0)=0$ is assumed. We may partially integrate (\ref{27int}) to obtain
\begin{equation}
3\frac{m}{R^3} = 4\pi \mu - \frac{4\pi}{R^3} \int^r_0\mu^{\prime}R^3 dr.
\label{3m/R3}
\end{equation}
\subsection{Junction conditions}
Outside $\Sigma^{(e)}$ (whose equation is $r=r_e=$ constant) we assume the Schwarzschild
spacetime, described by
\begin{equation}
ds^2=-\left(1-2\frac{M}{\rho}\right)dv^2-2d\rho dv+\rho^2(d\theta^2
+\sin^2\theta
d\phi^2) \label{1int},
\end{equation}
where $M$ denotes the total mass and $v$ is the retarded time.
The matching of the adiabatic fluid sphere to
Schwarzschild
spacetime, on the surface $r=r_e=$ constant (or $\rho=\rho(v)_e$ in the coordinates of (\ref{1int})), requires the continuity of the first and second differential forms (Darmois conditions), implying (see \cite{16N} for details)
\begin{eqnarray}
Adt\stackrel{\Sigma^{(e)}}{=}dv \left(1-2\frac{M}{\rho}\right), \label{junction1fn}\\
R\stackrel{\Sigma^{(e)}}{=}\rho(v), \label{junction1f2n}\\
m(t,r)\stackrel{\Sigma^{(e)}}{=}M, \label{junction1}
\end{eqnarray}
and
\begin{widetext}
\begin{eqnarray}
2\left(\frac{{\dot R}^{\prime}}{R}-\frac{\dot B}{B}\frac{R^{\prime}}{R}-\frac{\dot R}{R}\frac{A^{\prime}}{A}\right)
\stackrel{\Sigma^{(e)}}{=}-\frac{B}{A}\left[2\frac{\ddot R}{R}
-\left(2\frac{\dot A}{A}
-\frac{\dot R}{R}\right)\frac{\dot R}{R}\right]+\frac{A}{B}\left[\left(2\frac{A^{\prime}}{A}
+\frac{R^{\prime}}{R}\right)\frac{R^{\prime}}{R}-\left(\frac{B}{R}\right)^2\right],
\label{j2}
\end{eqnarray}
\end{widetext}
where $\stackrel{\Sigma^{(e)}}{=}$ means that both sides of the equation
are evaluated on $\Sigma^{(e)}$ (observe a misprint in equation (40) in \cite{16N} and a slight difference in notation).
Comparing (\ref{j2}) with (\ref{13}) and (\ref{14}) one obtains
\begin{equation}
P_r\stackrel{\Sigma^{(e)}}{=}0.\label{j3}
\end{equation}
As we mentioned in the introduction, the expansion-free models present an internal vacuum cavity. If we call $\Sigma^{(i)} $ the boundary surface between the cavity and the fluid, described by the equation $r=r_i=$ constant, then the matching of the Minkowski spacetime within the cavity to the fluid distribution, implies
\begin{eqnarray}
m(t,r)\stackrel{\Sigma^{(i)}}{=}0, \label{junction1i}\\
P_r\stackrel{\Sigma^{(i)}}{=}0.\label{j3i}
\end{eqnarray}
\subsection{Weyl tensor}
The Weyl tensor is defined through the Riemann tensor
$R^{\rho}_{\alpha \beta \mu}$, the Ricci tensor
$R_{\alpha\beta}$ and the curvature scalar $\cal R$, as
\begin{eqnarray}
C^{\rho}_{\alpha \beta \mu}=R^\rho_{\alpha \beta \mu}-\frac{1}{2}
R^\rho_{\beta}g_{\alpha \mu}+\frac{1}{2}R_{\alpha \beta}\delta
^\rho_{\mu}-\frac{1}{2}R_{\alpha \mu}\delta^\rho_\beta, \nonumber\\
+\frac{1}{2}R^\rho_\mu g_{\alpha \beta}+\frac{1}{6}{\cal
R}(\delta^\rho_\beta g_{\alpha \mu}-g_{\alpha
\beta}\delta^\rho_\mu). \label{34}
\end{eqnarray}
The electric part of Weyl tensor is defined by (the magnetic part vanishes due to the spherical symmetry)
\begin{equation}
E_{\alpha \beta} = C_{\alpha \mu \beta \nu} V^\mu V^\nu,
\label{elec}
\end{equation}
with the following non-vanishing components
\begin{eqnarray}
E_{11}=\frac{2}{3}B^2 {\cal E},\nonumber \\
E_{22}=-\frac{1}{3} R^2 {\cal E}, \nonumber \\
E_{33}= E_{22} \sin^2{\theta}, \label{ecomp}
\end{eqnarray}
where
\begin{eqnarray}
&&{\cal E}= \frac{1}{2 A^2}\left[\frac{\ddot R}{R} - \frac{\ddot B}{B} - \left(\frac{\dot R}{R} - \frac{\dot B}{B}\right)\left(\frac{\dot A}{A} + \frac{\dot R}{R}\right)\right]\\
&+& \frac{1}{2 B^2} \left[\frac{A^{\prime\prime}}{A} - \frac{R^{\prime\prime}}{R} + \left(\frac{B^{\prime}}{B} + \frac{R^{\prime}}{R}\right)\left(\frac{R^{\prime}}{R}-\frac{A^{\prime}}{A}\right)\right] - \frac{1}{2 R^2}.\nonumber
\label{E}
\end{eqnarray}
Observe that we may also write $E_{\alpha\beta}$ as
\begin{equation}
E_{\alpha \beta}={\cal E}\left(\chi_\alpha
\chi_\beta-\frac{1}{3}h_{\alpha \beta}\right). \label{52}
\end{equation}
Using (\ref{12}), (\ref{14}), (\ref{15}) with (\ref{18}) and (\ref{E}) we obtain
\begin{equation}
3\frac{m}{R^3}=4\pi (\mu-\Pi) - \cal{E}.
\label{mE}
\end{equation}
Finally, comparing (\ref{3m/R3}) with (\ref{mE}) we get
\begin{equation}
{\cal E} = \frac{4\pi}{R^3} \int^r_0\mu^{\prime}R^3 dr-4\pi \Pi,
\label{Emat}
\end{equation}
which exhibits the dependence of the Weyl tensor on energy-density inhomogeneity and local anisotropy of pressure.
\subsection{ Bianchi identities and an evolution equation for the Weyl tensor}
The two independent components of Bianchi identities for the system under consideration read (see \cite{4N} for details):
\begin{equation}
\dot{\mu}+( \mu + P_r)A\Theta-2\Pi \frac{\dot
R}{R}=0,\label{a}
\end{equation}
\begin{equation}
P_r ^\prime+(\mu +
P_r)\frac{A^\prime}{A}+2\Pi\frac{R^\prime}{R}=0.\label{b}
\end{equation}
Finally, the following evolution equation for the Weyl tensor will be used, which may be derived from the Bianchi identities \cite{17N,18N} (see also \cite{11N} for details)
\begin{equation}
\left[{\cal E}-4\pi( \mu-\Pi) \right]^{\dot{}}+3\left[{\cal E}-4\pi(
\mu+P_{\perp})\right]\frac{\dot R}{R}=0. \label{g}
\end{equation}
\section{THE EXPANSION-FREE FLUID}
If we assume the evolution to be expansion-free, i.e. $\Theta=0$,
then from (\ref{th}) we have
\begin{equation}
\frac{\dot B}{B}=-2\frac{\dot R}{R}\,\, \Rightarrow \,\,
B=\frac{h}{R^2},\label{th1}
\end{equation}
and after substituting (\ref{th1}) into (\ref{13}), we get
\begin{equation}
A=\frac{\dot R R^2}{\tau},\label{a1}
\end{equation}
where $\tau (t)$ and $h(r)$ are arbitrary functions of their arguments. Thus, without loss of generality, we shall take $h=1$ (observe that with this choice a unit constant with dimensions $[r^4]$ is assumed to multiply $dr^2$).
Next, using the equations (\ref{22}), (\ref{27}), (\ref{a}) and
(\ref{th1}), the physical variables $\mu$, $P_r$, and $\Pi$, can
be written through the mass function $m$ and the metric function $R$
as follows,
\begin{eqnarray}
4\pi \mu=\frac{m^\prime}{R^\prime R^2}, \label{mug}\\
4\pi P_r=-\frac{\dot m}{\dot R R^2},\label{pg}\\
\Pi=\frac{\dot \mu R}{2 \dot R}. \label{ppg}
\end{eqnarray}
On the other hand, combining (\ref{18}), (\ref{th1}) and
(\ref{a1}) we obtain,
\begin{equation}
m(t,r)=\frac{1}{2}\left(\frac{\tau ^2}{R^3}-R^{\prime 2}R^5+R\right). \label{mg}
\end{equation}
To completely determine the metric we need to give an extra
condition. In the next sections we provide some options.
\section{SOLUTION I}
The first family of solutions will be obtained by choosing the function $m(t,r)$ as follows \cite{19N},
\begin{equation}
2m(t,r)=jR+\frac{1}{3}kR^3+\frac{1}{5}lR^5, \label{mM}
\end{equation}
where $j(t)$, $k(t)$ and $l(t)$ are so far arbitrary functions of $t$.
Substituting (\ref{mM}) into (\ref{mug}) we
obtain for the energy density
\begin{equation}
8\pi \mu =\frac{j}{R^2}+k+lR^2. \label{muM}
\end{equation}
\noindent Then, taking the $r$ derivative of (\ref{mg}) and using
(\ref{mug}) and (\ref{muM}) we obtain,
\begin{equation}
j-1+kR^2+lR^4+3\frac{\tau^2}{R^4}+2 R^{\prime\prime}R^5+5R^{\prime 2}R^4=0.\label{eq0}
\end{equation}
Combining (\ref{mg}), (\ref{mM}) and
(\ref{eq0}), we get
\begin{equation}
4(j-1)+2kR^2+\frac{8}{5}lR^4+2R^{\prime\prime}R^5+8R^{\prime 2}R^4 =0,\label{eq1}
\end{equation}
which by introducing the new function $Z\equiv R^2$, can be transformed into
\begin{equation}
a+bZ+cZ^2+Z^{\prime\prime}Z^2+\frac{3}{2}Z^{\prime 2}Z=0,\label{eq2}
\end{equation}
where
\begin{equation}
a(t)\equiv 4(j-1),\quad b(t)\equiv 2k, \quad c(t)\equiv \frac{8}{5}l. \label{const}
\end{equation}
Finally, feeding back $Y(Z)=Z^{\prime 2}$ into (\ref{eq2}),
leads to the first order equation
\begin{equation}
\frac{dY}{dZ}+3\frac{Y}{Z}=-2\left(\frac{a}{Z^2}+\frac{b}{Z}+c\right),\label{eq3}
\end{equation}
and after integration we obtain
\begin{equation}
Y(Z)=Z^{\prime 2}
=-\left(\frac{a}{Z}+\frac{2}{3}b+\frac{1}{2}cZ\right).\label{eq4}
\end{equation}
This last equation can be integrated in its general
form, however in order to obtain simple models, in what follows we shall consider specific cases.
\subsection{a=0}
For $a=0$ ($j=1$), integration of
(\ref{eq4})produces,
\begin{equation}
Z=-\frac{4}{3}\frac{b}{c}-\frac{1}{8}c(r+\beta)^2,
\end{equation}
implying
\begin{equation}
R=\left[-\frac{4}{3}\frac{b}{c}-\frac{1}{8}c(r+\beta)^2\right]^{1/2}=\left[-\frac{5k}{3l}-\frac{l}{5}(r+\beta)^2\right]^{1/2},
\label{eq5}
\end{equation}
where $\beta (t)$ is an arbitrary function of t.
Then, using (\ref{muM}), (\ref{pg}) and
(\ref{ppg}) the physical variables, $\mu$, $P_r$ and $\Pi$, can be
expressed as
\begin{equation}
8\pi
\mu=\frac{1}{R^2}+k+lR^2,\label{muM1}
\end{equation}
\begin{equation}
8\pi P_r=-\frac{1+kR^2+lR^4}{R^2}-\frac{R}{\dot R}\left(\frac{\dot k}{3}+\frac{\dot l}{5}R^2\right),\label{pM1}\end{equation}
\begin{equation}
8\pi
\Pi=-\frac{1}{R^2}+R^2\left(l+\frac{\dot k+\dot l R^2}{2R\dot R}\right).\label{ppM1}
\end{equation}
This subfamily of solutions may satisfy Darmois conditions. Indeed junctions conditions (\ref{junction1}), (\ref{j3}) and (\ref{junction1i}), (\ref{j3i}) lead to two independent equations for the three functions $b(t)$, $c(t)$ and $\beta(t)$ which can be satisfied by any convenient choice of one of these functions. However for all the attempts we have carried out resulting expressions were extremely cumbersome and therefore we shall not pursue this direction further.
\subsection{b=c=0}
For $b(t)=c(t)=0$, integration of (\ref{eq4}) produces
\begin{equation}
Z=\left(\frac{3}{2}\sqrt{-a}r+\beta\right)^{2/3},
\end{equation}
or
\begin{equation}
R=\left(\frac{3}{2}\sqrt{-a}r+\beta\right)^{1/3}.\label{eq6}
\end{equation}
Then, using (\ref{muM}), (\ref{pg}) and
(\ref{ppg}) the physical variables $\mu$, $P_r$ and $\Pi$ read,
\begin{eqnarray}
8\pi\mu
=j\left(\frac{3}{2}\sqrt{-a}r+\beta\right)^{-2/3},\label{muM2}\\
8\pi P_r=2{\dot j}\left(\frac{3}{2}\sqrt{-a}r
+\beta\right)^{1/3}\left(\frac{1}{2}\frac{\dot a}{\sqrt{-a}}r
-\frac{2}{3}{\dot\beta}\right)^{-1} \nonumber\\
-j\left(\frac{3}{2}\sqrt{-a}r+\beta\right)^{-2/3},\label{pM2}\\
8\pi \Pi=-{\dot j}\left(\frac{3}{2}\sqrt{-a}r+\beta\right)^{1/3}
\left(\frac{1}{2}\frac{\dot a}{\sqrt{-a}}r-\frac{2}{3}{\dot\beta}\right)^{-1} \nonumber\\
-j\left(\frac{3}{2}\sqrt{-a}r+\beta\right)^{-2/3}.\label{ppM2}
\end{eqnarray}
It is a simple matter to check that this subfamily of solution does not satisfy Darmois conditions on the internal boundary surface $r=r_i$. Accordingly a thin shell should be present there.
\section{SOLUTION II}
The next family of solutions is found with the additional assumption $P_r=0$. Fluid spheres with tangential stresses alone, have a long and a venerable history starting with Lema\^{\i}tre seminal paper \cite{20N}. Afterwards, many authors have used such kind of fluid distribution for different purposes (see for example \cite{21N, 22N, 23N, 24N, 25N, 26N, 27N} and references therein).
Thus, assuming $P_r=0$, we find from (\ref{j})
\begin{equation}
\mu=\frac{C_1}{R^\prime R^2}\,\, \Rightarrow\,\, R^3=3\int
\frac{C_1}{\mu}dr+C_2,\label{01}
\end{equation}
where $C_1(r)$ and $C_2(t)$ are two functions of
integration. Using (\ref{mug}) in (\ref{01}), the former can be easily identified as
\begin{equation}
C_1=\frac{m'}{4\pi}.
\label{c1}
\end{equation}
Now, feeding back (\ref{th1}) in (\ref{a}) we obtain
for the tangential pressure $P_\perp$
\begin{equation}
P_\perp=-\frac{\dot \mu}{2}\frac{R}{\dot R}.\label{pmo}
\end{equation}
In what follows we shall impose different additional restrictions in order to obtain different subfamilies of models.
\subsection{$P_{\perp}=\alpha \mu$}
Assuming
\begin{equation}
P_r=0, \;\; P_{\perp}=\alpha\mu, \label{N3}
\end{equation}
where $\alpha$ is a constant, then (\ref{b}) becomes
\begin{equation}
\frac{{\dot R}^{\prime}}{\dot R}=2(\alpha -1)\frac{R^{\prime}}{R}, \label{N4}
\end{equation}
and after integration
\begin{equation}
{\dot R}=fR^{2(\alpha-1)}, \;\; R^{\prime}=gR^{2(\alpha-1)}. \label{N5}
\end{equation}
with $f(t)$ and $g(r)$ denoting arbitrary functions of their arguments. Then from (\ref{N5}) we obtain,
\begin{equation}
R^{3-2\alpha}=\psi(t)+\chi(r), \label{N6}
\end{equation}
with
\begin{equation}
\psi(t)=(3-2\alpha)\int fdt, \;\; \chi(r)=(3-2\alpha)\int gdr. \label{7}
\end{equation}
Without loss of generality we may choose $\tau(t)=f$, implying because of (\ref{a1}) and (\ref{N5}),
\begin{equation}
A=R^{2\alpha}. \label{N8}
\end{equation}
Considering (\ref{N3}) with (\ref{22}) we have
\begin{equation}
m=m(r), \label{N9}
\end{equation}
and from (\ref{18}) with (\ref{th1}) and (\ref{N8}) we obtain
\begin{equation}
m(r)=\frac{R}{2}\left(\frac{{\dot R}^2}{R^{4\alpha}}-g^2R^{4\alpha}+1\right). \label{N10}
\end{equation}
Matching conditions (\ref{junction1}) and (\ref{junction1i}) on the hypersurfaces $r=r_e$ and $r=r_i$, together with (\ref{N10}), produce
\begin{equation}
{\dot R}^2\stackrel{\Sigma^{(i)}}{=}R^{4\alpha}(g^2R^{4\alpha}-1), \label{N12}
\end{equation}
and
\begin{equation}
{\dot R}^2\stackrel{\Sigma^{(e)}}{=}R^{4\alpha}\left(\frac{2M}{R}+g^2R^{4\alpha}-1\right). \label{N12b}
\end{equation}
Equation (\ref{N12}) can be integrated for any given value of $\alpha$ (the particular case of dust, $\alpha=0$ has already been considered in \cite{5N}).
For $\alpha=1/2$ the solution reads
\begin{equation}
R\stackrel{\Sigma^{(i)}}{=}\frac{1}{g\cos(t+t_0)}, \label{N15}
\end{equation}
and from (\ref{N6}) evaluated on $\Sigma^{(i)}$,
\begin{equation}
\psi(t)\stackrel{\Sigma^{(i)}}{=}\frac{1}{g^2\cos^2(t+t_0)}+\chi. \label{N16}
\end{equation}
Thus the time dependence of all variables is fully determined. The radial dependence ($C_1$ or $\chi$) can be obtained from the initial data.
However, this solution is inconsistent for any $\alpha\geq 0$ with junction conditions satisfied on both hypersurfaces.
Indeed, from (\ref{N12}) and (\ref{N12b}) evaluated at $t=0$ we get
\begin{equation}
R_i^{4\alpha}(0)=\frac{1}{2g^2}\left[1+\left(1+4g^2\dot R_i^2(0)\right)^{1/2}\right],
\label{n1}
\end{equation}
and
\begin{eqnarray}
R_e^{4\alpha}(0)=\frac{1}{2g^2} \nonumber\\
\times\left\{1-\frac{2M}{R_e(0)}
+\left[\left(1-\frac{2M}{R_e(0)}\right)^2+4g^2\dot R_e^2(0)\right]^{1/2}\right\}.
\label{n2}
\end{eqnarray}
Now substracting (\ref{n2}) from (\ref{n1}) we obtain
\begin{widetext}
\begin{equation}
R_i^{4\alpha}(0)-R_e^{4\alpha}(0)=\frac{1}{2g^2}
\left[1+\left(1+4g^2{\dot R}_i^2(0)\right)^{1/2}\right] -\frac{1}{2g^2}\left\{1-\frac{2M}{R_e(0)}+\left[\left(1-\frac{2M}{R_e(0)}\right)^2
+4g^2{\dot R}_e^2(0)\right]^{1/2}\right\},
\label{n3}
\end{equation}
\end{widetext}
which should be negative for all $t$ (if $\alpha>0$) (or zero if $\alpha=0$) since obviously $R_e(t)>R_i(t)$. Unfortunately this is not the case.
In fact, from the definition of $U$, using (\ref{a1}) and (\ref{N8}) we obtain
\begin{equation}
U_i=\frac{f}{R_i^2}=\frac{\dot R_i}{R_i^{2\alpha}},
\label{n4}
\end{equation}
and
\begin{equation}
U_e=\frac{f}{R_e^2}=\frac{\dot R_e}{R_e^{2\alpha}}.
\label{n5}
\end{equation}
Since $\alpha<1$ and $R_e>R_i$ it follows from the above that
\begin{equation}
\dot R_i>\dot R_e.
\label{n6}
\end{equation}
Then using (\ref{n6}) in (\ref{n3}) we obtain at once that $R_i^{4\alpha}(0)>R_e^{4\alpha}(0)$ which is unacceptable.
Therefore such models require the presence of thin shells at either (or both) boundary surfaces.
\subsection{$\mu =\mu _0 C_1/r^2$}
Next, the following subfamily of solutions will be obtained under the assumption that the energy-density is separable. Thus we write
\begin{equation}
\mu = \frac{\mu _0 C_1}{r^2},\label{mumo}
\end{equation}
where $\mu_0(t)$ is an arbitrary function of its argument and $C_1(r)$ is given by (\ref{c1}).
Then substituting (\ref{mumo}) into (\ref{01}) and
(\ref{pmo}), we get
\begin{eqnarray}
R= \left(\frac{r^3}{\mu _0}+C_2\right)^{1/3}
\label{1tex}
\end{eqnarray}
\begin{eqnarray}
P_\perp=\frac{3}{2}\frac{\mu _0 \dot \mu _0 C_1}{r^2}\left(\frac{r^3+\mu_0C_2}
{{\dot\mu}_0 r^3-\mu _0^2 {\dot C}_2}\right),\label{1texbis}
\end{eqnarray}
and from (\ref{mg}) and (\ref{1tex}) we obtain
\begin{equation}
m=\frac{1}{2}\left[\frac{\tau^2}{R^3}-R\left(\frac{r^4}{\mu_0^2}-1\right)\right].
\label{op21}
\end{equation}
Evaluating this expression on $\Sigma^{(i)}$ by using (\ref{junction1i}) we obtain
\begin{equation}
R^4_i=\tau^2\left(\frac{r^4_i}{\mu_0^2}-1\right)^{-1},
\label{0p22}
\end{equation}
implying the restriction $r^4_i> \mu_0^2$ for all $t$.
Next, from the definition of $U$ and (\ref{a1}) we get
\begin{equation}
U=\frac{\tau}{R^2},
\label{0p23}
\end{equation}
implying
\begin{equation}
U^2_i=\frac{r^4_i}{\mu_0^2}-1.
\label{0p24}
\end{equation}
Thus, from the absence of superluminal velocities ($U<1$) and (\ref{0p24}) we need to impose
\begin{equation}
\sqrt{2}\mu_0>r^2_i>\mu_0.
\label{0p25}
\end{equation}
Hence from (\ref{0p23}) we have
\begin{equation}
U_e=U_i\left(\frac{R_i}{R_e}\right)^2,
\label{0p26}
\end{equation}
and since $R_e>R_i$ then from (\ref{0p26}) it follows that the inner boundary surface moves faster than the outer one.
Now, since $\tau$ is a function of $t$ which may be chosen arbitrarily, we put without loss of generality
\begin{equation}
\tau=R^2_i\dot R_i,\label{0p27}
\end{equation}
implying
\begin{equation}
A_i=1, \;\; \dot R_i=U_i, \;\; \tau=R^2_iU_i=U_e R^2_e.
\label{0p28}
\end{equation}
Let us now consider the junction conditions at the outer boundary surface $\Sigma^{(e)}$. Using (\ref{junction1}), (\ref{op21}) and (\ref{0p28}) we obtain
\begin{equation}
R_e=2M\left(U^2_e-\frac{r^4_e}{\mu_0}+1\right)^{-1},\label{0p29}
\end{equation}
imposing the restriction on $r_e$,
\begin{equation}
\sqrt{2} \mu_0>r^2_e>\mu_0,\label{0p30}
\end{equation}
which is the same as imposed on $r_i$.
Thus if we provide the time dependence of $\mu_0$ then (\ref{0p22}) allows to determine the time dependence of $R_i$, and using this result in (\ref{1tex}) we obtain $C_2(t)$. Finally evaluating (\ref{1tex}) on $\Sigma^{(e)}$ we get $R_e$.
Alternatively we may provide the time dependence of, either $R_i$ or $R_e$ which allows also to find the time dependence of all variables. We shall next exhibit a specific model obtained from the additional restriction of conformal flatness.
\subsubsection{Conformally flat models}
Instead of assuming any specific time dependence for any of the variables considered above, we shall assume here that the fluid is conformally flat, i.e. we shall assume ${\cal E}=0$ then we obtain from (\ref{mE}) (with $P_r=0$)
\begin{equation}
\mu+P_\perp=\frac{3m}{4 \pi R^3},\label{0p31}
\end{equation}
Next, evaluating (\ref{0p31}) at $\Sigma^{(i)}$, using (\ref{pmo}) and (\ref{mumo}), we obtain
\begin{equation}
R_i^2=\alpha \mu_0,\label{0p33}
\end{equation}
where $\alpha$ is a constant.
Feeding back (\ref{0p33}) into (\ref{0p24}) we obtain
\begin{equation}
\dot R_i^2=\left(\frac{\beta}{R_i^2}\right)^2-1, \label{0p34}
\end{equation}
with $\beta^2\equiv r^4_i \alpha^2$.
Thus
\begin{equation}
\dot R_i=\pm \left[\left(\frac{\beta}{R_i^2}\right)^2-1\right]^{1/2}, \label{0p35}
\end{equation}
where $+$ ($-$) corresponds to expanding (contracting) configurations.
The integral of (\ref{0p35}) is expressed through elliptic functions, and we shall not display explicitly the solution.
The time dependence of the remaining functions is obtained as in the previous case.
Finally, we need to determine $C_1(r)$, for doing that we may proceed as follows:
from (\ref{27}) and (\ref{mumo}) we have
\begin{equation}
m'=\frac{4\pi}{r^2}R'R^2\mu_0C_1.
\label{f}
\end{equation}
Taking the $r$ derivative of $m$ from (\ref{op21}), feeding it back into (\ref{f}) and evaluating the result at $t=0$ we obtain $C_1(r)$.
These solutions are regular everywhere (within the fluid distribution) and satisfy basic physical restrictions such as positivity of energy density distribution and $\mu>P_{\perp}$.
\section{CONCLUSIONS}
We have revisited the original Skripkin problem, but now allowing for a more general fluid distribution. This enabled us to find analytical models which satisfy Darmois junction conditions on both delimiting boundary surfaces.
Obviously, relaxing Darmois conditions on either (or both) boundary hypesurfaces enlarges the family of possible solutions, some of them are explicitly exhibited in previous sections (for models of voids within the thin wall approximation, see \cite{v4}-\cite{v6} and references therein).
The most important contribution in this manuscript is to illustrate the advantages of using the expansionfree condition for the modeling of cavity evolution. For that purpose we have integrated analytically the pertinent equations under a variety of circumstances, proving that such an integration could be achieved without undue difficulty. The obtained solutions were not intended to describe any specific astrophysical scenario but just to bring out the potential of the expansion--free condition to model situation where cavities are expected to appear.
Possibly our solutions could be applied as toy models of localized systems such as supernova explosion models. Most likely the combination of the vanishing expansion scalar condition with numerical integration of corresponding equations would produce more physically meaningful models.
It is worth mentioning the fact that in spite of the vanishing of the expansion scalar, the energy-density changes with time. This effect is due to the work done by the anisotropic force as it appears from (\ref{a}).
\begin{acknowledgments}
LH wishes to thank Fundaci\'on Empresas Polar for financial support. ADP acknowledges hospitality of the Departamento de F\'isica Te\'orica e Historia de la F\'\i sica at Universidad del Pa\'\i s Vasco. JO wishes to thank Universidad de Salamanca
for financial support under grant No. FK10.
\end{acknowledgments}
|
\section{Supplementary}
\section{Estimation of electric field enhancement:}
We consider a Gaussian laser beam with power $P$ and frequency
$\omega$ incident on a photonic crystal cavity. The power is
measured in front of the objective lens and the coupling
efficiency of the laser to the cavity is $\eta$. If the cavity
quality factor is $Q=\omega_0/\Delta\omega$, with cavity resonance
frequency $\omega_0$ and linewidth $\Delta\omega$, the energy
inside the cavity (for a laser resonant to the cavity) is
$W=P\eta/ \Delta\omega$. For an off-resonant cavity, where the
laser is detuned from the cavity by $\Delta$, the previous
expression for energy is multiplied by a Lorentzian:
\begin{equation}
f=\frac{1}{1+(2\Delta/\Delta\omega)^2}
\end{equation}
where $\Delta=\omega-\omega_0$ with $\omega_0$ being the resonance
frequency of the cavity. The energy in the cavity can also be
expressed as $\epsilon |E_{max}|^2 V_m$, where $\epsilon$ is the
permittivity of the medium, and $E_{max}$ is the electric field at
the point of maximum electric energy density, and $V_m$ is the
cavity mode volume. Equating the two expressions of energy, we can
write
\begin{equation}
\frac{P\eta} {\Delta\omega}
\frac{1}{1+(2\Delta/\Delta\omega)^2}=\epsilon |E_{max}|^2 V_m
\end{equation}
Using
\begin{equation}
\Delta\omega=\frac{\omega_0}{Q}=\frac{2\pi c}{Q\lambda_0}
\end{equation}
where $c$ is the velocity of light and $\lambda_0$ is the
resonance wavelength of the cavity, we can find that $E_{max}$:
\begin{equation}
\label{eqn_power_field} |E_{max}|=\sqrt{\frac{\eta P Q
\lambda_0}{2\pi c \epsilon
V_m}\frac{1}{1+(2\Delta/\Delta\omega)^2}}
\end{equation}
If the quantum dot is not located at the point of the maximum
electric field energy density, the electric field at its location
will be smaller than $E_{max}$ (and the spatial variation of the
E-field is determined by the mode pattern $\psi(x,y)$). Therefore,
the electric field at the location of the QD would be
\begin{equation}
|E_{cav}|=|E_{max}|\psi(x,y)
\end{equation}
On the other hand, when there is no cavity present, the intensity
$I$ of the light (assuming a Gaussian beam) incident on the GaAs
is given by
\begin{equation}
I=\frac{P}{2\pi \sigma_0^2}
\end{equation}
where $\sigma_0$ is the Gaussian beam radius of the laser. Also
the intensity of the laser is given by
\begin{equation}
I=\frac{1}{2}c\epsilon |E|^2
\end{equation}
Equating these two, the electric field is found to be
\begin{equation}
|E|=\sqrt{\frac{P}{c\epsilon\pi \sigma_0^2}}
\end{equation}
Assuming normal incidence on the air-GaAs interface
\begin{equation}
|E_{GaAs}|=\frac{2}{1+n}|E|
\end{equation}
where $n$ is the refractive index of GaAs. We note that the effect
of the reflection in the interface, is embedded in $\eta$ for the
analysis done for the cavity. From the above discussion, the
electric field sensed by the QD in the absence of the cavity has
the form
\begin{equation}
|E_{nocav}|=\frac{2}{1+n}\sqrt{\frac{P}{c\epsilon\pi \sigma_0^2}}
\end{equation}
Comparing the cavity and no-cavity case, we can find that the
electric field enhancement is given by
\begin{equation}
\frac{E_{cav}}{E_{nocav}}=\frac{1+n}{2}\sqrt{\frac{\eta Q
\lambda_0
W_0^2}{2V_m}\frac{1}{1+(2\Delta/\Delta\omega)^2}}\psi(x,y)
\end{equation}
When the laser is resonant with the cavity, the maximum field
enhancement for a linear three hole defect ($L_3$) cavity is $\sim
350$, assuming $\eta=1\%$, $Q=10000$,$\lambda_0=927$ nm,
$\sigma_0=3 \mu$ m, $V_m=0.8(\lambda_0/n)^3$, and the QD at the
field maximum, i.e., $\psi=1$. For a detuning of $4$ linewidths
(as is true for our experiment), the maximum enhancement is $\sim
40$. We note that this maximum enhancement can be increased by
using a better quality factor cavity, or lower mode volume.
Another way to increase the enhancement is increasing the coupling
efficiency $\eta$ by using a waveguide or a fiber coupled to the
cavity.
\section{Estimation of the QD Dipole Moment}
The data of Fig. $3$ a, b allows for order of magnitude estimation
of system parameters such as QD dipole moment and effective QD
electric field. Assuming a coupling efficiency of the Gaussian
laser beam to the PC cavity mode $\eta$, we can estimate the
maximum laser field amplitude $E$ at the position of the QD using
the Eqn. \ref{eqn_power_field}. From the linear fit in Fig. $3$ b,
we estimate the dipole moment $\mu_d$ of the QD to be be on the
order of $22$ Debye, with $\eta=1\%$ as obtained previously with
the same grating coupled cavity design
\cite{englund_toishi_perturbed_cavity}. For this dipole moment,
the maximum QD-cavity interaction strength $g/2\pi$ should be
$\sim 29$ GHz, assuming the QD is located at the electric field
maximum, thereby leading to the strong coupling. As mentioned
previously, we did not observe the anti-crossing of the QD and
cavity peaks in PL and thus believe that the actual value of $g$
is smaller than this calculated value most likely because the QD
is not located at the cavity electric field maximum.
\section{Theory of Bichromatic Driving}
In the theoretical description of our experiment, we calculate the
emission spectrum of the cavity under bichromatic driving of an
off resonantly coupled QD and measure how the intensity of the
cavity emission changes as a function of the probe laser detuning.
The bichromatic driving of a two-level system has been analyzed
before \cite{bichromatic_QD}. We use similar techniques to analyze
the driving of a two-level system such as a QD, coupled to an
off-resonant cavity. The dynamics of a driven QD-cavity system is
given by the Jaynes-Cummings Hamiltonian:
\begin{equation}
H=\omega_{cav} a^\dag a + \omega_{QD}\sigma^\dag \sigma +
g(\sigma^\dag a + \sigma a^\dag) + J\sigma + J^*\sigma^\dag
\end{equation}
where $\omega_{cav}$ and $\omega_{QD}$ are, respectively, the
cavity and the dot resonance frequency; $a$ and $\sigma$ are,
respectively, the annihilation operator for a cavity photon and
the lowering operator for the QD; $g$ is the coherent interaction
strength between the QD and the cavity and $J$ is the Rabi
frequency of the driving laser. For bichromatic driving, the
driving field $J$ consists of a strong pump laser with Rabi
frequency $J_1$ tuned to the QD frequency and a weak probe laser
with Rabi frequency $J_2$, which can be tuned to arbitrary
frequency, parameterized by the pump-probe detuning $\delta$:
\begin{equation}
J=J_1 e^{i\omega_{QD} t} + J_2 e^{i (\omega_{QD}+\delta) t}
\end{equation}
In a frame rotating with the pump laser frequency the Hamiltonian
is
\begin{equation}
H = H_0+H(t)
= \Delta a^\dag a + g(\sigma^\dag a + \sigma a^\dag) + J_1\sigma_x
+ J_2\left(e^{i\delta t}\sigma + e^{-i\delta
t}\sigma^\dagger\right)
\end{equation}
where $\Delta=\omega_{cav}-\omega_{QD}$ is the QD-cavity detuning.
We note that for a bichromatic driving, the Hamiltonian is always
time-dependent. To treat incoherent processes we use the master
equation \cite{book:quan_noise}:
\begin{eqnarray*}
\dot{\rho} &=& -i [H_0+H(t),\rho] +
\mathcal{D}\left(\sqrt{2\gamma}\sigma\right)\rho +
\mathcal{D}\left(\sqrt{2\kappa}a\right)\rho +\\
& & \mathcal{D}\left(\sqrt{2\gamma_r\bar{n}}a^\dagger\sigma\right)\rho
+\mathcal{D}\left(\sqrt{2\gamma_r(1+\bar{n})}a\sigma^\dagger\right)\rho
+ \mathcal{D}\left(\sqrt{2\gamma_d}\sigma^\dagger\sigma\right)
\end{eqnarray*}
where $\mathcal{D}\left(C\right)\rho$ is the Lindblad term $C\rho
C^\dagger-\frac{1}{2}\left(C^\dagger C\rho + \rho C^\dagger
C\right)$ associated with the collapse operator $C$. The first
two terms represent QD spontaneous emission with a rate $2\gamma$,
and cavity decay with a rate $2\kappa$. The two terms with
$\gamma_r$ represent a phonon mediated coupling between the cavity
and the QD \cite{majumdar_phonon_11}. The last term with
$\gamma_d$ phenomenologically describes pure dephasing of the QD.
We numerically calculate the emission spectrum of the cavity given
by the Fourier transform of the two-time correlation function of
the cavity field, proportional to $\langle a^\dagger (\tau)
a(0)\rangle$. Under the quantum regression theorem the
auto-correlation function is equal to tr$\{ a^\dagger M(\tau)\}$
where $M(\tau)$ obeys the master equation with initial condition
$a\rho(t\rightarrow\infty)$. The time dependence of the
Hamiltonian is such that the master equation can be cast in terms
of Liouvillian superoperators as
\begin{equation}\label{rhoEOM}
\dot{\rho}=\left(\mathcal{L}_0+\mathcal{L}_+e^{i\delta
t}+\mathcal{L}_-e^{-i \delta t}\right)\rho
\end{equation}
where
\begin{eqnarray*}
\mathcal{L}_0\rho &=& -i [H_0,\rho] +
\mathcal{D}\left(\sqrt{2\gamma}\sigma\right)\rho +
\mathcal{D}\left(\sqrt{2\kappa}a\right)\rho\\
&&+\mathcal{D}\left(\sqrt{2\gamma_r\bar{n}}a^\dagger\sigma\right)\rho
+\mathcal{D}\left(\sqrt{2\gamma_r(1+\bar{n})}a\sigma^\dagger\right)\rho
+ \mathcal{D}\left(\sqrt{2\gamma_d}\sigma^\dagger\sigma\right)\\
\mathcal{L}_+\rho &=& -i[\sigma,\rho] \\
\mathcal{L}_-\rho &=&-i[\sigma^\dag,\rho]
\end{eqnarray*}
This equation is solved by Floquet theory, by assuming a solution
of the form
$\rho(t)=\displaystyle\sum_{n=-\infty}^{\infty}{\rho_n(t) e^{i
n\delta t}}$. The number of terms in the expansion necessary to
obtain any level of precision is determined by the relative
strength of $J_1$ to $J_2$, and in this way the problem can be
considered perturbative in the probe strength. After Laplace
transforming the master equation, the method of continued
fractions is used to obtain the resonance fluorescence spectrum of
the cavity \cite{qotoolbox}. The height of the peak at the cavity
resonance is calculated as a function of the probe detuning
$\delta$. The criterion for the appearance of dressed states is
that the pump Rabi frequency $J_1$ should be higher than the QD
linewidth $2\gamma$. The inclusion of incoherent terms $\gamma_r$
and $\gamma_d$ effectively broadens the dot and alters this
condition, but below a certain critical value of $J_1$ the change
in the cavity height with probe detuning is a simple Lorentzian
with a linewidth on the order of the natural QD linewidth. Above
threshold, the dressed states are resolvable and cavity height
spectrum splits into two peaks in the experimental regime we
considered. Broadening of the peaks in the experiment beyond
theoretical prediction is caused by the spectral diffusion of the
QD. The parameters used for the simulations are: $\kappa/2\pi=17$
GHz, $\gamma/2\pi=1$ GHz, $\gamma_r/2\pi=.5$ GHz,
$\gamma_d/2\pi=3$ GHz, $\Delta=8\kappa$, $\bar{n}=1$. For the
simulation reported here we assume $g=0$, as the QD is not
strongly coupled to the cavity. Increasing $g$ makes the two peaks
more asymmetric.
\section{Numerical Simulation: Dependence of the splitting on pump power}
We show in the paper that we can probe the coherent interaction
between the QD and the resonant laser by monitoring the
off-resonant cavity emission (Fig. $1$ c in the paper). Both
theoretically and experimentally we observe two peaks at lower
pump powers. At higher pump power, we theoretically observe two
dips. In our experiment, however, we cannot reach this regime of
high pump power.
The separation between the peaks and dips increases linearly with
the pump Rabi frequency. However, from the theoretical plot, we
find that the peaks are separated by $4$ times the laser Rabi
frequency (Fig. \ref{Fig_suppl_peaks})and the dips are separated
by twice the laser Rabi frequency (Fig. \ref{Fig_suppl_dips}).
More detailed theoretical derivation will be provided in
\cite{papag_prep}.
\begin{figure}
\centering
\includegraphics[width=3in]{Fig_suppl_peaks.eps}
\caption{ The separation between the two peaks (as shown in Fig.
$1$ c in the paper) as a function of the laser Rabi frequency. The
slope of the linear fit is $\sim 4$.\label{Fig_suppl_peaks}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=3in]{Fig_suppl_dips.eps}
\caption{ The separation between the two dips (as shown in Fig.
$1$ c in the paper) as a function of the laser Rabi frequency. The
slope of the linear fit is $\sim 2$. \label{Fig_suppl_dips}}
\end{figure}
\section{Numerical Simulation: effect of $g$}
In the numerical simulation results presented in the paper, we
assumed $g=0$, i.e., no coherent interaction is present between
the QD and the cavity. Inclusion of $g$ makes the two peaks
asymmetric. Fig. \ref{Fig_suppl_effect_of_g} shows the cavity
emission for a pump power of $25$, with $g/2\pi$ ranging from $0$
to $10$. Here the cavity is at a shorter wavelength compared to
the QD, and we observe that the peak closer to cavity is not
enhanced. This observation is starkly different from the resonance
fluorescence measurement, where the peak close to the cavity is
enhanced, as observed in \cite{majumdar_phonon_11,hughes_mollow}.
This indicates again, that this way of measuring the coherent
interaction between the QD and the laser is akin to an absorption
measurement.
\begin{figure}
\centering
\includegraphics[width=3in]{Figure_suppl_g.eps}
\caption{ Cavity emission as a function of probe laser wavelength,
for different dot-cavity coupling $g$.
\label{Fig_suppl_effect_of_g}}
\end{figure}
The authors acknowledge financial support provided by the Army
Research Office, Office of Naval Research and National Science
Foundation. A.M. was supported by the Stanford Graduate Fellowship
(Texas Instruments fellowship). E.K. was supported by the
Intelligence Community (IC) Postdoctoral Research Fellowship.
|
\section{Introduction}
Recently, due to the discovery of topological insulators (TIs), the search for
new topological phases has attracted a lot of interests in the condensed matter physics community\cite{hasan2010,qi2010,qi2010a,moore2010}. Interestingly, topologically non-trivial phases not only exist in insulators, but also in superconductors (SCs), which are dubbed topological superconductors (TSCs). For TIs, the unique feature is the existence of topologically protected edge states (or surface states), and similarly, there are also gapless modes
at the edge (surface) of the TSCs which are Majorana fermions due to particle-hole symmetry.
Bound Majorana modes in a vortex core and chiral Majorana modes along the edge were first
proposed in $p_x$+$ip_y$ SCs\cite{read2000} and later on helical Majorana (HM) modes have been suggested to exist in TSCs with
time reversal symmetry\cite{qi2009b,tanaka2009,sato2009}.
Majorana fermions may serve as a building block for topological quantum computation
due to their non-Abelian statistics\cite{ivanov2001,nayak2008}. However, unlike TIs, up to now, the existence
of TSCs and Majorana fermions has not been experimentally confirmed.
The difficulty of the detection of Majorana fermions lies in
the fact that they are charge neutral and have no direct response to electromagnetic fields.
Therefore, recently it has been suggested to detect Majorana fermions through a Dirac-Majorana
converter, converting a pair of Majorana fermions into one Dirac fermion and vice versa\cite{fu2009,akhmerov2009}.
Such a Dirac-Majorana converter has been first proposed for the detection of chiral Majorana modes. However, for the detection of HM modes, realistic proposals barely exist\cite{footnote1}.
In this paper, we consider the heterostructure of a TI thin film sandwiched by two SCs.
Due to the proximity effect on both the top and bottom surfaces of the TI, it is found that,
if the s-wave pairing functions at the top and bottom surfaces have a relative $\pi$ phase shift,
a topological superconducting phase emerges in the TI thin film possessing a HM mode along its edges.
Based on such a configuration, we propose an experimental setup for a HM interferometer
and discuss the transport properties which can be used to distinguish it from other types of Majorana interferometers.
How can the envisioned heterostructure schematically shown in Fig.~\ref{fig:configuration}(a) be constructed based on existing materials?
Superconductivity can be realized by copper doped Bi$_2$Se$_3$\cite{wray2010}, TlBiTe$_2$\cite{yan2010},
and Bi$_2$Te$_3$ under pressure\cite{zhang2011}, which can be straightforwardly integrated into TI thin films based on Bi$_2$Se$_3$ or Bi$_2$Te$_3$. Let us describe the minimal model that captures the essential physics of such a system. We assume that the chemical potential lies in the bulk gap of the TI, then the low energy physics of the TI film
is described by two Dirac cones at the top and bottom surfaces, given
by the four band Hamiltonian\cite{liu2010,lu2010}
\begin{eqnarray}
&&\hat{H}_0=\sum_k\psi^{\dag}_kh(k)\psi_k ,\nonumber\\
&&h(k)
=m{\tau_x}+\hbar v_f\left( k_x\sigma_y\tau_z-k_y\sigma_x\tau_z \right)
\label{eq:hk}
\end{eqnarray}
with the field operator $\psi_k=[c_{1\uparrow},c_{1\downarrow},c_{2\uparrow},c_{2\downarrow}]^T$
, where 1(2) denotes the top (bottom) surface and $\uparrow$
($\downarrow$) denotes spin up (spin down); $v_f$ is the Fermi velocity and
$m$ the hybridization between the top and bottom surface states.
Here, the Pauli matrices $\sigma_i$ denote the spin and $\tau_i$ the opposite surfaces.
The Hamiltonian (\ref{eq:hk}) satisfies time reversal symmetry,
$T\hat{H}_0T^{-1}=\hat{H}_0$, with the time reversal operator
$T=i\sigma_yK$ ($K$ denotes complex conjugation).
The proximity effect from the s-wave SC for the top and bottom surfaces
can be taken into account by the Bogoliubov-de Gennes (BdG) Hamiltonian
\begin{eqnarray}
&&\hat{H}_{BdG}=\sum_k
\hat{\Psi}^\dag_kH_{BdG}(k)\hat{\Psi}_k,
\hat{\Psi}_k=\left(
\begin{array}{cc}
\psi_k\\(\psi^\dag_{-k})^T
\end{array}
\right),\nonumber\\
&&H_{BdG}(k)=\left(
\begin{array}{cc}
h(k)-\mu&\Delta\\
\Delta^\dag&-h^T(-k)+\mu
\end{array}
\right)
\label{eq:HamBdG}
\end{eqnarray}
with the s-wave pairing function $\Delta$ given by
\begin{eqnarray}
\Delta=\left(
\begin{array}{cc}
i\Delta_1\sigma_y&0\\
0&i\Delta_2\sigma_y
\end{array}
\right),
\label{eq:Delta}
\end{eqnarray}
where $\Delta_{1(2)}$ is for the top (bottom) surface.
The BdG Hamiltonian is invariant under charge conjugation $C=\lambda_xK$,
where Pauli matrices $\lambda_i$ denote particle and hole.
To preserve time reversal symmetry, the pairing function $\Delta_{1(2)}$
can only be real, indicating that the relative phase of the pairing functions between
the top and bottom surfaces can only be 0 or $\pi$.
The band dispersion of the Hamiltonian (\ref{eq:HamBdG}) can be directly solved and
we first look at some simplified limits. If the chemical potential $\mu=0$,
the band dispersion is given by
\begin{eqnarray}
E_{rts}=s\sqrt{\hbar^2v_f^2k^2+E_g^2}, E_g=\sqrt{m^2+\Delta_+^2}+t|\Delta_-|
\label{eq:eigHamBdG}
\end{eqnarray}
with $r,t,s=\pm$ and $\Delta_\pm=\frac{\Delta_1\pm\Delta_2}{2}$.
The above band dispersion shows double degeneracy and the two degenerate
branches ($r=\pm$) can be traced back to time reversal symmetry (Kramers partners).
We note that for the branches with $t=+$, there is always a gap for any non-zero $m$, $\Delta_\pm$, while
for the branches with $t=-$, the gap will close when $m^2+\Delta_+^2=\Delta_-^2$.
At the critical point, the band dispersion (\ref{eq:eigHamBdG}) becomes linear
and there are totally two Dirac cones ($r=\pm$) which are Kramers partners.
For one Kramers partner, a single Dirac cone will induce the change of the
Berry phase by $\pi$ between the $E_g>0$ and $E_g<0$ regimes, while for the other one,
the change of the Berry phase is $-\pi$. Therefore, similarly to the quantum spin Hall
case \cite{bernevig2006d}, a topological phase transition is expected to happen across the point $E_g=0$ in parameter space.
When the SC gap $\Delta_\pm$ is much smaller than the hybridization gap $m$,
the system should be a trivial SC. Thus, we expect that the system should be a topologically non-trivial SC
when $m^2+\Delta_+^2<\Delta^2_-$ as shown in Fig. \ref{fig:configuration}(b) (which will be confirmed below).
We note that the above condition implies if $\Delta_1\Delta_2<0$. Hence, the non-trivial
phase only exists when the pairing functions of the s-wave SCs for the top and
bottom surfaces have opposite signs (i.e. they form a Josephson junction with a $\pi$ phase shift).
To determine the type of TSC, we consider the limit $\Delta_+=0$,
in which we can transform the Hamiltonian (\ref{eq:HamBdG}) into the basis
$c_{\pm,\uparrow(\downarrow)}
=\frac{1}{\sqrt{2}}\left( c_{1,\uparrow(\downarrow)}\pm c_{2,\uparrow(\downarrow)} \right)$,
and find the obtained Hamiltonian is block diagonal with the two blocks related by time
reversal symmetry. Each block of the new Hamiltonian is exactly the Hamiltonian for
the chiral TSC discussed, for instance, in Ref.~\onlinecite{qi2010c}.
Therefore, our system is expected to be nothing but a TSC with HM edge modes. This can be seen
in analogy to the quantum spin Hall state with helical edge states consisting of two copies of the
chiral quantum anomalous Hall state\cite{bernevig2006d}.
To substantiate the above picture, we take into account another
limit ($m=0$) in which the two surfaces are decoupled from each other.
Then, for one surface, our Hamiltonian is the same as the one discussed by Fu and Kane in Ref.~\onlinecite{fu2008}.
According to Fu and Kane, a Josephson junction with a $\pi$ phase shift will induce a helical
Majorana wire along the one-dimensional junction. In this decoupled regime ($m=0$), the prediction by Fu and Kane is consistent with ours that a HM wire emerges at the edge of our sandwich structure if there is a $\pi$ phase shift across the SC-TI-SC Josephson junction.
Next, we directly solve the Hamiltonian (\ref{eq:HamBdG}) in a half infinite plane ($y<0$)
to obtain the effective Hamiltonian for the HM mode. We assume that the x-direction is translation invariant so that $k_x$ is a good quantum number.
We will first solve the eigen value problem at $k_x=0$
\begin{eqnarray}
H_{BdG}(k_x=0,-i\partial_y)\Phi(y)=E\Phi(y)
\label{eq:eigeneqn}
\end{eqnarray}
and then project the Hamiltonian (\ref{eq:HamBdG})
onto the subspace of the eigenstates at $k_x=0$. In order to impose open boundary conditions,
we need to add a quadratic term into the hybridization term $m=m_0+B\partial_y^2$,
where $B$ is considered to be a small positive number.
For simplicity, we only consider the case $\mu=0$, $\Delta_+=0$ and $\Delta_->m_0>0$.
With such a simplication, we search for the zero-energy state ($E=0$) with the ansatz
$\Phi\propto e^{\lambda y}\phi$ for the boundary conditions $\Phi(0)=0$ and $\Phi(y\rightarrow-\infty)=0$,
and find the doubly degenerate eigenstates
\begin{eqnarray}
\Phi_i=\frac{1}{N_0}\left( e^{-\lambda_1y}-e^{-\lambda_2y} \right)\phi_i,\qquad i=\pm,
\label{eq:eigenstate}
\end{eqnarray}
where $N_0$ is a normalization factor and $\lambda_{1,2}$ given by
\begin{eqnarray}
\lambda_{1,2}=\frac{1}{2B}\left( -\hbar v_f\pm\sqrt{\hbar^2v_f^2-4B(m_0+\Delta_-)} \right)
\label{eq:lambda}
\end{eqnarray}
which are real numbers, smaller than zero. The wave functions $\phi_{+}=\frac{1}{4}\left[1+i,1-i,-1-i,1-i,1-i,1+i,-1+i,1+i \right]^T$ and
$\phi_{-}=\frac{1}{4}\left[-1-i,1-i,-1-i,-1+i,-1+i,1+i,\right.$ $\left.-1+i,-1-i \right]^T$ are given in the basis
$|1,\uparrow,e\rangle, |1,\downarrow,e\rangle, |2,\uparrow,e\rangle, |2,\downarrow,e\rangle, |1,\uparrow,h\rangle,
|1,\downarrow,h\rangle, |2,\uparrow,h\rangle, |2,\downarrow,h\rangle$.
Since the $\phi_\pm$ are invariant under charge conjugation ($C\phi_{\pm}=\phi_\pm$)
and can be related to each other by time reversal ($T\phi_+=-\phi_-$ and $T\phi_-=\phi_+$),
they are expected to be HM modes.
Indeed, for nonzero $k_x$, we project the Hamiltonian
(\ref{eq:HamBdG}) into the subspace of $\Phi_\pm$, which can be done by
expanding the field operator $\hat{\Psi}$ as $\hat{\Psi}(k_x,y)=\sum_i\Phi_i(y)\hat{\gamma}_i(k_x)$ ($i=\pm$),
and obtain the effective Hamiltonian
\begin{eqnarray}
\hat{H}_{\rm eff}=\sum_{k_x}\hbar v_fk_x\Bigl( \hat\gamma_+^\dag(k_x)\hat\gamma_+(k_x)-\hat\gamma_-^\dag(k_x)\hat\gamma_-(k_x) \Bigr),
\label{eq:HMES_Heff}
\end{eqnarray}
which is exactly the HM mode as expected.
After determining the condition of the topological phase for $\mu=0$,
we can easily extend to the $\mu\neq 0$ regime. This is done by analyzing the $\mu\neq 0$ case and afterwards adiabatically connecting the result to the $\mu=0$ regime. By solving the bulk dispersion, the helical topological superconductor phase is found to exist
in the regime $|m|<\sqrt{\left( \frac{\mu^2}{\Delta^2_-}+1 \right)\left(
\Delta_-^2-\Delta_+^2\right)}$.
\begin{figure}
\begin{center}
\includegraphics[width=3.5in]{configuration.eps}
\end{center}
\caption{ (Color online) (a) Side view of the SC-TI-SC heterostructure; (b) phase diagram for the SC-TI-SC heterostructure
when the chemical potential $\mu=0$.
}
\label{fig:configuration}
\end{figure}
In order to confirm the HM mode experimentally, we consider the setup shown in Fig.~\ref{fig:interferometer}(a). It is well-known that for a TI film, it is possible to obtain the two-dimensional quantum spin Hall state
with helical edge states by tuning the thickness of the film due to quantum confinement\cite{liu2010,lu2010}.
Then, for the TI film with proper thickness, we introduce an s-wave superconductor on both the top and bottom surfaces
near the upper edge as shown by the pink regime in Fig.~\ref{fig:interferometer}(a) (marked by the label TSC).
By tuning the relative phase of the pairing functions
for the SCs on the top and bottom surfaces, we can obtain a TSC with
HM modes at the edges. Hence, at the top edge in Fig.~\ref{fig:interferometer}(a), we find one helical Dirac fermion from the left hand side (LHS) will be splitted into two Majorana fermions and then recombined into another helical Dirac fermion at the right hand side (RHS). Therefore, a HM interferometer is obtained.
Next, we will discuss the transport properties of the HM interferometer within the scattering matrix formalism.
The HM interferometer can be divided into three parts, a Y junction at the LHS is connected to another Y junction
at the RHS by two freely propagating Majorana wires of different length $L_1$ and $L_2$.
The Y junction regime consists of one helical Dirac fermion mode and two Majorana fermion
modes, as shown in Fig.~\ref{fig:interferometer}(b). We name the Dirac fermion mode $|\psi^\alpha_{i(o)}\rangle$ ($\alpha=e,h$) and the two Majorana modes $|\gamma^\alpha_{i(o)}\rangle$ ($\alpha=1,2$), where i (o) stands for incoming (outgoing) mode. Time reversal requires
$T|\psi^\alpha_{i(o)}\rangle=(-)|\psi^\alpha_{o(i)}\rangle$ and $T|\gamma^\alpha_{i(o)}\rangle=(-)|\gamma^\alpha_{o(i)}\rangle$,
while particle-hole symmetry gives $C|\gamma^\alpha_{\eta}\rangle=|\gamma^\alpha_{\eta}\rangle$ and $C|\psi^{e(h)}_\eta\rangle
=|\psi^{h(e)}_\eta\rangle$ ($\eta=i,o$). The scattering wave function takes the form
$|\Psi\rangle=\sum_{\alpha,\eta}\left( c_{\alpha,\eta}|\gamma^\alpha_\eta\rangle+d_{\alpha,\eta}|\psi^\alpha_\eta \rangle\right)$,
and consequently the S-matrix for the Y junction $S_Y$ is defined as
\begin{eqnarray}
\left(
\begin{array}{c}
c_{1,j,o}\\c_{2,j,o}\\d_{e,j,o}\\d_{h,j,o}
\end{array}
\right)=S_Y\left(
\begin{array}{c}
c_{1,j,i}\\c_{2,j,i}\\d_{e,j,i}\\d_{h,j,i}
\end{array}
\right),
\label{eq:Smatdef}
\end{eqnarray}
where $j=L,R$ for the left and right Y junction.
With the constraints from time reversal symmetry\cite{bardarson2008}, particle-hole symmetry,
and the unitary condition, $S_Y$ can be parametrized as
\begin{eqnarray}
S_Y=\left(
\begin{array}{cc}
\mathcal{R}_1&\mathcal{T}_1\\
-\mathcal{T}_1^T&-e^{-i\phi}\mathcal{R}_1^*
\end{array}
\right),
\label{eq:SmatY}
\end{eqnarray}
where
\begin{eqnarray}
\mathcal{R}_1=\left(
\begin{array}{cc}
0&r_1\\
-r_1&0\\
\end{array}
\right),
\mathcal{T}_1=\left(
\begin{array}{cc}
t_1&t_2\\
it^*_2&-it^*_1
\end{array}
\right),
\label{eq:TR}
\end{eqnarray}
$|t_1|^2+|t_2|^2+|r_1|^2=1$, $r_1$ is real and $t_2=t^*_1$. In the latter equations, $r_1$ refers to the reflection amplitude, and $t_{1(2)}$ are the transmission amplitudes depending on the microscopic details.
In the middle regime, the two HM
fermions are propagating freely, only picking up a phase shift, therefore the corresponding
S-matrix $S_m$ reads
\begin{eqnarray}
\left(
\begin{array}{c}
c_{1L,i}\\c_{2L,i}\\c_{1R,i}\\c_{2R,i}
\end{array}
\right)=S_m\left(
\begin{array}{c}
c_{1L,o}\\c_{2L,o}\\c_{1R,o}\\c_{2R,o}
\end{array}
\right),
\label{eq:SmatdefTI}
\end{eqnarray}
with
\begin{eqnarray}
&&S_m=\left(
\begin{array}{cc}
0&\mathcal{T}_{LR}\\
\mathcal{T}_{RL}&0
\end{array}
\right),
\mathcal{T}_{LR}=\left(
\begin{array}{cc}
e^{i\theta_{1}}&0\\
0&e^{i\theta_{2}}\\
\end{array}
\right),
\label{eq:Smatn2}
\end{eqnarray}
and $\mathcal{T}_{RL}=-\mathcal{T}_{LR}^T$ due to time reversal symmetry. The angles
$\theta_1$ and $\theta_2$ denote the phase shifts of the Majorana fermions
propagating along the two arms of the TSC. The phases (in the absence of trapped vortices in the TSC region) are given by $\theta_1 =k_x L_1 + \pi$, $\theta_2 =k_x L_2$,
where $L_{1,2}$ is the length of the upper (index 1) and the lower (index 2) arm in Fig.~\ref{fig:interferometer}(a). The $\pi$ phase shift, which we choose to add to $\theta_1$, comes from a Berry phase of a spin-1/2 rotation.
Combining the above S-matrix for three different regions, we can obtain the
total S-matrix $S_t$ defined as
\begin{eqnarray}
\left(
\begin{array}{c}
d_{eL,o}\\d_{hL,o}\\d_{eR,o}\\d_{hR,o}
\end{array}
\right)=S_t\left(
\begin{array}{c}
d_{eL,i}\\d_{hL,i}\\d_{eR,i}\\d_{hR,i}
\end{array}
\right),
\label{eq:Smatf2}
\end{eqnarray}
which connects the helical Dirac fermions on the LHS to those
on the RHS. The matrix elements of $S_t$ are named $s^{\alpha\beta}_{jl}$ where $\alpha,\beta=e,h$ denote electron or hole
and $j,l=L,R$ denote left or right lead. Then, the conductance is given by\cite{blonder1982,anantram1996,chung2010}
\begin{eqnarray}
&&G_{jl}=\frac{d\langle \hat{I}_j\rangle}{d V_l}=\frac{e^2}{h}\left[ \delta_{jl}-T^{ee}_{jl}+T^{he}_{jl} \right]
\label{eq:Conductance1}
\end{eqnarray}
where $T^{\alpha\beta}_{jl}=|s^{\alpha\beta}_{jl}|^2$. $G_{jl}$ represents the differential (i.e. non-linear) conductance for the current in lead $j$ with respect to the voltage in lead $l$. We only consider the zero temperature case for simplicity.
Here, $T^{ee}_{jj}$ and $T^{ee}_{jl}$ ($j\neq l$) are the electron reflection and transmission probabilities, respectively, while
$T^{eh}_{jj}$ and $T^{eh}_{jl}$ ($j\neq l$) are the Andreev reflection and transmission probabilities, respectively.
By solving $S_t$ from the expressions (\ref{eq:Smatdef})-(\ref{eq:Smatn2}), we can obtain the $T^{\alpha\beta}_{jl}$. We find $T^{ee}_{LL}$ and $T^{ee}_{RR}$ vanish because of time reversal symmetry.
The Andreev reflection, the electron transmission, and the
Andreev transmission probabilities are given by
\begin{eqnarray}
T^{he}_{LL}=T^{he}_{RR}=\frac{4r_1^2\sin^2\left( \frac{\theta_1+\theta_2}{2} \right)}{r_1^4+1-2r^2_1\cos(\theta_1+\theta_2)},
\label{eq:TheLLRR}\\
T^{ee}_{LR}=T^{ee}_{RL}
=\frac{4|t_1|^4\sin^2(\frac{\theta_1-\theta_2}{2})}{r_1^4+1-2r_1^2\cos(\theta_1+\theta_2)},
\label{eq:TeeLRRL}\\
T^{he}_{LR}=T^{he}_{RL}=\frac{4|t_1|^4\cos^2(\frac{\theta_1-\theta_2}{2}) }
{r_1^4+1-2r_1^2\cos(\theta_1+\theta_2)},
\label{eq:TheLRRL}
\end{eqnarray}
respectively, which satisfy the current conservation condition $\sum_{l\beta}T^{\alpha\beta}_{jl}=1$.
\begin{figure}
\begin{center}
\includegraphics[width=3.3in]{interferometer.eps}
\end{center}
\caption{ (Color online) (a) Proposed experimental setup for the HM interferometer; (b) Zoom into the Y junction, where Dirac fermions (full lines) are converted into Majorana fermions (dashed lines) and vice versa; (c) Dependence of the local conductance $G_{LL}=dI_L/dV_L$ and (d) the non-local conductance $G_{LR}=dI_L/dV_R$ on the phase factors $\theta_{1/2}$ (in units of $\frac{e^2}{h}$).}
\label{fig:interferometer}
\end{figure}
From the expressions (\ref{eq:TheLLRR})-(\ref{eq:TheLRRL}), we find that the current through our HM interferometer
is qualitatively different from both (i) the current through a TI-SC-TI junction\cite{adroguer2010} and (ii) the current through a chiral Majorana interferometer\cite{fu2009,akhmerov2009}. First, although there is no electron reflection in the HM interferometer, Andreev reflection still exists. This is similar to the TI-SC-TI junction\cite{adroguer2010} but different to the chiral Majorana interferometer\cite{fu2009,akhmerov2009}. Second, in our case, an electron can transmit both as an electron or a hole, which can be used to distinguish the HM interferometer from the TI-SC-TI junction\cite{adroguer2010}, where only electron transmission is allowed.
In the HM interferometer, the normal and Andreev transmission probability depend on the phase difference between the upper and lower arms. It can be tuned by changing the applied bias, by varying the lengths of the arms $L_{1/2}$, or by varying the number of vortices trapped in the TSC regime as recently suggested in Ref.~\onlinecite{beri2011}.
The dependences of the local conductance $G_{LL}$ and the non-local conductance $G_{LR}$ on the parameters $\theta_1$ and $\theta_2$
are shown in Fig.~\ref{fig:interferometer}(c) and (d), respectively. The observation of such an interference pattern would be a clear signature of HM modes in this setup.
To summarize, we have proposed a realistic setup for a HM interferometer which can be used
to confirm the existence of HM modes and helical TSCs. The core parts of our setup could, for instance,
be realized in Bi$_2$Se$_3$ thin films selectively doped by copper to turn layers within the structure superconducting.
A big advantage of our proposal is that the superconducting material and the TI material
are of the same type. Therefore, there should be no Schottky barriers between them maximizing the proximity effect. We have proposed to identify HM modes by measuring phase-dependent transport through a HM interferometer.
We would like to thank the Humboldt foundation (CXL), and the DFG-JST Research Unit ``Topological electronics''
(BT) for funding as well as B. Beri, M. Guigou, X.L. Qi, P. Recher, M. Wimmer, and S.-C. Zhang for interesting discussions.
|
\section{Introduction}
The superconducting quantum interference device \cite{Clarke} (SQUID) is recognized as the most sensitive magnetic-flux detector ever realized, and combines the physical phenomena of Josephson effect \cite{Josephson} and flux quantization \cite{Tinkham} to operate. SQUIDs are nowadays exploited in a variety of physical measurements \cite{gallop,greenberg} with applications spanning, for instance, from pure science to medicine and biology \cite{Clarke,kleiner}.
Recently, the interest in the development of nanoscale SQUIDs \cite{finkler,hao,troeman} has been motivated by the opportunity to exploit these sensors for the investigation of the magnetic properties of isolated dipoles \cite{ketchen,tilbrook2,tilbrook,huber,cleuziou}, with the ultimate goal to detect one single atomic spin, i.e., one Bohr magneton.
Here we theoretically analyze a hybrid superconducting interferometer
which exploits the phase dependence of the density of states (DOS) of a proximized metallic nanowire to achieve high sensitivity to magnetic flux.
The operation of a prototype structure based on this principle, the SQUIPT \cite{squipt}, has been recently reported.
We show that with a careful design transfer functions as large as a few mV$/\Phi_0$ and intrinsic flux noise $\sim 10^{-9}\Phi_0/\sqrt{\text{Hz}}$ can be achieved below 1 K. Limited dissipation joined with the opportunity to access single-spin detection
make
this structure attractive for the investigation of the switching dynamics of individual magnetic nanoparticles.
The paper is organized as follows. The model of the hybrid superconducting magnetometer is presented in Sec. II. The Josephson and quasiparticle current are calculated in Secs. III and IV, respectively. The flux resolution and device performance are finally presented in Sec. V, where we address briefly the feasibility of this structure as a single-spin detector. Sec. VI is devoted to the Conclusions.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{fig1.pdf}
\caption{\label{fig1} (Color online) (a) Scheme of the device. $L$ is the wire length whereas $w$ is the width of the superconducting tunnel junction (S$_2$) coupled to the middle of the N region. $\varphi$ is the macroscopic quantum phase difference in S$_1$, while $\Phi$ is the magnetic flux threading the loop. Furthermore, $R_t$ is the tunnel junction normal-state resistance and $I_{bias}$ is the current flowing through the structure.
}
\end{figure}
\section{Model}
The interferometer [sketched in Fig. \ref{fig1}(a)] consists of a diffusive normal metal (N) wire of length $L$ in good electric contact (i.e., ideal interface transmissivity) with two superconducting electrodes (S$_1$) which define a ring.
We assume the wire transverse dimensions to be much smaller than $L$ so that it can be considered as quasi-one-dimensional.
The contact with S$_1$ induces superconducting correlations in N through \emph{proximity} effect \cite{proximity,giaz,pot,pet1,pet2} which is responsible for the modification of the wire DOS \cite{gueron}. For lower-transparency NS$_1$ interfaces the proximity effect in the wire will be reduced thus weakening the effects described below.
In addition, a superconducting junction (S$_2$) of width $w$ and normal-state resistance $R_t$ is tunnel-coupled to the middle of the N region.
The loop geometry allows to change the
phase difference ($\varphi$) across the NS$_1$
boundaries through the application of a magnetic
field which modifies the wire DOS \cite{spivak}
and the transport through the tunnel junction \cite{lesueur,squipt}.
The proximity effect in the wire can be described with the quasiclassical Usadel equations \cite{proximity}.
The short-junction limit (i.e., for $\Delta_1\ll \hbar D/L^2=E_{Th}$, where $\Delta_1$ is the order parameter in S$_1$, $D$ is the wire diffusion constant, and $E_{Th}$ is the Thouless energy) will be considered in the following, since in such regime the Usadel equations allow an analytic expression for the wire DOS \cite{heikkila} thus simplifying the device analysis. In addition, the interferometer performance is optimized in this limit as proximity effect in the wire is maximized \cite{spivak,heikkila}.
Assuming a step-function form for the order parameter $\Delta_1$ \cite{selfconsistency}, i.e., constant in S$_1$ and zero in the N wire,
the wire DOS normalized to the DOS at the Fermi level in the absence of proximity effect is given by \cite{heikkila}
\begin{equation}
N_N(x,\varepsilon,T,\varphi)= \mbox{Re}\left\{\cosh\left[\theta(x,\varepsilon,T,\varphi)\right]\right\},
\end{equation}
where
\begin{equation}\label{teta}
\theta=\mbox{arcosh}[\alpha(\varepsilon,\varphi,T)\mbox{cosh}[2x\mbox{arcosh}[\beta(\varepsilon,\varphi,T)]]],
\end{equation}
\begin{equation}
\alpha=\sqrt{\varepsilon^2/[\varepsilon^2-\Delta_1^2(T)\mbox{cos}^2(\varphi/2)]}
\end{equation}
and
\begin{equation}
\beta=\sqrt{[\varepsilon^2-\Delta_1^2(T)\mbox{cos}^2(\varphi/2)]/[\varepsilon^2-\Delta_1^2(T)]}.
\end{equation}
In the above expressions, $\varepsilon$ is the energy relative to the chemical potential of the superconductors, $T$ is the temperature,
and $x\in [-L/2,L/2]$ is the spatial coordinate along the wire.
$N_N$ exhibits a minigap ($\varepsilon_g$)
\begin{equation}
\varepsilon_g(\varphi)=\Delta_1(T)|\mbox{cos}(\varphi/2)|
\end{equation}
for $|\varepsilon|\leq \varepsilon_g$ whose amplitude depends on $\varphi$ and is constant along the wire.
In particular, $\varepsilon_g=\Delta_1$ for $\varphi =0$ and decreases by increasing $\varphi$, vanishing at $\varphi=\pi$.
Finally, by neglecting the ring inductance the phase difference becomes $\varphi=2\pi\Phi/\Phi_0$, where $\Phi$ is the total flux through the loop area, and $\Phi_0=2.067\times 10^{-15}$ Wb is the flux quantum.
\section{Josephson current}
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{fig2.pdf}
\caption{\label{fig2} (Color online) Josephson current vs flux ($I_{eq}-\Phi$) characteristics calculated for a few values of temperature $T$ assuming $\Delta_1^0=0.1E_{Th}$ and $\Delta_1^0=4\Delta_2^0$. Thick lines represent the supercurrent calculated using the Ambegaokar-Baratoff formula Eq.~(\ref{Ambe}), whereas the thin lines are the supercurrent calculated for an extended tunnel junction with width $w=L/2$.
}
\end{figure}
At equilibrium a current through the system $I_{eq}$ can flow thanks to a direct Josephson coupling between the superconducting electrode S$_2$ [with order parameter equal to $\Delta_2(T) e^{i\chi_\text{S}}$ where $\chi_\text{S}$ is the macroscopic phase] and the proximized N wire.
We assume the BCS temperature dependence for $\Delta_{1,2}(T)$ with critical temperature $T_{\text{c1,2}}=\Delta_{1,2}^{0}/(1.764k_{\text{B}})$ where $\Delta_{1,2}^{0}$ is the zero-temperature order parameter in S$_{1,2}$, and
$k_{\text{B}}$ is the Boltzmann constant.
Since the junction NS$_2$ is extended in the $x$-direction one can calculate, in the tunneling limit, $I_{eq}$ using the quasiclassical approach~\cite{nazarov1,nazarov2,proximity} as the following integral
\begin{eqnarray}
& & I_{eq}(\varphi)=-\frac{1}{8ewR_t}\int_{-w/2}^{w/2} dx \\
& & \int_{-\infty}^{+\infty} d\epsilon \text{Tr} \{ \sigma_3 [G_{\text{R}}^\text{N}(\epsilon,x,\varphi),G_\text{R}^{\text{S}_2}(\epsilon)]
\tanh (\frac{\epsilon}{2k_BT}) \} ,\nonumber
\end{eqnarray}
where $\sigma_3$ is the third Pauli matrix, $[\cdot,\cdot]$ represent the commutator, and $e$ is electron charge.
Furthermore, $G_\text{R}^\text{N}(\epsilon,x,\varphi)$ is the position-dependent retarded Green's function on the N wire and $G_\text{R}^{\text{S}_2}(\epsilon)$ is the retarded Green's function of the electrode S$_2$.
They are defined as follows:
\begin{equation}
G_\text{R}^\text{N}(\epsilon,x,\varphi)=\left(
\begin{array}{cc}
\cosh \theta & \sinh \theta e^{i\chi} \\
-\sinh \theta e^{-i\chi} & -\cosh \theta
\end{array}
\right)
\end{equation}
and
\begin{equation}
G_\text{R}^{\text{S}_2}(\epsilon)=\frac{1}{\sqrt{\epsilon^2-\Delta_2^2}}
\left(
\begin{array}{cc}
\epsilon & \Delta_2 e^{i\chi_\text{S}} \\
-\Delta_2 e^{-i\chi_\text{S}} & -\epsilon
\end{array}
\right) ,
\end{equation}
where $\theta$ is defined in Eq.~(\ref{teta}) and
\begin{equation}\label{chi}
\chi=-\mbox{arctan}[\gamma(\varepsilon,\varphi,T)\mbox{tanh}[2x\mbox{arcosh}[\beta(\varepsilon,\varphi,T)]]],
\end{equation}
with
\begin{equation}
\gamma=\sqrt{[\varepsilon^2-\Delta_1^2(T)\mbox{cos}^2(\varphi/2)]}/[\Delta_1(T)\mbox{cos}(\varphi/2)].
\end{equation}
Note that the supercurrent $I_{eq}$ depends on the phase $\chi_S$ of the order parameter in the electrode S$_2$. We are interested in the critical current that we determine by fixing $\chi_S$ such that it gives the maximum supercurrent for a given value of $\varphi$.
In Fig.~\ref{fig2} the equilibrium critical current is plotted, for different values of temperature, as a function of $\varphi$ in units of $\Delta_2^0/(eR_t)$, assuming $\Delta^0_1=0.1E_{Th}$, $\Delta_1^0=4\Delta^0_2$ and $w=L/2$.
For the sake of comparison we also plot the supercurrent calculated through the Ambegaokar-Baratoff~\cite{ab} formula, relative to a point-like NS$_2$ junction:
\begin{eqnarray}\label{Ambe}
& & I_{eq}^\text{AB}=\frac{\pi\varepsilon_g(\varphi)\Delta_2(T) k_{\text{B}}T}{eR_t}\times \\
& & \sum_{l=0,\pm1,\cdots}\frac{1}{\sqrt{[\omega_l^2+\varepsilon_g^2(\varphi)][\omega_l^2+\Delta_2^2(T)]}} ,\nonumber
\end{eqnarray}
where $\omega_l=\pi k_{\text{B}}T(2l+1)$.
Interestingly, for our choice of parameters the difference between $I_{eq}$ and $I_{eq}^\text{AB}$ is hardly noticeable: the lateral spatial extension of the NS$_2$ junction along $x$ plays a marginal role.
As a matter of fact the supercurrent turns out to be negligible in the experiments reported in Refs.~\onlinecite{squipt,meschke}.
The reason for such Josephson current suppression is presently unclear. One possibility is that the system is brought out of equilibrium by voltage fluctuations which might originates from the measuring circuit. Such voltage fluctuations will drop mostly across the tunneling barrier, being the most resistive component of the system. As a result, the Josephson current will oscillates at high frequency, hindering the possibility of detection.
This fact is fortunate, since a supercurrent might prevent a correct voltage read-out of the device, which is not observed experimentally.
\section{Quasiparticle current}
The current through the tunnel junction biased at voltage $V$ is therefore dominated by the quasiparticles, and can be written as \cite{Tinkham}
\begin{equation}
I=\frac{1}{ewR_t}\int_{-w/2}^{w/2}dx\int d\varepsilon N_N(x,\varepsilon,\varphi)N_{S2}(\tilde{\varepsilon})F(\varepsilon,\tilde{\varepsilon}),
\end{equation}
where
\begin{equation}
N_{S2}(\varepsilon,T)=\frac{|\varepsilon|}{\sqrt{\varepsilon^2-\Delta_2(T)^2}}\Theta[\varepsilon^2-\Delta_2(T)^2]
\end{equation}
is the normalized DOS of the S$_2$ electrode,
$\tilde{\varepsilon}=\varepsilon-eV$,
$\Theta(y)$ is the Heaviside step function, $F(\varepsilon,\tilde{\varepsilon})=[f_0(\tilde{\varepsilon})-f_0(\varepsilon)]$,
and $f_0(\varepsilon)=[1+\mbox{exp}(\varepsilon/k_BT)]^{-1}$ is the Fermi-Dirac energy distribution.
In the following we set $\Delta_2^0=200\,\mu$eV and $\Delta_{1}^0=4\Delta_2^0=800\,\mu$eV as representative values for a structure exploiting aluminum (Al) and vanadium (V) as superconductors \cite{pascual,quaranta}, respectively, $w=L/2$ and $R_t=5$ M$\Omega$.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{fig3.pdf}
\caption{\label{fig3} (Color online) Interferometer quasiparticle current vs voltage ($I-V$) characteristics calculated for a few values of $\Phi$ at $T=0.1 T_{c2}$. $T_{c2}$ is the critical temperature of S$_2$, $I_{bias}$ is the current flowing through the device and $V(\Phi)$ is the resulting voltage modulation.
}
\end{figure}
Figure 3 shows the interferometer current vs voltage ($I$-$V$) characteristics calculated at $T=0.1T_{c2}$ for different values of the applied flux $\Phi$ \cite{gamma}.
In particular, for $\Phi=0$ the $I-V$ characteristic resembles that typical of a superconductor-insulator-superconductor junction (i.e., S$_1$IS$_2$ where I denotes an insulator) where the minigap in the wire is maximized [i.e., $\varepsilon_g=\Delta_1(T)$], and the onset for large quasiparticle current occurs at \cite{Tinkham} $V=[\Delta_1(T)+\Delta_2(T)]/e$.
For $\Phi=\Phi_0/2$ the characteristic is similar to that of a normal metal-insulator-superconductor junction (i.e., NIS$_2$) with $\varepsilon_g$ suppressed. The curves show a peak at $V=|\Delta_1(T)-\Delta_2(T)|/e$ which corresponds to the singularity appearing in the tunneling characteristic between different superconductors \cite{Tinkham}.
In a current-biased setup
the interferometer operates as a flux-to-voltage transducer providing a voltage response $V(\Phi)$ that depends on the bias current $I_{bias}$ fed through the tunnel junction [see Fig. 3].
For any $I_{bias}$, $V(\Phi)$ is determined by solving the equation $I_{bias}-I=0$.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{fig4.pdf}
\caption{\label{fig4} (Color online) (a) $V$ vs $\Phi$ calculated for several bias currents $I_{bias}$ at $T=0.1T_{c2}$. (b) $\mathcal{F}$ vs $\Phi$ calculated for the same $I_{bias}$ values and $T$ as in (a).
}
\end{figure}
Figure 4(a) shows $V(\Phi)$ at $T=0.1T_{c2}$ calculated for several $I_{bias}$ values. $V(\Phi)$ turns out to be maximized at the lower bias currents where the voltage swing obtains values as large as $4\Delta_2^0/e$, whereas it is gradually reduced by increasing $I_{bias}$. The interferometer performance is thus improved at low $I_{bias}$.
An important figure of merit of the interferometer is represented by the flux-to-voltage transfer function \cite{Clarke}
\begin{equation}
\mathcal{F}(\Phi)=\frac{\partial V}{\partial \Phi}
\end{equation}
which is shown in Fig. 4(b) for the same $I_{bias}$ values as in panel (a). In particular, $\mathcal{F}$ as large as $\simeq 12.5\Delta_2^0(e\Phi_0)^{-1}$ can be obtained at the lowest currents, whereas it is gradually suppressed at higher $I_{bias}$.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{fig5.pdf}
\caption{\label{fig5} (Color online) (a) $V$ vs $\Phi$ calculated for a few temperatures at $I_{bias}=1.0\Delta_2^0/eR_t$. (b) $\mathcal{F}$ vs $\Phi$ calculated for the same $T$ values and $I_{bias}$ as in (a).
}
\end{figure}
The role of the temperature is shown in Fig. 5(a) which displays $V(\Phi)$ calculated for several $T$ values at $I=1.0\Delta_2^0/(eR_t)$.
An increase in $T$ leads to a reduction of $V(\Phi)$ as well as to a suppression and smearing of the voltage swing.
This directly reflects on the transfer function, as displayed in Fig. 5(b). We note that even at $T=T_{c,2}$, i.e., when S$_2$ is driven into the normal state,
$\mathcal{F}$ as large as $\simeq 8.6\Delta_2^0(e\Phi_0)^{-1}$ can be achieved.
It follows that voltage swings up to 0.8 mV and $\mathcal{F}$ as large as 2.5 mV/$\Phi_0$ can be achieved with the suggested materials combination for $T\lesssim1$ K.
\section{Noise and device performance}
We now turn on discussing the noise properties of the interferometer.
In the actual current-biased setup an important quantity is represented by the voltage noise spectral density ($S_V$) defined as
\begin{equation}
S_V=R_d^2S_I,
\end{equation}
where $R_d=\partial V/\partial I$ is the tunnel junction dynamic resistance, and $S_I$ is the current noise spectral density (shot noise) given by \cite{golubev}
\begin{equation}
S_I=\frac{2}{wR_t}\int_{-w/2}^{w/2}dx\int d\varepsilon N_N(x,\varepsilon,\Phi)N_{S2}(\tilde{\varepsilon})M(\varepsilon,\tilde{\varepsilon}),
\end{equation}
where
\begin{equation}
M(\varepsilon,\tilde{\varepsilon})=f_0(\tilde{\varepsilon})[1-f_0(\varepsilon)]+f_0(\varepsilon)[1-f_0(\tilde{\varepsilon})].
\end{equation}
The intrinsic flux noise per unit bandwidth of the interferometer ($\Phi_{ns}$) is related to the voltage noise spectral density as \cite{Clarke}
\begin{equation}
\Phi_{ns}=\frac{\sqrt{S_V}}{|\mathcal{F}(\Phi)|}.
\end{equation}
Note that $\Phi_{ns}\propto \sqrt{R_t}$, as $S_V\propto R_t$ and $\mathcal {F}(\Phi)$ is independent of tunnel junction resistance.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{fig6.pdf}
\caption{ (Color online) (a) $\Phi_{ns}$ vs $I_{bias}$ calculated for different $\Phi$ values at $T=0.3$ K.
(b) $\Phi_{ns}$ vs $I_{bias}$ for $\Phi=0.3\Phi_0$ calculated at different temperatures.
(c) $P$ vs $I_{bias}$ calculated for different $\Phi$ values at $T=0.3$ K. In all calculations we set $\Delta_2^0=200\,\mu$eV, $\Delta_1^0=800\,\mu$eV, and $R_t=5$ M$\Omega$.
}
\label{fig6}
\end{figure}
Figure \ref{fig6}(a) shows $\Phi_{ns}$ versus $I_{bias}$ for several flux values at $T=300$ mK.
$\Phi_{ns}$ is a non-monotonic function of $I_{bias}$ with a minimum which depends, for each $\Phi$, on the bias current.
In particular, an increase in $\Phi$ leads to a general reduction of $\Phi_{ns}$ at low $I_{bias}$, while
its minimum moves toward lower bias current.
We stress that $\Phi_{ns}$ as low as $10^{-9}\,\Phi_0/\sqrt{\text{Hz}}$ or better can be achieved at this temperature in the $\sim 10...80$ pA range for suitable values of $\Phi$.
This good flux sensitivity stems from the low shot noise $S_I$ (which is peculiar to all-superconducting tunnel junctions) together with a small $R_d$ at the biasing point and large $\mathcal{F}(\Phi)$.
The temperature dependence is displayed in Fig. \ref{fig6}(b) where $\Phi_{ns}$ vs $I_{bias}$ is plotted for different $T$ values at $\Phi=0.3\Phi_0$.
Notably, the minimum of $\Phi_{ns}$ turns out to be quite insensitive to the temperature up to $\simeq 900$ mK.
Then, higher $T$ yields to a reduction of the current window suitable for high flux sensitivity and to an overall enhancement of $\Phi_{ns}$.
Furthermore, for $T\geq T_{c2}$ [see the line corresponding to $T=1.5$ K in Fig. 6(b)] $\Phi_{ns}$ is significantly degraded in the whole $I_{bias}$ range. This emphasizes the effectiveness of a superconducting tunnel probe for a drastic suppression of $\Phi_{ns}$.
The impact of dissipation $P=VI$ is displayed in Fig.~\ref{fig6}(c) which shows $P$ vs $I_{bias}$ for different $\Phi$ values at $T=0.3$ K.
$P$ can largely change by varying $\Phi$ and $I_{bias}$ as well.
In particular, in the $\sim 10...80$ pA current range, $P$ can vary from a few fW to some tens of fW.
Such a small power has the additional advantage to prevent substantial electron heating in the N wire \cite{rmp}. By contrast, in conventional SQUIDs dissipation is typically from two to five orders of magnitude larger \cite{Clarke,kleiner}. As $P\propto R_t^{-1}$, dissipation can be tailored by choosing a proper value of the tunnel junction resistance.
For a correct operation of the interferometer the two following conditions should be fulfilled: i) $2\pi I_c^0\mathcal{L}_G/\Phi_0\lesssim 1$ \cite{Tinkham} (where $I_c^0$ is the zero-temperature critical current of the S$_1$NS$_1$ Josephson junction, and $\mathcal{L}_G$ is the loop geometric inductance), and ii) $\mathcal{L}_k^{S_1}\ll \mathcal{L}_k^{N}$ \cite{lesueur,squipt} [where $\mathcal{L}_k^{S_1,N}\simeq \hbar R_{S_1,N}/\pi \Delta_1^0$ is the kinetic inductance \cite{Tinkham}, and $R_{S_1,N}$ is the normal-state resistance of S$_1$(N)].
Condition i), where $I_c^0=0.66\pi\Delta_1^0/eR_N$ \cite{heikkila}, ensures to avoid magnetic hysteresis whereas ii), which is equivalent to $R_{S_1}\ll R_N$, ensures that the phase difference set by $\Phi$ drops entirely at the wire ends thus allowing a full modulation of its DOS.
As an additional set of parameters we choose a silver (Ag) wire with $L=80$ nm, cross section $\mathcal{A}=30\times10$ nm$^2$, and $D=0.02$ m$^2$s$^{-1}$ which yield $R_N=L/(\mathcal{A}\nu_F e^2D)\simeq 5.2\,\Omega$, where $\nu_F=1\times 10^{47}$ J$^{-1}$m$^{-3}$ is the DOS at the Fermi level in Ag, $\Delta_1/E_{Th}\simeq 0.3$, and $I_c^0\simeq 318$ $\mu$A.
By choosing, for instance, a circular washer geometry \cite{Clarke} with $2r=150$ nm as internal diameter and external radius $\mathcal{R}\gg r$ we get $\mathcal{L}_G=2\mu_0 r\approx 0.19$ pH, where $\mu_0$ is the vacuum permeability, so that $2\pi I_c^0\mathcal{L}_G/\Phi_0\approx 0.18$.
Condition ii) can be fulfilled as well by choosing a suitable washer thickness and $\mathcal{R}$.
As $\mathcal{L}_G$ has to be kept small to satisfy condition i) it follows that the present structure could be suitable for the measurement of the magnetic properties of small isolated samples.
In this context, the magnetometer sensitivity ($\mathcal{S}_n$) to an isolated magnetic dipole placed at the center of the loop is approximtely given by \cite{ketchen,tilbrook}
\begin{equation}
\mathcal{S}_n=\frac{2r\Phi_{ns}}{\mu_0\mu_B},
\end{equation}
where $\mu_B$ is the Bohr magneton.
With our choice for $r$ and by coupling the device to a cryogenic voltage preamplifier \cite{kiviranta} (which we assume dominates the voltage noise) with $\sqrt{S_V^{pre}}\simeq 0.1$ nV$/\sqrt{\text{Hz}}$ yields a total flux noise $\Phi_{ns}^{tot}=\frac{\sqrt{S_V^{pre}}}{\text{max}|\mathcal{F}(\Phi)|}\simeq 40$ n$\Phi_0/\sqrt{\text{Hz}}$, leading to
$\mathcal{S}_n\approx 1$ atomic spin$/\sqrt{\text{Hz}}$ below 1 K. Furthermore, the best achievable energy resolution \cite{Clarke} would be $\mathcal{E}=\frac{(\Phi_{ns}^{tot})^2}{2\mathcal{L}_G}\simeq 170\hbar$.
\section{Conclusions}
In summary, we have theoretically investigated a hybrid superconducting magnetometer whose operation is based on magnetic flux-driven modulation of the density of states of a proximized metallic nanowire. In particular, we have shown that with suitable geometrical and material parameters the interferometer can provide large transfer functions (i.e., of the order of a few mV/$\Phi_0$) and intrinsic flux noise down to a few n$\Phi_0/\sqrt{\text{Hz}}$ below 1K. Furthermore, joined with limited power dissipation, the structure has the potential for the realization of sensitive magnetometers for the investigation of the switching dynamics of small spin populations.
\acknowledgments
We acknowledge L. Faoro, R. Fazio, P. Helist$\ddot{\text{o}}$, L. B. Ioffe, M. Kiviranta, M. Meschke, Yu. V. Nazarov, J. P. Pekola, and S. Pugnetti for fruitful discussions. The FP7 program ``MICROKELVIN'' and the EU project ``SOLID'' are acknowledged for partial financial support.
|
\section{Introduction}
\label{sec:intro}
The $\Lambda$ Cold Dark Matter ($\Lambda$CDM) cosmology is currently the most widely accepted and successful paradigm for describing the Universe \citep{Blumenthal1984,Davis1985, Gramann1988,Peebles2003}. Consequently, structure in the Universe forms hierarchically, as shown analytically and in simulations \citep[e.g.][]{PressSchechter1974,WhiteRees1978,Davis1985}. First, dark matter collapses into small halos, and later these collect as subhalos into galaxies, where gas cools into a disk to form stars. In especially dense regions of the Universe galaxies bind together gravitationally into galaxy clusters.
In a hierarchical Universe, substructure is expected to be invariant at all scales of interest. In some of the earliest simulations that were able to resolve substructures, \citet{Moore1999} found that dark matter-only simulations of galaxies and galaxy clusters had the same number of substructures relative to the total mass of the system. A comparison of the simulations to observations showed that the simulated galaxy cluster matched the quantity of substructure in the nearby Virgo cluster, but that the simulated galaxy had significantly more substructure than the Local Group. Using constrained simulations of a system similar to the Local Group, \citet{Klypin1999} found far more substructure than what has been observed. This discrepancy between $\Lambda$CDM and observations is known as the ``missing satellites problem''. \citet{Kravtsov2010} provides a recent review of the progress made towards solving this problem.
Large observational surveys have also discovered a new class of ultra-faint galaxies \citep{Willman2005,Belokurov2007, Koposov2008}. The detection of such galaxies slightly lessens the number of satellites that are ``missing'' from the Local Group. Early observations of their velocity dispersions \citep{Simon2007} show that stars may form in halos of lower mass than previously believed possible, although the typical star formation efficiency in these low mass halos must be extremely low given that the halo mass function rises steeply at these masses \citep{Tollerud2009,Guo2010b}. A key question is whether there is a minimum mass halo in which stars can form, and if there is significant scatter in the star formation efficiency at a given halo mass. The low mean star formation efficiency at these masses might be driven by halo-to-halo variation, or a steep, universal relation between star formation efficiency and halo mass. The ultra faint satellites have been detected down to $M_{V}\approx-2$, which corresponds to 100 $L_{\sun}$, below the resolution of the simulations studied here. It is impossible to give a full census of such objects from the simulations. However, we show that some small objects do form in the simulations.
Generally, there are two paths pursued to solve the missing satellites problem. One is to alter the cosmological paradigm. Examples of this include self-interacting dark matter in which subhalos are destroyed through self-annihilation \citep{Spergel2000}, initially warm dark matter out of which small structures do not form \citep{Dalcanton2001}, or removing small scale perturbations from the primordial power spectrum \citep{Zentner2003}. Recent gravitational lensing studies have discovered dark substructures \citep{Dalal2002,Mao2004}, so it appears $\Lambda$CDM is consistent with observations and we must find what physical mechanisms play the largest role in darkening small galactic halos. This leads to the other path, which is to consider the effects of baryonic physics, such as stellar feedback \citep{Dekel1986,MacLow1999} and UV ionisation \citep{Efstathiou1992,Quinn1996,Bullock2000}, which might render many satellites dark.
The four primary mechanisms that can remove mass from halos are:
\begin{itemize}
\item {\bf UV ionisation:} luminous objects emit UV radiation that ionises hydrogen and sets a background temperature above the virial temperature of the subhalo.
\item {\bf Ram pressure stripping: } as a satellite passes through the hot halo gas, the incident gas pressure becomes stronger than the gravitational force of the satellite and gas is thus removed.
\item {\bf Stellar feedback:} stellar winds and supernovae inject sufficient energy into the interstellar medium (ISM) of the small galaxy so that some or all of the ISM is ejected.
\item {\bf Tidal stripping:} as a satellite orbits close to a larger host galaxy, the tidal forces become sufficient to remove material. Unlike the other three mechanisms mentioned above this is the only one that can remove collisionless matter, namely dark matter and stars, as well as gas from a subhalo.
\end{itemize}
Previous efforts have been made at examining these mechanisms in detail. Early efforts were analytical due to the large dynamic range necessary to properly simulate substructures, but recent simulations have allowed a closer look at satellites. \citet{Dekel2003} compared careful observations of many dwarfs with an analytical model based on the effect of supernova feedback and found that supernova feedback defines the line between low and high luminosity dwarf galaxies. \citet{Kravtsov2004} used high resolution, cosmological simulations to study the role of tidal stripping in the mass evolution of satellites. They concluded that the combined effect of tides and ionisation could produce a Milky Way-like satellite luminosity function. \citet{Read2006} considered both supernovae-driven winds and ionisation in cosmological simulations and found that ionisation was critical to make their simulations agree with observed luminosity functions. \citet{Governato2007} found in another series of cosmological simulations that UV background dramatically reduced the number of luminous subhalos, but that stellar feedback was required to make the simulated luminosity functions the same as those observed. More recent cosmological simulations by \citet{Okamoto2010} have varied the strength of a kinetic supernova wind feedback to determine exactly how much energy is required to produce the observed luminosity function. \citet{Klimentowski2010} saw how tidal stripping determines a subhalo's baryon content and final morphology. \citet{Wadepuhl2011} introduced black holes into their simulations and found that the black holes are not massive enough in subhalos to have an effect on their luminosities. They did, however, find that wind-driven galactic outflows can reduce the number of high mass satellites, and cosmic rays can suppress the luminosity of low mass subhalos. Each of these models successfully fit the data by studying in detail one or two mechanisms, while we will consider all four within SPH simulations.
Recent semi-analytic models also show some success at reproducing the observed satellite luminosity function. In these, only the more massive subhalos (\citet{OkomotoFrenk2009}, \citet{Guo2010a}, \citet{Maccio2009}) retain stars. One semi-analytic model of an N-body simulation of a Milky Way-like halo \citep{Li2009} reveals luminous subhalos whose mass in dark matter spanned one order of magnitude, while the luminosity ranged over five orders of magnitude, matching observations. There were also many more dark matter-only subhalos present, whose mass spanned three orders of magnitude.
\citet{Mayer2006} pointed out the importance of the combined effects of each mechanism. They simulated individual satellites falling into a static gravitational potential filled with hot, dense gas. In these simulations, tidal forces excite star formation and thus stellar feedback, as well as reshaping the gas distribution so that it can be more easily stripped due to ram pressure. \citet{Mayer2006} called this combination of processes ``tidal stirring'' and found that it can remove enough gas from dwarf irregulars to turn them into gas-free dwarf spheroidals.
The purpose here is to discover which mechanisms tore the baryons off the subhalos, focusing on UV background, tidal stripping, ram pressure stripping, and stellar feedback. For the first time, we will explicitly track the causes behind the departure of individual gas particles from their subhalo in order to construct a comprehensive picture showing the relative strength of each gas loss mechanism. We only analyse the satellites inside the virial radius, so we expect the satellites we are analysing to be similar to dwarf spheroidal (dSph) galaxies, a population that dominates the satellite population of the Milky Way within $r_{vir}$. Dwarf spheroidals are gas-poor \citep[e.g. Fig 3 in][]{Grebel2003} containing as little as $10^{4}$M$_{\sun}$ to undetectable amounts, but generally they continued forming stars until recently \citep{Skillman2007}.
\S\ref{sec:method} establishes the background behind the tools we used in this work. \S\ref{sec:simlumefns} introduces the subhalo population of our host galaxy g15784 at $z=0$, while \S\ref{sec:histhalos} details the history of these subhalos and how we determine the causes of baryon loss. The concluding \S\ref{sec:conclu} discusses the implications of our findings and areas for future work.
\section{Method}
\label{sec:method}
We analyse the evolution of the subhalos of two galaxies (g15784 and g5664) from the McMaster Unbiased Galaxy Simulations (MUGS). The purpose of MUGS is to provide a sample of M$^*$ galaxies simulated using SPH at high resolution. A full description of MUGS can be found in \citet{Stinson2010}, but we briefly summarize it here. The MUGS sample is chosen from a 50 $h^{-1}$ Mpc volume of a WMAP3 $\Lambda$CDM universe ($H_{0}$=73 km s$^{-1}$ Mpc$^{-1}$, $\Omega_{m}$=0.24, $\Omega_{\Lambda}$=0.76, $\Omega_{baryon}$=0.04, and $\sigma_{8}$=0.79) \citep{Spergel2007}. It consists of a random selection from the galaxies with halo masses between $\approx5\times10^{11}$ M$_{\sun}$ and $\approx2\times10^{12}$~M$_{\sun}$ that did not evolve near to structures more massive than $5.0\times10^{11}$M$_{\sun}$ within 2.7 Mpc. While this would have eliminated the Milky Way from our sample, there is no evidence for a past interaction between the Milky Way and M31 and so the Milky Way's satellite population should be unaffected by its near neighbor. The sample is unbiased with regards to angular momentum, merger history, and less massive neighbors and it is hoped that the sample will reproduce the observed spread in galaxy properties.
The only bias is random. The selected galaxies are simulated with the commonly-used zoom technique that focuses resolution on individual galaxies while maintaining the large scale torques necessary to give galaxies their angular momentum. The initial dark matter, gas and star particle masses are $1.1\times 10^{6}$~M$_{\sun}$, $2.2 \times 10^{5}$~M$_{\sun}$ and $6.3 \times 10^{4}$~M$_{\sun}$ respectively. Each type of particle uses a constant gravitational softening length of 310 pc. Most of this paper will be dedicated to the subhalos of one of these galaxies, g15784, which has a mass of $1.43\times10^{12}$~M$_{\sun}$. Its disk has a mass of $3.27\times10^{10}$~M$_{\sun}$ and the bulge a mass of $5.49\times10^{10}$~M$_{\sun}$, based on kinematic decomposition \citep{Stinson2010}. We also utilize the $5.2\times10^{11}$~M$_{\sun}$ galaxy g5664 to see how the luminosity function changes depending on the presence of stellar feedback and UV background.
Outputs were at most 214 Myr apart, especially at lower redshift, with irregular outputs at key times. Outputs were much closer together at high redshifts, typically 107 Myr apart.
MUGS was run using the SPH code \textsc{gasoline} \citep{Wadsley2004}. \textsc{gasoline} includes low temperature metal cooling (described in \citet{Shen2010} and briefly here), UV background radiation, star formation, and physically-motivated stellar feedback. The metal cooling grid is constructed using CLOUDY (version 07.02, last described by \citet{Ferland1998}), assuming ionisation equilibrium. A uniform ultraviolet ionising background, adopted from Haardt \& Madau (in preperation; see \citet{HaardtMadau}), is used in order to calculate the metal cooling rates self-consistently. It starts to have an effect at $z=9.9$.
\subsection{Star Formation and Feedback}
The star formation and feedback recipes are the ``blastwave model" described in detail in \citet{Stinson2006}. They are summarized as follows. Gas particles must be dense ($n_{\rm min}=1.0~\mathrm{cm}^{-3}$) and cool ($T_{\rm max}$ = 15,000 K) to form stars. A subset of the particles that pass these criteria are randomly selected to form stars based on the commonly used star formation equation,
\begin{equation}
\frac{dM_{\star}}{dt} = c^{\star} \frac{M_{gas}}{t_{dyn}}
\end{equation}
where $M_{\star}$ is mass of stars created, $c^{\star}$ is a constant star formation efficiency factor, $M_{gas}$ is the mass of the gas particle spawning the star, $dt$ is how often star formation is calculated (1 Myr in all of the simulations described in this paper) and $t_{dyn}$ is the gas dynamical time. The constant parameter, $c^{\star}$, is tuned to 0.05 so that the simulated isolated Model Milky Way used in \citet{Stinson2006} matches the \citet{Kenn1998} Schmidt Law, and then $c^\star$ is left fixed for all subsequent applications of the code.
At the resolution of these simulations, each star particle represents a large group of stars (6.32 $\times 10^4$ M$_{\sun}$). Thus, each particle has its stars partitioned into mass bins based on the initial mass function presented in \citet{Kroupa1993}. These masses are correlated to stellar lifetimes as described in \citet{Raiteri1996}. Stars larger than $8~$M$_{\sun}$ explode as supernovae during the timestep that overlaps their stellar lifetime after their birth time. The explosion of these stars is treated using the analytic model for blastwaves presented in \citet{MO77} as described in detail in \citet{Stinson2006}. While the blast radius is calculated using the full energy output of the supernova, less than half of that energy is transferred to the surrounding ISM, $E_{SN}=4\times10^{50}$ ergs. The rest of the supernova energy is assumed to be radiated away.
To capture the behavior of clustered star formation, we stochastically determine when a star particle releases feedback energy,
\begin{eqnarray}
p & = & \frac{N_{SNII}~mod~N_{SNQ}}{N_{SNQ}} \\
N_{ESN} & = &\left\lfloor\frac{N_{SNII}}{N_{SNQ}}\right\rfloor + \left\{
\begin{array}{l l}
0 & \quad ,r \leq p \\
N_{SNQ} & \quad ,r > p \\
\end{array} \right.
\end{eqnarray}
where $N_{SNII}$ is the number of supernovae calculated to explode during that star formation timestep, $N_{SNQ}$ is the ``supernova quantum'', i.e. the number of supernova required per explosion (fixed at $30$, the number of supernovae expected from the star particles in our simulation), and $N_{ESN}$ is the total number of supernova explosions that will have their energy distributed during a the star formation timestep. If the probability, $p$, is greater than a random number, $r$, selected between 0 and 1, $N_{SNQ}$ supernovae's worth of energy is released. This causes SN energy to be released in quantized packets over the 35 Myr until the largest star remaining is $< 8$ M$_{\sun}$.
\subsection{Group Finding: Amiga Halo Finder}
In order to identify the host galaxy and its subhalos, we used the Amiga Halo Finder (AHF) \citep{Knollmann2009}. AHF is based on the spherical overdensity method for finding halos. It is able to identify density peaks using an adaptive mesh algorithm. Once the density peaks are identified, AHF cuts out halos (and subhalos) using isodensity contours. Particles belonging to subhalos are distinguished from those of the background halo using a simple unbinding procedure to determine whether the particles are gravitationally bound to the subhalo. We base our analysis on a minimum group size of 50 particles, which is $2.2\times10^{7}$~M$_{\sun}$ when the group only contains dark matter but could be less massive if it also contains gas and star particles. Our analysis is restricted to those satellites identified by AHF as lying inside the virial radius ($r_{vir}$) of the halo. For g15784, $r_{vir}=240$~kpc, while for g5664, $r_{vir}=152$~kpc.
\subsection{Merger Trees}
\label{sec:mt}
We traced the histories of each subhalo in the galaxy. First, we identified groups at every output 100 Myr apart with AHF and then traced the particles present in the subhalos at $z=0$ back through the simulation including any gas out of which stars formed. For the sake of clarity, let us call the subhalo of interest ``Alpha''. At each output, we note every group that contains Alpha's particles. The group that had the largest number of Alpha's particles is set as Alpha's progenitor at that output. In this way, we trace the properties of each subhalo through time, including mass, distance from the host galaxy, and temperature.
Because this method only depends on the number of particles, a subhalo can ``jump'' in position space between outputs, switching between subhalos with different groups of particles. Out of the higher mass subhalos, such behaviour was only observed in three of them. This does not affect our results because the jumps only occurred at high redshifts and between subhalos of comparable mass and position that were soon to merge. The pivotal point of analysis in each subhalo's history is its maximum mass, as will
be show, in \S \ref{sec:histhalos}, and it occurs well after any jumping behavior we see. Also, we will only look at the \textit{last} time that a particle leaves its subhalo in order to prevent a single gas particle from being double-counted. Therefore, any event of gas being ``lost'' due to subhalo jumping will automatically be removed from the analysis.
\section{Simulated Luminosity Functions}
\label{sec:simlumefns}
The first analysis we undertook was to compare the satellite luminosity function of g15784 to observations. We found that the cumulative number is similar to the Milky Way's, though there was an excess of high luminosity satellites and so the shape of the cumulative functions did not match. We also re-simulated a smaller galaxy, g5664, with and without the UV background and stellar feedback to compare the effects the presence these two mechanisms have on the subhalo population as a whole.
\subsection{The Main Halo: g15784}
\label{sec:g15784}
AHF found 107 satellites inside $r_{vir}$ of g15784. In order to determine the luminosity of each subhalo, we treated each star particle as its own stellar population, where MUGS allowed us to model a population of stars with a distribution of ages. We based the brightness of the stars on the luminosity grid provided by CMD 2.1 \citep{Leitherer1999,Marigo2008}. Using the grid, we performed a bilinear interpolation over the stellar ages and metallicities of each star particle and then summed the luminosities of all the star particles in each satellite to derive a stellar magnitude for the satellites. We neglect the effects of dust extinction since dwarf galaxies are low metallicity and rarely appear dust obscured \citep{Mateo1998,Lisenfeld1998}. This MUGS galaxy has an effective resolution of $2048^3$, and we will also compare g15784 to a lower resolution run with an effective resolution of $1024^3$.
Figure \ref{fig:cumemag} shows the cumulative luminosity function of the host galaxy's subhalo population in the V-band at $z=0$. The results show that down to $M_{V} \approx -6$, the dimmest subhalo in our simulation, the number of satellites is 23 compared with 20 to 21 for the Milky Way at a similar magnitude. Thus, when all the relevant baryonic processes are included the order-of-magnitude missing satellites problem \citep{Moore1999, Klypin1999} disappears as has been seen in other simulations of comparable resolution \citep{Okamoto2010, Knebe2010, Wadepuhl2011}.
For comparison to observations we include two lines for reference. This first is a recent compilation of the classical satellites and the new ultra-faint dwarf galaxies from \citet{Tollerud2008}. The census of ultra-faint dwarfs is certainly incomplete, both due to the limited sky coverage of the Sloan Digital Sky Survey and the difficulty in detecting very faint galaxies. These galaxies extend to lower luminosities than our simulations can resolve, where our faintest satellite has one star particle. The resolution of our simulations is not sufficient to make robust predictions regarding the new classes of ultra-faint dwarfs.
The second is \citet{Koposov2008} who modelled these effects and derived a corrected luminosity function, $\frac{dN}{dM_{V}}=10\times10^{0.1(M_{V}+5)}$ for $-18 \le M_{V} \le -2$, that would represent a theoretically complete set of satellite galaxies; we have also plotted this function in Figure~\ref{fig:cumemag}. Our host galaxy g15784 has a mass close to the Milky Way's, so it is reasonable that its cumulative number of luminous subhalos is comparable to the Milky Way's, as indeed it is. In other words, \emph{our simulated galaxy does not suffer from the missing satellite problem}. The major difference between the simulations and observations is an excess of brighter satellites in the simulations, and that our star formation recipe may form too many stars. This causes an extra ``knee'' in the shape of our luminosity function.
The dash-dotted line in Figure \ref{fig:cumemag} shows the luminosity function of a lower resolution simulation exactly the same as g15784, but with half the spatial resolution and an initial gas particle mass of $\approx 10^{6}$ M$_{\sun}$. From this it is clear that decreasing resolution decreases the number and luminosity of subhalos. The lowest luminosity subhalo, at $M_V \approx -8.2$, contains a single star; in our higher resolution run this magnitude corresponds to 10 star particles. We discuss resolution effects in \S \ref{sec:res}.
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig01.eps}}
\caption{The cumulative V-band luminosity function of the host galaxy's subhalos at $z=0$, in solid black. Observational data of the Milky Way as gathered by \citet{Tollerud2008} is shown in dashed red, while a theoretical completion-corrected function from \citet{Koposov2009} is shown in dashed blue. The black dash-dotted line is the luminosity function from a lower resolution run of the same galaxy where the initial gas particle mass is $\approx 10^{6}$ M$_{\sun}$.} (Note that the observed ultra-faint dwarfs extend to a much lower luminosity). Within the resolution limit, all three functions lie within an order magnitude of each other.
\label{fig:cumemag}
\end{figure}
Figure \ref{fig:msmtot} shows the baryonic mass of each subhalo as a function of total mass at $z=0$. The dashed line shows where subhalos that obey the cosmic baryon fraction would lie. The subhalos contain systematically fewer baryons below $5\times10^8$ M$_{\sun}$. This falloff is similar to the low mass dropoff found by \citet{McGaugh2010} in the observed baryonic Tully-Fisher relationship. In these lower mass halos, baryons are preferentially stripped. As we shall see, these halos have also lost a significant amount of dark mass. However, almost all the lower mass halos also have fewer than $10$ baryons.
Additionally, below $2\times10^9$ M$_{\sun}$ there are many halos that contain no baryon particles according to our simulation's resolution at all and are thus dark. Of the 23 satellites that contain baryons, only 10 contain gas with the maximum gas fraction being 4 \% of the total mass. This fraction might seem low if compared to high gas fractions found in isolated dwarf irregular galaxies \citep{Geha2006}, but the subhalos considered here are all within $r_{vir}$ and are more appropriately compared with the dwarf spheroidals presented in \citet{Grebel2003}, which contain little gas. We must stress that when we say that a subhalo ``has no baryons'' we mean in the sense of our simulation's resolution there are no baryon particles. Actual dark subhalos will always have at least some trace quantities of gas.
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig02.eps}}
\caption{Baryon mass versus total mass for our subhalo population. The symbols correspond to mass to light ratio (total mass divided by baryonic mass), and the dashed black line is where subhalos with the cosmic baryon fraction would lie. The lower bound of the mass-to-light ratio is inclusive. The horizontal dotted line is the most luminous of the satellites with fewer than $10$ baryon particles, corresponding to our resolution limit. Most striking here is that the luminous subhalos, the brightest of which follow the power law, and dark subhalos overlap in total mass.}
\label{fig:msmtot}
\end{figure}
Figure \ref{fig:cumemass} shows the cumulative mass function at $z=0$ for all the satellites as well as the subset that formed stars. The total satellite mass function is similar to the collisionless, dark matter-only simulations of \citet{Moore1999} and \citet{Klypin1999} while the mass function of luminous satellites is closer to what is observed. There is about an order of magnitude more subhalos in total than those that contain baryons. How these dark subhalos lost their gas and their stars, if they ever had any, is the key to understanding the missing satellites problem. We will investigate gas removal mechanisms, the reason some subhalos are light and some are dark, and why the only 23 subhalos that are luminous are not also the 23 most massive in \S \ref{sec:histhalos}.
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig03.eps}}
\caption{The cumulative mass function of our host galaxy's subhalos at $z=0$, divided into all subhalos in blue, and those with baryons in red. The total number of subhalos is over a hundred, but the total number containing baryons is almost an order of magnitude fewer.}
\label{fig:cumemass}
\end{figure}
\subsection{Effects of Stellar Feedback and Ultraviolet Background: g5664}
\label{sec:g5664}
One test to delve deeper into how the cumulative mass function compares with dark matter-only simulations while the cumulative number of luminous satellites is consistent with observations is to re-simulate a galaxy with and without the UV and stellar feedback. We re-simulated a second, smaller host galaxy, g5664, with a mass of $5.2\times10^{11}$ M$_{\sun}$, three times with different baryonic physics:
\begin{enumerate}
\item[(a)] with the standard MUGS simulation including UV and stellar feedback,
\item[(b)] with UV but no stellar feedback, and
\item[(c)] with neither UV nor stellar feedback.
\end{enumerate}
This galaxy contained half the number of gas and star particles as g15784 and thus was faster to run, which is why it was chosen for the parameter comparison. Since g15784 is more massive and has a larger $r_{vir}$, it will contain more and likely more massive satellites. Since the UV background and stellar feedback more strongly effect low mass galaxies, g15784 will contain more luminous satellites. However, this should not affect the relative comparison of the different resimulations of g5664. The relationship between these mechanisms and the mass of the satellites will be explored in \S \ref{sec:histhalos}, in addition to how these mechanisms affect the satellites throughout their history.
Figure \ref{fig:g5664mag} shows the cumulative luminosity function for the three simulations at $z=0$. It confirms the reduction in number of satellites. There about one-fourth as many in g5664 as in g15784. Nearly every satellite in the simulation run without feedback and UV (c) contains stars. Conversely, many fewer satellites contain stars in the simulation that includes feedback (a). It is apparent that the UV ionisation plays a large role in stopping stars from forming in many subhalos. When stellar feedback is added, there is little effect on subhalos brighter than $M_{V}\approx-15$, but fainter subhalos are only populated with a few stars whose feedback was effective at eliminating star formation for the rest of the simulation. The $M_{V}\approx-15$ threshold is similar to the flattening seen in Figure \ref{fig:cumemag} for both the simulated g15784 and the observed Local Group mass function.
The UV-only simulation (b) contains several medium luminosity subhalos but none that are extremely faint, stopping at $M_{V}\approx-12$. It is curious then that simulation (a), with both feedback and UV, lacks satellites between $-8 < M_{V} < -15$, but contains two very faint satellites at $M_{V} \approx -6$. The two luminosity functions diverge at $M_{V} \approx -15$, which corresponds to about $10^4$ star particles in all three runs of g5664. The origin of this situation is unclear from inspection of the luminosity functions alone. In \S \ref{sec:histhalos}, however, when we track the mechanisms of baryon loss through time an explanation for this phenomenon will arise showing that stellar feedback preferentially strips medium mass satellites.
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig04.eps}}
\caption{Cumulative luminosity function for g5664, comparing runs with three different conditions on the baryonic physics. The run with UV background and stellar feedback is in red (a). The run with UV background, but no feedback is in green (b). The run with neither UV background nor stellar feedback is in blue (c). Both feedback and the UV background reduce the total number of luminous subhalos.}
\label{fig:g5664mag}
\end{figure}
\subsection{Resolution}
\label{sec:res}
Satellite galaxies are the most difficult objects to resolve in simulations, and so they show the strongest resolution effects. Figure \ref{fig:cumemag} begins to show how resolution can affect satellites. It is harder to detect low mass satellites in lower resolution simulations, and effects such as tides and ram pressure stripping are more accurately captured as resolution increases.
Other authors have discussed the resolution effects on their satellite luminosity functions in simulations similar to ours. \cite{Libeskind2007} ran simulations at a lower resolution than ours, studying the satellite systems of several $\Lambda$CDM galaxies with their gas resolved to $\approx10^{6}$ M$_{\sun}$. They compared their luminosity functions to semi-analytic models of resolution over four different orders of magnitude and found a convergence at $M_{V}\approx-12$. \cite{Okamoto2010} studied the effects of several feedback models on the satellite populations on host galaxies around the same mass as our g15784 and with similar gas particle masses. Matching the circular velocity of their satellites to a power law, they concluded that their satellites were well resolved down to satellites with at least 10 star particles. \cite{Wadepuhl2011} compared a low resolution simulation (whose gas particle mass is $\approx10^{5}$ M$_{\sun}$ corresponding to our high resolution simulation) to their high resolution simulation with gas mass of $\approx10^{4}$ M$_{\sun}$ and found them convergent up to $M_{V}\approx-8$.
We compared our high resolution simulation of g15784, with initial gas particle mass $\approx 10^5$ M$_{\sun}$, to a low resolution simulation, with initial gas particle mass $\approx 10^6$ M$_{\sun}$, and found that their luminosity functions were similar in their region of overlap down to $M_{V} \approx -8.2$. The major difference is that at higher resolution there appears to be a slight overabundance of highly luminous subhalos. \citet{Christensen2010} examined the effects of resolution on galaxies ranging from $10^{13}$ to $10^{9}$ M$_{\sun}$ and found that $10^4$ gas particles at each galaxy's maximum mass are required before the star formation recipe used here converges.
Subhalos that formed stars in our high resolution g15784 simulation had their star formation drop off at about $10^4$ star particles at $z=0$, corresponding to the number of gas particles being at least $\approx 5\times10^3$ at maximum mass. The star count corresponds to $M_{V} \approx -15$ in our high resolution simulation. Based on \citet{Christensen2010}, we expect to see a decrease in star formation across this range. However, it is difficult to conclude that the reduction in star formation is solely due to resolution since it corresponds to the mass range between $10^9$ and $5\times10^{10}$ M$_{\sun}$ where feedback has its strongest effects. Our low resolution simulation of g15784 only has one subhalo at $M_{V} \approx -17$ with $\approx 7\times10^3$ star particles, and only two more around $M_{V} \approx -16$ with $\approx 10^3$ star particles. The discrepancy between high and low resolutions is explained by the lack of particles in the low resolution run.
Some of our subhalos contain many fewer particles than \cite{Christensen2010} suggest is necessary to resolve star formation, but we include them for completeness. We will still show and analyze all luminous and dark subhalos with comprehensive histories (as explained early in \S \ref{sec:histhalos}). With higher resolution, \cite{Christensen2010} suggest that these satellites would contain slightly more stars since gas could form a disk and become denser than the star formation threshold. However, we note that it is hard to draw conclusions without the higher resolution simulations because of the non-linear interaction of star formation and stellar feedback, particularly considering that the present simulation over-predicts the number of highly luminous satellites (Stinson et al. 2010).
\section{Gas Loss Mechanisms}
\label{sec:histhalos}
The previous section showed that both stellar and UV feedback play a significant role in the evolution of satellite galaxies. In order to gain an overall picture of how stellar and UV feedback work with ram pressure and tidal stripping, we will focus on g15784 and study the detailed, particle-by-particle mass loss history of its subhalos.
We traced the evolution of 85 of the 107 subhalos identified by AHF in g15784. We were unable to perform a detailed trace of every subhalo for two reasons:
\begin{enumerate}
\item[(1)] 17 low mass subhalos did not maintain 50 member particles throughout the simulation, which is $2.2\times10^{7}$M$_{\sun}$ when the group only contains dark matter but could be less massive if it also contains gas and star particles.
\item[(2)] 5 halos were spatially coincident with another subhalo during one output, and thus appeared to gain a large amount of mass. Since our analysis focused on cumulative baryon loss and the subhalos' maximum mass these sudden spikes in mass would invalidate any results including those halos.
\end{enumerate}
10 of the 85 subhalos we analysed contained both gas and stars at $z=0$, with total masses ranging from $5.6\times10^{8}$~M$_{\sun}$ to $3.3\times10^{10}$~M$_{\sun}$. Of the subhalos that did not have gas at $z=0$, 17 formed stars at some point but only 13 of these retained them until $z=0$. This leaves a total of 23 luminous subhalos at $z=0$, of which 13 have more than 50 baryon particles.
Figure \ref{fig:timeevall} shows examples of the time evolution of these subhalos' mass and distance to the host at each output, representing an upper limit on the subhalos' closest distance to the host. The top panel shows subhalo (a) that retains gas and stars at $z=0$, the middle panel shows subhalo (b) that retains stars but not gas at $z=0$, and the bottom panel (c) shows a dark satellite that has no baryon particles at $z=0$. Close passages to the center of the main halo most strongly removes gas and to a lesser extent, dark matter and stars through tidal stripping. Stars tend to sit at the centre of the subhalo and are less vulnerable to stripping than the dark matter around the subhalo's exterior. That is, subhalos that form stars before they lose their gas retain those stars until $z=0$. If stars are not formed before their first close passage then the subhalo will never form stars and becomes a dark satellite. The quantity of gas lost is also determined by the subhalo's proximity to the host. For example the luminous first and second subhalos (a) and (b) start out with similar mass, but the latter has a pericentre that is more than twice as close to the host and as a result, it loses all its gas by $z=0$. Central location is not the full story. The dark satellite (c) is farther from the host galaxy than either of the luminous ones and has a higher mass at $z=0$ than the subhalo (b), but it never forms stars and its mass in gas is much lower before being lost early in the simulation. There are more factors than a subhalo's mass at $z=0$ and its proximity to the host galaxy that decide if it is luminous at $z=0$, which we will explore in this section.
\begin{figure}
\centering
\resizebox{0.3\textheight}{!}{\includegraphics{fig05.eps}}
\caption{The time evolution of the distance to the host and the mass of different components of three subhalos. The black line corresponds to the distance to the host at each output on the right-hand axis on each plot. The coloured lines correspond to the masses on the left-hand axis: dark matter (blue), gas (red) and stars (green). The top subhalo (a) is massive enough to retain both gas and stars to $z=0$ (final mass: $3.9\times10^{9}$~M$_{\sun}$). The centre subhalo (b) retains stars but not gas at $z=0$ (final mass: $2.0\times10^{8}$~M$_{\sun}$). The bottom subhalo (c) loses all its baryons before $z=0$, making it a dark satellite (final mass: $7.2\times10^{8}$~M$_{\sun}$) despite the fact that it ends up heavier than the luminous subhalo in the centre panel.}
\label{fig:timeevall}
\end{figure}
\subsection{Ultraviolet Background}
Quantifying how much gas a subhalo loses due to the UV background involves some degree of subjectivity. Instead of merely counting gas that was in the subhalo, it will involve determining how much gas was never in the subhalo but ought to have been.
We confine our analysis of the evolution of subhalos to the time after $z=10.8$, when the first halos get larger than the minimum group finder particle limit and are virialized. One of the earliest mechanisms that removes gas from subhalos is the reionising UV background radiation that is emitted from the first luminous objects. This radiation makes its first impact at $z=9.9$ in these simulations.
Most of the subhalos cannot be identified until a few outputs after $z=9.9$. To enable the analysis of the effects of UV radiation, every dark matter particle within $r_{vir}$ is matched with a ``twin'' gas particle at the initial conditions. In our halo-by-halo analysis, dark matter twin particles are defined as subhalo members at the time their subhalo reaches its maximum mass. The evolution of the gas twins of these member particles is then traced from the earliest output onwards. The ensemble of the twin gas particles are referred to as the ``background gas'' later in this section.
The mass lost due to reionisation is defined as the gas that had a dark matter twin in a given subhalo, but itself was never contained in that subhalo. The reason that this gas was not in the subhalo is that subhalos are unable to contain gas with a temperature higher than the subhalo virial temperature ($T_{vir}$). Low mass subhalos will thus contain dark matter without its gas twin. Note that the twins are calculated from dark matter in the subhalo at the time of the subhalo's maximum mass and not reionisation. This is important because the way subhalo tracing works, only one subhalo is labelled as ``the'' subhalo at any given timestep. However, the larger subhalos reach their maximum mass as a result of the merger of several smaller subhalos. Determining the UV loss at a subhalo's maximum mass accounts for loss in each of the individual subhalos. Figure \ref{fig:tmmax_ss} shows that satellites reach their maximum mass prior to being captured by the main halo, typically immediately before.
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig06.eps}}
\caption{For all our subhalos, the time of maximum mass versus the time they join the main halo. Most of them reach their maximum mass shortly before becoming substructure. The dotted line marks where the time of maximum mass is equal to the substructure time, showing that none of the subhalos achieve their maximum mass after becoming substructure.}
\label{fig:tmmax_ss}
\end{figure}
Subhalos that did not virialize before reionisation generally do not contain gas. The subhalos that virialize after reionisation, but contain gas that goes on to form stars, do typically have gas without a dark matter twin in the subhalo. In the face of such a mismatch between gas and dark matter particle members we developed the following analysis to justify why the amount of background gas that ends up in the subhalo correlates to UV loss.
In order for a gas particle to have enough energy to overcome the potential of the subhalo and escape, its temperature must be greater than $T_{vir}$, where:
\begin{equation}
T_{vir}=\frac{2G\mu m_{p} M_{subhalo}}{3kR_{subhalo}}
\label{eqn:tvirial}
\end{equation}
where $G$ is the gravitational constant, $\mu$ is the mean molecular weight (where the typical value is 0.6 for ionised gas), $m_{p}$ is the proton mass, $k$ is Boltzmann's constant, and $M_{subhalo}$ is the subhalo's mass. $R_{subhalo}$ is the distance between the halo's centre of mass and the most distant member particle.
Figure \ref{fig:tempev} shows the evolution of the background mean temperature (twin particles) and $T_{vir}$ in halos (b) and (c) from Figure \ref{fig:timeevall}. The subhalo in the top panel (b) retains its baryons through reionisation because of its higher mass and consequently higher T$_{vir}$. While halo (c) is more massive at $z=0$ than (b), it is dark because the twin gas of its dark matter was hotter than $T_{vir}$. The subhalos that form earliest capture the most gas and form the most stars.
\begin{figure}
\centering
\resizebox{0.3\textheight}{!}{\includegraphics{fig07.eps}}
\caption{Evolution of background gas mean temperature of the background twin gas (blue) and $T_{vir}$ as given by Equation \ref{eqn:tvirial} (red) for subhalos (b) and (c) from Figure \ref{fig:timeevall}. Temperature corresponds to the left axis, and the subhalos' orbits (black, solid) to the right. The vertical dashed line corresponds to $z=9.9$ when the UV background is turned on. Even though the subhalo on the bottom is more massive at $z=0$, it is unable to retain as much gas as the lower mass subhalo on the top due to the fact that its background gas is consistently hotter than $T_{vir}$.}
\label{fig:tempev}
\end{figure}
Figure \ref{fig:bfractemp} shows how the ratio of $T_{vir}$ with the mean gas twin temperature effects the final baryon fraction. Specifically, we compare the maximum ratio of $T_{vir}$ to background gas mean temperature versus the baryon fraction at two times: $z=0$ and the time at which this temperature ratio is a maximum. The subhalos with T$_{vir}$/T$_{background} > 1$ at some point are more likely to be near the cosmic baryon fraction. With one exception, the subhalos that contain baryons at $z=0$ are a subset of these. The one exceptional satellite started out near to the host and gathered enough gas early in its formation to form enough stars to remain luminous over its subsequent several close passages to the host, even though it lost all its gas on its second passage. Since none of the subhalos that never contained their twins held onto any of their baryons, it appears that the ``twin'' particle analysis is robust. Therefore, we can safely define gas lost due to UV background as the mass of twin gas particles that never entered subhalo.
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig08.eps}}
\caption{The baryon fraction versus the maximum ratio between the virial and background gas temperatures that a subhalo ever obtains over its lifetime. The horizontal dashed black line represents the cosmic mean, 0.17, while the vertical dashed black line separates the subhalos that achieved a $T_{vir}$ higher than the background gas temperature on the right from those that did not on the left. The asterisks represent subhalos that have ten or more baryon particles at $z=0$, while Xs represent subhalos that have fewer than ten baryon particles at $z=0$, corresponding to the resolution limit found in Figure \ref{fig:cumemag}. Red shows the subhalos' baryon fraction at the time of their maximum ratio of $T_{vir}/T_{background}$, while blue shows the subhalos' baryon fraction at $z=0$. With one exception, every subhalo that is luminous at $z=0$ is to the right of the vertical line.}
\label{fig:bfractemp}
\end{figure}
\subsection{Ram Pressure Stripping}
One of the difficult aspects of this study is that the mass loss mechanisms involve the interaction between two gas phases with significantly different properties. This circumstance is one where SPH struggles. \citet{Agertz2007} showed that SPH has trouble modelling ram pressure stripping, particularly when the Kelvin-Helmholtz time is important. However, \citet{Mayer2006} showed that SPH can model stripping when $\tau_{\mathrm{dyn}} < \tau_{\mathrm{KH}}$, and our g15784 satellites typically fall into this regime (\citealt{Mayer2006} show that $\tau_{\mathrm{KH}} \ga 4$~Gyr when they reach their minimum at pericentre, compared to satellite dynamical times of $\tau_{\mathrm{dyn}} \approx 0.2$--$6$~Gyr).
With these caveats in mind, we do a classical ram pressure analysis \citep{Gunn1972} to see how close these simulations come to reality. In order to quantitatively measure the effect ram pressure stripping has on our subhalos, we use the criterion from \citet{Grebel2003}:
\begin{equation}
P_{ram} \approx \rho_{hhg} v_{subhalo}^{2} > \frac{\sigma_{subhalo}^{2}\rho_{gas}}{3}
\label{eqn:pram}
\end{equation}
where $P_{ram}$ is the ram pressure, $\rho_{hhg}$ is the gas density in the hot halo gas around the subhalo, $v_{subhalo}$ is the subhalo's velocity relative to the host galaxy, $\sigma_{subhalo}$ is the velocity dispersion of the gas in the subhalo, $\rho_{gas}$ is the average density of the gas in the subhalo. Here the subhalo's velocity dispersion is defined as:
\begin{equation}
\sigma_{subhalo}^{2}=\frac{3}{5}\frac{GM_{subhalo}}{R_{subhalo}}
\label{eqn:veldisp}
\end{equation}
where $M_{subhalo}$ is the subhalo's mass, and $R_{subhalo}$ is the subhalo's radius.
The gas density of the hot halo gas around each halo is defined as the average density of the $n$ nearest gas particles, where $n$ is twice the number of particles in the subhalo to a maximum of 4000. The density and temperature structure of the gaseous halo of g15784, through which the subhalos pass, is shown in Figure~\ref{fig:gden} with substructure removed.
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig09.eps}}
\caption{The gas temperature (red) and density (blue) profiles of g15784 at $z=0$ for the outer disk at $r \ge 30$kpc. Power laws fit to $\rho \propto r^{-1.7}$ (dashed purple) and $T \propto r^{-0.73}$ (dashed orange) are shown for reference.}
\label{fig:gden}
\end{figure}
Every gas particle that leaves during the timesteps when the ram pressure exceeds the internal pressure of the halo is classified as leaving due to ram pressure stripping except for those particles that qualified for stellar feedback (\S\ref{sec:sf}).
However, we also generated movies of twelve low-mass subhalos and visually inspected them to see if ram pressure or tides removed their gas. Figure \ref{fig:ram013} gives an example of a subhalo that loses its gas due to ram pressure. As a subhalo approaches the host galaxy, it enters the hot halo gas. In low enough mass subhalos, its gas gets left behind, cleanly separating from its dark matter. Figure \ref{fig:ram013} shows that the gas maintains the shape of the subhalo for a several tens of Myrs. This contrasts with the signature of tidal stripping where tidal forces elongate the matter ahead and behind the satellite's orbit. Typically, ram pressure stripping occurs farther out from the host than tidal stripping.
\begin{figure}
\centering
\begin{tabular}{cc}
\resizebox{0.45\linewidth}{!}{\includegraphics{fig10a.eps}} &
\resizebox{0.45\linewidth}{!}{\includegraphics{fig10b.eps}}
\end{tabular}
\caption{How ram pressure stripping removes gas from a $1.7\times10^{9}$~M$_{\sun}$ subhalo between $z=0.26$ (left) and $z=0.16$ (right). The dark matter (dark green), gas (light green) and stars (turquoise, none present here) are marked at the time of maximum mass, while the bottom layer (brown, covered the other layers) is the subhalo at the present output. Ram pressure stripping has affected this subhalo, removing all of the gas it pocessed at the time of maximum mass, but without strengthening the ram pressure stripping diagnosis by a factor of 10 the individual gas particles would not be marked as ram pressure stripped. The background colours denote the temperature of the surrounding gas, ranging from dark blue (colder, $10^{1.5}$~K) through to white (hotter, $10^{7}$~K).}
\label{fig:ram013}
\end{figure}
Our definition for ram pressure underestimated the effect that is visible in the simulations. This seems to be a numerical effect and requires a multiplication of the calculated ram pressure by a factor of 10. The factor of 10 may compensate for the use of the mean satellite gas pressure instead of the pressure in the outer regions where particles are getting removed. Figure \ref{fig:presev} shows how the subhalo's gas pressure and ram pressure evolves in two subhalos. The factor of 10 increases the ram pressure so that it is comparable with the subhalo's gas pressure. In the future, a comparison of higher-resolution simulations, as well as grid codes, to ours would be useful to see whether they exhibit results closer to the analytic determination.
\begin{figure}
\centering
\begin{tabular}{cc}
\resizebox{0.45\linewidth}{!}{\includegraphics{fig11a.eps}} &
\resizebox{0.45\linewidth}{!}{\includegraphics{fig11b.eps}}
\end{tabular}
\caption{Pressure versus time for subhalos (b) and (c) from Figure \ref{fig:timeevall}. Internal pressure is given by the solid red line, ram pressure from the hot halo gas is the blue dashed line, and ram pressure strengthened by a factor of 10 is the green dashed line. Visual inspection of the subhalos revealed that the one in the left panel was affected both by ram pressure stripping and later tidal stripping, while the subhalo on the right was affected predominantly by ram pressure stripping.}
\label{fig:presev}
\end{figure}
\subsection{Tidal Stripping}
Next, we examine mass loss due to tidal forces using a simple spherical approximation for the host halo and the satellite. For each subhalo, the tidal force particles feel from the host is compared with the force they feel from their internal gravity. The subhalo's tidal radius is the place at which its self gravity is less than the tidal force of the host galaxy (e.g \citet{Hayashi2003}). Due to Newton's Theorem, assuming spherical symmetry, one need only consider the mass interior to a given particle. The gravitational force on the particle from inside the particle's orbit is
\begin{equation}
F_{subhalo}=\frac{GM_{subhalo}(r)m_{particle}}{r^{2}}
\label{eqn:halograv}
\end{equation}
where $r$ is the distance from the particle to the centre of its subhalo, $m_{particle}$ is the mass of the particle, and $M_{subhalo}(r)$ is the subhalo's mass interior to $r$. The tidal force that the particle feels from the host, if the host galaxy is approximated by a point mass and assuming all its mass is contained inside the satellite, is the differential pull between the particle's position in the satellite and the satellite's centre:
\begin{equation}
\delta F_{tidal}=\frac{-2GM_{host}m_{particle}r}{R_{host}^{3}}
\label{eqn:halotide}
\end{equation}
where $R_{host}$ is the distance between the subhalo and the host, and $M_{host}$ is the mass of the host.
Therefore, the condition for when the particle feels a greater tidal force than gravitational from its own subhalo is given by:
\begin{equation}
\frac{M_{subhalo}(r)}{r^{3}}<\frac{2M_{host}}{R_{host}^{3}}
\label{eqn:tidecon}
\end{equation}
A particle that passes this test qualifies for tidal stripping. We emphasize that this a spherical approximation, given that the subhalos occasionally have their shape distorted, though the distortion is usually symetrical. We did try varying the strengh of the tidal force, as we had done for ram pressure stripping, but found that it did not change results much and so this method has some degree of robustness.
Figure \ref{fig:tide017} shows an example of a subhalo being tidally stripped of its gas. The tidal force pulls material out in leading and trailing arms. Additionally, material is stripped off the outside first before the material on the inside.
Regarding the possibility of tidal stirring, Figure \ref{fig:timeevall} shows that there was not significant star formation following close passages of satellites. During these close passages, dark matter is often stripped and a large fraction of gas is stripped. It is possible that these simulations are too low of a resolution to model tidal stirring like was seen in \citet{Mayer2006}.
\begin{figure}
\centering
\begin{tabular}{cc}
\resizebox{0.45\linewidth}{!}{\includegraphics{fig12a.eps}} &
\resizebox{0.45\linewidth}{!}{\includegraphics{fig12b.eps}}
\end{tabular}
\caption{How tidal stripping removes gas from a $2.0\times10^{8}$~M$_{\sun}$ subhalos between $z=1.0$ (left) and $z=0.8$ (right). The dark matter (dark green), gas (light green) and stars (turquoise) are marked at the time of maximum mass, while the bottom layer (brown, covered the other layers) is the subhalo at the present output. Note how the tidal stripping has already begun at $z=1.0$, and by $z=0.8$ the tidal tail is prominent. The dark matter has been compressed, while the stars sit more safely in the subhalo's centre. The background colours are gas temperature as in Figure \ref{fig:ram013}.}
\label{fig:tide017}
\end{figure}
\subsection{Stellar Feedback}
\label{sec:sf}
Often, much credit for the removal of baryons is given to stellar feedback \citep{Dekel1986, MacLow1999, Dekel2003}. Supernovae release large amounts of energy into the ISM, which can be sufficient to liberate gas from halo potential wells. Others have argued that the coupling between the stellar feedback and the ISM is insufficient to remove a significant amount of gas from subhalos. Determining how much gas stellar feedback removed in the simulations proved to be a challenging task.
One signature of stellar feedback in these simulations is that a gas particle has its cooling turned off, which allows it to maintain its high temperature due to the stellar energy release. When it gains sufficient kinetic energy, it can escape the gravitational potential of the subhalo. We found this method of tracking stellar feedback to be very limiting, however. Outputs were limited to approximately one every 200 to 100 Myr. Cooling is typically shut off for 50 Myrs, so between $1/2$ and $3/4$ of particles whose cooling was turned off would be missed by simply counting particles whose cooling was shut off during an output.
Another signature of stellar feedback is the release of metals into the surrounding interstellar medium. In our simulations, a star particle releases about $300 $ M$_{\sun}$ of metals from type II supernovae to the nearest 32 gas particles. The ejection is smoothed so that gas closer to the star receives more metals than particles further away. A typical gas particle will receive a few M$_{\sun}$ in metals from a star particle. Gas can also receive metals from other gas particles through diffusion. These metal transfers are typically $<1$ M$_{\sun}$.
So, when gas had an increase in metals of $\ge 5$ M$_{\sun}$, it is likely that it was in the neighborhood of stellar feedback and we classify it as having been lost from the halo due to stellar feedback.
This is a conservative estimate because there may also have been cases where gas directly heated by stellar feedback acquired sufficient pressure to push out different gas, a process called ``mass-loading''. This is common in dwarf galaxies \citep{Vacchia2008}. Stellar feedback could have augmented another mechanism like ram pressure stripping and gotten unlabelled as such.
\subsection{A Combination of Mechanisms}
\label{sec:combomech}
Using all the techniques described above, we now present a summary of which processes dominated accretion and loss of gas.
One confounding effect happened when massive subhalos pass through the pericentre of their orbit. Subhalos temporarily accreted a small quantity of gas and quickly lose it, possibly a numerical effect. Because of this, we did not count mass loss of particles that entered and left subhalos after they reached their maximum mass.
Some particles left the subhalo more than once. In order to avoid double counting, only the last method by which a particle entered or left its subhalo was counted. We noted which gas particles were converted into stars, so that we could identify what portion of the gas mass decrement was due to star formation and what portion was due to gas leaving the subhalo. The loss of many particles could not be classified, so mass loss is often classified as ``other'', emphasizing the schematic nature of this method.
Figures \ref{fig:mlosptcla} to \ref{fig:mlosptclc} show the evolutionary history of the same subhalos as shown in Figure \ref{fig:timeevall}. These histories are coloured to indicate the amount of gas lost due to each of the mechanisms described above, as a fraction of each subhalo's maximum mass. Most of the baryons for the most massive subhalo (a) turn into stars quickly, and, other than the initial UV ionisation which prevents a significant amount of gas from being captured, any gas that is lost is usually due to tidal stripping or stellar feedback. After UV ionisation the medium mass subhalo (b) that retains its stars loses its gas largely due to stellar feedback with smaller contributions by tidal and ram pressure stripping, though there is a large number of unclassified particles as well. The medium mass subhalo (c) that ends up as a dark satellite lost almost all of its gas due to the UV ionisation, with the small remainder being stripped by ram pressure.
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig13.eps}}
\caption{Mechanisms for baryon loss over time for the massive subhalo (a) (maximum mass: $7.4\times10^{9}$~M$_{\sun}$, final mass: $3.9\times10^{9}$~M$_{\sun}$) from Figure \ref{fig:timeevall}. Turquoise is the cumulative gas lost due to the UV background, red is the cumulative gas lost due to ram pressure stripping, dark green is the cumulative gas lost due to tidal stripping, light green the is the cumulative gas lost due to stellar feedback, purple is the cumulative gas lost due to undetermined causes, orange is the cumulative lost stars, blue is the gas at the current timestep, and yellow is the stars at the current timestep.}
\label{fig:mlosptcla}
\end{figure}
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig14.eps}}
\caption{Mechanisms for baryon loss over time for the medium mass subhalo (b) (maximum mass: $4.2\times10^{9}$~M$_{\sun}$, final mass: $2.0\times10^{8}$~M$_{\sun}$) from Figure \ref{fig:timeevall}. The colour scheme follows Figure \ref{fig:mlosptcla}.}
\label{fig:mlosptclb}
\end{figure}
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig15.eps}}
\caption{Mechanisms for baryon loss over time for the medium mass subhalo (c) (maximum mass: $1.4\times10^{9}$~M$_{\sun}$, final mass: $7.2\times10^{8}$~M$_{\sun}$) from Figure \ref{fig:timeevall}. The colour scheme follows Figure \ref{fig:mlosptcla}.}
\label{fig:mlosptclc}
\end{figure}
Figure \ref{fig:mlosptclall} combines all plots of type Figures \ref{fig:mlosptcla} to \ref{fig:mlosptclc} for all subhalos and shows as a fraction of each subhalo's maximum mass the gas lost due to each mechanism at $z=0$. These mass loss fractions are plotted as a function of maximum mass rather than final mass because the sequence of mechanisms appears more clearly (as shown in Figure \ref{fig:msmtot}). Figure \ref{fig:mlosptclall} shows that the massive subhalo (a) in Figure \ref{fig:mlosptcla} is no aberration. It is common for the most massive subhalos to efficiently form stars and for tidal stripping to play the most important role in their mass loss. Medium mass subhalos tend to be more dominated by stellar feedback, since they are massive enough to form stars but light enough that they are more susceptible to losing their gas. Following Figure \ref{fig:mlosptclc}, lower mass subhalos lose significant mass due to UV reionisation and then much of the remaining mass is stripped by ram pressure.
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig16.eps}}
\caption{Mechanisms for baryon loss at $z=0$ for all subhalos, as a function of the maximum mass that a subhalo was able to achieve. The colour scheme follows Figure \ref{fig:mlosptcla}. The various loss mechanisms are cumulative over time, while the total gas and stars are for the current time. All values are a fraction of the subhalo's maximum mass. The luminous subhalos show an excess of baryons over their lifetime, above the cosmic mean of 0.17. The cutoff between luminous and dark satellites happens at about $2.0\times10^{9}$~M$_{\sun}$.}
\label{fig:mlosptclall}
\end{figure}
A dichotomy of evolutionary scenarios appears in Figure \ref{fig:mlosptclall}. Halos less massive than $\approx2.0\times10^{9}$~M$_{\sun}$ lose their gas due mostly to UV reionisation. Higher mass halos formed stars, lost less mass to reionization, and lost gas due to a variety of other mechanisms. This distinction disappears when the satellites are classified by their final mass as in Figure \ref{fig:msmtot} where the populations of baryonless and luminous subhalos overlap in terms of total mass at $z=0$.
Table \ref{tab:mechpercent} summarizes the results of Figure \ref{fig:mlosptclall} by dividing the subhalos into three categories (ones with gas and stars at $z=0$, ones without gas at $z=0$ but had stars, and ones that never formed stars and have no gas at $z=0$). UV ionisation is the most prominent for all subhalos. For the most massive subhalos tidal stripping followed, while stellar feedback and ram pressure stipping had little impact. For the subhalos massive enough to form stars at some point but which did not retain gas at $z=0$, after UV ionisation, stellar feedback was the most prominet mechanism, while tidal and ram pressure stripping were close in magnitude. Finally, UV ionisation was the most important for the subhalos that never formed stars, with some impact from ram pressure stripping. Tidal stripping and stellar feedback were negligible. Note the caveat: since it is impossible in all cases to clearly distinguish the mechanism that leads to the loss of a gas particle, the boundaries between the mechanisms are not clearly defined and the percentages should, therefore, be take as indicative of the relative importance of the various gas loss mechanisms.
\begin{table}
\centering
\begin{tabular}{l|lll}
\hline
At $z=0$: & no gas and & no gas but & gas \\
& never stars & have/had stars & and stars \\ \hline \hline
Min Mass (M$_{\sun}$)& $1.10\times10^{8}$ & $7.79\times10^{8}$ & $3.14\times10^{9}$ \\
Max Mass (M$_{\sun}$)& $2.09\times10^{9}$ & $4.30\times10^{9}$ & $6.03\times10^{10}$ \\ \hline
UV (\%) & 15.73 & 11.19 & 6.33 \\
Ram (\%) & 1.10 & 1.54 & 0.26 \\
Tides (\%) & 0.12 & 1.87 & 1.93 \\
Sfb (\%) & 0.01 & 2.08 & 0.64 \\
\hline
\end{tabular}
\caption{The cumulative effect of each gas-loss mechanism (UV ionisation, ram pressure stripping, tidal stripping, and stellar feedback) given as an approximate percentage of the subhalos' maximum mass, averaged over all the subhalos in each category. The first category are the subhalos that had no gas at $z=0$ and never formed stars (represented by subhalo (c)), the second are those that have no gas at $z=0$ but had stars at some point in their history (represented by subhalo (b)), and the last category are those subhalos that retain both gas and stars at $z=0$ (represented by subhalo (a)). Also included are the minimum and maximum masses of the subhalos in each category.}
\label{tab:mechpercent}
\end{table}
For the lower mass subhalos, the amount of mass lost adds up nearly to the cosmic baryon fraction ($\approx 0.17$), in part because our analysis relied on pairing dark and gas particles. However, not only do the higher mass subhalos contain more baryons than the cosmic fraction, they contain more \emph{stars} than the cosmic fraction. To understand how the higher mass subhalos form so many stars, we investigated the origin of these stars. Figure \ref{fig:twintrack} shows the mass evolution of a $7.1\times10^{9}$~M$_{\sun}$ subhalo that ends up with more than the cosmic baryon fraction in stars. The mass evolution is divided into categories based on whether the particles were twins of the dark matter present at the maximum mass. While most of the stars formed from gas that was a twin of this dark matter, almost 10\% of the stars formed from gas that were twins of dark matter that were not members of this halo at its time of maximum mass, or any of the outputs immediately before and after the time of maximum mass.
\begin{figure}
\centering
\resizebox{0.45\textwidth}{!}{\includegraphics{fig17.eps}}
\caption{Time evolution of a $7.1\times10^{9}$~M$_{\sun}$ mass subhalo's matter, broken down into dark matter present (twin) and not present (not twin) at the time of maximum mass, gas twinned or not from the maximum mass dark matter, and stars that did and did not come from twinned gas. The dashed vertical line is the time of maximum mass. Despite the difference in magnitudes between the twin and non-twin dark matter, the twin and non-twin gas particles are comparable, suggesting that the dark matter draws on gas outside its region of origin.}
\label{fig:twintrack}
\end{figure}
Figure \ref{fig:snotgorg} shows how this extra gas (marked as light green) comes from a much wider region than the dark matter (marked as brown). While such accretion could be a numerical artifact of over-efficient gas cooling, it could also be a unique feature of satellites that orbit in high-density regions like a massive galaxy's hot halo. The mechanism that appears in the simulations is that high mass satellites quickly form stars out of gas that are twins of member dark matter particle. After they form stars and the stellar feedback cools down, there is less gas to provide pressure support to keep hot gas from the main halo out of the satellite. So this gas is accreted, cooled, and finally forms stars.
\begin{figure}
\centering
\begin{tabular}{cc}
\resizebox{0.45\linewidth}{!}{\includegraphics{fig18a.eps}} &
\resizebox{0.45\linewidth}{!}{\includegraphics{fig18b.eps}}
\end{tabular}
\caption{Snapshots at $z=6$ of a $7.1\times10^{9}$~M$_{\sun}$ mass subhalo (left) and a $2.4\times10^{9}$~M$_{\sun}$ mass subhalo (right). The subhalos are in brown, while light green is the gas that will produce all the non-twin stars that will end up in the subhalos. This gas will eventually converge into the subhalos. The subhalo on the left will end up in a tight orbit around the host, and hence the several shells of gas that it will draw upon. The background colours are gas temperature as in Figure~\ref{fig:ram013}.}
\label{fig:snotgorg}
\end{figure}
\section{Conclusions and Discussion}
\label{sec:conclu}
To gain insight into the missing satellites problem, we compared the satellite luminosity functions of two simulated galaxies from the MUGS project (g5664 and g15784) with late-type galaxies.
The cumulative number of luminous satellites in g15784 was only slightly higher than that observed in the Milky Way, though there were an excess of high luminosity satellites that created a ``knee'' in the satellite luminosity function that is not observed. Other SPH simulations of similar or lower resolution to ours have found that the missing satellites problem is no longer a matter of an order magnitude difference between the Local Group and simulated subhalo populations. When we compared our luminosity function of g15784 to a simulation of the same galaxy at a lower gas resolution both luminosity functions were similar in their area of overlap down to $M_V \approx -8.2$, though the low resolution run did suffer from having fewer gas particles than needed to properly resolve star formation and therefore it had fewer stars. A couple of our dwarfs had luminosities comparable to the recently discovered ultra-faint dwarfs, but at the resolution of these simulations, they had only one or two star particles, so it is impossible to draw any conclusions about the formation of fainter dwarf galaxies from these simulations.
The satellite mass function of g15784 revealed a large population of dark satellites. In our more massive galaxy g15784 ($1.4\times10^{12}$~M$_{\sun}$) the subhalos constituted 6.0\% of the host galaxy's mass, while g5664's subhalos were 4.4\% of the host galaxy's mass. This fits within the range that \citet{Dalal2002} found from probing substructure with gravitational lensing, between 0.6\% and 7.0\%.
We used two methods to determine how the dark satellites lost their baryons and the effect of negative feedback on the luminous satellites.
One method was to simulate the less massive galaxy than the Milky Way, g5664, using several different physical treatments. The simplest included no UV or stellar feedback. In the second, UV was added. These two simulations were compared with the standard MUGS simulation that included both UV and stellar feedback. The effect they had on the subhalo populations was significant. UV feedback alone stopped star formation in all the satellites with total masses less than $2\times10^{9}$~M$_{\sun}$. When stellar feedback was added, it reduced the luminosity of several additional subhalos so that only a couple of star particles formed in those subhalos before the feedback ejected all the remaining gas from the subhalos and eliminated the possibility of future star formation. In more massive subhalos, the stellar feedback had little impact on reducing the star formation efficiency. This unbalanced influence of the stellar feedback may have been due in part to the quantized feedback that was used in the MUGS simulations.
The second method was to analyse the individual evolution of satellites in a more massive halo. We made a comprehensive study of the mechanisms that remove matter from subhalos by defining criteria for mass loss due to the UV background, tidal stripping, ram pressure stripping, and stellar feedback. This analysis reiterated the strong impact ionisation had on low mass satellites. We used metals to track stellar feedback, and found that its impact was largest on subhalos of medium mass that had formed stars, with lesser impact on the highest mass subhalos, and no impact on the lower mass subhalos.
A strange phenomenon was apparent in the higher mass subhalos. These subhalos contained a higher fraction of their maximum mass in stars than the cosmic baryon fraction. Subsequent analysis showed that accretion of baryons was not confined to the same limited region from which dark matter was accreted, but from a larger region surrounding the subhalo and even from across the hot gaseous host halo as the subhalo moved through its orbit. While this may partially be another symptom of overcooling that has been long noted in simulations, it may also point to the enhanced baryonic accretion possible by subhalos in high density regions.
Stripping, either ram pressure or tidal, also plays a vital role in shaping the satellites that were analysed. Ram pressure removed whatever gas remained in small subhalos that had most of their gas removed during ionisation. In some cases, more gas was stripped from subhalos than would have been predicted based on a \citet{Gunn1972} analysis. Tidal stripping removed most of the gas from the more massive satellites, making them comparable with Local Group dSphs rather than dIrrs. The stripping became apparent once the subhalos crossed inside the virial radius.
Tidal stripping was important for the subhalos that had the closest encounters with the main galaxy. In some cases, tidal stripping removed enough of the outer layers of dark matter that the total mass of the satellites dropped below the ionisation mass limit of $2\times10^{9}$~M$_{\sun}$. Because tidal stripping reduces the total mass of subhalos by different amounts, it is critical to organize the satellites by their maximum mass rather than their mass at $z=0$ to see a continuous behavior in the baryon fraction as a function of mass. Many authors have noted the similarity in mass inferred in Local Group dSphs \citep{Bullock2000,Strigari2007,Pen2008} by extrapolating satellite total masses using NFW density profiles. Our simulations point out that because of tidal stripping, these satellites may no longer contain that much mass. However, those extrapolated masses may be similar to the maximum mass of the satellite, and the constant lower mass limit suggests a mass-dependent gas removal mechanism like ionisation. What is not clear in the simulations is why the efficiency of star formation varies so much from dSph down to ultra faint galaxies if they did form from subhalos that were all the same mass.
We should also extend this work to include all MUGS galaxies to determine if there are specific factors in the environment or history of individual galaxies that affect the satellite population, as well as compare the subhalo populations outside of the main halos to the ones that end up within the virial radius of their main halos.
\section*{Acknowledgements}
We thank SHARCNET for generously providing supercomputers without which the MUGS galaxies would not be possible, NSERC for funding, and the referee for helpful comments. HMPC thanks the Canadian Institute for Advanced Research for support. GSS was a CITA National Fellow during part of this study and a Jeremiah Horrocks fellow during the rest of the study.
\bibliographystyle{mn2e}
|
\section{Introduction}
The method of Very Long Baseline Interferometry (VLBI) first proposed
by \citet{r:mat65} allows us to derive the position of sources with
nanoradian precision (1 nrad $\approx$ 0.2~mas). The first catalogue of
source coordinates determined with VLBI contained 35~objects
\citep{r:first-cat}. Since then, hundreds of sources have been observed under
geodesy and astrometry VLBI observing programs at 8.6 and 2.3~GHz
(X and S bands) using the Mark3 recording system at the International VLBI
Service for Geodesy and Astrometry (IVS) network. Analysis of these
observations resulted in the ICRF catalogue of 608~sources \citep{r:icrf98}.
Later, over 6000 sources were observed in the framework of the VLBA Calibrator
Survey (VCS) program \citep{r:vcs1,r:vcs2,r:vcs3,r:vcs4,r:vcs5,r:vcs6}, the VLBA
regular geodesy RDV program \citep{r:rdv}, the VLBA Imaging and
Polarimetry Survey (VIPS) \citep{r:vips,r:astro_vips}, the VLBA Galactic plane
Survey (VGaPS) \citep{r:vgaps}, the on-going Australian Long Baseline Array
Calibrator Survey (LCS) \citep{r:lcs1}, and several other programs.
The number of extragalactic sources with positions determined from
analysis of observations under absolute astrometry or geodesy programs
reached 6455 by June 2011, and it continues to grow rapidly due
to analysis of new observations and an on-going campaign of in depth
re-analysis of old observations.
The catalogue of positions of all these compact extragalactic radio sources
determined with VLBI\footnote{Available at \web{http://astrogeo.org/rfc}.}
with accuracies in a range of 0.05--30~mas forms a dense grid on the sky
that can be used for many applications, such as differential astrometry,
phase-referencing VLBI observations of weak objects, space navigation,
Earth orientation parameter determination, and space geodesy.
To date, this position catalogue is the most precise astrometric catalogue.
However, this high accuracy of positions of listed objects can be exploited
{\it directly} only by applications that utilize the VLBI technique.
Applications that use different observational techniques can benefit from
the high accuracy of VLBI positions only {\it indirectly} by observing common
objects from the VLBI catalogue with instruments at other wavelengths.
For last three decades significant efforts were made for connecting
the VLBI position catalogue and existing optical catalogues made with
the use of ground instruments. An overview of the current status of
radio-optical connection and detailed analysis of the differences between
VLBI and optical source positions can be found in \citet{r:lfrq}.
According to them, the standard deviation of the differences between
the VLBI and optical catalogues is $\sim\!\!130$~mas. It was
shown by \citet{r:zah08} that when modern dedicated ground-based
observations are used, the differences are close to 30~mas.
This level of agreement between VLBI and optical positions roughly
corresponds to the position accuracy of common objects from ground optical
catalogues, typically at a level of 100~mas. The European Space Agency
space-born astrometry mission Gaia, scheduled to be launched in 2013,
according to \citet{r:gaia} promises to reach sub-mas accuracies
of determining positions of quasars of 16--20 magnitude that will rival
accuracies of absolute astrometry VLBI. Since position catalogues produced
with Gaia and VLBI will be completely independent, their mutual rotations,
zonal differences and possibly other systematic effects can be interpreted
as errors of one of the techniques after resolving the differences
due to a misalignment of centers of optic and radio images of quasars
and a frequency-dependent core-shift \citep{r:kov08,r:por09,r:sokol11}.
Investigation of systematic differences will be very important for the
assessment of the overall quality of Gaia results and, possibly, the errors
in the VLBI position catalogue.
This comparison will produce valuable results if 1)~it will be
limited to those common sources which VLBI positions are known with errors
smaller than several tenths of a milliarcsecond; 2)~the number of sources
will be large enough to derive meaningful statistics; and 3)~the sources
will be uniformly distributed over the sky. However, the number of quasars
that are a)~bright both in optical and radio wavelengths and therefore, can
be detected with both techniques (e.g. brighter than magnitude 18 as
suggested by \citet{r:mig03}) and b)~have a compact core, currently is
rather limited. Among 3946 radio sources with $\delta > -10\ifm{}^\circ\else${}^\circ$\fi$
observed with the VLBA in the absolute astronomy mode, 508 objects have
an association with a quasar or a BL~Lac object brighter than V~$18^m$
from the catalogue of \citet{r:vcv2010} within a $4''$ search radius.
It was realized in mid 2000s that the densification of the list of such
objects is desirable. A specific program for identifying new VLBI
sources in the northern hemisphere, suitable for aligning the VLBI and Gaia
coordinate systems, was launched in 2006 \citep{r:bou08} with the eventual
goal of deriving highly accurate position of sufficiently radio-bright
quasars from VLBI observations in the absolute astrometry mode. Since the
current VLBI position catalogue is complete to the correlated flux density
level of 200~mJy, the new candidate sources should necessarily be by a factor
of 2--4 weaker than that level. The original observing sample consisted
of 447 optically bright, relatively weak extragalactic radio sources with
declinations above $-10^{\circ}$. The detailed observing scheme of this
project is presented in \cite{r:bou08}. The first VLBI observations resulted
in the detection of 398 targets with the European VLBI Network (EVN)
\citep{r:bou10}, although no attempt to derive their positions of produce
images was made. VLBI observations of this sample in the absolute astrometry
mode promises to increase the number of optically bright radio sources with
precisely known positions by 80\%.
As a next step of implementing this program, a subset of 105 detected
sources was observed with the global VLBI network that comprises the VLBA and
EVN observing stations with the goal of revealing their morphology on
milliarcsecond scales from VLBI images \citep{r:bou11} for consecutive
screening the objects with structure that potentially may cause non-negligible
systematic position errors. I present here results of astrometric analysis
from this VLBI experiment. Observations and their analysis are described in
sections \ref{s:obs} and \ref{s:anal}. The position catalogue is presented
in \ref{s:cat}. Concluding remarks are given in section \ref{s:summ}.
\section{Observations}
\label{s:obs}
The observations used in this paper were carried out during a 48-hour
experiment GC030 on 7--9 March 2008 with a global VLBI array comprising
ten VLBA and 6 EVN stations ({\sc eflsberg}, {\sc hartrao},
{\sc medicina}, {\sc noto}, {\sc onsala60}, and {\sc dss63} for part
of the time), simultaneously at S and X bands. The data were recorded
at 512~Mbps. The schedule was prepared by ensuring a minimum of three
5 minute long scans of each target source, while minimizing the slewing time
from source to source. In total, 115 objects, including 105 target sources
and 10 strong calibrators were observed during a 48-hour observing session.
Three target objects were observed in 2 scans, 20 target objects were
observed in 3 scans, 43 target objects were observed in 4 scans,
26 objects were observed in 5 scans, 10 objects were observed in 6
scans, 2 objects were observed in 7 scans, and 1 object was observed
in 8 scans. Antennas spent 78\% time recording signal from target sources.
Although the overall goal of the observing program was absolute
astrometry, the design the GC030 experiment suffered several limitation and
was not favorable for determining sources coordinates with high accuracy.
First, the intermediate frequencies were selected to cover a continuous range
at both S and X-bands: 2.22699--2.29099~GHz and 8.37699--8.44099~GHz
respectively. There were two rationals behind selection that frequency setup
(P.~Charlot (2011), private communication). First, at the beginning of 2008,
the 512~Mbps mode was new. At that time, that setup was tested only for
a case of contiguously allocated intermediate frequencies (IFs).
It was not clear whether every non-VLBA station will be able to support
the wide-band mode. Since it happened in the past when a change in
frequency setup ruined experiments,
it was decided to stay on the safe side and make the schedule using
contiguously spread IFs. Second, it was known (for example, D.~Gordon,
private communication, 2010) that AIPS implementation of fringe fitting,
task FRING, does not produce correct group delays when the IFs are spread
over the wide band. As a workaround, all absolute astrometry/geodesy
experiments prior 2010 were processed using a two-step approach: first the
fringe fit was made using data from each IF individually, and then
group delays over entire band were computed using fringe phases from each
individual IF derived in the previous step. The drawback of that approach
is that a source should be detected at each IF individually, which raises
the detection limit by $\sqrt{N}$, where $N$ is the number of IFs at
each band (4 in our case). Since the target sources were expected to be
weak, it was important to avoid a degradation of the detection limit by
a factor of 2. Work for developing an alternative fringe fitting
procedure \citep{r:vgaps}, free from this drawback was underway in 2008,
when the experiment was scheduled, but not finished at that time.
Unfortunately, group delays determined with the contiguously frequency
setup are {\it one order of magnitude} less precise with respect to the
frequency allocation traditionally used for absolute astrometry work
with the VLBA.
The second limitation of the GC030 schedule for astrometry use
was a relatively rare observation of sources at low and high elevations for
better estimation of troposphere path delay in zenith direction. It was found
in the past that if to observe calibrator sources at low and high elevations
at each station every 1--2 hours, the reliability of estimates of the path
delay in the neutral atmosphere is significantly improved, and as a result,
systematic errors caused by mismodeling propagation effects are
reduced \citep{r:vcs3}.
The third limitation of the GC030 schedule was a small number of sources
observed in prior astrometry/geodesy programs at dual S/X bands:
only 19 objects. Observations of a large number of sources, typically
30--60 objects in a 24 hour experiment, overlapping with previous
observations helps to establish firmly the orientation of the array
and to link positions of new sources with positions of other objects.
Despite all these limitations, it was worth efforts to derive source
positions from such data since the a~priori positions of
these objects determined from Very Large Array (VLA)
observations \citep{r:first,r:nvss} were in the range of $0.03''$--$1''$.
\section{Data analysis}
\label{s:anal}
The data were correlated at the Socorro hardware VLBA correlator.
The correlator computed the spectrum of cross correlation and
autocorrelation functions with frequency setup of 0.25~MHz at
accumulation intervals of 1.048576~s long.
The procedure of further analysis is described in full details
in \citet{r:vgaps}. Here only a brief outline is given. At the first step,
the fringe amplitudes were corrected for the signal distortion in the sampler
and then calibrated according to measurements of system temperature and
elevation-dependent gain. Since the log files from VLBA sites for the second
half of the experiment were lost, no phase calibration was applied.
Then the group delay, phase delay rate, group delay rate, and fringe phase
were determined for all observations for each baseline at X and S bands
separately using the wide-band fringe fitting procedure. These estimates
maximize the amplitude of the sum of the cross-correlation spectrum
coherently averaged over
all accumulation periods of a scan and over all frequency channels in all IFs.
After the first run of fringe fitting, 12 observations at each baseline with
the strongest signal to noise ratios (SNR) were used to adjust the
station-based complex bandpass corrections, and the procedure of computing
group delays was repeated. This part of analysis is done with $\cal P\hspace{-0.067em}I\hspace{-0.067em}M\hspace{-0.067em}A$ \
software\footnote{Available at \web{http://astrogeo.org/pima}.}. Then the
results of fringe fitting were exported to the VTD/post-Solve VLBI analysis
software\footnote{Available at \web{http://astrogeo.org/vtd}.} for
interactive processing group delays with the SNR high enough to ensure
that the probability of false detection is less than 0.001.
This SNR threshold is $5.8$ for the GC030 experiment. Detailed description of
the method for evaluation of the detection threshold can be found in
\citep{r:vgaps}. Then, theoretical path delays were computed
according to the state-of-the art parametric model as well as their partial
derivatives, and small differences between group delays and theoretical
path delay were used for estimation of corrections to a parametric model
that describe the observations with least squares (LSQ). Coordinates
of target source, positions of all stations, except the reference one,
parameters of the spline that describes corrections to the a~priori path
delay in the neutral atmosphere in the zenith direction for all stations,
and parameters of another spline that describes the clock function
with the time span 1 hour were solved for in separate least square
solutions that used group delays at X and S bands individually.
Observations that deviated by more than
$3.5\sigma$ in the preliminary solution were identified and temporarily
eliminated, and additive corrections to a~priori weights were determined.
The most common reason for an observation to be marked as an outlier
is a misidentification of the main maximum of the two-dimensional
Fourier-transform of the cross-spectrum. Then the fringe fitting procedure
was repeated for observations marked as outliers. But this time
the group delay and phase delay rate were evaluated for these observations
in a narrow window of 4~ns wide centered around the predicted value
of group delay computed using parameters of the VLBI model adjusted in
the preliminary LSQ solution. New estimates of group delays for points
with the probabilities of false detection less than 0.1, which corresponds
to the SNR $> 4.6$ for the narrow fringe search window, were used in the
next step of the interactive analysis procedure. The observations marked
as outliers in the preliminary solution and detected in the narrow window
at the second round of the fringe fitting were tried again.
If the new estimate of the residual was within 3.5 formal uncertainties,
the observation was restored and used in further analysis.
Parameter estimation, elimination of remaining outliers and adjustments
of additive weight corrections were then repeated. In total, 16629 matching
pairs of X and S band group delays out of 22750 scheduled were used in the
solution. Each source was detected at both bands and had the number
of dual-band pairs in the range of 19--321.
The result of the interactive solution provided the clean dataset of
ionosphere-free linear combinations of X and S-band group delays with
updated weights. The dataset that was used for the final parameter estimation
utilized all dual-band S/X data acquired under absolute astrometry and space
geodesy programs from April 1980 through December 2010, including the data
from the GC030 experiment, in total 8 million observations. Thus,
the GC030 experiment was analyzed exactly the same way as over 5000 other
VLBI experiments, using the same analysis strategy that was used for
processing prior observations for ICRF, VCS, VGaPS, LCS, and K/Q survey
\citep{r:kq} catalogues. The estimated parameters are right ascensions and
declination of all sources, coordinates and velocities of all stations,
coefficients of B-spline expansion of non-linear motion for 17 stations,
coefficients of harmonic site position variations of 48 stations at 4
frequencies: annual, semi-annual, diurnal, semi-diurnal, and axis offsets
for 67 stations. Estimated variables also included Earth orientation
parameters for each observing session, parameters of clock function and
residual atmosphere path delays in the zenith direction modeled with
the linear B-spline with interval 60 and 20 minutes respectively.
All parameters were adjusted in a single LSQ run.
The system of LSQ equations has an incomplete rank and defines a family
of solutions. In order to pick a specific element from this family, I applied
the no-net rotation constraints on the positions of
212~sources marked as ``defining'' in the ICRF catalogue \citep{r:icrf98}
that required the positions of these sources in the new catalogue to have
no rotation with respect to their positions in the ICRF catalogue.
No-net rotation and no-net-translation constraints on site positions
and linear velocities were also applied. The specific choice of identifying
constraints was made to preserve the continuity of the new catalogue with
other VLBI solutions made during last 15 years.
The global solution sets the orientation of the array with respect to
an ensemble of $\sim\!\!\!5000$ extragalactic remote radio sources.
The orientation is defined by the continuous series of Earth orientation
parameters and parameters of the empirical model of site position
variations over 30 years evaluated together with source coordinates.
Common sources observed in the GC030 experiment as amplitude calibrators
provided a connection between the new catalogue and the old catalogue of
compact sources.
As a valuable by-product of GC030 observations, positions of {\sc dss63}
station were determined (see Table~\ref{t:dss63}). To my knowledge,
this is the only S/X experiment with participation of this station that
can be found in publicly accessible databases. Velocity of {\sc dss63}
was constrained to be the same as velocity of {\sc dss65} station that
is located in 1440 meters from {\sc dss63}.
Radio images of observed sources in both S and X bands were presented
in a graphical form in \citet{r:bou11}. In order to provide
a measure of source strengths at long and short baselines for predicting
the SNR in future observations, I made my own simplified amplitude analysis
and derived the median correlated flux densities at baseline projection
lengths shorter than 900~km and longer than 5000 km. This procedure is
described in details in \citet{r:lcs1}. It is outlined here briefly. First,
I computed the a~priori system equivalent flux density (SEFD) using system
temperatures and gain curves for each antenna. Fringe amplitudes for every
observation used in astrometric analysis, except those marked as outliers,
were converted to flux densities by multiplying them by the square root
of the product of the a~priori SEFDs of both stations of a baseline.
Then I adjusted multiplicative gain corrections from logarithms of ratios
of observed correlated flux densities of 10 amplitude calibrators to
their values predicted on the basis of publicly available brightness
distributions\footnote{Available at \web{http://astrogeo.org/vlbi\_images}}
using least squares. I applied these corrections to estimates of correlated
flux densities of observed sources and computed the median value at two
ranges of baseline projections. Comparison of estimates of median
correlated flux densities derived by this method with estimates of
correlated flux densities generated using images produced by a rigorous
self-calibration procedure for three 24~hour survey experiments VCS5
\citep{r:vcs5} showed that the accuracy of median correlated flux densities
estimated using the simplified method is at a level of 15\%.
These estimates of correlated flux densities are complementary to
source image statistics shown in \citet{r:bou11}, for instance, to their
total flux densities integrated from X- and S-band images.
\begin{table}[hb]
\caption{Position of {\sc dss63} station at the 2000.0 epoch determined from
the GC030 experiment. Its velocities listed in the right column
were constrained to be the same as velocities of station
{\sc dss65}.}
\label{t:dss63}
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{l @{\quad} r @{\quad} r}
& Position (m) & Velocity (mm/yr) \\
\hline
X & $ 4849092.429 \pm 0.025 $ & $ -1.63 \pm 0.23 $ \\
Y & $ -3601804.438 \pm 0.013 $ & $ 18.49 \pm 0.09 $ \\
Z & $ 4115109.146 \pm 0.024 $ & $ 10.62 \pm 0.24 $ \\
\end{tabular}
\end{table}
\section{The catalogue}
\label{s:cat}
\begin{table*}[t]
\caption{First 12 rows of the OBRS--1 source position catalogue.}
\label{t:cat}
\begin{tabular}{ l l r r r r r r r r r r @{\enskip} r}
\hline
\nntab{c}{IAU name} &
\nntab{c}{Source coordinates} &
\nnntab{c}{Position errors} &
&
\nntab{c}{$F_{corr}$ S-band} &
\nntab{c}{$F_{corr}$ X-band} &
Flag \\
B1950 &
J2000 &
\ntab{c}{$ \alpha $ } &
\ntab{c}{$ \delta $ } &
$ \sigma_\alpha $ &
$ \sigma_ \delta $ &
Corr &
\#pnt &
short &
unres &
short &
unres &
\\
& &
\ntab{l}{~hr~mn~sec} &
\ntab{l}{~~~~$^{\circ}$~~~~$^\prime$~~~~~$^{\prime\prime}$} &
\ntab{c}{mas} & \ntab{c}{mas} & & &
\ntab{c}{Jy} & \ntab{c}{Jy} &
\ntab{c}{Jy} & \ntab{c}{Jy} & \\
\hline
0003$+$123 & J0006$+$1235 & 00 06 23.056086 & $+$12 35 53.09833 & 0.26 & 0.42 & $ 0.107$ & 198 & 0.137 & 0.071 & 0.138 & 0.083 & X \\
0049$+$003 & J0052$+$0035 & 00 52 05.568998 & $+$00 35 38.14614 & 0.89 & 3.16 & $-0.631$ & 32 & 0.031 & 0.024 & 0.062 & 0.061 & \\
0107$-$025 & J0110$-$0219 & 01 10 13.160493 & $-$02 19 52.84055 & 0.65 & 2.26 & $-0.476$ & 98 & 0.074 & 0.069 & 0.061 & 0.058 & \\
0109$+$200 & J0112$+$2020 & 01 12 10.190819 & $+$20 20 21.76438 & 0.31 & 0.67 & $-0.305$ & 170 & 0.095 & 0.070 & 0.125 & 0.087 & \\
0130$-$083 & J0132$-$0804 & 01 32 41.126050 & $-$08 04 04.83517 & 1.38 & 1.57 & $-0.029$ & 64 & 0.104 & 0.083 & 0.065 & 0.058 & \\
0145$+$210 & J0147$+$2115 & 01 47 53.822855 & $+$21 15 39.72637 & 0.40 & 1.15 & $-0.413$ & 127 & 0.295 & 0.210 & 0.107 & 0.053 & \\
0150$+$015 & J0152$+$0147 & 01 52 39.610907 & $+$01 47 17.38264 & 0.86 & 2.75 & $-0.576$ & 70 & 0.046 & 0.047 & 0.048 & 0.042 & \\
0210$+$515 & J0214$+$5144 & 02 14 17.934429 & $+$51 44 51.94772 & 0.39 & 0.36 & $-0.064$ & 401 & 0.092 & 0.076 & 0.059 & 0.046 & X \\
0446$+$074 & J0449$+$0729 & 04 49 21.170617 & $+$07 29 10.69568 & 0.63 & 1.31 & $-0.147$ & 80 & 0.043 & 0.036 & 0.092 & 0.074 & \\
0502$+$041 & J0505$+$0415 & 05 05 34.769151 & $+$04 15 54.57316 & 2.02 & 5.40 & $-0.381$ & 41 & 0.036 & 0.051 & 0.032 & 0.036 & \\
0519$-$074 & J0522$-$0725 & 05 22 23.196279 & $-$07 25 13.47580 & 4.96 & 7.17 & $ 0.088$ & 19 & 0.043 & 0.043 & 0.053 & 0.038 & \\
0651$+$428 & J0654$+$4247 & 06 54 43.525947 & $+$42 47 58.73588 & 0.48 & 0.64 & $ 0.494$ & 168 & 0.098 & 0.101 & 0.075 & 0.077 & \\
\hline
\end{tabular}
\tablecomments{Table~\ref{t:cat} is presented in its entirety
in the electronic edition of the Astronomical Journal.
A portion is shown here for guidance regarding
its form and contents.
}
\end{table*}
I have determined positions of 105 sources observed in GC030 experiment.
They are listed in Table~\ref{t:cat}. Although positions of all 5336
astrometric sources were adjusted in the LSQ solution that included the
OBRS--1 sources, only coordinates of the 105 target sources observed during
GC030 experiment are presented in the table. The 1st and 2nd columns give
the IVS source name (B1950 notation) and IAU name (J2000 notation).
The 3rd and 4th columns give source coordinates at the equinox on
the J2000.0 epoch. Columns 5 and 6 give formal source position uncertainties
in right ascension and declination in mas (without $\cos\delta$ factor),
and column 7 gives the correlation coefficient between the errors in right
ascension and declination. The number of group delays used for position
determination is listed in column 8. Columns 9 and 10 provide the median
value of the correlated flux density in Jansky at S-band at baseline
projection lengths shorter than 900~km and at baseline projection lengths
longer than 5000~km. The latter estimate serves as a measure of the
correlated flux density of the unresolved component of a source.
Columns 11 and 12 provide the median of the correlated flux density at
X-band at baselines shorter than 900~km and longer than 5000~km. The last
column contains a cross-reference flag: V if a sources was observed in
VIPS campaign, X if it was observed at X/S bands in other absolute
astrometry campaign, and VX if it was observed in both.
Uncertainties in sources position that were observed only in the GC030
experiment range from 0.3~mas (\object{1345+735})
to 7.2~mas (\object{0519-074}) with the median 1.1~mas. The distribution
of the semi-major axes of position error ellipses is presented in
Figure~\ref{f:obrs1_hist}.
\begin{figure}[tbh]
\includegraphics[width=0.48\textwidth,clip]{obrs1_hist.eps}
\caption{The histogram of the semi-major axes of position
error ellipses among 105 target sources in the OBRS--1
catalogue.
\label{f:obrs1_hist}
}
\end{figure}
\section{Error analysis}
Among 105 target sources, 26 objects were observed in VIPS C-band
(5~GHz) program and 9 objects were observed in dual-frequency S/X VLBA
experiments under absolute astrometry programs. For comparison
purposes, I made a trial solution that used exactly the same
setup as the main solution, but excluded 9 common objects from the
GC030 experiment. The results of the comparison presented
in Table~\ref{t:diff} shows that, except for
declination of \object{2043+749}, the differences are within formal
uncertainties of the OBRS--1 catalogue. The formal uncertainties of
source positions are computed from standard deviations of group delay
estimates using the law of error propagation. Since the selection of
intermediate frequencies was unfavorable for a precise determination of
group delays and the sources were relatively weak, the thermal noise
dominates the error budget.
Realistic uncertainties of parameter adjustments can be evaluated
only by exploiting some redundancy in the data or by using additional
information. We do not have enough redundancy to evaluate rigorously the
level of systematic errors in the GC030 campaign. Comparison of positions
of 9~common sources indicates that systematic errors, if exist, do not
exceed 1~mas. Although only 10 atmosphere calibrators were included in
the schedule, 3--5 times less than in dedicated absolute astrometry observing
experiments, they were observed rather intensively. We can indirectly
estimate the level of systematic errors caused by the sparseness of the
distribution of calibrator sources by comparing the source distribution
in experiment GC030 with that in the prior VLBA Calibrator Survey program
VCS1 \citep{r:vcs1}. The azimuthal-elevation distribution of sources observed
in that campaign for a central VLBA antenna (Figure~3b in \citet{r:vcs1}) was
concentrated in a narrow band at the sky for 95\% of the sources, and
very few atmospheric calibrators outside that band were used. The reliability
of estimation of atmosphere path delays in zenith direction was significantly
compromised, and as a result, the formal uncertainties from the LSQ solution
had to be inflated by adding in quadrature the error floor
of 0.4~mas.
The VCS1 campaign can be considered as an extreme case of the
effect of the non-uniform distribution of observed sources. Both calibrator
sources and targets in GC030 were distributed more uniformly than in the VCS1
campaign. Analysis of estimates of residual atmosphere path delays does
not show abnormalities. I surmise tentatively that systematic
errors of the OBRS--1 catalogue are probably do not exceed 0.4~mas, which
is insignificant with respect to its random errors. I presented formal
uncertainties from the LSQ solution ``as is'', leaving investigation of
systematic errors in depth in the future when more observations in this
mode will be collected.
\begin{table*}[ht]
\caption{Differences between estimates of coordinates of 9 common target
sources determined using only GC030 observations and estimates from
other X/S VLBA absolute astrometry experiments.}
\label{t:diff}
\begin{tabular}{ l l r r r r r r r r}
\hline
\nntab{c}{Source name} & \nntab{c}{Position difference} &
\nntab{c}{XS source position} & \nnntab{c}{X/S position uncertainty} &
\phantom{$\bigl(\bigr)$} \\
B1950-name & J2000-name & $ \Delta \alpha \cos \delta $ &
\ntab{c}{$ \Delta\delta $} & Right ascension & Declination &
$ \sigma(\alpha)$ & $ \sigma(\delta)$ & Corr & \# Obs \\
& & \ntab{c}{mas} & \ntab{c}{mas} & \ntab{l}{~hr~mn~sec} &
\ntab{l}{~~~~$^{\circ}$~~~~$^\prime$~~~~~$^{\prime\prime}$} &
\ntab{c}{mas} & \ntab{c}{mas} & & \\
\hline
0003$+$123 & J0006$+$1235 & $ -0.4 \pm 0.6 $ & $ 0.0 \pm 0.8 $ & 00 06 23.05607 & +12 35 53.0983 & 0.3 & 0.5 & 0.124 & 62 \\
0210$+$515 & J0214$+$5144 & $ -1.3 \pm 0.6 $ & $ -0.1 \pm 0.8 $ & 02 14 17.93433 & +51 44 51.9475 & 0.8 & 0.5 & 0.296 & 99 \\
0708$+$742 & J0714$+$7408 & $ -0.3 \pm 0.4 $ & $ -0.2 \pm 0.4 $ & 07 14 36.12502 & +74 08 10.1440 & 0.6 & 0.2 & 0.085 & 88 \\
1721$+$343 & J1723$+$3417 & $ 0.3 \pm 0.6 $ & $ 0.3 \pm 1.1 $ & 17 23 20.79594 & +34 17 57.9652 & 0.2 & 0.4 & -0.432 & 76 \\
1759$+$756 & J1757$+$7539 & $ 0.8 \pm 0.8 $ & $ 2.8 \pm 1.6 $ & 17 57 46.35883 & +75 39 16.1800 & 1.3 & 0.4 & 0.375 & 276 \\
2043$+$749 & J2042$+$7508 & $ -0.6 \pm 0.6 $ & $ -3.7 \pm 1.2 $ & 20 42 37.30776 & +75 08 02.4415 & 1.4 & 1.1 & 0.211 & 45 \\
2111$+$801 & J2109$+$8021 & $ 0.6 \pm 1.7 $ & $ -0.6 \pm 2.2 $ & 21 09 19.16511 & +80 21 11.2264 & 9.8 & 2.2 & -0.032 & 13 \\
2316$+$238 & J2318$+$2404 & $ 0.3 \pm 0.3 $ & $ 0.1 \pm 0.7 $ & 23 18 33.96785 & +24 04 39.7496 & 0.3 & 0.4 & 0.050 & 72 \\
2322$+$396 & J2325$+$3957 & $ 0.3 \pm 0.4 $ & $ -0.5 \pm 0.8 $ & 23 25 17.86983 & +39 57 36.5084 & 0.2 & 0.2 & -0.274 & 118 \\
\hline
\end{tabular}
\end{table*}
\section{Discussion}
In the course of development of radio astrometry for last 40 years,
we learned that in order to derive precise source positions using the method
of absolute astrometry, a VLBI experiment should 1)~have intermediate
frequencies spread as wide as possible over the band(s); 2)~observe every
1--2~hours blocks of 3--5 sources with at least one source at
elevations $20\ifm{}^\circ\else${}^\circ$\fi$ above the horizon and one source at elevations $55\ifm{}^\circ\else${}^\circ$\fi$
above the horizon; 3)~collect enough bits for detection target sources
at long baselines.
Unfortunately, the selection of intermediate frequencies in the GC030
experiment did not satisfy the first condition. The choice of intermediary
frequencies is not very important for producing source images and observers
often record a continuous bandwidth. But this choice is critical for absolute
astrometry applications, since precision of group delay is reciprocal to the
variance of the frequencies in the band. The choice is especially important
for astrometry of weak sources, since unlike to observations of bright
sources when systematic errors dominate the error budget, the position
accuracy of weak sources is determined by the uncertainties of group delays
caused by the thermal noise. The frequency setup spread over 494~MHz used
in VLBA geodesy/astrometry RDV program \citep{r:rdv} had the uncertainties
of group delay by a factor of 11.1 smaller than in the GC030 experiment at
a given signal to noise ratio (SNR). The VLBA hardware allows to spread the
IFs over 1000~MHz that brings uncertainties of group delay down even further
by a factor of~2 \citep{r:wide-memo11}.
It should be stressed that there is no necessity to limit the spread of
intermediate frequencies for image experiments. One of the most extensive
dedicated imaging program, the VLBI Image and Polarization Survey (VIPS)
\citep{r:vips} used 4 IFs spread over 494~MHz in order to improve the
$uv$ coverage and to allow for rotation measure determinations \citep{r:tay05}.
Analysis of both VIPS and RDV observations provided excellent source
maps \citep{r:vips,r:rdv_astro,r:pus08}. Maps from absolute astrometry
observations typically have dynamic range 1:100--1:1000 (see \citet{r:vcs6}
and references therein). These maps allowed \citet{r:cha07} to determine
source structure indexes and make conclusions about suitability of sources
for precise astrometry.
The approach proposed by \citet{r:bou08} to run 3 observing campaigns
for an absolute astrometry program, first for detection, second for producing
source maps, third for deriving source positions deviates sharply from
the strategy used for last 40 years for determining positions of 6000 sources,
which used one campaign per program.
Our analysis of GC030 experiment shows that running 2 separate observing
campaigns for imaging and astrometry, which doubles requested observing time,
is not the best choice. Spreading the intermediate frequencies over 500~MHz
would reduce random errors of position estimates by a factor of 11, i.e. the
median position error would be 0.1~mas, without compromising imaging results.
With such precise group delays, position accuracy would be limited by
systematic errors. More intensive observations of troposphere calibrators
for mitigation systematic errors would require approximately 5--8\% additional
observing time according to \citet{r:vgaps}. That means that the goal of
the project could be reached by using 2 runs instead of 3, which
requires one half of requested resources.
Including {\sc eflsberg} in the array is beneficial, because this
station improves the baseline sensitivity at X-band by a factor of 4,
which is important for detecting weak sources. The benefit of using other
European stations and especially a station in South Africa which has almost
no mutual visibility with both American and European stations is less obvious.
In order to access the impact of other stations on the source position
estimates, I made a trial solution that excluded {\sc medicina},
{\sc hartrao}, {\sc noto}, {\sc onsala60}, and {\sc dss63} from GC030.
Comparison of position differences showed that they are within formal
uncertainties. An average increase of uncertainties of the trial solution
using the data from the restricted array was 20\% for right ascensions and
30\% for declinations. The median increase was 28\% and 42\% respectively.
Removing EVN stations from the array would, of course, degrade the quality
of images, but as analysis of other VLBA experiment showed, for instance
K/Q survey \citep{r:kq}, not to the level that would undermine their
usability for the goals of this specific project.
These are important lessons that we learned from analysis of
these observations.
\section{Summary}
\label{s:summ}
Analysis of the first dual-band S/X VLBA experiment of the campaign for
observing optically bright extragalactic radio sources allowed us to
determine positions of 105 target sources. Despite using the frequency setup
unfavorable for absolute astrometry, the position uncertainties ranged
from 0.3 to 7~mas with the median value of 1.1~mas. The sources were
relatively weak: the median correlated flux density at baselines longer
than 5000~km ranged from 25 to 190~mJy with the median value around 60~mJy
at both bands, which is a factor of 2 weaker than in the VLBA Calibrator
surveys. However, recording at 512~Mbps with integration length of 300~s
was sufficient to detect 73\% of the observations, including those
at long baselines.
A position accuracy of 1~mas is sufficient for using these
sources as phase calibrators, but not sufficient for drawing meaningful
conclusions from comparison of Gaia and VLBI positions. All the sources
will have to be re-observed with the wide-band frequency setup in order
to reach 0.1~mas level of accuracy.
In 2010--2011, the remaining 293 sources were observed
at the VLBA + EVN. These observations will help us to further extend
the position catalogue of optically bright radio sources.
\acknowledgements
It is my pleasure to thank G\`{e}raldine Bourda and Patric Charlot
for fruitful discussions. The National Radio Astronomy Observatory is
a facility of the National Science Foundation operated under cooperative
agreement by Associated Universities, Inc.
{\it Facilities:} \facility{VLBA (project code GC030)}.
|
\section{Introduction}
Since decades antennas are used in everyday devices in the radio and
microwave spectral range as a bridge between propagating radiation
and localized fields. In this spectral range semi-analytical models
exist \citep{Balanis1989a}, offering insight into interaction processes
and guiding engineers in antenna design. In the same time, for practical
antenna design and optimization well established numerical tools are
typically used \citep{Volakis2007}. Shifting antenna resonances towards
optical spectral range brings new challenges both from the fabrication
and the theoretical perspectives. At optical frequencies metal can
no longer be treated as perfect electric conductor and dimensions
of the antenna might be as small as several tens of nanometers \citep{Bharadwaj2009}.
Recent progresses in nanotechnology have enabled the fabrication of
optical antennas (nano-antennas) \citep{Bharadwaj2009}, \citep{Muhlschlegel2005}
and opened many exciting possibilities towards nano-antenna applications.
For example, it has been recently demonstrated that nano-antennas
can enhance \citep{Rogobete2007} and direct the emission of single
molecules \citep{Taminiau2008} and that they can play a key role
in sensing application \citep{Raschke2003}. Great potential in improving
the efficiency of solar-cells should also be mentioned \citep{Atwater2010}.
Design and optimization of optical antennas are mainly done using
general numerical Maxwell solvers \citep{myro2008}, which demand
huge computational resources. Therefore accurate analytical and semi-analytical
models predicting characteristics and performance of nano-antenna
are of great importance. There are just very few exact analytical
solutions available. Light scattering on metal spheres \citep{Bohren1998},
infinite long cylinders \citep{Bohren1998} and spheroids \citep{Sinha1977}
can be derived in closed analytical form. For the light scattering
problem involving a finite length nanowire, a Pocklington-like one-dimensional
(1D) integral equation can be introduced using equivalent surface
impedance method \citep{Hanson2006}. In this paper, we propose a
further development of this approach, which provides better accuracy
especially in the case of nano-antennas with small aspect ratio.
The paper is organized as follows. In section 2 the problem of light
scattering on a thin perfectly conducting wire is reviewed. Pocklington's
integral equation is introduced and extended to the case of a nanowire
of finite conductivity. We demonstrate how one can use the knowledge
of the exact solution of the problem of the plane wave scattering
on an infinite cylinder in order to improve accuracy of the surface
impedance method \citep{Hanson2006}. In Sections 3 and 4 a new numerical
method to solve the resulting one-dimensional (1D) integral equation
is introduced. The method involves a method of moments (MoM) like
discretization scheme, but does not require any specific boundary
conditions to be imposed at the nanowire ends. In section 5 numerical
calculations of scattering cross-sections for plane wave scattering
on gold nanowires with varying geometries is presented and compared
with numerically rigorous discrete dipole approximation (DDA) calculations
\citep{YURKIN2007a}. Section 6 summarizes the paper.
\section{Pocklington like equation}
\begin{figure}
\begin{centering}
\includegraphics[height=0.6\columnwidth]{fig1}
\par\end{centering}
\caption{\label{fig:geometry}Definition of the geometrical parameters, radius
$a$ and length $l$, of the scattering cylinder as well as the chosen
body centered coordinate system. Additionally the incident angle $\xi$
and the polarization basis vector $\mathbf{E}_{0}^{\parallel}$ are depicted in the incident plane.}
\end{figure}
In infinite free space the electric field $\mathbf{E}(\mathbf{r})$
generated by a time harmonic current density distribution $\mathbf{j}(\mathbf{r})$
(time dependence $e^{-i\omega t}$) enclosed in the finite volume
$V$ is given by \citep{Bladel2007} \begin{multline}
\mathbf{E}(\mathbf{r})=\mathbf{E}^{inc}(\mathbf{r})+i\omega\mu_{0}\left(\overleftrightarrow{\mathbf{I}}+\frac{1}{k^{2}}\nabla\otimes\nabla\right)\\
\int_{V}g(\mathbf{r},\mathbf{r}')\mathbf{j}(\mathbf{r}')\, d^{3}r'\label{eq:integral_equation_j}\end{multline}
where $\overleftrightarrow{\mathbf{I}}$ denotes the three-dimensional
unit tensor, $\otimes$ is the tensor product defined by $\left(\mathbf{a}\otimes\mathbf{b}\right)_{ij}=a_{i}b_{j}$,
\begin{equation}
g(\mathbf{r},\mathbf{r}')=\frac{e^{ik\left|\mathbf{r}-\mathbf{r}'\right|}}{4\pi\left|\mathbf{r}-\mathbf{r}'\right|}\label{eq:scalar_greens_function}\end{equation}
is the scalar Green's function and $k=\frac{\omega}{c}$ the free
space wave number. $\mathbf{E}^{inc}$ is an electric field due to
sources not contained in $V$. In scattering problems the driving
current density is not controlled from the outside, but instead induced
by $\mathbf{E}^{inc}$ which plays the role of an incident field in
this case. Spatial derivatives of the integral in (\ref{eq:integral_equation_j})
exist as long as the source current density $\mathbf{j}(\mathbf{r})$
satisfies the H\"{o}lder condition \citep{Kellogg1953}, $\left|\mathbf{j}\left(\mathbf{r}\right)-\mathbf{j}(\mathbf{r}')\right|\leq k\left|\mathbf{r}-\mathbf{r}'\right|^{\alpha}$
with $k>0$ and $0<\alpha\leq1$ for every pair $\mathbf{r},\mathbf{r}'\in V$.
In what follows we consider light scattering on a cylinder (wire)
with length $l$ and radius $a$ (Fig.~\ref{fig:geometry}). The
geometry of the problem including a body centered coordinate system
with $z$-axis parallel to the cylinder axis is depicted in figure~\ref{fig:geometry}.
The wave vector $\mathbf{k}$ enclosing the incident angle $\xi$
with the positive $z$-axis lies in the $xz$-plane (incident plane).
To excite longitudinal resonances only the projection of the incident
plane wave electric field on the incident plane $\mathbf{E}_{0}^{\parallel}=E_{0}^{\parallel}\left(\sin\xi\hat{\mathbf{z}}-\cos\xi\hat{\mathbf{x}}\right)$
have to be taken into account. The perpendicular polarization component can be
safely ignored. Under the assumption that the cylinder diameter is
much smaller than the free space wavelength, i.e. $ka\ll1$, the incident electric field
interacting with the cylinder can be viewed as a function depending
only on $z$\begin{eqnarray}
\mathbf{E}^{inc}(z) & \approx & \mathbf{E}_{0}^{\parallel}e^{-ikz\cos\xi}.\label{eq:incident_field-1}\end{eqnarray}
First we briefly review the derivation of Pocklington's equation for
scattering on a thin perfectly conducting wire \citep{Balanis1989a}.
In this case in cylindrical coordinates $\left\{ \rho,\phi,z\right\} $
the induced current density $\mathbf{j}(\mathbf{r})$ has only a $z$-component,
shows no $\phi$-dependence and exists solely on the antenna interface.
Then the induced current $I(z)$ is related to the current density
$\mathbf{j}(\mathbf{r})$ as\begin{equation}
\mathbf{j}(\mathbf{r})=\hat{\mathbf{z}}I(z)\frac{\delta(\rho-a)}{2\pi a}\label{eq:surface_current_density}\end{equation}
so that \begin{equation}
I(z)=\int_{0}^{2\pi}d\phi\int_{0}^{a}\rho d\rho\, j(\mathbf{r}).\label{eq:I(z)}\end{equation}
Using (\ref{eq:surface_current_density}) in (\ref{eq:integral_equation_j})
and performing the $\rho'$-integration one yields for the $z$-component
of the electric field on the cylinder surface the following integro-differential
equation\begin{multline}
E_{z}(a,z)=E_{z}^{inc}(z)+i\frac{\omega\mu_{0}}{2\pi}\left(1+\frac{1}{k^{2}}\frac{\partial^{2}}{\partial z^{2}}\right)\\
\int_{-\frac{l}{2}}^{\frac{l}{2}}dz'\int_{0}^{2\pi}d\phi'\, g_{a}(\phi',z-z')I(z').\label{eq:E_z(rho,z)}\end{multline}
where $g_{a}(\phi',z-z')=g(\mathbf{r},\mathbf{r}')$ with $\mathbf{r}=\left( a,0,z\right) $
and $\mathbf{r}'=\left( a,\phi',z'\right) $ in cylindrical coordinates.
While the antenna is perfectly conducting, the $z$-component of the
total field at the antenna surface has to vanish, $E_{z}(a,z)=0$.
Applying this boundary condition to Eq.~(\ref{eq:E_z(rho,z)}) one
can derive Pocklington's integro-differential equation in the form\begin{multline}
E_{z}^{inc}(z)=-i\frac{\omega\mu_{0}}{2\pi}\left(1+\frac{1}{k^{2}}\frac{\partial^{2}}{\partial z^{2}}\right)\\
\int_{-\frac{l}{2}}^{\frac{l}{2}}dz'\int_{0}^{2\pi}d\phi'\, g_{a}(\phi',z-z')I(z'),\label{eq:IE_pec}\end{multline}
which provides a general solution of the scattering problem \citep{Balanis1989a}.
Here the fact, that the antenna is perfectly conducting, is taken
into account twice, first by assuming a special form of the induced
current (\ref{eq:surface_current_density}) and second by enforcing
the boundary condition $E_{z}(a,z)=0$.
A typical nano-antenna at optical frequency range possesses high,
but finite conductivity. In this case the surface current approximation
(\ref{eq:surface_current_density}) is still applicable, while the
boundary condition $E_{z}(a,z)=0$ is generally not. In order to use
a Pocklington's like equation at this frequency range, one needs a
relationship between the field $E_{z}(a,z)$ and the current $I(z)$
at the antenna interface.
Assuming a long nanowire, i.e. $l\gg a$, it is reasonable to expect
that the internal electric field is separable similar to the solution
of the equivalent problem involving an infinite cylinder \citep{Bohren1998}.
Moreover, it can be shown that as long as the wire is electrically
thin, i.e. $ka\ll1$, and $\left|\epsilon_{r}\right|\gg1$ at least for the case of inclined incidents
($\xi\neq\frac{\pi}{2}$) the $\rho$- and $\phi$-components of the
internal field are negligible compared to the $z$-component of the
field and the $z$-component shows no $\phi$-dependence. Assuming
these conditions are fulfilled the internal electric field can be
approximated by \begin{equation}
\mathbf{E}(\mathbf{r})\approx\hat{\mathbf{z}}f(z)J_{0}\left(k_{\rho}\rho\right)\label{eq:field_ansatz}\end{equation}
where $J_{0}\left(k_{\rho}\rho\right)$ denotes the Bessel function
of the first kind, $k_{\rho}=k\sqrt{\epsilon_{r}-\cos^{2}\xi}$ with
relative permittivity $\epsilon_{r}$ of the wire and $f(z)$
an unknown $z$-dependent function giving the amplitude of the internal
field along the wire. A connection between the induced current density
and the internal electric field is given by means of the volume equivalence
theorem \citep{Balanis1989a} by\begin{equation}
\mathbf{j}(\mathbf{r})=-i\omega\epsilon_{0}\Delta\epsilon_{r}\mathbf{E}(\mathbf{r})\label{eq:surface equivalence theorem}\end{equation}
with $\Delta\epsilon_{r}=\epsilon_{r}-1$. Combining (\ref{eq:field_ansatz})
and (\ref{eq:surface equivalence theorem}) the total current through
the wire can be calculated \begin{align}
I(z) & =\int_{0}^{2\pi}d\phi\int_{0}^{a}\rho d\rho\, j_{z}(\rho,z)\nonumber \\
& =-i\omega\epsilon_{0}\Delta\epsilon_{r}2\pi a\frac{J_{1}\left(k_{\rho}a\right)}{k_{\rho}}f(z).\label{eq:I(z) infinite cylinder}\end{align}
Further comparing results of the integration in (\ref{eq:I(z) infinite cylinder})
with (\ref{eq:field_ansatz}) one can derive the following relation
between the electric field and the total current at the wire interface
\begin{equation}
E_{z}(a,z)=Z_{S}I(z)\label{eq:E_from_impedance}\end{equation}
with the surface impedance\begin{equation}
Z_{S}=i\frac{J_{0}(k_{\rho}a)k_{\rho}}{2\pi a\omega\epsilon_{0}\Delta\epsilon_{r}J_{1}(k_{\rho}a)}.\label{eq:surface impedance}\end{equation}
Using relation (\ref{eq:E_from_impedance}) the following Pocklington
like integro-differential equation for induced total current, the
surface impedance (SI) integro-differential equation \citep{Hanson2006},
can be obtained from equation (\ref{eq:E_z(rho,z)}) \begin{multline}
Z_{S}I(z)=E_{z}^{inc}(z)+i\frac{\omega\mu_{0}}{2\pi}\left(1+\frac{1}{k^{2}}\frac{\partial^{2}}{\partial z^{2}}\right)\\
\int_{-\frac{l}{2}}^{\frac{l}{2}}dz'\int_{0}^{2\pi}d\phi'\, g_{a}(\phi',z-z')I(z').\label{eq:SI-IE}\end{multline}
We propose a further improvement to the approximation (\ref{eq:SI-IE})
by releasing the solely surface current ansatz (\ref{eq:surface_current_density}).
In order to do that, we assume that the induced current density on
the right hand side of equation (\ref{eq:integral_equation_j}) can
be factorize similar to the internal field (\ref{eq:field_ansatz}).
In this way substituting (\ref{eq:field_ansatz}) in equation (\ref{eq:integral_equation_j})
both on the left hand side as boundary condition as well as by using
(\ref{eq:surface equivalence theorem}) on the right hand side to
rewrite the induced current density one obtains a self-consistent
integro-differential equation for the unknown amplitude $f(z)$ \begin{multline}
f(z)J_{0}\left(k_{\rho}a\right)=E_{z}^{inc}(z)+k^{2}\Delta\epsilon_{r}\left(1+\frac{1}{k^{2}}\frac{\partial^{2}}{\partial z^{2}}\right)\\
\int_{V}d^{3}r'\, g(a,z;\mathbf{r}')f(z')J_{0}\left(k_{\rho}\rho'\right)\label{eq:VC-IE}\end{multline}
where $g(a,z;\mathbf{r}')=g(\mathbf{r},\mathbf{r}')$ with $\mathbf{r}=\left( a,0,z\right) $
in cylindrical coordinates. This volume current (VC) integro-differential
equation takes into account both appropriate boundary conditions at
the wire interface and an appropriate volume current distribution
inside the wire.
\section{Discrete form of VC integral equation\label{sec:Solving-the-Volume}}
In order to solve numerically the integro-differential equations (\ref{eq:SI-IE})
and (\ref{eq:VC-IE}) one has to impose additional boundary conditions
at the nano-antenna edges. A common choice is to impose the total
current $I(z)$ to be equal to zero for $z=\pm l/2$ \citep{Hanson2006}.
For a solid wire with finite conductivity this choice is generally
not justified, while the total current can be discontinuous at the
wire edges \citep{Bladel2007}.
To overcome the requirement of additional boundary condition one has
to convert the integro-differential equations into purely integral
ones. To do that one has to bring the differential operator $\left(1+k^{-2}\partial_{z}^{2}\right)$
inside the integral. This procedure results in singularities of the
order $\left|\mathbf{r}-\mathbf{r}'\right|^{3}$ which are generally
not integrable over a volume. To treat these singularities we follow
the regularization scheme proposed by Lee et al. in Ref.~\citep{Lee1980}.
The regularized VC equation (\ref{eq:VC-IE}) reads: \begin{multline}
\left(1+\Delta\epsilon_{r}L_{33}\right)E_{z}(a,z)=E_{z}^{inc}(z)+\\
k^{2}\Delta\epsilon_{r}\left\{ \int_{V-V^{\star}}G_{33}(a,z;\mathbf{r}')E_{z}(\rho',z')d^{3}r'+\right.\\
\int_{V^{\star}}\Biggl[G_{33}(a,z;\mathbf{r}')E_{z}(\rho',z')-\\
\frac{1}{k^{2}}\frac{\partial^{2}}{\partial z^{2}}g_{0}(a,z;\mathbf{r}')E_{z}(a,z)\Biggr]\, d^{3}r'\Biggr\}\label{eq:volume_integral_equation}\end{multline}
with $E_{z}(\rho,z)=f(z)J_{0}\left(k_{\rho}\rho\right)$. Here $V^{\star}$
denotes a finite and arbitrary shaped principal volume $V^{\star}$
containing the singular point $\mathbf{r}=(a,0,z)$, $g_{0}(\mathbf{r},\mathbf{r}')=\lim_{k\rightarrow0}g(\mathbf{r},\mathbf{r}')$
is the static scalar Green's function, $G_{33}$ is the $zz$-element
of the dyadic Green's function \citep{Bladel2007} \begin{equation}
G_{33}(\mathbf{r},\mathbf{r}')=\left(1+\frac{1}{k^{2}}\frac{\partial^{2}}{\partial z^{2}}\right)g(\mathbf{r},\mathbf{r}'),\label{eq:dyadic_greens_function}\end{equation}
and $L_{33}$ the $zz$-element of the source dyadic \citep{Bladel2007}
\begin{equation}
L_{33}=\frac{1}{4\pi}\oint_{\partial V^{\star}}d^{2}r'\,\frac{\left(z'-z\right)\left(\hat{\mathbf{n}}\cdot\hat{\mathbf{z}}\right)}{\left|\mathbf{r}-\mathbf{r}'\right|^{3}}.\label{eq:source_dyadic}\end{equation}
The surface integration in (\ref{eq:source_dyadic}) has to be performed
over the surface $\partial V^{\star}$ enclosing the principal volume
$V^{\star}$, $\hat{\mathbf{n}}$ denotes the outer surface normal.
The main advantage of the regularization scheme \citep{Lee1980} is
(i) that all singularities disappear and (ii) that the principal volume
$V^{\star}$ can be finite and arbitrary shaped. We choose a cylinder
with length $\Delta$, radius $\frac{\Delta}{2}$ and center at the
singular point $\mathbf{r}=(a,0,z)$ as the principal volume (Fig.~\ref{fig:coordinate-trafo}).
We assume $\Delta$ to be small, such that both the Bessel function
$J_{0}\left(k_{\rho}\rho'\right)$ and the amplitude function $f(z')$
are approximately constant over $V^{\star}$. Taking that into account,
writing the volume integrals in the regularized VC equation (\ref{eq:volume_integral_equation})
in cylinder coordinates $\left\{ \rho,\phi,z\right\} $ and collecting
all terms containing $f(z)$ on the left and all terms containing $f(z'\neq
z)$ on the right hand side, one obtains a one-dimensional
integral equation in the form \begin{multline}
f(z)\Gamma\approx E_{z}^{inc}(z)+\int_{-\frac{l}{2}}^{z-\frac{\Delta}{2}}dz'f(z')\mathcal{L}(\left|z-z'\right|)+\\
\int_{z+\frac{\Delta}{2}}^{\frac{l}{2}}dz'f(z')\mathcal{L}(\left|z-z'\right|),\label{eq:volume_integral_1d-2}\end{multline}
with \begin{equation}
\Gamma=J_{0}\left(k_{\rho}a\right)\left\{ 1+\Delta\epsilon_{r}L_{33}-\eta_{in}\right\} -\eta_{out},\label{eq:def_Gamma}\end{equation}
\begin{multline}
\eta_{in}=4k^{2}\Delta\epsilon_{r}\int_{0}^{\frac{\Delta}{2}}d\tilde{z}\\
\left\{ \left(\int_{\phi_{min}}^{\pi}d\tilde{\phi}\int_{0}^{\frac{\Delta}{2}}\tilde{\rho}d\tilde{\rho}+\int_{\frac{\pi}{2}}^{\phi_{min}}d\tilde{\phi}\int_{0}^{\rho_{max}\left(\tilde{\phi}\right)}\tilde{\rho}d\tilde{\rho}\right)\right.\\
\left.\left(G_{33}(\mathbf{0},\tilde{\mathbf{r}})-\frac{1}{k^{2}}\frac{\partial^{2}}{\partial z^{2}}g_{0}(\mathbf{0},\tilde{\mathbf{r}})\right)\right\} ,\label{eq:def_eta_in}\end{multline}
\begin{multline}
\eta_{out}=4k^{2}\Delta\epsilon_{r}\int_{0}^{\frac{\Delta}{2}}d\tilde{z}\int_{\phi_{min}}^{\pi}d\tilde{\phi}\int_{\frac{\Delta}{2}}^{\rho_{max}(\tilde{\phi})}\tilde{\rho}d\tilde{\rho}\\
\left\{ J_{0}\left[k_{\rho}\rho'(\tilde{\rho},\tilde{\phi})\right]G_{33}(\mathbf{0},\tilde{\mathbf{r}})\right\} \label{eq:def_eta_out}\end{multline}
and
\begin{multline}
\mathcal{L}(\left|z-z'\right|)=2k^{2}\Delta\epsilon_{r}\int_{0}^{a}\rho'd\rho'J_{0}\left(k_{\rho}\rho'\right)\\
\int_{0}^{\pi}d\phi'G_{33}(a,z;\rho',z').\label{eq:def_L}\end{multline}
In (\ref{eq:def_eta_in}) the integration is performed over the principal
volume $V^{\star}$, while in (\ref{eq:def_eta_out}) over the corresponding
wire slice of thickness $\Delta$ centered at $z=z'$ but with excluded
principal volume $V^{\star}$. To calculate these two integrals a
new coordinate system with its center at the singular point has been
chosen Fig.~(\ref{fig:coordinate-trafo}). In the new coordinate
system the source dyadic (\ref{eq:source_dyadic}) can be explicitly
written as\begin{multline}
L_{33}=\frac{(\pi-\phi_{min})\left(2-\sqrt{2}\right)}{2\pi}+\frac{\left(\phi_{min}-\frac{\pi}{2}\right)}{\pi}-\\
\frac{\Delta}{\pi}\int_{\frac{\pi}{2}}^{\phi_{min}}d\tilde{\phi}\frac{1}{\sqrt{4\rho_{max}^{2}(\tilde{\phi})+\Delta^{2}}}.\label{eq:L33}\end{multline}
The radius vector $\rho'$ and the integration ranges in (\ref{eq:def_eta_in},\ref{eq:def_eta_out},\ref{eq:L33})
parametrically depend on the nanowire radius $a$ and the wire slice
thickness $\Delta$ and are given by\begin{equation}
\rho'(\tilde{\rho},\tilde{\phi})=\left|\left(\begin{array}{c}
a\\
0\end{array}\right)+\left(\begin{array}{c}
\tilde{\rho}\cos\tilde{\phi}\\
\tilde{\rho}\sin\tilde{\phi}\end{array}\right)\right|,\label{eq:rho_transformed}\end{equation}
\begin{align}
\rho_{max}(\tilde{\phi}) & =-2a\cos\tilde{\phi}\nonumber \\
\phi_{min} & =\arccos\left[-\frac{\Delta}{4a}\right].\label{eq:int_range_functions}\end{align}
Taking into account that the nanowire diameter is small in comparison
to the wavelength one can further simplify the integrals in (\ref{eq:def_eta_in})
and (\ref{eq:def_eta_out}) by expanding the Green's function $G_{33}$
in power of $k$ up to the linear term\begin{equation}
G_{33}^{NF}(R)\approx\frac{1}{4\pi}\left(\frac{3R_{z}^{2}-R^{2}}{k^{2}R^{5}}+\frac{R_{z}^{2}+R^{2}}{2R^{3}}+i\frac{2}{3}k\right)\label{eq:G33_NF}\end{equation}
where $\mathbf{R}=\mathbf{r}'-\mathbf{r}$. Using (\ref{eq:G33_NF})
one can analytically integrate equation (\ref{eq:def_eta_in}) over
$\tilde{z}$ and $\tilde{\rho}$ and equation (\ref{eq:def_eta_out})
over $\tilde{z}$ . Residual integration in (\ref{eq:def_eta_in}),
(\ref{eq:def_eta_out}) and (\ref{eq:L33}) has to be performed numerically.
The regularization scheme (\ref{eq:volume_integral_equation}) ensures,
that numerical integration converges. In this way $\Gamma$ in (\ref{eq:volume_integral_1d-2})
can be efficiently calculated once for given radius $a$ of the nanowire
antenna, $\Delta$ and wavelength.
Equation (\ref{eq:volume_integral_1d-2}) can be solved numerically
by a method of moments (MoM) approach. Specifically we choose a point
matching MoM scheme with pulse-function basis \citep{Orfanidis2010}.
That means we divide the nanowire into a set of slices with thickness
$\Delta$ and label each slice with an index $i$. For reasonable
thin slices the amplitude function $f(z)$ on each of them can be
approximated by its value at the center $f(z_{i})=\mathbf{f}_{i}$.
This leads to the $n=\frac{l}{\Delta}$ dimensional matrix equation
as discretized version of (\ref{eq:volume_integral_1d-2})\begin{equation}
\left(\Gamma\overleftrightarrow{\mathbf{I}}-\overleftrightarrow{\mathbf{M}}\right)\mathbf{f}=\mathbf{E}^{inc}\label{eq:1d_MoM}\end{equation}
where $\mathbf{E}_{i}^{inc}=E_{z}^{inc}(z_{i})$ and \begin{equation}
M_{ij}=\begin{cases}
\int_{z_{j}-\frac{\Delta}{2}}^{z_{j}+\frac{\Delta}{2}}dz'\mathcal{L}\left(\left|z_{i}-z'\right|\right) & \text{for }i\neq j\\
0 & \text{for }i=j\end{cases}\label{eq:M_ij}\end{equation}
which can be easily inverted numerically to yield the discrete set
of values $f(z_{i})$. While the Green's function depends on the distance
between two slices $\left|z-z'\right|$, only the first row $M_{1j}$
of the matrix $M$ has to be calculated. All other elements can be
filled using the rule $M_{ij}=M_{i-1,j-1}$.
Using the Green's function expansion (\ref{eq:G33_NF}) in the calculation
of the matrix elements $M_{1j}$ for $j<\frac{2a}{\Delta}$ the integration
over $z'$ can be done analytically. Only the integration over the
nanowire profile has to be done numerically. In the calculation of
the matrix elements $M_{1j}$ for $j>\frac{2a}{\Delta}$ the full
dyadic Green's function have to be used and so the complete volume
integral has to be done numerically. These integrations are done over
$\mathbf{r}'$-regions far from the singularity and they demonstrate
good convergence. An additional performance improvement can be achieved
by expanding the Bessel function $J_{0}$ for small arguments as \citep{Stegun1965}
\begin{equation}
J_{0}(x)\approx1-\frac{x^{2}}{4}+\frac{x^{4}}{64}+\mathcal{O}(x^{6}).\label{eq:bessel_expansion}\end{equation}
Having the amplitude function $f(z)$ calculated from (\ref{eq:1d_MoM})
the $z$-component of the induced electric field is given by the ansatz
(\ref{eq:field_ansatz}).
\begin{figure}
\begin{centering}
\includegraphics[width=0.8\columnwidth]{fig2}
\par\end{centering}
\caption{\label{fig:coordinate-trafo}Cross-section of the nanowire (big circle)
and the principal volume $V^{\star}$ (small circle) together with
the definition of the coordinate systems used to calculate integrals
in equation (\ref{eq:volume_integral_1d-2}).}
\end{figure}
\section{Discrete form of SI integral equation}
To solve the surface impedance integro-differential equation (\ref{eq:SI-IE})
the regularization and discretization scheme presented in section~\ref{sec:Solving-the-Volume}
can be also applied. A principal volume is chosen as a full slice
of the wire at position $z$ with thickness $\Delta$. Regularizing
the SI equation (\ref{eq:SI-IE}), writing it in cylinder coordinates
and collecting all terms containing $I(z)$ and $I(z'\neq z)$, one
obtains a one-dimensional integral equation in the form\begin{multline}
I(z)\Gamma^{S}\approx E_{z}^{inc}+\int_{-\frac{l}{2}}^{z-\frac{\Delta}{2}}dz'I(z')\mathcal{L}^{S}(\left|z-z'\right|)+\\
\int_{z+\frac{\Delta}{2}}^{\frac{l}{2}}dz'I(z')\mathcal{L}^{S}(\left|z-z'\right|)\label{eq:Iz_integral_equation_2nd}\end{multline}
with\begin{equation}
\mathcal{L}^{S}(\left|z-z'\right|)=i\frac{\omega\mu_{0}}{\pi}\int_{0}^{\pi}d\phi'\, G_{33}^{a}(\phi',z-z')\label{eq:LLS}\end{equation}
and\begin{multline}
\Gamma^{S}=Z_{S}+i\frac{\omega\mu_{0}}{2\pi}\left\{ \frac{1}{k^{2}}L_{33}^{S}-4\int_{0}^{\frac{\Delta}{2}}dz'\int_{0}^{\pi}d\phi'\right.\\
\left.\left[G_{33}^{a}(\phi',z')-\frac{1}{k^{2}}\frac{\partial^{2}}{\partial z^{2}}g_{0}^{a}(\phi',z')\right]\right\} ,\label{eq:GammaS}\end{multline}
where\begin{equation}
L_{33}^{S}=\frac{\Delta}{4\pi}\int_{0}^{2\pi}d\phi'\,\frac{1}{\sqrt{\left(\frac{\Delta}{2}\right)^{2}+4a^{2}\sin^{2}\left(\frac{\phi'}{2}\right)}^{3}}\label{eq:L33_1st}\end{equation}
and $G_{33}^{a}(\phi',z-z')=G_{33}(\mathbf{r},\mathbf{r}')$ as well
as $g_{0}^{a}(\phi',z-z')=\lim_{k\rightarrow0}g(\mathbf{r},\mathbf{r}')$
with $\mathbf{r}=\left\{ a,0,z\right\} $ and $\mathbf{r}'=\left\{ a,\phi',z'\right\} $
in cylinder coordinates. The source dyadic $L_{33}^{S}$ can be expressed
in terms of the complete elliptic integral of the second kind $E\left(m\right)$
\citep{Stegun1965} as \begin{equation}
L_{33}^{S}=\frac{1}{\pi}\frac{8}{16a^{2}+\Delta^{2}}E\left(-\frac{16a^{2}}{\Delta^{2}}\right),\label{eq:L33_2nd}\end{equation}
where $E\left(m\right)$ is defined by\begin{equation}
E\left(m\right)=\int_{0}^{\frac{\pi}{2}}\sqrt{1-m\sin^{2}\theta}\, d\theta.\label{eq:Def_elliptic_integral}\end{equation}
In $\Gamma^{S}$ (\ref{eq:GammaS}) the $z'$-integrations can be
performed analytically with the help of the near-field expansion of
the Green's function (\ref{eq:G33_NF}). The residual $\phi'$-integration
has to be done numerically once for given nanowire radius $a$, $\Delta$
and wavelength.
To solve (\ref{eq:Iz_integral_equation_2nd}) numerically we use a
point matching MoM scheme with pulse function basis \citep{Orfanidis2010}
as in section \ref{sec:Solving-the-Volume}. Again the nanowire should
be divided into a set of slices of thickness $\Delta$ yielding the
matrix equation\begin{equation}
\left(\Gamma^{S}\overleftrightarrow{\mathbf{I}}-\overleftrightarrow{\mathbf{M}}^{S}\right)\mathbf{I}=\mathbf{E}^{inc}\label{eq:SI_IE}\end{equation}
where\begin{equation}
M_{ij}^{S}=\begin{cases}
\int_{z_{j}-\frac{\Delta}{2}}^{z_{j}+\frac{\Delta}{2}}dz'\mathcal{L}^{S}\left(\left|z_{i}-z'\right|\right) & \text{for }i\neq j\\
0 & \text{for }i=j\end{cases}\label{eq:MS_ij}\end{equation}
and $I_{i}=I(z_{i})$. Similar to the case of volume current integral
equation considered in section \ref{sec:Solving-the-Volume} only
the first row of the matrix $\overleftrightarrow{\mathbf{M}}^{S}$
have to be calculated due to the symmetry of the Green's function.
Further for slices with $\left|z_{1}-z_{j}\right|\leq2a$ the near
field approximation of the Green's function (\ref{eq:G33_NF}) can
be used to perform the $z'$-integrations in (\ref{eq:MS_ij}) analytically.
The remaining integrations as well as the matrix inversion have to
be performed numerically. Good convergence of the integrals is ensured,
while the integrand is evaluated far from the singular point. Numerical
inversion of (\ref{eq:SI_IE}) results in the current $I_{i}=I(z_{i})$
at the discrete number of points $z_{i}$ along the nanowire. Finally
the $z$-component of the induced electric field is given via (\ref{eq:E_from_impedance}).
\section{Numerical results and discussion}
\begin{figure}
\begin{centering}
\includegraphics[width=0.9\columnwidth]{fig3}
\par\end{centering}
\caption{\label{fig:comparison_numeric}Scattering cross-section of a gold
nanowire ($l=200nm,\, a=10nm$) under slanting incidence ($\xi=\frac{\pi}{4}$)
calculated with different rigorous numerical methods as well as the
proposed volume current and surface impedance one-dimensional integral
equations. In the top panel (a) the first and in the bottom panel
(b) the third resonance peak are shown.}
\end{figure}
In this section the semi-analytical methods developed in sections
3 and 4 are evaluated and compared with numerically rigorous methods.
Plane wave scattering on a gold nanowire is considered in the optical
and near-infrared spectral range. The relative permittivity of gold
in this spectral range can be modeled by a free electron Drude model\begin{equation}
\epsilon_{r}(\omega)\approx\epsilon_{\infty}-\frac{\omega_{p}^{2}}{\omega^{2}+i\gamma\omega}\label{eq:drude_modell}\end{equation}
with parameters $\epsilon_{\infty}=9$, $\omega_{p}=1.36674\cdot10^{16}\, s^{-1}$
and $\gamma=7.59297\cdot10^{13}\, s^{-1}$\citep{myro2008}. For comparison
purposes we calculate the total scattering cross-section\begin{equation}
\sigma_{scat}=\frac{\oint_{S}d^{2}r\,\left|\mathbf{E}_{s}\right|^{2}}{\left|\mathbf{E}^{inc}\right|^{2}}\label{eq:sigma_scat_Def}\end{equation}
where the scattered electric field $\mathbf{E}_{s}$ is generated
by the induced current density $\mathbf{j}(\mathbf{r})$ given by
equation (\ref{eq:surface equivalence theorem}) in the case of the
VC and by equation (\ref{eq:surface_current_density}) in the case
of the SI integral equation model, respectively. In the far-field
zone the scattering field is purely transverse and one can use the
far-field dyadic Green's function\begin{equation}
\lim_{r\rightarrow\infty}\overleftrightarrow{\mathbf{G}}(\mathbf{r},\mathbf{r}')=\frac{e^{ik\left(r-\hat{\mathbf{r}}\cdot\mathbf{r}'\right)}}{4\pi r}\left\{ \overleftrightarrow{\mathbf{I}}-\hat{\mathbf{r}}\otimes\hat{\mathbf{r}}\right\} \label{eq:far_field_greens_function}\end{equation}
and integral relation\begin{equation}
\mathbf{E}_{s}(\mathbf{r})=i\omega\mu_{0}\int_{V}d^{3}r'\,\overleftrightarrow{\mathbf{G}}(\mathbf{r},\mathbf{r}')\mathbf{j}(\mathbf{r}')\label{eq:scattering_field}\end{equation}
to calculate the scattering cross-section. For the VC integral equation
model the scattering cross section can be written in terms of the
amplitude function $f(z)$\begin{multline}
\sigma_{scat}=\frac{k^{4}a^{4}\pi}{8\left|\mathbf{E}^{inc}\right|^{2}}\left|\Delta\epsilon_{r}\left(1-\frac{1}{8}a^{2}k^{2}\epsilon_{r}\right)\right|^{2}\\
\int_{0}^{\pi}d\theta\,\sin^{3}\theta\left|\int_{-\frac{l}{2}}^{\frac{l}{2}}dz'f(z')e^{-ik\cos\theta z'}\right|^{2}\label{eq:sigma_scat_VolMoM}\end{multline}
while for SI integral equation model in terms of the total current
$I(z)$\begin{multline}
\sigma_{scat}^{S}=\frac{\omega^{2}\mu_{0}^{2}}{8\pi\left|\mathbf{E}^{inc}\right|^{2}}\\
\int_{0}^{\pi}d\theta\,\sin^{3}\theta\left|\int_{-\frac{l}{2}}^{\frac{l}{2}}dz'I(z')e^{-ik\cos\theta z'}\right|^{2}.\label{eq:sigma_scat_SI-IE}\end{multline}
\begin{figure}
\begin{centering}
\includegraphics[width=0.9\columnwidth]{fig4}
\par\end{centering}
\caption{\label{fig:1st_resonances}Scattering cross-sections of gold nanowire
with fixed radius $a=10nm$ and different lengths. The full and dashed
lines show the VC and SI integral equation results, respectively.
The dotted line is the numerically rigorous reference calculated using
DDA method.}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[width=0.9\columnwidth]{fig5}
\par\end{centering}
\caption{\label{fig:map}Contours plot of the resonance frequencies of the
nanowires with different lengths under slanting incidence ($\xi=\frac{\pi}{4}$).
Top panel radius is $a=10nm$, bottom panel $a=20nm$. The full and
dashed lines show the VC and SI integral equation results, respectively.
The dots are represents the numerically rigorous DDA calculations.}
\end{figure}
We have performed a systematic convergence test of the developed one-dimensional
integral equation based methods for gold nanowires of different lengths
and radii. Spatial resolution of $\Delta=1nm$ (thickness of the individual
slices in the wire discretization) typically results in less than
1THz discrepancy with respect to the converged value of the scattering
cross-section maximum calculated at resolution as small as $\Delta=0.05nm$.
In what follows spatial resolution of $\Delta=1nm$ has been used
for all calculations based of the VC and SI integral equation models.
In figure \ref{fig:comparison_numeric} the first (top panel) and
the third (bottom panel) resonance peaks of the cross-section spectra
are shown for a gold nanowire ($a=10nm,\, l=200nm$) under slanting
incidence ($\xi=\frac{\pi}{4}$) with $\left|\mathbf{E}^{inc}\right|=1$.
Results of both different rigorous numerical methods and semi-analytical
1D ones are compared. For the rigorous numerical calculations we used
(i) HFSS, a commercial finite-element frequency-domain Maxwell solver
from ANSYS \citep{HFSS}, (ii) an in-house implementation of the finite
difference time domain method (FDTD) \citep{Taflove2000} and (iii)
ADDA, an open-source software package for calculating scattering parameters
using the discrete dipole approximation (DDA) algorithm \citep{Yurkin2007}.
The space discretization in the shown DDA and FDTD calculations was
set to $1nm$. Discrepancies among the scattering cross-section spectra
calculated using different rigorous three-dimensional (3D) Maxwell
solvers are comparable with discrepancies of these spectra with the
spectrum calculated using VC integral equation method. The accuracy
of the SI integral equation method is slightly worse but still very
reasonable.
Most important are the differences in used computational resources
and execution time. The calculation of one frequency point in VC (SI)
integral equation method requires approximately 2 (1) seconds on one
core of a workstation using Mathematica \citep{mathematica}. In contrast
DDA calculations requires around 8 minutes per frequency point on
the same workstation. HFSS needs around 6 minutes per frequency point
if the mesh is optimized for one frequency only and is reused without
optimization for 30 other frequencies. With FDTD one gets the complete
spectra in one run in around 250 minutes on one core. In conclusion
the newly proposed 1D semi-analytical methods provide a speed-up in
execution time close to 200 times compared to general 3D Maxwell solvers.
Additionally up to 100 times less RAM is required for the semi-analytical
calculations.
In figure~\ref{fig:1st_resonances} the position of the scattering
cross-section maxima are shown for gold nanowires with different aspect
ratios. Radius is $a=10nm$ and length varies from $l=50nm$ (shown
in the inset) to $l=300nm$. Normal incidence ($\xi=\frac{\pi}{2}$)
with $\left|\mathbf{E}^{inc}\right|=1$ is considered. VC integral equation method (full black line), SI integral
equation method (dashed line) and DDA (dotted line) are compared.
As it is expected the accuracy of both semi-analytical methods deteriorates
with decreasing aspect ratio. However even for considerable low aspect
ratio $\frac{5}{4}$ ($l=50nm$) the first resonance peak predicted
using approximate methods represents the rigorous numerical result
with a good relative accuracy (2\% of the central frequency).
In general VC integral equation method demonstrate better agreement
with the DDA calculations in comparison with the SI method (Fig.~\ref{fig:comparison_numeric}
and \ref{fig:map}). This can be best seen in figure~\ref{fig:map},
where the resonance frequencies of increasing order (from left to
right) against the wire length are depicted for nanowires with radius
$a=10nm$ (top panel) and $a=20nm$ (bottom panel). The VC integral
equation method (solid line), SI integral equation method (dashed
line) are presented. Dots represent DDA results. The brighter (darker)
curves are the resonances of even (odd) orders. All calculations are
done under slanting incidence ($\xi=\frac{\pi}{4}$). The better accuracy
of the VC integral equation method can be systematical traced in figure~\ref{fig:map}
especially for the structures with smaller aspect ratio.
\section{Conclusion}
Two alternative methods to solve the scattering problem on optical
nanowire antenna, the volume current integral equation (VC-IE) method
and the surface impedance integral equation (SI-IE) method are introduced.
In order to reduce the general 3D volume integral equation describing
the scattering problem to a simple semi-analytical 1D integro-differential
equation, both methods utilize solutions of the problem of plane wave
scattering on infinite cylinder. A regularization and discretization
scheme is proposed in order to transform integro-differential equations
into solely integral equation. This transformation enables to solve
the original problem without necessity to impose additional boundary
conditions at the nanowire edges. Numerical evaluation of the proposed
methods and their comparison with different numerically rigorous methods
is presented for scattering cross-section calculations. Gold nanowires
are analyzed at optical and near-infrared spectral range. The introduced
one-dimensional semi-analytical methods demonstrate good agreement
and superior numerical performance in comparison with rigorous numerical
methods.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
\label{sec:intro}
To date, $\sim275$ SNRs have been identified in the Galaxy (\citealt{green09}). The majority of these are detected as steep spectrum ($\alpha\sim-0.5$ where $S_{\nu}$=$\nu^{\alpha}$) extended radio synchrotron sources. Acceleration of the synchrotron electrons to GeV-TeV energies occurs at the shock boundary (\citealt{reynolds81}). The ambient gas is heated at the shock to X-ray emitting temperatures (\citealt{shklovskii68}; \citealt{grader70}; \citealt{chevalier77}), and radiative cooling of the gas in slower shocks produces optical emission in {H$\alpha$}, S{\footnotesize II}$~$~ and O{\footnotesize III}$~$~ lines (\citealt{fesen85}). Observations of the thermal X-ray and optical emission can be used to derive fundamental properties such as the age of the remnant and the local gas density. Unfortunately, the study of Galactic SNRs is impeded by two factors. The high absorption in the direction of many of these sources means that only radio emission is detected and much of the broad-band information is lost. In addition, the distance measurements to sources within the Galaxy are often uncertain by a factor of two, leading to an uncertainty of a factor of 4 in luminosity and 5.5 in supernova explosion energy (\citealt{borkowski01}). It is therefore desirable to find populations of SNRs in nearby galaxies at known distances, so that their properties can be determined with greater certainty.
The Small Magellanic Cloud (SMC) is a very good target for study, as its
proximity (60 kpc, \citealt{karachentsev04}) allows the
morphology of SNRs and SNR candidates to be resolved while the low foreground Galactic
absorption column ($6\times10^{20}$ cm$^{-2}$, \citealt{dickey90})
permits detection of emission in optical and soft X-ray bands.
23 sources classified as SNRs have been discovered in the SMC
(\citealt{filipovic05}; \citealt{payne07}). Most of these have been
found with associated soft X-ray emission, exhibiting a variety of
morphological structures. A synoptic study of the X-ray
properties of thirteen of these remnants was completed by
\citet{vdh04} (hereafter VDH04). \citet{filipovic08}
presented {\it XMM-Newton~\/} results from another three known and one candidate remnant.
These studies found that the majority of
SMC SNRs are likely in the Sedov-adiabatic phase of evolution.
\begin{table*}
\caption{Details of the {\it XMM-Newton~\/} observations covering IKT 16 used in this study.}
\centering
\begin{tabular}{ccccccccc}
\hline
Observation ID & Start Date & Filter$^{a}$ & \multicolumn{2}{c}{Pointing co-ordinates} & \multicolumn{3}{c}{Useful exposure (ks)} & Off-axis angle \\
& (yyyy-mm-dd) & pn/MOS1/MOS2 & RA (J2000) & Dec (J2000) & pn & MOS 1 & MOS2 & (arcmin) \\
\hline
0018540101 & 2001-11-20 & M/M/M & $00^h59^m26.8^s$ & $-72\deg09\arcm55\hbox{$^{\prime\prime}$~\/}$ & 23.4$^{b}$ & 18.1$^{b}$ & 18.0$^{b}$ & 9.3 \\
0084200101 & 2002-03-30 & T/M/M & $00^h56^m41.7^s$ & $-72\deg20\arcm11\hbox{$^{\prime\prime}$~\/}$ & 8.8 & 10.0 & 10.3 & 8.0 \\
0110000201 & 2000-10-17 & M/M/M & $00^h59^m26.0^s$ & $-72\deg10\arcm11\hbox{$^{\prime\prime}$~\/}$ & 14.3$^{b}$ & 16.7$^{b}$ & 16.8$^{b}$ & 9.1 \\
0212282601 & 2005-03-27 & C/M/M & $00^h59^m26.8^s$ & $-72\deg09\arcm54\hbox{$^{\prime\prime}$~\/}$ & 0.0 & 3.8 & 3.8 & 9.3 \\
0304250401 & 2005-11-27 & M/M/M & $00^h59^m26.8^s$ & $-72\deg09\arcm54\hbox{$^{\prime\prime}$~\/}$ & 15.9 & 17.4$^{b}$ & 17.4$^{b}$ & 9.3 \\
0304250501 & 2005-11-29 & M/M/M & $00^h59^m26.8^s$ & $-72\deg09\arcm54\hbox{$^{\prime\prime}$~\/}$ & 14.9 & 16.5$^{b}$ & 16.6$^{b}$ & 9.3 \\
0304250601 & 2005-12-11 & M/M/M & $00^h59^m26.8^s$ & $-72\deg09\arcm54\hbox{$^{\prime\prime}$~\/}$ & 10.6 & 16.4$^{b}$ & 16.2$^{b}$ & 9.3 \\
0500980201 & 2007-06-06 & T/M/M & $01^h00^m00.0^s$ & $-72\deg27\arcm00\hbox{$^{\prime\prime}$~\/}$ & 14.8 & 23.2 & 23.9$^{b}$ & 11.8 \\
0601210801 & 2009-10-09 & T/M/M & $00^h56^m15.5^s$ & $-72\deg21\arcm55\hbox{$^{\prime\prime}$~\/}$ & 23.0$^{b}$ & 24.6 & 24.6$^{b}$ & 10.4 \\
\hline
\end{tabular}
\\
$^{a}$ - T = thin filter, M = Medium filter, C = Filter-wheel Closed. \\
$^{b}$ - Observation used for spectral analysis. \\
\label{table:ikt16obs}
\end{table*}
IKT 16 was discovered as a candidate supernova remnant in the SMC by
\citet{inoue83} using {\it Einstein}, and its identity as a
shell-type SNR was confirmed through radio and {H$\alpha$}~ analysis
(\citealt{mathewson84}). VDH04 included this remnant as part of their
study of SMC SNRs. Using the earliest {\it XMM-Newton~\/} observation, they
found it to be X-ray faint, and its properties were hence poorly constrained.
A region of harder X-ray emission close to the centre of the remnant, whose
source was unclear, was also reported.
Nine {\it XMM-Newton~\/} observations
covering IKT 16 now
exist, giving an eightfold increase in exposure time.
In this paper we present a study of IKT 16 and the object located towards
its centre using X-ray data from {\it XMM-Newton}, radio data from the Australia
Telescope Compact Array (ATCA)
and optical images from the Magellanic Clouds Emission Line Survey (MCELS).
In \S\ref{sec:obs}, we outline the procedures
used to reduce and analyse {\it XMM-Newton~\/} data. We also discuss complementary radio and
optical data. In \S\ref{sec:results}, we present the results of spatial and
spectral analysis of IKT 16 and the X-ray hard object inside it. In
\S\ref{sec:disc}, we discuss the derived properties of IKT 16 and speculate
on the nature of the hard source. We present our conclusions in
\S\ref{sec:conc}.
\begin{figure*}
\centering
\includegraphics[angle=0,width=85mm]{xrayfinal3.eps}
\includegraphics[angle=0,width=91mm]{ratiofinalb.eps}
\includegraphics[angle=0,width=85mm]{radio20cmfinal.eps}
\includegraphics[angle=0,width=91mm]{radio3cmfinalb.eps}
\caption{
{\it Top left}: Lightly smoothed RGB {\it XMM-Newton~\/} EPIC image of IKT~16 (R=0.3--0.8 keV, G=0.8--1.2 keV, B=1.2--2.0 keV). The two regions marked are those used for spectral analysis for the SNR (white circle) and the bright source (black circle). The centre of the remnant is marked with a cross. The image is binned in $4\hbox{$^{\prime\prime}$~\/}\times4\hbox{$^{\prime\prime}$~\/}$ pixels, and is displayed logarithmically with a dynamic range of 20.
{\it Top right}: MCELS S{\footnotesize II}$~$/H$\alpha$ ratio of IKT~16. The image is binned in $2\hbox{$^{\prime\prime}$~\/}\times2\hbox{$^{\prime\prime}$~\/}$ pixels. The white circle indicates the extent of the SNR.
{\it Bottom left}: Lightly smoothed soft X-ray image (0.3--1 keV) overlaid with high-resolution 20~cm (1400~MHz) ATCA radio contours. The contours are at levels of 0.07, 0.14 and 0.28~mJy~beam$^{-1}$. The X-ray image is displayed logarithmically with a dynamic range of 20, and the X-ray colour bar is shown in units of log(pn+MOS) ct/pixel/20 ks.
{\it Bottom right}: Soft X-ray map overlaid with 3~cm (8640~MHz) ATCA radio contours. The contours are at levels of 0.25, 0.5 and 1.0~mJy~beam$^{-1}$. The beam sizes for the radio observations are shown in Table~\ref{table:radio}.
}
\label{fig:ikt16images}
\end{figure*}
\section{Observations and data reduction}
\label{sec:obs}
\subsection{X-ray observations}
\label{sec:xobs}
{\it XMM-Newton~\/} has observed IKT 16 serendipitously with EPIC in full-frame mode
on 9 occasions between 2000 and 2009, at off-axis angles between 8\hbox{$^{\prime}$~\/} and 12\hbox{$^{\prime}$}.
Eight of the nine observations were used
for analysis, with one (obsid 0212282601) discarded due to
persistent high background.
Details of the observations are shown in Table~\ref{table:ikt16obs}.
Data reduction was based on SAS v10.0.0\footnote{Science Analysis Software,
http://xmm.esac.esa.int/sas/}.
Datasets were screened for periods of high background
through the creation of full-field single event (PATTERN=0) lightcurves above 10 keV
(with an upper limit of 12 keV for pn data). MOS data were excluded when the
count-rate in a 25 second bin exceeded 0.4${\rm ~count~s^{-1}}$, with pn data excluded above
0.5${\rm ~count~s^{-1}}$. These thresholds were selected based on
quiescent levels in the observations. After filtering, useful exposure times
range between 9 and 25 ks.
For each observation, images and exposure maps were created in several energy
bands: 0.3--0.8 keV, 0.8--1.2 keV and 1.2--2.0 keV for production of 3-colour
X-ray images, and 0.3--1.0 keV for a broader-band soft X-ray image.
We used data from single and double pixel events in the pn camera
(patterns 0--4) and single to quadruple pixel events in MOS cameras
(patterns 0--12). The selection FLAG=0 was used to exclude spurious data
from bad pixels, hot columns and chip gaps, and the pixel size was set
at 4\hbox{$^{\prime\prime}$} $\times $4\hbox{$^{\prime\prime}$}. A constant particle rate,
estimated from the corners of each image not exposed to the sky, was
subtracted from each image. The images for each camera were then co-added
using the SAS task {\it emosaic} in order to account for differences
in pointing direction between the observations, and the exposure
maps were added in the same way. Finally, the mosaic image for each camera was
divided by the relevant exposure map and the images from the three cameras
summed to produce a flat-field image suitable for further spatial analysis.
Spectral analysis was performed on 15 exposures (3 pn, 5 MOS1 and 7 MOS2),
selected on the basis of exposure time and position on the detector. When studying
extended sources, it is necessary to be careful in defining source
and background areas to extract. In this case there are
two areas of interest: the central source (detected as a point source in our
X-ray observations) located at RA(J2000)=$00^h58^m16.7^s$ Dec=$-72\deg18\arcm06\hbox{$^{\prime\prime}$~\/}$
and the remainder of the supernova remnant. The shock
boundary of the SNR is well-described by a circular
region centred on RA=$00^h58^m22.4^s$ Dec=$-72\deg17\arcm52\hbox{$^{\prime\prime}$~\/}$
with a radius of $128\pm8$\hbox{$^{\prime\prime}$}, with the error due to
uncertainty in the best fit circle radius. We therefore defined
two source regions for each observation: a 20$\hbox{$^{\prime\prime}$~\/}$ radius area centred on
the bright source and a 128$\hbox{$^{\prime\prime}$~\/}$ radius area centred on our best estimate
of the centre of the SNR, excluding the previously defined source area.
These extraction regions are shown on Fig.~\ref{fig:ikt16images} (top left). For
the first region, a background region was extracted from a nearby position
on the same CCD outside the SNR. For the second, a nearby background region
of similar size
to the source region was found such that the average exposure time for source
and background regions was the same. This was done to ensure that the
contributions from the cosmic X-ray background (vignetted by the telescope)
and particle background (not vignetted to first order) remained in the correct
proportion when applied to the source spectrum. The SAS tasks {\it arfgen} and
{\it rmfgen} were used to generate appropriate Auxiliary Response Files (ARF)
and Response Matrix Files (RMF) for each observation, and the counts in
adjacent spectral bins were summed to a minimum of 20 counts per bin
using the Ftool\footnote{http://heasarc.gsfc.nasa.gov/ftools/} {\it grppha}.
\subsection{Optical observations}
\label{sec:optobs}
The Magellanic Cloud Emission Line Survey (MCELS) was carried out using the University of Michigan/CTIO Curtis Schmidt telescope. Using narrow band filters corresponding to {H$\alpha$} (6563\AA), S{\footnotesize II}$~$ (6724\AA) and O{\footnotesize III}$~$ (5007\AA) line emission, plus red and green continuum bands (used to subtract most of the stars from the images), images of diffuse structure across both Magellanic Clouds have been produced with an angular resolution of 2.4\hbox{$^{\prime\prime}$~\/}/pixel. An image of the S{\footnotesize II}$~$/{H$\alpha$}~ ratio across the remnant is presented in Fig.~\ref{fig:ikt16images} (top right). Further information about the MCELS is given in \citet{smith00}.
Integral Field Spectroscopy (IFS) was performed on 3~November, 2010 at Siding Springs Observatory using the 2.3-m Advanced Technology Telescope and its Wide Field Spectrograph (WiFeS). WiFeS provides a 25\hbox{$^{\prime\prime}$~\/} $\times$ 38\hbox{$^{\prime\prime}$~\/} field with 0.5\hbox{$^{\prime\prime}$~\/} per pixel spatial sampling along each of 25\hbox{$^{\prime\prime}$~\/} $\times$ 1\hbox{$^{\prime\prime}$~\/} slits. The output format match two 4096$\times$4096 pixel CCD detectors in each camera individually optimized for the blue and red ends of the spectrum. The 900~second single exposure was made in the central region of IKT~16 at position angle (east of north) 90 degrees under clear skies with seeing estimated at 1\arcsec. The exposure was made in classical equal mode\footnote{This is a principal WiFeS data accumulation mode in which data were accumulated in the red and blue cameras on the source for equal times.} using the RT560 beam splitter and 3000 Volume Phased Holographic (VPH) gratings. For these gratings, the blue (708~lines~mm$^{-1}$) range includes 3200--5900\AA\ while the red (398~lines~mm$^{-1}$) range includes 5300--9800\AA.
The data were reduced using the WiFeS data reduction pipeline based on NOAO (National Optical Astronomy Observatory) IRAF software. This data reduction package was developed from the Gemini IRAF package (\citealt{2003SPIE.4841.1581M}). The pipeline consists of four primary tasks: {\it wifes} to set environment parameters, {\it wftable} to convert single extension FITS file formats to Multi-Extension FiTS ones and create file lists used by subsequent steps, {\it wfcal} to process calibration frames including bias, flat-field, arc and wire; and {\it wfreduce} to apply calibration files and create data cubes for analysis.
\subsection{Radio observations}
\label{sec:radioobs}
We have extracted all archival radio-continuum observations of IKT~16 using the Australia Telescope Compact Array (ATCA) comprising observations at 4 radio wavelengths: 20, 13, 6 and 3~cm for ATCA (\citealt{dickel01}; \citealt{filipovic02}; \citealt{2004MNRAS.355...44P}; \citealt{1997A&AS..121..321F}). In addition we make use of the Molonglo Observatory Synthesis Telescope (MOST) SMC survey at 36~cm (843 MHz) \citep{turtle98}. The beam size for these observations ranges from 7{\hbox{$^{\prime\prime}$}} to 45{\hbox{$^{\prime\prime}$}}. IKT~16 was also observed with the ATCA as a part of projects C281 and C634. More information about observing procedure and other sources observed during these sessions can be found in \citet{bojicic07}, \citet{bojicic10} and \citet{2008SerAJ.176...59C}. Soft X-ray images of IKT~16 overlaid with 3~cm and high-resolution 20~cm radio contours are shown in Fig.~\ref{fig:ikt16images}, and the integrated flux densities detected from our radio observations of the SNR are shown in Table~\ref{table:radio}. The error in the measured integrated flux density (for the whole system and the bright source) is estimated as a quadrature sum from the local noise level (0.1~mJy~beam$^{-1}$) and the uncertainty in the gain calibration (10\%). For the SNR region (excluding the bright source) the error is calculated by adding the errors for the other two extraction regions in quadrature.
\begin{figure}[t]
\centering
\rotatebox{0}{\scalebox{0.5}{\includegraphics{ikt-new2.eps}}}
\caption{Radio spectra of the entire SNR (blue), the bright source near the centre
of the remnant (red) and the remainder of the remnant (green), constructed from
radio observations at 5 frequencies using MOST and ATCA. The best fit
power-law fits are shown, and the data is also shown in Table~\ref{table:radio}.
}
\label{fig:radiospec}
\end{figure}
\begin{table*}[t]
\caption{Radio Integrated Flux Density of IKT 16.}
\begin{center}
\begin{tabular}{ccccccl}
\hline
$\nu$ & $\lambda$ & Beam Size & S$_\mathrm{Total}$ & S$_\mathrm{Source}$ & S$_\mathrm{SNR}$ & Telescope/Project \\
(MHz) & (cm) & (arcsec) & (Jy) &(Jy) & (Jy) &\\
\hline
843 & 36 & 44.9$\times$43 & 0.087 & 0.045 & 0.042 & MOST\\
1371 & 20 & 7.05$\times$6.63 & 0.071 & 0.039 & 0.032 & C281\\
2378 & 13 & 7.02$\times$6.58 & 0.058 & 0.038 & 0.020 & C281\\
4798 & 6 & 30$\times$30 & 0.041 & 0.033 & 0.008 & C634, Parkes\\
8640 & 3 & 20$\times$20 & 0.030 & 0.028 & 0.002 & C634, Parkes\\
\hline
\end{tabular}
\end{center}
\label{table:radio}
\end{table*}
\section{Results}
\label{sec:results}
\subsection{X-ray}
\subsubsection{Morphology}
\label{sec:morph}
Fig.~\ref{fig:ikt16images} shows a 3-colour X-ray (0.3--0.8 keV, 0.8--1.2 keV and 1.2--2.0 keV) image of IKT~16, along with radio images from ATCA and optical data from MCELS. The X-ray images are lightly smoothed with a Gaussian kernel of width 4{\hbox{$^{\prime\prime}$}}. We define the boundary of the SNR to be where the surface brightness of the X-ray emission abruptly falls to the background level. Using this definition, the radius of the remnant is estimated to be $128\hbox{$^{\prime\prime}$~\/}\pm8${\hbox{$^{\prime\prime}$}}, with the error resulting from a 2-pixel uncertainty in fitting a circular region to the edge of the X-ray emission. At the assumed distance of the SMC (60~kpc), this corresponds to a radius of 37$\pm$3~pc, consistent with being an older remnant in the Sedov phase of evolution (\citealt{cox72}). The measured radius makes it the largest known SNR in the SMC. The shock boundary is well-defined in the north and west of the SNR, where the diffuse X-ray emission is strongest. We find small variations in the temperature measured across the remnant, with the region of strongest X-ray emission in the north slightly hotter than its surroundings. More tenuous emission is observed to the east of the remnant, and there is some evidence for greater extension of the SNR in this direction. Depending on the direction of measurement, the linear radius of the SNR varies between 122\hbox{$^{\prime\prime}$~\/} (in the N-S direction) and 144\hbox{$^{\prime\prime}$~\/} (NE-SW), implying that the remnant is not perfectly circular on the sky. These deviations are relatively small and our simplification of the geometry of the system does not have a large effect on the results we obtain. These variations do however lead to uncertainty in the position of the SNR centre, which limits the precision with which we can measure the dynamics of the system.
The bright source observed inside the SNR shell is significantly harder than the emission from the rest of the remnant, as seen in the 3-colour image and its X-ray spectrum (in Section~\ref{sec:xspec}). We find no evidence for this source being extended in {\it XMM-Newton~\/} measurements. We therefore limit its maximum X-ray extent to the FWHM of the detector (6{\hbox{$^{\prime\prime}$}}), which is 1.7~pc at the distance of the SMC. Assuming that the SNR shell can be described as spherically symmetric, the position of the bright source is offset from the centre of the shell by $30\hbox{$^{\prime\prime}$~\/}\pm8\hbox{$^{\prime\prime}$~\/}$ (with the error mostly due to uncertainty in the position of the SNR centre). Not taking into account possible projection effects, this corresponds to a linear distance of $8\pm2$~pc. We note that the bright source has been previously identified in a 9 ks {\it Chandra~\/} observation (OBSID 2948; \citealt{evans10}). The {\it Chandra~\/} X-ray position is consistent with that found here, but due to the extremely short observation time and large off-axis angle in that observation, the errors on the {\it XMM-Newton~\/} position are smaller. The observation also suggests that the source may be slightly extended (with an extent of 1.1\hbox{$^{\prime\prime}$}$\pm$0.6\hbox{$^{\prime\prime}$}). A longer on-axis observation is necessary to constrain this value precisely.
\subsubsection{Spectral analysis}
\label{sec:xspec}
Spectral analysis was performed using XSPEC v12.5.1 (\citealt{arnaud96}). For each of the relevant observations, source and background spectra were extracted as described in the previous section. All 15 spectra extracted for each of the two extraction regions were fitted simultaneously, allowing only for a renormalisation factor between observations from different cameras. To account for photoelectric absorption by interstellar gas, two absorption components were used in fitting: a {\it phabs} component fixed at the foreground Galactic value of $6\times10^{20}{\rm ~erg~s^{-1}}$ (\citealt{dickey90}) assuming elemental abundances of \citet{wilms00}, and a variable absorption ({\it vphabs}) component to account for absorption inside the SMC. This second component has metal abundances fixed at 0.2 solar, as is typical in the SMC (\citealt{russell92}).
X-ray emission from the central region was fit with an absorbed power-law, and emission from the remainder of the SNR was fit with a Sedov model (\citealt{borkowski01}). Due to the relatively broad point spread function of the {\it XMM-Newton~\/} EPIC detector, we estimate that for a point source, $\sim$25\% of the source counts will fall outside an extraction region of $20\hbox{$^{\prime\prime}$~\/}$. Therefore, an extra power-law component is introduced to the SNR fit, with index fixed to that obtained through fitting the central region and normalization set to account for this spillover. Similarly, as the central extraction region encompasses 2.5\% of the overall SNR region, an appropriately scaled Sedov component is added to the central source model, with parameters fixed to those derived for the SNR.
The free parameters of the Sedov model are the mean shock temperature, electron temperature just behind the shock front, metal abundances, ionization age (electron density behind the shock front multiplied by the remnant age) and normalization. Providing that the SNR has spherical symmetry and is in the Sedov phase of evolution, the fits obtained can be used to estimate several physical parameters. Using the distance to the SNR ($D$) and radius ($R$ in m, assuming a distance of 60~kpc), volume of X-ray emitting material ($V$ in m$^3$), normalization derived from the XSPEC fit ($N$), shock temperature ($T_S$ in keV) and baryon number per hydrogen atom ($r_m\approx1.4$, assuming a helium/hydrogen ratio of 0.1), it is possible to derive the electron density ($n_e$), age of the remnant ($t_{dyn}$), total emitting mass ($M$), initial explosion energy ($E_0$) and ionization age ($I_t$) through the following system of equations (VDH04, from \citealt{borkowski01}):
\begin{equation}
\label{eq:ne}
n_e = \sqrt{\frac{3D^2N}{10^{-24}R^3}}, ~$m$^{-3}
\end{equation}
\begin{equation}
\label{eq:age}
t_{dyn} = \frac{1.3\times10^{-14}R}{\sqrt{T_S}}, ~$yr$
\end{equation}
\begin{equation}
\label{eq:mass}
M = 5\times10^{-31}m_pr_mn_eV, \hbox{$\rm ~M_{\odot}$}
\end{equation}
\begin{equation}
\label{eq:energy}
E_0 = 2.64 \times 10^{-8} T_sR^3n_e, ~$erg$
\end{equation}
\begin{equation}
\label{eq:ion}
I_t = 4\times10^{-6} n_et_{dyn}, ~$m$^{-3} $s$
\end{equation}
\begin{figure*}
\centering
\rotatebox{270}{\scalebox{0.4}{\includegraphics{SNRspec.ps}}}
\rotatebox{270}{\scalebox{0.44}{\includegraphics{brightsourcespec.eps}}}
\caption{EPIC spectra from IKT 16 (left) and the bright source located near its centre (right). In both cases the solid line corresponds to the best-fit spectral model (see text and Table~\ref{table:specfit}), with the upper line showing the combined spectra from pn observations and the lower line combined spectra from MOS1 and MOS2 observations. The $\chi^{2}$ residuals with respect to the best-fitting model are also shown. }
\label{fig:xrayspec}
\end{figure*}
\begin{table*}
\caption{Parameters of the best-fit model to SNR and bright source regions.}
\centering
\begin{tabular}{lccccccc}
\hline
Region & SMC Absorption & Power-law Cont. & \multicolumn{2}{c}{Sedov model} & Goodness & Unabsorbed L$_{\rm X}$$^{a}$ \\
& $10^{21}$cm$^{-2}$ & Photon Index & Shock T (keV) & Ion. time ($10^{10}$cm$^{-3}$s) & of Fit & 0.5--10 keV \\
& & Normalization & Normalization & & $\chi^{2}$/dof & ($10^{35}{\rm ~erg~s^{-1}}$) \\
\hline
SNR & 3.4(f) & 1.58(f) & 1.03$\pm$0.12 & 6.1 & 1103/1032 & 1.6$\pm$0.4 \\
& & $1.0\times 10^{-5}$(f) & $1.4\pm0.3\times 10^{-4}$ & & & & \\
\\
Bright Source & 3.4$\pm$0.6 & 1.58$\pm$0.07 & 1.03(f) & 6.1(f) & 150/137 & 1.6$\pm$0.2 \\
& & $3.4\pm0.4\times 10^{-5}$ & $3.6\times 10^{-6}$(f) & & & \\
\\
\hline
\end{tabular}
\\
$^{a}$ - Includes correction for spillover of bright source photons into SNR extraction area and vice versa. \\
(f) - Parameter fixed for consistency between fit regions. \\
\label{table:specfit}
\end{table*}
We note that our distance estimate of 60 kpc to IKT 16 may not be entirely accurate due to the large line of sight extent of the SMC, which is measured to be 4-6 kpc in this region of the galaxy (\citealt{subramanian09}). Errors in the distance measurement will also impact the measured radius of the SNR ($\propto$ D) and its volume ($\propto$ D$^3$), and the physical properties derived from the above system of equations will be affected.
Details of the best fit spectra obtained are given in Table~\ref{table:specfit} and the combined pn and MOS spectra for both extraction regions are shown in Fig.~\ref{fig:xrayspec}. Due to the limited quality of the spectral data, shock and electron temperatures are assumed to be the same (valid as an evolved SNR should be approaching electron-ion equilibration). The metal abundances in the fits are fixed at 0.2 solar for the same reason. Since lines, whose presence is required to fit the ionization age, are not easily discernable in the X-ray spectrum, this parameter was calculated based on the above system of equations. Spectral fitting was then repeated with the calculated ionization age to derive new Sedov parameters, and this process was repeated until convergence was achieved. The best fit physical parameters from the Sedov model fit are shown in Table~\ref{table:sedov}. The best fit models found require a significant absorption component within the SMC in addition to foreground Galactic absorption. We initially left this absorption as a free parameter for both regions. In this case, we find the best fit Sedov model for the SNR to have an electron temperature of 0.71~keV with absorption inside the SMC of $4.6\times10^{21}$ cm$^{-2}$ (${\chi^{2}}=1100.7/1032$), outside the error bounds of the absorption found in the bright source spectrum. Setting the SNR absorption at the level found for the bright source does not significantly worsen the fit (${\chi^{2}}=1103.2/1033$), but it does alter the best fit temperature (see Table~\ref{table:specfit}). We fix the absorption in the SNR to the level of the bright source, as is expected if the two emission regions are physically connected, and derive the Sedov parameters for the SNR based on this fit (see Table~\ref{table:sedov}). If we instead fix the absorption for the bright source at the level found for the SNR, we obtain a significantly worse fit for this component (${\chi^{2}} = 164/137$). In this case, we find the power-law index of the source to be somewhat steeper ($\Gamma = 1.74\pm0.08$). For both components, the average flux across all observations is measured to be $2.7\times10^{-13}{\rm ~erg~cm^{-2}~s^{-1}}$ (0.5--10 keV), corresponding to an unabsorbed luminosity of $1.6\times10^{35}{\rm ~erg~s^{-1}}$. The flux for both components is found to be constant between observations to within $15\%$. There is thus no evidence for any long-term variability of the point source.
The X-ray spectrum of the bright source within the SNR is typical of many astrophysical sources. The possible nature of this source will be discussed in Section~\ref{sec:discsrc}.
\subsection{Optical}
\label{sec:optical}
Using data from MCELS, we find that the S{\footnotesize II}$~$/H$\alpha$ ratio found in the SNR is $\sim0.25$ (Fig.~\ref{fig:ikt16images}, top right), significantly below the standard ``minimum'' ratio typically found in radiative shocks. This implies that there is no evidence for radiative shocks around the SNR. In known SMC SNRs, this ratio is typically found to be higher than 0.4 (\citealt{payne07}).
Similar results are found from analysis of the central region of IKT~16. Using QFitsView3\footnote{Written by Thomas Ott and freely available at www.mpe.mpg.de/~ott/dpuser/index.html}, a one-dimensional spectrum of the diffuse region excluding all visible stars was created (Fig.~\ref{fig:ratio2}). Analysis required the identification of cosmic rays on the 2-dimensional spectra while the IRAF task {\it splot} allowed determination of emission line counts based on gaussian fits. Standard errors were calculated based on an average rms noise of 600 counts; propagation of error methods were used for the addition or division of line fluxes. Using this method, we find a S{\footnotesize II}$~$/{H$\alpha$}\ ratio of 0.23 with a 2.1\% error. It is quite possible that there are very faint nebulae interspersed between the stars that do not meet the criteria for shocked regions. We cannot exclude the possibility of an H{\small II}$~$ region being present in the line of sight.
\begin{figure}
\centering
\includegraphics[width=88mm]{diffusespectrum.eps}
\caption{WiFeS spectra of the central region of IKT~16.}
\label{fig:ratio2}
\end{figure}
\subsection{Radio}
\label{sec:radio}
The available radio images show IKT~16 to be a very radio-weak SNR. There is evidence of faint diffuse radio structure throughout the remnant. The images are dominated by a source corresponding to the X-ray bright source, which appears to extend a distance of $\sim30\hbox{$^{\prime\prime}$~\/}$ towards the centre of the remnant. There is no evidence for a radio point source located within this emission. Table~\ref{table:radio} shows the integrated flux density measurements at frequencies from 843~MHz -- 8640 MHz. The overall radio spectrum (Fig.~\ref{fig:radiospec}) is well-described by a power-law with index $\alpha=-0.45\pm0.02$. The radio spectral index of the central source is found to be flatter ($-0.19\pm0.02$) than that of the SNR itself ($-0.97\pm0.35$). The SNR spectral index is unusually steep, especially given that IKT~16 is most likely an older (evolved) SNR, due to its rather large size. Usually, a steep gradient like this would suggest a much younger and more energetic SNR. However, in this case, the steepness can be contributed to missing short spacings at higher radio-continuum frequencies (4800 and 8640~MHz) and therefore missing flux. Specifically at 8640~MHz (where the ATCA primary beam is $\sim$300\arcsec) this SNR edges would be positioned close to the primary beam boundary where the flux tends to be significantly uncertain. We note that if we omit the two highest frequency data points from the SNR fit, the radio spectral index is --0.72, which is more in line with that found in other SNRs in the SMC (\citealt{filipovic05}).
\section{Discussion}
\label{sec:disc}
The only significant study of the X-ray properties of this remnant to date has been completed by VDH04. The {\it XMM-Newton~\/} observations covering IKT~16 since publication of this paper have significantly improved the available statistics. This allows us to examine the properties of the SNR more accurately, as well as enabling study of the nature of the bright source located close to its centre.
\subsection{Properties of the SNR}
\label{sec:discsnr}
In the analysis of IKT~16 conducted by VDH04, the bright source emission area was excluded from analysis, and the remaining emission was fit with both a single-temperature non-equilibrium ionization (NEI) model and a Sedov model. As we have shown, simply excluding the source emission area may not produce an accurate measure of the properties of the SNR, since $\sim25\%$ of the flux detected through this method will be from the point source. As this source is significantly harder than the SNR as a whole, this unduly influences the results obtained. Additionally, the previous study estimates the radius of the SNR to be $100\hbox{$^{\prime\prime}$~\/}$ compared with $128\hbox{$^{\prime\prime}$~\/}$ here. The new estimate is based on mosaicing of all available observations, through which the X-ray shock boundary is more clearly defined in all directions. This changes the estimate of the X-ray emitting volume, which also affects the derived Sedov parameters.
The Sedov properties obtained for both studies are shown in Table~\ref{table:sedov}. The properties we find justify our use of the Sedov approximation, as the swept-up mass significantly exceeds the ejecta mass but radiative cooling has not occurred yet. The parameters we derive are significantly different from those in VDH04. The temperature we find is cooler, which we attribute to the different methods of dealing with the point source mentioned in the previous paragraph. Along with our larger estimate of the radius of the remnant, this leads to a higher dynamical age, larger swept-up mass and lower initial explosion energy (as expected from Eq.\ref{eq:ne}-\ref{eq:energy}).
In order to test whether this explains the difference in our results, we attempted to fit the SNR emission region with a Sedov model without accounting for spillover from the bright source. We find a good fit (${\chi^{2}} = 1099/1032$) with a higher shock temperature of $1.30\pm0.20$ keV. This is consistent with the value found by VDH04 ($1.76\pm0.65$ keV).
VDH04 noted that the initial explosion energy appears high in comparison to typical SNRs, which they suggest is due to the remnant resulting from collapse of a massive star within a low-density circumstellar cavity. Our estimate of the explosion energy ($\sim10^{51}{\rm ~erg~s^{-1}}$) is much more in line with that derived for other SNRs in the SMC.
Compared with other SNRs in the SMC, IKT~16 is notably X-ray faint and is detected at higher electron temperature ($\sim1.0$ keV vs. $\sim0.2$ keV). While this is the largest confirmed SNR in the SMC there are several larger cousins in the Large Magellanic Cloud (LMC) \citep{2009SerAJ.179...55C,2010ApJ...725.2281K}. The radio emission from the SNR is also found to be faint (32 mJy at 1400~MHz; Table~\ref{table:radio}). The large size of the remnant, considering that it is still in the Sedov phase of evolution, suggest that the density of the SNR environment is low. This may also contribute to its X-ray and radio faintness. The low density implied is consistent with a SNR
explosion inside a wind blown bubble, which are typically found inside molecular clouds.
Such sites are associated with massive star formation, which would increase the likelihood
of this system being a core-collapse SNR and the central source being associated with it.
\citet{muller10} have carried out a survey of CO emission across the northern SMC,
finding a series of molecular clouds near the position of IKT 16. Their presence strongly
supports this scenario. The apparent X-ray faintness of the remnant may be additionally
enhanced by high absorption attenuating the soft X-ray signal.
Using a map of H{\small I}$~$ density from \citet{stanimirovic99}, the total column density of the SMC in the direction of IKT~16 is $5.7\times10^{21}$cm$^{-2}$. The value we derive from spectral fitting is $\approx$60\% of this value, indicating that IKT 16 is located deep within the SMC.
The soft X-ray image shows that the distribution of X-ray emission appears somewhat uneven inside the remnant, with more emission observed towards the northern shock boundary than to the south.
The radio images also show more emission to the north of the remnant.
The 3-colour X-ray image (Fig.~\ref{fig:ikt16images}, top left) suggests that the X-ray emission in the north may be at a higher temperature than in the rest of the remnant. This may be due to a combination of several factors, including fluctuations in the SMC absorption column, a higher concentration of SNR ejecta increasing the local metallicity or true variations in the plasma temperature of the ISM. Due to the limited X-ray statistics, it is not possible to carry out spatially resolved spectral analysis to investigate this further.
\begin{table}[t]
\caption{Physical properties of IKT~16 from best fit Sedov model.}
\centering
\begin{tabular}{lcccc}
\hline
Parameter & Units & This paper & VDH04 \\
\hline
Temperature & keV & 1.03 & 1.76 \\
Radius & pc & 37 & 29 \\
Electron Density & cm$^{-3}$ & 0.03 & 0.05/0.04$^{a}$ \\
Dynamical age & yr & 14700 & 7500 \\
Swept-up mass & \hbox{$\rm ~M_{\odot}$} & 232 & 124 \\
Explosion energy & $10^{51}$ erg & 1.2 & 3.4 \\
\hline
\end{tabular}
\\
$^{a}$ - VDH04 derive electron and proton densities separately, assumed to be approximately the same here.
\label{table:sedov}
\end{table}
\subsection{Nature of the Bright source}
\label{sec:discsrc}
The X-ray spectrum of the bright source is consistent with that of several different astrophysical objects, including X-ray binary systems, pulsar wind nebulae and background active galactic nuclei. Here, we attempt to discriminate between these possibilities.
In fitting a power law model to the source, a significant absorption component is required in addition to foreground Galactic absorption. This suggests that it is not a Galactic object located along the line of sight.
It is possible, however, that the source is a background AGN. The average power-law photon index of a type I AGN is measured to be $\sim1.9$ (\citealt{pounds90}), somewhat softer than what we see here. However, the absorption required in our spectral fit ($3.4\pm0.6\times10^{21}$ cm$^{-2}$) is significantly lower than that of the SMC along our line of sight, making this unlikely. We therefore attempted to fit a more complex model commonly found to fit AGN spectra: a power law plus a 0.1 keV black body component (\citealt{turner89}). Here, we find a reasonable fit with an extra absorption component of $3\pm1\times10^{21}$ cm$^{-2}$. The X-ray flux detected in this model is $\sim3\times10^{-13}{\rm ~erg~cm^{-2}~s^{-1}}$.
Studies of type~I AGN typically show the fluxes in X-ray (0.5--10 keV) and optical bands to lie within approximately an order of magnitude of each other (\citealt{maccacaro88}). For a source with this observed X-ray flux, we therefore expect there to be an optical counterpart at a V magnitude of between 14.8 and 20.8. We searched the Optical Gravitational Lensing Experiment (OGLE) (\citealt{udalski98}) and Magellanic Clouds Photometric Survey (\citealt{zaritsky02}) to try to find the optical counterpart. The closest match found to the X-ray position (searching down to a V band optical magnitude of 21.5) is at a distance of $2.5\hbox{$^{\prime\prime}$~\/}$, 2 times the combined errors in optical and X-ray positions. However, the two closest optical sources in the catalogue have optical colours consistent with stars within the SMC. The closest source with optical colours atypical for a main sequence star is at a distance of $3.2\hbox{$^{\prime\prime}$~\/}$, still within 3 sigma of our X-ray source position. This is a possible optical counterpart to our source.
We can also estimate the probability of finding a background AGN coincident with the SNR position. Studies of the log N-log S distribution of AGN ({\it e.g.~\/} \citealt{rosati02}) show that above our estimate of the unabsorbed X-ray flux of this object ($\sim4\times10^{-13}{\rm ~erg~cm^{-2}~s^{-1}}$, 0.5-10 keV) we expect to find 3 AGN per square degree of the sky. The probability of finding one such source within 30\hbox{$^{\prime\prime}$~\/} of the centre of IKT 16 is 0.07\%. Although we are unable to rule out an AGN origin for this source, we conclude that it is statistically unlikely. We assume from this point that the source is located inside the SMC.
As shown in \S\ref{sec:results}, the SNR shell is well-described by a circle of radius $128\hbox{$^{\prime\prime}$~\/}$. Presuming that the well-defined profile of the western shock boundary can be used to find the centre of the remnant, the bright source is located $\approx30\pm8\hbox{$^{\prime\prime}$~\/}$ (a transverse distance of $8\pm2$~pc) away from this point. Providing it is associated with the supernova event, it has traversed this distance in the dynamical age of the remnant (14700 yr), at an average speed of $580\pm100$ km s$^{-1}$. The direction of travel is not necessarily perpendicular to us, so this is a lower limit. However, unless the direction of travel is significantly out of this plane, the velocity of the object will not be much greater than this value. Asymmetry in SNR explosions typically gives the resulting compact object a random space velocity of $350-550$ km s$^{-1}$ (\citealt{lyne94}), so it is reasonable to suggest that the bright X-ray source is related to this object. There are two known objects which may produce the emission observed: a pulsar wind nebula or a microquasar.
Pulsar wind nebulae (PWNe) are generated by dissipation of a pulsar's rotational energy through relativistic winds, producing synchrotron emission visible in radio to X-ray bands. For a review of the properties of these systems, see \citet{gaensler06}. $\sim50$ PWNe have been observed in the Galaxy and the LMC, and the majority are associated with known SNRs (``Composite'' SNR, \citealt{kaspi06}). The X-ray spectrum of known PWNe is well-described by a power-law of index ranging between $\Gamma\sim1.1-2.3$ (\citealt{cheng04}), consistent with the power law fit found here. The measured X-ray luminosity of $1.6\times10^{35}{\rm ~erg~s^{-1}}$ is also similar to that of known PWNe in the Galaxy and the LMC ({\it e.g.~\/} \citealt{gaensler03}; \citealt{porquet03}; \citealt{slane04}). In our observations, the maximum extension of the X-ray source was found to be $6\hbox{$^{\prime\prime}$~\/}$, which corresponds to a radius of $\sim1$ pc at the distance of the SMC. \citet{gelfand09} find that in the spherically symmetric case, a pulsar wind nebula inside a Sedov-phase remnant at an age of 15~kyr should be undergoing compression due to the SNR reverse shock to a radius of $<1$ pc. This would be detected as a point source in {\it XMM-Newton~\/} observations, consistent with our measurements. The radius measured in the {\it Chandra~\/} observation is $0.3\pm0.2$ pc, consistent with this scenario. In this case, the movement of the source away from the centre of the SNR shell implies that spherical symmetry no longer holds, but as the distance travelled is only $\sim10\%$ of the shell diameter, this scenario may still be valid. Unfortunately, the limited spatial resolution of {\it XMM-Newton~\/} and the distance to the SMC (60~kpc) preclude identification of a central pulsar, which would confirm the identity of this source as a pulsar wind nebula. The relatively low X-ray flux also prevents us from conducting timing analysis to search for pulsations.
The ATCA images in Fig.\ref{fig:ikt16images} show the source to have a radio extent of $\sim40\hbox{$^{\prime\prime}$~\/}\times30\hbox{$^{\prime\prime}$~\/}$, elongated in the direction of the centre of the remnant. This extent is significantly greater than that observed in X-rays. This may be due to the longer cooling time for radio emission compared to that of X-ray emission. The radio emission may therefore trace the path of the nebula as it moves from the centre of the SNR, as is observed in ``relic'' PWN ({\it e.g.~\/} Vela X, \citealt{frail97}). The radio spectral index of this source is measured to be $\sim-0.2$, consistent with the typical radio spectrum found in PWNe (\citealt{gaensler06}). The absence of a radio point source in this emission rules out the presence of an established radio pulsar. However, confirmed PWNe ({\it e.g.~\/} LMC~SNR~B0453-685 and LMC~SNR~N\,206; \citet{gaensler03b,2005ApJ...628..704W}; Crawford et al. (in prep)) have been found where the central pulsar is not detected in radio observations, so this does not discount a PWN origin.
The other possibility we consider here is a microquasar travelling through the SMC, where the extended radio emission is due to optically thin synchrotron radiation in relativistic jets. The X-ray emission in this case would come from an X-ray binary (XRB) system with a power-law of index 1.5--2.2, with radio jets observed at a spectral index between --0.8 and --0.4 ({\it e.g.~\/} \citealt{fender06}; \citealt{hannikainen06}). The X-ray power-law slope is consistent with this scenario, but the radio spectrum is somewhat flatter than expected ($\alpha\sim-0.2$). If it is associated with the SNR, and providing the western shock boundary can be used to estimate the original site of the supernova explosion, the velocity we infer for this object ($580\pm100$ km s$^{-1}$) is very high compared with the average kick velocity for even Low-Mass X-ray Binary (LMXB) systems ($180\pm80$ km s$^{-1}$; \citealt{brandt95}). In fact, the majority of known microquasars are High-Mass X-ray Binary (HMXB) systems, for which the corresponding typical kick velocity is 50 km s$^{-1}$ (also from \citet{brandt95}). It is possible that the object may be within the SMC but not associated with the SNR, in which case this velocity measurement is not applicable. The relative stability of the X-ray flux between observations is somewhat unexpected, as many X-ray binary systems will exhibit state changes on a timescale of 10 years. From the available X-ray and radio evidence, we conclude that it is unlikely that the bright source is a microquasar. We note that no LMXB systems are currently known in the SMC, and they are very rare in the Galaxy. This also decreases the likelihood of this scenario.
On the basis of current observations, it is most likely that the source is a PWN. However, we cannot entirely rule out a microquasar or AGN. A future deep {\it Chandra~\/} observation will be able to determine the X-ray extent of this source accurately, and will thus conclusively determine its nature.
\section{Conclusions}
\label{sec:conc}
We have carried out analysis of the spatial and spectral properties of the SMC SNR IKT~16, using archive and proprietary data from {\it XMM-Newton}, radio data from ATCA and optical images from MCELS. The properties of the SNR are consistent with it being in the Sedov-adiabatic phase of evolution. The dynamic age of the remnant and initial explosion energy are typical of the population of SNRs in the SMC.
We find the X-ray point source near the centre of IKT~16 to be non-thermal, with an X-ray spectral index of $\sim1.6$. This source is coincident with a radio source which extends a distance of $\sim30\hbox{$^{\prime\prime}$~\/}$ in the direction of the centre of the SNR. We conclude that the X-ray source is unlikely to be a background AGN, due to the small probability of spatial coincidence with the SNR and the absence of an obvious optical counterpart. Two astrophysical objects are considered which may explain the X-ray and radio emission observed: a pulsar wind nebula or microquasar. The flatness of the radio spectrum, the one-sided morphology of the radio emission and inferred velocity of the source strongly favour a PWN origin, which may be confirmed by a future {\it Chandra~\/} observation. If this is the case, IKT 16 is the second largest such system (after the Vela SNR) found to date.
\begin{acknowledgements}
The {\it XMM~\/}\ project is supported by the Bundesministerium f\"ur Wirtschaft und Technologie/Deutsches Zentrum f\"ur Luft- und Raumfahrt (BMWI/DLR, FKZ 50 OX 0001) and the Max-Planck Society. RS acknowledges support from the BMWI/DLR grant FKZ 50 OR 0907. SM and AT acknowledge financial contribution from the agreement ASI-INAF I/009/10/0. We used the Karma/MIRIAD software packages developed by the ATNF. The Australia Telescope Compact Array is part of the Australia Telescope which is funded by the Commonwealth of Australia for operation as a National Facility managed by the CSIRO. We thank the Magellanic Clouds Emission Line Survey (MCELS) team for access to the optical images.
We thank the anonymous referee for their comments, which have helped to improve this manuscript significantly.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
While elementary particles have
mass and charge,
the origin still remains unknown.
When the generation of masses
is related to symmetry breaking,
the corresponding problem is how symmetry is broken.
In the standard model, the Higgs mechanism
describes physics of elementary
particles very well up to the weak scale.
However, it has been found in the own framework
of the standard model that
this picture is not valid at higher
energy scales.
In addition, the standard model does not have
any explanation for quantization of charge.
Hence,
we need a deeper understanding for mass and charge,
or symmetry breaking and charge
beyond the standard model.
If extra dimensions are included in an effective
theory,
the strength of force would behave
unlike four dimensions.
The gauge bosons of the standard model
propagating in extra dimensions
can rapidly become strongly coupled and
form scalar bound states of quarks and leptons.
The self-breaking of the standard-model
gauge symmetry was proposed in Ref.~\cite{ArkaniHamed:2000hv}.
The authors proposed that
the existence of a Higgs doublet is a consequence of
the standard-model gauge symmetry and three generation
of quarks and leptons
provided the gauge bosons and fermions
propagate in appropriate extra dimensions compactified
at a TeV scale.
It has been shown earlier
that electroweak symmetry may be broken
by fields propagating in extra dimensions
in Ref.~\cite{Dobrescu:1998dg}\cite{Cheng:1999bg}.
Also in the Randall-Sundrum warped space
\cite{Randall:1999ee}\cite{Randall:1999vf},
transition of the strength of force
and the type of bound states have been studied
\cite{Davoudiasl:1999tf}
\cite{Uekusa:2011ud}.
As for quantization of charge,
grand unification
has attracted much attention.
The charges belong to subgroups of a unification group.
Quantum numbers for quarks and leptons
are fixed by the group structure.
The SU(3), SU(2) and U(1) couplings meet at a single
point and the couplings
are evaded from blowup of
the values by renormalization group evolution.
Beyond the unification scale,
the energy dependence of
the three couplings are shown
graphically as one identical line
due to contributions of twelve $X$ and $Y$ bosons.
Thus, a grand unification with extra dimensions
might contain the generation of masses
and the quantization of charges automatically.
In this paper,
we study most attractive channel
for an SU(5) grand unification in five dimensions.
Our setup is that components in
each SU(5) multiplet obey
the same boundary condition with respect
to extra dimensions
although gauge symmetry breaking
by boundary conditions might
be an interesting possibility
\cite{Kawamura:2000ev}
\cite{Kawamura:2000ir}.
The energy dependence
of couplings in gauge theories
with extra dimensions
has been found
\cite{Dienes:1998vh}
\cite{Dienes:1998vg}
and has been examined in detail
\cite{Bhattacharyya:2006ym}
\cite{Uekusa:2007im}.
In particular,
the energy dependence of
the three couplings should be shown as
one identical line
beyond the unification scale.
It has been emphasized in
Ref.~\cite{Uekusa:2008iz} that
this is not the case
for gauge symmetry breaking by boundary conditions
where the same boundary conditions do not span
irreducible representations of the unification group.
In order to keep gauge coupling
unification above the unification scale,
we consider
only the self-breaking for the
origin of symmetry breaking.
The fundamental fields are
a gauge boson and fermions.
A gauge boson belongs to {\bf 24}.
Quarks and leptons are included in
{\bf 5*} and {\bf 10} Dirac fermions
whose left-handed components have
zero mode.
If these fields are only objects
propagating in the bulk,
there are no
scalar bound states composed of zero mode.
When an anomaly-free set of fields with
zero mode for the right-handed component
is added,
we find scalar bound states
whose binding strengths are
the largest for
{\bf 24} and the second largest
for {\bf 5},
correspondingly to
the breaking pattern
SU(5) $\to$ SU(3)$\times$SU(2)$\times$U(1)
$\to$ SU(3)$\times$U(1).
The paper is organized as follows.
In Section~\ref{memo},
we start with the basics such as
the scenario,
framework, field content and group structure.
With consideration for anomaly-free sets,
scalar bound states and the binding strengths
are given
in Section~\ref{sec:mini}.
Here we find various composites.
The simple model to form {\bf 24} and {\bf 5}
for appropriate symmetry breaking
is proposed.
Furthermore, the composites {\bf 50} and {\bf 75}
are discussed for the doublet-triplet splitting.
We conclude in Section~\ref{concl}.
\section{Basics of the model \label{memo}}
We consider a gauge theory
with two branes on fixed points
of the orbifold $S^1/Z_2$ in five dimensions.
Bulk fields can have strong coupling compared with
the corresponding four-dimensional theory
so that the ordinary four-dimensional massless gluon
and quarks
does not give rise to gauge symmetry breaking
and that the five-dimensional effect
may change
potentials for composites.
If
non-vanishing vacuum expectation values
are generated,
symmetry is broken.
The pattern of symmetry breaking depends
on the attractive force of composites
and the masses of the constituents.
The analysis can be applied
for both of flat spacetime and warped spacetime.
In the present model, there are no fundamental scalar
fields.
The gauge boson $A_M$ has
the four-dimensional vector component $A_\mu$
and the four-dimensional scalar component $A_5$.
The boundary conditions at the location
of the branes are given
by Neumann for all $A_\mu$ and
Dirichlet for all $A_5$.
Quarks and leptons are included
as zero mode for {\bf 5*}
and {\bf 10} Dirac fermions,
$\psi_{5*}$ and $\psi_{10}$.
For the left-handed component,
the Neumann boundary condition is imposed.
Since singlet neutrinos do not produce attractive force,
they can be omitted in the present analysis.
For these boundary conditions,
fields which belong to
irreducible representations of SU(5)
obey the same boundary conditions
so that at higher energy scales
above the scale of
symmetry breaking by any condensation
the theory behave as SU(5).
For simplicity, the number of the generation is
chosen as one.
The source of symmetry breaking
is scalar bound states represented by
the form $\bar{\psi}_I\psi_J$ where
$I,J$ are denoted as the species of fermions.
The attractive force is dominantly originated
from exchange of a gauge boson between fermions
and is determined by the gauge coupling and the group structure.
As for the group structure, we take the notation
$\textrm{tr}(t_r^a t_r^b) = C(r)
\delta^{ab}$
where $t_r^a$ is denoted as
the representation matrices in the irreducible
representation $r$
and $C(r)$ is a constant for each representation $r$.
The relation
$d(r) C_2(r) = d(\textrm{Adj}) C(r)$
is fulfilled where
$C_2(r)$ is the quadratic Casimir operator
for each representation and
$d$ is the dimension of each representation.
The values $C$, $C_2$ and $d$ are summarized
for several irreducible representations
of SU($N$)
in Table~\ref{tab:sun}.
For $N=5$,
the diagonal generator
$t^8 =
\textrm{diag}(2, 2, 2, -3, -3 )
/(2\sqrt{15})$
plays the same role as the U(1) hypercharge
up to a overall factor.
We call its charge $Q_Y$.
For SU(3)$\times$SU(2)$\times$U(1),
$\psi_{5*}$ and $\psi_{10}$
are decomposed into
({\bf 3*}, {\bf 1})${}_{Q_Y=-\sqrt{15}/15}$
$\oplus$
({\bf 1}, {\bf 2})${}_{Q_Y=\sqrt{15}/10}$
and
({\bf 3}, {\bf 2})${}_{Q_Y=-\sqrt{15}/30}$
$\oplus$
({\bf 3*}, {\bf 1})${}_{Q_Y=2\sqrt{15}/15}$
$\oplus$
({\bf 1}, {\bf 1})${}_{Q_Y=-\sqrt{15}/5}$, respectively.
\begin{table}[htb]
\caption{SU($N$) property. Each box in Young tableaux
is
symmetric for the horizontal direction
and antisymmetric for the vertical direction.
\label{tab:sun}}
\begin{center}
\begin{tabular}{cc|ccc}
\hline \hline
&& $d$ & $C$ & $C_2$ \\
\hline
\framebox{}
&& $N$ & ${1\over 2}$ & ${N^2-1\over 2N}$ \\
\begin{picture}(5,15)(0,0)
\put(0,6.5){\framebox{}}
\put(0,0){\framebox{}}
\end{picture}
&&
${N(N-1)\over 2}$
& ${N-2\over 2}$
& ${(N+1)(N-2)\over N}$
\\
\begin{picture}(10,15)(0,0)
\put(0,0){\framebox{}}
\put(6.5,0){\framebox{}}
\end{picture}
&&
${N(N+1)\over 2}$
& ${N+2\over 2}$
& ${(N-1)(N+2)\over N}$
\\
\begin{picture}(5,25)(0,0)
\put(0,0){\framebox{}}
\put(0,6.5){\framebox{}}
\put(0,13){\framebox{}}
\end{picture}
&&
${N(N-1)(N-2)\over 6}$
& ${(N-2)(N-3)\over 4}$
& ${3(N-3)(N+1)\over 2N}$
\\
\begin{picture}(10,25)(0,0)
\put(0,0){\framebox{}}
\put(0,6.5){\framebox{}}
\put(6.5,6.5){\framebox{}}
\end{picture}
&&
${N(N^2-1)\over 6}$
& ${(N^2-3)\over 2}$
& ${3(N^2-3)\over 2N}$
\\
\begin{picture}(15,15)(0,0)
\put(0,0){\framebox{}}
\put(6.5,0){\framebox{}}
\put(13,0){\framebox{}}
\end{picture}
&&
${N(N+1)(N+2)\over 6}$
& ${(N+2)(N+3)\over 4}$
& ${3(N-1)(N+3)\over 2N}$
\\
\begin{picture}(5,15)(0,0)
\put(0,0){Adj}
\end{picture}
&&
$N^2-1$
& $N$
& $N$
\\
\hline
\end{tabular}
\end{center}
\end{table}
For the quark and lepton multiplets,
possible combinations of the bound states are
$\bar{\psi}_{5*}\psi_{5*}$,
$\bar{\psi}_{10}\psi_{10}$ and
$\bar{\psi}_{5*}\psi_{10}$.
They are written in terms of four-dimensional fields
dependent on the five-dimensional coordinates
as chirality mixing operators
$\bar{\psi}_{L}\psi_{R} +
\bar{\psi}_{R}\psi_{L}$.
Because $\psi_R$ consists of Kaluza-Klein massive
mode without zero mode,
it is difficult to have potentials
for bound states
to yield nonzero vacuum expectation values.
Hence, we need to add new fields with
zero mode for the right-handed component.
In order that $\psi_{5*}$ and $\psi_{10}$ do not
form brane mass terms together with the new fields,
candidates are except for {\bf 5*} and {\bf 10}.
A singlet is also excluded
because it does not contribute to attractive force.
For forming the fundamental {\bf 5} or
the anti-fundamental {\bf 5*}
scalar bound states together with
quark and lepton multiplets,
possible bulk fermions are {\bf 5},
{\bf 10*} and {\bf 24}.
For forming an adjoint {\bf 24} scalar bound state
together with quark and lepton multiplets,
possible bulk fermions are {\bf 15},
{\bf 40*} and {\bf 45}.
Added new fermions have zero mode only for
the right-handed component.
In general they are chiral in four dimensions.
Therefore new fields need to be added in
such a way that anomaly is canceled.
An anomaly coefficient $A(r)$ is defined by
\bea
\textrm{tr}\left[t^a_r\{t^b_r,t^c_r\}\right]
= {1\over 2} A(r) d^{abc}
\eea
where
$\{t_N^a, t_N^b\}
={1\over N} \delta^{ab} + d^{abc} t_N^c$
for SU($N$).
For $N=5$, the generators associated with
$Q_Y$ yield $d^{888}=-1/\sqrt{15}$.
The anomaly coefficients for the above irreducible
representations are given by
\bea
A({\bf 5}) = -{1\over 2},~
A({\bf 10*}) = {1\over 2},~
A({\bf 15}) = {9\over 2}, ~
A({\bf 24}) = 0, ~
A({\bf 40*}) = -8 ,~
A({\bf 45}) = 3 .
\eea
Thus the minimal anomaly-free set
to obtain {\bf 5} or {\bf 5*} scalar bound states
and a {\bf 24} scalar bound state
is ({\bf 10*}, {\bf 15},
{\bf 40*}, {\bf 45}) Dirac fermions.
In the next section, we will examine the binding
strengths for bound states in the most attractive
channel approximation.
\section{Composites and binding strengths
\label{sec:mini}}
In this section
we derive
possible bound states
and their quantum numbers and binding strengths
by adding bulk fields
with zero mode for the right-handed component, $\chi$.
Here $\bar{\psi} \chi$ includes
$\bar{\psi}_L^{(0)} \chi_R^{(0)}$
composed of only zero mode
and strong attractive force
can lead to potentials with phase transition.
First we examine
the minimal anomaly-free set.
Taking into account the pattern of
symmetry breaking more,
we study a simple anomaly-free set.
In addition, the {\bf 50} and {\bf 75}
bound states associated with the
doublet-triplet splitting are discussed.
\subsubsection*{The minimal anomaly-free set}
The quark and lepton multiplets are
$\psi_{5*}$ and $\psi_{10}$.
The minimal set of
addition fermions is given by
$\chi_{10*}$, $\chi_{15}$,
$\chi_{40*}$ and $\chi_{45}$.
It is necessary to examine
the binding strengths for eight possible scalar
bound states
\bea
\bar{\psi}_{5*} \chi_{10*} , ~
\bar{\psi}_{5*} \chi_{15} , ~
\bar{\psi}_{5*} \chi_{40*} , ~
\bar{\psi}_{5*} \chi_{45} ,
~
\bar{\psi}_{10} \chi_{10*} ,~
\bar{\psi}_{10} \chi_{15} , ~
\bar{\psi}_{10} \chi_{40*} ,~
\bar{\psi}_{10} \chi_{45} .
\label{psichi}
\eea
They are decomposed further in terms of
irreducible representations into twenty four
bound states
\bea
{\bf 5} \otimes {\bf 10*}
&\!\!\!=\!\!\!&
{\bf 5*} \oplus {\bf 45} ,
\\
{\bf 5} \otimes {\bf 15}
&\!\!\!=\!\!\!&
{\bf 35} \oplus {\bf 40} , \label{5-15}
\\
{\bf 5} \otimes {\bf 40*}
&\!\!\!=\!\!\!&
{\bf 10*} \oplus {\bf 15*} \oplus
{\bf 175^*_\textrm{\scriptsize A}} ,
\\
{\bf 5} \otimes {\bf 45}
&\!\!\!=\!\!\!&
{\bf 24} \oplus {\bf 75} \oplus {\bf 126} ,
\\
{\bf 10*} \otimes {\bf 10*}
&\!\!\! =\!\!\!&
{\bf 5} \oplus {\bf 45*} \oplus {\bf 50*} ,
\\
{\bf 10*} \otimes {\bf 15}
&\!\!\!=\!\!\!&
{\bf 24} \oplus {\bf 126} , \label{10-15}
\\
{\bf 10*} \otimes {\bf 40*}
&\!\!\!=\!\!\!&
{\bf 24} \oplus {\bf 75}
\oplus {\bf 126*} \oplus
{\bf 175^*_{\textrm{\scriptsize B}}},
\\
{\bf 10*} \otimes {\bf 45}
&\!\!\!=\!\!\!&
{\bf 10} \oplus {\bf 15}\oplus
{\bf 40*} \oplus {\bf 175} \oplus {\bf 210} ,
\eea
The binding strength for $\bar{\psi}\chi$
is given by \cite{Raby:1979my}\cite{ArkaniHamed:2000hv}
\bea
{1\over 2}g^2 \left[ C_2(\bar{\psi})
+ C_2 (\chi) - C_2(\bar{\psi}\chi) \right] .
\eea
The coupling constant $g$ is common
for types of fields in the present
model.
For SU(5), the values
of $C_2$ for the irreducible representations
appearing here are summarized in Table~\ref{tab:c2}.
\begin{table}[htb]
\caption{The quadratic Casimir operator $C_2$
for each representation $r$.
\label{tab:c2}}
\begin{center}
\vspace{-2ex}
\begin{tabular}{c|ccccc ccccc ccc}
\hline\hline
$r$ &
{\bf 5} &
{\bf 10} &
{\bf 15} &
{\bf 24} &
{\bf 35} &
{\bf 40} &
{\bf 45} &
{\bf 50} &
{\bf 75} &
{\bf 126} &
{\bf 175${}_{\textrm{\scriptsize A}}$} &
{\bf 175${}_{\textrm{\scriptsize B}}$} &
{\bf 210}
\\ \hline
$C_2$ &
${12\over 5}$ &
${18\over 5}$ &
${28\over 5}$ &
5 &
${48\over 5}$ &
${33\over 5}$ &
${32\over 5}$ &
${42\over 5}$ &
8 &
10 &
${48\over 5}$ &
12 &
${90\over 7}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
Using these values,
we find the binding strengths for the twenty four
bound states as shown
in Table~\ref{tab:bind}.
\begin{table}[htb]
\caption{Binding strengths for $\bar{\psi}\chi$
for the minimal set in unit of $g^2$.
\label{tab:bind}}
\vspace{-2ex}
\begin{center}
\begin{tabular}{ccc}
\hline\hline
Composite &
Constituents &
Binding \\
&&
strength
\\
\hline
{\bf 10} &
$\bar{\psi}_{10} \chi_{45}$ &
$16/5$
\\
{\bf 10*} &
$\bar{\psi}_{5*} \chi_{40*}$ &
$27/10$
\\
{\bf 24} &
$\bar{\psi}_{10} \chi_{40*}$ &
$13/5$
\\
{\bf 5} &
$\bar{\psi}_{10} \chi_{10*}$ &
$12/5$
\\
{\bf 15} &
$\bar{\psi}_{10} \chi_{45}$ &
$11/5$
\\
{\bf 24} &
$\bar{\psi}_{10} \chi_{15}$ &
$21/10$
\\
{\bf 24} &
$\bar{\psi}_{5*} \chi_{45}$ &
$19/10$
\\
{\bf 5*} &
$\bar{\psi}_{5*} \chi_{10*}$ &
$9/5$
\\
{\bf 15*} &
$\bar{\psi}_{5*} \chi_{40*}$ &
$17/10$
\\
{\bf 40*} &
$\bar{\psi}_{10} \chi_{45}$ &
$17/10$
\\
{\bf 75} &
$\bar{\psi}_{10} \chi_{40*}$ &
$11/10$
\\
{\bf 40} &
$\bar{\psi}_{5*} \chi_{15}$ &
$7/10$
\\ \hline
\end{tabular}
~
\begin{tabular}{ccc}
\hline\hline
Composite &
Constituents &
Binding \\
&&
strength
\\
\hline
{\bf 75} &
$\bar{\psi}_{5*} \chi_{45}$ &
$2/5$
\\
{\bf 45*} &
$\bar{\psi}_{10} \chi_{10*}$ &
$2/5$
\\
{\bf 175${}_{\textrm{\scriptsize A}}$}&
$\bar{\psi}_{10} \chi_{45}$ &
$1/5$
\\
{\bf 126*} &
$\bar{\psi}_{10} \chi_{40*}$ &
$1/10$
\\
{\bf 45} &
$\bar{\psi}_{5*} \chi_{10*}$ &
$-1/5$
\\
{\bf 175${}_{\textrm{\scriptsize A}}$*} &
$\bar{\psi}_{5*} \chi_{40*}$ &
$-3/10$
\\
{\bf 126} &
$\bar{\psi}_{10} \chi_{15}$ &
$-2/5$
\\
{\bf 126} &
$\bar{\psi}_{5*} \chi_{45}$ &
$-3/5$
\\
{\bf 50*} &
$\bar{\psi}_{10} \chi_{10*}$ &
$-3/5$
\\
{\bf 35} &
$\bar{\psi}_{5*} \chi_{15}$ &
$-4/5$
\\
{\bf 175${}_{\textrm{\scriptsize B}}$*} &
$\bar{\psi}_{10} \chi_{40*}$ &
$-9/10$
\\
{\bf 210} &
$\bar{\psi}_{10} \chi_{45}$ &
$-10/7$
\\ \hline
\end{tabular}
\end{center}
\end{table}
For the minimal field content, it is found that
the binding strengths for {\bf 10} and {\bf 10*} are
larger than that of the adjoint {\bf 24}.
Usually SU(5) symmetry is broken by
{\bf 24} or {\bf 75}.
The results given in
Table~\ref{tab:bind} means
SU(5) symmetry would not be broken desirably
unless there is some mechanism such that
the binding for {\bf 10} and {\bf 10*}
decreases effectively
or the binding for {\bf 24} increases.
In the following
we consider a non-minimal but simple anomaly-free set
where this problem does not arise.
\subsubsection*{The simple anomaly-free set}
As given in the previous section,
a {\bf 24} scalar bound state can be made of
{\bf 15}, {\bf 40*} and {\bf 45} Dirac fermions
with quark and lepton multiplets.
From Table~\ref{tab:bind}, it is read that
a composite {\bf 24} has the largest binding strength
only for a {\bf 15} $\chi$ field and
the other composites are largest for different
quantum numbers:
the composite {\bf 10*} for
a {\bf 40*} $\chi$ field and
the composite {\bf 10} for
a {\bf 45} $\chi$ field.
Therefore the fermion with zero mode for
the right-handed component
to form a {\bf 24} scalar bound state is uniquely
fixed as a {\bf 15} $\chi$ field.
One of the simplest anomaly-free fermion content
with {\bf 15}
is 9 $\chi_5$ and 1 $\chi_{15}$.
The composite are given by
$\bar{\psi}_{5*}\chi_5$
$\bar{\psi}_{5*}\chi_{15}$
$\bar{\psi}_{10}\chi_5$
and $\bar{\psi}_{10}\chi_{15}$
as well as 8 copies of $\chi_5$.
They are decomposed into eight
irreducible representations,
\bea
{\bf 5} \otimes
{\bf 5} &\!\!\!=\!\!\!&
{\bf 10} \oplus {\bf 15} ,
\\
{\bf 10*} \otimes {\bf 5} &\!\!\!=\!\!\!&
{\bf 5*} \oplus {\bf 45} ,
\eea
with Eqs.~(\ref{5-15}) and (\ref{10-15}).
For this simple set, the binding strengths are
given in Table~\ref{tab:binds}.
\begin{table}[htb]
\caption{Binding strengths for $\bar{\psi}\chi$
for the simple set in unit of $g^2$.
\label{tab:binds}}
\vspace{-2ex}
\begin{center}
\begin{tabular}{ccc}
\hline\hline
Composite &
Constituents &
Binding \\
&&
strength
\\
\hline
{\bf 24} &
$\bar{\psi}_{10} \chi_{15}$ &
$21/10$
\\
{\bf 5*} &
$\bar{\psi}_{10} \chi_{5}$ &
$9/5$
\\
{\bf 40} &
$\bar{\psi}_{5*} \chi_{15}$ &
$7/10$
\\
{\bf 10} &
$\bar{\psi}_{5*} \chi_{5}$ &
$3/5$
\\ \hline
\end{tabular}
~
\begin{tabular}{ccc}
\hline\hline
Composite &
Constituents &
Binding \\
&&
strength
\\
\hline
{\bf 45} &
$\bar{\psi}_{10} \chi_{5}$ &
$-1/5$
\\
{\bf 126} &
$\bar{\psi}_{10} \chi_{15}$ &
$-2/5$
\\
{\bf 15} &
$\bar{\psi}_{5*} \chi_{5}$ &
$-2/5$
\\
{\bf 35} &
$\bar{\psi}_{5*} \chi_{15}$ &
$-4/5$
\\ \hline
\end{tabular}
\end{center}
\end{table}
We find scalar bound states
whose binding strengths are
the largest for
{\bf 24} in $\bar{\psi}_{10} \chi_{15}$
and the second largest
for {\bf 5} in
$\bar{\chi}_{5}\psi_{10}$.
The largest strength for {\bf 24}
corresponds to the breaking of the
SU(5) to SU(3)$\times$SU(2)$\times$U(1).
After the SU(5) is broken,
nine {\bf 5} and {\bf 5*} composite scalar
fields contribute
for the breaking of
SU(2)$\times$U(1) to U(1).
Now we compare
the binding strengths given in Table~\ref{tab:binds}
with the results of the standard model gauge group.
When quarks $Q,U,D$, leptons $L,E,N$
and SU(3)$\times$SU(2)$\times$U(1) gauge bosons
propagate in five dimensions,
the binding strengths for the composites
with zero mode are given in Table~\ref{tab:bs}.
Here $Q,L$ are SU(2) doublets and
$U,D,E,N$ are SU(2) singlets.
\begin{table}[htb]
\begin{center}
\caption{Binding strength for the standard model gauge
group \cite{Uekusa:2011ud}.
\label{tab:bs}}
\begin{tabular}{c|cccc}
\hline\hline
Composite
& Constituents
& SU(3)$\times$ SU(2)$\times$U(1)
& Binding
& Binding for
\\
&
& representation
& strength
& $\sqrt{3\over 5}\,g_1 = g_2=g_3\equiv g$
\\
\hline
$H_1$ & $\bar{Q}U$
& $({\bf 1}, {\bf 2},{1\over 2})$
& ${4\over 3} g_3^2 + {1\over 9} g_1^2$
& (7/5) $g^2$
\\
$H_2$ & $\bar{Q}D$
& $({\bf 1}, {\bf 2},-{1\over 2})$
& ${4\over 3} g_3^2 -{1\over 18} g_1^2$
& (13/10) $g^2$
\\ \hline\hline
$S_{10}$ & $\bar{E}Q$
& $({\bf 3}, {\bf 2},{7\over 6})$
& $-{1\over 6}g_1^2$
& ($-1/10$) $g^2$
\\
$S_{11}$ & $\bar{L}U$
& $({\bf 3}, {\bf 2},{5\over 6})$
& 0 & 0
\\
$S_{12}$ & $\bar{L}D$
& $({\bf 3}, {\bf 2},{1\over 6})$
& ${1\over 6}g_1^2$
& (1/10) $g^2$
\\
$S_{13}$ & $\bar{L}E$
& $({\bf 1}, {\bf 2},-{1\over 2})$
& ${1\over 2} g_1^2$
& (3/10) $g^2$
\\
$S_{14}$
& \multicolumn{2}{l}{
$N$ or $\bar{N}$ is included}
& 0 & 0
\\
\hline
\end{tabular}
\end{center}
\end{table}
From Tables~\ref{tab:binds} and \ref{tab:bs},
the binding strengths for {\bf 24} and {\bf 5*}
are larger than that of the Higgs doublet $H_1$.
Because the self-breaking of the standard model
gauge group has been claimed
\cite{ArkaniHamed:2000hv}
\cite{Dobrescu:1998dg}
\cite{Cheng:1999bg},
the symmetry breaking associated
with the larger binding strengths
can occur.
In other words, the symmetry breaking of
SU(5) to color and electromagnetism can be expected.
\subsubsection*{On the composites 50, 50* and 75}
A solution to the doublet-triplet splitting
has been to break the
SU(5) symmetry by the real representation
{\bf 75} instead of {\bf 24}
and add {\bf 50} and {\bf 50*}
\cite{Masiero:1982fe}
\cite{Grinstein:1982um}.
Now we consider the composites
{\bf 50}, {\bf 50*} and {\bf 75}.
First we examine the composite {\bf 75}.
For quark-lepton multiplets
$\psi_{5*}$ and $\psi_{10}$,
fermions $\chi$
with the representations {\bf 10}, {\bf 40},
{\bf 45} and {\bf 50}
can form {\bf 75} bound states as
\bea
{\bf 5} \otimes {\bf 45}
&\!\!\!=\!\!\!& {\bf 24} \oplus {\bf 75} \oplus {\bf 126},
\label{45}
\\
{\bf 5} \otimes {\bf 50}
&\!\!\!=\!\!\!&
{\bf 75} \oplus {\bf 175_{\textrm{\bf\scriptsize B}}} ,
\\
{\bf 10*} \otimes {\bf 10}
&\!\!\!=\!\!\!&
{\bf 1} \oplus {\bf 24} \oplus {\bf 75} ,
\\
{\bf 10*} \otimes {\bf 40*}
&\!\!\!=\!\!\!&
{\bf 24} \oplus {\bf 75} \oplus
{\bf 126*} \oplus {\bf 175_{\textrm{\bf \scriptsize B}}*} .
\eea
Among these constituents,
only the {\bf 50} $\chi$ field
yields a {\bf 75} bound state
without the adjoint {\bf 24}.
The binding strengths
for the constituents $\bar{\psi}_{5*}\chi_{50}$
are 12/5 for the composite {\bf 75}
and $-3/5$ for the composite
{\bf 175${}_{\textrm{\scriptsize B}}$}.
This point is favorable.
However,
in addition to the coupling with $\psi_{5*}$,
$\bar{\psi}_{10}\chi_{50}$
needs to be taken into account
where
${\bf 10*} \otimes
{\bf 50} = {\bf 10} \oplus {\bf 175}
\oplus {\bf 315}$.
The binding strength
for the constituents $\bar{\psi}_{10}\chi_{50}$
are 21/5 for the composite {\bf 10}.
This means that the composite {\bf 75}
with the quark-lepton multiplets for the constituents
involves other composites bound
with larger strengths.
Next we examine the composites {\bf 50} and {\bf 50*}.
When the {\bf 5*} and {\bf 10} quark-lepton multiplets
and their conjugates coupled to
fermions with right-handed zero mode
form {\bf 50} or {\bf 50*} bound states,
the composites would include
{\bf 45} or {\bf 45*} bound states simultaneously.
Since the quadratic Casimir operators
have the relation $C_2({\bf 50}) < C_2({\bf 45})$,
the composite {\bf 45} is bound stronger than
the composite {\bf 50}.
This gives rise to SU(2) doublet mass terms.
It is because the representation {\bf 45} for SU(5)
has a color singlet and weak doublet state
for SU(3)$\times$SU(2)
and the composites
{\bf 45} and {\bf 75} can become an SU(5) singlet with
the composite {\bf 5} as seen in Eq.~(\ref{45}).
Therefore
doublet-triplet splitting by
the composites {\bf 50}, {\bf 50*} and {\bf 75}
would not occur minimally.
The model would need to be modified for
application of the idea.
\section{Conclusion \label{concl}}
If extra dimensions are included in an effective
theory,
the strength of force would behave
unlike four dimensions.
We have studied most attractive channel
for an SU(5) grand unification in five dimensions.
We have assumed that components in
each SU(5) multiplet obey
the same boundary condition with respect
to extra dimensions to keep the gauge coupling
unification above the unification scale.
In addition to quarks, leptons and gauge bosons,
$\psi_{5*}$, $\psi_{10}$ and $A_M$,
anomaly-free sets of fields $\chi$ with
zero mode for the right-handed component
lead to various scalar bound states.
The minimal anomaly-free set
to obtain {\bf 5} and {\bf 24} scalar bound states
is ($\psi_{5*}$, $\psi_{10}$, $A_M$)
plus ($\chi_{10*}$, $\chi_{15}$,
$\chi_{40*}$, $\chi_{45}$).
For this field content, the
binding strengths for {\bf 10} and {\bf 10*} have been found to be
larger than that of {\bf 24}.
Then, before SU(5) symmetry is broken to
the standard model gauge group by {\bf 24},
symmetry breaking can occur by {\bf 10} and
{\bf 10*}.
A simple way to overcome this problem
is to adopt ($\chi_5^{I=1,\cdots,9}$, $\chi_{15}$)
instead of the anomaly-free set
($\chi_{10*}$, $\chi_{15}$,
$\chi_{40*}$, $\chi_{45}$).
Here the {\bf 15} $\chi$ field is uniquely chosen
because it is the only representation
to form a {\bf 24} scalar bound state
whose binding strength is large
compared with other composites.
The largest strength for the composite {\bf 24}
corresponds to the breaking of the
SU(5) to the standard model gauge group.
After the SU(5) is broken,
{\bf 5} in $\bar{\psi}_{10}\chi_5$ contribute
for the breaking of
SU(2)$\times$U(1) to U(1).
In addition, we have found that
the binding strengths for ${\bf 24}$ and ${\bf 5}$
are larger than the known results for
the composite Higgs doublet
for the standard model gauge group.
We have shown that the symmetry breaking of
the grand unification group
to color and electromagnetism may occur
by {\bf 24} and {\bf 5}.
It would be more interesting
to identify other fermion sets to assure
the double-triplet splitting.
However, we have found that
the doublet-triplet splitting by
the composites {\bf 50}, {\bf 50*} and {\bf 75}
does not seem to occur minimally.
Finally,
our analysis can be developed in
various directions such as modification of
field contents.
It needs to be examined in more detail
how
masses and charges of
elementary
particles are derived appropriately.
\vspace{4ex}
\subsubsection*{Acknowledgments}
This work is supported by Scientific Grants
from the Ministry of Education
and Science, Grant No.~20244028.
\vspace*{10mm}
|
\section{\bf Introduction}
\label{sect1} Balancedness and low out-of-phase autocorrelation
values are properties which one expects to find in randomly
generated sequences. However, in some engineering applications
(e.g., CDMA communication, radar, cryptography etc.) the
so-called pseudo-random sequences (which are generated
deterministically) are preferable because of the possible
deviations and to avoid data storage requirements. Among these
sequences we distinguish the {\bf m}-sequences, the GMW sequences
and their derivatives (both binary and non-binary) and many other
types of sequences having the properties which mimic those of
truly random sequences (see, e.g., \cite{Golomb} --
\cite{Tilborg}). In this correspondence, the relationship between two of
these properties: the ideal two-level autocorrelation property and
the almost balancedness in $p-$ ary case, is investigated. In the
binary case, it has been proved that the number of ones and zeroes
in a sequence with the two-level autocorrelation differ by 1 (see,
e.g., \cite{Golomb} or \cite{Tilborg}[Chapter $3$]). A similar
situation is observed in the general case ($p$ arbitrary prime)
and it is treated algebraically in this article.
The paper is organized as follows. In next section, we recall some
definitions and preliminaries, and in Section $3$, we present
our results.
\section{\bf Definitions and Preliminaries}
\label{sect2} Let ${\bf a} = \{a_{n}\}, n = 0,1,\ldots,N-1$ be a
sequence of length $N$ with entries from some finite set ${\bf S}$. For
arbitrary $c \in {\bf S}$, define $\mu_{c} = \vert n :\; a_{n} =
c,\; 0 \leq n \leq N-1 \vert$, and call $\mu_{c}$ multiplicity of
$c$ (or frequency of appearance) in ${\bf a}$. Clearly, we have:
\begin{eqnarray}\label{eq.1}
\sum_{c \in {\bf S}}\mu_{c} = N
\end{eqnarray}
Sequences with entries from the finite field ${\bf GF}(p)$
for a prime $p$, are called sometimes $p-$ary sequences.
In this correspondence, we consider mostly pure periodic $p-$ary sequences.
In order to present our result we need to recall the following two
definitions.
\begin{definition}\label{balanced}
Let $p$ be a prime and $t$ be some positive integer. A $p-$ary
sequence ${\bf a}$ of period $N = pt-1$, is called almost balanced
if there exists $c \in {\bf GF}(p)$ with multiplicity $t-1$, while
all other elements of the field are of multiplicity $t$ in one
period of ${\bf a}$. We will also say that $c$ has an exceptional
frequency of appearance in a period of ${\bf a}$.
\end{definition}
Let $\omega$ denotes the $p-$th primitive root of unity $e^{2 \pi i/p}$. Then
${\cal R}_{p} = \{ 1,\omega,\ldots,\omega^{p-1}\}$ is the set of
all $p-$th roots of unity. By identifying ${\bf GF}(p)$ with
$\MZ_{p}$, one can put into correspondence to any $p-$ary sequence ${\bf a} = \{a_{n}\}, n =
0,1,\ldots$ the complex sequence ${\bf s} = \{s(n) = \omega^{a_{n}}\}, n = 0,1,\ldots$
whose entries are from ${\cal R}_{p}$.
Further on, we shall use the notation $\alpha^{*}$ for the
conjugate of a complex number $\alpha$.
\begin{definition}(see, e.g., \cite{NoGo})\label{def.2}
The complex sequence ${\bf s} = \{s(n)\}$ of period $N$ is said
to have the ideal two-level autocorrelation property if its
autocorrelation function $R_{\bf s}(k)$ satisfies the following:
\begin{eqnarray*}
{R_{\bf s}(k)} = {{\hspace*{-0.00001cm} N, \;\;\;\;\; {\rm if}\;\; k \equiv 0\; mod\;N,} \atopwithdelims \{. {\hspace*{-1cm} -1, \;\;\; {\rm otherwise,}}}
\end{eqnarray*}
where $R_{\bf s}(k)$ is defined as follows:
\begin{equation*}
R_{\bf s}(k) = \sum_{n=0}^{N-1}s(n)s^{*}(n+k).
\end{equation*}
\end{definition}
When ${\bf a}$ is a $p-$ary sequence with corresponding complex
sequence possessing the ideal two-level autocorrelation property,
we shall say that ${\bf a}$ also possesses this property.
\section{\bf Results}
\label{sect3}
We start with the following
\begin{lemma}\label{lemma1}
Let ${\bf a} = \{a_{n}\}, n = 0,1,\ldots$ be a $p-$ary sequence
${\bf a}$ of period $N$ having the ideal two-level autocorrelation
property. Then there exists a positive integer $t$ such that $N = pt-1$.
\end{lemma}
\begin{proof}
Let ${\bf s} = \{s(n) = \omega^{a_{n}}\}, n = 0,1,\ldots$ be the
complex sequence corresponding to ${\bf a}$. Define
${\bf u} = \{u(n) = s(n)s^{*}(n+1)\}, n = 0,1,...,N-1$.
Obviously, $u(n) \in {\cal R}_{p}$, and therefore $u(n) =
\omega^{k_{n}}$ for some $0 \leq k_{n} < p$. Then according to the
definition of the ideal two-level autocorrelation property, we
have: $R_{\bf s}(1) = \sum_{n=0}^{N-1}u(n) = -1$, or equivalently
$\sum_{n=0}^{N-1}\omega^{k_{n}} + 1 = 0$. Let $\mu_{k}$ be the
multiplicity of $\omega^{k}$ in ${\bf u}$, for $0 \leq k < p$ .
Hence, from the previous equation, we get:
\begin{eqnarray*}
\sum_{k=p-1}^{1}\mu_{k}\omega^{k} + \mu_{0} + 1 = 0
\end{eqnarray*}
Thus, $\omega$ is a root of the polynomial $f({\bf x}) =
\sum_{k=p-1}^{1}\mu_{k}{\bf x}^{k} + \mu_{0} + 1$. Since the minimal polynomial of
$\omega$ (over the field of rational
numbers $\MQ$) is the $p-$th cyclotomic polynomial $\Phi_{p}({\bf x}) = {\bf x}^{p-1} + {\bf x}^{p-2} +
\ldots +{\bf x} + 1$, applying the well-known fact from Abstract Algebra (see, e.g., \cite{Obr}),
we conclude that $\Phi_{p}({\bf x})$ divides $f({\bf x})$.
Further, since $p - 1 = deg(\Phi_{p}) \geq
deg(f)$ and $f$ has integer coefficients, there exists some
positive integer $t$ for which $\mu_{k} = t$ when $1 \leq k < p$
and $\mu_{0} + 1 = t$. Finally, by equation (\ref{eq.1}), we have $N =
\sum_{k=0}^{p-1}\mu_{k} = t-1 +(p-1)t = pt-1$, and the proof is
completed.
\end{proof}
\begin{remark} The above Lemma shows that a $p-$ary
sequence possessing the ideal two-level autocorrelation property
cannot be strictly balanced.
It is well known that if a binary sequence has
the ideal two-level autocorrelation, it must have a period $N$
with $N \equiv -1\;(mod\;4)$ (see, e.g., \cite{KimSong}). Of
course, this fact is stronger than the claim of Lemma
\ref{lemma1} in case $p=2$.
Also, in case $p > 2$, for all known periodic sequences with the
ideal two-level autocorrelation, the period $N$ is of the form
$p^{m}-1$ for some positive $m$, while in the binary case there
exist examples when it is of different type (see, e.g., \cite{Go}).
But here, we will not discuss the details of this topic.
\end{remark}
\begin{lemma}\label{lemma2}
Let $A_{n}, 0 \leq n < k$ be $k$ positive integers which obey the two equations:
\begin{eqnarray}\label{eq.3}
\sum_{n=0}^{k-1}(A_{n} - A_{n+1})^{2} = 2,\;if \; k \geq 2
\end{eqnarray}
\begin{eqnarray}\label{eq.4}
\sum_{n=0}^{k-1}(A_{n} - A_{n+2})^{2} = 2,\;if \;k \geq 4
\end{eqnarray}
where $n+j,\; j = 1,2$ are taken modulo $k$. Then there exists an
index $m:\;0 \leq m < k$ and a positive integer $A$, such that
$A_{m} = A \pm 1$, while all other $A_{n}$s are equal to $A$.
\end{lemma}
\begin{proof}
Imagine that $A_{0},A_{1}.\ldots,A_{k-1}$ are written alongside a
circle. Since the differences $A_{n}-A_{n+j},\;j=1,2$ are integer
numbers, in each of the given equations there exist exactly two
terms $(A_{n}-A_{n+j})^{2}$ equal to $1$, while all other terms
have to be equal to $0$. Thus equation (\ref{eq.3}) implies that the
$A_{n}$s are divided into two nonempty groups of equal adjacent
numbers and the common values (for each group its own) differ by
$1$ (So, the proof is completed in cases $k=2,3$). Moreover,
taking into account equation (\ref{eq.4}), we conclude that one of the
groups has to contain a single number, since otherwise four terms
$(A_{n}-A_{n+2})^{2}$ will be equal to $1$. This completes the
proof in case $k > 3$.
\end{proof}
Furthermore, we prove the following theorem.
\begin{theorem}\label{th.1}
Let ${\bf s} = \{s(n)\}, n = 0,1,\ldots$ be complex sequence of
period $N$ having the ideal two-level autocorrelation property.
Then it holds $\vert \sum_{n=0}^{N-1} s(n) \vert = 1$.
\end{theorem}
\begin{proof}
By straightforward computations taking into account the definition
of autocorrelation function, we have:
\begin{eqnarray*}
\sum_{k=0}^{N-1}R_{\bf s}(k) = \sum_{k=0}^{N-1}\sum_{n=0}^{N-1}s(n)s^{*}(n+k) =
\sum_{n=0}^{N-1}s(n)(\sum_{k=0}^{N-1}s^{*}(n+k)) =
\end{eqnarray*}
\begin{eqnarray*}
\sum_{n=0}^{N-1}s(n)(\sum_{k=0}^{N-1}s^{*}(k)) =
\sum_{n=0}^{N-1}s(n) \times \sum_{k=0}^{N-1}s^{*}(k) = \sum_{n=0}^{N-1}s(n) \times (\sum_{k=0}^{N-1}s(k))^{*} =
\vert \sum_{n=0}^{N-1} s(n) \vert^{2}.
\end{eqnarray*}
On the other hand, by Definition \ref{def.2} we have that:
$\sum_{k=0}^{N-1}R_{\bf s}(k) = 1$. So, $\vert \sum_{n=0}^{N-1} s(n) \vert^{2} = 1$ which completes the proof.
\end{proof}
From Theorem \ref{th.1}, one can immediately derive the
following.
\begin{corollary}\label{cor.1}
Let ${\bf a}$ be a $p-$ary sequence of period $N$, and $\mu_n$ be
the frequency of appearance of $n,\; 0 \leq n \leq p-1,$ in one
period of ${\bf a}$. If in addition, ${\bf a}$ has the ideal two-level autocorrelation property then it holds:
\begin{eqnarray}\label{eq.2}
\vert \sum_{n=0}^{p-1} \mu_{n}\omega^{n} \vert^{2} = 1.
\end{eqnarray}
\end{corollary}
Now, we will prove the main theorem of this correspondence.
\begin{theorem} \label{th.2}
Let the complex sequence ${\bf s} = \{s(n)\}, n = 0,1,\ldots$
corresponding to a $p-$ary sequence ${\bf a}$ of period $N$ has
the ideal two-level autocorrelation property. Then ${\bf a}$ is
almost balanced.
\end{theorem}
\begin{proof}
Let $\mu_{n}$ be the multiplicity of $n \in \MZ_{p}$ in one period
of ${\bf a}$. We will rewrite the left-hand side of the equation
(\ref{eq.2}) from Corollary \ref{cor.1} as follows:
\begin{eqnarray*}
\vert \sum_{n=0}^{p-1} \mu_{n}\omega^{n} \vert^{2} = (\sum_{n=0}^{p-1} \mu_{n}\omega^{n})(\sum_{n=0}^{p-1} \mu_{n}\omega^{n})^{*} = (\sum_{n=0}^{p-1} \mu_{n}\omega^{n})(\sum_{n=0}^{p-1} \mu_{n}\omega^{p-n}) = \pi(\omega)
\end{eqnarray*}
Further, arranging $\pi(\omega)$ according to the powers of $\omega$, we get:
\begin{eqnarray*}
\pi(\omega) = \sum_{n=0}^{p-1}B_{n}\omega^{p-n},
\end{eqnarray*}
where
\begin{eqnarray}\label{eq.5}
B_{n} = \sum_{k=0}^{p-1}\mu_{k}\mu_{k+n},\; n = 0,1,\ldots,p-1.
\end{eqnarray}
On the other hand, by equation (\ref{eq.2}), we have: $\pi(\omega) = 1,$ or
\begin{eqnarray*}
\sum_{n=1}^{p-1}B_{n}\omega^{p-n}+ \sum_{n=0}^{p-1}\mu_{n}^{2} - 1 = 0.
\end{eqnarray*}
So, $\omega$ is a root of polynomial $f({\bf x}) =
\sum_{n=1}^{p-1}B_{n}{\bf x}^{p-n}+ \sum_{n=0}^{p-1}\mu_{n}^{2} -
1$. Proceeding with polynomials $f$ and $\Phi_{p}$ like in the proof of
Lemma \ref{lemma1}, we conclude that there exists a positive
integer $C$ which satisfies the following equations:
\begin{eqnarray*}
B_{n} = C,\;for\; n = p-1,\ldots,2,1
\end{eqnarray*}
\begin{eqnarray*}
\sum_{n=0}^{p-1}\mu_{n}^{2} = C+1
\end{eqnarray*}
Subtracting from the last equation the two formers and taking into account equation (\ref{eq.5}), we get:
\begin{eqnarray*}
\sum_{n=0}^{p-1}\mu_{n}^{2} - B_{j} = {1 \over 2} \sum_{n=0}^{p-1}(\mu_{n} - \mu_{n+j})^{2} = 1,\;for\;j=1,2,
\end{eqnarray*}
or equivalently
\begin{eqnarray}
\sum_{n=0}^{p-1}(\mu_{n} - \mu_{n+j})^{2} = 2,\;for\;j=1,2.
\end{eqnarray}
Now, by Lemma \ref{lemma2}, it follows the existence of an index
$m:\;0 \leq m \leq p-1$ and a positive integer $\mu$ such that
$\mu_{m} = \mu \pm 1$, while all other $\mu_{n}$s are equal to
$\mu$. Finally, Lemma \ref{lemma1} excludes the possibility
that $\mu_{m} = \mu +1$ when $p > 2$ (Note that if $p=2$ both
possibilities look the same). Therefore, the sequence ${\bf a}$ is
almost balanced.
\end{proof}
To the best of our knowledge, all known constructions of $p-$ary
sequences, $p > 2$, possessing the ideal two-level autocorrelation
property have the peculiarity to provide almost balanced sequences
for which the frequency of appearance of zero is exceptional, i.e.
zero appears once less than the other elements of ${\bf GF}(p)$
(see, e.g., \cite{SchWe} -- \cite{FanDar}). Moreover, in another
two papers \cite{GongSong1} and \cite{GongSong2} (among other things)
it is claimed that "the balance property has been proved in \cite{LudGong}
assuming only the ideal two-level autocorrelation function (whenever $q = p > 2$ is an odd prime)".
And the balance property of a $p-$ary sequence of period $p^{m}-1$ is defined as
that zero appears $p^{m-1}-1$ times while any nonzero element of ${\bf GF}(p)$
appears $p^{m-1}$ times in one period.
But as the next theorem shows that feature is not common
for sequences with the ideal two-level autocorrelation property whatever might be the period.
\begin{theorem}\label{th.3}
Let ${\bf a^{\prime}} = \{a^{\prime}_{n}\}, n = 0,1,\ldots$ be a $p-$ary sequence
of period $N$ having the ideal two-level autocorrelation property
and let the element $c^{\prime} \in {\bf GF}(p)$ be with exceptional frequency of
appearance in one its period.
Define ${\bf a}$ as $\{a^{\prime}_{n} - c^{\prime} + c\}, n = 0,1,\ldots$,
where $c$ is an arbitrary element of ${\bf GF}(p)$.
Then the latter sequence satisfies the ideal two-level autocorrelation property too,
having the element $c$ with exceptional frequency of appearance in one period.
\end{theorem}
\begin{proof}
Since by Theorem \ref{th.2}, ${\bf a}^{\prime}$ is an almost balanced
sequence then ${\bf a}= \{a^{\prime}_{n} - c^{\prime} + c\}, n = 0,1,\ldots$ is an
almost balanced too, but of course, instead of $c^{\prime}$ the frequency
of appearance of $c = c^{\prime} - c^{\prime} + c$ is the
exceptional one. Let ${\bf s}$ and ${\bf s}^{\prime}$ be the
complex sequences corresponding to ${\bf a}$ and ${\bf
a}^{\prime}$, respectively. Denote by $w = \omega^{c - c^{\prime}}$
Then by the definition of autocorrelation function (see,
Definition \ref{def.2}) for any $k$, we have the following:
\begin{equation*}
R_{{\bf s}}(k) = \sum_{n=0}^{N-1}s(n){s}^{*}(n+k) = \sum_{n=0}^{N-1}(s(n)^{\prime}\; w) \times (s^{\prime\;*}(n+k)\; w^{*}) = R_{\bf s^{\prime}}(k),
\end{equation*}
since the product $ w \times w^{*}$ equals to $1$. Thus, the
autocorrelation function of ${\bf s}$ coincides with that one
of ${\bf s}^{\prime}$ and therefore ${\bf a}$ possesses the
ideal two-level autocorrelation property as well.
\end{proof}
In other words, the above theorem states that together with any $p-$ary
sequence satisfying the ideal two-level autocorrelation property there exists
a whole one-parametric family of cardinality $p$ containing sequences of this kind
(and which are not cyclic replicas of the primary sequence).
This fact might be useful to vary, for instance, CDMA communication.
\section*{\bf Acknowledgments}
The author would like to thank Stefan M. Dodunekov, Ivan N. Landjev and Svetla Nikova for helpful discussions and comments which substantially improve the presentation of the results.
|
\section{Introduction}
Establishing and understanding the normal-state
phase diagram of cuprates is of primary importance
in the quest to uncover the mechanism of
high-temperature superconductivity.
The discovery of quantum oscillations in underdoped YBa$_2$Cu$_3$O$_{y}$ (YBCO) revealed
that in the absence of superconductivity, suppressed by application of a large magnetic field,
the ground state in the underdoped region of the phase diagram is a metal whose Fermi surface contains a small closed
pocket.\cite{Doiron-Leyraud2007}
The negative Hall and Seebeck coefficients of YBCO at $T \to 0$ show this pocket to be electron-like.\cite{LeBoeuf2007,Chang2010}
The presence of an electron pocket in the Fermi surface of a hole-doped cuprate is the typical signature
of a Fermi-surface reconstruction caused by the onset of a new periodicity which breaks the
translational symmetry of the crystal lattice.\cite{Taillefer2009,Chakravarty2008}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.375\textwidth]{FIG1FINAL.eps}
\caption{
Phase diagram of YBCO, showing the zero-field superconducting transition temperature $T_c$
(dotted line, Ref.~\onlinecite{Liang2006}, extrapolated as dashed line above $p=0.18$)
and the pseudogap crossover temperature $T^\star$ detected by the
Nernst effect (squares; Ref.~\onlinecite{Daou2010}).
The onset of in-plane anisotropy in the Nernst coefficient below $T^\star$ shows that the pseudogap phase
is a state with broken rotational symmetry (BRS).\cite{Daou2010}
Once superconductivity is suppressed by a magnetic field, the normal state at $T \to 0$ is characterized
by a reconstructed Fermi surface,\cite{Doiron-Leyraud2007,LeBoeuf2007}
evidence of broken translational symmetry (BTS).
The temperature $T_{\rm H}$ below which the Hall coefficient $R_{\rm H}(T)$ starts to deviate downward is the first signature
of Fermi-surface reconstruction upon cooling (open circles; from Ref.~\onlinecite{LeBoeuf2011}).
The white down-pointing arrow locates the present study of Nernst anisotropy on the phase diagram.
The two dashed lines are a guide to the eye.
}
\label{fig:Phasediagram}
\end{center}
\end{figure}
In the doping phase diagram of YBCO (Fig.~1), the electron pocket exists at $T \to 0$ (in the absence of superconductivity)
throughout the range from $p = 0.083$ to at least $p = 0.152$.\cite{LeBoeuf2011}
We infer that translational symmetry is broken at $T = 0$ over at least that range,
by an ordered phase that has yet to be definitively identified.
The fact that the Seebeck coefficient of the cuprate La$_{1.8-x}$Sr$_x$Eu$_{0.2}$CuO$_4$ (Eu-LSCO) at $T \to 0$
is negative over the same range of doping as in YBCO,\cite{Laliberte2011}
and that stripe order -- a unidirectional modulation of spin and/or charge densities\cite{Kivelson2003,Vojta2009} --
prevails in Eu-LSCO over that doping range,\cite{Fink2010}
is compelling evidence that stripe order is responsible for the broken translational symmetry
and that Fermi-surface reconstruction is a generic property of hole-doped cuprates.
\begin{figure}
\includegraphics[width=0.4\textwidth]{FIG2FINAL.eps}
\caption{
Nernst coefficient $\nu$ of YBCO at a hole concentration (doping) $p=0.12$,
plotted as $\nu/T$ versus magnetic field $H$,
for different temperatures, as indicated.
Top: the temperature gradient $\Delta T$ is applied along the $a$ axis of the orthorhombic crystal structure.
Bottom: $\Delta T$ is along the $b$ axis.
Inset:
zoom on the data at high field. The saturation in $\nu$ vs $H$ above $H \simeq 26$~T
indicates that the positive contribution from superconducting fluctuations has become negligible, and
the data above 26 T represent the normal-state properties of YBCO at that doping.
}
\label{fig:nuvsH}
\end{figure}
At temperatures above the superconducting transition temperature $T_c$,
the normal-state phase diagram of cuprates is characterized by the pseudogap phase,
below a crossover temperature $T^\star$.\cite{Norman2005}
The Nernst effect was recently found to be a sensitive probe of the pseudogap phase,\cite{Cyr-Choiniere2009,Matusiak2009,Daou2010}
such that it can be used to detect $T^\star$, as shown in Fig.~1 for YBCO.
Measurements of the Nernst coefficient $\nu(T)$ in detwinned crystals of YBCO for a temperature gradient along the $a$-axis
and $b$-axis directions within the basal plane of the orthorhombic crystal structure revealed a strong in-plane anisotropy,
setting in at $T^\star$.\cite{Daou2010}
This showed that the pseudogap phase breaks the four-fold rotational symmetry of the CuO$_2$ planes,
throughout the doping range investigated, from $p = 0.08$ to $p = 0.18$.\cite{Daou2010}
In this paper, we investigate the impact of Fermi-surface reconstruction on this Nernst anisotropy,
by extending the previous Nernst study to low temperature.
We find that below 80 K the anisotropy falls rapidly, in close parallel with the fall in the Hall coefficient to negative values.
We infer that the Nernst anisotropy disappears because the small closed high-mobility electron pocket which dominates the
transport properties of YBCO at low temperature yields isotropic transport.
\begin{figure}
\includegraphics[width=0.4\textwidth]{FIG3FINAL.eps}
\caption{
Nernst coefficient $\nu$ of YBCO at $p=0.12$,
plotted as $\nu/T$ versus temperature $T$,
for different values of the magnetic field, as indicated.
Top: $\Delta T$ is along the $a$ axis.
Bottom: $\Delta T$ is along the $b$ axis.
The vertical dashed line marks the zero-field superconducting transition,
at $T_c = 66.0$~K.}
\label{fig:nuvsT}
\end{figure}
\section{Methods}
Measurements were performed on high-quality detwinned YBCO crystals
grown in a non-reactive BaZrO$_3$ crucible from
high-purity starting materials.\cite{Liang2000}
The oxygen content was set at $y = 6.67$ and the
dopant oxygen atoms were made to order into an
ortho-VIII superstructure, yielding a superconducting
transition temperature $T_c = 66.0$~K.
The hole concentration (doping) $p = 0.12$ was
determined from a relationship between $T_c$ and
the $c$-axis lattice constant \cite{Liang2006}.
\begin{figure}
\begin{center}
\includegraphics[width=0.4\textwidth]{FIG4FINAL.eps}
\caption{
Nernst coefficient $\nu$ of YBCO plotted as $\nu/T$ vs $T$ for
the two directions of temperature gradient
($\nu_a$ for $\Delta T \parallel a$, blue symbols; $\nu_b$ for $\Delta T \parallel b$; red symbols),
for two values of the magnetic field: $H=14$~T (circles) and 28~T (squares).
Full symbols correspond to normal-state data, in which the superconducting contribution to the
Nernst signal is negligible.
Note that the normal-state data for $\nu_b$ at 14 T and 28 T do not quite coincide because
of a slight field dependence of $\nu_b$, akin to magnetoresistance (see isotherm at 85 K in Fig.~2).
}
\label{fig:nunbvsT}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.45\textwidth]{FIG5FINAL.eps}
\caption{
Temperature dependence of the in-plane anisotropy in the normal-state Nernst coefficient
of YBCO at $p = 0.12$.
Top:
Nernst anisotropy difference $(\nu_a - \nu_b)/T$ vs $T$
(blue circles, left axis; from Ref.~\onlinecite{Daou2010}).
The difference starts to rise below the pseudogap temperature $T^\star$ (vertical dashed line).
Also shown is the Hall coefficient $R_{\rm H}(T)$, measured in a field $H = 10$~T
(continuous red curve, right axis).
Below $\sim 100$~K, $R_{\rm H}(T)$ drops precipitously to reach large negative values
at $T \to 0$.\cite{LeBoeuf2007,LeBoeuf2011}
We can define the onset of this drop as $T_{\rm H}$, the temperature below which
$R_{\rm H}(T)$ acquires downward curvature (see Ref.~\onlinecite{LeBoeuf2011}).
Bottom:
Normalized Nernst anisotropy defined as the dimensionless ratio
$\nu_a - \nu_b)/(\nu_a + \nu_b)$,
plotted vs $T$ below 120~K (full blue symbols; squares, $H = 28$~T, this work;
circles, $H = 14$~T, from Ref.~\onlinecite{Daou2010}).
(Because both $\nu_a(T)$ and $\nu_b(T)$ cross zero around 150~K, the ratio becomes ill-defined above 120~K.
To avoid this, we define the sum and difference relative to their value at $T^\star$ and plot
their ratio above 120~K (open blue circles); see Ref.~\onlinecite{Daou2010}.)
Below $\sim 80$~K, the anisotropy ratio drops rapidly to zero as $T \to 0$ in a way that precisely tracks
the drop in the Hall coefficient, shown here as tan$\theta_{\rm H}(T) \equiv \rho_{xy} / \rho_{xx}$
vs $T$ (continuous red curve, right axis), where the transverse ($\rho_{xy}$)
and longitudinal ($\rho_{xx}$) resistivities are measured in 30~T.
This reveals that the disappearance of the Nernst anisotropy at low temperature is due to the Fermi-surface reconstruction
that leads to the formation of a small closed electron pocket of high mobility (see text).
}
\label{fig:NvsRH}
\end{center}
\end{figure}
The Nernst effect, being the transverse voltage $V$ generated
by a longitudinal temperature difference $\Delta T$ in
a perpendicular applied magnetic field $H$,\cite{Ong2006,Behnia2009}
was measured in Sherbrooke up to 15 T and at the LNCMI in Grenoble up to 28 T.
In both cases, we used a
one-heater two-thermometer setup and the field was
applied along the $c$ axis of the orthorhombic crystal structure.
The Nernst signal was measured with the thermal gradient $\Delta T$ either along
the $a$ axis ($\Delta T_a$) or the $b$ axis ($\Delta T_b$),
and the Nernst coefficient $\nu$ is indexed as follows:
\begin{equation}
\nu_{a}=\frac{\alpha}{H}\frac{V_b}{\Delta T_a } \quad \textrm{and} \quad \nu_{b}=\frac{\alpha}{H}\frac{V_a}{\Delta T_b }
\end{equation}
where $\alpha=\ell/w$ is the ratio of sample length (between thermometer contacts) to
sample width.
\section{Results}
Before we present our results, it is important to emphasize that there are two different contributions to the Nernst effect in a superconductor:
1) a positive contribution from superconductivity (moving vortices and fluctuations of the superconducting order parameter);
2) a contribution from normal-state quasi-particles, which can be either positive or negative.
In YBCO, the two contributions can be readily separated because the quasi-particle contribution is negative, of opposite sign to the signal from superconducting
fluctuations.\cite{Daou2010}
Note also that the quasi-particle contribution to the Nernst coefficient $\nu(H)$ is mostly independent of field,
whereas the superconducting contribution is strongly dependent on field.\cite{Ong2006}
In the electron-doped cuprate Pr$_{2-x}$Ce$_x$CuO$_4$, for example, this difference in the field dependence was used to separate the two contributions,
both positive in this case.\cite{Li2007}
The amplitude of the quasi-particle contribution may be estimated from the following expression:\cite{Behnia2009}
\begin{equation}
|\frac{\nu}{T}| \simeq \frac{\pi^2}{3} \frac{k_{\rm B}}{e} \frac{\mu}{T_{\rm F}}
\end{equation}
where $k_{\rm B}$ is Boltzmann's constant, $e$ is the electron charge,
$\mu$ is the carrier mobility and $T_{\rm F}$ the Fermi temperature.
This relation, applicable in the $T \to 0$ limit,
was found valid within a factor of two for a wide range of metals.\cite{Behnia2009}
Its implication is that the Nernst effect is highly sensitive to Fermi-surface reconstructions
that produce pockets with a small Fermi energy ($\epsilon_{\rm F} \equiv k_{\rm B} T_{\rm F}$)
and a high mobility.
A good example of this is the heavy-fermion metal URu$_2$Si$_2$ where, upon cooling
below $17$~K, $\epsilon_{\rm F}$ drops by a factor of ten simultaneously with
a ten-fold rise in the mobility $\mu$.
As a consequence, $\nu/T$ rises by two orders of magnitude.\cite{Bel2004}
In Figs. 2 and 3, the Nernst coefficient $\nu$ of YBCO at $p = 0.12$
is plotted as $\nu/T$ vs $H$ and vs $T$, respectively,
for both $a$- and $b$-axis directions.
The high-field $b$-axis data are presented here for the first time,
while the low-field data\cite{Daou2010} and the high-field
$a$-axis data\cite{Chang2010} were reported previously.
We start by examining the isotherms (Fig.~2).
At $T < T_c = 66$~K, $\nu(H)$ shows the strong field dependence typical
of a superconductor:
1) at low field, $\nu = 0$ in the vortex solid phase;
2) at intermediate fields, $\nu$ rises to give a strong positive peak;
3) at higher field, the positive signal gradually decreases, until
such fields as $\nu(H)$ becomes flat, where the superconducting contribution
has become negligible.
At the highest field measured in our experiment, 28 T, this saturation has been reached
for all temperatures down to $\sim 10$~K, so that we may regard the state at 28 T (and above)
as the normal state.
At $T = 85$~K, $\nu_a(H)$ is seen to be totally flat and $\nu_b(H)$
increases very slightly (a form of normal-state magneto-resistance).
The positive (field-decreasing) superconducting contribution has become vanishingly small.
This shows that in a clean underdoped cuprate the regime of significant superconducting fluctuations
does not extend in temperature very far beyond $T_c$.
More quantitatively, the superconducting contribution to the Nernst coefficient $\nu/T$ in YBCO drops to
0.1\% of its peak value at $T_c$ by $T \simeq 1.35~T_c$.
In Fig.~3, we see that the normal-state $\nu/T$ at 28 T is independent of
temperature below $\sim 25$~K.
Its large negative value at $T \to 0$ is completely and unambiguously due to quasi-particles.
In Fig.~4, we compare the normal-state $\nu_a$ and $\nu_b$ as a function of temperature.
We see that the large anisotropy characteristic of the pseudogap phase disappears below $\sim 25$~K.
The in-plane anisotropy of the normal-state $\nu(T)$ is plotted in Fig.~5, as the difference
$\nu_a/T - \nu_b/T$ (top panel) and the normalized difference $(\nu_a - \nu_b)/(\nu_a + \nu_b)$ (bottom panel).
The Nernst anisotropy is seen to rise just below the pseudogap temperature $T^\star$, defined as the temperature below which
the $a$-axis resistivity drops below its linear dependence at high temperature.\cite{Daou2010}
Upon cooling, it continues to rise, until it reaches a maximal value of
$(\nu_a - \nu_b)/(\nu_a + \nu_b) \simeq 0.75$
({\it i.e.} $\nu_a/\nu_b \simeq 7$)
at $\sim 80$~K.
Upon further cooling, however, we now find that the anisotropy drops rapidly, with
$(\nu_a - \nu_b)/(\nu_a + \nu_b) \to 0$
({\it i.e.} $\nu_a/\nu_b \to 1$)
as $T \to 0$.
\section{Discussion}
To elucidate the cause of this dramatic drop in the Nernst anisotropy, we turn to the Hall coefficient $R_{\rm H}(T)$.
In Fig.~5, the normal-state Hall angle $\theta_{\rm H}$ is plotted as $\tan \theta_{\rm H}=\rho_{xy}/\rho_{xx}$,
the ratio of Hall to longitudinal resistivities.
Upon cooling, we see that the drop in $\tan \theta_{\rm H}(T)$ to negative values tracks precisely the drop in Nernst anisotropy.
(Note that from the Onsager relation, $\sigma_{xy} = - \sigma_{yx}$, $R_{\rm H}$ is independent of current direction in the basal plane.\cite{Segawa2004})
\subsection{Fermi-surface reconstruction}
\subsubsection{Electron pocket}
Soon after the discovery of quantum oscillations in YBCO,\cite{Doiron-Leyraud2007}
the fact that the oscillations were seen on top of a large background of negative Hall resistance
led to the interpretation that the oscillations come from orbits around an electron-like Fermi pocket.\cite{LeBoeuf2007}
This interpretation was later confirmed by the observation of a large negative Seebeck coefficient at low temperature.\cite{Chang2010}
Clinching evidence came recently from the quantitative agreement between the measured (negative) value of $S/T$ at $T \to 0$, on the one hand,
and the magnitude of $S/T$ expected from the Fermi energy inscribed in the quantum oscillations,
on the other hand, both obtained in YBCO at the same doping, namely $p=0.11$.\cite{Laliberte2011}
Therefore, in the doping interval from $p = 0.083$ to at least $p = 0.152$,\cite{LeBoeuf2011,Laliberte2011}
the Fermi surface of YBCO in its non-superconducting ground state contains a small closed electron pocket.
This pocket dominates the transport properties at low temperature, as discussed in detail in Ref.~\onlinecite{LeBoeuf2011}.
In particular, it produces a large (quasiparticle) Nernst signal.
Applying Eq.~1 to YBCO at $p=0.11$, where quantum oscillations give $T_{\rm F} = 410 \pm 20$~K
and $\mu = 0.02 \pm 0.006$~T$^{-1}$,\cite{Jaudet2008}
yields $|\nu/T| = 13 \pm 3$~nV/K$^2$T,
while the measured value at $p=0.11$ is
$\nu/T = -~13 \pm 3$~nV/K$^2$T,\cite{Laliberte2011}
in perfect agreement.
The somewhat smaller value at $p=0.12$, namely
$\nu/T \simeq -~7$~nV/K$^2$T as $T \to 0$ (Fig.~4),
is probably due to the lower mobility expected of samples in the ortho-VIII phase (with $y=6.67$)
compared to those in the ortho-II phase (with $y = 6.54$), consistent with the much weaker
quantum oscillations in the former.
The fact that $\nu_a \simeq \nu_b$ as $T \to 0$ shows that the electron pocket yields
transport properties that are isotropic in the plane.
This explains two features of the transport in YBCO.
The first is the jump in the in-plane anisotropy of the resistivity as the doping drops below $p=0.08$.\cite{Ando2002,Sun2004}
Indeed, it was recently discovered that the electron pocket disappears suddenly as the doping is reduced below a critical
value $p = 0.08$,\cite{LeBoeuf2011}
in the sense that for $p < 0.08$ both Hall\cite{LeBoeuf2011} and Seebeck\cite{Laliberte2011} coefficients depend weakly on temperature
and remain positive at $T \to 0$.
This change in Fermi-surface topology (or Lifshitz transition) coincides with a ten-fold increase in resistivity at $T \to 0$,\cite{LeBoeuf2011}
showing that the high-conductivity part of the Fermi surface has disappeared.
Once the high-mobility electron pocket is removed, the in-plane anisotropy
ratio $\rho_a/\rho_b$ rises (see Ref.~\onlinecite{LeBoeuf2011}).
The second feature is the rapid loss of in-plane anisotropy in the Nernst coefficient upon cooling.
As shown in Fig.~5, the precipitous drop in the anisotropy ratio $(\nu_a - \nu_b)/(\nu_a + \nu_b)$ below 80~K tracks closely the fall in the
Hall signal (plotted as $\tan \theta_{\rm H}$) towards large negative values.
As the electron pocket becomes increasingly dominant upon cooling, {\it i.e.} as its mobility $\mu \propto \tan \theta_{\rm H}$ increases,
the Nernst signal becomes increasingly isotropic.
\subsubsection{Stripe order}
The natural explanation for the emergence of an electron pocket in a hole-doped cuprate is the onset of a new periodicity that breaks
the translational symmetry of the lattice and thus reduces the Brillouin zone, causing a reconstruction of the large hole Fermi surface
into smaller pieces.\cite{Chakravarty2008}
A recent study that compares YBCO to the hole-doped cuprate Eu-LSCO found that the Seebeck coefficient behaves in essentially
identical fashion in the two materials, as a function of both temperature and doping:\cite{Laliberte2011}
$S/T$ drops to negative values (of very similar magnitude) below the same peak temperature,
the sign-change temperature $T_0^S$ is maximal at $p = 1/8$ in both cases,
the drop in $S/T$ disappears below the same critical doping $p = 0.08$.
So the same Fermi-surface reconstruction must be taking place in Eu-LSCO as in YBCO,
pointing to a generic mechanism of hole-doped cuprates.
Now in Eu-LSCO, charge modulations are observed by x-ray diffraction
at low temperature,\cite{Fink2010}
over the entire doping range where $S/T < 0$.\cite{Laliberte2011}
Spin modulations are most likely also present, as observed in the closely related material
La$_{1.6-x}$Sr$_x$Nd$_{0.4}$CuO$_4$ (Nd-LSCO).\cite{Ichikawa2000}
Called `stripe order', these spin and charge modulations break the translational symmetry of the CuO$_2$ planes,
and so will cause a reconstruction of the Fermi surface.
Calculations for stripe order at $p = 1/8$ show that an electron pocket will generically appear,\cite{Millis2007}
causing the quasiparticle Nernst signal to be strongly enhanced.\cite{Hackl2010}
It is then reasonable to infer that stripe order causes the Fermi-surface reconstruction in these
underdoped cuprates.
Because stripe order involves unidirectional spin and/or charge modulations, it also breaks the four-fold rotational
symmetry of the CuO$_2$ planes.\cite{Kivelson2003,Vojta2009}
So the reconstructed Fermi surface is expected to have strong in-plane anisotropy,
manifest in the calculations by the presence of quasi-1D open sheets.\cite{Millis2007}
However, if the conductivity of the relatively isotropic electron pockets is much higher, at low temperature, than that of these open sheets, the
inherent anisotropy of the latter will only show up in transport when the electron pocket disappears, as it does below $p=0.08$.
\subsection{The pseudogap phase}
We have focused so far on the non-superconducting ground state at $T \to 0$, with its stripe order and
reconstructed Fermi surface.
Let us now ask what happens when the temperature is raised.
The presence of the electron pocket persists at least as long as the Hall coefficient is negative.
In YBCO at $p=0.12$, $R_{\rm H}(T)$ changes from negative to positive at the sign-change
temperature $T_0^H = 70$~K, above $T_c = 66$~K.\cite{LeBoeuf2007,LeBoeuf2011}
Of course, the drop in $R_{\rm H}(T)$ starts at higher temperature, namely at the peak in $R_{\rm H}(T)$ near 90~K (see Fig.~5).
In fact, the onset of the downturn
is really at the temperature $T_{\rm H}$ where the curvature changes from upward at high temperature to downward at low temperature,
{\it i.e.} at the inflexion point.
At $p = 0.12$, $T_{\rm H} \simeq 120$~K (see Fig.~5).
In Fig.~1, this temperature $T_{\rm H}$ is plotted vs $p$ on the phase diagram of YBCO.
We see that it lies inside the pseudogap phase, between the crossover temperature $T^\star$ and the zero-field
superconducting temperature $T_c$.
This means that the electron pocket starts emerging at temperatures well above the onset of superconductivity,
and it does so even in small magnetic fields. In this sense, the onset of Fermi-surface reconstruction is not field-induced;
it is a property of the zero-field pseudogap phase.
If $T_{\rm H}$ marks the onset of Fermi-surface reconstruction in YBCO as detected in the Hall effect,
what corresponding characteristic temperature do we obtain from other transport properties?
From the Seebeck coefficient $S/T$ vs $T$, a similar characteristic temperature is obtained, with
$T_{\rm S} \simeq 100$~K at $p = 0.12$.\cite{Chang2010,Laliberte2011}
However, the Nernst coefficient, plotted as $\nu/T$ vs $T$, starts its drop to large negative values at a temperature $T_{\nu}$
which is significantly higher, namely $T_{\nu} \simeq 225$~K at $p=0.12$.\cite{Daou2010}
($T_{\nu}$ is independent of direction, the same whether it is measured in $\nu_a$ or $\nu_b$.\cite{Daou2010})
Calculations show that the quasiparticle Nernst effect is an extremely sensitive probe of Fermi-surface distortions such as would
arise from broken rotational symmetry.\cite{Hackl2009}
The value of $T_{\nu}$ is plotted as a function of doping in the phase diagram of Fig.~1.
We see that $T_{\nu} \simeq 2~T_{\rm H}$.
Now $T_{\nu}$ coincides with the temperature $T_{\rho}$ below which the in-plane ($a$-axis) resistivity $\rho_a(T)$ of YBCO deviates from
its linear temperature dependence at high temperature.
This $T_{\rho}$ is regarded as the standard definition of the pseudogap crossover temperature $T^\star$.\cite{Ito1993}
The fact that $T_{\nu} = T_{\rho}$ at all dopings shows that the drop in $\nu/T$ to negative values is a property of the pseudogap phase.
Given that the large value of $\nu/T$ at $T \to 0$ is firmly associated with the small high-mobility electron pocket in YBCO,
can $T_{\nu}$ therefore be regarded as the onset of Fermi-surface reconstruction as detected in the Nernst effect?
By the same token, can $T_{\rho}$ be regarded as the onset of incipient Fermi-surface reconstruction detected in the resisitivity?
If so, then the pseudogap phase would be the high-temperature precursor of the stripe-ordered phase present at low temperature.
One evidence in support of a stripe-precursor scenario is the fact that the enhancement of the Nernst coefficient $\nu/T$ below $T^\star$ is anisotropic,
that it breaks the rotational symmetry of the CuO$_2$ planes. This, of course, is a characteristic signature of stripe ordering.
Indeed, the sequence of broken symmetries, first rotational then translational, is expected in the gradual process of stripe ordering.\cite{Kivelson1998}
The sequence is called `nematic to smectic' ordering.
The fact that $T_{\nu}$ and $T_{\rho}$ are higher than $T_{\rm S}$ and $T_{\rm H}$, by roughly a factor 2, may come from the role of scattering in the various transport coefficients. Indeed, while $\nu$ and $\rho$ both depend directly on the scattering rate (or mobility $\mu$),
with $\nu \propto \mu$ and $\rho \propto 1/\mu$, $S$ and $R_{\rm H}$ do not (at least in a single band model).
In other words, if stripe fluctuations affect the transport primarily through the scattering rate, then we would expect $\rho$ and $\nu$ to be sensitive
to the onset of stripe fluctuations, but not $S$ and $R_{\rm H}$.
As we shall now see, similar precursor effects are observed in the iron-pnictide superconductors.
\subsection{Comparison with pnictide superconductors}
It is instructive to compare the cuprate superconductor YBCO with
the iron-pnictide superconductor Ba(Fe$_{1-x}$Co$_x$)$_2$As$_2$ (Co-Ba122).
In the parent compound BaFe$_2$As$_2$, a well-defined antiferromagnetic order sets in
below a critical temperature $T_{\rm N} = 140$~K.\cite{Canfield2010}
This order is unidirectional, with chains of ferromagnetically aligned spins along the $b$-axis of the orthorhombic
crystal structure alternating antiferromagnetically in the perpendicular direction ($a$-axis).
In other words, this is a form of `spin-stripe' order, which breaks both the translational and rotational symmetry of the original
tetragonal lattice (present well above $T_{\rm N}$).
As Co is introduced, $T_{\rm N}$ falls, and superconductivity appears, with $T_c$ peaking at the point where it crosses $T_{\rm N}$.
In other words, in the underdoped region the normal state is characterized by spin-stripe order for some range of temperature above $T_c$.
This order causes a reconstruction of the Fermi surface, which leads to a change in the in-plane resistivity:
a drop below $T_{\rm N}$ at low Co concentration $x$, an upturn at intermediate $x$.
However, the upturn starts well above $T_{\rm N}$.
One can define a temperature $T_\rho$ below which the roughly linear $T$ dependence at high temperature turns upwards.
For $3 \% < x < 5 \%$, $T_\rho \simeq 2$~$T_{\rm N}$.\cite{FisherSCIENCE2010}
Moreover, the rise in $\rho$ is anisotropic: a strong in-plane anisotropy appears at $T_\rho$.\cite{FisherSCIENCE2010}
At higher doping, above the quantum critical point where antiferromagnetic order disappears (in the absence of superconductivity),
both the upturn and the anisotropy in $\rho$ vanish.\cite{FisherSCIENCE2010}
The overall phenomenology is seen to be remarkably similar to that of YBCO.
The unidirectional order that breaks translational symmetry and reconstructs the Fermi surface at low temperature is preceded at high temperature
by a regime with strong in-plane transport anisotropy, evidence of broken rotational symmetry.
In Co-Ba122, it is very natural to view this regime as the nematic precursor to the smectic phase
at low temperature.
The analogy supports the view that the pseudogap phase in YBCO is just such a precursor to stripe order.
It is interesting to note that in Co-Ba122 the in-plane anisotropy in $\rho$ is not largest at $x=0$, where the order is strongest ($T_{\rm N}$ is highest).
Indeed, the ratio $\rho_b/\rho_a$ at $T \to 0$ is larger at $x > 2~\%$ than at $x=0$.\cite{FisherSCIENCE2010}
This could well be due to a short-circuiting effect similar to that observed in YBCO, whereby a small and isotropic Fermi pocket of high mobility dominates the
conductivity.
Indeed, the concentration $x \simeq 0.02$ in Co-Ba122 appears to be a Lifshitz critical point, as suggested
by ARPES experiments that
reveal the existence of a small hole pocket for $x<0.02$, and not above.\cite{KaminskiNATPHYS2010}
Such a pocket could short-circuit the in-plane anisotropy coming from other parts of the Fermi surface.
\section{Conclusion}
In summary, the Nernst effect is a highly sensitive probe of electronic transformations in
metals. In the cuprate superconductor YBCO, the onset of the pseudogap phase
at $T^\star$ causes a 100-fold enhancement of the quasiparticle Nernst coefficient,
which goes smoothly from
$\nu_b/T = +~0.07$~nV/K$^2$T at $T^\star$
to
$\nu_b/T = -~7$~nV/K$^2$T at $T \to 0$, in the normal (non-superconducting) state.
This enhancement is strongly anisotropic in the plane, showing that the pseudogap phase
breaks the rotational symmetry of the CuO$_2$ planes.
When the Fermi-surface reconstructs at low temperature, the formation of a high-mobility
electron pocket short-circuits this in-plane anisotropy.
The magnitude of $\nu/T$ at $T \to 0$ is in perfect agreement with the value
expected from the small closed Fermi pocket detected in quantum oscillations.
The negative sign proves that the signal comes from quasiparticles and
not superconducting fluctuations or vortices.
There is compelling evidence that the Fermi-surface reconstruction is caused by
a stripe order which breaks the translational symmetry of the CuO$_2$ planes at low temperature,\cite{Laliberte2011}
but also its rotational symmetry.
The sequence of broken symmetries upon cooling, first rotational then translational,
suggests a process of nematic-to-smectic stripe ordering, similar to that observed in
the iron-pnictide superconductor Co-Ba122, where a phase of spin-stripe antiferromagnetic order prevails
in the underdoped region of the temperature-concentration phase diagram.
The analogy suggests that the enigmatic pseudogap phase of hole-doped cuprates is also a high-temperature precursor
of a stripe-ordered phase, with unidirectional charge and/or spin modulations.
This nematic interpretation is consistent with the in-plane anisotropy of the spin fluctuation spectrum
detected by neutron scattering in underdoped YBCO.\cite{Hinkov2008}
\section{Acknowledgements}
We thank J. Corbin and J. Flouquet for assistance with measurements in Sherbrooke and at the LNCMI in Grenoble,
respectively.
J.C. was supported by Fellowships from the Swiss SNF and
FQRNT.
Part of this work was supported by Euromagnet under the EU contract RII3-CT-2004-506239.
C.P. and K.B. acknowledge support from the ANR project DELICE.
L.T. acknowledges support from the Canadian Institute for
Advanced Research, a Canada Research Chair, NSERC, CFI and
FQRNT.
|
\section{Introduction}
\label{Sec1}
The fluctuating electrodynamic near field close to the surface of dielectric bodies due to
thermal and quantum fluctuations inside that bodies has come to the fore
in the last decade. The growing interest of researchers
in the investigation of fluctuating near fields is accompanied by manifold
new possibilities to measure the interesting properties of such thermal near
fields~\cite{ShchegrovEtAl00,CarminatiGreffet99,C.HenkelEtAl2000,J.J.GreffetEtAl2002,J.J.GreffetEtAl2003,F.MarquierEtAl2004,K.JoulainEtAl2005},
which have been developed
in the last decade. For example, two of these techniques are the thermal radiation scanning tunneling
microscopy~\cite{Y.DeWileEtAl2006} and the usage of Bose-Einstein condensates~\cite{J.M.ObrechtEtAl2007}.
Moreover, near field scanning thermal microscopy~\cite{KittelEtAl05} (NSThM) is a new possibility to measure the
radiative heat transfer, which is itself related to the
properties of the fluctuating near field~\cite{Dorofeyev98,MuletEtAl01,KittelEtAl05} between dielectric
bodies~\cite{PolderVanHove71}. From the experimental point of view it should also
be helpful to study the near field of coated materials.
Not only the significance of the fluctuating near field in scanning probe techniques or
nanotechnological applications makes a theoretical investigation necessary and useful.
The electrodynamic near field is also of great theoretical interest, because it shows new
and unexpected physical properties. For example, it has been shown in
recent publications~\cite{CarminatiGreffet99} that coherent quasi-monochromatic evanescent waves can exist in
the thermal near field, although the latter is generated by fluctuating thermal sources.
In order to study near-field effects one may calculate different physical quantities such
as the cross-correlation tensor~\cite{CarminatiGreffet99}, the local density of
states (LDOS)~\cite{JoulainEtAl03} or the spectral energy density~\cite{ShchegrovEtAl00}
in the vicinity of the dielectric body, where this body is usually assumed to be a semi-infinite medium.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\epsfig{file=F1_coated.eps, width = 0.6\textwidth}
\caption{Sketch of the configuration used here: The bulk extends in the regime $z \leq -d$, the coating ranges from
$z = -d$ to $z=0$, and the half-space $z > 0$ is assumed to be vacuum.}
\label{Fig:Configuration}
\end{minipage}
\end{figure}
In this paper we will study how a coating influences the thermal electrodynamical
near field of a semi-infinite substrate (see fig.\ \ref{Fig:Configuration}). For that reason we calculate the energy density
above the coated body. We will show that an effect predicted for a free standing metallic film~\cite{Biehs2007}
can be retrieved by using a polar material as substrate, whereas for a metal substrate the thermal
near-field energy density changes dramatically. Both cases are discussed and understood with the help of surface
plasmon coupling within the coating. Furthermore we investigate how the coating on different substrates
influences the near-field radiative heat transfer.
The near-field radiative heat transfer was already disussed in such a slab configuration for
polar materials~\cite{A.NarayanaswamyC.Chen2003} for an application in thermophotovoltaics and briefly for
metal substrates coated with different metals~\cite{A.I.VolokitinB.N.J.Persson2001}. Here we explicitly
calculate the near-field radiative heat transfer between a semi-infinite and a coated semi-infinite body, showing
in detail that the physical mechanisms leading to different energy densities will leave their imprints in the near-field
radiative heat transfer. With this information at hand it could for example be possible to give a better understanding
of the signal measured with a NSThM, and to clarify the question whether that signal can be interpreted within a
dipole-model~\cite{Dorofeyev98,MuletEtAl01}
or whether it can be modelled as the heat transfer between a semi-infinite and a coated semi-infinite body. Furthermore, the results derived in
this and the preceding paper~\cite{Biehs2007} can also serve as a basis for the investigation
of near-field effects for coated materials, which are often used in experimental setups.
This paper is a direct follow-up of reference~\cite{Biehs2007}, so we refer the reader to that paper for a brief discussion
of Rytov's fluctuational electrodynamics~\cite{RytovEtAl89}. When considering the geometry given in
fig.\ \ref{Fig:Configuration}, it is in principle necessary to construct the dyadic Green's
function with observation point located in the regime $z > 0$, and with sources within the coating or the substrate, respectively.
Since we already determined the dielectric Green's
function for a dielectric film, as corresponding to the coating, in all details~\cite{Biehs2007}, the dyadic Green's function with
source currents within the coating can directly be taken from that reference.
The dyadic Green's function with
sources within the substrate can be constructed
in a straightforward way, so we present here only the results and
refer the interested reader to~\cite{Biehs2007} and~\cite{ChenToTai71}, respectively. For convenience we use
the same notation as in our preceding paper~\cite{Biehs2007}, and for comparability we use again for numerical
computations the Drude model for metals and the Reststrahlen formula for polar materials with material
parameters taken from~\cite{AshcroftMermin76,S.Adachi2004}.
This paper is organized in the following way: In section \ref{Sec2} we briefly discuss the thermal radiation of a coated material.
In section \ref{Sec3} we study the thermal near field of the coated material for different coatings and substrates and show in section
\ref{Sec4} how the observed effects can be interpreted with the surface plasmon polariton coupling inside the coating. Finally,
in the last section we calculate the near-field radiative heat transfer and discuss the influence of a metal coating.
\section{Thermal radiation}
\label{Sec2}
In this paper we are mainly interested in the evanescent near field
of the coated semi-infinite body, but for the sake of completeness we
also report the results for the radiative part. In order to derive
the thermal radiation of the coated semi-infinite body we calculate
the averaged $z$-component $\langle S_z \rangle$ of the Poynting vector outside the
layered system in fig.\ \ref{Fig:Configuration}, which is assumed to be
in local thermal equilibrium at temperature $T$, setting $\epsilon_3 = \epsilon_0$. Taking fluctuating
source currents inside the bulk medium (the substrate) with permittivity $\epsilon_1$
and inside the coating with permittivity $\epsilon_2$, which
contribute additively to the Poynting vector outside the layered system,
we get after a lengthy but straightforward calculation
\begin{equation}
\langle S_z \rangle = \int\!\!{\rm d} \omega\, \frac{E(\omega, \beta)}{(2 \pi)^2} \int\!\!\!{\rm d} \lambda \,
\lambda {\rm e}^{- 2 h_0'' z} \bigl( T_\perp^{{\rm total}} + T_\parallel^{{\rm total}}\bigr).
\label{Eq:Poynting_vector}
\end{equation}
The transmission coefficents $T^{{\rm total}}$ are given as the sum of the bulk and coating
transmission coefficients, $T^{{\rm b}} + T^{{\rm c}}$, for TM- and TE-polarization ($\parallel$ and $\perp$), respectively.
The transmission coefficients for the bulk contribution are given by
\begin{align}
T^{{\rm b}}_\perp &= 16 |h_2|^2 \frac{\Re(h_0) \Re(h_1)}{|D_\perp|^2}, \nonumber \\
T^{{\rm b}}_\parallel &= 16 |h_2|^2 \frac{|k_2|^4}{|k_1|^4} \frac{\Re(h_0) \Re(h_1 \overline{\epsilon}_{1})}{|D_\parallel|^2}
\label{Eq:Transmission_Bulk}
\end{align}
with $h_i = \sqrt{k_0^2 \epsilon_i - \lambda^2}$ for $i = 0, 1, 2$ and
\begin{equation}
D = a^{12} a^{02} {\rm e}^{-{\rm i} h_2 d} - b^{12} b^{02} {\rm e}^{{\rm i} h_2 d}.
\end{equation}
The coefficients $a$ and $b$ are defined as
\begin{align}
a^{ij}_\perp &:= h_i + h_j \label{a_te},\\
a^{ij}_\parallel &:= h_i \frac{\epsilon_j}{\epsilon_i} + h_j \label{a_tm},\\
b^{ij}_\perp &:= h_i - h_j \label{b_te},\\
b^{ij}_\parallel &:= h_i \frac{\epsilon_j}{\epsilon_i} - h_j.
\label{b_tm}
\end{align}
The transmission coefficients for the coating have already been calculated in~\cite{Biehs2007}
and can be stated as
\begin{align}
T_\perp^{{\rm c}} &= \frac{4 \Re(h_0)}{|D_\perp|^2} \biggl[ \Re(h_2) A_\perp
+ 2 \Im(h_2) B_\perp\biggr] \nonumber \\
T_\parallel^{{\rm c}} &= \frac{4 \Re(h_0)}{|D_\parallel|^2} \biggl[ \Re(h_2 \overline{\epsilon}_{r2}) A_\parallel
+ 2 \Im(h_2 \overline{\epsilon}_{r2}) B_\parallel\biggr]
\label{Eq:Transmission_Coating}
\end{align}
with
\begin{align}
A &= |a^{12}|^2 \bigl({\rm e}^{2 h_2'' d} - 1 \bigr)
\quad- |b^{12}|^2 \bigl({\rm e}^{- 2 h_2'' d} - 1\bigr), \\
B &= \Im\biggl(a^{12} \overline{b^{12}} \bigl({\rm e}^{- 2 {\rm i} h_2' d} - 1\bigr)\biggr),
\label{Eq:Definition_of_Coefficients}
\end{align}
where we have used the notation $h_i = h_i' + {\rm i} h_i''$.
Even though the transmission coefficients, which are rather complicated,
could be reformulated in term of Fresnel reflection coefficients~\cite{J.D.Jackson1999},
we will not perfom this procedure here, because in that case we
get different forms of transmission coefficients for the propagating and evanescent modes (cf.\ \cite{Biehs2007}),
i.e., we get four equations instead of the two given in
(\ref{Eq:Transmission_Bulk}) and (\ref{Eq:Transmission_Coating}), thus unnecessarily
inflating the formalism. But it should be kept in mind that the transmission coefficients,
which can be stated with one equation for the propagating part with $\lambda < k_0$
and the evanescent part with $\lambda > k_0$, behave in a quite different manner
for propagating and evanescent modes, respectively. This is a consequence of
the fact that $h_0$ is purely real for propagating modes or purely imaginary for evanescent modes,
$h_0={\rm i}\sqrt{\lambda^2 - k_0^2}\equiv{\rm i}\gamma$. Therefore, the evanescent component $T^{{\rm total}}_{\rm ev}$
does not contribute to the expression for the Poynting vector (\ref{Eq:Poynting_vector}), i.e.,
the Poynting vector covers information on the propagating modes only.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\epsfig{file=F2_Sges_platte_Au_auf_Pt_helvetica30.eps, width = 0.45 \textwidth}
\epsfig{file=F3_Sges_platte_Pt_auf_Au_helvetica30.eps, width = 0.45 \textwidth}
\caption{Left: Numerical result for the thermal radiation of a Pt-coated Au-substrate at temperature $T = 300 {\rm K}$
for different thickness $d$ of the coating, normalized to the black body value $S_{\rm BB}$ given
by the Stefan-Boltzmann law. The solid line is the contribution of the coating and
the dashed line that of the substrate, whith the sum of both being given by the dotted line.
Right: Here the role of the substrate and coating
are interchanged, so that this panel shows the thermal radiation of a Au-coated Pt-substrate at temperature $T = 300 {\rm K}$.}
\label{Fig:Poynting_Pt_Au}
\end{minipage}
\end{figure}
Before we present numerical results for the Poynting vector,
we specify the limiting values of the transmission coefficients
for the propagating modes for different layer thickness $d$, considering
the two cases $d \gg 1/h_2''$ and $d \ll 1/h_2''$, i.e., layers much thicker
or thinner than the skin depth of the coating material, given by
\begin{equation}
d_{\rm s} = \frac{1}{k_0 \Im(\sqrt{\epsilon_r2})} \approx \frac{1}{h_2''}.
\label{Eq:Skin_depth}
\end{equation}
For thin coatings with $d \ll d_s$ the
transmission coefficients $T^{\rm c}$ go linearly with thickness $d$
to zero (cf.\ \cite{Biehs2007}), whereas the transmission coefficents of the bulk
$T^{\rm b}$ converge to the transmission coeffient of a semi-infinite
body~\cite{PolderVanHove71} with permittivity $\epsilon_1$, i.e.,
\begin{align}
T^{{\rm total}}_\perp &\rightarrow T^{{\rm b}}_\perp \approx 4 \frac{\Re(h_1) \Re(h_0)}{|a_\perp^{01}|^2}, \nonumber \\
T^{{\rm total}}_\parallel &\rightarrow T^{{\rm b}}_\parallel \approx 4 \frac{\Re(h_1 \overline{\epsilon}_{r1}) \Re(h_0)}{|a_\parallel^{01}|^2}.
\label{Eq:Semi-infinite}
\end{align}
In contrast, for thick coatings with $d \gg d_s$ the transmission coefficents
of the bulk contributions go to zero and the transmission coefficients of the
coating converge to the transmission coeffient of a semi-infinite
body~\cite{PolderVanHove71} with permittivity $\epsilon_2$, which can
be derived from eq.\ (\ref{Eq:Semi-infinite}) by exchanging the index 1 with 2.
Therefore, the thermal radiation of a coated body given by propagating
modes only and being independent of $z$ (because $h_2'' = 0$ for propagating modes),
has different values for different thicknesses $d$ of the coating. In the limit that the coating
is very thick, i.e., $d \gg d_{\rm s}$, the radiation is that of a half-space filled with the coating
material only. In the other limit of very thin coating, i.e.\ $d \ll d_{\rm s}$, the radiation is that of
a half-space filled entirely with the bulk material. In general, the value of the Poynting vector always falls
between these two extremes. Thus it seems that the thermal radiation maximum found for free standing metallic films
of a certain thickness~\cite{Biehs2007} cannot be observed for coated materials.
This is illustrated in the left panel of fig.\ (\ref{Fig:Poynting_Pt_Au}), where there is a maximum in the contribution
of the coating, but this is overlayed by the bulk contribution.
\section{Thermal near field}
\label{Sec3}
Next, we discuss the non-radiative part of the fluctuating near field
in the vicinity of the coated substrate. To this end, we investigate
the energy density in the distance $z$ from the coated body, which can be written as
\begin{equation}
\langle u(z) \rangle = \int\!\!{\rm d}\omega\,\frac{E(\omega,\beta)}{(2 \pi)^2} \int\!\!{\rm d}\lambda\, \lambda \frac{\lambda_s^2}{2 \omega} {\rm e}^{- 2 h_0'' z} \frac{\bigl(T_\perp^{{\rm total}} + T_\parallel^{{\rm total}} \bigr)}{\Re(h_0)},
\label{Eq:Energy_density}
\end{equation}
with $\lambda_s^2 = 2 k_0^2$ for propagating modes with $\lambda < k_0$ and
$\lambda_s^2 = 2 \lambda^2$ for evanescent modes with $\lambda > k_0$.
Here the factor $\Re(h_0)$ appearing in the transmission coefficients (\ref{Eq:Transmission_Bulk})
and (\ref{Eq:Transmission_Coating}) is canceled out by the
denominator in eq.\ (\ref{Eq:Energy_density}), so that the energy
density contains information about the evanescent thermal near field. Due to these
evanescent modes the expression for the energy density becomes dependent on the distance $z$ from the layered system,
although the contribution of the propagating modes is again independent of the observation distance $z$.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\epsfig{file=F4_U_platte_Bi_auf_GaN_und_Al_tm_helvetica30.eps, width = 0.9 \textwidth}
\caption{Numerical results for the thermal near-field energy density $\langle u^{\rm c}_\parallel \rangle$ of a Bi-coating
on a GaN- and an Al-substrate with temperature $T = 300 {\rm K}$, as functions of
the observation distance $z$ from
the layered system, normalized to the corresponding black body value.
We plot here the results for different thicknessess $d$ of the coating material, with
the solid line giving the thermal energy density above a semi-infinite Bi medium. The dashed lines
give the TM-mode part of thermal energy density for $d = 5\cdot10^{-9} {\rm m}$,
and the dotted lines for $d = 1\cdot10^{-9} {\rm m}$. As indicated by the arrows,
for the case of the polar substrate GaN the energy density raises over that of the
semi-infinite Bi medium. On the other hand, the energy density for the metal substrate Al is diminished in
comparison to that of the semi-infinite Bi medium at distances $z \gg d$. }
\label{Fig:Dens_Bi_auf_GaN_bzw_Al_tm}
\end{minipage}
\end{figure}
Taking the limits for thin and thick coatings is in this case not easy, because for
the evanescent modes the transmission coefficients
of the coating $T^{\rm c}$ contain expressions depending
on $h_2'' d$ in the nominator and denominator which compete with each other, in
analogy to the behaviour discussed in ref.~\cite{Biehs2007} for a single thin dielectric layer.
In contrast, the limit of the transmission coefficients $T^{\rm b}$ for a thin coating with $h_2'' d \ll 1$
reduces for both propagating and evanescent modes to the expression (\ref{Eq:Semi-infinite}) and vanishes
for thick coatings, $h_2'' d \gg 1$.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\epsfig{file=F5_U_platte_Bi_auf_GaN_und_Al_te_helvetica30.eps, width = 0.9 \textwidth}
\caption{As fig.\ \ref{Fig:Dens_Bi_auf_GaN_bzw_Al_tm} but for $\langle u^{\rm c}_\perp \rangle$, i.e., for
the TE-mode contribution. Symbols are as in fig.\ \ref{Fig:Dens_Bi_auf_GaN_bzw_Al_tm}.
In this case the thermal energy density of the coatings falls below that of a semi-infinite
Bi medium for $d \ll d_{\rm s}$, such that the thermal energy density $\langle u^{\rm c}_\perp \rangle$
obtained with a metal substrate is smaller than that for a polar substrate at distances $z \gg d$.}
\label{Fig:Dens_Bi_auf_GaN_bzw_Al_te}
\end{minipage}
\end{figure}
In the evanescent-mode regime $\lambda > k_0$ the
energy density depends on $z$ or, more precisely, on $\exp(- 2 h_2'' z)$.
From this fact and the form of the transmission coefficients given in eq.\ (\ref{Eq:Transmission_Coating})
it appears reasonable to discuss the cases of thin and thick coatings, i.e., $h_2'' d \gg 1$
and $h_2'' d \ll 1$, in the regions $z \ll d$ and $z \gg d$ separately.
As follows from ref.~\cite{Biehs2007}, in the region $z \ll d$ the transmission
coefficients $T^{\rm c}$ take the same form as those for a half-space filled entirely with the coating material.
It can be shown that for $z \ll d$ the bulk contribution $T^{\rm b}$ becomes negligible.
This is a reasonable result, because the evanescent waves with the lateral wave vector $\lambda$
are damped at a length scale $\lambda z \approx 1$ above the layered system. Therefore for $z \ll d$
the near field is dominated by evanescent waves with $\lambda^{-1} \approx z \ll d$,
which do not carry information about the restriction due to the
finite layer thickness $d$. From that it seems to be clear that for $z \ll d$
one receives a result which coincides with that for a bulk made up of the
coating material, i.e., with the permittivity $\epsilon_2$.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.4\textwidth}
\epsfig{file=F6_Uges_platte_5e-9_Bi_auf_Al_bulk_coating.eps, width = 0.9 \textwidth}
\end{minipage}
\begin{minipage}[t]{0.4\textwidth}
\epsfig{file=F7_Uges_platte_5e-9_Bi_auf_GaN_bulk_coating.eps, width = 0.9 \textwidth}
\end{minipage}
\caption{Numerical results for the thermal near-field energy density $\langle u^{\rm c} \rangle$ and $\langle u^{\rm b} \rangle$
for a $5 {\rm nm}$ Bi-coating on a Al- (left) and GaN- (right) substrate, assuming $T = 300 {\rm K}$.
It is seen that for a thin coating with $d \ll d_{\rm s}$ on a metal substrate the energy density
above the layered structure
can be dominated by the contribution of the substrate for distances $z \gg d$, whereas this conclusion cannot be drawn
for a polar substrate.}
\label{Fig:Dens_Bi_auf_GaN_bulk_coating}
\end{figure}
For $d \ll d_{\rm s}$ and $z \gg d$ the situation is more complex, as far as TM modes are concerned.
The TM-mode contribution to the thermal energy density given by the coating is in that region
given by~\cite{Biehs2007}
\begin{equation}
\langle u^{{\rm c,ev}}_\parallel \rangle \approx \int\!\!{\rm d}\omega\, \frac{E(\omega,\beta)}{(2 \pi)^2} \frac{2}{z^3 \omega}
\int\!\!{\rm d} \eta\, \eta^2 \frac{\Im(r^{02}_\parallel) {\rm e}^{-2 \eta}}{|1 - r^{12}_\parallel r^{02}_\parallel (1 - 2 \eta \frac{d}{z})|^2}
\bigl[ 2 \eta \frac{d}{z} (1 + |r^{12}_\parallel|^2) \bigr],
\label{Eq:Energy_density_coating}
\end{equation}
where $r^{12}$ and $r^{02}$ are the usual Fresnel reflection coefficients~\cite{J.D.Jackson1999} for the interfaces
at $z = - d$ and $z = 0$, respectively, and $\eta \equiv \lambda z$.
Through these reflection coefficients, the energy density depends on the properties of bulk and coating material.
Let us restrict the following discussion to metal coatings, so that $|r^{02}| \approx 1$.
Now the energy density contribution of the
coating solely depends on the choice of bulk material. If we choose
as bulk material the vacuum or a polar material, i.e., $r^{12} = r^{02}$ or $r^{12} \approx r^{02}$,
then the expression for the energy density reduces to
\begin{equation}
\langle u^{{\rm c,ev}}_\parallel \rangle \approx \int\!\!{\rm d}\omega\, \frac{E(\omega,\beta)}{(2 \pi)^2}
\frac{4}{z^2 d \omega} \int\!\!{\rm d} \eta\, \eta \, \Im(r^{02}_\parallel) {\rm e}^{-2 \eta}.
\end{equation}
For a metal film or a coated polar material, respectively, we get a
$1/z^2$-dependence of the energy density, as discussed in~\cite{Biehs2007}.
(For a metal coating obeying the Hagens-Rubens approximation the power laws
derived in~\cite{Biehs2007} also give reasonable approximations for a polar substrate.)
In contrast, if we take a second metal as bulk material, then the
reflection coefficients $r^{12}$ between these two metals should be small, so
we can approximate the denominator in eq.\ (\ref{Eq:Energy_density_coating})
by 1. As a consequence we find a $1/z^4$-dependence of the energy
density for the TM-modes for coated metals. Therefore, the $1/z^3$-dependence
of $\langle u_\parallel \rangle$ provided by a half-space consisting solely of
the coating material for $z \ll d$ changes to a $1/z^2$- or $1/z^4$-dependence
for $z \gg d$ when considering a polar or metal bulk with a metal coating. In fig.\
\ref{Fig:Dens_Bi_auf_GaN_bzw_Al_tm} this splitting is shown for a Bi-coating
of different thicknesses $d$ on a GaN bulk and an Al bulk, respectively.
It is interesting to see that the contribution of the coating material
to the energy density $\langle u_\parallel^{{\rm c}} \rangle$ for polar bulk
materials becomes greater than its bulk value for distances $z \gg d$,
similar to what has been discussed for thin metall films in ref.~\cite{Biehs2007}.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\epsfig{file=F8_Uges_platte_5e-9_Bi_te.eps, width = 0.9 \textwidth}
\caption{Numerical results for the total thermal energy density $\langle u^{\rm total}_\perp \rangle$
above the layered structure for a substrate consisting of GaN, Pt or Al coated with a $5 {\rm nm}$ layer
of Bi at a temperature $T = 300 \,{\rm K}$.}
\label{Fig:Dens_Bi_te}
\end{minipage}
\end{figure}
Such a splitting can also be observed for the TE-mode part of the energy density
contribution of the coating material $\langle u_\perp^{{\rm c}} \rangle$ for
distances $z \gg d$, as shown in fig.\ \ref{Fig:Dens_Bi_auf_GaN_bzw_Al_te}.
But in contrast to $\langle u_\parallel^{{\rm c}} \rangle$ the energy density
of the coating material does never rise over its bulk value. From the numerical
result displayed in fig.\ \ref{Fig:Dens_Bi_auf_GaN_bzw_Al_te} one can infer that for a
coated polar bulk material $\langle u_\perp^{{\rm c}} \rangle$ again has
a $1/z^2$-dependence for $z \gg d$, whereas for coated metals there seems to be no
well-developed power law.
Now let us study the interplay of the contributions of the bulk or substrate and
that of the coating to the thermal energy density for $d \ll d_{\rm s}$.
From the discussion above it follows
that for $z \ll d$ it is always $\langle u^{\rm c} \rangle$ which dominates the total
energy density, with the value of $\langle u^{\rm c} \rangle$ coinciding
with its half-space value, i.e., being independent of the coating thickness $d$.
For distances $z \gg d$ it is {\itshape a priori} not clear whether the bulk or the coating
contribution dominates the energy density. However, one may expect for a
polar bulk material and a metal coating that the bulk contribution does not
play an important role because $|r^{12}| \approx 1$, whereas for a metal
bulk $|r^{12}|$ is small, so that waves generated by fluctuating source currents
in the bulk medium can propagate into and through the coating and therefore
contribute to the energy density in a much more significant way at distances
$z \gg d$. In fig.\ \ref{Fig:Dens_Bi_auf_GaN_bulk_coating} the numerical
plots for Al/Bi and GaN/Bi systems confirm this expectation.
Before finishing the discussion of the energy density, we give in
figs.\ \ref{Fig:Dens_Bi_te} and \ref{Fig:Dens_Bi_tm} two further numerically
computed plots of $\langle u_\parallel^{\rm total} \rangle$ and $\langle u_\perp^{\rm total} \rangle$ for a $5 {\rm nm}$ Bi-coating
on different bulk materials. One sees that the $1/z^2$- and
$1/z^4$-power laws of $\langle u^{{\rm c}}_\parallel \rangle$ derived above leave their imprints
in the TM-mode part of the total energy density. For the TE-mode part of the
energy density one has to distinguish between a metal-metal system
and a polar material-metal system, because for a metal-metal system
at $z \gg d$ the bulk contributions dominate the energy density, but for
a polar bulk material this region is dominated by the contribution of
the metal coating only.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\epsfig{file=F9_Uges_platte_5e-9_Bi_tm.eps, width = 0.9 \textwidth}
\caption{Numerical results for the thermal near-field energy density $\langle u^{\rm total}_\parallel \rangle$
above the layered structure for a substrate consisting of GaN, Pt or Al coated with a $5 {\rm nm}$ Bi
layer at temperature $T = 300 {\rm K}$.}
\label{Fig:Dens_Bi_tm}
\end{minipage}
\end{figure}
\section{Surface plasmon coupling}
\label{Sec4}
The rise in the TM-mode part of the energy density for a polar substrate coated with a metal can
be explained in terms of the low-frequency surface plasmon polariton resonance within
the coating. In the given geometry (see fig.\ \ref{Fig:Configuration}) the surface modes are given by the zeros of the
function~\cite{KliewerFuchs67,J.J.BurkeEtAl1986,H.Raether1980}
\begin{equation}
N_\parallel = 1 - r_\parallel^{12} r_\parallel^{02} {\rm e}^{-2 {\rm i} h_2 d} \equiv 0
\label{Eq:dispersion_generell}
\end{equation}
with $h_2^2 = k_0^2 \epsilon_{r2} - \lambda^2$. This function coincides with the denominator of $T^{\rm c}_\parallel$ (cf.\
$D_\parallel$ in eq.\ (\ref{Eq:Transmission_Coating})). For a non-magnetic material these surface modes are purely TM-polarized
and do exist for materials with a negative permittivity only~\cite{H.Raether1980}. For a polar substrate or bulk material
with a metal coating the Fresnel coefficient $r_\parallel^{12}$ can be approximated by $r_\parallel^{02}$ for all relevant
frequencies. Within this rough approximation the dispersion relation in eq.\ (\ref{Eq:dispersion_generell}) coincides with the
dispersion relation of a free standing metal film surrounded by a vacuum only. Therefore the conclusions drawn for a free
standing metal film~\cite{Biehs2007} can be applied to the metal-coated polar substrate.
It follows~\cite{H.Raether1980,Biehs2007} that for coatings thinner than the skin depth $d_{\rm s}$ the two degenerate surface plasmon
branches with the resonance frequency $\omega_s \approx \omega_{p}/\sqrt{2}$ split into two non-degenerate branches given by~\cite{H.Raether1980}
\begin{equation}
\omega_\pm = \frac{\omega_{p2}}{\sqrt{2}} \sqrt{1 \pm {\rm e}^{-\lambda d}},
\label{Eq:SPP_branches}
\end{equation}
where for convenience the plasma model is used to describe the permittivity. As expressed by eq.\ (\ref{Eq:SPP_branches}) the
resonance frequency of the high-frequency surface plasmon polariton branch $\omega_+$ goes to the plasma frequency $\omega_{p2}$
of the coating, and the resonance frequency of the low-frequency branch $\omega_-$ goes to zero for very
thin coatings, i.e., for $\lambda d \ll 1$. Due to the fact that the $\lambda$-integral for the energy density
in eq.\ (\ref{Eq:Energy_density}) is dominated by lateral wave vectors of the order
$\lambda \approx z^{-1}$, for $z \ll d$ the splitting of the surface plasmon branch cannot be observed, since
$\lambda d \gg 1$. In this case one obtains the same energy density as in the case of an infinitely thick coating.
On the other hand, for observation distances $z \gg d$ in the near field above the coated material the surface
plasmon coupling leads to a splitting of the surface plasmon branches, since
$\lambda d \ll 1$ in this case.
Therefore at these distances the resonance of the low-frequency branch $\omega_-$ will go to frequencies which
are accessible thermally, leading to an increase in the thermal near-field energy density.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\epsfig{file=F10_uspektrump_intermediate_bi_gan_20.eps, width = 0.9 \textwidth}
\caption{Plot of the LDOS of the TM modes of a layered system for different thickness $d$ of a Bi-coating
on a GaN-substrate. The frequencies are normalized to the plasma frequency of the coating material. }
\label{Fig:LDOS_Bi_GaN}
\end{minipage}
\end{figure}
In fig.\ \ref{Fig:LDOS_Bi_GaN} we plot the local density of states (LDOS) defined in~\cite{JoulainEtAl03} for the
TM-modes only. One observes how the resonance at $\omega_s$ splits into two resonances, where the high-frequency resonance
goes to $\omega_{p2}$ of the coating and the low-frequency resonance goes straight to zero.
Thus, it reaches the thermally accessible region
for thin coatings and increases the LDOS in that region, and therefore also the thermal near-field energy density
leading to the $z^{-2}$-power law.
For a metal substrate coated with a metallic material, the dispersion relation in eq.\ (\ref{Eq:dispersion_generell}) can be
approximated in the near-field region with $\lambda \gg k_0$ as
\begin{equation}
\frac{\epsilon_{r2} - \epsilon_{r1}}{\epsilon_{r2} + \epsilon_{r1}} \frac{\epsilon_{r2} - 1}{\epsilon_{r2} + 1} {\rm e}^{-2 \lambda d} = 1
\end{equation}
which leads again within the plasma model to two surface plasmon polariton branches. In this case, for $z \gg d$,
the resonance frequencies of the surface plasmon polariton branches go to the plasma frequency of the
coating $\omega_{p2}$ and the surface
plasmon resonance frequency $\omega_{p1}/\sqrt{2}$ for arbitrarily thin coatings. Therefore the surface plasmon polariton
coupling will not lead to an increase of the LDOS in the thermally accessible region, since for real metals the plasma
frequencies are much greater than the thermal frequency $\omega_{\rm th} \approx 10^{14}{\rm s}^{-1}$ at $T = 300 {\rm K}$.
It follows that the thermal near-field energy density is unaffected by the surface plasmon coupling, leading to values
below that of the semi-infinite body, and to a quite different $z^{-4}$-power law for metal substrates as previously shown
in fig.\ \ref{Fig:Dens_Bi_auf_GaN_bzw_Al_tm}.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\epsfig{file=F11_uspektrump_intermediate_bi_pt_20.eps, width = 0.9 \textwidth}
\caption{Plot of the LDOS of the TM modes of a layered system for different thicknessess $d$ of a Bi-coating
on a Pt-substrate. The frequencies are normalized to the plasma frequency of the coating material. }
\label{Fig:LDOS_Bi_Pt}
\end{minipage}
\end{figure}
In fig.\ \ref{Fig:LDOS_Bi_Pt} we plot the LDOS for the TM-modes for a Bi-coating on a Pt-substrate.
The splitting of the surface plasmon resonance into two resonances is clearly visible.
Here the high-frequency resonance goes to the plasma frequency of the coating and the low-frequency resonance
goes to the surface plasma resonance of the substrate given by
$\omega_{p1}/\sqrt{2} = 8\cdot10^{15} {\rm s}^{-1} = 0.27 \omega_{p2}$ with
$\omega_{p2} = 2.1\cdot10^{16}{\rm s}^{-1}$~\cite{AshcroftMermin76}.
\section{Thermal near-field radiation}
\label{Sec5}
In this last section we discuss the radiative near-field heat transfer between a semi-infinite body
and a coated semi-infinite body as sketched in fig.\ \ref{Fig:Configuration2}. Since the calculation follows the
well-established rules, we proceed directly to the result for the Poynting vector in this geometry, assuming
$T_1 \neq 0$ for the material at $z < 0$ and $T_3 \neq 0$ for the layered structure at $z > a$. With $\epsilon_2 = \epsilon_0$,
the result takes the form
\begin{equation}
\begin{split}
\langle S_{z} \rangle &= \int\!{\rm d} \omega\, \frac{E (\omega,T_1) - E (\omega,T_3)}{(2 \pi)^2} \biggl\{ \int_0^{k_0}\!\!\!{\rm d} \lambda\, \lambda \frac{(1 - |r_\perp^{21}|^2)(1 - |R_\perp|^2)}{|N_\perp'|^2} \\
&\quad + \int_{k_0}^\infty\!\!\!{\rm d} \lambda\, \lambda \frac{4 \Im(r_\perp^{21}) \Im(R_\perp) {\rm e}^{- 2 \gamma a}}{|N_\perp'|^2} \, + \, \parallel \biggr\},
\end{split}
\label{Eq:Polder_van_Hove_Schicht_kap5}
\end{equation}
where the symbol $\parallel$ abbreviates the corresponding expressions for the TM-modes, and
with the usual Fresnel coefficients $r_\perp$ and $r_\parallel$. In addition,
\begin{equation}
R = \frac{r^{23} + r^{34} {\rm e}^{2 {\rm i} h_3 d}}{1 - r^{34} r^{32} {\rm e}^{2 {\rm i} h_3 d}} \quad\text{and}\quad
N' = 1 - r^{21} R {\rm e}^{2 {\rm i} h_2 a}
\end{equation}
for TE- and TM-polarization, respectively. It can be easiliy checked that for $d \rightarrow \infty$ this expression
reduces to the Polder-van-Hove (PvH) result~\cite{PolderVanHove71} for the near-field radiative heat transfer
between two semi-infinite bodies.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\epsfig{file=F12_platte2bulk.eps, width = 0.6 \textwidth}
\caption{Sketch of the configuration for the near-field radiative heat transfer between a semi-infinite medium at
$z \leq 0$ and a coated semi-infinite medium at $z \geq a$.}
\label{Fig:Configuration2}
\end{minipage}
\end{figure}
It is well-known~\cite{PolderVanHove71} that the radiative heat transfer between two metals described by the
PvH expression~\cite{A.I.VolokitinB.N.J.Persson2001} is dominated by the TE-modes, whereas the radiative heat
transfer between a metal and a polar material or two polar materials,
respectively, is dominated by the TM-modes giving
\begin{equation}
\langle S_\parallel \rangle \propto \frac{1}{a^2} \quad\text{and}\quad \langle S_\perp \rangle \propto {\rm const}
\end{equation}
in the near-field region. Hence, the exponents of the $1/z^3$- and $1/z$-dependence of the TM- and TE-mode
parts of the thermal near-field energy density of a half-space are reduced by one.
It is to be expected that the radiative heat transfer between a semi-infinite body and a layered structure with
a thin coating of thickness $d \ll d_{\rm s}$ will again resemble the usual PvH expression for $a \ll d$,
since the energy density above the layered structure coincides in this case with that of a semi-infinite body consisting
of the coating material only. In the opposite case, for $a \gg d$, the radiative heat transfer should be
determined by the change in the thermal near-field energy density described in
the preceding section.
Furthermore one expects that when taking a metallic material for medium 1 the TE-modes of the layered
structure dominate the heat transfer, so that the radiative heat transfer should behave similar to the thermal
near-field energy density $\langle u_\perp^{\rm total} \rangle$ plotted in fig.\ \ref{Fig:Dens_Bi_te}.
Choosing Au for medium 1 we get the near-field radiative heat transfer plotted in fig.\
\ref{Fig:S_Halbraum_Au_Bi_GaN_bzw_Pt}. Indeed this figure fully confirms this expectation.
Moreover, using a metal substrate such as Pt for medium 4, the
radiative heat transfer rises over the PvH-result for a Au-Bi configuration, as is explained by the contribution of
the Pt-substrate, so that in this case the radiative heat transfer in the layered structure
is in prinicple, given by the PvH-result for a Au-Pt configuration.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\centering
\epsfig{file=F13_Sges_Au_hr_5nm_Bi_auf_GaN_bzw_Pt_30.eps, width=0.9\textwidth}
\caption{Numerical results for the near-field radiative heat transfer between a semi-infinite Au-body
with $T_1 = 300 {\rm K}$ and a coated semi-infinite GaN- or Pt-substrate with $T_3 = 0 {\rm K}$,
as functions of the gap width $a$.
The thickness of the Bi-coating is chosen to be $5 {\rm nm}$.}
\label{Fig:S_Halbraum_Au_Bi_GaN_bzw_Pt}
\end{minipage}
\end{figure}
On the other hand, choosing GaN for medium 1, we expect dominance of the TM-mode energy density depicted in
fig.\ \ref{Fig:Dens_Bi_tm}, but with reduced power laws for $a \gg d$, i.e., the $1/z^2$-power law should
lead to a radiative heat transfer proportional to $1/a$, whereas
the $1/z^4$-power law should lead to a radiative heat transfer proportional to $1/a^3$. This is exactly what is seen in the
numerical results plotted in fig.\ \ref{Fig:S_Halbraum_GaN_Bi_GaN_bzw_Pt}. Thus, it is possible to understand
the near-field radiative heat transfer qualitatively from the thermal energy density of the considered materials.
Even more interesting, the enhancement in the thermal near-field energy density due to surface plasmon polariton
coupling in the coating material can be observed in the radiative heat transfer in a slab geometry as sketched in fig.\ \ref{Fig:Configuration2}.
\begin{figure}[Hhbt]
\centering
\begin{minipage}[t]{0.9\textwidth}
\centering
\epsfig{file=F14_Sges_GaN_hr_5nm_Bi_auf_GaN_bzw_Pt_30.eps, width=0.9\textwidth}
\caption{Numerical results for the near-field radiative heat transfer between a semi-infinite GaN-body
with $T_1 = 300 {\rm K}$ and a coated semi-infinite GaN- or Pt-substrate with $T_3 = 0 {\rm K}$,
as functions of the gap width $a$.
The thickness of the Bi-coating is chosen to be $5 {\rm nm}$.}
\label{Fig:S_Halbraum_GaN_Bi_GaN_bzw_Pt}
\end{minipage}
\end{figure}
\section{Conclusions}
\label{Sec6}
In this paper, we have given a discussion of the thermal radiation and the thermal near-field energy density
of a metal-coated substrate. It has been shown that the maximum of the thermal radiation, which
is observed for free metal films at a certain thickness~\cite{Biehs2007}, does not appear for coated
materials, since for thin coatings the thermal radiation of the substrate hides this maximum.
On the other hand, the increase in the thermal near-field energy density of a free standing metal film~\cite{Biehs2007}
due to surface plasmon polariton coupling inside the metal coating has also been found for a coated substrate, when a polar material
is used as substrate. For metal coatings on metal substrates such an increase does not exist. Moreover, for metallic substrates the
thermal near-field energy density $\langle u_\parallel^{\rm total} \rangle$ for observation distances $z \gg d$ and
coating thickness $d \ll d_{\rm s}$ is some orders of magnitude smaller than for a polar substrate (with the same coating),
obeying a rather different power law. This difference in behaviour resulting from the interchange
of the substrate material can be explained with the surface plasmon polariton coupling: For a polar substrate the thermally accessible
LDOS will be enhanced due to the low-frequency surface plasmon resonance, which goes to zero frequency for arbitrarily thin coatings,
whereas for a metal substrate this resonance goes to the surface plasmon resonance of the substrate for arbitrarily thin coatings and
can therefore not be accessed thermally for plasma frequencies much greater than the thermal frequency.
In the last part we have shown that the differences investigated for the thermal near-field energy density of a coated material
leave their imprints in the near-field radiative heat transfer between a semi-infinite body and a coated semi-infinite body.
Using a metal or a polar material allows one to 'select' the TE- or TM-mode part of the thermal
near-field energy density of the coated material to dominate the radiative near-field heat transfer. Therefore, it is possible to
observe the TM-mode-enhancement due to surface plasmon polariton coupling inside the coating by thermal heat transfer experiments.
Due to the fact that the expressions for the near-field radiative heat transfer and the vacuum
friction~\cite{J.B.Pendry1997,A.I.VolokitinB.N.J.Persson2003} are fairly
similar, the discussed effect should also be observable for vacuum friction between coated materials.
Since a polarizable particle or an atom couples to the electric field, the radiative heat
transfer~\cite{Dorofeyev98,MuletEtAl01,A.I.VolokitinB.N.J.Persson2002}, the spontaneous emission
rate~\cite{J.M.Wylie1984,G.W.Ford1984,M.S.Tomas1995,H.T.DungEtAl2002} near a hot body and the thermal Casimir-Polder
potential~\cite{P.Milonni1994,C.HenkelEtAl2002,A.I.VolokitinB.N.J.Persson2002} should be proportional
to $\langle \mathbf{E}^2 \rangle \propto \langle u_\parallel \rangle$ in the near field, so that the discussed
enhancement of the TM-mode part of the thermal near field should
also enhance the near-field radiative heat transfer between a small particle and a coated material, the spontaneous emission rate
of an atom near a hot coated material, and the thermal Casimir-Polder potential, respectively. Moreover, the spin flip rate of
atoms~\cite{C.HenkelEtAl1999,P.K.Rekdal2004} above a layered structure, which is in principle proportional
to $\langle \mathbf{B}^2 \rangle \propto \langle u_\perp \rangle$, will also be changed by the use of thin coatings
on appropriate substrates. Furthermore, it appears possible that the coherence of the thermal near
field~\cite{CarminatiGreffet99,C.HenkelEtAl2000} can be controlled by the use of different metal coatings, since the surface
plasmon resonance frequency can be changed by the choice of the thickness of the coating. In this sense the
discussion of the thermal energy density has a much broader field of application than the radiative heat transfer and the vacuum friction.
The author acknowledges support from the Studienstiftung des deutschen Volkes. Furthermore he thanks O. Huth, F. R\"uting, D. Reddig, and
M. Holthaus for helpful discussions and kind criticism.
|
\section{Introduction}\label{sec1}
In this paper we study a Bayesian approach to estimating
a parameter $\m$ from an observation $Y$ following the model
\begin{equation}
\label{EqProblem}
Y = K\m+ \frac{1}{\sqrt{n}}Z.
\end{equation}
The unknown parameter $\m$ is an element
of a separable Hilbert space $H_1$, and is mapped into another
Hilbert space $H_2$ by a known, injective, continuous
linear operator $K\dvtx H_1 \to H_2$.
The image $K\m$ is perturbed by unobserved, scaled Gaussian white
noise $Z$.
There are many special examples of this infinite-dimensional regression model,
which can also be viewed as an idealized
version of other statistical models, including density estimation.
The inverse problem of estimating $\m$ has
been studied by both statisticians and numerical
mathematicians (see, e.g., \cite
{Donoho,Cavalier,Munk,RuymgaartII,RuymgaartIII,Stuart}
for reviews), but rarely from a theoretical Bayesian perspective;
exceptions are \cite{Cox} and \cite{Simoni}.
The Bayesian approach to (\ref{EqProblem}) consists of putting a prior
on the parameter~$\m$, and computing the posterior distribution. We
study Gaussian priors, which are conjugate to the model, so that the
posterior distribution is also Gaussian and easy to derive. Our interest
is in studying the properties of this posterior distribution, under
the ``frequentist'' assumption that the data $Y$ has been generated
according to the model (\ref{EqProblem}) with a~given ``true''
parameter $\m_0$. We investigate whether and at what rate the
posterior distributions contract to $\m_0$ as $n\ra\infty$ (as in
\cite{GGvdV}), but have as main interest the performance of
credible sets for measuring the uncertainty about the parameter.
A Bayesian \textit{credible set} is defined as a central region in the
posterior distribution of specified posterior probability, for
instance, 95\%. As a consequence of the Bernstein--von Mises theorem
credible sets for smooth \textit{finite-dimensional}
parametric models are asymptotically equivalent to confidence regions
based on the maximum likelihood estimator (see, e.g., \cite{vdVAS},
Chapter~10), under mild conditions on the prior. Thus, ``Bayesian
uncertainty'' is equivalent to ``frequentist uncertainty'' in these
cases, at least for large~$n$. However, there is no corresponding
Bernstein--von Mises theorem in nonparametric
Bayesian inference, as noted in \cite{Freedman}. The
performance of Bayesian credible sets in these situations has
received little attention, although in practice such sets are
typically provided as indicators of uncertainty, for instance, based
on the spread of the output of a (converged) MCMC run. The paper \cite{Cox}
did tackle this issue and came to the alarming conclusion
that Bayesian credible sets have frequentist coverage zero.
If this were true, many data analysts would
(justifiably) distrust the spread in the posterior distribution as a
measure of uncertainty. For other results see
\cite{Bontemps,Ghosal99,Ghosal00} and~\cite{Leahu}.
The model considered in \cite{Cox} is equivalent to
our model (\ref{EqProblem}), and a good starting point
for studying these issues. More precisely, the conclusion
of~\cite{Cox} is that \textit{for almost every parameter $\m_0$ from
the prior the
coverage of a credible set} (\textit{of any level}) \textit{is} 0. In the present paper
we show that this is only part of the story, and, taken by itself,
the conclusion is misleading. The coverage
depends on the true parameter $\m_0$ and the prior together,
and it can be understood in terms of a bias-variance trade-off,
much as the coverage of frequentist nonparametric procedures.
A nonparametric procedure that oversmoothes the truth (too big a bandwidth
in a~frequentist procedure, or a prior that puts too much weight
on ``smooth'' parameters) will be biased, and a confidence
or credible region based on such a procedure will be both
too concentrated and wrongly located, giving zero coverage. On the other
hand, undersmoothing does work (to a certain extent),
also in the Bayesian setup, as we show below.
In this light we reinterpret the conclusion
of \cite{Cox} to be valid only in the oversmoothed case (notwithstanding
a conjecture to the contrary in the Introduction of this paper;
see page 905, answer to objection~4). In the undersmoothed case
credible regions are conservative in general, with coverage
tending to 1. The good news is that typically they are
of the correct order of magnitude, so that they do give
a reasonable idea of the uncertainty in the estimate.
Of course, whether a prior under- or oversmoothes depends
on the regularity of~the true parameter. In practice, we may not
want to consider this known, and adapt the prior smoothness
to the data. In this paper we do consider the effect of changing
the ``length scale'' of a prior, but do not study data-dependent
length scales. The effect of setting the latter by, for example, an
empirical or full Bayes method will require further study.
Credible sets are by definition ``central regions'' in the
posterior distribution. Because the posterior distribution is a random
probability measure on the Hilbert space $H_1$, a ``central ball'' is
a natural shape of such a set, but it has the disadvantage that it is
difficult to visualize. If the Hilbert space is a function space,
then \textit{credible bands} are more natural. These correspond to
simultaneous credible intervals for the function at a point, and
can be obtained from the (marginal) posterior distributions of a set
of linear functionals. Besides the full posterior distribution, we
therefore study its marginals for linear functionals. The same issue
of the dependence of coverage on under- and oversmoothing arises, except
that ``very smooth'' linear functionals cancel the inverse nature
of the problem, and do allow a~Bernstein--von Mises theorem for a
large set of priors. Unfortunately point evaluations are usually
not smooth in this sense.
Thus, we study two aspects of inverse problems---recovering the full
parameter $\m$ (Section \ref{SectionFull}) and recovering linear
functionals (Section \ref{SectionLinear}). We obtain the rate of
contraction of the posterior distribution in both settings, in its
dependence on parameters of the prior. Furthermore, and most
importantly, we study the ``frequentist'' coverage of credible regions
for $\m$ in both settings, for the same set of priors. In the next
section we give a more precise statement of the problem, and in
Section~\ref{SectionPriorPosterior} we describe the priors that we
consider and derive the corresponding posterior distributions. In
Section \ref{SectionVolterra} we illustrate the results by
simulations and pictures in the particular example that $K$ is the
Volterra operator. Technical proofs are placed in
Sections~\ref{SectionProofs} and \ref{SectionTechnical}
at the end of the paper.
Throughout the paper $\langle\cdot,\cdot\rangle_1$ and \mbox{$\|\cdot\|_1$},
and $\langle\cdot,\cdot\rangle_2$ and \mbox{$\|\cdot\|_2$} denote the inner
products and norms of the Hilbert spaces $H_1$ and $H_2$. The adjoint
of an operator $A$ between two Hilbert spaces is denoted by $A^T$.
The Sobolev space $S^\b$ with its norm \mbox{$\|\cdot\|_\b$} is defined
in (\ref{EqDefSobolev}). For two sequences $(a_n)$ and~$(b_n)$ of numbers
$a_n\asymp b_n$ means that $|a_n/b_n|$ is bounded away from zero
and infinity as $n\ra\infty$, and $a_n\lesssim b_n$ means that
$a_n/b_n$ is bounded.
\section{Detailed description of the problem}
\label{SectionDetailed}
The noise process $Z$ in (\ref{EqProblem}) is the
standard normal or \textit{iso-Gaussian process}
for the Hilbert space $H_2$. Because this is
not realizable as a random element in $H_2$, the model (\ref{EqProblem})
is interpreted in process form (as in \cite{Munk}). The iso-Gaussian process
is the zero-mean Gaussian process $Z=(Z_h\dvtx h\in H_2)$ with covariance function
$\E Z_{h}Z_{h'}=\langle h,h'\rangle_2$, and the measurement
equation (\ref{EqProblem}) is interpreted in that we
observe a Gaussian process $Y=(Y_h\dvtx h\in H_2)$ with
mean and covariance functions
\begin{equation}
\label{EqMeanCovY}
\E Y_h=\langle K\m,h\rangle_2,\qquad
\operatorname{cov} (Y_h,Y_{h'})=\frac1n\langle h,h'\rangle_2.
\end{equation}
Sufficiency considerations show that it is statistically equivalent
to observe the subprocess $(Y_{h_i}\dvtx i\in\NN)$, for any orthonormal
basis $h_1,h_2,\ldots$ of $H_2$.
If the operator $K$ is compact, then the \textit{spectral decomposition}
of the self-adjoint operator $K^TK\dvtx H_1\to H_1$ provides a convenient
basis. In the compact case the operator $K^TK$ possesses countably
many positive eigenvalues $\k_i^2$ and there is a corresponding
orthonormal basis $(e_i)$ of $H_1$ of eigenfunctions (hence,
$K^TKe_i=\k_i^2e_i$ for $i\in\NN$; see, e.g., \cite{Rudin}). The
sequence $(f_i)$ defined by $Ke_i=\k_if_i$ forms an orthonormal
``conjugate'' basis of the range of $K$ in $H_2$. An element $\m\in
H_1$ can be identified with its sequence $(\m_i)$ of coordinates
relative to the eigenbasis $(e_i)$, and its image
$K\m=\sum_i\m_iKe_i=\sum_i\mu_i\k_if_i$ can be identified with its
coordinates $(\m_i\k_i)$ relative to the conjugate basis~$(f_i)$. If
we write $Y_i$ for $Y_{f_i}$, then (\ref{EqMeanCovY}) shows that
$Y_1,Y_2,\ldots$ are independent Gaussian variables with means $\E
Y_i=\mu_i\k_i$ and variance $1/n$. Therefore, a~concrete equivalent
description of the statistical problem is to \textit{recover the
sequence $(\m_i)$ from independent observations $Y_1,Y_2,\ldots$
with $N(\mu_i\k_i,1/n)$-distributions}.
In the following we do not require $K$ to be compact, but
we do assume the existence of an orthonormal basis of eigenfunctions of $K^TK$.
The main additional example we then cover is the \textit{white noise
model}, in which $K$ is the identity operator. The description of
the problem remains the same.
If $\k_i\ra0$, this problem is ill-posed, and the recovery of
$\m$ from $Y$ an \textit{inverse problem}. The ill-posedness can be
quantified by the speed of decay $\k_i\da0$. Although the tools are
more widely applicable, we limit ourselves to the \textit{mildly
ill-posed} problem (in the terminology of \cite{Cavalier}) and
assume that the decay is polynomial: for some $p\ge0$,
\[
\k_i \asymp i^{-p}.
\]
Estimation of $\m$ is harder if the decay is faster (i.e., $p$ is larger).
The difficulty of estimation may be measured by the minimax
risks over the scale of \textit{Sobolev spaces} relative to the
orthonormal basis $(e_i)$ of eigenfunctions of $K^TK$. For $\b>0$ define
\begin{equation}
\label{EqDefSobolev}
\|\m\|_\b= \sqrt{\sum_{i=1}^\infty\m_i^2i^{2\b}}
\qquad\mbox{if }
\m= \sum_{i=1}^\infty\m_ie_i.
\end{equation}
Then the Sobolev space of order $\b$ is
$S^\b= \{\m\in H_1\dvtx \|\m\|_\b<\infty\}$.
The minimax rate of estimation over the unit\vadjust{\goodbreak} ball of this space
relative to the loss $\|t-\m\|_1$ of an estimate $t$ for $\m$ is
$n^{-\b/(1+2\b+2p)}$.
This rate is attained by various ``regularization'' methods,
such as generalized \textit{Tikhonov} and \textit{Moore--Penrose}
regularization \cite{Mair,Bertero,Cavalier,Goldenshluger,Munk}.
The Bayesian approach is closely connected to these methods:
in Section~\ref{SectionPriorPosterior}
the posterior mean is shown to be a regularized estimator.
Besides recovery of the full parameter $\m$, we consider estimating
linear functionals $L\m$. The minimax rate for such functionals over
Sobolev balls depends on $L$ as well as on the parameter of the Sobolev
space. Decay of the coefficients of $L$ in the eigenbasis may alleviate
the level of ill-posedness, with rapid decay even bringing the
functional in the domain of ``regular'' $n^{-1/2}$-rate estimation.
\section{Prior and posterior distributions}
\label{SectionPriorPosterior}
We assume a mean-zero Gaussian prior for the parameter $\m$.
In the next three paragraphs we recall some essential facts on
Gaussian distributions on Hilbert spaces.
A \textit{Gaussian distribution} $N(\n,\Lambda)$
on the Borel sets of the Hilbert space
$H_1$ is characterized by a \textit{mean} $\n$, which
can be any element of $H_1$, and a \textit{covariance operator}
$\Lambda\dvtx H_1\to H_1$, which is a nonnegative-definite,
self-adjoint, linear operator
\textit{of trace class}: a compact operator with
eigenvalues~$(\l_i)$ that are summable $\sum_{i=1}^\infty\l
_i<\infty$
(see, e.g., \cite{Skorohod}, pages 18--20). A random element $G$ in
$H_1$ is
$N(\n,\Lambda)$-distributed if and only if
the stochastic process $(\langle G, h\rangle_1\dvtx h\in H_1)$
is a Gaussian process with mean and covariance functions
\begin{equation}
\label{EqProcessG}
\E\langle G, h\rangle_1=\langle\n, h\rangle_1,\qquad
\operatorname{cov} (\langle G, h\rangle_1,\langle G,
h'\rangle_1)
=\langle h, \Lambda h'\rangle_1.
\end{equation}
The coefficients $G_i=\langle G,\varphi_i\rangle_1$
of $G$ relative to an orthonormal eigenbasis $(\varphi_i)$
of $\Lambda$ are independent, univariate Gaussians
with means the coordinates $(\n_i)$ of the mean vector $\n$
and variances the eigenvalues $\l_i$.
The iso-Gaussian process $Z$ in (\ref{EqProblem}) may be thought of as
a $N(0,I)$-distributed Gaussian element, for $I$ the identity operator
(on $H_2$), but as $I$ is not of trace class, this distribution is not
realizable as a proper random element in $H_2$. Similarly, the data
$Y$ in (\ref{EqProblem}) can be described as having a~$N(K\m,
n^{-1}I)$-distribution.
For a stochastic process $W=(W_h\dvtx h\in H_2)$ and a
continuous, linear operator $A\dvtx H_2\to H_1$, we define the
transformation $AW$ as the stochastic process with coordinates
$(AW)_h=W_{A^T h}$, for $h\in H_1$. If the process $W$ arises as
$W_h=\langle W,h\rangle_2$ from a random element $W$ in the Hilbert
space $H_2$, then this definition is consistent with identifying the
random element $A W$ in $H_1$ with the process $(\langle
AW,h\rangle_1\dvtx h\in H_1)$, as in (\ref{EqProcessG}) with $G=AW$.
Furthermore, if~$A$ is a \textit{Hilbert--Schmidt} operator (i.e., $AA^T$
is of trace class), and $W=Z$ is the iso-Gaussian process, then the
process $AW$ can be realized as a random variable in $H_1$ with a
$N(0,AA^T)$-distribution.
In the Bayesian setup the prior, which we take
$N(0,\Lambda)$, is the marginal distribution of $\mu$,
and the noise $Z$ in (\ref{EqProblem}) is
considered independent of $\m$. The joint distribution of $(Y,\m)$
is then also Gaussian, and so is the conditional distribution of
$\m$ given $Y$, the posterior distribution of $\m$.
In general, one must be a bit careful with manipulating possibly
``improper'' Gaussian distributions (see \cite{Mandelbaum}), but
in our situation the posterior is a
proper Gaussian conditional distribution on $H_1$.
\begin{proposition}[(Full posterior)]\label{Posterior}
If $\m$ is $N(0,\Lambda)$-distributed and
$Y$ given $\m$ is $N(K\m, n^{-1}I)$-distributed, then
the conditional distribution of $\m$ given $Y$ is
Gaussian $N(AY, S_n)$ on $H_1$, where
\begin{equation}\label{PostCov}
S_n = \Lambda-A(n^{-1}I+K\Lambda K^T)A^T,
\end{equation}
and $A\dvtx H_2\to H_2$ is the continuous linear operator
\begin{equation}\label{A}
A=\Lambda^{1/2}\biggl(\frac1{n}I+\Lambda^{1/2}K^TK\Lambda
^{1/2}\biggr)^{-1}
\Lambda^{1/2}K^T=\Lambda K^T\biggl(\frac1nI+K\Lambda
K^T\biggr)^{-1}.\hspace*{-28pt}
\end{equation}
The posterior distribution is proper (i.e., $S_n$ has finite trace)
and equivalent (in the sense of absolute continuity) to the prior.
\end{proposition}
\begin{pf}
Identity (\ref{A}) is a special case of the
identity $(I+BB^T)^{-1} B=B(I+B^TB)^{-1}$, which is valid
for any compact, linear operator $B\dvtx H_1\to H_2$.
That $S_n$ is of trace class is a consequence of the fact
that it is bounded above by $\Lambda$ (i.e., $\Lambda-S_n$ is nonnegative
definite), which is of trace class by assumption.
The operator $\Lambda^{1/2}K^TK\Lambda^{1/2}\dvtx H_1\to H_1$
has trace bounded by $\|K^TK\|\tr(\Lambda)$ and hence
is of trace class. It follows that the variable
$\Lambda^{1/2}K^TZ$ can be defined as a random element in the
Hilbert space $H_1$, and so can $AY$, for~$A$ given by the first
expression in (\ref{A}). The joint distribution of $(Y,\m)$
is Gaussian with zero mean and covariance operator
\[
\pmatrix{
n^{-1}I + K\Lambda K^T & K \Lambda\cr
\Lambda K^T& \Lambda}.
\]
Using this with the second form of $A$ in (\ref{A}), we can
check that the cross covariance operator
of the variables $\m-AY$ and $Y$ (the latter viewed
as a~Gaussian stochastic process in $\RR^{H_2}$) vanishes and, hence,
these variables are independent.
Thus, the two terms in the decomposition $\m=(\m-AY) +AY$
are conditionally independent and degenerate given $Y$,
respectively. The distribution of $\m-AY$ is zero-mean
Gaussian with covariance operator $\Cov(\m-AY)=\Cov(\m)-\Cov(AY)$,
by the independence of $\m-AY$ and~$AY$.
This gives the form of the posterior distribution.
The final assertion may be proved by explicitly comparing
the Gaussian prior and posterior. Easier is to note that it
suffices to show that the model consisting of all
$N(K\m,n^{-1}I)$-distributions is dominated. In that case
the posterior can be obtained using Bayes' rule, which
reveals the normalized likelihood as a density relative to the
(in fact, \textit{any}) prior. To prove domination, we may consider
equivalently the distributions $\bigotimes_{i=1}^\infty N(\k_i\m_i,n^{-1})$
on $\RR^\infty$ of the sufficient statistic $(Y_i)$ defined
as the coordinates of $Y$ relative to the conjugate spectral
basis. These distributions, for $(\m_i)\in\ell_2$, are
equivalent to the distribution $\bigotimes_{i=1}^\infty N(0,n^{-1})$,
as can be seen with the help of Kakutani's theorem,
the affinity being $\exp(-\sum_i\k_i^2\m_i^2/8)>0$. (This
argument actually proves the well-known fact that
the Gaussian shift experiment obtained by translating
the standard normal distribution on $\RR^\infty$ over its RKHS
$\ell_2$ is dominated.)
\end{pf}
In the remainder of the paper we study the asymptotic behavior of the
posterior distribution, under the assumption that $Y= K\m_0 +
n^{-1/2}Z$ for a~fixed $\m_0 \in H_1$. The posterior is characterized
by its \textit{center} $AY$, the posterior mean, and its \textit{spread},
the posterior covariance operator $S_n$. The first depends on
the data, but the second is deterministic. From a frequentist-Bayes
perspective both are important: one would like the posterior mean
to give a good estimate for $\m_0$, and the spread to give a
good indication of the uncertainty in this estimate.
The posterior mean is a regularization,
of the Tikhonov type, of the naive estimator $K^{-1}Y$. It can also be
characterized as a penalized least squares estimator (see
\cite{Tikhonov,Mathe}): it
minimizes the functional
\[
\m\mapsto\|Y-K\m\|_2^2 + \frac1n\|\Lambda^{-1/2}\m\|^2_1.
\]
The penalty $\|\Lambda^{-1/2}\m\|_1$ is interpreted as $\infty$ if
$\m$ is not in the range of $\Lambda^{1/2}$. Because this range is
precisely\vspace*{1pt} the \textit{reproducing kernel Hilbert space} (RKHS) of the
prior (cf. \cite{vdVvZRKHS}),
with $\|\Lambda^{-1/2}\m\|_1$ as the RKHS-norm of $\m$, the
posterior mean also fits into the general regularization framework
using RKHS-norms (see \cite{Wahba}). In any case the posterior mean is a
well-studied point estimator in the literature on inverse problems.
In this paper we add a Bayesian interpretation to it, and are (more)
concerned with the full posterior distribution.
Next consider the posterior distribution of a linear functional $L\m$
of the parameter. We are not only interested in continuous, linear functionals
$L\m=\langle\m,l\rangle_1$, for some given $l\in H_1$, but also in
certain discontinuous functionals, such as point evaluation in a
Hilbert space of functions. The latter entail some technicalities. We
consider \textit{measurable linear functionals relative to the prior}
$N(0,\Lambda)$, defined in \cite{Skorohod}, pages 27--29, as Borel
measurable maps
$L\dvtx H_1\to\RR$ that are linear on a measurable linear subspace
$\underline H_1\subset H_1$ such that $N(0,\Lambda)(\underline
H_1)=1$. This definition is exactly right to make the marginal
posterior Gaussian.
\begin{proposition}[(Marginal posterior)]\label{PosteriorLinear}
If $\m$ is $N(0,\Lambda)$-distributed and
$Y$ given $\m$ is $N(K\m, n^{-1}I)$-distributed, then
the conditional distribution\vadjust{\goodbreak} of $L\m$ given $Y$ for
a $N(0,\Lambda)$-measurable linear functional $L\dvtx H_1\to\RR$ is
a Gaussian distribution $N(LAY, s_n^2)$ on $\RR$, where
\begin{equation}\label{PostCovL}
s_n^2 = (L\Lambda^{1/2})(L\Lambda^{1/2})^T-LA(n^{-1}I+K\Lambda K^T)(LA)^T,
\end{equation}
and $A\dvtx H_2\to H_2$ is the continuous linear operator defined
in (\ref{A}).
\end{proposition}
\begin{pf}
As in the proof of Proposition \ref{Posterior}, the first term
in the decomposition $L\m=L(\m-AY)+LAY$ is independent of $Y$.
Therefore, the posterior distribution is the marginal
distribution of $L(\m-AY)$ shifted by $LAY$. It suffices
to show that this marginal distribution is $N(0,s_n^2)$.
By Theorem 1 on page 28 in \cite{Skorohod}, there exists a sequence of
continuous linear maps $L_m\dvtx H_1\to\RR$ such that $L_mh\ra Lh$ for all
$h$ in a set with probability one under the prior
$\Pi=N(0,\Lambda)$. This implies that $L_m \Lambda^{1/2}h\ra
L\Lambda^{1/2}h$ for \textit{every} $h\in H_1$. Indeed, if $V=\{h\in
H_1\dvtx L_mh\ra
Lh\}$ and $g\notin V$, then $V_1:=V+ g$ and $V$ are disjoint measurable,
affine subspaces of $H_1$, where $\Pi(V)=1$. The range of $\Lambda^{1/2}$
is the RKHS of $\Pi$ and, hence, if $g$ is in this range, then
$\Pi(V_1)>0$, as $\Pi$ shifted over an element from its
RKHS is equivalent to $\Pi$. But then $V$ and $V_1$ are not disjoint.
Therefore, from the first definition of $A$ in (\ref{A}) we see that
$L_mA\ra LA$, and, hence, $L_m(\m-AY)\ra L(\m-AY)$, almost surely. As
$L_m$ is continuous, the variable $L_m(\m-AY)$ is normally distributed
with mean zero and variance
$L_mS_mL_m^T=(L_m\Lambda^{1/2})(L_m\Lambda^{1/2})^T-
L_mA(n^{-1}I+K\Lambda K^T)(L_mA)^T$, for $S_n$ given by
(\ref{PostCov}). The desired result follows upon taking the limit as
$m\ra\infty$.
\end{pf}
As shown in the preceding proof, $N(0,\Lambda)$-measurable linear
functionals $L$ automatically have the further property that
$L\Lambda^{1/2}\dvtx H_1\to\RR$ is a continuous linear map. This shows
that $LA$ and the adjoint operators $(L\Lambda^{1/2})^T$ and $(LA)^T$
are well defined, so that the formula for $s_n^2$ makes sense. If $L$
is a~\textit{continuous} linear operator, one can also write these
adjoints in terms of the adjoint\vspace*{1pt} $L^T$ of $L$, and express $s_n^2$ in
the covariance operator $S_n$ of Proposition~\ref{Posterior} as
$s_n^2=LS_nL^T$. This is exactly as expected.
In the remainder of the paper we study the full posterior distribution
$N(AY,S_n)$, and its marginals $N(LAY,s_n^2)$. We are particularly
interested in the influence of the prior on the performance of the
posterior distribution for various true parameters~$\m_0$. We study
this in the following setting.
\begin{assumption}\label{common}\label{PPP}
The operators $K^T K$ and $\Lambda$ have the same
eigenfunctions~$(e_i)$, with eigenvalues $(\k_i^2)$ and $(\lambda_i)$,
satisfying
\begin{equation}\label{SetupPPP}
\lambda_i = \t_n^2i^{-1-2\a},\qquad C^{-1}i^{-p} \leq\k_i \leq C i^{-p}
\end{equation}
for some $\a>0 $, $p \geq0$, $C\ge1$ and $\t_n > 0$ such that
$n\t_n^2\ra\infty$.
Furthermore, the true parameter $\m_0$ belongs to $S^\b$ for some $\b>0$:
that is, its coordinates~$(\m_{0,i})$ relative to $(e_i)$ satisfy
$\sum_{i=1}^\infty\m_{0,i}^2i^{2\b}<\infty$.
\end{assumption}
The setting of Assumption \ref{PPP} is a Bayesian extension of the
\textit{mildly ill-posed inverse problem} (cf. \cite{Cavalier}). We
refer to the parameter $\b$ as the ``regularity'' of the
true parameter $\m_0$. In the special case that $H_1$ is a function
space and $(e_i)$ its Fourier basis, this parameter gives smoothness
of $\m_0$ in the classical Sobolev sense. Because the coefficients
$(\m_i)$ of the prior parameter $\m$ are normally
$N(0,\l_i)$-distributed, under Assumption \ref{PPP} we have $\E
\sum_ii^{2\a'}\m_i^2=\t_n^2\sum_ii^{2\a'}\l_i<\infty$ if and
only if
$\a'<\a$. Thus, $\a$ is ``almost'' the smoothness of the parameters
generated by the prior. This smoothness is modified by the
\textit{scaling factor} $\t_n$. Although this leaves the relative sizes
of the coefficients $\m_i$, and hence the qualitative smoothness of
the prior, invariant, we shall see that scaling can completely alter
the performance of the Bayesian procedure. Rates $\t_n\da0$ increase,
and rates $\t_n\ua\infty$ decrease the regularity.
\section{Recovering the full parameter}
\label{SectionFull}
We denote by $\Pi_n(\cdot\,\given Y)$ the posterior distribution
$N(AY,S_n)$, derived in Proposition \ref{Posterior}. Our first
theorem shows that it contracts as $n\ra\infty$ to the true parameter
at a rate $\eps_n$ that depends on all four parameters $\a,\b,\t_n,p$
of the (Bayesian) inverse problem.
\begin{theorem}[(Contraction)]\label{ContrPPP}
If $\m_{0}$, $(\lambda_i)$, $(\k_i)$ and $(\t_n)$ are
as in Assumption~\ref{PPP}, then
$\E_{\m_0}\Pi_n(\m\dvtx \|\m-\m_0\|_1 \geq M_n \eps_n \given
Y) \ra0$,
for every $M_n\ra\infty$, where
\begin{equation}\label{Rate}
\eps_n =(n\t_n^2)^{-{\b}/({1+2\a+2p})\wedge1} + \t_n(n\t
_n^2)^{-{\a}/({1+2\a+2p})}.
\end{equation}
The rate is uniform over $\m_0$ in balls in $S^\b$. In particular:
\begin{longlist}[(iii)]
\item[(i)] If $\t_n \equiv1$, then $\eps_n = n^{-{(\a\wedge\b
)}/{(1+2\a+2p)}}$.
\item[(ii)] If $\b\leq1+2\a+ 2p$ and $\t_n \asymp n^{(\a-\b
)/(1+2\b+2p)}$,
then $\eps_n = n^{-\b/(1+2\b+2p)}$.
\item[(iii)] If $\b> 1+2\a+ 2p$, then $\eps_n \gg n^{-\b/(1+2\b+2p)}$,
for every scaling $\t_n$.
\end{longlist}
\end{theorem}
The minimax rate of convergence over a Sobolev ball $S^\b$ is of the
order $n^{-\b/(1+2\b+2p)}$ (see \cite{Cavalier}). By (i) of the
theorem the
posterior contraction rate is the same
if the regularity of the prior is chosen to match the
regularity of the truth ($\a=\b$) and the scale $\t_n$ is fixed.
Alternatively, the optimal rate is also attained by appropriately
scaling ($\t_n \asymp n^{(\a-\b)/(1+2\b+2p)}$, determined
by balancing the two terms in ${\varepsilon}_n$) a prior that is
regular enough ($\b\le1+2\a+2p$). In all other cases
(no scaling and $\a\not=\b$, or any scaling combined with a
rough prior $\b> 1+2\a+ 2p$), the contraction rate is
slower than the minimax rate.
That ``correct'' specification of the prior gives the optimal rate is
comforting to the true Bayesian. Perhaps the main message of the
theorem is that even if the prior mismatches the truth, it may be
scalable to give the optimal rate. Here, similar as found by
\cite{vdVvZRescaledGaussian} in a different setting, a smooth prior
can be scaled to make it ``rougher'' to any degree, but a rough prior
can be ``smoothed'' relatively little (namely, from $\a$ to any
$\b\le1+2\a+2p$). It will be of interest to investigate a full or
empirical Bayesian approach to set the scaling parameter.
Bayesian inference takes the spread in the posterior distribution as
an expression of uncertainty. This practice is not validated by (fast)
contraction of the posterior. Instead we consider the
frequentist coverage of credible sets. As the posterior
distribution is Gaussian, it is natural to center a~credible region
at the posterior mean. Different shapes of such a set could be
considered. The natural counterpart of the preceding theorem is to consider
balls. In the next section we also consider bands. (Alternatively, one
might consider ellipsoids, depending on geometry of the support of the
posterior.)
Because the posterior spread $S_n$ is deterministic, the radius is the
only degree of freedom when we choose a ball, and we fix it by the
desired ``credibility level'' $1-\g\in(0,1)$. A \textit{credible
ball} centered at the posterior mean $AY$ takes the form, where
$B(r)$ denotes a ball of radius $r$ around 0,
\begin{equation}
\label{EqCredReg}
AY+B(r_{n,\g}):=\{\m\in H_1 \dvtx \|\m-AY\|_1<r_{n,\g}\},
\end{equation}
where the radius $r_{n,\g}$ is determined so that
\begin{equation}
\label{EqRadius}
\Pi_n\bigl(AY+B(r_{n,\g})\given Y\bigr)=1-\g.
\end{equation}
Because the posterior spread $S_n$ is not dependent on the data,
neither is the radius~$r_{n,\g}$.
The frequentist \textit{coverage} or confidence of the set
(\ref{EqCredReg}) is
\begin{equation}
\label{EqCoverage}
\Pr_{\m_0}\bigl( \m_0\in AY+B(r_{n,\g})\bigr),
\end{equation}
where under the probability measure $\Pr_{\m_0}$
the variable $Y$ follows (\ref{EqProblem}) with $\m=\m_0$. We shall consider
the coverage as $n\ra\infty$ for fixed $\m_0$, uniformly
in Sobolev balls, and also along sequences $\m_0^n$ that change
with $n$.
The following theorem shows that the relation of the coverage to the
credibility level $1-\g$ is mediated by all parameters of the
problem. For further insight, the credible region is also compared
to the ``correct'' frequentist confidence ball $AY+B(\wt r_{n,\g})$,
which has radius $\wt r_{n,\g}$ chosen so that the probability in
(\ref{EqCoverage}) with $r_{n,\g}$ replaced by $\wt r_{n,\g}$ is
equal to $1-\g$.
\begin{theorem}[(Credibility)]\label{CrS}\label{CrNS}
Let $\m_{0}$, $(\l_i)$, $(\k_i)$, and $\t_n$
be as in Assumption~\ref{PPP}, and set $\wt\b= \b\wedge(1+2\a+2p)$.
The asymptotic coverage of the credible region (\ref{EqCredReg}) is:
\begin{longlist}[(iii)]
\item[(i)] 1, uniformly in $\m_0$ with $\|\m_0\|_\b\leq1$,
if $\t_n \gg n^{(\a-\wt\b)/(1+2\wt\b+2p)}$;
in this case $\wt r_{n,\g}\asymp r_{n,\g}$.
\item[(ii)] 1, for every fixed $\m_0 \in S^\b$, if $\b< 1+2\a+2p$ and
$\t_n \asymp n^{(\a-\wt\b)/(1+2\wt\b+2p)}$;
$c$,
along some $\m_{0}^{n}$ with ${\sup_n }\|\m_0^{n}\|_\b< \infty$,
if $\t_n \asymp n^{(\a-\wt\b)/(1+2\wt\b+2p)}$ (any $c\in[0,1)$).
\item[(iii)] 0, along some $\m_0^n$ with ${\sup_n }\|\m_0^{n}\|_\b<
\infty$,\vadjust{\goodbreak}
if $\t_n \ll n^{(\a-\wt\b)/(1+2\wt\b+2p)}$.
\end{longlist}
If $\t_n\equiv1$, then the cases \textup{(i), (ii)} and \textup{(iii)} arise if
$\a<\b$, $\a=\b$ and $\a>\b$, respectively. In case \textup{(iii)} the sequence
$\m_0^n$ can then be chosen fixed.
\end{theorem}
The theorem is easiest to interpret in the situation without scaling
(\mbox{$\t_n\equiv1$}). Then oversmoothing the prior [case (iii): $\a>\b$] has
disastrous consequences for the coverage of the credible sets,
whereas undersmoothing [case~(i): $\a<\b$] leads to conservative
confidence sets. Choosing a prior of correct regularity [case (ii):
$\a=\b$] gives mixed results.
Inspection of the proofs shows that the lack of coverage in case of
oversmoothing arises from a bias in the positioning of the posterior
mean combined with a posterior spread that is smaller even than in the
optimal case. In other words, the posterior is off mark, but believes
it is very right. The message is that (too) smooth priors
should be avoided; they lead to overconfident posteriors, which
reflect the prior information rather than the data, even if the amount
of information in the data increases indefinitely.
Under- and correct smoothing give very
conservative confidence regions (coverage equal to 1). However,
(i) and (ii) also
show that the credible ball has the same order of magnitude
as a correct confidence ball ($1\ge\wt r_{n,\g}/\allowbreak r_{n,\g}\gg0$), so
that the spread in the posterior does give the correct \textit{order} of
uncertainty. This at first sight surprising phenomenon is caused
by the fact that the posterior distribution concentrates near the boundary
of a~ball around its mean, and is not spread over the inside of the ball.
The coverage is 1, because this sphere is larger than
the corresponding sphere of the frequentist distribution of $AY$, even though
the two radii are of the same order.
By Theorem \ref{ContrPPP} the optimal contraction rate is obtained
(only) by a prior of the correct smoothness. Combining the two
theorems leads to the conclusion that priors that slightly undersmooth
the truth might be preferable. They attain a nearly optimal rate of contraction
and the spread of their posterior gives a reasonable sense of uncertainty.
Scaling\vspace*{1pt} of the prior modifies these conclusions. The optimal scaling
$\t_n\asymp n^{(\a-\b)/(1+2\a+2p)}$ found in Theorem \ref{ContrPPP},
possible if $\b<1+2\a+2p$, is covered in case (ii). This rescaling
leads to a balancing of square bias, variance and spread, and to
credible regions of the correct order of magnitude, although the
precise (uniform) coverage can be any number in $[0,1)$.
Alternatively, bigger rescaling rates are covered in case (i)
and lead to coverage 1. The optimal or slightly bigger rescaling
rate seems the most sensible. It would be interesting to extend
these results to data-dependent scaling.
\section{Recovering linear functionals of the parameter}
\label{SectionLinear}
We denote by $\Pi_n(\m\dvtx\break L\m\in\cdot\,\given Y)$ the posterior
distribution of the linear functional $L$, as described in
Proposition \ref{PosteriorLinear}. A continuous, linear functional
$L\dvtx
H_1\to\RR$ can be identified with an inner product $L\m=\langle\m,
l\rangle_1$, for some $l\in H_1$, and hence with a~sequence $(l_i)$ in
$\ell_2$ giving its coordinates in the eigenbasis $(e_i)$.
As shown in the proof of Proposition \ref{PosteriorLinear}, for $L$ in
the larger class of $N(0,\Lambda)$-measurable linear functionals, the
functional $L\Lambda^{1/2}$ is a continuous linear map on $H_1$
and hence can be identified with an element of
$H_1$. For such a functional $L$ we define a sequence $(l_i)$ by
$l_i=(L\Lambda^{1/2})_i/\sqrt{\l_i}$, for
$((L\Lambda^{1/2})_i)$ the coordinates of $L\Lambda^{1/2}$
in the eigenbasis. The assumption that~$L$ is a
$N(0,\Lambda)$-measurable linear functional implies that
$\sum_il_i^2\l_i<\infty$, but $(l_i)$ need not be contained
in $\ell_2$; if \mbox{$(l_i)\in\ell_2$}, then $L$ is
continuous and the definition of $(l_i)$ agrees with the definition in
the preceding paragraph.
We measure the smoothness of the functional $L$ by the size
of the coefficients~$l_i$, as $i\ra\infty$. First we assume that
the sequence is in $S^q$, for some~$q$.
\begin{theorem}[(Contraction)]
\label{LinContrPPP}
If $\m_{0}$, $(\l_i)$, $(\k_i)$ and $(\t_n)$ are
as in Assumption~\ref{PPP} and the representer $(l_i)$ of
the $N(0,\Lambda)$-measurable linear functional $L$ is contained in
$S^q$ for
$q\ge-\b$, then
$\E_{\m_0}\Pi_n(\m\dvtx |L\m- L\m_0| \ge\break M_n \eps_n \given
Y) \ra0$,
for every sequence $M_n\to\infty$, where
\[
\eps_n = (n\t_n^2)^{-({\b+q})/({1+2\a+2p})\wedge1}
+\t_n(n\t_n^2)^{-({1/2+\a+q})/({1+2\a+2p})\wedge(1/2)}.
\]
The rate is uniform over $\m_0$ in balls in $S^\b$. In particular:
\begin{longlist}[(iii)]
\item[(i)] If $\t_n\equiv1$, then
$\eps_n=n^{-(\b\wedge(1/2+\a)+q)/(1+2\a+2p)}\vee n^{-1/2}$.
\item[(ii)] If\vspace*{1pt} $q \leq p$ and $\b+q \leq1 + 2\a+ 2p$ and
$\t_n \asymp n^{({1/2+\a-\b})/({2\b+ 2p})}$, then
$\eps_n = n^{-(\b+q)/(2\b+2p)}$.
\item[(iii)] If $q\le p$ and $\b+ q > 1+2\a+2p$, then
$\eps_n \gg n^{-({\b+q})/({2\b+2p})}$ for every scaling $\t_n$.
\item[(iv)] If $q \geq p$ and
$\t_n\gtrsim n^{({1/2+\a-\wt\b+ p-q})/({2\wt\b+ 2q})}$,
where $\wt\b= \b\wedge(1+2\a+2p-q)$, then $\eps_n = n^{-1/2}$.
\end{longlist}
\end{theorem}
If $q\ge p$, then the smoothness of the functional $L$ cancels the
ill-posedness of the operator $K$, and estimating $L\m$ becomes a
``regular'' problem with an~$n^{-1/2}$ rate of convergence. Without
scaling the prior ($\tau_n\equiv1$), the posterior contracts at this
rate [see (i) or (iv)] if the prior is not too smooth
$(\a\le\b-1/2+q-p$). With scaling, the rate is also attained, with any
prior, provided the scaling parameter $\t_n$ does not tend to zero too
fast [see (iv)]. Inspection of the proof shows that too smooth priors
or too small scale creates a bias that slows the rate.
If $q<p$, where we take $q$ the ``biggest'' value such that $(l_i)\in
S^q$, estimating~$L\m$ is still an inverse problem. The minimax rate
over a ball in the Sobolev space $S^\b$ is known to be bounded above
by $n^{-(\b+q)/(2\b+2p)}$ (see
\cite{Donoho,DonohoLow,Goldenshluger}).
This rate is attained without scaling [see (i): $\tau_n\equiv1$] if
and only if the prior smoothness $\a$ is equal to the true smoothness
$\b$ minus $1/2$ ($\a+1/2=\b$). An intuitive explanation for this
apparent mismatch of prior and truth is that regularity of the
parameter in the Sobolev scale ($\m_0\in S^\b$) is not the appropriate
type of regularity for estimating a linear functional $L\m$. For
instance, the difficulty of estimating a function at a point is
determined by the regularity in a neighborhood of the point, whereas
the Sobolev scale measures global regularity over the domain. The fact
that a Sobolev space of order $\b$ embeds continuously in a H\"older
space of regularity $\b-1/2$ might give a quantitative explanation of
the ``loss'' in smoothness by $1/2$ in the special case that the
eigenbasis is the Fourier basis.
In our Bayesian context we draw the conclusion that the prior must be
adapted to the inference problem if we want to obtain the optimal
frequentist rate: for estimating the global parameter, a
good prior must match the truth ($\a=\b$), but for estimating a linear
functional a good prior must consider a Sobolev truth of order $\b$ as
having regularity $\a=\b-1/2$.
If the prior smoothness $\a$ is not $\b-1/2$, then the minimax rate
may still be attained by scaling the prior. As in the global problem,
this is possible only if the prior is not too rough [$\b+q\le1+2\a+2p$,
cases (ii) and (iii)]. The optimal scaling when this is possible [case
(ii)] is the same as the optimal scaling for the global problem
[Theorem \ref{ContrPPP}(ii)] \textit{after} decreasing $\b$ by $1/2$.
So the
``loss in regularity'' persists in the scaling rate. Heuristically
this seems to imply that a simple data-based procedure to set the
scaling, such as empirical or hierarchical Bayes, cannot attain
simultaneous optimality in both the global and local senses.
In the application of the preceding theorem, the functional $L$, and
hence the sequence $(l_i)$, will be given. Naturally, we apply the
theorem with $q$ equal to the largest value such that $(l_i)\in
S^q$. Unfortunately, this lacks precision for the sequences $(l_i)$
that decrease exactly at some polynomial order: a~sequence $l_i\asymp
i^{-q-1/2}$ is in $S^{q'}$ for every $q'<q$, but not in $S^q$. In the
following theorem we consider these sequences, and the slightly more
general ones such that $|l_i|\le i^{-q-1/2}\S(i)$, for some slowly
varying sequence $\S(i)$. Recall that $\S\dvtx [0,\infty)\to\RR$ is
\textit{slowly varying} if $\S(tx)/\S(t)\ra1$ as $t\ra\infty$, for
every $x>0$. [For these sequences $(l_i)\in S^{q'}$ for every
$q'<q$, $(l_i)\notin S^{q'}$ for $q'>q$,
and $(l_i)\in S^q$ if and only if $\sum_i\S^2(i)/i<\infty$.]
\begin{theorem}[(Contraction)]\label{LinContrPPPRV}
If $\m_{0}$, $(\l_i)$, $(\k_i)$ and $(\t_n)$ are
as in Assumption~\ref{PPP} and the representer $(l_i)$ of
the $N(0,\Lambda)$-measurable linear functional $L$
satisfies $|l_i|\le i^{-q-1/2}\S(i)$ for a slowly varying function $\S
$ and
$q>-\b$, then the result of Theorem \ref{LinContrPPP} is valid with
\begin{equation}
\label{EqTauLRV}
\eps_n = (n\t_n^2)^{-({\b+q})/({1+2\a+2p})\wedge1}\g_n
+\t_n(n\t_n^2)^{-({1/2+\a+q})/({1+2\a+2p})\wedge(1/2)}\d_n,\hspace*{-35pt}
\end{equation}
where, for $\r_n=(n\t_n^2)^{1/(1+2\a+2p)}$,
\begin{eqnarray*}
\g_n^2&=&\cases{ \S^2(\r_n), &\quad if $\b+q < 1+2\a+2p$,\vspace*{2pt}\cr
\displaystyle \sum_{i \leq\r_n}\frac{\S^2(i)}{i}, &\quad if $\b+q=1+2\a+2p$,\vspace*{2pt}\cr
1, &\quad if $\b+q>1+2\a+2p$,}\\
\d_n^2&=&\cases{\S^2(\r_n), &\quad if $q < p$,\vspace*{2pt}\cr
\displaystyle \sum_{i \leq\r_n}\frac{\S^2(i)}{i}, &\quad if $q=p$,\vspace*{2pt}\cr
1, &\quad if $q>p$.}
\end{eqnarray*}
This has the same consequences as in Theorem \ref{LinContrPPP},
up to the addition of slowly varying terms.
\end{theorem}
Because the posterior distribution for the linear
functional $L\m$ is the one-dimensional normal distribution $N(LAY,
s_n^2)$, the natural \textit{credible interval} for $L\m$ has endpoints
$LAY \pm z_{\g/2}s_n$, for $z_{\g}$ the (lower)
standard normal $\g$-quantile. The
\textit{coverage} of this interval is
\[
\Pr_{\m_0} (LAY+z_{\g/2}s_n\le L\m_0 \leq LAY -z_{\g
/2}s_n),
\]
where $Y$ follows (\ref{EqProblem}) with $\m=\m_0$.
To obtain precise results concerning coverage, we assume that $(l_i)$
behaves polynomially up to a slowly varying term, first in the
situation $q<p$ that estimating $L\m$ is an inverse problem. Let $\wt
\t_n$ be the (optimal) scaling
$\t_n$ that equates the two terms in the right-hand side of (\ref{EqTauLRV}).
This satisfies $\wt\t_n\asymp n^{({1/2+\a-\wt\b
})/({2\wt\b+ 2p})}\h_n$, for a slowly varying factor~$\h_n$, where
$\wt\b=\b\wedge(1+2\a+2p-q)$.
\begin{theorem}[(Credibility)]
\label{LinCrS}
Let $\m_{0}$, $(\l_i)$, $(\k_i)$ and $(\t_n)$ be as in
Assumption~\ref{PPP}, and let $|l_i|=i^{-q-1/2}\S(i)$
for $q < p$ and a slowly varying function~$\S$. Then the
asymptotic coverage of the interval $LAY \pm z_{\g/2}s_n$ is:
\begin{longlist}[(iii)]
\item[(i)] in $(1-\g,1)$, uniformly in $\m_0$ such that $\|\m_0\|
_\b\leq1$
if $\t_n \gg\wt\t_n$.
\item[(ii)] in $(1-\g,1)$, for every $\m_0\in S^\b$, if $\t
_n\asymp\wt\t_n$
and $\b+q < 1+2\a+2p$;
in $(0,c)$, along some $\m_0^{n}$ with $\sup_n\|\m_0^{n}\|
_\b< \infty$
if $\t_n\asymp\wt\t_n$ [any $c\in(0, 1)$].
\item[(iii)] $0$ along some $\m_0^n$ with $\sup_n\|\m_0^{n}\|_\b<
\infty$
if $\t_n \ll\wt\t_n$.
\end{longlist}
In case \textup{(iii)} the sequence $\m_0^n$ can be taken a fixed element
$\m_0$ in $S^\b$ if $\t_n\lesssim n^{-\d}\wt\t_n$ for some $\d>0$.
Furthermore, if $\t_n\equiv1$, then the coverage takes the form
as in \textup{(i), (ii)} and~\textup{(iii)} if $\a<\b-1/2$, $\a=\b-1/2$, and $\a>\b-1/2$,
respectively, where in case~\textup{(iii)} the sequence $\m_0^n$ can be taken
a fixed element.
\end{theorem}
Similarly, as in the nonparametric problem, oversmoothing leads to
coverage 0, while undersmoothing gives conservative intervals.
Without scaling the cut-off for under- or oversmoothing is at
$\a=\b-1/2$; with scaling the cut-off for the scaling rate is at the
optimal rate $\wt\t_n$.
The conservativeness in the case of undersmoothing is less
extreme for functionals than for the full parameter, as the coverage
is strictly between the credibility level $1-\g$ and 1.
The general message is the same: oversmoothing is disastrous for
the interpretation of credible band, whereas undersmoothing
gives bands that at least have the correct order of magnitude, in the
sense that their width is of the same order as the variance of the
posterior mean (see the proof). Too much undersmoothing is also
undesirable, as it leads to very wide confidence bands, and may cause
that $\sum_i l_i^2\lambda_i$ is no longer finite (see measurability property).
The results (i) and (ii) are the same for every $q < p$, even if $\tau
_n\equiv1$. Closer inspection would reveal that for a given $\mu_0$
the exact coverage depends on $q$ [and $\S(i)$] in a complicated way.
If $q\ge p$, then the smoothness of the functional $L$ compensates the
lack of smoothness of $K^{-1}$, and estimating $L\m$ is not a true
inverse problem. This drastically changes the performance of
credible intervals. Although oversmoothing again destroys their
coverage, credible intervals are exact confidence sets if the prior
is not too smooth. We formulate this in terms of a~Bernstein--von Mises
theorem.
The Bernstein--von Mises theorem for parametric models asserts that the
posterior distribution approaches a normal distribution centered at an
efficient estimator of the parameter and with variance equal to its
asymptotic variance. It is the ultimate link between Bayesian and
frequentist procedures. There is no version of this theorem
for infinite-dimensional parameters \cite{Freedman}, but the
theorem may hold for ``smooth'' finite-dimensional projections,
such as the linear functional $L\m$ (see~\cite{Castillo}).
In the present situation the posterior distribution of $L\m$ is
already normal by the normality of the model and the prior:
it is a $N(LAY,s_n^2)$-distribution by Proposition \ref{PosteriorLinear}.
To speak of a Bernstein--von Mises theorem, we also require the following:
\begin{longlist}[(iii)]
\item[(i)] That the (root of the) spread $s_n$ of the posterior distribution
is asymptotically equivalent to the standard deviation
$t_n$ of the centering variable $LAY$.
\item[(ii)] That the sequence $(LAY-L\m_0)/t_n$ tends in distribution to
a standard normal distribution.
\item[(iii)] That the centering $LAY$ is
an asymptotically efficient estimator of~$L\m$.
\end{longlist}
We shall show that (i) happens if and only if the
functional $L$ cancels the ill-posedness of the operator $K$,
that is, if $q\ge p$ in Theorem \ref{LinContrPPPRV}. Interestingly, the
rate of convergence $t_n$ must be $n^{-1/2}$ up to a slowly varying factor
in this case, but it could be strictly slower than $n^{-1/2}$ by a
slowly varying factor increasing to infinity.
Because $LAY$ is normally distributed by the normality of the model,
assertion (ii) is equivalent to saying that its bias tends to zero
faster than~$t_n$. This happens provided the prior does not oversmooth
the truth too much. For very smooth functionals ($q>p$) there is some
extra ``space'' in the cut-off for the smoothness, which (if the prior
is not scaled: $\t_n\equiv1$) is at $\a=\b-1/2+q-p$, rather than at
$\a=\b-1/2$ as for the (global) inverse estimating problem. Thus, the
prior may
be considerably smoother than the truth if the functional is very smooth.
Let \mbox{$\|\cdot\|$} denote the total variation norm between measures.
Say that $l\in\RV^q$ if $|l_i|=i^{-q-1/2}\S(i)$ for a slowly
varying function $\S$. Write
\[
B_n=\sup_{\|\m\|_\b\lesssim1}|LAK\m-L\m|
\]
for the maximal bias
of $LAY$ over a ball in the Sobolev space $S^\b$. Finally,
let $\wt\t_n$ be the (optimal) scaling $\t_n$
in that it equates the two terms in the right-hand side
of~(\ref{EqTauLRV}).
\begin{theorem}[(Bernstein--von Mises)]\label{TBvM-reg}
Let $\m_{0}$, $(\l_i)$, and $(\k_i)$ be as in Assumption \ref{PPP},
and let $l$ be the representer of the $N(0,\Lambda)$-measurable
linear functional~$L$:
\begin{longlist}[(iii)]
\item[(i)] If $l\in S^p$, then $s_n/t_n\ra1$; in this case
$nt_n^2\ra\sum_i l_i^2/\k_i^2$.
If $l\in\RV^q$, then $s_n/t_n\ra1$ if and only if $q\ge p$;
in this case $n\mapsto nt_n^2$ is slowly varying.
\item[(ii)] If $l\in S^q$ for $q\ge p$, then $B_n=o(t_n)$
if either $\t_n\gg n^{(\a+1/2-\b)/(2\b+2q)}$ or
($\t_n\equiv1$ and $\a<\b-1/2+q-p$).
If $l\in\RV^q$ for $q\ge p$, then $B_n=o(t_n)$ if
($\t_n \gg\wt\t_n$) or ($\t_n\equiv1$ and $\a<\b-1/2+q-p$)
or ($q=p$, $\t_n\equiv1$ and $\a=\b-1/2+q-p$)
or [$q> p$, $\t_n\equiv1$ and $\a=\b-1/2+q-p$ and $\S(i)\ra0$ as
$i\ra\infty$].
\item[(iii)] If $l\in S^p$ or $l\in\RV^p$ and $B_n=o(t_n)$,
then $\E_{\m_0}\|\Pi_n(L\m\in\cdot\,\given Y)-N(LAY$, $
t_n^2)\|\ra0$
and $(LAY-L\m_0)/t_n$ converges under $\m_0$
in distribution to a standard normal distribution, uniformly
in $\|\m_0\|_\b\lesssim1$. If $l\in S^p$, then this is also
true with $LAY$ and $t_n^2$ replaced by $\sum_iY_il_i/\k_i$ and
its variance $n^{-1}\sum_il_i^2/\k_i^2$.
\end{longlist}
In both cases \textup{(iii)},
the asymptotic coverage of the credible interval $LAY \pm z_{\g/2}s_n$
is $1-\g$, uniformly in $\|\m_0\|_\b\lesssim1$.
Finally, if the conditions under~\textup{(ii)} fail,
then there exists $\m_0^n$ with $\sup_n\|\m_0^n\|_\b<\infty$
along which the coverage tends to an arbitrarily low value.
\end{theorem}
The observation $Y$ in (\ref{EqProblem}) can be viewed as a reduction
by sufficiency of a random sample of size $n$ from the distribution
$N(K\m,I)$. Therefore, the model fits in the framework of
i.i.d. observations, and ``asymptotic efficiency'' can be defined in
the sense of semiparametric models discussed in, for example,
\cite{BKRW,vdV88} and \cite{vdVAS}. Because the model is shift-equivariant,
it suffices to consider local efficiency at $\m_0=0$. The
one-dimensional submodels $N(K(th),I)$ on the sample space
$\RR^{H_2}$, for $t\in\RR$ and a fixed ``direction'' $h\in H_1$, have
likelihood ratios
\[
\log\frac{dN(tKh,I)}{dN(0,I)}(Y)=tY_{Kh}-\frac12 t^2\|Kh\|_2^2.
\]
Thus, their \textit{score function} at $t=0$ is the $(Kh)$th coordinate
of a single observation $Y=(Y_h\dvtx h\in H_2)$, the \textit{score
operator} is the map $\tilde K\dvtx H_1\to L_2(N(0,I))$ given by
$\tilde Kh(Y)=Y_{Kh}$, and the \textit{tangent space} is the range of
$\tilde K$. [We denote the score operator by the same symbol~$K$ as in
(\ref{EqProblem}); if the observation $Y$ \textit{were} realizable
in~$H_2$, and not just in the bigger sample space~$\RR^{H_2}$, then
$Y_{Kh}$ would correspond to $\langle Y, Kh\rangle_2$ and, hence, the
score would be exactly $Kh$ for the operator in (\ref{EqProblem}) after
identifying $H_2$ and its dual space.] The adjoint of the score
operator restricted to the closure of the tangent space is the operator
$\tilde K^T\dvtx \overline{\tilde KH_1}\subset L_2 (N(0,I))\to H_1$
that satisfies $\tilde K^T(Y_g)=K^Tg$, where~$K^T$ on the right is the
adjoint of \mbox{$K\dvtx H_1\to H_2$}. The functional $L\m=\langle
l,\m\rangle_1$ has derivative $l$. Therefore, by \cite{vdV91}
asymptotically regular sequences of estimators exist, and the local
asymptotic minimax bound for estimating $L\m$ is finite, if and only if
$l$ is contained in the range of $K^T$. Furthermore, the variance bound
is $\|m\|_2^2$ for $m\in H_2$ such that $K^Tm=l$.
In our situation the range of $K^T$ is $S^p$, and if $l\in S^p$, then
by Theorem \ref{TBvM-reg}(iii) the variance of the posterior is
asymptotically equivalent to the variance bound and its centering can
be taken equal to the estimator $n^{-1}\sum_iY_il_i/\k_i$, which
attains this variance bound. Thus, the theorem gives a semiparametric
Bernstein--von Mises theorem, satisfying every of (i), (ii),~(iii) in
this case. If only $l\in\RV^p$ and not $l\in S^p$, the theorem still
gives a~Bernstein--von Mises type theorem, but the rate of convergence is
slower than $n^{-1/2}$, and the standard efficiency theory does not
apply.
\section{Example---Volterra operator}
\label{SectionVolterra}
The classical \textit{Volterra operator} $K\colon L^2[0$, $1] \to L^2[0,1]$ and
its adjoint $K^T$ are given by
\[
K\m(x) = \int_0^x \m(s) \,ds,\qquad
K^T\m(x) = \int_x^1 \m(s) \,ds.
\]
The resulting problem (\ref{EqProblem}) can also be written in
``signal in white noise'' form as follows: observe the process
$(Y_t\dvtx
t\in[0,1])$
given by $Y_t=\break\int_0^t\int_0^s \m(u)\, du \,ds+n^{-1/2}W_t$, for a Brownian
motion $W$.
The eigenvalues, eigenfunctions of $K^TK$ and conjugate basis
are given by (see~\cite{Halmos}), for $i=1,2,\ldots,$
\begin{eqnarray*}
\k_i ^2&=& \frac{1}{(i-1/2)^2\pi^2},\qquad
e_i(x) = \sqrt{2}\cos\bigl((i-1/2)\pi x\bigr),\\
f_i(x) &=& \sqrt{2}\sin\bigl((i-1/2)\pi x\bigr).
\end{eqnarray*}
The $(f_i)$ are the eigenfunctions of $KK^T$,
relative to the same eigenvalues, and
$Ke_i = \k_i f_i$ and $K^T f_i = \k_ie_i$, for every $i\in\NN$.
To illustrate our results with simulated data, we start by
choosing a true function~$\m_0$, which we expand as $\m_0 = \sum_i
\m_{0,i}e_i$ on the basis $(e_i)$. The data are the function
\[
Y = K\m_0 + \frac1{\sqrt n}Z = \sum_i \m_{0,i}\k_i f_i + \frac
1{\sqrt n}Z.
\]
It can be generated relative to the conjugate basis $(f_i)$ as a
sequence of independent Gaussian random variables $Y_1, Y_2, \ldots$
with $Y_i \sim N(\m_{0,i}\k_i, n^{-1/2})$. The posterior
distribution of $\m$ is Gaussian with mean $AY$ and covariance
operator $S_n$, given in Proposition \ref{Posterior}. Under
Assumption \ref{PPP} it can be represented in terms of
the coordinates $(\m_i)$ of $\m$ relative to the basis $(e_i)$
as (conditionally) independent Gaussian variables $\m_1, \m_2, \ldots
$ with
\[
\m_i \big|Y \sim N\biggl(\frac{n\l_i\k_iY_i}{1+n\l_i\k_i^2},\frac{\l
_i}{1+n\l_i\k_i^2}\biggr).
\]
The (marginal) posterior distribution for the function $\m$ at a point
$x$ is obtained by expanding $\m(x) = \sum_i \m_ie_i(x)$, and applying
the framework of linear functionals $L\m=\sum_il_i\m_i$
with $l_i =e_i(x)$. This
shows that
\[
\m(x)\big|Y \sim N\biggl(\sum_i \frac{n\l_i\k_iY_ie_i(x)}{1+n\l_i\k
_i^2},\sum_i \frac{\l_ie_i(x)^2}{1+n\l_i\k_i^2}\biggr).
\]
We obtained (marginal) posterior credible bands by
computing for every $x$ a central $95\%$ interval in the normal
distribution on the right-hand side.
Figure \ref{Figure1} illustrates these bands for
$n=1\mbox{,}000$. In every one of the 10 panels in the figure the black curve
represents the function $\m_0$, defined by the coefficients
$i^{-3/2}\sin(i)$ relative to $e_i$ ($\b=1$). The 10 panels
represent~10 independent realizations of the data, yielding 10
different realizations of the posterior mean (the red curves) and the
posterior credible bands (the green curves). In the left five
panels the prior is given by $\l_i=i^{-2\a-1}$ with $\a=1$, whereas in
the right panels the prior corresponds to $\a=5$. Each of the 10
panels also shows 20 realizations from the posterior
distribution.
\begin{figure}
\includegraphics{920f01.eps
\vspace*{-2pt}
\caption{Realizations of the posterior mean (red) and (marginal)
posterior credible bands (green), and 20 draws from
the posterior (dashed curves). In all ten panels $n=1\mbox{,}000$ and $\b=1$.
Left 5~panels: $\a=1$; right 5 panels: $\a=5$. True curve
(black) given by coefficients $\m_{0,i}=i^{-3/2}\sin(i)$.}
\label{Figure1}
\vspace*{-2pt}
\end{figure}
\begin{figure}
\includegraphics{920f02.eps}
\caption{Realizations of the posterior mean (red) and (marginal)
posterior credible bands (green), and 20 draws from
the posterior (dashed curves). In all\vspace*{1pt} ten panels $\b=1$.
Left~5 panels: $n=1\mbox{,}000$ and $\a=0.5,1,2,3,5$ (top to bottom);
right 5 panels: $n=10^8$ and $\a=0.5,1,2,3,5$ (top to bottom). True curve
(black) given by coefficients $\m_{0,i}=i^{-3/2}\sin(i)$.}
\label{Figure2}
\end{figure}
Clearly, the posterior mean is not estimating the true curve very well,
even for $n=1\mbox{,}000$. This is mostly caused by the intrinsic difficulty
of the inverse problem: better estimation requires bigger sample
size. A comparison of the left and right panels shows that
the rough prior ($\a=1$) is aware of the difficulty: it produces
credible bands that in (almost) all cases contain the true
curve. On the other hand, the smooth prior ($\a=5$)
is overconfident; the spread of the posterior distribution
poorly reflects the imprecision of estimation.
Specifying a prior that is too smooth relative to the true curve
yields a posterior distribution which gives both a bad reconstruction
and a misguided sense of uncertainty. Our theoretical results show
that the inaccurate quantification of estimation error remains even as
$n\ra\infty$.
The reconstruction, by the posterior mean or any other
posterior quantiles, will eventually converge to the true
curve. However, specification of a too smooth\vadjust{\goodbreak} prior will slow down
this convergence significantly. This is illustrated in
Figure \ref{Figure2}. Every one of its 10 panels is similarly
constructed as before, but now with $n=1\mbox{,}000$ and $n=10^8$ for the five
panels on the left-hand and right-hand side, respectively, and with
$\a=1/2,1,2,3,5$ for the five panels from top to bottom.
At first sight $\a= 1$ seems better (see the left column in Figure \ref{Figure2}),
but leads to zero coverage because of the mismatch close to the bump
(see the right column), while $\a= 1/2$ captures the bump.
For $n=10^8$ the posterior for this optimal prior has
collapsed onto the true curve, whereas the smooth posterior for $\a=5$
still has major difficulty in recovering the bump in the
true curve (even though it ``thinks'' it has captured the correct
curve, the bands having collapsed to a single curve in the\vadjust{\goodbreak}
figure).\looseness=-1
\section{Proofs}\vspace*{3pt}
\label{SectionProofs}
\subsection{\texorpdfstring{Proof of Theorem \protect\ref{ContrPPP}}{Proof of Theorem 4.1}}
The second moment of a Gaussian distribu\-tion on $H_1$ is equal to
the square norm of its mean plus the trace of its~co\-variance
operator.
Because the posterior is Gaussian $N(AY,S_n)$, it follows that
\[
\int\|\m-\m_0\|_1^2 \,d\Pi_n(\m\given Y)
=\|AY-\m_0\|_1^2+\tr(S_n).
\]
By Markov's inequality, the left-hand side is an upper bound to
$M_n^2{\varepsilon}_n^2\Pi_n(\m\dvtx\allowbreak \|\m-\m_0\|_1 \geq
M_n{\varepsilon}_n \given Y)$.
Therefore, it suffices to show that the expectation under~$\m_0$ of
the right-hand
side of the display is bounded by a multiple of~${\varepsilon}_n^2$.
The expectation of the
first term is the mean square error\vspace*{1pt} of the posterior mean~$AY$,
and can be written as the sum
$\|AK\m_0 - \m_0\|_1^2 +{n}^{-1}\tr(AA^T)$ of its square bias and
``variance.''
The second term $\tr(S_n)$ is deterministic. Under
Assumption \ref{common} the three quantities
can be expressed in the coefficients relative to the eigenbasis $(e_i)$
as
\begin{eqnarray}\label{SqBias}
\|AK\m_0 - \m_0\|_1^2
&=&\sum_i\frac{\m_{0,i}^2}{(1+n\l_i\k_i^2)^2}
\asymp\sum_i \frac{\mu_{0,i}^2}{(1+n\tau_n^2i^{-1-2\alpha
-2p})^2},\\
\label{Var}
\frac{1}{n}\tr(AA^T)
&=& \sum_i\frac{ n\l_i^2\k_i^2}{(1+n\l_i\k_i^2)^2}
\asymp\sum_{i} \frac{n\tau_n^4 i^{-2-4\alpha-2p}}{(1+n\tau
_n^2i^{-1-2\alpha-2p})^2},\\
\label{PostSpr}
\tr(S_n) &=& \sum_i\frac{\l_i}{1+n\l_i\k_i^2}
\asymp\sum_{i} \frac{\tau_n^2i^{-1-2\a}}{1+n\tau_n^2i^{-1-2\alpha-2p}}.
\end{eqnarray}
By Lemma\vspace*{1pt} \ref{NormSeries} (applied with $q=\b$, $t=0$, $u = 1+2\a+2p$,
$v=2$ and $N = n\t_n^2$), the first can be bounded by\vspace*{2pt}
$\|\m_0\|^2_\b(n\t_n^2)^{-({2\b})/({1+2\a+2p})\wedge2}$, which accounts
for the first term in the definition of ${\varepsilon}_n$.
By Lemma \ref{RVSeries} [applied with
$\S(i)=1$, $q=-1/2$, $t=2+4\a+2p$, $u=1+2\a+2p$, $v=2$, and $N = n\t_n^2$],
and again Lem\-ma~\ref{RVSeries} [applied with $\S(i)=1$, $q=-1/2$,
$t=1+2\a$, $u=1+2\a+2p$, $v=1$ and $N =n\t_n^2$],
both the second and third expressions are of the order the square of the
second term in the definition of ${\varepsilon}_n$.
The consequences (i) and (ii) follow by verification after
substitution of~$\t_n$ as given. To prove consequence (iii),
we note that the two terms in the definition of ${\varepsilon}_n$ are
decreasing and increasing in $\t_n$, respectively. Therefore,
the maximum of these two terms is minimized with respect to $\t_n$ by equating
the two terms.\vspace*{1pt} This minimum (assumed at
$\t_n =n^{-({1+\a+2p})/({3+4\a+6p})}$) is much bigger than $n^{-\b
/(1+2\b+2p)}$
if $\b>1+2\a+2p$.
\subsection{\texorpdfstring{Proof of Theorem \protect\ref{LinContrPPP}}{Proof of Theorem 5.1}}
By Proposition \ref{PosteriorLinear} the posterior distribution
is $N(LAY,s_n^2)$, and, hence, similarly as in the
proof of Theorem \ref{ContrPPP}, it suffices to show that
\[
\E_{\m_0} |LAY-L\m_0|^2 +s_n^2
=|LAK\m_0-L\m_0|^2+\frac1n\|LA\|_1^2+s_n^2
\]
is bounded above by a multiple of $\eps_n^2$. Under Assumption \ref{common}
the expressions on the right can be written
\begin{eqnarray}\label{LinBias}
LAK\m_0-L\m_0
&=& -\sum_{i} \frac{l_i\m_{0,i}}{1+n\l_i\k_i^2}
\lesssim\sum_i \frac{|l_i\m_{0,i}|}{1+n\t_n^2i^{-1-2\a-2p}},\\
\label{LinVar}
t_n^2:\!&=&\frac{1}{n}\|LA\|_1^2
= \sum_{i}
\frac{l_i^2n\l_i^2\k_i^2}{(1+n\l_i\k_i^2)^2}\nonumber\\[-9pt]\\[-9pt]
&\asymp& n\t_n^4 \sum_i \frac{l_i^2i^{-2-4\a-2p}}{(1+n\t
_n^2i^{-1-2\a-2p})^2},\nonumber\\[-2pt]
\label{LinPostSpr}
s_n^2 &=& \sum_{i} \frac{l_i^2\l_i}{1+n\l_i\k_i^2}
\asymp\t_n^2 \sum_{i} \frac{l_i^2 i^{-1-2\a}}{1+n\t_n^2i^{-1-2\a-2p}}.
\end{eqnarray}
By the Cauchy--Schwarz inequality the square of the bias
(\ref{LinBias}) satisfies
\begin{equation}
\label{EqSqBiasLinear}
|LAK\m_0-L\m_0|^2
\lesssim\|\m_0\|^2_\b\sum_i \frac{l_i^2i^{-2\b}}{(1+n\t
_n^2i^{-1-2\a-2p})^2}.
\end{equation}
By Lemma \ref{NormSeries}
(applied with $q=q, t=2\b, u=1+2\a+2p, v = 2$ and $N = n\t_n^2$)
the right-hand side of this display
can be further bounded by $\|\m_0\|_\b^2\|l\|_q^2$ times the square of
the first term in the sum of two terms that defines ${\varepsilon}_n$.
By Lemma \ref{NormSeries} (applied
with $q=q, t=2+4\a+2p, u = 1+2\a+2p, v=2$ and $ N = n\t_n^2$)
and again Lemma \ref{NormSeries} (applied with $q=q, t=1+2\a, u=1+2\a
+2p, v = 1$ and $N = n\t_n^2$), the right-hand sides of (\ref{LinVar})
and (\ref{LinPostSpr})
are bounded above by $\|l\|^2_q$ times the square of the second term
in the definition of ${\varepsilon}_n$.\looseness=-1
Consequences (i)--(iv) follow by substitution, and, in the
case of (iii), optimization over $\t_n$.\vspace*{-3pt}
\subsection{\texorpdfstring{Proof of Theorem \protect\ref{LinContrPPPRV}}{Proof of Theorem 5.2}}
This follows the same lines as the proof of Theorem \ref{LinContrPPP},
except that we use Lemma \ref{RVSeries}
(with $q\,{=}\,q, t\,{=}\,2\b$, \mbox{$u\,{=}\,1\,{+}\,2\a\,{+}\,2p$}, \mbox{$v\,{=}\,2$} and $N\,{=}\,n\t_n^2$)
and Lemma \ref{RVSeries}
(with $q\,{=}\,q, t\,{=}\,2\,{+}\,4\a\,{+}\,2p, u\,{=}\,1\,{+}\,2\a\,{+}\,2p$, \mbox{$v=2$} and $N = n\t_n^2$)
and again Lemma \ref{RVSeries}
(with $q=q, t=1+2\a, u=1+2\a+2p$, \mbox{$v = 1$} and $N = n\t_n^2$)
to bound the three terms
(\ref{LinVar})--(\ref{EqSqBiasLinear}).\vspace*{-3pt}
\subsection{\texorpdfstring{Proof of Theorem \protect\ref{CrNS}}{Proof of Theorem 4.2}}
Because the posterior distribution is $N(AY,\allowbreak S_n)$,
by Proposition \ref{Posterior}, the radius\vspace*{1pt} $r_{n,\g}$ in
(\ref{EqRadius}) satisfies
$\Pr(U_n< r_{n,\g}^2)=1-\g$,
for~$U_n$ a random variable distributed as the square norm
of an $N(0,S_n)$-variable. Under (\ref{EqProblem}) the variable
$AY$ is $N(AK\m_0,n^{-1}AA^T)$-distributed, and, thus, the coverage
(\ref{EqCoverage}) can be written as
\begin{equation}
\label{EqCoverageW}
\Pr(\|W_n+AK\m_0-\m_0\|_1\le r_{n,\g})
\end{equation}
for $W_n$ possessing\vspace*{1pt} a $N(0,n^{-1}AA^T)$-distribution.
For ease of notation let $V_n=\|W_n\|_1^2$.
The variables $U_n$ and $V_n$ can be represented as
$U_n=\sum_i s_{i,n}Z_i^2$ and $V_n=\sum_i t_{i,n}Z_i^2$,
for $Z_1,Z_2,\ldots$ independent standard normal variables,
and $s_{i,n}$ and $t_{i,n}$ the eigenvalues of $S_n$ and $n^{-1}AA^T$,
respectively, which satisfy
\begin{eqnarray*}
s_{i,n}&=&\frac{\l_i}{1+n\l_i\k_i^2}
\asymp\frac{\t_n^2i^{-2\a-1}}{1+n\t_n^2i^{-2\a-2p-1}},\\
t_{i,n}&=&\frac{ n\l_i^2\k_i^2}{(1+n\l_i\k_i^2)^2}
\asymp\frac{n\t_n^4i^{-4\a-2p-2}}{(1+n\t_n^2i^{-2\a-2p-1})^2},\\
s_{i,n}-t_{i,n}&=&\frac{ \l_i}{(1+n\l_i\k_i^2)^2}
\asymp\frac{\t_n^2i^{-2\a-1}}{(1+n\t_n^2i^{-2\a-2p-1})^2}.
\end{eqnarray*}
Therefore, by Lemma \ref{RVSeries} (applied with $\S\equiv1$ and $q=-1/2$;
always the first case),
\begin{eqnarray*}
\E U_n&=&\sum_i s_{i,n}\asymp\t_n^2(n\t_n^2)^{-{2\a}/({1+2\a
+2p})},\\
\E V_n&=&\sum_i t_{i,n}\asymp\t_n^2(n\t_n^2)^{-{2\a}/({1+2\a
+2p})},\\
\E(U_n-V_n)&=&\sum_i(s_{i,n}-t_{i,n})\asymp\t_n^2(n\t_n^2)^{-
{2\a}/({1+2\a+2p})},\\
\var U_n&=&2\sum_i s_{i,n}^2\asymp\t_n^4(n\t_n^2)^{-({1+4\a
})/({1+2\a+2p})},\\
\var V_n&=&2\sum_i t_{i,n}^2\asymp\t_n^4(n\t_n^2)^{-({1+4\a
})/({1+2\a+2p})}.
\end{eqnarray*}
We conclude that the standard deviations of $U_n$ and $V_n$ are
negligible relative to their means, and also relative to the
difference $\E(U_n-V_n)$ of their means. Because $U_n\ge V_n$, we
conclude that the distributions of $U_n$ and $V_n$ are asymptotically
completely separated: $\Pr(V_n\le v_n\le U_n)\to1$ for some $v_n$
[e.g., $v_n=\E(U_n+V_n)/2$]. The numbers $r_{n,\g}^2$ are
$1-\g$-quantiles of $U_n$, and, hence, \mbox{$\Pr(V_n\le r^2_{n,\g
}(1+o(1)))\ra
1$}. Furthermore, it follows that
\[
r_{n,\g}^2 \asymp\t_n^2(n\t_n^2)^{-{2\a}/({1+2\a+2p})}\asymp
\E U_n\asymp\E V_n.
\]
The square norm of the bias $AK\m_0-\m_0$ is given in
(\ref{SqBias}), where it was noted that
\[
B_n:={\sup_{\|\m_0\|_\b\lesssim1}}\|AK\m_0-\m_0\|_1 \asymp
(n\t_n^2)^{-\b/({1+2\a+2p})\wedge1}.
\]
The bias $B_n$ is decreasing
in $\t_n$, whereas $\E U_n$ and $\var U_n$ are increasing. The scaling
rate $\tilde\t_n\asymp n^{(\a- \wt\b)/(1+2\wt\b+2p)}$ balances the
square bias $B_n^2$ with the variance $\E V_n$ of the posterior mean,
and hence with $r_{n,\g}^2$.
Case (i). In this case $B_n \ll r_{n,\g}$. Hence,
$\Pr(\|W_n+AK\m_0-\m_0\|_1\le r_{n,\g})
\ge\Pr(\|W_n\|_1\le r_{n,\g}-B_n)
=\Pr(V_n\le r_{n,\g}^2(1+o(1)))\ra1$,
uniformly in the set of~$\m_0$ in the supremum defining $B_n$.
Case (iii). In this case $B_n \gg r_{n,\g}$. Hence,
$\Pr(\|W_n+AK\m_0^n-\m_0^n\|_1\le r_{n,\g}) \le
\Pr(\|W_n\|_1\ge B_n-r_{n,\g})\ra0$ for any sequence
$\m_0^n$ (nearly) attaining the supremum in the definition of
$B_n$. If $\t_n\equiv1$, then $B_n$ and $r_{n,\g}$ are both
powers of $1/n$ and, hence, $B_n \gg r_{n,\g}$ implies that
$B_n\gtrsim r_{n,\g} n^\d$, for some $\d>0$. The preceding argument
then applies for a fixed $\m_0$ of the form $\m_{0,i}\asymp
i^{-1/2-\b-{\varepsilon}}$, for small ${\varepsilon}>0$, that gives
a bias that is much
closer than $n^\d$ to $B_n$.
Case (ii). In this case $B_n\asymp r_{n,\g}$. If $\b<1+2\a+2p$, then by
the second assertion (first case) of Lemma \ref{NormSeries} the bias
$\|AK\m_0-\m_0\|_1$ at a fixed $\m_0$ is of strictly smaller order
than the supremum $B_n$. The argument of (i) shows that the asymptotic
coverage then tends to 1.
Finally, we prove the existence of a sequence $\m_0^n$ along which
the coverage is a given $c\in[0,1)$. The coverage (\ref{EqCoverageW})
with $\m_0$ replaced by $\m_0^n$ tends to $c$ if,
for $b_n=AK\m_0^n-\m_0^n$ and $z_c$ a standard normal quantile,
\begin{eqnarray}
\label{EqNormality}
\frac{\|W_n+b_n\|_1^2-\E\|W_n+b_n\|_1^2}{{\sdev}\|W_n+b_n\|_1^2}
&\weak& N(0,1),\\
\label{EqConvergenceRng}
\frac{r_{n,\g}^2-\E\|W_n+b_n\|_1^2}{{\sdev}\|W_n+b_n\|_1^2}
&\ra& z_c.
\end{eqnarray}
Because $W_n$ is mean-zero Gaussian, we have
$\E\|W_n+b_n\|_1^2=\E\|W_n\|_1^2+\|b_n\|_1^2$ and
${\var}\|W_n+b_n\|_1^2={\var}\|W_n\|_1^2+4\var\langle W_n,b_n\rangle_1$.
Here $\|W_n\|_1^2=V_n$ and the distribution of
$\langle W_n,b_n\rangle_1$ is zero-mean Gaussian with variance
$\langle b_n, n^{-1}AA^Tb_n\rangle_1$. With
$t_{i,n}$ the eigenvalues of $n^{-1}AA^T$, display (\ref
{EqConvergenceRng}) can be
translated in the coefficients $(b_{n,i})$ of $b_n$ relative
to the eigenbasis, as
\begin{equation}
\label{Hulp}
\frac{r_{n,\g}^2-\E V_n-\sum_i b_{n,i}^2}{\sqrt{\var V_n+4\sum
_it_{i,n}b_{n,i}^2}}
\ra z_c.
\end{equation}
We choose $(b_{n,i})$ differently in the cases that $\b\le1+2\a+2p$
and $\b\ge1+2\a+2p$, respectively. In both cases the sequence
has exactly one nonzero coordinate. We denote this coordinate by $b_{n,i_n}$,
and set, for numbers $d_n$ to be determined,
\[
b_{n,i_n}^2=r_{n,\g}^2-\E V_n-d_n\sdev V_n.
\]
Because $r_{n,\g}^2$, $\E V_n$ and $r_{n,\g}^2-\E V_n$ are of the
same order
of magnitude, and~$\sdev V_n$ is of strictly smaller
order, for bounded or slowly diverging $d_n$ the right-hand side of the
preceding display is equivalent to $(r_{n,\g}^2-\E V_n)(1+o(1))$.
Consequently, the left-hand side of (\ref{Hulp}) is equivalent to
\[
\frac{d_n \sdev V_n}{\sqrt{\var V_n+4t_{i_n,n}(r_{n,\g}^2-\E V_n)(1+o(1))}}.
\]
The remainder of the argument is different in the two cases.
Case $\b\le1+2\a+2p$. We choose $i_n\asymp(n\t_n^2)^{1/(1+2\a+2p)}$.
It can be verified that $t_{i_n,n}(r_{n,\g}^2-\E V_n)/\var V_n\asymp1$.
Therefore, for $c\in[0,1]$, there exists a bounded or slowly
diverging sequence $d_n$
such that the preceding display tends to $z_c$.\vadjust{\goodbreak}
The bias $b_n$ results from a parameter $\m_0^n$ such that
$b_{n,i}=(1+n \l_i\k_i^2)^{-1}(\m_0^n)_i$, for every $i$. Thus,
$\m_0^n$ also has exactly one nonzero coordinate, and this is
proportional to the corresponding coordinate of $b_n$, by the definition
of $i_n$. It follows that
\[
i_n^{2\b}(\m_0^n)^2_{i_n}\asymp i_n^{2\b} b_{n,i_n}^2
\lesssim i_n^{2\b}(r_{n,\g}^2-\E V_n)\asymp1
\]
by the definition of $\t_n$. It follows that $\|\m_0^n\|_\b\lesssim1$.
Case $\b\ge1+2\a+2p$. We choose $i_n=1$. In this case
$\t_n\ra0$ and it can be verified that
$t_{i_n,n}(r_{n,\g}^2-\E V_n)/\var V_n\ra0$. Also,
\[
(\m_0^n)_1^2\asymp(1+n\t_n^2)^2b_{n,1}^2\lesssim(1+n\t_n^2)^2\E V_n.
\]
This is $O(1)$, because $\t_n$ is chosen so that $\E V_n$
is of the same order as the square bias $B_n^2$, which is $(n\t_n^2)^{-2}$
in this case.
It remains to prove the asymptotic normality (\ref{EqNormality}).
We can write
\[
\|W_n+b_n\|_1^2-\E\|W_n+b_n\|_1^2
= \sum_i t_{i,n}(Z_i^2-1)+2b_{n,i_n}\sqrt{t_{{i_n},n}}Z_{i_n}.
\]
The second term is normal by construction. The first
term has variance $2\sum_i t_{i,n}^2$. With some effort it can be seen that
\[
\sup_i \frac{t_{i,n}^2}{\sum_i t_{i,n}^2}\ra0.
\]
Therefore, by a slight adaptation of the Lindeberg--Feller theorem
(to infinite sums), we have that $\sum_i t_{i,n}(Z_i^2-1)$ divided
by its standard deviation tends in distribution to the standard
normal distribution. Furthermore, the preceding display shows
that this conclusion does not change if the $i_n$th term is left out
from the infinite sum. Thus, the two terms converge jointly
to asymptotically independent standard normal variables, if scaled
separately by their standard deviations.
Then their scaled sum is also asymptotically
standard normally distributed.
\subsection{\texorpdfstring{Proof of Theorem \protect\ref{LinCrS}}{Proof of Theorem 5.3}}
Under (\ref{EqProblem}) the variable $LAY$ is
$N(LAK\m_0,\allowbreak t_n^2)$-distributed, for $t_n^2$ given in (\ref{LinVar}).
It follows that the coverage can be written, with $W$
a standard normal variable,
\begin{equation}
\label{EqCoverageLinW}
\Pr( |Wt_n+LAK\m_0-L\m_0|\le-s_n z_{\g/2}).
\end{equation}
The bias $LAK\m_0-L\m_0$ and posterior spread $s_n^2$ are
expressed as a series in~(\ref{LinBias}) and (\ref{LinPostSpr}).
In the proof of Theorem \ref{LinContrPPPRV}
$s_n$ and $t_n$ were seen to have the same order of magnitude, given by
the second term in ${\varepsilon}_n$ given in (\ref{EqTauLRV}),
with a slowly varying term $\d_n$ as given in the theorem,
\begin{equation}
\label{EqSnTnLin}
s_n\asymp t_n\asymp\t_n (n\t_n^2)^{-(1/2+\a+q)/(1+2\a+2p)}\d_n.
\end{equation}
Furthermore,\vspace*{1pt} $t_n \leq s_n$ for every $n$, as every term in the
infinite series (\ref{LinVar}) is $n\l_i\k_i^2/(1+n\l_i\k_i^2)\le1$
times the corresponding term in (\ref{LinPostSpr}).
Because $W$ is centered, the coverage (\ref{EqCoverageLinW}) is
largest if the bias $LAK\m_0-L\m_0$ is zero. It is then at least
$1-\g$, because $t_n\le s_n$; remains strictly smaller than $1$,
because $t_n
\asymp s_n$; and tends to exactly $1-\g$ iff $s_n/t_n \ra1$. By
Theorem~\ref{TBvM-reg}(i) the latter is impossible if $q < p$.
The analysis for nonzero $\m_0$ depends strongly on the size
of the bias relative to $t_n$.
The supremum of the bias satisfies, for $\g_n$ the slowly varying term
given in Theorem \ref{LinContrPPPRV},
\begin{equation}
\label{EqSupBiasLinear}
B_n:={\sup_{\|\m_0\|_\b\lesssim1}}|LAK\m_0-L\m_0|
\asymp(n\t_n^2)^{-((\b+q)/(1+2\a+2p))\wedge1}\g_n.
\end{equation}
That the left-hand side of (\ref{EqSupBiasLinear})
is smaller than the right-hand side was already shown
in the proof of Theorem \ref{LinContrPPPRV}, with
the help of Lemma \ref{RVSeries}. That this upper bound is
sharp follows by considering the sequence $\m_0^n$ defined by,
with $\tilde B_n$ the right-hand side of the preceding display,
\[
\m_{0,i}^{n} = \frac1{\tilde B_n} \frac{i^{-2\b}l_i}{1+n\t
_n^2i^{-1-2\a-2p}}.
\]
[This is the sequence that gives equality in the application
of the Cauchy--Schwarz inequality to derive (\ref{EqSqBiasLinear}).]
Using Lemma \ref{RVSeries}, it can be seen that $\|\m_0^n\|_\b
\lesssim1$
and that the bias at $\m_0^n$ is of the order $\tilde B_n$.
By Lemma \ref{LemmaTechnicalBias},
the bias at a \textit{fixed} $\m_0\in S^\b$ is of strictly smaller order
than the supremum $B_n$ if $\b+q<1+2\a+2p$.
The maximal bias $B_n$ is a decreasing function of the scaling
parameter~$\t_n$, while the standard deviation $t_n$ and root-spread
$s_n$ increase with $\t_n$.
The scaling rate $\wt\t_n$ in the statement of the
theorem balances $B_n$ with $s_n\asymp t_n$.
Case (i). If $\t_n\gg\wt\t_n$, then $B_n\ll t_n$. Hence, the bias
$LAK\m_0-L\m_0$ in (\ref{EqCoverageLinW}) is negligible relative
to $t_n\asymp s_n$, uniformly in $\|\m_0\|_\b\lesssim1$, and
the coverage is asymptotic to $\Pr( |Wt_n|\le-s_nz_{\g/2})$,
which is asymptotically strictly between $1-\g$ and~$1$.
Case (iii). If $\t_n\ll\wt\t_n$, then $B_n\gg t_n$.
If $b_n=LAK\m_0^n-L\m_0^n$ is
the bias at a~sequence $\m_0^n$ that (nearly) attains the supremum
in the definition of~$B_n$, then the coverage at $\m_0^n$ satisfies
$\Pr( |Wt_n+b_n|\le-s_n z_{\g/2})
\le\Pr( |Wt_n|\ge b_n-s_n |z_{\g/2}|)\ra0$,
as $b_n\asymp B_n\gg s_n$. By the same argument,
the coverage also tends to zero for a fixed $\m_0$
in $S^\b$ with bias $b_n=LAK\m_0-L\m_0\gg t_n$.
For this we choose $\m_{0,i}=i^{-\b}i^ql_i\tilde S(i)$ for a slowly
varying function such that $\sum_i \S^2(i)\tilde S^2(i)/i<\infty$.
The latter condition ensures that $\|\m_0\|_\b<\infty$.
By another application of Lemma \ref{RVSeries}, the bias
at $\m_0$ is of the order [cf. (\ref{LinBias})]
\[
\sum_i\frac{l_i\m_{0,i}}{1+n\t_n^2i^{-1-2\a-2p}}
=\sum_i \frac{(l_i\tilde S^{1/2}(i))^2i^{-\b+q}}{1+n\t_n^2i^{-1-2\a-2p}}
\asymp(n\t_n^2)^{-({\b+q})/({1+2\a+2p})\wedge1}\tilde\g_n,
\]
where, for $\r_n=(n\t_n^2)^{1/(1+2\a+2p)}$,
\[
\tilde\g_n^2=\cases{
\S^2(\r_n)\tilde S(\r_n), &\quad if $\b+q < 1+2\a+2p$,\vspace*{2pt}\cr
\displaystyle \sum_{i \leq\r_n}\frac{\S^2(i)\tilde S(i)}{i}, &\quad if
$\b+q=1+2\a+2p$,\vspace*{2pt}\cr
1, &\quad if $\b+q>1+2\a+2p$.}
\]
Therefore, the bias at $\m_0$ has the same form as the maximal
bias $B_n$; the difference is in the slowly varying factor $\tilde\g_n$.
If $\t_n\le\tilde\t_nn^{-\d}$, then $B_n\gtrsim t_nn^{\d'}$ for
some $\d'>0$
and, hence, $b_n\asymp B_n\tilde\g_n/\g_n\gg t_n$.
Case (ii). If $\t_n\asymp\wt\t_n$, then $B_n\asymp t_n$. If
$b_n=LAK\m_0^n-L\m_0^n$ is again the bias at a sequence $\m_0^n$ that
nearly assumes the supremum in the definition of $B_n$, we have that
$\Pr( |Wt_n+d b_n|\le-s_n z_{\g/2}) \le\Pr(
|Wt_n|\ge
db_n-s_n |z_{\g/2}|)$ attains an arbitrarily small value if $d$
is chosen sufficiently large. This is the coverage
at the sequence~$d\m_0^n$, which is bounded in $S^\b$.
On the other hand, the bias at a fixed
$\m_0\in S^\b$ is of strictly smaller order than the supremum~$B_n$,
and, hence, the coverage at a fixed $\m_0$ is as in case (i).
If the scaling rate is fixed to $\t_n\equiv1$, then it can be checked
from (\ref{EqSnTnLin}) and~(\ref{EqSupBiasLinear}) that $B_n\ll t_n$,
$B_n\asymp t_n$ and $B_n\gg t_n$ in the three cases $\a<\b-1/2$,
$\a=\b-1/2$ and $\a>\b-1/2$, respectively. In the first and third
cases the maximal bias and the spread differ by more than a polynomial term
$n^\d$; in the second case it must be noted that the slowly varying
terms $\g_n$ and $\d_n$ are equal [to $\S(\r_n)$]. It follows that the
preceding analysis (i), (ii), (iii) extends to this situation.
\subsection{\texorpdfstring{Proof of Theorem \protect\ref{TBvM-reg}}{Proof of Theorem 5.4}}
(i). The two quantities $s_n$ and $t_n$ are given as series in
(\ref{LinPostSpr}) and (\ref{LinVar}). Every term in the series
(\ref{LinVar}) is $n\l_i\k_i^2/(1+n\l_i\k_i^2)\le1$ times the
corresponding term in the series (\ref{LinPostSpr}). Therefore,
$s_n/t_n\ra1$ if and only if the series are determined by the terms for
which these numbers are ``close to'' 1, that is, $n\l_i\k_i^2$ is
large. More precisely, we show below that $s_n/t_n\ra1$ if and only
if, for every $c > 0$,
\begin{equation}
\label{Hulp1}
\sum_{n\l_i\k_i^2 \leq c} \frac{l_i^2\l_i}{1+n\l_i\k_i^2}
= o\biggl(\sum_i \frac{l_i^2\l_i}{1+n\l_i\k_i^2}\biggr).
\end{equation}
If $l\in S^p$, then the series on the left is as in
Lemma \ref{NormSeries}
with $q=p$, $u=1+2\a+2p$, $v=1$, $N=n\t_n^2$ and $t=1+2\a$. Hence,
$(t+2q)/u\ge v$, and the display follows from the final assertion
of the lemma. If $l_i=i^{-q-1/2}\S(i)$ for a slowly varying function
$\S$,
then the series is as in Lemma \ref{RVSeries}, with the
same parameters, and by the last statement
of the lemma the display is true if and only if
$(t+2q)/u\ge v$, that is, $q\ge p$.
To prove that (\ref{Hulp1}) holds iff $s_n/t_n\ra1$,
write $s_n^2=A_n+B_n$, for $A_n$ and $B_n$ the
sums over the terms in (\ref{LinPostSpr})
with $n\l_i\k_i^2> c$ and $n\l_i\k_i^2\le c$,
respectively, and, similarly, $t_n^2=C_n+D_n$. Then
\[
\frac{D_n}{B_n}\le\frac c{1+c}\le\frac{C_n}{A_n}\le1.
\]
It follows that
\[
\frac{t_n^2}{s_n^2}
=\frac{C_n+D_n}{A_n+B_n}
=\frac{C_n/A_n+(D_n/B_n)(B_n/A_n)}{1+B_n/A_n}
\le\frac{1+c/(1+c)(B_n/A_n)}{1+B_n/A_n}.
\]
Because $x\mapsto(1+rx)/(1+x)$ is strictly decreasing from 1 at $x=0$
to $r<1$ at $x=\infty$ (if $0<r<1$), the right-hand side
of the equation is asymptotically~1 if and only if $B_n/A_n\to0$,
and otherwise its liminf is strictly smaller.
Thus, $t_n/s_n\to1$ implies that $B_n/A_n\to0$.
Second,
\[
\frac{t_n^2}{s_n^2}
\ge\frac{C_n}{A_n+B_n}
=\frac{C_n/A_n}{1+B_n/A_n}\ge\frac{c/(1+c)}{1+B_n/A_n}.
\]
It follows that $\liminf t_n^2/s_n^2\ge c/(1+c)$ if $B_n/A_n\to0$.
This being true for every $c>0$ implies that $t_n/s_n\to1$.
\hphantom{i}(i) Second assertion.\vspace*{1pt} If $l\in S^p$, then we apply
Lemma \ref{NormSeries} with $q=p$, $t=1+2\a$, $u=1+2\a+2p$, $v=1$ and
$N=n\t_n^2$ to see that $s_n^2\asymp\t_n^2(n\t_n^2)^{-v}=n^{-1}$.
Furthermore, the second assertion of the lemma with $(uv-t)/2=p$ shows
that $ns_n^2\ra\|l\|_{p}^2=\sum_il_i^2/\k_i^2$ in the case that
$\k_i=i^{-p}$. The proof can be extended to cover the slightly more
general sequence $(\k_i)$ in Assumption~\ref{PPP}.
If $l\in\RV^q$, then we apply Lemma \ref{RVSeries} with $q=p$,
$t=1+2\a$, $u=1+2\a+2p$, $v=1$ and $N=n\t_n^2$ to see that
$s_n^2\asymp n^{-1}\sum_{i\le N^{1/u}}\S^2(i)/i$.\vspace*{1pt}
(ii) If $l\in S^q$, then the bias is bounded above in
(\ref{EqSqBiasLinear}), and in the proof of Theorem \ref{LinContrPPP} its
supremum $B_n$ over $\|\m_0\|_\b\lesssim1$ is seen to be bounded by
$(n\t_n^2)^{-(\b+q)/(1+2\a+2p)\wedge1}$, the first term in the
definition of ${\varepsilon}_n$ in the statement of this theorem. This
upper bound
is $o(n^{-1/2})$ iff the stated conditions hold.
[Here we use that $\S^2(N)\ll\sum_{i\le N}\S^2(i)/i$
as $N\ra\infty$, as noted in the proof of Lem\-ma~\ref{RVSeries}.]
The supremum of the bias $B_n$ in the case that $l\in\RV^q$ is given in
(\ref{EqSupBiasLinear}). It was already seen to be $o(t_n)$ if
$\t\gg\wt\t_n$ in the proof of case\vspace*{1pt} (i) of Theorem~\ref{LinCrS}.
If $\t_n=1$, we have that $B_n\asymp n^{-(\b+q)/(1+2\a+2p)\wedge
1}\g_n$,
for $\g_n$ the slowly varying factor
given in the statement of Theorem \ref{LinContrPPPRV}. Furthermore,
we have $s_n\asymp t_n\asymp n^{-1/2}\d_n$, for $\d_n$ the slowly varying
factor in the same statement. Under the present conditions,
\mbox{$\d_n\asymp1$} if $q>p$ and $\d_n^2\asymp\sum_{i\le\r_n}\S^2(i)/i$
if $q=p$. We can now verify that $B_n=o(t_n)$ if and only if the
conditions as stated hold.
(iii) The total variation distance between two Gaussian distributions
with the same expectation and standard deviations $s_n$ and $t_n$
tends to zero if and only if $s_n/t_n\ra1$. Similarly, the total
distance between two Gaussians with the same standard deviation $s_n$
and means $\m_n$ and $\n_n$ tends to zero if and only if
$\m_n-\n_n=o(s_n)$. Therefore, it suffices to show that $(LAY-\sum_i
Y_il_i/\k_i)/s_n\ra0$ if $l\in S^p$. Because the bias was already
seen to be $o(t_n)$ and $s_n\asymp n^{-1/2}$ if $l\in S^p$, it
suffices to show that $LAZ-\sum_i Z_il_i/\k_i\ra0$.
Under Assumption \ref{PPP} this difference is equal to
\[
\sum_i \frac{\k_i\l_il_iZ_i}{n^{-1}+\k_i^2\l_i}-\sum_iZ_i\frac
{l_i}{\k_i}
=\sum_i \frac{Z_il_i}{\k_i}\biggl(\frac1{1+n\k_i^2\l_i}\biggr).
\]
If $\sum_il_i^2/\k_i^2<\infty$, then the variance of this
expression is seen to tend to zero by dominated convergence.
The final assertion of the theorem follows along the
lines of the proof of Theorem \ref{LinCrS}.
\section{Technical lemmas}
\label{SectionTechnical}
\begin{lemma}\label{NormSeries}
For any $q \geq0$, $t \geq-2q$, $u >0$ and $ v\ge0$,
as $N \to\infty$,
\[
\sup_{\|\xi\|_q\le1}\sum_i \frac{\xi_i^2 i^{-t}}{(1+N i^{-u})^v}
\asymp N^{-((t+2q)/u)\wedge v}.
\]
Moreover, for every fixed $\xi\in S^q$, as $N \to\infty$,
\[
N^{((t+2q)/u)\wedge v}\sum_i \frac{\xi_i^2 i^{-t}}{(1+N i^{-u})^v}
\ra
\cases{
0, &\quad if $(t+2q)/u < v$,\cr
\|\xi\|^2_{(uv-t)/{2}}, &\quad if $(t+2q)/u \ge v$.}
\]
The last assertion remains true if the sum is limited
to the terms $i\le c N^{1/u}$, for any $c>0$.
\end{lemma}
\begin{pf}
In the range $i\le N^{1/u}$ we have $Ni^{-u}\le1+N i^{-u}\le2Ni^{-u}$,
while $1\le1+N i^{-u}\le2$ in the range $i>N^{1/u}$.
Thus, deleting either the first or second term, we obtain
\begin{eqnarray*}
\sum_{i \leq N^{1/u}} \frac{\xi_i^2 i^{-t}}{(1+N i^{-u})^v}
&\asymp&\sum_{i \leq N^{1/u}} \xi_i^2 i^{2q} \frac{i^{uv-t-2q}}{N^v}
\leq\|\xi\|_q^2 N^{-((t+2q)/u)\wedge v},\\
\sum_{i > N^{1/u}} \frac{\xi_i^2 i^{-t}}{(1+N i^{-u})^v}
&\asymp&\sum_{i > N^{1/u}} \xi_i^2i^{2q} i^{-t-2q}
\leq N^{-(t+2q)/u} \sum_{i > N^{1/u}}\xi_i^2i^{2q}.
\end{eqnarray*}
The inequality in the first line follows by bounding $i$ in $i^{uv-t-2q}$
by $N^{1/u}$ if $uv-t-2q>0$, and by 1 otherwise.
This proves the upper bound for the supremum.
The lower bound follows by considering the two sequences $(\xi_i)$
given by $\xi_i=i^{-q}$ for $i\sim N^{1/u}$ and $\xi_i=0$ otherwise
(showing that the supremum is bigger than $N^{-(t+2q)/u}$), and given
by $\xi_1=1$ and $\xi_i=0$ otherwise (showing that the supremum is
bigger than $N^{-v}$).
The second line of the preceding display shows that the sum over the
terms $i>N^{1/u}$ is $o(N^{-(t+2q)/u})$. Furthermore, the first line
can be multiplied by $N^{(t+2q)/u}$ to obtain
\[
N^{(t+2q)/u}\sum_{i \leq N^{1/u}} \frac{\xi_i^2 i^{-t}}{(1+N i^{-u})^v}
\asymp\sum_{i \leq N^{1/u}}\xi_i^2 i^{2q} \biggl(\frac
{i}{N^{1/u}}\biggr)^{uv-t-2q}.
\]
If $(t+2q)/u<v$, then $uv-t-2q>0$ and this tends to
zero by dominated convergence. Also,
\[
N^v\sum_i \frac{\xi_i^2 i^{-t}}{(1+N i^{-u})^v}
=\sum_i \xi_i^2i^{uv-t} \biggl(\frac{ N i^{-u}}{1+N i^{-u}}\biggr)^v.
\]
If $(t+2q)/u\ge v$, then $q\ge(uv-t)/2$ and, hence, $\xi\in
S^{{(uv-t)}/{2}}$, and
the right-hand side tends to $\sum_i \xi_i^2i^{uv-t}$ by dominated
convergence.
The final assertion needs to be proved only in the case that
$(t+2q)/u\ge v$,
as in the other case the whole sum tends to 0.
The sum over the terms $i>N^{1/u}$ was seen to be always
$o(N^{-(t+2q)/u})$, which is $o(N^{-v})$ if $(t+2q)/u\ge v$. The final
assertion for $c=1$ follows, because the sum over the terms $i\le N^{1/u}$
was seen to have the exact order $N^{-v}$ (if $\xi\not=0$). For general
$c$ the proof is analogous, or follows by scaling $N$.
\end{pf}
\begin{lemma}\label{RVSeries}
For any $t,v\ge0$, $u>0$, and $(\xi_i)$ such that
$|\xi_i|=i^{-q-1/2}\S(i)$ for $q > -t/2$ and a slowly varying function
$\S\dvtx (0,\infty)\to(0,\infty)$, as $N \ra\infty$,
\[
\sum_i \frac{\xi_i^2i^{-t}}{(1+N i^{-u})^v}
\asymp\cases{ N^{-(t+2q)/u}\S^2(N^{1/u}), &\quad if $(t+2q)/u<v$,\vspace*{2pt}\cr
\displaystyle N^{-v}\sum_{i \leq N^{1/u}}\S^2(i)\big/i, &\quad if $(t+2q)/u=v$,\vspace*{2pt}\cr
N^{-v}, &\quad if $(t+2q)/u>v$.}
\]
Moreover, for every $c>0$, the sum on the left
is asymptotically equivalent to the same sum restricted to the
terms $i\le cN^{1/u}$ if and only if $(t+2q)/u\ge v$.
\end{lemma}
\begin{pf}
As in the proof of the preceding lemma, we split the
infinite series in the sum over the terms $i\le N^{1/u}$ and $i>N^{1/u}$.
For the first part of the series
\[
\sum_{i \leq N^{1/u}} \frac{\xi_i^2 i^{-t}}{(1+N i^{-u})^v}
\asymp\sum_{i \leq N^{1/u}} \S(i)^2 \frac{i^{uv-t-2q-1}}{N^v}.
\]
If $uv-t-2q>0$ [i.e., $(t+2q)/u<v$], the right-hand side
is of the order $N^{-(t+2q)/u}\S^2(N^{1/u})$,
by Theorem 1(b) on page 281 in \cite{Feller},
while if $uv-t-2q<0$, it is of the order $N^{-v}$
by Lemma on page 280 in \cite{Feller}.
Finally, if $uv-t-2q=0$, then the right-hand side is identical to
$N^{-v}\sum_{i \leq N^{1/u}} {\S^2(i)}/{i}$.
The other part of the infinite series satisfies,
by Theorem 1(a) on page~281 in \cite{Feller},
\[
\sum_{i > N^{1/u}} \frac{\xi_i^2 i^{-t}}{(1+N i^{-u})^v}
\asymp\sum_{i > N^{1/u}} \S(i)^2 i^{-t-2q-1} \asymp N^{-(t+2q)/u}\S
^2(N^{1/u}).
\]
This is never bigger than the contribution of the first part of the sum,
and of equal order if $(t+2q)/u< v$.
If $(t+2q)/u>v$, then the leading polynomial term
is strictly smaller than $N^{-v}$. If $(t+2q)/u=v$,
then the leading term is equal to $N^{-v}$, but the slowly\vspace*{1pt}
varying part satisfies $\S^2(N^{1/u}) \ll\sum_{i \leq N^{1/u}}{\S
^2(i)}/{i}$,
by Theorem 1(b) on page 281 in \cite{Feller}. Therefore, in both
cases the preceding display is negligible relative to the
first part of the sum.
This proves the final assertion of
the lemma for $c=1$. The proof for general $c>0$ is analogous.
\end{pf}
By the Cauchy--Schwarz inequality, for any $\m\in S^{t/2}$,
\[
\biggl|\sum_i\frac{\xi_i\mu_i}{1+Ni^{-u}}\biggr|^2
\le\|\m\|_{t/2}^2\sum_i \frac{\xi_i^2i^{-t}}{(1+N i^{-u})^2}.
\]
The preceding lemma gives the exact order of the right-hand side.
The application of the Cauchy--Schwarz inequality is sharp, in that
there is equality for some \mbox{$\mu\in S^{t/2}$}. However, this $\mu$ depends
on $N$. For fixed $\mu\in S^{t/2}$ the left-hand side is strictly smaller
than the right-hand side.
\begin{lemma}
\label{LemmaTechnicalBias}
For any $t,u\ge0$, $\mu\in S^{t/2}$ and $(\xi_i)$ such that
$|\xi_i|=\break i^{-q-1/2}\S(i)$ for $0<t+2q <2u$ and a slowly varying function
$\S\dvtx (0,\infty)\to(0,\infty)$, as $N \ra\infty$,
\[
\sum_i \frac{|\xi_i\mu_i|}{1+N i^{-u}}
\ll N^{-(t+2q)/(2u)}\S(N^{1/u}).
\]
\end{lemma}
\begin{pf}
We split the series in two parts, and bound the denominator
$1+Ni^{-u}$ by $Ni^{-u}$ or $1$. By the Cauchy--Schwarz inequality,
for any $r>0$,
\begin{eqnarray*}
\biggl|\sum_{i\le N^{1/u}} \frac{|\xi_i\mu_i|}{N i^{-u}}\biggr|^2
&\le&\frac1{N^2}\sum_{i\le N^{1/u}} \frac{\S^2(i)i^r}{i}
\sum_{i\le N^{1/u}}\m_i^2i^{2u-2q-r}\\
&\asymp&\frac1{N^2}\S^2(N^{1/u})N^{r/u}\\
&&{}\times\sum_{i\le N^{1/u}}{\m_i^2i^t}\biggl(\frac{i}{N^{1/u}}
\biggr)^{2u-2q-r-t} N^{(2u-2q-r-t)/u},\\
\biggl|\sum_{i> N^{1/u}} \frac{|\xi_i\mu_i|}{1}\biggr|^2
&\le&\sum_{i> N^{1/u}} \frac{\S^2(i)}{i}i^{-2q}\sum_{i> N^{1/u}}\m_i^2
\asymp\S^2(N^{1/u})N^{-2q/u}\sum_{i> N^{1/u}}\m_i^2.
\end{eqnarray*}
The terms in the remaining series in the right-hand side of the first
inequality are bounded by $\m_i^2i^t$ and tend to zero pointwise
as $N\ra\infty$ if $2u-2q-r-t>0$.
If $t+2q<2u$, then there exists $r>0$ such that the latter
is true, and for this $r$ the sum tends to zero by the dominated
convergence theorem. The other terms collect to
$N^{-(t+2q)/(u)}\S^2(N^{1/u})$. The sum in the right-hand side of the
second inequality is bounded by
$\sum_{i>N^{1/u}}\m_i^2i^t N^{-t/u}=o(N^{-t/u})$.
\end{pf}
|
\section{Introduction}
Let $M$ be a projective algebraic manifold in $\mathbb CP^N$,
$N\geq\dim_{\mathbb C}M=n$. The hyperplane line bundle of $\mathbb
CP^N$ restricts to an ample line bundle $L$ on $M$, which is called
a polarization on $M$. A K\"ahler metric $g$ is called a polarized
metric, if the corresponding K\"ahler form represents the first
Chern class $c_1(L)$ of $L$ in $H^2(M,\mathbb Z)$. Given any
polarized K\"ahler metric $g$, there is a Hermitian metric $h_L$ on
$L$ whose Ricci form is equal to
$$\omega_g=\frac{\sqrt{-1}}{2\pi}\sum_{i,j=1}^n g_{i\overline{j}}dz_i\wedge dz_{\overline{j}}.$$
Let $\mathcal E$ be a holomorphic vector bundle on $M$ of rank $r$
with a Hermitian metric $h_{\mathcal E}$. Then for any holomorphic
sections $U_1, U_2\in H^0(M,\mathcal E(m))$ ($\mathcal
E(m):=\mathcal E \otimes L^m$), we have the pointwise metric
$\langle U_1(x), U_2(x)\rangle_{h_{\mathcal E}\otimes h_L^m}$ and
the $L^2$-metric
\begin{equation}
(U_1,U_2)=\int_M\langle U_1(x),U_2(x)\rangle_{h_{\mathcal E}\otimes
h_L^m}\frac{\omega_g^n}{n!}.
\end{equation}
Let $S_1,\dots,S_d$ be an orthonormal basis of $H^0(M,\mathcal
E\otimes L^m)$ in the $L^2$-metric. The {\it Bergman kernel} is
defined to be the following:
\begin{equation}
B_m(x):=\sum_{j=1}^d \langle\cdot, S_j(x)\rangle S_j(x) \in {\rm
End}(\mathcal E\otimes L^m).
\end{equation}
Note that $B_m(x)$ is independent of the choice of orthonormal
basis. In particular, when $\mathcal E=\mathbb C$,
\begin{equation}
B_m(x)=\sum_{j=1}^{d}\|S_{j}(z)\|^{2}_{h_{L}^m}.
\end{equation}
The following asymptotic expansion was first proved by Zelditch
\cite{Zel} and Catlin \cite{Cat}, motivated by the convergence of
Bergman metrics started in the paper of Tian \cite{Tia} (cf. also
\cite {Bou, Rua}) following a suggestion of Yau \cite{Yau}.
\begin{theorem}{\rm\bf (Zelditch, Catlin)} \label{tyz} With the notation above, there is an asymptotic
expansion when $m\rightarrow\infty$:
\begin{equation} \label{eqb10}
B_m(x)\sim a_{0}(x)m^{n}+a_{1}(x)m^{n-1}+a_{2}(x)m^{n-2}+\cdots
\end{equation} where $a_{k}(x)\in {\rm End}(\mathcal
E\otimes L^m)$. More precisely,
$$||B_m(x)-\sum_{j=0}^k a_j(x)m^{n-j}||_{C^\mu}\leq C_{k,\mu}m^{n-k-1},$$
where $C_{k,\mu}$ depends on $k,\mu$ and the manifold $M$.
\end{theorem}
In the case of $\mathcal E=\mathbb C$, Lu \cite{Lu} computed
$a_{k}(x)$ for $k\leq3$ by using the peak section method \cite{Tia}
and proved that each $a_{k}(x)$ is a polynomial of the curvature and
its covariant derivatives. The above theorem and Lu's computation of
$a_1(x)$ have played a crucial role in Donaldson's breakthrough work
\cite{Don}. The theorem was generalized to symplectic manifolds for
Boutet de Monvel-Guillemin's almost holomorphic sections by Shiffman
and Zelditch \cite{SZ}, and Borthwick and Uribe \cite{BU}. Dai, Liu
and Ma \cite{DLM} established the full off-diagonal expansion of the
Bergman kernel on orbifolds and symplectic manifolds for the
Spin$^c$ Dirac operator by using the heat kernel method. See also
\cite{Son, RT}.
There are alternative derivations of these $a_k$ by Engli\v{s}
\cite{Eng} and Loi \cite{Loi} using the Laplace integral, by Douglas
and Klevtsov using path integral \cite{DK} when $\mathcal E=\mathbb
C$. For general $\mathcal E$, X. Wang \cite{Wan} computed $a_1$ for
the first time; L. Wang \cite{WanL} computed $a_1,a_2$; Ma and
Marinescu \cite{MM3, MM, MM4} computed $a_1,a_2$ in the symplectic
case and presented a recursive method to compute coefficients of the
expansion of more general Bergman kernels as well as the kernel of
Toeplitz operators; Berman, Berndtsson and Sj\"ostrand \cite{BBS}
gave a proof of Theorem \ref{tyz} by microlocal analysis and showed
a recursive algorithm for computing higher order terms; Liu and Lu
\cite{LL} gave a proof of Theorem \ref{tyz} using the complex
geometric method and uncovered new geometric information about the
expansion. The excellent monograph by Ma and Marinescu \cite{MM2}
contains a comprehensive introduction to the asymptotic expansion of
the Bergman kernel and its applications.
In other contexts, the asymptotic expansion of the Bergman kernel
also plays an important role in the Berezin quantization on K\"ahler
manifolds \cite{Ber}. We will discuss the Berezin transform briefly
in Section \ref{berezin}. It was applied to define the Berezin
$\star$-product on K\"ahler manifolds. Reshetikhin and Takhtajan
\cite{RTa} obtained a formula of the Berezin $\star$-product in
terms of partition functions of Feynman graphs. More explicit
graph-theoretic formulae of the Berezin and Berezin-Toeplitz $\star$-products were obtained
in \cite{Xu, Xu2}. We remark that Kontsevich's celebrated formula
\cite{Kon} for a $\star$-product on Poisson manifolds was also
written as a summation over graphs.
In order to state the main results in this paper, we assume
$\mathcal E=\mathbb C$ and introduce some terminologies in graph
theory.
A {\it digraph} (directed graph) $G=(V,E)$ is defined to be a finite
set $V$ (whose elements are called vertices) and a multiset $E$ of
ordered pairs of vertices, called directed edges. Throughout the
paper, we allow a digraph to have loops and multi-edges. The
adjacency matrix $A=A(G)$ of a digraph $G$ with $n$ vertices is a
square matrix of order $n$ whose entry $A_{ij}$ is the number of
directed edges from vertex $i$ to vertex $j$. The {\it outdegree}
$\deg^+(v)$ and {\it indgree} $\deg^-(v)$ of a vertex $v$ are
defined to be the number of outward and inward edges at $v$
respectively.
A digraph $G$ is called {\it connected} if the underlying undirected
graph is connected, and {\it strongly connected} if there is a
directed path from each vertex in $G$ to every other vertex.
We call a directed graph $G=(V,E)$ {\it stable} if at each vertex
$v$ both the outdegree and indegree are no less than $2$. The set
of stable graphs will be denoted by $\mathcal G$. The {\it weight}
of $G$ is defined to be $|E|-|V|$. We will define a natural function
$z(G)$ on $\mathcal G$ such that the coefficients $a_k$ can be
written as a sum over $\mathcal G(k)$, the set of stable graphs of
weight $k$ (see Section \ref{secgraph} for details)
\begin{equation} \label{eqc}
a_k(x)=\sum_{G\in\mathcal G(k)} z(G)\cdot G, \quad z(G)\in\mathbb Q.
\end{equation}
So we may regard $z$ as a map from the set $\mathcal G$ of all
stable graphs to $\mathbb Q$, from which we can easily recover the
curvature-tensor expression of these $a_k$'s (see Example
\ref{tyza2} where $a_1$ and $a_2$ are computed, and Appendix
\ref{apthree} where $a_3$ is computed).
The main result of this paper is the following theorem.
\begin{theorem} \label{main} Let $G=(V,E)\in\mathcal G$ be a stable graph with the adjacency matrix $A$.
\begin{enumerate}
\item[i)] If $G$ is a disjoint
union of connected subgraphs $G=G_1\cup\dots\cup G_n$, then we have
\begin{equation}
z(G)=\prod_{j=1}^n z(G_j)/|Sym(G_1,\dots,G_n)|,
\end{equation}
where $Sym(G_1,\dots,G_n)$ denote the permutation group of these $n$
connected subgraphs.
\item[ii)] If $G$ is connected but not strongly connected, then
\begin{equation}
z(G) = 0.
\end{equation}
\item[iii)] If $G$ is strongly
connected, then
\begin{equation}
z(G) = -\frac{\det(A-I)}{|{\rm Aut}(G)|},
\end{equation}
where $I$ is the identity matrix.
\end{enumerate}
\end{theorem}
We computed $z(G)$ for $G\in\mathcal G(k),\, k\leq4$ using Loi's
recursive formula \eqref{eqloi}. They match with values computed by
the above theorem (see the appendix).
As pointed out by the referee, it would be interesting to see the
implication of the above theorem on the relation between the Bergman
kernel on K\"ahler manifolds and the path integral (cf. \cite{DK}).
We hope our work will find application in Fefferman's program of
studying the Bergman kernel of a strong pseudoconvex domain as an
analogy of the heat kernel of Riemannian manifolds (cf. \cite{Ale,
Eng, Fef, Hir, Xu}).
We also obtained some interesting combinatorial identities, for
example we proved the following formula for Bernoulli numbers $B_k$
in corollary \ref{bern}.
\begin{equation} \label{eqbern}
B_k=(-1)^{k} k\sum_{G\in\mathcal G(k)}\frac{\det(A-I)}{|{\rm
Aut}(G)|}\cdot\epsilon(G)\prod_{v\in V}(\deg^+(v)-1)! ,\quad k\geq1,
\end{equation}
where $\epsilon(G)$ is the number of Euler tours in $G$. Note that
in the right-hand side, $\epsilon(G)=0$ unless $G$ is connected and
balanced, and hence strongly connected. A digraph is called {\it balanced} if
$\deg^+(v)=\deg^-(v)$ for each vertex $v$.
\
\noindent{\bf Acknowledgements} The author is grateful to Professor
Kefeng Liu for his kind help over many years. The author benefited a
lot from Professor Shing-Tung Yau's seminars. The author thanks
Professor Xiaonan Ma for helpful comments. The author thanks the
referee for very helpful suggestions which greatly improved the
presentation of the paper.
\vskip 30pt
\section{Tensor calculus on K\"ahler manifolds} \label{sectensor}
Let $(M,g)$ be a K\"ahler manifold of dimension
$n$. Locally the K\"ahler form is given by
$$\omega_g=\frac{\sqrt{-1}}{2\pi}\sum_{i,j=1}^n
g_{i\overline{j}}dz_i\wedge dz_{\overline{j}}.$$
We will use the Einstein summation convention. The indices
$i,j,k,\dots$ run from $1$ to $n$, while Greek indices
$\alpha,\beta,\gamma$ may represent either $i$ or $\bar i$. Let
$\det g$ be the determinant of the Hermitian matrix $(g_{i\bar j})$
and $(g^{i\bar j})$ be the inverse of the matrix $(g_{i\bar j})$. We
also use the notation
\begin{equation}
g_{j\bar
k\alpha_1\alpha_2\dots\alpha_m}:=\partial_{\alpha_1\alpha_2\dots\alpha_m}g_{j\bar
k}.
\end{equation}
The curvature tensor is defined as
\begin{equation}\label{eqcur1}
R_{i\bar jk\bar l} =-g_{i\bar j k\bar l}+g^{m\bar p}g_{m\bar j\bar
l}g_{i\bar p k}.
\end{equation}
The Ricci tensor is
\begin{equation}\label{eqcur2}
R_{i\bar j}=g^{k\bar l}R_{i\bar jk\bar l}=-\partial_i\partial_{\bar
j}\log(\det g)
\end{equation}
and the scalar curvature is the trace of the Ricci curvature
\begin{equation}\label{eqcur3}
\rho=g^{i\bar j}R_{i\bar j}.
\end{equation}
The covariant derivative of a tensor field
$T_{\beta_1\dots\beta_q}^{\alpha_1\dots\alpha_p}$ is defined by
\begin{equation}\label{eqcur4}
T_{\beta_1\dots\beta_q;\gamma}^{\alpha_1\dots\alpha_p}=\partial_{\gamma}T_{\beta_1\dots\beta_q}^{\alpha_1\dots\alpha_p}-\sum_{i=1}^q
\Gamma_{\gamma\beta_i}^{\delta}T_{\beta_1\dots\beta_{i-1}\delta\beta_{i+1}\dots\beta_q}^{\alpha_1\dots\alpha_p}
+\sum_{j=1}^p\Gamma_{\delta\gamma}^{\alpha_j}T_{\beta_1\dots\beta_q}^{\alpha_1\dots\alpha_{j-1}\delta\alpha_{j+1}\dots\alpha_p},
\end{equation}
where the Christoffel symbols $\Gamma_{\beta\gamma}^\alpha=0$ except
for
\begin{equation}
\Gamma_{jk}^i=g^{i\bar l}g_{j\bar l k},\quad \Gamma_{\bar j\bar
k}^{\bar i}=g^{l\bar i}g_{l\bar j\bar k}.
\end{equation}
\begin{lemma} For tensors in K\"ahler geometry, we have the following identities:
\begin{enumerate}
\item[i)] The K\"ahler metric $g$ satisify
\begin{equation} \label{eqcur6}
\partial_i g_{j\bar k}=\partial_j g_{i\bar k},\quad \partial_{\bar
l}g_{j\bar k}=\partial_{\bar k} g_{j\bar l}.
\end{equation}
\item[ii)] The derivative of $g^{m\bar l}$ satisfy
\begin{equation}\label{eqcur7}
\partial_{\alpha}g^{m\bar l}=-g^{p\bar l}g^{m\bar q}g_{p\bar
q\alpha}.
\end{equation}
\item[iii)] The derivative of $\det g$ satisfy
\begin{equation}\label{eqdet}
\partial_{\alpha}\det g=\det g\cdot g^{m\bar l}g_{m\bar
l\alpha}.
\end{equation}
\item[iv)] The curvature tensor satisfy the first Bianchi identity
\begin{equation}
R_{i\bar jk\bar l}=R_{i\bar l k\bar j}=R_{k\bar j i\bar l}.
\end{equation}
\item[v)] The covariant derivatives of the curvature tensor
satisfy the second Bianchi identity
\begin{equation}
R_{i\bar j k\bar l;m}=R_{m\bar j k\bar l;i}=R_{i\bar j m\bar
l;k},\quad R_{i\bar j k\bar l;\bar p}=R_{i\bar p k\bar l;\bar
j}=R_{i\bar j k\bar p;\bar l}.
\end{equation}
\item[vi)] The Ricci formula gives the difference when we
interchange two covariant derivative indices
\begin{equation}\label{eqcur5}
T_{\beta_1\dots\beta_q;i\bar
j}^{\alpha_1\dots\alpha_p}-T_{\beta_1\dots\beta_q;\bar j
i}^{\alpha_1\dots\alpha_p}=\sum_{k=1}^q R_{\beta_k i\bar
j}^{\gamma}T_{\beta_1\dots\beta_{k-1}\gamma\beta_{k+1}\dots\beta_q}^{\alpha_1\dots\alpha_p}
-\sum_{l=1}^p R_{\gamma i\bar
j}^{\alpha_l}T_{\beta_1\dots\beta_q}^{\alpha_1\dots\alpha_{l-1}\gamma\alpha_{l+1}\dots\alpha_p},
\end{equation}
where $R_{\bar l i\bar j}^{\bar k}= -g^{m\bar k}R_{m\bar l i\bar j}$,
$R_{l i\bar j}^{k}= g^{k\bar m}R_{l\bar m i\bar j}$ and $R_{\bar l
i\bar j}^{k}= R_{l i\bar j}^{\bar k}=0$.
\end{enumerate}
\end{lemma}
Recall that around each point $x$ on a K\"ahler manifold, there
exists a normal coordinate such that
\begin{equation} \label{eqnormal}
g_{i\bar j}(x)=\delta_{ij}, \qquad g_{i\bar j k_1\dots
k_r}(x)=g_{i\bar j \bar l_1\dots\bar l_r}(x)=0
\end{equation}
for all $r\leq N\in \mathbb N$, where $N$ can be chosen arbitrary
large.
\begin{remark} \label{rm1}
Fix a normal coordinate around $x$ on a K\"ahler manifold. By
\eqref{eqcur1} and \eqref{eqcur4}, it is not difficult to see
inductively that any covariant derivative $R_{i\bar jk \bar l;
\alpha_1\dots \alpha_r}$ is canonically equal to a polynomial of
$g_{a\bar b \alpha}$ and their partial derivatives. By
\eqref{eqnormal}, any monomial containing $g_{i\bar j k_1\dots k_r}$
or $g_{i\bar j \bar l_1\dots\bar l_r}$ vanishes at $x$. For example
$$R_{i\bar jk \bar l}(x)=-g_{i\bar jk \bar l}(x),$$
$$R_{i\bar jk \bar l; p\bar q}(x)=-g_{i\bar jk \bar l p\bar q}(x)
+g^{s\bar t}(g_{p\bar j s\bar l}g_{i\bar q k\bar t}+g_{k\bar j s\bar
l}g_{p\bar q i\bar t}+g_{i\bar j s\bar l}g_{k\bar q p\bar t})(x),$$
$$R_{i\bar jk \bar l; p_1\dots p_r}(x)=-g_{i\bar jk \bar l
p_1\dots p_r}(x),\quad \forall\, r\geq 1,$$
$$R_{i\bar jk \bar l; \bar p_1\dots \bar
p_r}(x)=-g_{i\bar jk \bar l \bar p_1\dots \bar p_r}(x),\quad
\forall\, r\geq 1.$$
Note that these identities hold only at the point $x$. However they
can easily recover the original identities (hold in a neighborhood
of $x$), since from \eqref{eqcur1} and \eqref{eqcur4}, we see that
$g_{i\bar j k\bar l \beta_1\dots\beta_{r}}$ can be inductively
expressed as a canonical polynomial of covariant derivatives of
curvature tensors, denoted by $D(g_{i\bar j k\bar l
\beta_1\dots\beta_{r}})$. For example
$$D(g_{i\bar j k\bar l})=-R_{i\bar j k\bar l},$$
$$D(g_{i\bar j k\bar l p\bar q})=-R_{i\bar jk \bar l; p\bar q}+g^{s\bar t}(R_{p\bar j s\bar l}R_{i\bar q k\bar t}+R_{k\bar j s\bar
l}R_{p\bar q i\bar t}+R_{i\bar j s\bar l}R_{k\bar q p\bar t}).$$ In
general, we can inductively get
\begin{equation}
D(g_{i\bar j k\bar l \beta_1\dots\beta_{r}})=-R_{i\bar j k\bar l;
\beta_1\dots\beta_{r}}+\text{ covariant derivatives of lower order}.
\end{equation}
In Section \ref{secgraph}, this observation will allow us to store
Weyl invariants as polynomials of $g_{a\bar b c \bar d}$ and their
derivatives. The advantage is that we do not need to deal with the
problem of exchanging indices as in the Ricci formula
\eqref{eqcur5}.
\end{remark}
We will now define the concept of an admissible tree, which is a
directed tree with half-edges. We call the two trees in Figure
\ref{fig1} {\it primitive admissible trees}.
\begin{figure}[h]
$\xymatrix@C=7mm@R=5mm{\ar[dr]^{\bar k} && &\\ &\circ \ar[r]^{e} & v
\ar[ur]^j \ar[dr] &
\\\ar[ur] &&&}$
\qquad\quad $\xymatrix@C=7mm@R=5mm{ \ar[dr] & &\\ &\circ \ar[ur]
\ar[dr]&
\\ \ar[ur] & &}$
\caption{The atomic admissible tree and a trivial admissible tree}
\label{fig1}
\end{figure}
We can define three types of actions by an outward half-edge $i$ or
an inward half-edge $\bar i$ on a directed tree with half-edges. Let
us take the left-hand primitive admissible tree as an example:
\begin{enumerate}
\item The action on a vertex $v$ is to add the half edge to the
vertex (see Figure \ref{fig5}).
\begin{figure}[h]
\begin{tabular}{c}$\xymatrix@C=7mm@R=5mm{\ar[dr]^{\bar k} && &\\ &\circ
\ar[r]^{e} & v \ar[r] \ar[ur]^j \ar[dr] & i
\\\ar[ur] &&&}$ \end{tabular} \quad\text{ or
}\quad \begin{tabular}{c}$\xymatrix@C=7mm@R=5mm{\ar[dr]^{\bar k} && &\\
&\circ \ar[r]^{e} & v \ar[ur]^j \ar[dr] & {\bar i} \ar[l]
\\\ar[ur] &&&}$ \end{tabular}
\caption{Action on a vertex $v$} \label{fig5}
\end{figure}
\item The action on a half-edge with the same direction (i.e. $i$ can only act on outward half-edges and $\bar i$ only act on inward half-edges) is to
put them together on a new vertex (see Figure \ref{fig4}).
\begin{figure}[h]
\begin{tabular}{c}$\xymatrix@C=7mm@R=5mm{ \ar[r]^{\bar k} &\circ \ar[dr] & & &\\
\ar[ur]_{\bar i} & &\circ \ar[r]^e & v \ar[ur]^j \ar[dr]&
\\ & \ar[ur] & & &}$ \end{tabular}
\quad\text{ or }\quad
\begin{tabular}{c}$\xymatrix@C=7mm@R=5mm{ \ar[dr]^{\bar k} & & &\circ \ar[r]^j \ar[dr]_i &\\ &\circ \ar[r]^e & v \ar[ur] \ar[dr] & &
\\ \ar[ur] & & & &}$ \end{tabular}
\caption{Action on a half-edge} \label{fig4}
\end{figure}
\item The action on an edge $e$ is to generate a new vertex
to split $e$ and add the half-edge to the new vertex (see Figure
\ref{fig3}).
\begin{figure}[h]
\begin{tabular}{c}$\xymatrix@C=7mm@R=5mm{ \ar[dr]^{\bar k} & & & &\\ & \circ \ar[r] &\circ \ar[r] \ar[u]_i& v \ar[ur]^j \ar[dr]&
\\ \ar[ur]& & & &}$ \end{tabular}\quad\text{ or
}\quad
\begin{tabular}{c}$\xymatrix@C=7mm@R=5mm{ \ar[dr]^{\bar k} & & & &\\ & \circ \ar[r] &\circ \ar[r]& v \ar[ur]^j \ar[dr]&
\\ \ar[ur]& &\ar[u]_{\bar i} & &}$ \end{tabular}
\caption{Action on an edge $e$} \label{fig3}
\end{figure}
\end{enumerate}
For any two trees $T_1$ and $T_2$, we can join them by gluing an
outward half-edge of $T_1$ to an inward half-edge of $T_2$, or vice
versa.
\begin{definition}
An {\it admissible tree} (with half-edges) is defined to be the join
of finite number of trees which can be obtained by a finite number
of the above three actions on the primitive admissible trees in
Figure \ref{fig1}. For an admissible tree, we usually label its
outward half-edges and inward half-edges with distinct indices
$\{i,j,k,\dots\}$ and $\{\bar l,\bar m,\bar p,\dots\}$ respectively.
\end{definition}
An admissible tree is called {\it decomposable} if we can cut an
edge to get two admissible trees (see Figure \ref{fig2}). Otherwise,
it is called {\it indecomposable.}
\begin{figure}[h]
\begin{tabular}{c}$\xymatrix@C=7mm@R=5mm{& \ar[dr] & & \ar[dr] & &\\ & & \circ \ar[rr]^e \ar[dr]& & \circ \ar[ur] \ar[dr]&
\\ & \ar[ur]& & & &}$ \end{tabular} $\Longrightarrow$ \quad
\begin{tabular}{c}$\xymatrix@C=7mm@R=5mm{ \ar[dr] & &\\ &\circ \ar[r] \ar[dr]&
\\ \ar[ur] & &} \quad \xymatrix@C=7mm@R=5mm{ \ar[dr] & &\\ \ar[r] &\circ \ar[ur] \ar[dr]&
\\ & &}$ \end{tabular}
\caption{Decomposition by cutting an edge $e$} \label{fig2}
\end{figure}
\begin{definition} \label{defD}
For any $g_{i\bar j k\bar l \beta_1\dots\beta_{r}}$, as in remark
\ref{rm1}, there exists a canonical polynomial of covariant
derivatives of curvature tensor, denoted by $D(g_{i\bar j k\bar l
\beta_1\dots\beta_{r}})$, that coincides with $g_{i\bar j k\bar l
\beta_1\dots\beta_{r}}$ at the center of the normal coordinate. By
\eqref{eqcur1} and \eqref{eqcur4}, $D(g_{i\bar j k\bar l
\beta_1\dots\beta_{r}})$ may be written as a unique universal
polynomial of $g_{a\bar b\gamma_{1}\dots\gamma_t}$ with $t\geq 1$.
\end{definition}
For example,
$$D(g_{i\bar j k\bar l})=g_{i\bar j k\bar l}-g^{m\bar p}g_{m\bar j\bar l}g_{ik\bar p}.$$
Obviously, the above identity can be expressed in terms of
admissible trees
\begin{equation}
D\left(\begin{tabular}{c} \xymatrix@C=5mm@R=5mm{\ar[dr]^{\bar l} &&
\\ &\circ \ar[ur]^k \ar[dr]^i &
\\\ar[ur]^{\bar j} &&}
\end{tabular}\right) = \begin{tabular}{c}\xymatrix@C=5mm@R=5mm{\ar[dr]^{\bar l} &&
\\ &\circ \ar[ur]^k \ar[dr]^i &
\\\ar[ur]^{\bar j} &&}\end{tabular} - \begin{tabular}{c}\xymatrix@C=7mm@R=5mm{\ar[dr]^{\bar l} && &\\ &\circ \ar[r] & \circ \ar[ur]^k \ar[dr]^i &
\\\ar[ur]^{\bar j} &&&}\end{tabular}
\end{equation}
The map $D$ can be naturally extended to be defined on any
admissible tree with both indegree and outdegree $\geq2$ at each
vertex. We simply apply $D$ to each vertex while keeping index
pairings and expanding linearly.
Let $\mathcal T_m(\alpha_1,\dots,\alpha_r)$ denote the set of all
indecomposable admissible trees with $m$ vertices and $r$ half edges
labeled by the set of indices $\{\alpha_1,\dots,\alpha_r\}$.
\begin{lemma} \label{covar} We have
\begin{equation} \label{eqcur9}
D(g_{i\bar j k\bar l \beta_1\dots\beta_{r}})=\sum_{m=1}^{r+2}\
\sum_{T\in \mathcal T_m(i,\bar j, k,\bar l, \beta_1,\dots,\beta_r)}
(-1)^{m+1} T.
\end{equation}
\end{lemma}
\begin{proof} (Sketch) We first note that equation \eqref{eqcur7}
corresponds to the action on edges in Figure \ref{fig3} and the
multiplication by $\Gamma$ in equation \eqref{eqcur4} corresponds to
the action on half-edges in Figure \ref{fig4}.
By \eqref{eqcur1}, we have
\begin{equation}
R_{i\bar jk\bar l \beta_1\dots\beta_{r}} =-g_{i\bar j k\bar l
\beta_1\dots\beta_{r}}+(g^{m\bar p}g_{m\bar j\bar l}g_{ik\bar
p})_{\beta_1\dots\beta_{r}}
\end{equation}
The left-hand side is a derivative of $R_{i\bar jk\bar l}$. By
induction, we can see that $D((g^{m\bar p}g_{m\bar j\bar l}g_{ik\bar
p})_{\beta_1\dots\beta_{r}})$ is a sum over all decomposable
admissible trees with the set of indices $\{i,\bar j, k,\bar l,
\beta_1,\dots,\beta_r\}$ such that no irreducible component of the
tree can contain indices from both of the two sets $\{i,k\}$ and
$\{\bar j,\bar l\}$. We can apply \eqref{eqcur4} inductively to see
that all trees from $R_{i\bar jk\bar l \beta_1\dots\beta_{r}}$ are
admissible and get the desired equation \eqref{eqcur9}.
\end{proof}
\vskip 30pt
\section{The local Bergman kernel and the Berezin transform}
\label{berezin}
In this section, we consider only the case that $\mathcal E=\mathbb
C$. So the Bergman kernel $B_m(x)$ is a scalar function on $M$.
Now we introduce the local Bergman kernel following \cite{Eng}. Let
$\Omega$ be a bounded domain in $\mathbb C^n$ and $\Phi$ be a
K\"{a}hler potential for a K\"{a}hler metric $g$ on $\Omega$
satisfying
$$\omega_g=\frac{\sqrt{-1}}{2\pi}\partial\overline\partial\Phi.$$
To simplify, we may assume $\Omega$ is equipped with a normal
coordinate.
Let $\Phi(x,y)$ be an almost analytic extension to a neighborhood of
the diagonal, i.e. $\bar\partial_x\Phi$ and $\partial_y\Phi$ vanish
to infinite order for $x=y$ (cf. \cite{BS}). We can assume
$\overline{\Phi(x,y)}=\Phi(y,x)$. Consider the real valued function
$$D(x, y) = \Phi(x, x) + \Phi(y, y)-\Phi(x, y)-\Phi(y, x),$$
which is called the Calabi diastatic function \cite{Cal}. It is
easily seen that the function $D(x, y)\geq0$ and $D(x, y)=0$ if and
only $x=y$.
We need the following important result of Engli\v{s} \cite{Eng}.
\begin{theorem}{\rm(Engli\v{s})} \label{eng}
There is an asymptotic expansion for the Laplace integral
$$\int_{\Omega} f(y)e^{-mD(x,y)}\frac{\omega_g^n(y)}{n!} \sim \frac{1}{m^n}\sum_{j\geq0}m^{-j}R_j(f)(x),$$
where $R_j: C^\infty(\Omega)\rightarrow C^\infty(\Omega)$ are
explicit differential operators defined by
\begin{equation} \label{eqeng}
R_j f(x)=\frac{1}{\det g}\sum_{k=j}^{3j}\frac{1}{k!(k-j)!}L^k(f\det
g S^{k-j})|_{y=x},
\end{equation}
where $L$ is the (constant-coefficient) differential operator
$$L f(y)=g^{i\bar j}(x)\partial_i\partial_{\bar j} f(y)$$
and the function $S(x,y)$ satisfies $$S=\partial_\alpha
S=\partial_{\alpha\beta}S=\partial_{i_1 i_2\dots
i_m}S=\partial_{\bar i_1\bar i_2\dots \bar i_m}S=0 \quad \text{ at }
y=x,$$
$$\partial_{i\bar j\alpha_1\alpha_2\dots
\alpha_m}S|_{y=x}=-\partial_{\alpha_1\alpha_2\dots \alpha_m}g_{i\bar
j}(x).$$ In particular, we have
$$\begin{cases}
R_0=id
\\ R_1(f)=\Delta f-\frac12 f\rho.
\\ R_2(f)=\frac12\Delta^2f-\frac12 L_{Ric}f-\frac\rho
2\Delta f-\frac12(\rho_i f_{;i}+\rho_{\bar i}f_{;\bar i})\\\qquad
\qquad -(\frac13\Delta\rho-\frac18\rho^2-\frac16
|Ric|^2+\frac1{24}|R|^2)f.
\end{cases}$$
\end{theorem}
Here $L_{Ric}f=R_{j\bar i}f_{;i\bar j},\, |Ric|^2=R_{i\bar
j}R_{j\bar i},\, |R|^2=R_{i\bar j k\bar l}R_{j\bar i l\bar k}$.
Note that our convention of curvatures $R_{i\bar j k\bar l},
R_{i\bar j}, \rho$ all differ by a minus sign with that of
\cite{Eng}.
We need the following extension of Tian-Yau-Zelditch asymptotic
expansion \cite{Cat, Eng2, KS}.
\begin{theorem} \label{tyz2}
Let $M$ be a compact complex manifold endowed with a polarized
K\"{a}hler metric $g$. Let $B_m(x, y)$ denote an almost analytic
extension of $B_m(x)$ to an open neighborhood, say $U \times U$ of
the diagonal. Then, for $U$ sufficiently small, $B_m(x, y)$ admits
an asymptotic expansion (as $m\rightarrow +\infty$) of the form:
$$B_m(x, y)\sim\sum_{j\geq0} a_j(x,y)m^{n-j}.$$
\end{theorem}
For $\alpha > 0$, consider the weighted Bergman space of all
holomorphic function on $\Omega$ square-integrable with respect to
the measure $e^{-\alpha \Phi}\frac{w_g^n}{n!}$.
We denote by $K_\alpha(x,y)$ the reproducing kernel. As pointed out
by Engli\v s \cite{Eng}, it is often the case that the following
holds. (e.g. if $(\Omega, g_{i\bar j})$ is a bounded symmetric with
the invariant metric or whenever $\Omega$ is strongly pseudoconvex
with real analytic boundary.)
\begin{itemize}
\item[(1)] $K_\alpha(x,y)$ has an asymptotic expansion in a small
neighborhood of the diagonal
\begin{equation}\label{eqb1}
K_\alpha (x,y)\sim e^{\alpha \Phi (x,y)} \sum^\infty_{k=0} b_k (x,y)
\alpha^{n-k}.
\end{equation}
\item[(2)] For any neighborhood $U$ of a point $x \in \Omega$ and a
bounded measurable function $f$,
\begin{equation}\label{eqb2}
\int_{\Omega\backslash U} f(y)\frac{|K_\alpha
(x,y)|^2}{K_\alpha(x,x)} e^{-\alpha \Phi (y)}
\frac{w_g^n(y)}{n!}=o(\alpha^{-k}), \quad \forall\,
k\geq1.\end{equation}
\end{itemize}
Denote $b_k(x,x)$ by $b_k(x)$.
\begin{theorem} \label{berg}
For any $k\geq0$, we have $a_k(x)=b_k(x)$.
\end{theorem}
This theorem means that the global Bergman kernel can be
approximated by the local Bergman kernel. An analytic proof of this
theorem can be found in \cite{BBS}. Here we present a simple proof
which is implicit in \cite{Loi}.
The {\it Berezin transform} is the integral operator
\begin{equation}
I_\alpha f(x)=\int_\Omega
f(y)\frac{|K_\alpha(x,y)|^2}{K_\alpha(x,x)}e^{-\alpha
\Phi(y)}\frac{w_g^n(y)}{n!}.
\end{equation}
At any point for which $K_\alpha (x,x)$ invertible, the integral
converges for each bounded measurable function $f$ on $\Omega$.
Assume that \eqref{eqb1} and \eqref{eqb2} hold, then the Berezin
transform has an asymptotic expansion (see \cite{Eng})
\begin{equation}
I_\alpha f(x)=\sum^\infty_{k=0} Q_k f(x)\alpha^{-k},
\end{equation}
where $Q_k$ are differential operators.
We have that for $(x,y)$ near the diagonal,
\begin{equation} \label{eqb12}
\frac{|K_\alpha (x,y)|^2}{K_\alpha(x,x)} e^{-\alpha \Phi
(y)}=e^{-\alpha D(x,y)} \sum^\infty_{k=0} \tilde b_{k} (x,y)
\alpha^{n-k},
\end{equation}
where $\tilde b_0=1, \tilde b_1(x,y)=b_1(x,y)+b_1(y,x)-b_1(x,x),$
etc.
Fix a bounded neighborhood $U$ of $x$ such that \eqref{eqb12} holds.
Applying Theorem \ref{eng}, we get
\begin{equation} \label{eqb3}
Q_k f(x)=\sum^k_{j=0} R_j (\tilde b_{k-j}(x,y)f(y))|_{y=x},
\end{equation}
where the operators $R_j$ apply to the $y$-variable.
Since $Q_k f=0$ when $k>0$ and $f$ is analytic or anti-analytic,
setting $f=1$ in \eqref{eqb3}, we get
\begin{equation}\label{eqb4}
b_k(x)=-\sum^k_{j=1} R_j(\tilde b_{k-j}(x,y))|_{y=x}, \quad k\geq 1.
\end{equation}
The following result from quantization of K\"ahler manifolds can be
found in \cite{CGR, Loi}.
\begin{theorem} \label{qua}
In the same hypothesis of Theorem \ref{tyz2}, we have a globally
defined function on $M$. $$\psi_m
(x,y)=\frac{e^{-mD(x,y)}|B_m(x,y)|^2}{B_m (x) B_m(y)}.$$ Then
\begin{equation}\label{eqb5}
\int_M \psi_m(x,y) B_m(y)\frac{w_g^n(y)}{n!}=1
\end{equation}
Moreover, for any neighborhood $U$ of $x\in M$ and every smooth
function $f$ on $M$.
$$\int_{M\backslash U}\psi_m(x,y) B_m(y)\frac{w_g^n
(y)}{n!}=o(m^{-k}), \quad \forall\, k\geq1.$$
\end{theorem}
From \eqref{eqb5} and Theorem \ref{eng}, we get
$$\sum_{j=0}^k R_j(\tilde a_{k-j}(x,y))|_{y=x}=0,$$ where $\tilde
a_0=1, \tilde a_1(x,y)=a_1(x,y)+a_1(y,x)-a_1(x)$, etc. Note that
this is exactly the same recursive equation as \eqref{eqb4}. Since
$a_0(x)=b_0(x)=1$, we must have $a_k(x)=b_k(x)$ for all $k\geq0$. So
we conclude the proof of Theorem \ref{berg}.
\vskip 30pt
\section{The Bergman kernel of vector bundles}
Let $h(x)$ be a Hermitian metric on $\Omega\times\mathbb C^r$. For
$\alpha
> 0$, consider the weighted Bergman space of all holomorphic
function $f: \Omega\rightarrow \mathbb C^r$ square-integrable with
respect to the norm
\begin{equation}
||f||^2=\int_{\Omega}\langle f,f\rangle_{h} e^{-\alpha
\Phi}\frac{w_g^n}{n!}.
\end{equation}
Now the Bergman kernel $K_\alpha(x,y)$ takes value in matrices. The
results in the last section could be extended without difficulty to
the matrix case.
By the property of reproducing kernels, we have for $v\in \mathbb
C^r$,
\begin{equation} \label{eqb6}
\langle K_\alpha(x,x)f(x), v\rangle=\int_\Omega\langle
K_\alpha(x,y)f(y), K_\alpha(x,y)v\rangle e^{-\alpha\phi(y)}g(y)dy.
\end{equation}
We take $f(x)$ to be a constant function and take the inner product
to be the Hermitian metric $h_{\alpha\bar \beta}$. Let
$H=h_\alpha^\beta \in \Gamma(End(\mathcal {E}))$ and apply Theorem
\ref{eng} to \eqref{eqb6}, we get
\begin{equation} \label{eqb7}
a_k(x)=\sum_{\ell+i+j=k}R_{\ell}(a_i(x,y)a_j(y,x)H)|_{y=x}.
\end{equation}
\begin{remark}
The connection of Engli\v{s}' work and the asymptotic expansion of
Tian-Yau-Zelditch was studied by Loi \cite{Loi}, who obtained the
formula \eqref{eqb7} when $\mathcal E=\mathbb C$ using Theorem
\ref{qua}.
\end{remark}
From the curvature equation $$\Theta_{i\bar j}H=-H_{i\bar
j}+\partial_i H\partial_{\bar j}H,$$ we have
\begin{align}
(H_{i\bar i})_{/j\bar
j}&=-\Theta_{i\bar i / j\bar j}H-\Theta_{i\bar i}H_{j\bar j}+H_{i\bar j}H_{j\bar i} \nonumber\\
\label{eqcur} &=-\wedge \partial \bar \partial \wedge \Theta+\wedge
\Theta \wedge\Theta+\Theta_{i\bar j}\Theta_{j\bar i}.
\end{align}
Since $R_0$ is the identity operator, we have
\begin{equation} \label{eqb8}
a_k(x)=-R_k(H)-\sum_{\substack{i+j=k
\\i,j\geq1}}a_i(x)a_j(x)H-\sum_{\substack{\ell+i+j=k\\1\leq\ell\leq
k-1}}R_\ell(a_i(x,y)a_j(y,x)H)|_{y=x}.
\end{equation}
Then it is not difficult to compute the first few terms
\begin{equation}\label{eqtyz1} a_1(x)=-\Delta H+\frac{\rho}{2}H=\frac{\rho}{2}I+\wedge\Theta.
\end{equation}
Note that in the second equation, we used $H(0)=I$.
For $k=2$, we need a little more work.
$$a_2(x)=-R_2(H)-a_1(x)^2H-R_1(a_1(x,y)H)|_{y=x}-R_1(a_1(y,x)H)|_{y=x}.$$
We compute terms in the right-hand side one by one. By formula
\eqref{eqcur}, we have
\begin{align*}
R_2(H)=&\frac{1}{2}\Delta\Delta H-\frac{1}{2}R_{i\bar j}H_{j\bar i}-\frac{\rho}{2}\Delta H-(\frac{1}{3}\Delta\rho-\frac{1}{8}\rho^2-\frac{1}{6}|Ric|^2+\frac{1}{24}|R|^2)H\\
=&(-\frac{1}{2}\wedge\partial\bar\partial\wedge\Theta+\frac{1}{2}\Theta_{i\bar
j}\Theta_{j\bar
i}+\frac{1}{2}\wedge\Theta\wedge\Theta)+\frac{1}{2}R_{i\bar
j}\Theta_{j\bar
i}+\frac{\rho}{2}\wedge\Theta\\
&-(\frac{1}{3}\Delta\rho-\frac{1}{8}\rho^2-\frac{1}{6}|Ric|^2+\frac{1}{24}|R|^2)I.
\end{align*}
Substituting $a_1(x)$ computed in \eqref{eqtyz1}, we have
$$a_1(x)^2H=\frac{\rho^2}{4}I+\rho\wedge\Theta+\wedge\Theta\wedge\Theta.$$
From almost-analyticity of $a_1(x,y)$, we get
\begin{align*} R_1(a_1(x,y)H)|_{y=x}&=\Delta(a_1(x,y)H)|_{y=x}-\frac{\rho}{2}a_1(x)H\\
&=a_1(x)\Delta
H-\frac{\rho}{2}a_1(x)\\&=-\frac{\rho}{2}\wedge\Theta-\wedge\Theta\wedge\Theta-\frac{\rho^2}{4}I-\frac{\rho}{2}\wedge\Theta
\\&=-\frac{\rho^2}{4}I-\wedge\Theta\wedge\Theta-\rho\wedge\Theta.
\end{align*}
Similarly,
$R_1(a_1(y,x)H)|_{y=x}=-\frac{\rho^2}{4}I-\wedge\Theta\wedge\Theta-\rho\wedge\Theta$.
Summing up the above computation, we arrive at the desired value of
$a_2(x)$,
\begin{align}
a_2(x)=&\frac{1}{2}\Delta\wedge\Theta+\frac{1}{2}(\wedge\Theta)^2+\frac{\rho}{2}\wedge\Theta-\frac{1}{2}\Theta_{i\bar
j}\Theta_{j\bar i}-\frac{1}{2}R_{i\bar j}\Theta_{j\bar
i}\\&+(\frac{1}{3}\Delta\rho+\frac{1}{8}\rho^2-\frac{1}{6}|Ric|^2+\frac{1}{24}|R|^2)I,
\nonumber
\end{align}
where $\Delta\wedge\Theta=\wedge\partial\bar\partial\wedge\Theta$.
\begin{definition} {\rm (Lu \cite{Lu})}
Let $P$ be a $d$-th order covariant derivative of $R_{i\bar jk\bar
l}, R_{i\bar j},\rho,\Theta,\wedge\Theta$. Define the weight and the
order of $P$ to be the numbers $w(P)=(1+\frac d2)$ and ${\rm
ord}(P)=\frac d2$ respectively. The weight and order functions can
be extended additively to monomials of curvatures and their
derivatives.
\end{definition} For example,
$$w(R_{i\bar jk\bar
l})=1,\quad w(R_{i\bar j,k})=\frac32, \quad w(R_{i\bar j,k}R_{i\bar
j k\bar l})=\frac52,
$$
$${\rm ord}(R_{i\bar jk\bar
l})=0,\quad {\rm ord}(\rho_i R_{i\bar jk\bar l})=\frac12,\quad {\rm
ord}(\Delta\wedge\Theta)=1.$$
In the following, the word ``leading term'' will mean the sum of
terms with the highest order.
\begin{lemma} \label{coef1}
The leading term in $a_k$ is $\frac{k}{(k+1)!}\Delta^{k-1}\rho
I+\frac{1}{k!}\Delta^{k-1}\wedge\Theta$.
\end{lemma}
\begin{proof}
Let $u_k$ denote the coefficient of $\Delta^{k-1}\rho$ in
$R_{k}(1)$. Then from \eqref{eqeng},
$$R_{k}(1)=\frac{1}{\det g}\sum^{2k}_{i=k}\frac{1}{i!(i-k)!}L^{i}(\det g S^{i-k})|_{y=0},$$
here we take normal coordinate system around point $x=0$ on the
K\"ahler manifold $M$. It is not difficult to see that the only
terms that may contribute to $\Delta^{k-1}\rho$ must have $i=k$ or
$i=k+1$. So we have
$$u_k=-\frac{1}{k!}+\frac{1}{(k+1)!}=-\frac{k}{(k+1)!}.$$
From \eqref{eqb8}, we see that the coefficient of $\Delta^{k-1}\rho$
in $a_k$ equals $-u_k=\frac{k}{(k+1)!}$.
Similarly, we can prove that the coefficient of
$\Delta^{k-1}\wedge\Theta$ in $a_k$ equals $\frac{1}{k!}$.
\end{proof}
\begin{remark} Lemma \ref{coef1} was obtained by Liu and Lu \cite{LL} using peak section
method, improving the $\mathcal E=\mathbb C$ case proved in
\cite{LT}.
\end{remark}
Let $M=M_1 \times M_2$ a product of two projective manifolds
equipped with the product K\"ahler metric and the twisted bundles
$\mathcal E^{(1)}(m)\boxtimes \mathcal E^{(2)}(m)$. Denote the
Bergman kernel of $M$ by $B_m^{(M)}$. The following fact is
well-known.
\begin{lemma}\label{prodrel}
We have $B^{(M)}_m(x_1,x_2) = B^{(M_1)}_m(x_1) B^{(M_2)}_m(x_2)$.
\end{lemma}
\begin{proof}
We just need to note that if $(s_1, \cdots, s_i, \cdots)\ and \
(t_1, \cdots, t_j, \cdots)$ are orthonormal basis of $\mathcal
E^{(1)}(m)$ and $\mathcal E^{(2)}(m)$ respectively, then $(s_1 t_1,
\cdots, s_i t_j, \cdots)$ is an orthonormal basis of $\mathcal
E^{(1)}(m)\boxtimes \mathcal E^{(2)}(m)$.
\end{proof}
Consider the semigroup $N^\infty$ of
sequences ${\bold d}=(d_1,d_2,\dots)$ where $d_i$ are nonnegative
integers and $d_i=0$ for sufficiently large $i$. We sometimes also
use $(1^{d_1}2^{d_2}\dots)$ to denote $\bold d$. We will use the
following notation:
\begin{equation} \label{eqnotation}
|\bold d|:=\sum_{i\geq 1}i\cdot d_i,\quad ||\bold
d||:=\sum_{i\geq1}d_i, \quad {\bold d}!=\prod_{i\geq1}d_i!, \quad
\bold c\cdot\bold d:=\sum_{i\geq1}c_id_i.
\end{equation}
\begin{proposition}
Let $\bold{c, d, e}\in N^\infty$ and $|\bold c|+\bold d\cdot\bold
e+||\bold e||=k$. Let $u(\bold{c,d,e})$ denote the coefficient of
the weight $k$ monomial $(\text{where }d_j\ne d_{j+1}, \forall\,
j\geq1)$
$$\prod_{i\geq1}
(\Delta^{i-1}\rho)^{c_i}\prod_{j\geq1}(\Delta^{d_j}\wedge\Theta)^{e_j}$$
in $a_k$. Then we have
\begin{equation}\label{eqb9}
u(\bold{c,d,e})=\frac{1}{\bold c!}\prod_{i\geq
1}\left(\frac{i}{(i+1)!}\right)^{c_i}\cdot \frac1{\bold
e!}\prod_{j\geq1}\left(\frac1{(d_j+1)!}\right)^{e_j}.
\end{equation}
\begin{proof}
By Lemma \ref{coef1} and Lemma \ref{prodrel}, we have the following
relations:
\begin{align*}
c_i\cdot u(\bold{c,d,e})&=\frac{i}{(i+1)!} u(\bold c-\bm{\delta}_i,\bold{d,e}), \\
e_j\cdot u(\bold{c,d,e})&= \frac{1}{(d_j+1)!}u(\bold{c,d},\bold
e-\bm{\delta}_j),
\end{align*}
where $\bm{\delta}_i\in N^\infty$ denotes the sequence with $1$ at
the $i$-th place and zeros elsewhere.
From $u(\bold{0,0,0})=1$, we can recursively prove the desired
formula \eqref{eqb9} of $u(\bold{c,d,e})$.
\end{proof}
\end{proposition}
\vskip 30pt
\section{Weyl invariants and graphs}
\label{secgraph}
In the rest of the paper, we will restrict to the case $\mathcal
E=\mathbb C$, namely $\Theta=0$.
The $a_k$'s in \eqref{eqb10} are the so-called {\it Weyl invariants}
introduced by Fefferman \cite{Fef}. Consider the tensor products of
covariant derivatives of the curvature tensor $R_{i\bar j k \bar l;
p\cdots \bar q}$, e.g.
$$R_{ijk \bar l;p \bar q}\otimes\cdots \otimes R_{a \bar b c \bar d; \bar
e}.$$ The Weyl invariants are constructed by first pairing up the
unbarred indices to barred indices and then contracting all paired
indices.
For the sake of brevity, Weyl invariants such as
$$g^{{i_1}\bar j_1} g^{{i_2}\bar j_2}g^{{i_3}\bar j_3}g^{{i_4}\bar
j_4}g^{{i_5}\bar j_5} R_{{i_1}\bar j_{1} i_2 \bar j_3; i_5 \bar j_4}
R_{{i_3}\bar j_{2} i_4 \bar j_5 }$$ will be abbreviated as
\begin{equation} \label{eqweyl}
W:=R_{ i_1 \bar i_1 i_2 \bar i_3 ; i_5\bar i_4} R_{ i_3 \bar i_2 i_4
\bar i_5 },
\end{equation}
knowing that $(i_k, \bar i_k), 1\leq k\leq 5$ are paired indices to
be contracted.
It is useful to represent Weyl invariants as digraphs (also called
quivers), namely directed graphs possibly with loops and
multi-edges. We put curvature tensors as nodes and draw a directed
edge from $i_k$ to $\bar i_k$ for each $k$. For example, the
associated graph of the above Weyl invariant \eqref{eqweyl} is
depicted in Figure \ref{fig6}.
\begin{figure}[h]
$\xymatrix{
*+[o][F-]{1} \ar@/^1pc/[rr]^2
& &
\circ \ar@/^1pc/[ll]^2}
$ \caption{The associated graph of $W$} \label{fig6}
\end{figure}
In this paper, we represent a multidigraph as a weighted digraph.
The weight of a directed edge is the number of multi-edges. The
number attached to a vertex denotes the number of its self-loops. A
vertex without loops will be denoted by a small circle. The indegree
and outdegree of a vertex $v$ are denoted by $\deg^-(v)$ and
$\deg^+(v)$ respectively.
By Ricci formula, we cannot recover a Weyl invariant from its
associated graph, namely different Weyl invariants may have the same
associated graph. As mentioned in Remark \ref{rm1}, a remedy for the
discrepancy is to express the Weyl invariant in terms of derivatives
of the K\"ahler metric. No information will be lost and the
uniqueness of the associated graph is guaranteed by \eqref{eqcur6}.
\begin{definition} We call a vertex $v$ of a digraph $G$ semistable if we have
$$\deg^-(v)\geq1,\ \deg^+(v)\geq1,\ \deg^-(v)+\deg^+(v)\geq3.$$
$G$ is called {\it semistable} if each vertex of $G$ is semistable.
We call $v$ stable if $\deg^-(v)\geq2,\ \deg^+(v)\geq2$. A digraph
$G$ is {\it stable} if each vertex of $G$ is stable.
Let $\mathcal G_{r,s}$ denote the set of stable digraphs with $r$
vertices and $s$ edges. The set of semistable and stable graphs will
be denoted by $\mathcal G^{ss}$ and $\mathcal G$ respectively. The
associated graph of a Weyl invariant lies in $\mathcal G$. For any
$G\in \mathcal G_{r,s}$, its weight $w(G)$ is defined to be $s-r$.
We denote by $\mathcal G^{ss}(k)$ and $\mathcal G(k)$ respectively
the set of semistable and stable digraphs with weight $k$. Let
$\mathcal G_{con}(k)$ and $\mathcal G_{scon}(k)$ respectively be the
set of connected and strongly connected graphs in $\mathcal G(k)$.
We also define a special set of graphs:
$$\Lambda(k)=\{G\in\mathcal G_{scon}(k)\mid 1 \text{ is not an
eigenvalue of } A(G)\}.$$ This set is of interest in view of
Corollary \ref{eigenvalue}. We have computed the cardinalities of
these sets when $k\leq5$ in Table \ref{tb2}.
\end{definition}
\begin{table}[h]
\caption{Numbers of stable graphs} \label{tb2}
\begin{tabular}{|c||c|c|c|c|c|}
\hline
$k$ &$1$&$2$&$3$ &$4$ &$5$
\\\hline $|\mathcal G(k)|$ &$1$&$4$&$15$&$82$&$589$
\\\hline $|\mathcal G_{con}(k)|$ &$1$&$3$&$11$&$61$&$474$
\\\hline $|\mathcal G_{scon}(k)|$ &$1$&$3$&$10$&$51$&$373$
\\\hline $|\Lambda(k)|$ &$1$&$3$&$9$ &$45$&$316$
\\\hline
\end{tabular}
\end{table}
It is not difficult to see that $\mathcal G(k)= \cup_{j=1}^k
\mathcal G_{j,j+k}$ for any $k\geq1$.
\begin{remark}
It shall be interesting to find formulae, closed or recursive, for
the number of graphs in the above sets.
\end{remark}
We can write the coefficient $a_k$ as a polynomial of $g_{i\bar j
\beta_1\dots\beta_{r}},\, r\geq 1$, and consequently as a sum over
graphs in $\mathcal G^{ss}(k)$,
\begin{equation} \label{eqc}
a_k(x)=\sum_{G\in\mathcal G^{ss}(k)} z(G)\cdot G, \quad
z(G)\in\mathbb Q.
\end{equation}
So we may regard $z$ as a map from $\mathcal G^{ss}$ to $\mathbb Q$.
By Remark \ref{rm1} and Lemma \ref{covar}, it is enough to know only
those $z(G)$ of stable graphs $G\in\mathcal G$.
\begin{remark}
For convenience, we assume that the empty graph $\emptyset$ is the
unique element in $\mathcal G(0)$ and $z(\emptyset)=1$. This is
consistent with $a_0=1$.
\end{remark}
First we recall Loi's recursion formula \cite{Loi}
\begin{equation} \label{eqloi}
a_k(x)=-R_k(1)-\sum_{i+j=k \atop
i,j\geq1}a_i(x)a_j(x)-\sum_{\ell+i+j=k\atop 1\leq \ell\leq
k-1}R_\ell (a_i(x,y)a_j(y,x))|_{y=x}.
\end{equation}
It was pointed out to the author recently that essentially the same identity was also obtained independently
in \cite{Cha}.
\begin{remark}
Here we outline the strategy of our proof of the closed formulae for
$z(G)$ in Theorem \ref{main}. We will give a graph-theoretic
interpretation of Loi's formula in Proposition \ref{alltyz}, the key
ingredient is the graphical formulae for derivatives of $\det g$
proved in Lemma \ref{dettree}. Then we extract the coefficient
$z(G)$ for any given graph $G$ in Lemma \ref{single}, which will be
used to prove $z(G)=0$ for any connected but not strongly connected
graph $G$ in Proposition \ref{sink}. By specializing Lemma
\ref{single} to a strongly connected graph $G$, we prove in Lemma
\ref{single2} that $z(G)$ can be written as a summation over
equivalent classes of linear subgraphs of $G$, which implies the
explicit closed formula in Proposition \ref{strong} through a
classical result in spectral graph theory: the Coefficient Theorem.
\end{remark}
Next we introduce some notation. Let $\mathscr L$ be the set of
digraphs consisting of a finite number of vertex-disjoint simple
cycles (i.e. simple polygons without common vertex). The length of a
simple cycle is defined to be the number of its edges. For each
graph $L\in \mathscr L$, we can write $L$ as a finite increasing
sequence of nonnegative integers $[i_1,\dots, i_m]$, meaning $L$
consists of $m$ disjoint simple cycles, whose lengths are specified
by $i_1,\dots,i_m$. We define the index of $L$ to be
\begin{equation}
i(L)=m+i_1+\dots+i_m.
\end{equation}
Note that $[0]$ is just a single vertex and $[1]$ is a vertex with a
self-loop. If $0\notin L$, then $L$ is usually called a linear
digraph. Recall that a {\it linear directed graph} is a digraph in
which $\deg^+(v)=\deg^-(v)=1$ for each vertex $v$.
Given a set of indices ${\alpha_1,\dots,\alpha_r}$, denote by
$\mathscr L(\alpha_1,\dots,\alpha_r)$ the isomorphism classes of all
possible decorations of the vertices of $L\in \mathscr L$ with the
half-edges ${\alpha_1,\dots,\alpha_r}$ requiring each vertex to be
semistable. Two decorations of $L$ that differ by a graph
isomorphism are considered the same.
The following crucial lemma explains the graphical properties of the
partial derivatives of $\det g$.
\begin{lemma} \label{dettree}
We have
\begin{equation} \label{eqdettree}
\frac{1}{\det g}\partial_{\alpha_1}\dots\partial_{\alpha_r}\det g
=\sum_{\substack{L\in \mathscr L(\alpha_1,\dots,\alpha_r)\\0\notin
L}} (-1)^{i(L)} \cdot L.
\end{equation}
\end{lemma}
\begin{proof}
By \eqref{eqcur7} and \eqref{eqdet}, we have $$\frac{1}{\det
g}\partial_{\alpha}\det g=g_{j\bar j\alpha}= \xymatrix{ \circ
\ar@(ul,dl)[]|{} \ar@{-}[r]^\alpha &}.$$
$$
\frac{1}{\det g}\partial_{\beta}\partial_{\alpha}\det g =g_{i\bar
i\alpha}g_{j\bar j\beta}-g_{i\bar j\alpha}g_{j\bar i\beta}+g_{j\bar
j\alpha\beta}$$\bigskip
$$=\quad \xymatrix{ \circ \ar@(ur,ul)[]
\ar@{-}[d]^\alpha & \circ \ar@(ur,ul)[] \ar@{-}[d]^\beta \\&}\quad
-\quad \xymatrix{ \circ\ar@{-}[d]_\beta \ar@/^/[r] &
\circ\ar@{-}[d]^\alpha \ar@/^/[l]\\& }\quad + \xymatrix{ &\circ
\ar@(ur,ul)[] \ar@{-}[dr]^\alpha \ar@{-}[dl]_\beta & \\&&}.
$$
So equation \eqref{eqdettree} follows by an easy induction.
\end{proof}
\begin{definition} \label{config}
Let $L\in\mathscr L$, $G_1, G_2 \in \mathcal G^{ss}$. We denote by
$\mathscr G(L,G_1,G_2)$ the set of all different ways to add a
finite number of directed edges to the disjoint union of three
graphs $L,G_1,G_2$ satisfying that
\begin{enumerate}
\item[i)]
No head of these new edges is connected to vertices of $G_1$ and no
tail of these new edges is connected to vertices of $G_2$. Such
$G_1$ and $G_2$ are called source and sink respectively.
\item[ii)] These new edges may act on edges of $G_1$ and $G_2$ as illustrated
in Figure \ref{fig3}.
\item[iii)] The final resulting digraph is semistable.
\end{enumerate}
\end{definition}
For $\mathcal Z\in \mathscr G(L,G_1,G_2)$, we denote by $[\mathcal
Z]$ the resulted graph. Two configurations $[\mathcal Z_1],
[\mathcal Z_2]\in \mathscr G(L,G_1,G_2)$ are considered equal if
there is an automorphism between $[\mathcal Z_1]$ and $[\mathcal
Z_2]$ leaving $L$ invariant and keeping $G_1$ and $G_2$ fixed.
Given $k\geq 1$, we also define a function
\begin{equation} \label{eqF}
F_k(L,G_1,G_2)=\sum_{\substack{\mathcal Z\in\mathscr G(L,G_1,G_2)\\
\omega([\mathcal Z])=k}}\frac{(-1)^{i(L)+1}[\mathcal Z]}{|{\rm
Aut}(\mathcal Z)|},
\end{equation}
where ${\rm Aut}(\mathcal Z)$ is the subgroup of the group of
automorphism of $[\mathcal Z]$ leaving $L$ invariant and keeping
$G_1$ and $G_2$ fixed. We may extend $F_k$ as a bilinear function in
the second and third parameters.
\begin{proposition} \label{alltyz} Given $k\geq1$, we have
\begin{equation} \label{eqgraph}
a_k(x)=\sum_{\substack{L\in\mathscr L,\ G_1,G_2\in\mathcal G^{ss} \\
\omega(G_1)<k,\ \omega(G_2)<k }} z(G_1)z(G_2) F_k(L,G_1,G_2).
\end{equation}
Note that we allow $L,G_1,G_2$ to be empty.
\end{proposition}
\begin{proof} In fact, \eqref{eqgraph} is just a graph-theoretic interpretation of
Loi's formula \eqref{eqloi}
\begin{equation}\label{eqloi2}
a_k(x)=-\sum_{\ell+i+j=k\atop i< k,\, j< k}R_\ell
(a_i(x,y)a_j(y,x))|_{y=x}.
\end{equation}
We need to take a close look at the operators $R_j$ defined in
\eqref{eqeng}. Note that the function $S$ in \eqref{eqeng} plays
the role of an isolated vertex $[0]\in\mathscr L$, the cycle with
length zero. The linear subgraphs are contributed by derivatives of
$\det g$ as described in Lemma \ref{dettree}. Moreover, $G_1$ and
$G_2$ respectively represent $a_j(y,x)$ and $a_i(x,y)$ in the
right-hand side of \eqref{eqloi2}. The factorials in the denominator
of \eqref{eqeng} account for the removal of labels of the added
edges and isolated vertices in each configuration. With each term of
\eqref{eqeng} thus interpreted,
\eqref{eqgraph} follows readily from \eqref{eqloi2}.
\end{proof}
If $[\mathcal Z]=G$, we call $\mathcal Z$ a $G$-configuration (abbr.
$G$-config). For a given digraph $G$, to specify a $G$-configuration
is equivalent to specify three vertex-disjoint subgraphs $L, G_1,
G_2$, where $L$ is a linear directed subgraph of $G$ without
isolated vertex and $G_1, G_2\in\mathcal G^{ss}$ are respectively a
source and a sink. Note that the union of $L, G_1, G_2$ may not
contain all vertices of $G$. Let $\mathcal L$ be the set of all
linear directed subgraphs $L$ of $G$. We may simply write $\mathcal
Z=(L,G_1,G_2)\in G\text{-config}$ and define a function
\begin{equation} \label{eqF2}
F_G(\mathcal Z)=\frac{(-1)^{i(\mathcal Z)+1}}{|{\rm Aut}(\mathcal
Z)|},
\end{equation}
where $i(\mathcal Z)$ is the index of $\mathcal Z$ defined by
\begin{equation}
i(\mathcal Z)=v(L)+p(L)+x(\mathcal Z).
\end{equation}
Here $v(L)$ is the number of vertices in $L$, $p(L)$ is the number
of components in $L$ and $x(\mathcal Z)$ is the number of vertices
of $G$ not belong to $L, G_1, G_2$.
\begin{lemma} \label{single} For any $G\in \mathcal G^{ss}$, we have
\begin{equation}
z(G)=\sum_{\mathcal Z\in G\text{-config}}z(G_1)z(G_2)F_G(\mathcal
Z),
\end{equation}
where $\mathcal Z=(L,G_1,G_2)$ runs over all isomorphism classes of
$G$-configurations.
\end{lemma}
\begin{proof}
It follows from Proposition \ref{alltyz}.
\end{proof}
Now we treat the case that $G\in \mathcal G^{ss}(k)$ is strongly
connected. In the right-hand side of \eqref{eqloi}, the graphs from
the second term are not connected. The graphs from the last term are
not strongly connected, since it contains either a source or a sink.
We see that only the first term $-R_k(1)$ contributes to $z(G)$ and
all $G$-configurations are of the form $(L,\emptyset,\emptyset)$. It
is not difficult to see that the automorphism group of the
$G$-configuration $(L,\emptyset,\emptyset)$, denoted by ${\rm
Aut_G}(L)$, is the subgroup of ${\rm Aut}(G)$ leaving $L$ invariant.
Two $G$-configurations $(L_1,\emptyset,\emptyset)$ and
$(L_2,\emptyset,\emptyset)$ are equivalent if and only if
$L_1=h(L_2)$ for some $h\in {\rm Aut}(G)$. This defines an
equivalence relation $\sim$ on $\mathcal L$. So Lemma \ref{single}
specializes to the following formula for strongly connected graphs.
\begin{lemma} \label{single2} For any strongly connected graph $G\in \mathcal G^{ss}$, we have
\begin{equation}\label{eqzstrong}
z(G)=\sum_{L\in \mathcal L/\sim}\frac{(-1)^{|V|+1+p(L)}}{|{\rm
Aut}_G(L)|},
\end{equation}
where $L$ runs over the equivalence classes of linear subgraphs of
$G$ and $p(L)$ is the number of components in $L$.
\end{lemma}
\vskip 30pt
\section{Proof of Theorem \ref{main}}
\begin{proposition} \label{disjoint}
If $G\in \mathcal G^{ss}$ is not connected and we write $G$ as a
disjoint union of connected subgraphs $G=G_1\cup\dots\cup G_m$, then
we have
\begin{equation}
z(G)=\prod_{j=1}^m z(G_j)/|Sym(G_1,\dots,G_m)|,
\end{equation}
where $Sym(G_1,\dots,G_m)$ denote the permutation group of these $m$
connected subgraphs.
\end{proposition}
\begin{proof}
It follows from Lemma \ref{prodrel}.
\end{proof}
By the above proposition, we may restrict to compute $z(G)$ for
connected $G$. For a digraph $G=(V,E)$, we can partition $V$ into
strongly connected components, namely the maximal strongly connected
subgraphs of $G$. Among these strongly connected components, we have
at least one {\it sink} (a component without outgoing edges) and one
{\it source} (a component without incoming edges).
\begin{proposition} \label{sink}
Given a stable graph $G\in \mathcal G$ which is connected but not strongly connected, then we have
$z(G) = 0$.
\end{proposition}
\begin{proof} By definition, any sink or source of a stable graph $G$ must be at
least semistable.
Without loss of generality, we may assume that $C\in\mathcal G^{ss}$ is a sink of $G$.
We
have a nonempty subgraph $G'$ of $G$ containing all vertices not in
$C$, such that
there are $k$ arrows $e_1,\dots,e_k$ from $G'$ to $C$,
as depicted in Figure \ref{figsink}.
\begin{figure}[h]
$\xymatrix{ *+++[o][F-]{G'} \ar@/^/@{->} @< 3pt> [rr]^{e_1} \ar
@{-->}[rr] \ar@/_/@{->} @<-3pt> [rr]_{e_k} && *+++[o][F-]{C} }$
\caption{Decomposition of $G$ into a sink $C$ and $G'$}
\label{figsink}
\end{figure}
In this case, any $G$-configuration is a union of $\mathcal
Z'=(L',G_1',G_2')\in G'\text{-config}$ and $\mathcal
Z''=(L'',\emptyset,\emptyset)\in C\text{-config}$. Since $G$ is stable,
we can rule out Case (ii) of Definition \ref{config}.
Let ${\rm
Aut}_f(G)$ be the subgroup of ${\rm Aut}(G)$ fixing all edges
$e_1,\dots,e_k$. Let ${\rm Aut}_{f}(C)$ and ${\rm Aut}_{f}(\mathcal
Z'')$ be respectively the subgroups of ${\rm Aut}(C)$ and ${\rm
Aut}(\mathcal Z'')$
fixing the heads of the arrows $e_i,\,1\leq i\leq k$.
We define ${\rm Aut}_{f}(\mathcal Z')$ to be the subgroup of ${\rm
Aut}(\mathcal Z')$ fixing the tails of the arrows $e_i,\,1\leq i\leq
k$.
Given a $G'$-configuration $\mathcal Z'=(L',G_1',G_2')$ and a
$C$-configuration $\mathcal Z''=(L'',\emptyset,\emptyset)$, we
define the following set
\begin{equation}
S(\mathcal Z',\mathcal Z'')=\left\{\text{isomorphism classes in
}G\text{-config}\text{ of the form } \mathcal Z'\coprod p(\mathcal
Z''),\ p\in\frac{{\rm Aut} (C)}{{\rm Aut}_f (C)}\right\},
\end{equation}
where $p(\mathcal Z'')=(p(L''),\emptyset,\emptyset)\in
C\text{-config}$ and $\mathcal Z'\coprod p(\mathcal Z'')=(L'\coprod
p(L''),G_1',G_2')$.
First we assume that all automorphisms of $G$ fix these $e_i,\,1\leq
i\leq k$, namely ${\rm Aut}(G)= {\rm Aut}_f(G)$. By our assumption,
for $\mathcal Z=\mathcal Z'\coprod p(\mathcal Z'')\in S(\mathcal
Z',\mathcal Z'')$ we have ${\rm Aut}(\mathcal Z)={\rm
Aut}_f(\mathcal Z'){\rm Aut}_f(p(\mathcal Z''))$
It is enough to show that for any fixed $G'$-configuration $\mathcal
Z'=(L',G_1',G_2')$, the contribution to $z(G)$ from the
$G$-configurations $(L',G_1',G_2'\coprod C)$ and $\{S(\mathcal
Z',\mathcal Z'')\}_{\mathcal Z''\in C\text{-config}}$ add up to
zero.
We will proceed by induction on the weight of $G$. Namely we assume that if a semistable graph $H\in\mathcal G^{ss}$
is connected but not strongly connected and the weight of $H$
is less than the weight of $G$, then $z(H)=0$ (cf. Corollary \ref{conn}).
First we assume $G_2'=\emptyset$.
By Lemma \ref{single}, the contribution to $z(G)$ from the
$G$-configuration $\mathcal Z=(L',G_1',C)$ is
\begin{align}
&z(G_1')z(c)\frac{(-1)^{i(\mathcal Z')+1}}{|{\rm Aut}_f(\mathcal
Z')|}\times\frac{|{\rm Aut}(C)|}{|{\rm Aut}_{f}(C)|} \nonumber\\
=&z(G_1')\frac{(-1)^{i(\mathcal Z')+1}}{|{\rm Aut}_f(\mathcal
Z')|}\sum_{\mathcal Z''\in C\text{-config}}\frac{(-1)^{i(\mathcal
Z'')+1}}{|{\rm Aut}(\mathcal Z'')|}\times\frac{|{\rm Aut}(C)|}{|{\rm
Aut}_f(C)|}. \label{cont1}
\end{align}
The contribution to $z(G)$ from the set of $G$-configurations
$\{S(\mathcal Z',\mathcal Z'')\}_{\mathcal Z''\in C\text{-config}}$
is
\begin{equation} \label{cont2}
z(G_1')\frac{1}{|{\rm Aut}_f(\mathcal Z')|}\sum_{\substack{\mathcal
Z''\in C\text{-config}\\\mathcal Z'\coprod p(\mathcal Z'')\in
S(\mathcal Z',\mathcal Z'')}} \frac{(-1)^{i(\mathcal Z')+i(\mathcal
Z'')+1}}{|{\rm Aut}_f(p(\mathcal Z''))|}.
\end{equation}
We now show that the contributions to $z(G)$ in \eqref{cont1} and
\eqref{cont2} add up to zero, namely we need to prove that for any
$\mathcal Z'' \in C\text{-config}$
\begin{equation} \label{eqkey}
\frac{1}{|{\rm Aut}(\mathcal Z'')|}\times\frac{|{\rm Aut}(C)|}{|{\rm
Aut}_f(C)|}=\sum_{\mathcal Z'\coprod p(\mathcal Z'')\in S(\mathcal
Z',\mathcal Z'')}\frac{1}{|{\rm Aut}_f(p(\mathcal Z''))|}.
\end{equation}
This can be seen as follows: Denote by $H$ the set of all possible
distributions of the labeled heads of $e_i,\,1\leq i\leq k$ on $C$.
Note that some of these $e_i$ may have the same head on $C$. It is
obvious that $H$ is in one-to-one correspondence with ${\rm Aut}(C)/
{\rm Aut}_{f}(C)$.
We have the natural action of ${\rm Aut}(\mathcal Z'')$ on $H$, then
the set of orbits is just $S(\mathcal Z',\mathcal Z'')$ and the
isotropy group at $\mathcal Z=\mathcal Z'\coprod p(\mathcal Z'')\in
S(\mathcal Z',\mathcal Z'')$ is just ${\rm Aut}_f(p(\mathcal Z''))$.
If $G_2'\neq\emptyset$ and $G_2'$ is connected with a tail of some
$e_i$ ,then the contribution to $z(G)$ from the $G$-configurations
$(L',G_1',G_2'+ C)$ is zero by induction and all configurations in
$\{S(\mathcal Z',\mathcal Z'')\}_{\mathcal Z''\in C\text{-config}}$
are not allowed. So we may assume that there are no edges between
$G_2'$ and $C$, then we have the same cancellation by Proposition
\ref{disjoint}. Note that $G_2'+ C$ denote the induced subgraph whose
vertices are the union of vertices of $G_2'$ and $C$.
If ${\rm Aut}(G)\neq {\rm Aut}_f(G)$, the above argument still works
with appropriate modifications.
\end{proof}
\begin{remark}
We will give an alternative proof of Proposition \ref{sink} in
\cite{Xu}. As pointed out by the referee, Proposition \ref{sink}
should also follow from the path integral formula by Douglas and
Klevtsov \cite{DK}.
\end{remark}
\begin{corollary} \label{conn}
If $G\in \mathcal G^{ss}$ is connected but not strongly connected, then
$z(G) = 0$.
\end{corollary}
\begin{proof}
Given any stable graph $H\in \mathcal G$, if we apply Lemma \ref{covar} to each vertices of $H$, we get a sum of
semistable graphs of the same weight. Moreover, it is not difficult to see that if $H$ is strongly connected, all these
semistable graphs are also strongly connected. So we conclude the proof.
\end{proof}
\begin{example}
Let us illustrate the above proof by considering the following
semistalbe graph $G$ in Figure \ref{figzero}, which is only weakly
connected.
\begin{figure}[h]
$\xymatrix@C=3mm@R=7mm{
*+[o][F-]{1} \ar@/^0.5pc/[d]
& & &*+[o][F-]{1} \ar[dr]& \\
\circ\ar[urrr] \ar@/^0.5pc/[u] \ar[rr]
& & *+[o][F-]{1}\ar[ur] && *+[o][F-]{1} \ar[ll] } $
\caption{A semistable graph $G\in \mathcal G^{ss}(6)$}
\label{figzero}
\end{figure}
Let $C$ be its unique sink and $G'$ be its unique source. Then we
have $|{\rm Aut}(C)|=3$ and $|{\rm Aut}_f(C)|=1$. For all
$G$-configurations, we can check that the two sides of the equation
\eqref{eqkey} match, their values are listed in Table \ref{tb1}.
\begin{table}[h]
\caption{$G$-configurations}\label{tb1}
\begin{tabular}{|c|c|c|c|c|c|}
\hline \backslashbox{$L'$}{$L''$}
& $\emptyset$ & $[1]$ & $[1^2]$ & $[1^3]$ & $[3]$
\\
\hline $\emptyset$ & 1 & 3 & 3 & 1 & 1\\
\hline $[1]$ & 1 & 3 & 3 & 1 & 1\\
\hline $[2]$ & 1 & 3 & 3 & 1 & 1\\
\hline
\end{tabular}
\end{table}
\end{example}
The following theorem is called the coefficient theorem (see Theorem
1.2 in \cite{CDS}) from the spectral graph theory.
\begin{theorem} \label{cds}
Let $P_G (\lambda) = \lambda^n + c_1\lambda^{n-1} + \dots + c_n$
be
the characteristic polynomial of a digraph $G$ with $n$ vertices.
Then for each $i = 1,\dots,n$,
$$c_i = \sum_{L\in\mathcal L_i} (-1)^{p(L)}, $$ where $\mathcal L_i$ is the set of all
linear directed subgraphs $L$ of $G$ with exactly $i$ vertices;
$p(L)$ denotes the number of components of $L$.
\end{theorem}
\begin{proposition} \label{strong}
If $G=(V,E)\in \mathcal G^{ss}$ is strongly connected, then
\begin{equation} \label{eqstrong}
z(G) = (-1)^{|V|+1}\frac{\det(I-A)}{|{\rm Aut}(G)|}.
\end{equation}
\end{proposition}
\begin{proof}
Let $\mathcal L$ be the set of all linear directed subgraphs of $G$.
We have the natural action of ${\rm Aut}(G)$ on $\mathcal L$, given
by $h\cdot L=h(L)$ for $h\in {\rm Aut}(G), L\in\mathcal L$. From our
discussion at the end of Section \ref{secgraph}, the orbits of this
action are just $\mathcal L/\sim$ and the isotropy group at
$L\in\mathcal L$ is just ${\rm Aut}_G(L)$. By Theorem \ref{cds} and
Lemma \ref{single2}, we have
\begin{align*}
\det(I-A)=1+\sum_{i=1}^n c_i &=\sum_{L\in\mathcal
L/\sim}(-1)^{p(L)}|{\rm Aut}(G)\cdot
L|\\
&=|{\rm Aut}(G)|\sum_{L\in\mathcal L/\sim}\frac{(-1)^{p(L)}}{|{\rm
Aut}_G(L)|}\\
&=|{\rm Aut}(G)| (-1)^{|V|+1}z(G),
\end{align*}
which is just the equation \eqref{eqstrong}.
\end{proof}
\begin{corollary} \label{eigenvalue}
If $G \in \mathcal G^{ss}$ is strongly connected, then $z(G) = 0$ if
and only if $1$ is an eigenvalue of the characteristic polynomial of
$G$.
\end{corollary}
\begin{proof}
It follows from Proposition \ref{strong}.
\end{proof}
Propositions \ref{disjoint}, \ref{sink} and
\ref{strong} together imply Theorem \ref{main}.
\begin{proposition} \label{onevertex} If $G=\xymatrix{*+[o][F-]{k}}$, the digraph with
one vertex and $k\geq2$ self-loops, we have
\begin{equation}
z(G)=-\frac{k-1}{k!}.
\end{equation}
\end{proposition}
\begin{proof}
It follows from Lemma \ref{coef1} or Proposition \ref{strong}.
\end{proof}
\begin{proposition} \label{twovertex}
Let $G=\xymatrix{
*+[o][F-]{m} \ar@/^/[r]^i
&
*+[o][F-]{n} \ar@/^/[l]^j}
$ be a strongly connected stable graph with two vertices, namely
$ij\neq0$. Then we have
\begin{equation}
z(G) = \begin{cases} \dfrac{ij-(1-m)(1-n)}{2\cdot m!\, n!\, i!\, j!}
&\mbox{if } i=j \text{ and } m=n;
\\ \dfrac{ij-(1-m)(1-n)}{ m!\, n!\, i!\, j!} & \mbox{otherwise} .
\end{cases}
\end{equation}
\end{proposition}
\begin{proof}
The formula follows from Proposition \ref{strong}.
\end{proof}
\begin{example} \label{tyza2}
Fix a normal coordinate around $x\in M$. By Proposition
\ref{onevertex}, we get the well known
\begin{equation}\label{eqa1}
a_1=-\frac{1}{2}\left[\xymatrix{*+[o][F-]{2}}\right]=-\frac{1}{2}g_{i\bar
i j\bar j}=\frac{1}{2}R_{i\bar i j\bar j}=\frac{1}{2}\rho,
\end{equation}
where $(i,\bar i),\,(j, \bar j)$ are paired indices to be
contracted.
By Proposition \ref{onevertex}, Proposition \ref{twovertex} and
Proposition \ref{disjoint}, we have
\begin{equation}\label{eqa2}
a_2=-\frac{1}{3}\left[\xymatrix{*+[o][F-]{3}}\right]+\frac{1}{2}\left[\xymatrix{
*+[o][F-]{1} \ar@/^/[r]^1
&
*+[o][F-]{1} \ar@/^/[l]^1} \right ]+ \frac{3}{8}
\left[\xymatrix{ \circ \ar@/^/[r]^2 & \circ \ar@/^/[l]^2
}\right]+\frac{1}{8}\left[\xymatrix{*+[o][F-]{2}}\mid \xymatrix{*+[o][F-]{2}}\right].
\end{equation}
Written in terms of derivatives of the K\"ahler metric $g_{i\bar
j}$, we get
\begin{equation} \label{eqb11}
a_2=-\frac{1}{3}g_{i\bar i j\bar j k\bar k} +\frac{1}{2} g_{i\bar i
k\bar l}g_{j\bar j l\bar k}+\frac{3}{8}g_{i\bar j k\bar l}g_{j\bar i
l\bar k}+\frac{1}{8}g_{i\bar i j\bar j}g_{k\bar k l\bar l}.
\end{equation}
It is understood that $(i,\bar i),\,(j, \bar j),\,(k,\bar
k),\,(l,\bar l)$ are paired indices to be contracted.
Apply the operator $D$ (defined in Remark \ref{rm1}) to the
right-hand side of \eqref{eqb11} and use the identities
$$D(g_{i\bar j k\bar l})=-R_{i\bar j k\bar l},$$
\begin{align*}
D(g_{i\bar i j\bar j k\bar k})&=-R_{i\bar i j \bar j; k\bar
k}+R_{k\bar i s\bar j}R_{i\bar k j\bar s}+R_{j\bar j s\bar
i}R_{k\bar k i\bar s}+R_{i\bar i s\bar j}R_{k\bar k j\bar s}\\
&=-\Delta \rho+|R|^2+2|Ric|^2,
\end{align*}
we arrived at
\begin{equation}
a_2=\frac{1}{3}\Delta
\rho+\frac{1}{24}|R|^2-\frac{1}{6}|Ric|^2+\frac{1}{8}\rho^2,
\end{equation}
which is the same as the function $a_2(x)$ computed in \cite{Lu}. We
computed $a_3$ in Appendix \ref{apthree} and we can use tables at
Appendix \ref{apfour} to compute $a_4$.
\end{example}
\vskip 30pt
\section{Computations of $z(G)$}
In this section, we derive some explicit formulae of $z(G)$ .
\begin{figure}[h]$$
\xy/r2.7pc/: {*{\mathcal A_n}{}\xypolygon7{~><{} ~*{\circ}
~>>{_{2}}}}
\endxy \qquad \xy/r2.7pc/: {*{\mathcal B_n}{}\xypolygon7{~><{@{<->}} ~*{\circ} ~>>{_{1}^{1}}}}
\endxy \qquad \xy/r2.7pc/: {*{\mathcal C_n}{}\xypolygon7{~><{} ~*{\xybox{*{1}*\cir<2mm>{}}}
~>>{_{1}}}}
\endxy$$
\caption{Three types of stable graphs in $\mathcal G_{n,2n}$}
\label{figthree}
\end{figure}
\begin{proposition} Given $n\geq3$, for the three graphs in Figure \ref{figthree}, we have
\begin{align*}
z(\mathcal A_n)&=\frac{(-1)^n(2^n-1)}{2^n
n},\\
z(\mathcal
B_n)&=\frac{(-1)^n}{2n}+\sum_{i=0}^{[n/2]}(-1)^{n+i+1}\sum _{H\in
\mathscr C_{i,n-2i}} \frac{1}{{\rm Aut}(H)}\\&=\begin{cases}0, &
n\equiv 0\\(-1)^n/(2n), & n\equiv \pm 1\\3(-1)^n/(2n), & n\equiv \pm
2
\\2(-1)^n/n, &n\equiv 3 \end{cases}\mod\, 6,\\
z(\mathcal C_n)&=\frac{(-1)^n}{n}.
\end{align*}
Here $\mathscr C_{i,j}$ denote the set of $2$-colored undirected
cycle graphs with $i$ black vertices and $j$ white vertices and
${\rm Aut}(H)$ is the group of color-preserving automorphisms. Note the first equation for $z(\mathcal
B_n)$ holds only when $n\geq5$.
\end{proposition}
\begin{proof} Now we will compute $z(G)$ for $G=\mathcal A_n, \mathcal B_n, \mathcal C_n$.
Since these graphs are strongly connected, all $G$-configurations
are simply linear subgraphs $L$ of $G$. Two linear subgraphs $L_1$
and $L_2$ define the same $G$-configuration if there is an
automorphism $\varphi$ of $G$ such that $\varphi(L_1)=L_2$.
i) When $G=\mathcal A_n$, it is easy to see that there are only two
$\mathcal A_n$-configurations given by empty graph
$\emptyset$ and a $n$-cycle $[n]$, respectively. By \eqref{eqzstrong}, we have
\begin{align*}
z(\mathcal A_n)=\frac{(-1)^{n+1}}{|{\rm Aut}(\mathcal A_n)|}+\frac{(-1)^n}{|{\rm Aut}_{\mathcal A_n}([n])|}
&=\frac{(-1)^{n+1}}{2^n n}+\frac{(-1)^n}{n}\\
&=\frac{(-1)^n(2^n-1)}{2^n n}.
\end{align*}
ii) When $G=\mathcal B_n$, the contribution to $z(\mathcal B_n)$ by empty graph is
\begin{equation*}
\frac{(-1)^{n+1}}{{\rm Aut}(\mathcal B_n)}=\frac{(-1)^{n+1}}{2n},
\end{equation*}
where ${\rm Aut}(\mathcal B_n)$ is the dihedral group.
The contribution to $z(\mathcal B_n)$ by a $n$-cycle $[n]$ is
\begin{equation*}
\frac{(-1)^n}{|{\rm Aut}_{\mathcal B_n}([n])|}=\frac{(-1)^n}{n}.
\end{equation*}
All other $\mathcal B_n$-configurations are given by disjoint cycles of
length two
\begin{equation*}
[2^i], \quad 0\leq i\leq [n/2],
\end{equation*}
which are in one-to-one
correspondence with the set $\mathscr C_{i,n-2i}$ consisting of
$2$-colored undirected cycle graphs with $i$ black vertices and
$n-2i$ white vertices. Their contributions to $z(\mathcal B_n)$ are
equal to
\begin{equation*}
(-1)^{n+i+1}\sum _{H\in \mathscr C_{i,n-2i}} \frac{1}{{\rm Aut}(H)},
\end{equation*}
where ${\rm Aut}(H)$ is the group of color-preserving automorphism. Summing up, we get the first
equation for $z(\mathcal B_n)$. The second equation follows from a
computation of $\det(I-A(\mathcal B_n))$.
iii) When $G=\mathcal C_n$, we use $\det(I-A(\mathcal
C_n))=-1$.
\end{proof}
\begin{proposition} Let $n,m\geq2$.
\begin{enumerate}
\item[i)]
Let $K_n$ be the complete digraph with $n$ vertices and every pair
of vertices is an edge, including a loop at each vertex. Namely
every entry of the adjacency matrix of $K_n$ is $1$. Then
\begin{equation*}
z(K_n)=(-1)^{n}(n-1)/n!.
\end{equation*}
\item[ii)]
Let $D_n$ be the de Bruijn graph of degree $n$. The graph $D_n$ has
$2^{n-1}$ vertices, which are the sequences of $0$'s and $1$'s with
length $n-1$. There is an edge from $a_1a_2\dots a_{n-1}$ to
$b_1b_2\dots b_{n-1}$ if and only if $a_2a_3\dots
a_{n-1}=b_1b_2\dots b_{n-2}$. Then
\begin{equation*}
z(D_n)=1/2.
\end{equation*}
\item[iii)] Let $K_{m,n}(V_1,V_2)$ be the complete bipartite digraph with $|V_1|=m,|V_2|=n$. Every vertex
in $V_1$ has an arrow to every vertex in $V_2$ and vice versa. Then
$$z(K_{m,n})=\frac{mn-1}{(1+\delta_{m,n})m!n!},$$
where $\delta_{m,n}=0$ if $m\neq n$ and $\delta_{n,n}=1$.
\end{enumerate}
\end{proposition}
\begin{proof}
First note that $K_n\in\mathcal G(n^2-n)$, $D_n\in\mathcal
G(2^{n-1})$ and $K_{m,n}\in\mathcal G(2mn-m-n)$ all are strongly connected digraphs. It is well-known that
their characteristic polynomials are respectively
given by
\begin{gather}
\det(\lambda I-A(K_n))=\lambda^{n-1}(\lambda-n),\\
\det(\lambda I-A(D_n))=\lambda^{2^{n-1}-1}(\lambda-2),\\
\det(\lambda I-A(K_{m,n}))=(-1)^{m+n}\lambda^{m+n-2}(\lambda^2-mn).
\end{gather}
We also have $|{\rm Aut}(K_n)|=n!$, $|{\rm Aut}(D_n)|=2$ and $|{\rm Aut}(K_{m,n})|=(1+\delta_{m,n})m!n!$, we conclude
the proof by Proposition \ref{strong}.
\end{proof}
\begin{remark}
The computation of $z(G)$ is reduced to study the spectra and the
automorphism group of $G$. A detailed study of the spectral
properties of graphs can be found at \cite{CDS}, which also contains
a chapter on the relations between spectra of a graph and its
automorphism group.
\end{remark}
Let $U$ be the unit ball of $\mathbb C^N$ and $g_{i\bar j}$ is the
Bergman metric on it,
\begin{equation}
g_{i\bar j}(z)=\frac{(1-|z|^2)\delta_{ij}+\bar z_i
z_j}{(1-|z|^2)^2},
\end{equation}
we know that $U$ is a normal coordinate with center $0$ for
$g_{i\bar j}$ and the curvature is constant
\begin{equation} \label{equ1}
R_{i\bar j k\bar l}=-(g_{i\bar j}g_{k\bar l}+g_{i\bar l}g_{k\bar
j}).
\end{equation}
The weighted reproducing kernel for $(U,g_{i\bar j})$ is
\begin{equation}
K_\beta(z)=\frac{\Gamma(\beta)}{\pi^N\Gamma(\beta-N)}(1-|z|^2)^{-\beta}.
\end{equation}
We know that $K_\beta(z)$ has an asymptotic expansion \cite{Ber,
Eng}
\begin{equation}
K_\beta(z)=(\beta/\pi)^N e^{\beta\Phi(z)} \sum_{k=0}^\infty a_k(z)
\beta^{-k},
\end{equation}
where $a_k(z)$ is the $k$-th coefficient of Tian-Yau-Zelditch
expansion and the potential $\Phi(z)$ is given by $
\Phi(z)=-\log(1-|z|^2)$.
So in the case of unit ball equipped with the Bergman metric, $a_k$
equals the polynomial
\begin{equation}
P_k=\sum _{1\leq i_1< i_2<\dots < i_k \leq N} (-i_1)\cdots(-i_k).
\end{equation}
In particular,
$$P_1=-\frac{N^2}{2}-\frac{N}{2},\qquad P_2=\frac{N^4}{8}+\frac{N^3}{12}-\frac{N^2}{8}-\frac{N}{12}.$$
\begin{lemma} \label{bergball}
Let $g_{i\bar j}$ be the Bergman metric on the unit ball $U$ of
$\mathbb C^N$. Then $g_{i\bar j \alpha_1\alpha_2\dots\alpha_r}(0)$
is nonzero only if the number of barred and unbarred indices in
$\{\alpha_1,\alpha_2,\dots,\alpha_r\}$ are equal. In this case, we
have
\begin{equation}\label{equ2}
g_{i_1\bar j_1 i_2 \bar j_2\dots i_k \bar
j_k}(0)=(k-1)!\sum_{\sigma\in S_k} g_{i_1 \bar j_{\sigma(1)}}g_{i_2
\bar j_{\sigma(2)}}\dots g_{i_k \bar j_{\sigma(k)}}(0).
\end{equation}
\end{lemma}
\begin{proof} The first assertion can be proved by an induction. Let
us prove \eqref{equ2}. By \eqref{eqcur1} and \eqref{equ1}, on $U$ we
have
\begin{equation}\label{equ3}
g_{i_1\bar j_1 i_2\bar j_2}=g_{i_1\bar j_1}g_{i_2\bar
j_2}+g_{i_1\bar j_2}g_{i_2\bar j_1}+g^{m\bar n}g_{m\bar j_1\bar
j_2}g_{i_1\bar n i_2}.
\end{equation}
As we have discussed in Section \ref{sectensor}, both sides of the
above equation may be represented by trees with half-edges and
taking partial derivatives may be regarded as the action of
half-edges on trees.
Let us look at the coefficients of $g_{i_1 \bar j_{1}}g_{i_2 \bar
j_{2}}\dots g_{i_k \bar j_{k}}$ after taking partial derivatives
$\partial_{i_3\bar j_3\dots i_k\bar j_k}$ to the right-hand side of
\eqref{equ3}. In fact, all contributions come from taking paired
derivatives $\partial_{i_3\bar j_3}\dots \partial_{i_k\bar j_k}$
consecutively to the first term in the right-hand side of
\eqref{equ3}. The coefficient is easily seen to be $2\cdot 3 \cdots
(k-1)=(k-1)!$.
\end{proof}
A {\it cycle decomposition} of a digraph $G$ is a partition of the
edges of $G$ into edge-disjoint cycles (i.e. closed directed paths
having no common edge). Let $\mathscr C_G$ denote the set of all
cycle decomposition of $G$. It is well known that $G$ admits a cycle
decomposition if and only if it $\deg^+(v)=\deg^-(v)$ for each
vertex $v$ of $G$.
\begin{proposition}
Given $k\geq1$, we have
\begin{equation}
\sum_{G\in\mathcal G(k)} (-)^{n(G)}\frac{\det(A-I)}{|{\rm Aut}(G)|}
\prod_{v\in V}(\deg^+(v)-1)!\sum_{H\in \mathscr C_G} N^{p(H)} =P_k,
\end{equation}
where $v$ runs over all vertices of $G$, $n(G)$ is the number of
components of $G$ and $p(H)$ is the number of cycles in the
cycle decomposition $H\in \mathscr C_G$. The graph $G$ in the
left-hand side need only run over all balanced graphs.
\end{proposition}
\begin{proof} From $a_k=P_k$ and Lemma \ref{bergball}, we have
\begin{equation}
\sum_{G\in\mathcal G(k)} z(G)\prod_{v\in V}(\deg^+(v)-1)!\sum_{H\in
\mathscr C_G} N^{p(H)} =P_k.
\end{equation}
Note that if $G\in\mathcal G$ is a disjoint union of connected
subgraphs $G=G_1\cup\dots\cup G_n$, then by Theorem \ref{main} we
have
\begin{equation}
z(G)=\begin{cases}\frac{(-1)^n\det(A-I)}{|{\rm Aut}(G)|}, & \text{if all }
G_i \text{ are strongly connected};
\\
0,& \text{otherwise}.
\end{cases}
\end{equation}
So we get the desired formula.
\end{proof}
\begin{corollary}
The leading term of the polynomial $P_k$ is $ \frac{(-1)^k}{2^k
k!}N^{2k} $.
\end{corollary}
\begin{proof}
It follows from Proposition \ref{disjoint} and Proposition
\ref{onevertex}.
\end{proof}
Let $G$ be a digraph. An {\it Euler tour} is a directed closed path
in $G$ which visits each edge exactly once. As a generalization of
the ``Seven Bridges of K\"onigsberg'' problem, Euler showed that $G$
has an Euler tour if and only if $G$ is connected and
$\deg^+(v)=\deg^-(v)$ at every vertex $v$. We denote by
$\epsilon(G)$ the number of Euler tours in $G$ starting with a fixed edge.
\begin{corollary} \label{bern}
Given $k\geq1$, we have
\begin{equation} \label{eqeuler}
\sum_{G\in\mathcal G(k)} z(G)\cdot\epsilon(G)\cdot\prod_{v\in
V}(\deg^+(v)-1)!
=\frac{(-1)^{k+1}}{k}B_k,
\end{equation}
where $B_k$ is the $k$-th Bernoulli number: $B_0=1, B_1=-1/2,
B_2=1/6, B_3=0$.
\end{corollary}
\begin{proof} Using Barnes' asymptotic formula for $\Gamma$
functions, we have the following equality of power series (cf.
\cite{Eng})
\begin{equation}
\sum_{k=0}^\infty P_k x^k=\exp(q_1 x+q_2 x^2+\dots),
\end{equation}
where $q_1, q_2,\dots$ are polynomials in $N$,
\begin{equation}
q_j=\frac{1}{j(j+1)}\sum_{i=0}^j (-1)^{i+1}\binom{j+1}{i} B_i
N^{j+1-i}.
\end{equation}
Note the left-hand side of \eqref{eqeuler} is the coefficient of $N$
in $a_k$, which is equal to the coefficient of $N$ in $q_k$. The
latter is just $\frac{(-1)^{k+1}}{k}B_k$.
\end{proof}
It is interesting to note that the factor $\prod_{v\in
V}(\deg^+(v)-1)!$ also appears in the formula (cf. \cite[p.56]{Sta})
$$\epsilon(G)=\tau(G)\prod_{v\in
V}(\deg^+(v)-1)!,$$ where $\tau(G)$ is the number of oriented spanning
subtrees of $G$ with a fixed root. An oriented tree $T$ with a root $v$ means that
the underlying undirected graph is a tree and all arrows in $T$ point toward $v$.
Let us check \eqref{eqeuler}.
When $k=1$ and $k=2$, by \eqref{eqa1} and \eqref{eqa2}, the left-hand side of \eqref{eqeuler} respectively equals to
\begin{gather*}
-\frac{1}{2}\cdot1\cdot1=-\frac{1}{2}=B_1,\\
-\frac{1}{3}\cdot2\cdot2+\frac{1}{2}\cdot1\cdot1+\frac{3}{8}\cdot2\cdot1=-\frac{1}{12}=\frac{-B_2}{2}.
\end{gather*}
When $k=3$, by \eqref{eqa3} in Appendix \ref{apthree}, we have
\begin{gather*}
z_4\cdot1\cdot1+z_5\cdot2\cdot1+z_6\cdot1\cdot1+z_7\cdot3\cdot1+z_9\cdot2\cdot2\\
+z_{10}\cdot4\cdot2+z_{14}\cdot6\cdot6+z_{15}\cdot4\cdot1
=0=\frac{B_3}{3}.
\end{gather*}
\vskip 30pt
|
\section{Introduction}
According to our current understanding, the energy-matter content of the Universe is dominated by dark components: dark energy (DE) ($\sim 73\%$) and dark matter (DM) ($\sim 22\%$), with ordinary baryonic matter accounting for only $\sim 5\%$ of the total ($\sim$ critical) density \citep[e.g., see][]{2011ApJS..192...18K}. In contrast to the poorly understood DE -- the substance causing the Universe to expand in an accelerated fashion -- we have physically well-motivated models for the DM. Among those, the most promising scenario states that the DM of the Universe consists of thermal relic density of stable weakly interacting massive particles (WIMPs). It is quite miraculous that having particles with masses and annihilation cross sections set by the electroweak scale automatically provide the right DM density after freeze-out \citep{1996PhR...267..195J,2005PhR...405..279B}.
The WIMP hypothesis, along with its potentially observable phenomenology, has initiated strong effort in the particle- and astrophysics communities to try to find other than purely gravitational manifestations of DM. So far, we have good knowledge of DM only through its gravitational effects, starting from the scale of galaxies and galaxy clusters, up to the cosmologically largest observable scales \citep{1996PhR...267..195J,2005PhR...405..279B,2009arXiv0901.0632E}. However, as there is already an impressive list of ongoing and upcoming direct DM detection experiments along with various indirect means of detection \citep[see][for overview]{2010ARA&A..48..495F}, the hopes are very high that in the nearest future the mystery of DM might at last be solved. Indeed, the first signals from DM particles could potentially have already been detected: the (expected) annual modulation signal from DAMA/LIBRA \citep{2010EPJC...67...39B}, signals from the CDMS \citep{2011PhRvL.106m1302A} and CoGeNT \citep{2011PhRvL.106m1301A} nuclear recoil experiments, anomalies of the cosmic ray positrons as revealed by PAMELA satellite \citep{2009Natur.458..607A}, or positrons+electrons as obtained by the Fermi satellite \citep{2009PhRvL.102r1101A} and HESS atmospheric Cherenkov telescope \citep{2009A&A...508..561A}. While the cosmic ray positron anomaly can possibly be explained by TeV-scale DM \citep{2008PhRvD..78j3520B,2009PhLB..672..141B,2009NuPhB.813....1C,2009PhRvD..79a5014A,2009PhRvD..79h3528F}, the signal from CoGeNT calls for light WIMPs within the mass range $7-12$~GeV \citep{2011PhRvL.106m1301A}. Consistent analyses of combined data from CoGeNT and DAMA/LIBRA determine the light DM mass to be $6-8$~GeV \citep{2010PhRvD..82l3509H}. Although this mass range is probed by the CDMS \citep{2011PhRvL.106m1302A}, XENON10 \citep{2009PhRvD..80k5005A}, and XENON100 \citep{2010PhRvL.105m1302A,2011arXiv1103.0303X} experiments, interpretation of those results \citep{2010JCAP...02..014K,2010arXiv1011.5432S} requires an ability to reliably reconstruct nuclear recoils at very low energy \citep{2011PhRvD..83e5002S}, as well as precise knowledge of DM distribution and velocity in the local halo. Therefore the CoGeNT and DAMA/LIBRA hints of light DM cannot be ruled out unambiguously.
In addition, there is an independent positive claim of the existence of {\cal O}(10)~GeV mass DM. A recent study by \citet{2011arXiv1102.5095D} also suggests that annihilating DM with similarly low masses ($m_{{\rm DM}}=1-20$ GeV) may give a good match to the observed Fermi and WMAP haze \citep{2010ApJ...717..825D,2008ApJ...680.1222D}. However, all those claims depend strongly on the knowledge of the profile of the DM halo of our Galaxy and precise knowledge of local DM density and halo substructure.
Thus the several interesting claims of the existence of {\cal O}(10)~GeV mass DM call for model-independent tests of the light DM scenario. Since lower DM particle masses imply higher number densities ($n_{\rm DM}\propto \Omega_{\rm DM}h^2/m_{{\rm DM}}$), and as the energy input from annihilations scales as $\propto n_{{\rm DM}}^2$, one might expect strong constraints on annihilation cross section, which might possibly reach below the standard thermal production value of $\langle\sigma_A\upsilon\rangle\simeq 3\times10^{-26}$ cm$^3$s$^{-1}$ \citep{1996PhR...267..195J}. The constraints from gamma-ray measurements \citep{2009PhRvD..79d3507B,2009JCAP...03..009B,2009PhRvD..79h1303B,2009NuPhB.821..399C,2010NuPhB.831..178M,2010NuPhB.840..284C,2010JCAP...03..014P,2010JCAP...07..008H,2010PhRvD..82l3511B,2011JCAP...01..011A,2010PhRvD..82l3519V,2011PhRvD..83l3513Z} along with CMB bounds \citep{2009PhRvD..80b3505G,2009PhRvD..80d3526S,2009JCAP...10..009C,2009A&A...505..999H,2010PThPh.123..853K}, indicate that this might indeed be the case.
In this paper we investigate how well DM annihilation cross sections for WIMPs with masses $m_{{\rm DM}}=5-100$ GeV can be constrained with CMB measurements, in particular focusing on the lower end of this mass range. The main advantage of CMB over other indirect probes, like gamma ray and cosmic ray measurements, is that it is practically insensitive to the complications caused by the nonlinear evolution of the cosmic density field. As our analysis shows, for any realistic structure formation scenario the CMB bounds on annihilating DM arise solely around the redshifts of $z\sim 1000$, while the contribution from lower redshift cosmic structures is completely negligible.
CMB constraints on annihilating DM have been obtained in several earlier studies: e.g. \citet{2005PhRvD..72b3508P,2006MNRAS.369.1719M,2006PhRvD..74j3519Z,2009PhRvD..80b3505G,2009PhRvD..80d3526S,2009JCAP...10..009C,2009A&A...505..999H,2010PThPh.123..853K}. Even though most of these analyses have assumed a simple `on the spot' approximation for the energy deposition\footnote{I.e., energy input from DM annihilations is assumed to get instantaneously absorbed by the cosmic medium, with some efficiency factor $f$.}, more recent studies \citep{2009PhRvD..80d3526S,2009A&A...505..999H} followed the energy transport problem including various energy-loss mechanisms in a more realistic way. Compared to the analysis of \citet{2009PhRvD..80d3526S}, which partially relies on the previously derived `on the spot' results of \citet{2009PhRvD..80b3505G}, in this paper we perform a more elaborate treatment for the energy deposition joined to the analysis of the CMB data via Markov chain Monte Carlo calculations that incorporate the most recent WMAP likelihood code. Also, we make an attempt to unify the results from various annihilation channels and provide a simple and rather generic fitting formula for calculating CMB constraints on annihilation cross section $\langle\sigma_A\upsilon\rangle$ for a broad range of annihilating DM models.
Our paper is organized as follows. In Section 2 we give a brief description of the energy input from DM annihilation and provide a simple treatment for its propagation. The effect on CMB temperature and polarization fluctuations is investigated in Section 3. Section 4 presents our main results about current and future CMB constraints. Our summary is given in Section 5.
\section{Input signals and their propagation}\label{sec2}
As in \citet{2009NuPhB.813....1C,2011JCAP...03..051C}, we treat our input signals from DM annihilation in an as model independent a way as possible. In \citet{2011JCAP...03..051C} the two-particle annihilation channels to all Standard Model (SM) particles were considered: leptons, quarks, photons, gluons, weak-interaction gauge bosons, Higgs boson, and neutrinos. In addition, annihilations to four leptons via an intermediate new boson $V$ were considered. Such a treatment can be considered as model independent since realistic models can always be decomposed into these basic channels where the particular branching ratios between the channels are given by the underlying theoretical particle physics model.
\begin{figure}
\centering
\includegraphics[width=9.cm]{Figs/fig01.eps}
\caption{The redshift $z^{'}$ where the optical depth for photons reaches unity (i.e., $\tau(z,z^{'})=1$) for several `observer's redshifts': $z=0,10,500,1000$. Here the energy plotted is the photon energy at redshift $z$. The light gray region corresponds to the DM mass interval considered in this paper.}
\label{fig1}
\end{figure}
Since in our work we focus on DM particle masses below $100$ GeV, out of all of the above channels the following remain: DM DM $\rightarrow$ SM $\overline{{\rm SM}}$, where SM$=\{e,\mu,\tau,q,c,b,\gamma,g$\}; plus 4-lepton channels via $V$. Here $q$ denotes the light quarks $u$, $d$, and $s$. Because the masses of interest in this work are mostly below the masses of the electroweak gauge bosons, $W^{\pm}$ and $Z$, those channels are left out. For the same reason we also do not need to distinguish between left- and right-handed particles. Also, the neutrino channels in this case provide only trivial output; i.e. $100\%$ of the energy is carried away by neutrinos, which escape freely at the redshifts of interest, so are not treated any further. Even though the channels $\gamma$ and $g$ are included in our model-independent approach, these are strongly suppressed for realistic models since DM should not carry color or interact electromagnetically.
For all channels, the spectra of the emerging stable particles, $e$, $p$, $\gamma$, $\nu$, after treatment of several decays, parton showers, and hadronization were calculated using PYTHIA Monte Carlo\footnote{\url{http://home.thep.lu.se/~torbjorn/Pythia.html}} \citep{2008CoPhC.178..852S}. All input spectra are downloadable from \url{http://www.marcocirelli.net/PPPC4DMID.html}. For more details, and in particular for a discussion on the level of possible uncertainties, we refer the reader to \citet{2011JCAP...03..051C}.
Among the stable output particles, neutrinos propagate freely at the redshifts of interest, while $e^{\pm}$ immediately interact with the ubiquitous CMB photons and upscatter those to the gamma-ray energy range via the inverse Compton (IC) mechanism. The total output energy in hadrons ($p$ and $d$) is typically quite negligible. Only in gluon and quark channels does it reach up to $\sim 15\%$, and that only for the highest DM particle masses considered in this paper. We therefore did not model this component in detail. However, in Sect.~\ref{sec:protons} we give some arguments that at the redshifts of interest, i.e. $z\sim 1000$, probably the majority of energy released in $p$ and $d$ channels is directly converted into heat by the cosmic medium. Thus, we only need a detailed treatment for the photons: (A) the energetic ones originating directly from the annihilation event (prompt photons), (B) the softer IC photons created by $e^{\pm}$ upscattering CMB photons.
The processes and the corresponding cross sections relevant to the propagation of photons through the cosmic medium were taken from \citet{1989ApJ...344..551Z}. Starting from the lowest of energies these include (i) photoionization, (ii) Compton losses (on both bound and free electrons), (iii) pair production on matter, (iv) photon-photon scattering, and (v) pair production on ambient photon fields.
In Fig.~\ref{fig1} we show the redshifts $z^{'}$ where the optical depth $\tau(z,z^{'})=1$ for various redshifts of the observer: $z=0,10,500,1000$. As can be seen, for intermediate photon energies (depending on the redshift) there is a well-known X-ray/gamma-ray energy window where the photons can propagate freely over cosmologically large distances \citep[e.g., see][]{2004PhRvD..70d3502C}. For the cosmological radiation transfer it is crucial that this `transparency window' is properly modeled. Once the photon gets outside of this window (we take it to happen after the first interaction) we assume that the following cascade will be locally absorbed in a very short time. Moreover, the fractions $(1-f_{{\rm ion}})/3$ and $(1+2f_{{\rm ion}})/3$ of the total absorbed energy, are assumed to be going for ionization and heating, respectively. Here $f_{{\rm ion}}$ is the fraction of ionized hydrogen atoms, and a similar expression for helium can be used. Excitations of atoms are neglected. This approximation, motivated by the work of \citet{1985ApJ...298..268S}, has been widely used in several subsequent papers, e.g. \citet{2004PhRvD..70d3502C,2005PhRvD..72b3508P,2006MNRAS.369.1719M,2006PhRvD..74j3519Z,2009PhRvD..80d3529N}. However, it is clear that these simple expressions only provide a rough estimate for the correct energy deposition efficiencies. As mentioned by \citet{2010MNRAS.402.1195C}, the precise redshift dependence of the efficiency factors and the fraction of energy that goes into excitations of hydrogen and helium atoms need more careful consideration of the radiative transfer processes, including secondary low-energy photons and their feedback. These extra photons have the potential of further delaying recombination, hence affect the last scattering surface and CMB anisotropies \citep{2000ApJ...539L...1P}. We leave a more detailed investigation of these ambiguities to a future paper; however, later on in Section~\ref{sec4} we briefly comment on how much the omission of excitations using the rough prescription of \citet{2004PhRvD..70d3502C} changes our final results.
In Fig.~\ref{fig1} we also see that at high redshifts, as the density of the environment becomes much higher, the X-ray/gamma-ray transparency window starts to close. Thus at sufficiently high $z$, we would expect all the produced annihilation energy (excluding the energy stored in neutrinos, since these can freely leak out) to be absorbed locally. In the following we call the ratio of the locally produced to the locally absorbed energy the $f$-parameter. At high redshifts (but well after the neutrino decoupling) we expect the $f$-parameter to asymptote to the value given by $(1-f_{\nu})$, where $f_{\nu}$ is the fraction of energy carried away by neutrinos. An example for the $f$-parameters in the case of $\mu$ annihilation channel are shown in the upper lefthand panel of Fig.~\ref{fig2}. Since $\sim 60\%$ of the energy is carried away by neutrinos in the $\mu$-channel the expected asymptotic high-redshift $f$-parameter should be $\sim 0.4$, which is indeed the case. With this in mind, we see that robust model-independent results from the CMB analyses can be obtained for the DM masses below $\sim 100$~GeV. This is the reason we concentrate on light WIMPS in this work. For heavier WIMPs, the computation depends on more complicated details for energy absorption.
\section{Effect on CMB}
\begin{figure*}[ht!]
\centering
\includegraphics[width=18.5cm]{Figs/fig02.eps}
\caption{{\bf Lefthand column} from top to bottom: {\bf (i)} $f$-parameters for the $\mu$-channel assuming $m_{{\rm DM}}=10$ GeV (dark solid lines) and $m_{{\rm DM}}=100$ GeV (light solid lines). In both cases, the lowest curve out of the triple of lines corresponds to only the smooth background, the middle one includes halos with a lower mass cutoff of $10^{-6}M_{\odot}$, while the top one has $10^{-9}M_{\odot}$. The dotted lines represent high-redshift fits (valid for $z\gtrsim 170$) as given by \citet{2009PhRvD..80d3526S}. The dashed line shows the CMB visibility function. Here $zg(z){\rm d}\ln z$ gives the probability that the CMB photon last scattered in the logarithmic redshift interval ${\rm d}\ln z$ centered on redshift $z$. {\bf (ii)} Fraction of free electrons as a function of redshift for $m_{{\rm DM}}=10$ GeV and $100$ GeV, along with two values for $\langle\sigma_A\upsilon\rangle/(\langle\sigma_A\upsilon\rangle_{{\rm std}}m_{{\rm DM}}[{\rm GeV}])$: 1 and 10. Here $\langle\sigma_A\upsilon\rangle_{{\rm std}}=3\times10^{-26}$ cm$^{3}$s$^{-1}$ is the standard thermal cross section. For the meaning of each line, see the description given in the legend. The lowest dotted line represents the standard $\Lambda$CDM case with no additional energy input from the annihilating DM. {\bf (iii)} Matter temperature as a function of redshift. The models shown are exactly the same as in the panel above. The long dashed line compares the temperature of CMB. {\bf (iv)} The CMB visibility function $g(z)$ times redshift $z$ for the same models as shown in the above two panels. {\bf Righthand column} from top to bottom: {\bf (i)} Angular power spectra of CMB temperature fluctuations for the same models as already given above. {\bf (ii)} Temperature and E-mode polarization cross-spectra. {\bf (iii)} E-mode polarization spectra. In all of the righthand panels the points with errorbars show the 7-year measurements by the WMAP space mission.}
\label{fig2}
\end{figure*}
Because we are able to calculate $f$-parameters, we can go on to calculate the effect on cosmological recombination. To this end we modify the cosmological recombination code RECFAST \citep{1999ApJ...523L...1S} along the lines presented in \citet{2005PhRvD..72b3508P}. For more details see also \citet{2009A&A...505..999H}. Although more advanced cosmological recombination codes have recently been released\footnote{{\sc CosmoRec} (\url{www.Chluba.de/CosmoRec}) already provides a simple module that accounts for the effect of DM annihilation or more general energy injection, but this module is still being validated.} \citep{2010MNRAS.tmp.1876C, 2011PhRvD..83d3513A}, at this point we do not include any of the recently discuss corrections to the cosmological recombination process arising from detailed radiative transfer and atomic physics \citep[see][and references therein for overview]{2009ApJS..181..627F, 2010MNRAS.403..439R}.
The main effect of DM annihilations is a delay in recombination at $z\sim 1100$ and an increase in the low-redshift freeze-out tail. This changes the position and width of the last scattering surface and thus affects the CMB temperature and polarization anisotropies \citep[see, for example,][]{2004PhRvD..70d3502C, 2005PhRvD..72b3508P}. Since the input annihilation power scales as $\propto \rho_{DM}^2$, it is clear that structure formation leads to a significant boost over the average density squared $\bar{\rho}_{DM}^2$, i.e. $\langle \rho_{DM}^2(z)\rangle=B(z)\bar{\rho}_{DM}^2(z)$. The details of how we calculated the structure boost factors $B(z)$ can be found in \citet{2009A&A...505..999H}. The onset of structure formation at $z\sim 100$ is clearly visible in the upper lefthand panel of Fig.~\ref{fig2}. Here the three dark (light) solid curves, representing the results for the $\mu$ annihilation channel, correspond to $m_{{\rm DM}}=10$ GeV ($100$ GeV). From the bottom to top, the trio of lines in each case represent: (i) no structure formation, i.e. only a smooth background, (ii) including structures with the lower halo mass cutoff of $10^{-6}M_{\odot}$, (iii) structures with the cutoff of $10^{-9}M_{\odot}$. The halo mass function was assumed to have an analytic Sheth-Tormen form \citep{1999MNRAS.308..119S} and the halo mass-concentration relation followed \citet{2008MNRAS.391.1940M} description. For the full details of this calculations, again see \citet{2009A&A...505..999H}. The dotted lines represent the high-redshift fits (valid for $z\gtrsim 170$) for the $f$-parameters as given by \citet{2009PhRvD..80d3526S}. After adjusting slightly the high-redshift normalizations we see that the shape of their $f$-parameters at high redshifts agrees remarkably well with our results, even though our treatment is somewhat more simplified. The sharply peaked dashed curve at $z\sim 1000$ in Fig.~\ref{fig2} shows the Thomson visibility function \citep{1970Ap&SS...7....3S}. More specifically, the quantity $\mathcal{V}\sim z\, g(z){\rm d}\ln z$ gives the probability that the CMB photon last scattered in the logarithmic redshift interval ${\rm d}\ln z$ centered on redshift $z$. For the CMB calculations the values of the $f$-parameter matter only for the redshift range over which the visibility function is large, and in this range our results agree with the \citet{2009PhRvD..80d3526S} calculations to better than $5\%$ accuracy. We also tested the results for the $\tau$-channel and found very good agreement. Thus, our somewhat more simplistic calculation of the $f$-parameters compared to \citet{2009PhRvD..80d3526S} seems to be justified.
The other lefthand panels of Fig.~\ref{fig2} (from top to bottom) show the fraction of free electrons, matter temperature, and the Thomson visibility function as a function of redshift for two DM particle masses, $10$ GeV and $100$ GeV, along with annihilation cross sections $\{10,100\}$ times, and $\{100,1000\}$ times the standard thermal cross section $\langle\sigma_A\upsilon\rangle_{{\rm std}}=3\times10^{-26}$ cm$^{3}$s$^{-1}$. See the legend for the line definitions. The dotted lines show the standard behavior of ionization fraction, matter temperature and visibility function without any additional energy input from DM annihilation. In the panel for the matter temperature we also show the line for the temperature of the CMB.
Because the annihilation power scales in proportion to $\langle\sigma_A\upsilon\rangle n_{{\rm DM}}^2m_{{\rm DM}}\propto \langle\sigma_A\upsilon\rangle/m_{{\rm DM}}$, we would expect the case with $m_{{\rm DM}}=10$ GeV, and $\langle\sigma_A\upsilon\rangle=10\langle\sigma_A\upsilon\rangle_{\rm std}$ ($100\langle\sigma_A\upsilon\rangle_{\rm std}$) give comparable results to the case $m_{{\rm DM}}=100$ GeV and $\langle\sigma_A\upsilon\rangle=100\langle\sigma_A\upsilon\rangle_{\rm std}$ ($1000\langle\sigma_A\upsilon\rangle_{\rm std}$), as long as the $f$-parameter does not depend strongly on $m_{{\rm DM}}$. This is indeed approximately so, as can be seen from Fig.~\ref{fig2}. Even though the lines for the ionization fraction, matter temperature, and CMB visibility function are easily separated at lower redshifts, this simple scaling with $\langle\sigma_A\upsilon\rangle/m_{{\rm DM}}$ seems to hold very well near the peak of the visibility function, and so one would not expect to be able to clearly distinguish the models with the same value of $\langle\sigma_A\upsilon\rangle/m_{{\rm DM}}$ via CMB measurements alone\footnote{To achieve better distinction between various models, one might try to use the differences in the behavior of the low-$z$ matter temperature and its observable consequence on the 21 cm transition measurements of the neutral hydrogen. The effect of annihilating DM on the 21 cm signal has been discussed e.g. in \citet{2006MNRAS.369.1719M,2006PhRvD..74j3502F,2007MNRAS.377..245V,2009PhRvD..80d3529N}.}. Also, the additional annihilation boost from the structure formation is not expected to influence the CMB measurements. Only in the very extreme case where instead of \citet{2008MNRAS.391.1940M} mass-concentration relation we use a simple power-law extrapolation down to the very low halo masses one is able to cause an additional significantly high low-redshift peak in the visibility function. However, for any realistic mass-concentration relation, along with annihilation cross sections $\langle\sigma_A\upsilon\rangle$ that do not violate the CMB data, the contribution from the structure formation to the CMB signal is completely negligible. Also, the contribution to the low-redshift ionization fraction is very mild, and thus the annihilating DM models which are compatible with CMB measurements could only play a marginal role in reionizing the low-$z$ Universe. These results support our similar findings previously reported in \citet{2009A&A...505..999H}.
In the righthand panels of Fig.~\ref{fig2}, we show the temperature--temperature (TT), temperature--E-mode polarization (TE), and E-mode--E-mode (EE) angular power spectra. The points with error bars show the WMAP seven-year measurements \citep{2011ApJS..192...16L}. The dotted lines in all of the panels show theoretical predictions for the concordance $\Lambda$CDM model \citep{2011ApJS..192...18K}. As we might already expect, remembering the visibility function behavior as shown in the lowest lefthand panel, the models with the same values for $\langle\sigma_A\upsilon\rangle/m_{{\rm DM}}$ are hardly distinguishable. This is because the redshift dependence of the $f$-parameter over the width of the visibility function is small, and so only the average value around $z\sim 1100$ really matters, in agreement with the statements of \citet{2009PhRvD..80d3526S}.
As a final note in this section we point out that, if one increases $\langle\sigma_A\upsilon\rangle$ to high enough values and adds the structure formation boost, so that the low-redshift ionization fraction starts to get significantly approach one, the original RECFAST code, along with its DVERK solver, is not properly able to deal with the underlying set of stiff ODEs and simply breaks down. To be able to numerically treat these cases, one can use the significantly improved {\sc Recfast++} code\footnote{{\sc Recfast++} is part of more advanced recombination code {\sc CosmoRec} and can be downloaded from \url{www.Chluba.de/CosmoRec}.} \citep{2010MNRAS.tmp.1876C}, along with its stiff ODE solver. However, in our calculations the original RECFAST works fine, since the typically allowed values of $\langle\sigma_A\upsilon\rangle/m_{{\rm DM}}$ compatible with the CMB data are low enough.
\section{CMB constraints}\label{sec4}
\begin{figure*}[ht!]
\centering
\includegraphics[width=18.cm]{Figs/fig03.eps}
\caption{The WMAP7 parameter constraints for the $\mu$ annihilation channel. Along with two annihilation parameters $\log_{10}(m_{{\rm DM}})$ and $\langle\sigma_A\upsilon\rangle/m_{{\rm DM}}$, we also show the constraints on the other 6-parameter $\Lambda$CDM background model: $\Omega_{\rm B}h^2$, $\Omega_{\rm DM}h^2$, $H_0$, $\tau$, $n_{\rm S}$, $\log_{10}(A_{\rm S})$. From the inside out the colored 2D areas show the $1$-sigma and $2$-sigma regions after marginalization over the other parameters, and the dashed lines show the same regions for the basic 6-parameter $\Lambda$CDM without annihilating DM. The topmost panels in each column plot the marginalized 1D probability distributions for all of the parameters.}
\label{fig3}
\end{figure*}
Now we have all the ingredients available to calculate current and future CMB constraints on annihilating DM. We notice that the CMB constraints for the decaying DM are not competitive (e.g. when compared to the bounds available from the low-redshift gamma-ray measurements) due to $\propto \rho_{DM}$ scaling when compared to the $\propto \rho_{DM}^2$ scaling for the annihilating DM, and thus are not being discussed in this work.
For the currently existing CMB data, we used the latest power spectra, i.e. seven-year spectra, from the WMAP space mission \citep{2011ApJS..192...16L}. In reality the current CMB bounds can be somewhat tightened if one includes results from other smaller scale CMB measurements, in particular the measurements of the temperature power spectrum from ACBAR \citep{2009ApJ...694.1200R} and temperature and polarization spectra from QUaD \citep{2009ApJ...692.1247P}. In addition, data from ACT \citep{2011ApJ...729...62D} and SPT \citep{2010ApJ...719.1045L} could be added to the analysis, providing additional leverage on small scales\footnote{Since part of the effect of DM annihilation is degenerate with the effect of changing the spectral index of scalar perturbations, $n_{\rm S}$, small-scale CMB measurements help break this degeneracy.}. However, to keep our analysis clearer we decided to restrict ourselves to the WMAP7 data only. Regarding future CMB results we made predictions for the bounds available from the currently ongoing Planck mission, and also for the idealized noise-free experiment up to the multipole of $\ell_{\max}=2000$. As in \citet{2008arXiv0811.3918Z}, for Planck we assume $80\%$ sky coverage, beam size $\theta_{\rm FWHM}=7$ arcmin, and noisebias $N_{\ell}^{\rm TT}\simeq 10^{-4}$ $\mu$K$^2$ and $N_{\ell}^{\rm EE}\simeq 3.5\times10^{-4}$ $\mu$K$^2$ for TT and EE, respectively. For the ideal noise-free experiment, we assume a full sky coverage and the underlying uncertainties of the CMB fluctuations are solely due to a finite number of available fluctuation modes on the sky, i.e. due to cosmic variance, and so we call this type of idealized experiment `cosmic variance limited' (CVL) in the following. First steps towards CVL CMB measurements in both temperature and polarization down to small scales will become available from a combinations of {\sc SPTpol}\footnote{\url{http://pole.uchicago.edu/}} \citep{SPTpol} and {\sc ACTPol}\footnote{\url{http://www.physics.princeton.edu/act/}} \citep{ACTPol} along with Planck data, so that the CVL case discussed here provides a good guideline to what could become possible in the near future.
We generated the synthetic data for the Planck and CVL experiments as described in \url{http://cosmocoffee.info/}\footnote{See also \url{http://lpsc.in2p3.fr/perotto/} in case CMB lensing is needed.}. Along with WMAP7 likelihood code\footnote{\url{http://lambda.gsfc.nasa.gov/product/map/current/likelihood_info.cfm}}, they will be used in our Markov chain Monte Carlo (MCMC) parameter estimation. For the MCMC engine we used the publicly available CosmoMC tools\footnote{\url{http://cosmologist.info/cosmomc/}}\citep{2002PhRvD..66j3511L}, where the standard recombination modules were modified in order to allow additional energy input from annihilating DM. Our baseline concordance $\Lambda$CDM model \citep{2011ApJS..192...18K} is parameterized by six free parameters: $\Omega_{\rm B}h^2$, $\Omega_{\rm DM}h^2$, $H_0$, $\tau$, $n_{\rm S}$, $\log_{10}(A_{\rm S})$, where $\Omega_{\rm B}$ and $\Omega_{\rm DM}$ are density parameters for baryons and DM, $H_0=100\cdot h$ km/s/Mpc is the Hubble constant, $\tau$ is the reionization optical depth, and $A_{\rm S}$ and $n_{\rm S}$ are the amplitude and spectral index for adiabatic scalar perturbations, respectively. This minimal six-parameter set is extended by two additional parameters describing the annihilating DM: its mass $m_{{\rm DM}}$, and thermally averaged annihilation cross section $\langle\sigma_A\upsilon\rangle$. Since MCMC will give better results if model parameters are as uncorrelated as possible, we used the parameter pair $\{m_{{\rm DM}},\langle\sigma_A\upsilon\rangle/m_{{\rm DM}}\}$ in place of $\{m_{{\rm DM}},\langle\sigma_A\upsilon\rangle\}$.
As an example, we show the WMAP7 constraints on all of the eight parameters, assuming $\mu$ annihilation channel on Fig.~\ref{fig3}. With dashed lines we show parameter constraints for the six-parameter $\Lambda$CDM model without annihilating DM. The colored 2D areas in all of the panels display 1-sigma and 2-sigma regions marginalized over all of the other parameters. For all of the parameters the topmost panels in each column plot the marginalized 1D probability distributions. It is reassuring to see that after introducing two additional parameters, all the previous six parameters, albeit with small shifts, are still as precisely determined. The strongest shifts compared to the baseline six-parameter model are quite understandably seen for parameters $A_{\rm S}$ and $n_{\rm S}$, which both lead to the rise of the CMB power spectra if increased, and thus counterbalance the damping of the spectrum caused by the extra scattering off the additional free electrons created by the energy input from the DM annihilation.
Here it is important to note that these small shifts towards higher values of $A_{\rm S}$ and $n_{\rm S}$ are directed in the same direction as recently discussed recombination corrections from previously neglected standard physics processes \citep{2010MNRAS.403..439R, 2011arXiv1102.3683S}. Although biases in $A_{\rm S}$ and $n_{\rm S}$ arising from recombination physics are not significant for WMAP7 \citep{2011arXiv1102.3683S}, parameters derived from Planck data will be affected by several standard deviations if recombination corrections are neglected. This indicates that precise constraints on models with annihilating DM should probably account for these recombination corrections simultaneously.
\begin{figure}
\centering
\includegraphics[width=9.cm]{Figs/fig04.eps}
\caption{{\bf Upper panel:} $f$-parameters for all of the channels considered in this work and for several DM particle masses in the interval $m_{{\rm DM}}=5-100$ GeV. For all of the cases, the asymptotic high-redshift $f$-parameter values have been renormalized to be equal to one. {\bf Lower panel:} The same as above, after dividing with a `typical shape', $(f_{\max}(z)+f_{\min}(z))/2$, of the $f$-parameter curve.}
\label{fig4}
\end{figure}
In Fig. \ref{fig3} we only present the constraints for one particular channel. However, we would like to provide easily usable recipes to calculate bounds for all of the channels described in Section~\ref{sec2}. As seen before, even though the $f$-parameters vary somewhat if one changes the mass from $10$ GeV to $100$ GeV, the CMB spectra as shown in Fig.~\ref{fig2} were fairly insensitive to this change as long as the ratio $\langle\sigma_A\upsilon\rangle/m_{{\rm DM}}$ was kept the same. In the upper panel of Fig.~\ref{fig4} we plot the $f$-parameters for all of the channels and for several masses in the range $m_{{\rm DM}}=5-100$ GeV. Here we have renormalized the high-$z$ values to be equal to one. In the redshift range where the visibility function peaks (see the lowest lefthand panel of Fig.~\ref{fig2}), the variation among different DM models is not too large. To see this better, in the lower panel of Fig.~\ref{fig4} we have divided the above lines by a `typical shape' of the $f$-parameter given by $(f_{\max}(z)+f_{\min}(z))/2$, where $f_{\max}(z)$ and $f_{\min}(z)$ are the extremal values of $f(z)$ for a particular $z$. We see that, at most relevant redshifts, the variation among all the different models is around $15\%$. This indeed suggests that it might be possible to provide a unified approximate scheme for calculating the bounds for all of the annihilation channels.
To bracket all the possibilities, we calculated the bounds assuming the $f$-parameter to have a typical shape $(f_{\max}(z)+f_{\min}(z))/2$ along with two extremal values $f_{\min}(z)$ and $f_{\max}(z)$. Of course, instead of two annihilation parameters we then only have to deal with one additional parameter: $\langle\sigma_A\upsilon\rangle/m_{{\rm DM}}$. We took the asymptotic values for the $f$-parameter to have values of $0.25,0.50,0.75$, and $1.00$. This asymptotic value, as explained above, is simply $(1-f_{\nu})$, where $f_{\nu}$ is the fraction of the total energy carried away by neutrinos. It turned out that the resulting 1-sigma and 2-sigma constraints on $(1-f_{\nu})\langle\sigma_A\upsilon\rangle/m_{{\rm DM}}$ for the above four values of $(1-f_{\nu})$ are the same with $(1-2)\%$ accuracy. Our final results (averaged over these small variations due to changing $(1-f_{\nu})$) can thus be given in the form
\begin{equation}\label{eq1}
(1-f_{\nu})\frac{\langle\sigma_A\upsilon\rangle\,\left[3\times10^{-26}\,{\rm cm}^3{\rm s}^{-1}\right]}{m_{{\rm DM}}\,\left[{\rm GeV}\right]}< r\,,
\end{equation}
where the values of $r$ for 1-sigma and 2-sigma bounds are given in Table~\ref{tab1}. One can see that the approximate bounds we provide are typically accurate at about the $20-30\%$ level. We tested that, indeed, the ranges of $r$ given in Table~\ref{tab1} fully cover the values of $r$ for all of the channels considered in this work. Two examples for the case of 1-sigma upper bounds are shown in Fig.~\ref{fig5}. Here the upper panel corresponds to the $\mu$-channel ($f_{\nu}\simeq 0.61$) and the lower one to the $e$-channel ($f_{\nu}\simeq 0.02$). We chose the above two channels because these are the two extreme cases among all of the channels treated in this paper. Consequently, they bracket the expected results for any realistic annihilating DM, given as a superposition over the basis channels. The solid lines show the bounds calculated directly through the MCMC analysis, while the shaded regions represent the ranges as obtained from Eq. (\ref{eq1}) and Table~\ref{tab1}. Indeed, the solid lines are fully covered by the shaded regions, as it should be. The vertical gray stripe indicates the range of WIMP masses ($m_{{\rm DM}}=6-8$ GeV) that provide a good fit to CoGeNT and DAMA/LIBRA data \citep{2010PhRvD..82l3509H}. The vertical dotted line marks the lowest DM particle mass $5$~GeV used in our PYTHIA simulations. This cut-off is not physical, but occurs because PYTHIA does not work below that energy. Therefore, the extrapolations of our results are shown below 5~GeV DM mass. Note that directly calculated lines for the $\mu$- and $e$-channel have slightly steeper slopes than given by the shaded regions, which increase as $\propto m_{{\rm DM}}$. As the $f$-parameters generally fall off more slowly for lower $m_{{\rm DM}}$, this behavior is also typical of the other channels. Thus, for lower $m_{{\rm DM}}$ values, one should actually get slightly stronger bounds than calculated directly from Eq. (\ref{eq1}), and so our extrapolations shown in Fig.~\ref{fig5} are somewhat conservative.
\begin{table}
\centering
\caption{$1$-sigma and $2$-sigma values of $r$ in Eq. (\ref{eq1}) and their uncertainties for WMAP, Planck, and CVL CMB experiments.}
\label{tab1}
\begin{tabular}{l|l|l|}
& \multicolumn{2}{c|}{$r$} \\
\cline{2-3}
& {\footnotesize 1-sigma} & {\footnotesize 2-sigma} \\
\hline
WMAP & ${\bf {\scriptstyle 0.073}^{{\scriptscriptstyle +0.021}}_{{\scriptscriptstyle -0.013}}}$ & ${\bf {\scriptstyle 0.191}^{{\scriptscriptstyle +0.066}}_{{\scriptscriptstyle -0.031}}}$ \\
Planck & ${\bf {\scriptstyle 0.0160}^{{\scriptscriptstyle +0.0055}}_{{\scriptscriptstyle -0.0022}}}$ & ${\bf{\scriptstyle 0.0326}^{{\scriptscriptstyle +0.0096}}_{{\scriptscriptstyle -0.0055}}}$ \\
CVL & ${\bf {\scriptstyle 0.0071}^{{\scriptscriptstyle +0.0020}}_{{\scriptscriptstyle -0.0013}}}$ & ${\bf {\scriptstyle 0.0137}^{{\scriptscriptstyle +0.0031}}_{{\scriptscriptstyle -0.0032}}}$
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[width=9.cm]{Figs/fig05.eps}
\caption{WMAP, Planck, and CVL 1-sigma constraints on the $\langle\sigma_A\upsilon\rangle-m_{{\rm DM}}$ plane for $\mu$ (upper panel) and $e$ (lower panel) annihilation channels. The solid lines show the upper bounds on annihilation cross section as determined directly through full MCMC calculations. The shaded regions around solid lines show the results from the simple recipe of Eq. (\ref{eq1}) with values of $r$ taken from Table~\ref{tab1}. The vertical gray stripe shows the range of WIMP masses ($m_{{\rm DM}}=6-8$ GeV) that provide a good fit to CoGeNT and DAMA/LIBRA data \citep{2010PhRvD..82l3509H}. The vertical dotted line marks the lowest DM particle mass $5$ GeV available for PYTHIA simulations. Extrapolations are shown below that value.}
\label{fig5}
\end{figure}
We see that for WMAP7, depending on the annihilation model, the limiting DM particle mass below which the upper bound on the annihilation cross section drops below the standard thermal production value is in the range $4.5-10$ GeV, while the corresponding numbers reachable for the Planck and CVL experiment are $19-43$ GeV and $(45-100)$ GeV, respectively. Since the $\mu$-channel represents our most conservative case, one can instead say that, according to currently available CMB data, the annihilation cross section should be below the standard value of $3\times10^{-26}$ cm$^3$s$^{-1}$ as long as $m_{{\rm DM}}\lesssim 5$ GeV. \footnote{The numbers given here correspond to the 1-sigma upper bounds.} Thus, for the CoGeNT and DAMA/LIBRA best-fit mass region $6-8$ GeV, the standard thermal production cross section is still compatible with the CMB measurements. This could quite possibly be changed soon as Planck results become available. Of course one should keep in mind that the cross section bounds given here directly apply to redshifts of $z\sim 1000$, and if the cross section depends on velocity (as in the case of P-wave annihilation or Sommerfeld-enhanced scenario), one should be careful in converting these numbers to the values relevant at $z=0$. Also, note that the standard annihilation cross section $3\times10^{-26}$ cm$^3$s$^{-1}$, with respect to what we are comparing our CMB bounds to, provides the desired thermal relic density (i.e., $\Omega_{\rm DM}\sim 0.3$) only if WIMPs are annihilating through S-wave processes.
Except for a typical assumption of S-wave annihilation our results are largely model-independent, so all the particle-physics scenarios motivated by DAMA/LIBRA and CoGeNT results \citep{2010PhRvD..81k5005F,2010PhRvD..82d3522A,2010JCAP...08..018C,2010PhLB..692...65F,2010PhRvD..82c5019B,2010PhRvD..82l3509H,2010PhRvD..82g5004F,2010arXiv1004.0691E,2010PhRvD..82i5011B,2011PhRvD..83h3511C,2011PhLB..702..216B} including theoretically well-motivated particle physics models that predict light DM, such as the MSSM \citep{2010PhRvD..81k7701F,2010PhRvD..81k1701K,2011PhRvD..83a5001F,2010PhRvD..81j7302B} and the NMSSM \citep{2011JCAP...01..028K,2010arXiv1009.0549B,2010arXiv1009.2555G,2011PhRvL.106l1805D}, are stringently tested by WMAP and Planck.
A recent study by \citet{2011arXiv1102.5095D} suggests that annihilating DM with masses in the range $m_{{\rm DM}}=1-20$ GeV (i.e. compatible with CoGeNT and DAMA best-fit region), along with annihilation cross section $\langle\sigma_A\upsilon\rangle\sim 9\times10^{-25}$ ${\rm cm}^3{\rm s}^{-1}$ (i.e. $30$ times the standard thermal relic cross section), may give a good match to the observed Fermi and WMAP haze. From Fig.~\ref{fig5} it is clear that, for the DM particle masses in this range, such high annihilation cross sections certainly conflict with available CMB data. If the required boost by a factor of $\sim 30$ is obtained via the Sommerfeld enhancement \citep{2009NuPhB.813....1C,2009PhRvD..79a5014A,2010JCAP...02..028S}, then at $z\sim 1000$ the expected cross section is at least as large as the value given above, so the model could be in even bigger trouble. The alternative claims \citep{2009arXiv0910.2998G,2011PhLB..697..412H} of light DM annihilations in our Galaxy will also be stringently tested.
Finally, it would be interesting to compare the bounds given in \citet{2009PhRvD..80d3526S}, which are based on the results obtained by \citet{2009PhRvD..80b3505G}, with our values. However, since \citet{2009PhRvD..80b3505G} used `on the spot' approximation for the energy deposition, the comparison can only be approximate. The WMAP5 2-sigma bound as given by Eq. (6) of \citet{2009PhRvD..80d3526S} can be cast in the form $\langle\sigma_A\upsilon\rangle\,\left[3\times10^{-26}\,{\rm cm}^3{\rm s}^{-1}\right]/m_{{\rm DM}}\,\left[{\rm GeV}\right]<0.12/f$. To compare this with our WMAP7 results, we have to set $f_{\nu}=0$ in Eq. (\ref{eq1}) and, in the above relation, use a typical value for the $f$-parameter at $z\sim 1000$, which from Fig.~\ref{fig4} is $f\sim 0.85$. Thus, the value for $r$ that should be directly comparable to the 2-sigma WMAP value given in Table~\ref{tab1} is $0.12/0.85\sim 0.14$. Considering the differences in the treatment for the energy deposition, this value agrees reasonably well with our result. However, there are significantly greater differences if one compares the forecasts for the Planck and CVL experiments. In our case Planck and CVL would tighten the 2-sigma bound of WMAP by a factor of $\sim 6$ and $\sim 14$, respectively. The corresponding numbers ($\sim 13$ and $\sim 40$, respectively) from \citet{2009PhRvD..80b3505G} are certainly more optimistic. For CVL, some of this discrepancy is surely due to the higher $\ell_{\max}$ value assumed in \citet{2009PhRvD..80b3505G}: $\ell_{\max}=2500$ compared to our $\ell_{\max}=2000$. Even though we are comparing here the results derived from WMAP5 and WMAP7, the difference is expected to be negligible because the accuracy of the cosmological parameters with an additional two years of WMAP data improves only very moderately (typically somewhere between $5-15\%$), and thus cannot be the cause of the above discrepancy.
NOTE: After the first version of this work was submitted, a new paper by \citet{2011PhRvD..84b7302G} appeared where the authors have performed a similar analysis (now also using WMAP7 data), this time finding results that agree with ours more closely: typical deviations are now within factors $1.2-2$, with their bounds on annihilation cross section being somewhat stronger. These relatively small deviations are actually quite negligible keeping in mind several approximations used in both analyses.
To see how much the omission of the additional excitations of atoms could change our results, we also performed several calculations where the exciting Ly-$\alpha$ photons are treated along the lines presented in \citet{2004PhRvD..70d3502C}. In agreement with the claim in \citet{2011PhRvD..84b7302G}, we find that the bounds on annihilation cross section are getting tighter only up to $\sim 10\%$; i.e., the excitations seem to have only a relatively weak effect with the current prescription.
\subsection*{Note on protons}
\label{sec:protons}
Although for most of the channels and, in particular, for the lower end of the considered DM particle mass range, the number of protons produced is completely negligible, for quark and gluon channels the contribution can reach $\sim 15\%$ of the total energy input if one has masses at the higher end of the considered range. As this is still only a relatively mild contribution, we did not attempt any detailed modeling for the proton component and simply assumed that all of this energy is absorbed as heat by the cosmic medium. A simple justification is the following. The energy loss rate for protons with the energies of interest in this paper is dominated by proton-proton scattering. Thus the loss rate $\Gamma\equiv -\frac{1}{E}\frac{{\rm d}E}{{\rm d}t}\simeq n_{\rm p}c\sigma_{\rm pp}K_{\rm pp}$, where $n_{\rm p}$ is the number density of target protons, $\sigma_{\rm pp}$ the scattering cross section, and $K_{\rm pp}$ the inelasticity parameter. The cross section $\sigma_{\rm pp}$ depends only weakly on proton energy with typical values of $30-40$ mbarn and inelasticity parameter $K_{\rm pp}\simeq 0.5$ for the energies of interest in this work (see e.g. \citealt{2008MNRAS.387..987W}). At redshifts $z\sim 1000$ we therefore get $\Gamma\simeq 10^{-13}$ s$^{-1}$ for the energy loss rate. Comparing this to the expansion rate $H$ at the same redshift, $H\simeq 0.4\times 10^{-13}$ s$^{-1}$, we see that $\Gamma \gtrsim H$, so one might expect a significant fraction of the proton energy to be absorbed.
\section{Summary}
In this paper we have calculated the existing and future CMB constraints on annihilating DM assuming the S-wave annihilation mode to be dominant. Our results can be summarized as follows.
\begin{itemize}
\item[$\bullet$] In full agreement with our earlier findings, first presented in \citet{2009A&A...505..999H}, we confirm that in case of any realistic halo mass-concentration relation, along with annihilation cross sections $\langle\sigma_A\upsilon\rangle$ that do not violate the CMB data, the contribution from structure formation to the CMB signal is completely negligible. Also, the contribution to the low-redshift ionization fraction is very mild, and thus the annihilating DM models compatible with CMB measurements could only play a minor role in helping to reionize the low-$z$ Universe. Therefore for low-mass DM the CMB constraints can be considered
almost free of cosmological uncertainties.
\item[$\bullet$] At the same time, the particle physics uncertainties on the DM annihilation channels can all be described with one single parameter $f_\nu$,
the energy fraction carried away by neutrinos. Therefore the DM annihilations to two muons represents the least stringently tested DM scenario.
\item[$\bullet$] As our main results, in the form of Eq. (\ref{eq1}) and Table~\ref{tab1}, we provided a simple recipe for estimating the upper bounds on annihilation cross section, valid for a broad range of DM models. Two examples for 1-sigma upper bounds for the $\mu$- and $e$-channels (which are the two extremal cases among all of the channels considered in this paper; i.e., those upper bounds bracket the expected results for more general annihilating DM models, given as a superposition over the basis channels) are shown in Fig.~\ref{fig5}.
\item[$\bullet$] For the DM particle masses $m_{{\rm DM}}=6-8$ GeV, which give best fits to CoGeNT and DAMA/LIBRA data, current CMB data is still compatible with a standard thermal relic annihilation cross section $3\times10^{-26}$ cm$^3$s$^{-1}$ only if the annihilations are dominantly into the $\mu$ or $\tau$ (and corresponding 4 lepton) channels. All other annihilation channels already now conflict with the CMB data.
\item[$\bullet$] The sensitivity of Planck space mission allows one to test all the DM annihilation channels definitively. If the Planck mission will not find signals of annihilating DM in the CMB spectrum, all the light DM scenarios motivated by DAMA/LIBRA, CoGeNT, Fermi haze, and WMAP haze will be stringently tested if DM is a thermal relic.
\item[$\bullet$] Our findings strongly disfavor a claim that annihilating DM with $m_{{\rm DM}}\sim 1-20$ GeV and $\langle\sigma_A\upsilon\rangle\sim 9\times10^{-25}$ ${\rm cm}^3{\rm s}^{-1}$ (i.e. $\sim 30$ times above the standard value) could be a cause for Fermi and WMAP haze.
\item[$\bullet$] The last conclusion applies to all DM scenarios with large boost factors from the Sommerfeld enhancement. Because CMB is sensitive to DM annihilations at $z\sim 1000$, at that time the Sommerfeld enhancement of annihilation cross section should have been at least as large as today. Therefore, for light DM scenarios the boost of DM annihilation in our Galaxy can only come form the halo substructure.
\item[$\bullet$] Our MCMC calculations assumed standard 6-parameter $\Lambda$CDM cosmology, which was extended by two additional parameters describing the annihilating DM: $m_{{\rm DM}}$ and $\langle\sigma_A\upsilon\rangle/m_{{\rm DM}}$. It turns out that, after introducing these 2 additional degrees of freedom, all the previous 6 parameters, albeit with small shifts (the most noticeable of which being the ones for $A_{\rm S}$ and $n_{\rm S}$), were still as precisely determined.
\item[$\bullet$] The weakest point of the analysis presented in this paper is the frequently used simple approximation about how the input energy of the cascading particles gets partitioned between ionizations and heating of the environment. For better treatment of this issue one possibly has to rely on detailed Monte Carlo calculations. This, however, we leave for a possible future study.
\end{itemize}
\acknowledgements{We thank Marco Cirelli, Mario Kadastik, and Alessandro Strumia for discussions and our referee for useful comments and suggestions. GH acknowledges the support provided through a visiting scientist fellowship at the MPA. This work was supported by the ESF ERMOS Postdoctoral Research Grant 35, SF0060067s08, SF0690030s09, ESF8005, ESF8943, ESF8090, MTT8, MJD52, and by EU FP7-INFRA-2007-1.2.3 contract No 223807. JC is also very grateful for additional financial support from the Beatrice~D.~Tremaine fellowship 2010.}
\bibliographystyle{aa}
|
\subsection{System preparation}
Our experiments began with nearly pure $\approx 1.8\times 10^5$
atom $\Rb87$ BECs in the $\ket{F=1,m_F=-1}$ state\cite{Lin2009a}
confined in a crossed optical dipole trap. The trap consisted of a
pair of $1064\nm$ laser beams propagating along $\hat{x}-\hat{y}$
($1/e^2$ radii of $w_{\hat{x}+\hat{y}}\approx 120\micron$ and
$w_{\hat z}\approx 50\micron$) and $-\hat{x}-\hat{y}$ ($1/e^2$ radii
of $w_{\hat{x}-\hat{y}}\approx w_{\hat z}\approx 65\micron$).
We prepared equal mixtures of $\ket{F=1,m_F=-1}$ and $\ket{1,0}$
using an initially off resonant rf magnetic field $B_{\text{rf}}(t)
\hat{x}$. We adiabatically ramped $\delta$ to $\delta \approx0$ in
$15\ms$, decreased the rf coupling strength $\Omega_{\text{rf}}$ to
about $150\Hz \ll \hbar\omega_q$ in $6\ms$, and suddenly turned off
$\Omega_{\text{rf}}$, projecting the BEC into an equal
superposition of $\ket{m_F=-1}$ and $\ket{m_F=0}$. We subsequently
ramped $\delta$ to its desired value in $6\ms$ and then linearly
increased the intensity of the Raman lasers from zero to
the final coupling $\Omega$ in $70\ms$.
\subsection{Magnetic fields}
Three pairs of Helmholtz coils, orthogonally aligned along
$\hat{x}+\hat{y}$, $\hat{x}-\hat{y}$ and $\hat{z}$, provided bias
fields $(B_{x+y},B_{x-y},\text{and } B_{z})$. By monitoring the
$\ket{F=1,m_F=-1}$ and $\ket{1,0}$ populations in a nominally
resonant rf dressed state, prepared as above, we observed a
short-time (below $\approx10$ minutes) RMS field stability $g
\mu_{\rm B} B_{\rm RMS}/h\lesssim 80\Hz$. The field drifted slowly
on longer time scales (but changed abruptly when unwary colleagues
entered through our laboratory's ferromagnetic doors). We compensated for
the drift by tracking the rf and Raman resonance conditions.
Due to the small energy scales involved in the experiment, it was
crucial to minimize magnetic field gradients. We detected stray
gradients by monitoring the spatial distribution of
$\ket{m_F=-1}$-$\ket{m_F=0}$ spin mixtures after TOF. Small
magnetic field gradients caused this otherwise miscible mixture
to phase separate along the direction of the gradient. We canceled
the gradients in the $\hat x\!-\!\hat y$ plane with two pairs of
anti-Helmholtz coils, aligned along $\hat x\!+\!\hat y$ and $\hat x\!-\!\hat
y$, to $g\mu_B B'/h\lesssim 0.7\Hz/\mu\text{m}$.
\end{methodssummary}
|
\section{Introduction}
The Klein-Gordon equation is a relativistic version of the Schr\"odinger equation. It describes the motion of a quantum scalar or pseudoscalar field, a field whose quanta are spinless particles. The Klein-Gordon equation describes the quantum amplitude for finding a point particle in various places, cf. for instance \cite{D,S}. It can be expressed in the form
$$
(\Delta -\alpha^2-\frac{1}{c^2} \frac{\partial^2}{\partial t^2})u(x;t) = 0,
$$
where $\Delta := \sum_{i=1}^3 \frac{\partial^2}{\partial x_i^2}$ is the usual Euclidean Laplacian in $\R^3$ and $\alpha = \frac{mc}{\hbar}$. Here, $m$ represents the mass of the particle, $c$ the speed of light and $\hbar$ is the Planck number.
\par\medskip\par
Since a long time it is well known that any solution of the Dirac equation, which describes the spinning electron, satisfies the Klein-Gordon equation. However, the converse is not true. In the time-independent case the homogeneous Klein-Gordon equation simplifies to a screened Poisson equation of the form
$$
(\Delta -\alpha^2) u(x) = 0.
$$
The resolution of this equation provides the first step to solve the more complicated time-dependent case. For this reason, the study of the time-independent solutions is very important. In this paper we therefore focus exclusively on the time-independent equation.
As explained extensively in the literature, see for example \cite{GS2,KC} and elsewhere, the quaternionic calculus has proven to be a useful tool to study the time-independent Klein-Gordon equation. It allows us to factorize the Klein-Gordon operator elegantly by $$\Delta - \alpha^2 =- (D -i \alpha) (D + i\alpha)$$ where $D := \sum_{i=1}^3 \frac{\partial }{\partial x_i} e_i$ is the Euclidean Dirac operator associated to the spatial variable $x=x_1 e_1+x_2 e_2 + x_3 e_3$. In the quaternionic calculus the basis elements $e_1,e_2,e_3$ behave like algebraically the quaternionic basic units ${\bf i}$, ${\bf j}$ and ${\bf k}$. The study of the solutions of the original scalar second order equation is thus reduced to study vector valued eigensolutions to the first order Dirac operator associated to purely imaginary eigenvalues. For eigensolutions of the first order Euclidean Dirac operator it was possible to develop a powerful higher dimensional version of complex function theory, see for instance \cite{GS2,KC,Zhenyuan2,Ry94}. By means of these function theoretical methods it was possible to set up fully analytic representation formulas for the solutions to the homogeneous and inhomogeneous Klein-Gordon in the three dimensional Euclidean space in terms of quaternionic integral operators.
\par\medskip\par
In our recent papers \cite{KraKG,ConKraMMAS2009}, we have shown how these techniques can be carried over to the context of conformally flat oriented cylinders and tori in $\R^3$, and more generally in $\R^n$. The fundamental solution of the Klein-Gordon operator has been constructed from elementary spinor valued multiperiodic solutions of the Klein-Gordon operator.
\par\medskip\par
The aim of this paper is to construct with analogous pseudo-multiperiodic series constructions the fundamental solution of the Klein-Gordon operator on some important examples non-orientable conformally flat manifolds. In the context of this paper we focus on higher dimensional generalizations of the M\"obius strip and the Klein bottle. These manifolds play an important role in general quantum field theory and relativity theory, but also in string and $M$-theory, see for instance \cite{BBS}.
\par\medskip\par
In contrast to the case of the oriented tori and cylinder that we considered earlier, which are examples of spin manifolds, we cannot construct the fundamental solution of the Klein-Gordon equation in terms of spinor valued sections that are in the kernel of $D - i\alpha$. First of all due to the lack of orientability, we cannot construct spinor bundles over these manifolds. Secondly, it is not possible either to construct non-vanishing solutions in the class Ker $D-i\alpha$ in $\R^n$ that have the additional pseudo periodic property to descend properly to these manifolds. A successful way is to start directly from special classes of harmonic functions that take values in bundles of the $^+Pin(n)$ group or $^-Pin(n)$ group. As described for instance in \cite{BD,BRS,KT} the consideration of pin structures is the most logical analogue to spin structures in the non-oriented cases. Pin structures are crucially applied in unoriented superstring theory where non-orientable worldsheets are considered, cf. for instance \cite{BM}.
By means of special classes of pseudo-multiperiodic harmonic functions we develop series representations for the Green's kernel of the Klein-Gordon operator for some $n$-dimensional generalizations of the M\"obius strip and the Klein bottle with values in different pin bundles. These functions represent a generalization of the Weierstra{\ss} $\wp$-function to the context of these geometries.
As applications we present some integral formulas that provide us with the basic stones for doing harmonic analysis in this geometrical context. This paper provides an extension to our prior works \cite{KraNonorientable} and \cite{KRV} in which we developed analogous formulas for the Laplacian and for the Schr\"odinger operator on these manifolds.
\par\medskip\par
The case of the Klein bottle has interesting particular features. As a consequence of the compactness of the Klein bottle we can prove that every solution of the Klein-Gordon operator having atmost unsessential singularities can be expressed as a finite linear combination of the fundamental solution and a finite amount of its partial derivatives. The only entire solution of the Klein-Gordon equation on the Klein bottle is the function $f \equiv 0$.
\section{Pin structures on conformally flat manifolds}
Conformally flat manifolds are $n$-dimensional Riemannian manifolds that possess atlases whose transition functions are conformal maps in the sense of Gauss. For $n > 3$ the set of conformal maps coincides with the set of M\"obius transformations. In the case $n=2$ the sense preserving conformal maps are exactly the holomorphic maps. So, under this viewpoint we may interpret conformally flat manifolds as higher dimensional generalizations of holomorphic Riemann surfaces. On the other hand, conformally flat manifold are precisely those Riemannian manifolds which have a vanishing Weyl tensor.
\par\medskip\par
As mentioned for instance in the classical work of N. Kuiper \cite{Kuiper}, concrete examples of conformally flat orbifolds can be constructed by factoring out a simply connected domain $X$ by a Kleinian group $\Gamma$ that acts discontinuously on $X$. In the cases where $\Gamma$ is torsion free, the topological quotient $X/\Gamma$, consisting of the orbits of a pre-defined group action $\Gamma \times X \to X$, is endowed with a differentiable structure. We then deal with examples of conformally flat manifolds.
\par\medskip\par
In the case of oriented manifolds it is natural to consider spin structures. In the non-oriented case, this is not possible anymore. However, one can consider pin structures instead. For details about the description of pin structures on manifolds that arise as quotients by discrete groups. we refer the reader for instance to \cite{BD}. See also \cite{BRS} and \cite{KT} where in particular the classical M\"obius strip and the classical Klein bottle has been considered.
\par\medskip\par
A classical way of obtaining pin structures for a given Riemannian manifold is to look for a lifting of the principle bundle associated to the orthogonal group $O(n)$ to a principle bundle for the pin groups $^{\pm}Pin(n)$. As described in the above cited works, the group $^+Pin(n):=Pin(n.0)$ is associated to the Clifford algebra $Cl_{n,0}$ of positive signature $(n,0)$. The Clifford algebra $Cl_{n,0}$ is defined as the free algebra modulo the relation $x^2= q_{n,0}(x)$ ($x \in \R^n$) where $q_{n,0}$ is the quadratic form defined by $q_{n,0}(e_i)=+1$ for all basis vectors $e_1,\ldots,e_n$ of $\R^n$. For particular details about Clifford algebras and their related classical groups we also refer the reader to \cite{p}. Next we recall that the group
$^-Pin(n) := Pin(0,n)$ is associated to the Clifford algebra $Cl_{0,n}$ of negative signature $(0,n)$. Here the quadratic form $q_{n,0}$ is replaced by the quadratic form $q_{0,n}$ defined by $q_{0,n}(e_i) =-1$ for all $i=1,\ldots,n$.
Topologically both groups are equivalent, however algebraically they are not isomorphic, cf. for example~\cite{KT}. The more popular $Spin(n)$ group is a subgroup of $^{\pm}Pin(n)$ of index $2$. Here we have $Spin(n):=Spin(0,n) \cong Spin(n,0)$. $Spin(n)$ consists exactly of those matrices from $^{\pm}Pin(n)$ whose determinant equals $+1$.
The groups $^{\pm}Pin(n)$ double cover the group $O(n)$.
So there are surjective homomorphisms $^{\pm}\theta: ^{\pm}Pin(n)\rightarrow ^{\pm}Pin(n)$ with kernel ${\mathbb{Z}}_{2}=\{\pm 1\}$. Adapting from Appendix C of \cite{pp}, where spin structures have been discussed, this homomorphism gives rise to a choice of two local liftings of the principle $O(n)$ bundle to a principle $^{\pm}Pin(n)$ bundle. The number of different global liftings is given by the number of elements in the cohomology group $H^{1}(M,{\mathbb{Z}}_{2})$. See \cite{lm} and elsewhere for details. These choices of liftings give rise to different pinor bundles over $M$. We shall explain their explicit construction on the basis of the examples that we consider in the following two sections.
\section{Higher dimensional M\"obius strips}
In all that follows $\{e_1,\ldots,e_n\}$ stands for the standard basis in $\R^n$.
\par\medskip\par
Let $k \in \{1,\ldots,n\}$ and $\underline{x}= x_1 e_1 + \cdots + x_k e_k$ be a reduced vector from $\mathbb{R}^k$. Let
$\Lambda_k:= \mathbb{Z} e_1 + \cdots + \mathbb{Z} e_k$ be the $k$-dimensional orthonormal lattice.
Suppose further that $\underline{v}:= m_1 e_1 + \cdots + m_k e_k$ is a vector from that lattice $\Lambda_k \subset \mathbb{R}^k$.
\par\medskip\par
Now it is natural to introduce higher dimensional analogoues of the M\"obius strip by the factorization
$$
{\cal{M}}_k^{-} = \mathbb{R}^n/\sim
$$
where $\sim$ is defined by the map $$(\underline{x}+\underline{v},x_{k+1},\ldots,x_{n-1},x_n) \mapsto (x_1,\ldots,x_k,x_{k+1},\ldots,x_{n-1},{\rm sgn}(\underline{v})x_n).$$
Here, for $\underline{v} = m_1 e_1 + \cdots + m_k e_k$ we write
${\rm sgn}(\underline{v}) = \left\{ \begin{array}{cc} 1 & {\rm if}\; \underline{v} \in 2 \Lambda_k \\ -1 & {\rm if}\; \underline{v} \in \Lambda_k \backslash 2 \Lambda_k. \end{array} \right. $
The manifold ${\cal{M}}_k^{-}$ consists of the orbits of the group action $\Lambda_k \times \R^n \to \R^n$, where the action is defined by
$$
\underline{v} \circ x := (\underline{x}+\underline{v},x_{k+1},\ldots,x_{n-1},{\rm sgn}(\underline{v}) x_n).
$$
The classical M\"obius strip is obtained in the case $n=2,k=1$ where the pair $(x_1+1,x_2,X)$ is mapped to $(x_1,-x_2,X)$ after one period. Due to the switch of the minus sign in the $x_n$-component we really deal in that context with non-orientable manifolds. The manifolds ${\cal{M}}_k^-$ are no spin manifolds.
\par\medskip\par
However, analogously to the case of a spin manifold, we can set up several distinct pinor bundles on these manifolds. An extensive algebraic description of all pin structures on the classical M\"obius strip can be found for instance in \cite{BRS}. A simple way to obtain different pin structures is to consider certain decompositions of the period lattices, as we did for the oriented cylinders and tori in our previous works, cf. e.g. \cite{KraRyan2}.
The decomposition of the lattice $\Lambda_k$ into the direct sum of the sublattices $\Lambda_l:= \mathbb{Z} e_1 + \cdots + \mathbb{Z} e_l$ and $\Lambda_{k-l}:= \mathbb{Z} e_{l+1} + \cdots + \mathbb{Z} e_k$ gives rise to $2^k$ different pinor bundles, namely by mapping the tupel $$(\underline{x}+\underline{v},x_{k+1},\ldots,x_n,X) \; {\rm to}\; (\underline{x}, x_{k+1},...,x_{n-1},{\rm sgn}(v) x_n,(-1)^{m_1+\cdots+m_k}X).$$
Here $X \in Cl_n(\mathbb{C})$, where $Cl_n(\mathbb{C})$ stands for the complexification of the Clifford algebra $Cl_{n,0}$ or $Cl_{0,n}$, respectively.
\par\medskip\par
In order to describe the fundamental solution of the time-independent Klein-Gordon operator on the manifolds ${\cal{M}}_k^-$ we recall from standard literature (see for instance \cite{Zhenyuan2}) that the fundamental solution to the Klein-Gordon operator $\Delta-\alpha^2$ in Euclidean flat space $\R^n$ is given by
\begin{equation}
E_{\alpha}(x) = - \frac{i \pi}{2 \omega_n \Gamma(n/2)} (\frac{1}{2}i\alpha)^{\frac{1}{2}n-1} |x|^{1-\frac{1}{2}n}
H_{\frac{n}{2}-1}^{(1)}(i\alpha |x|).
\end{equation}
In this formula, $\omega_n$ stands for the surface measure of the unit sphere in $\R^n$ while $H_{m}^{(1)}$ denotes the first Hankel function with parameter $m$. Furthermore, we choose the root of $\alpha^2$ such that $\alpha > 0$. Since the Hankel function $H_{\frac{n}{2}-1}^{(1)}$ has a point singularity of the order $\frac{1}{2}n-1$ at the origin, the function $E_{\alpha}(x)$ has a point singularity of order $n-2$ at the origin. This is a consequence of the prefactor $|x|^{1-\frac{1}{2}n}$.
\par\medskip\par
We now shall see that we can construct the Green's function of the Klein-Gordon operator on ${\cal{M}}_k^-$ by summing the fundamental solution $E_{\alpha}(x)$ over the period lattice $\Lambda_k$ in a way involving the appropriate minus sign in the $x_n$-component in exactly those terms of the series which are associated to lattice points where ${\rm sgn}(\underline{v}) = -1$. To proceed in this direction we first prove:
\begin{lemma}\label{conv}
Let $\alpha \in \R \backslash \{0\}$. The series
\begin{equation}
\wp_{\alpha}^{{\cal{M}}_k^-}(x):=\sum\limits_{\underline{v} \in \Lambda_k} E_{\alpha}(\underline{x}+\underline{v},x_{k+1},\ldots,x_{n-1},{\rm sgn}(\underline{v})x_n)
\end{equation}
converges uniformely in each compact subset of $\R^n\backslash \Lambda_k$.
\end{lemma}
\begin{proof}
To prove the convergence we use the well-known asymptotic formula
\begin{equation}\label{asy}
H_m^{(1)}(\alpha |x|_2) \sim \sqrt{\frac{2}{\pi
|\alpha||x|_2}} e^{i \alpha|x|_2},\quad {\rm for}\;|x|\;{\rm large},
\end{equation}
also used in \cite{ConKraMMAS2009}, which holds for all positive parameters $m \in \frac{1}{2} \mathbb{N}$ whenever $|x|$ is sufficiently large. From formula (\ref{asy}) we obtain the asymptotic estimate
$$
|E_{\alpha}(x)| \le c \frac{e^{-\alpha|x|}}{|x|^{(n-1)/2}}
$$
which then holds for sufficiently large values for $|x|$, where $c$ is a real constant.
Next we decompose the lattice $\Lambda_k$ in terms of the union of the sets $\Lambda_k = \bigcup_{m=0}^{+\infty} \Omega_m$ where
$$
\Omega_m := \{ \underline{v} \in \Lambda_k \mid |\underline{v}|_{max} = m\}.$$
$|\cdot|_{max}$ denotes the usual maximum norm of a vector. In order to avoid ambiguities we use the notation $|\cdot|_2$ for the Euclidean norm within this proof.
We further consider the following subsets of this lattice $L_m := \{
\underline{v} \in \Lambda_k \mid |\underline{v}|_{max} \le m\}$. Obviously, the
set $L_m$ contains exactly $(2m+1)^n$ points. Hence, the
cardinality of $\Omega_m$ is $\sharp \Omega_m = (2m+1)^n -
(2m-1)^n$. The Euclidean distance between the set $\Omega_{m+1}$
and $\Omega_m$ has the value $d_m := dist_2(\Omega_{m+1},\Omega_m)
= 1$.
\par\medskip\par
Now let us consider an arbitrary compact subset ${\cal{K}} \subset \R^n$.
Then there exists a positive $r \in \R$ such that all $x \in
{\cal{K}}$ satisfy $|x|_{max} \le |x|_2 < r$.
Suppose now that $x$ is a point of ${\cal{K}}$. To show the
normal convergence of the series, we can leave out without loss of
generality a finite set of lattice points. We consider without loss of
generality only the summation over those lattice points that
satisfy $|\underline{v}|_{max} \ge [r]+1$. In view of $|x +
\underline{v}|_2 \ge |\underline{v}|_2 - |x|_2 \ge
|\underline{v}|_{max}-|x|_2 = m - |x|_2 \ge m - r$ we
obtain
\begin{eqnarray*}
& & \sum\limits_{m=[r]+1}^{+\infty} \sum\limits_{\omega \in \Omega_m} |E_{\alpha}(\underline{x}+\underline{v},x_{k+1},\ldots,x_{n-1},{\rm sgn}(\underline{v}))|_2\\
& \le & c \sum\limits_{m=[r]+1}^{+\infty} \sum\limits_{\omega \in \Omega_m}
\frac{e^{-|\alpha||\underline{x}+\underline{v}+ x_{k+1}e_{k+1} + \cdots + {\rm sgn}(\underline{v}) x_n e_n|_2}}
{\sqrt{(|\underline{x}+\underline{v}+x_{k+1} + x_{k+1}e_{k+1} + \cdots + {\rm sgn}(\underline{v}) x_n e_n|_2)^{n-1}}}\\
& \le & c \sum\limits_{m=[r]+1}^{+\infty} [(2m+1)^n - (2m-1)^n]
\frac{e^{-|\alpha|(r-m)}}{(m-r)^{(n-1)/2}},
\end{eqnarray*}
where $c$ is an appropriately chosen positive real constant,
because $m - r \ge [r]+1-r > 0$. This sum clearly is absolutely uniformly convergent. Hence,
the series $\wp_{\alpha}^{{\cal{M}}_k^-}(x)$,
which can be written as
$$
\wp_{\alpha}^{{\cal{M}}_k^-}(x) :=
\sum\limits_{m=0}^{+\infty}\sum\limits_{\underline{v} \in \Omega_m}
E_{\alpha}(\underline{x} + \underline{v},x_{k+1},\ldots,x_{n-1},{\rm sgn}(\underline{v})x_n),
$$
converges normally on $\R^n \backslash \Lambda_k$.
\end{proof}
{\bf Remark}. The case $\alpha=0$ is the harmonic case treated in \cite{KraNonorientable}.
\par\medskip\par
Next we may establish that
\begin{lemma}\label{reg}
The function $\wp_{\alpha}^{{\cal{M}}_k^-}(x)$ is an element from Ker $\Delta-\alpha^2$ on $\R^n\backslash \Lambda_k$.
\end{lemma}
\begin{proof}
The main point is to observe that each term
$$
E_{\alpha}(\underline{x} + \underline{v},x_{k+1},\ldots,{\rm sgn}(\underline{v})x_n)
$$
is an element from Ker $\Delta-\alpha^2$ in $\mathbb{R}^n\backslash\{(\underline{v},0)\}$. This is trivial for $\underline{v}=\underline{0}$, since $E_{\alpha}(x)$ is the fundamental solution of $\Delta-\alpha^2$. Next, since we differentiate twice to each variable (also w.r.t. $x_n$), the term
$E_{\alpha}(\underline{x},\ldots,x_{n-1},\pm x_n)$ turns out to be an element from Ker $\Delta-\alpha^2$, too. Finally, due to the invariance of $\Delta-\alpha^2$ under translations of the form $x \mapsto x + \underline{v}$, each term $E_{\alpha}(\underline{x} + \underline{v},x_{k+1},\ldots,{\rm sgn}(\underline{v})x_n)$ then is an element from that kernel, due to the previous arguments.
Now we may conclude by Weierstra{\ss}' convergence theorem that the entire series $\wp_{\alpha}^{{\cal{M}}_k^-}(x)$ is an element from that kernel in all points of $\R^n \backslash \Lambda_k$, due to the normal convergence.
\end{proof}
The next step is to observe
\begin{lemma}\label{periodic}
The function $\wp_{\alpha}^{{\cal{M}}_k^-}(x)$ satisfies
\begin{equation}
\wp_{\alpha}^{{\cal{M}}_k^-}(\underline{x}+\underline{\eta},x_{k+1},\ldots,x_n) = \wp_{\alpha}^{{\cal{M}}_k^-}(\underline{x},x_{k+1}.\ldots,x_{n-1},{\rm sgn}(\underline{\eta})x_n) \quad\quad \forall \underline{\eta} \in \Lambda_k.
\end{equation}
\end{lemma}
\begin{proof}
Take an arbitrary $\underline{\eta} \in \Lambda_k$. Then
\begin{eqnarray*}
& & \wp_{\alpha}^{{\cal{M}}_k^-}(\underline{x}+\underline{\eta},x_{k+1},\ldots,x_n)\\ &=& \sum\limits_{\underline{v} \in \Lambda_k} E_{\alpha}(\underline{x}+\underline{\eta}+\underline{v},x_{k+1},\ldots,x_{n-1},{\rm sgn}(\underline{v})x_n)\\
&=& \sum\limits_{\underline{v} \in \Lambda_k} E_{\alpha}(\underline{x}+(\underline{v}+\underline{\eta}),x_{k+1},\ldots,x_{n-1},{\rm sgn}( \underline{\eta}){\rm sgn}(\underline{v}+\underline{\eta})x_n)\\
&=& \sum\limits_{\underline{v'} \in \Lambda_k} E_{\alpha}(\underline{x}+\underline{v'},x_{k+1},\ldots,x_{n-1},{\rm sgn} (\underline{\eta}){\rm sgn}(\underline{v'})x_n)\\
&=& \wp_{\alpha}^{{\cal{M}}_k^-}(\underline{x},\ldots,x_{n-1},{\rm sgn}(\underline{\eta})x_n),
\end{eqnarray*}
where we applied a direct rearrangement argument.
\end{proof}
Now let $p_-$ be the well-defined projection map from the universal covering space $\R^n$ down to the manifold ${\cal{M}}_k^- = \R^n/\sim$.
As a consequence of Lemma~\ref{periodic} we may now conclude that the function $\wp_{\alpha}^{{\cal{M}}_k^-}(x)$ descends to a well-defined pinor section ${\wp'}_{\alpha}^{{\cal{M}}_k^-}(x'):= p_{-}(\wp_{\alpha}^{{\cal{M}}_k^-}(x))$ on the manifold ${\cal{M}}_k^-$ which takes its values in the trivial pinor bundle. Here, we use the notation $x':= p_{-}(x) = x \;{\rm mod}\; \sim$. The projection map $p'$ induces canonically a Klein-Gordon operator ${\Delta'}_{\alpha} = p_-(\Delta-\alpha^2)$ on ${\cal{M}}_k^-$. As a consequence of Lemma~\ref{reg}, the section ${\wp'}_{\alpha}^{{\cal{M}}_k^-}(x')$ then is in Ker ${\Delta'}_{\alpha}$. Finally, we are in position to establish the main result of this paragraph
\begin{theorem}
Consider the decomposition of the lattice $\Lambda_k = \Lambda_{l} \oplus \Lambda_{k-l}$ for some $l \in \{1,\ldots,k\}$ and write a lattice point $\underline{v} \in \Lambda_k$ in the form $\underline{v} = m_1 e_1 + \cdots+ m_l e_l + m_{l+1}e_{l+1} + \cdots + m_k e_k$ with integers $m_1,\ldots,m_k \in \mathbb{Z}$.
Let ${\cal{E}}^{(q)}$ be the specific pinor bundle on ${\cal{M}}_k^-$ defined by the map
$$(\underline{x}+\underline{v},x_{k+1},\ldots,x_n,X) \; {\rm to}\; (\underline{x}, x_{k+1},...,x_{n-1},{\rm sgn}(v) x_n,(-1)^{m_1+\cdots+m_k}X).$$ The fundamental solution to the Klein-Gordon operator on ${\cal{M}}_k^-$ for pinor sections with values in the pinor bundle ${\cal{E}}^{(q)}$ can be expressed by
\begin{equation}
{E'}_{\alpha,q}(x') = p_-\Bigg(\sum\limits_{\underline{v} \in \Lambda_l \oplus \Lambda_{k-l}} (-1)^{m_1+\cdots+m_l} E_{\alpha}(\underline{x}+\underline{v},x_{k+1},\ldots,x_{n-1},{\rm sgn}(v)x_n) \Bigg).
\end{equation}
\end{theorem}
In the case of the trivial bundle we shall simply write ${E'}_{\alpha}(x')$ in what follows.
\begin{proof}
With the modification of the appropriate minus sign induced by the factor $(-1)^{m_1+\cdots+m_l}$ in each term of the series we ensure in combination with the previous lemmas that ${E'}_{\alpha}(x')$ actually is a well-defined section on ${\cal{M}}_k^-$ and it takes values in the chosen pinor bundle ${\cal{E}}^{(q)}$. With analogous arguments as used in the proof of Lemma~\ref{reg}, this section is in the kernel of the Klein-Gordon operator. Since the expression $E_{\alpha}(x-y)$ is the Green's kernel of the usual Klein-Gordon operator in the Euclidean flat space $\R^n$, its canonical projection ${E'}_{\alpha}(x')$ takes over this role on the manifold ${\cal{M}}_k^-$.
It reproduces each section in the kernel of $\Delta'_{\alpha}$ by the Green's integral.
\end{proof}
We can say more. These explicit formulas for the fundamental solutions allow us to express the solutions of the homogeneous and inhomogeneous Klein-Gordon problem with and without boundary conditions.
\par\medskip\par
Let us suppose that $D' \subset {\cal{M}}_k^-$ is a bounded domain with a $C^2$-boundary. Let $\nu$ denote the unit normal vector to the boundary $\partial D$ directed into the exterior of $D'$. Assume further that $u'$ is a $C^2$-section in the kernel of ${\Delta'}_{\alpha} = p_-(\Delta-\alpha^2)$ with values in the pinor bundle ${\cal{E}}^{(q)}$. Since ${\Delta'}_{\alpha}$ is a scalar operator, we may restrict to consider each scalar component function ${u_j}'$ of the pinor valued section $u'$ separately.
Suppose that $u$ has a normal derivative on $\partial D'$ in the sense that the limit
$$
\frac{\partial {u_j}'}{\partial \nu}(y') = \lim\limits_{h \to 0^+} \nu(y') \cdot {\rm grad}\; {u_j}'(y'-h\nu(y')),\quad y' \in \partial D'
$$
exists uniformely on $\partial D'$ for each scalar component function ${u_j}'$. Then we can infer by classical harmonic analysis arguments that each scalar component function satisfies
\begin{equation}\label{homcomp}
{u_j}'(x') = \int\limits_{\partial D'} \Big[ \frac{\partial {u_j}'}{\partial \nu}(y'){E'}_{\alpha,q}(x'-y') - {u_j}'(y')
\frac{\partial {E'}_{\alpha,q}(x'-y')}{\partial \nu(y')}\Big] dS(y).
\end{equation}
Therefore, the whole pinor valued section $u'$ satisfies
\begin{equation}\label{hom}
{u'}(x') = \int\limits_{\partial D'} \Big[ \frac{\partial {u'}}{\partial \nu}(y'){E'}_{\alpha,q}(x'-y') - {u'}(y')
\frac{\partial {E'}_{\alpha,q}(x'-y')}{\partial \nu(y')}\Big] dS(y).
\end{equation}
Next, as a further consequence we can directly present the solution of the inhomogeneous Klein-Gordon problem $\Delta'_{\alpha} u' = f'$ in $D'$ with no-boundary condition in terms of the convolution integral over the fundamental section ${E'}_{\alpha,q}$. Under the weaker condition that $f'$ is an ${\cal{E}}^{(q)}$ valued pinor section belonging to $W^{p}_{2}(D')$ for some $p > 0$ we have
\begin{equation}\label{inh}
u'(x') = \int\limits_{D'} {E'}_{\alpha,q}(x'-y') f'(y') dV(y').
\end{equation}
Here, and in the following $W^{p}_2(D')$ stands for the Sobolev space consisting of the $p$-times weakly differentiable ${\cal{E}}^{(q)}$-valued sections belonging to $L_2(D')$.
\par\medskip\par
In view of the linearity of the Klein-Gordon operator, we obtain from combining the latter statement with a weaker version of the previous one where we work in the same Sobolev spaces as in \cite{GS2}, the following statement.
\begin{theorem} (Inhomogeneous Klein-Gordon boundary value problem on ${\cal{M}}_k^-$).
Let $D' \subset {\cal{M}}_k^-$ be a bounded domain with a $C^2$-boundary. Suppose that $f' \in W^{p}_2(D')$ and that $g' \in W^{p+3/2}_2(\partial D')$. Then the boundary value problem $\Delta'_{\alpha} u' = f'$ in $D'$ with $u'|_{\partial D'} = g'$ has a unique solution in $W^{p+2,loc}_2$ of the form
\begin{eqnarray*}
u'(x') &=& \int\limits_{D'} {E'}_{\alpha,q}(x'-y') f'(y') dV(y') \\&+& \int\limits_{\partial D'} \Big[ \frac{\partial u'}{\partial \nu}(y'){E'}_{\alpha,q}(x'-y') - u'(y')
\frac{\partial {E'}_{\alpha,q}(x'-y')}{\partial \nu(y')}\Big] dS(y).
\end{eqnarray*}
\end{theorem}
\section{Higher dimensional generalizations of the Klein bottle}
Now we turn to discuss analogous constructions for higher dimensional generalizations of the Klein bottle. To leave it simple we consider an $n$-dimensional normalized lattice of the form $\Lambda_n:=\mathbb{Z} e_1 + \cdots + \mathbb{Z} e_n$.
\par\medskip\par
We introduce higher dimensional generalization of the classical Klein bottle by the factorization
$$
{\cal{K}}_n:=\mathbb{R}^n/\sim^*
$$
where $\sim^*$ is now defined by the map
$$
(\underline{x} + \sum_{i=1}^{n-1} m_i e_i +(x_n + m_n) e_n) \mapsto(x_1,\cdots,x_{n-1},(-1)^{m_n} x_n).
$$
The manifolds ${\cal{K}}_n$ can be described as the set of orbits of the group action $\Lambda_n \times \R^n \to \R^n$ where the action is now defined by
$$
v \circ x := (\sum\limits_{i=1}^{k-1} x_i e_i+\sum_{i=1}^{k-1}m_i e_i+((-1)^{m_k} x_k + m_k) e_k,x_{k+1},\ldots,x_{n-1}, x_n),
$$
where $v = m_1 e_1 + \cdots + m_n e_n$ is a lattice point from $\Lambda_n$.
Here, and in the remaining part of this section, $\underline{x}$ denotes a shortened vector in $\mathbb{R}^{n-1}$. In the case $n=2$ we obtain the classical Klein bottle. Notice that in contrast to the M\"obius strips treated in the previous section, here the minus sign switch occurs in one of the component on which the period lattice acts, too. As for the M\"obius strips we can again set up distinct pinor bundles. See also \cite{BRS} where pin structures of the classical four-dimensional Klein-bottle have been considered.
By decomposing the complete $n$-dimensional lattice $\Lambda_n$ into a direct sum of two sublattices $\Lambda_n = \Omega_l \oplus \Lambda_{n-l}$ we can again construct $2^n$ distinct pinor bundles by considering the maps
$$
(\underline{x} + \sum_{i=1}^{n-1} m_i e_i, x_n + m_n,X) \mapsto(x_1,\cdots,x_{n-1},(-1)^{m_n} x_n,(-1)^{m_1+\cdots+m_l}X).
$$
By similar arguments as applied in the previous section we may establish
\begin{theorem}
Consider the decomposition of the lattice $\Lambda_n = \Lambda_{l} \oplus \Lambda_{n-l}$ for some $l \in \{1,\ldots,n\}$ and write a lattice point $v \in \Lambda_n$ in the form $v = m_1 e_1 + \cdots+ m_l e_l + m_{l+1}e_{l+1} + \cdots + m_n e_n$ with integers $m_1,\ldots,m_n \in \mathbb{Z}$.
Let ${\cal{E}}^{(q)}$ be the pinor bundle on ${\cal{K}}_n$ defined by the map $$
(\underline{x} + \sum_{i=1}^{n-1} m_i e_i, x_n + m_n,X) \mapsto(x_1,\cdots,x_{n-1},(-1)^{m_n} x_n,(-1)^{m_1+\cdots+m_l}X).
$$
The fundamental solution of the Klein-Gordon operator on ${\cal{K}}_n$ (induced by $p_*(\Delta-\alpha^2)$ ) for sections with values
in the pinor bundle ${\cal{E}}^{(q)}$ can be expressed by
\begin{equation}
{E'}_{\alpha,q}(x') = p_*\Bigg(\sum\limits_{v \in \Lambda_l \oplus \Lambda_{n-l}} (-1)^{m_1+\cdots+m_l} E_{\alpha}(\sum_{i=1}^{n-1} (x_i + m_i) e_i,(-1)^{m_n}x_n+x_m) \Bigg),
\end{equation}
where $p_*$ now denotes the projection from $\R^n$ to ${\cal{K}}_n = \R^n/\sim^*$. The symbol $'$ represents the image under $p_*$ up from now.
\end{theorem}
To the proof, without loss of generality we consider the trivial bundle, as the arguments can easily be adapted to the other bundles that we also considered, namely by taking into account the parity factor $(-1)^{m_1+\cdots+m_l}$. Completely analogously to the proof of Lemma~\ref{conv} we may establish the normal convergence of the series
\begin{equation}
\wp_{\alpha}^{{\cal{K}}_n}(x) := \sum\limits_{v \in \Lambda_{n}} E_{\alpha}(\sum_{i=1}^{n-1} (x_i + m_i) e_i,(-1)^{m_n}x_n+x_m)
\end{equation}
in $\R^n \backslash \Lambda_n$. With the same argumentation as in Lemma~\ref{reg} we may establish that the function $\wp_{\alpha}^{{\cal{K}}_n}$ is an element from Ker $\Delta-\alpha^2$ in $\R^n \backslash \Lambda_n$. The main point that remains to show is
\begin{lemma} For all $k:=k_1 e_1 + \cdots + k_n e_n \in \Lambda_n$ we have
$$
\wp_{\alpha}^{{\cal{K}}_n}(x+k) = \wp_{\alpha}^{{\cal{K}}_n}(x_1,\ldots,x_{n-1},(-1)^{k_n} x_n).
$$
\end{lemma}
\begin{proof}
To prove this statement it is important to use the following decomposition
\begin{eqnarray*}
\wp_{\alpha}^{{\cal{K}}_n}(x) &=& \sum\limits_{(m_1,\ldots,m_n) \in \mathbb{Z}^n} E_{\alpha}(x_1+m_1,\ldots,x_{n-1} + m_{n-1},(-1)^{m_n} x_n + m_n) \\
&=& \sum\limits_{(m_1,\ldots,m_{n-1}) \in \mathbb{Z}^{n-1},m_n \in 2 \mathbb{Z}} E_{\alpha}(x_1+m_1,\ldots,x_{n-1} + m_{n-1}, x_n + m_n) \\
&=& \sum\limits_{(m_1,\ldots,m_{n-1}) \in \mathbb{Z}^{n-1},m_n \in 2 \mathbb{Z}+1} E_{\alpha}(x_1+m_1,\ldots,x_{n-1} + m_{n-1}, -x_n + m_n).
\end{eqnarray*}
First we note that
$$
\wp_{\alpha}^{{\cal{K}}_n}(x_1+k_1,\ldots,x_{n-1}+k_{n-1}, x_n) = \wp_{\alpha}^{{\cal{K}}_n}(x_1,\ldots,x_n)
$$
for all $(k_1,\ldots,k_{n-1}) \in \mathbb{Z}^{n-1}$.
This follows by the direct series rearrangement
\begin{eqnarray*}
& & \wp_{\alpha}^{{\cal{K}}_n}(x_1+k_1,\ldots,x_{n-1}+k_{n-1},x_n)\\
&=& \sum\limits_{(m_1,\ldots,m_n) \in \mathbb{Z}^n} E_{\alpha}(x_1+k_1+m_1,\ldots,x_{n-1}+k_{n-1}+m_{n-1},(-1)^{m_n} x_n + m_n)\\
&=& \sum\limits_{(p_1,\ldots,p_n) \in \mathbb{Z}^n} E_{\alpha} (x_1+p_1,\ldots,x_{n-1} + p_{n-1},(-1)^{p_n} x_n+p_n)
\end{eqnarray*}
where we put $p_i:= m_i + k_i \in \mathbb{Z}$ for $i=1,\ldots,n-1$ and $p_n := m_n$. Notice that rearrangement is allowed because the series converges normally on $\R^n\backslash \Lambda_n$.
\par\medskip\par
It thus suffices to show
$$
\wp_{\alpha}^{{\cal{K}}_n}(x_1,\ldots,x_{n-1},x_n+1)= \wp_{\alpha}^{{\cal{K}}_n}(x_1,\ldots,x_{n-1},-x_n).
$$
We observe that
\begin{eqnarray*}
& &\wp_{\alpha}^{{\cal{K}}_n}(x_1,\ldots,x_{n-1},x_n+1) \\
&=& \sum\limits_{(m_1,\ldots,m_n) \in \mathbb{Z}^n} E_{\alpha}(x_1+m_1,\ldots,x_{n-1}+m_{n-1},(-1)^{m_n}(x_n+1)+m_n)\\
&=& \sum\limits_{(m_1,\ldots,m_{n-1}) \in \mathbb{Z}^{n-1},m_n \in 2 \mathbb{Z}} E_{\alpha}(x_1+m_1,\ldots,x_{n-1}+m_{n-1}, x_n+\underbrace{m_{n}+1}_{odd})\\
&+& \sum\limits_{(m_1,\ldots,m_{n-1}) \in \mathbb{Z}^{n-1},m_n \in 2 \mathbb{Z}+1} E_{\alpha}(x_1+m_1,\ldots,x_{n-1}+m_{n-1}, -x_n+\underbrace{m_{n}-1}_{even})\\
&=& \sum\limits_{(p_1,\ldots,p_{n-1}) \in \mathbb{Z}^{n-1},p_n \in 2 \mathbb{Z}+1} E_{\alpha}(x_1+p_1,\ldots,x_{n-1}+p_{n-1}, x_n+p_n)\\
&=& \sum\limits_{(p_1,\ldots,p_{n-1}) \in \mathbb{Z}^{n-1},q_n \in 2 \mathbb{Z}} E_{\alpha}(x_1+p_1,\ldots,x_{n-1}+p_{n-1}, -x_n+q_n)\\
&=& \wp_{\alpha}^{{\cal{K}}_n}(x_1,\ldots,x_{n-1},-x_n).
\end{eqnarray*}
The fact that
$$
\wp_{\alpha}^{{\cal{K}}_n}(x_1,\ldots,x_{n-1},x_n+k_n) = \wp_{\alpha}^{{\cal{K}}_n}(x_1,\ldots,x_{n-1},(-1)^{k_n} x_n)
$$
is true for all $k_n \in \mathbb{Z}$ now follows by a direct induction argument on $k_n$.
\end{proof}
With this property we may infer that $\wp_{\alpha}^{{\cal{K}}_n}$ descends to a well defined section on ${\cal{K}}_n$ by applying the projection $p_*(\wp_{\alpha}^{{\cal{K}}_n})$. The result of this projection will be denoted by ${E'}_{\alpha}(x')$. ${E'}_{\alpha}(x')$ is the canonical skew symmetric periodization of $E_{\alpha}(x)$ that is constructed in such a way that it descends to the manifold. Therefore, the reproduction property of ${E'}_{\alpha}(x'-y')$ on ${\cal{K}}_n$ follows from the reproduction property of the usual Green's kernel $E_{\alpha}(x-y)$ in Euclidean space, where we apply the usual Green's integral formula for the Klein-Gordon operator.
\par\medskip\par
{\bf Remarks}. The fundamental solution of the Klein-Gordon operator on the usual Klein bottle in two real variables (for pinor sections with values in the trivial bundle) has the form
$$ p_*\Bigg(\sum\limits_{v \in \Lambda_{2}} E_{\alpha}((x_1 + m_1),(-1)^{m_2}x_2+m_2) \Bigg).$$
In terms of this formula for the fundamental solution of the Klein-Gordon operator on the manifolds ${\cal {K}}_n$ we can deduce similar representation formulas for the solutions to the general inhomogeneous Klein-Gordon problem with prescribed boundary conditions on these manifolds as presented at the end of the previous section in the context of the M\"obius strips.
\par\medskip\par
The fact that the manifolds ${\cal{K}}_n$ are compact manifolds has some interesting special function theoretical consequences. We shall see that we can express any arbitrary solution of the Klein-Gordon equation with unessential singularities on these manifolds as a finite sum of linear combinations of the fundamental solution ${E'}_{\alpha}$ and its partial derivatives. To proceed in this direction, we start to establish
\begin{lemma}
Let $\alpha \neq 0$. Suppose that $f: \mathbb{R}^n \to \mathbb{C}$ is an entire solution of $(\Delta-\alpha^2)f=0$ on the whole $\R^n$. If $f$ additionally satisfies
\begin{equation}\label{Kinv}
f(x_1+m_1,\cdots,x_n+m_n) = f(x_1,\ldots,x_{n-1},(-1)^{m_n} x_n)
\end{equation}
for all $(m_1,\ldots,m_n)\in \mathbb{Z}^n$, then $f$ vanishes identically on $\R^n$.
\end{lemma}
\begin{proof}
Since $f$ satisfies the relation $$f(x_1+m_1,\cdots,x_n+m_n) = f(x_1,\ldots,x_{n-1},(-1)^{m_n} x_n),$$ it takes all its values in the $n$-dimensional period cell $[0,1]^{n-1}\times [0,2]$, because
$$f(x_1+m_1,x_2+m_2,\ldots,x_{n-1}+m_{n-1},x_n+2m_n)=f(x_1,x_2,\ldots,x_{n-1},x_n)$$ for all $(m_1,\ldots,m_n) \in \mathbb{Z}^n$.
The set $[0,1]^{n-1} \times [0,2]$ is compact. Since $f$ is an entire solution of $(\Delta-\alpha^2)f=0$ on the whole $\R^n$, it is in particular continuous on $[0,1]^{n-1}\times [0,2]$. Consequently, $f$ must be bounded on $[0,1]^{n-1}\times [0,2]$ and therefore it must be bounded over the whole $\R^n$, too.
\par\medskip\par
Since $f$ is an entire solution of $(\Delta-\alpha^2)f=0$, it can be expanded into a Taylor series of the following form, compare with \cite{ConKraMMAS2009},
$$
f(x) = \sum\limits_{q=0}^{\infty} |x|^{1-q-n/2} J_{q+n/2-1}(\alpha|x|) H_q(x).
$$
This Taylor series representation holds in the whole space $\R^n$.
\par\medskip\par
Here, $H_q(x)$ are homogeneous harmonic polynomials of total degree $q$. These are often called spherical harmonics, cf. for example~ \cite{M}.
Since the Bessel $J$ functions are exponentially unbounded away from the real axis, $f$ can only be bounded if all spherical harmonics $H_q$ vanish identically. Hence, $f \equiv 0$.
\end{proof}
Notice that all constant functions $f \equiv C$ with $C \neq 0$ are not solutions of $(\Delta-\alpha^2)f=0$. As a direct consequence we obtain
\begin{corollary}
There are no non-vanishing entire solutions of $(\Delta-\alpha^2)f=0$ on the manifolds ${\cal{K}}_n$ (in particular on the Klein bottle ${\cal{K}}_2$).
\end{corollary}
This is a fundamental consequence of the compactness of the manifolds ${\cal{K}}_n$. Notice that this argument cannot be carried over to the context of the manifolds that we considered in the previous section, since those are not compact.
\par\medskip\par
{\bf Remark}. The statement can be adapted to the harmonic case $\alpha=0$. In this case one has a Taylor series expansion of the simpler form
$$
f(x) = \sum\limits_{q=0}^{\infty} H_q(x),
$$
where only the spherical harmonics of total degree $q=0,1,\ldots$ are involved. The only bounded entire harmonic functions are constants. Applying the same argumentation leads to the fact that the only harmonic solutions on ${\cal{K}}_n$ are constants.
\par\medskip\par
We can say more. We shall now explain how we can use the function $\wp^{{{\cal K}}_n}(x)$ as basic function to construct any arbitrary solution to the Klein-Gordon operator on ${\cal{K}}_n$.
\par\medskip\par
In order to make the paper self-contained we briefly recall from \cite{KraHabil,Nef2} the classification of the singularities and the definition of the order of singularities, adapted here to the context of harmonic functions in $\R^n$. We start with
\begin{definition}
A point $\tilde{x} \in \R^n$ is called a left regular point of a harmonic function $f$, if there exists an $\varepsilon > 0$ such that $f$ is harmonic in the open ball $B(\tilde{x},\varepsilon):=\{x \in \R^n \mid |x-\tilde{x}|_2 < \varepsilon\}$.
Otherwise, $\tilde{x}$ is called a singular point of $f$. $\tilde{x}$ is called an isolated singularity of $f$ if it is a singular point of $f$ for which one can find an $\varepsilon > 0$ such that $f$ is harmonic in $B(\tilde{x},\varepsilon)\backslash\{\tilde{x}\}$.
\end{definition}
Next suppose that $U \subset \R^n$ is an open set and that $S \subset U$ is a closed subset. Let us consider around each $s \in S$ a ball with radius $\rho > 0$ and let us denote by $H_{\rho}$ the hypersurface of the union of all balls centered at each $s \in S$ with radius $\rho$. In the case where $S$ is just a single isolated point, $H_{\rho}$ is simply the sphere centered at $s$ with radius $\rho$. If $S$ is a rectifiable line, or, more generally, a $p$-dimensional manifold with boundary, then $H_{\rho}$ is the surface of a tube domain.
Let us denote by $J(\rho)$ the limit inferior of the volumes of all closed orientable hypersurfaces $H_{R}$ with continuous normal field that contains $S$ in the interior where we suppose that
$$
\inf\{|s-\tilde{x}|_2; s \in S, \tilde{x} \in H_{R} \} \ge \rho.
$$
In terms of these notions one can now give the following definition which was given in \cite{KraHabil} for Clifford algebra valued monogenic functions.
\begin{definition}
Let $U \subset \R^n$ be an open set and let $S \subset U$ be a closed subset. Suppose that $f$ is harmonic in $U \backslash S$ and that $f$ has singularities in each $s \in S$. Let $H_{\rho} \subset U$ be a hypersurface as defined above.
The singular point $s \in S$ is then called an unessential singularity of $f$ if there is an $r \in \mathbb{N}$ and an $M>0$ such that
\begin{equation}\label{Hrho}
|\rho^r J(\rho) f(\tilde{x}) |_2< M \quad \quad \tilde{x} \in H_{\rho}.
\end{equation}
In this case the minimum of the parameters $r$ is called the order of the singular point $s$. If no finite $r$ with this property can be found, then $s$ is called an essential singularity.
\end{definition}
In the case where $S$ is a isolated point singularity, a rectifiable line or a manifold with boundary, one can substitute the condition (\ref{Hrho}) by
$$
|\rho^r f(\tilde{x}) |_2 < M.
$$
\par\medskip\par
Suppose now that $f$ is a function with the invariance behavior (\ref{Kinv}) that satisfies $(\Delta - \alpha^2) f = 0$ in $\R^n
\backslash \Lambda_n$. Let ${\bf m} := (m_1,\ldots,m_n) \in \mathbb{N}_0^n$ be
a multi-index with length $|{\bf m}|:= m_1 + \cdots + m_n$. Then
the function $f_{\bf m}(x) = \frac{\partial^{|{\bf
m}|}}{\partial x_1^{m_1} \cdots \partial x_n^{m_n}} f(x)$ satisfies (\ref{Kinv}) too, and it is also in Ker $(\Delta -
\alpha^2)$. In particular, when taking the function $$\wp_{\alpha,q}^{{\cal{K}}_n}(x) :=
\sum\limits_{v \in \Lambda_l \oplus \Lambda_{n-l}} (-1)^{m_1+\cdots+m_l} E_{\alpha}(\sum_{i=1}^{n-1} (x_i + m_i) e_i,(-1)^{m_n}x_n+x_m)
,$$
then the functions $\wp_{\alpha,q;{\bf m}}^{{\cal{K}}_n}(x) :=
\frac{\partial^{|{\bf m}|}}{\partial x_1^{m_1} \cdots \partial x_n^{m_n}}
\wp_{\alpha,q}^{{\cal{K}}_n}(x)$ have the invariance behavior (\ref{Kinv})
and satisfy $(\Delta-\alpha^2) \wp_{\alpha,q;{\bf m}}^{{\cal{K}}_n}(x) = 0$ at
each point of $\R^n \backslash \Lambda_n$. In each lattice point they
have an isolated pole of order $n-2+|{\bf m}|$. In view of the
translation invariance of the operator $\Delta-\alpha^2$, we can
construct functions satisfying $(\ref{Kinv})$ that have poles in a given
set of points $a_i + \Lambda_n$ of order $N_i$ ($i=1,\ldots,l)$ with
$N_i \ge n-2$ by making the construction
\begin{equation}
\label{con} \sum\limits_{i=1}^l \wp_{\alpha,q;{\bf N}_i}^{{\cal{K}}_n}(x -
a_i) b_i
\end{equation}
where ${\bf N}_i$ is a multi-index of length $N_i$ and where $b_i$
are arbitrary elements from $\mathbb{C}$.
Due to the compactness of the fundamental period cell, one can only construct functions in Ker $\Delta - \alpha^2$ satisfying (\ref{Kinv}) with a finite number of isolated singularities.
In contrast to the classical harmonic case, it is possible to construct non-vanishing sections on the manifolds ${\cal{K}}_n$ in the kernel of the Klein-Gordon operator with only one point singularity of order $n-2$ on the whole manifold. The most important example is ${E'}_{\alpha,q}(x')$.
Now we may formulate two main representation theorems that fully describe all sections on ${\cal{K}}_n$ in the kernel of the Klein-Gordon operator with non-essential singularities.
\begin{theorem}\label{th2}
Let ${a}_1,{a}_2,\ldots,{a}_p \in \R^n \backslash \Lambda_n$ be a
finite set of points that are incongruent modulo $\Lambda_n$. Suppose
that $f: \R^n \backslash\{{a}_1+\Lambda_n,\ldots,{a}_p+\Lambda_n\} \to
{\mathbb{C}}$ is an element from Ker $\Delta-\alpha^2$ satisfying (\ref{Kinv}) which has at most isolated poles at
the points ${a}_i$ of the order $K_i$ (where the order is defined as in \cite{KraHabil}). Then there exist
numbers $b_1,\ldots,b_p \in \mathbb{C}$ such that
\begin{equation}f(x)= \sum\limits_{i=1}^p \sum\limits_{m=0}^{K_i-(n-2)}\sum\limits_{m=m_1+m_2+\cdots+m_n}
\Bigg[\wp_{\alpha,q;{\bf m}}^{{\cal{K}}_n}({\bf x}-{a}_i)b_i\Bigg].
\end{equation}
\end{theorem}
\begin{proof} Since $f$ is supposed to be in Ker $\Delta-\alpha^2$
with isolated poles at the points $a_i$ of order $K_i$ its singular
parts of the local Laurent series expansions are of the form
$E_{\alpha,q;{\bf m}}(x-{a}_i)b_i$ at each point $a_i +
\Omega$, where $E_{\alpha,q,{\bf m}}(y):=
\frac{\partial^{|{\bf m}|}}{\partial y_1^{m_1} \cdots \partial y_n^{m_n}}
E_{\alpha,q}(y)$. As a sum of functions with the properties of satisfying $(\ref{Kinv})$ and of being in Ker $\Delta-\alpha^2$ , the expression
$$
g(x) = \sum\limits_{i=1}^p \sum\limits_{m=0}^{K_i-(n-2)}\sum\limits_{m=m_1+m_2+\cdots+m_n} \Bigg[\wp_{\alpha,q,{\bf m}}^{{\cal{K}}_n}(x-{a}_i)b_i\Bigg]
$$
also satisfies (\ref{Kinv}) and has also the same principal parts as
$f$. Hence, the function $h:=g-f$ also satisfies (\ref{Kinv}) and is in Ker $\Delta-\alpha^2$. However, $h$ has no singular parts, since
these are canceled out. The function $h$ is hence an entire solution in Ker $\Delta-\alpha^2$ which satisfies (\ref{Kinv}) and must
therefore vanish as a consequence of the preceding lemma.
\end{proof}
We can adapt this theorem to the case of dealing with functions satisfying (\ref{Kinv}) that have non-isolated singularities in the
following way:
\begin{theorem}\label{repthm2}
Suppose that $f$ is a $\mathbb{C}$-valued function satisfying (\ref{Kinv}) and $(\Delta- \alpha^2) f = 0$ in $\R^n$ except in a finite number of components of non-isolated singular sets $S_1,\ldots,S_l$ of the
orders $N(S_1),\ldots,N(S_l)$ (as defined in \cite{KraHabil}) which
are supposed to be incongruent modulo $\Lambda_n$. Then there exists
functions $b_{\bf j}:S_j \to \mathbb{C}$ of bounded variation
such that
$$
f(x) = \sum\limits_{i=1}^l \sum\limits_{{\bf j} \in {\bf J}^{(i)}} \Big(\int_{S_i} \wp_{\alpha,q;{\bf j}}^{{\cal{K}}_n}(x - c^{(i)}) d[b_{\bf j}(c^{(i)})]\Big),
$$
where the integral has to be understood as Lebesgue-Stieltjes
integral as in \cite{KraHabil,Nef2}. Here we denote by ${\bf
J}^{(i)}$ the set of those multi indices ${\bf j}$ for which the
functions $b_{\bf j}^{(i)}(c^{i})$ do not vanish identically.
\end{theorem}
To establish this more general version one can adapt the arguments
of the proof to the previous theorem and the arguments using
Lebesgue-Stieltjes integrals as applied in the context of
monogenic $n$-fold periodic functions with non-isolated
singularity sets in \cite{KraHabil}, Chapter 2.
\par\medskip\par
From Theorem~\ref{repthm2} we may now readily derive by applying the projection linear map $p_*$ the following statement.
\begin{theorem}\label{repthm3}
Let $S_1',\ldots,S_l' \subset {\cal{K}}_n$ be closed subsets.
Suppose that $f'$ is an ${\cal{E}}^{(q)}$-valued section on ${\cal{K}}$ that satisfies the Klein-Gordon equation $\Delta'_{\alpha} f'=0$ on the manifold except in a finite number of components of non-isolated singular sets $S_1',\ldots,S_l'$ of the
orders $N(S_1'),\ldots,N(S_l')$. Then there exist sections $b_{\bf j}':S_j' \to {\cal{E}}^{(q)}$ of bounded variation
such that
$$
f'(x') = \sum\limits_{i=1}^l \sum\limits_{{\bf j} \in {\bf J}^{(i)}} \Big(\int_{S_i'} p_*\Big(\wp_{\alpha,q;{\bf j}}^{{\cal{K}}_n}(x' - {c'}^{(i)})\Big) d[b_{\bf j}'({c'}^{(i)})]\Big),
$$
where we use the same notation as in Theorem~\ref{repthm2}.
\end{theorem}
In the case of dealing with essential singularities the latter series is extended over infinitely many multi-indices ${\bf j}$.
\section{Acknowledgements} The author is very thankful to Dr. Denis Constales from Ghent University for the fruitful discussions on the asymptotics of the fundamental solution of the Klein-Gordon operator in $\R^n$.
|
\section{Introduction}
\label{sec:intro}
In the theory of phase transitions it is often helpful to study models in a
range of dimensions ranging from above the ``upper critical dimension'',
$d_u$, where
mean-field critical behavior is expected, to below the ``lower critical
dimension'', $d_l$, where fluctuations destroy the transition. However, it has been
difficult to do this numerically for spin glasses, since $d_u = 6$ is
quite large, and slow dynamics
prevents more than a few thousand spins being equilibrated at low temperature
$T$. It follows that, at and above $d_u$,
one cannot study a sufficient range of sizes to perform the necessary
finite-size scaling analysis.
As a result, there has been a lot of recent attention on long-range models in
one-dimension, in which the interactions fall off with a power of the
distance. Such models have a venerable history going back to
Dyson~\cite{dyson:69,dyson:71}, who considered a ferromagnet with
interactions falling off like $1/r^\sigma$, and found a
paramagnet-ferromagnet transition for $1 < \sigma
\leq 2$. Kotliar et al.~\cite{kotliar:83} studied the spin glass version of
this model, which has received a lot of attention numerically in the last few
years~\cite{katzgraber:05,katzgraber:09,leuzzi:08,moore:10,leuzzi:09}.
There are few analytical results on spin glasses beyond mean field theory. For
the long-range models, Kotliar et al.~\cite{kotliar:83} computed critical
exponents in an expansion away
from the point where mean field theory occurs ($\epsilon$-expansion),
but, as we shall see, this is
poorly converged. Hence, most of what we know has come from numerical work.
Much of this, including numerics on long-range spin glass models, has
studied the Ising version. However, there are also reasons to study models with
vector spins, such as the Heisenberg (3-component) model.
One motivation is
that Kawamura~\cite{kawamura:87} proposed that there are two separate
transitions in vector spin glasses, a spin glass transition at $T=T_{SG}$ and
a ``chiral glass transition'' at higher temperature
$T_{CG}$, involving a freezing of vortex-like variables called chiralities.
While the original scenario had
$T_{SG} = 0$, it now appears that $T_{SG}$ is non-zero in three or more
dimensions, but the question of whether
$T_{SG} < T_{CG}$, or whether there is a single transition at which both types
of ordering occur, is still
open~\cite{hukushima:00,hukushima:05,matsubara:01,endoh:01,nakamura:02,picco:05,lee:03,lee:07,campos:06,fernandez:09b,viet:10}.
A second motivation for studying the Heisenberg, rather than Ising, spin glass
is that it is possible to study larger sizes, see for example
Ref.~\cite{fernandez:09b}, which should be helpful in a finite-size scaling
analysis. In a second paper in this series~\cite{sharma:11b}, we will
investigate whether there is a de Almeida-Thouless~\cite{almeida:78} (AT) line
of transitions in a magnetic field for
Heisenberg spin glasses. This follows our recent work~\cite{sharma:10} which
shows that there \textit{is} an AT line for vector spin glasses provided one
considers a random field. The ability to study larger sizes will be
particularly useful for the AT-line study.
Here we present data for the zero field transition for the Heisenberg spin
glass for values of the parameter $\sigma$ corresponding to (i) the mean-field
regime, (ii) the non-mean-field regime, and (iii) the borderline case where
the transition disappears. Most of our results find no evidence for separate
spin-glass and chiral-glass transitions. However, for one set of parameters in
the non-mean-field regime, the data indicates that $T_{CG}$ may be
somewhat greater than $T_{SG}$.
Whether this result remains valid in the thermodynamic limit, will require
future studies on significantly larger systems.
The plan of this paper is as follows. In Sec.~\ref{sec:model} we describe the
model that we study, while in Sec.~\ref{sec:numerics} we gives some technical details
of the simulations. The results are presented in Sec.~\ref{sec:results}
and our conclusions summarized in Sec.~\ref{sec:conclusions}.
\section{Model}
\label{sec:model}
We consider the Hamiltonian
\begin{equation}
\mathcal{H} = -\sum_{\langle i, j \rangle} J_{ij} \mathbf{S}_i \cdot \mathbf{S}_j,
\label{Ham}
\end{equation}
where $\mathbf{S}_i$ are classical $3$-component Heisenberg spins of length
$1$, and the interactions
$J_{ij}$ are independent variables with zero mean and a variance
which falls off with a power of the distance between the
spins,
\begin{equation}
[J_{ij}^{2}]_{av} \propto \frac{1}{r_{ij}^{2\sigma}},
\label{J2}
\end{equation}
where $[\cdots]_{av}$ means an average over disorder. In the version used in
early studies~\cite{katzgraber:05},
every spin interacts with every other spin with a strength which falls off, on
average, like Eq.~\eqref{J2}. However, this means that the CPU time per sweep
varies as $N^2$, rather than $N$, so large sizes can not be studied.
This problem was solved by Leuzzi et. al.~\cite{leuzzi:08} who proposed a
model in which,
instead of the \textit{magnitude} of the interaction falling off with distance like
Eq.~\eqref{J2}, it is the \emph{probability} of
there being a non-zero interaction between sites $(i,j)$ which falls off,
and when an
interaction does exist, its
variance is independent of $r_{ij}$. The mean
number of non-zero interactions from a site, which we call $z$,
can be fixed, and we
take $z = 6$.
To generate the set of pairs $(i,j)$ that have an interaction with the desired
probability
we choose spin $i$ randomly, and then choose
$j \ (\ne i)$ at distance $r_{ij}$ with probability
\begin{equation}
p_{ij} = \frac{r_{ij}^{-2\sigma}}{\sum_{j\, (j\neq i)}r_{ij}^{-2\sigma}} \, ,
\label{pij}
\end{equation}
where, for $r_{ij}$, we put the sites on a circle and use the distance of the
chord, i.e.
\begin{equation}
r_{ij}=\frac{N}{\pi}\sin\left[\frac{\pi}{N}(i-j)\right].
\end{equation}
If $i$ and $j$ are already connected, we repeat the process until we find a pair
which has not been connected before. We then connect $i$ and $j$ with an
interaction picked from a Gaussian interaction
whose mean is zero and whose standard deviation is $J$, which set equal to 1.
This process is
repeated precisely $N_b = z N / 2 $ times.
The result is that each pair $(i, j)$ will be
connected with a probability $P_{ij}$ which must satisfy the condition
$N \sum_j P_{ij} = N z$ since $P_{ij}$ only depends on $|i-j|$, $P_{ii}=0$, and
there are precisely $N z /2 $ connected pairs.
It follows that, for a fixed site $i$,
\begin{equation}
\sum_j [\,J_{ij}^{2}\,]_{av} = J^2 \sum_j P_{ij} = J^2 z \, .
\label{sumJij2}
\end{equation}
Note that $P_{ij}$ is different from $z \times p_{ij}$ in Eq.~\eqref{pij} because
of the constraint
that no bond can occur twice.
The mean-field spin glass
transition temperature for $m$-component vector spins is given
by~\cite{almeida:78b}
\begin{equation}
T_{SG}^{MF} = \frac{1}{m} \left(\sum_j [\,J_{ij}^{2}\,]_{av}\right)^{1/2}
= {\sqrt{z} \over m} \, J \, ,
\label{TcMF}
\end{equation}
where the last equality follows from Eq.~\eqref{sumJij2}.
We set $J=1$ so that, for the situation here,
\begin{equation}
J = 1,\ z=6,\ m=3\, ,
\end{equation}
we have
\begin{equation}
T_{SG}^{MF} = {\sqrt{6} \over 3} \simeq 0.816\, ,
\end{equation}
the same as for the nearest-neighbor Heisenberg spin glass on a simple cubic
lattice.
By varying $\sigma$ one finds different types of
behavior~\cite{kotliar:83,katzgraber:03}, as shown in Table~\ref{tab:ranges}.
For $\sigma < 1/2$ the model is non-extensive (for instance the mean-field
transition temperature in Eq.~\eqref{TcMF} diverges) unless the interactions are
scaled by an inverse power of the system size. We will call this ``infinite
range''. The extreme limit of this, $\sigma=0$, is the
Sherrington-Kirkpatrick~\cite{sherrington:75} model, whose exact solution was
found by Parisi~\cite{parisi:79,parisi:80,parisi:83}. In fact, it has been
suggested~\cite{mori:11} (see also Ref.~\onlinecite{mori:10}) that, in the
thermodynamic limit, the behavior of the model
is \textit{identical} to that of the SK model for the \textit{whole range}
$0 \le \sigma < 1/2$.
The model is extensive for $\sigma > 1/2$ and a finite temperature transition
is expected for $\sigma < \sigma_l$, where the ``lower critical'' value is
\begin{equation}
\sigma_l = 1\, .
\end{equation}
The transition is in the mean-field universality
class~\cite{kotliar:83} for $\sigma < \sigma_u$, where the ``upper
critical'' value is
\begin{equation}
\sigma_u = 2/3\, .
\end{equation}
For $\sigma_u < \sigma < \sigma_l$,
there is a finite-temperature transition with non mean-field
critical exponents.
In this paper we will study both
mean-field and non mean-field regions. Finally for $\sigma > \sigma_l$ the
there is no transition at finite temperature.
\begin{table}[!tb]
\caption{
A summary of the behavior for different ranges of $\sigma$ in one space
dimension and at zero field.
Infinite
range means that $\sum_{j\, (j\ne i)} J_{ij}^2$ diverges unless the bonds $J_{ij}$
are scaled by an inverse power of the system size.
The behavior is mean-field like
for $\sigma < \sigma_u$ where $\sigma_u = 2/3$, and a finite-temperature
transition no longer occurs for $\sigma > \sigma_l$ where $\sigma_l = 1$.
\label{tab:ranges}
}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\fill}} l l }
\hline
\hline
$\sigma$ & Behavior \\
\hline
$\sigma = 0$ & SK model \\
$0 < \sigma \le 1/2$ & Infinite range \\
$1/2 < \sigma < 2/3$ & Mean-field with $T_{\rm SG} > 0$ \\
$2/3 < \sigma < 1$ & Non-mean-field with $T_{\rm SG} > 0$ \\
$1 < \sigma $ & $T_{\rm SG} = 0$ \\
\hline
\hline
\end{tabular*}
\end{table}
\section{Numerical Setup}
\label{sec:numerics}
We perform large scale Monte-Carlo simulations for
$\sigma = 0.6,0.75,0.85$ and $1$. From the previous section we note that
$\sigma = 0.60$ is in the mean-field regime,
$\sigma = 0.75$ and $0.85$
are in the non mean-field regime, and $\sigma = 1$ is the borderline case,
$\sigma = \sigma_l$,
beyond which there is no transition. A plausible scenario is that
$T_{SG}=0$ for $\sigma = 1$, though the possibility that $T_{SG}$ is non-zero
cannot be ruled out \textit{a priori}.
Table \ref{simparams} lists the
parameters of the simulation.
\begin{table}
\caption{
Parameters of the simulations. $N_{\rm samp}$ is the number of samples,
$N_{\rm equil}$ is the number of overrelaxation Monte Carlo sweeps for
equilibration for each of the $2 N_T$ replicas for a single sample. The same
number of sweeps is done in the measurement phase, with a measurement
performed every four overrelaxation sweeps.
The number of heatbath sweeps is equal to
10\% of the number of
overrelaxation sweeps. $T_{\rm min}$ and $T_{\rm max}$ are the
lowest and highest temperatures simulated,
and $N_T$ is the number of temperatures
used in the parallel tempering.
\label{simparams}
}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\fill}}r r r r r r r r }
\hline
\hline
$\sigma$ & $N$ & $N_{\rm samp} $ & $N_{\rm equil}$ & $T_{\rm min}$ &
$T_{\rm max}$ & $N_{T}$ \\
\hline
$0.6$ & $128$ & $16000$ & $128$ & $0.20$ & $0.70$ & $40$ \\
$0.6$ & $256$ & $16000$ & $256$ & $0.20$ & $0.70$ & $40$ \\
$0.6$ & $512$ & $16000$ & $512$ & $0.20$ & $0.70$ & $40$ \\
$0.6$ & $1024$ & $16000$ & $1024$ & $0.20$ & $0.70$ & $40$ \\
$0.6$ & $2048$ & $16000$ & $2048$ & $0.20$ & $0.70$ & $40$ \\
$0.6$ & $4096$ & $6100$ & $4096$ & $0.30$ & $0.70$ & $40$ \\
$0.6$ & $8192$ & $1000$ & $8192$ & $0.30$ & $0.70$ & $50$ \\
$0.6$ & $16384$ & $500$ & $16384$ & $0.35$ & $0.70$ & $55$ \\
$0.6$ & $32768$ & $400$ & $32768$ & $0.35$ & $0.70$ & $60$ \\
\hline
$0.75$ & $128$ & $8000$ & $128$ & $0.20$ & $0.55$ & $40$ \\
$0.75$ & $256$ & $8000$ & $256$ & $0.20$ & $0.55$ & $40$ \\
$0.75$ & $512$ & $8000$ & $512$ & $0.20$ & $0.55$ & $40$ \\
$0.75$ & $1024$ & $3000$ & $1024$ & $0.20$ & $0.55$ & $40$ \\
$0.75$ & $2048$ & $3000$ & $2048$ & $0.20$ & $0.55$ & $40$ \\
$0.75$ & $4096$ & $3000$ & $4096$ & $0.20$ & $0.55$ & $40$ \\
$0.75$ & $8192$ & $1100$ & $8192$ & $0.20$ & $0.55$ & $50$ \\
$0.75$ & $16384$ & $500$ & $16384$ & $0.25$ & $0.55$ & $55$ \\
$0.75$ & $32768$ & $400$ & $32768$ & $0.25$ & $0.55$ & $60$ \\
\hline
$0.85$ & $128$ & $16000$ & $512$ & $0.09$ & $0.30$ & $40$ \\
$0.85$ & $256$ & $16000$ & $1024$ & $0.09$ & $0.30$ & $40$ \\
$0.85$ & $512$ & $10000$ & $2048$ & $0.09$ & $0.30$ & $40$ \\
$0.85$ & $1024$ & $10000$ & $8192$ & $0.14$ & $0.22$ & $20$ \\
$0.85$ & $2048$ & $8000$ & $16384$ & $0.14$ & $0.22$ & $20$ \\
$0.85$ & $4096$ & $4000$ & $32768$ & $0.14$ & $0.22$ & $36$ \\
$0.85$ & $8192$ & $2000$ & $65536$ & $0.15$ & $0.21$ & $20$ \\
$0.85$ & $16384$ & $1700$ & $131072$ & $0.16$ & $0.21$ & $20$ \\
\hline
$1.0$ & $128$ & $2000$ & $2048$ & $0.03$ & $0.10$ & $10$ \\
$1.0$ & $256$ & $2000$ & $4096$ & $0.03$ & $0.10$ & $10$ \\
$1.0$ & $512$ & $2000$ & $16384$ & $0.02$ & $0.10$ & $40$ \\
$1.0$ & $1024$ & $1200$ & $524288$ & $0.017$ & $0.08$ & $40$ \\
$1.0$ & $2048$ & $500$ & $2097152$ & $0.017$ & $0.08$ & $60$ \\
\hline
\hline
\end{tabular*}
\end{table}
\subsection{Equilibration}
As discussed in earlier work~\cite{katzgraber:01,katzgraber:09}
there is a convenient test for
equilibration with Gaussian interactions, namely the relationship
\begin{equation}
U = {J^2 \over T} \, {z \over 2} \, (q_l - q_s) \, ,
\label{equiltest}
\end{equation}
is valid in equilibrium but the two sides approach their common equilibrium
value from opposite directions as equilibrium is approached. Here
$U = -(1/N)[\sum_{\langle i, j \rangle} \epsilon_{ij} J_{ij} \langle {\bf S}_i \cdot
{\bf S}_j \rangle ]_{\rm av} $ is
the average energy per spin,
$q_l = (1/N_{b})\sum_{\langle i, j \rangle} \epsilon_{ij}
[ \langle
{\bf S}_i \cdot {\bf S}_j \rangle^2]_{\rm av}$ is the ``link overlap'', and
$q_s = (1/N_{b})\sum_{\langle i, j \rangle} \epsilon_{ij}
[\langle ({\bf S}_i \cdot {\bf S}_j)^2
\rangle]_{\rm av}$, where $N_{b}=z N/2$,
and $\epsilon_{ij} = 1$ if there is a bond between $i$ and $j$ and is zero
otherwise.
Equation~(\ref{equiltest}) is easily derived by integrating by parts the
expression for the average energy with respect to $J_{ij}$ since it has
a Gaussian distribution. Note that in the numerics we set $J = 1$.
We determine both sides of
Eq.~\eqref{equiltest} for different numbers
of Monte Carlo sweeps (MCS) which increase
in a logarithmic manner, each value being twice the previous one. In all
cases we
average over the last half of the sweeps. We consider the data to be
equilibrated, if, when averaging over a large number of samples,
Eq.~\eqref{equiltest} is satisfied for at least the last two points.
\subsection{Simulation Technology}
To equilibrate the system in as small a number of sweeps as
possible, with the minimum amount of CPU time, we perform three types of
Monte Carlo sweeps~\cite{lee:07,campos:06,fernandez:09b}.
The workhouse of our simulation is the ``Microcanonical'' sweep~\cite{alonso:96}
(also known as an ``over-relaxation'' sweep).
We sweep sequentially through the lattice, and, at
each site, compute the local field on the spin, $\mathbf{H}_i = \sum_j J_{ij}
\mathbf{S}_j$. The new value for the spin on site $i$ is taken to be
its old value reflected about $\mathbf{H}$, i.e.
\begin{equation}
\mathbf{S}'_i = -\mathbf{S}_i + 2\, {\mathbf{S}_i \cdot \mathbf{H}_i \over
H_i^2}\, \mathbf{H}_i \, .
\label{reflect}
\end{equation}
These sweeps are microcanonical
because they preserve energy. They are very fast because the
operations are simple and no random numbers are needed. For reasons
that are not fully understood, it also seems that they ``stir up'' the
spin configuration very efficiently~\cite{campos:06} and the system
equilibrates faster than if one only uses ``heatbath'' updates,
described next, see e.g.~Fig.~9 of Ref.~[\onlinecite{pixley:08}].
We also need to do heatbath sweeps in order to change the energy.
As for the
microcanonical case, we sweep sequentially through the lattice.
We take the direction of the local field $\mathbf{H}_i$,
to be the
polar axis for the spin on
site $i$. We compute the polar and azimuthal angle of the new spin
direction relative
to the local field by the requirement that this direction occurs with the
Boltzmann probability, see Ref.~[\onlinecite{lee:07}] for details.
Finally we perform parallel tempering sweeps~\cite{hukushima:96,marinari:98b}
to prevent the system from being trapped in local minima at low temperature.
We take $N_T$ copies of the system with the same bonds but at
a range of
different temperatures.
The minimum temperature, $T_{\rm min} \equiv T_1$,
is the low temperature where one
wants to investigate the system (below $T_{SG}$ in our case), and the
maximum,
$T_{\rm max} \equiv T_{N_T}$,
is high enough that the
the system equilibrates very fast
(well above $T_{SG}$ in our case). A parallel tempering sweep consists
of swapping the temperatures of the spin configurations at a pair of
neighboring temperatures, $T_i$ and $T_{i+1}$, for $i = 1, 2, \cdots ,
T_{N_T - 1}$ with a probability that satisfies the detailed balance
condition. The Metropolis probability for this is~\cite{hukushima:96}
\begin{equation}
\quad P(T\ \mbox{swap}) = \left\{
\begin{array}{ll}
\exp(\Delta \beta \, \Delta E), & (\mbox{if} \ \Delta \beta \, \Delta E
< 0), \\
1, & (\mbox{otherwise}),
\end{array}
\right.
\label{PTswap}
\end{equation}
where $\Delta \beta= 1/T_i - 1/T_{i+1}$ and $\Delta E = E_i - E_{i+1}$,
in which $E_i$ is the energy of the copy at temperature $T_i$. In this
way, a given set of spins (i.e.~a copy)
performs a random walk in temperature space.
We perform one parallel tempering sweep for
every ten overrelaxation sweeps. Since there are two copies of spins
\textit{at each
temperature}, indicated by labels ``$(1)$'' and ``$(2)$'' in Eq.~\eqref{qmunu}
below, we actually perform parallel tempering sweeps among the set of $N_T$
copies labeled ``$(1)$'' and, separately, among the set of $N_T$ copies
labeled ``$(2)$''.
\subsection{Quantities Measured}
The main quantities measured in this simulation are the spin glass
susceptibility $\chi_{SG}$, and the chiral glass susceptibility $\chi_{CG}$,
at wavevectors $k=0$, and $k=2\pi/N$, and from these we obtain the two corresponding
correlation lengths, $\xi_{SG}$ and $\xi_{CG}$. The spin
glass order parameter, $q^{\mu\nu}({k})$, at wave vector ${k}$, is defined to
be
\begin{equation}
q^{\mu\nu}({k}) = {1 \over N} \sum_{i=1}^N S_i^{\mu(1)} S_i^{\nu(2)}
e^{i {k} \cdot {R}_i},
\label{qmunu}
\end{equation}
where $\mu$ and $\nu$ are spin components, and ``$(1)$'' and ``$(2)$''
denote two
identical copies of the system with the same interactions. We run two copies
of the system at each temperature in order to evaluate quantities such as the
spin glass susceptibility, defined in Eq.~\eqref{chisgk} below, without bias.
From this we
determine the wave vector dependent
spin glass susceptibility $\chi_{SG}({k})$ by
\begin{equation}
\chi_{SG}({k}) = N \sum_{\mu,\nu} [\langle \left|q^{\mu\nu}({k})\right|^2 \rangle ]_{\rm av} ,
\label{chisgk}
\end{equation}
where $\langle \cdots \rangle$ denotes a thermal average and
$[\cdots ]_{\rm av}$ denotes an average over disorder. The spin glass correlation
length is then determined from
\begin{equation}
\xi_{SG} = {1 \over 2 \sin (k_\mathrm{min}/2)}
\left({\chi_{SG}(0) \over \chi_{SG}({k}_\mathrm{min})} - 1\right)^{1/(2\sigma-1)},
\label{chisg}
\end{equation}
where ${k}_\mathrm{min} = (2\pi/L)$.
For the Heisenberg spin glass, Kawamura
defines the local chirality in terms of three spins on a line as
follows~\cite{hukushima:00}:
\begin{equation}
\kappa_i = {\bf S}_{i+1} \cdot {\bf S}_i \times {\bf S}_{i-1}.
\label{eq:chiral_heis}
\end{equation}
The chiral glass susceptibility is then given by
\begin{equation}
\label{chicg}
\chi_{CG}({k}) = N [\langle \left| q_{c}({k})\right|^2
\rangle ]_{\rm av} ,
\end{equation}
where the chiral overlap $q_{c}({\bf k})$ is given by
\begin{equation}
\label{qc}
q_{c}({k}) = {1 \over N} \sum_{i=1}^N \kappa_i^{(1)} \kappa_i^{(2)}
e^{i {k} \cdot {R}_i}.
\end{equation}
We define
the chiral correlation length by
\begin{equation}
\label{xi_c}
\xi_{CG} = {1 \over 2 \sin (k_\mathrm{min}/2)}
\left({\chi_{CG}(0) \over \chi_{CG}({k}_\mathrm{min})} - 1\right)^{1/(2\sigma-1)}.
\end{equation}
As will be revealed in the next section, three of the four quantities defined
above, $\chi_{SG},\xi_{SG},$ and $\chi_{CG}$ may be used in a
finite-size-scaling analysis to locate and analyze the phase transition.
\subsection{Finite-Size Scaling}
According to finite-size
scaling~\cite{fss:gtlcd}, the correlation length of the finite-system varies,
near the transition temperature $T_c$,
as
\begin{subequations}
\label{eq:xiscale}
\begin{align}
\label{xi_nonmf}
{\xi \over N} &= {\mathcal X} [ N^{1/\nu} (T - T_c) ]
\; , \ (2/3 \le \sigma < 1),
\\
{\xi \over N^{\nu/3}} &= {\mathcal X} [ N^{1/3} (T - T_c) ]
\;,\ (1/2 <\sigma \le 2/3),
\label{xi_mf}
\end{align}
\end{subequations}
in which $\nu$, the correlation length exponent, is given, in the mean-field
regime, by $\nu = 1/(2\sigma - 1)$. We will use Eq.~\eqref{eq:xiscale}
for both the spin glass correlation length $\xi_{SG}$, in which $T_c$
will be set to $T_{SG}$, and the chiral glass correlation length $\xi_{CG}$,
in which $T_c$ will be set to $T_{CG}$. It follows that,
if there is a transition at $T = T_c$,
data for $ {\xi}/{N}$ ($ {\xi}/{N^{\nu/3}}$ in the mean-field region)
for different system sizes $N$ should cross at $T_c$.
We also present data for $\chi_{SG} \equiv \chi_{\rm SG}(0)$, which
has the finite-size scaling form
\begin{subequations}
\label{eq:chisgscale}
\begin{align}
\label{chi_nonmf}
{\chi_{\rm SG} \over N^{2 -\eta}} &= {\mathcal C}[N^{1/\nu} (T - T_c)]
\; , \ (2/3 \le \sigma < 1),
\\
{\chi_{\rm SG} \over N^{1/3}} &= {\mathcal C}[N^{1/3} (T - T_c)]
\; , \ (1/2 < \sigma \le 2/3).
\label{chi_mf}
\end{align}
\end{subequations}
Hence curves of $\chi_{\rm SG}/N^{2 - \eta}$ ($\chi_{\rm SG} /
N^{1/3}$ in the mean-field regime) should also intersect. This is particularly
useful for long-range models since $\eta$ is given by the simple expression $2
- \eta = 2 \sigma - 1$ {\it exactly}. However, we do not know the exponent
corresponding to $\eta$ for the chiral glass susceptibility, so we will not
use this quantity in the finite-size scaling analysis.
In practice, there are corrections to this finite-size-scaling, so
data for different sizes do not all intersect at the exactly the same temperature.
Including leading corrections to scaling, the intersection
temperature $T^{*}(N,2N)$ for sizes $N$ and $2N$ varies
as~\cite{binder:81b,ballesteros:96a,hasenbusch:08b,larson:10}
\begin{equation}
T^{*}(N,2N) = T_{c} + \frac{A}{N^{\lambda}},
\label{Tstar}
\end{equation}
where $A$ is the amplitude of the leading correction, and the
exponent $\lambda$ is given by
\begin{equation}
\lambda = \frac{1}{\nu}+\omega
\label{lambda}
\end{equation}
where $\omega$ is the leading correction to scaling
exponent.
\section{Results}
\label{sec:results}
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{both_0.6.eps}
\caption{(Color online)
The main figure shows
data for $\xi_{SG}/N^{5/3}$, in which the power of $N$ is chosen following
Eq.~\eqref{xi_mf} with $\nu=1/(2\sigma-1)$,
as a function of $T$ for
different system sizes for $\sigma=0.6$. The inset shows data for
$\chi_{SG}/N^{1/3}$, in which the power of $N$ is chosen following Eq.~\eqref{chi_mf}.
}
\label{fig:both_0.6}
\end{center}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{Tstar_0.6.eps}
\caption{(Color online)
The intersection temperatures, $T^{*}(N,2N)$, obtained from the data
in Fig.~\ref{fig:both_0.6}, for
$\xi_{SG}/N^{5/3}$ and $\chi_{SG}/N^{1/3}$ for $\sigma = 0.6$,
as a
function of $N^{-\lambda}$, with $\lambda = 0.467$ determined from
Eq.~\eqref{lambda} (which is valid
in the MF regime). A fit using all 8 points from $\xi_{SG}$
gives $T_{SG} = 0.564 \pm 0.002$, while a fit using the data for the largest
5 pairs of sizes from $\chi_{SG}$ gives
$T_{SG} = 0.562 \pm 0.002$.
}
\label{fig:Tstar_0.6}
\end{figure}
\subsection{$\sigma = 0.6$ (mean-field regime)}
As shown in Table~\ref{tab:ranges}, $\sigma=0.6$
lies well inside the mean-field regime.
Hence, according to Eq.~\eqref{xi_mf}, results for
$\xi_{SG}/N^{\nu/3}$ should intersect at $T_{SG}$ with
$\nu$ set equal to $1/(2\sigma-1)$.
The data is shown in the main part of
Fig.~\ref{fig:both_0.6}. The intersections do not occur at
precisely the same temperature, but fitting the intersection temperatures to
Eq.~\eqref{Tstar} is helped by the fact that we know
$\lambda \equiv \frac{5}{3}-2\sigma$ in the MF regime~\cite{larson:10},
which gives a value $0.467$ here.
A straight line fit of $\xi_{SG}/N^{\nu/3}$ against $N^{-\lambda}$,
shown in Fig.~\ref{fig:Tstar_0.6}, gives $T_{SG}=0.564 \pm 0.002$.
The inset to Fig.~\ref{fig:both_0.6}
shows data for $\chi_{SG}/N^{1/3}$, which should also intersect
at $T_{SG}$ according to
Eq.~\eqref{chi_mf}.
This time, we find that corrections to scaling are well described
Eq.~\eqref{Tstar}
but only if we consider just
the largest five pairs of sizes. The fit to Eq.~\eqref{Tstar},
shown in Fig.~\ref{fig:Tstar_0.6}, gives $T_{SG}=0.562 \pm 0.002$, which is
consistent with that obtained from the spin glass correlation
length.
We have also measured the chiral glass correlation length.
However, we find
that
$\xi_{CG}/N^{\nu/3} \lesssim 10^{-12}$ in the vicinity of $T_{SG}$.
Hence chiralities can not play an important role in this range of temperature.
\begin{figure}
\includegraphics[width=9cm]{both_deriv_0.6.eps}
\caption{(Color online)
The main figure is a log-log plot of the logarithmic derivative of $\xi_{SG}$
for $\sigma=0.6$
for different sizes $N$ evaluated at the intersection temperatures $T^\star(N,
2N)$ shown in Fig.~\ref{fig:Tstar_0.6}. According to
Eq.~\eqref{xi_mf} the slope is expected to be $1/3$. The
best fit is a little smaller than this, indicating that corrections to scaling are still
present for these sizes. The inset is the same but for $\chi_{SG}$.
}
\label{fig:both_deriv_0.6}
\end{figure}
According to Eqs.~\eqref{xi_mf} and \eqref{chi_mf} the argument of the scaling
functions is $N^{1/3}(T - T_c)$. Hence, at $T_{SG}$ the, logarithmic derivative
of $\xi_{SG}$ and $\chi_{SG}$ should vary as $N^{1/3}$. As we have seen, the
data do not all intersect at the same temperature, and so we evaluate the
derivatives at the intersection temperatures $T^\star$ plotted in
Fig.~\ref{fig:Tstar_0.6}. The results are shown in
Fig.~\ref{fig:both_deriv_0.6}. We get a power of 0.28 from $\xi_{SG}$ and 0.29
from $\chi_{SG}$,
in both cases a little less than
1/3, indicating that there are still some corrections to scaling even for
these large sizes.
\subsection{$\sigma = 0.75$ (non mean-field regime)}
\begin{figure}
\includegraphics[width=9cm]{both_0.75.eps}
\caption{(Color online)
The main figure shows data for different sizes for $\xi_{SG}/N$ at $\sigma =
0.75$ which,
according to Eq.~\eqref{xi_nonmf}, should intersect at $T_{SG}$. The inset
shows data for $\chi_{SG}/N^{2-\eta}$ (with $2-\eta = 2\sigma - 1$), which should
also intersect at $T_{SG}$ according to Eq.~\eqref{chi_nonmf}.
}
\label{fig:both_0.75}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{Tstar_0.75.eps}
\caption{(Color online)
The intersection temperatures for $\sigma = 0.75$. The all 8 data points
from $\chi_{SG}$ were fitted
to Eq.~\eqref{Tstar}, with the result that $\lambda = 0.44 \pm 0.13$. The same
exponent was then used to fit the results for
the largest 7 (pairs of) sizes from
$\xi_{SG}$, for which sub-leading
corrections to scaling seem to be more important. The resulting values for the
transition temperature are $T_{SG} = 0.354 \pm 0.005$
from $\chi_{SG}$, and
$T_{SG} = 0.359 \pm 0.003$ from $\xi_{SG}$.
}
\label{fig:Tstar_0.75}
\end{figure}
For $\sigma = 0.75$ we are no longer in the MF regime.
Hence, according to Eqs.~\eqref{xi_nonmf} and
\eqref{chi_nonmf}, $\xi_{SG}/N$
and $\chi_{SG}/N^{2-\eta}$ (with $2-\eta = 2 \sigma - 1$), should
intersect at $T_{SG}$. The data
are shown in Fig.~\ref{fig:both_0.75} and the resulting intersection
temperatures are shown in Fig.~\ref{fig:Tstar_0.75}.
We fit the intersection temperatures to Eq.~\eqref{Tstar}, but unfortunately
we do not know the value of the exponent $\lambda$ outside the MF region, and
have to include it as fit parameter.
The fit is therefore to
a non-linear function of the parameters. We determine the fit
parameters
using the Levenberg-Marquardt algorithm~\cite{press:92}.
The data from
$\chi_{SG}$ is better behaved than the data from $\xi_{SG}$ so we use the
former to determine the exponent $\lambda$ and then fix this value in the fit
(which did not include the smallest size)
to the data from $\xi_{SG}$. This procedure is justified since the exponent
giving the leading correction to scaling, $\lambda$, is universal, though the
amplitude of this correction ($A$ in Eq.~\eqref{Tstar}) is non-universal.
The results are $\lambda = 0.44 \pm 0.13$,
$T_{SG} = 0.359 \pm 0.003$ from $\xi_{SG}$, and $T_{SG} = 0.354 \pm 0.005$
from $\chi_{SG}$. The two estimates for $T_{SG}$ agree within the error bars.
The data for $\xi_{CG}/N$ in the region of the spin glass transition
temperature is very small, around $10^{-4}$. Hence, as for $\sigma = 0.6$,
chiralities do not play an important role in the transition.
According to Eqs.~\eqref{xi_nonmf} and \eqref{chi_nonmf}, adapted to include
corrections to finite-size scaling, the logarithmic
derivative of $\xi_{SG}$ and $\chi_{SG}$ should vary as $N^{1/\nu}$ at
$T^\star(N, 2N)$. The plots in Fig.~\ref{fig:both_deriv_0.75} yield
$1/\nu_{SG} = 0.25$ from $\xi_{SG}$ and $0.29$ from $\chi_{SG}$. The
difference presumably comes from corrections to scaling.
Kotliar et al.~\cite{kotliar:83} calculated critical exponents to leading
order in $\epsilon = \sigma - 2/3$ with the result
\begin{equation}
{1 \over \nu} = {1 \over3}\, \left(1 - 12 \epsilon + O(\epsilon^2) \right) \, .
\label{eps_exp}
\end{equation}
The large coefficient of $\epsilon$ indicates the expansion becomes poorly
converged well before the ``lower critical'' value $\epsilon = 1/3$ ($\sigma =
1$). Even for the present value of $\sigma$,
which corresponds to $\epsilon= 1/12$, Eq.~\eqref{eps_exp} gives $1/\nu = 0$,
and so is not useful for comparison with the numerics.
\begin{figure}
\includegraphics[width=9cm]{both_deriv_0.75.eps}
\caption{(Color online)
The main figure is a log-log plot of the logarithmic derivative of $\xi_{SG}$
for $\sigma = 0.75$
for different sizes $N$ evaluated at the intersection temperatures $T^\star(N,
2N)$ shown in Fig.~\ref{fig:Tstar_0.75}. According to
Eq.~\eqref{xi_nonmf} the slope is expected to be $1/\nu$.
The inset is the same but for $\chi_{SG}$.
}
\label{fig:both_deriv_0.75}
\end{figure}
\subsection{$\sigma = 0.85$ (non mean-field regime)}
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{both_0.85.eps}
\caption{(Color online)
The main figure shows data for $\xi_{SG}/N$,
as a function of $T$ for different system sizes for $\sigma=0.85$.
The inset shows data for $\chi_{SG}/N^{2-\eta}$ with $2-\eta = 2\sigma - 1$.
}
\label{fig:both_0.85}
\end{center}
\end{figure}
\begin{figure}
\includegraphics[width=9cm]{xi_CG_0.85.eps}
\caption{(Color online)
Data for $\xi_{CG}/N$, the spin glass correlation length divided by system
size, as a function of $T$ for different system sizes for $\sigma=0.85$.
}
\label{fig:xi_CG_0.85}
\end{figure}
\begin{figure}
\includegraphics[width=\columnwidth]{Tstar_0.85.eps}
\caption{(Color online)
The intersection temperatures for $\sigma = 0.85$. The data from $\chi_{SG}$
was fitted
to Eq.~\eqref{Tstar}, with the result that $\lambda = 0.99 \pm 0.13$. This is
the value used to plot all the data.
The same
exponent was then used to fit the largest four sizes in the
data from $\xi_{SG}$, for which sub-leading corrections appear to be very
large.
The full Levenberg-Marquardt fit to the data for $\xi_{CG}$ was unable to determine
$\lambda$ with any precision, see text.
However, in estimating $T_{CG}$ and its error
bar from data for $\xi_{CG}$ we
allowed $\lambda$ to vary.
The resulting values for the
transition temperatures are
$T_{SG} = 0.167 \pm 0.001$ from $\chi_{SG}$ (all 7 points),
$T_{SG} = 0.166 \pm 0.004$ from $\xi_{SG}$ (4 points), and
$T_{CG} = 0.190 \pm 0.006$ from $\xi_{CG}$ (5 points).
We emphasize that, in these estimates, we fixed the
value of $\lambda$ only for the data from $\xi_{SG}$.
}
\label{fig:Tstar_0.85}
\end{figure}
For $\sigma = 0.85$ we are further in the non-mean-field region. Among the
different models studied here, this is the one which is most similar to
a short-range model in three dimensions.
The spin
glass data is shown in Fig.~\ref{fig:both_0.85}. Chiral correlations are
larger than for $\sigma
= 0.6$ and $0.75$, so we show data
for $\xi_{CG}$ in Fig.~\ref{fig:xi_CG_0.85}.
As for $\sigma = 0.75$, to extrapolate the intersection temperatures to the
thermodynamic limit, we resort to Levenberg-Marquardt fits with
three parameters. Using data from $\chi_{SG}$ we find $\lambda_{SG} = 0.99 \pm
0.13$. The data
from $\xi_{CG}$ is insufficient to determine $\lambda_{CG}$ since the fits give
$\lambda_{CG} = 0.79 \pm 0.74$, i.e. the error bar is as large as the best
estimate. Nonetheless, in estimating $T_{CG}$ and its error bar from the data for
$\chi_{CG}$, we allow $\lambda$ to vary. For $\xi_{SG}$ there
appear to be very large sub-leading corrections, so
we fixed the value of $\lambda_{SG}$ to that obtained from $\chi_{SG}$ when
fitting the results from $\xi_{SG}$.
Results for $T^\star(N, 2N)$ against $1/N^\lambda$ and fits are shown
in Fig.~\ref{fig:Tstar_0.85}. In the \textit{plot}, for all data
we use the value of $\lambda$ determined from $\chi_{SG}$. However,
we again emphasize that,
in the
\textit{fit} to the $\chi_{CG}$ data, we estimated $T_{CG}$ and its error bar by
allowing $\lambda$ to vary.
From the fits, we
find $T_{SG} = 0.167 \pm 0.001 $ from $\chi_{SG}$,
$T_{SG} = 0.166 \pm 0.004$ from $\xi_{SG}$, and
$T_{CG} = 0.190 \pm 0.006$ from $\xi_{CG}$.
The two estimates of $T_{SG}$ agree
with each other but are lower than $T_{CG}$. This would imply spin-chirality
decoupling for $\sigma = 0.85$. However, we note that the data for the spin
glass correlation length appears, at intermediate sizes, to be
extrapolating to a value for
$T_{SG}$ of around 0.19 (our value for $T_{CG}$) but then,
for the largest sizes, veers down to about
$0.167$ (very close to our value for $T_{SG}$ obtained from $\chi_{SG}$).
Hence we cannot rule out the possibility that a similar
``crossover'' may occur for the chiral glass correlation length data, but at
even larger sizes. If so, then spin-chirality decoupling would not occur at
the largest scales. We also note that the actual values of
$\xi_{CG}/N$ shown in Fig.~\ref{fig:xi_CG_0.85} are still very small in the
vicinity of
$T_{CG}$, about 1/50 of
the value of $\xi_{SG}/N$ around the transition,
see Fig.~\ref{fig:both_0.85}. Hence we are very far the regime with $\xi_{CG}
> \xi_{SG}$ which will ultimately occur in the presence of spin-chirality
decoupling.
Figure \ref{fig:both_deriv_0.85} shows results for the logarithmic derivative
of $\xi_{SG}$ and $\chi_{SG}$ evaluated at $T^\star(N, 2N)$. Fits give
$1/\nu$ = 0.22 from $\xi_{SG}$ and 0.29 from $\chi_{SG}$.
The curvature in the data for $\xi_{SG}$ indicates strong
finite-size corrections. Presumably these
corrections are also the reason for the
difference between the two estimates for $1/\nu$.
Figure \ref{fig:xi_CG_deriv_0.85} shows similar data but for $\xi_{CG}$. The
best fit gives $1/\nu = 0.21$, which is close to the estimate from $\xi_{SG}$.
Note that the coefficient of $N^{1/\nu}$,
$14.01$, is much larger than the corresponding value,
3.18, for $\xi_{SG}$, presumably to compensate for the overall value of
$\xi_{CG}$ being much less than that of $\xi_{SG}$ in the vicinity of the
intersection temperatures $T^\star$.
\begin{figure}
\includegraphics[width=9cm]{both_deriv_0.85.eps}
\caption{(Color online)
The main figure is a log-log plot of the logarithmic derivative of $\xi_{SG}$
for $\sigma = 0.85$
for different sizes $N$
evaluated at the intersection temperatures $T^\star(N,
2N)$ shown in Fig.~\ref{fig:Tstar_0.85}. According to
Eq.~\eqref{xi_nonmf} the slope is expected to be $1/\nu$.
The inset is the same but for $\chi_{SG}$.
}
\label{fig:both_deriv_0.85}
\end{figure}
\begin{figure}
\includegraphics[width=9cm]{xi_CG_deriv_0.85.eps}
\caption{(Color online)
Similar to Fig.~\ref{fig:both_deriv_0.85} but for $\xi_{CG}$.
}
\label{fig:xi_CG_deriv_0.85}
\end{figure}
\subsection{$\sigma = 1\quad (= \sigma_l)$}
\begin{figure}
\includegraphics[width=9cm]{xi_SG_1.0.eps}
\caption{(Color online)
Data for $\xi_{SG}/N$, the spin glass correlation length divided by system
size, as a function of $T$ for different system sizes, for $\sigma=1.0$. The
inset shows the intersection temperatures $T^\star(N, 2N)$, as well as
a log-log fit assuming $T_{SG}=0$. The fit works well for $N=128,256$ and $512$.
There is no intersection
for the two lowest sizes, $N=1024$ and $2048$ for $T$ greater than the lowest
temperature we could simulate, $0.017$. This temperature is well
\textit{below} the value of the fit extrapolated to $N=1024$, which is about
0.042. Hence the intersection temperatures actually fall off \textit{faster}
at large sizes than shown in the fit.
}
\label{fig:corrsg1}
\end{figure}
\begin{figure}
\includegraphics[width=9cm]{xi_CG_1.0.eps}
\caption{(Color online)
Data for $\xi_{CG}/N$, the chiral glass correlation length divided by system
size, as a function of $T$ for different system sizes, for $\sigma=1.0$. The
inset is a log-log fit to the intersection temperatures
$T^\star(N, 2N)$, assuming $T_{CG}=0$. The data is quite consistent with this
behavior.
}
\label{fig:corrcg1}
\end{figure}
\begin{figure}
\includegraphics[width=9cm]{chi_SG_1.0.eps}
\caption{(Color online)
Data for $\chi_{SG}/N$, the spin glass susceptibility divided by system size
(the power $2-\eta$ is equal to 1 here), as a function of $T$ for different
system sizes for $\sigma=1.0$. There are no intersections for the
temperature-range in which the simulation is conducted.
}
\label{fig:khisg1}
\end{figure}
It is known from the early work of Kotliar et al.~\cite{kotliar:83} that
$\sigma = 1$ is the ``lower critical'' value $\sigma_l$, above which there is no spin
glass transition. Interestingly, Viet and
Kawamura~\cite{viet:10} claim that chiral
glass ordering persists to slightly \textit{larger} values of $\sigma$.
Testing this claim is one of our main motivations for performing
simulations at $\sigma = 1$.
Figure \ref{fig:corrsg1} shows the finite scaling for $\xi_{SG}$. In the inset,
we show a log-log plot of $T^{*}(N,2N)$ versus $N$ and include a straight-line fit for
$N = 128, 256$ and $512$. This fit works well. We find no intersections in the
range of $T$ that we can equilibrate $(T \ge 0.17)$ for $N=1024$ and 2048.
Hence the data is well consistent with $T_{SG} =0$.
Figure \ref{fig:corrcg1} shows the corresponding figure for
$\xi_{CG}$. Again, the inset shows log-log fits assuming that the transition
temperature, $T_{CG}$ in this case, is zero.
The fit is satisfactory and indicates that we do not find a
finite value for $T_{CG}$ at $\sigma = 1$, in contrast to the claim of
Viet and Kawamura~\cite{viet:10}.
Figure \ref{fig:khisg1} shows the data for $\chi_{SG}$. There are no
intersections at all in the range of temperature that we can equilibrate,
consistent with the conclusion from the $\xi_{SG}$ data that $T_{SG} =
0$.
\section{Conclusions}
\label{sec:conclusions}
Our primary motivation to study
the Heisenberg spin glass in one dimension with long-range
interactions which fall off with the power of the distance, is to test
Kawamura's
spin-chirality decoupling scenario in which $T_{CG} > T_{SG}$, and his
subsequent prediction~\cite{viet:10}
that chiral glass ordering persists for $\sigma >
\sigma_l$, where $\sigma_l = 1$ is the ``lower critical'' value for the spin
glass transition, with a finite value of $T_{CG}$ \textit{at} $\sigma = 1$.
For $\sigma = 1$ we find $T_{CG} = T_{SG} =0$ in contrast to Viet and
Kawamura~\cite{viet:10}. For most of the other values of $\sigma$
our data is well consistent with a single phase transition.
However, for $\sigma = 0.85$ the
best fits for the sizes we can equilibrate indicate $T_{CG} > T_{SG}$, see
Fig.~\ref{fig:Tstar_0.85}. Interestingly that figure shows very strong
subleading corrections to finite-size scaling in the data for $\xi_{SG}$ since
it only fits Eq.~\eqref{Tstar} for the largest sizes. At
intermediate sizes the intersection temperatures seem to be heading towards
the chiral glass transition temperature obtained from the $\xi_{SG}$ data,
but then dip down for the largest
sizes. We therefore cannot rule out that similar behavior
might occur for the chiral glass correlation length but at even larger length
scales. In this case there would be no spin-chirality decoupling. We also
note that, for a given size and temperature, $\xi_{CG}$ remains considerably
smaller than $\xi_{SG}$ in the vicinity of the intersection temperatures
$T^\star$, compare the main part of Fig.~\ref{fig:both_0.85}
with Fig.~\ref{fig:xi_CG_0.85}.
Hence the data is very far from the regime with
$\xi_{CG} > \xi_{SG}$ which ultimately prevails for $T_{SG} < T< T_{CG}$
if there is spin-chirality decoupling.
In a subsequent paper~\cite{sharma:11b}, we will investigate,
for the same models, under what
circumstances an AT line of transitions occurs in a magnetic field.
\acknowledgments
We acknowledge support from the NSF under Grant DMR-0906366. AS also acknowledges
partial support from DOE under Grant No.~FG02-06ER46319. We are grateful
for a generous allocation of computer time from the Hierarchical Systems
Research Foundation.
|
\section{Understanding 3-separations in the context of regular elements}
Let $M$ be a 3-connected non-regular binary matroid such that $M$ is the 3-sum of matroids $M_1$ and $M_2$ where $|E(M_1)|, |E(M_2)|\ge 7$, $E(M_1)\cap E(M_2)=T$ and $T$ is a triangle in $M_1$ and $M_2$. Assume that $e\in E(M_1)-E(M_2)$ is a regular element of $M$.
\bigskip
\noindent {\bf Lemma 2.1.} {\it The element $e$ is not spanned by $E(M_2)-E(M_1)$ in $M$.}
\bigskip
\noindent {\bf Proof.} Suppose $e$ is spanned by $E(M_2)-E(M_1)$ in $M$. Then $e$ is spanned by $T$ in $M_1$ and so $e$ is in parallel to some element $t\in T$. By hypothesis, $M\backslash e$ is regular. Observe that $M_1\backslash e$ and $M_2$ are regular because:
\begin{enumerate}
\item[(i)] when $|E(M_1)|>7$, $M\backslash e$ is the 3-sum of $M_1\backslash e$ with $M_2$; and
\item[(ii)] when $|E(M_1)|=7$, $M_1\backslash e$ has 6 elements and is isomorphic to $M(K_4)$. So $M\backslash e$ is obtained from $M_2$ after a $\Delta-Y$ operation along the triangle $T$.
\end{enumerate}
But $M_1$ is obtained from $M_1\backslash e$ by adding $e$ in parallel with $t$. Therefore $M_1$ and $M_2$ are regular; a contradiction because the class of regular matroids is closed under 3-sums. Thus $e$ is not spanned by $E(M_2)-E(M_1)$ in $M$. $\qed$
\bigskip
\noindent {\bf Lemma 2.2.} {\it The element $e$ is not spanned by $E(M_2)-E(M_1)$ in $M^*$.}
\bigskip
\noindent {\bf Proof.} If $N_i$ is obtained from $M_i$ by a $\Delta - Y$ operation along the triangle $T$, then $M^*$ is the 3-sum of $N_1^*$ and $N_2^*$. Applying Lemma 2.1, we conclude that $e$ is not spanned by $E(M_2)-E(M_1)$ in $M^*$. $\qed$
\bigskip
In the next result, we describe how the presence of a regular element in $M_1$ impacts the structure of $M$. We prove that one of two situations occur: either $M_1$ is non-regular with $e$ as a regular element and $M_2$ is regular or $M_2$ is non-regular and $M_1$ is a small matroid with with a specific structure. In the latter situation we prove that $E(M_1)-T=T'\cup T^*$ where $T'$ is a triangle and $T^*$ is a triad such that $e\in T\cap T^*$ and $E(M_1)-E(M_2)$ is closed in $M$. Since $M$ is binary, a triangle and triad must intersect in an even number of elements. This means $M_1$ has just 7 elements, one of which is parallel with an element of $T$.
\bigskip
\noindent {\bf Lemma 2.3.} {\it
\begin{enumerate}
\item[(i)] $M_2$ is a regular matroid; or
\item[(ii)] there is a triangle $T'$ and a triad $T^*$ of $M$ such that $e\in T'\cap T^*$ and $E(M_1)-T=T'\cup T^*$.
\end{enumerate}
Moreover,
\begin{enumerate}
\item[(iii)] when (i) happens, $M_1$ is a non-regular matroid having $e$ as a regular element; and
\item[(iv)] when (ii) happens, $E(M_1)-E(M_2)$ is closed in $M$.
\end{enumerate}}
\bigskip
\noindent {\bf Proof.} Assume that (i) does not hold, that is,
\begin{equation}\label{27.01.11.a}
\hbox{$M_2$ is non-regular.}
\end{equation}
First, we establish that
\begin{equation}\label{26.01.11.a}
r(M_1)=3\hbox{ or ${\rm si}(M/e)$ is not 3-connected.}
\end{equation}
Suppose that $r(M_1)\ge 4$ and ${\rm si}(M/e)$ is 3-connected. If $T'$ is a triangle of $M$ containing $e$, then, by Lemma 2.1, $|E(M_2)\cap T'|\le 1$. Therefore we may assume that ${\rm si}(M/e)=M/e\backslash X$, for $X\subseteq E(M_1)-T$. If $M_1/e\backslash X\simeq M(K_4)$, then $M_2$ is obtained from ${\rm si}(M/e)$ after a $Y-\Delta$ operation along the triad $E(M_1)-(e\cup X\cup T)$. So $M_2$ is regular; a contradiction to~\eqref{27.01.11.a}. If $M_1/e\backslash X\not\simeq M(K_4)$, then ${\rm si}(M/e)$ is the 3-sum of $M_1/e\backslash X$ and $M_2$. As ${\rm si}(M/e)$ is regular, it follows that $M_2$ is regular; a contradiction to~\eqref{27.01.11.a}. We have~\eqref{26.01.11.a}.
If $N_i$ is obtained from $M_i$ by a $\Delta - Y$ operation along the triangle $T$, then $M^*$ is the 3-sum of $N_1^*$ and $N_2^*$. Note that Lemma 2.3(i) holds for the decomposition $M=M_1\bigtriangleup M_2$ if and only if Lemma 2.3(i) holds for the decomposition $M^*=N_1^*\bigtriangleup N_2^*$. The analogous statement occurs when we replace (i) by (ii). Therefore, the dual of~\eqref{26.01.11.a} becames
\begin{equation}\label{07.02.11.a}
r(N_1^*)=3\hbox{ or $[{\rm co}(M\backslash e)]^*={\rm si}(M^*/e)$ is not 3-connected.}
\end{equation}
By Bixby's Theorem [5, 8.4.6], ${\rm si}(M/e)$ or ${\rm co}(M\backslash e)$ is 3-connected. By~\eqref{26.01.11.a} and~\eqref{07.02.11.a}, $r(M_1)=3$ or $r(N_1^*)=3$. Taking the dual when necessary, we may assume that
\begin{equation}\label{07.02.11.b}
r(M_1)=3.
\end{equation}
Next, we prove the following claim:
\bigskip
\noindent {\bf Claim:} $M_1$ does not have a minor $N$ such that $T$ and $T'=E(N)-T$ are triangles of $N$, $e\not\in E(N)=T\cup T'$ and $r(N)=2$.
\bigskip
Suppose that $N$ exists, say $N=M_1\backslash X/Y$. By hypothesis, $e\in X\cup Y$ and so $M\backslash X/Y$ is regular. Moreover, $M\backslash X/Y$ is isomorphic to $M_2$. Thus $M_2$ is regular; a contradiction to~\eqref{27.01.11.a}. Therefore the claim holds.
\bigskip
If ${\rm si}(M_1)\simeq F_7$, then $M_1/e$ is a rank-2 matroid. By Lemma 2.1, $M_1/e$ has $T$ as a triangle. We have a contradiction by the Claim because every parallel class of $M_1/e$ is non-trivial. Hence, by~\eqref{07.02.11.b}, ${\rm si}(M_1)\simeq M(K_4)$. In particular, $T^*=E(M_1)-{\rm cl}_{M_1}(T)$ is a triad of $M_1$. By Lemma 2.1, $e\in T^*$, say $T^*=\{e,e_1,e_2\}$. Let $f_1,\dots,f_k$ be the elements of ${\rm cl}_{M_1}(T)-T$. For each $i$, there is $t_i\in T$ such that $\{f_i,t_i\}$ is a parallel class of $M_1$. By the Claim, $k\le 2$. Next, we establish that
\begin{equation}\label{26.01.11.b}
k=1.
\end{equation}
As $|E(M_1)|\ge 7$ and $|E(M_1)-{\rm cl}_{M_1}(T)|=3$, it follows that $k\ge 1$. If~\eqref{26.01.11.b} does not hold, then $k=2$. In $M_1/e$, by the Claim, $e_i$ is in parallel with $f_j$, say $e_i$ is in parallel with $f_i$, for both $i$. Therefore $T_i=\{e,e_i,f_i\}$ is a triangle of $M$, for both $i$, and so $T_1\bigtriangleup T_2\bigtriangleup\{f_1,f_2,t_3\}=\{e_1,e_2,t_3\}$, where $T=\{t_1,t_2,t_3\}$ is a triangle of $M_1$. Thus $N=M_1\backslash e/e_1$ is a minor of $M_1$ contrary to the Claim. Thus~\eqref{26.01.11.b} holds. By the claim $e_1$ or $e_2$ is in parallel with $f_1$ in $M_1/e$, say $e_1$. That is, $T'=\{e,e_1,f_1\}$ is a triangle of $M_1$ and so of $M$. We have (ii).
Assume that (i) happens, that is, $M_2$ is regular. Thus $M_1$ is non-regular because $M$ is non-regular. To conclude (iii) we need to prove only that $e$ is a regular element of $M_1$. By the proof of Theorem 1.5, there are disjoint subsets $Y$ and $Z$ of $E(M_2)-E(M_1)$ such that $N=M_2\backslash Y/Z$ is a 6-element matroid such that $T''=E(N)-T$ is a triangle of $N$ and, for each $f\in T$, there is an $f''\in T''$ such that $\{f,f''\}$ is a circuit of $N$. So $M\backslash Y/Z$ is isomorphic to $M_1$ --- this isomorphism fix each element of $E(M_1)-E(M_2)$ and sends $f''$ into $f$, for each $f''\in T''$. As both $M\backslash e$ and $M/e$ are regular, it follows that $(M\backslash e)\backslash Y/Z\simeq M_1\backslash e$ and $(M/e)\backslash Y/Z\simeq M_1/e$ are regular. That is, $e$ is a regular element of $M_1$. We have (iii).
Assume that (ii) happens. If $E(M_1)-E(M_2)$ spans an element $g$ of $E(M_2)-E(M_1)$ in $M$, then $[E(M_1)-E(M_2)]\cup g$ is a 3-separating set for $M$. Using the 3-separation induced by this set, we can decompose $M$ as the 3-sum of matroids $M_1'$ and $M_2'$ such that $E(M_1')=[E(M_1)-E(M_2)]\cup g\cup T''$ and $T''=E(M_1')\cap E(M_2')$. Note that, in $M_1'$, the element $g$ is in parallel with some element of $T''$. In particular, $M_1'\backslash g\simeq M_1$ is regular. So $M_1'$ is regular; a contradiction to this lemma. Thus $E(M_1)-E(M_2)$ is closed in $M$. $\qed$
\bigskip
Now that we have shown $M$ has a clearly defined structure, we want to say more about the second situation. Recall that $R(M)$ is the set of regular elements.
For a triangle $T'$ and triad $T^*$ of $M$, we say that $T',T^*$ is an {\it undesired fan} if $T'\cap T^*\cap R(M)\not=\emptyset$. Note that $\{T'\cup T^*, E(M)-(T'\cup T^*)\}$ is an exact 3-separation for $M$ and by Theorem 1.4, it is possible to decompose $M$ as a 3-sum using it. In the next lemma we show that the presence of an undesired fan implies the existence of two regular elements.
\bigskip
\noindent {\bf Lemma 2.4.} {\it If $T',T^*$ is an undesired fan in $M$ such that $E(M_1)-E(M_2)=T'\cup T^*$, then $T'\cap T^*\subseteq R(M)$. Moreover, if $T^*-T'=\{f\}$, then $M/f$ is a $3$-connected non-regular matroid such that $T'\cap T^*\subseteq R(M/f)$.}
\bigskip
\noindent {\bf Proof.} Suppose that $T'=\{e,e',t\}$, $T^*=\{e,e',f\}$ and $e\in R(M)$. In $M/e'$, $t$ and $e$ are in parallel. As $M\backslash e$ and so $M/e'\backslash e$ is regular, it follows that $M/e'$ is regular because $M/e'$ is obtained from $M/e'\backslash e$ by adding $e$ in parallel with $t$. Using duality, we conclude that $M\backslash e'$ is regular. Hence $e'$ is a regular element of $M$ and so $T'\cap T^*\subseteq R(M)$.
Next, observe that $E(M_1)=T'\cup T^*\cup T$ and $E(M_2)=[E(M)-(T'\cup T^*)]\cup T$. As $M_1$ is regular, it follows that $M_2$ is non-regular. By Lemma 2.3, $f$ does not belong to a triangle of $M$. So $M/f$ is 3-connected because ${\rm si}(M/f)$ is 3-connected. But $M/f\simeq M_2$ because $M_1/f$ has three non-trivial parallel classes each containing one element of $T'$ and another of $T$. The result follows because $R(M)\subseteq R(M/f)$. $\qed$
\bigskip
In the next lemma, we prove that, when this happens, it is possible to uncontract $f$ keeping the property of these two regular elements.
\bigskip
\noindent {\bf Lemma 2.5.} {\it Let $N$ be a $3$-connected non-regular binary matroid having different regular elements $e$ and $e'$. Suppose that $T'$ is a triangle of $N$ such that $e,e'\in T'$ and $\{e,e'\}$ is not contained in a triad of $N$. If $M$ is a one-element binary lift of $N$, say $M/f=N$, such that $\{e,e',f\}$ is a triad of $M$, then $e$ and $e'$ are regular elements of $M$ (and $M$ is $3$-connected).}
\bigskip
\noindent {\bf Proof.} Observe that ${\rm si}(M/e)=M/e\backslash e'$. But, in $M\backslash e'$, $e$ and $f$ are in series. So $M/e\backslash e'\simeq M/f\backslash e'=N\backslash e'$ and ${\rm si}(M/e)$ is regular. Thus $M/e$ is regular. As $M\backslash e/f=N\backslash e$, it follows that $M\backslash e/f$ is regular and so $M\backslash e$ is regular. That is, $e$ is a regular element of $M$. A similar argument holds with $e'$. $\qed$
\bigskip
\section {The number of regular elements in a matroid}
Next we prove a result on the number of regular elements in a binary non-regular matroid. Observe that, $F_7^*$ has two single-element extensions $S_8$ and $AG(3, 2)$. The matroid $AG(3, 2)$ has one single-element extension $Z_4$. The matroid $S_8$ has two single-element extensions, $Z_4$ and $P_9$. Observe further that $F_7$ and $F_7^*$ have seven regular elements and $P_9$ has four regular elements. $AG(3, 2)$ has zero regular elements and consequently so do all its 3-connected extensions and coextensions.
\bigskip
\noindent {\bf Lemma 3.1.} {\it Let $M$ be a $3$-connected non-regular binary matroid. If $|E(M)|\ge 9$, then $|R(M)|=0, 1, 2$ or $4$. Moreover, if $|R(M)|=4$, then $R(M)$ is both a circuit and a cocircuit of $M$.}
\bigskip
\noindent {\bf Proof.} Since $|E(M)|\ge 9$, $M$ must have $P_9$ or $P_9^*$ as a minor and since $|R(P_9)|=4$, it follows that $|R(M)|\le 4$.
Suppose $|R(M)|=3$. Choose a minimal counter-example $M$. First observe that $|E(M)|\ge 10$ and by Theorem 1.3 $M$ is not internally 4-connected. By Theorem 1.4, we can decompose $M$ as the 3-sum of matroids $M_1$ and $M_2$ such that $E(M_1)\cap E(M_2)=T$ and $E(M_1)\cap R(M)\neq \phi$. If Lemma 2.3 (ii) occurs and $f\in T^*-T'$, then by Lemma 2.4 and the choice of $M$, $M/f$ has $R(M)\cup g$ as a circuit-cocircuit, for some element $g$. Note that $[R(M)\cup g]\bigtriangleup T^*$ is a triad of $M$ and $[R(M)\cup g]\bigtriangleup T'$ is a triangle of $M$ whose intersection contains a regular element. Therefore, by Lemma 2.4 the intersection has two regular elements ($g$ is the other regular element); a contradiction. Thus Lemma 2.3(i) occurs. Observe that $R(M)$ is contained in a circuit-cocircuit of $M_1$ consisting of regular elements avoiding $T$. Thus every element in this circuit-cocircuit is also a regular element of $M$; a contradiction. Thus we proved that $M$ cannot have exactly three regular elements.
Next, suppose $|R(M)|=4$, but $R(M)$ is not a circuit cocircuit. Choose a minimum counterexample. As before, $|E(M)|\ge 10$ and by Theorem 1.3 $M$ is not internally 4-connected. By Theorem 1.4, we can decompose $M$ as the 3-sum of matroids $M_1$ and $M_2$ such that $E(M_1)\cap E(M_2)=T$ and $E(M_1)\cap R(M)\neq \phi$. If Lemma 2.3(ii) occurs, $f\in T^*-T'$, then, by Lemma 2.4, $M/f$ has the same regular elements as $M$. By the choice of $M$, $R(M)$ is a circuit-cocircuit of $M/f$. As $R(M)\cup f$ contains a triad of $M$, it follows that $R(M)\cup f$ is not a circuit of $M$. Thus $R(M)$ is a circuit-cocircuit of $M$.
We may assume by Lemma 2.3 (i) that $M_2$ is regular, $M_1$ is non-regular, and $|R(M)|\subseteq E(M_1)$. By the choice of $M$ if $|E(M_1)|\ge 9$, $R(M)$ is a circuit-cocircuit of ${\rm si}(M_1)$ and therefore of $M$; a contradiction. Thus $M_1$ has at most 8 elements. Since ${\rm si}(M_1)$ is non-regular, ${\rm si}(M_1)$ is isomorphic to $F_7$ or $S_8$. In both cases, $R(M)$ is a circuit-cocircuit of this matroid. $\qed$
\bigskip
Using the previous lemma, we can refine the second part of Lemma 2.4.
\bigskip
\noindent {\bf Lemma 3.2.} {\it Let $M$ be a $3$-connected non-regular binary matroid with $|E(M)|\ge 10$ and suppose $T,T^*$ is an undesired fan of $M$ such that $T^*-T=\{f\}$. Then $M/f$ is a non-regular $3$-connected matroid such that $R(M/f)=R(M)$.}
\bigskip
\noindent {\bf Proof.} We argue by contradiction. Since $T\cap T^*\subseteq R(M)$, it follows from Lemma 2.4 that $|R(M)|\ge 2$. Lemma 3.1 implies that $|R(M/f)|$ is 2 or 4. Moreover, $R(M/f)$ is a circuit-cocircuit of $M/f$. So $|R(M)|=2$ and $R(M/f)=4$.
Since $T^*\subseteq R(M/f)\cup f$, it follows that $R(M/f)$ is also a circuit-cocircuit of $M$. Therefore $T'=T\bigtriangleup R(M/f)$ is a triangle of $M$ and $T'^*=T^*\bigtriangleup R(M/f)$ is a triad of $M$. But $T'$ is a triangle of $M/f$ containing two regular elements of $M/f$ such that no triad of $M/f$ contains these two elements. By Lemma 2.5 these two elements are also regular in $M$. Hence $R(M/f)=R(M)$; a contradiction. $\qed$
\bigskip
A 3-separation $\{X,Y\}$ for a 3-connected matroid is said to be trivial provided $|X|=3$ or $|Y|=3$.
\bigskip
\noindent {\bf Lemma 3.3.} {\it Let $M$ be a $3$-connected non-regular binary matroid such that $|R(M)|\ge 1$. If any non-trivial $3$-separation for $M$ has the union of a triangle and a triad of a undesired fan as one of its sets, then $M$ is isomorphic to $S_8$, $F_7$ or $F_7^*$.}
\bigskip
\noindent {\bf Proof.} If $|E(M)|\le 8$, then the result holds. Therefore, suppose that $|E(M)|\ge 9$. First assume that $M$ has just one non-trivial 3-separation. By Theorem 1.4, $M$ is the 3-sum of matroids $M_1$ and $M_2$ such that $E(M_1)-E(M_2)$ is the union of the triangle and the triad of the undesired fan. Thus $E(M_1)\cap R(M)\neq \phi$. Observe that Lemma 2.3 (ii) holds in this case. By the uniqueness of the 3-separation for $M$, $M_2$ is internally 4-connected. By Theorem 1.3, $M_2$ is isomorphic to $F_7$. Thus $|E(M)|=8$; a contradiction. Hence $M$ has at least two non-trivial 3-separations.
Let $T_1,T_1^*$ and $T_2,T_2^*$ be different undesired fans of $M$. For $i\in\{1,2\}$, set $Z_i=T_i\cap T_i^*$. By Lemmas 2.4, 3.1, and orthogonality, $R(M)=Z_1\cup Z_2$ is a circuit-cocircuit of $M$. In particular, $Z_1$ and $Z_2$ are unique and these are the unique undesired fans of $M$. If $T_1-T_1^*=\{t\}$ and $T_1^*-T_1=\{f\}$, then $T_2=Z_2\cup t$ and $T_2^*= Z_2\cup f$ because $T_1\bigtriangleup T_2=T_1^*\bigtriangleup T_2^*=R(M)$. Observe that $Z_1\cup Z_2\cup\{f,t\}$ is a 3-separating set for $M$. Thus $|E(M)|=9$ because $M$ has only two non-trivial 3-separations. Hence $M$ is isomorphic to $P_9$ or $P_9^*$; a contradiction. $\qed$
\section{The main result}
In this section we give the proof of Theorem 1.1.
\bigskip
\noindent {\bf Proof of Theorem 1.1.} First, we prove the ``only if'' part. If $M$ is non-binary, then by Theorem 1.2 we may conclude that $M\cong U_{2,4}$. Therefore suppose $M$ is binary and non-regular. Assume that $|R(M)|\ge 2$. If $M$ is an internally 4-connected matroid other than $F_7$ and $F_7^*$, then by Theorem 1.3 $M$ has a minor isomorphic to $M(E_5)$, $S_{10}$, $S_{10}^*$, $T_{12}\backslash e$, or $T_{12}/e$. Observe that $M(E_5)$ and $T_{12}\backslash e$, and $T_{12}/e$ have zero regular elements and $S_{10}$ and $S_{10}^*$ have one regular element; a contradiction because $R(M)\subseteq R(N)$. Thus $M\cong F_7$ and $F_7^*$.
We may now assume that $M$ is not internally 4-connected. By Lemma 3.3, $S_8$ is the unique matroid having all non-trivial 3-separations induced by the union of a triangle and a triad of some undesired fan. The result follows in this case. Therefore, we can assume that $M$ has a 3-separation such that none of its sets is the union of a triangle and a triad in a undesired fan, say $\{X_1,X_2\}$. By Theorem 1.4 there are 3-connected matroids (up to parallel elements with the common triangle) $M_1$ and $M_2$ such that $M$ is the 3-sum of $M_1$ and $M_2$ and, for $i\in\{1,2\}$, $E(M_i)=X_i\cup T$. By definition, $T$ is the common triangle between $M_1$ and $M_2$. By Lemma 2.3 we may assume that $M_1$ is non-regular and $M_2$ is regular. Moreover, $R(M)\subseteq X_1$. We may assume that $M_1$ is also 3-connected (the elements in parallel with elements of $T$, if them exist, are in $M_2$) By Lemmas 2.1 and 2.2, $T$ does not span any element of $R(M)$ in $M_1$ or $M_1^*$. Thus by induction we have three possibilities:
First, suppose $M_1$ is isomorphic to $F_7$ or $S_8$. The result follows because $M$ is the 3-sum of a matroid isomorphic to $F_7$ or $S_8$ (that is $M_1$) with a regular matroid (that is $M_2$).
Second, suppose $M_1$ is the 3-sum of matroids $N_1$ and $N_2$ along a triangle $T'$ such that $R(M)\subseteq E(N_1)$; $T'$ does not span any element of $R(M)$ in $N_1$; and $N_1$ is isomorphic to $F_7$ or $S_8$ and $N_2$ is regular (We may assume that $T'\cap E(M_2)=\emptyset$.) If $|E(N_2)\cap T|\ge 2$, then $T\subseteq E(N_2)$ because an element of $E(N_1)-E(N_2)$ spanned by $E(N_2)-E(N_1)$ in $M_1$ must be in parallel with some element of $T'$ in $N_1$. In this subcase, $M$ is the 3-sum of $N_1$ and the regular matroid obtained by doing the 3-sum of $N_2$ and $M_2$ along the triangle $T$. The result follows in this case. Thus we may assume that $|E(N_2)\cap T|\le 1$. As any two triangle of $N_1$ meet (recall that $N_1$ is isomorphic to $F_7$ or $S_8$), it follows that $E(N_2)\cap T=\{t\}$. Thus $t$ is in parallel with ane element $t'$ of $T'$ in $N_2$. Let $N_1'$ be the matroid obtained from $N_1$ by adding $t$ in parallel with $t'$. Note that $T$ is a triangle of $N_1'$. Thus $N_1'$ is isomorphic to $F_7^p$ or $S_8^p$. Moreover, $M$ is the 3-sum of $N_1'$ with $N_2\backslash t$ and $M_2$. The result also follows in this case.
Third, suppose there are matroids $N$, $N_1$, and $N_2$ such that:
\begin {enumerate}
\item $N$ has elements $t_1$ and $t_2$ in parallel;
\item $N\backslash t_1$ is isomorphic to $F_7$ or $S_8$;
\item $E(N_1)$ and $E(N_2)$ are disjoint;
\item $T_i=E(N)\cap E(N_i)$ is a triangle in both $N$ and $N_i$, for both $i\in\{1,2\}$;
\item $t_i\in T_i$, for both $i\in\{1,2\}$;
\item $N_1$ and $N_2$ are regular and 3-connected (up to some parallel elements with elements of $T_1$ and $T_2$ respectively);
\item $(T_1\cup T_2)\cap E(M_2)=\emptyset$; and
\item $M_1$ is the 3-sum of $N,N_1$ and $N_2$.
\end {enumerate}
\noindent We begin by showing that $|E(N_i)\cap T|\le 1$, for both $i\in\{1,2\}$. If $|E(N_i)\cap T|\ge 2$, say $i=2$, then $E(N_2)-T_2$ spans $T$ in $M_1$. As $t_1$ and $t_2$ are the only elements of $N$ in parallel, it follows that $T\subseteq E(N_2)-T_2$, otherwise the unique element belonging to $E(N_2)-T_2$ would be in parallel in $N$ with some element of $T_2$ and this element is not $t_1$. Hence $M$ is the 3-sum of $N,N_1$ and $N_2'$, where $N_2'$ is the 3-sum of $N_2$ and $M_2$ along $T$. The result follows, by induction. Thus we may assume that $|E(N_i)\cap T|\le 1$, for both $i\in\{1,2\}$. Moreover, when $|E(N_i)\cap T|=1$, say $E(N_i)\cap T=\{a_i\}$, $a_i$ is in parallel with some element $a_i'\in T_i$ in $N_i$. If $A_i=\{a_i\}$, when this happens, and $A_i=\emptyset$ otherwise, then $M_1$ is the 3-sum or $N'\backslash[\{a_1,a_2\}-(A_1\cup A_2)]$ with $N_1\backslash A_1$ and $N_2\backslash A_2$, where $N'$ is obtained from $N$ by adding, for both $i\in\{1,2\}$, $a_i$ in parallel with $a_i'$. As $T$ does not span any element of $R(M)$ in $N'$, by Lemma 2.1, and $|R(M)|\ge 2$, it follows that $T$ spans $T_1$ or $T_2$, say $T_2$. That is, each element of $T$ is in parallel with some element of $T_2$ in $N'$. We can transfer these elements for $N_2$ and we arrive at the previous case.
Finally, to see the ``if'' part, we use Lemmas 3.2 and 2.5 to reduce the $S_8$ case to the $F_7$ case in the 3-sums. The $F_7$ case is easy to verify. $\qed$
\bigskip
\noindent {\bf References}
\begin{enumerate}
\item S. R. Kingan, Binary matroids without prisms, prism duals, and cubes, {\it Discrete Mathematics}, {\bf 152} (1996), 211-224.
\item S. R. Kingan, A generalization of a graph result by D. W. Hall, {\it Discrete Mathematics} {\bf 173}, (1997) 129-135.
\item S. R. Kingan and M. Lemos, Almost-graphic matroids, {\it Advances in Applied Mathematics}, {\bf 28} (2002), 438 - 477.
\item J.G. Oxley, On nonbinary 3-connected matroids, {\it Trans. Amer. Math. Soc.} {\bf 300} (1987) 663?-679.
\item J. G. Oxley, {\it Matroid Theory}, (1992), Oxford University Press, New York.
\item P. D. Seymour, Decomposition of regular matroids, {\it J. Combin. Theory Ser. B } {\bf 28 } (1980), 305-359.
\item X. Zhou, On internally 4-connected non-regular binary matroids, {\it J. Combin. Theory Ser. B} {\bf 91 } (204), 327-343.
\end{enumerate}
\end {document} |
\section{Introduction}
\label{Intro}
There has been a great deal of progress in understanding the world-sheet theory of the string moving on some particular curved spacetimes like those involved in the basic AdS/CFT correspondence (see for example the series of review articles \cite{Beisert:2010jr}). The reason is that, under special circumstances, when the spacetime is of the form ${\mathbb R}_t\times F/G$, with $F/G$ a symmetric space like $S^n$ or ${\mathbb C}P^n$, the world-sheet theory is a non-relativistic integrable system. The excitations of the string are known as giant magnons~\cite{Hofman:2006xt,Dorey:2006dq,Chen:2006gea,Gaiotto:2008cg,Grignani:2008is,Abbott:2009um,Hollowood:2009sc}, which are soliton-like solutions on the string world-sheet, and in many cases the exact factorizable S-matrix is already known~\cite{Staudacher:2004tk,Beisert:2005tm,Beisert:2006qh,Arutyunov:2006yd,Ahn:2008aa}. More precisely the giant magnons are kink solutions that correspond to open strings and have to be put together to make closed string configurations.
It was noted a long time ago that the gauge-fixed theory of the string on such spacetimes is classically equivalent to a relativistic $1+1$-dimensional integrable QFT~\cite{Tseytlin:2003ii,Mikhailov:2005qv,Mikhailov:2005sy} known as a symmetric space sine-Gordon (SSSG) theory. These theories arise as the result of imposing the Pohlmeyer reduction on a sigma model with target space a symmetric space $F/G$~\cite{Pohlmeyer:1975nb}, and their Lagrangian formulation was originally proposed in~\cite{Bakas:1995bm} (for a recent review see~\cite{Miramontes:2008wt} and references therein). Integrable systems typically have more than one compatible symplectic structures, and it is known that at the classical level
the string sigma model and the SSSG theories have different symplectic structures \cite{Mikhailov:2006uc}. However, it has been suggested that the SSSG theory which is classically equivalent to superstrings on $AdS_5\times S^5$ could also be equivalent at the quantum level~\cite{Grigoriev:2007bu,Mikhailov:2007xr} (see also~\cite{Grigoriev:2008jq,Roiban:2009vh,Hoare:2009rq,Hoare:2009fs,Iwashita:2010tg,Hoare:2010fb}).
In this note we summarize some recent progress in understanding the SSSG theories and their relation to the string sigma models~\cite{Hollowood:2010dt}. We shall focus on the particular example of the SSSG theory associated to the symmetric space
\EQ{
S^5=F/G=\frac{SO(6)}{SO(5)}\ ,
}
which is relevant to the bosonic sector of the string moving on $AdS_5\times S^5$. We shall ignore the fermionic sector of these theories in the present letter, however, on the string side it is only the full theory with all the fermionic fields present which is expected to be a finite theory. For present purposes, where we only work at the level of semi-classical effects, we can ignore the issue of fully quantum effects and consequently ignore the fermions.\footnote{The symmetric space sine-Gordon theories with fermions has been considered in \cite{Schmidtt:2010bi}, and the solitons are constructed and investigated in \cite{Hollowood:2011fq}.}
The main paper \cite{Hollowood:2010dt} describes the generalization to any Symmetric Space of Type~I in Cartan's classification. The approach adopted in this work is to use the algebraic formalism of the symmetric space, and to this end we work with group- or algebra-valued fields
in $F=SO(6)$, or $\mathfrak{so}(6)$. The subgroup $G=SO(5)$, or algebra $\mathfrak{so}(5)$, is defined as the subgroup/algebra fixed by the involution
\EQ{
\sigma_-(f)=\theta f\theta\ ,\qquad\theta=\text{diag}(-1,1,1,1,1,1)\ ,
}
and in the following we will need the triplet of groups $SO(4)\subset SO(5)\subset SO(6)$ embedded as follows
\EQ{
SO(4)=\left(\begin{array}{cccccc} 1 & 0 & 0 &0&0&0\\ 0& 1 & 0&0&0&0\\ 0 & 0 & * &*&*&*\\ 0 & 0 & * &*&*&*\\ 0 & 0 & * &*&*&*\\ 0 & 0 & * &*&*&*\end{array}\right)\ ,\qquad
SO(5)=\left(\begin{array}{cccccc} 1 & 0 & 0 &0&0&0\\ 0& * & *&*&*&*\\ 0 & * & * &*&*&*\\ 0 & * & * &*&*&*\\ 0 & * & * &*&*&*\\ 0 & * & * &*&*&*\end{array}\right)\ .
\label{hgg}
}
\noindent
{\bf The Symmetric Space Sine-Gordon Theories}
For the symmetric space $S^5$, the SSSG theory is a gauged WZW theory for $SO(5)/SO(4)$ (where the anomaly free vector subgroup $H=SO(4)$ with $\gamma\to U\gamma U^{-1}$ is gauged) deformed with a kind of mass (or potential) term. The action takes the form
\EQ{
S=S_\text{gWZW}[\gamma,A_\mu] +S_\text{bt}[\phi,A_\mu] -\frac k{2\pi}\int d^2x\,{\rm Tr}\left(\Lambda
\gamma^{-1}\Lambda\gamma-\Lambda^2\right)\ ,
\label{ala}
}
and is invariant under the gauge transformations
\SP{
\gamma\to U\gamma U^{-1}\,,\qquad A_\mu \to U\big(A_\mu +\partial_\mu\big)U^{-1}\,, \qquad U\in H\,.
\label{gaugeT}
}
The mass term involves the element $\Lambda$ of the Lie algebra $\mathfrak{so}(6)$, which up to conjugation takes the form
\EQ{
\Lambda=m\left(\begin{array}{cccccc} 0 & -1 & 0 &0&0&0\\ 1& 0 & 0&0&0&0\\ 0 & 0 & 0&0&0&0\\ 0 & 0 & 0 &0&0&0\\ 0 & 0 & 0 &0&0&0\\ 0 & 0 & 0 &0&0&0\end{array}\right)\ .
\label{lam}
}
Note that $H=SO(4)$, defined as in \eqref{hgg}, is the stability group of $\Lambda$.
Here, $S_\text{gWZW}[\gamma,A_\mu]$ is the usual gauged WZW action for $G/H$ with level $k$,
\EQ{
S_\text{gWZW}[\gamma,A_\mu]&=-\frac k{4\pi}\int d^2x\,{\rm Tr}\,\Big[
\gamma^{-1}\partial_+\gamma\,\gamma^{-1}\partial_-\gamma+2A_+\partial_-\gamma\gamma^{-1}\\ &~~~~~~~~~
-2A_-\gamma^{-1}\partial_+\gamma-2\gamma^{-1}A_+\gamma A_-+2A_+A_-\Big]
\\ &~~~~~~~~~+\frac{k}{24\pi}\int d^3x\,\epsilon^{abc}{\rm Tr}\,\Big[\gamma^{-1}\partial_a\gamma\,
\gamma^{-1}\partial_b\gamma\,\gamma^{-1}\partial_c\gamma\Big]\ ,
\label{gWZW}
}
and
\EQ{
S_\text{bt}[\phi,A_\mu]=-\frac{k}{4\pi} \int d^2x\> \epsilon^{\mu\nu}\partial_\mu {\rm Tr}\big(A_\nu \phi \big)\ ,\qquad \gamma = e^\phi\ .
\label{Topol}
}
is a (total derivative) boundary term which does not contribute to the equations of motion.
Below we note some features of this theory:
(i) It can be thought of as a deformation of the CFT given by the gauged WZW model by the particular relevant operator corresponding to the mass term~\cite{Bakas:1993xh,Bakas:1995bm,CastroAlvaredo:2000kq}.
(ii) Classically, the vacuum is degenerate and, for $A_\mu=0$,
the potential term in~\eqref{ala} has a space of minima given by constant
group elements $\gamma\in H=SO(4)$. Hence, classically at least, there could be a Higgs effect since $\gamma(x=\pm\infty)\in SO(4)$ and the $SO(4)$ global gauge symmetry is generically spontaneously broken. However, there are solitons in the theory in the form of kinks which have a gapless spectrum whose existence means that in the functional integral one should integrate over the boundary values $\gamma(x=\pm\infty)$. This has the effect of restoring the $SO(4)$ global symmetry associated to gauge transformations. In other words, the kinks carry $SO(4)$ global charge and so are dyonic objects; namely, non-abelian Q-balls \cite{Safian:1987pr}.
At the quantum level the continuous spectrum of kinks becomes quantized and the Q-ball states transform in non-trivial representations of $SO(4)$. In particular, in these theories the gauge symmetry is not confined and physical states carry ``colour".
(iii) The SSSG theories are integrable. This can be seen by
writing the equations-of-motion of the SSSG equations in Lax form, that is as a zero curvature condition for a connection that depends on an auxiliary
complex spectral parameter $z$:
\EQ{
{\cal L}_\mu=\partial_\mu+{\cal A}_\mu(x;z)\ ,\qquad[{\cal L}_\mu(z),{\cal L}_\nu(z)]=0\ ,
\label{zcc}
}
where
\SP{
{\cal L}_+(z)&= \partial_++\gamma^{-1}\partial_+\gamma+\gamma^{-1}A_+\gamma-z
\Lambda
\ ,\\[5pt]
{\cal L}_-(z)&= \partial_-+A_--z^{-1}\gamma^{-1}\Lambda\gamma\ .
}
The existence of the Lax connection implies integrability in a way which is completely standard.
In addition, the equation-of-motion of the gauge field yields the additional constraints
\SP{
\Big(\gamma^{\mp1}\partial_\pm\gamma^{\pm1}
+\gamma^{\mp1}A_\pm\gamma^{\pm1}\Big)^\perp=A_\pm\ ,
\label{gco2}
}
where $\perp$ is a projection onto the Lie algebra of H, which in this case is $\mathfrak{so}(4)\subset\mathfrak{so}(6)$.
Since the equations-of-motion imply that $A_\mu$ is a flat connection, one can fix the gauge on-shell by choosing $A_\mu=0$ (fixing the gauge off-shell is described in \cite{Hoare:2009fs,Hoare:2010fb}).
(iv) Since the fields do not fall-off at $x=\pm\infty$, the WZ term requires careful treatment. In particular, it cannot strictly speaking be defined as an integral over a three-dimensional space with the two-dimensional spacetime as a boundary, and its definition in~\eqref{gWZW} should be taken to be schematic. One way to unambiguously define the action is, as in~\cite{Hoare:2009fs}, to use the condition of gauge invariance to pin down the expansion of the WZ term in terms of $\phi$, with $\gamma=e^\phi$. This prescription requires to supplement the action with the boundary term~\eqref{Topol} in order to make the leading order contribution gauge invariant. To spell this out, notice that
\EQ{
&S_\text{gWZW}=-\frac k{2\pi}\int d^2x\,{\rm Tr}\,\Big(
A_+\partial_-\phi -A_-\partial_+\phi
+[A_+, A_-]\phi + \cdots\Big)
} while
\EQ{
S_\text{gWZW}+S_\text{bt}=-\frac{k}{2\pi} \int d^2x\,{\rm Tr}\,\Big(
[\partial_+ + A_+, \partial_- +A_-]\phi + \cdots\Big)\,,
}
which shows that $S_\text{gWZW}+S_\text{bt}$ is indeed invariant under~\eqref{gaugeT} at leading order. The explicit expression for the expansion of the full action~\eqref{ala} in powers of $\phi$ can be found in~\cite{Hoare:2009fs}.
(v) The constraints~\eqref{gco2} have the interpretation of the vanishing, on-shell, of what is naively the Noether current corresponding to global gauge transformations. In fact this is just an example of the theorem of Hilbert and Noether that the current associated to a local symmetry vanishes on-shell, but crucially up to a topological contribution which, in our case, is fixed by the boundary term~\eqref{Topol}.\footnote{For a discussion of these issues in a modern context see, for example, \cite{Julia:1998ys,Silva:1998ii,Julia:2000er} and references therein.} The full expression for the current is
\SP{
{\cal J}_\pm=\frac{k}{4\pi} \Big(-(\gamma^{\mp1}\partial_\pm\gamma^{\pm1}+\gamma^{\mp1}A_\pm\gamma^{\pm1} -A_\pm)^\perp\pm\partial_\pm\phi^\perp\Big)\ ,
\label{ncor}
}
which exhibits that the Noether charge emerges as a kink charge,
\EQ{
{\cal Q}=\int dx\, {\cal J}^0= \frac{k}{4\pi}\int dx\,\partial_1\phi^\perp=
\frac{k}{4\pi}\big(\phi^\perp(\infty)-\phi^\perp(-\infty)\big)\ .
\label{chgs1}
}
\noindent
{\bf The String Sigma Model}
Once the string world-sheet theory for strings on ${\mathbb R}_t\times S^5$ is suitably gauge fixed, what remains is a sigma model with a target space $S^5\simeq SO(6)/SO(5)$ that can be
formulated in terms of a group-valued field $f\in SO(6)$ and a gauge field $B_\mu\in\mathfrak{so}(5)$ with a gauge symmetry
\EQ{
f \rightarrow f U^{-1}\>, \qquad
B_\mu \rightarrow U(B_\mu +\partial_\mu)U^{-1}\>,\qquad U\in SO(5)\>.
\label{GaugeTrans}
}
The action takes the form
\EQ{
S[f,B_\mu]= -\frac{\sqrt\lambda}{4\pi}\int d^2x\, \mathop{\rm Tr} \bigl(J_\mu J^\mu\bigr)\>,
\label{LagSM}
}
where the current $J_\mu = f ^{-1}\partial_\mu f -B_\mu
\rightarrow U J_\mu U^{-1}$ is covariant under gauge transformations, and $\lambda$ is the 't~Hooft coupling. The string sigma model also involves imposing the Virasoro constraints which, up to conjugation, take the form
\EQ{
J_+=-\Lambda\ ,\qquad J_-=-\gamma^{-1}\Lambda\gamma\ ,
\label{virc}
}
where the field $\gamma$ takes values in $G=SO(5)$.
The Virasoro constraints can be written as the auxiliary linear system \cite{Miramontes:2008wt}
\EQ{
\Big(\partial_++B_+-\Lambda\Big)f^{-1}=\Big(\partial_-+B_--\gamma^{-1}\Lambda\gamma\Big)f^{-1}=0\ .
\label{sml}
}
This sigma model has the vacuum solution
\EQ{
f_0=\exp(-2t\Lambda)\,.
\label{vacu}
}
Physically, it corresponds to a point-like string orbiting around the great circle in $S^5$ picked out by the element $\Lambda$. With our choice of $\Lambda$ in \eqref{lam} the motion is in the $(1,2)$ plane.
The sigma model has soliton solutions which, in their most general form, are known as dyonic giant magnons~~\cite{Hofman:2006xt,Dorey:2006dq,Chen:2006gea}. These solutions are also kinks because they describe open strings whose endpoints at $x=\pm\infty$ are at distinct points on $S^5$.
Notice that contrary to the SSSG theory the gauge symmetry is not realized in the spectrum, rather it is confined and one can use an equivalent manifestly gauge invariant formalism as in \cite{Hollowood:2009tw} by considering the gauge invariant field ${\cal F}=\theta f\theta f^{-1}$. As mentioned above, we shall ignore quantum effects and the running of the coupling in the sigma model.
\noindent
{\bf Relation between the string sigma model and the SSSG theory}
The sigma model and the SSSG theory are related via their equations-of-motion. In order to find the relation, notice that the Lax formulation of the SSSG equations \eqref{zcc} are the consistency conditions for the linear system
\EQ{
{\cal L}_\mu(z) \Upsilon(z)=0\ .
\label{LinProb}
}
It is useful to express
\EQ{
\Upsilon(z)=\chi(z)\Upsilon_0(z)\ ,
}
where, in the on-shell gauge $A_\mu=0$,
\EQ{
\Upsilon_0(x;z)=\exp\big[(zx^++z^{-1}x^-)\Lambda\big]
\label{dvs}
}
is the vacuum solution of the linear problem corresponding to $\gamma_0=1$. Then,
\EQ{
\gamma=\chi(0)^{-1}\ .
}
Comparing this linear system with \eqref{sml}, it follows that
$\Lambda$ and $\gamma$ are identified with the same quantities in the SSSG equations, and that the gauge fields in the sigma model are identified via
\EQ{
B_+=\gamma^{-1}\partial_+\gamma+\gamma^{-1}A_+\gamma\ ,\qquad B_-=A_-\ .
\label{Match}
}
The sigma model equations-of-motion then imply that $\gamma$ satisfies the SSSG equations-of-motion.
In addition, the group field of the sigma model is simply
\EQ{
f=\Upsilon(1)^{-1}=\Upsilon_0(1)^{-1}\chi(1)^{-1}\ .
\label{xll}
}
The fact that the expression for $f$ involves fixing the spectral parameter manifests the fact that the sigma model is not relativistically invariant. The reason is that the spectral parameter $z$ transforms as $z\to e^\vartheta z$ under a Lorentz transformation, and so setting it to 1 breaks Lorentz symmetry explicitly. In \eqref{xll}, notice that $f_0=\Upsilon_0(1)^{-1}$ is the vacuum solution \eqref{vacu}.
The implication is that there is one underlying integrable system which is expressed in two different ways, one as a relativistic QFT and the other as the non-relativistic theory on the string world sheet.
\section{The Q-ball Kinks and Dyonic Giant Magnons}
\label{Dressing}
A soliton solution of the integrable system is a Q-ball kink of the SSSG equations or a dyonic giant magnon of the string sigma model.\footnote{We shall reserve the term ``soliton" for the generic solution of the integrable system, and use ``Q-ball kink" and ``dyonic giant magnon" for its expression in the SSSG theory, via $\gamma$, and the sigma model, via $f$, respectively.} The explicit solutions are characterized by the dressing factor $\chi(z)$ having a small number of poles on the complex $z$ plane, four in the present case~\cite{Hollowood:2009sc}. In general, the pattern of the poles and the relation between the residues is determined by the particular symmetric space:
\EQ{
\chi(z)=1+\frac{{\boldsymbol F}_j(\Gamma^{-1})_{ji}{\boldsymbol F}_i^\dagger}{z-\xi_i}\ .
}
For $S^5$, ${\boldsymbol F}_i$ are four complex 6-vectors given by
\EQ{
{\boldsymbol F}_i=\Upsilon_0(\xi_i^*){\boldsymbol\varpi}_i\ ,\qquad i=1,\ldots,4\ ,
}
where ${\boldsymbol\varpi}_i$ are four constant 6-vectors given by
\EQ{
{\boldsymbol\varpi}_i=\{{\boldsymbol\varpi},{\boldsymbol\varpi}^*,\theta{\boldsymbol\varpi},\theta{\boldsymbol\varpi}^*\},
\label{vwr}
}
with the constraint
\EQ{
{\boldsymbol\varpi}\cdot{\boldsymbol\varpi}=0\ .
}
Fixing various redundancies, for a single soliton the constant 6-vector ${\boldsymbol\varpi}$ can be written as
\EQ{
{\boldsymbol\varpi}={\boldsymbol v}+{\boldsymbol\Omega}\ .
}
Here, ${\boldsymbol v}$ is one of the non-null eigenvectors of $\Lambda$. Since it does not matter which one is chosen, we take
\EQ{
{\boldsymbol v}=\frac1{\sqrt{2}}(1,-i,0,0,0,0)\ .
}
The other vector ${\boldsymbol\Omega}$ is a unit complex vector in the 4-dimensional subspace picked out by the subalgebra $\mathfrak{so}(4)\subset\mathfrak{so}(6)$, so that ${\boldsymbol\Omega}$ is a null eigenvector of $\Lambda$ with the additional constraint ${\boldsymbol\Omega}\cdot{\boldsymbol\Omega}=0$. The positions of the four poles are
\EQ{
\xi_i=\{\xi,\xi^*,-\xi,-\xi^*\}
\label{Poles}
}
and, finally, the matrix $\Gamma$ reads
\EQ{
\Gamma_{ij}=\frac{{\boldsymbol F}_i^*\cdot {\boldsymbol F}_j}{\xi_i-\xi^*_j}\ .
}
The data $({\boldsymbol\Omega},\xi)$ are the parameters associated to
the solution. The complex variable $\xi$ determines the energy and momentum of the soliton whereas ${\boldsymbol\Omega}$ is a genuine collective coordinate which labels solutions of the same energy and momentum. Writing $\xi =e^{-\vartheta -iq}$, the $t$ and $x$ dependence of the the soliton follow from
\EQ{
{\boldsymbol F}_1= \Upsilon_0(\xi^\ast) {\boldsymbol\varpi}=\exp\left[2imt' \cos q - 2mx' \sin q\right]{\boldsymbol v}+{\boldsymbol\Omega}\ ,
\label{Fvector}
}
where we have introduced the boosted cordinates
\EQ{
t' = t\cosh\vartheta -x\sinh\vartheta\ ,\qquad
x' = x\cosh\vartheta -t\sinh\vartheta
}
which identify $\vartheta$ as the rapidity.
The soliton carries a non-trivial moduli space corresponding to
the ``polarization" vector ${\boldsymbol\Omega}$. We can write ${\boldsymbol\Omega}=\frac1{\sqrt2}({\boldsymbol\Omega}_1+i{\boldsymbol\Omega}_2)$, where ${\boldsymbol\Omega}_i$ are two real orthonormal vectors.
The moduli space is swept out by the action of the global $SO(4)$ part of the gauge symmetry group identified with a particular (co-)adjoint of the form
$Uh_{\boldsymbol\Omega} U^{-1}$, for
\EQ{
h_{\boldsymbol\Omega}=i({\boldsymbol\Omega}\BOmega^\dagger-{\boldsymbol\Omega}^*{\boldsymbol\Omega}^t)={\boldsymbol\Omega}_1{\boldsymbol\Omega}_2^t-{\boldsymbol\Omega}_2{\boldsymbol\Omega}_1^t \in \mathfrak{so}(4)\ .
\label{defh}
}
Up to conjugation, we may take ${\boldsymbol\Omega}=\frac1{\sqrt2}(0,0,1,-i,0,0)$, which gives an orbit of the form
\EQ{
U\MAT{0&0&0&0&0&0\\ 0&0&0&0&0&0\\ 0&0&0&-1&0&0\\ 0&0&1&0&0&0\\ 0&0&0&0&0&0\\ 0&0&0&0&0&0}U^{-1}\ ,
}
where $U\in SO(4)$ as in \eqref{hgg}.
This identifies the orbit as the real Grassmannian
\EQ{
\mathfrak M=\frac{SO(4)}{SO(2)\times SO(2)}\simeq S^2\times S^2\ .
\label{fgt}
}
On the SSSG side, $q$ determines the mass and charges of the Q-ball kink. Physically, distinct solutions are given by restricting $q\in(0,\frac\pi2)$ and they have mass
\EQ{
M=\frac{4km}\pi\sin q\ .
}
Their kink charge is
\EQ{
\gamma^{-1}(x=\infty)\gamma(x=-\infty)=\exp\left[-4qh_{\boldsymbol\Omega}\right]\ .
}
Assuming that $\gamma(-\infty)$ and $\gamma(+\infty)$ commute, which is true for these configurations, the kink charge corresponds to
\EQ{
{\cal Q}=\frac{k}{4\pi}\big(\phi^\perp(\infty)-\phi^\perp(-\infty)\big) =\frac{k q}{\pi}\,h_{\boldsymbol\Omega}
}
in~\eqref{chgs1}. Therefore, since this is the Noether charge under global gauge transformations, these solutions can actually be understood both as kinks and as non-abelian $Q$-balls~\cite{Safian:1987pr}.
On the sigma model side, the dyonic giant magnon is characterized by its charge under the global symmetry corresponding to left multiplication $f\to Uf$, $U\in F$, relative to the vacuum solution. The conserved current for left multiplication is $D_\mu f\,f^{-1}$, and so
\EQ{
\Delta{\cal Q}_L=\int_{-\infty}^\infty dx\,\big(D_0f\,f^{-1}-D_0f_0\,f_0^{-1}\big)\ ,
}
where the vacuum solution is
\EQ{
f_0=\exp(-2t\Lambda)\ .
}
The calculation of this charge is described in Appendix~\ref{AppA}. It consists of two distinct contributions:
\EQ{
\Delta{\cal Q}_L = J_1m^{-1}\Lambda -J_2h_{{\boldsymbol\Omega}}\ ,
\label{NewCharge2}
}
with
\EQ{
J_1=\frac{r^2+1}r\Big|\sin\frac p2\Big|\ ,\qquad J_2=\frac{r^2-1}r\Big|\sin\frac p2\Big|\ ,
}
where the parameter $re^{ip/2}$ is related to $\xi$ via
\EQ{
re^{ip/2}=\frac{1-\xi}{1+\xi}\ .
}
These charges satisfy the relation
\EQ{
J_1 = \sqrt{J_2^2 + 4\sin^2\frac{p}{2}}\>.
\label{Dispersion}
}
In the AdS/CFT context, the components $J_1$ and $J_2$ are identified, up to scaling, with $\Delta-J$ and $Q$, respectively, where $\Delta$ is the scaling dimension of the associated operator in the CFT, and $J$ and $Q$ are two conserved $U(1)$ $R$-charges:
\EQ{
\Delta -J = \frac{\sqrt\lambda}{2\pi}\> J_1\>, \qquad
Q = \frac{\sqrt\lambda}{2\pi}\> J_2\>,
}
where $\lambda$ is the 't~Hooft coupling. Then,~\eqref{Dispersion} becomes the celebrated dispersion relation
\EQ{
\Delta -J = \sqrt{Q^2 + \frac\lambda{\pi^2}\sin^2\frac{p}{2}}\>.
\label{Celebrated}
}
In the rest frame, $\vartheta=0$ or $p=\pi$,
\EQ{
\Delta{\cal Q}_L=2m^{-1}\Lambda\operatorname{cosec} q-2h_{\boldsymbol\Omega}\cot q
\label{GMcharge}
}
and so
\EQ{
Q=\frac{\sqrt\lambda}\pi\cot q\ .
\label{RCharge}
}
It is important to notice that the internal moduli space of the soliton takes the form of a (co-)adjoint orbit of $SO(4)$.\footnote{For compact Lie groups there is no distinction between adjoint and co-adjoint orbits.} For the Q-ball kink, this is interpreted as the group of global gauge transformations, while for the giant magnon this is interpreted as the $H=SO(4)$ subgroup of left multiplications of $f$ (modulo gauge transformations) which fix the vacuum $f_0$. In particular the action which fixes the vacuum corresponds to the adjoint action $f\to UfU^{-1}$ which includes a compensating global gauge transformation (right multiplication).\footnote{In the gauge invariant formulation in terms of ${\cal F}=\theta f\theta f^{-1}$, the symmetry action is always vector-like, ${\cal F}\to U{\cal F}U^{-1}$.} The physical interpretation of the moduli space $\mathfrak M$ is that the dyonic motion takes place in a plane perpendicular to the plane picked out by the orbital motion of the vacuum solution. The orientation of this plane is then described by the real Grassmannian $\mathfrak M$ in \eqref{fgt}.
\section{Semi-Classical Quantization of the Dyons}
\label{SCQD}
The soliton that we have constructed is a non-abelian dyon since it carries a charge under a nonabelian global symmetry group $H=SO(4)$. The charge is function of the continuous
parameter $q\in(0,\frac\pi2)$ and also of the polarization vector ${\boldsymbol\Omega}$ that determines the orientation of the motion inside $SO(4)$ and corresponds to the real Grassmannian in \eqref{fgt}.
Since the soliton is a periodic classical solution, one way to quantize it semi-classically is to use the Bohr-Sommerfeld rule. Remarkably, it leads to different results for the Q-ball kinks and the dyonic giant magnons.
If $\phi$ denotes the soliton solution in its rest frame, then $S[\phi]+MT=2\pi{\EuScript N}$ with ${\EuScript N}=1,2,\ldots$, where the time period is $T=\pi/(m\cos q)$.
For the dyonic giant magnon, this gives~\cite{Chen:2006gea}
\EQ{
\text{Dyonic Giant Magnon:}\qquad\qquad\cot q=\frac{\pi{\EuScript N}}{\sqrt\lambda}\ ,\qquad {\EuScript N}=1,2,\ldots,\infty
\label{chr}
}
and so the $U(1)_R$ charge~\eqref{RCharge} is quantized in integer units, $Q={\EuScript N}$,
which is the known spectrum of the quantized dyonic magnons~\cite{Chen:2006gea}. For the Q-ball kink, the Bohr-Sommerfeld quantization leads to a different quantization of $q$; namely\footnote{The $S^n$ Q-ball kink has an obvious embedding in the $SO(4)/SO(3)$ SSSG theory for which $H=SO(2)$ is abelian.
In Appendix~\ref{AppB} we show that this yields the complex sine-Gordon theory for which the kink becomes a conventional
$Q$-ball~\cite{Miramontes:2004dr}. The Bohr-Sommerfeld rule applied to the complex sine-Gordon theory dyon then gives rise to the quantization that follows in~\eqref{iuuCSG}~\cite{Dorey:1994mg}.
}
\EQ{
\text{Q-Ball Kink:}\qquad\qquad q=\frac{\pi{\EuScript N}}{2k}\ ,\qquad{\EuScript N}=1,2,\ldots,k\ .
\label{iuuCSG}
}
Notice that the dyonic giant magnon spectrum is infinite while the Q-ball kink spectrum is truncated because of the finite range of the parameter $q$. Of course the Bohr-Sommerfeld rule is strictly speaking only valid in the semi-classical regime, so ${\EuScript N}$ of order $\sqrt\lambda$ and $k$, respectively.
Although the Bohr-Sommerfeld rule gives the energy levels in both cases, it does not reveal the symmetry multiplets at each level. In order to uncover this structure we must proceed in a different way. The idea is to find an effective description of the polarization degree-of-freedom ${\boldsymbol\Omega}$ which takes values in the moduli space $\mathfrak M$.
In order to motivate the way that we quantize the dyons, it is worth a digression into the quantization of solitons more generally.
Let us consider a hypothetical theory with fields $\phi(x^\mu)$ that has a soliton solution $\phi(X^i;x^\mu)$ having an internal moduli space $\tilde{\mathfrak M}=S^n$ on which the group $SO(n+1)$ acts as a global symmetry of the theory. In the conventional way of proceeding, the existence of a moduli space of solutions, whose coordinates $X^i$ are the collective coordinates of the soliton, means that the equations-of-motion have a set of zero modes
\EQ{
\delta_i\phi=\frac{\partial\phi(X^i;x^\mu)}{\partial X^i}
}
one for each of the collective coordinates. In the conventional setting, for example, for monopoles in Yang-Mills-Higss theories in $3+1$-dimensions, following the philosophy of Manton \cite{Manton:1981mp}, one allows the collective coordinates to depend on time $X^i\to X^i(t)$, and then substitutes $\phi(X^i(t);x^\mu)$ into the action of the theory to extract an effective action for a 1-dimensional theory along the soliton's world line:
\EQ{
S_\text{eff}[X^i]=S[\phi(X^i(t))]=\int dt\,\int dx\,\mathscr{L}(\phi(X^i;x^\mu))\ .
}
Assuming that the theory has a quadratic kinetic term, the effective quantum mechanical action that results contains terms which can be linear and quadratic in time derivative:
\EQ{
S_\text{eff}[X^i]=\int dt\, \Big[{\cal Q}_i(X)\dot X^i+\tfrac12g_{ij}(X)\dot X^i\dot X^j + \cdots\Big]\ ,
\label{sole}
}
where
\EQ{
{\cal Q}_i(X)=\int dx\, \dot\phi\,\delta_i\phi
}
is the charge of the soliton under the symmetry variation $\delta X^i$ and
$g_{ij}(X)$ is a metric on $\tilde{\mathfrak M}$ given by the inner-product of the zero modes,
\EQ{
g_{ij}(X)=\int dx\,\delta_i\phi\,\delta_j\phi\ .
\label{frr}
}
In many situations the soliton is a static solution and carries no charge. In this case the term linear in $\dot X^i$ is absent and the effective theory is a quantum mechanical sigma model on the moduli space.
To be concrete, and to bring the discussion as close as possible to the present setting, let us suppose that $\tilde{\mathfrak M}=S^n$ with a natural action of $SO(n+1)$ corresponding to some global symmetry of the parent theory.\footnote{Following the analogy, in the present setting, $n=3$ and $S^3$ is the subspace of $S^5$ perpendicular to the plane picked out by the vacuum solution.} The Euler-Lagrange equations that follow from the quantum mechanical sigma model
\EQ{
S_\text{eff}[X^i]=\int dt\, \tfrac12g_{ij}(X)\dot X^i\dot X^j
\label{sole2}
}
are simply the geodesic equations for the Riemannian manifold $(\tilde{\mathfrak M},g_{ij})$. For the $S^n$ example, the solutions are motions around great circles which carry arbitrary angular momentum, and the motion gives the classical solution with $SO(n+1)$ charge, in other words the excitations are dyonic.
This system can easily be quantized: the quantum Hamiltonian is the Laplacian on $S^n$ and the states are spherical harmonics.
This approach is only valid when the correction to the mass is large compared with one-loop quantum corrections but small with respect to the mass of the soliton. This latter requirement is needed in order that the back reaction of the motion on the soliton is small.
In principle, however, we can include all the effects of the back reaction of the internal motion at the classical level by finding more general dyonic generalizations of the
original soliton in the parent theory. In the present context it is integrability that allows us to write down the exact dyonic solutions.
The dyonic solutions will have a continuous parameter $\ell$ which is the magnitude of the angular momentum of the motion and also other parameters that determine
the axis of the rotation, or ``polarization", that is 2 orthonormal vectors ${\boldsymbol\Omega}_i$. In other words, $\{{\boldsymbol\Omega}_i\}$ determine the orientation of a plane and therefore the polarization degree-of-freedom takes values in the real Grassmannian
\EQ{
\mathfrak M=\frac{SO(n+1)}{SO(2)\times SO(n-1)}\ .
\label{fgt2}
}
One can also think of this moduli space as a (co-)adjoint orbit of the symmetry group $SO(n+1)$. In fact we can imagine generating the dyon solution from the soliton by performing a time-dependent symmetry transformation, schematically $\phi\to U(t)\phi$ with $U(t)\in SO(n+1)$, and then computing the back-reaction to the motion exactly.
The mass of the dyon will be a function of $\ell$ and the $SO(n+1)$ charge will be of the form
\EQ{
{\cal Q}=\frac{\ell}2h_{\boldsymbol\Omega}\ ,\qquad{\boldsymbol\Omega}=\frac1{\sqrt2}\big({\boldsymbol\Omega}_1+i{\boldsymbol\Omega}_2\big)
\label{Charge}
}
where $h_{\boldsymbol\Omega}$ is defined as in \eqref{defh}.
The effective description of the dyon solution then takes the form \eqref{sole} but, now, the term linear in $\dot X^i$ is non-vanishing and becomes the dominant term in the semi-classical expansion. We can write the effective description in terms of a time-dependent symmetry transformation of the static solution $U(t)$, in the form~\cite{Hollowood:2010dt}\footnote{There is a slight subtlety here in that the moduli space of the dyon $\mathfrak M$ is not exactly the moduli space of the static solution $\tilde{\mathfrak M}$ because for the time-dependent solutions translations in $t$ act on $\tilde{\mathfrak M}$, and $\mathfrak M$ is the quotient of this action.}
\EQ{
S_\text{eff}[U]=\int dt\,{\rm Tr}\left(U^{-1}\frac{dU}{dt}\,{\cal Q}\right)\ .
\label{cvc2}
}
Below we shall show how to quantize this system. However, it is clear that the equations-of-motion require $U(t)$ to be a constant, and consequently, on shell, the polarization ${\boldsymbol\Omega}$ is fixed. This is why the quantized dyon does not carry more than one charge and also the collective coordinate dynamics does not contribute to the mass of the dyon since the Hamiltonian vanishes.
In addition, as we show later, the angular momentum must be quantized precisely as $\ell\equiv {\EuScript N}=1,2,\dots$. It turns out that this quantization rule is identical to the Bohr-Sommerfeld quantization of the dyon; however, the bonus of this method is that we can compute the multiplet structure of the levels.
We shall find that the states of a given ${\EuScript N}$ transform in the rank-$\EuScript N$ symmetric representation of $SO(n+1)$. The mass of the dyon is then just the classical mass but with the quantized values of $\ell$ inserted.
The semi-classical quantization of the dyonic giant magnon and Q-ball kink fits exactly into the story above apart from the fact that in the case of the SSSG theory there is no original uncharged soliton solution to start with. The parameter $q$ determines the charge of the soliton as in \eqref{Charge}.
For the Q-ball kink, it plays the r\^ole of $\ell$ as in \eqref{nxn} below. Then, as $q\to0$,
the mass of the kink
goes to 0 as $\ell\to0$ and so solitons with small charge are not bona-fide semi-classical objects; rather they are actually the perturbative excitations of the theory~\cite{Hollowood:2010dt}. For the dyonic giant magnon the situation is more complicated due to the fact that the system has no relativistic invariance. In this case, as $\ell$ goes to zero, with $\ell$ as in \eqref{nzn} below, the dyonic magnon becomes an ordinary giant magnon which is a semi-classical object when $\sin\frac p2$ is of order 1 (in~\eqref{Celebrated}, $Q$ is small but $\Delta -J$ is large), but becomes a perturbative excitation as $p\to0$ (both $Q$ and $\Delta -J$ are small). In both situations, however,
the direct quantization of the dyonic solution as described above is valid for large enough $\ell$.
The claim is that when we take the dyon solution and transform it with a time-dependent $SO(4)$ transformation, substitute it into the action and perform the spatial integral we obtain an effective action of the form \eqref{cvc2}. For the Q-ball kinks of the SSSG theory this was shown in \cite{Hollowood:2010dt} with the result
\EQ{
\text{Q-Ball Kink:}\qquad\qquad \ell(q)=\frac{2kq}\pi\ .
\label{nxn}
}
Here, $k^{-1}$ plays the r\^ole of the coupling.
Now we show that we get a similar action for the dyonic giant magnon. According to~\eqref{Match}, in the $A_\mu=0$ on-shell gauge we have $B_-=0$ and $B_+=\gamma^{-1}\partial_+\gamma$, and we take
\EQ{
f\longrightarrow U(t)fU(t)^{-1}\ ,\qquad U(t)\in SO(4)\ .
}
So for
\EQ{
J_+=f^{-1}\partial_+f-B_+=f^{-1}\partial_+f-\gamma^{-1}\partial_+\gamma\ ,
}
and given that $\gamma\to U(t)\gamma U(t)^{-1}$ also,
we obtain
\EQ{
J_+ &\longrightarrow U\Big(J_+ + f^{-1}[U^{-1}\dot U,f]-\gamma^{-1}[U^{-1}\dot U,\gamma]\Big)U^{-1}\ ,\\
J_- &\longrightarrow U\Big(J_- + f^{-1}[U^{-1}\dot U,f]\Big)U^{-1}\ .
}
Then, to leading order in the semi-classical approximation, we can work to linear order in $\dot U$,
\EQ{
\delta S&=-\frac{\sqrt\lambda}{4\pi}\int d^2x\,{\rm Tr}\big(\delta J_+J_-+J_+\delta J_-\big)\\
&=-\frac{\sqrt\lambda}{4\pi}\int d^2x\,{\rm Tr}\Big[(f^{-1}U^{-1}\dot Uf-\gamma^{-1}U^{-1}\dot U\gamma)J_-\\
&\qquad\qquad+(f^{-1}U^{-1}\dot Uf-U^{-1}\dot U)J_+\Big]\\
&=-\frac{\sqrt\lambda}{4\pi}\int_{-\infty}^\infty dt\,{\rm Tr}\Big(U^{-1}\dot U
\,\Delta{\cal Q}_L\Big)
}
where, using the Virasoro constraints \eqref{virc} and $fJ_\mu f^{-1}=D_\mu f\,f^{-1}$, we have identified
\EQ{
\int_{-\infty}^\infty dx\, \big(
D_0f\,f^{-1}-D_0f_0\,f_0^{-1}\big)=\Delta{\cal Q}_L\ .
}
Therefore, just like the SSSG case, we have an effective quantum mechanical action
\EQ{
S_\text{eff}[U]=\frac{\ell(q)}2\int_{-\infty}^\infty dt\,{\rm Tr}\Big(U^{-1}\frac{dU}{dt}
\,h_{\boldsymbol\Omega}\Big)\ .
}
where, using~\eqref{GMcharge},
\EQ{
\text{Dyonic Giant Magnon:}\qquad\qquad\ell(q)=\frac{\sqrt\lambda}{\pi}\cot q\ .
\label{nzn}
}
We now proceed to a quantization of an action of the form \eqref{cvc2}. Before proceeding in earnest, it is worthwhile making some comments about such theories. Unlike the soliton effective theory \eqref{sole}, the action in \eqref{cvc2} does not involve metric data of the moduli space $\mathfrak M$. In fact it lies in the class of ``topological" or Chern-Simons quantum mechanics defined and investigated in \cite{Dunne:1989hv,Howe:1989uk,Ivanov:2003qq}. Such a system is defined by a manifold $\mathfrak M$ with a symplectic 2-form $F$ which can be written locally as $F=dA$. The quantum mechanical system has the form
\EQ{
S=\frac \ell2 \int f^*A\ ,
}
where $f^*A$ is the pull-back of $A$ to a one cycle (the world-line in our case). The normalization of $F$ is determined by $\int _{\mathfrak M}F\wedge\cdots\wedge F=4\pi$ and $\ell$ is a coupling constant. For the SSSG theories all the spaces $\mathfrak M$ are actually homogeneous K\"ahler manifolds, in which case the sympletic form can be taken to be the K\"ahler form and the action can be written in terms of the K\"ahler potential $K$,
\EQ{
S=\frac \ell2\int dt\,\Big(-i\dot z^i\frac{\partial K}{\partial z^i}+\text{c.c.}\Big)\ .
}
where $(z^i,\bar z^i)$ are a set of complex coordinates for $\mathfrak M$. A rather beautiful way to quantize the theory is to use ``analytic quantization" as described in \cite{Ivanov:2003qq}. Wavefunctions are sections of holormorphic line bundles over $\mathfrak M$ with curvature $\ell F/(4\pi)$ which must be integral for consistency. Given the normalization of $F$ this requires that $\ell$ is quantized in integer units. Below we shall explain how to quantize the particular example \eqref{fgt}, or the more general \eqref{fgt2}, which are homogeneous K\"ahler manifolds, in a more pedestrian
way.
First of all, we can think of the time-dependence via the vector ${\boldsymbol\Omega}$, by identifying ${\boldsymbol\Omega}(t)=U(t){\boldsymbol\Omega}_0$, where ${\boldsymbol\Omega}_0$ is some fixed reference vector.
In this case, the effective action takes the form
\EQ{
S_\text{eff}=-\frac{i\ell}2\int dt\,\Big({\boldsymbol\Omega}^*\cdot\frac{d{\boldsymbol\Omega}}{dt}-{\boldsymbol\Omega}\cdot\frac{d{\boldsymbol\Omega}^*}{dt}\Big)=-i\ell\int dt\,{\boldsymbol\Omega}^*\cdot\frac{d{\boldsymbol\Omega}}{dt}
}
and implicitly we have the constraints ${\boldsymbol\Omega}^*\cdot{\boldsymbol\Omega}=1$ and ${\boldsymbol\Omega}\cdot{\boldsymbol\Omega}=0$ as well as the identification ${\boldsymbol\Omega}\sim e^{i\alpha}{\boldsymbol\Omega}$.
It is useful to relax the constraints and
enlarge the phase space to ${\mathbb C}^4$ since then the Poisson brackets are trivial:
\EQ{
\{{\boldsymbol\Omega}_i,{\boldsymbol\Omega}^*_j\}=\frac{i}{\ell}\delta_{ij}\ .
}
The way to reduce the larger phase space proceeds via a K\"ahler quotient. This starts by noticing that
the $U(1)$ symmetry ${\boldsymbol\Omega}\to e^{i\alpha}{\boldsymbol\Omega}$
is a Hamiltonian symmetry generated by
$\Phi={\boldsymbol\Omega}^*\cdot{\boldsymbol\Omega}$. The physical phase space corresponds to restricting ${\mathbb C}^4$ to the level set
\EQ{
\Phi={\boldsymbol\Omega}^*\cdot{\boldsymbol\Omega}=1
}
and performing a quotient by the $U(1)$ symmetry, as well as imposing the constraint ${\boldsymbol\Omega}\cdot{\boldsymbol\Omega}=0$.
In the quantum theory, we can replace the Poisson brackets by commutators
involving the operators $\hat{\boldsymbol\Omega}_i$ and $\hat{\boldsymbol\Omega}_i^\dagger$:
\EQ{
[\hat{\boldsymbol\Omega}_i,\hat{\boldsymbol\Omega}_j^\dagger]=\frac{1}{\ell}\delta_{ij}
\label{nco}
}
and build a Hilbert space by treating the former as annihilation operators and the latter as creation operators. The generator of the Hamiltonian symmetry
\EQ{
\hat\Phi=\hat{\boldsymbol\Omega}^\dagger\cdot\hat{\boldsymbol\Omega} =\frac{\hat{\EuScript N}}\ell
}
is proportional to the number operator $\hat{\EuScript N}$ and the constraint $\hat\Phi=1$, along with the quantization of the occupation number, implies the quantization of $\ell$:
\EQ{
\ell={\EuScript N}=1,2,\ldots\ .
\label{iuu}
}
The Hilbert space is spanned by the states\footnote{Notice that the quotient by $U(1)$ is trivial at the level of the Hilbert space.}
\EQ{
\hat{\boldsymbol\Omega}_{i_1}^\dagger\hat{\boldsymbol\Omega}_{i_2}^\dagger\cdots\hat{\boldsymbol\Omega}_{i_{\EuScript N}}^\dagger|0\rangle\ ,
}
However, there is the additional constraint ${\boldsymbol\Omega}\cdot{\boldsymbol\Omega}=0$ to impose. Since this is holomorphic
it can be implemented directly at the level of the Fock space by removing by hand states which are of the form $\sum_i\cdots\hat{\boldsymbol\Omega}_i^\dagger\cdots\hat{\boldsymbol\Omega}_i^\dagger\cdots |0\rangle$. One recognizes this as the process of ``removing traces" that is well-known in the Young Tableaux approach to the orthogonal groups. The remaining states form a representation space for the rank-${\EuScript N}$ symmetric representations of $SO(4)$.
The construction we have presented has an interesting interpretation as ``fuzzy geometry''~\cite{Balachandran:2005ew}. This follows from the fact that in the quantum theory the coordinates of $\mathfrak M$ do not commute as in \eqref{nco}.
However, as ${\EuScript N}$ increases the non-commutativity
gets less marked and the fuzzy geometry
becomes a closer approximation of the classical geometry in the limit ${\EuScript N}\to\infty$, which clearly requires $k\to\infty$, the semi-classical limit.
One can see now that the quantization of $q$ in both cases is equivalent to the Bohr-Sommerfeld rule in \eqref{chr} and \eqref{iuuCSG}.
In both cases, the states come in the ${\EuScript N}$-rank symmetric representations of $SO(4)$, but the giant magnon tower is unbounded whereas the Q-ball kink tower is bounded to have height $k$. In addition, for the giant magnons, the quantization of ${\EuScript N}$ is the expected quantization of the $R$ charge~\cite{Chen:2006gea}.
Correspondingly, as a result of the quantization of $q$, the continuous spectrum of classical Q-ball kinks becomes discrete,
\EQ{
M=\frac{4km}\pi\sin\Big(\frac{\pi{\EuScript N}}{2k}\Big)\ ,\qquad
{\EuScript N}=1,2,\ldots,k\ .
}
Strictly speaking the semi-classical analysis only applies when ${\EuScript N}$ is of order $k$ and $\sqrt\lambda$ for the two cases, respectively. However, the results appear to apply also for small $\EuScript N$. For the SSSG theory, the exact S-matrix constructed for the case $F/G={\mathbb C}P^n$ in~\cite{Hollowood:2010rv} suggests that in the full quantum theory there is simply a finite renormalization of $k$, which is a well-known feature of the WZW theory.
The states at the bottom of the tower ${\EuScript N}=1$, which transform in the vector representation of $SO(4)$, correspond to the perturbative excitations of the SSSG Lagrangian~\cite{Hollowood:2010dt}. In particular, the gapless excitations in the classical theory get a mass gap at the quantum level.
\vspace{0.5cm}
\acknowledgments
TJH would like to acknowledge the support of STFC grant
ST/G000506/1.
\noindent
JLM acknowledges the support of MICINN
(grants FPA2008-01838 and\break
FPA2008-01177), Xunta de Galicia (Consejer\'\i a de Educaci\'on and INCITE09.296.035PR), the
Spanish Consolider-Ingenio 2010
Programme CPAN (CSD2007-00042), and FEDER.
\noindent
We would both like to
thank Arkady Tseytlin and Ben Hoare for discussions and comments on an earlier draft.
\startappendix
\Appendix{A Tale of Two Dressings}
\label{AppA}
In this appendix, we relate the sigma model dressing formalism, as described in \cite{Hollowood:2009tw},\footnote{In order to compare with that reference one must re-scale $\Lambda\to2\Lambda$.} with the SSSG dressing formalism described in detail in \cite{Hollowood:2010dt} and summarized in Section~\ref{Dressing}. Quantities in the sigma model dressing formalism, if denoted with the same letter as in the SSSG formalism, will be highlighted with a tilde.
In the sigma model formalism in terms of the gauge invariant field ${\cal F}=\theta f\theta f^{-1}$, the central quantity is $\Psi(\lambda)$, which is the solution of the linear system
\EQ{
\Big[\partial_\pm-\frac{\partial_\pm{\cal F}\,{\cal F}^{-1}}{1\pm\lambda}\Big]
\Psi(\lambda)=0\ ,
}
(the spectral parameter $\lambda$ here is not to be confused with the 't~Hooft coupling).
The dressing method starts with
\EQ{
\Psi(\lambda)=\tilde\chi(\lambda)\Psi_0(\lambda)\ ,
}
where
\EQ{
\Psi_0(\lambda)=\exp\Big[\frac{2x^+}{1+\lambda}\Lambda+\frac{2x^-}{1-\lambda}
\Lambda\Big]
}
is the ``vacuum" solution. Then, the dressing transformation takes the form
\EQ{
\tilde\chi(\lambda)=1+
\frac{\tilde{\boldsymbol F}_j\tilde\Gamma^{-1}_{ji}\tilde{\boldsymbol F}_i^\dagger}{\lambda-\lambda_i}\ .
}
Here, $\tilde{\boldsymbol F}_i$ are complex 6-vectors given by
\EQ{
\tilde{\boldsymbol F}_i=\Psi_0(\lambda_i^*){\boldsymbol\varpi}_i\ ,
}
where the ${\boldsymbol\varpi}_i$ are the same constant vector as in the SSSG formalism \eqref{vwr},
\EQ{
\tilde\Gamma_{ij}=\frac{\tilde{\boldsymbol F}_i^*\cdot \tilde{\boldsymbol F}_j}{\lambda_i-\lambda^*_j}\ ,
}
and the poles are related to those in~\eqref{Poles} by means of~\eqref{RelPole}.
The gauge invariant field is given by
\EQ{
{\cal F}=\Psi(0)\ ,
}
and the SSSG field $\gamma$ is given in the two formalisms by \cite{Hollowood:2009tw}
\EQ{
\gamma=\chi(0)^{-1}={\cal F}_0^{-1/2}\tilde\chi^{-1}(1)\tilde\chi(-1){\cal F}_0^{1/2}\ ,
\label{Gamma}
}
where ${\cal F}_0=\Psi_0(0)$ is the vacuum solution. One finds that the two formalisms are simply related via
\EQ{
\Upsilon_0(z)={\cal F}_0^{-1/2}\Psi_0(\lambda)\ ,
}
which requires
\EQ{
z=\frac{1-\lambda}{1+\lambda}\ .
}
This means that
\EQ{
{\boldsymbol F}_i={\cal F}_0^{-1/2}\tilde{\boldsymbol F}_i\ ,\qquad
\xi_i=\frac{1-\lambda_i}{1+\lambda_i}\ .
\label{RelPole}
}
Furthermore, one can show that
\EQ{
-\frac1{\xi_j}\Gamma^{-1}_{ij}=\frac2{(1-\lambda_j)(1+\lambda_i^*)}\tilde\Gamma^{-1}_{ij}\ ,
}
from which it follows
\EQ{
\gamma^{-1}=\chi(0)=1-\frac{{\boldsymbol F}_i\Gamma^{-1}_{ij}{\boldsymbol F}^\dagger_j}{\xi_j}&={\cal F}_0^{-1/2}\Big(1+\frac{2\tilde{\boldsymbol F}_i\tilde\Gamma^{-1}_{ij}{\boldsymbol F}_j^\dagger}
{(1-\lambda_j)(1+\lambda_i^*)}\Big){\cal F}_0^{1/2} \\[5pt]
&={\cal F}_0^{-1/2}\tilde\chi^{-1}(-1)\tilde\chi(1){\cal F}_0^{1/2}\ ,
}
which reproduces~\eqref{Gamma}.
The current associated to the left action $f\to Uf$, with $U\in H$, is $D_\mu f\, f^{-1}$.
Using the equations of motion of $B_\mu$, one can easily show that
\EQ{
D_\mu f\,f^{-1}=\frac12\theta\partial_\mu{\cal F}\,{\cal F}^{-1}\theta\ .
}
Moreover, in \cite{Hollowood:2009tw} it was shown that
\EQ{
\int_{-\infty}^\infty dx\,\big(\partial_0{\cal F}\,{\cal F}^{-1}-\partial_0{\cal F}_0\,
{\cal F}_0^{-1}\big)
=\tilde F_i\tilde\Gamma^{-1}_{ij}\tilde F_j^\dagger\Big|_{x=\infty}-
\tilde F_i\tilde\Gamma^{-1}_{ij}\tilde F_j^\dagger\Big|_{x=-\infty}
}
and from this one finds~\cite{Hollowood:2009sc}
\EQ{
\Delta{\cal Q}_L=-\frac i2\big(\lambda-\lambda^{-1}-\lambda^*+\lambda^{*-1}\big)
m^{-1}\Lambda+\frac i2\big(\lambda+\lambda^{-1}-\lambda^*-\lambda^{*-1}\big)h_{\boldsymbol\Omega}\ ,
}
where $\lambda=(1-\xi)/(1+\xi)= r e^{ip/2}$.
\Appendix{$S^n$ Kinks as Complex Sine-Gordon Q-balls}
\label{AppB}
The $F/G=S^n$ Q-ball kinks have an obvious embedding in $SO(4)/SO(3)$. This yields the complex sine-Gordon theory where the kinks become conventional Q-balls~\cite{Miramontes:2004dr,Bowcock:2008dn}. In this appendix we shall write the SSSG action corresponding to a Q-ball kink in terms of the Lagrangian of the complex sine-Gordon theory
\SP{
{\cal L}_\text{CSG}[\psi,\lambda]= \frac{\partial_\mu\psi \partial^\mu\psi^\ast}{1-\psi\psi^\ast} -\lambda \psi\psi^\ast\,.
\label{CSGlag}
}
The resulting expression provides the identification of the Bohr-Sommerfeld quatization of kinks with the quantization of Q-balls in the complex sine-Gordon theory~\cite{Dorey:1994mg} that leads to~\eqref{iuuCSG}. Remarkably, the result makes use of the boundary term~\eqref{Topol} in a rather non-trivial way.
For $\gamma\in SO(3)\subset SO(4)$ and $A_\pm \in{\mathfrak so}(3)$, we will use the following parameterization of Euler-angle type:
\EQ{
\gamma = e^{(\alpha+\beta) r_1} \,e^{\theta r_3}\, e^{(\alpha-\beta) r_1}\ ,\qquad
A_\pm = a_\pm r_1\,,
\label{Param}
}
where
\EQ{
r_1= \MAT{0&0&0&0\\0&0&0&0\\0&0&0&1\\0&0&-1&0}=\boldsymbol{e}_3 \boldsymbol{e}_4^t -\boldsymbol{e}_4 \boldsymbol{e}_3^t\,,\qquad
r_3= \MAT{0&0&0&0\\0&0&1&0\\0&-1&0&0\\0&0&0&0}\boldsymbol{e}_2 \boldsymbol{e}_3^t -\boldsymbol{e}_3 \boldsymbol{e}_2^t\ ,
}
and $\boldsymbol{e}_a$ denotes the $n+1$~column vector with components $(\boldsymbol{e}_a)_i=\delta_{i,a}$.
Then, a particular embedding of the $SO(4)/SO(3)$ Q-ball kink into $S^n=SO(n+1)/SO(n)$ is specified by the polarization vector ${\boldsymbol\Omega}=\frac{1}{\sqrt{2}}({\boldsymbol\Omega}_1+i{\boldsymbol\Omega}_2)$ by means of
\SP{
r_1 \longrightarrow {\boldsymbol\Omega}_1 {\boldsymbol\Omega}_2^t - {\boldsymbol\Omega}_2 {\boldsymbol\Omega}_1^t =h_{\boldsymbol\Omega}\,,\qquad
r_3 \longrightarrow \boldsymbol{e}_2 {\boldsymbol\Omega}_1^t - {\boldsymbol\Omega}_1 \boldsymbol{e}_2^t
\,.
}
For $SO(4)/SO(3)$, the SSSG action~\eqref{ala} is invariant under abelian $H=SO(2)$ vector gauge transformations, which correspond to
\EQ{
\beta \to \beta +\rho\,,\qquad a_\pm \to a_\pm-\partial_\pm \rho\,,
}
while $\theta$ and $\alpha$ remain invariant. These transformations also leave invariant the combinations of fields
\EQ{
b_\pm = a_\pm + \partial_\pm \beta\,.
}
Then,
\SP{
S_\text{gWZW}=&\frac k{4\pi}\int d^2x\, \Big[2\partial_+\theta \partial_-\theta
+ 8 \cos^2(\theta/2)\Big( \partial_+\alpha \partial_-\alpha
+b_+\partial_-\alpha - b_-\partial_+\alpha\Big) \\[5pt]
&\qquad+8\sin^2(\theta/2)\,b_+b_-
+8\cos^2(\theta/2) \big(\partial_+\alpha \partial_-\beta - \partial_+\beta \partial_-\alpha\big) \Big]\\[5pt]
-&\frac k{\pi}\int d^3x\,\epsilon^{abc} \,\partial_a\big[\cos^2(\theta/2) \,\partial_b\alpha \,\partial_c\beta\Big]\,.
\label{Action0}
}
The last contribution corresponds to the WZ term, whose form has to be determined by the conditon of gauge invariance. In this case, it can be partially fixed by removing all the terms that depend explicitly on $\beta$. The result is
\SP{
S_\text{gWZW}=&\frac k{4\pi}\int d^2x\, \Big[2\partial_+\theta \partial_-\theta
+8 \cot^2(\theta/2)\partial_+\alpha \partial_-\alpha
\\[5pt]
&
\hspace{-0.5cm}
+8\sin^2(\theta/2)\big(b_+ -\cot^2(\theta/2)\partial_+\alpha\big)\,
\big(b_- +\cot^2(\theta/2)\partial_-\alpha\big)
+ \epsilon^{\mu\nu} \partial_\mu F_\nu\Big]\,,
\label{Action1}
}
where $F_\mu$ parameterizes the remaining ambiguities coming from the WZ term.
If we ignore the boundary term~\eqref{Topol}, we can take $F_\mu=0$. Then, using the equations of motion of $a_\pm$,
\SP{
b_\pm = \pm \cot^2(\theta/2)\partial_\pm\alpha\,,
}
the SSSG action becomes
\SP{
&S_\text{gWZW}[\gamma,A_\mu] -\frac k{2\pi}\int d^2x\,{\rm Tr}\left(\Lambda
\gamma^{-1}\Lambda\gamma-\Lambda^2\right)\\[5pt]
&\qquad\qquad
=\frac k{2\pi}\int d^2x\,\Big({\cal L}_\text{CSG}[\cos(\theta/2)e^{i\alpha},-4m^2] - 4m^2\Big)\,,
\label{CSGneg}
}
which involves the complex sine-Gordon Lagrangian with negative mass term. This Lagrangian has a $U(1)$ degenerate set of vacua with $|\psi|=1$. At rest, its soliton solutions are time independent (non-dyonic) kinks that interpolate between two different vacua~\cite{Lund:1976ze,Lund:1977dt} (see also~\cite{Miramontes:2004dr}). In addition, notice that this Lagrangian does not
have a good expansion in terms of fields around their vacuum values due to the $\cot^2 \theta$ term
in~\eqref{Action1}, or $(1-|\psi|^2)^{-1}$ in~\eqref{CSGlag}.
However, the full SSSG action does include the boundary term~\eqref{Topol}. Then, the condition of gauge invariance fixes\footnote{Notice that ${\rm Tr}\big(r_1\,\phi\big)$ is gauge invariant. Moreover, using the Baker-Campbell-Hausdorff formula,
\SP{
{\rm Tr}\big(r_1\,\phi\big)= -4\alpha + \frac{1}{3} \alpha\, \theta^2+\cdots\,,
\label{BCH}
}
and all the terms in the ellipsis are proportional to $\theta^2$.
}
\SP{
F_\mu = - {\rm Tr}\big(r_1\,\phi\big)\,\partial_\mu\beta,
}
and the true SSSG action~\eqref{ala} reads
\SP{
S=&\frac k{4\pi}\int d^2x\, \Big[2\partial_+\theta \partial_-\theta
+8 \cot^2(\theta/2)\partial_+\alpha \partial_-\alpha
-8m^2 \sin^2(\theta/2)\\[5pt]
&
\hspace{-0.5cm}
+8\sin^2(\theta/2)\big(b_+ -\cot^2(\theta/2)\partial_+\alpha\big)\,
\big(b_- +\cot^2(\theta/2)\partial_-\alpha\big)
- \epsilon^{\mu\nu} \partial_\mu\big(b_\nu\, {\rm Tr}(r_1\,\phi)\big)\Big]\,.
\label{Action2}
}
Now, we can use the equations of motion of $a_\pm$ and $\alpha$,
\EQ{
b_\pm = \pm \cot^2(\theta/2)\partial_\pm\alpha\,,\qquad
\partial_+ b_- -\partial_-b_+=0\,,
\label{eom1}
}
to write $b_\pm$ in terms of a new field $\varphi$ as follows
\EQ{
b_\pm = \pm \cot^2(\theta/2)\partial_\pm\alpha = \partial_\pm\varphi\,.
\label{eom2}
}
This provides an explicit relation between the SSSG action for the $S^n$ kink Q-ball and the complex sine-Gordon Lagrangian~\eqref{CSGlag} with positive mass term
\SP{
S =
\frac k{2\pi}\int d^2x\, \Big({\cal L}_\text{CSG}[\sin(\theta/2)e^{i\varphi},+4m^2] -\frac{1}{2} \epsilon^{\mu\nu}\partial_\mu\big[\big({\rm Tr}(r_1\,\phi)+4\alpha\big)\, \partial_\nu\varphi\big]\Big)\,.
\label{CSGpos}
}
This Lagrangian has a non-degenerate vacuum at $|\psi|=0$, and its soliton solutions are $Q$-balls that carry $U(1)$ Noether charge~\cite{Getmanov:1977hk} (see also~\cite{Miramontes:2004dr}). For these solutions at rest, $\varphi$ only depends on $t$ and ${\rm Tr}(r_1\,\phi)+4\alpha$ vanishes at $x=\pm\infty$. Therefore, the last term in~\eqref{CSGpos} vanishes and the SSSG action for a kink in $S^n$ is equal to the action of a complex sine-Gordon $Q$-ball, which is the result used in Section~\ref{SCQD} to quantize the $S^n$ SSSG kinks by means of the Bohr-Sommerfeld rule, eq.~\eqref{iuuCSG}.
Finally, we can use the particular example of the $SO(4)/SO(3)$ SSSG theory to illustrate the need to supplement the SSSG action with the boundary term~\eqref{Topol}. Consider the action~\eqref{Action2} for $\gamma\in H=SO(2)$, which corresponds to $\theta=0$,
and leave the normalization of the boundary term free.\footnote{Notice that $\beta$ is not a good coordinate around $\theta=0$, in the same way that the polar angle is not a good coordinate around $r=0$.} Then, using~\eqref{Action0} and~\eqref{BCH},
\SP{
S_\text{gWZW} &+ NS_\text{bt}=\frac {2k}{\pi}\int d^2x\, \Big[\partial_+\alpha \partial_-\alpha
+a_+\partial_-\alpha - a_-\partial_+\alpha\\[5pt]
&
+N\big(\partial_+(a_-\,\alpha)- \partial_-(a_+\,\alpha)\big)\Big] -\frac k{\pi}\int d^3x\,\epsilon^{abc} \,\partial_a\big[\partial_b\alpha \,\partial_c\beta\Big]\,,
}
which shows that the naive WZ term vanishes.
Then, the choice $N=1$ is singled out as the only one that ensures gauge invariance:
\SP{
S_\text{gWZW} + S_\text{bt}=&\frac {2k}{\pi}\int d^2x\, \Big[\partial_+\alpha \partial_-\alpha
+\big(\partial_+ a_--\partial_- a_+\big)\alpha\Big]\,. \\[5pt]
}
|
\section{Introduction}\label{section1}
This paper studies orthogonal frequency-division multiple access
(OFDMA) based cellular systems with universal frequency reuse, in
which adjacent cells share the same frequency band for simultaneous
transmission to improve the spectrum efficiency. However, for such
systems, the inter-cell interference (ICI) control becomes crucial,
which has recently drawn significant attention (see, e.g.,
\cite{Gesb10} and references therein). In general, there are two
approaches to cope with the ICI in multi-cell systems \cite{Gesb10}:
\emph{interference coordination} and \emph{network MIMO}
(multiple-input multiple-output). The former approach mitigates the
ICI via cooperative resource allocation across different cells based
on their shared channel state information (CSI), where the latter
approach utilizes the ICI via baseband-level signal cooperation for
joint encoding and/or decoding at the base stations (BSs). Although
promising from the viewpoint of theoretical performance, network
MIMO requires the baseband time synchronization as well as message
sharing among different BSs, which is challenging to implement for
existing cellular systems. As such, in this paper, we focus our
study on the interference coordination approach.
Existing resource allocation schemes (e.g.,
\cite{Vent09,Yu2010,Da2011ICC}) for multi-cell OFDMA systems have
adopted a \emph{subcarrier-separation} approach, whereby all
subcarriers (SCs) are treated as separate dimensions for user
selection and power allocation. However, a recent study
\cite{Cada2008Insepa} has revealed that for a \emph{parallel
interference channel} (PIC) consisting of parallel Gaussian
interference subchannels, joint precoding over parallel subchannels
improves the sum capacity over the case without precoding for some
specially designed channel realizations, i.e., the PIC is in general
\emph{non-separable}. Motivated by this finding, in this paper, we
study a new approach to design cooperative resource allocation for
multi-cell OFDMA, by exploiting frequency-domain precoding over
parallel SCs.
The precoding technique used for the PIC in \cite{Cada2008Insepa} is
known as \emph{interference alignment} (IA). By properly aligning
the interference at the receivers via linear precoding at the
transmitters, for $K$-user Gaussian interference channels, it is
shown in \cite{Cada2008Aug} that the IA technique achieves a
sum-rate multiplexing gain of $K/2$ per time, frequency or antenna
dimension. In other words, IA enables interference-free
communications for all the users provided that each user utilizes
only half of the available degrees of freedom (DoF). Although IA has
been largely investigated in cases with time-domain symbol extension
or with spatial beamforming via multiple antennas, how to exploit
frequency-domain IA to design efficient OFDMA resource allocation in
a multi-cell scenario remains open and has been rarely studied in
the literature, to our best knowledge.
In this paper, we investigate the problem of exploiting IA in
cooperative resource allocation in OFDMA-based cellular systems. For
the purpose of exposition, we study the OFDMA downlink transmission
in a simplified three-cell system with universal frequency reuse.
All the BSs and user terminals are assumed to be each equipped with
a single antenna, and the system is thus modeled by a \emph{SISO
(single-input single-output) interfering broadcast channel (BC)}.
With the objective of maximizing the system weighted sum-rate, the
joint optimization of frequency-domain precoding via IA, SC
scheduling with user selection, and power allocation is studied in
this paper. From the numerical experiments under a symmetric channel
setup (with unit average channel gain for all in-cell direct links
and the same average channel gain for all cross-cell interference
links), we find that the IA-based resource allocation scheme
demonstrates its advantages over the traditional scheme without
frequency-domain precoding only when the cross-cell link has a
comparable strength as the direct link, and the receiver
signal-to-noise (SNR) is sufficiently large. Motivated by this
observation, a hybrid scheme is proposed for practical cellular
systems with heterogenous channel conditions. In this hybrid scheme,
the total spectrum is divided into two subbands, over which the
IA-based scheme and the traditional scheme are applied for resource
allocation to users located in the \emph{cell-intersection region}
and \emph{cell-non-intersection region}, respectively. It is shown
by simulation results that this hybrid scheme can be flexibly
designed to exploit the downlink IA gains for OFDMA-based cellular
systems with cooperative interference control.
\begin{figure}[!t]
\centering
\includegraphics[width=80mm]{Fig1.eps} \\
\vspace{-12pt}
\caption{System model for a cluster of three cells.}
\label{fig.1}
\end{figure}
\section{System Model}\label{section2}
As shown in Fig. \ref{fig.1}, for the purpose of exposition, we
consider a three-cell system, in which all cells share the same
frequency band for simultaneous downlink transmission. In each cell,
OFDMA is assumed for the downlink transmission, where the total
shared bandwidth is equally divided into $N$ SCs indexed by $n \in
\Lambda {\rm{ = }}\left\{ {1,2,...,N} \right\}$. The users in each
cell are scheduled for transmission over orthogonal SCs, while the
users' SC allocation are allowed to change from one scheduling
time-slot to another. A slow fading environment is assumed, in which
all the channels involved in the system are constant during each
scheduling slot, but can vary from slot to slot. In one particular
slot, each SC is assumed to be used by at most one user inside each
cell due to OFDMA, but allowed to be shared by users from different
cells due to universal spectrum sharing. We use $\Omega = \left\{
{1,2,3} \right\}$ to represent the set of three cells, and denote
the user associated with Cell $m$ as ${k_m} \in {\Delta _m}{\rm{ =
}}\left\{ {1,2,...,{K_m}} \right\}, m \in \Omega$, where $K_m$ is
the total number of users in Cell $m$. In addition, the complex
downlink baseband channel response from the BS of Cell $j$ to user
$k_m$ in Cell $m$ at SC $n$ is denoted as $h_{j{k_m}}^n$. Each BS is
assumed to have a total transmit power constraint given by
$P_m^{{\rm{tot}}}, m \in \Omega$. Furthermore, we use the matrix
$\bf{\pi}$ of size $3\times N$ to denote the SC allocation in the
whole system, of which the $(m,n)$-th element, denoted by $\pi _m^n$
with $m\in\Omega, n\in\Lambda$, indicates the user assigned to SC
$n$ in Cell $m$.
Note that for all the resource allocation schemes presented in the
sequel, we assume that there exists a central controller that
collects all the CSI in the network via dedicated feedback channels
or backhaul networks, and is thus enabled to perform a centralized
resource allocation.
\section{Traditional Resource Allocation}\label{section3}
In this section, we present one traditional scheme for cooperative
resource allocation in the studied three-cell system, which is based
on the subcarrier-separation principle, i.e., all the SCs are
treated as separate dimensions for user selection and power
allocation. First, supposing that $\pi _m^n=k_m$, the baseband
signal received by user $k_m$ at SC $n$ is written as
\begin{equation}\label{equ.1}
y_{{k_m}}^n = h_{m{k_m}}^nx_m^n + \sum\limits_{j = 1,j \ne m}^3
{h_{j{k_m}}^nx_j^n} + z_{{k_m}}^n,
\end{equation}
where $x_m^n$ is the complex symbol transmitted by BS $m$ at SC $n$,
${E}[{{{|{x_m^n}|}^2}}] = p_m^n$ with $p_m^n$ denoting the power
allocated, and $z_{{k_m}}^n$ is the circularly symmetric complex
Gaussian (CSCG) noise at the receiver with zero mean and variance
${\sigma ^2}$.
The signal-to-interference-plus-noise-ratio (SINR) of user $k_m$ in
Cell $m$ at SC $n$ is given by
\begin{equation}\label{equ.2}
SINR_{m{k_m}}^n = \frac{{p_m^ng_{m{k_m}}^n}}{{1 + \sum\limits_{j =
1,j \ne m}^3 {p_j^ng_{j{k_m}}^n} }},
\end{equation} where $g_{j{k_m}}^n = {{{{| {h_{j{k_m}}^n} |}^2}} \mathord{/
{\vphantom {{{{| {h_{j{k_m}}^n}|}^2}} {{\sigma ^2}}}}
\kern-\nulldelimiterspace} {{\sigma ^2}}}$. Then, the maximum achievable transmission rate at this SC is given by
(normalized to be in bps/Hz)
\begin{equation}\label{equ.3}
r_{m{k_m}}^n = \frac{1}{N}{\log _2}\left( {1 + SINR_{m{k_m}}^n}
\right),
\end{equation}
Hence, the problem of maximizing the weighted sum-rate of all users
in the system can be expressed as
\begin{equation}\label{equ.4}
\begin{array}{l}
\mathop {\max }\limits_{{\bf{p}} \succeq 0,{\bf{\pi }} \in \Pi } {\rm{ }}U = \sum\limits_{n = 1}^N {\sum\limits_{m = 1}^3 {{\omega _{m{k_m}}}r_{m{k_m}}^n} } \\
s.t.{\rm{ }}\sum\limits_{n = 1}^N {p_m^n} \le P_m^{{\rm{tot}}},{\rm{ }}m \in \Omega \\
\end{array}
\end{equation}
where ${\bf{p}} = \left\{ {{{\bf{p}}^1},...,{{\bf{p}}^N}} \right\}$
is a $3$-by-$N$ power allocation matrix consisting of column vectors
${{\bf{p}}^n} = {\left[ {p_1^n,p_2^n,p_3^n} \right]^T}, n \in
\Lambda$, $\Pi$ denotes the set of all possible SC allocation
matrices for ${\bf \pi}$, and ${\omega _{m{k_m}}}$ is the
(non-negative) weight of user $k_m$ in Cell $m$.
The problem in (\ref{equ.4}) can be shown to be non-convex over
${\bf{p}}$ due to the non-concave rate functions in (\ref{equ.3})
even for a fixed ${\bf \pi}$, and is thus in general non-convex with
arbitrary ${\bf \pi}$. Nevertheless, the Lagrange duality method can
be applied to this problem to obtain a set of close-to-optimal
solutions, which, as verified by our numerical experiments, usually
converge to the optimal solutions when the number of SCs becomes
large (i.e., the duality gap converges to zero as $N\rightarrow
\infty$). More specifically, the partial Lagrangian of problem
(\ref{equ.4}) is given by
\begin{equation}\label{equ.5}
L\left( {{\bf{p}},{\bf{\pi }},{\bf{\lambda }}} \right) = U +
{{\bf{\lambda }}^T}\left( {{{\bf{P}}^{\rm{tot}}} - \sum\limits_{n =
1}^N {{{\bf{p}}^n}} } \right),
\end{equation} where ${{\bf{P}}^{{\rm{tot}}}} = {\left[
{P_1^{{\rm{tot}}},P_2^{{\rm{tot}}},P_3^{{\rm{tot}}}} \right]^T}$ is
the power constraint vector, and ${\bf{\lambda }} = {\left[
{{\lambda _1},{\lambda _2},{\lambda _3}} \right]^T}$ is the vector
of non-negative dual variables. The Lagrange dual function then
becomes
\begin{equation}\label{equ.6}
f\left( {\bf{\lambda }} \right) = \mathop {\max }\limits_{{\bf{p}}
\succeq 0,{\bf{\pi }} \in \Pi } L\left( {{\bf{p}},{\bf{\pi
}},{\bf{\lambda }}} \right).
\end{equation} Thus, the dual problem can be defined as
\begin{equation}\label{equ.7}
\mathop {\min }\limits_{{\bf{\lambda }} \succeq 0} f\left(
{\bf{\lambda }} \right).
\end{equation}
For a given vector $\bf{\lambda}$, $f\left( {\bf{\lambda }} \right)$
can be simplified as
\begin{equation}\label{equ.8}
f\left( {\bf{\lambda }} \right) = \mathop {\max }\limits_{{\bf{p}}
\succeq 0,{\bf{\pi }} \in \Pi } \sum\limits_{n = 1}^N {{f_n}} +
{{\bf{\lambda }}^T}{{\bf{P}}^{{\rm{tot}}}},
\end{equation} where
\begin{equation}\label{equ.9}
{f_n} = \sum\limits_{m = 1}^3 {\left( {{\omega _{{k_m}}}r_{m{k_m}}^n - {\lambda _m}p_m^n} \right)}.
\end{equation}
The maximization problem in (\ref{equ.8}) is thus decoupled into
parallel per-SC based subproblems, which are given by
\begin{equation}\label{equ.10}
\mathop {\max }\limits_{{{\bf{p}}^n} \succeq 0,{{\bf{\pi }}^n} \in
{\Pi ^n}} {f_n},{\rm{ }}n \in \Lambda,
\end{equation} where ${{\bf{\pi }}^n} = {\left[ {\pi _1^n,\pi _2^n,\pi _3^n} \right]^T} \in {\Pi
^n}$ with ${\Pi ^n}$ denoting the $n$th column of ${\Pi}$. The
solutions to the problems in (\ref{equ.10}) have been studied in
e.g. \cite{Vent09,Yu2010,Da2011ICC} by various iterative methods;
the details are thus omitted for brevity. Last, subgradient-based
methods such as the ellipsoid method \cite{Da2011ICC} can be used to
iteratively search for the optimal ${\bf{\lambda }}$ in the dual
problem to make the corresponding power allocation solutions of the
problems in (\ref{equ.10}) satisfy all the per-BS power constraints.
Note that the above joint user SC and power allocation scheme is
based upon the premise that there is no frequency-domain precoding
over parallel SCs at each BS. Next, we will propose an IA-based
resource allocation scheme with frequency-domain precoding.
\section{IA-Based Resource Allocation}\label{section4}
In this section, we propose a new scheme for cooperative multi-cell
OFDMA downlink resource allocation with IA-based frequency-domain
precoding. Specifically, two non-adjacent SCs, denoted by
$(n_1,n_2)$, are grouped to form the $n$th SC-pair, which supports
simultaneous downlink transmission to three users, denoted by
$(k_1,k_2,k_3)$, each being selected from one of the three cells,
via linear transmit precoding over the two chosen SCs. Note that for
notational convenience, we use $n$ to denote the SC-pair index in
this section, where $n \in \tilde \Lambda {\rm{ = }}\left\{
{1,2,...,{N \mathord{\left/ {\vphantom {N 2}} \right.
\kern-\nulldelimiterspace} 2}} \right\}$ (assume that $N$ is an even
number).
Without loss of generality, we consider the baseband signals for
user $k_1$ over the $n$th SC-pair $(n_1,n_2)$, which are given by
\begin{equation}\label{equ.11.1}
y_{{k_1}}^{{n_1}} = h_{1{k_1}}^{{n_1}}x_1^{{n_1}} +
h_{2{k_1}}^{{n_1}}x_2^{{n_1}} + h_{3{k_1}}^{{n_1}}x_3^{{n_1}} +
z_{{k_1}}^{{n_1}},
\end{equation}
\begin{equation}\label{equ.11.2}
y_{{k_1}}^{{n_2}} = h_{1{k_1}}^{{n_2}}x_1^{{n_2}} +
h_{2{k_1}}^{{n_2}}x_2^{{n_2}} + h_{3{k_1}}^{{n_2}}x_3^{{n_2}} +
z_{{k_1}}^{{n_2}},
\end{equation}
where $x_m^{{n_1}} = v_m^{{n_1}}s_m^{{n}},x_m^{{n_2}} =
v_m^{{n_2}}s_m^{{n}},m \in \Omega $ with $v_m^{{n_1}},v_m^{{n_2}}$
being the precoder weights for user $k_m$ of Cell $m$ over the $n$th
SC-pair, and $s_m^n$ being the information-bearing symbol for user
$k_m$. Together, (\ref{equ.11.1}) and (\ref{equ.11.2}) can be
further expressed in the matrix form shown as follows:
\begin{equation}\label{equ.12}
{\bf{y}}_{{k_1}}^n = {\bf{H}}_{1{k_1}}^n{\bf{v}}_1^ns_1^n +
{\bf{H}}_{2{k_1}}^n{\bf{v}}_2^ns_2^n +
{\bf{H}}_{3{k_1}}^n{\bf{v}}_3^ns_3^n + {\bf{z}}_{{k_1}}^n,
\end{equation}
where ${\bf{v}}_m^n = {\left[ {v_m^{{n_1}},v_m^{{n_2}}} \right]^T}$
is the precoding vector for user $k_m$,
\begin{equation}\label{equ.12.H}
{\bf{H}}_{m{k_1}}^n = \left[ {\begin{array}{*{20}{c}}
{h_{m{k_1}}^{{n_1}}} & 0 \\
0 & {h_{m{k_1}}^{{n_2}}} \\
\end{array}} \right],m \in \Omega,
\end{equation}
is the equivalent (diagonal) channel matrix from BS $m$ to user
$k_1$, and ${\bf{z}}_{{k_1}}^n$ is the noise vector for user $k_1$,
all defined for the $n$th SC-pair. Based on the analysis for user
$k_1$, the downlink transmission from the three BSs to their
respective users over the $n$-th SC pair can be represented by
\begin{equation}\label{equ.13}
{\bf{y}}_{{k_m}}^n = {\bf{H}}_{m{k_m}}^n{\bf{v}}_m^ns_m^n +
\sum\limits_{j = 1,j \ne m}^3 {{\bf{H}}_{j{k_m}}^n{\bf{v}}_j^ns_j^n}
+ {\bf{z}}_{{k_m}}^n,
\end{equation}
$ m\in\Omega, n\in\tilde{\Lambda}$. At the receiver of user $k_m$,
an interference suppression vector ${\bf{u}}_m^n = \left[
{u_m^{{n_1}},u_m^{{n_2}}} \right]$ is pre-multiplied with the
received signal to eliminate the inter-cell interference, which
leads to the following equivalent channel model:
\begin{equation}\label{equ.14}
\tilde y_{{k_m}}^n = {\bf{u}}_m^n{\bf{y}}_{{k_m}}^n = \tilde
h_{m{k_m}}^ns_m^n + \sum\limits_{j = 1,j \ne m}^3 {\tilde
h_{j{k_m}}^ns_j^n} + \tilde z_{{k_m}}^n,
\end{equation}
where
\begin{equation}\label{equ.14.h}
\tilde h_{m{k_m}}^n = {\bf{u}}_m^n{\bf{H}}_{m{k_m}}^n{\bf{v}}_m^n,
\end{equation}
\begin{equation}\label{equ.14.RI}
\tilde h_{j{k_m}}^n = {\bf{u}}_m^n{\bf{H}}_{j{k_m}}^n{\bf{v}}_j^n, j
\in \Omega, j \ne m,
\end{equation}
\begin{equation}\label{equ.14.z}
\tilde z_{{k_m}}^n = {\bf{u}}_m^n{\bf{z}}_{{k_m}}^n.
\end{equation}
Note that (\ref{equ.14}) bears a similar expression as
(\ref{equ.1}), with $\tilde h_{m{k_m}}^n, \tilde h_{j{k_m}}^n$ being
equivalent channel gains for direct and inter-cell links,
respectively. Assuming that ${\bf{v}}_m^n,{\bf{u}}_m^n, m \in
\Omega, n\in\tilde{\Lambda}$ are all of unit norm and
${E}[|s_m^{{n}}|^2]=\tilde{p}_m^n$, the achievable rate over the
$n$th SC-pair for user $k_m$ is given by
\begin{equation}\label{equ.15}
\tilde r_{m{k_m}}^n = \frac{1}{N}{\log _2}\left( {1 + \frac{{{\tilde
p}_m^n\tilde g_{m{k_m}}^n}}{{1 + \sum\limits_{j = 1,j \ne m}^3
{\tilde{p}_j^n\tilde g_{j{k_m}}^n} }}} \right),
\end{equation} where $\tilde g_{m{k_m}}^n = {{{{| {\tilde h_{m{k_m}}^n}|}^2}}
\mathord{{\vphantom {{{{| {\tilde h_{m{k_m}}^n} |}^2}} {{\sigma
^2}}}} /\kern-\nulldelimiterspace} {{\sigma ^2}}}$.
Since searching for the optimal SC-pairs over $\Lambda$ is
computationally prohibitive, we use a pre-determined SC-pairing
method as follows: for the $n$th SC-pair, the two SCs are given by
$n_1=n, n_2=n+N/2$. This paring method is to maximize the frequency
gap between the two SCs in each pair, so as to maximally exploit the
frequency-domain channel diversity. With the above SC-pairs, the
weighted sum-throughput of the three-cell system is obtained by
solving the following problem:
\begin{equation}\label{equ.16}
\begin{array}{l}
\mathop {\max }\limits_{{\bf{\tilde p}} \succeq 0,{\bf{\tilde \pi }} \in \tilde \Pi } {\rm{ }}\tilde U = \sum\limits_{n = 1}^{{N \mathord{\left/
{\vphantom {N 2}} \right.
\kern-\nulldelimiterspace} 2}} {\sum\limits_{m = 1}^3 {{{\omega }_{m{k_m}}}\tilde r_{m{k_m}}^n} } \\
s.t. {\ } \sum\limits_{n = 1}^{{N \mathord{\left/
{\vphantom {N 2}} \right.
\kern-\nulldelimiterspace} 2}} {{\tilde p}_m^n} \le P_m^{{\rm{tot}}},{\rm{ }}m \in \Omega \\
\end{array}
\end{equation}
where ${\bf{\tilde p}} = \left\{ {{{\bf{\tilde
p}}^1},...,{{\bf{\tilde p}}^{{N \mathord{\left/ {\vphantom {N 2}}
\right. \kern-\nulldelimiterspace} 2}}}} \right\}$ with ${\bf{\tilde
p}}^n=[\tilde{p}_1^n,...,\tilde{p}_3^n]^T, n\in\tilde{\Lambda}$, and
${\bf{\tilde \pi}}$ is a $3$-by-$N/2$ matrix with its $n$th column
vector ${{\bf{\tilde \pi }}^n} \in {\tilde \Pi ^n}$ specifying the
users $(k_1,k_2,k_3)$ selected from the three cells for the $n$th
SC-pair. Note that ${\tilde \Pi}$ and ${\tilde \Pi^n}$ are similarly
defined as $\Pi$ and $\Pi^n$ in Section III, respectively, while it
is worth noting that $n$ here refers to the SC-pair index instead of
the SC index in Section III.
Similarly as the derivation given in Section III, the problem in
(\ref{equ.16}) can be decoupled into parallel per-SC-pair based
resource allocation subproblems in the dual domain, i.e.,
\begin{equation}\label{equ.18}
\mathop {\max }\limits_{{{\bf{\tilde p}}^n} \succeq 0,{{{\bf{\tilde
\pi }}}^n} \in {{\tilde \Pi }^n}} {\tilde f_n} = \sum\limits_{m =
1}^3 {\left( {{{\omega }_{m{k_m}}}\tilde r_{m{k_m}}^n - {{\tilde
\lambda }_m}\tilde{p}_m^n} \right)} ,{\rm{ }}n \in \tilde \Lambda,
\end{equation} where ${\bf{\tilde \lambda }} = {\left[ {{{\tilde \lambda }_1},{{\tilde \lambda }_2},{{\tilde \lambda }_3}}
\right]^T}$ is the vector of non-negative dual variables associated
with the per-BS power constraints.
Consider first the case when perfect IA is achievable, i.e., all the
inter-cell interference gains in (\ref{equ.14.RI}) become zero. In
this case, the residual interference (RI) terms in the denominator
of (\ref{equ.15}) vanish, which results in decoupled user power
optimization in (\ref{equ.18}) for each SC-pair. Specifically, by
letting
\begin{equation}\label{equ.19}
\frac{{\partial {{\tilde f}_n}}}{{\partial \tilde{p}_m^n}} =
\frac{{{{\omega }_{m{k_m}}}}}{N}\frac{{\tilde g_{m{k_m}}^n}}{{\ln
2\left( {1 + \tilde{p}_m^n\tilde g_{m{k_m}}^n} \right)}} - {\tilde
\lambda _m} = 0,
\end{equation}
we obtain the following optimal power allocation:
\begin{equation}\label{equ.20}
\tilde{p}_m^n = {\left( {\frac{{{{\omega }_{m{k_m}}}}}{{{{\tilde
\lambda }_m}N\ln 2}} - \frac{1}{{\tilde g_{m{k_m}}^n}}} \right)^ +
}, m\in\Omega, n\in\tilde{\Lambda},
\end{equation}
which resembles the conventional ``water-filling'' solution.
However, to our best knowledge, no closed-form expressions for
${\bf{v}}_m^n$ and ${\bf{u}}_m^n$ that achieve the perfect IA have
been found for our studied scenario. Thus, we resort to the existing
``Distributed Interference Alignment'' algorithm in \cite{Goma2008}
to compute the transmitter precoding and receiver interference
suppression vectors given the CSI for any set of three users and
SC-pair. Based on our numerical experiments, it is found that with
this algorithm, small but non-zero RI usually exists for the
equivalent three-user $2\times 2$ MIMO interference channel given by
(\ref{equ.13}), probably due to the particular structure of diagonal
direct-link and cross-link channel matrices. In this case with the
RI, the problem in (\ref{equ.16}) can be solved similarly as that in
(\ref{equ.4}) in the case of traditional scheme with virtually $N/2$
instead of $N$ SCs. For brevity, the details are omitted.
\begin{figure}[!t]
\centering
\includegraphics[width=80mm]{Fig2.eps} \\ \vspace{-6pt}
\caption{System throughput under a symmetric channel setup with $h= 1$.}
\label{fig.2} \vspace{-6pt}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=80mm]{Fig3.eps} \\ \vspace{-6pt}
\caption{System throughput under a symmetric channel setup with $h=0.1$.}
\label{fig.3} \vspace{-8pt}
\end{figure}
Next, we compare the performance of the tradition resource
allocation scheme in Section III and the newly proposed IA-based
scheme in terms of the system sum-rate (i.e., ${\omega_{m{k_m}}}=1,
\forall m\in\Omega, k_m\in\Delta_m$). In order to gain intuition
from the numerical results, we assume an ideal symmetric channel
model for the three-cell system in Fig. \ref{fig.1}, in which
$h_{j{k_m}}^n$'s are modeled as independent CSCG random variables
$\sim \mathcal{CN}\left( {0,1} \right)$ for $j=m$, and $\sim
\mathcal{CN}\left( {0,h} \right)$ for $j\neq m$, respectively. In
addition, it is assumed that $N=64$, $K_m=4, m\in\Omega$,
$P_m^{{\rm{tot}}}$'s are same for all three BSs, and $\sigma^2=1$.
Fig. \ref{fig.2} shows the sum-rate versus the average direct-link
SNR at each user's receiver with $h=1$, i.e., the direct-link has
the same average power gain as the cross-link in the symmetric
channel setup. It is observed that the IA-based scheme with the RI
performs slightly worse than that without the RI (assuming perfect
IA and thus neglecting the RI terms in (\ref{equ.15})). In addition,
the IA-based scheme is observed to perform better than the
traditional scheme when SNR is larger than 15dB, and eventually
achieve a DoF gain of $3/2$ compared to the traditional scheme when
SNR goes to infinity. This DoF gain can be explained as follows: For
the IA-based scheme (without RI), three data streams are transmitted
without mutual interference over each SC-pair consisting of two SCs,
while for the traditional scheme, when the SNR is sufficiently high,
the optimal SC allocation tends to be orthogonal, i.e., each SC is
assigned to only one user from a particular cell, while no users
from the other two cells are assigned to the same SC.
Similar observations are obtained in Fig. \ref{fig.3} for the
symmetric channel setup with $h=0.1$. However, in this case with
weak cross-link interference, the SNR threshold beyond which the
IA-based scheme outperforms the traditional scheme increases from
15dB in the previous case of $h=1$ to 35dB in the present case. In
addition, the IA-based scheme is observed to perform notably worse
than the traditional scheme from low to moderate SNRs (0-20dB) that
are typical for cellular systems. This performance gap is because
the IA-based scheme always transmits three data streams over two SCs
regardless of the SNR, while the traditional scheme can support up
to 3 users each from one cell at any single SC when the inter-cell
interference is sufficiently weak as in the present case.
\section{Hybrid Scheme}\label{section5}
The numerical results for the symmetric channel setup reveal that
the IA-based scheme with frequency-domain precoding provide
throughput gains over the traditional scheme without precoding only
when the cross-link has a comparable strength with the direct-link
and the operating SNR is sufficiently high. The above conditions are
rarely satisfied in practical cellular systems due to randomized
user locations and distance-dependent signal attenuation. This
motivates our following hybrid scheme for resource allocation in a
three-cell system with heterogenous channel conditions (cf. Fig.
\ref{fig.4}): For the users located within the six sectors (two from
each cell) that lie in the intersection region of the three cells,
namely \emph{cell-intersection region} (CIR), a dedicated set of
SCs, denoted by $\Phi$, is reserved over which the IA-based scheme
is applied, since in this region the aforementioned operating
conditions for IA are more likely to be satisfied; in contrast, for
the rest of the users located in the \emph{cell-non-intersection
region} (CNIR), the remaining set of SCs, denoted by $\Phi'$, is
allocated over which the traditional scheme is applied. For example,
we can design the sizes of $\Phi$ and $\Phi'$ to be proportional to
the number of sectors in the CIR and CNIR, respectively, i.e., $\Phi
= \left\{ {1,2,...,\left\lfloor {{N \mathord{\left/
{\vphantom {N 6}} \right.
\kern-\nulldelimiterspace} 6}} \right\rfloor ,...,1 + \left\lfloor {{N \mathord{\left/
{\vphantom {N 2}} \right.
\kern-\nulldelimiterspace} 2}} \right\rfloor ,} \right.$ $\left. {2 + \left\lfloor {{N \mathord{\left/
{\vphantom {N 2}} \right.
\kern-\nulldelimiterspace} 2}} \right\rfloor ,...,\left\lfloor {{{2N} \mathord{\left/
{\vphantom {{2N} 3}} \right.
\kern-\nulldelimiterspace} 3}} \right\rfloor } \right\}$, and thus $\Phi^\prime = \Lambda - \Phi$.
Similar to the problems formulated in (\ref{equ.4}) and
(\ref{equ.16}), the following problem maximizes the system weighted
sum-rate in the case of hybrid scheme (thus similarly solvable):
\begin{eqnarray*}
\mathop {\max }\limits_{{{\bf{p}}_h} \succeq 0,{{\bf{\pi }}_h} \in
{\Pi _h}} {U_h} = \sum\limits_{n \in \tilde \Phi }^{}
{\sum\limits_{m = 1}^3 {{\omega _{m{k_m}}}\tilde r_{m{k_m}}^n} } +
\sum\limits_{n \in \Phi '}^{} {\sum\limits_{m = 1}^3 {{\omega
_{m{k_m}}}r_{m{k_m}}^n} }
\end{eqnarray*}
\begin{equation}\label{equ.23}
s.t.{\rm{ }}\sum\limits_{n \in \tilde \Phi }^{} {\tilde{p}_m^n} +
\sum\limits_{n \in \Phi '}^{} {p_m^n} \le P_m^{{\rm{tot}}},{\rm{
}}m \in \Omega
\end{equation} where ${r_{m{k_m}}^n}$ and ${\tilde r_{m{k_m}}^n}$
are given in (\ref{equ.3}) and (\ref{equ.15}) respectively; the
power and SC allocations are similarly defined as
${{\bf{p}}_h},{{\bf{\pi }}_h}$, and $\tilde \Phi = \left\{
{1,2,...,\left\lfloor {{N \mathord{\left/ {\vphantom {N 6}} \right.
\kern-\nulldelimiterspace} 6}} \right\rfloor } \right\}$ is used to
index the SC-pairs in $\Phi$.
Fig. \ref{fig.5} shows the system sum-rates for various schemes in a
three-cell system with heterogenous channels with distance-dependent
attenuation (the attenuation factor is two). One additional
traditional scheme, namely \emph{orthogonal frequency partition}
(OFP), is also considered, which allocates each cell $1/3$ of the
total spectrum for orthogonal downlink transmission. It is assumed
that $N=256$, $K_m=12, m\in\Omega$, and the SNR shown in the figure
is the average SNR measured at the cell edge. It is observed that
the hybrid scheme can exploit the advantages of both traditional and
IA-based resource allocation schemes for low and high SNRs,
respectively, and thus achieve an overall more balanced performance.
It is also worth noting that by changing the relative sizes of
$\Phi$ and $\Phi'$, alternative designs of the hybrid scheme can be
flexibly obtained to maximize the throughput for a given SNR value.
\begin{figure}[!t]
\centering
\includegraphics[width=78mm]{Fig4.eps} \\
\caption{Cell and subcarrier partition for the hybrid scheme.}
\label{fig.4}\vspace{-6pt}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=80mm]{Fig5.eps} \\
\caption{System throughput of a three-cell system with heterogenous channel conditions.}
\label{fig.5}\vspace{-6pt}
\end{figure}
\section{Conclusion
This paper studies the downlink cooperative interference control in
cellular OFDMA systems. A new IA-based resource allocation scheme is
proposed, which jointly optimizes the frequency-domain precoding,
subcarrier user selection, and power allocation to maximize the
system throughput. For practical cellular systems with heterogenous
channel conditions, a hybrid scheme is proposed to exploit the
downlink IA gains.
|
\section{Introduction}
Acceleration of energetic particles is important for a wide range of
astrophysical environments, from stellar magnetospheres, accretion disk/jet
systems, supernova remnants and gamma ray bursts to clusters of galaxies.
Several mechanisms for particle acceleration have been discussed in the
literature which include varying magnetic fields in compact sources, stochastic
second order Fermi process in turbulent interstellar and intracluster media, and
the first order Fermi process behind shocks. An alternative, less explored
mechanism so far, involves particle acceleration in magnetic reconnection sites,
and this will be the focus of the present work which is a first of a series of
papers on this subject. For a comprehensive recent review on particle
acceleration mechanisms the reader is referred to \cite{melrose09}.
Magnetic reconnection may occur when two magnetic fluxes of opposite polarity
encounter each other. In the presence of finite magnetic resistivity, the
converging magnetic lines annihilate at the discontinuity surface and a current
sheet forms there. It is common knowledge that magnetic fields stay frozen in
highly conductive fluids. Estimates show that Ohmic diffusion for astrophysical
scales is absolutely negligible, contributing to the strongly entrenched view
that interacting magnetic fields of opposite polarity change their topology at
low speeds. This is determined by the magnetic reconnection speed, and the
famous example of two oppositely directed magnetic fluxes in Ohmic contact
undergoing slow Sweet-Parker \citep{sweet58,parker57} process of magnetic flux
annihilation is frequently invoked. In a more generic context, the magnetic
reconnection rate is the speed at which two magnetic flux tubes, which are
pushed against each other, can pass through altering the initial field topology.
This is a situation that can be frequently encountered in astrophysical fluids
with complex motions.
In the Sweet-Parker model, it has been shown that particles can accelerate due
to the induced electric field in the reconnection zone \citep{litvinenko03}.This
$one-shot$ acceleration process, however, is constrained by the narrow thickness
of the acceleration zone which has to be larger than the particle Larmor radius
and by the strength of the magnetic field. Therefore, the efficiency of this
process is rather limited. Besides, it also does not predict a power-law
spectrum, as generally observed for cosmic rays.
Observations have always been suggestive that magnetic reconnection can happen
at a high speed in some circumstances, in spite of the theoretical difficulties
in explaining it. For instance, the phenomenon of solar flares suggests that
magnetic reconnection should be first slow in order to ensure the accumulation
of magnetic flux and then suddenly become fast in order to explain the observed
fast release of energy. A model that can naturally explain this and other
observational manifestations of magnetic reconnection was proposed by
\cite{lazarian99}. The model appeals to the ubiquitous astrophysical turbulence
as a universal trigger and controller of fast reconnection. The predictions of
the model have been successfully tested in numerical simulations by
\cite{kowal09} which confirmed that this speed is of the order of the Alfv\'en
speed in the presence of weakly stochastic magnetic field fluctuations.
An important consequence of fast reconnection of turbulent magnetic
fields\footnote{This model of fast reconnection does not impose limits on the
amplitude of magnetic perturbations. They can be small and the magnetic flux
tubes require just to have a weak turbulent noise in order to produce fast
magnetic reconnection, i.e., the reconnection which does not depend on Ohmic
resistivity.} is the formation of a thick volume filled with reconnected small
magnetic flux loops. Now, if such turbulent flow is immersed in a current sheet
formed by two large scale converging magnetic flux tubes (such as, e.g., in the
Sweet-Parker configuration), then the three-dimensional magnetic fluctuations or
loops will contract and scatter test particles, presenting favorable conditions
for energetic particle acceleration in a first order Fermi process. In other
words, the particle may bounce back and forth between these converging magnetic
mirrors formed by oppositely directed magnetic fluxes moving towards each other
with the velocity corresponding to the reconnection speed. This has been first
described in \cite{degouveia05} \cite[see also][]{lazarian05} for the situation
when there is no back reaction of the accelerated particles on the reconnected
magnetic flux\footnote{This mechanism is in contrast to the second order Fermi
acceleration which is frequently discussed in terms of the particle acceleration
by turbulence generated by reconnection \citep{larosa06}, \cite[see also][in
prep.]{kowal11}.}. Later, \cite{drake06} appealed to a similar process, but
within a collisionless reconnection scenario. In their model, the contraction
of two-dimensional loops is controlled by the firehose instability that arises
in the particle-in-cell (PIC) simulations containing both electrons and ions
\citep{drake10}. In the present work, we show that this type of acceleration
can be present in a pure magnetohydrodynamical scenario as well, where the
pressure is fully isotropic and the contraction of islands is determined by
their interactions.
Magnetic reconnection is ubiquitous in astrophysical circumstances and therefore
it is expected to induce acceleration of particles in a wide range of
astrophysical environments including galactic and extragalactic ones. For
instance, the process has been already discussed for the production of
ultra-high energy cosmic rays \citep{degouveia00,degouveia01}, acceleration of
particles in gamma ray bursts \citep{lazarian03,zhang11}, microquasars
\citep{degouveia05} and astrophysical jet-accretion disks in general
\citep{degouveia10a,degouveia10b}. In particular, in the case of relativistic
jets, a diagram of the magnetic energy rate released by violent reconnection as
a function of the black hole (BH) mass spanning $10^9$ orders of magnitude was
derived and demonstrates that the magnetic reconnection power is more than
sufficient to explain the observed radio outbursts, from microquasars to low
luminous active galactic nuclei (AGN) \citep{degouveia10a}.
More recently, the acceleration in reconnection regions has obtained
observational support. It was suggested in \cite{lazarian09} that anomalous
cosmic rays measured by Voyagers are, in fact, accelerated in the reconnection
regions of the magnetopause \citep[see also][]{drake10}. Such a model explains
why Voyagers did not see any signatures of acceleration passing the Solar system
termination shock. In a separate development, \cite{lazarian10} have appealed
to the energetic particle acceleration in the wake produced as the Solar system
moves through interstellar gas to explain the excess of cosmic rays of the range
of both sub-TeV and multi-TeV energies in the direction of the magnetotail.
Note, that due to the 11-year solar magnetic cycle the accumulation of magnetic
reversals in the wake is unavoidable.
The implications of the acceleration process in reconnection sites are expected
to be even much wider. Numerical two-dimensional (2D) simulations recently
presented in \cite{drake10} confirmed the high efficiency of particle
acceleration in regions of magnetic reconnection. However, we will show here
that the process of acceleration happens rather differently in two and three
dimensional (3D) situations. The 3D geometry shows a wider variety of
acceleration regimes and this calls for much more detailed studies of the
acceleration.
In Section~\ref{sec:methods} we describe the methodology of the particle
acceleration studies presented here. In Section~\ref{sec:results} we show the
results obtained from these studies, in Sections~\ref{sec:discussion} and
\ref{sec:summary} we discuss the results and draw the main conclusions of this
work.
\section{Methodology}
\label{sec:methods}
In order to study particle acceleration in magnetic reconnection sites, we
performed numerical simulations solving the isothermal magnetohydrodynamic (MHD)
equations in two and three dimensions (2D and 3D, respectively). To compare our
MHD results with those obtained with the Particle in Cell (PIC) code in
\cite{drake10} we intentionally, in our 2D simulations, have reproduced their
set up of eight Harris current sheets in a periodic box. The initial density
profile is chosen in such a way that the total (gas plus magnetic) pressure is
uniform. Initially, we imposed random weak velocity fluctuations to this
environment in order to enable spontaneous reconnection events and the
development of magnetic islands. We evolved the system in 2D (both without and
with a guide field along the third dimension defined by Z direction) and in 3D
domains. After the development of the magnetic loops (which in a 2D geometry
form islands) in an underlying plasma configuration with defined density,
velocity and magnetic field profiles, we injected test particles at a given
snapshot and integrated their trajectories solving the equation of motion for
each charged particle
\begin{equation}
\frac{d}{d t} \left( \gamma m \vc{u} \right) = q \left( \vc{E} + \vc{u} \times
\vc{B} \right) , \label{eq:ptrajectory}
\end{equation}
where $m$, $q$ and $\vc{u}$ are the particle mass, electric charge and velocity,
respectively, $\vc{E}$ and $\vc{B}$ are the electric and magnetic fields,
respectively, $\gamma \equiv \left( 1 - u^2 / c^2 \right)^{-1/2}$ is the Lorentz
factor, and $c$ is the speed of light. In the studies below we assume that the
charged particles are protons.
\begin{figure*}[ht]
\center
\includegraphics[width=\textwidth]{figs/f1.eps}
\caption{Topology of the magnetic field represented as the gray texture with
semi-transparent color maps representing locations where the parallel and
perpendicular particle velocity components are accelerated for a 2D model with
$B_z = 0.0$ at time $6.0$ in the code units. The red and green colors
correspond to regions where either parallel or perpendicular acceleration
occurs, respectively, while the yellow color shows locations where both types of
acceleration occur. The parallel component increases in the contracting islands
and in the current sheets as well, while the perpendicular component increases
mostly in the regions between current sheets. White boxes show regions that are
more carefully analyzed in this paper. The simulation was performed with the
resolution 8192x4096. We injected 10,000 test particles in this snapshot with
the initial thermal distribution with a temperature corresponding to the sound
speed of the MHD model. \label{fig:locations}}
\end{figure*}
In the MHD simulations the electric field $\vc{E}$ is generated either by the
magnetized plasma or by resistive effects and can be obtained directly from the
Ohm's law equation, i.e.,
\begin{equation}
\vc{E} = - \vc{v} \times \vc{B} + \eta \vc{j} , \label{eq:efield}
\end{equation}
where $\vc{v}$ is the plasma velocity, $\vc{j} \equiv \nabla \times \vc{B}$ is
the current density, and $\eta$ is the Ohmic resistivity coefficient.
In our studies we are not interested in the acceleration by the electric field
resulting from resistivity effects, therefore we have neglected the last term in
Equation~(\ref{eq:efield}) in the trajectory integration. The charged particles
then feel an electric field $\vc{v} \times \vc{B}$ besides the magnetic field.
Substituting the Ohm's law, the equation of motion can be rewritten as
\begin{equation}
\frac{d}{d t} \left( \gamma m \vc{u} \right) = q \left[ \left( \vc{u} - \vc{v}
\right) \times \vc{B} \right] . \label{eq:trajectory}
\end{equation}
The particle equation of motion (Eq.~\ref{eq:trajectory}) was integrated using
the 6$^{th}$ order implicit Runge-Kutta-Gauss (RKG) method
\cite[see][e.g.]{sanz-serna94} with a fixed time step $dt=10^{-7}$. The RKG
methods are known to conserve the particle energy and momentum in very long
integrations \citep{sanz-serna94} in contrast to the standard 4$^{th}$ order
Runge-Kutta method with the adaptive time step based on the 5$^{th}$ order error
estimator \cite[see][e.g.]{press92}, which is very commonly used.
The interpolation of the local values of the plasma velocity $\vc{v}$ and
magnetic field $\vc{B}$ at each step of the integration has been done using
cubic interpolation \citep{lekien05} with our own discontinuity detector based
on a total variation diminishing (TVD) limiter. We performed also the
integration using linear interpolation which gave us essentially the same
statistical results, but with much faster integration times.
In the current study we do not include particle energy losses, therefore test
particles can gain or lose energy only through the interactions with the moving
magnetized plasma and its fluctuations. The inclusion of radiative and
non-radiative losses and also the back reaction of the particles on the plasma
is planned for future studies.
For simplicity, we assume the speed of light to be 20 times the Alfv\'en speed
$V_A$, which defines our plasma in a non-relativistic regime. The mean density
is assumed to be 1 atomic mass unit per cubic centimeter which is compatible
with the diffuse interstellar medium (ISM) density. The results are presented
in units normalized by the assumed light speed and time unit (which is 1 hour in
our simulations).
\section{Results}
\label{sec:results}
\subsection{Parallel and Perpendicular Acceleration in 2D Models}
\label{sec:multi-layer}
Figure~\ref{fig:locations} presents an evolved 2D configuration of the magnetic
field structure with magnetic islands. We clearly see the merging of islands in
some locations and the resulting deformations and/or contractions which provide
appropriate conditions for particle acceleration in a similar way to the results
obtained with the PIC code in \cite[][and references therein]{drake10}. In
general, we distinguish two kinds of accelerating zones where one or both
velocity components (parallel and perpendicular to the magnetic field) can
increase.
In Figure~\ref{fig:locations}, in order to demonstrate the zones which are
favorable for a particular type of acceleration, we superimposed on top of the
gray texture which represents the magnetic topology, a semitransparent color map
showing the locations where acceleration increases the parallel (red) or the
perpendicular (green) velocity component. In yellow areas, both types of
acceleration can occur depending on the speed and direction of the test particle
with respect to the orientation of the magnetic field lines.
We see that the increase of the parallel velocity component is mostly observed
within deformed islands and in current sheets (see the red and yellow zones in
Figure~\ref{fig:locations}), while the increase of the perpendicular component
is observed mostly near and within the islands and between current sheets (see
the green and yellow zones in Figure~\ref{fig:locations}). This complex
behavior is related to the degree of island deformation and the particle
direction and speed. If the island contracts, the particle can increase its
parallel component, however, when the island increases its deformation the
preferable acceleration usually takes place in the perpendicular direction.
Interestingly, we can find islands in which the acceleration does not happen at
all (see, e.g., the islands above and below the region R3 in
Figure~\ref{fig:locations}). This is due to the fact that these islands are not
undergoing contraction or deformation processes.
\subsection{Distribution of the Accelerated Particles}
\label{sec:acc_dist}
In Figure~\ref{fig:particle_histogram} we present the energy distribution of all
particles of the domain of Figure~\ref{fig:locations} at a time corresponding to
1 hour after the injection, when the particles are still accelerating. The red
and blue lines show the number of particles which increase their parallel and
perpendicular velocity components, respectively, as a function of the kinetic
energy. In addition, we show the initial thermal distribution of particles
(black line).
\begin{figure}[t]
\center
\includegraphics[width=0.5\textwidth]{figs/f2.eps}
\caption{Particle energy distributions in the full domain of
Figure~\ref{fig:locations}. Blue and red histograms show the number of
particles accelerating their perpendicular and parallel velocity components,
respectively, at one hour after the particles injection in the system. The
black line exhibits the initial thermal distribution of the injected particles.
\label{fig:particle_histogram}}
\end{figure}
In the particle energy distribution corresponding to the perpendicular
acceleration component (blue), we distinguish two modes, one in the low energy
range which is related to the initially injected thermal distribution with an
extended non-thermal tail, and the other corresponding to the high energy range.
For the particle energy distribution corresponding to the parallel acceleration
component, we see only one distribution mode with a maximum at high energies. In
the low energy range with energies smaller than the particle rest mass energy
(around $10^{-2}$), there is a clear dominance of the perpendicular
acceleration. This creates an anisotropy between the two velocity distributions
with respect to the magnetic field. However, once the particles become
relativistic and their kinetic energies become comparable to or larger than the
particle rest mass energy, the anisotropy disappears and the acceleration in
both directions is equally efficient.
Figure~\ref{fig:particle_histograms} presents the energy distributions of
accelerated particles in the selected regions R1 to R6, as shown in
Figure~\ref{fig:locations}. Regions R1 to R3 correspond to islands where both
parallel and perpendicular accelerations are observed, while regions R4 to R6
correspond to current sheets.
\begin{figure*}[ht]
\center
\includegraphics[width=0.32\textwidth]{figs/f3a.eps}
\includegraphics[width=0.32\textwidth]{figs/f3b.eps}
\includegraphics[width=0.32\textwidth]{figs/f3c.eps}
\includegraphics[width=0.32\textwidth]{figs/f3d.eps}
\includegraphics[width=0.32\textwidth]{figs/f3e.eps}
\includegraphics[width=0.32\textwidth]{figs/f3f.eps}
\caption{Particle energy distributions in the selected regions of
Figure~\ref{fig:locations}. Blue and red histograms correspond to the number of
particles which accelerated their perpendicular and parallel velocity
components, respectively. These histograms are created for all accelerating
particles passing through this region from the injection moment until 1 hour.
\label{fig:particle_histograms}}
\end{figure*}
The energy distributions for regions R1 to R3 are shown in the top panels of
Figure~\ref{fig:particle_histograms}. We notice a dominance of the
perpendicular acceleration in the low energy range, similarly as seen in
Figure~\ref{fig:particle_histogram}. As we move to higher energies we see a
trend to form a saturated tail. The efficiency of the acceleration of the
parallel component increases with the particle energy and at the highest
energies the number of particles accelerating in the parallel and perpendicular
directions is comparable.
In the bottom panels of Figure~\ref{fig:particle_histograms} we present the
particle energy distributions for regions within current sheets which are
labeled as R4, R5, and R6 in Figure~\ref{fig:locations}. In these cases, we
also observe an anisotropy between the perpendicular and parallel components in
the low energy range with a dominance of the perpendicular acceleration. As in
the case of the whole domain, this anisotropy decreases with increasing energy
and, beyond values around the particle rest mass energy both component
distributions become comparable. The high energy range tails of these
distributions are steeper than those of the top diagrams of the figure which
correspond to magnetic islands. This is especially well seen in the energy
distribution of region R6. This indicates that the number of accumulated
particles in the high energy range is smaller and therefore the acceleration,
although still significant, is less efficient in current sheets than within the
islands. This is apparently due to the fact that (at least in the 2D case) the
islands are able to retain the particles trapped for longer times. What could
be the dominant mechanism(s) for the acceleration in these zones?
The acceleration in the contracting/deforming islands is suggestive of first
order Fermi processes with the particles bouncing back and forth between
converging mirrors as described in \cite{degouveia05} and \cite{drake10}, while
within and between the current sheets the acceleration mechanism is not as
clear. In the following sections, we will examine the acceleration mechanism in
these different sites in more detail.
\subsection{Acceleration in Contracting Islands}
\label{sec:island}
In Figure~\ref{fig:event}, we show an example of a test proton which is trapped
in a contracting island and accelerates increasing its parallel component. The
left panel of the figure shows the topology of the magnetic field in a region
with a contracting island (R1 in Fig.~\ref{fig:locations}) with the trajectory
of the proton superimposed. As long as the proton remains trapped in the island
it orbits around the island center. In the middle panel we plot the evolution
of the parallel and perpendicular velocity components (red and green lines,
respectively), and the kinetic energy evolution (blue line) of the test
particle. In the right panel, we plot the change of the kinetic energy with the
particle X coordinate. While trapped in the island, the particle orbits around
the center and increases its energy after each cycle. Its parallel speed
increases while the gyro rotation slows down. This results in an exponential
growth of the kinetic energy of the particle (see middle panel). If we take a
closer look in the change of the kinetic energy with the position we see that
the increase of energy happens only when the particle moves across the right
part of the island. The left panel showing the magnetic field topology
indicates that the field lines in this part are contracting due to the
interaction with a small island which is merging with the central one. At the
same time, the particle gains more energy after each pass in this zone.
We clearly see how easily the results of \cite{drake10} can be reproduced using
the MHD approximation, confirming that the process of acceleration in the
islands is not restricted to the collisionless physics described by PIC codes.
MHD codes present an easier way to study the physics of particle acceleration
numerically.
\subsection{Acceleration Near and Within Current Sheets}
\label{sec:current_sheets}
We know from shock acceleration theory that particles are injected upstream and
allowed to convect into the shock, while diffusing in space so as to effect
multiple shock crossings, and thereby gain energy through the first order Fermi
process. Now, besides this mechanism, they may also experience shock drift
acceleration, which is attributed to the grad-B drift when the particle
encounters an abrupt change in magnetic fields. The origin of this effect is
due to the net work done on a charge by the Lorentz force (Eq.~\ref{eq:efield})
in a zone of non-uniform large scale magnetic field. The principal equation
governing this is the scalar product of the particle velocity (or momentum) and
the acceleration by the convective electric field, $-q \vc{v} \times \vc{B}$.
In uniform magnetic fields, the energy gain and loss acquired during a
gyroperiod exactly cancel, so in result no net work is done, $\Delta W = 0$.
In contrast, when the gyromotion of a charged particle straddles a
discontinuity, the sharp field gradient induces an asymmetry in the time spent
by the charge in either side of the discontinuity, so that energy gain and loss
do not compensate each other. The compressive nature of the field discontinuity
biases the net work done to positive increments in shock encounters between
upstream excursions, and it can be shown that $\Delta W = q E_x dx$, where the
$\vc{v} \times \vc{B}$ drifts lie in the x-direction, in other words, this
energy gain scales linearly with displacement along the drift coordinate x
(i.e., along the discontinuity). Thus the energy gained by a particle depends
on how far it drifts {\em along} the front \cite[see, e.g.,][and references
therein]{baring09,lugones11}.
\begin{figure*}[ht]
\center
\includegraphics[width=0.40\textwidth]{figs/f4a.eps}
\includegraphics[width=0.30\textwidth]{figs/f4b.eps}
\includegraphics[width=0.25\textwidth]{figs/f4c.eps}
\caption{The case of a contracting island where the particles accelerate
efficiently (region R1 of Fig.~\ref{fig:locations}). In the left panel we show
the trajectory of a test proton trapped in a contracting island. We see two
small magnetic islands on both sides of the central elongated island which are
merging with it. This process results in the contraction of the central island.
In the middle panel we show the evolution of the parallel and perpendicular
speeds of the test proton (red and green lines, respectively) and the evolution
of the particle energy (blue line). In the right panel it is shown the
change of the particle kinetic energy with the X coordinate. The proton
orbiting around the center of the magnetic island increases its energy increment
after each orbit. \label{fig:event}}
\end{figure*}
\begin{figure*}[ht]
\center
\includegraphics[width=0.40\textwidth]{figs/f5a.eps}
\includegraphics[width=0.30\textwidth]{figs/f5b.eps}
\includegraphics[width=0.25\textwidth]{figs/f5c.eps}
\caption{The case of acceleration near and within a single current sheet with a
Sweet-Parker configuration where the particles accelerate efficiently (like
region R6 of Fig.~\ref{fig:locations}). The left panel shows the trajectory of
a test proton approaching the diffusion region. The color of trajectory
corresponds to the particle energy (which increases from red to yellow and then
finally to white when the particle reaches the current sheet). The middle panel
shows the evolution of the parallel and perpendicular speeds of the test proton
(red and green lines, respectively) and the evolution of the particle energy
(blue line). The right panel shows the change of the particle kinetic energy
with the X coordinate. Outside the current sheet the proton drifts under the
effect of the magnetic field gradients and then when it arrives in the current
sheet it bounces back and forth between the converging magnetic fluxes of
opposite polarity while drifting along the magnetic field.
\label{fig:sp_acceleration}}
\end{figure*}
Similarly, in the current sheet zones (regions R4 to R6) we may be identifying
two distinct mechanisms of acceleration, either first order Fermi acceleration
in contracting/merging islands which are just forming there \cite[or even the
simple particle scattering between the converging flows entering both sides of
the current sheet, as described in][]{degouveia05}, or a drift acceleration, as
described above. The islands in current sheets are smaller than those of
regions R1, R2 and R3, what results in smaller acceleration rates as we see in
the lower panels of Figure~\ref{fig:particle_histograms}. In the zones above
and below the current sheets it is possible that we see predominantly a drift
acceleration driven by non-uniformities of the magnetic field. Generally, this
effect is less efficient than the first order Fermi process in
merging/contracting islands and results in smaller acceleration rates.
In order to better understand these distinct acceleration mechanisms, we have
also explored the details of the acceleration of a test particle near and within
a single (Sweet-Parker shaped) current sheet. Figure~\ref{fig:sp_acceleration}
shows the trajectory and energy evolution of this test particle. We see that
before the particle reaches the current sheet discontinuity it is drifted by the
plasma inflow and the increasing gradient of B as it approaches the current
sheet. When it enters the discontinuity (the white part of the trajectory in
the left panel), it bounces back and forth several times and gains energy (which
increases exponentially as shown in the middle panel of
Figure~\ref{fig:sp_acceleration}) due to head-on collisions with the converging
flow, on both sides of the magnetic discontinuity \cite[i.e., in a first order
Fermi process, as described in][]{degouveia05}. At the same time it drifts
along the magnetic lines which eventually allow it to escape from the
acceleration region. Therefore, we see two mechanisms: a drift acceleration
(dominating outside of the current sheet) and first order Fermi acceleration
inside the current sheet. These processes naturally depend on the initial
particle gyroradius, since it determines the amount of time the particle remains
in the acceleration zone before escaping.
Finally, we may also argue that turbulence above and below the current sheets in
Figure~\ref{fig:locations}, far from the islands and diffusion regions, favor
second order Fermi acceleration mechanisms with particles being scattered by
approaching and receding magnetic irregularities. Nevertheless, the first order
Fermi processes occurring within the islands and current sheets dominate the
overall particle acceleration in the system.
\subsection{The Role of a Guide Field: 2D vs. 3D simulations}
\label{sec:2d_vs_3d}
The results presented in the previous sections were obtained for 2D models
without a guide field. This means that in this case the magnetic lines creating
the islands are closed and a charged particle can be trapped indefinitely in
such an island. The presence of a guide field normal to the plane of
Figure~\ref{fig:locations} opens the magnetic loops and allows the charged
particles to travel freely in the out-of-plane direction. Moreover, the islands
evolve much slower in the presence of a strong guide field.
\begin{figure}[ht]
\center
\includegraphics[width=0.5\textwidth]{figs/f6a.eps}
\includegraphics[width=0.5\textwidth]{figs/f6b.eps}
\includegraphics[width=0.5\textwidth]{figs/f6c.eps}
\caption{Kinetic energy evolution of a group of 10$^4$ protons in 2D models of
reconnection with a guide field strength $B_z$=0.0 and 0.1 (top and middle
panels, respectively). In the bottom panel a fully 3D model with initial
$B_z$=0.0 is presented. The colors show how the parallel (red) and
perpendicular (blue) components of the particle velocities increase with time.
The energy is normalized by the rest proton mass energy. The background
magnetized flow with multiple current sheet layers is at time 4.0 in Alfv\'en
time units in all models. \label{fig:energy_2d_3d}}
\end{figure}
In Figure~\ref{fig:energy_2d_3d}, we present the time evolution of the kinetic
energy of the particles which have their parallel and perpendicular (red and
blue points, respectively) velocity components accelerated for three models of
reconnection. The upper panel shows the energy evolution for a 2D model without
the guide field (as in the models studied in the previous sections). Initially,
the particles pre-accelerate by increasing their perpendicular velocity
component only. However, later there is an exponential growth of energy mostly
due to the acceleration of the parallel component which stops after the energy
reaches values of 10$^3$--10$^4$~$m_p$ (where $m_p$ is the proton rest mass
energy). From that level on, particles accelerate their perpendicular component
only with smaller linear rate in a log-log diagram. The middle panel shows the
kinetic energy evolution in a 2D model with a weak guide field $B_z$=0.1 normal
to the plane of Figure~\ref{fig:locations}. In this case, there is also an
initial slow acceleration of the perpendicular component followed by the
exponential acceleration of the parallel velocity component. However, due to the
presence of a weak guide field, the parallel component accelerates further to
higher energies at a similar rate as the perpendicular one. This implies that
the presence of a guide field removes the restriction seen in the 2D model
without a guide field and allows the particles to increase their parallel
velocity components as they travel along the guide field, in open loops rather
than in confined 2D islands. This result is reassured by the 3D model in the
bottom panel of Figure~\ref{fig:energy_2d_3d}, where no guide field is necessary
as the MHD domain in fully three-dimensional. In this case, we clearly see a
continuous increase of both components, which suggests that the particle
acceleration behavior changes significantly when 3D effects are considered,
where open loops replace the closed 2D reconnecting islands.
Considering the parametrization we have chosen for our models, the gyroradius of
a proton becomes comparable to the size of the box when its Lorentz factor
reaches a value of a few times 10$^4$. The largest islands in the system can
have sizes of a few tenths of the size of the box. These rough estimates help
us to understand the energy evolution in Figure~\ref{fig:energy_2d_3d} and the
transition from an exponential to a much slower (linear) growth rate in the
energy around 10 hours. We note that in the diagrams of this figure, the energy
is normalized by the rest mass value, so that in fact, it is the Lorentz factor
that is plotted. In the case with absence of a guide field (top panel of
Fig.~\ref{fig:energy_2d_3d}), the exponential parallel acceleration stops right
before the energy value 10$^4$ is reached. After this, the rate of acceleration
significantly decreases. This occurs because the Larmor radius of the particles
has become larger than the sizes of biggest islands. Therefore, from this level
on the particles cannot be confined anymore within the islands and the first
order Fermi acceleration ceases. After that, there is a much slower drift
acceleration (of the perpendicular component only) caused by the gradients of
the large scale magnetic fields. If a guide field is inserted in such a system,
the picture is very similar, except for one detail. Now, since the particles
are able to travel along the guide field, their parallel velocity component also
continues to increase after the 10$^4$ threshold (see the middle panel of
Fig.~\ref{fig:energy_2d_3d}). Of course, in the 3D model, the particles follow
the same trend (bottom panel of Fig.~\ref{fig:energy_2d_3d}).
Figure~\ref{fig:energy_guide_field} exhibits the dependence of the acceleration
rate on the out-of-plane guide field strength. It shows the particle kinetic
energy distribution for three models of reconnection with different strengths of
the guide field ($B_z$=0.0, 0.5, and 1.0 for the top, middle, and bottom panels,
respectively) 1 hour after the injection. The initial injected thermal
distribution is represented by the blue line. The particles trajectory
integration was performed at the same MHD snapshot, at time 4.0 (in units of
Alfv\'en time), for all particles in these models. We note that at the final
state the particle distributions have developed high energy tails which depart
from the initial thermal distribution. However, both the number of accelerated
particles to higher energies and the maximum energy at the final time interval
strongly depends on the strength of the guide field.
\section{Discussion}
\label{sec:discussion}
In this paper we have investigated the process of particle acceleration in
reconnection zones considering test particles inserted in an isothermal
magnetohydrodynamic domain containing a series of reconnection layers, without
including kinetic effects. There were a few goals that motivated our approach.
The first goal was to study if the contraction of magnetic islands, which
develop in 2D reconnection layers with finite resistivity, can also occur when
the gas pressure is isotropic, as is the case in the MHD regime. In earlier
studies, Drake and collaborators \citep{drake10} invoked kinetic effects such as
the firehose and mirror instabilities to control the development and contraction
of the islands in current sheets. In the present work, after obtaining the
formation of contracting/deformed islands in current sheets in a nearly
incompressible MHD domain, the second goal was to investigate the acceleration
of test particles particularly in the direction parallel to the large scale
magnetic fields during island contractions. We were able to identify several
regions where such acceleration occurs both in a 2D environment with and without
a guide field and in a 3D environment. Also, we were able to identify the
nature of the acceleration mechanisms in the different acceleration regions, not
only within the contracting islands, but also near and inside the current
sheets. We found that first order Fermi processes occur in the islands and also
within the current sheets, while drift acceleration due to magnetic field
gradients seems to be dominating outside the current sheets.
\begin{figure}[ht]
\center
\includegraphics[width=0.5\textwidth]{figs/f7a.eps}
\includegraphics[width=0.5\textwidth]{figs/f7b.eps}
\includegraphics[width=0.5\textwidth]{figs/f7c.eps}
\caption{Kinetic energy distribution of the particles for three 2D models with
different strengths of the out-of-plane guide field $B_z$ = 0.0, 0.5, and 1.0
(top, middle and bottom panels, respectively.) at a time corresponding to 1 hour
after the injection. The initial particle thermal distribution is given by the
blue line. \label{fig:energy_guide_field}}
\end{figure}
Our results for particle acceleration in 2D MHD models were quite similar to
those obtained by \cite{drake10}, i.e., during an island contraction, in our
case resulting from the merge with other islands, a particle trapped in it can
accelerate and increase its energy exponentially and its parallel velocity
component grows while the perpendicular one undergoes a net decrease. During
this process the energy gain is proportional to the particle energy. These are
all necessary conditions in a Fermi like process and a close examination of the
motion of a test particle in a contracting island shows clear evidence of the
first order Fermi acceleration process.
Similarly, exponential acceleration of the parallel velocity component was also
detected in regions within the current sheets. A detailed examination of a test
particle trajectory within and near a current sheet revealed that outside the
discontinuity the particle experiences drift acceleration due to magnetic field
gradients and within the current sheet it goes back and forth between the
converging flows on both sides of the discontinuity also gaining energy
exponentially, just as described in \cite{degouveia05}, while drifting along the
magnetic field direction.
\cite{drake10} have claimed that particle acceleration in an X-point
reconnection in a 3D environment, even in the presence of a guide field, should
behave like in 2D systems. However, we have shown that in the presence of an
out-of-plane guide field, this picture changes completely \cite[see
also][]{drake10}. While in 2D MHD models without a guide field the parallel
acceleration stops after reaching a certain energy threshold, in 2D MHD models
with a guide field this constraint is removed, and particles can continue
increasing their parallel speed as they travel along the guide field.
Furthermore, in fully 3D MHD models with no guide field the acceleration of the
particles exhibits the same trend as in 2D models with a guide field, i.e.,
there is no constraint on the acceleration of the parallel speed. This result
is very important as it implies that the overall picture of particle
acceleration in 3D reconnection can be very distinct from that in 2D
reconnection. Also, it can offer some important clues for particle acceleration
in more complex domains, such as in the presence of 3D turbulence where
stochastic reconnection is involved \cite[see][in prep.]{kowal11}.
Another important result of the present work was to demonstrate that the
investigated acceleration mechanisms in reconnection sites (both drift and Fermi
processes) can be present in any environment whose evolution can be approximated
by a nearly ideal MHD description as employed here, therefore, without the
necessity of invoking kinetic instabilities or anomalous resistivity effects to
control the pressure anisotropy or the reconnection rate. The nearly
non-resistive MHD approach with test particle injection offers a possibility of
exploring more realistic systems. Even though PIC codes in general allow the
study of plasma processes in greater detail, they present some disadvantages
with regard to an MHD description, such as the necessity of a much higher
complexity in the 3D modeling, for instance, of turbulence.
A final remark is in order. \cite{onofri06} investigated particle acceleration
in reconnection zones and concluded that MHD should not be a good approximation
to describe the whole process of acceleration. However, their 3D numerical
simulations were performed in a fully resistive MHD regime. Therefore, they
obtained very efficient particle acceleration due to the high electric field
induced by resistivity and an absorption of most of the available magnetic
energy by the electrons in a very small fraction of the characteristic time of
the MHD simulation. This led them to conclude that resistive MHD codes are
unable to represent the full extent of particle acceleration in 3D reconnection.
In the present work, we have investigated particle acceleration in a nearly
ideal MHD regime where only numerical resistivity is present. In this case, the
contribution of a resistivity induced electric field is negligible when compared
to the advection component, namely, the electric field resulting from the plasma
motion in the magnetized medium, $\vc{E} = - q \vc{v} \times \vc{B}$. Moreover,
since there is no important dissipation effects, our model is scale independent
and allows for dimensionless units, so that it is straightforward to rescale the
configuration to reproduce different astrophysical environments. The main
drawback of the numerical resistivity is the creation of a ''hole'' in the
center of the magnetic islands, since the numerical modeling cannot handle
properly the strongly curved magnetic lines over a few cells. As a consequence,
numerical dissipation removes part of the magnetic flux in this region.
The importance of the present study of acceleration in the MHD regime is
motivated by the fact that magnetic reconnection in 3D becomes fast according to
the model in \cite{lazarian99} \cite[see also][]{lazarian04}. This model has
been successfully tested via numerical simulations in \cite{kowal09} which
confirmed that thick reconnection layers form, where magnetic energy is
transferred into energy of contracting loops \footnote{These thick layers were
also confirmed by \cite{ciaravella08}.}. In the present paper, we have
confirmed that these loops (or magnetic irregularities) can act as the places of
efficient particle acceleration providing the support for the mechanism first
discussed in \cite{degouveia05}. At the same time, we see that the process of
acceleration does not amount to only the process of first order Fermi
acceleration outlined in the aforementioned paper. Our simulations show a
complex interplay of different acceleration processes, which motivates for
further studies of the acceleration in the presence of realistic turbulent
reconnection \cite[see][]{kowal09}.
The present study has clearly an exploratory character. It testifies that the
acceleration in reconnection regions may be more complex than it may be inferred
from earlier simplified 2D studies. Further study of the acceleration in the
MHD regime is absolutely essential, as ubiquitous astrophysical turbulence is
expected to induce magnetic reconnection all over astrophysical volumes and
should be explored in more detail. In this situation the acceleration of
particles by reconnection may play a vital role, the extend of importance of
which can be evaluated from further research \citep{kowal11}.
\section{Summary}
\label{sec:summary}
The results of the paper can be very briefly summarized as follows:
\begin{itemize}
\item Advances in the understanding of magnetic reconnection in the MHD regime,
in particular, turbulent magnetic reconnection in \cite{lazarian99} model
motivate the studies of whether the reconnection in this regime can accelerate
energetic particles.
\item Contracting magnetic loops in magnetic reconnection in 2D, in the MHD
regime, provides the acceleration which successfully reproduces the results
obtained earlier with more complicated PIC codes, which proves that the
acceleration in reconnection regions is a universal process which is not
determined by the details of plasma physics.
\item Acceleration of energetic particles in 2D and 3D shows substantial
differences, which call for focusing on realistic 3D geometries of reconnection.
Our study also shows that apart from the first order Fermi acceleration,
additional acceleration processes interfere.
\end{itemize}
\acknowledgements
GK and EMGDP acknowledge the support by the FAPESP grants no. 2006/50654-3 and
2009/50053-8, and the CNPq grant no. 300083/94-7, and AL thanks the NSF grant
AST 0808118, NASA grant NNX09AH78G and the support of the Center for Magnetic
Self Organization. This research was also supported by the project TG-AST080005N
through TeraGrid resources provided by Texas Advanced Computing Center
(TACC:http://www.tacc.utexas.edu). Part of the computations presented here were
performed on the GALERA supercomputer in the Academic Computer Centre in
Gda\'nsk (TASK:http://www.task.gda.pl/).
|
\section{Introduction}
The idea that a perturbatively nonrenormalizable
theory could be consistently defined in the UV limit at a nontrivial fixed point (FP),
often called ``asymptotic safety'', is theoretically very attractive,
especially when applied to quantum gravity \cite{weinberg}.
Much progress in this direction has come from the direct
application of Renormalization Group (RG) techniques to gravity.
A particularly useful tool has been the non-perturbative
Functional Renormalization Group Equation (FRGE) \cite{wetterich},
defining an RG flow on a theory space which consists of all diffeomorphism invariant
functionals of the metric $g_{\mu\nu}$.
It defines a one parameter family of effective field theories with actions
$\Gamma_k(g_{\mu\nu})$ depending on a coarse graining scale (or ``cutoff'') $k$,
and interpolating between a ``bare'' action (for $k\to\infty$)
and the ordinary effective action (for $k\to 0$).
The application of such concepts to a cosmological framework is based to the following logic \cite{rw1}.
If we want to study the quantum evolution of the cosmic scale factor,
we should in principle use the full effective action $\Gamma(g_{\mu\nu})$,
which, as we mentioned above,
coincides with $\Gamma_k(g_{\mu\nu})$ in the limit $k\to 0$.
However, our knowledge of this functional is rather poor.
One way of gaining some traction on this issue is to observe
that the Hubble parameter appears as a mass in propagators.
Thus, the contributions of quantum fluctuations with wavelenghts
greater than $H^{-1}$ are suppressed.
As a result, the functional $\Gamma_k$ at $k$ comparable
to $H$ should be a reasonable approximation for the same functional at $k=0$.
We do not know $\Gamma_k$ much better than the full effective action,
but we can easily calculate the dependence of some terms in $\Gamma_k$ on $k$.
By doing so we effectively take into account nonlocal terms
that would be very hard to calculate otherwise.
Previous investigations along these lines, in particular \cite{bore},
have been based on the Einstein-Hilbert (EH) truncation.
It is important to establish that the results obtained there
persist when further operators are included in the truncation.
We know that at the FP the coefficients of these terms are not
very small, but their presence does not seem to affect the values
of the cosmological constant and Newton's constant too much.
In other words, the FP that is is seen in the EH truncation
seems to be robust.
The question then is to see if this stability of the FP against
the inclusion of new terms
is reflected in the stability of the corresponding solutions.
This question is important because
the values of the couplings at the FP are fixed and as a consequence
there are no free parameters to be varied.
We will see that the (power-law or exponential) inflationary solutions
are indeed stable against the inclusion of new terms, but establishing
this fact requires including a rather large number of terms.
\section{Asymptotic Safety}
It is well known that general relativity can be treated as an effective field theory. This means that it is possible to compute quantum effects due to graviton loops, as long as the momenta of the particles in the loops are cut off at some scale. The results are independent of the structure of any “ultraviolet completion”, and therefore constitute genuine low energy predictions of any quantum theory of gravity.
Given for a field $\phi$ an effective action $\Gamma(\phi)$, it can always be written as a sum
\begin{equation}\label{infaction}
\Gamma(\phi)=\sum g_i{\cal O}_i(\phi)
\end{equation}
that runs over the infinitely many operators consistent with the undelying symmetries of the theory. Given an energy scale $k$, one can define \emph{dimensionless} couplings $\tilde{g}_i=g_i k^{-d_i}$, where $[g_i]=[k]^{d_i}$. The resulting picture is an infinite dimensional space parametrized by the $\tilde{g}_i$'s, where a Renormalization Group flow can be seen as a vector field that shifts any action along a trajectory
\begin{equation}
\Gamma(\phi)\rightarrow\Gamma_k(\phi)=\sum\tilde{g}_i(k)k^{d_i}{\cal O}_i(\phi)\;.
\end{equation}
The particular realization of RG that was used in this work is the one $\grave{a}\,la$ Wilson, based on the idea that the effective action describing physical phenomena at a momentum scale $k$ can be thought of as the result of having integrated out all fluctuations of the field with momenta larger than $k$. This can be achieved adding to the bare action ${\cal S}$ a term
\begin{equation}
\Delta{\cal S}_k=\int dx\,\phi(x)R_k\,\phi(x)
\end{equation}
where $R_k$ is called the cutoff kernel (see \cite{cpr1} for details). This term acts as a mass for the fluctuations with momenta smaller than $k$. Following the usual procedure of background field method, we define a $k$-dependent generating functional of connected Green functions by
\begin{equation}
e^{-W_k(J)}=\int D\phi e^{-(S(\phi)+\Delta{\cal S}_k(\phi))-\int dx J\phi}
\end{equation}
and the effective action as its Legendre transform
\begin{equation}
\Gamma_k(\phi)=W_k(J)-\int dx J\phi-\Delta{\cal S}_k(\phi)\;.
\end{equation}
One can show that it obeys the so called functional renormalization group equation (FRGE)
\begin{equation}\label{frge}
k\frac{d\Gamma_k}{dk}=\frac{1}{2}\textrm{Tr}\left(\Gamma_k^{(2)}+R_k\right)^{-1}k\frac{dR_k}{dk}
\end{equation}
where $\Gamma_k^{(2)}$ is the inverse propagator of the graviton. From the r.h.s. of (\ref{frge}) one can then extract the $\beta$-functions of the couplings, as it is done in \cite{cpr1}, and study the RG flow of the theory towards the UV limit $k\rightarrow\infty$.
\pagebreak
A theory is then said to be asymptotically safe if
\begin{itemize}
\item for any coupling $\tilde{g}_i$ there exists a finite UV limit $\tilde{g}_i^\ast$ that is called a fixed point
\item the stability matrix $M_{ij}=\left.\frac{\partial\beta_{\tilde{g}_i}}{\partial\tilde{g}_j}\right|_\ast$ evaluated in the fixed point has a finite number of negative eigenvalues $m_i$ (in the following, it will come useful to define critical exponents $\theta_i=-m_i$).
\end{itemize}
If both the requests are satisfied one finds that the \emph{critical surface}, the locus of the points that are attracted towards the FP, is finite dimensional. This means that only a finite number of experiments is needed to univocally identify a trajectory, and the theory is thereby predictive.
In \cite{cpr1} this procedure was applied to a (truncated) gravitational action consisting of a sum of powers of the curvature scalar $R$
\begin{equation}\label{gravaction}
\Gamma(g)=\int d^4x\sqrt{|g|}\,\sum_{i=0}^n g_iR^i
\end{equation}
with $n=2,\dots,8$, and it was further extended to $n=9,10$ in \cite{bcp}. For every $n$ a fixed point was found, with only three positive critical exponents. To understand if a finite truncation is a reliable approximation of the action (\ref{infaction}), one can study the stability of the position of the fixed point {$\tilde{g}_i^\ast$} and of the values of the critical exponents $\theta_i$: these values should stay almost constant with respect to the inclusion of new operators in the action. In figure (\ref{FPthetas}) it is clearly visible that this seems to be the case, so that our finite sum is a reliable truncation.
\begin{center}
\begin{figure}[ht]\label{FPthetas}
\includegraphics[width=6.5cm,height=10cm]{position.jpg}
\includegraphics[width=6.5cm,height=10cm]{critical.jpg}
\caption{Position of the fixed point and values of the critical exponents as functions of the truncation order $n$.}
\end{figure}
\end{center}
\section{Asymptotically safe cosmology}
It is now time to apply the above described formalism to a cosmological background. There are two main issues that are worth stressing, the first concerning the possibility to obtain any inflationary solution, and the second about the resulting stability of such solutions with respect to the inclusion of new terms in the action. We will address them after having defined a key ingedient of our analysis: the identidication of the RG scale $k$ with a characteristic scale of the described system.
As it was said before, in the Wilsonian approach the cutoff scale is the momentum scale at which the process is taking place. This means that, in a Friedmann-Robertson-Walker metric, it should be set to the order of magnitude of the Hubble rate $H(t)$. Another argument can be found in the comparison with the framework of quantum field theory in curved spaces, where the Hubble rate appears as a mass term in the propagator of the field. We then set $k=\xi H$, where $\xi={\cal O}(1)$ is a free parameter that in principle can be tuned to offset the scheme dependence carried by the cutoff scale.
\subsection{Inflationary cosmology}
We have now all the ingredients to extract the desired inflationary solutions. From the action
\begin{equation}
\Gamma(g)=\frac{1}{16\pi G}\int d^4x\sqrt{|g|}\,(f(R)-\Lambda)
\end{equation}
where $f(R)=\sum f_i R^i$, we derive the modified Einstein equations that, after having imposed the FRW symmetry, appear as (modified) Friedmann equations
\begin{eqnarray}
{\cal A}(H)&=&8\pi G\rho+\Lambda\\
{\cal B}(H)&=&8\pi G\rho(1-3w)+4\Lambda
\end{eqnarray}
where $w$ si the equation of state of the ordinary matter and ${\cal A}(H)$, ${\cal B}(H)$ are two functions of the Hubble rate whose shape depends on the chosen $f(R)$, and can be rearranged in the form
\begin{eqnarray}\label{fried2}
{\cal B}(H)&=&(1-3w){\cal A}(H)+3(1+w)\Lambda\\
\rho&=&\frac{1}{8\pi G}\left({\cal A}(H)-\Lambda\right)\;.
\end{eqnarray}
Before solving the equations, we have to perform the above mentioned cutoff identification: the couplings are rewritten as
\begin{equation}
\Lambda=\tilde\Lambda k^2\ ;\quad
G=\tilde G/k^2\ ;\quad
f_i=\tilde f_i k^{2-2i}
\end{equation}
and, because we are interested in describing the dynamics of the very early universe, are assumed to lie in the UV fixed point. With these substitutions, we look for solutions of the (\ref{fried2}) in the form of a de Sitter ($H(t)=\bar H$ and $a(t)\propto e^{\bar Ht})$) and a power law ($H(t)=p/t$ and $a(t)\propto t^p$). It comes out that the system admits the former only for discrete values $\xi_{dS}$ of the free parameter, while we found continuous solutions of the latter type, with $p=p(\tilde\Lambda^\ast,\tilde G^\ast,\tilde f_i^\ast,w,\xi)$. As a final remark, the functions $p(\xi)$ show vertical asymptotes, located exactly at the values $\xi_{dS}$: de Sitter is then recovered as a limit of power law.
In figure (\ref{neq8}) the power law exponents $p$ obtained for $n=8$ are plotted as functions of $\xi$. There is in particular one growing solution that is greater than 1 (\emph{i.e.} $\ddot{a}/a>0$) for most values of the free parameter.
\begin{center}
\begin{figure}[ht]\label{neq8}
\includegraphics[width=8cm,height=6cm]{ext8.pdf}
\caption{Power law exponents as functions of the parameter $\xi$. The asymptote is clearly visible for $\xi\simeq3.9$.}
\end{figure}
\end{center}
\subsection{Reliability analysis}
As it was said before, the UV fixed point of the gravitational action (\ref{gravaction}) tends to stabilize as one widens the truncation. The issue is then whether the values of the solutions will show the same behaviour. We thus define a reliability criterion based on the following logic: taken the dimensionless curvature scalar $\tilde R=Rk^{-2}$, we write the dimensionless action with truncation $n$ evaluated at the fixed point $\left.\tilde f(\tilde R)\right|_n$ and we compare it with the same object in $n-1$. A truncation is then said to be reliable if
\begin{equation}
\left|\frac{\left.\tilde f(\tilde R)\right|_n-\left.\tilde f(\tilde R)\right|_{n-1}}{\left.\tilde f(\tilde R)\right|_{n-1}}\right|<5\%
\end{equation}
and in general this is only true for a limited range of values of $\tilde R$. One can schematically rewrite this as $\tilde R<c$, where
\begin{center}
\begin{tabular}{|c|r|r|r|r|r|r|r|r|r|r|}
\hline
$n$ & 1& 2& 3& 4& 5& 6& 7& 8& 9& 10\\
\hline
$c$ & 0.40& 0.24& 0.45& 1.07& 1.21& 0.90& 0.79& 0.92& 1.09& 1.09\\
\hline
\end{tabular}
\end{center}
The table shows that the reliable range tends to increase, so that a larger truncation is more likely to correctly approximate the infinite truncation $n\to\infty$.
For a power law solution, the constraint on the dimensionless curvature scalar can be translated into a constraint on the values of the exponents, namely
\begin{equation}
c\xi^2\gtrsim6\left(2+\dot{H}/H^2\right)=6\left(2-1/p\right)
\end{equation}
that can be plotted together with the exponents themselves, as it is shown in figure (\ref{rel}) on the left. From the plot it comes clear that only a small portion of the growing solution is to be considered reliable, so we decided to extend the truncation to $n=10$. The result is plotted in figure (\ref{rel}) on the right, that shows a great improvement in the reliability of the solutions.
\begin{center}
\begin{figure}[ht]\label{rel}
\includegraphics[width=7cm,height=5.25cm]{sol8ext.jpg}
\includegraphics[width=7cm,height=5.25cm]{sol10ext.jpg}
\caption{Power law exponents as functions of $\xi$ for $n=8$ (left) and $n=10$ (right). Areas to the right of the dashed lines are the reliability regions, while the dotted line indicates the $n=1$ solution described in \cite{bore}.}
\end{figure}
\end{center}
\section{Conclusions}
The results obtained here can be summarized saying that, if we identify the cutoff with a multiple of the Hubble parameter, inflationary power law solutions (i.e. with exponent $p>1$) exist for some range of values of $\xi$. The dependence on such parameter is strong, with the exponents diverging at some values $\xi_{dS}$. Exactly at those points, the theory admits also de Sitter solutions. Provided that $\xi$ lies in the above range, that the starting point is close enough to the FP and that $p>1$, it should always be possible to have a sufficient number of $e$-foldings.
Moreover, it has been shown that the increasing stability of the fixed point with respect to the widening of the truncation seems to be reflected in an increasing reliability of the cosmological solutions obtained using the truncated action. The overall conclusion that one can draw is that we found a strong hint of the convergence of the series in our fixed point action, so that any result coming from low order truncations should not be too misleading.
\section*{References}
|
\section*{Introduction} \sk
Let $A$ be a local ring with the maximal ideal $\mm$ and $M$ be a finitely generated $A$-module. Denote by $G_I(M) = \oplus_{n\geq 0}I^nM/I^{n+1}M $ the associated graded module of $M$ with respect to an $\mm$-primary ideal $I$. It is well-known that one can use the Castelnuovo-Mumford regularity $\reg(G_I(M))$ of $G_I(M)$ to bound other invariants of $M$ such as Hilbert coefficients, the postulation number, ... (see, e.g., \cite{T}, \cite{Va}, \cite{RTV}). Therefore, bounding $\reg(G_I(M))$ is an important problem. This problem was completely solved by Rossi, Trung and Valla for the case $M=A$ and $I= \mm$ in \cite{RTV} and by Linh for the general case \cite{L1}, where bounds on $\reg(G_I(M))$ were given in terms of the dimension $d$ of $M$ and of the so-called extended degree $D(I,M)$ (see Definition \ref{A1}).
Another important subject associated to $I$ is the so-called fiber cone $F_\mm(I) = \oplus_{n\geq 0}I^n/\mm I^n$. Inspired by the mentioned results of \cite{RTV} and \cite{L1}, it is natural to ask whether one can bound the Castelnuovo-Mumford regularity $\reg(F_\mm(I))$ of $F_\mm(I)$ in terms of $D(I,A)$ and $d$? One of the main ideas of \cite{RTV} and \cite{L1} is to reduce the problem of bounding the Castelnuovo-Mumford regularity $\reg(G_I(M))$ in the case $\depth M>0$ to bounding the so-called geometric Castelnuovo-Mumford regularity of $G_I(M)$. However this approach does not work for fiber cones, because a similar result to \cite[Theorem 5.2]{H} does not hold in this case. Fortunately, thank to a recent work by Rossi and Valla \cite{RV}, we can use the associated graded modules of some good filtrations of $M$ to solve the problem.
The notion of good $I$-filtrations $\FilM = \{ M_n\}_{n\ge 0}$ of $M$ was already considered in \cite{Bour} and \cite{AM} (see Definition \ref{Fil-Def}). Even in the case $M=A$ the class good filtrations of submodules is larger than the usual class of good filtrations of ideals. Recently, Rossi and Valla showed in \cite{RV} that one can not only extend many classical results to a filtered module (i.e. a module with a good filtration), but also use filtered modules to study the Hilbert function of fiber cones. Our work here once more shows the usefulness of this concept in studying fiber cones.
In order to bound $\reg(F_\mm(I))$ we first extend results of \cite{RTV,L1} to the associated graded module $G(\FilM) = \oplus_{n\ge 0}M_n/M_{n+1}$ of $\FilM$. The techniques we use are similar to that in \cite{RTV,L1}. The new point here is that we also use the maximal generating degree $r(\FilM)$ of the graded module $G(\FilM)$ over the standard graded ring $G_I(A)$ (see Definition \ref{A2}) and clarify its influence to the Castelnuovo-Mumford regularity of associated graded modules by a hyperplane section or by passing to $\bar{M} = M/H^0_\mm(M)$. As a consequence, we see that the bounds in this general case (see Theorem \ref{A3}) look similar to that in the classical case in \cite{L1}. The bounds are given in terms of $r(\FilM), \ d$ and in terms of $D(I,M)$, which does not depend on the filtration $\FilM$. Then using an exact sequence connecting fiber cones and associated graded modules of some good filtrations and a relation between Hilbert coefficients of these filtrations given in \cite{RV} we can also bound the Castelnuovo-Mumford regularity of a fiber cone associated to any good filtration in terms of $d,\ D(I,A)$ and $r(\FilM)$ (see Theorem \ref{F3}).
In this paper we also consider the graded case. For this situation we can apply the approach of \cite{RTV} and \cite{L1} only in the case $I$ being generated by homogeneous elements of the same degree. A new point here is that we can give a bound on $\reg(G(\FilM))$ in terms of $\reg(M)$ (see Theorem \ref{C3}). Note that $\reg(M)$ is in general not only much smaller than an arbitrary extended degree $D(I,M)$ of $M$ (see \cite{DGV, Va, Na}), but it is also much easier to compute. In the general case the above mentioned approach is not applicable, because $I$ does not contain a generic homogeneous element. To overcome the difficulty we pass to the localization and first give a bound in terms of the so-called homological degree (Theorem \ref{B3}). Combining with a recent result of \cite{CHH} we can then get a bound in terms of $\reg(M)$ (see Theorem \ref{B5}), provided $A$ is a polynomial ring over a field. However the bound in this general setting is much worse than the one in Theorem \ref{C3}.
The paper is divided into three sections. In Section \ref{Local} we bound the Castelnuovo-Mumford regularity of a filtered module in the local case. The graded case is considered in Section \ref{Graded}. In the last section \ref{Fiber} we derive bounds on the Castelnuovo-Mumford regularity of the fiber cone of a filtered module.
\section{Regularity of associated graded rings: Local case} \label{Local}
Let $A$ be a Noetherian local ring with an infinite residue field $K:= A/\mm$ and $M$ a finitely generated $A$-module. (Although the assumption $K$ being infinite is not essential, because we can tensor $A$ with $K(t)$.) First we recall some basic facts on filtered modules from \cite[Section III.3]{Bour}, \cite[Chapter 10]{AM} and \cite[Section 1]{RV}.
\begin{Definition} \label{Fil-Def} {\rm Given a proper ideal $I$. A chain of submodules
$$\FilM:\ M = M_0 \supseteq M_1 \supseteq M_2 \supseteq \cdots \supseteq M_n \supseteq \cdots $$
is called an {\it $I$-filtration} of $M$ if $IM_i \subseteq M_{i+1}$ for all $i$, and a {\it good $I$-filtration} if $IM_i = M_{i+1}$ for all sufficiently large $i$. A module $M$ with a filtration is called a} filtered module.
\end{Definition}
Thus $\{I^nM\}$ is a good $I$-filtration. Note that the notion of $I$-filtration of submodules is different from that of filtration of ideals. The sequence $A \supset \mm \supset \mm I \supset \mm I^2 \supset \cdots $ is a good $I$-filtration of submodules of $A$, but it is in general not a filtration of ideals in $A$, because $\mm^2 \not\subseteq \mm I$.
If $N$ is a submodule of $M$, then by Artin-Rees Lemma, the sequence $\{N\cap M_n\}$ is a good $I$-filtration of $N$ and we will denote it by $\FilM \cap N$. The sequence $\{M_n+N/N\}$ is a good $I$-filtration of $M/N$ and will be denoted by $\FilM/N$.
In this paper we always assume that $I$ is an $\mm$-primary ideal and $\FilM$ is a good $I$-filtration. The {\it associated graded module} to the filtration $\FilM$ is defined by
$$G(\FilM) = \bigoplus _{n\geq 0}M_n/M_{n+1} .$$
We also say that $G(\FilM)$ is the associated ring of the filtered module $M$. This is a finitely generated graded module over the standard graded ring $G:= G_I(A) := \oplus_{n\ge 0}I^n/I^{n+1}$ (see \cite[Proposition III.3.3]{Bour}).
\begin{Definition}\label{A2}{\rm Let $\FilM$ be a good $I$-filtration of $M$. We set
$$r = r(\FilM) = \min \{t \geq 0 \mid M_{n+1} = IM_n \ \ \text{for all} \ \ n \geq t \}.$$}
\end{Definition}
In particular, in the $I$-adic case, $r(\{I^nM\}) = 0$. Note that $r$ is always finite, and $M_{r+j} = I^jM_r$ for all $j\ge 0$. This means $\{ M_n\}_{n\ge r}$ is of form of an $I$-adic filtration of $M_r$. In other words, $r$ is the largest generating degree of $G(\FilM)$ as a graded module over $G$. Together with well-known facts on the Castelnuovo-Mumford regularity (see, e.g. \cite[Paragraph 1.3]{BS}) this explains why $r$ will naturally occur in our bounds on $\reg(G(\FilM))$, which is recalled below.
The {\it Castelnuovo-Mumford regularity} of a finitely generated graded module $E = \oplus _{n \in \ZZ} E_n$ over a standard $\NN$-graded ring $R = \oplus _{n\geq 0} R_n$ is defined as the number
$$\reg(E) = \max \{ a_i(E) + i \mid i \geq 0 \},$$
where
$$ a_i(E) =
\begin{cases}
\sup \{n|\ H_{R_+}^i(E)_n \ne 0 \} \ \ \ \ \text{if} \ \ H_{R_+}^i(E) \ne 0 ,\\
-\infty \ \hskip 3.6cm \text{if} \ \ H_{R_+}^i(E) = 0,
\end{cases}
$$
and $R_+ = \oplus _{n> 0} R_n.$ We will write $\reg(G(\FilM))$ to mean the Castelnuovo-Mumford regularity of $G(\FilM)$ being a graded module over the standard graded ring $G$.
An element $x\in R_1$ is called (linear) {\it filter-regular } on $E$ if $(0:_Ex)_n = 0$ for all $n\gg 0$. From this one can show that $0:_Ex \subseteq H^0_{R_+}(E)$, and hence $(0:_Ex)_n =0$ for all $n> \reg(E)$.
Each element $a\in A$ has a natural image, denoted by $a^*$, in $G_I(A)$. Thus, if $x\in I\setminus I^2$, then $x^*$ is a filter-regular element on $G(\FilM)$ if and only if $(M_{n+2}: x)\cap M_n = M_{n+1}$ for all $n> \reg(G(\FilM))$.
\begin{Lemma} \label{Fil-Property}
Let $x\in I\setminus \mm I$ be an element such that $x^*$ is a filter-regular on $G(\FilM)$ and let $a= \reg (G(\FilM))$. Then
\begin{itemize}
\item[(i)] $r(\FilM) \le a$.
\item[(ii)] $xM \cap M_n = xM_{n-1}$ for $n \ge a+1$.
\item[(iii)] $M_{n+1} : x = M_n + (0_M:x)$ and $(0_M:x) \cap M_{n+1} = 0$ for $n\ge a$.
\end{itemize}
\end{Lemma}
\begin{pf} As mentioned above, $r(\FilM)$ is the largest generating degree of $G(\FilM)$. Hence (i) is a well-known fact. The statements (ii) and (iii) are just filtration version of \cite[Lemma 4.4]{T2} and \cite[Proposition 4.6]{T2}, respectively. Eventually, (ii) and (iii) characterize the filter-regular property of $x^*$, see \cite[Theorem 1.5]{RV}.
\end{pf}
We are interested in bounding $\reg(G(\FilM))$. In the case of $I$-adic filtration, i.e. $M_n = I^nM$ for all $n \geq 0$, this problem was solved by Rossi, Trung and Valla provided $M=R$ and $I=\mm$ (see \cite{RTV}) and by Linh in the general case. The bound was given in terms of the so-called extended degree of $M$ w.r.t. $I$ (see \cite[Theorem 4.4]{L1}). We now recall the definition of this notion.
\begin{Definition} \label{A1} {\rm (see \cite[Section 3]{L1})
An {\it extended degree} $D(I,M)$ of $M$ w.r.t. an $\mm$-primary ideal $I$ is a numerical function satisfying the following properties:
(i) $D(I,M) = D(I,M/L) + \ell (L)$, where $L = H^0_\mm(M)$,
(ii) $D(I,M) \geq D(I,M/xM)$ for a generic element $x\in I \setminus \mm I$ on $M$,
(iii) $D(I,M) = e(I,M)$ if $M$ is a Cohen-Macaulay $A$-module,
where $e(I ,M)$ denotes the multiplicity of $M$ w.r.t. $I.$ }
\end{Definition}
This notion extends that of extended degrees of graded modules in \cite{DGV} and \cite[Definition 9.4.1]{Va}. Assume that $A$ is a homomorphic image of a Gorenstein ring $S$ with $\dim S = s$. Then, as an example of extended degrees we can take the homological degree which is defined recursively as follows
\begin{gather}
\hdeg(I,M) := e(I,M) + \displaystyle \sum_{i=0}^{d-1} {d-1 \choose
i}\hdeg(I,\Ext_S^{s+i+1-d}(M,S)) \label{E:hdeg}
\end{gather}
if $d= \dim M > 0$, and $\hdeg(I,M) =\ell(M)$ if $d= 0$. This follows from \cite[Theorem 9.4.2]{Va}.
For filtered modules, Linh's result \cite[Theorem 4.4]{L1} can be modified as follows:
\begin{Theorem}\label{A3}
Let $M$ be a finitely generated $A$-module with $\dim M = d \geq 1$, $\FilM = \{ M_n\}_{n\geq 0}$ a good $I$-filtration of $M$ and $D(I,M)$ an arbitrary extended degree of $M$ with respect to $I$. Then
\rm{(i)} \ $\reg(G(\FilM) )\leq D(I,M) +r(\FilM) - 1 \ {\mathrm{if}} \ d = 1$,
\rm{(ii)} $\reg(G(\FilM)) \leq [D(I,M) + r (\FilM)+1]^{3(d-1)!-1} - d\ {\mathrm{if}}\ d\geq 2$.
\end{Theorem}
Note that the bound in (ii) is simpler and in most cases a little bit better than the one in \cite[Theorem 4.4]{L1}. The proof is similar to that in \cite{L1}, but it needs some modifications. The main reason for the modifications is the fact that unlike the $I$-adic case, the module $G(\FilM)$ is not generated in degree $0$. Therefore below we formulate the required modifications and only give a proof when it is necessary.
We call
$$H_{\FilM}(n) = \ell(M/M_{n + 1})$$
the Hilbert-Samuel function of $M$ w.r.t $\FilM$. Since $H_{\FilM}(n) = \ell(M/I^{n + 1 - r}M_r)$ for all $n \geq r$, this function agrees with a polynomial - called the Hilbert-Samuel polynomial and denoted by $P_{\FilM}(n)$ - for $n \gg 0$. The Hilbert-Samuel function $\ell(M/I^{n+1}M)$ and its Hilbert-Samuel polynomial of $M$ w.r.t. an $\mm$-primary ideal $I$ are denoted by $H_{I,M}(n)$ and $P_{I,M}(n)$, respectively. Whenever $R_0$ is an Artinian local ring, we denote the Hilbert function $\ell_{R_0}(E_n)$ and the Hilbert polynomial of a finitely generated graded module $E$ over a graded ring $R$ by $h_E(n)$ and $p_E(n)$, respectively. Note that
\begin{equation} \label{Hh} H_{\FilM}(n) = \sum_{j=0}^n h_{G(\FilM)}(j).\end{equation}
\begin{Lemma}\label{A4} {\rm (cf. \cite[Lemma 3.5]{L1})}
Let $x \in I \setminus \mm I$ such that the initial form $x^*$ of $x$ in $G_I(A)$ is a filter-regular element on $G(\FilM)$. Let $N = M/xM$. Then
$$p_{G(\FilM)}(n) \leq H_{I,N}(n)$$
for $n \geq \reg(G(\FilM/xM)).$
\end{Lemma}
\begin{pf}
By a filtration version of the so-called Singh's formula (see \cite[Lemma 1.9]{RV}) we have
$$h_{G(\FilM)}(n) = H_{\FilM/xM}(n) - \ell(M_{n+1} : x/M_n).$$
Using Lemma \ref{Fil-Property}(iii) and essentially the same proof of \cite[Lemma 3.5]{L1} we get
$$p_{G(\FilM)}(n) \leq H_{\FilM/xM}(n)$$
for $n \geq \reg(G(\FilM/xM))$. Since $I^{n+1}M \subseteq M_{n+1}$, we have
$$H_{\FilM/xM}(n) = \ell(M/M_{n+1} + xM) \leq \ell(M/I^{n+1}M + xM) = H_{I,N}(n) .$$
\end{pf}
Recall that a subideal $Q\subseteq I$ is said to be a reduction of $I$ if $I^{n+1} = QI^n$ for $n\gg 0$. If $Q$ is a reduction of $I$, which does not properly contain a reduction of $I$, then $Q$ is called a {\it minimal reduction} of $I$. The second statement of the following result is \cite[Theorem 3.6]{L1}. However our proof here is much shorter.
\begin{Lemma}\label{A5}
Let $\dim M=d\geq 1$ and $I$ be an $\mm$-primary ideal. Then
{\rm (i)} $\ell(M/I^{n+1}M)\leq \binom{n+d}{d}\ell(M/QM),$
where $Q$ is a minimal reduction ideal of $I(A/\Ann(M)),$
{\rm (ii)} $\ell(M/I^{n+1}M)\leq \binom{n+d}{d}D(I,M).$
\end{Lemma}
\begin{pf}
(i) Let $Q = (x_1, ..., x_d)$. From the epimorphism
$$ B := (M/QM)[x_1, ..., x_d] \longrightarrow \bigoplus_{n\geq 0} Q^nM/Q^{n+1}M,$$
we get that
$$\ell(M/I^{n+1}M) \leq \ell(M/Q^{n+1}M) \leq \sum_{i=0}^{i=n} \ell(B_i) \leq \binom{n+d}{d}\ell(M/QM).$$
(ii) We may choose $x_1, ..., x_d \in I$ such that $x_i$ is a generic element on
$M/(x_1, ..., x_{i-1})M$. Then $Q = (x_1, ..., x_d)$
is a minimal reduction of $I(A/\Ann(M))$.
By (ii) and (iii) of Definition \ref{A1}, we get
$$D(I,M) \geq D(I,M/(x_1, ..., x_d)M) = \ell(M/(x_1, ..., x_d)M) = \ell(M/QM).$$
Hence the statement follows from (i).\end{pf}
Recall that the geometric Castelnuovo-Mumford regularity of a graded $R$-module $E$ is
$$\geom(E) = \max \{a_i(E) + i \mid i \geq 1 \}$$
(see \cite{RTV}). The following result is a module version of \cite[Corollary 5.3]{H} and the proof is essentially the same (see also \cite[Proposition 3.5]{HZ} and \cite[Lemma 4.2]{L1}).
\begin{Lemma} \label{A6}
Let $M$ be a finitely generated $A$-module such that $\depth M > 0.$ Then
$$\reg(G(\FilM)) = \geom(G(\FilM)).$$
\end{Lemma}
The next result is an extension of \cite[Lemma 4.3]{L1} to the filtration case.
\begin{Lemma}\label{A7}
Let $\overline{M} = M/H^0_\mm(M)$. Denote the filtration $\FilM/ H^0_\mm(M)$ of $\overline{M}$ by $\overline{\FilM}$. Then
$$\reg(G(\FilM)) \leq \max \{ \reg(G(\overline{\FilM}));\ r(\FilM)\} + \ell(H^0_\mm(M)).$$
\end{Lemma}
\begin{pf}
Let $L = H^0_\mm(M)$ and
$$K = \bigoplus_{n\geq 0}(M_{n+1} + M_n \cap L)/M_{n+1} \cong \bigoplus_{n \geq 0} \frac{M_n \cap L}{M_{n+1}\cap L}.$$
By the Artin-Rees theorem there is an integer $c$ such that
$$M_{n+1} \cap L \subseteq I^{n+1-r}M \cap L \subseteq I^{n+1-r-c}L = 0$$
for $n \gg 0.$ Hence, $\ell(K) = \ell(L)$. It is clear that we have an exact sequence
\begin{gather}
0 \longrightarrow K \longrightarrow G(\FilM) \longrightarrow G(\overline{\FilM}) \longrightarrow 0.
\label{E:A7}
\end{gather}
Let $p = \max \{ \reg(G(\overline{\FilM}));\ r \}$. Then there is an integer $p + 1 \leq m \leq p + \ell(L) + 1$ such that $K_m = 0$. Since $m > \reg(G(\overline{\FilM}))$, from (\ref{E:A7}) it then implies that $H^i_{G_I(A)_+}(G(\FilM))_{m-i} = 0$ for all $i \geq 0$. Note that $G(\FilM)$ is generated over $G_I(A)$ by elements of degrees $\leq r \leq m -1$. Hence, by \cite[Lemma 2.1]{Na} $\reg(G(\FilM)) \leq m -1 \leq p + \ell(L)$, as required.
\end{pf}
We can now prove Theorem \ref{A3}. As said above, it is essentially the same as the proof of \cite[Theorem 4.4]{L1}. Hence, in the proof below we only give the details when modifications are needed.
\vskip0.5cm
\noindent {\it Proof of Theorem \ref{A3}}.
Let $G=G_I(A)$, $L = H^0_\mm(M)$ and $x \in I \setminus \mm I$ be a generic element on $M$. Let $\overline{M} = M / L$ and recall that $ r : = r(\FilM)$. By Lemma \ref{A7},
$$\reg(G(\FilM)) \leq \max \{ \reg(G(\overline{\FilM}));\ r\} + \ell(L).$$
By (i) of Definition \ref{A1}, $D(I,M) = D(I,\overline{M}) + \ell(L)$. Hence, we only need to show the following statements:
(i') $\reg(G(\overline{\FilM})) \leq D(I,\overline{M}) + r - 1 \ {\mathrm{if}} \ d = 1$,
(ii') $\reg(G(\overline{\FilM})) \leq [D(I,\overline{M}) + r + 1]^{3(d-1)!-1} - d\ {\mathrm{if}}\ d\geq 2$.
\noindent Replacing $M$ by $\overline{M}$, we may assume that $\depth M > 0$. Set $D= D(I,M)$. By Lemma \ref{A6},
\begin{equation}
\reg(G(\FilM)) = \geom (G(\FilM)).\label{E:A3}
\end{equation}
If $d = 1$ then $M$ is a Cohen-Macaulay module, $G(\FilM)$ is a $G$-module of dimension one generated by elements of degree at most $r$. Hence, by Lemma \ref{A6} and \cite[Lemma 2.2]{L1}, we have
$$\reg(G(\FilM)) = \geom(G(\FilM)) =a_1(G(\FilM)) + 1 \leq e(G(\FilM)) + r - 1 = e(I,M) +r - 1 .$$
The last equality follows from \cite[Proposition 11.4(iii)]{AM}. Hence, by Definition \ref{A1}(ii), $\reg(G(\FilM)) \le D + r-1$.
Note that $r(\FilM/ xM) \le r$ and $D >0$. If $d \geq 2$, let $N = M/xM$ and $m: = \max \{ r;\ \reg(G(\FilM/xM)) \}$. Using the following exact sequence
\begin{equation} \label{Filsq} 0 \rightarrow \bigoplus_{n\ge 1}\frac{xM \cap M_n}{xM_{n-1} + xM\cap M_{n+1}} \rightarrow G(\FilM)/x^*G(\FilM) \rightarrow G(\FilM/xM) \rightarrow 0,
\end{equation}
and Lemma \ref{Fil-Property}(ii) we get $\geom(G(\FilM)/x^*G(\FilM)) = \geom(G(\FilM/xM))$. Hence,
$$\geom(G(\FilM)/x^*G(\FilM)) \leq m.$$
Since $G(\FilM)$ is generated by elements of degree at most $r \leq m$, by \cite[Theorem 2.7]{L1},
$$\geom(G(\FilM)) \leq m + p_{G(\FilM)}(m).$$
By Lemma \ref{A4} and Lemma \ref{A5} (ii) and the fact that $D(I,N) \leq D$, we get
\begin{equation}
\geom(G(\FilM)) \leq m + D(I,N){m+d-1\choose d-1} \leq m+ D{m+d-1\choose d-1}.\label{E:A3b}
\end{equation}
If $d = 2$, then by (i) of the theorem
$$m = \max \{ r;\ \reg(G(\FilM/xM)) \} \leq D(I,N) + r - 1 \leq D + r -1.$$
Since $r\ge 0$, by (\ref{E:A3}) and (\ref{E:A3b}), we get
\begin{align*}
\reg(G(\FilM)) = \geom(G(\FilM)) & \leq m + D(m+1) \\
& \leq D^2 +(r + 1)D + r - 1 \\
& \leq (D + r + 1)^2 - 2.
\end{align*}
Let $d \geq 3$. The case $m=0$ is trivial, so we may assume $m>0$. From (\ref{E:A3b}) we get
\begin{gather}\geom(G(\FilM)) \leq m + D{m+d-1 \choose d-1} \leq D(m + 1)^{d - 1} -1. \label{E:A3c}
\end{gather}
By the induction hypothesis we may assume that
$$ m \leq [D(I, N) + r + 1]^{3(d-2)! -1} - d+1\leq (D + r + 1)^{3(d-2)! - 1} - d+ 1.$$
Hence, by (\ref{E:A3}) and (\ref{E:A3c}),
$$\reg(G(\FilM)) \leq [D + r + 1]^{3(d-1)! - 1} - d.$$
\ \hfill $\square$
\vskip0.5cm
If we write
$$P_\FilM (t) = \sum_{i=0}^d (-1)^i e_i(\FilM){t+d-i \choose d-i},$$
then the integers $e_i(\FilM)$ are called {\it Hilbert coefficients} of $\FilM$ (see \cite[Section 1]{RV}). Note that $e_0(\FilM),...,e_{d-1}(\FilM)$ are the Hilbert coefficients of the graded module $G(\FilM)$. In the last section we need an estimation of Hilbert coefficients. This was done for the $\mm$-adic filtration of a ring in \cite[Theorem 4.1]{RTV} and extended to the module case in \cite[Theorem 3.1]{L2}. However the proof of \cite[Theorem 3.1]{L2} has a gap. Namely, in the case $d=1$ there was used the following wrong inequality: $|(e+1)e(I,M) - \ell(M/I^{r+1}M)| \le |(e+1)e(I,M) - (r+1)|$. Therefore we give here the proof of the following result, which moreover improves \cite[Theorem 3.1]{L2}.
\begin{Theorem} \label{Hilb} Let $M$ be a finitely generated $A$-module with $\dim M = d \geq 1$ and $\FilM = \{ M_n\}_{n\geq 0}$ a good $I$-filtration of $M$ . Then
\item[(i)] $e_0(\FilM) = e(I,M) \le D(I,M)$;
\item[(ii)] $|e_1(\FilM) | \le (D(I,M) + r(\FilM) -1)D(I,M)$;
\item[(iii)] $| e_i(\FilM)|\le (D(I,M) + r (\FilM) +1)^{3i! -i +1} $ if $i\ge 2$.
\end{Theorem}
\begin{pf} (i): By \cite[Proposition 11.4(iii)]{AM} $e_0(\FilM) = e(I,M)$. By Definition \ref{A1}, $e(I,M) \le D(I,M)$.
(ii) -(iii): From the Grothendieck-Serre formula
$$h_{G(\FilM)}(n) - p_{G(\FilM)}(n) = \sum_{j=0}^d(-1)^j H^j_{G_+}({G(\FilM)})_n,$$
it follows that $h_{G(\FilM)}(n) = p_{G(\FilM)}(n)$ for all $n>\reg(G(\FilM))$.
By (\ref{Hh}) it follows that
\begin{equation} \label{Ph} \ell(M/M_{m+1}) = \sum_{i=0}^d (-1)^i e_i(\FilM){m+d-i \choose d-i}
\end{equation}
for any $m\ge \reg(G(\FilM))$. For simplicity we set $r= r(\FilM)$, $D:= D(I,M)$ and $e_i := e_i(\FilM)$.
Assume that $d=1$. Using Theorem \ref{A3}(i) and putting $m = D+r -1$ into the equality (\ref{Ph}), we have
$$e_1 =(D+r) e_0 - \ell(M/M_{D+r}).$$
Since $M_{n} = I^{n-r}M_r$ for $n\ge r$ and $M_r \neq 0$,
$$\ell(M/M_{D+r})\ge \ell(M_r/IM_r) + \cdots + \ell(I^{D-1}M_r/I^D M_r) \ge D.$$
By the first claim, this implies
$$e_1 \le (D+r)e_0 - D \le D (D+r) - D= D(D+r-1).$$
On the other hand, since $r\ge 0$, by Lemma \ref{A5}(ii),
$$- e_1 =- (D+r) e_0 + \ell(M/M_{D+r})\le D(D+r) - (D+r) \le D(D+r-1).$$
Hence $|e_1| \le D(D+r-1)$ and the case $d=1$ is proven.
Let $d\ge 2$. First we assume that $\depth M >0$. Let $x\in I\setminus \mm I$ be a generic element. Then $x^*\in G$ is a filter-regular element on $G(\FilM)$. By
\cite[Proposition 1.10]{RV}, $e_i = e_i(\FilM/xM)$ for all $i<d$. Note that $0 \le r(\FilM/xM) \le r$ and by Definition \ref{A1}(ii), $D(I, M/xM) \le D$. Using the induction hypothesis we then get
\begin{equation}\label{EHilb0} |e_1| \le D(D+r- 1) \ \ \text{and} \ \ | e_i| \le (D + r +1)^{3i! -i +1} \ \text{for} \ 2\le i \le d-1.\end{equation}
To prove the inequality for $e_d$, we set $\mu = (D+r+1)^{3(d-1)!-1}$. By Theorem \ref{A3}, $\reg(G(\FilM) ) \le \mu -d$. Since $\reg(G(\FilM)) \ge r\ge 0$, $\mu \ge d$. Hence putting $m= \mu -d$ into (\ref{Ph}), we have
\begin{eqnarray}\nonumber |e_d| & = |(\ell(M/M_{\mu -d +1}) - e_0{\mu -d +d \choose d}) + \sum_{i=1}^d (-1)^i e_i{\mu -d +d-i \choose d-i}| \\
&\nonumber \le |\ell(M/M_{\mu -d +1}) - e_0{\mu \choose d}| + \sum_{i=1}^{d-1} | e_i|{\mu -i \choose d-i} \\
& \le \max\{\ell(M/M_{\mu -d +1}), \ e_0{\mu \choose d}\} + \sum_{i=1}^{d-1} | e_i|{\mu -i \choose d-i}. \label{EHilb1}
\end{eqnarray}
Note that ${\mu \choose d} \le \mu^d$. By (i) and Lemma \ref{A5}(ii) it yields
\begin{equation} \label{EHilb2} \max\{\ell(M/M_{\mu -d +1}), \ e_0{\mu \choose d}\} \le D\mu^d. \end{equation}
Further, by (\ref{EHilb0})
\begin{equation} \label{EHilb3}
|e_1|{\mu -1 \choose d-1} \le D(D+r-1)\mu^{d-1},
\end{equation}
and
\begin{equation} \label{EHilb4}
\sum_{i=2}^{d-1} | e_i|{\mu -i \choose d-i} \le \sum_{i=2}^{d-1} (D+r+1)^{3i! -i +1}\mu^{d-i} \le \mu^{d-1}\sum_{i=0}^{d-2}\frac{1}{2^i} < 2\mu^{d-1}.
\end{equation}
Since $D(D+r- 1) + 2< (D+r+1)^2 \le \mu $, from (\ref{EHilb1}) - (\ref{EHilb4}) we obtain
$$|e_d| \le D\mu^d + \mu^d = (D+1)(D+r+1)^{3d! -d} \le (D+r+1)^{3d! - d+1}.$$
Finally assume that $\depth M =0$. Set $L = H^0_\mm(M)$, $\bar{M} = M/L$ and $\overline{\FilM} = \FilM/L$. Then
$$\begin{array}{ll} \ell(\frac{M}{M_{n+1}}) & = \ell(\frac{M}{M_{n+1}+L}) + \ell(\frac{M_{n+1} + L}{M_{n+1}}) = \ell(\frac{M}{M_{n+1}+L}) + \ell(\frac{L}{L\cap M_{n+1}})\\
& = \ell(\frac{M}{M_{n+1}+L}) +\ell(L)
\end{array}$$
for $n\gg 0$ (by Lemma \ref{Fil-Property}(iii)). Hence $e_i = e_i(\overline{\FilM})$ for $i\le d-1$ and $e_d = e_d(\overline{\FilM}) +(-1)^d\ell(L)$. Since $D = D(I, \bar{M}) + \ell(L)$ and $r(\overline{\FilM}) \le r$, we get
$$\begin{array}{ll}
|e_1| & = e_1(\overline{\FilM}) \le D(I,\bar{M}) (D(I,\bar{M})+r-1) \le D(D+r-1)\\
| e_i| & = e_i(\overline{\FilM}) \le (D(I, \bar{M}) + r +1)^{3i! -i +1} \le (D + r +1)^{3i! -i +1} \ \text{for} \ 2\le i \le d-1,\\
| e_d| & \le e_d(\overline{\FilM}) +\ell(L) \le (D(I, \bar{M}) + r +1)^{3d! -id+1} + \ell(L) \le (D + r +1)^{3d! -d+1}.
\end{array}$$
\end{pf}
\section{Regularity of associated graded rings: Graded case} \label{Graded}
Let $A = \oplus _{n \geq 0}A_n$ be a Noetherian standard graded algebra over an Artinian local ring $(A_0, \mm_0)$, i.e. $A = A_0[A_1]$. As usual, we assume that $A_0 / \mm_0$ is infinite. We denote the maximal homogeneous ideal $\mm_0 \oplus (\oplus_{n \geq 1} A_n)$ of $A$ by $\mm$. Let $M = \oplus_{n\in \ZZ}M_n$ be a finitely generated graded $A$-module of dimension $d$ and $\FilM = \{ \Mc_n\}$ a good $I$-filtration consisting of homogeneous submodules of $M$, where $I$ is a homogeneous $\mm$-primary ideal of $A$. Our goal is to bound $\reg(G(\FilM))$ in terms of $\reg(M)$ and $r(\FilM)$ in the case $A$ is a polynomial ring over a field.
More generally, let $A$ be a homomorphic image of a Gorenstein graded algebra $S$ with $\dim S =s$. We define $e(I,M)$ to be the degree of the graded module $G_I(M)$, or equivalently, $(d-1)!$ times of the leading coefficient of the Hilbert polynomial $h_{G_I(M)}(t)$. Then, as in the local case, one can define the homological degree $\hdeg(I,M)$ by formula (\ref{E:hdeg}). In particular, if $I = \mm$, we set
\begin{gather}
\hdeg(M) := \hdeg(\mm,M)= e(M) + \displaystyle \sum_{i=0}^{d-1} {d-1 \choose
i}\hdeg(\mm, \Ext_R^{s + i +1 -d}(M,R)).\label{E:hdeg2}
\end{gather}
In fact, this definition was first given in \cite{DGV} and \cite[Definition 9.4.1]{Va}.
We now bound $\reg(G(\FilM))$ in terms of $r(\FilM)$ and $\hdeg(M)$. Our idea is to reduce to the local case.
\begin{Lemma}\label{B1}
Let $I \subset A$ be a homogeneous $\mm$-primary ideal. Then
$$\hdeg(I_\mm,M_\mm) \leq \ell(A/I)^d\hdeg(M).$$
\end{Lemma}
\begin{pf}
Let $p = \ell(A/I) = \ell((A/I)_\mm)$. There is a composition series
$$0 = L_0 \subset L_1 \subset ... \subset L_p = (A/I)_\mm.$$
Hence, $\mm^p A_\mm \subseteq I_\mm$. By \cite[Lemma 1.3]{L2}, we have
$$\hdeg(I_\mm,M_\mm) \leq p^d \hdeg(\mm A_\mm,M_\mm).$$
Since the $\Ext$ functor commutes with the localization (see, e.g. \cite[Theorem 9.50]{Ro}), from the recursive formulas (\ref{E:hdeg}) and (\ref{E:hdeg2}) and the fact that $e(E) = e(E_\mm)$ for any finitely generated graded $A$-module $E$, it follows that $\hdeg(\mm A_\mm,M_\mm) = \hdeg(M)$. Hence, $\hdeg(I_\mm,M_\mm) \leq p^d \hdeg(M).$
\end{pf}
\begin{Theorem}\label{B3}
Let $\FilM$ be a good $I$-filtration of a graded $A$-module $M$ of dimension $d$. Then
\rm{(i)} $\reg(G(\FilM)) \leq \ell(A/I)\hdeg(M) + r(\FilM) - 1 \ {\mathrm{if}} \ d = 1$,
\rm{(ii)} $\reg(G(\FilM)) \leq [\ell(A/I)^d\hdeg(M) + r (\FilM) +1]^{3(d-1)!-1} -d\ {\mathrm{if}}\ d\geq 2$.
\end{Theorem}
\begin{pf} Denote by $\RR = A \oplus I \oplus I^2 \oplus ...$ the Rees algebra of $A$ w.r.t. $I$ and $\RR_+ = \oplus _{n\geq 1} I^n$. Then we can consider $G(\FilM)$ as a finitely generated module over $\RR$. If $E = \oplus_{n \in \ZZ} E_n$ is a graded module over $\RR$, we denote by $E_\mm$ and $(E_n)_\mm$ the localization of $E$ and $E_n$ as $A$-modules with respect to the multiplicative set $A \setminus \mm$. We can consider $G_I(A)$ and $G(\FilM)$ as graded modules over $\RR$. Then it is easy to see that $(G(\FilM))_\mm \cong G(\FilM_\mm)$ and $(G_I(A))_\mm \cong G_{I_\mm}(A_\mm)$. This implies
$$\reg(G(\FilM)) = \reg(G(\FilM_\mm)).$$
On the other hand, by Lemma \ref{B1},
$$\hdeg(I_\mm,M_\mm) \leq \ell(A/I)^d\hdeg(M).$$
We also have $r(\FilM_\mm) \leq r(\FilM).$ Hence, applying Theorem \ref{A3} to the homological degree we get the statements of the theorem.
\end{pf}
An important consequence of Theorem \ref{B3} is
\begin{Corollary}\label{B4}
Let $I$ be an $\mm$-primary homogeneous ideal of a polynomial ring $A = K[x_1,...,x_n]$ over an infinite field $K$. Then
\rm{(i)} $\reg(G_I(A)) \leq \ell(A/I) - 1 \ {\mathrm{if}} \ d = 1$,
\rm{(ii)} $\reg(G_I(A)) \leq (\ell(A/I) +1)^{3(d-1)!-1} - d\ {\mathrm{if}}\ d\geq 2$.
\end{Corollary}
If $M$ is an arbitrary graded module over a polynomial ring $A$, then we can bound $\reg(G(\FilM))$ in terms of $\reg(M),\ r(\FilM)$ and some other invariants of $M$ as follows:
\begin{Theorem}\label{B5}
Let $M$ be a finitely generated graded module of dimension $d$ over a polynomial ring $A = K[x_1,...,x_n]$. Let $i(M)$ denote the initial degree of $M$ (i.e. $i(M) = \min\{ p\mid M_p \ne 0 \}$) and $\mu(M)$ the minimal number of generators of $M$. Then
\rm{(i)} $\reg(G(\FilM)) \leq \ell(A/I)\mu (M)[\reg(M) - i(M) +1]^n +r(\FilM) - 1 \ {\mathrm{if}} \ d = 1$,
\rm{(ii)} $\reg(G(\FilM)) \leq [\ell(A/I)^d (\mu (M)(\reg(M) - i(M) +1)^n)^{2^{(d - 1)^2}} + r(\FilM)+1]^{3(d - 1)! - 1} - d\ {\mathrm{if}}\ d\geq 2$.
\end{Theorem}
\begin{pf}
By \cite[Theorem 5.1]{CHH},
\begin{align*}
\hdeg(M) &\leq \biggl[\mu(M) \binom{\reg(M) - i(M) +n}{n}\biggr]^{2^{(d-1)^2}} \\
& \leq [\mu(M) (\reg(M) - i(M) +1)^n]^{2^{(d - 1)^2}}.
\end{align*}
Hence, the statement follows from Theorem \ref{B3}.
\end{pf}
In the rest of this section we assume that $\FilM$ is a good $I$-filtration, where $I$ is an $\mm$-primary ideal generated by homogeneous elements of the same degree. Under this setting we will give a bound on $\reg(G(\FilM))$ which is much better than the ones in Theorem \ref{B5} and holds for any standard graded ring $A$. In this new setting, we have
\begin{Lemma}\label{C1}
Let $i(M)$ denote the initial degree of $M$. Let $\overline{M} := M/H^0_\mm(M)$ and $\overline{\FilM} := \FilM/H^0_\mm(M)$. Then
$$ \reg(G(\FilM)) \leq \max\{\reg(G(\overline{\FilM})); \ \reg(M) -i(M) +r(\FilM) \}.$$
\end{Lemma}
\begin{pf}
As in the proof of the Lemma \ref{A7}, we have the following exact sequence
\begin{gather}
0\longrightarrow K\longrightarrow G(\FilM) \longrightarrow G(\overline{\FilM}) \longrightarrow 0,\label{E:C1}
\end{gather}
where
$$K = \bigoplus_{n \geq 0} \frac{\Mc_{n+1}+\Mc_n\cap H^0_\mm(M)}{\Mc_{n+1}}$$
is a module of finite length.
Let $n \geq \reg(M) - i(M) + r + 1$, where $r := r(\FilM)$. Note that $\reg(M) \geq i(M)$ and $I$ is generated by homogeneous elements of degree at least one. Hence,
$$ \Mc_n = I^{n-r}\Mc_r \subseteq I^{n-r}M \subseteq \bigoplus _{p \geq \reg(M) + 1}M_p.$$
Since $H^0_\mm(M) \subseteq M$ and $H^0_\mm(M)_p = 0$ for all $p \geq \reg(M) + 1$, we get that $\Mc_n \cap H^0_\mm(M) = 0$. Hence, $K_n = 0$. From this fact and the exact sequence (\ref{E:C1}) we get the statement of the lemma.
\end{pf}
\begin{Lemma}\label{C2}
Let $x \in I \setminus \mm I$ be a homogeneous element. Assume that the initial form $x^*$ of $G_I(A)$ is a filter-regular element on $G(\FilM)$. Then $x$ is a filter-regular element on $M$.
\end{Lemma}
\begin{pf}
The element $x^*$ is a filter-regular element on $G(\FilM)$ means that
\begin{gather}
(\Mc_{n+2}:x) \cap \Mc_n = \Mc_{n + 1}\label{E:C2}
\end{gather}
for all $n \geq n_0$, where $n_0$ is a certain fixed number. Let $u \in (0:_Mx)_p$. Since $M$ is finitely generated and $M/\Mc_{n_0}$ is of finite length, it follows that $u \in \Mc_{n_0}$ when $p \gg 0$. Then, by (\ref{E:C2}),
$$ u \in (0:_Mx) \cap \Mc_{n_0} \subseteq (\Mc_{n_0+2}:x) \cap \Mc_{n_0} = \Mc_{n_0+1}.$$
By induction it implies that
$$ u \in \bigcap_{q \geq n_0} \Mc_q = \bigcap_{q \gg 0} I^{q-r}\Mc_r = 0.$$
This means $(0:_Mx)_p = 0$ for $p \gg 0$, or equivalently $x$ is a filter-regular element on $M$.
\end{pf}
We can now improve Theorem \ref{B5} in the case of an equi-generated ideal $I$ as follows:
\begin{Theorem}\label{C3}
Assume that $I$ is generated by elements of degree $\Delta \ge 1$. Let $Q$ be a homogeneous minimal reduction of $I(A/\Ann(M))$. Let $i(M)$ denote the initial degree of $M$. Then
\rm{(i)} \ $\reg(G(\FilM)) \leq \ell(M/QM) + r (\FilM) + \reg(M) - i(M) - 1 \ {\mathrm{if}} \ d = 1$,
\rm{(ii)} $\reg(G(\FilM)) \leq [\ell(M/QM) + r(\FilM) + \reg(M) -i(M) + (d-1)\Delta ]^{3(d-1)!-1} - d\ {\mathrm{if}}\ d\geq 2$.
\end{Theorem}
\begin{pf} The main idea is the same as in the proof of \cite[Theorem 4.4]{L1}.
By Lemma \ref{C1}, it suffices to consider the case $\depth M > 0$. Let $r:= r(\FilM)$.
(i) If $d = 1$, then $M$ is a Cohen-Macaulay module. Hence, by \cite[Lemma 2.2]{L1}, Lemma \ref{A6} and \cite[Proposition 11.4(iii)]{AM}, we get
\begin{align*}
\reg(G(\FilM)) & = \geom(G(\FilM)) = a_1( G(\FilM)) + 1\leq e(G(\FilM)) + r -1 \\
& = e(I,M) + r-1 = e(Q,M) +r - 1 \leq \ell(M/QM) +r - 1.
\end{align*}
(ii) $d \geq 2$. It is clear that $Q$ is generated by $d$ elements of degree $\Delta $. Then one can find a minimal basis $\{x_1, ..., x_d \}$ of $Q$ such that the initial form $x^*_1$ in $G$ is a filter-regular element on $G(\FilM)$ (see \cite[Lemma 3.1]{T}). Note that all elements $x_1,...,x_d$ have degree $\Delta \geq 1$. Let $x = x_1$. Then $N:= M/xM$ is again a graded module. Let $m = \max\{r; \reg(G(\FilM/xM)) \}$. Using the exact sequence (\ref{Filsq}) we then get $\geom(G(\FilM)/x^*G(\FilM)) = \geom(G(\FilM/xM)) \le m$. Note that $(x_2, ... x_d)$ is a minimal reduction of $I(A/\Ann(N))$. Hence, by \cite[Theorem 2.7]{L1}, Lemma \ref{A4} and Lemma \ref{A5} (i) we get
\begin{align}
\geom(G(\FilM)) &\leq m + \ell(N/(x_2, ..., x_d)N) \binom{m + d - 1}{d-1}\notag \\
&\leq m+ \ell(M/QM)\binom{m + d - 1}{d-1},\label{E:C4}
\end{align}
which is similar to (\ref{E:A3b}).
Note that $\reg(N) \leq \reg(M) + \Delta -1, \ r (\FilM/xM)\le r$ and $i(N) \geq i(M)$. Let $d = 2$. By the induction hypothesis we have
\begin{align*}
\reg(G(\FilM/xM)) & \leq \ell(N/(x_2,...,x_d)N) + r + \reg(N) -i(N) -1 \\
& \leq \ell(M/QM) + r+\reg(N)-i(M)-1\\
& \le \ell(M/QM) +r+\reg(M)-i(M)+\Delta - 2.
\end{align*}
Hence,
$$m \le \ell(M/QM) +r+\reg(M)-i(M)+\Delta -2 .$$
Together with (\ref{E:C4}) and Lemma \ref{A6} this yields
$$\reg(G(\FilM)) = \geom(G(\FilM)) \leq [\ell(M/QM)+r+\Delta +\reg(M) -i(M)]^2-2.$$
If $d\geq 3$, then again by the induction hypothesis we get
\begin{align*}
m & \leq [\ell(N/(x_2, ... x_d)N)+r+(d-2)\triangle+\reg(N)-i(N)]^{3(d-2)!-1} -d+1\\
& \leq [\ell(M/QM)+r+(d-1)\triangle+\reg(M)-i(M)]^{3(d-2)!-1} -d+1.
\end{align*}
Hence, using (\ref{E:C4}) we can complete the proof of (ii).
\end{pf}
\section{Regularity of fiber cones} \label{Fiber}
In this section we assume that $\FilM = \{ M_n\}$ is a good $I$-filtration of a finitely generated module $M$ over a local ring $(A,\mm)$, where $I$ is an $\mm$-primary ideal. Given an ideal $\qq$ containing $I$, we define
$$F_\qq(\FilM) := \oplus_{n\ge 0}M_n/\qq M_n,$$
and call it the {\it fiber cone} of $\FilM$ with respect to $\qq$. This notion was introduced in \cite[Section 5]{RV} (see also \cite{JV}). If $\FilM$ is the $I$-adic filtration of $A$ and $\qq=\mm$, then this is the classical fiber cone $F_\mm(I) = \oplus_{n\ge 0}I^n/\mm I^n$ of $I$. Note that $F_\qq(\FilM)$ is a graded module over $G = G_I(A)$.
The purpose of this section is to give a bound for the Castelnuovo-Mumford regularity $\reg (F_\qq(\FilM))$ in terms of $D(I,A)$ and $r(\FilM)$. We will apply the results of Section \ref{Local}. Following \cite[(3)]{RV} we define a new good $I$-filtration
$$\qq M:\ M \supseteq \qq M \supseteq \qq M_1 \supseteq \cdots \supseteq \qq M_n \supseteq \cdots $$
\begin{Lemma}\label{F1}
Let $M$ be a finitely generated $A$-module with $\dim M = d \geq 1$, $\FilM = \{ M_n\}_{n\geq 0}$ a good $I$-filtration of $M$ and $D(I,M)$ an arbitrary extended degree of $M$ with respect to $I$. Assume that $I\subseteq \qq$ and $M_{n+1} \subseteq \qq M_n$ for all $n\ge 0$. Then
\rm{(i)} \ $a_0(F_\qq(\FilM) )\leq D(I,M) +r (\FilM) \ {\mathrm{if}} \ d = 1$,
\rm{(ii)} $a_0(F_\qq(\FilM)) \leq (D(I,M) + r(\FilM) +2)^{3(d-1)!-1} - d\ {\mathrm{if}}\ d\geq 2$.
\end{Lemma}
\begin{pf} Since $M_{n+1} \subseteq \qq M_n$ for all $n\ge 0$, by \cite[Proposition 5.1]{RV} we have the following exact sequence of $G$-graded modules
$$ 0 \rightarrow F_\qq(\FilM) \rightarrow G(\qq \FilM) \rightarrow N(-1) \rightarrow 0,$$
where $N = \oplus_{n\geq 0}\qq M_n/M_{n+1}$. Therefore
$$ a_0(F_\qq(\FilM) )\leq a_0(G(\qq \FilM)) \leq \reg(G(\qq\FilM)).$$
Note that $r(\qq \FilM) \le r(\FilM) +1 $. Hence the claim now follows from Theorem \ref{A3}.
\end{pf}
The study of Hilbert coefficients of $\FilM$ in Section \ref{Local} allows us to bound the Hilbert coefficients of the fiber cone $F_\qq(\FilM)$.
\begin{Proposition}\label{F2}
Under the assumption of Lemma \ref{F1} we have
\begin{itemize}
\item[(i)] $e_0(F_\qq(\FilM)) \leq 2D(I,M)(D(I,M) + r(\FilM) ).$
\item[(ii)] $|e_i (F_\qq(\FilM)) | \leq 2 (D(I,M) + r(\FilM) +2)^{3(i+1)! -i }$ if $1\le i \le d-1$.
\end{itemize}
\end{Proposition}
\begin{pf}
It was shown in \cite[(24)]{RV} that
$$e_i(F_\qq(\FilM)) = e_i(\FilM) + e_{i+1}(\FilM) - e_{i+1}(\qq \FilM),$$
for all $0\le i \le d-1$. Let $r:= r(\FilM)$. Since $r(\qq \FilM) \le r+1$, by Theorem \ref{Hilb}, we get
$$\begin{array}{ll}
e_0(F_\qq(\FilM)) &\le |e_0(\FilM)| + |e_1(\FilM) | + |e_1(\qq \FilM)|\\
&\le D+ D(D + r-1) + D(D+r) = 2D(D + r ),
\end{array}$$
where $D= D(I,M)$, and for $1\le i\le d-1$:
$$\begin{array}{ll}
| e_i(F_\qq(\FilM)) | &\le |e_i(\FilM)| + |e_{i+1}(\FilM) | + |e_{i+1}(\qq \FilM)|\\
&\le (D + r+1)^{3i! - i +1} + (D+r+1)^{3(i+1)! - i} + (D + r+2)^{3(i+1)! - i} \\
&\le 2(D + r+2)^{3(i+1)! - i}.
\end{array}$$
\end{pf}
\begin{Theorem}\label{F3}
Let $M$ be a finitely generated $A$-module with $\dim M = d \geq 1$, $\FilM = \{ M_n\}_{n\geq 0}$ a good $I$-filtration of $M$. Assume that $I\subseteq \qq$ and $M_{n+1} \subseteq \qq M_n$ for all $n\ge 0$. Then
\begin{itemize}
\item[(i)] $\reg(F_\qq(\FilM)) \le 2D(I,M)(D(I,M) + r(\FilM) ) + r(\FilM) -1$ if $d=1$;
\item[(ii)] $\reg(F_\qq(\FilM)) \le (D(I,M) +r(\FilM)+2)^2 + D(I,M)^2 - 3$ if $d= 2$;
\item[(iii)] $\reg(F_\qq(\FilM)) \le (D(I,M) +r(\FilM)+2)^{3(d-1)! - 1} - d$ if $d\ge 3$.
\end{itemize}
\end{Theorem}
\begin{pf} We do induction on $d$. Set $D= D(I,M)$ and $r=r(\FilM)$. Let $d=1$. By \cite[Lemma 2.2]{L1} and Proposition \ref{F2}(i) we have
$$ a_1(F_\qq(\FilM)) + 1 \le e_0(F_\qq(\FilM)) + r-1 \le 2D(D+r ) + r-1.$$
By Lemma \ref{F1} it yields
$$\begin{array}{ll} \reg(F_\qq(\FilM)) & = \max\{ a_0(F_\qq(\FilM));\ a_1(F_\qq(\FilM)) + 1\} \\
&\le \max\{ D+r; \ 2D(D + r )+r-1 \} = 2D(D+r ) + r-1.
\end{array}$$
Now let $d\ge 2$. Note that both $G(\FilM)$ and $F_\qq(\FilM)$ are modules over $G = G_I(A)$. Hence there exists a generic element $x\in I\setminus \mm I$ such that $x^*\in G$ is a filter-regular element on $G(\FilM)$ as well as a filter-regular element on $F_\qq(\FilM)$ (cf. \cite[Proposition 2.2]{RV}). Then
$$F_\qq(\FilM) / x^* F_\qq(\FilM) \cong \frac{M}{\qq M} \oplus (\oplus_{n\ge 0} \frac{M_n}{\qq M_n + xM_{n-1}}) ,$$
and $$ F_\qq(\FilM/xM) = \oplus_{n\ge 0} \frac{M_n}{\qq M_n + xM \cap M_n}.$$
Hence we have an exact sequence of $G$-modules:
$$0 \rightarrow K \rightarrow F_\qq(\FilM)/ x^*F_\qq(\FilM) \rightarrow F_\qq(\FilM/xM) \rightarrow 0,$$
where
$$K = \oplus_{n\ge 1} \frac{\qq M_n + xM \cap M_n}{\qq M_n + x M_{n-1}}.$$
By Lemma \ref{Fil-Property}(ii) $xM \cap M_n = xM_{n-1}$ for all $n> \reg (G(\FilM))$. Therefore $K$ is a module of finite length and $\reg(K) \le \reg(G(\FilM))$.
The above exact sequence gives
$$\begin{array}{ll}
\reg (F_\qq(\FilM)/ x^*F_\qq(\FilM) ) & = \max\{\reg(K); \ \reg(F_\qq(\FilM/xM))\}\\
& \le \max\{ \reg(G(\FilM)); \ \reg(F_\qq(\FilM/xM))\}.
\end{array}$$
By \cite[Proposition 20.20]{E},
$$\reg (F_\qq(\FilM)) = \max\{ a_0(F_\qq(\FilM)) ,\ \reg(F_\qq(\FilM)/x^*F_\qq(\FilM))\}.$$
Hence
\begin{equation} \label{EF3}
\reg (F_\qq(\FilM) ) \le \max\{a_0(F_\qq(\FilM));\ \reg(G(\FilM)); \ \reg(F_\qq(\FilM/xM))\}.
\end{equation}
Note that $r(\FilM/xM) \le r$ and by Definition \ref{A1}(ii) $D(I,M/xM) \le D$.
Using Theorem \ref{A3}, the inequality (\ref{EF3}) and Lemma \ref{F1} we get
$$\begin{array}{ll}
\reg (F_\qq(\FilM)) & \le \max\{ (D+r+2)^2 - 2; \ (D+r+1)^2 - 1; \ 2D(D+r )+r -1 \}\\
& < (D+r+2)^2 + D^2 - 3\end{array}$$
if $d=2$,
$$ \begin{array}{ll}
\reg (F_\qq(\FilM)) & \le \max\{ (D+r+2)^5 - 3; \ (D+r+1)^5 - 3; (D+r+2)^2 + D^2 -3 \} \\
& = (D+r+2)^5 - 3\end{array}$$
if $d=3$, and
$$ \begin{array}{ll}
\reg (F_\qq(\FilM)) & \le \max\{ (D+r+2)^{3(d-1)!-1}- d; \ (D+r+1)^{3(d-1)!-1} - d; \\
& \hskip2cm (D+r+2)^{3(d-2)!-1}- d+1 \}\\
& = (D+r+2)^{3(d-1)!-1}- d\end{array}$$
for all $d\ge 4$.
\end{pf}
As an immediate consequence of the above theorem we get the following bound for the Castelnuovo-Mumford regularity of the classical fiber cone of an $\mm$-primary ideal.
\begin{Corollary} \label{F4} Let
Let $I$ be an $\mm$-primary ideal of $d$-dimensional local ring $A$. Then
\begin{itemize}
\item[(i)] $\reg(F_\mm (I)) \le 2 D(I,A)^2 -1$ if $d=1$;
\item[(ii)] $\reg(F_\mm(I)) \le 2 D(I,A)^2 +4 D(I,A) + 1$ if $d= 2$;
\item[(iii)] $\reg(F_\mm(I)) \le (D(I,A) +2)^{3(d-1)! - 1} - d$ if $d\ge 3$.
\end{itemize}
\end{Corollary}
In the graded case we can apply the method in Section \ref{Graded} to bound the Castelnuovo-Mumford regularity of $F_\qq(\FilM)$. We formulate here only one result.
\begin{Proposition} Assume that $A$ is a homomorphic image of a Gorenstein graded algebra, $M$ is a finitely generated graded $A$-module with $\dim M = d \geq 1$, $I\subseteq \qq$ are graded $\mm$-primary of $A$, and $\FilM = \{ \Mc_n\}_{n\geq 0}$ is a good $I$-filtration of graded submodules of $M$ such that $\Mc_{n+1} \subseteq \qq \Mc_n$ for all $n\ge 0$. Then
\begin{itemize}
\item[(i)] $\reg(F_\qq(\FilM)) \le 2 \ell(A/I) \hdeg (I,M)(\ell(A/I) \hdeg(I,M) + r(\FilM)) + r(\FilM) -1$ if $d=1$;
\item[(ii)] $\reg(F_\qq(\FilM)) \le (\ell(A/I)^2 \hdeg(I,M) +r(\FilM) +2)^2 + \ell(A/I)^4 \hdeg(I,M)^2 - 3$ if $d= 2$;
\item[(iii)] $\reg(F_\qq(\FilM)) \le (\ell(A/I)^d \hdeg(I,M) +r(\FilM)+2)^{3(d-1)! - 1} - d$ if $d\ge 3$.
\end{itemize}
\end{Proposition}
\vskip0.5cm
\noindent {\bf Acknowledgment}: The authors would like to thank
the referee for his/her valuable comments.
|
\section{Introduction}
\label{sec:introduction}
\textbf{Problem setup} Recent studies and real-world incidents have
demonstrated the inability of the power grid to ensure a reliable service
in the presence of network failures and possibly malignant actions
\cite{ARM-RLE:10,MA:11}.
Besides failures and attacks on the {\em physical} power grid
infrastructure, the envisioned future smart grid is also prone to {\em
cyber} attacks on its communication layer. In short, cyber-physical
security is a fundamental obstacles challenging the smart grid vision.
A classical mathematical model to describe the grid on the
transmission level is the so-called {\em structure-preserving power
network model}, which consists of the dynamic {\em swing equation}
for the generator rotor dynamics, and of the algebraic
\emph{load-flow} equation for the power flows through the network
buses \cite{PWS-MAP:98}. In this work, we consider the linearized
small signal version of the structure-preserving model, which is
composed by the linearized swing equation and the {\em DC power flow
equation}. The resulting linear continuous-time descriptor model
of a power network has also been studied for estimation and security
purposes in \cite{ES:04,ADDG-ST:10,FP-AB-FB:10u}.
\begin{figure}
\centering
\includegraphics[width=.9\columnwidth]{./IEEE14}
\caption{For the here represented IEEE 14 bus system, if the
voltage angle of one bus is measured exactly, then a cyber
attack against the measurements data is always detectable by our
dynamic detection procedure. In contrary, as shown in
\cite{YL-MKR-PN:09}, a cyber attack may remain undetected by a
static procedure if it compromises as few as four measurements.}
\label{fig:ieee14}
\end{figure}
\textbf{From static to dynamic detection} Existing approaches to
security and stability assessment are mainly based upon static
estimation techniques for the set of voltage angles and magnitudes at
all system buses, e.g., see \cite{AA-AGS:04}. Limitations of these
techniques have been often underlined, especially when the network
malfunction is intentionally caused by an omniscient attacker
\cite{YL-MKR-PN:09,AT-AS-HS-KHJ-SSS:10}. The development of security
procedures that exploit the dynamics of the power network is
recognized~\cite{NB-TB-AB-VB-GC-DC-AF-LF-MGL-BFW-JNW:92} as an
outstanding important problem. We remark that the use of static state
estimation and detection algorithms has been adopted for many years
for several practical and technological reasons. First, because of the
low bandwidth of communication channels from the measuring units to
the network control centers, continuous measurements were not
available at the control centers, so that the transient behavior of
the network could not be captured. Second, a sufficiently accurate
dynamic model of the network was difficult to obtain or tune, making
the analysis of the dynamics even harder. As of today, because of
recent advances in hardware technologies, e.g., the advent of {\em
Phasor Measurement Units} and of large bandwidth communications, and
in identification techniques for power system parameters
\cite{AC-JHC-AS:11}, these two limitations can be overcome. Finally, a
dynamic estimation and detection problem was considered much harder
than the static counterpart. We address this theoretic limitation by
improving upon results presented in \cite{FP-AB-FB:09b,SS-CH:10a} for
the security assessment of discrete time dynamical networks.
\textbf{Literature review on dynamic detection} Dynamic security has
been approached via heuristics and expert systems, e.g., see
\cite{KYL:08}. Shortcomings of these methods include reliability and
accuracy against unforeseen system anomalies, and the absence of
analytical performance guarantees. A different approach relies on
matching a discrete-time state transition map to a series of past
measurements via Kalman filtering, e.g., see
\cite{ACZDS-JCSDS-AMLS:02,UAK-MDI-JMFM:10} and the references
therein. Typically, these transition maps are based on heuristic
models fitted to a specific operating point
\cite{ACZDS-JCSDS-AMLS:02}. Clearly, such a pseudo-model poorly
describes the complex power network dynamics and suffers from
shortcomings similar to those of expert systems methods.
In \cite{UAK-MDI-JMFM:10}, the state transition map is chosen more
accurately as the linearized and Euler-discretized power network
dynamics. The local observability of the resulting linear
discrete-time system is investigated in \cite{UAK-MDI-JMFM:10}, but in
the absence of unforeseen attacks.
Finally, in \cite{DK-XF-SL-TZ-KLBP:10} a graph-theoretic framework is
proposed to evaluate the impact of cyber attacks on a smart grid and
empirical results are given.
Recent approaches to dynamic security consider continuous-time power
system models and apply dynamic techniques
\cite{ES:04,ADDG-ST:10,FP-AB-FB:10u,AT-HS-KHJ:10}. While
\cite{AT-HS-KHJ:10} adopts an overly simplified model neglecting the
algebraic load flow equations, the references
\cite{ES:04,ADDG-ST:10,FP-AB-FB:10u} use a more accurate network
descriptor model.
In \cite{ADDG-ST:10} different failure modes are modeled as instances
of a switched system and identified using techniques from hybrid
control. This approach, though elegant, results in a severe
combinatorial complexity in the modeling of all possible
attacks.
In our earlier work \cite{FP-AB-FB:10u}, under the assumption of
generic network parameters, we state necessary and sufficient
conditions for identifiability of attacks based on the network
topology. Finally, in \cite{ES:04} dynamical filters are designed to
isolate certain predefined failures of the network components. With
respect to this last work, we assume no a priori knowledge of the set
of compromised components and of their compromised behavior. Our
results generalize and include those of \cite{ES:04}.
\textbf{Contributions} This paper's contributions are fourfold. %
First, we provide a unified modeling framework for dynamic power networks
subject to cyber-physical attacks. For our model, we define the notions of
{\em detectability} and {\em identifiability} of an attack by its effect on
output measurements. Informed by the classic work on geometric control
theory~\cite{WMW:85,HLT-AS-MH:01}, our framework includes the
\emph{deterministic static detection problem} considered in
\cite{YL-MKR-PN:09,AT-AS-HS-KHJ-SSS:10}, and the prototypical
\emph{stealth} \cite{DG-HS:10}, \emph{(dynamic) false-data injection}
\cite{YM-BS:10b}, and \emph{replay attacks} \cite{YM-BS:10a} as
special cases.
Second, we focus on the descriptor model of a power system and we show the
fundamental limitations of static and dynamic detection and identification
procedures. Specifically, we show that static detection procedures are
unable to detect any attack affecting the dynamics, and that attacks
corrupting the measurements can be easily designed to be undetectable. On
the contrary, we show that undetectability in a dynamic setting is much
harder to achieve for an attacker. Specifically, a cyber-physical attack is
undetectable if and only if the attackers' input signal excites uniquely
the zero dynamics of the input/output system. (As a complementary result,
our work \cite{FP-AB-FB:10u} gives necessary and sufficient graph-theoretic
conditions for the absence of zero dynamics, and hence for the absence of
undetectable attacks.) %
Third, we propose a detection and identification procedure based on
geometrically-designed residual filters. Under the assumption of attack
identifiability, our method correctly identifies the attacker set
independently of its strategy. From a system-theoretic perspective, correct
identification is implied by the absence of zero dynamics in our proposed
identification filters. Our design methodology is applicable to linear
systems with direct input to output feedthrough, and it generalizes the
construction presented in \cite{MAM-GCV-ASW-CM:89}.
Fourth and finally, we illustrate the potential impact of our theoretical
results on the standard IEEE 14 bus system (cf. Fig. \ref{fig:ieee14}). For
this system it is known~\cite{YL-MKR-PN:09} that an attack against the
measurement data may remain undetected by a static procedure if the
attacker set compromises as few as four measurements. We show here instead
that such an attack is always detectable by our dynamic detection procedure
provided that at least one bus voltage angle or one generator rotor angle
is measured exactly.
We conclude with two remarks on our contributions.
First, our results (the notions of detectability and identifiability,
the fundamental limitations of static versus dynamic monitoring, and
the geometric design of detection and identification filters) are
analogously and immediately applicable to arbitrary index-one
descriptor systems, thereby including any linear system
$\dot{x}=Ax+Bu$, $y=Cx+Du$, with attack signal $u$. Second, although
we treat here the noiseless case, it is well known~\cite{MB-IVN:93}
that our deterministic detection filters are the key ingredient,
together with Kalman filtering and hypothesis testing, in the design
of statistical identification methods.
\textbf{Organization} Section \ref{sec:model} presents the descriptor
system model of a power network, our framework for the modeling of
cyber-physical attacks, and the detection and identification
problem. Section \ref{sec:static_dynamic} states the fundamental
limitations of static and dynamic detection procedures. Section
\ref{sec:dynamic_detection} presents the residual filters for dynamic
detection and identification. Section \ref{sec:example} contains the IEEE
14 bus system case study.
\section{Cyber-physical attacks on power networks}
\label{sec:model}
\subsection{Structure-preserving power network model with cyber and
physical attacks}
We consider the linear small-signal version of the classical
structure-preserving power network model \cite{PWS-MAP:98}. This descriptor
model consists of the {\it dynamic linearized swing equation} and the {\it
algebraic DC power flow equation}. A detailed derivation from the
nonlinear structure-preserving power network model can be found, for
instance, in \cite{ES:04,FP-AB-FB:10u}.
Consider a connected power network consisting of $n$ generators
$\{g_{1},\dots,g_{n}\}$, their associated $n$ generator terminal buses
$\{b_{1},\dots,b_{n}\}$, and $m$ load buses
$\{b_{n+1},\dots,b_{n+m}\}$. The interconnection structure of the
power network is encoded by a connected admittance-weighted graph. The
generators $g_{i}$ and buses $b_{i}$ form the vertex set of this
graph, and the edges are given by the transmission lines
$\{b_{i},b_{j}\}$ weighted by the susceptance between buses $b_{i}$
and $b_{j}$, as well as the internal connections $\{g_{i},b_{i}\}$
weighted by the transient reactance between each generator $g_{i}$ and
its terminal bus $b_{i}$. The Laplacian matrix associated to the
admittance-weighted graph is the symmetric matrix $\left[
\begin{smallmatrix}
\subscr{\mathcal L}{gg} & \subscr{\mathcal L}{gl}\\
\subscr{\mathcal L}{lg} & \subscr{\mathcal L}{ll}
\end{smallmatrix}
\right] \in \mathbb R^{(2n+m) \times (2n+m)}$, where the first $n$
entries are associated with the generators and the last $n+m$ entries
correspond to the buses. The differential-algebraic model of the power
network is given by the linear continuous-time descriptor\,system
\begin{align}
\label{eq: power network descriptor system model}
E\dot x(t) = A x(t) + P(t),
\end{align}
where the state $x=[\delta^\transpose \; \omega^\transpose \;
\theta^\transpose]^\transpose \in \mathbb R^{2n + m}$ consists of the
generator rotor angles $\delta \in \mathbb R^{n}$, the frequencies $\omega
\in \mathbb R^{n}$, and the bus voltage angles $\theta \in \mathbb
R^{m}$. The input term $P(t)$ is due to {\em known} changes in mechanical
input power to the generators or real power demand at the
loads. Furthermore, the descriptor system matrices are
\begin{align}\label{a_matrix}
E=
\begin{bmatrix}
I & 0 & 0\\
0 & M & 0\\
0 & 0 & 0
\end{bmatrix}
,\,
A=-
\begin{bmatrix}
0 & -I & 0\\
\subscr{\mathcal L}{gg} & \subscr{D}{g} & \subscr{\mathcal L}{gl}\\
\subscr{\mathcal L}{lg} & 0 & \subscr{\mathcal L}{ll}
\end{bmatrix},
\end{align}
where $M$ (resp. $D_\textup{g}$) is the diagonal matrix of the
generators' inertias (resp. damping constants).
The dynamic and algebraic equations of the linear descriptor system
\eqref{eq: power network descriptor system model} are classically referred
to as the linearized swing equation and the DC power flow equation,
respectively. Notice that the initial condition of system \eqref{eq: power
network descriptor system model} needs to obey the algebraic constraint
$\subscr{\mathcal L}{lg} \delta(0) + \subscr{\mathcal L}{ll} \theta(0) =
P_{\theta}(0)$, where $P_{\theta}(0)$ is the vector containing the entries
$\{2n+1,\dots,2n+m\}$ of $P(0)$. Finally, we assume the parameters of the
power network descriptor model \eqref{eq: power network descriptor system
model} to be known, and we remark that they can be either directly
measured, or estimated through dynamic identification techniques, e.g., see
\cite{AC-JHC-AS:11}.
Throughout the paper, the assumption is made that a combination of the
state variables of the descriptor system \eqref{eq: power network
descriptor system model} is being continuously measured over
time. Let $C \in \mathbb R^{p \times n}$ be the output matrix and let
$y(t) = C x(t)$ denote the $p$-dimensional measurements
vector. Moreover, we allow for the presence of {\em unknown}
disturbances affecting the behavior of the plant \eqref{eq: power
network descriptor system model}, which, besides reflecting the
genuine failure of network components, can be the effect of a
cyber-physical attack against the network. We classify these
disturbances into \emph{state attacks}, if they show up in the
measurements vector after being integrated through the network
dynamics, and \emph{output attacks}, if they corrupt directly the
measurements vector.\footnote{Because of the linearity of \eqref{eq:
power network descriptor system model}, the known input $P(t)$ can
be neglected, since it does not affect the detectability of unknown
input attacks.}
The network dynamics in the presence of a cyber-physical attack can be
written as
\begin{align}
\label{eq: cyber_physical_fault}
\begin{split}
E \dot x(t)&=Ax(t) +
\underbrace{
\begin{bmatrix}
F & 0
\end{bmatrix}
}_{B}
\underbrace{
\begin{bmatrix}
f(t) \\ \ell(t)
\end{bmatrix}
}_{u(t)}
\,,\\
y(t)&=Cx(t) +
\underbrace{
\begin{bmatrix}
0 & L
\end{bmatrix}
}_{D}
\underbrace{
\begin{bmatrix}
f(t) \\ \ell(t)
\end{bmatrix}
}_{u(t)}
\,.\\
\end{split}
\end{align}
The input signals $f(t)$ and $\ell(t)$ are referred to as {\em state}
and {\em output attack modes}, respectively. The attack modes are
assumed to be unknown and piece-wise continuous functions of time of
dimension $2n+m$ and $p$, respectively, and they act through the full
rank matrices $F \in \mathbb R^{(2n+m) \times (2n+m)}$ and $L \in
\mathbb R^{p \times p}$.
For notational convenience, and without affecting generality, we
assume that each state and output variable can be independently
compromised by an attacker. Therefore, we let $F$ and $L$ be the
identity matrices of dimensions $2n+m$ and $p$.
The attack mode $u(t)$ depends upon the specific attack profile. In
the presence of $k \in \mathbb{N}_0$, $k \le 2n+m+p$, attackers
indexed by the {\em attack set} $K \subseteq \until{2n+m+p}$, the
corresponding (vector) attack mode $t \mapsto u_{K}(t) \in \mathbb
R^{2n+m+p}$ has exactly $k$ nonzero entries $u_{K,i}(t)$ for $i \in
K$.
Accordingly,\,\,the pair $(B_{K},D_{K})$ is called {\em attack signature}, where $B_K$
and $D_K$ are the submatrices of $B$ and $D$ with columns indexed by
$K$.
The model \eqref{eq: cyber_physical_fault} is very general, and it can
capture the occurrence of several concurrent contingencies in the
power network, which are caused either by components failure or
external attacks.\footnote{Genuine failures are a subcase of
intentional cyber-physical attacks.} For instance,
\begin{enumerate}
\item a change in the mechanical power input to generator $i$
(resp. in the real power demand of load $j$) is described by the
attack signature $(B_i,0)$ (resp. $(B_{2n+j},0)$), and a non-zero
attack mode $u_{n+i}(t)$ (resp. $u_{2n+j} (t)$);
\item a line outage occurring on the line $\{r,s\}$ is modeled by the
signature $([B_r \; B_s],[0 \; 0])$ and a non-zero mode $[u_r(t) \;
u_s(t)]^\transpose$ \cite{ES:04}; and
\item the failure of sensor $i$, or the corruption of the
$i$-th measurement by an attacker is captured by the signature
$(0,D_{2n+m+i})$ and a non-zero mode $u_{2n+m+i} (t)$.
\end{enumerate}
\subsection{Notions of detectability and identifiability for attack sets}
\label{sec:setup}
In this section we present the problem under investigation and we
recall some definitions. Observe that a cyber-physical attack may
remain undetected from the measurements if there exists a normal
operating condition of the network under which the output would be the
same as under the perturbation due to the attacker. Let $x(x_0,u,t)$
denote the network state trajectory generated from the initial state
$x_0$ under the attack signal $u(t)$, and let $y(x_0,u,t)$ be the
output sequence for the same initial condition and input. Throughout
the paper, let $T \subseteq \mathbb{R}_{\ge 0}$ denote the set of time
instants at which the presence of attacks against the network is
checked.
\begin{definition}[\bf Undetectable attack
set]\label{undetectable_input}
For the linear descriptor system \eqref{eq: cyber_physical_fault},
the attack set $K$ is \emph{undetectable} if there exist initial
conditions $x_1 , x_2 \in \mathbb{R}^{2 n + m}$, and an attack mode $u_K
(t)$ such that, for all $t \in T$, $y(x_1,u_K,t) =
y(x_2,0,t)$.
\end{definition}
\smallskip
A more general concern than detection is identifiability of attackers,
i.e., the possibility to distinguish from measurements between the
action of two distinct attacks.
\begin{definition}[\bf Unidentifiable attack set]\label{unidentifiable_input}
For the linear descriptor system \eqref{eq: cyber_physical_fault}, the
attack set $K$ is \emph{unidentifiable} if there exists an attack set
$R$, with $|R|\le|K|$ and $R\neq{K}$, initial conditions $x_K , x_R \in
\mathbb{R}^{2 n + m}$, and attack modes $u_K(t)$, $u_R (t)$ such that, for all
$t \in T$, $y(x_K,u_{K},t) = y(x_R,u_{R},t)$.
\end{definition}
\smallskip
Of course, an undetectable attack is also unidentifiable, since it
cannot be distinguished from the zero input. The converse does not
hold. The security problem we consider in this paper is as follows.
\noindent{\em Problem:}
{(\bf Attack detection and identification)}
For the linear descriptor system \eqref{eq: cyber_physical_fault},
design an attack detection and identification procedure.
Definitions~\ref{undetectable_input} and~\ref{unidentifiable_input} are
immediately applicable to arbitrary constrol systems subjects to external
attacks. Before proposing a solution to the Attack detection and
identification Problem, we motivate the use of a dynamic detection and
identification algorithm by characterizing the fundamental limitations of
static and dynamic procedures.
\section{Limitations of static and dynamic procedures for detection and
identification}\label{sec:static_dynamic}
The objective of this section is to show that some fundamental
limitations of a static detection procedure can be overcome by
exploiting the network dynamics. We start by deriving a reduced state
space model for a power network, which is convenient for illustration
and analysis purposes.
\subsection{Kron-reduced representation of a power network}
For the system \eqref{eq: cyber_physical_fault}, let $F = \left[
\begin{smallmatrix}
F_{\delta}^\transpose &
F_{\omega}^\transpose &
F_{\theta}^\transpose
\end{smallmatrix}
\right]^\transpose
$,
$L =
\left[
\begin{smallmatrix}
L_{\delta}^\transpose & L_{\omega}^\transpose &
L_{\theta}^\transpose
\end{smallmatrix}
\right]^\transpose $, and
$
C =
\left[
\begin{smallmatrix}
C_\delta & C_\omega & C_\theta
\end{smallmatrix}
\right]$, where the partitioning reflects the state $x =
[\delta^\transpose \; \omega^\transpose \;
\theta^\transpose]^\transpose$. Since the network Laplacian matrix is
irreducible (due to connectivity), the submatrix $\subscr{\mathcal L}{ll}$
in \eqref{a_matrix} is invertible and the bus voltage angles
$\theta(t)$ can be expressed via the generator rotor angles
$\delta(t)$ and the state attack mode $f(t)$ as
\begin{align}
\label{eq: elimination of bus voltages}
\theta (t) = - \subscr{\mathcal L}{ll}^{-1} \subscr{\mathcal L}{lg} \delta(t) -
\subscr{\mathcal L}{ll}^{-1} F_\theta f(t)
.
\end{align}
Hence, the descriptor system \eqref{eq: cyber_physical_fault} is of
index one \cite{ES:04}. The elimination of the algebraic variables
$\theta(t)$ in the descriptor system \eqref{eq: cyber_physical_fault}
leads to the state space system
\begin{align}
\label{eq: power network Kron-reduced model}
\begin{bmatrix}
\dot \delta \\ \dot \omega
\end{bmatrix}
=&\;
\underbrace{
\begin{bmatrix}
0 & I \\ - M^{-1} \bigl( \subscr{\mathcal L}{gg} - \subscr{\mathcal L}{gl} \subscr{\mathcal L}{ll}^{-1}
\subscr{\mathcal L}{lg} \bigr) & -M^{-1}\subscr{D}{g}
\end{bmatrix}
}_{\tilde A}
\begin{bmatrix}
\delta \\ \omega
\end{bmatrix}
\nonumber\\&\; +
\underbrace{
\begin{bmatrix}
F_\delta & 0\\
M^{-1} F_\omega - M^{-1} \subscr{\mathcal L}{gl} \subscr{\mathcal L}{ll}^{-1} F_{\theta} & 0
\end{bmatrix}
}_{\tilde B} u
\,,\\
y (t) =&\;
\underbrace{
\begin{bmatrix}
C_\delta - C_\theta \subscr{\mathcal L}{ll}^{-1} \subscr{\mathcal L}{lg} & C_\omega
\end{bmatrix}
}_{\tilde C}
\begin{bmatrix}
\delta \\ \omega
\end{bmatrix}
+
\underbrace{
\begin{bmatrix}
-C_\theta \subscr{\mathcal L}{ll}^{-1} F_\theta & L
\end{bmatrix}
}_{\tilde D} u. \nonumber
\end{align}
This reduction of the passive bus nodes is known as Kron reduction in
the literature on power networks and circuit theory
\cite{FD-FB:11d}. In what follows, we refer to \eqref{eq: power
network Kron-reduced model} as the {\em Kron-reduced
system}. Accordingly, for each attack set $K$, the attack signature
$(B_K,D_K)$ is mapped to the corresponding signature $(\tilde B_K,
\tilde D_K)$ in the Kron-reduced system through the transformation for
the matrices $B$ and $D$ described in \eqref{eq: power network
Kron-reduced model}. Clearly, for any state trajectory of the
Kron-reduced \eqref{eq: power network Kron-reduced model}, the
corresponding state trajectory of the (non-reduced) descriptor power
network model \eqref{eq: cyber_physical_fault} can recovered by
identity \eqref{eq: elimination of bus voltages}.
We point out the following subtle but important facts, which are
easily visible in the Kron-reduced system \eqref{eq: elimination of
bus voltages}. First, a state attack $F_{\theta} f(t)$ on the buses
affects directly the output $y(t)$. Second, for a connected bus
network, the lower block of $\tilde A$ is a fully populated Laplacian
matrix, and $\subscr{\mathcal L}{ll}^{-1}$ and $\subscr{\mathcal L}{gl}
\subscr{\mathcal L}{ll}^{-1}$ are both positive matrices
\cite{FD-FB:11d}. As one consequence, an attack on a single bus
affects the {\em entire} network and not only the locally attacked
node or its vicinity. Third and finally, the mapping from the input
signal $u(t)$ and the initial condition $x(0)$ (subject to the
constraint \eqref{eq: elimination of bus voltages} evaluated at $t =
0$) to the output signal $y(t)$ of the descriptor system \eqref{eq:
cyber_physical_fault} coincides with the corresponding input and
initial state to output map of the associated Kron-reduced system
\eqref{eq: power network Kron-reduced model}. Hence, the definition of
identifiability (resp. detectability) of an attack set is analogous
for the Kron-reduced system \eqref{eq: power network Kron-reduced
model}, and we can directly state the following lemma.
\begin{lemma}{\bf(Equivalence of detectability and identifiability under Kron reduction):}
\label{lemma:equivalence}
For the power network descriptor system \eqref{eq:
cyber_physical_fault}, the attack set $K$ is identifiable
(resp. detectable) if and only if it is identifiable
(resp. detectable) for the associated Kron-reduced system \eqref{eq:
power network Kron-reduced model}.
\end{lemma}
Following Lemma \ref{lemma:equivalence}, we study detectability and
identifiability of attacks against the power network descriptor model
\eqref{eq: cyber_physical_fault} by analyzing the associated
Kron-reduced system \eqref{eq: power network Kron-reduced model}.
\subsection{Fundamental limitations of a Static Detector}
By \emph{Static Detector}, or, with the terminology of
\cite{AA-AGS:04}, \emph{Bad Data Detector}, we denote an algorithm
that uses the network measurements to check for the presence of
attacks at some predefined instants of time, and without exploiting
any relation between measurements taken at different time instants. By
Definition \ref{undetectable_input}, an attack is undetectable by a
Static Detector if and only if, for all time instances $t$ in a
countable set $T$, there exists a vector $\xi(t)$ such that $y(t) =
\tilde C \xi(t)$. Without loss of generality, we set $T = \mathbb
N$. Loosely speaking, the Static Detector checks whether, at a
particular time instance $t \in \mathbb N$, the measured data is
consistent with the measurement equation, for example, the power flow
equation at a bus. Notice that our definition of Static Detector is
compatible with \cite{YL-MKR-PN:09}, where an attack is detected if
and only if the residual $r(t) = y(t) - \tilde C
[\hat\delta(t)^{\transpose} \;
\hat\omega(t)^{\transpose}]^{\transpose}$ is nonzero for some $t \in
\mathbb N$, where $[\hat\delta(t)^{\transpose} \;
\hat\omega(t)^{\transpose}]^{\transpose} = \tilde{C}^{\dag}
y(t)$.
If $r(t) \neq 0$, then a malfunction is detected, and it is undetected
otherwise.\footnote{Similar conclusion can be drawn for the case of
noisy measurements.}
\begin{theorem}
\emph{(\bf Static detectability of cyber-physical attacks)}
\label{Theorem: Static detectability of cyber-physical attacks}
For the power network descriptor system \eqref{eq:
cyber_physical_fault} and an attack set $K$, the following two statements are
equivalent:
\begin{enumerate}
\item the attack set $K$ is undetectable by a Static Detector;
\item there exists an attack mode $u_K (t)$ such that, for some
$\delta (t)$ and $\omega(t)$, at every $t \in \mathbb N$ it holds
\begin{align}\label{eq:cond2_static}
\tilde C
\begin{bmatrix}
\delta(t)\\
\omega(t)
\end{bmatrix} + \tilde D u_K(t) = 0\,,
\end{align}
where $\tilde C$ and $\tilde D$ are as in \eqref{eq: power network
Kron-reduced model}.
\end{enumerate}
Moreover, there exists an attack set $K$ undetectable by a Static Detector if and only if there exist $x \in \mathbb{R}^{2n}$ and $g \in \mathbb{R}^{|K|}$ such that $\tilde C x + \tilde D_K g =
0$.
\end{theorem}
\smallskip
Before presenting a proof of the above theorem, we highlight that a
necessary and sufficient condition for the equation
\eqref{eq:cond2_static} to be satisfied is that $L \ell(t) \in \operatorname{Im}
(C)$ at all times $t \in \mathbb N$, where $\ell(t)$ is the output
attack mode, i.e., the vector of the last $p$ components of
$u_K(t)$. Hence, statement (ii) in Theorem \ref{Theorem: Static
detectability of cyber-physical attacks} implies that {\em no} state
attack can be detected by a static detection procedure, and that an
undetectable output attack exists if and only if $\operatorname{Im} (D_{K}) \cap
\operatorname{Im} (C) \neq \{ 0 \}$.
\begin{pfof}{Theorem \ref{Theorem: Static detectability of
cyber-physical attacks}}
%
As previously discussed, the attack $K$ is undetectable by a Static
Detector if and only if for each $t \in \mathbb N$ there exists
$\delta(t)$, $\omega(t)$, and $u_{K}(t)$ such that
\begin{align*}
r(t) = y(t) - \tilde C
\tilde C^{\dagger}
y(t) = (I-\tilde C \tilde C^{\dagger}) \left( \tilde C
\begin{bmatrix}
\delta(t) \\ \omega(t)
\end{bmatrix}
+ \tilde D u_K(t) \right)
\end{align*}
vanishes. Consequently, $r(t) = (I - \tilde C \tilde C^{\dagger})
\tilde D u_K(t)$, and the attack set $K$ is undetectable
if and only if $\tilde Du_{K}(t) \in \operatorname{Im}(\tilde C)$, which is
equivalent to statement (ii). The last necessary and sufficient condition
in the theorem follows from (ii).
\end{pfof}
We now focus on the static identification problem. Following Definition
\ref{unidentifiable_input}, the following result can be asserted.
\begin{theorem}\emph{(\bf Static identification of cyber-physical attacks)}
\label{Theorem: Static identifiability of cyber-physical attacks}
For the power network descriptor system \eqref{eq: cyber_physical_fault}
and an attack set $K$, the following two statements are equivalent:
\begin{enumerate}
\item the attack set $K$ is unidentifiable by a Static Detector;
\item there exists an attack set $R$, with $|R|\le|K|$ and $R\neq{K}$,
and attack modes $u_K (t)$, $u_R (t)$, such that, for some $\delta (t)$
and $\omega(t)$, at every $t \in \mathbb N$, it holds
\begin{align*}
\tilde C
\begin{bmatrix}
\delta(t)\\
\omega(t)
\end{bmatrix} + \tilde D \left( u_K(t) + u_R(t) \right)= 0\,,
\end{align*}
where $\tilde C$ and $\tilde D$ are as in \eqref{eq: power network
Kron-reduced model}.
\end{enumerate}
Moreover, there exists an attack set $K$ unidentifiable by a Static
Detector if and only if there exists an attack set $\bar K$, $|\bar
K| \le 2 |K|$, which is undetectable by a Static Detector.
\end{theorem}
\smallskip
Similar to the fundamental limitations of static detectability in
Theorem \ref{Theorem: Static detectability of cyber-physical attacks},
Theorem \ref{Theorem: Static identifiability of cyber-physical
attacks} implies that, for instance, state attacks cannot be
identified and that an undetectable output attack exists if and only
if $\operatorname{Im} (D_{\bar K}) \cap \operatorname{Im} (C) \neq \{ 0 \}$.
\begin{pfof}{Theorem \ref{Theorem: Static identifiability of cyber-physical attacks}}
Because of the linearity of the system \eqref{eq:
cyber_physical_fault}, the unidentifiability condition in
Definition \ref{unidentifiable_input} is equivalent to
$y(x_K-x_R,u_{K}-u_{R},t) = 0$, for some initial condition $x_K$,
$x_R$, and attack mode $u_K (t)$, $u_R(t)$. The equivalence between
statements (i) and (ii) follow. The last statement follows from
Theorem \ref{Theorem: Static detectability of cyber-physical
attacks}.
\end{pfof}
\subsection{Fundamental limitations of a Dynamic Detector}
In the following we refer to a security system having access to the
{\em continuous} time measurements signal $y(t)$, $t \in \mathbb{R}_{\ge
0}$, as a {\em Dynamic Detector}. As opposed to a Static Detector, a
Dynamic Detector checks for the presence of attacks at every instant
of time $t \in \mathbb{R}_{\geq 0}$. By Definition
\ref{undetectable_input}, an attack is undetectable by a Dynamic
Detector if and only if there exists a network initial state $\xi(0)
\in \mathbb{R}^{2n}$ such that $y(t) = \tilde C e^{\tilde A t}\xi(0)$ for
all time instances $t\in \mathbb{R}_{\ge 0}$. Intuitively, a Dynamic
Detector is harder to mislead than a Static Detector.
\begin{theorem}
\emph{(\bf Dynamic detectability of cyber-physical attacks)}
\label{Theorem: Dynamic detectability of cyber-physical attacks}
For the power network descriptor system \eqref{eq:
cyber_physical_fault} and an attack set $K$, the following two
statements are equivalent:
\begin{enumerate}
\item the attack set $K$ is undetectable by a Dynamic Detector;
\item there exists an attack mode $u_K (t)$ such that, for some
$\delta (0)$ and $\omega(0)$, at every $t \in \mathbb{R}_{\ge 0}$, it
holds
\begin{align*}
\tilde C
e^{\tilde A t}
\begin{bmatrix}
\delta(0)\\
\omega(0)
\end{bmatrix}
\! + \tilde C \! \int_0^t \! e^{\tilde A (t-\tau)} \tilde B u_K (\tau)
d\tau
=- \tilde D u_K(t),
\end{align*}
where $\tilde A$, $\tilde B$, $\tilde C$, and $\tilde D$ are as in
\eqref{eq: power network Kron-reduced model}.
\end{enumerate}
Moreover, there exists an attack set $K$ undetectable by a Dynamic
Detector if and only if there exist $s \in \mathbb{C}$, $g \in
\mathbb{R}^{|K|}$, and $x \in \mathbb{R}^{2n}$, $x \neq 0$, such that
$(sI-\tilde A)x - \tilde B_K g = 0$ and $\tilde C x + \tilde D_K g =
0$.
\end{theorem}
\smallskip
Before proving Theorem \ref{Theorem: Dynamic detectability of
cyber-physical attacks}, some comments are in order. First, state
attacks {\em can be detected} in the dynamic case. Second, an attacker
needs to inject a signal which is consistent with the network dynamics
at every instant of time to mislead a Dynamic Detector. Hence, as
opposed to the static case, the condition $L \ell(t) \in \operatorname{Im} (C)$
needs to be satisfied for every $t \in \mathbb{R}_{\ge 0}$, and it is only
necessary for the undetectability of an output attack. Indeed, for
instance, state attacks can be detected even though they automatically
satisfy the condition $0 = L \ell(t) \in \operatorname{Im} (C)$. Third and
finally, according to the last statement of Theorem \ref{Theorem:
Dynamic detectability of cyber-physical attacks}, the existence of
invariant zeros for the Kron-reduced system $(\tilde A,\tilde B_K,
\tilde C, \tilde D_K)$ is equivalent to the existence of an
undetectable attack mode $u_K (t)$.\footnote{For the system $(\tilde
A,\tilde B_K, \tilde C, \tilde D_K)$, the value $s \in \mathbb{C}$ is
an invariant zero if there exists $x \in \mathbb{R}^{2n}$, with $x \neq
0$, $g \in \mathbb{R}^{|K|}$, such that $(sI-\tilde A)x - \tilde B_K g =
0$ and $\tilde C x + \tilde D_K g = 0$. For a linear dynamical
system, the existence of invariant zeros is equivalent to the
existence of zero dynamics \cite{HLT-AS-MH:01}.}
%
As a consequence, for the absence of undetectable cyber-physical
attacks, a dynamic detector performs better that a static detector,
while requiring, possibly, fewer measurements. A related example is in
Section \ref{sec:example}.
\begin{pfof}{Theorem \ref{Theorem: Dynamic detectability of
cyber-physical attacks}}
By Definition \ref{undetectable_input} and linearity of the system
\eqref{eq: power network Kron-reduced model}, the attack mode $u_K(t)$
is undetectable by a Dynamic Detector if and only if there exists
$[\delta(0)^{\transpose} \; \omega(0)^{\transpose}]^{\transpose}$ such
that $y([\delta(0)^{\transpose} \;
\omega(0)^{\transpose}]^{\transpose},u_K,t) = 0$ for all $t \in
\mathbb R_{\geq 0}$. Hence, statements (i) and (ii) are equivalent.
Following condition (ii) in Theorem \ref{Theorem: Dynamic
detectability of cyber-physical attacks}, an attack $u_K(t)$ may
remain undetected to a Dynamic Detector if and only if
$u_K(t)$ is an input-zero for some initial condition.
\end{pfof}
We now focus on the identification problem.
\begin{theorem}
\emph{(\bf Dynamic identifiability of cyber-physical attacks)}
\label{Theorem: Dynamic identifiability of cyber-physical attacks}
For the power network descriptor system \eqref{eq:
cyber_physical_fault}, the following two statements are equivalent:
\begin{enumerate}
\item the attack set $K$ is unidentifiable by a Dynamic Detector;
\item there exists an attack set $R$, with $|R| \le |K|$ and $R\neq{K}$ if
$|R| = |K|$, and attack modes $u_K(t)$, $u_R(t)$, such that, for some
$\delta (0)$ and $\omega(0)$, at every $t \in \mathbb{R}_{\ge 0}$, it holds
\begin{align*}
&\tilde C
e^{\tilde A t}
\begin{bmatrix}
\delta(0)\\
\omega(0)
\end{bmatrix}
\! + \tilde C \! \int_0^t \! e^{\tilde A (t-\tau)} \tilde B \left( u_K
(\tau) + u_R(\tau) \right) d\tau
\\
&= - \tilde D \left(u_K(t) + u_R(t) \right)
\end{align*}
where $\tilde A$, $\tilde B$, $\tilde C$, and $\tilde D$ are as in
\eqref{eq: power network Kron-reduced model}.
\end{enumerate}
Moreover, there exists an attack set $K$ unidentifiable by a Dynamic
Detector if and only if there exists an attack set $\bar K$,
$|\bar K| \le 2 |K|$, which is unidentifiable by a Dynamic Detector.
\end{theorem}
\begin{pf}
Notice that, because of the linearity of the system \eqref{eq:
cyber_physical_fault}, the unidentifiability condition in
Definition \ref{unidentifiable_input} is equivalent to the condition
$y(x_K-x_R,u_{K}-u_{R},t) = 0$, for some initial condition $x_K$,
$x_R$, and attack mode $u_K (t)$, $u_R$. The equivalence
between statements (i) and (ii) follows.
\end{pf}
In other words, the existence of an unidentifiable attack set of
cardinality $k$ is equivalent to the existence of invariant zeros for
the system $(\tilde A,\tilde B_{\bar K}, \tilde C, \tilde D_{\bar
K})$, for some attack set $\bar K$ with $|\bar K| \le 2k$. A careful
reader may notice that condition (ii) in Theorem 3.4 is hard to verify
because of its combinatorial complexity:
one needs to certify the absence of invariant zeros for all possible
distinct pairs of $|K|$-dimensional attack sets. Then, a conservative
verification of condition (ii) requires $\binom{2n+m+p}{2|K|}$
tests. In \cite{FP-AB-FB:10u} we partially address this complexity
problem by presenting an intuitive and easy to check graph-theoretic
condition for a given network topology and generic system parameters.
\begin{remark}(\emph{\bf Stealth, false-data injection, and replay
attacks}) The following prototypical attacks can be modeled and
analyzed through our theoretical framework:
\begin{enumerate}
\item stealth attacks, as defined in \cite{DG-HS:10}, correspond to
output attacks satisfying $D_K u_K(t) \in \operatorname{Im}(C)$;
\item (dynamic) false-data injection attacks, as defined in
\cite{YM-BS:10b}, are output attacks rendering the unstable modes
(if any) of the system unobservable. These unobservable modes are
included in the invariant zeros set; and
\item replay attacks, as defined in \cite{YM-BS:10a}, are state and
output attacks satisfying $\operatorname{Im}(C) \subseteq \operatorname{Im}(D_K)$, $B_K
\neq 0$. The resulting system may have an infinite number of
invariant zeros: if the attacker knows the system model, then it
can cast very powerful undetectable attacks.
\end{enumerate}
In \cite{YM-BS:10a}, a monitoring signal (unknown to the attacker)
is injected into the system to detect replay attacks. It can be
shown that, if the attacker knows the system model, and if the
attack signal enters additively as in \eqref{eq:
cyber_physical_fault}, then the attacker can design undetectable
attacks without knowing the monitoring signal. Therefore,
the fundamental limitations presented in Section
\ref{sec:static_dynamic} are also valid for \emph{active} detectors,
which are allowed to inject monitoring signals to reveal
attacks.\oprocend
\end{remark}
\section{Design of dynamic detection and identification
procedures}\label{sec:dynamic_detection}
\subsection{Detection of attacks}
We start by considering the attack detection problem, whose
solvability condition is in Theorem \ref{Theorem: Dynamic
detectability of cyber-physical attacks}.
We propose the following residual filter to detect cyber-physical
attacks.
\begin{theorem}[\bf Attack detection filter]\label{Proposition:
Attack detection filter}
Consider the power network descriptor system \eqref{eq:
cyber_physical_fault} and the associated Kron-reduced system
\eqref{eq: power network Kron-reduced model}. Assume that the attack
set is detectable and that the network initial state $x(0)$ is
known. Consider the {\em detection filter} \begin{align}
\begin{split}
\dot w(t) &= (\tilde A + G \tilde C) w(t) - Gy(t) , \\
r(t) &= \tilde C w(t) - y(t) ,
\end{split}
\label{eq: detection filter}
\end{align}
where $w(0) = x(0)$, and $G \in \mathbb R^{2n \times p}$ is such that
$\tilde A + G \tilde C$ is a Hurwitz matrix. Then $r (t) = 0$ at all
times $t \in \mathbb{R}_{\ge 0}$ if and only if $u(t) = 0$ at all times $t
\in \mathbb{R}_{\ge 0}$.
\end{theorem}
\begin{pf}
Consider the error $e(t) = w(t)-x(t)$ between the states of the
filter \eqref{eq: detection filter} and the Kron-reduced system
\eqref{eq: power network Kron-reduced model}. The error dynamics
with output $r(t)$ are then
\begin{align}
\begin{split}
\dot e(t) &= (\tilde A + G \tilde C) e(t) - (\tilde B + G\tilde D )u(t) , \\
r(t) &= \tilde C e(t) - \tilde D u(t) ,
\end{split}
\label{eq: detection filter - error dynamics}
\end{align}
where $e(0) = 0$. Clearly, if the error system \eqref{eq: detection
filter - error dynamics} has no invariant zeros, then $r(t)=0$ for all
$t \in \mathbb{R}_{\ge 0}$ if and only if $u(t) = 0$ for all $t \in \mathbb{R}_{\ge
0}$ and the claimed statement is true. The error system \eqref{eq:
detection filter - error dynamics} has no invariant zeros if and only
if there exists no triple $(s,\bar w,g) \in \mathbb C \times \mathbb
R^{2n} \times \mathbb R^{p}$ satisfying
\begin{equation}
\begin{bmatrix}
sI - (\tilde A+G \tilde C) & \tilde B + G \tilde D \\
\tilde C & -\tilde D
\end{bmatrix}
\begin{bmatrix}
\bar w \\ g
\end{bmatrix}
=
\begin{bmatrix}
0 \\ 0
\end{bmatrix}
\label{eq: pencil for error system}
\,.
\end{equation}
The second equation of \eqref{eq: pencil for error system} yields
$\tilde C x = \tilde D g$. Thus, by substituting $\tilde C x$ by
$\tilde D g$ in the first equation of \eqref{eq: detection filter -
error dynamics}, the set of equations \eqref{eq: pencil for error
system} can be equivalently written as
\begin{equation*}
\begin{bmatrix}
sI - \tilde A & \tilde B \\
\tilde C & -\tilde D
\end{bmatrix}
\begin{bmatrix}
\bar w \\ g
\end{bmatrix}
=
\begin{bmatrix}
0 \\ 0
\end{bmatrix}
\,.
\end{equation*}
Finally, note that the solution $(s,-\bar w,g)$ to the above set of
equations yields an invariant zero, zero state, and zero input for
the Kron-reduced system \eqref{eq: power network Kron-reduced
model}. By the detectability assumption, the Kron-reduced system
\eqref{eq: power network Kron-reduced model} has no zero dynamics. We
conclude that the error system \eqref{eq: detection filter - error
dynamics} has no zero dynamics, and the statement is true.
\end{pf}
In summary, the implementation of the residual filter~\eqref{eq: detection
filter} guarantees the detection of any detectable attack set.
\subsection{Identification of attacks}
We now focus on the attack identification problem, whose solvability
condition is in Theorem \ref{Theorem: Dynamic identifiability of
cyber-physical attacks}. Unlike the detection case, the
identification of the attack set $K$ requires a combinatorial
procedure, since, a priori, $K$ is one of the $\binom{2n+m+p}{|K|}$
possible attack sets. As key component of our identification
procedure, we propose a residual filter to determine whether a
predefined set coincides with the attack set.
We next introduce in a coordinate-free geometric way the key elements
of this residual filter based on the notion of condition-invariant
subspaces \cite{HLT-AS-MH:01}. Let $K$ be a $k$-dimensional attack
set, and let $\tilde B_K$, $\tilde D_K$ be as defined right after the
Kron reduced model~\eqref{eq: power network Kron-reduced model}. Let
$[V_{K}^\transpose \; Q_{K}^\transpose]^\transpose \in \mathbb R^{p
\times p}$ be an orthonormal matrix such that
\begin{align*}
V_{K} = \operatorname{Basis}(\operatorname{Im}(\tilde D_K)), \text{ and } Q_{K} =
\operatorname{Basis}(\operatorname{Im}(\tilde D_K)^\perp),
\end{align*}
and let
\begin{align}
\label{eq: Bz Bk}
B_{Z} = \tilde B_{K} (V_{K} \tilde D_K)^\dag, \text{ and }\bar B_{K}
= \tilde B_{K}(I-D_{K}D_{K}^{\dag}).
\end{align}
Define the subspace $\mathcal{S}^* \subseteq \mathbb{R}^{2n}$ to be the smallest
\mbox{$\bigl( \tilde A - \tilde B_{K} (V_K \tilde D_K)^\dag V_K \tilde
C \,,\, \operatorname{Ker}(Q_{K} \tilde C) \bigr)$-conditioned invariant} subspace
containing $\operatorname{Im}(\bar B_K)$, and let $J_K$ be an output injection
matrix such that
\begin{align}
\label{eq: S star}
(\tilde A - \tilde B_{K} (V_K \tilde D_K)^\dag V_K \tilde C +J_K
Q_{K} \tilde C) \mathcal{S}^* \subseteq \mathcal{S}^*.
\end{align}
Let $P_{K}$ be an orthonormal projection matrix onto the quotient
space $\mathbb{R}^{2n} \setminus \mathcal{S}^*$, and let
\begin{align}
\label{eq: Ak}
A_{K} = P_{K} (\tilde A - \tilde B_{K} (V_K \tilde D_K)^\dag V_K
\tilde C + J_{K} Q_{K} \tilde C)P_{K}^{\transpose}.
\end{align}
Finally, let $H_K$ and the unique $M_K$ be such that
\begin{align}
\label{eq:definition H and M}
\begin{split}
\operatorname{Ker}(H_K Q \tilde C) &= \mathcal{S}^* + \operatorname{Ker}(Q \tilde C),\text{ and }\\
H_K Q \tilde C &= M_K P_{K}.
\end{split}
\end{align}
\begin{theorem}[\bf Attack identification filter]
\label{Proposition: Attack identification filter}
Consider the power network descriptor system \eqref{eq:
cyber_physical_fault} and the associated Kron-reduced system
\eqref{eq: power network Kron-reduced model}. Assume that the attack
set $K$ is identifiable and that the network initial state is known.
Consider the
\emph{identification filter}
\begin{align}
\label{eq: identification filter}
\begin{split}
\dot w_{K} (t)
=&
(A_{K} + G_{K} M_{K}) w_{K} (t)
\\ &
+ \bigl( P_K B_{Z} V_{K} - (P_{K} J_{K} + G_{K} H_{K}) Q \bigr) y (t)
,\\
r_{K} (t)
=&
M_{K} w_{K} (t) - H_{K} Q y(t),
\end{split}
\end{align}
where $w_{K} (0) = P_{K} x(0)$, and $G_K \in \mathbb R^{2n \times
p}$
is such that $A_K + G_K M_K$ is a Hurwitz matrix. Then $r_{K} (t) =
0$ at all times $t \in \mathbb{R}_{\ge 0}$ if and only if $K$ equals the
attack set.
\end{theorem}
\smallskip
Note that the residual $r_K(t)$ is identically zero if the attack
set coincides with $K$, even if the attack input is nonzero.
\begin{pfof}{Theorem \ref{Proposition: Attack identification filter}}
Let $R$ be an attack set with $|R|\le|K|$ and $R \neq K$.
With the output transformation
$[z_1,z_2] = [V_{K}y,Q_{K}y]$, the Kron-reduced system \eqref{eq:
power network Kron-reduced model} becomes
\begin{align}
\label{eq: power network Kron-reduced model under attack - 1}
\begin{split}
\dot x (t) &= \tilde A x(t) +
\begin{bmatrix}
\tilde B_K & \tilde B_{R}
\end{bmatrix}
\begin{bmatrix}
u_K(t) \\ u_{R} (t)
\end{bmatrix}
,\\
z_1(t) &= V_K \tilde C x (t) + V_K \tilde D_K u_K(t) + V_K \tilde D_{R} u_{R}(t),\\
z_2(t) &= Q_K \tilde C x (t) + Q_K \tilde D_{R} u_{R}(t).
\end{split}
\end{align}
Note that the attack set $K$ affects only the output
$z_{1}(t)$. The output equation for $z_{1}(t)$ can be solved for
$u_{K}(t)$ as
\begin{multline*}
u_{K}(t)
=
(V_{K} \tilde D_{K})^{\dag} (z_{1}(t) - V_{K} \tilde C x(t))
\\ - (V_{K} \tilde D_{K})^{\dag} V_{K} \tilde D_{R} u_{R}(t) + \subscr{u}{hom}(t)
\,,
\end{multline*}
where $\subscr{u}{hom}(t) \in \operatorname{Ker}(V_{k} \tilde D_{K}) = \operatorname{Ker}( \tilde
D_{K})$ and $u_{R}(t)$ are unknown signals, while $z_{1}(t)$ is
known. The Kron-reduced system \eqref{eq: power network Kron-reduced
model under attack - 1} can equivalently be written with unknown
inputs $u_{K}(t)$ and $u_{R}(t)$, known input $z_{1}(t)$, and output
$z_{2}(t)$ as
\begin{align}
\label{eq: power network Kron-reduced model under attack - 2}
\begin{split}
\dot x (t) =&
(\tilde A - \tilde B_{K} (V_K \tilde D_K)^\dag V_K \tilde C) x(t)
\\&+
\begin{bmatrix}
B_Z & \bar B_{K} & \bar B_{R}
\end{bmatrix}
\begin{bmatrix}
z_1(t) \\ u_K(t) \\ u_{R} (t)
\end{bmatrix}
,\\
z_2(t) =& Q_K \tilde C x (t) + Q_K \tilde D_{R} u_{R}(t) ,
\end{split}
\end{align}
where $B_Z$ and $\bar B_K$ are as in \eqref{eq: Bz Bk}, and
\begin{align*}
\bar B_{R} = \tilde B_{R} - \tilde B_{K} (V_{K} \tilde
D_{K})^{\dag} V_{K} \tilde D_{R}.
\end{align*}
Let $\mathcal{S}^*$ and $J_K$ be as in \eqref{eq: S star},
and consider the orthonormal change of coordinates given by $T_{K} =
[W_K^\transpose \; P_K^\transpose] \in \mathbb R^{2n \times 2n}$,
where $W_K^\transpose$ is a basis of $\mathcal{S}^*$, $P_K$ is a projection
matrix onto the quotient space $\mathbb{R}^{2n} \setminus \mathcal{S}^*$, and
$T_{K}^{-1} = T_{K}^\transpose$. In the new coordinates
$[\xi_1,\xi_2] = [W_{K}x,P_{K}x]$, system \eqref{eq: power network
Kron-reduced model under attack - 2} reads as
\begin{align}
\label{eq: power network Kron-reduced model under attack - 3}
\begin{split}
\begin{bmatrix}
\dot \xi_1 \\ \dot \xi_2
\end{bmatrix}
&\!=\!
\begin{bmatrix}
\hat A_{11} & \hat A_{12}\\
0 & \hat A_{22}
\end{bmatrix}\!
\begin{bmatrix}
\xi_1 \\ \xi_2
\end{bmatrix}
\!+\!\!
\begin{bmatrix}
\hat B_{11} & \hat B_{12} & \hat B_{13}\\
\hat B_{21} & 0 & \hat B_{23}
\end{bmatrix}\!
\begin{bmatrix}
z_1 \\u_K \\ u_{R}
\end{bmatrix}
\!,\\
z_{2}(t)
&\!=\!
\begin{bmatrix}
\hat C_{1} & \hat C_{2}
\end{bmatrix}\!
\begin{bmatrix}
\xi_1 \\ \xi_2
\end{bmatrix}
\!+\!
\hat D u_{R}(t)
\!.
\end{split}
\end{align}
The zero pattern in the system and input matrix of \eqref{eq: power
network Kron-reduced model under attack - 3} arises due to the
invariance properties of $\mathcal{S}^*$, which contains $\operatorname{Im}(\bar B_K)$.
%
For the system \eqref{eq: power network Kron-reduced model
under attack - 3} we propose the filter
\begin{align}
\dot w_{K} (t) &\!=\! (\hat A_{22} + G_{K} M_{K})
w_{K} (t) + \hat B_{21} z_1(t) - G_{K} H_{K} z_2 (t),
\nonumber\\
r_{K} (t) &\!=\! M_{K} w_{K} (t) - H_{K} z_2(t),
\label{eq: identification filter - nice coordinates}
\end{align}
where $G_K$ is chosen such that $\hat A_{22} + G_{K} M_{K}$ is a
Hurwitz matrix. Let $H_K$ and $M_K$ be as in \eqref{eq:definition H
and M}, which, in these coordinates, coincides with $H_K \hat C_1
= 0$ and $H_K \hat C_2 = M_K$. Define the filter error $e(t) =
w_K(t) - \xi_2(t)$, then the residual filter \eqref{eq:
identification filter - nice coordinates} written in error
coordinates is
\begin{align*}
\dot e(t) =& (\hat A_{22} + G_{K} M_{K}) w_{K} (t) -\hat A_{22} \xi_2 (t)
- \hat B_{23} u_{R} (t)\\ &- G_{K} H_{K} \big([ \hat C_{1} \; \hat C_{2}]
\xi(t) + \hat D u_{R}(t) \big) \\
=& (\hat A_{22} + G_{K} M_{K}) e(t) - (\hat B_{23} + G_{K} H_{K} \hat D) u_{R}(t) \\
r_K(t) =& M_K e(t) - H_K \hat D u_{R}(t).
\end{align*}
It can be shown that $(\hat A_{22} + G_{K} M_{K}, - (\hat B_{23} +
G_{K} H_{K} \hat D), M_K, - H_K \hat D )$ has no zero dynamics, so
that the residual $r_K (t)$ is not affected by $K$, and every
nonzero signal $u_R (t)$ is detectable from $r_K (t)$. Consequently,
$r_K(t)$ is identically zero if and only if $K$ is the attack
set. Finally, in original coordinates, the filter \eqref{eq:
identification filter - nice coordinates} takes the form
\eqref{eq: identification filter}.
\end{pfof}
For an attack set $K$, we refer to the signal $r_K (t)$ in the
filter~\eqref{eq: identification filter} as the residual associated
with $K$. A corollary result of Theorem \ref{Proposition: Attack
identification filter} is that, if only an upper bound on the
cardinality of the attack set is known, then the residual $r_K (t)$ is
nonzero if and only if the attack set is contained in $K$. We now
summarize our identification procedure, which assumes the knowledge of
the network initial condition and of an upper bound $k$ on the
cardinality of the attack set $K$:
\begin{enumerate}
\item design an identification filter for each possible subset of
$\until{2n+m+p}$ of cardinality $k$;
\item monitor the power network by running each identification filter;
\item the attack set $K$ coincides with the intersection of the attack
sets $Z$ whose residual $r_Z (t)$ is identically zero.
\end{enumerate}
\begin{remark}\emph{(\bf Detection and identification filters for unknown initial condition)}
\label{Remark: detection filter for unknown initial condition}
If the network initial state is not available, then an arbitrary
initial state $w(0) \in \mathbb R^{2n}$ can be chosen. Consequently,
the filters performance becomes asymptotic, and some attacks may
remain undetected or unidentified. For instance, if the eigenvalues of
the detection filter matrix have been assigned to have real part
smaller than $c < 0$, with $c \in \mathbb{R}$, then, in the absence of
attacks, the residual $r(t)$ exponentially converges to zero with rate
less than $c$. Hence, only inputs $u(t)$ that vanish faster or equal
than $e^{-ct}$ can remain undetected by the filter \eqref{eq:
detection filter}. Alternatively, the detection filter can be
modified so as to converge in a predefined finite time
\cite{AVM-HTT:94}. In this case, every attack signal is detectable
after a finite transient.
\oprocend
\end{remark}
\begin{remark}\emph{(\bf Detection and identification in the presence
of process and measurement noise)}
The detection and identification filters here presented are a
generalization to dynamical systems with direct input to output
feedthrough of the devices presented in
\cite{MAM-GCV-ASW-CM:89}. Additionally, our design guarantees the
absence of invariant zeros in the residual system, so that
\emph{every} attack signal affect the corresponding
residual. Finally, if the network dynamics are affected by noise,
then an optimal noise rejection in the residual system can be
obtained by choosing the matrix $G$ in \eqref{eq: detection filter}
and $G_K$ in \eqref{eq: identification filter} as the Kalman gain
according to the noise statistics. \oprocend
\end{remark}
\section{A numerical study}\label{sec:example}
The effectiveness of our theoretic developments is here demonstrated
for the IEEE 14 bus system reported in Fig. \ref{fig:ieee14}.
Let the IEEE 14 bus power network be modeled as a descriptor model of
the form \eqref{eq: cyber_physical_fault}, where the network matrix
$A$ is as in \cite{rdz-cem-dg:11}. Following \cite{YL-MKR-PN:09}, the
measurement matrix $C$ consists of the real power injections at all
buses, of the real power flows of all branches, and of one rotor angle
(or one bus angle). We assume that an attacker can independently
compromise every measurement, except for the one referring to the
rotor angle, and that it does not inject state attacks.
Let $k \in \mathbb{N}$ be the cardinality of the attack set. From
\cite{YL-MKR-PN:09} it is known that, for a Static Detector, an
undetectable attack exists if $k \ge 4$. In other words, due to the
sparsity pattern of $C$, there exists a signal $u_K(t)$, with (the
same) four nonzero entries at all times, such that $D u_K(t) \in
\operatorname{Im} (C)$ at all times. By Theorem \ref{Theorem: Static
detectability of cyber-physical attacks} the attack set $K$ remains
undetected by a Static Detector through the attack mode $u_K (t)$. On
the other hand, following Theorem \ref{Theorem: Dynamic detectability
of cyber-physical attacks}, it can be verified that, for the same
output matrix $C$, and independent of the value of $k$, there exists
\emph{no} undetectable (output) attack set.
\section{Conclusion}\label{sec:conclusion}
For a power network modeled via a linear time-invariant descriptor
system, we have analyzed the fundamental limitations of static and
dynamic attack detection and identification procedures. We have
rigorously shown that a dynamic detection and identification method
exploits the network dynamics and outperforms the static counterpart,
while requiring, possibly, fewer measurements. Additionally, we have
described a provably correct attack detection and identification
procedure based on dynamic residuals filters, and we have illustrated
its effectiveness through an example of cyber-physical attacks against
the IEEE 14 bus system.
\bibliographystyle{ieeeconf}
|
\section{Introduction}
The central force problem in
non-relativistic classical mechanics is one of the most useful topics
in physics. Closely linked with the central force problem is the
Keplerian orbit theory which is a cornerstone for understanding
planetary motions in the solar system or motion of electrons near the
nucleus. In classical mechanics there is an important theorem called
the Bertrand's theorem which proposes that there can only be two types
of central potentials, the Coulomb type and the simple harmonic type,
which can produce stable, circular orbits for particles moving around
the potential source. A good presentation of the Bertrand's theorem can be
found in Ref.~\cite{goldstein}. The present article tries to
generalize the results of Bertrand's theorem when the orbiting
particle can have relativistic velocities.
In this article we first set up the relativistic orbit equation for a
particle in a central potential presumed to be dependent on the radial
coordinate only. The relativistic central force orbits were previously
studied in Refs.~\cite{boyer, tork, reut, frommert}. A brief
description of the central force problem in a relativistic setting in
a Coulomb potential was presented in the book on classical theory of
fields by Landau and Lifshitz \cite{landau}. Before one starts the
main analysis about the stability of orbits of relativistic particles
in a central force potential it is better to specify the assumptions
one makes in arriving at definite results. In the present article we
use the same assumptions and the approximations as utilized by Boyer
in Ref.~\cite{boyer} and Landau in Ref.~\cite{landau}. In the
specific references cited above, none of them present a Lorentz
covariant treatment of the relativistic central force problem. The
main reason being that all of them assumes a central potential $V(r)$
where $r=|{\bf r}|$ is the distance between the source and the
orbiting particle. The form of the potential only depends on the
position coordinates of the orbiting particle. The form of $V(r)$ is
not Lorentz covariant. In such cases the results of the whole analysis
is valid in a particular frame where the origin of the coordinate
system coincides with the potential center.
The references cited above assumes the particle which produces the
potential $V(r)$ to be static in the specific coordinate system
utilized by the observer. If the source of the potential does not have
any velocity then the retarded nature of the interactions, owing to
the finite velocity of light, does not complicate the calculation of
the orbit of the relativistic particle. A specific example will make
the point clear. In classical electrodynamics if the source of the
Coulomb potential $V(r)$ moves with a velocity ${\bf v}_s$ and the
orbiting particle has a velocity ${\bf v}$ then the potential $V(r)$
gets a relativistic correction. The magnitude of the lowest order
relativistic correction to the Coulomb potential was calculated by
Darwin in 1920 and it looks like
$$\frac{V(r)}{2c^2}\left[{\bf v}_s \cdot {\bf v} + \frac{({\bf
v}_s\cdot{\bf r})({\bf v}\cdot{\bf r})}{r^2}\right]\,.$$ For a
better understanding of the Darwin correction one can look at
Ref.~\cite{jackson}. In our case ${\bf v}_s=0$ and consequently there
will be no relativistic modification of $V(r)$. More over we do
not consider any general relativistic effects due to $V(r)$ into
account. We briefly comment on the general relativistic generalization
of the central force problem in section \ref{reltd}. In the present
article the background space-time is assumed to be flat.
In the article it will be shown that the stability condition of the
perturbed orbits around a stable circular orbit gives rise to a
non-linear differential equation for the central potential. The
Newtonian or the Coulomb potential satisfies the resulting
differential equation with some restrictions on the possible value of
the angular momentum of the orbiting particle. Except the Newtonian
potential solution we present a more general solution of the
differential equation for the potential which can give rise to stable,
circular orbit for relativistic particles. This solution gives rise to
a force which is not common in physics except its Newtonian inverse
square law limit. The equation of the orbit of a relativistic particle
in such a non-trivial force shows that the orbit will precess and the
precession angle can be calculated.
Unlike the non-relativistic case, in the relativistic case there exist
no radial effective potential minimizing which we can obtain the
radius of a circular orbit. In the relativistic case a first order
perturbation from a circular orbit is enough to determine the
stability criterion of the orbit. In the non-relativistic case one
uses higher order perturbations from a circular orbit to specify the
form of the potential. In the relativistic case the general solution
of the form of the potential from first order perturbation from
circular orbit is such that all higher order corrections becomes
irrelevant. As a consequence of this fact the general form of the
potential which can produce stable, circular orbits for relativistic
particles contains more parameters than the corresponding expressions
of non-relativistic potentials.
The material in the article is presented in the following manner. The
second section sets the conventions and derives the orbit equation of
a relativistic particle in a central orbit. Section \ref{stability}
generalizes the Bertrand's theorem for the relativistic case. In this
section the stability condition for the circular orbits will be
interpreted as a non-linear differential equation for the
potential. The solutions of the non-linear stability equation will
also be derived in section \ref{stability}. In section \ref{reltd} the
connection of the present work with some related works which were
existing in the literature are discussed. This section gives a wider
view for the readers who really want to understand the stability of
orbits in special relativity and general relativity. The last section
\ref{concl} summarizes the important points presented in the article.
\section{The orbit equation}
\label{orbit}
In this section we derive the orbit equation of the relativistic
particle in presence of a potential $V(r)$ which is purely a function
of the radial coordinate. The Lagrangian of a relativistic particle
of mass $m$ in presence of a continuous radial potential $V(r)$ is
\begin{eqnarray}
{\mathcal L} &=& -mc^2\sqrt{1-{v^2}/{c^2}} - V(r)\,,
\label{lagrange}
\end{eqnarray}
where the velocity of the particle ${\bf v}$ in plane polar
coordinates is given as
$${\bf v}=\dot{r}\hat{e}_r + r\dot{\theta}\hat{e}_\theta\,,$$ where
$\hat{e}_r$ , $\hat{e}_\theta$ are the mutually orthogonal unit
vectors along the radial and the angular directions. The form of the
Lagrangian in Eq.~(\ref{lagrange}) immediately shows that the angular
momentum
\begin{eqnarray}
L=\frac{\partial {\mathcal L}}{\partial \dot{\theta}}=
mr^2 \gamma \dot{\theta}\,,
\label{angmom}
\end{eqnarray}
is a constant, where
$$\gamma=\frac{1}{\sqrt{1-{v^2}/{c^2}}}\,.$$
The total energy $E$ of the particle in presence of the potential $V(r)$ is
\begin{eqnarray}
E= m c^2 \gamma + V(r)\,.
\label{e1}
\end{eqnarray}
Although from the definition of $\gamma$ it looks like that it is a
function of $r$, $\dot{r}$ and $\dot{\theta}$ but it can be shown that
in a central force field $\gamma$ is only a function of the radial
coordinate $r$. The reason for such behavior of $\gamma$ can be
understood from the following reason. As energy and angular momentum
are constant functions of $r$, $\dot{r}$ and $\dot{\theta}$ we can use
the conservation conditions of $E$ and $L$ to re-express $\dot{r}$ and
$\dot{\theta}$ as functions of $r$, $E$ and $L$. As $E$ and $L$ are
constants so in a central force field $\dot{r}$ and $\dot{\theta}$ are
functions of $r$ alone. Consequently $\gamma$ is only a function
of $r$. In special relativity the energy of the particle in a central
potential can also be written as
\begin{eqnarray}
E=\sqrt{p^2 c^2 + m^2 c^4}+V(r)\,,
\label{energy}
\end{eqnarray}
where $p=|{\bf p}|$,
\begin{eqnarray}
{\bf p}=m\gamma {\bf v}&=&m\gamma(\dot{r}\hat{e}_r +
r\dot{\theta}\hat{e}_\theta)\nonumber\\
&=&p_r \hat{e}_r + p_\theta \hat{e}_\theta\,,
\label{momentum}
\end{eqnarray}
and
$$p_r=m\gamma \dot{r}\,,\,\,\,\,\,\, p_\theta=m\gamma
r\dot{\theta}=\frac{L}{r}\,.$$ As because
$({p_r}/{p_\theta})=({\dot{r}}/{r\dot{\theta}})$, we
have $$p_r=\frac{L}{r^2}\frac{dr}{d\theta}\,.$$ With the above
information on the various momentum components we can now rewrite
Eq.~(\ref{energy}) as
\begin{eqnarray}
(E-V)^2 = \left(\frac{L}{r^2}\frac{dr}{d\theta}\right)^2 c^2
+ \frac{L^2c^2}{r^2} + m^2c^4\,.
\label{energy1}
\end{eqnarray}
Instead of $r$ we use the variable $u=\frac{1}{r}$ in terms of which
Eq.~(\ref{energy1}) becomes
$$(E-V)^2=L^2c^2\left(\frac{du}{d\theta}\right)^2 + u^2L^2c^2 + m^2c^4\,.$$ If
we differentiate the last equation with respect to $\theta$ and then
divide the resulting equation by $du/d\theta$ we obtain the desired
equation of the orbit of a particle of mass $m$ possessing momentum
${\bf p}$ moving in the presence of a general central potential $V(r)$
as
\begin{eqnarray}
\frac{d^2 u}{d \theta^2} + u = \frac{(V-E)}{L^2 c^2}\frac{dV}{du}\,.
\label{orbit_eqn}
\end{eqnarray}
Using Eq.~(\ref{e1}) we can rewrite the above equation in the form
\begin{eqnarray}
\frac{d^2 u}{d \theta^2} + u =
-\frac{m\gamma}{L^2}\frac{dV}{du}\,.
\label{gorbit_eqn}
\end{eqnarray}
Writing $L=\gamma\ell$, where
$\ell=mr^2\dot{\theta}$ is the non-relativistic angular momentum, the
above equation in the non-relativistic limit ($\gamma \to 1$)
transforms exactly to the form we get in a conventional
non-relativistic treatment of the problem as given in Ref.~\cite{goldstein}.
\section{Circular, Stable closed orbits}
\label{stability}
Lets define
\begin{eqnarray}
J(u)\equiv \frac{(V-E)}{L^2c^2}\frac{dV}{du}\,.
\label{defj}
\end{eqnarray}
Suppose Eq.~(\ref{orbit_eqn}) admits a circular orbit of radius
$r_0=1/u_0$. For small perturbations around this circular orbit we
can Taylor expand $J(u)$ around $u_0$. Keeping up to first order terms
in the perturbation of $u$ we get
\begin{eqnarray}
J(u)=J(u_0) + (u-u_0)\left(\frac{dJ}{du}\right)_{u_0}\,.
\label{pertbj}
\end{eqnarray}
Noting that $J(u_0)=u_0$ for the circular orbit, we can now write
Eq.~(\ref{orbit_eqn}) as
\begin{eqnarray}
\frac{d^2 u}{d \theta^2} + (u-u_0) = (u-u_0)\left(\frac{dJ}{du}\right)_{u_0}\,.
\nonumber
\end{eqnarray}
If we define $x \equiv u-u_0$ then the above equation can be written as
\begin{eqnarray}
\frac{d^2 x}{d \theta^2} + \zeta^2 x = 0\,,
\label{stbld}
\end{eqnarray}
where $\zeta^2$ is defined as
\begin{eqnarray}
\zeta^2 \equiv 1- \left(\frac{dJ}{du}\right)_{u_0}\,.
\label{betasqrd}
\end{eqnarray}
From Eq.~(\ref{stbld}) it is clear that if the orbit of the
relativistic particle in a general central potential has to be stable
then $\zeta^2 > 0$ and if the orbit has to be closed then $\zeta$ must
be a rational number.
\subsection{A differential equation for the potential $V(r)$ producing
stable and closed circular orbits}
The rational number $\zeta$ as predicted, in Eq.~(\ref{betasqrd}),
from the stability criterion of closed circular orbits in the central
force problem is an interesting input in the theory. The interesting
property about this rational number is that it is a constant and so it
does not depend on the details of the orbit which one tries to
perturb. The reason for the constancy of $\zeta$ is the following. For
any circular orbit with radius $r_0$ a specific $\zeta$ specifies the
number of undulations of the perturbed orbit. If $\zeta$ is a rational
number then the number of undulations of the perturbed orbit will be
such that they form a closed geometrical structure. Now suppose one
takes another circular orbit of radius $r_0+\delta r$ where $\delta r
\ll r_0$. If $\zeta$ has a different value on this orbit then the
number of undulations due to a perturbation will be different. In the
limit $\delta r \to 0$ in a continuous manner the two unperturbed
circular orbits tends to each other but the number of undulations on
the circular orbits will not match as $\zeta$ is not a continuous
variable but can only have discrete rational values. Consequently the
number of cycles of the perturbations will change discontinuously with
radius and the perturbed orbits cannot be closed at this
discontinuity. As we are only interested in stable, closed orbits we
can conclude that $\zeta$ must be a constant and not change discretely
with $r$. The discussion on the constancy of $\zeta$ as given above
closely follows the analysis given in Ref.~\cite{goldstein} where the
author gives a nice discussion on the role of $\zeta$ in the case of
non-relativistic orbits.
As $\zeta$ is a constant and must not depend upon the choice of $u_0$
or $x$ one can interpret Eq.~(\ref{betasqrd}) as an independent
differential equation by itself,
\begin{eqnarray}
1- \left(\frac{dJ}{du}\right)=\zeta^2\,.
\label{nbetasqrd}
\end{eqnarray}
whose solutions would give us information about the general form of the
central potential $V(r)$. Using Eq.~(\ref{defj}) we can write the last
equation as
\begin{eqnarray}
(V-E)\frac{d^2 V}{d u^2} + \left(\frac{dV}{du}\right)^2
=L^2 c^2 (1-\zeta^2)\,,
\label{maineqn}
\end{eqnarray}
which is a non-linear second order differential equation. The right
hand side of the above equation is a constant which can be written as
\begin{eqnarray}
d=L^2c^2(1-\zeta^2)\,.
\label{ddef}
\end{eqnarray}
Eq.~(\ref{maineqn}) admits multiple solutions for $V$. The constant
$E$ is the total energy of the particle.
It is interesting to note that the differential equation for the
potential stemming from the stability of closed, circular orbits in
the relativistic case does not have a non-relativistic
analogue. Although the orbit equation Eq.~(\ref{gorbit_eqn}) has a
proper non-relativistic limit the same cannot be said about
Eq.~(\ref{maineqn}). The reason for such behavior can be seen clearly
if we rewrite Eq.~(\ref{maineqn}) in a slightly different way. From the
expression of the energy of the particle in the central force field $\gamma$
can always be written as $(E-V)/mc^2$. As the total energy is a constant in the
present case we must have $d\gamma/du=-(1/mc^2)dV/du$. Consequently
Eq.~(\ref{maineqn}) can also be written as
\begin{eqnarray}
\gamma\frac{d^2 \gamma}{du^2}+\left(\frac{d\gamma}{du}\right)^2
=\frac{L^2(1-\zeta^2)}{m^2c^2}\,,
\label{geqna}
\end{eqnarray}
which gives a differential equation of $\gamma$. The equation above
obviously does not have a well defined non-relativistic limit. The
relativistic stability condition produces an ill-defined
non-relativistic limit due to the fact that in the relativistic case
$J(u)$ as given in Eq.~(\ref{defj}) depends upon the velocity of the
orbiting particle\footnote{The $V-E$ in $J(u)$ is proportional to
$\gamma$ which depended upon the velocity of the particle.}. In the
non-relativistic $J(u)$ was purely a function of the radial coordinate
of the orbiting particle. A perturbation from the circular orbit in
the relativistic case consists of two kinds of perturbations. One is
related to the change in position of the particle from its previous
orbit and the other is the change in velocity from the
velocity it had previously on the circular orbit. In the
non-relativistic case only a radial perturbation from the circular
orbit fixes the shape of the stability condition. As because the
stability condition of the orbit depends upon velocity of the
relativistic particle and the corresponding non-relativistic stability
condition does not depend upon the velocity of the particle, the
non-relativistic limit of Eq.~(\ref{maineqn}) or Eq.~(\ref{geqna}) is
not well defined.
In the case of non-relativistic motion we know that the inverse square
law potential and the simple harmonic potential has the capability to
produce stable, closed circular orbits. In the present case to get the
forms of the potentials which can produce stable, closed orbits we
have to solve Eq.~(\ref{maineqn}). As it is a non-trivial equation we
will first try to see whether the the potentials which produced
stable, closed orbits in the non-relativistic regime still satisfy
Eq.~(\ref{maineqn}). Let us try to see whether any power law solution
of the form
\begin{eqnarray}
V(u)=-\alpha u^\tau\,,
\label{plaw}
\end{eqnarray}
where $\alpha>0$ satisfies Eq.~(\ref{maineqn}). In the above equation
$\alpha$ and $\tau$ are constants. If we substitute the above form of
the potential in Eq.~(\ref{maineqn}) we get
$$u^{2(\tau-1)}\left\{\alpha^2\tau(\tau-1) +
\alpha^2\tau^2\right\} -
u^{\tau-2}\left\{E\alpha\tau(\tau-1)\right\}=d\,.$$ This
directly shows that the above relation can be valid for any $u$
only if $\tau=1$, when $\alpha^2=d$ or
\begin{eqnarray}
L=\frac{\alpha}{c\sqrt{1-\zeta^2}}\,,
\label{lcond}
\end{eqnarray}
on using Eq.~(\ref{ddef}). In this case we see that choosing $\tau=1$ in
Eq.~(\ref{plaw}) we get the Coulomb or
Newtonian potential. The last equation shows that for stable,
circular orbits the particles angular momentum must satisfy some
condition. Eq.~(\ref{lcond}) implies that $\zeta^2 < 1$, and as
$\zeta^2 > 0$ for a stable orbit, we have
\begin{eqnarray}
0 < \zeta^2 < 1\,.
\label{betalim}
\end{eqnarray}
The above equation gives
\begin{eqnarray}
L > \frac{\alpha}{c}\,,
\label{ajp}
\end{eqnarray}
giving a lower bound on the angular momentum of the orbiting
particle. This lower bound of the orbiting particle was previously
obtained in a different way by T. H. Boyer in Ref.~\cite{boyer}. It
must be noted here that except $\tau=1$ no other values of $\tau$ are
allowed in the potential which can produce stable circular orbits of
relativistic particles . In non-relativistic mechanics we do also have
the harmonic-oscillator potential corresponding to $\tau=-2$ and
$\alpha<0$ in Eq.~(\ref{plaw}), but interestingly relativistic effects
forbid this value of $\tau$.
\subsection{The general solution of the differential equation for the
potential and the nature of orbits}
We can find out the general form of the force which can produce
stable, circular relativistic orbits. Noticing that the left hand side
of Eq.~(\ref{maineqn}) can also be written as
$$\frac{d^2}{du^2}\left[\frac{(V-E)^2}{2}\right]\,,$$
it can be easily shown that
\begin{eqnarray}
V(r)-E= - \sqrt{d\left(b+\frac{1}{r}\right)^2 + a}\,,
\label{nvform}
\end{eqnarray}
satisfies Eq.~(\ref{maineqn}) where $d$ is as given in
Eq.~(\ref{ddef}) and $b$ and $a$ are two other dimensional,
integration constants. For an attractive force $b > 0$ and $d > 0$ but
$a$ can have any sign. If we assume that as $r\to \infty$, $V(r)\to 0$
then we get a relation between the constants $d$, $b$ and $a$ as
\begin{eqnarray}
E=\sqrt{db^2+a}\,.
\label{dba}
\end{eqnarray}
From Eq.~(\ref{nvform}) we get the force acting on the particle,
$${\bf F}=-\nabla V(r)\,,$$
as
\begin{eqnarray}
{\bf F}=-\frac{d\left(b+\frac1r\right)}
{r^2\sqrt{d\left(b+\frac1r\right)^2+a}}\,\,\hat{r}\,,
\label{force}
\end{eqnarray}
From the form of the force and Eq.~(\ref{dba}) we immediately see that
if $a=0$ we have $b=E/\sqrt{d}$ and we get back the Newtonian or the
Coulombic potential. From the form of the potential as written in
Eq.~(\ref{plaw}) we can furthermore identify $\alpha=\sqrt{d}$ and
consequently when $a=0$ we have $b=E/\alpha$. If $a\ne 0$ then the form of the
force is non-trivial. The form of the force as given in
Eq.~(\ref{force}) cannot be reduced to the harmonic oscillator force
in any limits of the constants. This shows that special relativistic
effects do not allow stable circular orbits in the presence of a force
which is proportional to the negative of the displacement
vector. Although the force expression in Eq.~(\ref{force}) is
mathematically interesting but in physics we do not encounter such a
force, except the $a=0$ limit.
From the expression of $V-E$ as given in Eq.~(\ref{nvform}) we get
\begin{eqnarray}
J(u) &\equiv& \frac{(V-E)}{L^2c^2}\frac{dV}{du}
= \frac{d}{L^2c^2}(u+b)\,.
\label{jup}
\end{eqnarray}
yielding
\begin{eqnarray}
\frac{d^2 u}{d\theta^2} +u= \frac{d}{L^2c^2}(u+b)\,,
\label{forbit}
\end{eqnarray}
which gives the orbit equation of the relativistic particle which is
acted on by a force given by Eq.~(\ref{force}). As in general
$d=L^2c^2(1-\zeta^2)$ for a stable closed orbit where
$\zeta$ must be a rational number, we get
\begin{eqnarray}
\frac{1}{r}=\frac{1}{R}\cos(\zeta \theta) + \frac{b(1-\zeta^2)}{\zeta^2}\,,
\label{prec}
\end{eqnarray}
where
\begin{eqnarray}
R=Lc\zeta\left[\frac{b^2L^2c^2(1-\zeta^2)}{\zeta^2}+a-m^2c^4\right]^{-1/2}\,.
\end{eqnarray}
The equation of the orbit in Eq.~(\ref{prec}) shows that in the most general case we will
have precession of the orbits dictated by the condition $(2\pi+\delta\theta)\zeta=2\pi$, which predicts that the
orbit precesses by an angle
\begin{eqnarray}
\delta\theta=\frac{2\pi(1-\zeta)}{\zeta}\,,
\end{eqnarray}
per orbit.
\subsection{The case of large perturbations}
Till now we have utilized first order perturbation from a circular
orbit as described in Eq.~(\ref{pertbj}) in the beginning of this
section. To include higher order perturbations from a circular orbit
we require more terms in the Taylor series expansion of $J(u)$ in
Eq.~(\ref{pertbj}). The second order effects will come from terms
proportional to $(d^2J/du^2)_{u_0}$. If the general form of the
potential $V(r)$ satisfies Eq.~(\ref{nvform}) then it is immediately
clear from Eq.~(\ref{jup}) that all derivatives of $J(u)$, except the
first, vanishes. Consequently in the relativistic case it is
impossible to restrict the constants $\zeta$, $a$ and $b$ by higher
order perturbation terms to the circular orbit. For higher order
perturbations from circular orbits the form the potential as given in
Eq.~(\ref{nvform}) remains the same.
\section{Connection of the present work with some related works}
\label{reltd}
One of the findings of the present article is related to the absence
of stable, circular orbits for relativistic particles in presence of a
harmonic oscillator potential. The trajectories of relativistic
particles in a three dimensional harmonic oscillator potential has
been studied previously by L.~Homorodean in
Ref.~\cite{homorodean}. The method followed in the referred work is
completely equivalent to the one followed in the present work. It is
interesting to note that in Homorodean's analysis the general shape of
the orbit in the relativistic case is not an ellipse, or a circle, but
a rosette shaped curve. In presence of the oscillator potential the
angular momentum of the orbiting particle with a specific energy has
an upper bound. The trajectory of the relativistic particle can only
be a circle when it has the highest angular momentum for a fixed
energy. In Ref.~\cite{homorodean} the author does not give any
information about the stability of the orbits. In the non-relativistic
limit the orbit of the particle can be circular. In the light of the
findings in Ref.~\cite{homorodean} of Homorodean the prediction of the
absence of a stable, circular orbit in the oscillator potential is a
sensible result.
Bertrand's theorem in non-relativistic classical mechanics has
inspired some authors to propose a space-time (a metric to be precise)
where any bounded trajectory of a particle is periodic in nature. This
kind of a space-time is named as Bertrand space-time. The works of
Perlick, Ballesteros, Enciso, Herranz and Ragnisco, in
Refs.~\cite{perlick, enciso}, try to generalize the results of the
classical Bertrand's theorem on a flat 3-space to a curved
3-manifold. In Ref.~\cite{perlick} the author found that a specific
form of a space-time which is asymptotically flat can support
Keplerian orbits. The asymptotically flat Bertrand space cannot support
closed trajectories expected in an oscillator type of potential. One
of the findings of the present article predicts that even in flat
space relativistic effects forbid closed, stable trajectories of
particles in presence of an oscillator potential.
\section{Conclusion}
\label{concl}
The outline of the article is based on the well known
Bertrand's theorem on central potentials and orbits of particles as
described in most of the classical mechanics books. Like the
non-relativistic case the relativistic particle's orbit around a
potential source takes place in a plane where the angular momentum and
presumably the energy of the orbiting particle remains constant. The
main difference between the non-relativistic orbits and relativistic
orbits crops up in the orbit equation itself. Unlike the
non-relativistic case in the relativistic case the orbit equation
depends upon the total energy of the particle. The main aim of the
article was to find out possible forms of central potentials which can
produce stable circular orbits for relativistic particles. The
stability condition for the orbits can be transformed to a non-linear
differential equation for the central potential. It is seen that one
of the solutions of the non-linear differential equation for the
central potential is just the normal Coulomb potential. But relativity
affects the properties of the orbits by curtailing the angular
momentum values beyond a certain limit. Except the Coulomb potential
solution we find that the stability equation has another general
mathematically interesting solution which is unlike any potential
which we use in conventional physics. In a specific limit the general
solution reproduces the Newtonian or Coulomb form. In the relativistic
version of the central force problem we lack some restrictions on the
potential which can produce stable circular orbits. In the
non-relativistic version minimizing the effective potential one can
figure out the radius of the circular orbits and higher order
perturbation corrections to the stability condition of the orbits
could be used for unravelling the exact nature of the potential. In
the relativistic version none of those restrictions remain and
consequently the general solution of the potential contains some
constants whose values cannot be analytically calculated.
An important fact which comes out from the article is
about the non-existence of the harmonic oscillator potential as a
solution of the stability equation. In non-relativistic treatment of
the Bertrand's theorem it is well known that only two kinds of potentials
can produce stable circular orbits, one is of the Coulomb type and the
other is of the harmonic oscillator type. The Coulomb form of the
potential passes the stability test for circular orbits but the
harmonic type does not.
The article was focussed on some mathematical properties of
relativistic particle orbits in a central potential. Before we finally
conclude it is pertinent to say some thing on the practical side of
the relativistic central force problem. Atomic physics always remains
a store house of exciting phenomena and one of the places where one
may like to apply the tools of relativistic central force problems
lies inside the atom. This fact was discussed in Ref.~\cite{boyer}.
People have studied about the Schr\"{o}dinger equation and the Dirac
equation in presence of the Coulomb potential. It can be quite
interesting to study the analogous problems using the potential
presented in this article instead of the Coulomb potential. This
attempt can seriously shed some light on the physics of the atoms.
More over as there exists some work on the general relativistic
generalization of Bertrand's theorem one may expect that in the
simplest of the situations, where space-time remains flat, the results
of the present work can be applied for the orbits of very fast moving
bodies interacting via Newtonian gravity with a massive source.
\vskip .1cm
\noindent
{\bf Acknowledgement:}\,\,The authors acknowledge the illuminating
comments from H.~C.~Verma after he read the initial manuscript. Many
of his suggestions were implemented while preparing the final version
of the manuscript.
|
\section{Introduction}
This article is devoted to the study of the Cauchy problem of Hartree equation
\begin{equation}\label{eq:H}\tag{H}
\left\{
\begin{aligned}
&i \partial_t u + \frac12 \Delta u = \eta (|x|^{-\nu}*|u|^2) u, \\
&u(0)=u_0,
\end{aligned}
\right.
\end{equation}
where $(t,x) \in {\mathbb{R}}\times {\mathbb{R}}^d$, $d\geqslant1$, and $\eta\in{\mathbb{R}}$.
The function space to which the initial data $u_0$ belong will be specified later.
We treat the case where the exponent $\nu$ is negative.
More specifically, let us consider $\nu \in [-2,0)$.
To make notation clear, we introduce $\gamma=-\nu\in (0,2]$
and $\lambda=-\eta$, and consider
\begin{equation}\label{eq:nH}\tag{nH}
\left\{
\begin{aligned}
&i \partial_t u + \frac12 \Delta u = -\lambda (|x|^{\gamma}*|u|^2) u, \\
&u(0)=u_0.
\end{aligned}
\right.
\end{equation}
In what follows, we call \eqref{eq:nH} as \emph{negative Hartree equation}
and distinguish it from \eqref{eq:H} by assuming $\nu,\gamma>0$.
The Hartree equation \eqref{eq:H} and the
negative Hartree equation \eqref{eq:nH} are generalized models of
Schr\"odinger-Poisson system
\begin{equation}\label{eq:SP}\tag{SP}
\left\{
\begin{aligned}
&i \partial_t u + \frac12 \Delta u = V_{\mathrm{P}} u, \\
&-\Delta V_{\mathrm{P}} = |u|^2, \\
&u(0)=u_0.
\end{aligned}
\right.
\end{equation}
When $d\geqslant 3$ the potential $V_{\mathrm{P}}$ is given by
$V_{\mathrm{P}}(x)=c_d (|x|^{-(d-2)}*|u|^2)$,
where $c_d$ is a positive constant.
Hence \eqref{eq:SP} corresponds to the special case of
\eqref{eq:H} such that $\eta= c_d>0$
and $\nu = d-2 >0$.
Hartree equation is extensively studied after \cite{GV-MZ} (see \cite{CazBook} and references therein).
On the other hand, when $d\leqslant2$ the potential $V_{\mathrm{P}}$ has a different form:
\[
V_{\mathrm{P}}(x) = \left\{
\begin{aligned}
&-\frac1{2\pi}(\log|x|*|u|^2), \quad (d=2), \\
& -\frac1{2}(|x|*|u|^2), \quad (d=1).
\end{aligned}
\right.
\]
One sees that that the nonlinearity grows at the spatial infinity.
In particular, when $d=1$ \eqref{eq:SP} corresponds to
\eqref{eq:H} with $\eta=-1/2<0$ and $\nu=-1<0$, that is,
to \eqref{eq:nH} with $\lambda=1/2>0$ and $\gamma=1>0$.
Then, the negative Hartree equation \eqref{eq:nH} appears as a generalized model
with respect to $d$ and $\gamma$.
It turns out that the nonlinear interaction is defocussing (or repulsive) if $\lambda>0$,
and focusing (or attractive) if $\lambda<0$.
This article is a consequence of \cite{Ma2DSPe} in which
global well-posedness of \eqref{eq:SP} for dimensions $d=1$ and $d=2$
is shown in an energy class
(see also \cite{DR-JMP,StrSIAMMA,StiMMMAS} for one dimensional case).
The global well-posedness of \eqref{eq:nH} for $\gamma \in (0,1]$
follows by adapting the arguments in \cite{Ma2DSPe} (see Theorem \ref{thm:appendix1}).
Hence, in this article we concentrate on
$\gamma \in (1,2]$, in which case
the growth rate of the nonlinear potential is higher
than in the previous results.
Another type of local existence result on \eqref{eq:SP} for $d=2$
is established in \cite{Ma2DSP}.
\subsection{\texorpdfstring{Main result 1 - the case $\gamma<2$}{Main result 1}}
We first sate our result for $\gamma\in (1,2)$.
Throughout this article,
we use the notation $\Jbr{x}:=(1+|x|^2)^{1/2}$ for $x\in {\mathbb{R}}^d$
and denote by ${\mathbb{F}}$ the Fourier transform in ${\mathbb{R}}^d$;
\[
{\mathbb{F}} f (\xi)=(2\pi)^{-d/2}\int_{{\mathbb{R}}^d}e^{-ix\cdot\xi}f(x)dx.
\]
For nonnegative $s$ and $r$, we define a function space $\Sigma^{s,r}$ by
\[
\Sigma^{s,r}= \Sigma^{s,r}({\mathbb{R}}^d) := \{f \in H^s({\mathbb{R}}^d) | \Jbr{x}^r f(x) \in L^2({\mathbb{R}}) \}
\]
with a norm $\norm{f}_{\Sigma^{s,r}}^2:= \norm{f}_{H^s({\mathbb{R}}^d)}^2+ \norm{\Jbr{x}^{r}f}_{L^2({\mathbb{R}}^d)}^2$, where $H^s$ stands for the Sobolev space.
Let us introduce an \emph{energy}
\[
E[u(t)] = \frac12 \Lebn{\nabla u(t)}2^2
- \frac{\lambda}{4} \iint_{{\mathbb{R}}^d\times {\mathbb{R}}^d} |x-y|^{\gamma}
|u(t,x)|^2|u(t,y)|^2 dxdy,
\]
a \emph{center of mass}
\[
X[u(t)] = \int_{{\mathbb{R}}^d} y |u(t,y)|^2 dy,
\]
and a \emph{momentum}
\[
P[u(t)] = \operatorname{Im} \int_{{\mathbb{R}}^d} \overline{u(t,y)} \nabla u(t,y) dy
=\int_{{\mathbb{R}}^d} \xi |{\mathbb{F}} u(t,\xi)|^2 d\xi.
\]
Notice that $E[u]$, $X[v]$, and $P[w]$ make sense for $u \in \Sigma^{1,\gamma/2}$,
$v\in \Sigma^{0,1/2}$, and $w\in H^{1/2}$, respectively.
\begin{theorem}\label{thm:main1}
Let $d\geqslant 1$, $\gamma \in (1,2)$, and $\lambda \in {\mathbb{R}}$.
Then, \eqref{eq:nH} is globally well-posed.
More precisely,
for $u_0 \in \Sigma^{1,\gamma/2}$, there exists a global solution
$u\in C({\mathbb{R}};\Sigma^{1,\gamma/2}) \cap C^1({\mathbb{R}};(\Sigma^{1,\gamma/2})^{\prime})$ to \eqref{eq:nH}.
The solution conserves the mass $\Lebn{u(t)}2$, the energy $E[u(t)]$, and
the momentum $P[u(t)]$.
The solution is unique in the following class:
\[
\left\{ u\in C({\mathbb{R}},\Sigma^{1,\gamma/2}) \Bigm| P[u(t)] = \mathrm{const.} \right\}.
\]
\end{theorem}
\begin{remark}
\begin{enumerate}
\item
We say that uniqueness holds unconditionally if it holds
under $C({\mathbb{R}};\Sigma^{1,\gamma/2})$.
In this theorem, the uniqueness of \eqref{eq:nH} holds conditionally since
the conservation of momentum is additionally required.
For the unconditional uniqueness of nonlinear Schr\"odinger equation
with power nonlinearity, we refer the reader to \cite{FPT-CM,FT-CCM,KaJAM95,WT-HMJ}.
\item Conservation of $P[u(t)]$ is equivalent to $X[u(t)]\equiv M ({\bf a}t+{\bf b})$, where
$M$, ${\bf a}$, and ${\bf b}$ are defined in \eqref{def:M} and \eqref{def:ab} below.
\item We have the following bound on $L^2$-norms of $\nabla u(t)$ and $\Jbr{x}^{\gamma/2}u(t)$
(see also Remark \ref{rmk:expgrow}):
\begin{equation}\label{eq:dispdu}
\Lebn{\nabla u(t)}2 \leqslant \left\{
\begin{aligned}
&C\Jbr{t}^{\frac{\gamma}{2-\gamma}} && \text{if }\lambda>0,\\
&C && \text{if }\lambda<0,
\end{aligned}
\right.
\end{equation}
\begin{equation}\label{eq:dispxu}
\Lebn{\Jbr{x}^{\frac{\gamma}2} u(t)}2 \leqslant \left\{
\begin{aligned}
&C\Jbr{t}^{\frac{\gamma}{2-\gamma}} && \text{if }\lambda>0,\\
&C\Jbr{t}^{\frac{\gamma}2} && \text{if }\lambda<0.
\end{aligned}
\right.
\end{equation}
\item We show this theorem in a general framework
by replacing $\lambda |x|^\gamma$ with an abstract potential
in a class of divergent functions
(Theorem \ref{thm:general}).
\end{enumerate}
\end{remark}
\begin{remark}
A formula $|x|^{-d+\alpha} * = \Lambda(\alpha) (-\Delta)^{-\frac\alpha2}$
is known for $0<\alpha<d$,
where $\Lambda(\alpha)=2^{\alpha-d/2}\Gamma(\alpha/2)/\Gamma(d/2-\alpha/2)$
(see \cite{StBook}).
If $\gamma \in (0,2)$ and if $\gamma-2>-d$ then we have
\[
-\Delta (|x|^\gamma*|u|^2) = -\gamma(\gamma-2+d) (|x|^{\gamma-2}*|u|^2)
=\Lambda(d+\gamma)(-\Delta)^{-\frac{d+\gamma}{2}+1}|u|^2.
\]
In a sense, this implies $|x|^\gamma*=\Lambda(d+\gamma) (-\Delta)^{-\frac{d+\gamma}2}$ ,
which is an extension of the above formula to $\alpha=d+\gamma>d$.
\end{remark}
\subsection{\texorpdfstring{Main Result 2 - the case $\gamma=2$}{Main result 2}}
We next consider the case $\gamma=2$.
There exists an explicit solution in this case.
Moreover, uniqueness property holds without conservation of momentum.
Assume $u_0 \in \Sigma^{1,1}$ and introduce the following vectors and numbers.
We first let
\begin{equation}\label{def:M}
M= \norm{u_0}_{L^2}^2
\end{equation}
and define constant vectors
\begin{align}
{\bf a} &{}:= M^{-1}P[u_0],&
{\bf b} &{}:= M^{-1}X[u_0],
\label{def:ab}
\end{align}
which represents the (scaled) momentum and
the center of mass, respectively. Let
\begin{equation}\label{eq:cde}
\begin{aligned}
c&{}:=\Lebn{\nabla u_0}2^2, &
d&{}:= \operatorname{Im}\int \overline{u_0} x\cdot \nabla u_0 dx , &
e&{}:=\Lebn{x u_0 }2^2.
\end{aligned}
\end{equation}
For $\omega \in {\mathbb{R}}$, we set
$\mathcal{U}_{\omega}(t) = e^{it(\frac{\Delta}{2}+ \frac{\omega}{2}|x|^2)}$.
An integral representation of $\mathcal{U}_{\omega}(t)$ is known
as Mehler's formula.
For a vector ${\bf a}$,
let $\tau_{\bf a}$ and $\pi_{\bf a}$ be translation operators defined by
$(\tau_{\bf a}f)(x)=f(x-{\bf a})$ and $(\pi_{\bf a} f)(x)=e^{ix\cdot {\bf a}}f(x)$,
respectively.
\begin{theorem}\label{thm:main2}
Let $d\geqslant1$, $\gamma=2$ and $\lambda=\pm1/2$.
Then, \eqref{eq:nH} is globally well-posed in $\Sigma^{1,1}$.
Moreover, the uniqueness holds unconditionally.
Furthermore, the unique solution of \eqref{eq:nH} is given by
\[
u(t) = \left\{
\begin{aligned}
&\exp\left({i\frac{|{\bf a}|^2}2t + i\psi_+(t)}\right) \tau_{{\bf a}t+{\bf b}}\pi_{{\bf a}}\mathcal{U}_{M}(t) \pi_{-{\bf a}}
\tau_{-{\bf b}} u_0, &&\text{if }\lambda=1/2, \\
&\exp\left({i\frac{|{\bf a}|^2}2t + i\psi_-(t)}\right) \tau_{{\bf a}t+{\bf b}}\pi_{{\bf a}}
\mathcal{U}_{-M}(t) \pi_{-{\bf a}}
\tau_{-{\bf b}} u_0, &&\text{if }\lambda=-1/2,
\end{aligned}
\right.
\]
where
\begin{equation*}
\begin{aligned}
\psi_+(t) ={}&
\frac{c -M|{\bf a}|^2 +M(e-M|{\bf b}|^2) }{4M^{3/2}}\sinh (\sqrt{M}t)\cosh (\sqrt{M}t)
\\&{}
+ \frac{d-M{\bf a}\cdot{\bf b}}{2M}\sinh^2 (\sqrt{M}t)
+ \frac{- c + M|{\bf a}|^2 +M(e - M|{\bf b}|^2)}{4M} t
\end{aligned}
\end{equation*}
and
\[
\begin{aligned}
\psi_-(t) ={}& \frac{c-M|{\bf a}|^2 -M (e- M|{\bf b}|^2)}{4M^{3/2}}\sin (\sqrt{M}t)\cos (\sqrt{M}t) \\
&{} + \frac{M{\bf a}\cdot{\bf b} - d}{2M} \sin^2 (\sqrt{M}t)
+ \frac{- c+M|{\bf a}|^2 -M(e-M|{\bf b}|^2)}{4M} t .
\end{aligned}
\]
\end{theorem}
By a scaling argument, the general $\lambda>0$ case and $\lambda<0$ case are reduced to the
case $\lambda=1/2$ and the case $\lambda=-1/2$, respectively.
\begin{remark}
From the explicit representation of the solution, we can deduce that
the nonlinearity causes the following three effects on large time behavior
of the solution.
First is the nonlinear phase $e^{i\psi_{\pm}(t)}$.
It is known that if $\nu>1$ (if $\gamma<-1$) then the nonlinear dynamics
is compared with free one
but if $\nu\leqslant 1$ (if $\gamma \geqslant -1$) then it is not and a phase correction
must be taken into account (cf. long-range scattering \cite{GO-CMP}).
It seems that $e^{i\psi_{\pm}(t)}$ is a correction of this kind.
Furthermore, the linear dynamics of which the nonlinear dynamics can be regarded
as a perturbation changes from $e^{it\Delta/2}$ into $e^{it(\Delta\pm M|x|^2)/2}$.
This is the second respect.
It is important to remark that this modified linear dynamics depends
on the mass of the solution.
This phenomena occurs at least for $\gamma>0$.
When $\gamma=0$, a solution of \eqref{eq:nH} is given by
$u(t)=e^{-i\lambda t M} e^{it\Delta/2}u_0$, which is a free dynamics with a phase
correction.
Third is the translation in both Fourier and physical spaces,
which depends only on the momentum ${\bf a}$ and the center of mass ${\bf b}$.
This represents the motion of the center of mass and is involved at least for $\gamma>1$.
\end{remark}
\begin{remark}\label{rmk:expgrow}
Let $\gamma=2$ and $\lambda=1/2$.
A calculation shows
\begin{multline*}
\Lebn{ \nabla u }2^2=\Lebn{(\cosh(\sqrt{M}t)\nabla + i\sqrt{M} \sinh(\sqrt{M}t)x) u_0}2^2 \\
+M\abs{\bf a}^2 - M\abs{ \cosh(Mt){\bf a}+ \sqrt{M}\sinh (\sqrt{M}t){\bf b} }^2
\end{multline*}
and
\begin{multline*}
\Lebn{ \sqrt{M}x u }2^2=\Lebn{\left({\sinh(\sqrt{M}t)}\nabla + i\sqrt{M}\cosh(\sqrt{M}t)x\right) u_0}2^2 \\
+M^2\abs{{\bf a}t+{\bf b}}^2 - M\abs{ \sinh(Mt){\bf a}+ \sqrt{M}\cosh (\sqrt{M}t){\bf b} }^2.
\end{multline*}
Hence, $\Lebn{ \nabla u (t)}2=O(e^{|t|})$ and $\Lebn{ x u (t)}2=O(e^{|t|})$ as $|t|\to\infty$
for a suitable data (for example, $u_0(x)=e^{-|x|^2}$).
These growth rates are much faster than those for free solutions; $\Lebn{\nabla e^{i{t\Delta}/2}u_0}2=O(1)$
and $\Lebn{x e^{i{t\Delta}/2}u_0}2=O(|t|)$ as $|t|\to\infty$.
This is because the nonlinearity, which is regarded as a repulsive quadratic potential and a remainder,
accelerates the dispersion.
Carles studies effects of repulsive quadratic potentials in \cite{CaSIAMMA}.
The above exponential growths for $\gamma=2$ are, in a sense, equalities of \eqref{eq:dispdu}
and \eqref{eq:dispxu} in the limit $\gamma\uparrow2$.
If a similar acceleration occurred for $\gamma<2$ and $\lambda>0$,
it seems reasonable that the time
growths of $\Lebn{\nabla u}2$ and $\Lebn{xu}2$ are faster than those of free solutions
as in \eqref{eq:dispdu} and \eqref{eq:dispxu}.
\end{remark}
\subsection{\texorpdfstring{Transformation of \eqref{eq:nH}}{Transformation of (nH)}}
What is difficult when we solve \eqref{eq:nH} is the fact that the nonlinear potential
$(|x|^\gamma*|u|^2)$ grows at the spatial infinity.
For this, it is hard to apply a usual perturbation argument to the corresponding integral equation.
To overcome this respect, we introduce a transformation of \eqref{eq:nH}.
Let us now observe this with a formal computation.
We consider the case $\gamma=2$ as a model.
The equation is then
\[
i\partial_t u + \frac12 \Delta u = -\lambda (|x|^2*|u|^2)u.
\]
The right hand side is equal to
\[
-\lambda |x|^2 \Lebn{u(t)}2^2 u(x)
+ 2\lambda x \cdot X[u] u(x)
-\lambda \int_{{\mathbb{R}}^d} |y|^2|u(y)|^2 dy u(x).
\]
As long as $\lambda \in{\mathbb{R}}$, we can expect that $\Lebn{u(t)}2$ is conserved.
Hence the first term is regarded as $-\lambda M |x|^2 u(x)$, where $M$ is
as in \eqref{def:M}.
Now, $u$ solves
\begin{equation}\label{eq:aux1}
i\partial_t u + \frac12 \Delta u + \lambda M |x|^2u
= 2\lambda x \cdot X[u]u
-\lambda \int_{{\mathbb{R}}^d} |y|^2|u(y)|^2 dy u.
\end{equation}
Although the right hand side of this equation is still divergent,
the main part of the nonlinearity is removed and so
the growth rate is not $O(|x|^2)$ any longer but $O(|x|^1)$ as $|x|\to\infty$.
This argument is introduced in \cite{Ma2DSPe}.
Now, let us go one step further.
We next observe from \eqref{eq:nH} that
\[
\frac{d}{dt}X[u(t)] = P[u(t)]
\]
follows. Similarly, by a formal calculation, one verifies that
$\frac{d}{dt}P[u(t)]=0$.
Thus, integrating twice gives us
\[
X[u(t)] = M({\bf a}t + {\bf b}),
\]
where ${\bf a}$ and ${\bf b}$ are defined in \eqref{def:ab}.
Now, we introduce a new unknown
\begin{equation}\label{def:ut}
\widetilde{u}(t,x) = e^{-i \frac{|{\bf a}|^2}{2}t}
(\pi_{{-\bf a}} \tau_{-{\bf a}t-{\bf b}} u)(x) .
\end{equation}
Namely, we work with the center of mass frame.
Then, one verifies that $\widetilde{u}$ also solves \eqref{eq:aux1} and
$X[\widetilde{u}(t)] \equiv 0$.
These facts imply that $\widetilde{u}$ is a solution to
\begin{equation}\label{eq:aux2}
i\partial_t \widetilde{u} + \frac12 \Delta \widetilde{u} + \lambda M |x|^2\widetilde{u}
= -\lambda \int_{{\mathbb{R}}^d} |y|^2|\widetilde{u}(y)|^2 dy \widetilde{u}.
\end{equation}
Now, the right hand side is bounded with respect to $x$.
Let us further set
\begin{equation}\label{def:w}
w(t,x)= \widetilde{u}(t,x) \exp\left(-i \lambda \int_0^t \int_{{\mathbb{R}}^2} |y|^2 |\widetilde{u}(s,y)|^2 dyds\right).
\end{equation}
Then, $w$ solves a linear Schr\"odinger equation $i\partial_t w + \frac12 \Delta w + \lambda M|x|^2 w=0$.
Applying inverses of \eqref{def:w} and \eqref{def:ut},
we obtain an explicit solution of \eqref{eq:nH}.
The argument in the case $\gamma \in (1,2)$ is similar.
We introduce $\widetilde{u}$ as in \eqref{def:ut} and try to solve
a \emph{modified Hartree equation}
\begin{equation}\label{eq:mH}\tag{mH}
\left\{
\begin{aligned}
&i \partial_t \widetilde{u} + \frac12 \Delta \widetilde{u} + \lambda M |x|^{\gamma}\chi(|x|) \widetilde{u} \\
&\qquad\qquad {}= -\lambda \int_{{\mathbb{R}}^d} \left(|x-y|^{\gamma}-|x|^\gamma\chi(|x|) + \gamma \Jbr{x}^{\gamma-2}x\cdot y\right)|\widetilde{u}(y)|^2dy \widetilde{u}, \\
& \widetilde{u}(0)=\pi_{-\bf a}\tau_{-\bf b} u_0
\end{aligned}
\right.
\end{equation}
instead of \eqref{eq:aux2},
where $\chi$ is a smooth non-decreasing function such that
$\chi(r)=0$ for $r\leqslant 1$ and $\chi(r)=1$ for $r\geqslant2$.
It will turn out that
\eqref{eq:mH} can be solved in a standard way because
the growth of the nonlinearity of \eqref{eq:nH} is
successfully removed by the transformation.
It is important to note that \eqref{eq:nH} is \emph{not} a perturbation of
free equation $i\partial_t \psi + (1/2)\Delta \psi=0$ any more but
of $i\partial_t \psi + (1/2)\Delta \psi + \lambda M |x|^{\gamma}\chi(|x|) \psi=0$,
which involves a linear potential.
Oh considered in \cite{OhJDE} the Cauchy problem of
nonlinear Schr\"odinger equation with a divergent potential and
$L^2$-subcritical power-type nonlinearity
(see also \cite{CazBook}).
In particular, the case where the potential is a quadratic polynomial
is extensively studied.
We refer the reader to \cite{CaAHP,CaSIAMMA,CaDCDS,CMS-SIAM,KVZ-CPDE,WZ-NA,ZxFM}.
The rest of this article is organized as follows:
We prove Theorem \ref{thm:main2} in the next Section,
and Theorem \ref{thm:main1} in Section 3.
\section{\texorpdfstring{Proof of Theorem \ref{thm:main2}}{Proof of Theorem 1.4}}
Let us prove our theorem for an equation with a harmonic potential
\begin{equation}\label{eq:2H}
i\partial_t u + \frac12 \Delta u + \frac{\eta}2 |x|^2 u = -\frac\zeta2 (|x|^2*|u|^2)u, \quad
u(0) = u_0\in \Sigma^{1,1}({\mathbb{R}}^d),
\end{equation}
where $\eta$ and $\zeta$ are real constants.
For $\omega\in {\mathbb{R}}$ and ${\bf a},{\bf b}\in {\mathbb{R}}^d$, we define an ${\mathbb{R}}^d$-valued
function $g_\omega(t)$ as
\begin{equation}\label{def:gomega}
g_\omega (t) =
\left\{
\begin{aligned}
&{\bf a}\frac{\sinh (\sqrt\omega t)}{\sqrt{\omega}} +
{\bf b}\cosh (\sqrt\omega t) , &&\omega>0, \\
&{\bf a}t + {\bf b}, &&\omega=0 ,\\
&{\bf a}\frac{\sin (\sqrt{|\omega|} t)}{\sqrt{|\omega|}} +
{\bf b}\cos (\sqrt{|\omega|} t), &&\omega<0.
\end{aligned}
\right.
\end{equation}
Notice that $g_\omega(t)$ is a solution to $g_\omega^{\prime\prime}(t)=\omega g_\omega(t)$
with $g(0)={\bf b}$ and $g^\prime(0)={\bf a}$.
\begin{theorem}\label{thm:harmonic}
\begin{enumerate}
\item Let $d\geqslant1$, $\eta\in {\mathbb{R}}$, and $\zeta\in {\mathbb{R}}$.
Then, \eqref{eq:2H} is globally well-posed in $\Sigma^{1,1}$.
The uniqueness holds unconditionally.
Moreover, $X[u(t)]=M g_\eta(t)$ holds.
\item For a data $u_0 \in \Sigma^{1,1}({\mathbb{R}}^d)$,
define $M$ by \eqref{def:M} and set $\omega = \eta + \zeta M$.
Let $g_\iota(t)$ be defined in \eqref{def:gomega} with a parameter $\iota \in {\mathbb{R}}$
and the data ${\bf a}$ and ${\bf b}$ given by \eqref{def:ab}.
Then, the unique solution to \eqref{eq:2H} is written as
\[
u(t,x) = e^{i\Psi_{\eta,\zeta}(t)}[ \tau_{g_\eta(t)} \pi_{g_\eta^\prime(t)}
\pi_{-g_\omega^\prime(t)} \tau_{-g_\omega(t)} \mathcal{U}_\omega(t)u_0](x)
\]
with
\[
\Psi_{\eta,\zeta}(t) = \frac12(g_\eta(t)\cdot g_\eta^\prime(t) - g_\omega(t)\cdot
g_\omega^\prime(t))
-\frac{\zeta M}2\int_0^t |g_\omega(s)|^2 ds + \frac\zeta2 \psi_\omega(t),
\]
where $\psi_\omega$ is defined with $c$, $d$, and $e$ given by \eqref{eq:cde}
as follows:
\begin{itemize}
\item If $\omega>0$ then
\begin{align*}
\psi_\omega(t) ={}&
\frac{c+\omega e}{2\omega^{3/2}}\sinh(\sqrt{\omega}t) \cosh(\sqrt{\omega}t)
+\frac{d}{\omega}\sinh^2(\sqrt{\omega}t) -\frac{c-\omega e}{2\omega}t
;
\end{align*}
\item if $\omega=0$ then
\[
\psi_0(t) = \frac13 ct^3 + dt^2 + et;
\]
\item if $\omega<0$ then
\begin{align*}
\psi_\omega(t) ={}& -\frac{
c +\omega e
}{2|\omega|^{3/2}}\sin(\sqrt{|\omega|}t) \cos(\sqrt{|\omega|}t)
+\frac{d}{|\omega|}\sin^2(\sqrt{|\omega|}t)
+\frac{
c -\omega e
}{2|\omega|}t.
\end{align*}
\end{itemize}
\end{enumerate}
\end{theorem}
\begin{remark}
Thanks to Proposition \ref{prop:translation} below,
Theorem \ref{thm:main2} immediately follows by taking
$\eta=0$ and $\zeta=\pm1$ (and so $\omega=\pm M$).
\end{remark}
\begin{remark}
In general, momentum of a solution is not conserved in the presence of a linear potential.
Indeed, the solution of \eqref{eq:2H} given in this theorem
satisfies $P[u(t)]=M g^\prime_\eta (t)$.
This is not conserved unless $\eta=0$ (or $u_0$ satisfies ${\bf a}={\bf b}=0$).
\end{remark}
\begin{remark}
Up to a translation in both physical and Fourier spaces and a
nonlinear phase, the solution behaves as $\mathcal{U}_\omega(t)u_0$.
In particular, we have $\Lebn{u(t)}p=\Lebn{\mathcal{U}_\omega(t)u_0}p$ for all $p$.
It is worth pointing out that
not $\eta$ but the exponent $\omega = \eta + \zeta M$ decides the linear profile
of the solution.
This means that, from the view point of change of the dispersive property,
the nonlinearity has the same effect as by the linear potential.
Recall that, however, the motion of the center of mass $X[u(t)]$ is
governed only by the linear potential.
An interesting case would be $\eta\zeta<0$.
In this case, there exists a critical mass $M_c = -\eta/\zeta$
such that the sign of $\omega$ changes at this value.
If $\Lebn{u_0}2^2=M_c$ then
the effect of the linear potential is partially removed by the nonlinearity so that
the solution of \eqref{eq:2H} is
a solution of the free Schr\"odinger equation $e^{i\frac{\Delta}2t}u_0$
up to a translation and a nonlinear phase.
\end{remark}
\begin{proof}
Let us first consider the equation
\begin{equation}\label{eq:aux4eta}
i \partial_t w + \frac12 \Delta w + \frac{\omega}2 |x|^2 w= 0,
\quad w(0)
=\pi_{-{\bf a}} \tau_{-{\bf b}} u_0.
\end{equation}
Obviously, a solution is given by $w(t)=\mathcal{U}_{\omega}(t)\pi_{-{\bf a}} \tau_{-{\bf b}} u_0$.
One verifies that $\frac{d}{dt}X[w(t)] = P[w(t)]$ and
\[
\frac{d}{dt}P[w(t)] = \omega X[w(t)].
\]
Since $X[w(0)] = \int (y-{\bf b})|u_0|^2 dy =0$
and $P[w(0)]=P[u_0]-M {\bf a}=0$ hold,
it follows that $X[w(t)] \equiv 0$.
Set
\begin{equation}\label{def:iw}
\widetilde{u}(t,x) = w(t,x) \exp \left(
i\frac\zeta2 \int_0^t \int_{{\mathbb{R}}^d} |y|^2|w(s,y)|^2 dy ds
\right).
\end{equation}
Now, it is easy to see that $\widetilde{u}$ solves
\begin{equation}\label{eq:aux3eta}
i \partial_t \widetilde{u} + \frac12 \Delta \widetilde{u}
+ \frac\omega2 |x|^2 \widetilde{u}
= -\frac\zeta2 \int_{{\mathbb{R}}^d} |y|^2|\widetilde{u}(y)|^2 dy \widetilde{u}
\end{equation}
and $\widetilde{u}(0)=w(0)=\pi_{-{\bf a}} \tau_{-{\bf b}} u_0$.
Since $|\widetilde{u}(t,x)|=|w(t,x)|$, it also holds that
$X[\widetilde{u}(t)]\equiv X[w(t)]\equiv0$.
We introduce $g_\eta(t)$ by \eqref{def:gomega} with
the data ${\bf a}$ and ${\bf b}$ given in \eqref{def:ab}.
Recall that $g_\eta^{\prime\prime} = \eta g_\eta(t)$.
Hence,
\begin{multline}\label{eq:aux2eta}
i \partial_t \widetilde{u} + \frac12 \Delta \widetilde{u}
+ \frac\omega2 |x|^2 \widetilde{u} +(\eta g_\eta(t)-g_\eta^{\prime\prime}(t))\cdot x \widetilde{u} \\
= \zeta x \cdot X[\widetilde{u}] \widetilde{u}
-\frac\zeta2 \int_{{\mathbb{R}}^d} |y|^2|\widetilde{u}(y)|^2 dy \widetilde{u}.
\end{multline}
Since $\omega = \eta + \zeta M = \eta + \zeta \Lebn{\widetilde{u}(t)}2^2$,
this equation is equivalent to
\[
i \partial_t \widetilde{u} + \frac12 \Delta \widetilde{u}
+ \frac\eta2 |x+g_\eta(t)|^2 \widetilde{u}
-g_\eta^{\prime\prime}(t)\cdot x \widetilde{u} -\frac{\eta}2 \abs{g_\eta(t)}^2 \widetilde{u}
= -\frac\zeta2 \widetilde{u} \int |x-y|^2 |\widetilde{u}(y)|^2 dy .
\]
Hence, if we define $u$ by
\begin{equation}\label{def:iut}
{u}(t,x) = e^{\frac{i}2\int_0^t (|g^\prime_\eta(s)|^2 + \eta |g_\eta(s)|^2 )ds}
( \tau_{g_\eta(t)}\pi_{g^\prime_\eta(t)} \widetilde{u})(x),
\end{equation}
then $u$ solves \eqref{eq:2H}.
It also holds that $u(0)=\tau_{{\bf b}}\pi_{{\bf a}} \widetilde{u}(0)=
\tau_{{\bf b}}\pi_{{\bf a}} \pi_{-{\bf a}} \tau_{-{\bf b}} u_0=u_0$.
Combining \eqref{def:iut} and \eqref{def:iw}, we conclude that
\begin{align}
u(t,x)={}&
e^{i \frac{1}{2}\int_0^t (|g^\prime_\eta(s)|^2 +\eta |g_\eta(s)|^2)ds}
( \tau_{g_\eta(t)} \pi_{g^\prime_\eta(t)} \widetilde{u})(x)\nonumber\\
={}&
e^{i \widetilde{\Psi}_{\eta,\zeta}(t)}
( \tau_{g_\eta(t)} \pi_{g^\prime_\eta(t)}
\mathcal{U}_\omega(t)\pi_{-{\bf a}} \tau_{-{\bf b}} u_0)(x), \label{eq:explicitsol}
\end{align}
where
\begin{align*}
\widetilde{\Psi}_{\eta,\zeta}(t)={}& \frac{1}2\int_0^t (|g^\prime_\eta(s)^2| +\eta |g_\eta(s)|^2) ds + \frac\zeta2 \int_0^t \int_{{\mathbb{R}}^d} |y|^2|w(s,y)|^2 dy ds\\
={}& \frac12(g_\eta(t)\cdot g_\eta^\prime(t) - {\bf a}\cdot{\bf b})+ \frac\zeta2 \int_0^t \int_{{\mathbb{R}}^d} |y|^2|w(s,y)|^2 dy ds
\end{align*}
Then, Propositions \ref{prop:translation} and \ref{prop:phaseformula} show the stated
representation of the solution.
Now, let us proceed to the proof of the uniqueness.
Suppose that $u_1\in C({\mathbb{R}};\Sigma^{1,1})$ is a solution of \eqref{eq:2H}
in $(\Sigma^{1,1})^\prime$ sense.
By the equation, we see that $u_1 \in C^1 ({\mathbb{R}};(\Sigma^{1,1})^\prime)$ and so that
$\Lebn{u_1}2$ is conserved and
$\frac{d}{dt} X[u_1(t)] = P[u_1(t)] \in C({\mathbb{R}})$ holds.
By Proposition \ref{prop:motion} below, we obtain $P[u_1(t)] \in C^1({\mathbb{R}})$ and $\frac{d}{dt}P[u_1(t)]=\eta X[u_1(t)]$.
Then, $X[u_1(t)] = M g_\eta(t)$.
Let us introduce
\begin{equation*}
\widetilde{u}_1(t,x) =
e^{-\frac{i}2\int_0^t (|g^\prime_\eta(s)|^2+\eta|g_\eta(s)|^2)ds}
(\pi_{-g^\prime_\eta(t)} \tau_{-g_\eta(t)}u_1)(x) ,
\end{equation*}
which is the inverse transform of \eqref{def:iut}.
Then, $\widetilde{u}_1$ solves \eqref{eq:aux2eta}. Since
\[
X[\widetilde{u}_1(t)] = \int (y-g_\eta(t))|u_1(t,y)|^2 dy
=X[u_1(t)] - M g_\eta(t)=0,
\]
$\widetilde{u}$ is also a solution to \eqref{eq:aux3eta}.
Now, let us further introduce
\begin{equation*}
w_1(t,x) = \widetilde{u}_1(t,x) \exp \left(
-i\frac\zeta2 \int_0^t \int_{{\mathbb{R}}^d} |y|^2|\widetilde{u}_1(s,y)|^2 dy ds
\right).
\end{equation*}
This is the inverse transform of \eqref{def:iw} since
$|w_1(t,x)|=|\widetilde{u}_1(t,x)|$.
Then, $w_1$ solves \eqref{eq:aux4eta} and so $w_1(t,x)=w(t,x)$.
Applying \eqref{def:iut} and \eqref{def:iw}, we conclude that
$u_1$ is identical to $u$.
\end{proof}
\begin{proposition}\label{prop:translation}
Let $\kappa\in {\mathbb{R}}$ and ${\bf a}$, ${\bf b} \in {\mathbb{R}}^d$.
Define $g_\kappa(t)$ by \eqref{def:gomega}.
Then, it holds for all $t\in {\mathbb{R}}$ that
\[
\mathcal{U}_{\kappa}(t) \pi_{-{\bf a}} \tau_{-{\bf b}}
=
e^{-\frac{i}2(g_\kappa(t)\cdot g_\kappa^\prime(t) - {\bf a}\cdot{\bf b})}
\pi_{-g_\kappa^\prime(t)} \tau_{-g_\kappa(t)} \mathcal{U}_{\kappa}(t).
\]
\end{proposition}
\begin{proof}
Fix $\varphi \in \Sigma^{1,1}$.
Let us set
$w(t)=\mathcal{U}_{\kappa}(t) \pi_{-{\bf a}} \tau_{-{\bf b}} \varphi$.
Then,
\[
i\partial_t w + \frac12 \Delta w + \frac\kappa2 |x|^2 w = 0,\quad
w(0) = \pi_{-\bf a} \tau_{-\bf b} \varphi.
\]
Now, we introduce
$
\widetilde{w}(t,x) = e^{\frac{i}2(g_\kappa(t)\cdot g_\kappa^\prime(t) - {\bf a}\cdot{\bf b})}
[\tau_{g_\kappa(t)} \pi_{g_\kappa^\prime(t)} w(t)](x).
$
One easily verifies that
$i\partial_t \widetilde{w} + \frac12 \Delta \widetilde{w} + \frac\kappa2 |x|^2 \widetilde{w} = 0$
and $\widetilde{w}(0) = \varphi$ hold,
which implies $\widetilde{w}(t)=\mathcal{U}_\kappa(t) \varphi$.
Hence,
\[
\mathcal{U}_{\kappa}(t) \varphi
= e^{\frac{i}2(g_\kappa(t)\cdot g_\kappa^\prime(t) - {\bf a}\cdot{\bf b})}
\tau_{g_\kappa(t)} \pi_{g_\kappa^\prime(t)} \mathcal{U}_{\kappa}(t) \pi_{-\bf a} \tau_{-\bf b} \varphi
\]
is valid for arbitrary $\varphi\in \Sigma^{1,1}$.
Alternatively, let $\zeta=0$ in \eqref{eq:explicitsol}.
\end{proof}
\begin{proposition}\label{prop:phaseformula}
Let $w(t)=\mathcal{U}_\omega (t)\pi_{-\bf a} \tau_{-\bf b} u_0$.
Define $\psi_\omega(t)$ as in Theorem \ref{thm:harmonic}.
Then,
\[
\int_0^t \int_{{\mathbb{R}}^d} |y|^2 |w(t)|^2 dy ds =
\psi_\omega(t) -M \int_0^t |g_\omega(s)|^2 ds .
\]
\end{proposition}
\begin{proof}
By the previous proposition, we obtain
\[
\Lebn{xw(t)}2^2 = \Lebn{x\tau_{-g_\omega(t)} \mathcal{U}_\omega(t)u_0}2^2
=\Lebn{x\mathcal{U}_\omega(t)u_0}2^2 - M|g_\omega(t)|^2,
\]
where we have used $X[\mathcal{U}_\omega(t)u_0] = M g_\omega(t)$.
It therefore suffices to show that $\psi_\omega(t)=\int_0^t \Lebn{x\mathcal{U}_\omega(s)u_0}2^2ds$.
Assume $\omega>0$.
It is well known that $\mathcal{U}_{\omega}(t)$ is decomposed as
$\mathcal{U}_{\omega}(t)= \mathcal{M}_{\omega}(t)\mathcal{D}_{\omega}(t)
\mathcal{F} \mathcal{M}_{\omega}(t)$,
where $\mathcal{M}_{\omega}(t)$ is a multiplication operator defined by
\[
\mathcal{M}_{\omega}(t)= \exp\left(i\frac{\sqrt\omega}2 \coth(\sqrt\omega t) |x|^2\right)
\]
and $\mathcal{D}_{\omega}(t)$ is a dilation operator defined by
\[
(\mathcal{D}_{\omega}(t)f)(x)= \left( \frac{\sqrt\omega}{i \sinh (\sqrt\omega t)} \right)^{\frac{n}2}f\left(\frac{\sqrt\omega x}{\sinh(\sqrt\omega t)}\right).
\]
Hence,
\begin{align*}
\Lebn{x\mathcal{U}_\omega(t)u_0}2^2
={}& \left(\frac{\sinh (\sqrt\omega t)}{\sqrt\omega}\right)^2 \Lebn{x\mathcal{F}\mathcal{M}_{\omega}(t) u_0}2^2 \\
={}&\left(\frac{\sinh (\sqrt\omega t)}{\sqrt\omega}\right)^2 \Lebn{\nabla\left(\mathcal{M}_{\omega}(t) u_0\right)}2^2 \\
={}& \left(\frac{\sinh (\sqrt\omega t)}{\sqrt\omega}\right)^2 \Lebn{\nabla u_0+i\sqrt\omega\coth(\sqrt\omega t)x u_0 }2^2 \\
={}& \frac{c}{\omega} \sinh^2(\sqrt\omega t)
+ \frac{2 d}{\sqrt\omega} \sinh(\sqrt\omega t) \cosh(\sqrt\omega t) + e \cosh^2(\sqrt\omega t)
\end{align*}
and so
\begin{align*}
\psi_\omega(t)
= \left(\frac{c+e\omega}{2\omega^{3/2}}\right)\sinh(\sqrt\omega t) \cosh(\sqrt\omega t)
+\frac{d}{\omega }\sinh^2(\sqrt\omega t)
-\left(\frac{c-e\omega }{2\omega }\right)t,
\end{align*}
where $c$, $d$, and $e$ are constants defined in \eqref{eq:cde}.
The proof for $\omega\leqslant 0$ is similar.
We omit details.
\end{proof}
\begin{proposition}\label{prop:motion}
Let $d\geqslant1$ and $\eta,\zeta\in{\mathbb{R}}$. Let $u_0\in \Sigma^{1,1}$.
Let $u\in C({\mathbb{R}}; \Sigma^{1,1})$ solve \eqref{eq:2H} in $(\Sigma^{1,1})^\prime$ sense.
Then, $P[u(t)]$ is a continuously differentiable function of time
and $\frac{d}{dt}P[u(t)]=\eta X[u(t)]$ holds.
\end{proposition}
\begin{proof}
Since $u\in C({\mathbb{R}};\Sigma^{1,1})$, we see that $X[u(t)] \in C({\mathbb{R}})$.
By \eqref{eq:2H}, $\Lebn{u(t)}2=\Lebn{u_0}2$
and $\frac{d}{dt}X[u(t)]=P[u(t)] \in C({\mathbb{R}})$ hold.
Set
\[
v(t,x)= u(t,x) \exp\left( -i\frac{\zeta}{2}\int_0^t \int_{{\mathbb{R}}^d}|y|^2|u(t,s)|^2 dy\, ds \right).
\]
We have $|v(t,x)|=|u(t,x)|$ and so $X[v(t)]=X[u(t)]$
and $\Lebn{v(t)}2=\Lebn{u(t)}2=\Lebn{u_0}2$.
Hence, one sees that $v$ is a solution to
\[
i\partial_t v + \frac12 \Delta v + \frac\omega2 |x|^2 v = \lambda X[v(t)] \cdot x v, \quad
v(0)=u_0,
\]
where $\omega=\eta+\zeta M$.
Let us introduce a function $H(t)\in C^3({\mathbb{R}})$ as follows:
\[
H(t)= -\zeta \int_0^t \frac{e^{\sqrt\omega (t-s)}+ e^{-\sqrt\omega (t-s)}}{2}
\int_0^s X[u(\sigma)] d\sigma\, ds ,
\]
where $\frac{e^{\sqrt\omega t}+ e^{-\sqrt\omega t}}{2} = \cos (\sqrt{|\omega|}t)$
if $\omega<0$.
Remark that $H$ is a solution of
$H^{\prime\prime}(t)=\omega H(t) - \zeta X[u(t)]$ with $H(0)=H^\prime(0)=0$.
Further set
\[
v(t,x) = \exp\left(-\frac{i}{2}\int_0^t (|H^\prime(s)|^2 + \omega |H(s)|^2) ds\right)
[\pi_{H^\prime(t)} \tau_{H(t)} w(t)](x).
\]
Then,
$w$ solves
$ i\partial_t w + \frac12 \Delta w + \frac{\omega}{2}|x|^2 w
=0$ with $w(0) = u_0$, and so
$X[w(t)] = M g_\omega (t) \in C^\infty({\mathbb{R}})$ follows.
Hence, we conclude that
\[
X[u(t)]=X[v(t)] = \int_{{\mathbb{R}}^d} (y+H(t)) |w(t,y)|^2 dy = M g_\omega(t) + M H(t) \in C^3({\mathbb{R}}),
\]
which gives us $P[u(t)] = \frac{d}{dt}X[u(t)]\in C^2({\mathbb{R}})$. Moreover,
\begin{align*}
\frac{d}{dt}P[u(t)]=\frac{d^2}{dt^2} X[u(t)] = {}&\frac{d^2}{dt^2} X[w(t)] + M H^{\prime\prime}(t) \\
={}& M \omega g_\omega(t) + M (\omega H(t) - \zeta X[u(t)]) \\
={}& \omega (M g_\omega(t)+M H(t)) - \zeta M X[u(t)] \\
={}& (\omega - \zeta M) X[u(t)] \\
={}& \eta X[u(t)].
\end{align*}
\end{proof}
\section{\texorpdfstring{Proof of Theorem \ref{thm:main1}}{Proof of Theorem 1.1}}
We shall prove Theorem \ref{thm:main1} in a general framework.
Consider a \emph{generalized Hartree equation}
\begin{equation}\label{eq:gH}\tag{$\mathrm{gH}$}
\left\{
\begin{aligned}
&i \partial_t u + \frac{1}2 \Delta u = - ((V+R)* |u|^2)u , \text{ in } {\mathbb{R}}^{1+d},\\
& u(0,x)=u_0,
\end{aligned}
\right.
\end{equation}
where $V$ and $R$ are real-valued functions of $x\in{\mathbb{R}}^d$.
A pair $(q,r)$ is admissible if $2\leqslant q,r\leqslant \infty$ satisfy the relation
$2/q=\delta(r):=d(1/2-1/r) \in [0,1]$
(however, we exclude the case $(d,q,r)=(2,2,\infty)$).
For nonnegative numbers $p$, $q$ and $r$, define
\begin{equation*}
W^{p,q}_{V,r} := \{ f \in L^r({\mathbb{R}}^d);
\Jbr{\nabla}^{p}f, \, \Jbr{V(\cdot)}^{q/2} f \in L^r({\mathbb{R}}^d) \}
\end{equation*}
with a norm
\[
\norm{f}_{{W}^{p,q}_{V,r}} := \Lebn{\Jbr{\nabla}^{p} f}r + \Lebn{\Jbr{V(\cdot)}^{q/2} f}r.
\]
Let $\Sigma^{p,q}_V:=W^{p,q}_{V,2}$.
Take a vector-valued function $W$ and set
\begin{equation}\label{def:K}
K(x,y)=V(x-y)- V(x) + y\cdot W(x) .
\end{equation}
The assumptions on the potential $V$ and $R$ are the following.
\begin{assumption}
Suppose that $V:{\mathbb{R}}^d\to{\mathbb{R}}$ is a smooth function
satisfying the following properties:
\begin{itemize}
\item[$(\rm{V1})$]
$\partial^\alpha V(x) \in L^\infty({\mathbb{R}}^d)$
for all index $\alpha$ with $|\alpha|\geqslant 2$;
\item[$(\rm{V2})$] There exist constants $C>0$ and $\kappa\in[0,1)$ such that
$\abs{\nabla V(x)} \leqslant C\Jbr{V(x)}^{\kappa/2}$;
\item[$(\rm{V3})$] There eixsts a vector-valued function $W$ such that
$\partial^\alpha W(x) \in L^\infty({\mathbb{R}}^d)$ for all $|\alpha|\geqslant 2$ and
$K$ defined by \eqref{def:K} satisfies
\begin{equation*}
\sup_x\abs{ K(x,y)} + \sup_x\abs{ \nabla_x K(x,y)} \leqslant C \Jbr{V(y)};
\end{equation*}
\item[$(\rm{V4})$] There exists a constant $C>0$ such that $\Jbr{x} \leqslant C \Jbr{V(x)}$.
\end{itemize}
\end{assumption}
\begin{assumption}
Assume that $R$ satisfies the following.
\begin{itemize}
\item[$(\rm{R1})$] $R \in L^\zeta({\mathbb{R}}^d) + L^\infty({\mathbb{R}}^d)$, where
$\zeta \in [1,\infty]$ and $\zeta>d/4$.
\item[$(\rm{R2})$] $R^+:=\max(R,0)\in L^\theta({\mathbb{R}}^d) + L^\infty({\mathbb{R}}^d)$, where
$\theta \in [1,\infty]$ and $\theta>d/2$.
\end{itemize}
\end{assumption}
\begin{remark}
\begin{enumerate}
\item Roughly speaking, $V$ and $R$ denote a divergent part and a remainder
of a (divergent) potential $V+R$, respectively.
When we prove Theorem \ref{thm:main1}, we choose
$V=\lambda |x|^\gamma \chi(x)$ and $R=\lambda|x|^\gamma (1-\chi(x))$, where
$\chi$ is a smooth radial function such that $\chi\equiv1$ for $|x|\geqslant2$
and $\chi\equiv0$ for $|x|\leqslant1$.
Notice that $\lambda|x|^\gamma$ itself does not satisfy $(\rm{V1})$.
\item $\Sigma^{1,1}_V \subset \Sigma^{1,1/2}$ as long as Assumption $(\rm{V4})$ is satisfied,
under which condition $X[u]$ is well-defined for $u \in \Sigma^{1,1}_V$.
\item If $(\rm{R1})$ is satisfied then $(R*|u|^2)u \in (H^1)^\prime$ for $u\in H^1$.
An example of $R$ satisfying $(\rm{R1})$ is the $H^1$-subcritical
Hartree-type nonlinearity; $\eta |x|^{-\nu}$ with $\eta\in{\mathbb{R}}$ and $0<\nu<\min(4,d)$.
\end{enumerate}
\end{remark}
The main result of this section is the following.
\begin{theorem}\label{thm:general}
Suppose Assumptions $(\rm{V1})$, $(\rm{V2})$, $(\rm{V3})$, $(\rm{V4})$, $(\rm{R1})$, and $(\rm{R2})$ are satisfied.
Then, \eqref{eq:gH} is globally well-posed in $\Sigma^{1,1}_V$.
More precisely,
for $u_0\in \Sigma^{1,1}_{V}$, there exists a global solution $u$ of \eqref{eq:gH}
in $C({\mathbb{R}}; \Sigma^{1,1}_{V})\cap C^1({\mathbb{R}};(\Sigma^{1,1}_V)^\prime)$ which
satisfies
$u \in L^q_{\mathrm{loc}}({\mathbb{R}};W^{1,1}_{V,r})$
for all admissible pair $(q,r)$ and
conserves the mass $\Lebn{u(t)}2$, the \emph{energy}
\begin{equation}\label{def:genergy}
E[u(t)] = \frac12 \Lebn{\nabla u(t)}2^2 - \frac14
\iint_{{\mathbb{R}}^{d+d}} (V(x-y)+R(x-y)) |u(t,x)|^2 |u(t,y)|^2 dxdy,
\end{equation}
and the momentum $P[u(t)]$.
The solution is unique in the class
\[
\left\{ u\in L^\infty({\mathbb{R}},\Sigma^{1,1}_V)\cap L^{\frac{8\zeta}d}_{\mathrm{loc}}({\mathbb{R}},W^{1,1}_{V,\frac{4\zeta}{2\zeta-1}}) \Bigm| P[u(t)] = \mathrm{const.} \right\}.
\]
Moreover, we have the following estimates:
If $V \leqslant 0$ then
\[
\Lebn{\nabla u(t)}2 \leqslant C, \quad
\Lebn{\Jbr{V(\cdot)}^{1/2} u(t)}2\leqslant C\Jbr{t}^{\frac1{2-\kappa}};
\]
otherwise
\[
\norm{\nabla u(t)}_{L^2}
\leqslant C \Jbr{t}^{\frac{1}{1-\kappa}}, \quad
\Lebn{\Jbr{V(\cdot)}^{1/2} u(t)}2\leqslant C\Jbr{t}^{\frac1{1-\kappa}},
\]
where $\kappa$ is the number defined in Assumption $(\rm{V2})$.
\end{theorem}
We now apply the transform observed in the introduction.
Then, the problem boils down to the Cauchy problem of
the following modified version of the \eqref{eq:gH}:
\begin{equation}\label{eq:gmH}\tag{$\mathrm{mgH}$}
\left\{
\begin{aligned}
&i \partial_t u + \frac{1}2 \Delta u + M V(x) u= - u \int_{{\mathbb{R}}^d} K(\cdot,y) |u(y)|^2 dy -(R*|u|^2)u,\\
& u(0)=u_0 \in \Sigma^{1,1}_V,
\end{aligned}
\right.
\end{equation}
where $M$ is as in \eqref{def:M} and
$K$ is defined by \eqref{def:K} with $W$ given in Assumption $(\rm{V3})$.
If $(\rm{V3})$ is satisfied then $u\int K(\cdot,y)|u(y)|^2 dy \in \Sigma^{j,1}_V$ for
$u\in \Sigma^{j,1}_V$, where $j=0,1$.
Denote $A:=\frac12\Delta + M V(x)$.
It is well known that $A$ is an essentially self-adjoint operator on $L^2$ as long as
Assumption $(\rm{V1})$ holds.
Moreover, we have Strichartz's estimate under this condition (\cite{YajCMP}).
\begin{lemma}[Strichartz's estimate]
Suppose $(\rm{V1})$. For any fixed $T>0$, the following properties hold:
\begin{itemize}
\item
Suppose $\varphi \in L^2({\mathbb{R}}^2)$.
For any admissible pair $(q,r)$, there exists a constant $C=C(T,q,r)$
such that
\[
\norm{e^{itA} \varphi}_{L^q((-T,T);L^r)} \leqslant C \Lebn{\varphi}2.
\]
\item Let $I \subset (-T,T)$ be an interval and $t_0 \in \overline{I}$.
For any admissible pairs $(q,r)$ and $(\gamma,\rho)$,
there exists a constant $C=C(t,q,r,\gamma,\rho)$ such that
\[
\norm{\int_{t_0}^t e^{i(t-s)A} F(s) ds}_{L^q(I;L^r)}
\leqslant C \norm{F}_{L^{\gamma^\prime}(I; L^{\rho^\prime})}
\]
for every $F \in L^{\gamma^\prime}(I; L^{\rho^\prime})$.
\end{itemize}
\end{lemma}
\begin{lemma}\label{lem:commutatorA}
Let $A=\frac12\Delta + M V(x)$.
Suppose that Assumption $(\rm{V1})$ holds.
Denote by $e^{itA}$ a one-parameter group generated by $A$.
Let $F$ be an arbitrary weight function such that $\nabla F$ and $\Delta F$ are bounded.
Then, for all $f \in \Sigma^{1,1}_V$,
\[
e^{- itA} \nabla e^{itA} f= \nabla f + iM\int_0^t e^{-isA} \left(\nabla V\right) e^{isA} f ds
\]
and
\[
e^{-itA}F e^{itA} f= F f + i\int_0^t e^{-isA}
\left(\nabla F\cdot \nabla + \frac12 (\Delta F)\right)e^{isA} f ds.
\]
\end{lemma}
By means of this lemma, it immediately follows from $(\rm{V1})$ and $(\rm{V2})$
that
\[
\norm{e^{itA}\phi}_{L^\infty((-T,T),\Sigma^{1,1}_V)}
\leqslant \norm{\phi}_{\Sigma^{1,1}_V} + CT \norm{e^{itA}\phi}_{L^\infty((-T,T),\Sigma^{1,1}_V)} + CT\Lebn{\phi}2.
\]
This implies $\norm{e^{itA}\phi}_{\Sigma^{1,1}_{V}} \leqslant C(|t|,M)\norm{\phi}_{\Sigma^{1,1}_{V}}$.
A similar argument shows
$\norm{e^{itA}\phi}_{\Sigma^{k,k}_{V}} \leqslant C(k,|t|,M)\norm{\phi}_{\Sigma^{k,k}_{V}}$
for any positive integer $k$.
Further, as in \cite[Lemma 2.11]{DR-JMP}, $e^{itA}\phi$
is continuous in $\Sigma^{k,k}_V$
with respect to $M$ for each $t$ and $\phi\in \Sigma^{k,k}_V$.
\subsection{\texorpdfstring{Local well-posedness of \eqref{eq:gmH}}{Local well-posedness of (mgH)}}
We first give a unique local solution to \eqref{eq:gmH}.
Throughout this subsection and the next subsection
we suppose that the constant $M$ in the operator $A$ is not
necessarily equal to $\Lebn{u_0}2^2$.
In what follows, we write $L^p((-T,T);X)=L^p_T X$, for short.
The integral form of \eqref{eq:gmH} is
\begin{multline}\label{def:Q}
u(t) = Q[u]:= e^{itA}u_0 +i \int_0^t e^{i(t-s)A} u(s)\int_{{\mathbb{R}}^d} K(x,y)|u(s,y)|^2dy ds\\
+i \int_0^t e^{i(t-s)A} ((R*|u|^2)u)(s) ds.
\end{multline}
\begin{proposition}
Suppose that Assumptions $(\rm{V1})$, $(\rm{V2})$, $(\rm{V3})$, and $(\rm{R1})$ are satisfied.
Let $u_0\in \Sigma^{1,1}_V$.
Define a Banach space
\[
\mathcal{H}_{T,\delta} := \{f \in L^\infty((-T,T);\Sigma^{1,1}_V);\,
\norm{f}_{\mathcal{H}_T} \leqslant \delta \},
\]
where
\[
\norm{f}_{\mathcal{H}_T} :=
\norm{f}_{L^\infty_T {\Sigma}^{1,1}_V}
+ \norm{f}_{L^q_T W^{1,1}_{V,r}}
\]
with $q=8\zeta/d$ and $r=4\zeta/(2\zeta-1)$.
Then,
there exists $\delta_0$ depending only on $\norm{u_0}_{\Sigma^{1,1}_V}$
such that for any $\delta \in [\delta_0, \infty)$,
the operator $Q$ given in \eqref{def:Q} is a contraction map
from $\mathcal{H}_{T,\delta}$ to itself for suitable $T=T(\delta)>0$.
\end{proposition}
\begin{proof}
We prove that $Q$ is a contraction map in two steps.
\subsubsection*{Step 1}
Fix $\delta>0$.
We show that the existence of $T$ such that
$Q[u] \in \mathcal{H}_{T,\delta}$ as long as $u \in \mathcal{H}_{T,\delta}$.
Let $u \in \mathcal{H}_{T,\delta}$.
We first establish estimates on $R$.
Set $R=R_1 + R_2$ with $R_1 \in L^\zeta$ and $R_2\in L^\infty$.
One sees from Young's inequality, H\"odler's inequality, and Sobolev's embedding that
\begin{align*}
\norm{\mathcal{A}((R_1*|u|^2)u)}_{L^{q^\prime}_TL^{r^\prime}}
&{}\leqslant C T^{1-\frac{d}{4\zeta}}\Lebn{R_1}\zeta \norm{\nabla u}_{L^\infty_T L^2}^{\frac{d}{2\zeta}}
\norm{u}_{L^\infty_T L^2}^{2-\frac{d}{2\zeta}} \norm{\mathcal{A}u}_{L^q_TL^r}\\
&{}\leqslant CT^{1-\frac{d}{4\zeta}} \delta^3
\end{align*}
and
\begin{align*}
\norm{\mathcal{A}((R_2*|u|^2)u)}_{L^1_TL^2}
&{}\leqslant C T \Lebn{R_2}\infty \norm{u}_{L^\infty_T L^2}^2
\norm{\mathcal{A}u}_{L^\infty_T L^2}\\
&{}\leqslant CT \delta^3,
\end{align*}
where $\mathcal{A}=\mathrm{Id}$, $\nabla$ or $\Jbr{V(x)}^{1/2}$.
By Strichartz's estimate, we have
\begin{multline*}
\norm{Q[u]}_{L^\infty_T L^2} + \norm{Q[u]}_{L^q_T L^r} \leqslant
C\Lebn{u_0}2 + C\norm{u\int_{{\mathbb{R}}^d} K(\cdot,y)|u(y)|^2dy}_{L^1_TL^2}\\
+C\norm{(R_1*|u|^2)u}_{L^{q^\prime}_TL^{r^\prime}}+
C\norm{(R_2*|u|^2)u}_{L^1_TL^2}.
\end{multline*}
Since $\sup_x |K(x,y)| \leqslant C\Jbr{V(y)}$ holds
by Assumption $(\rm{V3})$, we have
\begin{align*}
\Lebn{u\int_{{\mathbb{R}}^d} K(\cdot,y)|u(y)|^2dy}2
\leqslant{}&\norm{u(x)K(x,y)|u(y)|^2}_{L^2_x({\mathbb{R}}^d; L^1_y({\mathbb{R}}^d))}\\
\leqslant{}&\norm{u(x)K(x,y)|u(y)|^2}_{L^1_y({\mathbb{R}}^d; L^2_x({\mathbb{R}}^d))}\\
\leqslant{}& C \norm{u}_{L^2} \norm{\Jbr{V(\cdot)}^{1/2}u}_{L^2}^2.
\end{align*}
Take $L^1_T$ norm to yield
\[
\norm{u\int_{{\mathbb{R}}^d} K(\cdot,y)|u(y)|^2dy}_{L^1_TL^2} \leqslant C
T\norm{u}_{L^\infty_T L^2} \norm{u}_{L^\infty_T \Sigma^{1,1}_V}^2.
\]
Hence
\begin{equation}\label{eq:estQ1}
\norm{Q[u]}_{L^\infty_T L^2}+\norm{Q[u]}_{L^q_T L^r}
\leqslant C\Lebn{u_0}2+
C(T+T^{1-\frac{d}{4\zeta}})\delta^3.
\end{equation}
\smallbreak
We next estimate $\nabla Q[u]$.
By Lemma \ref{lem:commutatorA}, one sees that
\begin{align*}
\nabla Q [u] ={}& e^{itA} \nabla u_0 + i \int_0^t e^{i(t-s)A} \nabla \left(u(s)\int_{{\mathbb{R}}^d} K(x,y)|u(s,y)|^2dy\right) ds \\
&{} + i \int_0^t e^{i(t-s)A} \nabla \left( (R*|u|^2)u \right)ds \\
&{}
+ i M\int_0^t e^{i(t-s)A} (\nabla V) Q[u] (s)ds.
\end{align*}
By Strichartz's estimate,
\begin{align*}
&\norm{\nabla Q [u]}_{L^\infty_TL^2} + \norm{\nabla Q [u]}_{L^q_TL^r}\\
&{}\leqslant C\norm{\nabla u_0}_{L^2} + C\norm{\nabla \left(u(s)\int_{{\mathbb{R}}^d} K(x,y)|u(s,y)|^2dy\right)}_{L^1_TL^2}
\\&\quad{}
+C\norm{\nabla ((R_1*|u|^2)u)}_{L^{q^\prime}_TL^{r^\prime}}
+C\norm{\nabla ((R_2*|u|^2)u)}_{L^1_TL^2}
\\&\quad{}
+CM \norm{(\nabla V)Q[u]}_{L^1_T L^2}.
\end{align*}
Using Assumption $(\rm{V3})$, we obtain
\begin{align*}
&\norm{\left( \int_{{\mathbb{R}}^d} (\nabla K(\cdot,y))|u(y)|^2dy\right)u}_{L^1_TL^2}
\leqslant CT\norm{\Jbr{V(\cdot)}^{1/2}u}_{L^\infty_TL^2}^2 \norm{u}_{L^\infty_TL^2},\\
&\norm{(\nabla u)\int_{{\mathbb{R}}^d} K(\cdot,y)|u(y)|^2dy}_{L^1_TL^2}
\leqslant CT\norm{\Jbr{V(\cdot)}^{1/2}u}_{L^\infty_TL^2}^2 \norm{\nabla u}_{L^\infty_TL^2}.
\end{align*}
It follows from Assumption $(\rm{V2})$ that
\[
\norm{(\nabla V)Q[u]}_{L^1_T L^2}
\leqslant C\norm{\Jbr{V(\cdot)}^{1/2}Q[u]}_{L^1_T L^2}.
\]
We hence deduce that
\begin{multline}\label{eq:estQ2}
\norm{\nabla Q[u]}_{L^\infty_TL^2} +
\norm{\nabla Q[u]}_{L^q_TL^r} \\
\leqslant C\Lebn{\nabla u_0}2
+ C(T+T^{1-\frac{d}{4\zeta}})\delta^3 + C M T \norm{\Jbr{V(\cdot)}^{1/2}Q[u]}_{L^\infty_TL^2}.
\end{multline}
\smallbreak
Let us proceed to the estimate of $\Jbr{V(x)}^{1/2}Q[u]$.
Apply the second identity of Lemma \ref{lem:commutatorA} with
$F=\Jbr{V(x)}^{1/2}$ to yield
\begin{align*}
&\Jbr{V(x)}^{1/2}Q[u] \\
&{}= e^{itA} \Jbr{V(x)}^{1/2} u_0 \\
&\quad{} + i\int_0^t e^{i(t-s)A} \Jbr{V(x)}^{1/2} u(s)\int_{{\mathbb{R}}^d} K(x,y)|u(s,y)|^2dy ds\\
&\quad{}+ i\int_0^t e^{i(t-s)A} \Jbr{V(x)}^{1/2} (R*|u|^2)u ds\\
&\quad{}+i \int_0^t e^{i(t-s)A}\left(\nabla \Jbr{V(x)}^{1/2} \cdot \nabla + \frac12(\Delta \Jbr{V(x)}^{1/2})\right)Q[u](s)ds.
\end{align*}
Now, Assumptions $(\rm{V1})$ and $(\rm{V3})$ give us that
\[
\abs{\nabla \Jbr{V(x)}^\frac12} \leqslant C, \quad
\abs{\Delta \Jbr{V(x)}^\frac12} \leqslant C.
\]
Hence,
\begin{multline}\label{eq:estQ3}
\norm{\Jbr{V(\cdot)}^{1/2}Q[u]}_{L^\infty_T L^2}
+ \norm{\Jbr{V(\cdot)}^{1/2}Q[u]}_{L^q_T L^r}\\
\leqslant C\Lebn{\Jbr{V(\cdot)}^{1/2} u_0}2
+ C(T+T^{1-\frac{d}{4\zeta}})\delta^3 \\
+CT(\norm{\nabla Q[u]}_{L^\infty_TL^2}+\norm{Q[u]}_{L^\infty_TL^2}).
\end{multline}
From \eqref{eq:estQ1}, \eqref{eq:estQ2}, and \eqref{eq:estQ3},
we finally reach to
\[
\norm{Q[u]}_{\mathcal{H}_T}
\leqslant C_1 \norm{u_0}_{\Sigma^{1,1}_V}
+ C_2 (T+T^{1-\frac{d}{4\zeta}})\delta^3
+ C_3 T \norm{Q[u]}_{\mathcal{H}_T}.
\]
Letting $T$ so small that $C_3 T \leqslant 1/2$, we see
$\norm{Q[u]}_{\mathcal{H}_T}
\leqslant 2C_1 \norm{u_0}_{\Sigma^{1,1}_V}
+ 2C_2 (T+T^{1-\frac{d}{4\zeta}})\delta^3$.
We now choose $\delta_0:= 3C_1 \norm{u_0}_{\Sigma^{1,1}_V}$.
Then, for any $\delta\geqslant \delta_0$,
$Q$ maps $\mathcal{H}_{T,\delta}$ to itself, provided $T$
is so small that $2C_2(T+T^{1-\frac{d}{4\zeta}})\delta^2
\leqslant 1/3$.
\subsubsection*{Step 2}
We next show that $Q$ is a contraction map.
In the same way as in Step 1, we obtain
\begin{equation}\label{eq:contraction}
\norm{Q[u] -Q[v]}_{\mathcal{H}_{T}}
\leqslant C_4 (T+T^{1-\frac{d}{4\zeta}}) (\norm{u}_{\mathcal{H}_{T}}^2+\norm{v}_{\mathcal{H}_{T}}^2)\norm{u-v}_{\mathcal{H}_{T}}
\end{equation}
for small $T$ and $u,v \in \mathcal{H}_{T,\delta}$.
Letting $T$ so small that $2C_4(T+T^{1-\frac{d}{4\zeta}})\delta^2\leqslant 1/2$ if necessary,
we conclude that
$\norm{Q[u] -Q[v]}_{\mathcal{H}_{T}}
\leqslant (1/2) \norm{u-v}_{\mathcal{H}_{T}}$.
\end{proof}
\begin{proposition}
Suppose that Assumptions $(\rm{V1})$, $(\rm{V2})$, $(\rm{V3})$, $(\rm{V4})$, and $(\rm{R1})$ are satified.
Let $u_0\in \Sigma^{2,2}_V$.
Define a Banach space
\[
\mathcal{I}_{T,\delta} := \{f \in L^\infty((-T,T);\Sigma^{2,2}_V);\,
\norm{f}_{\mathcal{I}_T}
\leqslant \delta \},
\]
where
$\norm{f}_{\mathcal{I}_T} :=
\norm{f}_{L^\infty_T \Sigma^{2,2}_V}
+ \norm{f}_{L^q_T W^{2,2}_{V,r}}
$
with $q=8\zeta/d$ and $r=4\zeta/(2\zeta-1)$.
Then, there exists $\delta_0$ depending only on $u_0$
such that for any $\delta \in [\delta_0, \infty)$,
the operator $Q$ given in \eqref{def:Q} is a contraction map
from $\mathcal{I}_{T,\delta}$ to itself for suitable $T=T(\delta)$.
\end{proposition}
The proof is done in a similar way.
\begin{theorem}\label{thm:LWPgmH}(1)
Suppose that Assumptions $(\rm{V1})$, $(\rm{V2})$, $(\rm{V3})$, and $(\rm{R1})$ are satisfied.
Then, for any $u_0 \in \Sigma^{1,1}_V$, there exists
$T=T(\norm{u_0}_{\Sigma^{1,1}_V})$ and a unique solution
$u \in C([-T,T],\Sigma^{1,1}_V)
\cap L^{8\zeta/d}([-T,T],W^{1,1,}_{V,\frac{4\zeta}{2\zeta-1}})$ of \eqref{eq:gmH}.
The solution belongs to
$ C^1([-T,T],(\Sigma^{1,1}_V)^\prime) $ and
$ L^q([-T,T];W^{1,1}_{V,r})$ for all admissible pair $(q,r)$,
and conserves mass $\Lebn{u(t)}2$.
Moreover, the solution depends continuously on the data.
(2) If Assumptions $(\rm{V1})$, $(\rm{V2})$, $(\rm{V3})$, $(\rm{V4})$, and $(\rm{R1})$ are satisfied and if
\begin{equation}\label{eq:neutrality}
\Lebn{u_0}2^2=M,\quad X[u_0] = P[u_0] =0
\end{equation}
are fulfilled, then the solution given in (1)
conserves energy $E[u(t)]$ defined in \eqref{def:genergy} and momentum $P[u(t)]$.
In other words, the Cauchy problem \eqref{eq:gmH} is locally well-posed in
$\widetilde\Sigma^{1,1}_{V}:=\{u_0\in \Sigma^{1,1}_{V}| u_0 \text{ satisfies }\eqref{eq:neutrality}\}$.
Moreover, the solution $u$ solves \eqref{eq:gH}.
\end{theorem}
\begin{remark}
The theorem holds even if we replace Assumption $(\rm{V2})$ with a weaker one;
(V2') $|\nabla V(x)| \leqslant C \Jbr{V(x)}^{1/2}$.
\end{remark}
\begin{proof}
For $u_0 \in \Sigma^{1,1}_V$, we choose $\delta$ and $T$ so that
$Q[u]$ becomes a contraction map from $\mathcal{H}_{T,\delta}$ to itself.
Then, we obtain a unique solution $u \in C([-T,T],\Sigma^{1,1}_V)\cap L^{8\zeta/d}([-T,T],W^{1,1,}_{V,\frac{4\zeta}{2\zeta-1}})$
of integral version of \eqref{eq:gmH}.
From the equation \eqref{eq:gmH}, we see that
$u \in C^1((-T,T),(\Sigma^{1,1}_V)^\prime)$.
Strichartz's estimate gives a bound in $L^q([-T,T],W^{1,1}_{V,r})$ for all admissible pair.
As in \eqref{eq:contraction}, one gets
\begin{equation}\label{eq:approximation}
\norm{u -v}_{\mathcal{H}_{T}}
\leqslant C \norm{u(0) -v(0)}_{\Sigma^{1,1}_V} +
C(T+T^{1-\frac{d}{4\zeta}}) (\norm{u}_{\mathcal{H}_{T}}^2+\norm{v}_{\mathcal{H}_{T}}^2)\norm{u-v}_{\mathcal{H}_{T}}
\end{equation}
for any two solutions $u$ and $v$ of \eqref{eq:gmH}.
From this estimate, we deduce continuous dependence of the solution on the data.
Now, a use of \eqref{eq:gmH} shows that
\[
\frac{d}{dt}\Lebn{u(t)}2^2 = 2\operatorname{Re} \Jbr{u,\partial_t u}_{\Sigma^{1,1}_V,( \Sigma^{1,1}_V)^\prime}=0,
\]
which completes the proof of the first statement of the theorem.
We now suppose that Assumption $(\rm{V4})$ holds and $u_0$ satisfies \eqref{eq:neutrality}.
To justify the momentum conservation, we use a regularization argument.
Let $\{u_{0,n}\}_{n}$ be a sequence of functions in $\Sigma^{2,2}_V$ which
satisfies $\Lebn{u_{0,n}}2^2=M$ and
converges to $u_0$ in $\Sigma^{1,1}_V$ as $n\to\infty$,
and let $u_n \in C([-\widetilde{T}_n,\widetilde{T}_n],\Sigma^{2,2}_V)
\cap C^1((-\widetilde{T}_n,\widetilde{T}_n),L^2)$ be
corresponding solutions of \eqref{eq:gmH} with $u_n(0)=u_{0,n}$.
Notice that $\widetilde{T}_n \geqslant T/2$ for large $n$ since
$\norm{u_{0,n}}_{\Sigma^{1,1}_V}$ converges to
$\norm{u_{0}}_{\Sigma^{1,1}_V}$ as $n\to\infty$.
Thanks to \eqref{eq:approximation}, $u_n$ converges to $u$ in
\begin{equation}\label{eq:app-convergence}
C([-T/2,T/2],\Sigma^{1,1}_V) \cap C^1((-T/2,T/2),(\Sigma^{1,1}_V)^\prime).
\end{equation}
It therefore holds for each $n$ that
\begin{align*}
\frac{d}{dt}
P[u_n(t)]
={}& \operatorname{Im} \int_{{\mathbb{R}}^d} \overline{\partial_t u_n(t,y)}
\nabla u_n(t,y) dy\\
{}&+ \operatorname{Im} \int_{{\mathbb{R}}^d} \overline{u_n(t,y)}
\nabla \partial_t u_n(t,y) dy,
\end{align*}
which makes sense because $\nabla \partial_t u_n \in H^{-1}$.
Integrating by parts and plugging \eqref{eq:gmH}, we see that
\begin{align}\label{eq:Pconservation1}
\frac{d}{dt}P[u_n(t)] = \int_{{\mathbb{R}}^d} |u_n(t,x)|^2 \nabla \left(W(x)
\cdot X[u_n(t)] \right)dx.
\end{align}
Notice that the right hand side makes sense because $|\nabla W_j(x)| \leqslant C\Jbr{V(x)}^{\frac12}$ is valid
and $\Sigma^{1,1}_V\subset \Sigma^{1,1/2}$ holds by virtue of $(\rm{V4})$,
where $W_j$ denotes the $j$-th component of $W$.
Indeed, take $y={\bf e}_j$ in \eqref{def:K}
and differentiate with respect to $x$ to yield
$\nabla W_j(x)=\nabla K(x,{\bf e}_j)-\nabla V(x-{\bf e}_j)+\nabla V(x)$.
By Assumptions $(\rm{V1})$, $(\rm{V2})$, and $(\rm{V3})$, one obtains $|\nabla W_j(x)| \leqslant C\Jbr{V(x)}^{1/2}$.
We also have
\begin{equation}\label{eq:Pconservation2}
\frac{d}{dt}X[u_n(t)]
= P[u_n(t)].
\end{equation}
Two estimates \eqref{eq:Pconservation1} and \eqref{eq:Pconservation2} give us
\begin{multline*}
\frac{d}{dt} \left( \abs{X[u_n(t)]}^2 + |P[u_n(t)]|^2 \right)\\
\leqslant C\left(1+\norm{u_n(t)}_{\Sigma^{0,1}_{V}}^2\right) \left( \abs{X[u_n(t)]}^2 + |P[u_n(t)]|^2 \right).
\end{multline*}
Applying Gronwall's lemma, we conclude that
\begin{equation}\label{eq:pseudomomentum}
\abs{X[u_n(t)]}^2 + |P[u_n(t)]|^2
\leqslant C \delta^2 e^{T(1+\delta^2)} \norm{u_n - u_0}_{\Sigma^{1,1}_V}^2
\end{equation}
for $t\in(-T,T)$,
where we have used the estimates
\begin{align*}
\abs{X[u_{n}(0)]}
&{}=\abs{X[u_{0,n}]-X[u_0]}=\abs{\int_{{\mathbb{R}}^d} y (|u_{0,n}(y)|^2 - |u_{0}(y)|^2)dy} \\
&{}\leqslant C\norm{u_{0,n}-u_0}_{\Sigma^{1,1}_V}(\norm{u_{0,n}}_{\Sigma^{1,1}_V}
+\norm{u_0}_{\Sigma^{1,1}_V})
\end{align*}
and
\begin{align*}
\abs{P[u_n(0)]}
&{}=\abs{P[u_{0,n}]-P[u_0]}=\abs{\int_{{\mathbb{R}}^d} \xi (|{\mathbb{F}} u_{0,n}(\xi)|^2 - |{\mathbb{F}} u_{0}(\xi)|^2)d\xi} \\
&{}\leqslant \norm{u_{0,n}-u_0}_{\Sigma^{1,1}_V}(\norm{u_{0,n}}_{\Sigma^{1,1}_V}
+\norm{u_0}_{\Sigma^{1,1}_V}).
\end{align*}
Convergence of $u_n$ in the topology \eqref{eq:app-convergence} allows us to claim
$X[u(t)] \equiv P[u(t)]\equiv0$ for $t\in (-T/2,T/2)$.
By this fact and the mass conservation, the right hand side of \eqref{eq:gmH}
becomes
\begin{multline*}
-u \left( (V*|u|^2)- V(x) \Lebn{u_0}2^2
+W (x) \cdot X[u]\right) \\
= - (V*|u|^2)u +M V(x) u-0.
\end{multline*}
Therefore, $u$ solves \eqref{eq:gH}.
Let us again consider the above sequence $\{u_{0,n}\}_n \in \Sigma^{2,2}_V$ approximating $u_0$
and the corresponding sequence of solutions $\{u_n\}_n$ to \eqref{eq:gmH}.
Conservation of mass allows us to rewrite \eqref{eq:gmH} into
\[
i\partial_t u_n + \frac12 \Delta u_n = -((V+R)*|u_n|^2) u_n - W \cdot X[u_n] u_n.
\]
Multiplying this equality by $\overline{\partial_t u_n}$ and integrating real parts
of the resulting terms, we get
\[
\frac{d}{dt}E[u_n(t)]= \operatorname{Re} \int W u_n\overline{\partial_t u_n} dx \cdot X[u_n].
\]
This calculation makes sense because $\overline{\partial_t u_n}$
is a continuous $L^2$-valued function and $u_n$ satisfies
\eqref{eq:gmH} in $L^2$ sense.
Since $\operatorname{Re} \int W_j u_n\overline{\partial_t u_n} dx=\frac12\operatorname{Im}\int \overline{u_n}\nabla W_j \cdot \nabla u_ndx$,
one sees from \eqref{eq:pseudomomentum} that
\[
\abs{E[u_n(t)] - E[u_{0,n}]}
\leqslant \frac{C\delta^3 }{1+\delta^2}e^{\frac{T}2(1+\delta^2)}\norm{u_n - u_0}_{\Sigma^{1,1}_V}
\to 0
\]
as $n\to\infty$ for $t\in (-T,T)$.
By means of the convergence of $u_n$ in \eqref{eq:app-convergence},
$|E[u(t)]-E[u(0)]|\leqslant |E[u(t)]-E[u_{n}(t)]|+ |E[u_n(t)]-E[u_{0,n}]| + |E[u_{0,n}]-E[u_{0}]|\to 0$ as $n\to\infty$ for $t\in (-T/2,T/2)$,
which gives us the desired conservation law.
\end{proof}
\subsection{\texorpdfstring{Global well-posedness of \eqref{eq:gmH}}{Global well-posedness of (mgH)}}
We next extend the above solution of \eqref{eq:gH} to whole real line ${\mathbb{R}}$.
\begin{lemma}\label{lem:estVu}
Suppose Assumption $(\rm{V2})$ is satisfied.
Let $u\in C((-T,T);\Sigma^{1,1}_V)$ be a solution to \eqref{eq:gH} with
data $u_0 \in \Sigma^{1,1}_V$.
Then, it holds that
\[
\norm{\Jbr{V(\cdot)}^{1/2}u}_{L^\infty_t L^2}
\leqslant C (\norm{u_0}_{\Sigma^{1,1}_V}) (1+ (|t|\norm{\nabla u}_{L^\infty_tL^2}
)^{\frac1{2-\kappa}} )
\]
for $t\in(-T,T)$, where $\kappa$ is the number defined in Assumption $(\rm{V2})$.
\end{lemma}
\begin{proof}
Let us estimate $\Lebn{\sqrt{|V|}u}2$ because $\Lebn{u}2$ is conserved.
For $r>0$, we have
\begin{align*}
\frac{d}{dt} \Lebn{{\bf 1}_{\{|x|\leqslant r\}}\sqrt{|V|} u(t)}2^2
&{}= 2\operatorname{Re} \Jbr{{\bf 1}_{\{|x|\leqslant r\}} |V| u, \partial_t u }
= \operatorname{Im} \Jbr{{\bf 1}_{\{|x|\leqslant r\}}|V| u, \Delta u }\\
&{}= \operatorname{Im} \int_{|x|\leqslant r} (\overline{\nabla u} \cdot \nabla |V|) u dx,
\end{align*}
which implies
\[
\abs{\frac{d}{dt} \Lebn{{\bf 1}_{\{|x|\leqslant r\}}\sqrt{|V|} u(t)}2^2}
\leqslant \Lebn{{\bf 1}_{\{|x|\leqslant r\}}|\nabla V| u(t)}2 \Lebn{\nabla u(t)}2
\]
Integrating in time and substituting $|\nabla V|\leqslant C\Jbr{V}^{\kappa/2}$ give us
\begin{multline*}
\norm{{\bf 1}_{\{|x|\leqslant r\}}\sqrt{|V|} u}_{L^\infty_tL^2}^2 \leqslant
C\norm{u_0}_{\Sigma^{1,1}_V}^2 \\
+ |t|\norm{\nabla u}_{L^\infty_tL^2} \norm{{\bf 1}_{\{|x|\leqslant r\}}\Jbr{V}^{1/2} u}_{L^\infty_tL^2}^\kappa \norm{{\bf 1}_{\{|x|\leqslant r\}} u}_{L^\infty_tL^2}^{1-\kappa}.
\end{multline*}
Since $r$ is arbitrary, we pass to the limit $r\to \infty$ and reach to
\begin{equation}\label{eq:estVuind}
\norm{\Jbr{V}^{1/2} u}_{L^\infty_tL^2}^2 \leqslant
C\norm{u_0}_{\Sigma^{1,1}_V}^2
+ C|t|\norm{\nabla u}_{L^\infty_tL^2}\Lebn{u_0}2^{1-\kappa}
\norm{\Jbr{V}^{1/2} u}_{L^\infty_tL^2}^\kappa.
\end{equation}
It follows from Young's inequality that
\begin{multline*}
C|t|\norm{\nabla u}_{L^\infty_tL^2} \Lebn{u_0}2^{1-\kappa}
\norm{\Jbr{V}^{1/2} u}_{L^\infty_tL^2}^\kappa\\
\leqslant \frac12 \norm{\Jbr{V}^{1/2} u}_{L^\infty_tL^2}^2
+ C_{\kappa}(|t|\norm{\nabla u}_{L^\infty_tL^2}\Lebn{u_0}2^{1-\kappa})^{\frac2{2-\kappa}}.
\end{multline*}
Plugging this estimate to \eqref{eq:estVuind}, we conclude that
\[
\norm{\Jbr{V}^{1/2} u}_{L^\infty_tL^2}^2
\leqslant C\norm{u_0}_{\Sigma^{1,1}_V}^2
+ C\Lebn{u_0}2^{2-\frac2{2-\kappa}}(|t|\norm{\nabla u}_{L^\infty_tL^2})^{\frac2{2-\kappa}}.
\]
\end{proof}
A blow-up criterion immediately follows from this lemma.
\begin{corollary}[Blow-up criterion]
Suppose Assumptions $(\rm{V1})$, $(\rm{V2})$, $(\rm{V3})$, $(\rm{V4})$,
and $(\rm{R1})$ and \eqref{eq:neutrality} are satisfied.
Let $u\in C((-T_{\mathrm{min}},T_{\mathrm{max}});\Sigma^{1,1}_V)$ be a
maximal solution to \eqref{eq:gH}.
If $T_{\mathrm{max}}<\infty$ (resp. $T_{\min}<\infty$) then
$\Lebn{\nabla u(t)}2\to\infty$ as $t \uparrow T_{\max}$
(resp. as $t \downarrow -T_{\min}$).
\end{corollary}
Now, the following lemma shows that \eqref{eq:gmH} is globally well-posed
in $\widetilde\Sigma^{1,1}_{V}$.
More precisely, the solution of \eqref{eq:gmH} given
in Theorem \ref{thm:LWPgmH} (2) never blows up in finite time.
\begin{lemma}[Global $\dot{H}^1$-bound]\label{lem:H1bound}
Suppose $(\rm{V1})$, $(\rm{V2})$, $(\rm{V3})$, $(\rm{V4})$, $(\rm{R1})$, and $(\rm{R2})$.
Let $u_0\in \widetilde\Sigma^{1,1}_{V}$ and let
$u\in C((-T_{\mathrm{min}},T_{\mathrm{max}});\Sigma^{1,1}_V)$ be a corresponding
maximal solution to \eqref{eq:gmH} which conserves energy $E[u(t)]$
and momentum $P[u(t)]$.
Then, we have the following bounds for $t\in(-T_{\mathrm{min}},T_{\mathrm{max}})$:
\begin{itemize}
\item If $V \leqslant 0$ then, $\Lebn{\nabla u(t)}2 \leqslant C$;
\item otherwise, $\norm{\nabla u(t)}_{L^2}
\leqslant C \Jbr{t}^{\frac1{1-\kappa}}$,
where $\kappa$ is the number defined in Assumption $(\rm{V2})$.
\end{itemize}
\end{lemma}
\begin{proof}
If $V\leqslant 0$ then
\begin{align*}
\Lebn{\nabla u(t)}2^2 \leqslant{}& 2E[u(t)] + \frac12 \Lebn{ (R^+* |u|^2) |u|^2}1\\
\leqslant{}& 2E[u_0] + C\Lebn{R^+}\theta
\Lebn{\nabla u(t)}2^{\frac{d}{\theta}}\Lebn{u_0}2^{4-\frac{d}{\theta}}
+ \frac12 \Lebn{R^+}\infty \Lebn{u_0}2^4.
\end{align*}
This gives us $ \Lebn{\nabla u(t)}2 \leqslant C $ since $\theta>d/2$.
Otherwise, we have
\begin{multline*}
\Lebn{\nabla u(t)}2^2
\leqslant 2E[u_0] + \frac12
\iint_{{\mathbb{R}}^{d+d}} V(x-y) |u(t,x)|^2 |u(t,y)|^2 dxdy\\
+\frac12 \Lebn{ (R^+* |u|^2) |u|^2}1.
\end{multline*}
Since $u_0 \in \widetilde\Sigma^{1,1}_{V}$,
we have $X[u(t)]=0$.
Therefore, Assumption $(\rm{V3})$ yields
\begin{align*}
&\iint_{{\mathbb{R}}^{d+d}} V(x-y) |u(t,x)|^2 |u(t,y)|^2 dxdy\\
&{}= \iint_{{\mathbb{R}}^{d+d}} K(x,y) |u(t,x)|^2 |u(t,y)|^2 dxdy\\
&{}\quad + \int_{{\mathbb{R}}^d}V(x)|u(t,x)|^2dx \int_{{\mathbb{R}}^d}|u(t,y)|^2dy\\
&{}\quad + \int_{{\mathbb{R}}^d}W(x)|u(t,x)|^2dx \cdot X[u(t)] \\
&{}\leqslant C \Lebn{u_0}2^2 \Lebn{\Jbr{V(\cdot)}^{1/2}u(t)}2^2.
\end{align*}
Since $u$ solves \eqref{eq:gH}, plugging the estimate of Lemma \ref{lem:estVu},
one sees that
\[
\norm{\nabla u}_{L^\infty_tL^2}^2 \leqslant C + C|t|^{\frac{2}{2-\kappa}}
\norm{\nabla u}_{L^\infty_tL^2}^{\frac2{2-\kappa}}
+C\Lebn{\nabla u(t)}2^{\frac{d}{\theta}}.
\]
By Young's inequality,
\begin{multline*}
\norm{\nabla u}_{L^\infty_tL^2}^2
\leqslant C + \left(\frac14 \norm{\nabla u}_{L^\infty_tL^2}^2 + C_{\kappa}
\left(C |t|^{\frac2{2-\kappa}}\right)^{\frac{2-\kappa}{1-\kappa}}\right)\\
+\left(\frac14 \norm{\nabla u}_{L^\infty_tL^2}^2 + C_\theta \right).
\end{multline*}
Thus, we conclude that
$\norm{\nabla u}_{L^\infty_tL^2} \leqslant C \Jbr{t}^{\frac1{1-\kappa}}$.
\end{proof}
\subsection{\texorpdfstring{Proof of Theorem \ref{thm:general}}{Proof of Theorem 3.4}}
We have shown that \eqref{eq:gmH} is globally well-posed in $\widetilde\Sigma^{1,1}_{V}$.
Let us next extend the global existence of solution to \eqref{eq:gH}
for a general data, that is, for a data which \emph{does not necessarily} satisfy
\eqref{eq:neutrality}.
\begin{proof}[Proof of Theorem \ref{thm:general}]
Take a nonzero $u_0 \in \Sigma^{1,1}_V$ and define a positive constant $M$
and $d$-dimensional vectors ${\bf a}$ and $ {\bf b}$ as in \eqref{def:M} and \eqref{def:ab}, respectively.
Set $v_0(x) =(\pi_{-\bf a}\tau_{-\bf b} u_0) (x)$.
One easily verifies that $v_0$ satisfies \eqref{eq:neutrality}.
Namely, $v_0 \in \widetilde\Sigma^{1,1}_{V}$.
Then, applying Theorem \ref{thm:LWPgmH} (2) and Lemma \ref{lem:H1bound},
we obtain a global solution
$\widetilde{u}$ of \eqref{eq:gH} which conserves the mass, the energy, and the momentum.
The solution depends continuously on $v_0$, and so on $u_0$
(Recall that $e^{itA}\phi$ is continuous with respect to the parameter $M$
in $A$).
We now define a function $u$ by $u= \exp({i \frac{|{\bf a}|^2}{2}t})
\tau_{{\bf a}t+{\bf b}} \pi_{{\bf a}} \widetilde{u}$
as in \eqref{def:ut}.
Then, $u$ belongs to the same class as $\widetilde{u}$ and solves \eqref{eq:gH} with $u(0)=u_0$.
The solution $u$ conserves the mass because
\[
\Lebn{u(t)}2 = \Lebn{\widetilde{u}(t)}2 = \Lebn{v_0}2 =\Lebn{u_0}2.
\]
Similarly,
\begin{align*}
\Lebn{\nabla u(t)}2^2
&{}=\Lebn{\nabla \widetilde{u}(t) + i {\bf a}\widetilde{u}(t)}2^2 \\
&{}=\Lebn{\nabla \widetilde{u}(t)}2^2 - 2{\bf a}\cdot
P[\widetilde{u}(t)]
+ |{\bf a}|^2 \Lebn{\widetilde{u}(t)}2^2
\end{align*}
and
\begin{align*}
\iint_{{\mathbb{R}}^{2d}}V{(x-y)} |u(t,x)|^2|u(t,y)|^2 dxdy
= \iint_{{\mathbb{R}}^{2d}}V{(x-y)} |\widetilde{u}(t,x)|^2|\widetilde{u}(t,y)|^2 dxdy
\end{align*}
hold. These give us the conservation of energy
\begin{multline*}
E[u(t)] = E[\widetilde{u}(t)] -
{\bf a}\cdot P[\widetilde{u}(t)]
+ \frac12|{\bf a}|^2 \Lebn{\widetilde{u}(t)}2^2 \\
=E[v_0] - {\bf a}\cdot P[v_0] + \frac12|{\bf a}|^2 \Lebn{v_0}2^2
= E[u_0]
\end{multline*}
and the conservation of momentum
\begin{align*}
P[u(t)]
&{}= \operatorname{Im} \int_{{\mathbb{R}}^d}\overline{\widetilde{u}(t,y)}(\nabla \widetilde{u}(t,y) +i{\bf a}\widetilde{u}(t,y))dy \\
&{}= P[\widetilde{u}(t)] +{\bf a}\Lebn{\widetilde{u}(t)}2^2
=P[v_0] +{\bf a}M = P[u_0].
\end{align*}
The estimate on $\Lebn{\nabla u(t)}2$ is given in Lemma \ref{lem:H1bound},
and then the estimate on $\Lebn{\Jbr{V(\cdot)}^{1/2} u(t)}2$
follows from Lemma \ref{lem:estVu}.
So far, we have shown all the statement
except for the uniqueness.
Let $w \in L^\infty([-T,T],\Sigma^{1,1}_V)\cap L^{8\zeta/d}([-T,T], W^{1,1}_{V,\frac{4\zeta}{2\zeta-1}})$ be another
solution of \eqref{eq:gH} with $w(0)=u_0$ which conserves $P[w(t)]$.
Then, $\partial_t w \in L^\infty((-T,T),(\Sigma^{1,1}_V)^\prime)$ holds from \eqref{eq:gH}.
We define
$\widetilde{v} \in L^\infty([-T,T],\Sigma^{1,1}_V)\cap L^{8\zeta/d}_{\mathrm{loc}}({\mathbb{R}}, W^{1,1}_{V,\frac{4\zeta}{2\zeta-1}})$
as in \eqref{def:ut}.
Then, $\widetilde{v}$ is also a solution of \eqref{eq:gH} and conserves the momentum.
Since $X[\widetilde{v}(0)]=P[\widetilde{v}(0)]=0$,
we can see that $X[\widetilde{v}(t)] \equiv 0$.
Now, $\partial_t \widetilde{v} \in L^\infty((-T,T),(\Sigma^{1,1}_V)^\prime)$.
Multiply \eqref{eq:gH} by $\overline{\widetilde{v}}$ and integrate its imaginary part
to yield
$\Lebn{\widetilde{v}(t)}2^2=\Lebn{\widetilde{v}(0)}2^2=\Lebn{u_0}2^2$.
Then, we deduce that $\widetilde{v}$ solves \eqref{eq:gmH}.
By the uniqueness of \eqref{eq:gmH}, we obtain $\widetilde{v}=v$.
Back to the transform \eqref{def:ut}, this implies $w=u$.
\end{proof}
\subsection{\texorpdfstring{Proof of Theorem \ref{thm:main1}}{Proof of Theorem 1.1}}\label{subsec:appl}
Now, we are in a position to complete the proof of our main theorem.
Set
\begin{equation}\label{def:VR}
V(x)=\lambda |x|^\gamma \chi(x), \quad R(x)=\lambda|x|^\gamma (1-\chi(x)),
\end{equation}
where $\chi$ is a smooth radial non-decreasing (with respect to $|x|$)
function such that $\chi\equiv1$ for $|x|\geqslant2$
and $\chi\equiv0$ for $|x|\leqslant1$.
One immediately sees that $R \in L^\infty$ and Assumptions $(\rm{R1})$ and $(\rm{R2})$ are fulfilled
with $\zeta=\theta=\infty$.
Thus, Theorem \ref{thm:main1} follows from Theorem \ref{thm:general}
if we prove that $V$ satisfies Assumptions
$(\rm{V1})$, $(\rm{V2})$, $(\rm{V3})$, and $(\rm{V4})$.
We shall demonstrate merely $(\rm{V3})$ since the others are trivial.
Remark that $(\rm{V2})$ holds with $\kappa = \frac{2(\gamma-1)}{\gamma}$ and
that $\kappa<1$ if and only if $\gamma<2$.
\begin{lemma}\label{lem:esttK}
Let $\gamma \in (1,2]$ and
\[
\widetilde{K}(x,y) = |x-y|^{\gamma} - |x|^{\gamma}+ \gamma |x|^{\gamma -2} x\cdot y.
\]
There exists a positive constant
$C$ depending only on $\gamma $ such that
\[
\sup_{x \in {\mathbb{R}}^d} |\widetilde{K}(x,y)|\leqslant C |y|^{\gamma }.
\]
\end{lemma}
\begin{proof}
The case $y=0$ is trivial. We hence fix $ {\mathbb{R}}^d \ni y \neq 0$.
It immediately follows that
\[
\sup_{x, |x|\leqslant 2|y|} |\widetilde{K}(x,y)| \leqslant 3^\gamma|y|^\gamma + 2^\gamma |y|^\gamma + \gamma 2^{\gamma-1}|y|^{\gamma}
\leqslant C |y|^\gamma.
\]
We now consider the case $|x| \geqslant 2|y|$.
An elementary computation shows that $K$ is written as
\begin{align*}
\widetilde{K}(x,y) ={}& \int_0^1 \frac{\partial}{\partial a}|x-ay|^\gamma da + \gamma |x|^{\gamma -2} x\cdot y \\
={}& -\int_0^1 \gamma |x-ay|^{\gamma-2} y\cdot(x-ay) da + \gamma |x|^{\gamma -2} x\cdot y \\
={}& \gamma |y|^2 \int_0^1 a |x-ay|^{\gamma-2} da
- \gamma x\cdot y\int_0^1 \int_0^1\frac{\partial}{\partial b}|x-bay|^{\gamma-2} db da \\
={}& \gamma |y|^2 \int_0^1 a |x-ay|^{\gamma-2} da \\
&{} + \gamma(\gamma-4) x\cdot y\int_0^1 \int_0^1 |x-bay|^{\gamma-4}ay\cdot(x-bay) db da\\
=:{}& \widetilde{K}_1(x,y) + \widetilde{K}_2(x,y).
\end{align*}
For any integer $m\geqslant2$, it holds that
\[
\sup_{x, m|y| \leqslant |x| \leqslant (m+1)|y|}|\widetilde{K}_1(x,y)|
\leqslant \gamma |y|^2 \int_0^1 a ((m-1)|y|)^{\gamma-2}da=\frac{\gamma}{2(m-1)^{2-\gamma}}|y|^\gamma.
\]
Therefore,
\[
\sup_{x,|x|\geqslant2|y|}|\widetilde{K}_1(x,y)| \leqslant \sup_{m\geqslant 2}\frac{\gamma}{2(m-1)^{2-\gamma}}|y|^\gamma
=\frac{\gamma}{2}|y|^\gamma
\]
Similarly, we have
\begin{align*}
&\sup_{x, m|y| \leqslant |x| \leqslant (m+1)|y|}|\widetilde{K}_2(x,y)|\\
&{}\leqslant \gamma(4-\gamma) (m+1)|y|^2\int_0^1 \int_0^1 ((m-1)|y|)^{\gamma-4}a((m+1)|y|^2) db da \\
&{}=\frac{\gamma(4-\gamma)(m+1)^2}{2(m-1)^{4-\gamma}}|y|^\gamma.
\end{align*}
Since $\sup_{m\geqslant 2}(m+1)^2(m-1)^{4-\gamma}=3^2$, we conclude that
\[
\sup_{x,|x|\geqslant2|y|}|\widetilde{K}_2(x,y)|
\leqslant \sup_{m\geqslant 2}\frac{\gamma(4-\gamma)(m+1)^2}{2(m-1)^{4-\gamma}}|y|^\gamma
=\frac{9\gamma(4-\gamma)}{2}|y|^\gamma,
\]
which completes the proof.
\end{proof}
\begin{proposition}
Let $V$ be as in \eqref{def:VR}.
If $\gamma \in (1,2)$ then $V$ satisfies Assumption $(\rm{V3})$
with $W(x)=-\lambda\gamma x \Jbr{x}^{\gamma-2}$.
More precisely, if we put
\[
K(x,y) =\lambda \chi(|x-y|)|x-y|^{\gamma} - \lambda\chi(|x|)|x|^{\gamma}+\lambda \gamma \Jbr{x}^{\gamma -2} x\cdot y,
\]
then there exists a positive constant
$C$ depending only on $\gamma $ and $\lambda$ such that
\[
\sup_{x \in {\mathbb{R}}^d} |K(x,y)|\leqslant C \Jbr{y}^{\gamma }, \quad
\sup_{x \in {\mathbb{R}}^d} |\nabla_x K(x,y)|\leqslant C \Jbr{y}.
\]
\end{proposition}
\begin{proof}
For simplicity, let $\lambda=1$.
Let $\widetilde{K}$ be as in Lemma \ref{lem:esttK}.
We deduce that
\[
\sup_{x \in {\mathbb{R}}^d}|\widetilde{K}(x,y)-K(x,y)|\leqslant
2^{1+\gamma}
+ \sup_{x \in {\mathbb{R}}^d}|\gamma (|x|^{\gamma -2} - \Jbr{x}^{\gamma -2}) x\cdot y|.
\]
Let us estimate the second term of the right hand side.
An elementary calculation shows
\[
|x|^\nu -\Jbr{x}^\nu
= -\int_0^1 \partial_a (a+|x|^2)^{\frac{\nu}2} da
= -\frac{\nu}{2}\int_0^1 (a+|x|^2)^{\frac{\nu}2-1} da
\]
for any $\nu$, and so
\[
\sup_{|x|\geqslant 1} |\gamma (|x|^{\gamma -2} - \Jbr{x}^{\gamma -2}) x\cdot y|
\leqslant \frac{\gamma(2-\gamma)}{2} |y|
\leqslant C\Jbr{y}.
\]
It is obvious that
\[
\sup_{|x|\leqslant 1} |
\gamma (|x|^{\gamma -2} - \Jbr{x}^{\gamma -2}) x\cdot y|
\leqslant C\Jbr{y}.
\]
Then, the first inequality follows from Lemma \ref{lem:esttK}.
Let us proceed to the second inequality.
Notice that
\begin{multline*}
\nabla_x K(x,y)
= \gamma|x-y|^{\gamma-2}(x-y)\chi(|x-y|) +|x-y|^\gamma \nabla \chi(|x-y|)
\\
-\gamma|x|^{\gamma-2}x\chi(|x|)- |x|^\gamma \nabla \chi(|x|)\\
+\gamma\Jbr{x}^{\gamma-2}y +\gamma(\gamma-2)\Jbr{x}^{\gamma-4}(x\cdot y)x.
\end{multline*}
Since $\nabla \chi(|x|) =0$ for $|x|\leqslant1$ and $|x|\geqslant2$, it holds that
\[
\sup_{x \in {\mathbb{R}}^d} ||x-y|^\gamma \nabla \chi(|x-y|)|
+ ||x|^\gamma \nabla \chi(|x|)| \leqslant 2^{\gamma+1}\Lebn{\nabla \chi}\infty .
\]
We also deduce that
\[
\sup_{x \in {\mathbb{R}}^d}
|\Jbr{x}^{\gamma-2}y|+ |\Jbr{x}^{\gamma-4}(x\cdot y)x|
\leqslant C|y|
\]
for $\gamma\in (1,2) $.
Now, It holds that
\begin{align*}
&|x-y|^{\gamma-2}(x-y)\chi(|x-y|) - |x|^{\gamma-2}x\chi(|x|) \\
&{}= \int_0^1 \partial_a |x-ay|^{\gamma-2}(x-ay)\chi(|x-ay|) da\\
&{}= (2-\gamma)\int_0^1 |x-ay|^{\gamma-4}y\cdot(x-ay)(x-ay)\chi(|x-ay|) da \\
&\quad - \int_0^1 |x-ay|^{\gamma-2}y\chi(|x-ay|) da \\
&\quad - \int_0^1 |x-ay|^{\gamma-3}y\cdot(x-ay) \chi^\prime(|x-ay|) da.
\end{align*}
Notice that $\sup_{x\in {\mathbb{R}}^d}\abs{|x|^{\gamma-2}\chi(|x|)}\leqslant 1$.
Hence, the first term of the right hand side of above equality is estimated as
\begin{align*}
&\sup_{x\in {\mathbb{R}}^d}\abs{\int_0^1 |x-ay|^{\gamma-4}y\cdot(x-ay)(x-ay)\chi(|x-ay|) da}\\
&\quad \leqslant |y| \sup_{0\leqslant a\leqslant1}\sup_{x\in {\mathbb{R}}^d}|x-ay|^{\gamma-2} \chi(|x-ay|) \leqslant |y|.
\end{align*}
One can obtain similar estimates for other terms. Thus, we conclude that
\begin{equation*}
\sup_{x \in {\mathbb{R}}^d}
\abs{|x-y|^{\gamma-2}(x-y)\chi(|x-y|) - |x|^{\gamma-2}x\chi(|x|)} \leqslant C|y|.
\end{equation*}
\end{proof}
|
\section{Introduction}
Massive stars strongly disturb the interstellar medium (ISM) via
emission of ionizing photons, stellar winds, and supernova explosions.
These processes produce overpressured hot bubbles
that expand into ambient interstellar clouds.
At the same time, a shock wave sweeps up the ambient clouds into
a dense shell.
The expanding shell strongly influences the dynamics of the ISM.
Especially, it is often supposed to
trigger the formation of stars through the gravitational fragmentation \citep{EL77}.
\citet{HI05,HI06} investigated the dynamical expansion of the HII region, the
photodissociation region, and the swept-up shell, solving
the UV and far-UV radiative transfer and thermal and
chemical processes in the one-dimensional (1D) hydrodynamics code.
They showed that the shell becomes cold and dense enough for the
gravitational instability (GI) to take place owing to the reformation of molecules destructed
by far-UV photons.
Numerous authors have discovered evidences for
the star formation in shell-like molecular clouds around hot bubbles.
\citet{C06,C07} have compiled a catalogue of $\sim600$ shells from
data in {\it Spitzer}-GLIMPSE survey.
Recently, \citet{Detal10} have investigated 102 samples identified as shells on the
{\it Spitzer}-GLIMPSE images at 8.0$\mu$m.
They found that 86\% of the shells are associated with ionized gases, or HII regions.
They obtained statistical properties of the triggered star formation, and
suggested that more than a quarter of the shells may have triggered the formation of massive stars.
This suggests that the trigger star formation process may be important in the massive
star formation.
To understand the triggered star formation,
it is important to investigate how and when the expanding shell fragments through the GI.
\citet{E94} and \citet{Wetal94} derived a simple dispersion relation taking into account
the mass accretion and the dilution effect of perturbations owing to the expansion.
They assumed the thin-shell approximation, and
neglect the boundary effect of the contact discontinuity (CD).
Recently, \citet[][hereafter Paper I]{IIT11} investigated stability of expanding shells
taking into account asymmetric density profiles with imposing an approximate shock boundary
condition on the leading surface and the CD boundary condition on the trailing surface.
They found that the dispersion relation and eigen-function strongly depends on
the boundary conditions and the degree of asymmetry of the density profile.
Although many authors have studied fragmentation of shells
by using linear analysis, it is still uncertain how and when shells fragment.
This is because the stability analysis of the evolving shell is
difficult to perform without any approximations.
Therefore, time-dependent multi-dimensional simulations are crucial to investigate it.
To resolve propagating thin shells all the time,
the Eulerian adaptive mesh refinement (AMR) technique in the mesh
code or the smoothed particle hydrodynamics (SPH) have been used.
However, since enormous meshes or SPH particles are required
to resolve the thickness,
researches based on numerical simulations are less advanced.
\citet{DBW07} investigated the gravitational fragmentation of the shell driven
by the expansion of the HII region by using three-dimensional (3D)
SPH taking into account the radiative transfer of ionizing photons.
In their simulation, its thickness was not resolved sufficiently.
\citet{Detal09} have investigated fragmentation of shells with fixed
mass that expand into a hot rarefied gas by using
two different numerical schemes, the Eulerian AMR method
(the {\sc flash} code) and the SPH method.
They considered shells confined by a constant pressure on both surfaces.
In the more realistic situation where shells are confined by
the CD and the shock front (SF),
multi-dimensional simulations including the
self-gravity have not been performed yet.
In this paper, we perform the 3D simulation to investigate
the fragmentation process of expanding shells through GI.
As numerical method, we adopt the 3D SPH method.
In order to attain high resolution enough to resolve the thickness,
we calculate not the whole but a part of the shell.
The outline of the paper is as follows:
in Section \ref{sec:simulations}, we describe our model and simulations in detail.
Brief review of Paper I is presented in Section \ref{sec:review}.
The results of our simulations are shown in Section \ref{sec:sph result}.
In Sections \ref{sec:discuss} and \ref{sec:summary}, we present discussions and summaries of
our paper.
\section{Simulations}\label{sec:simulations}
\subsection{Numerical Methods}\label{sec:num method}
We adopt the SPH method \citep[e.g.,][]{M92}.
Since SPH is Lagrangian particle method, it is one of the
best method for problems where a wide low density region
and geometrically thin shell coexist.
Instead of additional viscosity to handle shocks,
we use the ``Godunov SPH'' where the results of the Riemann problem
are used to calculate the interaction between SPH particles \citep{I02}.
The tree method \citep{BH86} is used to calculate self-gravitational force.
\subsubsection{Treatment of Contact Discontinuity}\label{sec:sph}
In the standard SPH method \citep{M92},
the $i$-th particle has the mass of $m_i$, and its density
is given by
\begin{equation}
\rho_i = \sum_k m_k W({\bf x}_i-{\bf x}_k,\bar{h}_{ik}),
\label{eoc sph}
\end{equation}
where $W({\bf x},h)$ is kernel function and we use the Gaussian kernel, $h$ is
the smoothing length, and $\bar{h}_{ik}$ is an average between
$h_i$ and $h_k$.
The equation of motion of the $i$-th particle is given by
\begin{equation}
\frac{\Delta {\bf v}_i}{\Delta t}
= - \sum_k m_k \left( \frac{P_i}{\rho_i^2} + \frac{P_k}{\rho_k^2} \right)
\frac{\partial}{\partial {\bf x}_i}W({\bf x}_i - {\bf x}_k,
\bar{h}_{ik}).
\label{eom sph}
\end{equation}
The standard SPH has a shortcoming when calculating the CD
between a hot and a cold gas \citep{A07}.
\citet{A07} reported that the artificial tension stabilizes
the Kelvin-Helmholtz instability with a different densities in
the standard SPH.
In contrast, \citet{C10} have shown that
Godunov SPH can describe the Kelvin-Helmholtz instability
reasonably well.
In addition,
\citet{TM05} and \cite{HA06} proposed a new method that treats the CD naturally in
the context of a multi-phase fluid.
They modified Equations (\ref{eoc sph}) and
(\ref{eom sph}) to treat the CD as follows:
\begin{equation}
\rho_i = m_i \sum_k W({\bf x}_i-{\bf x}_k,\bar{h}_{ik}),
\label{eoc sph modi}
\end{equation}
and
\begin{eqnarray}
\frac{\Delta {\bf v}_i}{\Delta t}
&=&
- \frac{1}{m_i}\sum_k \left[ P_i\left(\frac{m_i}{\rho_i}\right)^2
+ P_k\left(\frac{m_k}{\rho_k} \right)^2\right]\nonumber\\
&&\hspace{3cm}\times\frac{\partial}{\partial {\bf x}_i}
W({\bf x}_i - {\bf x}_k.
\bar{h}_{ik}).
\label{eom sph modi}
\end{eqnarray}
The term of $m_i/\rho_i$ represents the volume of $i$-th particle $\sim h_i^3$.
Therefore, if we determine particle mass of hot and cold gases so that its smoothing length
distributes smoothly across the CD, the terms inside the bracket in Equation (\ref{eom sph modi})
is smooth across the CD, leading to much better behavior at the CD.
Their new method is very useful in problems where
the position of the CD is known in advance as in our modeling.
Equation (\ref{eom sph modi}) satisfies the relation of
$m_i\Delta {\bf v}_i = - m_k \Delta{\bf v}_k$, suggesting that the momentum conservation
is guaranteed.
Furthermore,
we apply the Godunov technique to Equation (\ref{eom sph modi}) as
\begin{eqnarray}
\frac{\Delta {\bf v}_i}{\Delta t}
&=&
- \frac{1}{m_i}\sum_k P^*_{ik}\left[ \left(\frac{m_i}{\rho_i}\right)^2
+ \left(\frac{m_k}{\rho_k} \right)^2\right]\nonumber\\
&&\hspace{3cm}\times\frac{\partial}{\partial {\bf x}_i}
W({\bf x}_i - {\bf x}_k,
\bar{h}_{ik}).
\label{eom sph godunov}
\end{eqnarray}
where $P^*_{ik}$ is the result of the Riemann problem between
the $i$-th and the $k$-th particles \citep{I02}.
\subsection{Models}\label{sec:models}
Massive stars emit ultraviolet photons ($h\nu>13.6$ eV) and produce
HII regions around them.
Here, we consider a massive star that emits ionizing photons with the photon number luminosity
$Q_\mathrm{UV}\:[\mathrm{s}^{-1}]$
into the ambient gas with the uniform density of $\rho_\mathrm{E}=mn_\mathrm{E}$, where
$n_\mathrm{E}$ and $m$ are the number density and the mean mass of the ambient gas particle,
respectively.
We construct as simple model as possible
to concentrate on the physics of GI of expanding shells.
Figure \ref{setup} illustrates the schematic picture of our model.
In this paper, we do not calculate the
radiative transfer of ionizing photons but introduce hot and cold gases that
are assumed to evolve keeping their temperatures constant.
The hot gas is occupied in $r<R_\mathrm{CD}$ (the dotted region), and the thin shell is in
$R_\mathrm{CD}<r<R_\mathrm{SF}$, where $R_\mathrm{CD}$ and
$R_\mathrm{SF}$ are positions of the CD and the SF, respectively.
The pre-shock ambient cold gas in $r>R_\mathrm{SF}$ is assumed to be uniform.
In our model, the inner boundary is set at $r=R_\mathrm{b}$ inside the hot
bubble $(R_\mathrm{b}<R_\mathrm{CD})$.
In the HII region, the detailed balance between the recombination and the ionization
is approximately established all the time.
Therefore, the interior pressure of the HII region is given by
\begin{equation}
P_\mathrm{int} = \rho_\mathrm{E} c_\mathrm{II}^2 \left( \frac{R_\mathrm{ST}}{R_\mathrm{CD}} \right)^{3/2},
\label{Pext}
\end{equation}
where $R_\mathrm{ST}$ is the Str{\"o}mgren radius,
\begin{equation}
R_\mathrm{ST} = \left( \frac{3Q_\mathrm{UV}}{4\pi \alpha_\mathrm{B} n_\mathrm{E}^2} \right)^{1/3}.
\label{stremgren}
\end{equation}
We assume that the pressure of the HII region is spatially constant because of the high temperature.
The hot gas at $r=R_\mathrm{b}$ is pushed by the interior pressure $P_\mathrm{int}$.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f1.eps}
\end{center}
\caption{
Schematic picture of our model.
The hot gas occupying the dotted region
is embedded by the cold gas.
}
\label{setup}
\end{figure}
\subsection{Calculation Domain}\label{sec:domain}
If we simulate the overall shell, the required number of
SPH particles, $N_\mathrm{tot}$, is too large to calculate.
Therefore, to save $N_\mathrm{tot}$,
we calculate a part of the shell.
The schematic picture of the calculation domain is shown in
Figure \ref{fig:region}, and
is designated by
$|\tan^{-1}(y/x)|\le\theta_\mathrm{b}$ and $|\tan^{-1}(z/x)|\le\theta_\mathrm{b}$.
Since the solid angle subtended by the calculation domain
from the center $O$ is given by
\begin{equation}
\Omega_\mathrm{b} \simeq 4 \theta_\mathrm{b}^2\;\;\;\mathrm{for}\;\;\theta_\mathrm{b}\ll1,
\end{equation}
the total number of SPH particles can be reduced
by a factor of $\Omega_\mathrm{b}/4\pi\sim\theta_\mathrm{b}^{2}/\pi$
compared with the calculation of the whole shell.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f2.eps}
\end{center}
\caption{Schematic picture of the calculation domain.}
\label{fig:region}
\end{figure}
\subsection{Simulation Units}
In this paper, for convenience, the units of the time, length,
and mass scales are taken to be
\begin{equation}
t_0 = \sqrt{\frac{3\pi}{32G\rho_\mathrm{E}}} =
1.6\;n_\mathrm{E,3}^{-1/2}\;\mathrm{Myr},
\label{t scale}
\end{equation}
\begin{equation}
R_0 = \left( \frac{7 c_\mathrm{II} t_0}{\sqrt{12}} \right)^{4/7} R_\mathrm{ST}^{3/7}
= 5.9\;
Q_\mathrm{UV,49}^{1/7}\:
n_\mathrm{E,3}^{-4/7}\;\mathrm{pc},
\label{r scale}
\end{equation}
and
\begin{equation}
M_0 = \rho_\mathrm{E} R_0^3
= 5.0\times10^3\;
Q_\mathrm{UV,49}^{3/7}\:
n_\mathrm{E,3}^{-5/7}\;M_\odot,
\label{m scale}
\end{equation}
respectively, where $Q_\mathrm{UV,49}=Q_\mathrm{UV}/10^{49}\:\mathrm{s}^{-1}$,
and $n_\mathrm{E,3}=n_\mathrm{E}/10^3\:\mathrm{cm^{-3}}$.
The dependence of the expansion law on the parameters $(Q_\mathrm{UV},\:n_\mathrm{E})$ are
approximately eliminated by using the non-dimensional quantities normalized by
$t_0$, $R_0$, and $M_0$ (see Figure 1 in Paper I).
\subsection{Initial and Boundary Conditions}\label{sec:ini bou}
It is difficult to resolve
very thin shell in early evolutionary phase even if
we calculate the part of the shell (see Figure \ref{fig:region}).
Thus, we use a grid-based 1D hydrodynamical
code to describe the early evolutionary phase of the HII region
expansion, and we switch to 3D SPH at $t=0.4t_0$.
As a grid-based numerical method, we use
the 1D spherically symmetric Lagrangian
Godunov method \citep{vL97}.
The innermost mesh at $R_\mathrm{b}$ is pushed by the interior
pressure given by Equation (\ref{Pext}).
When the 1D simulation reaches $t/t_0=0.4$,
the radial profiles of physical quantities
are used as the initial condition of the 3D calculations.
We smooth the distribution of the sound speed across the CD
in the scale of the smoothing length.
It is assumed that
the temperature of the hot gas is 64 times as high as
that of the cold gas.
The specific value of the
hot gas temperature does not change our conclusion as long as
it is much larger than the temperature of the cold gas.
We calculate the expansions of HII regions around the
41$M_{\odot}$ (the high-mass (HM) model) and 19$M_\odot$ (the low-mass (LM) model) stars
that are embedded by the uniform ambient gas of $n_\mathrm{E}=10^3$ cm$^{-3}$.
Simulation parameters are listed in Table \ref{table:model para}.
The temperature of the cold gas $T_\mathrm{c}$ is assumed to be $10$ K.
The opening angles of simulation domains are set to
$\theta_\mathrm{b}=2\pi/26$ and $2\pi/14$ for
the HM and the LM models, respectively, so that
the calculation domains contain a single wavelength of the most
unstable mode predicted from the thin-shell linear analysis by \citet{E94}.
The mass of SPH particles are set
so that the initial thickness is resolved by five SPH particles.
Since the thickness increases with time,
the relative resolution becomes better as the shell expands.
In the later phase, the thickness is resolved by
$\sim15$ particles.
The total number of SPH particles for the HM and the LM models are
$4.00\times10^6$ and $2.26\times10^6$, respectively.
Corrugation-type perturbations are put into the shell at
the initial state.
We displace the SPH particles so that their densities do not change,
or ${\bf \nabla}\cdot \Delta {\bf x}_i=0$,
where $\Delta{\bf x}_i$ is the displacement of the $i$-th particle.
The functional form of the
displacement of the shell is assumed to be $\propto -\cos (l\phi)$,
where $\phi = \tan^{-1}(y/x)$.
Therefore, the shell has the negative displacement at $\phi=0$.
The SPH particles are displaced keeping their velocity,
the sound speed, and the mass fixed.
We concentrate on the evolution of a single mode in each calculation.
Dependence on $l$ is investigated by many runs.
The moving boundary condition at $r=R_\mathrm{b}(t)$
is realized by using ``ghost particles'' \citep{TMS94}.
The time evolution of $R_\mathrm{b}(t)$ is obtained by the results of
the 1D simulations.
At the four surface areas ($y=\pm x\tan\theta_\mathrm{b}$, $z
=\pm x\tan\theta_\mathrm{b}$)
in the pyramid-shaped calculation domain (see Figure \ref{fig:region}),
the rotational periodic boundary condition is imposed.
Since the gravitational force is a long-range force,
we have to take into account the particles in the whole solid angle.
However, we only have information in the part of the solid angle
as shown in Section \ref{sec:domain}.
In this paper, the gravitational force is evaluated in the following method.
Figure \ref{fig:grav region} shows the schematic picture of the gas sphere viewed from the $x$-direction for the
case with $\theta_\mathrm{b}=2\pi/10$.
The latitude and longitude lines are plotted at interval of $\theta_\mathrm{b}$.
The region is surrounded by the thick solid lines represents the calculation domain (see Figure \ref{fig:region}).
We rotate the calculation domain $i\theta_\mathrm{b}$ degrees in the $\theta$-direction, where
$i=-\pi/(2\theta_\mathrm{b}), \ldots, \pi/(2\theta_\mathrm{b})$.
At each rotation step in the $\theta$-direction, we rotate the domain $k\theta_\mathrm{b}$ degrees
in the $\phi$-direction, where $k=1,\ldots,2\pi/\theta_\mathrm{b}$.
After that, the information in the whole solid angle can be obtained.
However, for $i\ne0$, the rotated domains in the $\phi$-direction have overlaps.
This comes from the fact that the pyramid cannot fill 3D space.
For example, the dashed regions in Figure \ref{fig:grav region} correspond to the rotated domains in
the $\theta$-direction.
We remove the overlapped SPH particles with $|\tan^{-1}(y/x)|>\theta_\mathrm{b}$ from
the rotated domains. In the simulation, we rotate not SPH particles but the tree structure, and
calculate the gravitational force from the rotated tree structures at each rotation step.
By this method, supposed particle distribution in the whole solid angular $4\pi$ is
not cyclic completely. However, the effect of the approximation is negligible since
$\theta_\mathrm{b}$ is very small in this paper.
For $i=\pm1$ (near the equatorial plane), the number of removed SPH particle is
negligible compared with the total number.
On the other hand, for large $|i|$ (near the poles), the fraction of removed SPH particles becomes large.
However, since the rotated domain is very far from the calculation domain,
the detail particle distribution is not important.
In this paper, we neglect the gravitational force outside the SF
to prevent the ambient gas from collapsing toward the center.
We find the radius $R_\mathrm{grav}$ of the most distant particle from the center that
has density $\rho>\rho_\mathrm{th}$, where $\rho_\mathrm{th}$ is a threshold density \citep{Betal09}.
In our simulations, we adopt $\rho_\mathrm{th}=1.1\rho_\mathrm{E}$.
Only SPH particles within $R_\mathrm{grav}$ are assumed to feel the gravitational force that is calculated
above method. This means that there is the discontinuity in the gravitational force at
the $R_\mathrm{grav}$. In the Appendix, we investigate the effect of the discontinuity, and
confirm that it does not influence our results as long as $R_\mathrm{grav}$ is inside the shock transition layer.
The main reason is that the pressure suddenly changes in the shock transition layer, suggesting that
the pressure gradient is much larger than the jump in the gravitational force.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f3.eps}
\end{center}
\caption{
Schematic picture viewed from the $x$-direction for the case with $\theta_\mathrm{b}=2\pi/10$.
The latitude and longitude lines are plotted at interval of $\theta_\mathrm{b}$.
The region is surrounded by the thick solid lines represents the calculation domain
(see Figure \ref{fig:region}).
The dashed regions correspond to rotated domains in the $\theta$-direction.
}
\label{fig:grav region}
\end{figure}
\begin{table*}
\centering
\begin{tabular}{cccccc}
\hline
Model & $M_*(M_\odot)^a$ & $R_\mathrm{ST}(\mathrm{pc})$ & $R_0(\mathrm{pc})$ & $T_0$(Myr) & $M_0(M_\odot)$ \\
\hline
\hline
High mass (HM)& 41 & 0.56 & 5.5 & 1.6 & $4.1\times10^3$ \\%& $0.087$ \\
\hline
Low mass (LM)& 19 & 0.25 & 3.9 & 1.6 & $1.46\times10^3$ \\%& $0.12$ \\
\hline
\end{tabular}
{\footnotesize
\begin{flushleft}
$^a$ Mass of the star \\
\end{flushleft}
}
\caption{Parameters of Our Models}
\label{table:model para}
\end{table*}
\subsection{Measures of Fluctuations}\label{sec:focus}
In order to evaluate growth of perturbations in the SPH calculations,
we introduce indicators for fluctuations of density and the position of the CD.
The solid angle $\Omega_\mathrm{b}$ is divided into $N_\Omega$ cells, and
the direction of the $i$-th cell is defined by the
unit vector ${\bf n}_i$ $(1\le i\le N_\Omega)$.
The radial profiles of the physical quantities along ${\bf n}_i$
are obtained from the SPH calculations.
Here, we define the angle-dependent
maximum density $\rho_\mathrm{max}({\bf n}_i)$ along ${\bf n}_i$.
The position of the CD along ${\bf n}_i$ is determined as the
position where the temperature coincides with a critical value
that is set to $\sqrt{2}T_\mathrm{c}$.
The specific choice of the critical
value is not important
in our results.
We average $\rho_\mathrm{max}({\bf n}_i)$ and $R_\mathrm{CD}({\bf n}_i)$
over all directions.
The mean value of $Q=(\rho_\mathrm{max}$, $R_\mathrm{CD})$
is defined by
\begin{equation}
\langle Q \rangle = \frac{1}{N_\Omega}\sum_{i} Q({\bf n}_i).
\label{Qave}
\end{equation}
As the indicators of the fluctuations,
we evaluate the dispersions of these quantities
\begin{equation}
\Delta Q \equiv \sqrt{\frac{1}{N_\Omega}\sum_{i}
\left\{ Q({\bf n}_i) - \langle Q\rangle\right\}^2}.
\label{delta Q}
\end{equation}
\subsection{Unperturbed Evolution}\label{sec:without per}
As a test calculation, we present the evolution of the shell
without initial perturbations in the HM model.
Figure \ref{fig:surpos} shows the time evolution
of $\langle R_\mathrm{CD}\rangle$ (Equation (\ref{Qave}))
evaluated from the SPH simulation (the open circles).
For comparison, we plot the result of the 1D simulation by the solid line.
One can see that the SPH simulation can reproduce the results
of the 1D calculation very well.
Figures \ref{fig:1dim} show snapshots of density (the upper panel)
and pressure (the lower panel) profiles
for $t/t_0=0.5$, 0.7, 1.0, and 1.26.
The abscissae indicate the distance from the CD at each time.
Only SPH particles in $|y|/R_0<0.05$ and $|z|/R_0<0.05$ are plotted
by the dots.
For comparison, the density profiles in the 1D simulation are
superimposed by the thick gray lines in the upper panel.
It is seen that the result of the SPH calculation agrees with that of
the 1D simulation very well.
Our 3D simulation clearly represents the structure and evolution of
the shell although it is very thin.
Note that the pressure profiles in the lower panel of Figure \ref{fig:1dim}
are smooth at the CD $(r-\langle R_\mathrm{CD}\rangle=0)$ thanks to
the modified Godunov SPH shown in Section \ref{sec:sph}.
In the standard SPH, a wiggle in pressure profile appears at the CD.
In order to see whether the pressure profile is smooth even when perturbation is added,
the density and pressure profiles for $t=1.0$ and $l=52$ are plotted in
Figure \ref{fig:1dim per}.
The detailed investigation of the results with perturbation is shown in Section \ref{sec:sph result}.
The abscissa is the distance from $\langle R_\mathrm{CD}\rangle$.
We plot SPH particles in $\phi_0-\Delta \phi<\phi<\phi_0+\Delta \phi$ and $|z|/R_0<0.05$, where
$\Delta \phi=0.001\theta_\mathrm{b}$.
Figures \ref{fig:1dim per}(a) and (b) correspond to $\phi=0$ where $\delta \rho$
has the maximum value and $\phi_0=\theta_\mathrm{b}/2$ where $\delta \rho$ has the minimum value,
respectively. One can clearly see that the pressure profile is smooth at the CD even with perturbation.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f4.eps}
\end{center}
\caption{
Time evolution of $\langle R_\mathrm{CD}\rangle$. The open circles indicate
the result of the SPH calculation. The solid line
indicates the result of the 1D calculation.
}
\label{fig:surpos}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f5.eps}
\end{center}
\caption{
Snapshots of density (the upper panel) and pressure
(the lower panel) profiles for
$t/t_0=0.5$, 0.7, 1.0, and 1.26.
Only SPH particles in $|y|/R_0<0.05$ and $|z|/R_0<0.05$ are plotted by
the dots.
The abscissae are the distance from the CD.
The thick gray lines in the upper panel indicate the results of the
1D simulation.
}
\label{fig:1dim}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f6.eps}
\end{center}
\caption{
Snapshots of the density and pressure profiles for $t/t_0=1.0$ and $l=52$.
The abscissae are the distance from the average position of the CD.
SPH particles in $\phi_0-\Delta \phi<\phi<\phi_0+\Delta \phi$ and $|z|/R_0<0.05$ are plotted,
where $\phi_0=$(a)0, (b)$\theta_\mathrm{b}/2$, and $\Delta \phi=0.001\theta_\mathrm{b}$.
}
\label{fig:1dim per}
\end{figure}
\section{Prediction by Linear Analysis}\label{sec:review}
\citet{IIT11} performed a linear analysis taking into account the thickness
of the shell as in Figure \ref{fig:1dim}
and imposing the approximate SF and the CD boundary conditions.
They neglected evolutionary effect, i.e., expansion and mass accretion through
the SF. At each instant of time, they solved eigen-value problem
and obtained the growth rate $\Gamma(l,t)$ as
a function of the time and the angular wavenumber, where
$\Gamma$ is the imaginary part of $\omega(k,t)$ in Paper I.
The angular wavenumber $l$ can be expressed as $2\pi R_\mathrm{s}/k$ by using
the wavenumber $k$ in Paper I, where $R_\mathrm{s}$ is the shell radius.
\subsection{Time Evolution and Scaling Law of Density Profile}\label{sec:den pro}
The density profile of the decelerating shell is asymmetric with respect to the density peak.
\citet{IIT11} developed a semi-analytic method for deriving the time evolution of the
density profile by assuming that the shell reaches the hydrostatic equilibrium
among the pressure gradient, the inertia force owing to the deceleration, and
the radial gravitational force toward the center \citep{WF02}.
\citet{IIT11} found that the time evolution of the density profile
can be divided into three evolutionary phases, deceleration-dominated, intermediate,
and self-gravity-dominated phases (see Figure 2 in Paper I).
In the early phase (deceleration-dominated phase),
the density peak is at the SF because the inertia force is larger than the gravitational force.
As the shell expands, the inertia force decreases while the gravitational force increases.
Thus, the density peak moves from the SF toward the CD as shown in Figure \ref{fig:1dim}.
In the intermediate phase, the density peak is closer to the SF than the CD.
In the later phase (self-gravity-dominated phase), the density peak is closer to the CD than
the SF.
It is also found that the density profiles
for various sets of $(n_\mathrm{E},\;Q_\mathrm{UV})$
are characterized by a single parameter, that is the
typical Mach number,
\begin{equation}
{\cal M}_0
= \frac{4}{7}\frac{R_0}{c_\mathrm{s} t_0}= 7\;
Q_\mathrm{UV,49}^{1/7}\:
T_\mathrm{c,10}^{-1/2}\:
n_\mathrm{E,3}^{-1/14},
\label{mach}
\end{equation}
where $T_\mathrm{c,10}=T_\mathrm{c}/10$ K.
They found the development of the GI is strongly influenced by
two effects, asymmetry of the density profile and boundary conditions.
\subsection{Asymmetry of the Density Profile}\label{sec:asym}
The asymmetry of the density profile strongly influences the GI, especially,
in the self-gravity-dominated phase (for example, the shell at $t/t_0=1.26$ in Figure \ref{fig:1dim}).
In this later phase, the distance from the density peak to the SF is larger than $H_0$, where
$H_0=c_\mathrm{s}/\sqrt{2\pi G\rho_{00}}$
is the scale height of the shell and $\rho_{00}$ is the peak density.
On the other hand, the distance from the density peak to the CD is smaller than $H_0$.
The gas tends to collect toward the density peak because the unperturbed gravitational
potential has the minimum value there.
The gas near the SF can collapse toward the density peak because the sound wave cannot
travel from the density peak to the SF. This mode is ``compressible mode''.
On the other hand, the gas near the CD cannot collapse to
the peak. However, the GI can proceed even there through the deformation of the CD
that makes the gravitational potential deeper.
This mode is ``incompressible mode'' \citep{EE78, LP93}.
Therefore, the GI of the shell in the later phase has properties of
compressible and incompressible modes.
Moreover, \citet{IIT11} found that the growth rate is enhanced compared with symmetric case through
the combination of the GI and the Rayleigh-Taylor instability.
\subsection{Boundary Conditions}\label{sec:boundary}
Expanding shell is confined by the SF on the leading surface and
by the CD on the trailing surface.
Dispersion relations of \citet[][E94]{E94} and \citet{Wetal94} are essentially
based on the dispersion relation of the layer confined by the SF on both surfaces.
The shock-confined layer is stabilized
by the tangential flow generated by obliqueness of the SF compared with
the layer confined by thermal pressure of hot gas.
Therefore, the linear analysis in Paper I taking into account SF+CD boundary conditions
gives larger growth rate compared with E94.
In summary, as described in Sections \ref{sec:asym} and \ref{sec:boundary},
Paper I predicts that
the GI begins to grow earlier than the prediction from E94.
The development of the GI in the later phase is accompanied by
the significant deformation of the CD.
\section{Results}\label{sec:sph result}
\subsection{The Case with High-mass Central Star}
First, we present the results of the HM model with perturbations.
We calculate four runs for the case with $l=26$, 52, 78, and 156.
Figures \ref{fig:model1_growth}(a), (b), (c), and (d)
represent the time evolution of the perturbation amplitude
for the wavenumber $l=26$, 52, 78, and 156, respectively.
The thick solid and the thick dashed lines correspond to
the density perturbation $\Delta \rho_\mathrm{max}/\langle\rho_\mathrm{max}\rangle$ and
the deformation of the CD
$30\times\Delta R_\mathrm{CD}$, respectively.
In each panel, $\Delta \rho_\mathrm{max}/\langle\rho_\mathrm{max}\rangle$ and
$\Delta R_\mathrm{CD}$ grow in the similar way.
Perturbations with $l=52-78$ grow with the largest growth rate.
For larger angular wavenumber $l=156$ and smaller
angular wavenumber $l=26$, they grow more slowly.
The prediction from E94 is also superimposed by the thin dashed line in each panel.
It is found that in our 3D results, perturbations begin to grow earlier than the
prediction of E94.
Let us compare the results of SPH simulations with the linear analysis in Paper I.
However, Paper I neglected the evolutionary effects, i.e., expansion of the shell
and the mass accretion through the SF.
To include these effects approximately, we modify the instantaneous growth rate from
$\Gamma$ to $\Gamma_\mathrm{evo}$ as follows:
\begin{equation}
\Gamma_\mathrm{evo}(l,t) \equiv -3\frac{V_\mathrm{s}(t)}{R_\mathrm{s}(t)}
+ \sqrt{\left( \frac{V_\mathrm{s}(t)}{R_\mathrm{s}(t)} \right)^2 + \Gamma(l,t)^2},
\label{mod}
\end{equation}
where $V_\mathrm{s}$ is the velocity of the shell.
This modification is inspired by E94.
The time evolution of $R_\mathrm{s}(t)$ and $V_\mathrm{s}(t)$ is determined by the thin-shell model
shown in Section 2 in Paper I.
One can see that the terms of $V_\mathrm{s}/R_\mathrm{s}$
correspond to evolutionary effect of the shell that stabilizes the GI.
The prediction based on
Paper I is superimposed by the thin solid line in each panel of
Figure \ref{fig:model1_growth}. It is evaluated by
$\exp\left[\int_t \Gamma_\mathrm{evo}(l,t') \mathrm{d} t'\right]$.
From Figure \ref{fig:model1_growth},
one can see that the linear analysis of Paper I describes the growth of perturbations very well.
Figure \ref{fig:den_l52} shows the time sequence of
cross sections through the $z=0$ plane for $l=52$.
The color scale indicates density, and
the arrows represent relative velocity to
the fluid with the maximum density.
The abscissa and the ordinate axes correspond to the
azimuthal angle $\phi$,
and the distance from the CD divided by $\langle R_\mathrm{CD}\rangle$,
respectively.
Figure \ref{fig:den_l52}(a) represents the initial condition
where the shell around $\phi=0$ is displaced to the negative radial direction.
Figure \ref{fig:den_l52}(b) shows that the tangential flow
is generated by the obliqueness of the SF, and
collects the gas around $\phi=0$.
As a result, enhanced pressure pushes the SF outward in the
radial direction (see Figure \ref{fig:den_l52}(c)).
The gravitational potential around $\phi=0$ becomes deep, leading to
gas accumulation (see Figure \ref{fig:den_l52}(d)).
This mode of the GI is similar to the incompressible mode in the sense that
the deformation develops at the boundary \citep{EE78,LP93}.
In the later phase (Figure \ref{fig:den_l52}(e)),
the CD highly deforms while the SF is almost flat.
This is consistent with prediction of Paper I (see Section \ref{sec:review}).
Figure \ref{fig:den_l52}(e) is very similar to Figure 11 in Paper I that shows the
eigenfunction in parameters of the HM model at $t/t_0=1.3$ for $l=52$.
Figure \ref{fig:den_l156} is the same as Figure \ref{fig:den_l52} but
for larger wavenumber case, $l=156$.
The initial state is shown in Figure \ref{fig:den_l156}(a).
Because of large wavenumber, the tangential flow collects the gas
more quickly in Figure \ref{fig:den_l156}(b) that is similar to Figure
\ref{fig:den_l52}(c).
In this moment, since the SF at $\phi=0$ deforms to the positive radial direction,
the gas moves away from $\phi=0$
by the tangential flow. As a result, at $t/t_0=0.8$ (Figure \ref{fig:den_l156}(c)),
the SF and the CD are almost flat while the velocity field exists inside the shell.
The density distribution at $t/t_0=1.0$
(Figure \ref{fig:den_l156}(d))
has the opposite phase compared to the initial state.
In this moment, the perturbation begins to grow but the growth rate is smaller than that
for $l=52$.
The CD deforms largely at the final state in both of
Figures \ref{fig:den_l52} and \ref{fig:den_l156}.
In this sense, the behavior of the GI is similar to that for $l=52$.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f7.eps}
\end{center}
\caption{
Time evolution of perturbations for the HM model.
Each panel corresponds to $l=$ (a) 26, (b) 52, (c) 78, and (d) 156.
The thick solid and the thick dashed lines represent the density perturbation
$\Delta \rho_\mathrm{max}/\langle\rho_\mathrm{max}\rangle$ and the deformation of the CD
$30\times\Delta R_\mathrm{CD}$, respectively.
The thin solid lines correspond to the prediction from Paper I.
The thin dashed lines correspond to the prediction based on \citet{E94}.
}
\label{fig:model1_growth}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f8.eps}
\end{center}
\caption{
Cross sections through the $z=0$ plane for $l=52$.
Each panel shows the shell at
$t/t_0=(a)0.4$, (b)0.6, (c)0.8, (d)1.0, and (d)1.3, respectively.
The color scale represents the density, and arrows are velocities relative to
the fluid at the density peak.
The abscissa and the ordinate axes correspond to the azimuthal angle $\phi=\tan^{-1}(y/x)$,
and $R/\langle R_\mathrm{CD}\rangle-1$, respectively.
}
\label{fig:den_l52}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f9.eps}
\end{center}
\caption{
Same as Figure \ref{fig:den_l52} but for $l=156$.
}
\label{fig:den_l156}
\end{figure}
\subsection{The Case with Low Mass Central Star}
Next, we present results of the LM model.
Because the central star is less energetic,
the Mach number of the expanding shell is smaller than that in the
HM model,
suggesting that the shell is geometrically thicker.
Therefore, perturbations are expected to grow more slowly than the HM model.
This feature is confirmed in Figure
\ref{fig:model2_growth} that is similar to
Figure \ref{fig:model1_growth}.
The most unstable mode is found in smaller wavenumber $l\sim28$ and
the growth rate is smaller than the HM model.
We found that perturbations begin to grow earlier than the
prediction of E94 also in this case.
The linear analysis in Paper I describes the GI very well.
Figure \ref{fig:den_l56}
shows the cross section through $z=0$ for $l=56$ at $t/t_0=1.5$.
The significant deformation of the CD is seen as in the HM model.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f10.eps}
\end{center}
\caption{
The same as Figure \ref{fig:model1_growth} but for the LM model.
Each panel corresponds to $l=$(a)14, (b)28, (c)56, and (d)84.
}
\label{fig:model2_growth}
\end{figure}
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f11.eps}
\end{center}
\caption{
Cross section through $z=0$ plane for $l=56$ at $t/t_0=1.5$ for the
LM model.
The color scale represents the density, and the arrows
show velocities relative to
the fluid at the density peak.
The abscissa and ordinate axes are the same as Figure \ref{fig:den_l52}.
}
\label{fig:den_l56}
\end{figure}
\section{Discussion}\label{sec:discuss}
\subsection{Growth Rate of Gravitational Instability}
\begin{figure*}[htpb]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=8cm]{f12a.eps}
&
\includegraphics[width=8cm]{f12b.eps}
\end{tabular}
\end{center}
\caption{
Contour maps of (a)$t_\mathrm{bgn}$, (b)$R_\mathrm{s,bgn}$,
(c)$\lambda_\mathrm{bgn}$, and (d)$M_\mathrm{bgn}$
in the $(\log_{10}n_\mathrm{E},\;\log_{10}Q_\mathrm{UV})$ plane.
The open circles correspond to the parameters that are used
in Figure \ref{fig:begin cs}.
}
\label{fig:begin}
\end{figure*}
We presented only two examples of the HM and the LM models
in Section \ref{sec:sph result}.
In this section, we predict the GI in the shells with various parameters
$(n_\mathrm{E},\;Q_\mathrm{UV},\;T_\mathrm{c})$ by using the
results of the linear analysis presented in Paper I.
As shown in Figures \ref{fig:model1_growth} and \ref{fig:model2_growth},
it is found that the results of the SPH simulations are well described by the
growth rate, $\Gamma_\mathrm{evo}$.
Using this growth rate, one can predict the time $t_\mathrm{bgn}$ when
the GI begins to grow by the condition $\Gamma_\mathrm{evo}(l_\mathrm{bgn},\;
t_\mathrm{bgn})=0$, where $l_\mathrm{bgn}$ is the angular wavenumber of the
most unstable mode at $t=t_\mathrm{bgn}$.
This condition is rewritten as
\begin{equation}
\sqrt{8}\frac{V_\mathrm{s}}{R_\mathrm{s}} =
\Gamma(l_\mathrm{bgn},\:t_\mathrm{bgn}),
\label{frag criterion}
\end{equation}
where the left-hand side corresponds to
the evolutionary rate owing to expansion and mass accretion,
the right-hand side corresponds to the growth rate of the GI.
At early phase, the evolutionary rate is larger than the growth rate of the GI,
indicating that the shell is stable.
As the shell expands, $V_\mathrm{s}/R_\mathrm{s}$ decreases while $\Gamma$ increases.
Therefore, the GI begins to grow when the growth rate of the GI is larger than
the evolutionary rate.
The radius of the shell at $t=t_\mathrm{bgn}$ is defined by $R_\mathrm{s,bgn}$.
Figure \ref{fig:begin}(a), (b), (c), and (d) show
contour maps of $t_\mathrm{bgn}$, $R_\mathrm{s,bgn}$,
$\lambda_\mathrm{max}\equiv2\pi R_\mathrm{s,bgn}/l_\mathrm{bgn}$, and
$M_\mathrm{bgn}\equiv\Sigma(t=t_\mathrm{bgn})
\pi (\lambda_\mathrm{bgn}/2)^2$ in the $(n_\mathrm{E},\;Q_\mathrm{UV})$ plane derived
by using the linear analysis, respectively.
Here, $T_\mathrm{c}$ is assumed to be 10 K.
From Figures \ref{fig:begin}, one can see that the
$t_\mathrm{bgn}$, $R_\mathrm{s,bgn}$, $\lambda_\mathrm{bgn}$,
and $M_\mathrm{bgn}$ decrease with increasing $n_\mathrm{E}$.
This can be understood by scaling of
the typical number density of the shell, $n_\mathrm{sh}=n_\mathrm{E} {\cal M}_0^2$ that
corresponds to the shocked gas density.
From Equation (\ref{mach}), we have $n_\mathrm{sh}\propto n_\mathrm{E}^{6/7}$.
Therefore, for larger $n_\mathrm{E}$, the shell becomes denser and unstable earlier
(Figure \ref{fig:begin}(a)),
and the most unstable scale is smaller (Figures \ref{fig:begin}(c) and (d)).
Next, let us see the dependence of $t_\mathrm{bgn}$, $R_\mathrm{s,bgn}$, $\lambda_\mathrm{bgn}$,
and $M_\mathrm{bgn}$ on $Q_\mathrm{UV}$.
Since the central star is more energetic for larger $Q_\mathrm{UV}$,
the Mach number of the expanding shell $\left({\cal M}_0\propto Q_\mathrm{UV}^{1/7}\right)$
is larger, indicating that
the shell is denser and is more unstable.
Therefore, $t_\mathrm{bgn}$, $\lambda_\mathrm{bgn}$, and
$M_\mathrm{bgn}$ decreases with increasing $Q_\mathrm{UV}$.
On the other hand, $R_\mathrm{s,bgn}$ increases with $Q_\mathrm{UV}$ because
the expansion velocity increases with $Q_\mathrm{UV}$.
\begin{figure}[htpb]
\begin{center}
\includegraphics[width=8cm]{f13.eps}
\end{center}
\caption{
Dependence of (a)$t_\mathrm{bgn}/t_0$, (b)$R_\mathrm{s,bgn}/R_0$, and
(c)$l_\mathrm{max}$ on the typical Mach number ${\cal M}_0$.
The open circles correspond to the results of linear analysis.
The solid lines indicate the analytic formulae.
}
\label{fig:begin cs}
\end{figure}
From Figures \ref{fig:begin}, since the contour lines are almost straight in
the plane $(\log_{10}n_\mathrm{E},\:\log_{10}Q_\mathrm{UV})$,
quantities are expected to have power-law dependence on $Q_\mathrm{UV}$ and
$n_\mathrm{E}$, or equivalently on ${\cal M}_0$.
In Figure \ref{fig:begin cs},
we plot $t_\mathrm{bgn}/t_0$, $R_\mathrm{s,bgn}/R_0$, and $l_\mathrm{bgn}$
as a function of ${\cal M}_0$
for various parameters that correspond to
the open circles in Figure \ref{fig:begin}.
The open circles correspond to the results of the linear analysis.
From Figures \ref{fig:begin cs}, one can see that the results of
the linear analysis are well described by the analytic formulae,
$t_\mathrm{bgn}/t_0=1.62{\cal M}_0^{-0.75}$,
$R_\mathrm{s,bgn}/R_0=1.2{\cal M}_0^{-0.375}$,
and $l_\mathrm{bgn}=3.2{\cal M}_0^{1.5}$ that
are plotted by the solid lines. Using Equation (\ref{mach}), the analytic formulae
are rewritten as
\begin{equation}
t_\mathrm{bgn} = 0.6\:Q_\mathrm{UV,49}^{-0.11}\:
T_\mathrm{c,10}^{0.375}\:
n_\mathrm{E,3}^{-0.45}\;\mathrm{Myr},
\label{tbegin}
\end{equation}
\begin{equation}
R_\mathrm{s,bgn} = 3.4\:
Q_\mathrm{UV,49}^{0.09}\:
T_\mathrm{c,10}^{0.19}\:
n_\mathrm{E,3}^{-0.54}\;\mathrm{pc},
\label{rbegin}
\end{equation}
and
\begin{equation}
l_\mathrm{bgn} = 54\;
Q_\mathrm{UV,49}^{0.21}\:
T_\mathrm{c,10}^{-0.75}\:
n_\mathrm{E,3}^{-0.11}.
\label{lbegin}
\end{equation}
The wavelength $\lambda_\mathrm{bgn}$ and
the mass $M_\mathrm{bgn}$ that correspond to $l_\mathrm{bgn}$
are given by
\begin{equation}
\lambda_\mathrm{bgn}
= 0.4\;
Q_\mathrm{UV,49}^{-0.125}\:
T_\mathrm{c,10}^{0.94}\:
n_\mathrm{E,3}^{-0.44}\;\mathrm{pc},
\label{lambegin}
\end{equation}
and
\begin{equation}
M_\mathrm{bgn} = 3.5\;
Q_\mathrm{UV,49}^{-0.16}\:
T_\mathrm{c,10}^{2.06}\:
n_\mathrm{E,3}^{-0.42}\;\mathrm{M_\odot},
\label{mbegin}
\end{equation}
\citet[][W94]{Wetal94} also derived similar formulae by
using the thin-shell linear analysis similar to E94.
The dependence of Equations (\ref{tbegin})-(\ref{mbegin})
on parameters $Q_\mathrm{UV}$, $T_\mathrm{c}$, and $n_\mathrm{E}$
shows close agreement with that in W94
although the numerical factors are different.
Our results show that the GI begins to grow earlier and
as a result,
the fragment mass is smaller than their results.
In detail, $t_\mathrm{bgn}$, $R_\mathrm{bgn}$, and $\lambda_\mathrm{bgn}$
are roughly half of those in W94, and
$M_\mathrm{bgn}$ is roughly 1/8 of that in W94.
This is because the stabilization effect of the SF does not work
on the trailing surface that is the CD as shown in Paper I.
As pointed out in W94, the properties of fragments
are insensitive to $Q_\mathrm{UV}$, and
mainly depend on $n_\mathrm{E}$ and $T_\mathrm{c}$.
Indeed, in Equations (\ref{lambegin}) and (\ref{mbegin}),
the dependence of $\lambda_\mathrm{bgn}$ and $M_\mathrm{bgn}$ on $n_\mathrm{E}$ are
close to the Jeans length and the Jeans mass of the shell $(\lambda_\mathrm{J},\;M_\mathrm{J})
\propto n_\mathrm{sh}^{-1/2} \propto n_\mathrm{E}^{-3/7}$.
\subsection{Comparison with Previous Studies}
\citet{E89} investigated the GI in a decelerating shocked layer.
He numerically integrated perturbation equations with respect to time
by taking into account
time evolution of the layer and by averaging physical quantities
over the thickness.
He has taken into account SF and CD boundary conditions
on the SF and CD, respectively.
He called the boundary effect on the CD ``pinching force''.
In his linear analysis,
the destabilizing effect of the pinching force does not work efficiently.
One reason is that he used Equation (3) in Paper I as the
pressure of the HII region although
the geometry is plane-parallel.
Therefore, the pressure at the CD is underestimated, leading
to underestimate the pinching force of the CD.
The other reason is the geometry of the gravitational potential.
Asymmetry of density profile of spherical shell is qualitatively different from
that of plane-parallel layer.
In the plane-parallel geometry, the static symmetric layer peaks at the mid-plane.
If the layer decelerates, the density peak is always closer to the SF than the CD.
Therefore, the tangential flow behind SF erodes the density perturbation more efficiently.
On the other hand, for the case with shells,
the peaks can be closer to the CD than the SF (see Section \ref{sec:den pro}).
From these reasons, he derived lower growth rates than our results.
\citet{Detal09} simulated the gravitational fragmentation of expanding shells
confined by temporary constant thermal pressure of hot rarefied gases on both sides.
In their calculation, the motion of the shell is determined so that the pressure at
the leading surface is the same as that at the trailing surface.
Therefore, the density peak is always around the mid-plane of the shell, and
the density profile is almost symmetric.
In their calculation, the column density decreases with time because the shell expands keeping the mass fixed.
Therefore, the pressures at the boundaries approach to the peak pressure.
They found that the confining pressure accelerates fragmentation in the later phase, and
describe this effect as ``pressure-assisted'' gravitational fragmentation.
This mode is the same as the incompressible mode.
Recently, taking into account the effect of the thickness of the shell
approximately,
\citet{WDPW10} established a semi-analytic linear analysis that
explains results of their SPH simulations.
Different from them, in Paper I, we have taken into account the effect of SF+CD boundary condition
approximately.
\subsection{Other Instabilities}
We discuss other potentially important effects that are not analyzed
in this paper.
\subsubsection{Vishniac Instability}
\citet{V83} found a hydrodynamical overstability on decelerating shells
confined by the SF on the leading surface and by the CD on the trailing surface
by using thin-shell linear analysis.
However, the Vishniac instability (VI) did not appear to occur in our simulations.
This can be explained by the stabilizing effect by expansion.
\citet{VR89} obtained an analytic dispersion relation of the VI of
the expanding shell without
self-gravity by taking into account the thickness approximately.
In Section 6 in Paper I, they discussed the VI by applying the analytic
dispersion relation to the shell driven by the HII region.
They found that the VI is expected not to be important in the later phase ($t/t_0>0.5$)
since its growth rate is comparable to the expansion rate of the shell radius.
Therefore, the VI was not found in our simulation.
However, they also found that before the beginning time of our simulations ($t/t_0<0.4$),
the VI is expected to grow rapidly since the Mach number is large
(see Figure 13 in Paper I). The perturbations quickly reach the nonlinear
regime and saturate as a subsonic transverse flow \citep{MN93}.
As a results, a small scale subsonic turbulence whose angular wavelength of
$l=10^2\sim10^3$ is expected to be induced. The consequence of VI may correspond to the
increase of the initial perturbations in our 3D simulation.
\subsubsection{Thermal Instability}\label{sec:TI}
In this paper, we assume that the fluid evolves keeping its temperature fixed.
However, in real interstellar medium (ISM), shock-heated gas cools via radiative cooling.
It is well known that the cooling ISM is often thermally unstable.
In such a case, during the cooling condensation, the layer is expected to
fragment into dense clouds with various scales \citep{IT09}.
\citet{KI02} investigated the propagation of a shock wave into the warm
neutral medium. They found that cold cloudlets move with supersonic
velocity dispersion in surrounding
warm gas in the postshock region.
The velocity dispersion is larger than the sound speed of cloudlets while
it is smaller than the sound speed of surrounding warm gas.
Since the shocked molecular cloud is also thermally unstable, the thermal instability
drives supersonic turbulence inside the shell, and
cold molecular cloudlets move with the supersonic velocity dispersion.
This situation is quite different from the isothermal gas that has been considered
in this paper.
The supersonic turbulence is expected to be important in contrast to
the subsonic turbulence
driven by the VI, and considerably modify the evolution.
The mass of the cold molecular cloudlets increases with time through the accretion of
surrounding gas and the coalescence between cloudlets.
The supersonic turbulence may provide effective turbulent pressure,
while coalescence of cold cloudlets may decrease turbulent pressure.
We will address the effect of the thermal instability on the
GI of the expanding shells in our next work.
\subsection{Ionization Front}
In the paper, we did not solve the radiative transfer equation of the ionizing photons.
Instead, we assumed that the trailing surface is the CD instead of the
ionization front (IF).
There are two effects of the ionization on the GI.
One is that a fraction of the shell becomes ionized gas as HII region expands.
Therefore, our simulations overestimate the column density of the shell.
The swept-up mass by the SF is given by
\begin{equation}
M_\mathrm{sw} = \frac{4\pi}{3}\rho_\mathrm{E} R_\mathrm{s}^3.
\label{swept mass}
\end{equation}
On the other hand, the ionized mass is given by
\begin{equation}
M_\mathrm{ion} = \frac{4\pi}{3}\rho_\mathrm{i} R_\mathrm{s}^3,
\label{ion mass}
\end{equation}
where $\rho_\mathrm{i}$ is the density of the ionized gas, and
we neglect the difference between radii of the IF and SF.
The real shell mass is given by $M_\mathrm{sw}-M_\mathrm{ion}$.
Inside the HII region, the balance between the ionization and the recombination
is approximately established. Therefore, the density of the HII region becomes
\begin{equation}
\rho_i
= \left( \frac{3m^2 Q_\mathrm{UV}}{4\pi \alpha_\mathrm{B} R_\mathrm{s}^3} \right)^{1/2}
= \rho_\mathrm{E}\left( \frac{R_\mathrm{ST}}{R_\mathrm{s}} \right)^{3/2},
\label{ion den}
\end{equation}
where $m$ is the mean weight of the gas particles and we use Equation (\ref{stremgren}).
Therefore, from Equations (\ref{swept mass})-(\ref{ion den}),
the ratio of $M_\mathrm{ion}$ to $M_\mathrm{sw}$ is
\begin{equation}
{\cal R}\equiv \frac{M_\mathrm{ion}}{M_\mathrm{sw}}
= \left( \frac{R_\mathrm{ST}}{R_\mathrm{s}} \right)^{3/2}.
\end{equation}
If ${\cal R}\ll 1$, the effect of the ionization is negligible.
In Figure 1 in Paper I (also see Figure \ref{fig:surpos}), the Str{\"o}mgren radius is
as large as $0.1R_0$ in various parameters $(Q_\mathrm{UV},\;n_\mathrm{E})$.
At the beginning time of our simulations, $t=0.4t_0$ $(R_\mathrm{s}\sim0.6R_0)$, ${\cal R}\sim
(0.1/0.6)^{3/2}=0.07$. As the shell expands, ${\cal R}$ decreases.
At $t=t_0$, ${\cal R}\sim (0.1/10)^{3/2}\sim 0.03$.
Therefore, the contribution of the ionization is limited to only several percent.
The other is that the IF induces an instability on the shell.
The D-type IF is analogous to a combustion front that
is unstable to corrugational deformations \citep{LL87}.
\citet{V62} found that the D-type IF is unstable.
In this instability, the most unstable scale is comparable to the thickness
of the IF.
Moreover, \citet{GF95} showed that the VI is strongly modified by the IF.
The shell rapidly fragments and finger-like
structures are generated in their two-dimensional
simulations.
Note, however, that they assumed the power-law density distribution $\propto r^{-w}$ of
the ambient gas with the uniform density core in the center with $0.2$ pc,
where parameter $w$ is set to 2.
The expansion of the IF depends on the power-law index $w$ in the 1D model.
If $w>1.5$, even if the SF is emerged in front of the IF in the uniform density core, the
IF quickly gets ahead of the SF, and eventually
the whole of the shell and the ambient gas are ionized.
Moreover, \citet{H07} investigated the expansion of the IF and the dissociation front,
and showed that the star formation is expected to be suppressed when $w>1$.
This is because the column density decreases as the shell expands and the
FUV photon easily escapes in front of the shock.
Therefore, model by \citet{GF95} corresponds to the case where
the triggered star formation is not simply expected.
We need to investigate the expansion of the IF for the case with the shallower density
profile $w<1$ in detail
using the radiative multi-dimensional simulations taking into account the self-gravity.
Moreover, \citet{W99} discovered shadowing
instability in the R-type IF before emerging of the SF.
This instability may also disturb the shell in the very early phase of evolution.
\subsection{Magnetic Field}
It is well known that the magnetic field is important in the dynamics of the ISM although
it is neglected in this paper.
\citet{NIM98} investigated the GI of a pressure-confined
isothermal layer threaded by
uniform magnetic field by using linear analysis.
They found that the magnetic field cannot stabilize the GI.
Moreover, they found the layer fragment in filamentary clouds whose
longitudinal axis
is parallel or perpendicular to the magnetic field
depending on the thickness of the layer.
However, their analysis is restricted to the static magnetized layer.
It is important to investigate stability of magnetized shocked layer or shell
in dynamical modeling.
\section{Summary}\label{sec:summary}
In this paper, we have investigated the GI of the expanding shell using
3D numerical simulation with high resolution.
We summarize our results as follows:
\begin{enumerate}
\item The GI begins to grow earlier than the
prediction from the linear analysis
based on the thin-shell approximation \citep{E94, Wetal94}.
During the development of the GI,
the contact discontinuity highly deforms while the shock front remains almost flat.
These all features are consistent with the prediction
from \citet{IIT11}.
\item The GI is expected to begin to grow at an epoch
when the growth rate of the GI becomes larger than
the evolutionary rate owing to the expansion and the mass accretion (see Equation (\ref{frag criterion}))
Using the results of the linear analysis,
we have derived useful approximate analytic formulae (Equations (\ref{tbegin})-(\ref{mbegin}))
for the fragment scale
and the epoch when the GI starts.
\end{enumerate}
This paper and Paper I revisit the fragmentation of the expanding shells by using both
the 3D simulation and the linear analysis.
Since the gas is subject to heating owing to far-UV photon, the gas temperature is
larger than 10 K \citep{HI05,HI06}.
When $T_\mathrm{c}=30$ K ($50$ K), the GI with mass scale of
$M_\mathrm{bgn}\sim 34\:M_\odot\:$ ($96$ $M_\odot$)
begins to grow at $t_\mathrm{bgn}\sim0.9$ Myr (1.1 Myr)
($Q_\mathrm{UV}=10^{49}$ s$^{-1}$, $n_\mathrm{E}=10^3$
cm$^{-3}$) from Equations (\ref{tbegin}) and (\ref{mbegin}).
The fragment mass is expected to be larger than $M_\mathrm{bgn}$ because
it increases through the gas accretion.
Moreover, perturbations with larger scale than the most unstable mode
is also unstable. Therefore, the larger scale perturbations gradually grows, and
small scale fragments may coalesce into larger one, and
the total number of fragments may decrease.
The predicted masses are roughly comparable
to (or are slightly smaller than) observational
masses of dense cores (dozens $\sim$ hundreds M$_\odot$) \citep[e.g.,][]{Detal03,Zetal06}
although our model is based on very idealized situation, such as uniform ambient
gas and isothermal equation of state.
\acknowledgments
We thank the referee for very careful reading and
many constructive comments that improved this paper significantly.
Numerical computations were carried out on PC cluster at Osaka University Cybermedia Center,
and Cray XT4 at the CfCA of National Astronomical Observatory of Japan.
This work was supported by Grants-in-Aid for Scientific Research
from the MEXT of Japan (K.I.:22864006; S.I.:18540238 and 16077202),
and Research Fellowship from JSPS (K.I.: 21-1979).
The page charge of this paper is supported by CfCA
|
\section{Introduction}
\em Categorical quantum mechanics \em \cite{AC} aims to recast quantum
mechanical notions in terms of symmetric monoidal categories with
additional structure. One layer of extra structure, compactness
\cite{KellyLaplaza}, encompasses the well-known Choi-Jamiolkowski
isomorphism. Compactness is itself subsumed by the much richer
commutative Frobenius algebra structure \cite{CarboniWalters}, which
governs classical data, observables, and certain tripartite states
\cite{CPav, CD, CES2, CK}. In this symmetric monoidal form,
quantum mechanics enjoys:
\begin{itemize}
\item an \em operational interpretation \em by making sequential and
parallel composition of systems and processes the basic
connectives of the language \cite{ContPhys};
\item an intuitive \em diagrammatic calculus \em \cite{ContPhys} via the
Penrose-Joyal-Street diagrammatic calculus for symmetric monoidal
categories \cite{Penrose, JS}, augmented with Kelly and Laplaza's
coherence result for compact categories, and Lack's work on
distributive laws \cite{Lack};
\item a \em logical underpinning \em \cite{RossThesis} via the closed
structure resulting from compactness.
\end{itemize}\par\noindent
The last allows the application of automated reasoning techniques to
quantum mechanics \cite{DD, quanto, DK}. A prototype software
implementation, {\tt quantomatic}, already exists and is jointly
developed in Edinburgh and Oxford.
Categorical quantum mechanics has meanwhile been successful in
solving problems in quantum information \cite{DP} and quantum
foundations \cite{CES2}, where other methods and structures failed
to be adequate. Key to these results is the description of \em
interacting basis structures \em in \cite{CD}. The language of that
paper consists of a pair of abstract bases or \em basis
structures\em, which are, again in abstract terms, mutually
unbiased, and an abstract generalisation of phases relative to
bases. This formalism has been implemented in {\tt quantomatic},
and is expressive enough to universally model any linear map
$f:\mathbb{Q}^{\otimes n}\to \mathbb{Q}^{\otimes m}$, where
$\mathbb{Q}=\mathbb{C}^2$. On the other hand, if we restrict the
language to the two basis structures only it becomes very poor,
describing no more than 2 qubit states.
This brings us to the subject of this paper. In \cite{CK} two of
the authors introduced pairs of interacting commutative Frobenius
algebras that do not model bases, but the tripartite GHZ and W
states \cite{DVC}. Both these states can indeed be endowed with the
structure of a commutative Frobenius algebra, yielding a \em GHZ
structure \em and a \em W structure \em as we recall in Section
\ref{sec:ghz-w}. The main point of this paper is that the language
consisting of the GHZ structure (which is essentially the same as a
basis structure) and the W structure is already rich enough to
encode rational arithmetic, with the exception of additive inverses.
Now an infinite number of qubit states can be described,
corresponding to the rational numbers of the arithmetic system. We
demonstrate this in Section \ref{sec:arithmetic}. In Section
\ref{sec:additive-inverses} we extend the GHZ/W-calculus with one
basic graphical element which then allows additive inverses to be
captured. Section \ref{sec:automation} addresses the issue of how
to implement the calculus within the {\tt quantomatic} software.
We assume that the reader is familiar with the diagrammatic calculus
for symmetric monoidal categories \cite{JS,SelingerSurvey}, which is
also reviewed in \cite{CK}. We also assume that the reader is
familiar with the (very) basics of finite dimensional Hilbert spaces
and Dirac notation as used in quantum computing.
\section{Frobenius Algebras and the GHZ/W-calculus}\label{sec:ghz-w}
Fix a symmetric monoidal category $({\bf V},\otimes,I,\sigma)$. Throughout this paper, we shall
define morphisms in $\bf V$ using the graphical notation defined in \cite{SelingerSurvey}. In this
notation, `wires' correspond to objects and vertices, and `boxes' correspond to morphisms. We shall
express composition vertically, from top to bottom, and the monoidal product as (horizontal)
juxtaposition of graphs. When wires are not labeled, they are assumed to represent a fixed object,
$Q$.
\begin{example}
A canonical example throughout will be ${\bf FHilb}$, the category of finite-dimensional Hilbert
spaces and linear maps. In this case, $\otimes$ is the usual tensor product, $\sigma$ the swap map
$v \otimes w \mapsto w \otimes v$, $I := \mathbb C$ and $Q := \mathbb C^2$, the space of qubits. We
shall also refer the ``projective'' category of finite-dimensional Hilbert spaces, ${\bf FHilb}_p$,
whose objects are the same as ${\bf FHilb}$ and whose arrows are linear maps, taken to be
equivalent iff they differ only by a non-zero scalar.
\end{example}
\subsection{Commutative Frobenius Algebras}
A \em commutative Frobenius algebra \em (CFA) consists of an internal commutative monoid $(Q, \dotmult{small dot}\ , \dotunit{small dot})$ and an internal cocommutative comonoid $(Q, \dotcomult{small dot}\ , \dotcounit{small dot})$ that interact via the Frobenius law:
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0.5, -1.5) {};
\node [style=none] (1) at (1.5, -1.5) {};
\node [style=none] (2) at (3, -1.5) {};
\node [style=none] (3) at (4, -1.5) {};
\node [style=dot] (4) at (1.5, -2) {};
\node [style=dot] (5) at (3.5, -2) {};
\node [style=none] (6) at (2.5, -2.25) {=};
\node [style=dot] (7) at (1, -2.5) {};
\node [style=dot] (8) at (3.5, -2.5) {};
\node [style=none] (9) at (1, -3) {};
\node [style=none] (10) at (2, -3) {};
\node [style=none] (11) at (3, -3) {};
\node [style=none] (12) at (4, -3) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (8) to (12.center);
\draw (9.center) to (7);
\draw (5) to (2.center);
\draw (4) to (7);
\draw (11.center) to (8);
\draw (4) to (1.center);
\draw (5) to (3.center);
\draw (5) to (8);
\draw[bend left=15] (7) to (0.center);
\draw[bend left=15, looseness=1.25] (4) to (10.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
One can show that any connected graph consisting only of $\dotmult{small dot}, \dotunit{small dot}, \dotcomult{small dot}, \dotcounit{small dot}, \sigma$
and $1_Q$ depends only upon the number of inputs, outputs, and loops. As such, it can be reduced to
a canonical normal form:
\[
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-2, 4.25) {};
\node [style=none] (1) at (-1, 4.25) {};
\node [style=none] (2) at (0, 4.25) {};
\node [style=none] (3) at (2, 4.25) {};
\node [style=dot] (4) at (-1.5, 3.75) {};
\node [style=dot] (5) at (-1, 3.25) {};
\node [style=none] (6) at (-0.75, 3) {};
\node [style=none] (7) at (-0.5, 2.75) {\small ...};
\node [style=none] (8) at (-0.25, 2.5) {};
\node [style=dot] (9) at (0, 2.25) {};
\node [style=dot] (10) at (0, 1.5) {};
\node [style=dot] (11) at (0, 0.75) {};
\node [style=none] (12) at (0, 0.25) {};
\node [style=none] (13) at (0, 0) {\small ...};
\node [style=none] (14) at (0, -0.25) {};
\node [style=dot] (15) at (0, -0.75) {};
\node [style=dot] (16) at (0, -1.5) {};
\node [style=dot] (17) at (0, -2.25) {};
\node [style=none] (18) at (-0.25, -2.5) {};
\node [style=none] (19) at (-0.5, -2.75) {\small ...};
\node [style=none] (20) at (-0.75, -3) {};
\node [style=dot] (21) at (-1, -3.25) {};
\node [style=dot] (22) at (-1.5, -3.75) {};
\node [style=none] (23) at (-2, -4.25) {};
\node [style=none] (24) at (-1, -4.25) {};
\node [style=none] (25) at (0, -4.25) {};
\node [style=none] (26) at (2, -4.25) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw[bend left=60, looseness=1.25] (10) to (11);
\draw[bend right=60, looseness=1.25] (15) to (16);
\draw[bend right=60, looseness=1.25] (10) to (11);
\draw (16) to (17);
\draw (9) to (10);
\draw (4) to (5);
\draw (9) to (3.center);
\draw (1.center) to (4);
\draw (25.center) to (21);
\draw (21) to (20.center);
\draw (14.center) to (15);
\draw (8.center) to (9);
\draw (5) to (6.center);
\draw (23.center) to (22);
\draw (24.center) to (22);
\draw (0.center) to (4);
\draw (22) to (21);
\draw (2.center) to (5);
\draw (17) to (26.center);
\draw[bend left=60, looseness=1.25] (15) to (16);
\draw (18.center) to (17);
\draw (11) to (12.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
In any connected graph, loops are counted as the total number of edges that can be removed without
disconnecting the graph. We shall use `spider' notation to represent graphs of Frobenius algebras
using vertices of any arity. We express any connected graph as above with $m$ inputs, $n$ outputs,
and no loops as a single vertex of the same colour:
\[
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (3, 3.25) {$\overbrace{\qquad\qquad\qquad\qquad}^m$};
\node [style=none] (1) at (1, 2.5) {};
\node [style=none] (2) at (2, 2.5) {};
\node [style=none] (3) at (3, 2.5) {};
\node [style=none] (4) at (5, 2.5) {};
\node [style=dot] (5) at (1.5, 2) {};
\node [style=none] (6) at (-1, 1.75) {$\overbrace{\qquad\qquad}^m$};
\node [style=dot] (7) at (2, 1.5) {};
\node [style=none] (8) at (2.25, 1.25) {};
\node [style=none] (9) at (-2, 1) {};
\node [style=none] (10) at (-1.5, 1) {};
\node [style=none] (11) at (0, 1) {};
\node [style=none] (12) at (2.5, 1) {\small ...};
\node [style=none] (13) at (-0.75, 0.75) {\small ...};
\node [style=none] (14) at (2.75, 0.75) {};
\node [style=dot] (15) at (3, 0.5) {};
\node [style=dot] (16) at (-1, 0) {};
\node [style=none] (17) at (1, 0) {=};
\node [style=dot] (18) at (3, -0.5) {};
\node [style=none] (19) at (-0.75, -0.75) {\small ...};
\node [style=none] (20) at (2.75, -0.75) {};
\node [style=none] (21) at (-2, -1) {};
\node [style=none] (22) at (-1.5, -1) {};
\node [style=none] (23) at (0, -1) {};
\node [style=none] (24) at (2.5, -1) {\small ...};
\node [style=none] (25) at (2.25, -1.25) {};
\node [style=dot] (26) at (2, -1.5) {};
\node [style=none] (27) at (-1, -1.75) {$\underbrace{\qquad\qquad}_n$};
\node [style=dot] (28) at (1.5, -2) {};
\node [style=none] (29) at (1, -2.5) {};
\node [style=none] (30) at (2, -2.5) {};
\node [style=none] (31) at (3, -2.5) {};
\node [style=none] (32) at (5, -2.5) {};
\node [style=none] (33) at (3, -3.25) {$\underbrace{\qquad\qquad\qquad\qquad}_n$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (14.center) to (15);
\draw[bend right=15] (10.center) to (16);
\draw (29.center) to (28);
\draw (28) to (26);
\draw (30.center) to (28);
\draw[bend right=15] (16) to (22.center);
\draw (5) to (7);
\draw (15) to (4.center);
\draw[bend left] (16) to (23.center);
\draw (3.center) to (7);
\draw (20.center) to (18);
\draw (15) to (18);
\draw (7) to (8.center);
\draw[bend right] (9.center) to (16);
\draw (2.center) to (5);
\draw[bend left=15] (11.center) to (16);
\draw (31.center) to (26);
\draw (18) to (32.center);
\draw[bend right] (16) to (21.center);
\draw (1.center) to (5);
\draw (26) to (25.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
We give two of these graphs special names. The \emph{cup} is defined as $\dotcup{small dot}$ and the \emph{cap} is defined as $\dotcap{small dot}$. These induce a compact structure, since
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (-0.75, 0.75) {};
\node [style=none] (1) at (1.75, 0.75) {};
\node [style=dot] (2) at (-4.25, 0.5) {};
\node [style=dot] (3) at (-0.75, 0.25) {};
\node [style=none] (4) at (0.25, 0.25) {};
\node [style=none] (5) at (-4.75, -0) {};
\node [style=none] (6) at (-2.75, -0) {};
\node [style=none] (7) at (-2, -0) {$=$};
\node [style=none] (8) at (1, -0) {$=$};
\node [style=none] (9) at (-1.25, -0.25) {};
\node [style=dot] (10) at (-0.25, -0.25) {};
\node [style=dot] (11) at (-3.25, -0.5) {};
\node [style=dot] (12) at (-0.25, -0.75) {};
\node [style=none] (13) at (1.75, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw[out=0, in=270] (11) to (6.center);
\draw[out=90, in=180] (5.center) to (2);
\draw (3) to (10);
\draw[out=0, in=180] (2) to (11);
\draw (0) to (3);
\draw (1.center) to (13.center);
\draw (9.center) to (3);
\draw (12) to (10);
\draw (10) to (4.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
\subsection{Phases}
\begin{definition}{\rm\cite{CD}}\label{Phasedef}
Given a CFA on an object $A$, a morphism $f:A\to A$ is a \em phase \em if we have
\begin{equation}\label{eq:phases1}
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-0.5, 1.5) {};
\node [style=none] (1) at (0.5, 1.5) {};
\node [style=square box] (2) at (-0.5, 0.5) {$f$};
\node [style=none] (3) at (0.5, 0.5) {};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=none] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw [bend right] (2.center) to (4);
\draw [bend right] (4) to (3.center);
\draw (3.center) to (1);
\draw (4) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-0.5, 1.5) {};
\node [style=none] (1) at (0.5, 1.5) {};
\node [style=square box] (2) at (0.5, 0.5) {$f$};
\node [style=none] (3) at (-0.5, 0.5) {};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=none] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (3.center);
\draw [bend right] (3.center) to (4);
\draw [bend right] (4) to (2.center);
\draw (2.center) to (1);
\draw (4) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-0.5, 1.5) {};
\node [style=none] (1) at (0.5, 1.5) {};
\node [style=dot] (2) at (0, 0.5) {};
\node [style=square box] (3) at (0, -0.5) {$f$};
\node [style=none] (4) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (0) to (2);
\draw [bend right=30] (2) to (1);
\draw (2) to (3.center);
\draw (3.center) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\eeq
\end{definition}
Equivalently, phases can be described as module endomorphisms, where $\dotmult{small dot}$ is considered as a left (or right) module over itself.
\begin{proposition}\label{prop:phases-states}
A phase $f:A\to A$
can be equivalently defined as a morphism of the form:
\begin{equation}\label{eq:phases2}
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw (2.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ := \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=dot] (2) at (0, 0) {};
\node [style=white dot] (3) at (1, 1) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw (2.center) to (1);
\draw [bend right=60] (2.center) to (3.center);
\end{pgfonlayer}
\end{tikzpicture}}
\eeq
for some element $\psi:{\rm I}\to A$.
\end{proposition}
\begin{proof}
Given eqs.~(\ref{eq:phases1}) we have
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=square box] (2) at (0, 0) {$f$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw (2.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=dot] (2) at (0, -0.5) {};
\node [style=square box] (3) at (0, 0.5) {$f$};
\node [style=dot] (4) at (1, 1.5) {};
\node [style=none] (5) at (1, 0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (3.center);
\draw (3.center) to (2.center);
\draw (2.center) to (1);
\draw [bend right=45] (2.center) to (5.center);
\draw (5.center) to (4.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=dot] (2) at (0, -0.5) {};
\node [style=square box] (3) at (1, 0.5) {$f$};
\node [style=dot] (4) at (1, 1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw (2.center) to (1);
\draw (3.center) to (4);
\draw [bend right=60] (2.center) to (3.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=small white dot,font=\footnotesize] (3) at (1, 1) {$f \circ \dotunit{small dot}$};
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=dot] (2) at (0, -0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=60] (2.center) to (3.center);
\draw (2) to (1);
\draw (2) to (0);
\end{pgfonlayer}
\end{tikzpicture}}
\]
where we used unitality of the CFA, and conversely, given eq.~(\ref{eq:phases2}), eqs.~(\ref{eq:phases1}) straightforwardly follow by associativity and commutativity of the CFA.
\end{proof}
\begin{proposition}\label{prop:invphase}
The inverse of a phase is a phase.
\end{proposition}
\begin{proof}
Setting $
\raisebox{-4mm}{
\begin{tikzpicture}[scale=0.5]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1) {};
\node [style=square box] (1) at (0, 0) {$f$};
\node [style=none] (2) at (0, -1) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (1);
\draw (1) to (2.center);
\end{pgfonlayer}
\end{tikzpicture}
}
:=
\left(\raisebox{-4mm}{
\begin{tikzpicture}[scale=0.5]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1) {};
\node [style=white dot] (1) at (0, 0) {$\psi$};
\node [style=none] (2) at (0, -1) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (1);
\draw (1) to (2.center);
\end{pgfonlayer}
\end{tikzpicture}
}\right)^{-1}
$ we have
\[
\begin{tikzpicture}[scale=0.8]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-5.75, 1.5) {};
\node [style=none] (1) at (-4.75, 1.5) {};
\node [style=none] (2) at (-3.25, 1.5) {};
\node [style=none] (3) at (-2.25, 1.5) {};
\node [style=none] (4) at (-0.75, 1.5) {};
\node [style=none] (5) at (0.25, 1.5) {};
\node [style=none] (6) at (1.75, 1.5) {};
\node [style=none] (7) at (2.75, 1.5) {};
\node [style=square box] (8) at (-5.75, 0.75) {$f$};
\node [style=none] (9) at (-4.75, 0.75) {};
\node [style=square box] (10) at (-3.25, 0.75) {$f$};
\node [style=none] (11) at (-2.25, 0.75) {};
\node [style=square box] (12) at (-0.75, 0.75) {$f$};
\node [style=dot] (13) at (-5.25, 0) {};
\node [style=none] (14) at (-4, 0) {$=$};
\node [style=dot] (15) at (-2.75, 0) {};
\node [style=none] (16) at (-1.5, 0) {$=$};
\node [style=none] (17) at (1, 0) {$=$};
\node [style=white dot] (18) at (-0.75, -0.25) {$\psi$};
\node [style=none] (19) at (0.25, -0.25) {};
\node [style=none] (20) at (1.75, -0.25) {};
\node [style=none] (21) at (2.75, -0.25) {};
\node [style=white dot] (22) at (-2.75, -0.75) {$\psi$};
\node [style=dot] (23) at (-0.25, -1) {};
\node [style=dot] (24) at (2.25, -1) {};
\node [style=square box] (25) at (-2.75, -1.75) {$f$};
\node [style=square box] (26) at (-0.25, -1.75) {$f$};
\node [style=square box] (27) at (2.25, -1.75) {$f$};
\node [style=none] (28) at (-5.25, -2.5) {};
\node [style=none] (29) at (-2.75, -2.5) {};
\node [style=none] (30) at (-0.25, -2.5) {};
\node [style=none] (31) at (2.25, -2.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (22) to (25);
\draw (9.center) to (1.center);
\draw[bend right] (20.center) to (24);
\draw (6.center) to (20.center);
\draw (26) to (30.center);
\draw[bend right] (24) to (21.center);
\draw[bend right] (13) to (9.center);
\draw (7.center) to (21.center);
\draw (27) to (31.center);
\draw[bend right] (15) to (11.center);
\draw (0.center) to (8);
\draw[bend right] (10) to (15);
\draw (4.center) to (12);
\draw[bend right] (8) to (13);
\draw[bend right] (23) to (19.center);
\draw (19.center) to (5.center);
\draw (23) to (26);
\draw (13) to (28.center);
\draw (24) to (27);
\draw[bend right] (18) to (23);
\draw (2.center) to (10);
\draw (25) to (29.center);
\draw (15) to (22);
\draw (11.center) to (3.center);
\draw (12) to (18);
\end{pgfonlayer}
\end{tikzpicture}
\]
\end{proof}
\subsection{GHZ/W calculus}
In this paper, we are concerned not only with general CFAs, but two specific cases, depending on the behaviour of the loops. We refer to $\dotmult{small dot} \circ \dotcomult{small dot}$ as the \emph{loop map} of a CFA.
\begin{definition}{\rm\cite{CK}}
A \emph{GHZ-structure} is a special commutative Frobenius algebra; that is, a commutative Frobenius algebra where the loop map is equal to the identity:
\[
\begin{tikzpicture}[dotpic,yshift=5mm]
\node [dot] (a) at (0,0) {};
\node [dot] (b) at (0,-1) {};
\draw [bend left] (a) to (b);
\draw [bend right] (a) to (b);
\draw (0,0.5) to (a) (b) to (0,-1.5);
\end{tikzpicture} = \
\begin{tikzpicture}[dotpic]
\draw (0,1) -- (0,-1);
\end{tikzpicture}
\]
\end{definition}
These GHZ-structures have also been referred to as \emph{basis structures}, for example in \cite{CES2}, because of their strong connection to bases in finite-dimensional vector spaces. See Theorem \ref{basisthm} below.
\begin{definition}{\rm\cite{CK}}
A \emph{W-structure} is an anti-special commutative Frobenius algebra. This is commutative Frobenius algebra whose loop map obeys the following equation:
\begin{equation}\label{eq:antispec}
\circl\ \begin{tikzpicture}[dotpic]
\node [dot] (a) at (0,0.5) {};
\node [dot] (b) at (0,-0.5) {};
\draw [bend left] (a) to (b);
\draw [bend right] (a) to (b);
\draw (0,1) to (a) (b) to (0,-1);
\end{tikzpicture}\ \ =
\begin{tikzpicture}[dotpic]
\node [dot] (a) at (0,0.7) {};
\node [dot] (b) at (0,-0.7) {};
\draw (0,1.2) to (a) (b) to (0,-1.2);
\draw (a) to [downloop] ();
\draw (b) to [uploop] ();
\end{tikzpicture}
\end{equation}
where we use the following short-hand notation:
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (1, 0.75) {};
\node [style=none] (1) at (5.25, 0.75) {};
\node [style=none] (2) at (3.25, 0.5) {};
\node [style=dot] (3) at (1, 0.25) {};
\node [style=dot] (4) at (5.25, 0.25) {};
\node [style=dot] (5) at (-1, 0) {};
\node [style=none] (6) at (0, 0) {=};
\node [style=dot] (7) at (3.25, 0) {};
\node [style=none] (8) at (4.25, 0) {=};
\node [style=dot] (9) at (1, -0.25) {};
\node [style=dot] (10) at (5.25, -0.25) {};
\node [style=none] (11) at (-1, -0.5) {};
\node [style=none] (12) at (1, -0.75) {};
\node [style=dot] (13) at (5.25, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw[bend left=300, looseness=1.25] (10) to (4);
\draw[bend right=300, looseness=1.25] (3) to (9);
\draw[out=45, in=135, loop] (5) to ();
\draw (4) to (1.center);
\draw[bend left=60, looseness=1.25] (10) to (4);
\draw (9) to (12.center);
\draw (0) to (3);
\draw[bend right=60, looseness=1.25] (3) to (9);
\draw[out=-45, in=-135, loop] (7) to ();
\draw (5) to (11.center);
\draw (7) to (2.center);
\draw (13) to (10);
\end{pgfonlayer}
\end{tikzpicture}
\]
\end{definition}
This distinction essentially comes down to whether the loop map is singular or invertible.
\begin{lemma}\label{lem:iso-scfa}
If the loop map of a CFA is an isomorphism, the CFA can be made special via a phase.
\end{lemma}
\begin{proof}
Consider a CFA $(\dotmult{small dot}, \dotunit{small dot}, \dotcomult{small dot}, \dotcounit{small dot})$. Since its loop map is a phase, by Proposition \ref{prop:invphase} so is the inverse of the loop map, which we denote $f$. Then
$(\,\dotmult{small dot}\, , \, \dotunit{small dot}\,,
\raisebox{-4.5mm}{
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-8, 1.5) {};
\node [style=square box] (1) at (-8, 0.75) {$f$};
\node [style=dot] (2) at (-8, 0) {};
\node [style=none] (3) at (-8.5, -0.5) {};
\node [style=none] (4) at (-7.5, -0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2) to (4.center);
\draw (1) to (2);
\draw (0.center) to (1);
\draw (2) to (3.center);
\end{pgfonlayer}
\end{tikzpicture}}
,\,\cololli\,)$ is easily seen to be a special CFA.
\end{proof}
\begin{lemma}[Herrmann \cite{Hermann}]
If the loop of a CFA is disconnected, i.e.~factors over the tensor unit, then it obeys eq.~{\rm(\ref{eq:antispec})}, that is the CFA is necessarily anti-special.
\end{lemma}
The following is an example of a GHZ-structure in ${\bf FHilb}$:
\begin{equation}\label{GHZ-SCFA}
\begin{split}
\dotmult{small white dot} & = \ket{0}\bra{00} + \ket{1}\bra{11} \qquad\qquad
\dotunit{small white dot} = \sqrt{2}\, \ket{+} := \ket{0}+\ket{1} \\
\dotcomult{small white dot} & = \ket{00}\bra{0} + \ket{11}\bra{1} \qquad\qquad
\dotcounit{small white dot} = \sqrt{2} \bra{+} := \bra{0}+\bra{1}
\end{split}\vspace{-1.5mm}
\end{equation}
and we also have an example of a W-structure in ${\bf FHilb}$:
\begin{equation}\label{W-ACFA}
\begin{split}
\dotmult{small dot} & = \ket{1}\bra{11} + \ket{0}\bra{01} + \ket{0}\bra{10}
\qquad\qquad\qquad
\dotunit{small dot} = \ket 1\qquad\qquad \\
\dotcomult{small dot} & = \ket{00}\bra{0} + \ket{01}\bra{1} + \ket{10}\bra{1}
\qquad\qquad\qquad
\dotcounit{small dot} = \bra 0\qquad\qquad
\end{split}
\end{equation}
Note that the cups for these CFAs do not coincide:
\[
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (1.5, 0.75) {};
\node [style=white dot] (1) at (-1.5, 0.25) {};
\node [style=white dot] (2) at (1.5, 0.25) {};
\node [style=none] (3) at (0, 0) {$:=$};
\node [style=none] (4) at (-2, -0.25) {};
\node [style=none] (5) at (-1, -0.25) {};
\node [style=none] (6) at (1, -0.25) {};
\node [style=none] (7) at (2, -0.25) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2);
\draw (7.center) to (2);
\draw[bend right=45] (5.center) to (1);
\draw (6.center) to (2);
\draw[bend left=45] (4.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}\ = \ket{00}+\ket{11}
\qquad\qquad
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (1.5, 0.75) {};
\node [style=dot] (1) at (-1.5, 0.25) {};
\node [style=dot] (2) at (1.5, 0.25) {};
\node [style=none] (3) at (0, 0) {$:=$};
\node [style=none] (4) at (-2, -0.25) {};
\node [style=none] (5) at (-1, -0.25) {};
\node [style=none] (6) at (1, -0.25) {};
\node [style=none] (7) at (2, -0.25) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2);
\draw (7.center) to (2);
\draw[bend right=45] (5.center) to (1);
\draw (6.center) to (2);
\draw[bend left=45] (4.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}\ = \ket{01}+\ket{10}
\]
However, the composition of a cap from one CFA with a cup from the other yields the Pauli X, or `NOT', gate:
\[
\begin{tikzpicture}[dotpic,scale=0.5]
\node [bn] (0) at (0,1.5) {};
\node [bn] (1) at (0,-1.5) {};
\draw (0)-- node[tick]{-} (1);
\end{tikzpicture}
:=\
\begin{tikzpicture}[dotpic,yshift=-5mm,scale=0.5]
\node [bn] (b0) at (-1,2) {};
\node [dot] (0) at (0,0) {};
\node [white dot] (1) at (1.5,1) {};
\node [bn] (b1) at (2.5,-1) {};
\draw (b0) to [out=-90,in=180] (0) (0) to [out=0,in=180] (1) (1) to [out=0,in=90] (b1);
\end{tikzpicture} \ =\
\begin{tikzpicture}[dotpic,yshift=-5mm,scale=0.5]
\node [bn] (b0) at (-1,2) {};
\node [white dot] (0) at (0,0) {};
\node [dot] (1) at (1.5,1) {};
\node [bn] (b1) at (2.5,-1) {};
\draw (b0) to [out=-90,in=180] (0) (0) to [out=0,in=180] (1) (1) to [out=0,in=90] (b1);
\end{tikzpicture}
\ =\
\left(\begin{array}{cc}
0 & 1 \\
1 & 0
\end{array}\right)
\]
These CFAs respectively induce the following tripartite states:
\[
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0, 0.75) {};
\node [style=white dot] (1) at (0, 0) {};
\node [style=white dot] (2) at (-0.5, -0.5) {};
\node [style=none] (3) at (-1, -1) {};
\node [style=none] (4) at (0, -1) {};
\node [style=none] (5) at (1, -1) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (4.center) to (2);
\draw (3.center) to (2);
\draw (2) to (1);
\draw (0) to (1);
\draw (5.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}
\ = \ket{000}+\ket{111} = \ket{\textit{GHZ}\,}
\qquad
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 0.75) {};
\node [style=dot] (1) at (0, 0) {};
\node [style=dot] (2) at (-0.5, -0.5) {};
\node [style=none] (3) at (-1, -1) {};
\node [style=none] (4) at (0, -1) {};
\node [style=none] (5) at (1, -1) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (4.center) to (2);
\draw (3.center) to (2);
\draw (2) to (1);
\draw (0) to (1);
\draw (5.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}
\ = \ket{100}+\ket{010}+\ket{001} = \ket{\textit{W\,}}
\]
As the name suggests, the associated tripartite state of the above GHZ-structure is a GHZ state,
and that of the W-structure is a W state. Furthermore, Theorem \ref{GHZ/Wthm} asserts that for
qubits, the associated tripartite state of \emph{any} GHZ-structure (resp. W-structure) is a GHZ
state (resp. W state), up to local operations.
\begin{theorem}{\rm\cite{CK}}\label{GHZ/Wthm}
For any special (respectively~anti-special) CFA on a qubit in ${\bf FHilb}$, the induced tripartite
state is SLOCC-equivalent to $\ket{\textit{GHZ}\,}$ (respectively~$\ket{\textit{W\,}}$). Furthermore, any tripartite state
$\ket\Psi$ either induces a special or anti-special CFA-structure, depending on whether it is
SLOCC-equivalent to $\ket{\textit{GHZ}\,}$ or to $\ket{\textit{W\,}}$.
\end{theorem}
Theorem \ref{basisthm} justifies the alternative name \em basis structure \em for GHZ-structures.
\begin{theorem}{\rm\cite{Aguiar}}\label{basisthm}
Special commutative Frobenius algebras on a finite-dimensional Hilbert space ${\cal H}$ are in 1-to-1 correspondence with (possibly non-orthogonal) bases for ${\cal H}$.
\end{theorem}
For any special CFA, phases are matrices that are diagonal in the corresponding basis. The
corresponding $|\psi\rangle$ (as in proposition \ref{prop:phases-states}) lies on the equator of
the Bloch sphere, justifying the name `phases'.
We can also consider interactions between a GHZ-structure and a W-structure.
\begin{definition}{\rm\cite{CK}}
A GHZ- and a W-structure form a \em GHZ/W-pair \em if the following equations hold:
\begin{center}
\raisebox{4mm}{
\begin{tikzpicture}[dotpic,scale=0.5]
\node [bn] (0) at (0,1.5) {};
\node [bn] (1) at (0,-1.5) {};
\draw (0)-- node[tick]{-} (1);
\end{tikzpicture}
\ $:=$\ \
\begin{tikzpicture}[dotpic,yshift=-5mm,scale=0.5]
\node [bn] (b0) at (-1,2) {};
\node [dot] (0) at (0,0) {};
\node [white dot] (1) at (1.5,1) {};
\node [bn] (b1) at (2.5,-1) {};
\draw (b0) to [out=-90,in=180] (0) (0) to [out=0,in=180] (1) (1) to [out=0,in=90] (b1);
\end{tikzpicture}
\ $\stackrel{\alpha}{=}$\ \
\begin{tikzpicture}[dotpic,yshift=-5mm,scale=0.5]
\node [bn] (b0) at (-1,2) {};
\node [white dot] (0) at (0,0) {};
\node [dot] (1) at (1.5,1) {};
\node [bn] (b1) at (2.5,-1) {};
\draw (b0) to [out=-90,in=180] (0) (0) to [out=0,in=180] (1) (1) to [out=0,in=90] (b1);
\end{tikzpicture}
}
\qquad
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (2.5, 1.5) {};
\node [style=none] (1) at (5, 1.5) {};
\node [style=dot] (2) at (-3, 1.25) {};
\node [style=dot] (3) at (-0.75, 1) {};
\node [style=dot] (4) at (0.25, 1) {};
\node [style=none] (5) at (-1.75, 0.75) {$\stackrel{\beta}{=}$};
\node [style=none] (6) at (3.75, 0.75) {$\stackrel{\gamma}{=}$};
\node [style=white dot] (7) at (2.5, 0.5) {};
\node [style=white dot] (8) at (5, 0.5) {};
\node [style=white dot] (9) at (-3, 0.25) {};
\node [style=none] (10) at (-0.75, 0) {};
\node [style=none] (11) at (0.25, 0) {};
\node [style=none] (12) at (-3.5, -0.25) {};
\node [style=none] (13) at (-2.5, -0.25) {};
\node [style=none] (14) at (1.75, -0.25) {};
\node [style=none] (15) at (3.25, -0.25) {};
\node [style=none] (16) at (4.25, -0.25) {};
\node [style=none] (17) at (5.75, -0.25) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (7) to (15.center);
\draw (2) to (9);
\draw (12.center) to (9);
\draw (9) to (13.center);
\draw (8) to node[tick]{-} (17.center);
\draw (16.center) to node[tick]{-} (8);
\draw (0.center) to node[tick]{-} (7);
\draw (3) to (10.center);
\draw (1.center) to (8);
\draw (14.center) to (7);
\draw (4) to (11.center);
\end{pgfonlayer}
\end{tikzpicture}
\qquad\ \ \raisebox{3mm}{\ensuremath{\circl \dottickunit{small dot} \stackrel{\xi}{=}\, \lolli}}
\end{center}
\end{definition}
By eqs.~($\beta$, $\gamma$) we also have:\vspace{-5mm}
\begin{center}
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (-1.5, -0.75) {};
\node [style=none] (1) at (0, -1.25) {$\stackrel{\beta'}{=}$};
\node [style=dot] (2) at (1.25, -1.25) {};
\node [style=dot] (3) at (2.25, -1.25) {};
\node [style=white dot] (4) at (-1.5, -1.75) {};
\node [style=none] (5) at (-2, -2.25) {};
\node [style=none] (6) at (-1, -2.25) {};
\node [style=none] (7) at (1.25, -2.25) {};
\node [style=none] (8) at (2.25, -2.25) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (4);
\draw (3) to node[tick]{-} (8.center);
\draw (5.center) to (4);
\draw (2) to node[tick]{-} (7.center);
\draw (4) to (6.center);
\end{pgfonlayer}
\end{tikzpicture}
\end{center}
\subsection{Plugging}
Since we are often concerned with objects in a monoidal category that are finitary in nature, we can deduce many new identities using a technique we call \emph{plugging}.
\begin{definition}\label{def:plugging-set}
A set of points $\{ \psi_i : I \rightarrow Q \}$ form a \emph{plugging set} for $Q$ if they suffice to distinguish maps from $Q$. That is, for all objects $A$ and maps $f,g : Q \rightarrow A$,
\[\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-3.5, 1.5) {};
\node [style=none] (1) at (-1.5, 1.5) {};
\node [style=white dot, font=\footnotesize] (2) at (2.5, 1.5) {$\psi_i$};
\node [style=white dot, font=\footnotesize] (3) at (4.5, 1.5) {$\psi_i$};
\node [style=none, font=\footnotesize] (4) at (-3.75, 1.25) {$Q$};
\node [style=none, font=\footnotesize] (5) at (-1.75, 1.25) {$Q$};
\node [style=square box] (6) at (-3.5, 0) {$f$};
\node [style=none] (7) at (-2.5, 0) {=};
\node [style=square box] (8) at (-1.5, 0) {$g$};
\node [style=none] (9) at (0, 0) {$\Leftrightarrow$};
\node [style=none] (10) at (1.5, 0) {$\forall i .$};
\node [style=square box] (11) at (2.5, 0) {$f$};
\node [style=none] (12) at (3.5, 0) {=};
\node [style=square box] (13) at (4.5, 0) {$g$};
\node [style=none, font=\footnotesize] (14) at (-3.75, -1.25) {$A$};
\node [style=none, font=\footnotesize] (15) at (-1.75, -1.25) {$A$};
\node [style=none, font=\footnotesize] (16) at (2.25, -1.25) {$A$};
\node [style=none, font=\footnotesize] (17) at (4.25, -1.25) {$A$};
\node [style=none] (18) at (-3.5, -1.5) {};
\node [style=none] (19) at (-1.5, -1.5) {};
\node [style=none] (20) at (2.5, -1.5) {};
\node [style=none] (21) at (4.5, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (18.center);
\draw (1.center) to (19.center);
\draw (3) to (21.center);
\draw (2) to (20.center);
\end{pgfonlayer}
\end{tikzpicture}\]
\end{definition}
When we prove a graphical identity by showing two maps are not distinguished by a plugging set, we
call this `proof by plugging.' Also note that we can extend such proofs to maps of the form $f : Q
\otimes A \rightarrow B$ or $f' : A \rightarrow Q \otimes B$ by using the Frobenius caps and cups
when $Q$ has a CFA $(\dotmult{small dot}, \dotunit{small dot}, \dotcomult{small dot}, \dotcounit{small dot})$ and $A$ a CFA $(\dotmult{small white dot}, \dotunit{small white dot},
\dotcomult{small white dot}, \dotcounit{small white dot})$,
\[
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-2.25, 1.75) {};
\node [style=white dot] (1) at (4.25, 1.75) {};
\node [style=none] (2) at (-4, 1.5) {};
\node [style=none] (3) at (1.5, 1.5) {};
\node [style=none, font=\footnotesize] (4) at (-4.25, 1.25) {$Q$};
\node [style=none, font=\footnotesize] (5) at (-3.25, 1.25) {$A$};
\node [style=none, font=\footnotesize] (6) at (1.75, 1.25) {$Q$};
\node [style=none, font=\footnotesize] (7) at (3.25, 1.25) {$A$};
\node [style=none] (8) at (-3, 1) {};
\node [style=none] (9) at (-1.5, 1) {};
\node [style=none] (10) at (3.5, 1) {};
\node [style=none] (11) at (5, 1) {};
\node [style=none] (12) at (-4, 0.5) {};
\node [style=none] (13) at (-3, 0.5) {};
\node [style=none] (14) at (3.5, 0.5) {};
\node [style=square box, minimum width=1 cm] (15) at (-3.5, 0) {$f$};
\node [style=square box, minimum width=1 cm] (16) at (3.5, 0) {$f'$};
\node [style=none] (17) at (-3.5, -0.5) {};
\node [style=none] (18) at (3, -0.5) {};
\node [style=none] (19) at (4, -0.5) {};
\node [style=none] (20) at (1.5, -0.75) {};
\node [style=none] (21) at (3, -0.75) {};
\node [style=none, font=\footnotesize] (22) at (-3.75, -1.25) {$B$};
\node [style=none, font=\footnotesize] (23) at (4.25, -1.25) {$B$};
\node [style=none] (24) at (-3.5, -1.5) {};
\node [style=none] (25) at (-1.5, -1.5) {};
\node [style=dot] (26) at (2.25, -1.5) {};
\node [style=none] (27) at (4, -1.5) {};
\node [style=none] (28) at (5, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2.center) to (12.center);
\draw (8.center) to (13.center);
\draw[out=0, in=90] (0) to (9.center);
\draw (17.center) to (24.center);
\draw[out=180, in=90] (1) to (10.center);
\draw (11.center) to (28.center);
\draw (9.center) to (25.center);
\draw (10.center) to (14.center);
\draw (19.center) to (27.center);
\draw[out=180, in=90] (0) to (8.center);
\draw[out=0, in=90] (1) to (11.center);
\draw (21.center) to (18.center);
\draw[out=0, in=-90] (26) to (21.center);
\draw (20.center) to (3.center);
\draw[out=180, in=-90] (26) to (20.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
The axioms of a GHZ/W-pair suffice to prove the following lemma for Hilbert spaces.
\begin{lemma}[\cite{CK}]\label{lem:dot-lolli-2d}
For a GHZ/W-pair on $H$ in ${\bf FHilb}$ with $\dim(H) \geq 2$, the points $\dotunit{small dot}$ and $\dottickunit{small dot}$ span a 2-dimensional space; hence for $H=\mathbb{C}^2$ the points $\dotunit{small dot}$ and $\dottickunit{small dot}$ form a basis.
\end{lemma}
Motivated by this fact, we assume the that $\{ \dotunit{small dot}, \dottickunit{small dot} \}$ forms a plugging set for $Q$. More explicitly:
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (-6, 0.75) {};
\node [style=dot] (1) at (-4.5, 0.75) {};
\node [style=dot] (2) at (-2.5, 0.75) {};
\node [style=dot] (3) at (-1, 0.75) {};
\node [style=none] (4) at (2, 0.75) {};
\node [style=none] (5) at (3.5, 0.75) {};
\node [style=square box] (6) at (-6, 0) {$f$};
\node [style=none] (7) at (-5.25, 0) {$=$};
\node [style=square box] (8) at (-4.5, 0) {$g$};
\node [style=none] (9) at (-3.5, 0) {$\wedge$};
\node [style=square box] (10) at (-2.5, 0) {$f$};
\node [style=none] (11) at (-1.75, 0) {$=$};
\node [style=square box] (12) at (-1, 0) {$g$};
\node [style=none] (13) at (0.5, 0) {$\Leftrightarrow$};
\node [style=square box] (14) at (2, 0) {$f$};
\node [style=none] (15) at (2.75, 0) {$=$};
\node [style=square box] (16) at (3.5, 0) {$g$};
\node [style=none] (17) at (-6, -0.75) {};
\node [style=none] (18) at (-4.5, -0.75) {};
\node [style=none] (19) at (-2.5, -0.75) {};
\node [style=none] (20) at (-1, -0.75) {};
\node [style=none] (21) at (2, -0.75) {};
\node [style=none] (22) at (3.5, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (10) to (19.center);
\draw (0) to (6);
\draw (2) to node[tick]{-} (10);
\draw (16) to (22.center);
\draw (8) to (18.center);
\draw (1) to (8);
\draw (6) to (17.center);
\draw (12) to (20.center);
\draw (14) to (21.center);
\draw (4.center) to (14);
\draw (5.center) to (16);
\draw (3) to node[tick]{-} (12);
\end{pgfonlayer}
\end{tikzpicture}
\]
\section{Arithmetic from a GHZ/W-pair}\label{sec:arithmetic}
Given a GHZ/W-pair, we can extract an arithmetic system. First, we establish some preliminary results.
\subsection{Properties of GHZ-phases}
Below, all phases are GHZ-phases. When relying on plugging, we have the following:
\newcommand{\psidot}{
\raisebox{5mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0, 0.5) {$\psi$};
\node [style=dot] (1) at (0, -0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (1);
\end{pgfonlayer}
\end{tikzpicture}}}
\newcommand{\psitickdot}{
\raisebox{5mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0, 0.5) {$\psi$};
\node [style=dot] (1) at (0, -0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (1.center);
\end{pgfonlayer}
\end{tikzpicture}}}
\newcommand{\smallpsidot}{
\raisebox{-2.5mm}{\begin{tikzpicture}[scale=0.5]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0, 0.5) {\footnotesize$\!\psi$};
\node [style=dot] (1) at (0, -0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (1);
\end{pgfonlayer}
\end{tikzpicture}}}
\newcommand{\smallpsitickdot}{
\raisebox{-2.5mm}{\begin{tikzpicture}[scale=0.5]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0, 0.5) {\footnotesize$\!\psi$};
\node [style=dot] (1) at (0, -0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (1.center);
\end{pgfonlayer}
\end{tikzpicture}}}
\begin{theorem}\label{thmdelta_1}
\[
\raisebox{-8mm}{
\psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-0.5, 1.5) {};
\node [style=none] (1) at (0.5, 1.5) {};
\node [style=dot] (2) at (0, 0.5) {};
\node [style=white dot] (3) at (0, -0.5) {$\psi$};
\node [style=none] (4) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (0) to (2);
\draw [bend right=30] (2) to (1);
\draw (2) to (3.center);
\draw (3.center) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ \stackrel{\delta_1}{=}\ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-0.75, 1.5) {};
\node [style=none] (1) at (0.75, 1.5) {};
\node [style=white dot] (2) at (-0.75, 0.5) {$\psi$};
\node [style=white dot] (3) at (0.75, 0.5) {$\psi$};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=none] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw [bend right] (2.center) to (4);
\draw [bend right] (4) to (3.center);
\draw (3.center) to (1);
\draw (4) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\end{theorem}
\begin{proof}
Plugging $\dotunit{small dot}$ to one input:
\[
\raisebox{-8mm}{
\psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (-0.5, 1.25) {};
\node [style=none] (1) at (0.5, 1.5) {};
\node [style=dot] (2) at (0, 0.5) {};
\node [style=white dot] (3) at (0, -0.5) {$\psi$};
\node [style=none] (4) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (0) to (2);
\draw [bend right=30] (2) to (1);
\draw (2) to (3.center);
\draw (3.center) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=white dot] (3) at (0, 0) {$\psi$};
\node [style=none] (4) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (3.center);
\draw (3.center) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (-0.5, -0.5) {};
\node [style=none] (1) at (0.5, 1.5) {};
\node [style=dot] (2) at (0, -1) {};
\node [style=white dot] (3) at (0.5, 0) {$\psi$};
\node [style=none] (4) at (0, -1.5) {};
\node [style=white dot] (5) at (-0.5, 1) {$\psi$};
\node [style=dot] (6) at (-0.5, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (0) to (2);
\draw [bend right=30] (2) to (3.center);
\draw (1) to (3.center);
\draw (2.center) to (4);
\draw (5) to node[tick]{-} (6.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.25) {};
\node [style=none] (1) at (0.75, 1.5) {};
\node [style=dot] (2) at (0, -1) {};
\node [style=white dot] (3) at (0.75, 0) {$\psi$};
\node [style=none] (4) at (0, -1.5) {};
\node [style=white dot] (5) at (-1.5, 1) {$\psi$};
\node [style=white dot] (6) at (-0.75, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (6) to (2);
\draw [bend right=30] (5) to (6) (6) to (0);
\draw [bend right=30] (2) to (3.center);
\draw (1) to (3.center);
\draw (2.center) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (-0.75, 1.25) {};
\node [style=none] (1) at (0.75, 1.5) {};
\node [style=white dot] (2) at (-0.75, 0.5) {$\psi$};
\node [style=white dot] (3) at (0.75, 0.5) {$\psi$};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=none] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw [bend right] (2.center) to (4);
\draw [bend right] (4) to (3.center);
\draw (3.center) to (1);
\draw (4) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\]
Plugging $\dottickunit{small dot}$ to one input of both sides:
\[
\raisebox{-8mm}{
\psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (-0.5, 1.25) {};
\node [style=none] (1) at (0.5, 1.5) {};
\node [style=dot] (2) at (0, 0.5) {};
\node [style=white dot] (3) at (0, -0.5) {$\psi$};
\node [style=none] (4) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (0) to node[tick]{-} (2);
\draw [bend right=30] (2) to (1);
\draw (2) to (3.center);
\draw (3.center) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=dot] (1) at (0, 1) {};
\node [style=dot] (2) at (0, 0.5) {};
\node [style=white dot] (3) at (0, -0.5) {$\psi$};
\node [style=none] (4) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (1) (2) to node[tick]{-} (3);
\draw (3.center) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0.5, 0.5) {};
\node [style=none] (1) at (0.5, 1.5) {};
\node [style=dot] (5) at (0.5, 0) {};
\node [style=white dot] (2) at (0, -1) {};
\node [style=white dot] (3) at (-0.5, 0) {$\psi$};
\node [style=none] (4) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (1);
\draw [bend right=30] (3.center) to (2.center);
\draw [bend right=30] (2.center) to node[tick]{-} (5.center);
\draw (2.center) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\psidot \ \psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=dot] (1) at (0, 0.25) {};
\node [style=dot] (2) at (0, -0.25) {};
\node [style=none] (3) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (1) (2) to node[tick]{-} (3);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (-0.75, 1.5) {};
\node [style=none] (1) at (0.75, 1.5) {};
\node [style=white dot] (2) at (-0.75, 0.5) {$\psi$};
\node [style=white dot] (3) at (0.75, 0.5) {$\psi$};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=none] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (2);
\draw [bend right] (2.center) to (4);
\draw [bend right] (4) to (3.center);
\draw (3.center) to (1);
\draw (4) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (-0.5, 1.25) {};
\node [style=white dot] (1) at (-1.5, 1) {$\psi$};
\node [style=white dot] (2) at (-1, 0) {};
\node [style=dot] (3) at (0, -1) {};
\node [style=white dot] (4) at (1, 0) {};
\node [style=white dot] (5) at (0.5, 1.25) {$\psi$};
\node [style=none] (6) at (1.5, 1.5) {};
\node [style=none] (7) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (1) to (2) (2) to node[tick]{-} (0);
\draw [bend right=30] (5) to (4) (4) to (6);
\draw [bend right=30] (2) to (3) (3) to (4);
\draw (7) to (3);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\psidot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (2) at (-1, 0) {};
\node [style=dot] (3) at (0, -1) {};
\node [style=white dot] (4) at (1, 0) {};
\node [style=white dot] (5) at (0.5, 1.25) {$\psi$};
\node [style=none] (6) at (1.5, 1.5) {};
\node [style=none] (7) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (5) to (4) (4) to (6);
\draw [bend right=30] (2) to node[tick]{-} (3) (3) to (4);
\draw (7) to (3);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\psidot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (4) at (0, 0) {};
\node [style=white dot] (5) at (-0.5, 1.25) {$\psi$};
\node [style=none] (6) at (0.5, 1.5) {};
\node [style=dot] (7) at (0, -0.5) {};
\node [style=dot] (8) at (0, -1) {};
\node [style=none] (9) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (5) to (4) (4) to (6);
\draw (7) to node[tick]{-} (4) (8) to node[tick]{-} (9);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\psidot \ \psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=dot] (1) at (0, 0.25) {};
\node [style=dot] (2) at (0, -0.25) {};
\node [style=none] (3) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (1) (2) to node[tick]{-} (3);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\end{proof}
\begin{theorem}\label{thmdelta_2}
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (2.center);
\draw (2.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ \stackrel{\delta_2}{=}\ \
\raisebox{-8mm}{
\psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0)--(1);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\end{theorem}
\begin{proof}
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (2.center);
\draw (2.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0) {};
\node [style=white dot] (3) at (-1, 1) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (2.center);
\draw (2.center) to (1);
\draw [bend right=30] (3) to (2);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0)--(1);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\end{proof}
Note that for $\dotmult{small white dot}$-phases we have:
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0) {$\frac{1}{\psi}$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw (2.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ := \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (2.center);
\draw (2.center) to node[tick]{-} (1);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0) {};
\node [style=white dot] (3) at (1, 1) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (2.center);
\draw (2.center) to node[tick]{-} (1);
\draw [bend right=60] (2.center) to (3.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0) {};
\node [style=white dot] (3) at (1, 1) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw (2.center) to (1);
\draw [bend right=60] (2.center) to node[tick]{-} (3.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=small white dot,font=\footnotesize] (3) at (1, 1) {${\tickpsi}$};
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, -0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=60] (2.center) to (3.center);
\draw (1.center) to (2.center);
\draw (0.center) to (2.center);
\end{pgfonlayer}
\end{tikzpicture}}
\]
The particular choice of notation $\frac{1}{\psi}$ is justified below, and will play a key role in this paper.
\begin{theorem}\label{thmdelta_3}
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=small white dot,font=\footnotesize] (2) at (0, 0.75) {$\psi$};
\node [style=small white dot,font=\footnotesize] (3) at (0, -0.5) {$\frac{1}{\psi}$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (2.center);
\draw (2.center) to (3.center);
\draw (3.center) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ \stackrel{\delta_3}{=}\ \
\raisebox{-8mm}{
\psidot \ \psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0)--(1);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\end{theorem}
\begin{proof}
Plugging $\dotunit{small dot}$ into the input:
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=small white dot,font=\footnotesize] (2) at (0, 0.75) {$\psi$};
\node [style=small white dot,font=\footnotesize] (3) at (0, -0.5) {$\frac{1}{\psi}$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (2.center);
\draw (2.center) to (3.center);
\draw (3.center) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0.75) {};
\node [style=white dot] (3) at (0, -0.5) {};
\node [style=white dot] (4) at (-1, 1.25) {$\psi$};
\node [style=white dot] (5) at (-1, 0) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (2.center);
\draw (2.center) to (3.center);
\draw (3.center) to (1.center);
\draw [bend right=30] (4) to (2) (5) to node[tick]{-} (3);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\psidot \ \psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0)--(1);
\end{pgfonlayer}
\end{tikzpicture}}
\]
Plugging $\dottickunit{small dot}$ into the input:
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=small white dot,font=\footnotesize] (2) at (0, 0.75) {$\psi$};
\node [style=small white dot,font=\footnotesize] (3) at (0, -0.5) {$\frac{1}{\psi}$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (2);
\draw (2.center) to (3.center);
\draw (3.center) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0.75) {};
\node [style=white dot] (3) at (0, -0.5) {};
\node [style=white dot] (4) at (-1, 1.25) {$\psi$};
\node [style=white dot] (5) at (-1, 0) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (2);
\draw (2.center) to (3.center);
\draw (3.center) to (1.center);
\draw [bend right=30] (4) to (2) (5) to node[tick]{-} (3);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{
\psidot \ \psitickdot \
\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) -- node[tick]{-} (1);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\end{proof}
In a setting like ${\bf FHilb}_p$, where we ignore cancelable scalar multipliers, and provided
that the scalars $\smallpsidot$ and $\smallpsitickdot$ in the equation are cancelable, eqs.~($\delta_1$, $\delta_2$, $\delta_3$) simplify to:
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-0.5, 1.5) {};
\node [style=none] (1) at (0.5, 1.5) {};
\node [style=dot] (2) at (0, 0.5) {};
\node [style=white dot] (3) at (0, -0.5) {$\psi$};
\node [style=none] (4) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (0) to (2);
\draw [bend right=30] (2) to (1);
\draw (2) to (3.center);
\draw (3.center) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ \stackrel{\delta_1}{=}\ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-0.75, 1.5) {};
\node [style=none] (1) at (0.75, 1.5) {};
\node [style=white dot] (2) at (-0.75, 0.5) {$\psi$};
\node [style=white dot] (3) at (0.75, 0.5) {$\psi$};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=none] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw [bend right] (2.center) to (4);
\draw [bend right] (4) to (3.center);
\draw (3.center) to (1);
\draw (4) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\qquad \qquad \qquad
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, 0) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (2.center);
\draw (2.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ \stackrel{\delta_2}{=}\ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0)--(1);
\end{pgfonlayer}
\end{tikzpicture}}
\qquad \qquad \qquad
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=small white dot,font=\footnotesize] (2) at (0, 0.75) {$\psi$};
\node [style=small white dot,font=\footnotesize] (3) at (0, -0.5) {$\frac{1}{\psi}$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (2.center);
\draw (2.center) to (3.center);
\draw (3.center) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ \stackrel{\delta_3}{=}\ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0)--(1);
\end{pgfonlayer}
\end{tikzpicture}}
\]
Examples of phases for which some of these simplified equations fail to hold are:
\[
\begin{tikzpicture}[scale=0.4]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-7, 1.5) {};
\node [style=none] (1) at (-3.25, 1.5) {};
\node [style=none] (2) at (3.25, 1.5) {};
\node [style=none] (3) at (7, 1.5) {};
\node [style=dot] (4) at (-6, 0.75) {};
\node [style=dot] (5) at (4.25, 0.75) {};
\node [style=dot] (6) at (-3.25, 0.5) {};
\node [style=dot] (7) at (7, 0.5) {};
\node [style=none] (8) at (-4.75, 0) {$=$};
\node [style=none] (9) at (5.5, 0) {$=$};
\node [style=white dot] (10) at (-7, -0.5) {};
\node [style=dot] (11) at (-3.25, -0.5) {};
\node [style=white dot] (12) at (3.25, -0.5) {};
\node [style=dot] (13) at (7, -0.5) {};
\node [style=none] (14) at (-7, -1.5) {};
\node [style=none] (15) at (-3.25, -1.5) {};
\node [style=none] (16) at (3.25, -1.5) {};
\node [style=none] (17) at (7, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (10);
\draw (11) to (15.center);
\draw (16.center) to (12);
\draw (14.center) to (10);
\draw (17.center) to node[tick]{-} (13);
\draw[bend right=45, looseness=0.75] (10) to (4);
\draw[bend right=45, looseness=0.75] (12) to node[tick]{-} (5);
\draw (1.center) to node[tick]{-} (6);
\draw (7) to (3.center);
\draw (2.center) to (12);
\end{pgfonlayer}
\end{tikzpicture}
\]
\subsection{Natural Number Arithmetic}\label{sec:natural-numbers}
Assume that we are given a GHZ/W-pair. In particular, we have two internal commutative monoids $(\dotmult{small dot}, \dotunit{small dot})$ and $(\dotmult{small white dot}, \dotunit{small white dot})$. We will now consider the induced commutative monoids on elements:
\[
\left(
\raisebox{-3mm}{\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 1.25) {$\psi$};
\node [style=none] (1) at (-0.5, 0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}
\raisebox{1.5mm}{,}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 1.25) {$\phi$};
\node [style=none] (1) at (-0.5, 0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}}
\right)
\mapsto
\raisebox{-4mm}{\begin{tikzpicture}[scale=0.8]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 1.25) {$\psi$};
\node [style=white dot] (1) at (0.5, 1.25) {$\phi$};
\node [style=dot] (2) at (0, 0.75) {};
\node [style=none] (3) at (0, 0.25) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2);
\draw (1) to (2);
\draw (2) to (3.center);
\end{pgfonlayer}
\end{tikzpicture}}
\qquad\qquad
\left(
\raisebox{-3mm}{\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 1.25) {$\psi$};
\node [style=none] (1) at (-0.5, 0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}
\raisebox{1.5mm}{,}
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 1.25) {$\phi$};
\node [style=none] (1) at (-0.5, 0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}}
\right)
\mapsto
\raisebox{-4mm}{\begin{tikzpicture}[scale=0.8]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 1.25) {$\psi$};
\node [style=white dot] (1) at (0.5, 1.25) {$\phi$};
\node [style=white dot] (2) at (0, 0.75) {};
\node [style=none] (3) at (0, 0.25) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2);
\draw (1) to (2);
\draw (2) to (3.center);
\end{pgfonlayer}
\end{tikzpicture}}
\]
We will call $\dotmult{small dot}$ applied to elements \em addition \em and $\dotmult{small white dot}$ applied to elements
\em multiplication\em, for reasons that will become apparent shortly. Similarly, we call $\dotunit{small dot}$
the \em unit for addition \em and $\dotunit{small white dot}$ the \em unit for multiplication\em. By Theorem
\ref{thmdelta_1} we have a distributivity law, up to a scalar, partially explaining our choices of
the names addition and multiplication for the monoids:
\begin{corollary}\label{cor-dist}
\[
\begin{tikzpicture}[scale=1]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-2.25, 1) {$\phi$};
\node [style=white dot] (1) at (-1.25, 1) {$\varphi$};
\node [style=white dot] (2) at (0.5, 1) {$\psi$};
\node [style=white dot] (3) at (1.25, 1) {$\phi$};
\node [style=white dot] (4) at (2.25, 1) {$\psi$};
\node [style=white dot] (5) at (3, 1) {$\varphi$};
\node [style=white dot] (6) at (-2.75, 0.5) {$\psi$};
\node [style=dot] (7) at (-1.75, 0.5) {};
\node [style=white dot] (8) at (1.25, 0.5) {};
\node [style=white dot] (9) at (2.25, 0.5) {};
\node [style=white dot] (10) at (-3.5, 0.25) {$\psi$};
\node [style=white dot] (11) at (-2.25, 0) {};
\node [style=none] (12) at (-0.25, 0) {$=$};
\node [style=dot] (13) at (1.75, 0) {};
\node [style=dot] (14) at (-3.5, -0.25) {};
\node [style=none] (15) at (-2.25, -0.5) {};
\node [style=none] (16) at (1.75, -0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (7) to (11);
\draw (3) to (8);
\draw (11) to (15.center);
\draw (9) to (13);
\draw (10) to node[tick]{-} (14);
\draw (6) to (11);
\draw (2) to (8);
\draw (8) to (13);
\draw (1) to (7);
\draw (4) to (9);
\draw (0) to (7);
\draw (13) to (16.center);
\draw (5) to (9);
\end{pgfonlayer}
\end{tikzpicture}
\]
\end{corollary}
Moreover, we can use these to do concrete arithmetic on the natural numbers. We start by defining an encoding for the natural numbers:
\[
\begin{tikzpicture}[scale=1]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-5.75, -1) {$0$};
\node [style=dot] (1) at (-3.75, -1) {};
\node [style=white dot] (2) at (-0.75, -1) {$n\!+\!1$};
\node [style=white dot] (3) at (1.25, -1) {$n$};
\node [style=white dot] (4) at (2.25, -1) {};
\node [style=none] (5) at (-4.75, -1.5) {$=$};
\node [style=none] (6) at (0.25, -1.5) {$=$};
\node [style=dot] (7) at (1.75, -1.5) {};
\node [style=none] (8) at (-5.75, -2) {};
\node [style=none] (9) at (-3.75, -2) {};
\node [style=none] (10) at (-0.75, -2) {};
\node [style=none] (11) at (1.75, -2) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (4) to (7);
\draw (1) to (9.center);
\draw (3) to (7);
\draw (0) to (8.center);
\draw (7) to (11.center);
\draw (2) to (10.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
From hence forth, we shall assume we are working in a category with no non-trivial invertible (i.e.
non-zero) scalars, such as ${\bf FHilb}_p$. Thus, we shall drop any invertible scalars.
Furthermore, we shall assume scalar (i.) is invertible for all $n$ and scalar (ii.) is invertible
for all $n \neq 0$:
\[
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-1, 0.5) {$n$};
\node [style=white dot] (1) at (2, 0.5) {$n$};
\node [style=none] (2) at (-2, 0) {(i.)};
\node [style=none] (3) at (1, 0) {(ii.)};
\node [style=dot] (4) at (-1, -0.5) {};
\node [style=dot] (5) at (2, -0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (1) to (5);
\draw (0) to node[tick]{-} (4);
\end{pgfonlayer}
\end{tikzpicture}
\]
That $\dotmult{small dot}$ is the normal addition operation for these numbers follows immediately from their
definition and associativity of $\dotmult{small dot}$. We can also show that $\dotmult{small white dot}$ is the normal
multiplication operation (noting first that the encoding of $1$ is $\dotunit{small white dot}$, and hence the unit
of $\dotmult{small white dot}$):
\[
\raisebox{-8mm}{\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 1.25) {$n$};
\node [style=white dot] (1) at (0.5, 1.25) {$m$};
\node [style=white dot] (2) at (0, 0.75) {};
\node [style=none] (3) at (0, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (2.center);
\draw (1.center) to (2.center);
\draw (2.center) to (3);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0.5, 1.75) {$\mdots$};
\node [style=none] (4) at (0.5, 1.25) {};
\node [style=white dot] (1) at (-0.5, 1.25) {$n$};
\node [style=white dot] (2) at (0, 0.75) {};
\node [style=none] (3) at (0, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (4) to (2);
\draw (1) to (2);
\draw (2) to (3);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ \stackrel{\delta_1}{=}\ \
\raisebox{-8mm}{$\overbrace{
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-1, 2) {$n$};
\node [style=white dot] (1) at (-0.5, 2) {};
\node [style=white dot] (2) at (-0.75, 1.5) {};
\node [style=white dot] (3) at (1, 2) {};
\node [style=white dot] (4) at (0.5, 2) {$n$};
\node [style=white dot] (5) at (0.75, 1.5) {};
\node [style=none] (6) at (0, 1.5) {$\ldots$};
\node [style=dot] (7) at (0, 0.75) {};
\node [style=none] (8) at (0, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2);
\draw (1) to (2);
\draw (3) to (5);
\draw (4) to (5);
\draw (7) to (2);
\draw (7) to (5);
\draw (7) to (8);
\end{pgfonlayer}
\end{tikzpicture}}^m$}
\ \ = \ \
\raisebox{-8mm}{$\overbrace{
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.75, 1.5) {$n$};
\node [style=white dot] (1) at (0.75, 1.5) {$n$};
\node [style=none] (2) at (0, 1.25) {$\ldots$};
\node [style=dot] (3) at (0, 0.75) {};
\node [style=none] (4) at (0, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (3);
\draw (1) to (3);
\draw (3) to (4);
\end{pgfonlayer}
\end{tikzpicture}}^m$}
\ \ = \ \
\raisebox{-8mm}{
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0, 1.25) {$n m$};
\node [style=none] (1) at (0, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}}
\]
The distributivity law stated earlier now translates into the normal distributivity of multiplication over addition in the natural numbers, up to a scalar.
\subsection{Multiplicative inverses}\label{sec:rationals}
By Theorem \ref{thmdelta_3} we have that $\tick$ is also an inverse for $\dotmult{small white dot}$, up to a scalar:
\begin{corollary}\label{cor-inverse}
\[
\begin{tikzpicture}[scale=1]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-2.75, 0.5) {$\psi$};
\node [style=white dot] (1) at (-1.75, 0.5) {$\psi$};
\node [style=white dot] (2) at (1.5, 0.5) {};
\node [style=white dot] (3) at (0, 0.25) {$\psi$};
\node [style=white dot] (4) at (0.75, 0.25) {$\psi$};
\node [style=white dot] (5) at (-2.25, 0) {};
\node [style=none] (6) at (-1, 0) {$=$};
\node [style=dot] (7) at (0, -0.25) {};
\node [style=dot] (8) at (0.75, -0.25) {};
\node [style=none] (9) at (-2.25, -0.5) {};
\node [style=none] (10) at (1.5, -0.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (1) to node[tick]{-} (5);
\draw (0) to (5);
\draw (3) to (7);
\draw (5) to (9.center);
\draw (4) to node[tick]{-} (8);
\draw (2) to (10.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
\end{corollary}
Hence, we have an encoding for the multiplicative inverses of the natural numbers:
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0.5, 1) {$\frac{1}{n}$};
\node [style=none] (1) at (0.5, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}
\ \ \, \raisebox{6.5mm}{:=}\,
\overbrace{
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0.75, 1.5) {};
\node [style=none] (1) at (1.25, 1.5) {$\ldots$};
\node [style=white dot] (2) at (1.75, 1.5) {};
\node [style=dot] (3) at (1.25, 1) {};
\node [style=none] (4) at (1.25, 0.25) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2) to (3);
\draw (0) to (3);
\draw (3) to node[tick]{-} (4.center);
\end{pgfonlayer}
\end{tikzpicture}}^n
\, \raisebox{6.5mm}{=}\, \ \
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0.5, 1) {$n$};
\node [style=none] (1) at (0.5, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[tick]{-} (1.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
Thoughtout this paper, we shall assume any natural number occurring in the denominator is not equal
to $0$. This allows us to encode positive fractions in the following form:
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0.5, -1.25) {$n$};
\node [style=white dot] (1) at (2, -1.25) {$\frac{1}{m}$};
\node [style=white dot] (2) at (3.5, -1.25) {$\frac{1}{m}$};
\node [style=white dot] (3) at (5, -1.25) {$n$};
\node [style=white dot] (4) at (-1.25, -1.75) {$\frac{n}{m}$};
\node [style=none] (5) at (-0.25, -2) {$=$};
\node [style=white dot] (6) at (1.25, -2) {};
\node [style=none] (7) at (2.75, -2) {$=$};
\node [style=white dot] (8) at (4.25, -2) {};
\node [style=none] (9) at (-1.25, -2.75) {};
\node [style=none] (10) at (1.25, -2.75) {};
\node [style=none] (11) at (4.25, -2.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2) to (8);
\draw (4) to (9.center);
\draw (8) to (11.center);
\draw (0) to (6);
\draw (3) to (8);
\draw (6) to (10.center);
\draw (1) to (6);
\end{pgfonlayer}
\end{tikzpicture}
\]
where the second equality follows from associativity and commutativity of the GHZ-structure.
\begin{remark}
It should be noted that the construction of this encoding depends on the choice of numerator and denominator,
and not just on the rational number being represented. Therefore, we should demonstrate that the actual point
depends only on the number represented. We start by noting that, by corollary \ref{cor-inverse},
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0, 0.75) {$n$};
\node [style=white dot] (1) at (1.5, 0.75) {$\frac{1}{n}$};
\node [style=white dot] (2) at (-1.75, 0.25) {$\frac{n}{n}$};
\node [style=white dot] (3) at (3.75, 0.25) {};
\node [style=none] (4) at (-0.75, -0) {$=$};
\node [style=white dot] (5) at (0.75, -0) {};
\node [style=none] (6) at (2.25, -0) {$=$};
\node [style=none] (7) at (-1.75, -0.75) {};
\node [style=none] (8) at (0.75, -0.75) {};
\node [style=none] (9) at (3.75, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2) to (7.center);
\draw (3) to (9.center);
\draw (0) to (5);
\draw (5) to (8.center);
\draw (1) to (5);
\end{pgfonlayer}
\end{tikzpicture}
\]
where we have ignored the scalar in corollary \ref{cor-inverse}, since it is cancellable
for the points that we have used to encode the natural numbers.
We know that if $\frac{n}{m} = \frac{n'}{m'}$, there
is a $p$ such that $n = n'p$ and $m = m'p$, so it follows easily that
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-5, 0.75) {\scriptsize$n$};
\node [style=white dot] (1) at (-3.5, 0.75) {$\frac{1}{m}$};
\node [style=white dot] (2) at (-2.25, 0.75) {\scriptsize$n'p$};
\node [style=white dot] (3) at (-0.75, 0.75) {$\frac{1}{m'p}$};
\node [style=white dot] (4) at (0.5, 0.75) {\scriptsize$n'$};
\node [style=white dot] (5) at (1.25, 0.75) {\scriptsize$p$};
\node [style=white dot] (6) at (2, 0.75) {$\frac{1}{m'}$};
\node [style=white dot] (7) at (3, 0.75) {$\frac{1}{p}$};
\node [style=white dot] (8) at (4.25, 0.75) {$\frac{n'}{m'}$};
\node [style=white dot] (9) at (5.75, 0.75) {$\frac{p}{p}$};
\node [style=white dot] (10) at (-6.25, 0.25) {$\frac{n}{m}$};
\node [style=white dot] (11) at (6.75, 0.25) {$\frac{n'}{m'}$};
\node [style=none] (12) at (-5.5, -0) {$=$};
\node [style=white dot] (13) at (-4.25, -0) {};
\node [style=none] (14) at (-2.75, -0) {$=$};
\node [style=white dot] (15) at (-1.5, -0) {};
\node [style=none] (16) at (0, -0) {$=$};
\node [style=white dot] (17) at (1.75, 0) {};
\node [style=none] (18) at (3.5, -0) {$=$};
\node [style=white dot] (19) at (5, -0) {};
\node [style=none] (20) at (6, -0) {$=$};
\node [style=none] (21) at (-6.25, -0.75) {};
\node [style=none] (22) at (-4.25, -0.75) {};
\node [style=none] (23) at (-1.5, -0.75) {};
\node [style=none] (24) at (1.75, -0.75) {};
\node [style=none] (25) at (5, -0.75) {};
\node [style=none] (26) at (6.75, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (7) to (17);
\draw (9) to (19);
\draw (4) to (17);
\draw (1) to (13);
\draw (17) to (24.center);
\draw (19) to (25.center);
\draw (8) to (19);
\draw (6) to (17);
\draw (5) to (17);
\draw (2) to (15);
\draw (10) to (21.center);
\draw (11) to (26.center);
\draw (0) to (13);
\draw (13) to (22.center);
\draw (3) to (15);
\draw (15) to (23.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
\end{remark}
\begin{example}\label{ex:invfac}
All the usual properties of fractions follow in a straightforward manner from the axioms of rational arithmetic that we have proved. For example,
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0.75, 1.5) {$\frac{n}{m}$};
\node [style=white dot] (1) at (2.25, 1.5) {$\frac{n'}{m'}$};
\node [style=white dot] (2) at (4.25, 1) {$\frac{n n'}{m m'}$};
\node [style=white dot] (3) at (1.5, 0.75) {};
\node [style=none] (4) at (3.25, 0.75) {$=$};
\node [style=none] (5) at (1.5, 0) {};
\node [style=none] (6) at (4.25, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (3);
\draw (1) to (3);
\draw (2) to (6.center);
\draw (3) to (5.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
is immediate from the associativity commutativity of the GHZ-structure, and we also have:
\[
\raisebox{-8mm}{\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=small white dot] (0) at (-0.5, 1.5) {$\frac{n}{m}$};
\node [style=small white dot] (1) at (0.5, 1.5) {$\frac{n'}{m'}$};
\node [style=dot] (3) at (0, 0.5) {};
\node [style=none] (4) at (0, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (3.center);
\draw (1.center) to (3.center);
\draw (3.center) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\ \stackrel{\delta_3}{=}\
\raisebox{-8mm}{\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=small white dot] (0) at (-1.5, 2) {$\frac{n}{m}$};
\node [style=small white dot] (1) at (1.5, 2) {$\frac{n'}{m'}$};
\node [style=small white dot] (2) at (-0.5, 2) {$\frac{m'}{m'}$};
\node [style=small white dot] (3) at (0.5, 2) {$\frac{m}{m}$};
\node [style=white dot] (4) at (-1, 1) {};
\node [style=white dot] (5) at (1, 1) {};
\node [style=dot] (6) at (0, 0.5) {};
\node [style=none] (7) at (0, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (4.center);
\draw (2.center) to (4.center);
\draw (1.center) to (5.center);
\draw (3.center) to (5.center);
\draw (4.center) to (6.center);
\draw (5.center) to (6.center);
\draw (6.center) to (7);
\end{pgfonlayer}
\end{tikzpicture}}
\ = \
\raisebox{-8mm}{\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=small white dot] (0) at (-1.5, 2) {\scriptsize$n m'$};
\node [style=small white dot] (1) at (1.5, 2) {\scriptsize$m n'$};
\node [style=small white dot] (2) at (-0.5, 2) {$\frac{1}{m m'}$};
\node [style=small white dot] (3) at (0.5, 2) {$\frac{1}{m m'}$};
\node [style=white dot] (4) at (-1, 1) {};
\node [style=white dot] (5) at (1, 1) {};
\node [style=dot] (6) at (0, 0.5) {};
\node [style=none] (7) at (0, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (4.center);
\draw (2.center) to (4.center);
\draw (1.center) to (5.center);
\draw (3.center) to (5.center);
\draw (4.center) to (6.center);
\draw (5.center) to (6.center);
\draw (6.center) to (7);
\end{pgfonlayer}
\end{tikzpicture}}
\ \stackrel{\delta_1}{=}\
\raisebox{-8mm}{\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=small white dot] (0) at (-0.5, 2) {\scriptsize$n m'$};
\node [style=small white dot] (1) at (0.5, 2) {\scriptsize$m n'$};
\node [style=small white dot] (2) at (0.75, 1) {$\frac{1}{m m'}$};
\node [style=dot] (3) at (0, 1.25) {};
\node [style=white dot] (4) at (0, 0.5) {};
\node [style=none] (5) at (0, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (3.center);
\draw (1.center) to (3.center);
\draw (3.center) to (4.center);
\draw (4.center) to (2.center);
\draw (4.center) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\ = \
\raisebox{-8mm}{
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=small white dot] (0) at (0, 1.5) {$\frac{n m' + m n'}{m m'}$};
\node [style=none] (1) at (0, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\end{example}
where we used distributivity and ignored the scalars; this is fine as the
scalars in corollary \ref{cor-dist} are cancellable for all the points
we are considering.
\section{Additive inverses}\label{sec:additive-inverses}
The last thing we need to actually produce a field of fractions is additive inverses.
Suppose we have an involutive operation $\cross$ that is a phase for $\dotmult{small white dot}$:
\[
\begin{tikzpicture}[dotpic]
\node [style=none] (0) at (0,-1) {};
\node [style=white dot] (1) at (0,-0.25) {};
\node [style=none] (2) at (-0.5,0.5) {};
\node [style=none] (3) at (0.5,0.5) {};
\draw (0) to node[pauli z]{$\times$} (1) (1) to (2) (1) to (3);
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\node [style=none] (0) at (0,-1) {};
\node [style=white dot] (1) at (0,-0.25) {};
\node [style=none] (2) at (-0.5,0.5) {};
\node [style=none] (3) at (0.5,0.5) {};
\draw (0) to (1) (1) to node[pauli z]{$\times$} (2) (1) to (3);
\end{tikzpicture}
\]
Suppose further that $\{ \dotunit{small white dot},\ \dotcrossunit{small white dot} \}$ forms a plugging set:
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-6, 0.75) {};
\node [style=white dot] (1) at (-4.5, 0.75) {};
\node [style=white dot] (2) at (-2.5, 0.75) {};
\node [style=white dot] (3) at (-1, 0.75) {};
\node [style=none] (4) at (2, 0.75) {};
\node [style=none] (5) at (3.5, 0.75) {};
\node [style=square box] (6) at (-6, 0) {$f$};
\node [style=none] (7) at (-5.25, 0) {$=$};
\node [style=square box] (8) at (-4.5, 0) {$g$};
\node [style=none] (9) at (-3.5, 0) {$\wedge$};
\node [style=square box] (10) at (-2.5, 0) {$f$};
\node [style=none] (11) at (-1.75, 0) {$=$};
\node [style=square box] (12) at (-1, 0) {$g$};
\node [style=none] (13) at (0.5, 0) {$\Leftrightarrow$};
\node [style=square box] (14) at (2, 0) {$f$};
\node [style=none] (15) at (2.75, 0) {$=$};
\node [style=square box] (16) at (3.5, 0) {$g$};
\node [style=none] (17) at (-6, -0.75) {};
\node [style=none] (18) at (-4.5, -0.75) {};
\node [style=none] (19) at (-2.5, -0.75) {};
\node [style=none] (20) at (-1, -0.75) {};
\node [style=none] (21) at (2, -0.75) {};
\node [style=none] (22) at (3.5, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (14) to (21.center);
\draw (8) to (18.center);
\draw (4.center) to (14);
\draw (0) to (6);
\draw (6) to (17.center);
\draw (12) to (20.center);
\draw (16) to (22.center);
\draw (5.center) to (16);
\draw (2) to node[pauli z]{$\times$} (10);
\draw (1) to (8);
\draw (10) to (19.center);
\draw (3) to node[pauli z]{$\times$} (12);
\end{pgfonlayer}
\end{tikzpicture}
\]
Then this will act as an additive inverse operation.
\begin{lemma}\label{lem:pauli-z-homom}
\[
\begin{tikzpicture}[dotpic]
\node [style=none] (0) at (0,-1) {};
\node [style=dot] (1) at (0,-0.25) {};
\node [style=none] (2) at (-0.5,0.5) {};
\node [style=none] (3) at (0.5,0.5) {};
\draw (0) to node[pauli z]{$\times$} (1) (1) to (2) (1) to (3);
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\node [style=none] (0) at (0,-1) {};
\node [style=dot] (1) at (0,-0.25) {};
\node [style=none] (2) at (-0.5,0.5) {};
\node [style=none] (3) at (0.5,0.5) {};
\draw (0) to (1) (1) to node[pauli z]{$\times$} (2) (1) to node[pauli z]{$\times$} (3);
\end{tikzpicture}
\]
\end{lemma}
\begin{proof}
\[
\begin{tikzpicture}[dotpic]
\node [style=none] (0) at (0.5,1) {};
\node [style=none] (1) at (-0.5,1) {};
\node [style=dot] (2) at (0,0.25) {};
\node [style=none] (3) at (0,-0.5) {};
\draw (0) to node[pauli z]{$\times$} (2) (1) to node[pauli z]{$\times$} (2) (2) to (3);
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-0.5, 1.25) {};
\node [style=none] (1) at (0.5, 1.25) {};
\node [style=white dot] (2) at (-1.25, 1) {};
\node [style=white dot] (3) at (1.25, 1) {};
\node [style=white dot] (4) at (-0.5, 0.5) {};
\node [style=white dot] (5) at (0.5, 0.5) {};
\node [style=dot] (6) at (0, 0) {};
\node [style=none] (7) at (0, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw[bend right] (2) to (4);
\draw (4) to (0.center);
\draw[bend right] (5) to (3);
\draw (4) to node[pauli z]{$\times$} (6);
\draw (5) to node[pauli z]{$\times$} (6);
\draw (6) to (7.center);
\draw (5) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-0.5, 1.25) {};
\node [style=none] (1) at (0.5, 1.25) {};
\node [style=white dot] (2) at (-1.25, 1) {};
\node [style=white dot] (3) at (1.25, 1) {};
\node [style=white dot] (4) at (-0.5, 0.5) {};
\node [style=white dot] (5) at (0.5, 0.5) {};
\node [style=dot] (6) at (0, 0) {};
\node [style=none] (7) at (0, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (6) to (7.center);
\draw (4) to (0.center);
\draw[bend right] (5) to node[pauli z]{$\times$} (3);
\draw (4) to (6);
\draw (5) to (1.center);
\draw[bend right] (2) to node[pauli z]{$\times$} (4);
\draw (5) to (6);
\end{pgfonlayer}
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\node [style=none] (0) at (0.5,1) {};
\node [style=none] (1) at (-0.5,1) {};
\node [style=dot] (2) at (0,0.5) {};
\node [style=white dot] (3) at (0,0) {};
\node [style=white dot] (4) at (0.5,0.5) {};
\node [style=none] (5) at (0,-0.5) {};
\draw (0) to (2) (1) to (2) (2) to (3) (3) to (5);
\draw [bend right] (3) to node[pauli z]{$\times$} (4);
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\node [style=none] (0) at (0.5,1) {};
\node [style=none] (1) at (-0.5,1) {};
\node [style=dot] (2) at (0,0.5) {};
\node [style=white dot] (3) at (0,-0.25) {};
\node [style=white dot] (4) at (0.5,0.25) {};
\node [style=none] (5) at (0,-0.75) {};
\draw (0) to (2) (1) to (2) (2) to node[pauli z]{$\times$} (3) (3) to (5);
\draw [bend right] (3) to (4);
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\node [style=none] (0) at (0.5,1) {};
\node [style=none] (1) at (-0.5,1) {};
\node [style=dot] (2) at (0,0.25) {};
\node [style=none] (3) at (0,-0.5) {};
\draw (0) to (2) (1) to (2) (2) to node[pauli z]{$\times$} (3);
\end{tikzpicture}
\]
\end{proof}
\begin{lemma}\label{lem:pauli-z-zero}
\[
\begin{tikzpicture}[dotpic]
\node [style=dot] (0) at (0,0.25) {};
\node [style=none] (1) at (0,-0.5) {};
\draw (0) to node[pauli z]{$\times$} (1);
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\node [style=dot] (0) at (0,0.25) {};
\node [style=none] (1) at (0,-0.5) {};
\draw (0) to (1);
\end{tikzpicture}
\]
\end{lemma}
\begin{proof}
\[
\begin{tikzpicture}[dotpic]
\node [style=dot] (0) at (0,0.25) {};
\node [style=none] (1) at (0,-0.5) {};
\draw (0) to node[pauli z]{$\times$} (1);
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\node [style=dot] (0) at (-0.5,0.5) {};
\node [style=dot] (1) at (0.5,0.5) {};
\node [style=dot] (2) at (0,0) {};
\node [style=none] (3) at (0,-0.5) {};
\draw[bend right] (0) to node[pauli z]{$\times$} (2) (2) to (1);
\draw (2) to (3);
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\node [style=dot] (0) at (-0.5,0.5) {};
\node [style=dot] (1) at (0.5,0.5) {};
\node [style=dot] (2) at (0,0) {};
\node [style=none] (3) at (0,-0.5) {};
\draw[bend right] (0) to (2) (2) to node[pauli z]{$\times$} (1);
\draw (2) to node[pauli z]{$\times$} (3);
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\node [style=dot] (0) at (0,0.5) {};
\node [style=none] (1) at (0,-0.5) {};
\draw (0) to node[pauli z,pos=0.3]{$\times$} node[pauli z,pos=0.7]{$\times$} (1);
\end{tikzpicture}
\ \ = \ \
\begin{tikzpicture}[dotpic]
\node [style=dot] (0) at (0,0.25) {};
\node [style=none] (1) at (0,-0.5) {};
\draw (0) to (1);
\end{tikzpicture}
\]
\end{proof}
\begin{lemma}\label{lem:phi-minus-phi-inv}
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (2) at (-0.75, 0.5) {$\psi$};
\node [style=white dot] (3) at (0.75, 0.5) {$\psi$};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=none] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right] (2.center) to (4);
\draw [bend right] (4) to node[pauli z]{$\times$} (3);
\draw (4) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (2) at (-0.75, 0.5) {$\psi$};
\node [style=white dot] (3) at (0.75, 0.5) {$\psi$};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=none] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right] (2.center) to (4);
\draw [bend right] (4) to node[pauli z]{$\times$} (3);
\draw (4) to node[pauli z]{$\times$} (5);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\end{lemma}
\begin{proof}
Follows immediately from Lemma \ref{lem:pauli-z-homom} and the fact that $\cross$ is an involution.
\end{proof}
\begin{theorem}\label{thm:add-inv}
The cross is additive inverse:
\[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.75, 0.5) {$\psi$};
\node [style=white dot] (1) at (0.75, 0.5) {$\psi$};
\node [style=dot] (2) at (0, -0.5) {};
\node [style=none] (3) at (0, -1.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw[bend right] (2) to node[pauli z]{$\times$} (1);
\draw[bend right] (0) to (2);
\draw (2) to (3.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 0.25) {$\psi$};
\node [style=white dot] (1) at (0.5, 0.25) {$\psi$};
\node [style=dot] (2) at (0, -0.5) {};
\node [style=dot] (3) at (1.5, -0.5) {};
\node [style=white dot] (4) at (0, -1) {};
\node [style=none] (5) at (1.5, -1.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (3) to (5.center);
\draw[bend right] (2) to node[pauli z]{$\times$} (1);
\draw[bend right] (0) to (2);
\draw (2) to (4);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\end{theorem}
\begin{proof}
Proof by plugging: Recall that
\begin{tikzpicture}[dotpic]
\node [style=dot] (0) at (0,0.25) {};
\node [style=white dot] (1) at (0,-0.25) {};
\draw (0) to (1);
\end{tikzpicture} $= 1_{\mathbf{I}}$.
\begin{enumerate}[a)]
\item \[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (2) at (-0.75, 0.5) {$\psi$};
\node [style=white dot] (3) at (0.75, 0.5) {$\psi$};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=white dot] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right] (2.center) to (4);
\draw [bend right] (4) to node[pauli z]{$\times$} (3);
\draw (4) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 0.25) {$\psi$};
\node [style=white dot] (1) at (0.5, 0.25) {$\psi$};
\node [style=dot] (2) at (0, -0.5) {};
\node [style=dot] (3) at (1.5, -0.5) {};
\node [style=white dot] (4) at (0, -1) {};
\node [style=white dot] (5) at (1.5, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2) to (4);
\draw[bend right] (2) to node[pauli z]{$\times$} (1);
\draw[bend right] (0) to (2);
\draw (3) to (5.center);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\item \[
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (2) at (-0.75, 0.5) {$\psi$};
\node [style=white dot] (3) at (0.75, 0.5) {$\psi$};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=white dot] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right] (2.center) to (4);
\draw [bend right] (4) to node[pauli z]{$\times$} (3);
\draw (4) to node[pauli z]{$\times$} (5);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ \stackrel{Lem.\ref{lem:phi-minus-phi-inv}}{=} \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (2) at (-0.75, 0.5) {$\psi$};
\node [style=white dot] (3) at (0.75, 0.5) {$\psi$};
\node [style=dot] (4) at (0, -0.5) {};
\node [style=white dot] (5) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right] (2.center) to (4);
\draw [bend right] (4) to node[pauli z]{$\times$} (3);
\draw (4) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ = \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 0.25) {$\psi$};
\node [style=white dot] (1) at (0.5, 0.25) {$\psi$};
\node [style=dot] (2) at (0, -0.5) {};
\node [style=dot] (3) at (1.5, -0.5) {};
\node [style=white dot] (4) at (0, -1) {};
\node [style=white dot] (5) at (1.5, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2) to (4);
\draw[bend right] (2) to node[pauli z]{$\times$} (1);
\draw[bend right] (0) to (2);
\draw (3) to (5.center);
\end{pgfonlayer}
\end{tikzpicture}}
\ \ \stackrel{Lem.\ref{lem:pauli-z-zero}}{=} \ \
\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 0.25) {$\psi$};
\node [style=white dot] (1) at (0.5, 0.25) {$\psi$};
\node [style=dot] (2) at (0, -0.5) {};
\node [style=dot] (3) at (1.5, -0.5) {};
\node [style=white dot] (4) at (0, -1) {};
\node [style=white dot] (5) at (1.5, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2) to (4);
\draw[bend right] (2) to node[pauli z]{$\times$} (1);
\draw[bend right] (0) to (2);
\draw (3) to node[pauli z]{$\times$} (5.center);
\end{pgfonlayer}
\end{tikzpicture}}
\]
\end{enumerate}
\end{proof}
In the case of ${\bf FHilb}$, this operation is the Pauli Z gate multiplied by $-1$:
\[
\cross
\ \ = \ \
\left(\begin{array}{cc}
-1 & 0 \\
0 & 1
\end{array}\right)
\]
In this case, we have
\[
\begin{tikzpicture}[dotpic]
\node [style=white dot] (0) at (0,0.5) {};
\node [style=none] (1) at (0,-0.5) {};
\draw (0) -- node[pauli z]{$\times$} (1);
\end{tikzpicture}
\ \ = \ \
-\sqrt{2} \ket{-}
\ \ = \ \
\ket{1} - \ket{0}
\]
We can now naturally define:
\[
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (1) at (0, -1) {};
\node [style=white dot] (2) at (0, 1) {$-\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2.center) to (1);
\end{pgfonlayer}
\end{tikzpicture}
\ \ := \ \
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (1) at (0, -1) {};
\node [style=white dot] (2) at (0, 1) {$\psi$};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2.center) to node[pauli z]{$\times$} (1);
\end{pgfonlayer}
\end{tikzpicture}
\]
and, in particular,
\[
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0.5, 1) {$-\frac{n}{m}$};
\node [style=none] (1) at (0.5, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0.center) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}
\ \ \, \raisebox{6.5mm}{:=}\,
\begin{tikzpicture}
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0.5, 1) {$\frac{n}{m}$};
\node [style=none] (1) at (0.5, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to node[pauli z]{$\times$} (1.center);
\end{pgfonlayer}
\end{tikzpicture}
\]
Thus, we have reconstructed all of the axioms for the field of rational numbers. All of the expected identities involving $\cross$ then follow from the field axioms.
\section{!-boxes and automation}\label{sec:automation}
Graph rewriting is a computation process in which graphs are transformed by various \emph{rewrite
rules}. A rewrite rule can be thought of as a `directed graphical equation'. For example, the
``specialness'' equation from section \ref{sec:ghz-w} could be expressed as a graph rewrite rule:
\[
L: \begin{tikzpicture}[dotpic,yshift=5mm]
\node [dot] (a) at (0,0) {};
\node [dot] (b) at (0,-1) {};
\draw [bend left] (a) to (b);
\draw [bend right] (a) to (b);
\draw (0,0.5) to (a) (b) to (0,-1.5);
\end{tikzpicture} \Rightarrow \
R: \begin{tikzpicture}[dotpic]
\draw (0,1) -- (0,-1);
\end{tikzpicture}
\]
A graph rewrite rule $L \Rightarrow R$ can be applied to a graph $G$ by identifying a
\emph{matching}, that is, a monomorphism $m : L \rightarrow G$. The image of $L$ under $m$ is then
removed and replaced by $R$. This process is called \emph{double pushout (DPO) graph rewriting}. A
detailed description of how DPO graph rewriting can be performed on the graphs described in this
paper is available in \cite{DK}.
It is also useful to talk not only about `concrete' graph rewrite rules, but also \emph{pattern
graph} rewrite rules, which can be used to express an infinite set of rewrite rules. We form
pattern graphs using \emph{!-boxes} (or `bang-boxes'). These boxes identify portions of the graph
that can be replicated any number of times. More precisely, the set of concrete graphs represented
by a pattern graph is the set of all graphs (containing no !-boxes) that can be obtained by
performing any sequence of these four operations:
\begin{itemize}
\item \texttt{COPY}: copy a !-box and its incident edges
\item \texttt{MERGE}: merge two !-boxes
\item \texttt{DROP}: remove a !-box, leaving its contents
\item \texttt{KILL}: remove a !-box and its contents
\end{itemize}
For example, the following pattern graph represents the encoding of a natural number given in section \ref{sec:natural-numbers}:
\[\left\llbracket\ \
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=box vertex] (0) at (0, 0.75) {};
\node [style=white dot] (1) at (0, 0.75) {};
\node [style=dot] (2) at (0, 0) {};
\node [style=none] (3) at (0, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2) to (3.center);
\draw (1) to (2);
\end{pgfonlayer}
\end{tikzpicture}\ \ \right\rrbracket =
\left\{\
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 0) {};
\node [style=none] (1) at (0, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (1.center);
\end{pgfonlayer}
\end{tikzpicture}\ ,\ \
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (0, 0.75) {};
\node [style=dot] (1) at (0, 0) {};
\node [style=none] (2) at (0, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (1) to (2.center);
\draw (0) to (1);
\end{pgfonlayer}
\end{tikzpicture}\ ,\ \
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.5, 0.75) {};
\node [style=white dot] (1) at (0.5, 0.75) {};
\node [style=dot] (2) at (0, 0) {};
\node [style=none] (3) at (0, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (1) to (2);
\draw (2) to (3.center);
\draw (0) to (2);
\end{pgfonlayer}
\end{tikzpicture},\ \
\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=white dot] (0) at (-0.75, 0.75) {};
\node [style=white dot] (1) at (0, 0.75) {};
\node [style=white dot] (2) at (0.75, 0.75) {};
\node [style=dot] (3) at (0, 0) {};
\node [style=none] (4) at (0, -0.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (2) to (3);
\draw (0) to (3);
\draw (3) to (4.center);
\draw (1) to (3);
\end{pgfonlayer}
\end{tikzpicture},\ \
\ldots\
\right\} :=
\left\{
\point{0},\ \point{1},\ \point{2},\ \point{3},\ \ldots
\right\}\]
Note how not only the vertices are duplicated, but also all of the edges connected to those vertices.
A pattern graph rewrite rule is a pair of pattern graphs with the same inputs and outputs.
Furthermore, there is a bijection between the !-boxes occurring on the LHS and the RHS. When one of
the four operations is performed to a !-box on the LHS, the same is performed to its corresponding
!-box on the RHS.
We can rewrite (the natural numbers versions of) equations $\delta_1$, $\delta_2$, and $\delta_3$ as pattern graph rewrite rules.
\[
L:\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-0.5, 1.5) {};
\node [style=none] (1) at (0.5, 1.5) {};
\node [style=dot] (2) at (0, 0.5) {};
\node [style=white dot] (3) at (0, -0.5) {};
\node [style=none] (4) at (0, -1.5) {};
\node [style=white dot] (5) at (1, 0.5) {};
\node [style=box vertex,inner sep=1.5mm] (6) at (1, 0.5) {};
\node [style=dot] (7) at (1, 0) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (0) to (2);
\draw [bend right=30] (2) to (1);
\draw (2) to (3);
\draw (3) to (4);
\draw [bend right=30] (3) to (7);
\draw (7) to (5);
\end{pgfonlayer}
\end{tikzpicture}}
\Rightarrow\ \
R: \raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (-1.25, 1.5) {};
\node [style=white dot] (1) at (-0.25, 1.5) {};
\node [style=white dot] (2) at (0.25, 1.5) {};
\node [style=none] (3) at (1.25, 1.5) {};
\node [style=dot] (4) at (-0.25, 1) {};
\node [style=dot] (5) at (0.25, 1) {};
\node [style=white dot] (6) at (-0.75, 0) {};
\node [style=white dot] (7) at (0.75, 0) {};
\node [style=dot] (8) at (0, -1) {};
\node [style=none] (9) at (0, -1.5) {};
\node [style=box vertex,minimum height=2mm,minimum width=6mm] (10) at (0, 1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw [bend right=30] (0) to (6.center);
\draw [bend right=30] (6.center) to (4.center);
\draw (4.center) to (1);
\draw [bend right=30] (6.center) to (8.center);
\draw [bend right=30] (8.center) to (7.center);
\draw [bend right=30] (5.center) to (7.center);
\draw [bend right=30] (7.center) to (3);
\draw (5.center) to (2);
\draw (8.center) to (9);
\end{pgfonlayer}
\end{tikzpicture}}
\qquad \qquad
L:\raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\node [style=white dot] (2) at (0, -0.5) {};
\node [style=white dot] (3) at (1, 1) {};
\node [style=dot] (4) at (1, 0.5) {};
\node [style=box vertex,inner sep=1.5mm] (5) at (1, 1) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0) to (2.center);
\draw (2.center) to (1);
\draw [bend right=30] (2.center) to (4.center);
\draw (4.center) to (3.center);
\end{pgfonlayer}
\end{tikzpicture}}
\Rightarrow\ \
R: \raisebox{-8mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=dot] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0)--(1);
\end{pgfonlayer}
\end{tikzpicture}}
\qquad \qquad
L:\begin{tikzpicture}[dotpic]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=white dot] (1) at (1.25, 1.5) {};
\node [style=box vertex, minimum width=9 mm, minimum height=2 mm] (2) at (1.75, 1.5) {};
\node [style=white dot] (3) at (2.25, 1.5) {};
\node [style=white dot] (4) at (0.5, 1) {};
\node [style=white dot] (5) at (1.5, 1) {};
\node [style=dot] (6) at (0.75, 0.5) {};
\node [style=dot] (7) at (1.75, 0.5) {};
\node [style=white dot] (8) at (0, 0) {};
\node [style=white dot] (9) at (0, -1) {};
\node [style=none] (10) at (0, -1.75) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (9) to (10.center);
\draw[bend right=15] (4) to (6);
\draw (8) to (9);
\draw (0.center) to (8);
\draw[out=30, in=258] (6) to (1);
\draw[bend right=15] (5) to (7);
\draw[bend right] (9) to node[tick]{-} (7);
\draw[bend right] (8) to (6);
\draw[bend right=15] (7) to (3);
\end{pgfonlayer}
\end{tikzpicture}
\Rightarrow\ \
R: \raisebox{-9mm}{\begin{tikzpicture}[scale=0.6]
\begin{pgfonlayer}{nodelayer}
\node [style=none] (0) at (0, 1.5) {};
\node [style=none] (1) at (0, -1.5) {};
\end{pgfonlayer}
\begin{pgfonlayer}{edgelayer}
\draw (0)--(1);
\end{pgfonlayer}
\end{tikzpicture}}
\]
These equations only apply to encodings of the natural numbers, not for arbitrary inputs $\psi$.
However, we showed in sections \ref{sec:natural-numbers}, \ref{sec:rationals}, and
\ref{sec:additive-inverses} that even those weaker equations suffice to recover the usual
identities for fraction arithmetic. Also note that the extra white vertices in $\delta_3'$
eliminate the case of $\frac{0}{0} \neq 1$.
So why bother expressing graphical identities as graph rewrite rules? Graph rewriting can be
automated! {\tt quantomatic}~\cite{quanto} is a automatic graph rewriting tool developed by two of
the authors. It is specifically designed to work with the kinds of diagram described in this paper
and to perform pattern graph rewriting. It remains to be seen what new insights can be obtained by
adding the new graphical identities derived in this paper to {\tt quantomatic}.
\section{Closing Remarks}
In previous work two of the authors showed that the main difference between the GHZ state and the W state, or more precisely, the induced GHZ structure and W structure, boils down to the value of the loop map of these CFAs:
\[
\frac{\mbox{GHZ}}{\mbox{W}}\ \ =\ \ \frac{\begin{tikzpicture}[dotpic,yshift=5mm]
\node [dot] (a) at (0,0) {};
\node [dot] (b) at (0,-1) {};
\draw [bend left] (a) to (b);
\draw [bend right] (a) to (b);
\draw (0,0.5) to (a) (b) to (0,-1.5);
\end{tikzpicture} = \
\begin{tikzpicture}[dotpic]
\draw (0,1) -- (0,-1);
\end{tikzpicture}
}{
\circl\ \begin{tikzpicture}[dotpic]
\node [dot] (a) at (0,0.5) {};
\node [dot] (b) at (0,-0.5) {};
\draw [bend left] (a) to (b);
\draw [bend right] (a) to (b);
\draw (0,1) to (a) (b) to (0,-1);
\end{tikzpicture}\ \ =
\begin{tikzpicture}[dotpic]
\node [dot] (a) at (0,0.7) {};
\node [dot] (b) at (0,-0.7) {};
\draw (0,1.2) to (a) (b) to (0,-1.2);
\draw (a) to [downloop] ();
\draw (b) to [uploop] ();
\end{tikzpicture}
}
\]
In this paper, by focussing on the interaction of these two structures, we were able to establish a connection with the operations of basic arithmetic:
\[
\frac{\mbox{W}}{\mbox{GHZ}}=\frac{+}{\times}
\]
More specifically, the diagrammatic language of these structures was sufficient to encode the positive rational numbers (and, with a minor extension, the whole field of rational numbers).
In the process of highlighting this encoding, we identified a surprising fact. The distributive law
governing the interaction of addition and multiplication in arithmetic also captures the
interaction of the GHZ-structure and W-structure. Future work includes exploiting this interaction
in the study of multipartite quantum entanglement, which brings us back to the initial motivation
for crafting a compositional framework to reason about multipartite states.
|
\section{Introduction}
\label{s:Introduction}
Heavy quarkonia provide a convenient testing ground for analytical
approaches to hadron properties in QCD, both in the vacuum and in the
presence of a (deconfined) medium~\cite{Brambilla:2010cs}.
In the latter case, the original idea of a direct link between charmonium
suppression and deconfinement~\cite{Matsui:1986dk} has been refined into
more involved predictions for the behaviors of the various states with
rising temperature (for recent reviews see Refs.~\cite{Rapp:2008tf,%
Kluberg:2009wc,Rapp:2009my}).
Many predictions are formulated in a static picture, most noticeably in
terms of threshold temperatures, above which a given state is entirely
``melted'', while it remains intact below.
More dynamical approaches to the dissociation and formation or
recombination of quarkonia in a medium have been considered in various
classical kinetic frameworks:
\`a la Boltzmann~\cite{LevinPlotnik:1995un,Polleri:2003kn,Yan:2006ve}, in
Langevin or Fokker--Planck descriptions~\cite{Patra:2001th,Young:2008he}
or through rate equations~\cite{Grandchamp:2003uw,Zhao:2011cv}.
A common feature of these studies is their focus on charmonia.
This is quite natural, since this corresponds to most of the existing
experimental results.
Now, the general expectation is that $c\bar c$ pairs can only exist in a
quark-gluon plasma (QGP) as either a $J/\psi$ or an unbound system.
Accordingly, the above mentioned studies consider these two possibilities
only.
On the other hand, it is thought that several bottomonia are still bound in
a QGP just above the deconfinement temperature.
This opens a richer spectrum of possible behaviors for $b\bar b$ pairs,
especially if transitions can be induced between different bound states.
In this Letter, we wish to take this latter possibility seriously, and
to investigate some of the consequences of this ansatz, restricting
ourselves to inner degrees of freedom.
To this effect, we shall hereafter perform a Gedankenexperiment:
we put at time $t=0$ a bound $Q\bar Q$ pair in a static, infinitely large
QGP at temperature $T$.
Then we let the system evolve, and follow the populations of the various
quarkonium states in time.
In Section~\ref{s:Model-method} we introduce our model for the heavy
quarkonia, the QGP and their interaction, as well as for the resulting
dynamical equations.
Section~\ref{s:Results} contains our results for the evolving populations
of quarkonium states.
Further results, in particular involving the external degrees of freedom of
quarkonia, will be presented in a longer, more technical
publication~\cite{BG_inprep}.
Eventually, we discuss both our model with its underlying assumptions and
our results in Section~\ref{s:Discussion}.
\section{Model and method}
\label{s:Model-method}
In this Section, we describe our model for the heavy quarkonium states and
the medium in which they are immersed, as well as for the coupling between
both.
Then we briefly introduce the equations that govern the dynamics of the
$Q\bar Q$-state populations.
\subsection{Quarkonium in a quark-gluon plasma as a \{system + reservoir\}
dissipative system}
\label{s:Model}
Our purpose in the present study is to investigate the evolution of heavy
quarkonium states under the influence of the thermal degrees of freedom of
the plasma, and especially gluons.
Thus, we do not consider (light) quarks, which would interact with the
heavy quark or antiquark through non-thermal gluons: our medium is a gluon
plasma.
We model this plasma as a static bath of harmonic excitations, whose
frequencies span a continuum $\{\omega_\lambda\}$.
The bath is treated as a thermal reservoir, whose thermodynamic properties
are not affected by the transitions between the various states of the
embedded quarkonia.
The modeling of heavy quarkonia is far more delicate than that of the
plasma.
In this Letter, we wish to identify generic behaviors that follow from
allowing quark-antiquark pairs to be in different (quasi-)bound states in a
deconfined medium, between which medium-induced transitions are permitted.
In that view, we deliberately adopt an admittedly simplistic quarkonium
model, which relies on a minimal number of parameters, deferring more
realistic modeling to further studies.
In particular, we make a number of assumptions, which we shall further
discuss in Section~\ref{s:Discussion}.
Our first model assumption is the existence of a non-relativistic in-medium
quark-antiquark potential~\cite{Digal:2001iu,Wong:2004zr,Arleo:2004ge,%
Alberico:2006vw,Cabrera:2006wh,Mocsy:2007yj} that admits bound states,
thereby neglecting any imaginary part in the potential~\cite{Laine:2006ns}.
To make the model simple, we consider an attractive Coulomb potential
$V(r)=-\alpha/r$, with its well-known spectroscopy.\footnote{We
take $\alpha=0.4$ for $c\bar c$ pairs and a smaller $\alpha\simeq 0.25$
for $b\bar b$ pairs, to account for the running of the coupling constant
and the smaller size of the ground state.}
As we want to investigate the possible influence of transitions between
levels, we depart from the exact Coulombic spectroscopy for the states that
are bound in the vacuum: $J/\psi$, $\chi_c$ and $\psi'$ on the one hand,
$\Upsilon$, $\chi_b$, $\Upsilon'$, $\chi'_b$ and $\Upsilon''$ on the other
one~\cite{PDG10}.\footnote{For given $S$ and $L$ quantum numbers, we
consider for simplicity a single (2$L$+1)-fold degenerated state.}
Those states are modeled with Coulomb wave functions, yet with the same
energy with respect to the ground 1$S$-state as measured in the vacuum.
This is clearly an approximation, which allows us to lift the degeneracy
between e.g. 2$S$ and 1$P$ states, and thereby permit single-gluon
transitions between them.
Obviously, a single gluon cannot induce a transition between color-neutral
$Q\bar Q$ states.
We thus further assume that our spectrum of color-singlet states is
accompanied by a parallel spectrum of color-octet states, which for the
sake of simplicity we shall denote similarly to their singlet
counterparts.
Eventually, we have to model the unbound $Q\bar Q$ pairs, which should
normally constitute a continuum.
Consistency with our model for in-medium bound states through a real
potential implies that the states in that continuum have higher energies
than the bound states, i.e.\ the latter are {\em always\/} energetically
favored, which is far from granted.
This is even more a problem with the Coulomb potential, which admits
arbitrarily large bound states with high principal quantum number --- which
in the static Debye-screening picture would appear as unbound.
To get rid of those states, we fix the {\em dissociation\/} threshold by
considering bound states of the Coulomb potential as describing unbound
pairs: 2$S$ and 1$P$ states (resp.\ 3$S$ and 2$P$ states) and the higher
excited states for $c\bar c$ (resp.\ $b\bar b$) pairs.
A minimal approach to mimic the unbound character of such states consists
in forbidding transitions from them back to the bound ones.
In our computations, we have fully discarded the scattering solutions of
the Coulomb potential, and considered two or three levels of
``dissociated'' states, with at least two states per level, to estimate the
error on our result.
To leave room for the possible ``recombination'' of heavy quark and
antiquark into a quarkonium state~\cite{Thews:2000rj}, we also slightly
modified the model by allowing transitions from the lowest dissociated
states to bound ones.
The plots we present in Section~\ref{s:Results} are for results from this
variant of our model.
Further plots will be shown elsewhere~\cite{BG_inprep}.
Eventually, we need to specify the coupling between a $Q\bar Q$ pair and
the plasma.
We assume {\em dipolar coupling\/}, that is, the gluons only interact
through their chromoelectric field.
Incidentally, we need also assume that the Bohr frequencies between
$Q\bar Q$ states are included in the continuum of bath frequencies, so that
transitions between states can be induced.
\subsection{Evolution of the $Q\bar Q$ states in the plasma}
Now that we have specified the ingredients of our model, we can turn to the
time evolution.
Further details will be given in a longer publication~\cite{BG_inprep},
here we shall merely outline the calculation.
We expect that the state of the quark-antiquark system at a given time
should be a statistical superposition of (vacuum) eigenstates.
Then a natural approach is to use the master-equation formalism~\cite[%
Chapter 4]{CDG2}.
Within the decorrelation approximation --- i.e., technically, assuming that
the density matrix of the whole system factorizes into the product of the
density matrix $\rho^{Q\bar Q}$ of the $Q\bar Q$ pair and that of the
plasma at every time ---, which amounts to considering an expansion up to
second order in the coupling potential between the quark-antiquark pair and
the plasma, the {\em populations\/} (diagonal elements) of $\rho^{Q\bar Q}$
are governed by the coupled Einstein equations
\begin{equation}
\label{eq:Einstein}
\frac{{\rm d}\rho^{Q\bar Q}_{ii}}{{\rm d} t}(t) =
-\sum_{k\neq i}\Gamma_{i\to k}\rho^{Q\bar Q}_{ii}(t) +
\sum_{k\neq i}\Gamma_{k\to i}\rho^{Q\bar Q}_{kk}(t),
\end{equation}
where the transition rates $\Gamma_{i\to k}$ between $Q\bar Q$ levels
follow from Fermi's golden rule (except for those we set to zero, to mimic
the continuum, as explained above).
These rates involve a sum over the states of the QGP, weighted by their
respective probabilities.
This introduces a dependence of all $\Gamma_{i\to k}$ on the plasma
temperature $T$.
In the following Section, we present solutions to these evolution equations
for the $c\bar c$ and the richer $b\bar b$ systems.
\section{Results}
\label{s:Results}
The Gedankenexperiment we discussed in Section~\ref{s:Introduction} amounts
to picking out an initial condition at $t=0$ for Eqs.~\eqref{eq:Einstein},
for instance $\rho^{Q\bar Q}_{ii}(t\!=\!0)=1$ for the ground 1$S$ state and
$0$ for the excited levels, and to solve the coupled system.
\begin{figure*}[t!]
\includegraphics*[width=0.495\linewidth]{charmonia_vs_t}
\includegraphics*[width=0.495\linewidth]{bottomonia_vs_t}
\caption{\label{fig:t-dependences}Time dependence of the populations of
bound, recombining and unbound states.
Left: $c\bar c$ pairs at $T=2T_c$;
right: $b\bar b$ pairs at $T=5T_c$.
Here and in Fig.~\ref{fig:T-dependences}, $T_c=170$~MeV.}
\end{figure*}
A first observation, independent from the initial condition, follows
directly from the structure of the coupled evolution equations:
the latter are not diagonal.
As a consequence, the populations corresponding to the vacuum eigenstates
of the potential do not constitute an eigenstate of the ``evolution
operator'' for the vector of populations that can be read off
Eqs.~\eqref{eq:Einstein};
rather, they are linear combinations of the latter.
That is, the different quarkonium states do not evolve independently from
each other, but they are coupled together by the medium.
For instance, even if $\Upsilon'$ is constantly either dissociated or
decaying into $\chi_b$, at the same time it is recreated, with different
rates, through the excitation of $\chi_b$ or --- when we allow for that
possibility --- through recombination of an unbounded $b\bar b$ pair.
Thus we cannot have the total disappearance of a state in the plasma at
finite times, as implied by static approaches to quarkonium suppression.
Let $\Gamma$ denotes the smaller (in absolute value) of the eigenvalues of
the evolution operator for the populations.
The medium-induced mixing of states is such that after some transient
behavior, which depends on the specific choice of initial condition, the
populations $\rho^{Q\bar Q}_{ii}$ of all bound states (and of the unbound
states that are allowed to recombine) evolve with the same characteristic
time scale $\Gamma^{-1}$.
We illustrate this in Fig.~\ref{fig:t-dependences}, which shows the time
evolution of the populations of $c\bar c$ (left) and $b\bar b$ (right)
states at respectively $2T_c$ and $5T_c$.
The characteristic evolution times for these systems at the chosen
temperatures within our model are respectively 3.4 and 3.8~fm/$c$.
Quite obviously, for a given system $\Gamma^{-1}$ decreases with rising
temperature.
\begin{figure}[b]
\includegraphics*[width=\linewidth]{bottomonia-ratios_vs_T}
\caption{\label{fig:T-dependences}Temperature dependence of the ratios
of the populations of excited $b\bar b$ states to that of the ground
state ($\Upsilon$) in the stationary regime (see text).
Curves: corresponding ratios in a thermally equilibrated system.}
\end{figure}
Past the transient regime, the populations of the various bound and
recombining states are ``equilibrated'' with each other, in the sense that
the population ratios remain stable.
(There is however no strict equilibrium, since $Q\bar Q$ pairs are
consistently lost to non-recombining states).
These ``stationary'' population ratios depend on the plasma temperature, as
shown for $b\bar b$ pairs in Fig.~\ref{fig:T-dependences}.
The stationary ratios differ significantly from those for thermally
equilibrated levels, which is easily understandable.
The thermal ratios are those which make Eqs.~\eqref{eq:Einstein} stationary
with {\em all\/} transition rates fulfilling the detailed balance condition
\begin{equation}
\label{eq:detailed-balance}
\Gamma_{i\to k}\,{\rm e}^{-E_i/T} = \Gamma_{k\to i}\,{\rm e}^{-E_k/T}\quad
\forall\, i,k.
\end{equation}
Here, the condition is obeyed only by bound ($\Upsilon$, $\chi_b$,
$\Upsilon'$) and recombining ($\chi'_b$, $\Upsilon''$) states, but
not by the unbound ones, from which there are no transition.
It is thus normal that the resulting stationary ratios diverge from the
thermal ratios.
We have checked that when we do not suppress back transitions from the
unbound states, but set them according to condition~\eqref{%
eq:detailed-balance}, then the stationary abundance ratios are the same
as in thermal equilibrium.
Given the crudeness of our model, which we wish to further discuss in the
next Section, we have not attempted at this stage to make predictions for
observables in real nucleus-nucleus collisions, which would anyway
necessitate some extra modeling of the plasma kinetics as well as
accounting for the hadronic phase.
\section{Summary and discussion}
\label{s:Discussion}
We have examined the dynamics of the populations of heavy quarkonium states
in a static QGP within the quantum-mechanical master-equation formalism,
assuming that some elements of the quarkonium spectroscopy survive in the
plasma, in particular that different bound states exist, between which
transitions can be induced by the medium.
Under this assumption, we find that the bound states are mixed together by
the medium, so that they all evolve with the same characteristic time
scale.
This differs from the usual picture of sequential melting, inasmuch as no
bound state can totally disappear while others would survive.
In addition, we find that after some transient regime, the population
ratios remain stationary, at values determined by the plasma temperature,
yet different from the ratios for quarkonium states in thermal equilibrium.
While the fact that vacuum eigenstates are no longer eigenstates in the
QGP is model-independent, the actual predictions for the abundance ratios
depend on the model parameters, and could be used to constrain the latter
from experimental results.
Let us now discuss our model.
Its key ingredient in our eyes is the ansatz of medium-induced transitions
--- with a large enough rate compared to the inverse of the plasma lifetime
--- between bound quarkonium states.
This is the element which couples the evolutions of the various bound
states together, irrespective of the details of their modeling or of the
mechanism responsible for the transitions.
The latter is important inasmuch as it selects which states are coupled
together.
Here we wanted to consider, in analogy to quantum optics, transitions
induced by the gauge bosons.
This choice forced us to invoke some hazy ``spectrum of color-octet
states'', paralleling that of color-singlets --- which is far from being
granted, given the repulsive nature of the color-octet channel.
Instead, we could have conjured non-perturbative effects, like some kind of
soft color interactions~\cite{Edin:1995gi}, to instantly turn color octets
into singlets in the medium.
Such explanations seem to us to be as disputable as our choice in the
present context.
Alternatively, one could think of non-color-exchanging processes, as
quasi-elastic collisions with off-shell quarks or gluons or scattering of
photons, provided the latter happen with a sufficient rate.\footnote{See
Ref.~\cite{Marasinghe:2011bt}, which appeared while this Letter was being
finalized, for a study of the photoionization of the $J/\psi$ by large
magnetic fields in a QGP.}
If one accepts the possibility of medium-induced transitions between bound
states, then the models for medium, quarkonia and their interaction that we
have used are purposely the simplest ones one can think of, yet still
realistic enough to illustrate some plausible phenomena.
The orders of magnitude we obtain for the typical time scales for the
evolution of quarkonium populations are reasonable, which justifies our
choice {\em a posteriori\/}.
The model allows us to compute transition rates between bound states (which
is the reason why we have kept the Coulomb wave functions although with
``wrong'' energies), while these are not known in more realistic models,
since such transitions have not been investigated before.
For the dissociation widths of the $1S$ states, one could use the known
results at leading~\cite{BhanotPeskin} or next-to-leading order~\cite{%
Park:2007zza}: this would represent an improvement, yet only a partial
one.
Note that the dissociation process that we consider, as everyone does, is a
{\em classical\/} process, in which energy is transferred to the $Q\bar Q$
pair.
This is inherent to the description by a real potential similar to the
vacuum one.
It might turn out that a better description of the transition from bound
$Q\bar Q$ state to unbound quark and antiquark in a QGP should
involve some tunneling through a barrier, as studied in hadronic matter in
Ref.~\cite{Kharzeev:1995ju}.
The master-equation approach we have used relies on a few assumptions,
which we shall detail elsewhere~\cite{BG_inprep}.
In short, these amount to assuming --- as is also done in Boltzmann,
Langevin or Fokker--Planck formalisms --- that the typical time
scale of
plasma correlations is small against the characteristic time scale of the
$Q\bar Q$-plasma interaction, i.e.\ the formalism implicitly rests on a
``weak-coupling'' assumption.
We are investigating alternate approaches that do not make use of this
hypothesis~\cite{DBG_inprep}, yet this seems only feasible at the cost of
some alternative approximations.
Eventually, we have assumed dipolar coupling between a $Q\bar Q$ pair and
the plasma, discarding chromomagnetic or quadrupolar and higher order
chromoelectric couplings.
This is a large-wavelength approximation --- which might be disputable for
gluons that should resolve the structure of the bound states --- that can
be released when using a more realistic model of the quarkonia in the
plasma.
\section{Acknowledgments}
We gratefully thank J.-P.\ Blaizot, J.-Y.\ Ollitrault and H.\ Satz for
enlightening discussions.
C.\ G. acknowledges support from the Deutsche Forschungsgemeinschaft under
grant GRK 881.
|
\section{Introduction}
Quarter of a century after the dramatical discovery of high
temperature superconductivity (HTS) the problem of its mechanism
remains one of the longest-standing puzzles in the history of science.
The cornucopia of ideas is immense -- almost every quantum process was
tested whether it is the long sought mechanism of HTS. In parallel, the
intensive experimental research made cuprates into the most
investigated materials. First-principles electronic structure
calculations play an important role in understanding the physics of
these materials. The Fermi surface of these highly anisotropic
crystals is almost cylindrical with rounded-square cross-section.
Fifteen years after the beginning of the cuprates era the eye of the
professionalist has uncovered a subtle correlation~\cite{Pavarini:01}
between the shape of the Fermi contour and the critical temperature
\tc\ for optimally hole doped cuprates. Pavarini \textit{et al.}~\cite{Pavarini:01}
considered many hole doped cuprates:
Ca$_2$CuO$_2$Cl$_2$,
La$_2$CuO$_4$,
Bi$_2$Sr$_2$Cu$_2$O$_6$,
Tl$_2$Ba$_2$CuO$_6$,
Pb$_2$Sr$_2$Cu$_2$O$_6$,
TlBaLaCuO$_5$,
HgBa$_2$CuO$_4$,
LaBa$_2$Cu$_3$O$_7$,
La$_2$CaCu$_2$O$_6$,
Pb$_2$Sr$_2$YCu$_3$O$_8$,
YBa$_2$Cu$_3$O$_7$,
Tl$_2$Ba$_2$CaCu$_2$O$_8$,
Tl$_2$Ba$_2$Ca$_2$Cu$_3$O$_{10}$,
HgBa$_2$Ca$_2$Cu$_3$O$_8$,
HgBa$_2$CaCu$_2$O$_6$;
the list can be extended but we believe that a universal correlation was discovered.
In an acceptable approximation the electronic band structure can be
described by the four-band Linear Combination of Atomic Orbitals (LCAO) model with on-site energies
$\epsilon_\mathrm{s}, \epsilon_\mathrm{p}, \epsilon_\mathrm{d}$, and hopping parameters $t_\mathrm{sp},
t_\mathrm{pd}, t_\mathrm{pp}.$ In this approximation the Hilbert space spans over
the Cu~3d$_{x^2-y^2}$, Cu~4s, O~2$p_x$ and O~2$p_y$ valence
orbitals. The range parameter $r\equiv1/2(1+s)$ which determines the shape
of the Fermi contour is determined by the LCAO parameters and the
Fermi energy $E_\mathrm{F}$
\be
s(E_\mathrm{F})=(\epsilon_\mathrm{s}-E_\mathrm{F})(E_\mathrm{F}-\epsilon_\mathrm{p})/(2t_\mathrm{sp})^2.
\ee
The correlation between \tc\ and $r$ from the work by Pavarini
\textit{et al.}~\cite{Pavarini:01} is reproduced in
\Fref{fig_crucial}. How such a correlation can possibly exist?
Superconductivity is, certainly, created by some interaction, whilst
the electronic band structure describes the properties of independent
electrons. Is it then possible to uncover an unknown mechanism of
interaction investigating only the properties of noninteracting
particles? The \tc--$r$ correlation covers the whole temperature range
of HTS, but now ten years it remains unexplained. We suppose that this
correlation is as important for HTS, as was the isotope effect for the
conventional phonon superconductors half a century ago. The aim of the
present work is to interpret the correlation reported by
Pavarini~\textit{et al.}~\cite{Pavarini:01} in the framework of some of
the models of HTS.
\begin{figure}[t]
\centering
\includegraphics[height=6cm]{./fig1.eps}
\caption{Critical temperature of hole doped cuprates versus band
structure parameter $r$ by Pavarini~\textit{et al.}~\cite{Pavarini:01}.
The theoretical curve is calculated according to \Eqref{BCS} (see text for details).}
\label{fig_crucial}
\end{figure}
\section{{\slshape{T}}$_{\bm{\textsf{c}}}$--{\slshape{r}} correlation within the s--d theory}
Let us introduce a small modification of the $s$ parameter,
$\tilde{s}=\displaystyle \frac{t_\mathrm{sp}}{t_\mathrm{pd}}s.$ Its reciprocal value
\be
\frac{1}{\tilde{s}(\epsilon)} =
\frac{4t_\mathrm{pd}t_\mathrm{sp}}{(\epsilon_\mathrm{s}-\epsilon)(\epsilon-\epsilon_\mathrm{p})},
\label{eq:stilde}
\ee
has an energy denominator typical for the perturbation theory
as applied to the secular equation of the generic four-band model
\begin{equation}
\begin{pmatrix}
\epsilon_\mathrm{d} & 0 & t_\mathrm{pd}s_x & -t_\mathrm{pd}s_y \\
0 & \epsilon_\mathrm{s} & t_\mathrm{sp}s_x & t_\mathrm{sp}s_y \\
t_\mathrm{pd}s_x & t_\mathrm{sp}s_x & \epsilon_\mathrm{p} & t_\mathrm{pp}s_x s_y \\
-t_\mathrm{pd}s_y & t_\mathrm{sp}s_y & -t_\mathrm{pp}s_x s_y & \epsilon_\mathrm{p}
\end{pmatrix}
\begin{pmatrix}
D_\mathbf{p} \\ S_\mathbf{p} \\ X_\mathbf{p} \\ Y_\mathbf{p}
\end{pmatrix}
=\epsilon_\mathbf{p}
\begin{pmatrix}
D_\mathbf{p} \\ S_\mathbf{p} \\ X_\mathbf{p} \\ Y_\mathbf{p}
\end{pmatrix},
\end{equation}
where $\epsilon_\mathbf{p}$ is the electron energy for the conducting
d-band, $s_x=2\sin(p_x/2),$ $s_y=2\sin(p_y/2),$ and the wave function is normalized
$D^2_\mathbf{p}+S^2_\mathbf{p}+X^2_\mathbf{p}+Y^2_\mathbf{p}=1.$
The approximate solution for small hopping amplitudes
\begin{equation}
\begin{pmatrix}
D_\mathbf{p} \\ S_\mathbf{p} \\ X_\mathbf{p} \\ Y_\mathbf{p}
\end{pmatrix}
\approx
\left(
\begin{array}{c}
1 \\
\left(s_x^2-s_y^2\right) / 4\tilde{s}(\epsilon)\\
\displaystyle\frac{t_\mathrm{pd}}{\epsilon-\epsilon_\mathrm{p}}s_x\\
-\displaystyle
\frac{t_\mathrm{pd}}{\epsilon-\epsilon_\mathrm{p}}s_y
\end{array}\right),
\end{equation}
has a simple interpretation: in the initial approximation we have a
pure Cu~3d state with $D_\mathbf{p}\approx 1,$ in first
approximation we have a linear dependence from the $t_\mathrm{pd}$ amplitude
on O~2p levels $X_\mathbf{p},\; Y_\mathbf{p} \propto t_\mathrm{pd}$, and
finally in second approximation, the amplitude on the Cu~4s orbital
$S_\mathbf{p}\propto t_\mathrm{pd}t_\mathrm{sp}$ is included by the second virtual
transition proportional to $t_\mathrm{sp}$. In short, the Cu~4s amplitude of
the conduction band can be phrased as 3d--to--4s--by--2p.
As it was concluded by Pavarini~\textit{et al.}~\cite{Pavarini:01}
that materials with lower $\epsilon_\mathrm{s}$ tend to be those with higher
observed values of $T_{\mathrm{c}\,\,max}$. In the materials with
higher $T_{\mathrm{c}\,\,max}$ the axial orbital is almost Cu~4s. What is the
simplest interpretation of these observations? There is an emerging
consensus that superconductivity in cuprates is created by some
exchange interaction. The Pavarini \textit{et al}~\cite{Pavarini:01}
correlation gives that higher $T_{\mathrm{c}\,\,\mathrm{max}}$ is determined by
the highest Cu~4s amplitude $S_\mathbf{p}$. Now we have to recall that
the most usual 3d--4s intra-atomic exchange has one of the
largest amplitudes in condensed matter physics.
We also have to point out that transition metal ion-ion exchange integrals $J_{ii},$
which create antiferromagnetism in insulator phase
are smaller than intra atomic exchange integral $J_{sd}.$
The intensive investigations of the Kondo effect have demonstrated
that the anti-ferromagnetic sign of the two electrons s--d
exchange is the rule and the ferromagnetic sign is an exception.
Incorporated in the BCS scheme
for the calculation of the order parameter $\Xi,$
the anti-ferromagnetic sign gives
pairing in singlet channel with momentum dependent gap
\be
\Delta_\mathbf{p}(T)=\Xi(T)\chi_\mathbf{p},
\qquad \chi_\mathbf{p} \equiv S_\mathbf{p}D_\mathbf{p}.
\ee
This gap is included in the fermion excitation energy
\be E_{\mathbf{p}} \equiv (\eta_{\mathbf{p}}^2 +
\Delta^2_{\mathbf{p}})^{1/2},\qquad \eta_{\mathbf{p}} \equiv
\epsilon_\mathbf{p} -E_\mathrm{F},
\ee
and for the temperature dependent order parameter $\Xi(T)$ we have the
standard BCS equation
\be
\label{BCS}
2J_\mathrm{sd} \left<\frac{\chi_{\mathbf{p}}^2}{2E_{\mathbf{p}}}
\tanh\left(\frac{E_{\mathbf{p}}}{2k_\mathrm{B} T}\right) \right> = 1,
\quad
\left<f_\mathbf{p}\right> \equiv \int_{0}^{2\pi}
\int_{0}^{2\pi} \frac{dp_x \, dp_y}{(2\pi)^2}f(\mathbf{p})
\ee
where $\langle\dots\rangle$ denotes momentum--space averaging over the
Brillouin zone. For a pedagogical derivation of the BSC gap equation
in the present notations see the textbook~\cite{Mishonov:11} and references therein.
Slightly below the critical temperature where the order parameter
disappears $\Xi(\tc -0)=0$ the gap equation reads
\be
\label{Tc}
\quad \left<\frac{\chi_{\mathbf{p}}^2}{\eta_{\mathbf{p}}}
\tanh\left(\frac{\eta_{\mathbf{p}}}{2k_\mathrm{B} \tc}\right) \right>
= \frac{1}{J_\mathrm{sd}}.
\ee
Supposing that $J_\mathrm{sd}$, being an intra-atomic process, is weakly
temperature dependent we can determent its value using band parameters
$\epsilon_\mathrm{s},$ $\epsilon_\mathrm{d},$ $\epsilon_\mathrm{p},$ $t_\mathrm{sp},$ $t_\mathrm{pd},$
$t_\mathrm{pp},$ $E_\mathrm{F}$ and known \tc. Then we can calculate, for
example, the dependence of $T_{\mathrm{c}\,\,\mathrm{max}}$ by position of the Cu~4s
levels $\epsilon_\mathrm{s}$. One illustrative example is shown in
\Fref{fig_logTc}. The almost linear dependence has to have a simple
qualitative interpretation. Let us try to reveal this simplicity using
the BCS interpolation formula
\be k_\mathrm{B} T_\mathrm{c} =
1.14\,\hbar\omega_\mathrm{D}\mathrm{e}^{-1/N(0)V}.
\label{BCSformula}
\ee
In the present case $\omega_\mathrm{D}$ is an energy parameter of
order of the bandwidth, $N(E_\mathrm{F})$ is electronic density of states per
spin at the Fermi level and
\be
V = J_\mathrm{sd}
\frac{4t_\mathrm{pd}t_\mathrm{sp}}{(\epsilon_\mathrm{s}-E_\mathrm{F})(E_\mathrm{F}-\epsilon_\mathrm{p})}
= J_\mathrm{sd}/\tilde{s}(E_\mathrm{F}).
\ee
\begin{figure}[t]
\centering
\includegraphics[height=7cm]{./fig2.eps}
\caption{Logarithm of \tc\ (K) versus $\epsilon_{\text{Cu}\,4\mathrm{s}}$. Science
starts with simplicity -- the almost linear behavior corresponds to
the well-known BCS formulae given by Eqs.~(\ref{BCSformula}) and (\ref{lnTc}).}
\label{fig_logTc}
\end{figure}
The interpolation BCS formula gives
\be
\label{lnTc}
\ln(\tc) \approx \ln(1.14\omega_\mathrm{D}) -
\tilde{s}(E_\mathrm{F})/[N(E_\mathrm{F})J_\mathrm{sd}],
\ee
which according to \Eqref{eq:stilde}
correlations reported by Pavarini~\textit{et al.}~\cite{Pavarini:01}
are actually correlations between the critical temperature
$T_{\mathrm{c}\,\,\mathrm{max}}$ and the BCS coupling constant
$J_\mathrm{sd}/\tilde{s}(E_\mathrm{F})$. This general correlation is
typically perturbed by stripes, inhomogeneities, magnetic phenomena,
structural phase transition, apex oxygen, chemical substitution, and
many other accessories of HTS cuprates. Nevertheless, they can be
clearly seen for all hole doped cuprates. This qualitative agreement
gives hope that four-band model with incorporated s--d exchange can
become a standard model for superconductivity of cuprates. The key
role of the $\tilde{s}(E_\mathrm{F})$ parameter reveals why cuprates
are unique for reaching high-\tc. Imagine that we can easily tune the
position of the Cu~4s level $\epsilon_\mathrm{s}$. The pre-exponential
factor $\omega_\mathrm{D}$ is so big that we can easily reach a
\emph{sauna-temperature} superconductivity if $\epsilon_\mathrm{s}$ is small
enough. Maximal \tc\ is reached upon 3d--2p--4s hybridization.
For the Cu--O duet we have maximal triple coincidence of levels which
ensures the success of the CuO$_2$ plane. For all other combinations
of transition metal with a chalcogenide $\tilde{s}$ is much bigger.
For 25 years many ways to decrease $\epsilon_\mathrm{s}$ were
empirically found: apex oxygen, bilayer hopping, pressure, etc.
Perhaps only metastable artificial layers are not completely
investigated.
\section{Computational method}
The parent CuO$_2$ layer is an insulator with a half-filled conduction
band. Doping with $\tilde{p}$ holes per Cu atom results in metalization
with hole filling $f=\frac{1}{2}+\tilde{p}.$ The optimal doping
corresponds to $\tilde{p}_\mathrm{max}=0.16$ and
$f_\mathrm{max}=0.66$. We will use 66\% hole filling for all examples
in the present work. The Fermi level is determined by the condition
\be
f=\left<\theta(\epsilon_\mathbf{p}>E_\mathrm{F})\right>,\quad \mbox{where}\quad
\theta(\epsilon_\mathbf{p}>E_\mathrm{F}) =
\begin{cases}
1 & \text{if } E_\mathbf{p} > E_\mathrm{F},\\
0 & \text{if } E_\mathbf{p} < E_\mathrm{F}.
\end{cases}
\ee
The parameters of the four-band model can be determined by comparison
with first-principles electronic structure calculations. For example,
using the $\Gamma$ point one can determine the on-site energies
$\epsilon_\mathrm{d}=\epsilon(p_x=0,p_y=0)$ for the conduction band.
In the present paper we use as an illustration a set of parameters (in eV)
similar to Ref.~\cite{Pavarini:01}
\begin{gather}
\epsilon_\mathrm{s}=5.4,\quad \epsilon_\mathrm{p}=-1, \quad \epsilon_\mathrm{d}=0,\\
t_\mathrm{sp}=2,\quad t_\mathrm{pd}=1.5, \quad t_\mathrm{pp}=0.2 \nonumber.
\end{gather}
eV Assuming $T_\mathrm{c,\,\,max}=90$~K for $\epsilon_\mathrm{s}=5.4$
we obtain, according to \Eqref{Tc}, $J_\mathrm{sd}=2.44\;\mathrm{eV}.$
Then for so fixed $J_\mathrm{sd}$ we can calculate the correlations of
$T_\mathrm{c,\,\,max}$ with the range parameter
\be
r\equiv \frac{1}{2(1+s)}.
\ee
Let us discuss briefly our finding.
Within the four-band model one can derive an exact equation for
constant energy contours
\be
-2t(\epsilon)[\cos(p_x)+\cos(p_y)] +4t'(\epsilon)\cos(p_x)\cos(p_y)
=q(\epsilon),
\ee
where $t(\epsilon),$ $t'(\epsilon)$ and $q(\epsilon)$ are polynomial
functions of energy. The ratio of the effective intra-layer hopping
parameters
\be \frac{t'(E_\mathrm{F})}{t(E_\mathrm{F})}\propto
r(E_\mathrm{F}),
\ee
describes small variations of the shape of the Fermi contour
influenced by the position of the Cu~4s level $\epsilon_\mathrm{s}$.
\section{Conclusions}
Electronic structure experts should be proud that after many years of
systematic research band calculations have revealed that \tc\ depends
on the Cu~4s level. This numerical experiment is actually the
crucial one for understanding the mechanism of HTS. The $r$ parameter
is introduced in electronic structure calculations in such a way, that
for the first time we have a mechanism of a physical phenomenon
possibly revealed by computer.
We advocate a conventional theoretical explanation of the \tc-$r$
correlations~\cite{Pavarini:01} which has no alternative among other
theories of HTS. We have to wait for the appearance of some other
descriptions, because this important correlation covers a hole range
of HTS. The LCAO model was only a tool for the theoretical analysis
of the mechanism of HTS. It should be interesting to determine
$S_\mathbf{p}$ and $D_\mathbf{p}$ directly from partial wave analysis
at the muffin spheres directly by the electronic band calculations.
The interpolating $\chi_\mathbf{p}=S_\mathbf{p} D_\mathbf{p}$ function
can be directly substituted in the BCS gap equation (\ref{BCS}) and
the equation for \tc, \Eqref{Tc}. For the cuprates $\chi_\mathbf{p}$
well describes the experimental data for the gap anisotropy~$\Delta_\mathbf{p}$
on the Fermi surface. It remains to be seen if
this mechanism is applicable only to cuprates.
The iron-based pnictides are suspected to pursue an Oscar for
supporting role. It will be extremely interesting to probe if
\Eqref{Tc} can describe the common trends for the \tc\ and gap
anisotropy of ferro-pnictides. In this regards, finally we would like
to re-analyze the ``boost'' that the physics of superconductivity got
from the iron pnictides. Very recently, Mazin~\cite{mazin10} recalled the
Matthias' rules, well known in the physicists’ folklore: 1) A high
symmetry is good; cubic symmetry is the best. 2) A high density of
electronic states is good. 3) Stay away from oxygen. 4) Stay away
from magnetism. 5) Stay away from insulators. 6) Stay away from
theorists.
In order to emphasize the common properties of cuprates and iron
pnictides we feel it compelling to rephrase these rules: (1) A high
symmetry is good, square symmetry is the best, layered structure
ensures empty s-band. (2) Having empty s-band, the Fermi level falls
in the narrow d-band which ensures high density of states. (3) The
oxygen and pnictide p-orbitals are perfect ``go-between'' for the
transition metal 3d and 4s orbitals. These oxidants create significant
4s polarization of the conduction 3d band. (4) We are just in the
epicenter of magnetism; the s-d exchange amplitude which creates the
ferromagnetism of iron, possibly creates the superconductivity of iron
pnictides and cuprates. (5) Having narrow 3d band in the case of
half-filling we observe metal-insulator transition. Transition to
insulator phase is a hint for high density of states at the Fermi
level for a doped compound. (6) Favorites of great socialists like
Lenard, Kapica and Stark, have given similar advice. In order to
surmount this weakness it is necessary to remedy the inferiority
complex of the band calculators. One Nobel prize is a good initial
dose of this treatment. Up to now there is no Nobel prize awarded to
computational physics, but the considered in this paper $T_c$-$r$
correlations are a typical physical phenomenon whose nature is
revealed by numerical calculations. Prediction of an artificial
structure with high $\tilde{s}$ parameter, further synthesis by
solid-state chemists and measurement of $T_c$ may successfully
initiate a new direction of intensive research.
In conclusion, we briefly re-state the main property of high-$T_c$
superconductivity---the triple coincidence of the d-, s-, and p-levels
which ensures big enough BCS coupling constant. This triple
coincidence is analogous to the sequence of vernal equinox, full moon,
and Sunday (the end of the winter season, the end of the moon month
and the end of the week): a triple holiday creating the Great-day. In
this sense the high-$T_c$ superconductivity is the Easter of condensed
matter physics.
The authors are thankful to Prof.~Ivan Zhelyazkov for a critical
reading of the manuscript and to Alvaro de R\'ujula for his interest and
comments during the scientific conference in memory of Matey Mateev,
held in Sofia, 11--12 April 2011.
|
\section{Introduction}
In the last ten years, several studies have focused on the determination of the chemical composition of open clusters with ages
$\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle > 100$ Myr; less attention has been paid to abundances of $\sim 10-100$ Myr old clusters (\citealt{dorazirandich09}, and
references therein), and to the chemical pattern of young ($\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle < 10$ Myr) clusters and star forming regions (hereafter, SFRs;
\citealt{Jamesetal2006, santosetal2008, biazzoetal2011}, and references therein). Instead, elemental abundances of very young clusters
and SFRs provide a powerful tool to investigate the possible common origin of different subgroups, to investigate scenarios of triggered
star formation, and to trace the chemical composition of the solar neighborhood and Galactic thin disk at recent times.
Also, whilst the correlation between stellar metallicity and the presence of giant planets around old solar-type stars is well
established (\citealt{Gonzalez1998, Santosetal2001, johnsonetal2010}) and seems to be primordial (\citealt{Gillietal2006}),
the presence of a metallicity-planet connection at early stages of planet formation is still a matter of debate. On the one
hand, the efficiency of dispersal of circumstellar (or proto-planetary) disks, the planet birthplace, is predicted to depend on
metallicity (\citealt{ercolanoclarke2010}). In a recent study, \cite{yasuietal2010} find that the disk fraction in significantly
low-metallicity clusters ([O/H]$\sim -0.7$) declines rapidly in $<1$~Myr, which is much faster than the value of $\sim 5-7$ Myr
observed in solar-metallicity clusters. They suggest that, since the shorter disk lifetime reduces the time available for planet formation,
this could be one of the reasons for the strong planet-metallicity correlation. On the other hand, \cite{Cusanoetal2011} find
that a significant fraction of the oldest pre-main sequence stars in the SFR Sh 2-284 have preserved their accretion discs/envelopes, despite
the low-metallicity environment. Moreover, none of the SFRs with available metallicity determination is metal rich, challenging a full
understanding of the metallicity-planet connection at young ages. Additional
measurements of [Fe/H] in several star forming regions and young clusters are thus warranted.
In a recent paper, we have presented an homogeneous determination of the abundance pattern in the Orion Nebula Cluster (ONC) and the
OB1b subgroup (\citealt{biazzoetal2011}). We report here an abundance study of two other young groups belonging to the Orion complex;
namely, the 25~Orionis and $\lambda$~Orionis clusters. Although the two clusters have been extensively observed and analyzed, no abundance
measurement is available for 25~Ori, while for $\lambda$~Ori the metallicity of only one member has been derived
(namely, Dolan 24; \citealt{dorazietal09}). The metallicity of these two clusters will allow us to put further constrains on the metal
distribution in the whole Orion complex; furthermore, the age of 25~Ori ($\sim 7-10$ Myr; \citealt{bricenoetal05}) and $\lambda$~Ori
($\sim 5-10$ Myr; \citealt{dolanmathieu2002}) is comparable to the maximum lifetime of accretion disks as derived from near-IR studies
and can be considered indicative of the timescale for the formation of giant planets. For these reasons, measurements of [Fe/H] and
other elements in such clusters are important and timely.
The 25~Ori group is a concentration of T Tauri stars (TTS) in the Orion OB1a subassociation, roughly surrounding the B2e star
25~Ori; this stellar aggregate was first recognized by \cite{bricenoetal05} during the course of the CIDA Variability Survey of
Orion (CVSO) as a distinct feature centered at $\alpha_{\rm J2000.0}=$5$^{\rm h}$23$^{\rm m}$ and
$\delta_{\rm J2000.0}=$1$^{\degr}$45$'$. The 25~Orionis group was also reported by \cite{kharchenkoetal2005} as ASCC-16 in their
list of 109 new open clusters identified through parallaxes, proper motions, and photometric data. \cite{bricenoetal07} determined
radial velocities of the cluster candidates, finding that the 25~Ori
group constitutes a distinct kinematic group from the other Orion subassociations. The color-magnitude diagram, the
lithium equivalent widths, and the fraction of CTTS ($\sim$6\%) are consistent with the 25~Ori group being older than Ori
OB1b (age $\sim4-5$ Myr; \citealt{bricenoetal07}). Finally, the same authors derive a cluster radius of $\sim 7$ pc
centered 23.6$'$ southeast of the star 25~Ori, with some indication of extension farther north.
The $\lambda$~Orionis cluster, discovered by \cite{gomezlada1998}, is distributed over an area of 1 square degree around
the O8 III star $\lambda^1$~Ori (at a distance of 400 pc; \citealt{murdinpenston1977}). As suggested by
\cite{dolanmathieu1999, dolanmathieu2001}, the star formation process in this region began some 8--10 Myr ago in the central region
and has since been accelerating, followed by an abrupt decline $\sim$1--2 Myr ago due to a supernova (SN) explosion
(\citealt{cunhasmith1996}). CO surveys clearly show that the central region of the cluster, composed by eleven OB stars,
has been largely vacated of molecular gas. A ring of neutral and molecular hydrogen with a diameter of $\sim$9\degr\,surrounds
an \ion{H}{ii} region, and is likely the consequence of the sweeping of interstellar gas by the \ion{H}{ii} region
(see, e.g., \citealt{langetal2000}, and references therein).
In Sect.~\ref{sec:obs} we describe the target selection, observations, and data reduction. In Sect.~\ref{sec:analysis}
the radial velocity and elemental abundance measurements are determined, while the discussion and conclusions are presented
in Sect.~\ref{sec:discussion} and \ref{sec:conclusion}.
\section{Observations and data analysis}
\label{sec:obs}
\subsection{Target selection}
\label{sec:sample}
We selected four K2-K7 members of 25~Ori from the \cite{bricenoetal07} sample, and four additional stars located about 1 degree to
the north-east (Brice\~no et al. 2011, in preparation). These four targets, candidate pre-main sequence (PMS) stars in the CVSO
dataset, were included for follow-up observations with the Hectospec and Hectochelle spectrographs at the 6.5m MMT (Arizona), in order
to investigate the extent of the 25~Ori aggregate. As for $\lambda$~Ori, we selected four stars from the \cite{dolanmathieu1999} sample
and two candidates from \cite{barradoetal2004}.
The distribution in the sky of the sample stars is shown in Figs.~\ref{fig:25Ori_map} and \ref{fig:lambdaOri_map} overlaid on CO
maps\footnote{http://www.cfa.harvard.edu/mmw/MilkyWayinMolClouds.html}. Both samples lie away from giant molecular clouds and
dark clouds, i.e. far from regions where star formation is actively occurring.
In Table~\ref{tab:literature} we list the sample stars along with information from the literature. In particular, in Columns 1--5 we
give for the 25~Ori stars the name, $VJK$ magnitude, spectral type, and object class from \cite{Cutri2003}, and Brice\~no et al.
(2005; 2007; 2011, in preparation), while for the $\lambda$~Ori stars the name, $VJK$ magnitude,
spectral type, and disk property are from \cite{dolanmathieu1999}, \cite{Cutri2003}, \cite{Zacha2004}, \cite{saccoetal2008},
and \cite{hernandezetal2010}. In the last Column of both samples, radial velocities from \cite{dolanmathieu1999}, \cite{bricenoetal07},
and \cite{saccoetal2008} are given.
Both samples were selected with the following criteria: $i)$ no evidence of extremely high accretion; $ii)$ no evidence of high rotational
velocity ($v\sin i \protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle > 30$ km/s), and $iii)$ no evidence of binarity.
\begin{figure*}
\begin{center}
\includegraphics[width=13cm,angle=-90]{25Ori_map.eps}
\caption{Spatial distribution of the 25~Ori stars (squares) in Galactic coordinates overlaid on the velocity-integrated CO
J$=1\rightarrow$0 emission map (\citealt{langetal2000}). The position of 25~Ori, the Orion Belt stars and $\sigma$~Ori is shown
by open triangles.}
\label{fig:25Ori_map}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[width=13cm,angle=-90]{lambdaOri_map.eps}
\caption{As in Fig.~\ref{fig:25Ori_map}, but for the $\lambda^1$~Ori stars (triangles). The position of $\lambda$~Ori is shown by
an open square ($l\sim195$\degr, $b\sim-$12\degr).}
\label{fig:lambdaOri_map}
\end{center}
\end{figure*}
\subsection{Observations and data reduction}
The observations were carried out in October 2008--January 2010 with {\sc flames} (\citealt{pasquinietal02}) on UT2 Telescope. We
observed both 25~Ori and $\lambda$~Ori using the fiber link to the red arm of {\sc uves}. We allocated four fibers to the 25~Ori
sample, and two or four fibers to the $\lambda$~Ori sample (depending on the pointing), leaving from two to six fibers for sky
acquisition. We used the CD\#3 cross-disperser covering the range 4770--6820 \AA~at the resolution $R=47\,000$, allowing us to select
70+9 \ion{Fe}{i}+\ion{Fe}{ii} lines, as well as 57 lines of other elements (Sect.~\ref{sec:abundance}).
In the end, the 25~Ori targets were observed in two different pointings, each including four stars, with no overlap. For each
pointing, we obtained four 45 minutes exposures, resulting in a total integration time of 3 hours each. Two pointings
were also necessary to observe the $\lambda$~Ori targets. The pointings included four and two stars, and each field was observed
three times, for a total exposure time of 2.3 hours each. The log book of the observations is given in Table~\ref{tab:observations}.
For a detailed description of the instrumental setup and data reduction, we refer to Biazzo et al. (2011; and references therein).
Ultimately, the typical signal-to-noise ratio ($S/N$) of the co-added spectra in the region of the Li-$\lambda$6708 \AA~line
was 30--80 for 25~Ori and 30--150 for $\lambda$~Ori (see Table~\ref{tab:radvel_lithium}).
\setlength{\tabcolsep}{2.5pt}
\begin{table}[h]
\caption{Sample stars.}
\label{tab:literature}
\tiny
\begin{center}
\begin{tabular}{lccclcc}
\hline
\hline
~\\
\multicolumn{7}{c}{25~Ori}\\
\hline
Star$^a$ & $V$ & $J$ & $K$ & Sp.T. & Class$^b$ & $V_{\rm rad}$ \\
& (mag) & (mag) & (mag) & & & (km/s) \\
\hline
CVSO35 & 14.72 & 11.46 & 10.35 & K7 & C & 18.8 \\
CVSO207 & 13.63 & 11.35 & 10.56 & K4 & W & 24.2 \\
CVSO211 & 13.98 & 11.78 & 11.06 & K5 & W & 19.4 \\
CVSO214 & 13.42 & 11.43 & 10.74 & K2 & W & 19.1 \\
CVSO259 & 14.93 & 12.25 & 11.40 & K7.25 & W & ... \\
CVSO260 & ... & 12.57 & 11.70 & M0.6 & W & ... \\
CVSO261 & ... & 12.71 & 11.83 & M0 & W & ... \\
CVSO262 & 13.84 & 11.26 & 10.45 & ... & ... & ... \\
\hline
~\\
\end{tabular}
\begin{tabular}{lccclcc}
\multicolumn{7}{c}{$\lambda$~Ori}\\
\hline
Star$^a$ & $V$ & $J$ & $K$ & Sp.T. & Class$^b$ & $V_{\rm rad}$ \\
& (mag) & (mag)& (mag) & & & (km/s) \\
\hline
DM07 & 13.01 & 10.80 & 10.13 & ... & DL & 25.48 \\
DM14 & 15.18 & 12.07 & 11.19 & K7 & DL & 25.66, 27.72 \\
DM26 & 13.02 & 10.95 & 10.30 & ... & EV & 24.96 \\
DM21 & 14.80 & 11.66 & 10.84 & ... & DL & 23.65 \\
CFHT6 & 13.97 & 11.54 & 10.65 & ... & DL & ... \\
CFHT12 & 14.09 & 11.82 & 10.80 & ... & DL & ... \\
\hline
\end{tabular}
\end{center}
$^a$ Star names from the following sources: CVSO = CIDA Variability Survey of Orion (\citealt{bricenoetal05});
DM = \cite{dolanmathieu1999}; CFHT = Canada-France-Hawaii Telescope
optical survey (\citealt{barradoetal2004}).\\
$^b$ W: Weak-Lined T Tauri; C: Classical T Tauri; DL: Disk-Less; EV: EVolved disk.
\end{table}
\normalsize
\setlength{\tabcolsep}{4.pt}
\begin{table}
\caption{Log of the observations.}
\label{tab:observations}
\begin{center}
\begin{tabular}{cccccc}
\hline
\hline
$\alpha$ & $\delta$ & Date & UT & $t_{\rm exp}$ & \#\\
(h:m:s) & (\degr: $^\prime$ : \arcsec) & (d/m/y) & (h:m:s) & (s) & (stars) \\
\hline
~\\
\multicolumn{6}{c}{25~Ori}\\
\hline
05:25:14 & 01:45:32 & 12/10/2008 & 07:17:02 & 2775 & 4\\
05:25:14 & 01:45:32 & 16/01/2009 & 02:20:23 & 2775 & 4\\
05:25:14 & 01:45:32 & 17/01/2009 & 00:54:27 & 2775 & 4\\
05:25:14 & 01:45:32 & 19/01/2009 & 00:37:23 & 2775 & 4\\
05:28:06 & 02:51:36 & 21/01/2009 & 00:52:22 & 2775 & 4\\
05:28:06 & 02:51:36 & 21/01/2009 & 01:40:26 & 2775 & 4\\
05:28:06 & 02:51:36 & 27/11/2008 & 04:58:42 & 2775 & 4\\
05:28:06 & 02:51:36 & 29/11/2008 & 06:27:28 & 2775 & 4\\
\hline
~\\
\multicolumn{6}{c}{$\lambda$~Ori}\\
\hline
05:34:33 & 09:44:01 & 12/12/2009 & 03:16:07 & 2775 & 4\\
05:34:33 & 09:44:01 & 12/12/2009 & 04:08:18 & 2775 & 4\\
05:34:33 & 09:44:01 & 14/12/2009 & 03:50:03 & 2775 & 4\\
05:35:07 & 09:54:55 & 14/12/2009 & 04:49:03 & 2775 & 2\\
05:35:07 & 09:54:55 & 15/12/2009 & 03:11:49 & 2775 & 2\\
05:35:07 & 09:54:55 & 07/01/2010 & 03:33:58 & 2775 & 2\\
\hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[h]
\caption{S/N ratios, radial velocities, luminosities, and masses.}
\label{tab:radvel_lithium}
\tiny
\begin{center}
\begin{tabular}{lccccc}
\hline
\hline
~\\
\multicolumn{6}{c}{25~Ori}\\
\hline
Star & $S/N$& $V_{\rm rad}$ & Notes$^a$ & $L$&$M^{\rm in}$\\
& & (km/s) & & ($L_{\odot}$)& ($M_{\odot}$)\\
\hline
CVSO35 & 70 & 19.8$\pm$1.2 & M & ... & ...\\
CVSO207 & 70 & 18.6$\pm$0.6 & M & 0.64& 1.06\\
CVSO211 & 70 & 21.0$\pm$0.4 & M & 0.44& 0.92\\
CVSO214 & 70 & 20.1$\pm$0.2 & M & 0.62& 1.02\\
CVSO259 & 50 & 18.9$\pm$0.8 & M & 0.27& 0.74\\
CVSO260 & 30 & 19.2$\pm$0.6 & M & ... & ... \\
CVSO261 & 30 & 19.8$\pm$0.6 & M & ... & ... \\
CVSO262 & 80 & 19.2$\pm$1.0 & M & 0.67& 1.07 \\
\hline
~\\
\end{tabular}
\begin{tabular}{lccccc}
\multicolumn{6}{c}{$\lambda$~Ori}\\
\hline
Star & $S/N$& $V_{\rm rad}$ & Notes$^a$ & $L$ & $M^{\rm in}$\\
& & (km/s) & & ($L_{\odot}$)& ($M_{\odot}$) \\
\hline
DM07 & 150 & 25.3$\pm$0.2 & M & 1.63& 1.35\\
DM14 & 60 & 26.6$\pm$0.4 & M & 0.46& 0.77\\
DM26 & 120 & 28.0$\pm$0.2 & M & 1.44& 1.22\\
DM21 & 50 & 25.4$\pm$1.2 & M & 0.67& 0.86\\
CFHT6 & 30 & 26.8$\pm$0.6 & PB& 0.79& 0.82\\
" & " & 31.1$\pm$0.5 & " & " & " \\
CFHT12 & 60 & $-$3.3$\pm$2.5 & NM& ... & ...\\
\hline
\end{tabular}
\end{center}
$^a$ M = Member; NM = Non-Member; PB = Probable Binary.
\end{table}
\normalsize
\section{Analysis and results}
\label{sec:analysis}
\subsection{Radial velocity measurements and membership}
\label{sec:rad_vel}
We measured radial velocities (RVs) in each observing night, to identify possible non-members and binaries. Since we could not
acquire any template spectrum, we first chose for both 25~Ori and $\lambda$~Ori the stars with the spectra at highest $S/N$,
low $v\sin i$ ($<$15 km s$^{-1}$), available RV measurements from the literature, and without any accretion signature. In particular,
we considered CVSO211 for 25~Ori and DM14 for $\lambda$~Ori and followed the method explained in \cite{biazzoetal2011}.
We obtained $V_{\rm rad}=$21.0$\pm$0.4 km s$^{-1}$ for CVSO211, while for DM14 we found $V_{\rm rad}=$26.6$\pm$0.4 km s$^{-1}$.
These values are close to the values of $V_{\rm rad}=$19.4$\pm$0.5 km s$^{-1}$ and $V_{\rm rad}=$27.72$\pm$0.18 km s$^{-1}$
obtained by \cite{bricenoetal07} and \cite{saccoetal2008}, respectively. Using these two templates, we measured the heliocentric RV
of all our targets using the task {\sc fxcor} of the
IRAF\footnote{IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of the
Universities for Research in Astronomy, inc. (AURA) under cooperative agreement with the National Science Foundation.} package
{\sc rv}, following the prescriptions given by \cite{biazzoetal2011}.
All our stars were characterized by single peaks of the cross-correlation function. For all measurements with RV of
different nights in agreement within $2\sigma$, we computed a mean RV as the weighted average of the different values.
Only the star CFHT6 in $\lambda$~Ori shows evidence of binarity, with RV measurements around $V_{\rm rad}=$26.8 km s$^{-1}$
(in December 2009) and $V_{\rm rad}=$31.1 km s$^{-1}$ (in January 2010). Final radial velocities are listed in
Table~\ref{tab:radvel_lithium}.
We find that all the 25~Ori stars are members with a cluster distribution centered at $<V_{\rm rad}>=19.6\pm0.8$ km s$^{-1}$
(Fig.~\ref{fig:vrad_distr}, Table~\ref{tab:radvel_lithium}), i.e. very close to the peak of $<V_{\rm rad}>=19.7\pm1.7$ km s$^{-1}$
found by \cite{bricenoetal07}. In particular, the four new PMS stars off to the north-east of 25~Ori (CVSO259,260,261,262)
share the same radial velocity of the cluster members from \cite{bricenoetal07}, consistent with being members of this
stellar aggregate.
Although four stars represent a small number statistics, this result supports the suggestion made by \cite{bricenoetal07}
that 25~Ori extends further north than implied by the map in their Fig. 5. The full photometric census of 25~Ori will be
presented in Brice\~no et al. (2011, in preparation).
All the $\lambda$~Ori stars are members, with the exception of CFHT12, as also suggested
by \cite{barradoetal2004} from photometric selection criteria. Excluding CFHT6 and CFHT12, we find
a mean $\lambda$~Ori RV of $V_{\rm rad}=$26.3$\pm$1.3 km s$^{-1}$, in good agreement with the cluster distribution
centered at $<V_{\rm rad}>=27.03\pm0.49$ km s$^{-1}$ (\citealt{saccoetal2008}) .
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=9.2cm]{vrad_distr_25Ori_lambdaOri.ps}
\end{tabular}
\caption{Radial velocity distributions (solid lines) of all the targets in 25~Ori ({\it upper panel}) and $\lambda$~Ori
({\it lower panel}). The dashed lines represent the distribution obtained by \cite{bricenoetal07} for 47 stars in 25~Ori
and by \cite{saccoetal2008} for 45 members in $\lambda$~Ori.}
\label{fig:vrad_distr}
\end{center}
\end{figure}
\subsection{Final sample for elemental analysis}
\label{sec:final_sample}
Unfortunately, some of the stars are not suitable for abundance analysis. In particular, CVSO35 is a very cool accreting
star (classified as CTTS by \citealt{bricenoetal05}), for which the abundance determination through equivalent widths is not
reliable. Moreover, CVSO260 and CVSO261 have low $S/N$ spectra ($\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle < 40$) and low effective temperature
($T_{\rm eff}\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <$4000 K). In these $T_{\rm eff}$ regime, strong molecular bands do not allow one to derive accurate EWs.
As a result, we performed abundance measurements in ten stars, five in 25~Ori and
five in $\lambda$~Ori. In Figs.~\ref{fig:spectra_25} and \ref{fig:spectra_l} we display the co-added spectra of these stars
in the spectral region (6220--6260 \AA) containing several features used to derive the effective temperature through
line-depth ratios (Sect. \ref{sec:initial_parameters}), and to measure the elemental abundances (Sect. \ref{sec:abundance}).
\begin{figure*}
\begin{center}
\begin{tabular}{c}
\resizebox{\hsize}{!}{\includegraphics{spectra_25Ori.ps}}\\
\end{tabular}
\caption{Portion of spectra of the targets in 25~Ori used for the abundance determination. The values of $T_{\rm eff}^{\rm SPEC}$
are given for each star.}
\label{fig:spectra_25}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\begin{tabular}{c}
\resizebox{\hsize}{!}{\includegraphics{spectra_lambdaOri.ps}}\\
\end{tabular}
\caption{As in Fig.~\ref{fig:spectra_25}, but for $\lambda$~Ori. }
\label{fig:spectra_l}
\end{center}
\end{figure*}
\subsection{Initial temperature and surface gravity}
\label{sec:initial_parameters}
Since effective temperatures from the literature were not available for all our targets, we used as initial
the values ($T_{\rm eff}^{\rm LDR}$) derived using the method based on line-depth ratios (\citealt{Gray1991}).
We followed the prescriptions given by \cite{biazzoetal2011} and obtained the values listed in Table~\ref{tab:param_chem}
and plotted in Fig.~\ref{fig:teff_ldr_spec}.
As for surface gravity, the initial gravity ($\log g^{\rm in}$ in Table~\ref{tab:param_chem}) was obtained from the relation
between mass, luminosity, and effective temperature: $\log g=4.44+\log (M/M_\odot) +4 \log (T_{\rm eff}/5770)-\log (L/L_\odot)$.
Effective temperatures were set to our $T_{\rm eff}^{\rm LDR}$ (see Table~\ref{tab:param_chem}). Stellar luminosities
(Table~\ref{tab:radvel_lithium}), were computed from a relationship between the absolute bolometric magnitude, the $J$ absolute
magnitude ($M_{J}$), and the $(V-K)$ color index given by \cite{kenyonhartmann95} for late-type stars,
considering $M_{\rm bol}^{\odot} = 4.64$ as solar bolometric magnitude (\citealt{cox2000}). The value of $M_{J}$ was obtained adopting
a distance of 330 pc for 25~Ori (\citealt{bricenoetal05}) and 400 pc for $\lambda$~Ori (\citealt{murdinpenston1977}),
respectively. We used as mean interstellar extinctions the values of $A_V=0.29$ given by \cite{bricenoetal07} for 25~Ori and
$A_J=0.106$ for $\lambda$~Ori (\citealt{barradoetal2004}). Initial masses ($M^{\rm in}$ in Table~\ref{tab:radvel_lithium}) were
estimated from the isochrones of \cite{pallastahler1999} at given stellar $M_{J}$ and $T_{\rm eff}^{\rm LDR}$.
The HR diagrams for both samples are shown in Fig.~\ref{fig:HRD}, together
with the \cite{pallastahler1999} tracks and isochrones. The 25~Ori stars show an age around 10 Myr, with the exception
of one member, which is marginally older. Three of the $\lambda$~Ori targets are 3--5 Myr old, while the other
two are slightly older. Both samples show ages in agreement with previous studies (see \citealt{dolanmathieu2001}
for $\lambda$~Ori and \citealt{bricenoetal07} for 25~Ori).
\begin{figure}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=9cm]{Teff_our_25Ori_lambdaOri.ps}
\end{tabular}
\caption{Spectroscopic effective temperatures ($T_{\rm eff}^{\rm SPEC}$) derived by imposing the excitation
equilibrium versus effective temperatures obtained through line-depth
ratios ($T_{\rm eff}^{\rm LDR}$). The squares and triangles refer to 25~Ori and $\lambda$~Ori, respectively.}
\label{fig:teff_ldr_spec}
\end{center}
\end{figure}
\begin{sidewaystable*}
\setlength{\tabcolsep}{1.2pt}
\caption{Astrophysical parameters and elemental abundances derived from spectroscopic analysis.}
\label{tab:param_chem}
\tiny
\begin{center}
\begin{tabular}{lcccccrrrrrrrrrrrl}
\hline
\hline
Star&$T_{\rm eff}^{\rm LDR}$&$T_{\rm eff}^{\rm SPEC}$& $\tiny{\log g}^{\rm in}$&$\tiny{\log g}^{\rm SPEC}$&$\xi$&[\ion{Fe}{i}/H]&[\ion{Fe}{ii}/H]&[Na/Fe]&[Al/Fe]&[Si/Fe]&[Ca/Fe]&[\ion{Ti}{i}/Fe]&[\ion{Ti}{ii}/Fe]&$<$[Ti/Fe]$>$&[Ni/Fe]\\
& (K) & (K) & & & \tiny{(km/s)} & & & & & & & & & & & \\
\hline
~\\
\multicolumn{16}{c}{25~Ori}\\
\hline
CVSO207& 4447$\pm$48 & 4600 & 4.2& 4.1 & 1.7 &$-$0.10$\pm$0.09(34)& $-$0.07$\pm$0.11(4) & 0.03$\pm$0.10(1)& 0.16$\pm$0.11(2) & 0.06$\pm$0.21(5) & 0.05$\pm$0.09(3) & $-$0.02$\pm$0.16(6)& 0.00$\pm$0.10(2) & $-$0.01$\pm$0.13&$-$0.04$\pm$0.11(13)\\
CVSO211& 4399$\pm$47 & 4550 & 4.3& 4.2 & 1.4 &$-$0.05$\pm$0.09(43)& $-$0.03$\pm$0.13(3) & $-$0.04$\pm$0.09(1)& 0.08$\pm$0.12(2) & 0.08$\pm$0.24(5) & 0.01$\pm$0.10(3) & $-$0.03$\pm$0.12(7)& 0.10$\pm$0.10(2) & 0.04$\pm$0.13 &$-$0.02$\pm$0.12(15) \\
CVSO214& 4787$\pm$52 & 4900 & 4.3& 4.2 & 2.1 &$-$0.08$\pm$0.06(41)& $-$0.06$\pm$0.09(5) & 0.08$\pm$0.06(1)& 0.12$\pm$0.13(2) & 0.10$\pm$0.10(3) & 0.06$\pm$0.08(3) & 0.01$\pm$0.10(9)& 0.00$\pm$0.08(3) & 0.01$\pm$0.13 &$-$0.02$\pm$0.13(22) \\
CVSO259& 4011$\pm$42 & 4050 & 4.2& ... & 1.4 & 0.01$\pm$0.16(32)& ... & $-$0.26$\pm$0.16(1)& $-$0.06$\pm$0.22(2) & 0.06$\pm$0.16(2) & $-$0.20$\pm$0.16(2) & $-$0.28$\pm$0.19(8)& ... & ... & 0.01$\pm$0.21(6) \\
CVSO262& 4453$\pm$150 & 4450 & 4.1& ... & 1.9 &$-$0.01$\pm$0.09(23)& ... & ... & $-$0.05$\pm$0.09(2) & ... & $-$0.04$\pm$0.15(1) & $-$0.21$\pm$0.21(3)& ... & ... &$-$0.02$\pm$0.17(2) \\
~\\
25~Ori & & && & &$-$0.05$\pm$0.05 & $-$0.05$\pm$0.02 & 0.02$\pm$0.06 & 0.08$\pm$0.08 & 0.08$\pm$0.02 & 0.02$\pm$0.05 & & & 0.01$\pm$0.02 &$-$0.02$\pm$0.02 \\
\hline
~\\
\multicolumn{16}{c}{$\lambda$~Ori}\\
\hline
DM07 & 4956$\pm$67& 5050 & 4.1& 4.2 & 2.1 & 0.00$\pm$0.07(39)& 0.02$\pm$0.04(3)& 0.06$\pm$0.08(1)& 0.15$\pm$0.01(2)& 0.13$\pm$0.03(4)&$-$0.03$\pm$0.09(2)& 0.00$\pm$0.09(7)&0.12$\pm$0.17(4)& 0.06$\pm$0.20 &$-$0.01$\pm$0.07(14)\\
DM14 & 4037$\pm$64& 4000 & 4.0& ... & 1.4 & 0.01$\pm$0.08(31)& ... &$-$0.26$\pm$0.09(1)& $-$0.13$\pm$0.05(2)& 0.18$\pm$0.19(4)&$-$0.21$\pm$0.09(2)&$-$0.23$\pm$0.12(6)&0.45$\pm$0.14(2)& 0.11$\pm$0.18 & 0.08$\pm$0.13(16)\\
DM21 & 4167$\pm$55& 4300 & 4.0& ... & 1.4 & 0.02$\pm$0.08(19)& ... &$-$0.13$\pm$0.13(1)& $-$0.01$\pm$0.09(2)& 0.06$\pm$0.43(2)&$-$0.09$\pm$0.13(1)&$-$0.05$\pm$0.08(3)& ... &... &$-$0.03$\pm$0.15(3)\\
DM26 & 5128$\pm$40& 5300 & 4.2& ... & 2.3 & 0.01$\pm$0.10(21)& ... & ... & 0.23$\pm$0.01(2)& 0.04$\pm$0.18(3)& ... & 0.06$\pm$0.10(5)& ... & ...& 0.01$\pm$0.13(5)\\
CFHT6 & 4099$\pm$58& 4000 & 3.8& ... & 1.6 & $-$0.01$\pm$0.14(18)& ... &$-$0.23$\pm$0.16(1)& $-$0.20$\pm$0.09(2)& 0.14$\pm$0.17(1)&$-$0.26$\pm$0.14(2)&$-$0.29$\pm$0.20(3)&0.26$\pm$0.19(1)& $-$0.02$\pm$0.28 & 0.04$\pm$0.17(5)\\
~\\
$\lambda$~Ori & & & && & 0.01$\pm$0.01 & 0.02$\pm$0.04 & 0.06$\pm$0.08 & 0.19$\pm$0.06 & 0.11$\pm$0.06 & $-$0.03$\pm$0.09 & & & 0.05$\pm$0.06 &0.02$\pm$0.04 \\
\hline
\end{tabular}
\end{center}
\end{sidewaystable*}
\normalsize
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{HRD_25Ori_Palla.ps}
\includegraphics[width=9cm]{HRD_lambdaOri_Palla.ps}
\caption{HR diagram of the members of 25~Ori ({\it upper panel}; squares) and $\lambda$~Ori ({\it lower panel}; triangles)
where the abundance analysis was done. The \cite{pallastahler1999} PMS evolutionary tracks are displayed with
the labels representing their masses (solid lines). Similarly, the birthline and isochrones (from 1 to 30 Myr) are shown with
dash-dotted and dotted lines, respectively.}
\label{fig:HRD}
\end{center}
\end{figure}
\subsection{Stellar abundance measurements}
\label{sec:abundance}
Stellar abundances were measured following the prescriptions given by \cite{biazzoetal2011}. We briefly list
the main steps, but we refer to that paper for a detailed description of the method.
\begin{itemize}
\item We performed abundance analysis in local thermodynamic equilibrium (LTE) conditions using equivalent widths
and the 2002 version of MOOG (\citealt{sneden1973}). The radiative and Stark broadening were treated in a standard way,
while, for collisional broadening, we used the \cite{unsold1955} approximation.
\item \cite{Kuru93} and Brott \& Hauschildt (2010, private comm.) grid of plane-parallel model atmospheres were used for
warm ($>$4400 K) and cool ($\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <$4400 K) stars, respectively.
\item Equivalent widths were measured by direct integration or by Gaussian fit using the IRAF {\sc splot} task.
\item We adopted the line list of \cite{biazzoetal2011}, adding other useful iron lines (see Table~\ref{tab:other_iron_lines})
taken from \cite{santosetal2008}.
\item After removing lines with $EW>150$ m\AA, likely saturated and most affected by the treatment of damping, we
applied a $2\sigma$ clipping rejection.
\item Although we do not expect high levels of veiling in our sample, we searched for any possible excess emission. We
found that all stars had veiling levels consistent with zero, confirming that this quantity is not an issue
in our abundance measurements.
\item Our study is performed differentially with respect to the Sun. We analyzed the {\sc uves} solar spectrum adopting
$T_{\rm eff}=5770$ K, $\log g=4.44$, $\xi=1.1$ km s$^{-1}$ (see \citealt{randichetal2006}). We obtained
$\log n{\rm (\ion{Fe}{i})}=7.52\pm0.02$ and $\log n{\rm (\ion{Fe}{i})}=7.51\pm0.02$ with Kurucz and Brott \& Hauschildt
model atmospheres, respectively. For the iron solar abundance of each line, see Biazzo et al. (2011, and references therein),
and Table~\ref{tab:other_iron_lines}. The solar abundances of all the other elements are listed in Table~\ref{tab:radvel_lithium}
of \cite{biazzoetal2011}, since we used the same line list.
\end{itemize}
\begin{table}
\caption{Wavelength, excitation potential, oscillator strength, equivalent width, and abundance obtained with
ATLAS (\citealt{Kuru93}) and GAIA (Brott \& Hauschildt 2010, private comm.) models for the iron lines taken
from \cite{santosetal2008} are listed.}
\label{tab:other_iron_lines}
\begin{center}
\begin{tabular}{ccrccc}
\hline
\hline
$\lambda$ & $\chi$ & $\log gf$ & $EW$ & $\log n_{\rm ATLAS}$& $\log n_{\rm GAIA}$\\
(\AA) & (eV) & & (m\AA) & & \\
\hline
5322.05 & 2.28 & $-$2.90 & 62.3 & 7.49 & 7.47 \\
5852.22 & 4.55 & $-$1.19 & 41.2 & 7.50 & 7.48 \\
5855.08 & 4.61 & $-$1.53 & 22.2 & 7.46 & 7.45 \\
6027.06 & 4.08 & $-$1.18 & 63.7 & 7.48 & 7.46 \\
6079.01 & 4.65 & $-$1.01 & 48.3 & 7.54 & 7.52 \\
6089.57 & 5.02 & $-$0.88 & 36.1 & 7.51 & 7.49 \\
6151.62 & 2.18 & $-$3.30 & 51.1 & 7.47 & 7.47 \\
6608.03 & 2.28 & $-$3.96 & 17.8 & 7.46 & 7.46 \\
6646.94 & 2.61 & $-$3.94 & 11.0 & 7.51 & 7.51 \\
6710.32 & 1.48 & $-$4.82 & 17.2 & 7.49 & 7.50 \\
\hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{Stellar parameters}
\label{sec:stellar_parameters}
Effective temperatures were also determined by imposing the condition that the \ion{Fe}{i} abundance does not depend on the line
excitation potentials. The initial value of the effective temperature ($T_{\rm eff}^{\rm LDR}$), and the final values
($T_{\rm eff}^{\rm SPEC}$) are listed in Table~\ref{tab:param_chem} and plotted in Fig.~\ref{fig:teff_ldr_spec}. The temperatures
obtained with both methods are in good agreement (mean difference of $\sim$100 K), with $T_{\rm eff}^{\rm SPEC}$ higher on average
than $T_{\rm eff}^{\rm LDR}$.
The microturbulence velocity $\xi$ was determined by imposing that the \ion{Fe}{i} abundance is independent on the line equivalent
width. The initial value was set to 1.5 km\,s$^{-1}$, and the final values are listed in Table~\ref{tab:param_chem}.
The surface gravity was determined by imposing the ionization equilibrium of \ion{Fe}{i}/\ion{Fe}{ii}. With the exception of
CVSO259 and CVSO262 with few suitable lines in the spectra, we were able to
constrain the gravity for the stars in 25~Ori ($\log g^{\rm SPEC}$ in Table~\ref{tab:param_chem}). In the case of $\lambda$~Ori, only DM07 shows enough
\ion{Fe}{ii} lines to allow a spectroscopic measurement of the gravity.
\subsubsection{Errors}
\label{sec:abun_errors}
As described in \cite{biazzoetal2011}, elemental abundances are affected by random (internal) and systematic (external) errors.
In brief, sources of internal errors include uncertainties in atomic parameters, stellar parameters, and line equivalent widths,
where the former should cancel out when the analysis is carried out differentially with respect to the Sun, as in our case.
Errors due to uncertainties in stellar parameters were estimated by assessing errors in $T_{\rm eff}$, $\xi$, and $\log g$, and
by varying each parameter separately, while keeping the other two unchanged. We found that variations in $T_{\rm eff}$ larger than
60 K would introduce spurious trends in $\log n{\rm (\ion{Fe}{i})}$ versus $\chi$; variations in $\xi$ larger than 0.2 km s$^{-1}$
would result in significant trends of $\log n{\rm (\ion{Fe}{i})}$ versus $EW$, and variations in $\log g$ larger than 0.2 dex would
lead to differences between $\log n{\rm (\ion{Fe}{i})}$ and $\log n{\rm (\ion{Fe}{ii})}$ larger than 0.05 dex. Errors in abundances
due to uncertainties in stellar parameters are summarized in Table~\ref{tab:errors} for one of the coolest stars (DM14) and for the
warmest target (DM07).
Random errors in EW are well represented by the standard deviation around the mean abundance determined from all the lines. When only
one line was measured, the abundance error is the standard deviation of three independent EW measurements. In Table~\ref{tab:param_chem}
these errors are listed together with the number of lines used (in brackets).
As for the external error, the biggest one is given by the abundance scale, which is mainly influenced by errors in
model atmospheres. As pointed out by \cite{biazzoetal2011}, this error source is negligible, because we used for both
clusters the same procedure and instrument set-up to measure the elemental abundances.
\begin{table*}
\caption{Internal errors in abundance determination due to uncertainties in stellar parameters for one of the coolest
star (namely, DM14) and for the warmest star (DM07). Numbers refer to the differences between
the abundances obtained with and without the uncertainties in stellar parameters.}
\label{tab:errors}
\begin{center}
\begin{tabular}{lccc}
\hline
\hline
DM14 & $T_{\rm eff}$=4000 K & $\log g=4.0$ & $\xi=1.4$ km/s\\
\hline
$\Delta$ & $\Delta T_{\rm eff}=-/+60$ K & $\Delta \log g=-/+0.2$ & $\Delta \xi=-/+0.2$ km/s\\
\hline
$[$\ion{Fe}{i}/H$]$ & 0.05/$-$0.03 & $-$0.04/0.06 & 0.05/$-$0.04 \\
$[$Na/Fe$]$ & $-$0.09/0.06 & 0.07/$-$0.09 & $-$0.03/0.01 \\
$[$Al/Fe$]$ & $-$0.05/0.03 & 0.02/$-$0.03 & $-$0.03/0.01 \\
$[$Si/Fe$]$ & 0.04/$-$0.06 & $-$0.05/0.04 & $-$0.05/0.03 \\
$[$Ca/Fe$]$ & $-$0.08/0.07 & 0.07/$-$0.07 &0.01/$-$0.01 \\
$[$\ion{Ti}{i}/Fe$]$ & $-$0.07/0.06 & 0.03/$-$0.04 & 0.04/$-$0.06 \\
$[$\ion{Ti}{ii}/Fe$]$ & 0.02/$-$0.03 & $-$0.08/0.06 & $-$0.03/0.02 \\
$[$Ni/Fe$]$ & 0.00/$-$0.01 & $-$0.02/0.02 & $-$0.01/0.01 \\
\hline \\
DM07 & $T_{\rm eff}$=5050 K & $\log g=4.2$ & $\xi=2.1$ km/s\\
\hline
$\Delta$ & $\Delta T_{\rm eff}=-/+60$ K & $\Delta \log g=-/+0.2$ & $\Delta \xi=-/+0.2$ km/s\\
\hline
$[$\ion{Fe}{i}/H$]$ &$-$0.01/0.02 & $-$0.02/0.02 & 0.06/$-$0.05 \\
$[$\ion{Fe}{ii}/H$]$& 0.05/$-$0.04 & $-$0.14/0.10 & 0.03/$-$0.02 \\
$[$Na/Fe$]$ & $-$0.03/0.01 & 0.03/$-$0.05 & $-$0.05/0.03\\
$[$Al/Fe$]$ & $-$0.02/0.01 & 0.03/$-$0.03 & $-$0.03/0.03\\
$[$Si/Fe$]$ & 0.02/$-$0.04 & $-$0.04/0.00 & $-$0.05/0.03\\
$[$Ca/Fe$]$ & $-$0.03/0.02 & 0.04/$-$0.04& $-$0.02/0.01\\
$[$\ion{Ti}{i}/Fe$]$ & $-$0.06/0.04 & 0.02/$-$0.02 & 0.01/$-$0.01\\
$[$\ion{Ti}{ii}/Fe$]$ & 0.03/$-$0.03 & $-$0.08/0.06 & $-$0.05/0.04\\
$[$Ni/Fe$]$ & 0.01/$-$0.02 & $-$0.03/0.01 & $-$0.03/0.02\\
\hline \\
\end{tabular}
\end{center}
\end{table*}
\subsection{Elemental abundances}
We measured the abundances of iron, sodium, aluminum, silicon, calcium, titanium, and nickel. Our final abundances are listed
in Table~\ref{tab:param_chem}.
In Fig.~\ref{fig:elements_abund} we show the [Fe/H] and [X/Fe] ratios as a function of $T_{\rm eff}^{\rm SPEC}$ for
25~Ori and $\lambda$~Ori. The figure shows that there is no star-to-star variation in all elements
in both groups with the exceptions of the elements where NLTE effects are present, namely Na, Al, Ca,
and Ti. In particular, due to their young age, cluster stars are characterized by high levels of chromospheric
activity and are more affected by NLTE over-ionization, which cause a decreasing trend towards lower temperatures of
neutral species of elements with low ionization potential (see \citealt{dorazirandich09, biazzoetal2011}, and
references therein, for thorough discussions of this issue).
For 25~Ori, the mean iron abundance is $<$[Fe/H]$>=-0.05\pm 0.05$, while for $\lambda$~Ori we find $<$[Fe/H]$>$$=0.01\pm 0.01$.
Fig.~\ref{fig:elements_abund} also shows that stars of 25~Ori and $\lambda$~Ori are characterized by homogeneous, close-to-solar
elemental abundances.
\begin{figure*}
\begin{center}
\begin{tabular}{c}
\resizebox{\hsize}{!}{\includegraphics{elements_abund_25Ori_lambdaOri.ps}}\\
\end{tabular}
\vspace{-1.5cm}
\caption{[X/Fe] versus $T_{\rm eff}^{\rm SPEC}$ for our 25~Ori (squares) and $\lambda$~Ori (triangles) members.
Mean abundances values for elements affected by NLTE effects were computed considering the stars with
$T_{\rm eff}^{\rm SPEC} > 4400$ K, while $\pm1\sigma$ dispersions are shown by dash-dotted lines for 25~Ori
and dotted lines for $\lambda$~Ori. The dashed and dash-dotted lines show the $\pm1\sigma$ values from
the mean 25~Ori and $\lambda$~Ori abundances, respectively. The horizontal error bar in all plots represents the
typical uncertainty in $T_{\rm eff}^{\rm SPEC}$.}
\label{fig:elements_abund}
\end{center}
\end{figure*}
\section{Discussion}
\label{sec:discussion}
\subsection{Metallicity distribution in Orion}
\label{sec:met_distr}
In Fig.~\ref{fig:orion_iron_distr} {and in Table~\ref{tab:sfr_iron_abun}} we show the [Fe/H] distribution of the Orion subgroups,
namely the Orion Nebula Cluster (ONC), OB1b, OB1a (\citealt{biazzoetal2011}), $\sigma$~Ori
(\citealt{gonzalez-hernandez2008}), $\lambda$~Ori, and 25~Ori (this work). As can be seen, the Orion complex seems to be
slightly inhomogeneous, with the iron abundance ranging from $-$0.13$\pm$0.03 (the ONC) to 0.01$\pm$0.01 ($\lambda$~Ori), and
an internal dispersion of $\sim 0.05$ dex. Interestingly, while $\lambda$~Ori has an iron abundance identical to solar,
all the Orion subgroups are below solar. It is also remarkable that the youngest object, the ONC, has the lowest
value at all. These results seem to exclude a direct and significant amount of contamination between neighboring (or adjacent)
regions, or an increase of iron abundance with age, as na\"{\i}vely expected in a SN-driven abundance enriched scenario
(see \citealt{biazzoetal2011}, and references therein). This is also supported by the velocity dispersion observed in the
Orion subgroups, which is around 0.5--2.7 km/s, or $\sim$0.5--2 pc/Myr (see Table~\ref{tab:sfr_iron_abun}). This means
that the members of the subgroups would have moved at most (if on ballistic orbits) by only a few pc for the ONC
and up to $\sim$10 pc for the oldest ones. Excluding systematic biases in the analysis, this small dispersion in Orion may be
the result of several effects. $i)$ Different and independent episodes of star formation between $\lambda$~Ori and
the other Orion subgroups analyzed. The former formed $\sim 8-10$ Myr ago from a supernova explosion close to the
present $\lambda^1$~Ori position (\citealt{dolanmathieu1999, dolanmathieu2001}), while the other regions are the result of
sequential star formation triggered by supernova type-II events in the OB1a association (\citealt{preibzinne2006}).
$ii)$ Large-scale formation processes on $\sim 1$ kpc scale may lead to chemically inhomogenous gas inside the giant
molecular cloud (\citealt{elmegreen1998}). A quick onset of star formation in clumps of the molecular clouds born under
turbulent conditions implies that any non-uniformity of abundances may enter the cloud, from an initial non-uniform
pre-cloud gas or from the background Galactic gradient. In this case there is not enough time for uniform mixing before star
formation, and the range of cloud metallicity inhomogeneity may reach a level of $\pm 0.05$ dex. Even if the spatial scales
involved in the formation of the Orion complex are most likely smaller (\citealt{hartmannburkert2007}), the presence
of inhomogeneous gas in the interstellar medium could contribute to the observed small dispersions.
These results are encouraging, because for many years our knowledge about the chemical composition in the Orion complex
from early-B main sequence stars (\citealt{cunhalamb92, cunhalamb94}) and late-type stars (\citealt{cunhaetal98, dorazietal09})
was characterized by highly inhomogeneous abundances, with group-to-group differences of $\sim 0.40$
dex in oxygen, $\sim 0.30-0.40$ dex in silicon, and $\sim 0.10$ dex in iron. These large spreads have been interpreted as
the chemical signature of self-enrichment inside the Orion OB association by ejecta of type-II SNe from massive
stars. Our results change this picture. Also, the new revision of the abundances of B-type stars in Orion
OB1a, b, c, d performed by \cite{simondiaz2010} and the good agreement between our own and the \cite{santosetal2008}
values for low-mass stars indicate a very low dispersion in the abundances of oxygen, silicon, nickel, and iron,
with the Orion Nebula Cluster representing the most metal-poor subgroup. Note that, as pointed out by \cite{simondiaz2010}
and \cite{biazzoetal2011}, the small group-to-group elemental
abundance dispersions are smaller than the internal errors in measurements.
Why is the ONC so significantly different from the other regions is an interesting question that should be addressed
in the future. The cluster is located at the tip of a loose molecular cloud that contains an extended population
of embedded and optically visible low-mass stars (\citealt{megeathetal2005}). Their properties have been recently
analyzed by \cite{fangetal2009}, who provide useful candidates to be observed for abundance measurements. Then, one
can verify if the ONC is just an anomaly or if it shows the same abundance of the present Giant Molecular Cloud. Similarly,
the situation in $\lambda$~Ori is favorable to study a possible link between what we find close to the $\lambda^1$~Ori
star and the chemical abundances of the current generation of young stars located in the vicinity of the molecular
gas clumps shown in Fig.~\ref{fig:lambdaOri_map}. For example, \cite{dolanmathieu2001} provide the properties
of the CTTS/WTTS in these regions that can then be observed for metallicity measurements.
\begin{figure}
\begin{center}
\includegraphics[width=9cm]{feh_spatial_distr_Orion_bis.ps}
\caption{Iron abundance distribution of members of the Orion complex (ONC, OB1b, OB1a, $\sigma$~Ori,
25~Ori, and $\lambda$~Ori) taken from the most recent literature (\citealt{gonzalez-hernandez2008, biazzoetal2011})
and this work. Symbols with different size denote three different metallicity bins: [Fe/H]$\le-0.10$,
$-0.09\le$[Fe/H]$\le0.00$, and [Fe/H]$> 0.00$, from the smallest to the largest size. The central bin
corresponds to the average metallicity of the Orion complex $\pm 1\sigma$. The mean metallicity of each
subgroup is written in parenthesis. The circle outlines the \cite{bricenoetal05} boundary of the
Orion OB subgroups.}
\label{fig:orion_iron_distr}
\end{center}
\end{figure}
\begin{table*}
\caption{Mean iron abundance of low-mass stars, velocity dispersion, and age range of the Orion subgroups.}
\label{tab:sfr_iron_abun}
\tiny
\begin{center}
\begin{tabular}{crlr|cl|cl}
\hline
\hline
SFR & [Fe/H] & Reference & \# stars & $\delta V_{\rm rad}$ & Reference & Age & Reference\\
& & & & (km/s) & & (Myr) & \\
\hline
$\sigma$~Ori & $-$0.02$\pm$0.09 & \cite{gonzalez-hernandez2008} & 8 & 0.9 & \cite{saccoetal2008} & 2--3 & \cite{walteretal2008}\\
ONC & $-$0.13$\pm$0.03 & \cite{biazzoetal2011} & 10 & 3.1 & \cite{fureszetal2008}& 2--3 & \cite{darioetal2010}\\
OB1b & $-$0.05$\pm$0.05 & \cite{biazzoetal2011} & 4 & 1.9 & \cite{bricenoetal07} & 4--6 & \cite{bricenoetal07}\\
OB1a/25~Ori & $-$0.08$\pm$0.15 & \cite{biazzoetal2011} & 1 & $\sim$2 & \cite{bricenoetal07} & 7--10& \cite{bricenoetal05} \\
25~Ori & $-$0.05$\pm$0.05 & This work & 5 & 1.7 & \cite{bricenoetal07} & 7--10 & \cite{bricenoetal05}\\
$\lambda$~Ori & 0.01$\pm$0.01 & This work & 5 & 0.5 & \cite{saccoetal2008} & 5--10 & \cite{dolanmathieu2002}\\
\hline
\end{tabular}
\end{center}
\end{table*}
\subsection{SFR abundances in the Galactic disk}
Each component of our Galaxy (bulge, halo, thin and thick disk) presents a characteristic abundance pattern of elements (such
as iron-peak and $\alpha$-elements) with respect to iron. The observed differences reflect a variety of star formation
histories. One would expect that SFRs show a similar pattern as thin disk stars, unless the gas from which they have
formed has undergone a peculiar enrichment. In order to check if this is indeed true, we compare [X/Fe] ratios versus
[Fe/H] trends for SFRs (\citealt{kingetal2000, santosetal2008, gonzalez-hernandez2008, biazzoetal2011, dorazietal2011})
with studies of nearby field stars of the thin Galactic disk (\citealt{soubirangirard2005}).
Fig.~\ref{fig:alphaFe_FeH} shows an overall good agreement between the [X/Fe] versus [Fe/H] distribution of SFRs and
nearby field stars. A more detailed discussion on the different groups of elements is provided below:
\paragraph{The $\alpha$ elements Si, Ca, and Ti}
~\\
The $\alpha$ elements are mostly produced in the aftermath of explosions of type II supernovae, with small contribution from
type Ia SNe. On average, [Si/Fe] in SFRs seems to be slightly higher than the Sun (Fig.~\ref{fig:alphaFe_FeH}). Given the
dispersion in Si abundance of the SFRs, it is fully consistent with the distribution of the field stars of the thin disk.
The behaviour of [Ca/Fe] and [Ti/Fe] ratios versus [Fe/H] is similar to those reported by \cite{soubirangirard2005} for
field stars with $-0.2\protect\raisebox{-0.5ex}{$\:\stackrel{\textstyle <$[Fe/H]$\ltsim0.2$.
The slight overabundance in [$\alpha$/Fe] of the SFRs might be the result of the contributions of recent local enrichment
due to SNIIe ejecta (occurred in particular in the Orion complex) or of local accretion from enriched gas pre-existing in
the nearby thin disk (\citealt{elmegreen1998}). In particular, the slight overabundance in Si of SFRs might reflect
its larger production during SNII events compared to the calcium and titanium elements (\citealt{woosleyweaver1995}).
\paragraph{The iron-peak element nickel}
~\\
Most of the iron-peak elements are synthesized by SNIa explosions. The nickel abundance is very close to the solar value,
as found for field stars with [Fe/H] between $-0.4$ and 0.0 (Fig.~\ref{fig:alphaFe_FeH}).
\paragraph{Other elements: Na and Al}
~\\
Sodium and aluminum are thought to be mostly a product of Ne and Mg burning in massive stars, through the NeNa and MgAl
chains. The [Na/Fe] ratio does not show a trend with [Fe/H], with a mean value close to that of field stars at the same [Fe/H].
Similarly, the Al abundance does not show trend with [Fe/H], but at slightly above-solar values, as for field stars
in the solar neighborhood. Its phenomenological behaviour is very similar to the $\alpha$-elements, as indicated by
\cite{McWilliam1997}.
\begin{figure*}
\begin{center}
\begin{tabular}{c}
\includegraphics[width=16cm]{alphaFe_FeH_aver_SFR.ps}
\end{tabular}
\caption{[X/Fe] versus [Fe/H] for SFRs with known high-resolution abundances. The references are: Biazzo et al. (2011; B11),
D'Orazi et al. (2011; D11), Santos et al. (2008; S08), King et al. (2000; K00), and Gonz\'alez-Hern\'andez et al. (2008;
GH08). The errors in [Fe/H] are around 0.02--0.12 dex. The dashed lines mark the
solar abundances. The filled dots in the background represent the [X/Fe] distribution of nearby field stars in the
Galactic thin disk (\citealt{soubirangirard2005}).}
\label{fig:alphaFe_FeH}
\end{center}
\end{figure*}
\section{Conclusions}
\label{sec:conclusion}
In this paper we have derived homogeneous and accurate abundances of Fe, Na, Al, Si, Ti, Ca, and Ni of two
very young clusters, namely 25~Orionis and $\lambda$~Orionis. Our main results can be summarized as follows:
\begin{itemize}
\item Stellar properties have been derived for 14 low-mass stars over a relatively wide temperature interval.
Four new PMS candidates show kinematics consistent with membership in the 25~Ori cluster.
\item $\lambda$~Ori and 25~Ori have mean iron abundances of $0.01\pm0.01$ and $-0.05\pm0.05$, respectively, without the presence
of metal-rich stars.
\item Whereas this should be confirmed based on larger number statistics, our results suggest that no
star-to-star variation is present for all elements, with a high degree of homogeneity inside each stellar group.
\item Elemental abundances of $\lambda$~Ori, 25~Ori, and other known SFRs agree with those of the Galactic thin disk.
\item The Orion complex seems to be characterized by a small group-to-group dispersion, consistent with
a different star formation history of each subgroup and an initially inhomogeneous interstellar gas.
\item Finally, we note that the close-to-solar metallicity of both 25~Ori and $\lambda$~Ori reinforces the conclusion that none of
the SFRs in the solar neighborhood is metal-rich and that metal-rich stars hosting giant planets are likely migrated from the
inner part of the Galactic disk to their current location (\citealt{santosetal2008, haywood2008}).
\end{itemize}
\begin{acknowledgements}
The authors are very grateful to the referee for a careful reading of the paper.
KB thanks the INAF--Arcetri Astrophysical Observatory for financial support. Thomas Dame kindly provided the CO maps. This research has made
use of the SIMBAD database, operated at CDS (Strasbourg, France).
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\label{sect:Introduction}
There has been much recent interest in the so-called $f(T)$ gravity theory
as an alternative to dark energy for explaining the acceleration of the
universe \cite{Bengochea:2009, Yu:2010a, Myrzakulov:2010a, Tsyba:2010,
Linder:2010, Yi:2010b, Kazuharu:2010a, Kazuharu:2010b, Myrzakulov:2010b,
Yu:2010c, Karami:2010, Dent:2010, Li:2010, Zheng:2010, Dent:2011,
Sotiriou:2010, Zhang:2011}. This theory is a generalisation of the
teleparallel gravity \cite{Einstein, Aldrovandi} created by replacing $T$,
the lagrangian of teleparallel gravity, by a function $f(T$). It uses the
curvature-free Weitzenbock connection \cite{Weitzenbock} to define the
covariant derivative instead of the conventional torsionless Levi-Civita
connection of general relativity
Teleparallel gravity (see Ref.~\cite{Aldrovandi} for a review and references
therein) has a set of four tetrad (or vierbein) fields which form the
orthogonal bases for the tangent space at each point of spacetime. They are
the dynamical variables and play the role of the metric tensor field in
general relativity. The vierbeins are parallel vector fields, which gave the
theory the descriptor ``teleparallel". It is dynamically equivalent to
general relativity and so is not really an alternative to it, but a
reformulation which allows for a different interpretation: gravity is not
due to curvature, but to torsion.
The generalisation to $f(T)$ gravity can be viewed as an phenomenological
extension of teleparallel gravity (which is the special case $f(T)=T$),
inspired by the $f(R)$ generalization (see Ref.~\cite{Sotiriou:2008rp} for a
review) of general relativity. However, it has the advantage over $f(R)$
gravity that its field equations are second-order instead of fourth-order
(although it is know that, even though it leads to fourth-order equations,
f(R)$ gravity can be ghost free). Yet, it also possesses disadvantages.
Although $f(R)$ gravity is probably not the low-energy limit of some
fundamental theory, it does include models that can be motivated by
effective field theory. In contrast, $f(T)$ gravity seems at this stage to
be just an \emph{ad hoc} generalization.
Another serious disadvantage of $f(T)$ gravity that was pointed out very
recently in Ref.~\cite{Li:2010,Sotiriou:2010} is that it does not respect
local Lorentz symmetry. From a theoretical perspective this is a rather
undesirable feature and experimentally there are stringent constraints. A
Lorentz-violating theory is only attractive if the violations are small
enough to avoid detection andit leads to some other significant
pay-off. So far, the only such pay-off that has been suggested is that $f(T)$
gravity might provide an alternative to conventional dark energy in general
relativistic cosmology.
The specific models that have been considered in the literature \cit
{Bengochea:2009, Yu:2010a, Myrzakulov:2010a, Tsyba:2010, Linder:2010,
Yi:2010b, Kazuharu:2010a, Kazuharu:2010b, Myrzakulov:2010b, Yu:2010c,
Karami:2010, Dent:2010} are rather special as they are tailored to
reproduce the late-time accelerated expansion of the universe without a
cosmological constant. However, to do so, a parameter in these models is
required to be tuned to a very low value, comparable with the observed value
of the cosmological constant. Thus, given the lack of clear theoretical
motivation for these models, it is rather questionable if this can really be
considered to be a resolution of the cosmological constant problem.
Nonetheless, given the attention these models have attracted recently, it
seems worthwhile to address their viability as alternatives to general
relativity in the field of cosmology itself, which was their initial
motivation. We do so by going beyond background cosmology and considering
linear perturbations. We will show that these models behave very differently
from the $\Lambda $CDM model on large scales, and are, therefore, very
unlikely to be suitable alternatives to it.
Cosmological perturbations in these models have been considered recently in
Refs.~\cite{Dent:2010, Zheng:2010, Dent:2011} as well. However, in these papers the field equations are written with only
partial derivatives and look quite different from the usual Einstein
equations. Here, we will present a covariant version of the field equations of $f(T)$ gravity with a clear correspondence to the Einstein's equations. We will use them to derive the field
equations for a perturbed Friedman--Lema\^itre--Robertson--Walker (FLRW) universe. As $f(T)$ gravity has different
dynamical variables (the tetrad fields) from general relativity
or $f(R)$ gravity (rank-2 tensorial metric field), we end up with
new degree(s) of freedom. This fact has been neglected before, for example,
in Refs.~\cite{Dent:2010, Dent:2011}. Here we will show how the new degree of freedom
arise in the perturbed field equations, and numerically assess its effects on the linear-perturbation
observables, such as the CMB and matter power spectra.
The paper is arranged as follows: in Sect.~\ref{sect:Equations} we derive
the field equations for $f(T)$ gravity in its original form and rewrite
it in the covariant form. In Sect.~\ref{sect:pert} we give a detailed
introduction to the method of deriving the covariant and gauge-invariant
linear perturbation equations for $f(T)$ gravity theory, and list those
equations. We focus on a specific model with a power-law functional form for
$f(T)$ in Sect.~\ref{sect:num}, and study its background cosmology and the
growth of large-scale structure. We summarise and conclude in Sect.~\re
{sect:con}. Throughout this work we use the metric convention $(+,-,-,-)$
and set $c=1$ and $\kappa =8\pi G$, where $G$ is the gravitational constant.
\section{The $f(T)$ Model and its Equations}
\label{sect:Equations}
In this section we give a brief description of the $f(T)$ model and a
detailed derivation of its field equations. In contrast to previous works,
which wrote these equations in terms of partial derivatives of the tetrads, we
shall do this by expressing them in terms of the Einstein tensor plus
relevant covariant derivatives of the vierbein field. This approach makes
the equations closely resemble their counterparts in GR and provides a basis
for the derivation of the perturbation equations in the and gauge covariant
formalism, which is our final goal.
Since $f(T)$ gravity is a simple generalisation of teleparallel gravity
theory, we shall briefly introduce the latter (for a comprehensive review
see \cite{Aldrovandi}).
\subsection{Ingredients of Teleparallel Gravity}
\label{subsect:model}
In teleparallel gravity we have the vierbein, or tetrad, fields, $\mathbf{h
_{a}\left( x^{\mu }\right) $, as our dynamical variables; Latin indices
a,b,\cdots $ run from $0$ to $3$ and label tangent-space coordinates; Greek
indices $\mu ,\nu ,\cdots $ run from $0$ to $3$ and label spacetime
coordinates. The $h_{a}^{\mu }$ are both spacetime vectors and Lorentz
vectors in the tangent space. As the former (indexed by $\mu $), they are
the dynamical fields of gravitation, as the latter (indexed by $a$), they
form an orthonormal basis for the tangent space at each spacetime point.
The metric tensor of spacetime, $g_{\mu\nu}$, is given by
\begin{eqnarray}\label{eq:metric}
g_{\mu\nu} = \eta_{ab}h^a_\mu h^b_\nu
\end{eqnarray}
where $\eta_{ab}={\rm diag}(1,-1,-1,-1)$ is the Minkowski metric
for the tangent space. From this relation it follows that
\begin{eqnarray}
h_a^\mu h_\nu^a\ =\ \delta^\mu_\nu, \ \ \ h_a^\mu h^b_\mu\ =\ \delta^b_a,
\end{eqnarray}
where Einstein convention of summation has been used.
Eq.~(\ref{eq:metric}) implies that in this model
$g_{\mu\nu}$, $h_a^\mu$ and $h_\mu^a$ are all dependent on each
other, which is important for the derivation of
the field equations by variation.
General relativity is built on the
Levi-Civita connection of the metric
\begin{eqnarray}\label{eq:levi-civita}
\Gamma^\alpha_{\beta\gamma} &\equiv& \frac{1}{2}g^{\alpha\lambda}\left(g_{\lambda\beta,\gamma} + g_{\lambda\gamma,\beta} - g_{\beta\gamma,\lambda}\right),
\end{eqnarray}
where a comma is used to denote a partial derivative
($_{,\mu}\equiv\partial/\partial x^\mu$). This connection has
nonzero curvature but zero torsion. Teleparallel
gravity, or the teleparallel interpretation of general relativity, instead makes use of the Weitzenbock connection (tilded to
distinguish from $\Gamma^\alpha_{\beta\gamma}$)
\begin{eqnarray}\label{eq:weizenbock}
\tilde{\Gamma}^\alpha_{\ \beta\gamma} &\equiv& h^\alpha_b\partial_\gamma h^b_\beta\ =\ -h^b_\beta\partial_\gamma h^\alpha_b
\end{eqnarray}
which has a zero curvature but nonzero torsion. The
torsion tensor reads
\begin{eqnarray}\label{eq:torsion}
T^\alpha_{\ \beta\gamma} &\equiv& \tilde{\Gamma}^\alpha_{\ \gamma\beta} - \tilde{\Gamma}^\alpha_{\ \beta\gamma}\
=\ h_b^\alpha\left(\partial_\beta h_\gamma^b - \partial_\gamma h^b_\beta\right).
\end{eqnarray}
The difference between the Levi-Civita and Weitzenbock
connections, neither of which is a spacetime tensor, is a spacetime tensor, and is
known as the contorsion tensor:
\begin{eqnarray}\label{eq:contorsion}
K^\rho_{\mu\nu} &\equiv& \tilde{\Gamma}^\rho_{\ \mu\nu} - \Gamma^\rho_{\mu\nu}\
=\ \frac{1}{2}\left(T_{\mu\ \nu}^{\ \rho} + T_{\nu\ \mu}^{\ \rho} - T^{\rho}_{\ \mu\nu}\right).
\end{eqnarray}
It is worth pointing out at this point that, based on the definition we have given above, the Weitzenbock connection, the torsion tensor and the contorsion tensor are not local Lorentz scalars ({\em i.e.}~they do not remain invariant under a local Lorentz transformation in tangent space) even though they do not have any tangent space indices.\footnote{Teleparallelism assumes the existence of a class of frames where the spin connection is zero and in which the Weitzenbock connection assume the form given in Eq.~(\ref{eq:weizenbock}). We choose to work in one of these frame. One could introduce a Lorentz covariant formulation of the theory at the level of the action, but this would only change appearances \cite{Sotiriou:2010}.} This is the root of the lack of Lorentz invariance in generalized teleparallel theories of gravity.
The Lagrangian density of teleparallel gravity is given by
\begin{eqnarray}\label{eq:T}
\mathcal{L}_T &\equiv& \frac{h}{16\pi G}T\,,
\end{eqnarray}
where
\begin{eqnarray}
T = \frac{1}{4}T^{\rho\mu\nu}T_{\rho\mu\nu} + \frac{1}{2}T^{\rho\mu\nu}T_{\nu\mu\rho} - T_{\rho\mu}^{\ \ \rho}T^{\nu\mu}_{\ \ \ \nu},
\end{eqnarray}
and $h\equiv\sqrt{-g}$ is the determinant of $h_a^{\alpha}$
with $g$ being the determinant of the metric $g_{\mu\nu}$. After adding the matted Lagrangian density $\mathcal{L}_m$, variation with respect to the tetrad yields the field equations
\begin{eqnarray}\label{eq:field_eqn_tel}
\partial_\rho \left(hh_a^\nu S_{\nu}^{\ \lambda\rho}\right)&-&hh^\rho_a S^{\mu\nu\lambda}T_{\mu\nu\rho} \nonumber\\
&+& \frac{1}{4}hh^\lambda_a S^{\rho\mu\nu}T_{\rho\mu\nu}=8\pi G\Theta_a^{\lambda}
\end{eqnarray}
with $\Theta_a^{\lambda}$ related to the usual energy-momentum tensor
$\Theta_{\mu\nu}$ by
$\Theta^{\mu\nu}\equiv\eta^{ab}\Theta_{a}^{\nu}h_b^\mu$ and
\begin{eqnarray}\label{eq:S}
S^{\rho\mu\nu} &\equiv& K^{\mu\nu\rho} - g^{\rho\nu}T^{\sigma\mu}_{\ \ \ \sigma} + g^{\rho\mu}T^{\sigma\nu}_{\ \ \ \sigma}.
\end{eqnarray}
\subsection{Field Equation for $f(T)$ Gravity}
\label{subsect:f(T)}
The idea of $f(T)$ gravity is simply to promote the $T$ in the
Lagrangian to become an arbitrary function of $T$:
\begin{eqnarray}
\mathcal{L}_T &\rightarrow& \mathcal{L}\ =\ \frac{h}{16\pi G}f(T).
\end{eqnarray}
The field equations are straightforward generalizations of those of standard teleparallel gravity just given above:
\begin{eqnarray}\label{eq:field_eqn}
f_T\left[\partial_\rho \left(hh_a^\nu S_{\nu}^{\
\lambda\rho}\right) - hh^\rho_a
S^{\mu\nu\lambda}T_{\mu\nu\rho}\right] \nonumber\\ +
f_{TT}hh_a^\nu S_{\nu}^{\ \lambda\rho}\partial_\rho T +
\frac{1}{2}hh^\lambda_a f(T) &=& 8\pi G\Theta_a^{\lambda}
\end{eqnarray}
where $f_T\equiv\partial f(T)/\partial T$ and
$f_{TT}\equiv\partial^2f(T)/\partial T^2$.
Obviously, if $f(T)=T+\Lambda$ with $\Lambda$ a constant, then
Eq.~(\ref{eq:field_eqn}) simply reduces to
Eq.~(\ref{eq:field_eqn_tel}).
\subsection{Covariant Version of the Field Equations}
\label{subsect:gr-form}
The field equation Eqs.~(\ref{eq:field_eqn_tel}) and (\ref{eq:field_eqn}) are
written in terms of the tetrad and partial derivatives and appear very different
from Einstein's equations. This makes comparison with general
relativity rather difficult. In this subsection we show that Eq.~(\ref{eq:field_eqn_tel}) can be written in terms of the metric only and it then becomes Einstein's equation. We also present an equation relating $T$ with the Ricci scalar of the metric $R$. These will make the equivalence between teleparallel gravity and general relativity clear.
On the other hand, the tetrad cannot be eliminated completely in favour of the metric in Eq.~(\ref{eq:field_eqn}), because of the lack of local Lorentz symmetry, but we will show that the later can be brought in a form that closely resembles
Einstein's equation. This form is more suitable for introducing the covariant and gauge invariant formalism for cosmological perturbations.
First, let us note that although in
Sect.~\ref{subsect:model} the tensors were all written in terms of
partial derivatives, they could be rearranged so that all the
partial derivatives are replaced with covariant derivatives
compatible with the metric $g_{\mu\nu}$, {\it i.e.}, $\nabla_\alpha$
where $\nabla_\alpha g_{\mu\nu}=0$. In particular, we would have
\begin{eqnarray}
T^{\alpha}_{\ \beta\gamma} &=& h_b^\alpha\left(\partial_\beta h_\gamma^b - \Gamma^\sigma_{\beta\gamma}h^b_\sigma - \partial_\gamma h^b_\beta + \Gamma^\sigma_{\gamma\beta}h^b_\sigma\right)\nonumber\\
&=& h_b^\alpha\left(\nabla_\beta h_\gamma^b - \nabla_\gamma
h^b_\beta\right),
\end{eqnarray}
where we have used the fact that $\Gamma^\alpha_{\beta\gamma}$ is
symmetric in the subscripts $\beta, \gamma$. From this it can be
readily checked that
\begin{eqnarray}\label{eq:K_cov}
K^\beta_{\ \gamma\alpha} &=& h^\beta_a\nabla_\alpha h^a_\gamma,\\
\label{eq:S_cov}
S_\alpha^{\ \beta\gamma} &=& \eta^{ab}h^\beta_a\nabla_\alpha h_b^\gamma + \delta^\gamma_\alpha\eta^{ab}h^\mu_a\nabla_\mu h_b^\beta\nonumber\\&&\qquad - \delta^\beta_\alpha\eta^{ab}h^\mu_a\nabla_\mu h^\gamma_b
\end{eqnarray}
and clearly
\begin{eqnarray}
K^{\alpha\beta\gamma} &=& K^{[\alpha\beta]\gamma},\nonumber\\
T^{\alpha\beta\gamma} &=& T^{\alpha[\beta\gamma]},\nonumber\\
S^{\alpha\beta\gamma} &=& S^{\alpha[\beta\gamma]},\nonumber
\end{eqnarray}
in which the square brackets mean anti-symmetrisation, and also
\begin{eqnarray}
K^{\mu}_{\ \rho\mu}\ =\ T^{\mu}_{\ \mu\rho}, \ \ \ K^{\rho\mu}_{\
\ \ \mu}\ =\ T^{\mu\rho}_{\ \ \ \mu}.\nonumber
\end{eqnarray}
These relations are useful in the calculation below.
Next, from the relation between $\Gamma^\alpha_{\beta\gamma}$ and
$\tilde{\Gamma}^\alpha_{\beta\gamma}$ as given in
Eq.~(\ref{eq:contorsion}) and the fact that the
curvature tensor associated with the Weitzenbock connection
$\tilde{\Gamma}^\alpha_{\beta\gamma}$ vanishes, we can
write the Riemann tensor for the connection
$\Gamma^\alpha_{\beta\gamma}$ as \cite{Aldrovandi}
\begin{eqnarray}
R^\rho_{\ \mu\lambda\nu} &=& \partial_\lambda\Gamma^\rho_{\mu\nu} - \partial_\nu\Gamma^{\rho}_{\mu\lambda} + \Gamma^{\rho}_{\sigma\lambda}\Gamma^\sigma_{\mu\nu} - \Gamma^\rho_{\sigma\nu}\Gamma^\sigma_{\mu\lambda}\nonumber\\
&=& \nabla_\nu K^\rho_{\ \mu\lambda} - \nabla_\lambda K^\rho_{\
\mu\nu} + K^\rho_{\ \sigma\nu}K^\sigma_{\ \mu\lambda} -
K^{\rho}_{\ \sigma\lambda}K^\sigma_{\ \mu\nu}\nonumber
\end{eqnarray}
and the corresponding Ricci tensor and Ricci scalar are
\begin{eqnarray}\label{eq:Ricci}
R^\rho_{\ \lambda}
&=& \nabla_\mu K^{\rho\mu}_{\ \ \ \lambda} - \nabla_\lambda K^{\rho\mu}_{\ \ \ \mu} + K^\rho_{\ \sigma\mu}K^{\sigma\mu}_{\ \ \ \lambda} - K^{\rho}_{\ \sigma\lambda}K^{\sigma\mu}_{\ \ \ \mu}\nonumber\\
\label{eq:Ricci2}R &=& K^{\mu\rho}_{\ \ \ \mu}K^\nu_{\ \rho\nu} - K^{\mu\nu\rho}K_{\rho\nu\mu} - 2\nabla^\mu\left(T^\nu_{\ \mu\nu}\right)\nonumber\\
&=& -T - 2\nabla^\mu\left(T^\nu_{\ \mu\nu}\right).
\end{eqnarray}
This last equation implies that the $T$ and $R$ differ only by a covariant divergence of a spacetime vector. Therefore, the Einstein-Hilbert action and the teleparallel action ({\em i.e.}~the action constucted with the Lagrangian density given in Eq.~(\ref{eq:T})) will both lead to the same field equations and are dynamically equivalent theories.
We can show this equivalence directly at the level of the field equations. With the aid of the equations listed above, it can be shown, after some
algebraic manipulations, that
\begin{equation}\label{eq:einstein-tensor}
hh_a^\rho G^\lambda_{\ \rho}
= \partial_\xi\left(hh_a^\rho S_\rho^{\ \lambda\xi}\right) - hh_a^\xi S^{\mu\nu\lambda}T_{\mu\nu\xi}+ \frac{1}{2}hh_a^\lambda T,
\end{equation}
where $G_{\mu\nu}$ is the Einstein tensor.
Substituting this equation into Eq.~(\ref{eq:field_eqn_tel}) and rearranging, we
obtain Einstein's equations. If we do the same for
Eq.~(\ref{eq:field_eqn}) we get
\begin{eqnarray}\label{eq:modified_einstein_eqn}
f_TG_{\mu\nu} &+& \frac{1}{2}g_{\mu\nu}\left[f(T)-f_TT\right] \nonumber\\&&\qquad+
f_{TT}S_{\nu\mu\rho}\nabla^\rho T = 8\pi G\Theta_{\mu\nu}.
\end{eqnarray}
Eq.~(\ref{eq:modified_einstein_eqn}) can be taken as the
starting point of the $f(T)$ modified gravity model, and it has a
structure similar to the field equation of the $f(R)$ gravity.
Note that when $f(T)=T$ general relativity is exactly recovered, as expected.
\section{The Linear Perturbation Equations}
\label{sect:pert}
In order to study the evolution of linear perturbations in the $f(T)$
gravity, we have to linearise the field equations. Usually, this is achieved
by writing all quantities in terms of the metric perturbation variables, and
for this we have to use some metric ansatz. In $f(T)$ gravity, however, it
is the vierbein, rather than the metric, that is the fundamental field, and
it has 16 rather than 10 independent components. Usually, the six additional
components correspond to local Lorentz symmetry, but, as mentioned
previously $f(T)$ gravity is not invariant under that symmetry.
Consequently, specifying a metric ansatz does not necessarily fix all the
tetrad components \cite{Li:2010}, and one needs to specify an ansatz for the
tetrad itself and derive the metric perturbations thereafter.
Other subtleties emerge here. For example, one cannot write the metric in
some familiar gauges ({\em e.g.}~conformal Newtonian) \textit{a priori}, as these
gauges are obtained by gauging away certain degrees of freedom in the metric
fields. However, in $f(T)$ gravity the degrees of freedom are different and
the lack of local Lorentz invariance means that we can only gauge away 4 of
the 16 components of the tetrad due to the invariance of the action under
spacetime coordinate transformations.
We follow Ref.~\cite{GR3+1} and derive the perturbation equations in the $3+1
$ formalism, in which all the quantities are covariant and gauge invariant.
This formalism deals with physical quantities directly and does not need to
make a metric ansatz \textit{a priori}. It is appropriate to use it given
that we have derived the field equation Eq.~(\ref{eq:modified_einstein_eqn})
in the covariant form. It has proved quite useful in studies of perturbation
evolution in modified gravity theories \cite{Li:pfr-1, Li:pfr-2, Li:mfr,
Li:mgb, Li:aether}.
\subsection{Covariant and Gauge Invariant Perturbation Equations in General Relativity}
\label{subsect:GR3+1}
The $3+1$ decomposition makes spacetime splits of physical quantities with respect to the 4-velocity $u^{\alpha}$ of an observer. The projection tensor $H_{\alpha\beta}$ is defined as $H_{\alpha\beta}=g_{\alpha\beta}-u_{\alpha}u_{\beta}$ and can be used to obtain covariant tensors perpendicular to $u$. For example, the covariant spatial derivative $\hat{\nabla}$ of a tensor field $T_{\mu\cdot\cdot\cdot\nu}^{\beta\cdot\cdot\cdot\gamma}$ is defined as
\begin{eqnarray}\label{eq:spatial_deriv}
\hat{\nabla}^{\alpha}T_{\mu\cdot \cdot \cdot\nu}^{\beta\cdot \cdot \cdot \gamma}\equiv
H_{\rho}^{\alpha}H_{\sigma}^{\beta}\cdot \cdot \cdot \ H_{\delta}^{\gamma}H_{\mu}^{\lambda}\cdot \cdot \cdot \
H_{\nu}^{\xi}\nabla ^{\rho}T_{\lambda\cdot \cdot \cdot \xi}^{\sigma\cdot \cdot \cdot \delta}.
\end{eqnarray}
The energy-momentum tensor and covariant derivative of the 4-velocity are
decomposed respectively as
\begin{eqnarray}\label{eq:decomp1}
\Theta_{\alpha\beta} &=& \pi_{\alpha\beta}+2q_{(\alpha}u_{\beta)}+\rho u_{\alpha}u_{\beta}-pH_{\alpha\beta},\\
\label{eq:decomp2}\nabla _{\alpha}u_{\beta} &=&
\sigma_{\alpha\beta}+\varpi_{\alpha\beta}+\frac{1}{3}\theta
H_{\alpha\beta}+u_{\alpha}A_{\beta}.
\end{eqnarray}
In the above expressions, $\pi_{\alpha\beta}$ is the projected symmetric trace-free (PSTF)
anisotropic stress, $q_{\alpha}$ the heat flux vector, $p$ the isotropic
pressure, $\sigma _{\alpha\beta}$ the PSTF shear tensor, $\varpi _{\alpha\beta}=\hat{\nabla}_{[\alpha}u_{\beta]}$ the vorticity, $\theta = \nabla^{\alpha}u_{\alpha} = 3\dot{a}/a$ ($a$ is the mean expansion scale factor) the expansion scalar, and $A_{\alpha}=\dot{u}_{\alpha}$ the acceleration; the overdot denotes time derivative expressed as $\dot{\phi}=u^{\alpha}\nabla_{\alpha}\phi $, brackets mean antisymmetrisation, and parentheses symmetrization. The 4-velocity normalization is chosen to be $ u^{\alpha}u_{\alpha}=1$. The quantities $\pi_{\alpha\beta}, q_{\alpha}, \rho, p$ are referred to as \emph{dynamical} quantities and $\sigma_{\alpha\beta}, \varpi_{\alpha\beta}, \theta, A_{\alpha}$ as \emph{kinematical} quantities. Note that the dynamical quantities can also be obtained from the energy-momentum tensor $\Theta_{\alpha\beta}$ through the relations
\begin{eqnarray}\label{eq:dynamic_quantities}
\rho &=& \Theta_{\alpha\beta}u^{\alpha}u^{\beta},\nonumber\\
p &=& -\frac{1}{3}H^{\alpha\beta}\Theta_{\alpha\beta},\nonumber\\
q_{\alpha} &=& H_{\alpha}^{\mu}u^{\mu}\Theta_{\mu\nu}, \nonumber \\
\pi_{\alpha\beta} &=& H_{\alpha}^{\mu}H_{\beta}^{\nu}\Theta_{\mu\nu}+pH_{\alpha\beta}.
\end{eqnarray}
Decomposing the Riemann tensor and making use the Einstein equations, after linearisation we
obtain five constraint equations \cite{GR3+1}:
\begin{eqnarray}
\label{eq:constraint_varpi} 0 &=& \hat{\nabla}^{\mu}(\varepsilon _{\ \ \mu\nu}^{\alpha\beta}u^{\nu}\varpi_{\alpha\beta});\\
\label{eq:constraint_q} \kappa q_{\alpha} &=& -\frac{2\hat{\nabla}_{\alpha}\theta}{3} + \hat{\nabla}^{\beta}\sigma_{\alpha\beta}+\hat{\nabla}^{\beta}\varpi_{\alpha\beta};\ \ \ \\
\mathcal{B}_{\alpha\beta} &=& \left[ \hat{\nabla}^{\mu}\sigma_{\nu(\alpha}+\hat{\nabla}^{\mu}\varpi_{\nu(\alpha}\right] \varepsilon_{\beta)\rho\mu}^{\ \ \ \ \nu}u^{\rho};\\
\label{eq:constraint_phi} \hat{\nabla}^{\beta}\mathcal{E}_{\alpha\beta} &=& \frac{1}{2}\kappa \left[\hat{\nabla}^{\beta}\pi_{\alpha\beta}+\frac{2}{3}\theta q_{\alpha}+\frac{2}{3}\hat{\nabla}_{\alpha}\rho\right];\\
\hat{\nabla}^{\beta}\mathcal{B}_{\alpha\beta} &=& \frac{1}{2}\kappa\left[\hat{\nabla}_{\mu}q_{\nu}+(\rho+p)\varpi_{\mu\nu}\right]\varepsilon_{\alpha\beta}^{\ \ \mu\nu}u^{\beta},
\end{eqnarray}
and five propagation equations:
\begin{eqnarray}
\label{eq:raychaudhrui} \dot{\theta}+\frac{1}{3}\theta^{2}
-\hat{\nabla}^{a}A_{a}+\frac{\kappa }{2}(\rho +3p) &=& 0; \\
\label{eq:propagation_sigma} \dot{\sigma}_{\alpha\beta}+\frac{2}{3}\theta\sigma_{\alpha\beta}-\hat{\nabla}_{\langle\alpha}A_{\beta\rangle }+\mathcal{E}_{\alpha\beta}+\frac{1}{2}\kappa\pi_{\alpha\beta} &=& 0; \\
\dot{\varpi}+\frac{2}{3}\theta \varpi -\hat{\nabla}_{[\alpha}A_{\beta]} &=& 0; \\
\label{eq:propagation_phi} \frac{1}{2}\kappa\left[\dot{\pi}_{\alpha\beta}+\frac{1}{3}\theta\pi_{\alpha\beta}\right] - \frac{1}{2}\kappa\left[(\rho+p)\sigma_{\alpha\beta}+\hat{\nabla}_{\langle\alpha}q_{\beta\rangle}\right] \nonumber\!\!\!\!\!\!\!\!\!\!\!\!\!\\
-\left[\dot{\mathcal{E}}_{\alpha\beta}+\theta\mathcal{E}_{\alpha\beta}-\hat{\nabla}^{\mu}\mathcal{B}_{\nu(\alpha}\varepsilon_{\beta)\rho\mu}^{\ \ \ \ \nu}u^{\rho}\right] &=& 0;\ \ \ \ \\
\dot{\mathcal{B}}_{\alpha\beta}+\theta\mathcal{B}_{\alpha\beta}+\hat{\nabla}^{\mu}\mathcal{E}_{\nu(\alpha}\varepsilon_{\beta)\rho\mu}^{\ \ \ \ \nu}u^{\rho}\nonumber\\+\frac{\kappa}{2}\hat{\nabla}^{\mu}\mathcal{\pi}_{\nu(\alpha}\varepsilon_{\beta)\rho\mu}^{\ \ \ \ \nu}u^{\rho} &=& 0.
\end{eqnarray}
Here, $\varepsilon_{\alpha\beta\mu\nu}$ is the covariant permutation tensor, $\mathcal{E}_{\alpha\beta}$ and $\mathcal{B}_{\alpha\beta}$ are respectively the electric and magnetic parts of the Weyl tensor $\mathcal{W}_{\alpha\beta\mu\nu}$, defined by $\mathcal{E}_{\alpha\beta}=u^{\mu}u^{\nu}\mathcal{W}_{\alpha\mu\beta\nu}$ and $\mathcal{B}_{\alpha\beta}=-\frac{1}{2}u^{\mu}u^{\nu}\varepsilon_{\alpha\mu}^{\ \ \rho\sigma}\mathcal{W}_{\rho\sigma\beta\nu}$. Note that the angle bracket denotes taking the trace-free part of a quantity.
Using the definition of the projected derivatives, it can be shown that
\begin{eqnarray}
[\hat{\nabla}_\mu,\hat{\nabla}_\nu]v_\rho &=& - 2\varpi_{\mu\nu}\dot{v}_\rho +H^\alpha_\mu H^\beta_\nu H^\gamma_\rho R_{\alpha\beta\gamma}^{\ \ \ \ \lambda}v_\lambda\nonumber\\
&&+\left(\hat{\nabla}_\mu u_\rho\hat{\nabla}_\nu u^\lambda-\hat{\nabla}_\mu u^\lambda\hat{\nabla}_\nu u_\rho\right)v_\lambda
\end{eqnarray}
for any projected vector field $v_\rho$ ($u^\rho v_\rho = 0$). In the absence of vorticity $\varpi_{\mu\nu}$ (which is true up to first order in perturbation because we are considering the scalar mode only), the above equation can be written as
\begin{eqnarray}
[\hat{\nabla}_\mu,\hat{\nabla}_\nu]v_\rho &\equiv& -\hat{R}_{\mu\nu\rho}^{\ \ \ \ \lambda}v_\lambda
\end{eqnarray}
where $\hat{R}_{\mu\nu\rho\lambda}$ is the spatial 3-curvature tensor defined in analogy to the Riemann curvature tensor in the 4D spacetime (the minus sign is conventional). We can then define the corresponding Ricci scalar of the hyperspace perpendicular to the 4-velocity in the usual way: $\hat{R}=\hat{R}_{\mu\nu}^{\ \ \mu\nu}$. With the Einstein equation it is easy to find
\begin{eqnarray}\label{eq:spatial_ricci}
\hat{R} &\approx& 2\kappa\rho - \frac{2}{3}\theta ^{2}.
\end{eqnarray}
The spatial derivative of $\hat{R}$, $\eta _{\alpha} \equiv \frac{1}{2}a\hat{\nabla}_{\alpha}\hat{R}$, is then given as
\begin{eqnarray}\label{eq:constraint_eta}
\eta _{\alpha} &=& \kappa\hat{\nabla}_{\alpha}\rho - \frac{2a}{3}\theta\hat{\nabla}_{\alpha}\theta,
\end{eqnarray}
and its propagation equation by
\begin{eqnarray}\label{eq:propagation_eta}
\dot{\eta}_{\alpha}+\frac{2\theta}{3}\eta_{\alpha} &=& -\frac{2}{3}\theta a\hat{\nabla}_{\alpha}\hat{\nabla}\cdot A - a\kappa\hat{\nabla}_{\alpha}\hat{\nabla}\cdot q.
\end{eqnarray}
Finally, there are the conservation equations for the energy-momentum tensor:
\begin{eqnarray} \label{eq:energy_conservation}
\dot{\rho}+(\rho +p)\theta +\hat{\nabla}^{\alpha}q_{\alpha} &=& 0,\\
\label{eq:heat_flux_evolution} \dot{q}_{\alpha}+\frac{4}{3}\theta q_{\alpha}+(\rho+p)A_{\alpha}-\hat{\nabla}_{\alpha}p+\hat{\nabla}^{\beta}\pi_{\alpha\beta} &=& 0.
\end{eqnarray}
As we are considering a spatially-flat universe,\footnote{See Ref.~\cite{Ferraro:2011us} for a discussion on hyperspherical and hyperbolic universes.} the spatial
curvature must vanish on large scales and so in the background $\hat{R}=0$. Thus,
from Eq.~(\ref{eq:spatial_ricci}), we obtain
\begin{eqnarray}
\frac{1}{3}\theta^{2} &=& \kappa\rho.
\end{eqnarray}
This is the Friedmann equation in general relativity, and the other
background equations can be obtained by taking the zero-order parts of Eqs.~(\ref{eq:raychaudhrui}, \ref{eq:energy_conservation}), yielding:
\begin{eqnarray}
\dot{\theta}+\frac{1}{3}\theta ^{2}+\frac{\kappa }{2}(\rho +3p) &=& 0, \\
\label{eq:background_energy_conservation} \dot{\rho}+(\rho +p)\theta &=& 0.
\end{eqnarray}
\subsection{Generalisation to the $f(T)$ Gravity}
\label{subsect:fT3+1}
In order to make best use of the formulae obtained for general relativity,
we can consider the modifications to the Einstein equation in $f(T)$ gravity
as a new effective energy-momentum tensor $\Theta _{\mu \nu }^{eff}$ in
addition to that of the fluid matter, $\Theta _{\mu \nu }^{f}$. Eq.~(\ref{eq:modified_einstein_eqn}) can then be rewritten as
\begin{eqnarray}
G_{\mu\nu} &=& \kappa\left(\Theta^f_{\mu\nu} + \Theta^{eff}_{\mu\nu}\right)
\end{eqnarray}
in which
\begin{eqnarray}
\kappa\Theta^{eff}_{\mu\nu} &\equiv& -\frac{f_T-1}{f_T}\kappa\Theta^f_{\mu\nu} - \frac{1}{2f_T}g_{\mu\nu}\left[f-f_TT\right]\nonumber\\ &&- \frac{f_{TT}}{f_T}S_{\nu\mu\rho}\nabla^{\rho}T.
\end{eqnarray}
As already mentioned, here we have to work with the tetrad and not just the metric, so the setup will be slightly different than that of general relativity.
Since we intend to investigate the perturbation evolution in an almost
Friedmann universe, let us first consider an exact Friedmann universe: there
is no special spatial direction and the fundamental observer's world line is
in the time direction. Assuming that the comoving observer's frame is
aligned with the frame defined by the tetrad in tangent space we have $h_
\underline{0}}^{\mu }=u^{\mu }$, where $u^{\mu }$ is the 4-velocity of the
fundamental observer, and the $h_{\underline{i}}^{\mu }$ ($\underline{i
=1,2,3$) are three orthonormal vectors in the 3-space of the fundamental
observer (here we use an underline to denote components of the Lorentz
index). If we define $U_{a}\equiv h_{a}^{\mu }u_{\mu }$, then $1=g_{\mu \nu
}u^{\mu }u^{\nu }=\eta _{ab}h_{\mu }^{a}h_{\nu }^{b}u^{\mu }u^{\nu }=\eta
_{ab}U^{a}U^{b}$. Note that in this case $U^{a}=\delta _{\underline{0}}^{a}$.
In an almost Friedmann universe, the above symmetry is at best an
approximation, and $h_{\underline{0}}^{\mu }$ will not coincide exactly with
$u^{\mu }$ but could differ slightly. Instead of $U^{a}=(1,\mathbf{0})$, we
will have $U^{a}=(U^{\underline{0}},U^{\underline{i}})$ where the $U^
\underline{i}}$ are small, and $\eta _{ab}U^{a}U^{b}=1$ implies that $U^
\underline{0}}=1$ up to first order in perturbation. As $U^{\underline{0
}=h_{\mu }^{\underline{0}}u^{\mu }$, we can write $h_{\mu }^{\underline{0
}=u_{\mu }+\ae _{\mu },$ where $\ae _{\mu }$ is a perturbation vector and
u^{\mu }\ae _{\mu }=0$. As it will turn out, all the information we need to
know about $h_{\mu }^{\underline{i}}$ is that $h_{\mu }^{\underline{i
}u^{\mu }=U^{\underline{i}}$ is first order in perturbation and $h_{\mu }^
\underline{i}}$. This suffices to show that, to this order of perturbation,
\ae _{\mu }$ is the only new physical degree of freedom with respect to
general relativity. It was expected to appear due to the lack of local
Lorentz invariance \cite{Li:2010}. Detailed calculations in support of these
statements, as well as explicit derivations of the perturbative expressions
for quantities entering the field equations can be found in the Appendix~\re
{appen:expression}. Here, we will only quote the results of these
calculations. Obviously, the only quantities that are not already present in
general relativity are $T$ and $S_{\nu \mu \rho }\nabla ^{\rho }T$. Up to
first order in perturbations, we have
\begin{eqnarray}
\label{perT}
T &\approx& -\frac{2}{3}\theta^2-\frac{4}{3}\theta\hat{\nabla}\cdot\ae,\\
\label{perS}
S_{\nu\mu\rho}\nabla^{\rho}T &\approx& \frac{2}{3}\dot{T}\left(\theta+\hat{\nabla}\cdot\ae\right)H_{\mu\nu}-\frac{1}{2}\dot{T}u_{\nu}\hat{R}_{\mu}\nonumber\\
&&- \dot{T}\left(\sigma_{\mu\nu}+\varpi_{\mu\nu}+\hat{\nabla}_{\langle\mu}\ae_{\nu\rangle}+\hat{\nabla}_{[\mu}\ae_{\nu]}\right)\nonumber\\
&&-\frac{2}{3}\theta u_\mu\hat{\nabla}_{\nu}T.
\end{eqnarray}
Here $\hat{\nabla}\cdot\ae\equiv\hat{\nabla}^{\mu}\ae_{\mu}$ and $\hat{R}_{\mu}$ satisfies $\hat{\nabla}^\mu\hat{R}_\mu=\hat{R}$.
Using the definitions given in Eq.~(\ref{eq:dynamic_quantities}),
it is straightforward to obtain
\begin{eqnarray}
\label{eq:rho_eff}\kappa\rho^{eff} &\approx&
-\frac{1}{f_T}\left[(f_T-1)\kappa\rho^{f}+\frac{1}{2}\left(f-f_TT\right)\right],
\end{eqnarray}
\begin{eqnarray}
\label{eq:p_eff}\kappa p^{eff} &\approx&
-\frac{1}{f_T}\left[(f_T-1)\kappa p^{f}-\frac{1}{2}\left(f-f_TT\right)\right]\nonumber\\
&&+\frac{2}{3}\frac{1}{f_T}f_{TT}\dot{T}\left(\theta+\hat{\nabla}\cdot\ae\right),
\end{eqnarray}
\begin{eqnarray}
\label{eq:q_eff_1}\kappa q^{eff}_{\alpha} &\approx&
-\frac{1}{f_T}\left[(f_T-1)\kappa q^{f}_{\alpha} -
\frac{1}{2}f_{TT}\dot{T}\hat{R}_{\alpha}\right],\\
\label{eq:q_eff_2}&\approx& -\frac{1}{f_T}\left[(f_T-1)\kappa
q^{f}_{\alpha} -
\frac{2}{3}f_{TT}\theta\hat{\nabla}T\right],\\
\label{eq:pi_eff}\kappa\pi^{eff}_{\alpha\beta} &\approx&
-\frac{1}{f_T}\Big[(f_T-1)\kappa\pi^{f}_{\alpha\beta}-f_{TT}\dot{T}\left(\sigma_{\alpha\beta}
+\hat{\nabla}_{\langle\alpha}\ae_{\beta\rangle}\right)\Big].\nonumber\\
\end{eqnarray}
up to first order in perturbation. There are two different
expressions for $q^{eff}_{\alpha}$, which is because the quantity
$S_{\nu\mu\rho}\nabla^{\rho}T$ is not symmetric {\it a priori}, but the field equations require its antisymmetric part to vanish.
We are also interested in the density and pressure perturbations,
and these can be obtained by differentiating
Eqs.~(\ref{eq:rho_eff}, \ref{eq:p_eff}):
\begin{eqnarray}
\label{eq:delta_rho_eff}\kappa\hat{\nabla}_\alpha\rho^{eff}
&\approx&
\frac{1}{f_T}\left[(1-f_T)\kappa\hat{\nabla}_{\alpha}\rho^{f}+f_{TT}T\hat{\nabla}_{\alpha}T\right],\\
\label{eq:delta_p_eff}\kappa\hat{\nabla}_\alpha p^{eff} &\approx&
-\frac{1}{f_T}\Bigg[(f_T-1)\kappa\hat{\nabla}_{\alpha}p^{f}
+\frac{8}{9}\theta^2\dot{\theta}f_{TTT}\hat{\nabla}_{\alpha}T\nonumber\\
&& -\frac{4}{3}\left(\dot{\theta}+\frac{2}{3}\theta^2\right)f_{TT}\hat{\nabla}_{\alpha}T
-\frac{2}{3}f_{TT}\theta\left(\hat{\nabla}_{\alpha}T\right)^{\cdot}
\nonumber\\&&+\frac{8}{9}\theta^2\dot{\theta}f_{TT}A_{\alpha}\Bigg].
\end{eqnarray}
Eqs.~(\ref{eq:q_eff_1}, \ref{eq:q_eff_2}, \ref{eq:pi_eff},
\ref{eq:delta_rho_eff}, \ref{eq:delta_p_eff}), together with the
equations given in Sect.~\ref{subsect:GR3+1}, are all we need to
study the perturbation evolution in $f(T)$ gravity.
\subsection{Scalar Equations in $f(T)$ Gravity}
\label{subsect:scalar_eqn}
Our formalism has so far been as general as possible. Now we will focus
exclusively on scalar perturbations and perform the following harmonic
expansions of our perturbation variables
\begin{eqnarray} \label{eq:HarmonicExpansion}
\hat{\nabla}_{\alpha}\rho = \sum_{k}\frac{k}{a}\mathcal{X}Q_{\alpha}^{k},\qquad \hat
\nabla}_{\alpha}p = \sum_{k}\frac{k}{a}\mathcal{X}^{p}Q_{\alpha}^{k} \nonumber \\
q_{\alpha} = \sum_{k}qQ_{\alpha}^{k},\qquad \pi_{\alpha\beta} =
\sum_{k}\Pi Q_{\alpha\beta}^{k},\qquad
\nonumber \\
\hat{\nabla}_{\alpha}\theta = \sum_{k}\frac{k^{2}}{a^{2}}\mathcal{Z}Q_{\alpha}^{k}, \qquad \sigma_{\alpha\beta} = \sum_{k}\frac{k}{a}\sigma Q_{\alpha\beta}^{k} \nonumber \\
\hat{\nabla}_{\alpha}a = \sum_{k}khQ_{\alpha}^{k},\qquad
A_{\alpha} = \sum_{k}\frac{k}{a}AQ^{k}_{\alpha} \nonumber \\
\ae_{\alpha} = \sum_{k}\ae Q^{k}_{\alpha},\qquad \eta_{\alpha} = \sum_{k}\frac{k^{3}}{a^{2
}\eta Q_{\alpha}^{k} \nonumber \\
\mathcal{E}_{\alpha\beta} = -\sum_{k}\frac{k^{2}}{a^{2}}\phi
Q_{\alpha\beta}^{k}
\end{eqnarray}
in which $Q^{k}$ is the eigenfunction of the comoving spatial Laplacian
a^{2}\hat{\nabla}^{2}$ satisfying
\begin{eqnarray}
\hat{\nabla}^{2}Q^{k} &=& \frac{k^{2}}{a^{2}}Q^{k}.\nonumber
\end{eqnarray}
$Q_{\alpha}^{k},Q_{\alpha\beta}^{k}$ are given by
$Q_{\alpha}^{k}=\frac{a}{k}\hat{\nabla}_{\alpha}Q^{k},
Q_{\alpha\beta}^{k}=\frac{a}{k}\hat{\nabla}_{\langle
\alpha}Q_{\beta\rangle }^{k}$.
In terms of the above harmonic expansion coefficients,
Eqs.~(\ref{eq:constraint_q}, \ref{eq:constraint_phi},
\ref{eq:propagation_sigma}, \ref{eq:propagation_phi},
\ref{eq:constraint_eta}, \ref{eq:propagation_eta}) can be
rewritten as \cite{GR3+1}
\begin{eqnarray}
\label{eq:constraint_q2} \frac{2}{3}k^{2}(\sigma - \mathcal{Z})
&=& \kappa qa^{2},\\
\label{eq:constraint_phi2} k^{3}\phi &=& -\frac{1}{2}\kappa
a^{2}\left[k(\Pi+\mathcal{X})+3\mathcal{H}q\right],\\
\label{eq:propagation_sigma2} k(\sigma' + \mathcal{H}\sigma) &=&
k^{2}(\phi+A)-\frac{1}{2}\kappa a^{2}\Pi,\\
\label{eq:propagation_phi2} k^{2}(\phi'+\mathcal{H}\phi) &=&
\frac{1}{2}\kappa a^{2}\left[k(\rho+p)\sigma+kq-\Pi'
-\mathcal{H}\Pi\right],\ \ \ \ \\
\label{eq:constraint_eta2} k^{2}\eta &=& \kappa\mathcal{X}a^{2} -
2k\mathcal{H}\mathcal{Z},\\
\label{eq:propagation_eta2} k\eta' &=& -\kappa qa^{2} -
2k\mathcal{H}A
\end{eqnarray}
in which $\mathcal{H}\equiv a'/a=\frac{1}{3}a\theta $ and a prime
denotes the derivative with respect to the conformal time $\tau$
($ad\tau =dt$). Also, Eq.~(\ref{eq:heat_flux_evolution}) and the
spatial derivative of Eq.~(\ref{eq:energy_conservation}) become
\begin{eqnarray}
\label{eq:heat_flux_evolution2} q' + 4\mathcal{H}q + (\rho+p)kA -
k\mathcal{X}^{p} +
\frac{2}{3}k\Pi &=& 0,\\
\label{eq:energy_conservation2} \mathcal{X}' + 3h'(\rho+p) +
3\mathcal{H}(\mathcal{X}+\mathcal{X}^{p}) + kq &=& 0
\end{eqnarray}
We shall always neglect the superscript
$^{\mathrm{tot}}$ for the total dynamical quantities and add
appropriate superscripts for individual matter species. Note that
\begin{eqnarray}
h' &=& \frac{1}{3}k\mathcal{Z} - \mathcal{H}A.
\end{eqnarray}
and $\rho, p, \mathcal{X}, \mathcal{X}^{p}, q, \Pi$ with
superscripts $^{f}$ or $^{eff}$ are the total quantities (fluid
matter plus correction terms). The harmonic coefficients
$\mathcal{X}^{eff}, \mathcal{X}^{p,eff}, q^{eff}, \Pi^{eff}$ can
be derived from Eqs.~(\ref{eq:delta_rho_eff},
\ref{eq:delta_p_eff}, \ref{eq:q_eff_1}, \ref{eq:q_eff_2},
\ref{eq:pi_eff}) such that
\begin{align}
\label{eq:delta_rho_tot}
&f_T\kappa\left(\mathcal{X}^{f}+\mathcal{X}^{eff}\right)a^2
= \kappa\mathcal{X}^{f}a^2 +
24\frac{f_{TT}}{a^2}k\mathcal{H}^{3}(\mathcal{Z}+\ae),\\
\label{eq:delta_p_tot}
&f_T\kappa\left(\mathcal{X}^{p,f}+\mathcal{X}^{p,eff}\right)a^2
= \kappa\mathcal{X}^{p,f}a^2\nonumber\\&\qquad\qquad-\frac{f_{TT}}{a^2} [8k\mathcal{H}\left(3\mathcal{H}'-\mathcal{H}^2\right)(\mathcal{Z}+\ae)
\nonumber\\&\qquad\qquad+8k\mathcal{H}^2(\mathcal{Z}+\ae)' +24\mathcal{H}^2\left(\mathcal{H}'-\mathcal{H}^2\right)A]\nonumber\\&\qquad\qquad+96\frac{f_{TTT}}{a^4}k\mathcal{H}^3\left(\mathcal{H}'-\mathcal{H}^2\right)(\mathcal{Z}+\ae),
\\
\label{eq:q_tot_a}
& f_T\kappa\left(q^{f}+q^{eff}\right)a^2
= \kappa
q^{f}a^2 -8\frac{f_{TT}}{a^2}k^2\mathcal{H}^{2}(\mathcal{Z}+\ae)\\
\label{eq:q_tot_b}
& \qquad\qquad\qquad\qquad\;= \kappa q^{f}a^2 -
12\frac{f_{TT}}{a^2}k\mathcal{H}\left(\mathcal{H}'-\mathcal{H}^2\right)\eta,\\
\label{eq:pi_tot}
& f_T\kappa\left(\Pi^{f}+\Pi^{eff}\right)a^2
= \kappa\Pi^{f}a^2\\&\qquad\qquad\qquad\qquad-12\frac{f_{TT}}{a^2}k\mathcal{H}\left(\mathcal{H}'-\mathcal{H}^2\right)(\sigma+\ae).\nonumber
\end{align}
This completes our derivation of the scalar mode covariant and
gauge-invariant perturbation equations for $f(T)$ gravity, and
we have one extra dynamical degree of freedom $\ae$. It is now
straightforward to choose a gauge, and as an example the
perturbation equations in the conformal Newtonian gauge are given
in Appendix~\ref{appen:newtonian}.
\begin{figure*}
\includegraphics[scale=1]{background.eps}
\caption{(Colour online) The background evolution for the $f(T)$
gravity model with $f(T)=T-\mu^{2(1+n)}/(-T)^n$. {\it Upper-left
Panel}: the fractional energy densities for matter ($\Omega_m$),
radiation ($\Omega_r$) and the effective dark energy
($\Omega_{\mathrm{DE}}=1-\Omega_m-\Omega_r$), as functions of the
cosmic scale factor $a$, which is normalised to $1$ today. {\it
Upper-right Panel}: the total effective equation of state
$w_{eff}=-1-\frac{2\dot{H}}{3H^2}$, as a function of $a$. {\it
Lower-left Panel}: the ratio between the Hubble expansion rates
for the $f(T)$ gravity model and for the $\Lambda$CDM paradigm, as
a function of $a$. {\it Lower-right Panel}: $f_T-1$ as a function of
$a$. Here, results are shown for $n=0$ (black solid curve), $0.1$ (green dotted curve), $-0.1$ (cyan dashed curve), $0.2$ (purple dash-dotted curve) and $-0.2$ (pink dash-triple-dotted curve). Note that $n=0$ corresponds
to the $\Lambda$CDM paradigm. The relevant physical parameters are
$\Omega_m=0.257, \Omega_r=8.0331\times10^{-5}$ and
$H_0=71.9$~km/s/Mpc.} \label{fig:background}
\end{figure*}
\begin{figure}
\includegraphics[scale=0.49]{ae.eps}
\caption{The time-evolution of frame-independent quantity
$\epsilon\equiv\ae+\sigma$, for the model with $f(T)=T-\mu^{2(1+n)}/(-T)^n$
and $n=0.1$, on different length scales, characterised respectively by
$k/(h~\mathrm{Mpc}^{-1})=10^{-4}$ (solid curve), $10^{-3}$ (dotted curve),
$10^{-2}$ (dashed curve) and $10^{-1}$ (dash-dotted curve).} \label{fig:ae}
\end{figure}
\begin{figure*}
\includegraphics[scale=1]{linpert.eps}
\caption{(Colour online) The power spectra for the large-scale structure of the $f(T)$
gravity model with $f(T)=T-\mu^{2(1+n)}/(-T)^n$. {\it Upper-left
Panel}: the CMB spectrum for different values of $n$ -- $0$ (black
solid curve), $0.1$ (green dotted curve), $-0.1$ (cyan dashed
curve), $0.2$ (purple dash-dotted curve) and $-0.2$ (pink
dash-triple-dotted curve). {\it Upper-right Panel}: the same as
the upper-left panel, but for the matter power spectrum at
redshift $0$ (today). {\it Lower-left Panel}: the late-time
evolution of the dark matter density contrast
$\Delta_{\mathrm{CDM}}$ on different scales (as indicated besides
the curves); three values of $n$ have been considered -- $n=0.0$
(solid curves), $0.1$ (dotted curves) and $-0.1$ (dashed curves).
{\it Lower-right Panel}: the same as the lower-left panel, but for
the late-time evolution of the gravitational potential $\phi$ on
different scales. The physical parameters are the same as listed
in the caption of Fig.~\ref{fig:background}, and three species of
massless neutrinos are used.} \label{fig:linpert}
\end{figure*}
\section{Numerical Results}
\label{sect:num}
For a quantitative analysis of the evolution of cosmological
perturbations in $f(T)$ gravity, one needs to consider a concrete class of models. Since the motivation for considering $f(T)$ gravity was based on the suggestion that it could account for the late time cosmic speed-up without the need for dark energy, it makes sense to restricts ourselves to models that exhibit this property (we have expressed out reservations about the theoretical motivation of a general $f(T)$ theory in the Introduction). Thus, we focus on the class of models that can be parametrized as
\begin{eqnarray}
f(T) &=&T-\frac{\mu^{2(n+1)}}{(-T)^n}
\end{eqnarray}
where $n$ is some real number. The $\mu$ parameter will be fixed to such a value so that the model can reproduce the late time accelerated expansion of the universe. The minus sign in $(-T)^n$ has also be chosen with some foresight, as $T=-\frac{2}{3}\theta^2=-6H^2<0$ in background cosmology. Our aim is to examine if this particular class of model which can reproduce the background cosmological evolution of the $\Lambda$CDM model is also compatible with large scale structure evolution.
Such a Lagrangian has been studied previously by
\cite{Linder:2010, Zheng:2010} but in different contexts.
\subsection{Background Evolution}
In background cosmology, the modified Friedman equation is given as
\begin{eqnarray}
3H^2 &=& \kappa\left(\rho^f+\rho^{eff}\right)\nonumber\\
&=& \frac{1}{f_T}\kappa\rho^f - \frac{1}{2f_T}\left(f-f_TT\right).
\end{eqnarray}
Using the fact that $T=-6H^2$, this equation could be written as
\begin{eqnarray}
-T-(1+2n)\frac{\mu^{2(n+1)}}{(-T)^n} &=& 2\kappa\rho^f,
\end{eqnarray}
according to which we could fix $\mu$ by assuming that the present
fractional energy density for ``dark energy" is $\Omega_\Lambda$:
\begin{eqnarray}
\mu^{2(1+n)} &=& \frac{1}{1+2n}\Omega_\Lambda\left(6H_0^2\right)^{1+n}.
\end{eqnarray}
Here $H_0$ is the present Hubble expansion rate. The
modified Friedman equation can then take the form
\begin{eqnarray}\label{eq:modified_friedman}
3H^2 &=& \kappa\rho^f+3\Omega_\Lambda H_0^2\left(\frac{H_0}{H}\right)^n.
\end{eqnarray}
Here the second term in the right-hand side represents the energy
density of an effective dark energy component. Given the value for
$n$, we can solve the algebraic equation
Eq.~(\ref{eq:modified_friedman}) to find the expansion rate of the
Universe at any earlier time.
We have considered five different values for $n$, with $n=0.0,\pm 0.1,\pm 0.2
$, and summarised the results for the background evolution in Fig.~\re
{fig:background}. The upper left panel shows the fractional energy densities
for matter, radiation and effective dark energy respectively. The black
solid curve ($n=0$) is the $\Lambda $CDM paradigm.
$H_0/H$ increases until
it reaches its current value $1.0$, so a positive $n$ (green dotted
and purple dash-dotted curves; same below) means the energy
density of dark energy was lower in the past. The opposite is true for a negative $n$ (cyan dashed and
pink dash-triple-dotted curves; same below). This behaviour is as predicted by
Eq.~(\ref{eq:modified_friedman}).
For positive values of $n$ the energy density of the ``dark energy"
increases in time, which implies that its pressure-density ratio should be
less than $-1$. Given that we normalise the ``dark energy" fractional energy density by its value today, at earlier times it will be lower in the $f(T)$ gravity model than in $\Lambda$CDM and so the universe will expands slower than in the latter. The effect on the total effective pressure-density ratio of all matter species, which is defined as $w_{eff}\equiv=-1-\frac{2\dot{H}}{3H^2}$, is shown in
the upper-right panel of Fig.~\ref{fig:background}.
Meanwhile, since for positive values of $n$ the dark energy (and
therefore the total energy) density was lower in the past than in
the $\Lambda$CDM paradigm, the Hubble expansion rate for the
former must be lower too, as can be seen from the lower-left panel
of Fig.~\ref{fig:background}. Note that at very early times the
expansion rates in these two models are almost the same, because
the effect of the $f(T)$ correction (or the cosmological constant)
is negligible then.
Finally, we shall find that the quantity $f_T=df/dT$ is important
in the $f(T)$ gravity model and so have plotted its evolution in
the lower-right panel of Fig.~\ref{fig:background}. Clearly
\begin{eqnarray}
f_T &=& 1-n\frac{\mu^{2(n+1)}}{(-T)^{n+1}}
\end{eqnarray}
and therefore must be negative for positive values of $n$,
and vice versa. Again, at very early times
$f_T-1\approx0$ because
$|f(T)-1|\ll|T|$, and the deviation of $f_{T}$ from unity
only becomes large at late times.
These results show that as long as $|n|$ is close enough to
$0$, the deviation of the $f(T)$ gravity model from $\Lambda $CDM is small
but the background expansion rate could provide a weak
constraint on the model parameter $n$.
\subsection{CMB and Large-scale Structure}
Having fixed $\mu$ in order to reproduce the desired background evolution, we are ready to consider the evolution of linear
perturbations. These could place much more stringent constraints on the
model parameters.
In the section above we gave the covariant and gauge invariant
linear perturbation equations for general $f(T)$ models. In order to solve these equation numerically we must specify a gauge (or reference frame). As
usual, we choose to work in the CDM frame (that is, the reference frame of an observer
comoving with dark matter fluid), which is characterised by
$v_{\mathrm{CDM}}=A=0$, where $v_{\mathrm{CDM}}$ is the peculiar
velocity of the dark matter fluid and $A$ is the acceleration of
the observer.
Next, we need to determine the behaviour of the new degree of freedom
$\ae$. In most modified gravity theories, this will be governed by a
dynamical equation. In the $f(T)$ gravity, however, its value is given by a
constraint equation.
This is a consequence of the fact that the right-hand side of Eq.~(\ref{eq:field_eqn}) is not {\em a priori} antisymmetric, but it is required to be as a consequence of the field equations. This leads to the two different expressions in
Eqs.~(\ref{eq:q_tot_a}) and (\ref{eq:q_tot_b}), which imply that
\begin{eqnarray}
k\mathcal{H}(\mathcal{Z}+\ae) &=& \frac{3}{2}\left(\mathcal{H}'-\mathcal{H}^2\right)\eta\,.
\end{eqnarray}
This equation can be used to determine $\ae$ in terms of $\mathcal{Z}, \eta$ and background
quantities.
We can then eliminate $\ae$ in all the relevant
perturbation equations. Nonetheless, it is interesting to see how the new degree
of freedom $\ae$ evolves in time on different length scales, and this
is shown in Fig.~\ref{fig:ae}. Since $\ae$ is not a gauge invariant quantity,
what we have plotted is $\epsilon\equiv\ae+\sigma$, which is gauge invariant. Note that
in the conformal Newtonian gauge, in which
$\sigma=0$ (c.f.~Appendix~\ref{appen:newtonian}), the quantity $\epsilon$ coincides with $\ae$. We show
the results for $n=0.1$ in Fig.~\ref{fig:ae}. We see that $\epsilon$ decreases
in time, and the decrease becomes more rapid as one moves to smaller scales
(bigger $k$'s). Therefore, we expect any deviations from the $\Lambda$CDM model to be more important on large scales than
on small scales. We will confirm this below.
We can now examine the growth of the dark-matter density contrast in the
context of the $f(T)$ gravity model. For simplicity, we shall assume that
the universe is filled with dark matter only, which is a fair approximation
at late times.
Taking the spatial derivative of
the Raychaudhuri equation one gets \cite{Li:2009}
\begin{eqnarray}
\label{interim}
&&k\mathcal{Z}' + k\mathcal{HZ} -k^2A\\&&\qquad\qquad + 3\left(\mathcal{H}'-\mathcal{H}^2\right)A = -\frac{1}{2}\left(\mathcal{X}+3\mathcal{X}^p\right)a^2,\nonumber
\end{eqnarray}
in which $k$ is the wavenumber and $\mathcal{X}, \mathcal{X}^p$
include contributions from both the dark matter and the $f(T)$
corrections. In the CDM frame $A=0$, and the conservation
equation for dark matter gives $\Delta'=-k\mathcal{Z}$, where
$\Delta=\mathcal{X}_{\mathrm{DM}}/\rho_{\mathrm{DM}}$ is the dark
matter density contrast. Then, Eq.~(\ref{interim}) can be rewritten, by
manipulating our set of perturbation equations, as
\begin{eqnarray}\label{eq:delta_dm}
\Delta''+\left(1-2C\right)H\Delta' &=& \frac{\kappa\rho_{\mathrm{DM}}a^2}{f_T}\left[\frac{1}{2}+C\right]\Delta
\end{eqnarray}
with $C$ defined by
\begin{eqnarray}
C &\equiv& 216\frac{f_{TTT}/a^4}{f_T}\frac{\mathcal{H}^2\left(\mathcal{H}'-\mathcal{H}^2\right)^2}{k^2-36\frac{f_{TT}/a^2}{f_T}\mathcal{H}^2\left(\mathcal{H}'-\mathcal{H}^2\right)}\nonumber\\
&&-216\frac{\left[\frac{f_{TT}/a^2}{f_T}\mathcal{H}\left(\mathcal{H}'-\mathcal{H}^2\right)^2\right]^2}{k^2-36\frac{f_{TT}/a^2}{f_T}\mathcal{H}^2\left(\mathcal{H}'-\mathcal{H}^2\right)}\nonumber\\
&&+\frac{f_{TT}/a^2}{f_T}\frac{156\mathcal{H}^2\mathcal{H}'-24\mathcal{H}\mathcal{H}''-60\mathcal{H}^3-48\mathcal{H}'^2}{k^2-36\frac{f_{TT}/a^2}{f_T}\mathcal{H}^2\left(\mathcal{H}'-\mathcal{H}^2\right)}.\nonumber
\end{eqnarray}
Clearly, on very small scales, where $k\gg \mathcal{H}
\mathcal{H}^{\prime }/\mathcal{H}$ and $\mathcal{H}^{\prime \prime }
\mathcal{H}^{2}$ we have $C\rightarrow 0$ and Eq.~(\ref{eq:delta_dm})
reduces to that in the $\Lambda$CDM model, only with the
value of the gravitational constant rescaled by $1/f_T$.
On very large scales, in contrast, we can neglect $k^{2}$ in the expression
for $C$, and Eq.~(\ref{eq:delta_dm}) becomes very complicated, leading to
large deviations from $\Lambda $CDM.
One should also be able to derive an evolution equation for the
gravitational potential $\phi$ defined in
Eq.~(\ref{eq:HarmonicExpansion}) (indeed this will be easier if we use the
Newtonian gauge potentials given in Appendix~\ref{appen:newtonian}), but we
shall not do that here.
In Fig.~\ref{fig:linpert} we show some results for the linear perturbation
evolutions in the $f(T)$ model studied here. Clearly both the CMB and matter
power spectra (for all choices of $n$ except for $n=0$ which corresponds to
\Lambda $CDM) blow up on
large angular scales (small $\ell$ or small $k$), which is
consistent with the above analysis that the evolution of matter
density perturbations (and therefore the gravitational potential)
on large scales is very different from the $\Lambda$CDM
predictions. On small scales, however, the $f(T)$ model gives
similar predictions as $\Lambda$CDM, which is as expected.
To see more clearly how the growth of the dark matter
density contrast and the growth of the gravitational
potential have been modified, we have plotted them in the
lower panels of Fig.~\ref{fig:linpert}. For $\Delta _{\mathrm{CDM}}$, the
difference between the $f(T)$ models (with $n=\pm 0.1$) and the $\Lambda $CDM is within $\sim 10\%$ on small scales ($k>0.001h$~Mpc$^{-1}$) because
the effective gravitational constant is rescaled and the cosmic expansion
rate is modified as well. But on very large scales ($k<0.0001h$~Mpc$^{-1}$), the difference becomes very significant.
The same happens to $\phi$.
These results are expected to remain qualitatively true for other choices
for the function $f(T)$, if they are made so as to explain the late-time
acceleration of the universe. This can be seen from the expression
for $C$, which shows that the
large-scale deviation from $\Lambda$CDM is inevitable whenever
$f_{TT}$ and/or $f_{TTT}$ are nonzero.
The results suggest that $f(T)$ gravity models which are proposed as an alternative to
dark energy could face severe difficulties in being compatible with observations regarding large scale evolution. The expectation that that linear perturbation
analysis gives better constraints than the consideration of
background cosmology alone is clearly confirmed here as well.
\section{Summary and Conclusions}
\label{sect:con}
In summary, we have given the modified Einstein equations for general $f(T)$
gravity models in a covariant formalism, and derived the covariant and
gauge-invariant perturbation equations in the $3+1$ formalism. The
perturbation equations take full account of the extra degrees of freedom in
the $f(T)$ gravity theory (the importance of which was first discussed in
Ref.~\cite{Li:2010}) up to linear order. The equations in specific gauges can
then be obtained straightforwardly as shown in Appendix~\ref{appen:newtonian}.
For a general $f(T)$ theory it turns out that
no new degrees of freedom
appear at the background level, and the modified Friedmann equation
is simply a nonlinear algebraic equation in the Hubble rate $H$ that can easily be solved numerically.
At the linear order in perturbation there is a new vector degree of freedom (as a consequence of the lack of
local Lorentz symmetry, as pointed out in Ref.~\cite{Li:2010}). However, at this order the equations include no time derivatives of this vector, which just satisfies a constraint equation.
After developing the general formalism and deriving the perturbed equation at linear order, we restricted our attention to scalar perturbations. We then considered a broad class of $f(T)$ theories which are representative examples of models that could account for the late-time acceleration of the universe, as proposed in the literature. We studied in detail their background
cosmology and the evolution of linear perturbations. We were able to determine the new degree of freedom algebraically in
terms of other curvature perturbation quantities.
We also derived the evolution equation for the dark-matter density
contrast $\Delta$ in a dark-matter-dominated universe, and showed
that it resembles that of $\Lambda$CDM on small scales, but gets
significantly modified on large scales. The
large-scale CMB and matter power spectra blow up, signalling a serious viability problem for any
$f(T)$ models that are able to account for the accelerated
expansion of the universe at the background level. We have argued that this conclusion is
robust and holds true for other choices of $f(T)$ unless
$f_{TT}=f_{TTT}=0$ at late times.
Our result clarifies the effects of the new degree of freedom in the $f(T)$ gravity model
at the linear perturbation level, and we have seen here that only one extra degree of freedom arises. An interesting
question is whether further degrees of freedom will enter into the field equations, and if so, whether they are well behaved, when followed beyond
linear perturbation. This will be investigated elsewhere.
|
\section{Introduction}
Over the past few years there has been spectacular progress~\cite{MHVoneloop, NMHVoneloop, BCDKS, FiveLoop, KorchemskyOneLoop,TwoLoopSixPt, CachazoLeadingSingularityAndCalcs, Spradlin3loop, VerguAnd2L6pNMHV, Grassmannians, AHIntegrands,TwistorWilsonLoopAndDualCSW} in writing down the integrands of loop-level scattering amplitudes of the ${\cal N}=4$ super Yang-Mills (sYM) theory~\cite{NeqFourDefinition} in the planar leading-color sector, or 't Hooft limit. Similarly, for the integrated planar multiloop amplitudes there are recent important advances, notably the all-loop resummation of four and five-particle amplitudes proposed by Anastasiou, Bern, Dixon, Kosower, and Smirnov~\cite{BDS}, and by the fully analytic evaluation of the six-particle two-loop maximally-helicity-violating amplitude~\cite{AnalyticSixPoint}; beyond that, there are many hints, both at weak and strong coupling~\cite{MagicIdentities, AldayMaldacena, AmplitudeWilsonLoopDuality, DualSuperconformalSymmetry, Yangian, MassiveRegulatorProgress, LoopPolytopes, BeisertProgram, correlatorsAndamplitudes}, that the planar sector may soon be solved.
The subleading color contributions to {${\cal N}=4$~sYM}\ amplitudes are less well studied---save for the one loop case, where all subleading-color amplitudes may be obtained from the planar sector ones (see e.g. \cite{MHVoneloop,treeLoopcolorDecomp}). Beyond one loop, the literature contains only the full non-planar contributions of the four-point amplitude, worked out at two~\cite{BRY}, three~\cite{superfinite, CompactThree,BCJloops} and four~\cite{Neq44np} loops at the level of the integrand. Notably, these four-point amplitudes also happen to be some of the few amplitudes known to be valid in higher dimensions $D>4$ \cite{Neq44np,SixDim} in the dimensionally oxidized ${\cal N}=4$ theory (dimensional reductions of the $D=10$ sYM theory).
In this contribution to the review, we will discuss some generally applicable methods for constructing multiloop multiparticle integrands of amplitudes, valid for {${\cal N}=4$~sYM}\ in four and higher dimensions, including non-planar contributions. Much of what is covered also applies to less supersymmetric theories, including pure QCD; however, the focus and examples in this chapter will be on the {${\cal N}=4$~sYM}\ theory. We will exhibit the general organization of unitarity cuts and amplitudes, specifically the maximal cut method~\cite{FiveLoop,CompactThree,Neq44np} and the imposition of a duality between color and kinematic structures of amplitudes~\cite{BCJ,BCJloops}. Interestingly, the duality allows one to recognize that all contributing terms of the amplitude are fully specified by some small set of independent ``master graphs''. A discussion which only applies to {${\cal N}=4$~sYM}\ is the review of special recursive cuts known as ``two-particle cuts''~\cite{BRY,BDDPR,Neq44np}, and ``box cuts''~\cite{Neq44np}, relying on the fact that all four-point amplitudes in {${\cal N}=4$~sYM}~factorize into a product of the corresponding tree amplitude times a universal function.
Loop amplitudes of massless theories in four dimensions suffer from infrared divergences (and frequently also ultraviolet divergences). A standard approach for avoiding these infrared divergences is to dimensionally regulate the amplitudes, that is, to work in $D=4-2\epsilon$ with $\epsilon<0$.
In this scheme, it is thus critical to compute loop amplitudes in $D>4$ dimensions. The methods presented in this review are well suited for this task.
Throughout this contribution we make use of the unitarity method~\cite{UnitarityMethod} and generalized unitarity~\cite{GeneralizedUnitarity,GeneralizedUnitarity2,BCF}, which we consequently will briefly outline. However, deeper understanding of details, as well as applications of this approach, may be found in some of the parallel chapters of this review, {\it e.g.} Bern and Huang~\cite{ZviYutinReview}, Britto~\cite{BrittoReview}, and Ita~\cite{ItaReview}.
The organization is as follows. In \sect{PreliminariesSection}, we set up notation and organization of multiloop integrands and cuts. In \sect{MaximalCutsSection}, the method of maximal cuts is discussed in some detail, focusing on both the systematic identification of needed cuts as well as the computation of these using tree amplitudes as input. Complimentary to the method of maximal cuts, in \sect{RecursiveCutsSection}, we review two particularly simple classes of multiloop cuts, relevant to the {${\cal N}=4$~sYM}\ theory: {\it two-particle cuts}, and {\it box cuts}. Finally, in \sect{BCJ}, we discuss the set of powerful constraints that one can impose on multiloop gauge theory amplitudes by considering the color--kinematics duality.
\section{Preliminaries}
\label{PreliminariesSection}
\subsection{Integral representations for amplitudes}
To systematically employ the unitarity method~\cite{UnitarityMethod, GeneralizedUnitarity,GeneralizedUnitarity2}, it is convenient to organize the amplitude using a sufficiently general integral representation.
We note that one can express any local massless gauge theory amplitude in $D$ dimensions in terms of local interactions and non-local propagators $\sim 1/p^2$, regardless of the particle content ({\it cf.} Feynman gauge propagators and vertices). With this in mind, one can write any amplitude as
\begin{equation}
{\cal A}_n^{(L)} \sim \sum_{i\in\Gamma} \, \int \frac{d^{LD}\ell}{(2\pi)^{LD}} \, \frac{\rm local}{p_{i_1}^2p_{i_2}^2p_{i_3}^2\cdots p_{i_m}^2}\,,
\label{localRep}
\end{equation}
where $d^{LD}\ell=\prod_{j=1}^{L}d^D\ell_j$ is the usual Feynman diagram integral measure of $L$ independent $D$-dimensional loop momenta $\ell_j^\mu$, and $\Gamma$ is the set of all $L$-loop graphs, with $n$ external legs, counting all relabeling of external legs. Corresponding to each internal line (edge) of the $i^{\rm th}$ graph, we associate a propagator $1/p_{i_l}^2$, which is a function of the independent internal and external momenta, $\ell_j$ and $k_j$, respectively. The local numerator functions, here only schematically indicated, include information about the color, kinematics and states, {\it etc}. This representation thus mimics the behavior of Feynman diagrams, but as we will see the local numerators need not be calculated using off-shell Feynman rules. Instead it is often more efficient to indirectly infer these numerator functions by imposing general constraints and by matching to physical on-shell information gleaned from generalized unitarity cuts.
Clearly the way of expressing an amplitude in \eref{localRep} is not unique. One can always add more propagators to any graph by multiplying by unity, expressed as $p_{i_l}^2/p_{i_l}^2$, absorbing the numerator $p_{i_l}^2$ into the graph's local numerator factor without changing its locality. To remove this freedom, we can insist that the graphs we use have only cubic (trivalent) interactions. That is, we can hide the contact terms that one usually encounters in gauge theories through the repeated multiplication of $p_{i_l}^2/p_{i_l}^2$ factors until every graph has the same number of propagators as a cubic graph: $n+3L-3$. Thus, we use a more constrained representation of the amplitude,
\begin{equation}
{\cal A}_n^{(L)}=\sum_{i\in\Gamma_3} \, \int \frac{d^{LD}\ell}{(2\pi)^{LD}} \frac{1}{S_i} \, \frac{N_i C_i}{p_{i_1}^2p_{i_2}^2p_{i_3}^2\cdots p_{i_{m}}^2}\,,
\label{cubicRep}
\end{equation}
where $\Gamma_3$ is the set of all purely cubic graphs, and $m=n+3L-3$ for all graphs. Furthermore, we have, for convenience, separated out the kinematic parts $N_i$ and the color parts $C_i$ of the local numerator factor, and normalized the numerators by dividing out by the standard (Bose) symmetry factor $S_i$ of each graph.
One might well wonder if it is always possible to factorize the kinematic and color terms for each graph. Indeed, this requirement imposes severe constraints on the local numerator. Nonetheless, when all particles are in the adjoint representation of the gauge group, as is the situation for the theories we will consider, this factorization is possible~\cite{BCJ}. It can be seen from the Feynman rules of gauge theory: if all particles are in the adjoint the only color structure that can appear in cubic vertices are the gauge group structure constants $f^{abc}$. Hence, every purely cubic Feynman diagram has a unique color factor associated to it. For the quartic contact vertices of gauge theory, there are three different color structures: $f^{a_1a_2b}f^{ba_3a_4}$ and its permutations. If we promote such a contact term to be a part of a cubic graph using $f^{a_1a_2b}f^{ba_3a_4} \rightarrow p^2_{12}/p^2_{12}f^{a_1a_2b}f^{ba_3a_4}$, then this term is proportional to the same color factor as the non-contact contribution in the $p^2_{12}=(p_1+p_2)^2$ channel. Thus, by this construction, the color factor of each cubic graph in \eref{cubicRep} is unique, up to a normalization factor. They are given by
\begin{equation}
C_i=\prod_{j\in V_{i}}\tilde f^{a_{j_1}a_{j_2}a_{j_3}}\,,
\label{ColorDef}
\end{equation}
where the product runs over all the cubic vertices in the $i^{\rm th}$ graph. Here, we use structure constants $\tilde f^{abc}$ with a normalization differing from normal conventions, $\tilde f^{abc}= i \sqrt{2} f^{abc} = \Tr([T^a, T^b] T^c)\,$, and with hermitian generators $T^a$ normalized via $\Tr(T^a T^b) = \delta^{ab}$.
The kinematic local numerators $N_i$ are by definition polynomials in the loop momenta but may be taken to be either polynomials or rational functions of external kinematic variables depending on the convenience of the situation at hand. Needless to say, $N_i$ also depends on the polarizations, helicities and spinors, {\it etc}, according to the details of the external states. The maximal degree of the polynomial in loop momenta can be determined by dimensional analysis: an $n$-particle gauge theory amplitude in $D=4$ has dimension $4-n$; thus, after accounting for the dimension of the integral measure and propagators, $N_i$ has dimension $n+2L-2$. This dimensionality coincides with the number of cubic vertices in each graph, as is expected of a gauge theory where each cubic vertex has one power of momenta. The dimension of $N_i$ then has to be distributed between the factors of external and internal momenta. This means that in a generic gauge theory, one would expect that the loop momentum dependence in $N_i$ comes in monomials of the form
\begin{equation}
1, \ell^{1}, \ell^{2},\ldots, \ell^{n+2L-3},\ell^{n+2L-2}~~ \in~~ N_i\,,
\label{monomials}
\end{equation}
where $\ell^{p}$ is a mnemonic for any Lorentz tensor constructed out of $p$ factors of loop momenta $\ell_j^\mu$. For special gauge theories, the highest degree monomials are absent for all $N_i$. This property can be attributed to the good ultraviolet behavior of the theory, as the ultraviolet regime is indeed the corner of phase space in which $\ell_j^\mu \rightarrow \infty$.
Although the ultraviolet behavior of a theory is interesting in itself, for the purposes of this review we will only note that such behavior constrains the form of the amplitude. In the absence of other constraints, the fewer powers of $\ell$ that one must consider, the smaller the space of possible numerator polynomials will be. This can greatly facilitate amplitude construction. The set of well-known gauge theories with good ultraviolet behavior includes many supersymmetric Yang-Mills theories. In particular, maximally supersymmetric {${\cal N}=4$~sYM}\ is known to reduce the maximum power of the monomials \eref{monomials} by at least four relative to pure Yang-Mills theory~\cite{NeqFourFiniteness,MS,BDDPR,HoweStelleRevisited}. Theories with ${\cal N}$-fold supersymmetry are expected to reduce the maximal power of $\ell$ by at least ${\cal N}$ in \eref{monomials}, see {\it e.g.}~\cite{UnexpCanc}. In some cases, the cancellations of the high powers of $\ell$ are even stronger than this naive counting~\cite{MS,Neq44np, DoubleTraceArguments}.
For completeness, we also mention an alternative integral representation frequently used in the literature. For amplitudes with many external legs (compared to the number of spacetime dimensions), it can be cumbersome to use an integral representation with only cubic vertices. For example, the standard one-loop integral basis in four dimensions consists of box, triangle and bubble integrals; no higher polygon integrals are needed for massless $D=4$ theories (see {\it e.g.}~\cite{BrittoReview}). A similar basis is expected to exist at the two-loop level~\cite{Kosower2Lbasis}. To allow for such cases, we use a sufficiently general (over-complete) integral representation
\begin{equation}
{\cal A}_n^{(L)}=\sum_{k\in {\rm TB}}G_{k} \sum_{i\in\Gamma} \, R_{i,k} \int \frac{d^{LD}\ell}{(2\pi)^{LD}} \, \frac{P_{i,k} }{p_{i_1}^2p_{i_2}^2p_{i_3}^2\cdots p_{i_{m}}^2}\,,
\label{convRep}
\end{equation}
where the first sum runs over color factors $G_k$ in the trace basis (TB), of say $SU(N_c)$, and the second sum runs over $n$-point $L$-loop graphs as in \eqn{localRep}. The $R_{i,k}$ are rational functions of external momenta, and $P_{i,k}$ are polynomials of the independent loop momenta. As desired, one can generally restrict the number of propagators $m \le \lceil D\rceil L$ ($\lceil \bullet \rceil$ is the ceiling function), because integrals with $m>\lceil D\rceil L $ are reducible ({\it e.g.} using Melrose, Van-Neerven-Vermasseren, or Bern-Dixon-Kosower reduction and similar methods~\cite{integralReductionsMNVBDK}), although the exact upper bound depends critically on the subtleties of dimensional regularization at each loop order. Similar to the representation \eref{cubicRep} the ultraviolet behavior of the theory under consideration imposes useful constraints on the form of $P_{i,k}$. For example, the good ultraviolet behavior of {${\cal N}=4$~sYM}\ can be used to show that only scalar boxes contribute at one loop in $D=4$ \cite{MHVoneloop}; thus, one can set $P_{i,k}=1$ for all boxes and $P_{i,k}=0$ for all other one-loop integrals in this theory.
As a recent alternative to the mentioned general integral representations, we remark on an interesting new set of chiral integrals that have been introduced specifically for planar {${\cal N}=4$~sYM}\ using momentum twistor space notation~\cite{AHIntegrands}. Indeed, using special properties of a theory, such as dual (super-)conformal symmetry~\cite{MagicIdentities, DualSuperconformalSymmetry} (or Yangian structure~\cite{Yangian}) of {${\cal N}=4$~sYM}, one may construct more constrained integral representations, which consequently increases the power of any amplitude calculation~\cite{BCDKS, FiveLoop, KorchemskyOneLoop,TwoLoopSixPt}. In similar spirit, the duality between color and kinematics~\cite{BCJ,BCJloops} discussed in \sect{BCJ}, is conjectured to provide an constraining organizing principle for the integrals in more general gauge theories.
We note that the color--kinematics duality will critically rely on the cubic integral representation \eref{cubicRep}, whereas in sections \ref{MaximalCutsSection} and \ref{RecursiveCutsSection}, which treat generalized unitarity cuts, either integral representation \eref{cubicRep} or \eref{convRep} is applicable.
\subsection{Generalized unitarity cuts}
Unitarity of the $S$-matrix requires that loop amplitudes must satisfy certain properties relating them to sums over products of lower-loop amplitudes. The goal of the unitarity method~\cite{UnitarityMethod, GeneralizedUnitarity,GeneralizedUnitarity2} is to construct a compact expression that obeys all constraints required by unitarity for a particular amplitude. For massless theories it is sufficient to satisfy all generalized unitarity cuts in $D$ dimensions, this guarantees that no logarithmic or rational terms can go undetected. Strictly speaking only a subset of these, a "spanning set" of cuts~\cite{Neq44np}, are necessary to guarantee the satisfaction of all other cuts.
To identify contributions to the amplitude, we will compute generalized unitarity cuts on the level of the integrand, expressing them as a product of on-shell subamplitudes,
\begin{equation}
i^c \sum_{\rm states}A_{(1)}\,A_{(2)}\,A_{(3)}\cdots A_{(m)}\,,
\label{cutdef}
\end{equation}
where $A_{(i)}$ are the subamplitudes, which may be either tree or lower-loop amplitudes. The number of cut lines for each cut (the number of on-shell $p_i^2=0$ internal legs) is denoted by $c$.
The cuts can be computed directly from the theory by feeding in the corresponding subamplitudes and summing over the intermediate states, as in \eref{cutdef}. They may, of course, also be computed from graph-based integrand representations of the full loop amplitude, \eref{cubicRep} or \eref{convRep}, when available. In this latter case, the sum over intermediate states has already been carried out, and the generalized cut can be formally obtained by replacing the propagators of the cut lines by delta functions,
\begin{equation}
\frac{1}{p_i^2} \rightarrow 2\pi \delta(p_i^2)\,,
\label{cutdef2}
\end{equation}
and collecting the terms with the most delta functions. Since the full integrand of the amplitude under consideration is typically {\it a priori} unknown, the main purpose of these types of cuts is to apply them to an ansatz of the amplitude.
The unitarity method instructs us to reconstruct amplitudes using the information from the set of all generalized unitarity cuts of a theory. This set is overcomplete; consequently, there are different strategic approaches on how to most efficiently extract the relevant information for building the amplitudes. For example, one might well consider a strategy of using the fewest number of cuts, with each individual cut thereby providing the most information about a given amplitude. The most extreme such situation is exemplified by the one-particle cut of an $n$-point $L$-loop amplitude, relating it to a $(n+2)$-point $(L-1)$-loop amplitude sewn with itself. This cut has the minimal number of cut propagators, and in principle contains the full information of all the other generalized cuts of the $L$-loop amplitude (ignoring any subtleties with possible collinear and soft divergences of the cut line). Without an organizing principle, disentangling this information can be prohibitive. In the planar-limit of~{${\cal N}=4$~sYM}, dual conformal symmetry (or Yangian invariance) appears to provide such an organizing principle~\cite{AHIntegrands}. See also~\cite{BrittoSingleCut} for a general discussion of single-line cuts at the one-loop level.
In the next section, we consider a different strategy, using the method of maximal cuts. Rather than maximizing the information coming from each cut, we maximize the number of cuts and similarly the number of cut lines, ensuring that each cut provides as small an amount of information as possible, thus facilitating the piece-by-piece construction of the amplitude.
\section{Maximal cuts in $D=4$ and general dimension}
\label{MaximalCutsSection}
The method of maximal cuts~\cite{FiveLoop,CompactThree,Neq44np} offers a particularly efficient means for determining multiloop amplitudes. In this method, we start from those generalized cuts that have the maximum number of cut propagators (maximal cuts). These cuts provide an initial ansatz for the amplitude. Next we systematically correct the ansatz by considering the information provided by near-maximal cuts. These sequentially reduce the number of cut propagators one by one. In this process, all potential contact contributions are determined.
Generalized unitarity has a wide range of applicability: amplitudes of generic theories in generic spacetime dimension can be constructed. Since maximal and near-maximal cuts are special cases of generalized unitarity cuts, they can be used in many theories and space-time dimensions. However, there are some mild assumptions necessary if one wishes to organize around cubic vertices and cut all propagators: it is important that massless on-shell three-point amplitudes are non-vanishing and non-singular~\cite{BCF,BuchbinderCachazo}, for appropriate choices of complex cut loop momenta~\cite{GoroffSagnotti,WittenTopologicalString}. Moreover, it is important that the space-time is of dimension $D\ge 4$, in order to allow for one to impose both momentum conservation and on-shell conditions of the legs meeting at a three-point vertex without having collinear momenta\footnote{These assumptions are needed if one wants to cut down to cubic vertices. In theories that have no on-shell three point vertices amplitudes may instead be organized around quartic vertices, which has no such restrictions~\cite{ThreeDimAmpl}.}.
The strategy of maximizing the number of cut propagators is known to be a powerful method for computing one-loop {${\cal N}=4$~sYM}\ amplitudes in four dimensions, because
only box integrals appear~\cite{MHVoneloop}. As observed by
Britto, Cachazo and Feng~\cite{BCF}, taking a quadruple cut,
where all four propagators in a box integral are cut, freezes the
four-dimensional loop integration. This allows the kinematic coefficient of the box
to be straightforwardly computed directly in terms of the cut. The use of complex
momenta, as suggested by twistor space
theories~\cite{WittenTopologicalString}, makes it possible to define
massless three-vertices and thereby to use quadruple cuts to determine
the coefficients of all box integrals including those with massless
external legs.
The idea behind the quadruple cuts has been generalized by Buchbinder and Cachazo~\cite{BuchbinderCachazo} to two loops using hepta- and
octa-cuts of the double box integral topology. Although the two-loop four-point double-box integral only has seven physical propagators, it secretly enforces an additional spurious eighth constraint, yielding an octa-cut which localizes the integration
completely. These types of cuts are now known as a part of the ``leading-singularity''
technique, which is similar to the maximal cut method in that it is based on cutting
a maximal or near-maximal number of
propagators~\cite{CachazoSkinner,CachazoLeadingSingularityAndCalcs},
but in addition it may make use of extra conditions from spurious
singularities that are special to four dimensions.
The strength of the method of maximal cuts has most clearly been seen in amplitudes that have a large number of multiloop integrals. For example, the four-point amplitudes of {${\cal N}=4$~sYM}\ have been evaluated using maximal and near-maximal cuts at five loops~\cite{FiveLoop} in the planar sector, and similarly for the full non-planar amplitudes at three~\cite{CompactThree} and four loops~\cite{Neq44np}, the latter containing 50 distinct integrals, each with multiple terms.
\subsection{Working definition of maximal and near-maximal cuts}
\label{maxcutdefSection}
To make the concept of maximal cuts clear, we will introduce some proper definitions. First we note that one can represent generalized cuts~\eref{cutdef} as graphs: vertices (nodes) representing on-shell subamplitudes, and lines (edges) representing on-shell propagators. Maximal cuts are simply those generalized cuts (graphs) that have the most possible on-shell cut conditions, given three numbers: $n$ the number of external legs, $L$ the number of loops and $D$ the number of dimensions of the loop momenta. For any generalized cut, there are two competing upper limits to the number of cut conditions $c$ that one can impose:
\begin{eqnarray}
c\le n+3L-3\,, \hskip 1cm {(\rm cubic~graph~limit)} \nonumber \\
c\le DL\,, \hskip 1cm~~~~~~~~~~(\mbox{loop~momentum~freeze-out})
\label{MCbound}
\end{eqnarray}
where the first limit comes from counting the internal lines of a cubic $n$-point $L$-loop graph, and the second by counting the degrees of freedom of the independent loop momenta. It is clear that a cut consisting of only cubic vertices can have no more physical singularities as the integrand is completely local. Similarly, a cut where all the degrees of freedom of the loop momenta are frozen can have no more singularities in the internal phase space.
We define maximal cuts to be those generalized cuts which saturate one of the above bounds\footnote{In some unitarity cuts, the ``loop momentum freeze-out'' bound, which counts the total degrees of freedom of the loop momenta, is not a good measure of what happens locally in each subgraph. For example, a direct product of a one-loop triple cut and a one-loop penta-cut would be a two-loop maximal cut in four dimensions, since it is an octa-cut. However, since the one-loop penta-cut does not exist in four dimensions, one could argue that it should not be counted. Nonetheless, in order to simplify our definition~\eref{MCdef}, we simply count such situations as valid cuts that happens to evaluate to zero.}, that is,
\begin{equation}
c_{\rm MC}={\rm min}\{n+3L-3,DL\}\,.
\label{MCdef}
\end{equation}
For example, for $D=4$ and $L=1$ the maximal cut is equal to the quadruple cut $c_{\rm MC}=4$. For $n=4$ and $L=2$ it is the hepta-cut $c_{\rm MC}=7$, {\it etc}. (For comparison, we note that if we work in four dimensions, then for loop orders $L \le n-3$ the maximal cuts coincide with the leading singularity cuts as both have $4L$ cut lines.)
To define the set of near-maximal cuts, we consider those cuts with fewer cut lines than the maximal cuts $c<c_{\rm MC}$. In particular, a next-to-maximal cut has $c=c_{\rm MC}-1$ cut lines, and a next-to-next-to-maximal cut have $c=c_{\rm MC}-2$ cuts, and so on. For example, in this classification, we refer to both one-loop triple cuts in $D=4$ and quadruple cuts in $D=5$ as next-to-maximal cuts (for $n>4$).
\subsection{An algorithm for the maximal cut method}
We give here a general outline for constructing the loop amplitudes using the method of maximal cuts.
In words, the algorithm is as follows:
\begin{enumerate}
\item Enumerate all possible maximal cuts given the multiplicity $n$, loop order $L$, and dimension $D$. For example, if $n\le D$, or more generally $L>(n-3)/(D-3)$, the maximal cuts are in one-to-one correspondence with the cubic graphs of loop order $L$ and multiplicity $n$ (as follows from \eref{MCdef}).
\item Compute the maximal cuts at the level of the integrand ({\it i.e.} \eref{cutdef}). If the maximal cuts do not freeze the loop momenta, make sure to capture this residual dependence using fully Lorentz covariant variables of the loop momenta (Lorentz products or Levi--Civita tensor contractions).
\item Using the information from the maximal cuts, construct an initial ansatz for the amplitude. By the Lorentz covariant construction, the expressions obtained from the maximal cuts are automatically pushed to off-shell loop momenta (the validity of any such continuation will be ensured by the consecutive near-maximal cuts).
\item Enumerate the next-to-maximal cuts. Calculate each cut using \eref{cutdef}, then compare to the result of each cut as obtained from the amplitude ansatz constructed in the previous step. If the difference between the true cut and the cut of the ansatz is nonzero, then correct the ansatz by incorporating said difference.
\item Continue iterating the previous step for the next-to-next-to-maximal cuts, and so on; each step with one fewer cut condition imposed.
\end{enumerate}
The iteration can be stopped at any point when the amplitude ansatz appears to receive no more corrections. A proof that the ansatz is correct will follow from evaluating it on a complete set of ``spanning cuts''~\cite{Neq44np}, which are cuts that are known to fully determine $n$-point $L$-loop amplitudes of any theory (see~\cite{ZviYutinReview} for more details). Alternatively, the iteration may be continued by removing cut conditions one by one until the iteration naturally terminates by generating its own set of complete ``spanning cuts''.
If we represent the amplitude using the integrals in \eqn{cubicRep}, then the amplitude is entirely specified by the integral numerators $N_i$. The idea of the above algorithm is then that the numerators can be expanded as\footnote{It should be noted that this expansion is not unique. For example, contributions of the amplitude corresponding to contact terms can be shuffled between different graphs without affecting the amplitude, nor the generalized unitarity cuts. The algorithm of the maximal cut method simply gives a representation of the amplitude that satisfies all unitarity cuts; thus, much of the reparametrization freedom of the full amplitude is present in this representation.}
\begin{equation}
N_i=N_i^{\rm MC}+N_i^{\rm NMC}+N_i^{\rm N^2MC}+\ldots\,,
\label{expansion}
\end{equation}
where $N_i^{\rm MC}$ is the initial ansatz determined by the maximal cuts, and the corrections from the near-maximal cuts take the schematic form
\begin{equation}
N_i^{\rm NMC}=\sum_{j=1}^{m} p^2_{i_j} P_{i;j}\,, \hskip 5mm N_i^{\rm N^2MC}=\sum_{j,k=1}^{m} p^2_{i_j}p^2_{i_k} P_{i;j,k}\,, \hskip 5mm \cdots\,,
\end{equation}
where the $p^2_{i_j}$ are inverse propagators and $P_{i;j}$ and $P_{i;j,k}$ are polynomials fixed by the next-to- and next-to-next-to-maximal cuts, respectively. The contributions from next-to-maximal cuts must be multiplied by at least one inverse propagator since these contributions must vanish on maximal cuts where all $p^2_{i_j}=0$. Similarly, the corrections from the next-to-next-to-maximal cuts be multiplied by at least two inverse propagators since they vanish on all next-to-maximal cuts, and so on. Furthermore, since the inverse propagators carry dimension, the dimension (or degree) of the local polynomials must reduce for each iteration step, yet by locality remain non-negative. This is why the algorithm naturally terminates.
In the following sections, we will review how to compute maximal cuts using gauge theory tree amplitudes as input.
\subsection{Cuts with on-shell three-vertices}
In this part, we direct our attention to the evaluation of the maximal cuts. First, we will consider maximal cuts consisting of only three-point tree amplitudes
\begin{equation}
i^c \sum_{\rm states}A^{{\rm tree}}_{3,(1)}\,A^{{\rm tree}}_{3,(2)}\,A^{{\rm tree}}_{3,(3)}\cdots A^{{\rm tree}}_{3,(m)}\,,
\end{equation}
where the number of cut conditions is given by $c=n+2L-2$. Many of the interesting features of this cut are exposed by studying the kinematics of a single on-shell tree $A^{{\rm tree}}_{3}$.
Consider three massless momenta $p$, $q$ and $r=-p-q$ meeting at a three-point vertex, as in \fig{ChiralVertexFigure}; one has
the following constraints:
\begin{equation}
p^2=q^2 =p\cdot q=0\,,
\label{pqcond}
\end{equation}
where the orthogonality $p\cdot q=0$ follows from the masslessness of the third leg $r^2=(p+q)^2=2p\cdot q=0$. For real momenta in Minkowski signature $(1,D-1)$, these equations have only degenerate solutions where the momenta are collinear, $p\,\|\, q$. For nondegenerate solutions we may either work in $(2,D-2)$ signature, or more generally, we may explicitly complexify the momentum space (the $(2,D-2)$ signature can be thought of as a special case of complexfied momenta). For example, an explicit nondegenerate solution can, choosing a convenient Lorentz frame, take the form\footnote{Given the two $D$-dimensional momenta $p$ and $q$ that satisfies \eref{pqcond}, we introduce a massive reference vector $v$ such that $v\cdot q= 0$ and $v\cdot p\neq 0$, which can always be found when $p$ and $q$ are not collinear. By a Lorentz transformation, we may bring $v$ and $p$ to the forms $v^\mu=(\bullet,0,\ldots,0)$ and $p^\mu=(\bullet,\bullet,0,\ldots,0)$. From the conditions $v\cdot q= 0$ and $p\cdot q= 0$, we conclude that in this frame $q^\mu=(0,0,\bullet,\bullet,\ldots,\bullet)$. Next we perform a Lorentz transformation in all but the first two spacetime directions, which leaves $p$ unaffected and brings $q$ to the form $q^\mu=(0,0,\bullet,\bullet,0, \ldots,0)$.}
\begin{equation}
p^\mu=(1,1,0,0,0,\ldots,0)\,, \hskip 1cm q^\mu=(0,0,1,i,0,\ldots,0) \,.
\label{embededPQ}
\end{equation}
Interestingly the solutions in $D<4$ are strictly degenerate; thus the on-shell three-vertex is only supported by kinematics in $D\ge 4$. The case $D=4$ is the most interesting, not only because it is the spacetime we notice in our day-to-day lives but also because it is the marginal case where the phase space of the kinematics is the most constrained, yet nondegenerate. We will see the consequences of this below.
In four dimensions, we use the standard spinor helicity notation (see {\it e.g.}~\cite{DixonTASI,OnShellMethodsReview}) to solve kinematic constraints of the three-point vertex~\cite{BCF}; the momentum bispinor form $p=\lambda_p\widetilde{\lambda}_p$ and $q=\lambda_q\widetilde{\lambda}_q$ ensures the null conditions, and the orthogonality condition becomes
\begin{equation}
0 = 2p \cdot q = \spa{p}.{q} \spb{q}.{p} \,.
\label{oscond}
\end{equation}
For real momenta in Minkowski signature, $\lambda_{p}$ and ${\widetilde{\lambda}}_{p}$ are complex conjugates of
each other (up to a sign). Hence, if $\spa{p}.q$ vanishes then
$\spb{p}.q$ must also vanish and the momenta are collinear. If the momenta are taken to be complex, however, the two
spinors $\lambda_{p}$ and ${\widetilde{\lambda}}_{p}$ are independent. This
gives two independent solutions to \eqn{oscond},
\begin{equation}
\spa{p}.q=0 \hskip 1 cm \hbox {\rm or} \hskip 1 cm
\spb{p}.q=0 \,,
\label{TwoBranches}
\end{equation}
implying that $\lambda_{p}$ and $\lambda_{q}$ are proportional or, in the latter case, that $\widetilde{\lambda}_{p}$ and $\widetilde{\lambda}_{q}$ are proportional. Since $r=\lambda_r\widetilde{\lambda}_r=-\lambda_p\widetilde{\lambda}_p-\lambda_q\widetilde{\lambda}_q$, it then follows that there are overall two possible solutions:
all $\lambda$s proportional, and hence all $\spa{i}.{j}$ vanishing, or all $\widetilde{\lambda}$s proportional, and hence all $\spb{i}.j$ vanishing. We conclude that in $D=4$, the momentum-space support of the three-vertex exhibits a remarkable twofold branching or chirality. We will denote the vertices that are supported by $\lambda$ kinematics by $(+)$ and the vertices supported by $\widetilde{\lambda}$ kinematics by $(-)$, as illustrated in \fig{ChiralVertexFigure}.
\begin{figure}
\centerline{\epsfxsize 3.5 truein \epsfbox{ChiralVertex.eps}} \caption{In four dimensions, the kinematics of the on-shell three-point vertex comes in two chiral classes: the $(+)$\ vertex is supported on $\lambda$s (thus all $\widetilde{\lambda}$s are parallel and their $\spb{i}.{j}$ vanish), and the $(-)$\ vertex is supported on $\widetilde{\lambda}$s (thus all $\lambda$s are parallel and their $\spa{i}.{j}$ vanish). } \label{ChiralVertexFigure}
\end{figure}
\begin{figure}
\centerline{\epsfxsize 2.7 truein \epsfbox{ChiralProduct.eps}} \caption{In a $D=4$ cut, one can encounter two three-vertices of the same chirality joined by a common line ({\it e.g.} see cuts in \fig{PentaBoxFigure}(b) and \ref{non-planarFigure}). For the purpose of solving the kinematics, one can treat this configuration as an effective four-point vertex of definite chirality (note that a corresponding amplitude is not defined). All intermediate channels of this effective kinematic vertex are on-shell, as all legs are proportional to the same spinor.} \label{ChiralProductFigure}
\end{figure}
The chiral branching implies that the on-shell three-point amplitudes of {\it any massless theory} in $D=4$ have the following kinematic dependence:
\begin{eqnarray}
A_3^{(+)} \equiv A_3(\lambda_1,\lambda_2,\lambda_3)=a \spa{1}.{2}^b\spa{2}.{3}^c\spa{1}.{3}^d, \nonumber \\
A_3^{(-)} \equiv A_3(\widetilde{\lambda}_1,\widetilde{\lambda}_2,\widetilde{\lambda}_3)=a \spb{1}.{2}^b\spb{2}.{3}^c\spb{1}.{3}^d,
\label{ThreeVertexPlus}
\end{eqnarray}
where $a,b,c$ and $d$ are numbers which may be determined by calculation ($b,c,d$ alone may be determined by Lorentz symmetry~\cite{BenincasaCachazo}). For example, the pure-gluon amplitudes, where leg 1 has opposite helicity to that of leg 2 and 3, are
\begin{eqnarray}
A_3^{(+),{\rm tree}}(1^+,2^-,3^-)=i\frac{ \spa{2}.{3}^3}{ \spa{1}.{2} \spa{3}.{1}} \,,\nonumber \\
A_3^{(-),{\rm tree}}(1^-,2^+,3^+)=-i\frac{ \spb{2}.{3}^3}{ \spb{1}.{2} \spb{3}.{1}}\,.
\label{gluonThreeVertex}
\end{eqnarray}
Note that the $A_3^{(+),{\rm tree}}(1^+,2^-,3^-)$ amplitude vanishes for the kinematics of a $(-)$\ vertex, as the numerator vanishes faster than the denominator $\sim 0^3/0^2$. Similarly, the $A_3^{(-),{\rm tree}}(1^-,2^+,3^+)$ amplitude vanishes on the $(+)$\ vertex. The same is true for all three-particle amplitudes in massless gauge and gravity theories, as these amplitudes are completely local objects (the spinor--helicity formalism obscures this fact).
As we are interested in cuts that involve more than one on-shell three vertex, we should also understand the properties of configurations of several such vertices. One immediate simple ``transitivity'' property is: if two three-vertices of the same chirality are connected by a line with momentum $\lambda \widetilde{\lambda}$, then the two vertices are supported on the same chiral branch; thus, all momenta of the two vertices are proportional to a single spinor: $\lambda $ or $\widetilde{\lambda}$ depending on the chirality. As illustrated in \fig{ChiralProductFigure}, this implies that one can treat the kinematic configuration as an effective chiral four-point vertex. Similarly, if one encounters a connected chain of three-point vertices of the same chirality, one may treat it as an effective chiral higher-multiplicity vertex. All intermediate channels of such an effective kinematic vertex are on-shell, as all legs are proportional to the same spinor (note that there exists no known corresponding effective amplitude, the vertex here is used only as a kinematic tool).
\begin{figure}
\centerline{\epsfxsize 4 truein \epsfbox{QuadrupleCut.eps}} \caption{A quadruple cut of a four-point amplitude in $D=4$, and a penta-cut of a five-point amplitude in $D>4$. In $D=4$, we use $(+)$\ and $(-)$\ labels on three-point vertices to specify the chirality, but in $D>4$ the chirality is not a well-defined notion.} \label{QuadrupleCutFigure}
\end{figure}
In figures \ref{QuadrupleCutFigure}, \ref{PentaBoxFigure} and \ref{non-planarFigure} we illustrate typical cuts that are built out of only on-shell three point vertices. The four-particle quadruple cut in \fig{QuadrupleCutFigure}(a) uses four-dimensional momenta and must be built with vertices of alternating chirality~\cite{GeneralizedUnitarity2,BCF}. (The non-alternating configurations (not illustrated) correspond to degenerate situations where the momenta of two external legs are ``half-collinear''; that is, they have a common spinor, similar to the situation in \fig{ChiralProductFigure}.)
In $D>4$, the penta-cut in \fig{QuadrupleCutFigure}(b) is well defined, but unlike the $D=4$ case, the three vertices no longer come with definite chirality; therefore, we simply mark the vertices by a light shade (to avoid confusing them with tiny loops).
In \fig{PentaBoxFigure}, we show examples of two-loop cuts in $D=4$. Because the hepta-cut (a) enforces seven on-shell propagators $l_i^2=0$, one out of the eight degrees of freedom of the loop momenta is unconstrained (see discussion below). For the $D=4$ octa-cut of the five-point amplitude, \fig{PentaBoxFigure}(b), the number of on-shell conditions matches that of the number of loop momentum parameters; thus, the loop momenta is frozen by the cut, similar to the situation of the one-loop quadruple cut.
\begin{figure}
\centerline{\epsfxsize 5 truein \epsfbox{PentaBox.eps}} \caption{A hepta-cut of a four-point amplitude, and an octa-cut of a five-point amplitude, both in $D=4$.} \label{PentaBoxFigure}
\end{figure}
\begin{figure}
\centerline{\epsfxsize 4.6 truein \epsfbox{NonPlanar.eps}} \caption{A non-planar hepta-cut of a two-loop four-point amplitude, and a non-planar deca-cut of a three-loop amplitude, both in $D=4$.} \label{non-planarFigure}
\end{figure}
Before we conclude this section by working out a concrete example of a {${\cal N}=4$~sYM}\ maximal cut, it is convenient to recall some simplifying properties of state sums in maximal cuts~\cite{FiveLoop}. If a gluon of helicity $(+1)$ is attached to a chiral $(+)$ vertex, then the corresponding fermion and scalar amplitudes vanish, only the pure gluon amplitude has support:
\begin{eqnarray}
A_3^{(+),{\rm tree}}(1_g^+,2_f,3_f)=A_3^{(+),{\rm tree}}(1_g^+,2_s,3_s)=0\,, \nonumber \\
A_3^{(+),{\rm tree}}(1_g^+,2_g^{-},3_g^{-})\neq 0\, .
\label{vanishing}
\end{eqnarray}
The same is true for a gluon of helicity $(-1)$ that is attached to a chiral $(-)$ vertex. It is easy to understand these vanishings from the point of view of charge conservation. For a massless fermion the helicity has to be conserved in gluon--fermion interactions; thus, in our all out-going notation the helicity must flip. For example, $A_3^{(+),{\rm tree}}(1_g^+,2^-_f,3^-_f)$ is forbidden by helicity conservation. The case $A_3^{(+),{\rm tree}}(1_g^+,2^+_f,3^-_f)$ is allowed by helicity conservation, but it vanishes by virtue of having the wrong chirality. Similarly for a complex scalar its $U(1)$ charge has to be conserved in gluon--scalar interactions, implying that only amplitudes of the form $A_3^{(+),{\rm tree}}(1_g^+,2^+_s,3^-_s)$ are valid (where in this case $\pm$ ``scalar helicity'' refers to the scalar and its complex conjugate), but again this amplitude vanishes because of the chirality mismatch. The only nonvanishing amplitude is thus $A_3^{(+),{\rm tree}}(1_g^+,2^-_g,3^-_g)$, which has the right chirality. This ``projection property'' allows for some great simplifications of the maximal cuts by fine tuning the external states, such that all, or almost all, of the complicated particle and state sums in the loops drop out. In fact, \eref{vanishing} is generic to any massless gauge theory (and gravity); however, the most useful application is to the {${\cal N}=4$~sYM}\ theory where the amplitudes assemble into superamplitudes that are relatively insensitive to the choice of external states.
If we classify the various state-sum cut contributions according to the helicity configuration of the particles (thus ignoring the particle species labels), then the greatest simplification that can arise in a nontrivial kinematic solution is the restriction to a single allowed helicity configuration
for the internal lines. We will refer to this configuration as a ``singlet''. In this case only gluons can propagate inside the
diagram, as in \fig{singlet}(a). Fermions or scalars are not allowed because in each loop there is at least one vertex of the wrong
chirality type. The second-strongest simplification allows two helicity configurations. In such configurations, the particle content is purely gluonic except for one loop in which any particle type can propagate, as shown in
\fig{singlet}(b). (A single fermionic loop always allows two helicity
assignments, corresponding to interchanging fermion and antifermion,
and the same is true for a loop of complex scalars.)
\begin{figure}
\centerline{\epsfxsize 5.5 truein \epsfbox{singlet.eps}}
\caption{A singlet hepta-cut (a) and one of the two helicity
configurations (b) in a non-singlet cut. The latter allows
gluons, fermions and scalars to propagate in the loop indicated by a
dashed circle. The other configuration is obtained by flipping all
the helicity signs of the lines in this loop. All lines in this figure
are cut and carry on-shell momenta. The arrows in (a) refer to the direction
of momentum flow.}
\label{singlet}
\end{figure}
Now we continue with an explicit example of a maximal cut with only cubic vertices. Consider the two-loop four-point hepta-cut in the singlet configuration shown in \fig{singlet}(a). Although for analytical computations is not strictly necessary to solve the on-shell cut conditions, for illustrative purposes, we do so here. The seven on-shell conditions $l_i^2=0$, together with
momentum conservation at the six cubic vertex give a set of equations with the solution
\begin{equation}
\begin{array}{llll}
l_1=\xi \lambda_1 \widetilde{\lambda}_2 \,, & l_2=l_1-k_2\,, &l_3=l_1+k_1\,, \\
l_4= \lambda_4 \widetilde{\lambda}_3 \frac{\vphantom{\tilde A}\spa{3}.{l_2}}{\spa{l_2}.{4}}\,, & l_5=k_3+l_4\,,& l_6=l_4-k_4\,, & l_7=l_5-l_2\,,\\
\end{array} \label{ksolution}
\end{equation}
where $\xi$ is an arbitrary complex parameter, corresponding to the
remaining degree of freedom in the integration not frozen by the
hepta-cut. The procedure to obtain this solution is quite simple. First we note that $l_1$ has to be proportional to the external spinors of the vertices that it connects to, namely, $l_1\propto \lambda_1 \widetilde{\lambda}_2$. We may try to find the overall scale of this momentum, but we observe that there is one degree of freedom unfixed by this cut, therefore, we may account for this by introducing an arbitrary complex scale $\xi$ as in \eqn{ksolution}. Thus, we consider $l_1$ solved. Next we note that both $l_2$ and $l_3$ are on shell simply by momentum conservation. Therefore, we proceed by solving $l_4$. Similar to $l_1$ it has to take the form $l_4= \xi' \lambda_4 \widetilde{\lambda}_3 $. Again by momentum conservation, the neighboring momenta $l_5$ and $l_6$ become on-shell using this ansatz. However, the momentum $l_7$ is off shell for generic values of $\xi$ and $\xi'$. Indeed the condition $0=l_7^2=(k_3+l_4-l_2)^2$ is the equation that allows us to solve for one of these parameters, say $\xi'$. The reader may check that this equation is linear in $\xi'$ and that the solution takes the form given in \eqn{ksolution}.
The hepta-cut of \fig{singlet}(a), for external gluons, is given by
\begin{eqnarray}
A_4^{(2)}\Big|_{\rm 7\hbox{-}cut}&=&i^7 A^{\tag1}_{(1)} A^{\tag2}_{(2)}A^{\tag3}_{(3)}A^{\tag4}_{(4)}
A^{\tag5}_{(5)}A^{\tag6}_{(6)}\nonumber \\
&=&-i \frac{\spb{l_1}.{l_3}^3}{\spb{l_1}.{1}\spb{1}.{l_3}} \frac{\spa{l_1}.{l_2}^3}{\spa{l_1}.{2}\spa{2}.{l_2}} \frac{\spa{l_5}.{l_4}^3}{\spa{l_5}.{3}\spa{3}.{l_4}} \nonumber \\ && \times \frac{\spb{l_6}.{l_4}^3}{\spb{l_6}.{4}\spb{4}.{l_4}} \frac{\spb{l_2}.{l_5}^3}{\spb{l_2}.{l_7}\spb{l_7}.{l_5}} \frac{\spa{l_3}.{l_6}^3}{\spa{l_3}.{l_7}\spa{l_7}.{l_6}} \nonumber \\
&=& -i s^2 t \frac{\spa{1}.{4}^4}{\spa{1}.{2}\spa{2}.{3}\spa{3}.{4}\spa{4}.{1}} = -s^2tA^{\rm tree}_4(1^-,2^+,3^+,4^-) \,,
\label{singletstatement}
\end{eqnarray}
where $A^{\tag1}_{(i)}$ schematically denotes the three-point subamplitudes of \fig{singlet}(a), and $s=(k_1+k_2)^2$, $t=(k_2+k_3)^2$ are the familiar Mandelstam variables. The intermediate steps in this calculation can be checked numerically using the solution \eref{ksolution}, or by simply massaging the analytical expressions using basic properties of spinor products. (For example, in order to simplify seemingly complicated products of tree amplitudes, it may be useful to keep in mind that maximal cuts are always local objects in the loop momenta.) The reader may confirm that \eqn{singletstatement} holds for any (nonzero)
value of the arbitrary parameter $\xi$.
As \eqn{singletstatement} is a unitarity cut in the ${{\cal N}=4}$ theory, we should have summed over all particle states as well as helicities in the loops; however, as argued this particular kinematic configuration, with conveniently chosen external helicities, is a singlet cut. There is only a single purely gluonic contribution to the state sum! For a complete treatment of the two-loop four-point hepta-cut, one should also consider other kinematic solutions corresponding to other choices of chirality configurations for the cubic vertices. In general, this will result in non-singlet cuts, thus requiring the full supersymmetric multiplet of states. For such cases, it is useful to employ the four-dimensional on-shell superspace formalism of~\cite{Nair,GGK,FreedmanGenerating,DualSuperconformalSymmetry,KorchemskyOneLoop,AHCKGravity}; for additional details of on-shell superspace in multiloop cuts, see~\cite{SuperSum}. However, for the purpose of this review, it is sufficient to note that the cuts of the other kinematic solutions give the same result as \eref{singletstatement}, as shown by \cite{BuchbinderCachazo}.
We have now computed the hepta-cut in one of the possible channels of the planar two-loop four-point amplitude; this allows us to write up an ansatz for the Feynman-like diagram in this ``horizontal'' channel: a double box scalar integral $I^{\rm (P)}(s,t)$ times the coefficient we computed in the cut $-s^2tA^{\rm tree}_4(1^-,2^+,3^+,4^-)$ . There is a second possible channel corresponding a ``vertical'' double box integral $I^{\rm (P)}(t,s)$, see \fig{twoloopFigure}. The hepta-cut corresponding to this diagram can be computed the same way as the first hepta-cut \eref{singletstatement}. The result, which we will simply quote, is remarkably similar to \eref{singletstatement}, the only difference is a exchange $s \leftrightarrow t$ (note that the helicity amplitude $A^{(2)}_4(1^-,2^+,3^+,4^-)$ has no cyclic symmetry or flip symmetry that automatically enforces this). Finally, with the results of the maximal cuts in our hands we can write down an ansatz for the full planar amplitude:
\begin{equation}
A^{(2)}_4(1,2,3,4)=-A_{4}^{{\rm tree}}(1,2,3,4)\left(s^2 t I^{\rm (P)}(s,t)+s t^2 I^{\rm (P)}(t,s)\right)\,,
\label{2loopamp}
\end{equation}
where we have suppressed the helicity assignment $(1^-,2^+,3^+,4^-)$ that we have been considering, the coupling constant, and the usual planar color structure of adjoint particles. If, rather, we were to compute the maximal cuts of the remaining four point amplitudes, for other gluon helicity assignments $(1^\pm,2^\pm,3^\pm,4^\pm)$, or for the external fermions and scalars, the result is the same as \eref{2loopamp} (this fact can be attributed to special properties of the four-point function of {${\cal N}=4$~sYM}\ which hold in all dimensions $D\le10$, see \sect{RecursiveCutsSection}).
\begin{figure}[t]
\centerline{\epsfxsize 5 truein \epsfbox{twoloop.eps}} \caption{An ansatz for the planar four-point two-loop {${\cal N}=4$~sYM}\ amplitude in terms of the scalar integrals $I^{\rm (P)}(s,t)$ and $I^{\rm (P)}(t,s)$, as given by the computation of the maximal cuts. An overall factor of $-A^{\rm tree}_4(1,2,3,4)$ is suppressed. Other cuts, such as near maximal cuts, are needed to detect any terms missing from this ansatz. (However, no corrections are to be found as indeed this is the full planar amplitude~\cite{BRY}.) } \label{twoloopFigure}
\end{figure}
\subsection{On-shell three-vertices in higher dimensions.}
In $D>4$, the structure of a single three-point vertex is quite similar to the $D=4$ case, as the three momenta $p$, $q$, and $r$ can be embedded in a four-dimensional subspace (see \eref{embededPQ}) where the details are the same as above. However, the two-fold branching is no longer a meaningful notion in $D>4$ as one may smoothly connect the two chiral solutions of a given subspace by a rotation into the dimensions orthogonal to the subspace. Also the simple $D=4$ forms of the amplitudes cannot be used in four-dimensional subspaces, as the $D=4$ spinors (and amplitudes) transform nontrivially under the full $D$-dimensional Lorentz rotations. Instead, for the $D$-dimensional vertices one may simply use the fully covariant form of the amplitude, {\it e.g.} for three gluons,
\begin{equation}
A_3^{{\rm tree}}(1,2,3)=\varepsilon_1\cdot \varepsilon_2 ~ \varepsilon_3^\mu (k_1-k_2)_\mu~+~{\rm cyclic}\,,
\label{CovariantThreeVertex}
\end{equation}
which is simply the familiar Feynman gauge vertex contracted with the gluon polarizations.
Alternatively one may find useful the explicit generalizations of the spinor--helicity notation in $D=6$ \cite{Donal,SixDimSusy,SixDim} and in $D=10$ \cite{Donal10}.
Since much of the properties of the $D=4$ three-point vertex carries over to higher dimensions, it should also be true that there are particular choices of gluon polarizations that will remove the contributions of the fermion and scalar amplitudes. Indeed, choosing the polarization vector of the gluon to be proportional to the momentum of one of the neighboring lines is the correct $D$-dimensional generalization of \eref{vanishing},
\begin{eqnarray}
A_3^{{\rm tree}}(1_g,2_f,3_f)=A_3^{{\rm tree}}(1_g,2_s,3_s)=0\,, \hskip 1.5cm (\varepsilon_1^{\mu}\propto k_2^{\mu}) \nonumber \\
A_3^{{\rm tree}}(1_g,2_g,3_g)\neq 0\,.
\end{eqnarray}
This is easily seen to be true from the fermion vertex $\overline{\psi}_3 \varepsilon_1^\mu \gamma_\mu \psi_2$, because of the Dirac equation $k_2^\mu\gamma_\mu \psi_2=0$ , and from the scalar vertex $\varepsilon_1^\mu(k_2-k_3)_\mu$, because $k_i \cdot k_j=0$ for three-point kinematics.
It should be noted that in some cases it is convenient to carry out the method of maximal cuts (or other schemes of generalized unitarity) in four dimensions where the amplitudes entering the cuts are quite simple and, after arriving at a well-behaved ansatz, verify completeness using higher-dimensional spanning cuts. In particular, for four-particle amplitudes in the {${\cal N}=4$~sYM}\ theory, this strategy has been successfully carried out through four loops~\cite{CompactThree,Neq44np}.
\subsection{Cuts with higher-multiplicity vertices}
Maximal cuts do not only include generalized cuts with three-point vertices; according to the definition \eref{MCdef}, higher multiplicity vertices are allowed. This situation occurs, for example, in the five-point one-loop quadruple cut in four dimensions, displayed in \fig{NearMaxFigure}(a). The four-point vertex may contain a propagator corresponding to a pentagon integral, but the constraints of four-dimensional spacetime prohibits all five propagators to be on-shell simultaneously (for non-degenerate external momenta). The number of cut conditions is saturated by the one-loop quadruple cut in four dimensions, and thus quadruple cuts are considered maximal in this spacetime.
However, for $D>4$, the five-point one-loop amplitude evaluated on the penta-cut exists and is considered a maximal cut, and the quadruple cut, displayed in \fig{NearMaxFigure}(b), would be considered a next-to-maximal cut, as per the discussion in \sect{maxcutdefSection}.
\begin{figure}
\centerline{\epsfxsize 4.5 truein \epsfbox{NearMax.eps}} \caption{A quadruple cut in $D=4$ is a maximal cut, whereas a quadruple of the same five-point amplitude in $D>4$ is a near-maximal cut, or more precisely a next-to-maximal cut.} \label{NearMaxFigure}
\end{figure}
From a technical point of view, the process of carrying out maximal cuts and near-maximal cuts that involve higher-multiplicity subamplitudes is quite similar to the case of maximal cuts involving only three-point vertices. The main distinction is that higher-point vertices have no definite chirality, not even in four dimensions. This implies that the phase space of a four dimensional cut with many higher point vertices is in general smoother than four-dimensional cuts with only three-point vertices. The chiral branching of the four dimensional kinematics entering a cut is entirely due to the presence of four-dimensional on-shell thee-point vertices.
For an example of a near-maximal cut with a four-point subamplitude, we will consider the two-loop four-point amplitude again. We wish to show that the ansatz \eref{2loopamp} receives no more corrections at the level of the next-to-maximal hexa-cuts. First we list the different cut topologies that can arise at the next-to-maximal level. Starting from the maximal cut topology in \fig{singlet}, we may relax the cut condition of any one of the on-shell lines. This gives the three cut topologies in shown in \fig{nearmaximalFigure}. As before, these cuts have to be considered for all possible kinematic solutions (assignment of the vertex chirality) and for all possible channels (cyclic and anticyclic permutations of the external states). We will work out the \fig{nearmaximalFigure}(a) hexa-cut explicitly. Note that the vertex chirality and labeling is the same as for \fig{singlet}(a) except that momentum $l_7$ is no longer on-shell. We may recycle the kinematic solution \eref{ksolution}, with the only difference that $\xi'$ is now arbitrary, and thus $l_7$ is off-shell:
\begin{equation}
\begin{array}{llll}
l_1=\xi \lambda_1 \widetilde{\lambda}_2 \,, & l_2=l_1-k_2\,, &l_3=l_1+k_1\,, \\
l_4= \xi' \lambda_4 \widetilde{\lambda}_3 \,, & l_5=k_3+l_4\,,& l_6=l_4-k_4\,, & l_7=l_5-l_2\,.
\end{array} \label{ksolution2}
\end{equation}
The hexa-cut is given by
\begin{eqnarray}
A_4^{(2)}\Big|_{\rm 6\hbox{-}cut}\hskip-2mm&=&
i^6 A^{\tag1}_{(1)} A^{\tag2}_{(2)}A^{\tag3}_{(3)}A^{\tag4}_{(4)}
A^{\tag5}_{(5')}\nonumber \\&=&
-i \frac{\spb{l_1}.{l_3}^3}{\spb{l_1}.{1}\spb{1}.{l_3}} \frac{\spa{l_1}.{l_2}^3}{\spa{l_1}.{2}\spa{2}.{l_2}}\frac{\spa{l_5}.{l_4}^3}{\spa{l_5}.{3}\spa{3}.{l_4}} \frac{\spb{l_6}.{l_4}^3}{\spb{l_6}.{4}\spb{4}.{l_4}} \frac{\spa{l_3}.{l_6}^3}{\spa{l_3}.{l_2}\spa{l_2}.{l_5}\spa{l_5}.{l_6}} \nonumber \\
&=& -i s^2 t \frac{1}{l_7^2} \frac{\spa{1}.{4}^4}{\spa{1}.{2}\spa{2}.{3}\spa{3}.{4}\spa{4}.{1}}\nonumber \\
&=& -s^2t \frac{1}{l_7^2} A^{\rm tree}_4(1^-,2^+,3^+,4^-) \,,
\label{nexttomax}
\end{eqnarray}
where again $A^{\tag1}_{(i)}$ schematically denotes the three- and four-point subamplitudes of the cut, in this case \fig{nearmaximalFigure}(a). The result is surprisingly similar to the hepta-cut \eref{singletstatement}, the only difference is the appearance of the propagator $1/l_7^2$. This a positive result showing that our ansatz for the amplitude passed the test. Indeed the reader may check that the same hexa-cut on the scalar double box integral in our convention \eref{cutdef2} is
\begin{equation}
I^{\rm (P)}(s,t)\Big|_{\rm 6-cut} = \frac{1}{l_7^2} \,,
\end{equation}
explaining the appearance of this propagator factor. Finally, we note that evaluating all the hexa-cuts of \fig{nearmaximalFigure} for the various kinematic solutions and channels, informs us that the ansatz \eref{2loopamp} receives no more corrections at the next-to-maximal cut level. This is a strong indication that the ansatz for the amplitude may be complete. To ensure that the ansatz is correct, one would have to evaluate it on a complete set of cuts in $D$ dimensions, such as the two- and three-particle unitarity cuts (done in ref.~\cite{BRY} in $D=4$). This we will not do explicitly here (although the next section discusses the two-particle cut); we simply note that \eref{2loopamp} is indeed the full planar amplitude~\cite{BRY}.
\begin{figure}
\centerline{\epsfxsize 6 truein \epsfbox{nearmaximal.eps}} \caption{ Planar next-to-maximal cuts of the four-point two-loop amplitude. } \label{nearmaximalFigure}
\end{figure}
\section{Some recursive $D$-dimensional cuts}
\label{RecursiveCutsSection}
In the previous section, we discussed unitarity cuts which rely on cutting all or almost all propagators of every internal loop. As a contrast, here we will consider cuts where many of the propagators are left uncut. Specifically, will review two types of very useful cuts, particular to four-point {${\cal N}=4$~sYM}, namely the well-known ``two-particle cut'' of the four-point amplitude~\cite{BRY,BDDPR}, and the more recent ``box cut''~\cite{Neq44np}; both cuts can be expressed very simply, and in all generality, in $D$ dimensions, in terms of lower-loop expressions. These cuts are useful for computing contributions to the $D=4-2\epsilon$ dimensionally regulated amplitudes; in this case the regularization scheme implied is the four-dimensional-helicity (FDH) version of dimensional reduction~\cite{DimensionalReductionScheme}. And similarly, for amplitudes in $D\le 10$, the corresponding theory implied is the dimensional reduction of the ten dimensional ${\cal N}=1$ sYM theory.
Before introducing the two cuts, we will review a structure present in the four-point amplitudes of {${\cal N}=4$~sYM}, which plays an important role in the subsequent discussion, due to the especially simple dependence on the external states. It is known that all {${\cal N}=4$~sYM}\ four-point amplitudes can be expressed in a common factorized form~\cite{Neq44np}
\begin{equation}
{\cal A}_4^{(L)}(1,2,3,4)
= g^{2L + 2} {\cal K}(1,2,3,4) \, U^{(L)}(1,2,3,4)\,,
\label{FourPointFactorization}
\end{equation}
where ${\cal A}_4$ represents the full color-dressed $D\le10$-dimensional amplitude ({\it cf.} the color-stripped partial amplitude $A_4$), the kinematic prefactor ${\cal K}$ is the same in all loop-diagrams, and the loop-specific factor $U$ is independent of the states of the external particles. The $D$-dimensional {${\cal N}=4$~sYM}\ amplitudes are defined as amplitudes of the ten-dimensional ${\cal N}=1$ sYM theory dimensionally reduced to $D$ dimensions.
All of the ${{\cal N}=4}$ state dependence in ${\cal A}_4$ ($16^4$ configurations of states) is carried by the kinematic prefactor
\begin{equation}
{\cal K}(1,2,3,4) \equiv s_{12} \, s_{23} \,A_4^{{\rm tree}}(1,2,3,4)\,,
\label{Kprefactor}
\end{equation}
where $A_4^{{\rm tree}}(1,2,3,4)$ is the canonically color-ordered $D$-dimensional tree amplitude. Note that multiplying the tree by $s_{12} s_{23}=(k_1+k_2)^2(k_2+k_3)^2$ renders the resulting expression fully crossing symmetric; hence, the ordering of the arguments of ${ \cal K}$ indicates no specific color order.
In four dimensions, a compact form of ${ \cal K}$ can be written down using anti-commuting parameters $\eta$ in
the four-dimensional on-shell superspace formalism~\cite{Nair,GGK,FreedmanGenerating,DualSuperconformalSymmetry,KorchemskyOneLoop,AHCKGravity},
\begin{eqnarray}
A_{4}^{{\rm tree}}(1,2,3,4)\Bigr|_{D=4} &=&
{i\delta^{(8)}(\lambda_1 \eta_1+\lambda_2 \eta_2
+\lambda_3 \eta_3+\lambda_4 \eta_4)
\over \spa{1}.{2}\spa{2}.{3}\spa{3}.{4}\spa{4}.{1}} \,,
\label{SuperAmplitude} \\
{\cal K}(1,2,3,4)\Bigr|_{D=4} &=&
-i \, {\spb1.2\spb3.4 \over \spa1.2\spa3.4}
\, \delta^{(8)}(\lambda_1 \eta_1+\lambda_2 \eta_2
+\lambda_3 \eta_3+\lambda_4 \eta_4) \,.
\label{SuperAmplitudeII}
\end{eqnarray}
It is not difficult to verify that $(\spb1.2\spb3.4)/(\spa1.2\spa3.4)$
is symmetric under exchange of any two legs. Similar compact superspace expressions for the four-point amplitude exist in $D=6$~\cite{SixDimSusy} using generalized spinor-helicity notation~\cite{Donal}, and in $D=10$~\cite{Mafra} using Berkovits' pure spinor formalism~\cite{Berkovits}.
Related to this, ${\cal K}$ represents the color-stripped four-point (linearized)
matrix elements of the local operator $\Tr F^4$, plus its
supersymmetric partners. Therefore, ${\cal K}$ is a natural prefactor
to extract from the four-point amplitude in {${\cal N}=4$~sYM}.
All of the color dependence of the amplitude is carried by the external-state-independent ``universal factor'' $U^{(L)}$. This factor at tree level, $U^{(0)}$, is given by~\cite{Neq44np},
\begin{equation}
{U}^{(0)}(1,2,3,4) =
\biggl(
\frac{\tilde f^{a_4a_1b}\tilde f^{ba_2a_3}} {s_{12} s_{23}}
+\frac{\tilde f^{a_3a_1b}\tilde f^{ba_2a_4}} {s_{12} s_{13}}
\biggr)\,.
\label{TreeUniversalFactor1}
\end{equation}
In general, the universal factor $U^{(L)}$ is a sum of $L$-loop integrals. The integrands entering $U^{(L)}$ are rational functions of momentum invariants involving the loop and external momenta. The universal factors for $L=1,2,3,4$, including planar and non-planar contributions, may be found by in~\cite{Neq44np}, and planar $L=5$ in~\cite{FiveLoop}.
The four-point factorization \eref{FourPointFactorization} makes it possible to immediately write down simple $D$-dimensional expression for cuts involving this amplitude, usually in terms of lower-loop numerator factors. We will now review two related special classes of cuts, which owe their simple structure to this property. These cuts can be applied to both planar and non-planar amplitudes, making them especially powerful.
\subsection{Two-particle cuts of the four-point {${\cal N}=4$~sYM}\ amplitude}
\label{TwoParticleCutSection}
\begin{figure}[tbh]
\centerline{\epsfxsize 1.9 truein \epsfbox{TwoParticleCut.eps}}
\caption[a]{\small A two-particle cut that may be used to construct
contributions to the $(L+L'+1)$-loop amplitude from those at
$L$ and $L'$ loop orders.}
\label{TwoParticleCutFigure}
\end{figure}
A regular {\it two-particle cut} of a multiloop four-point amplitude has the form shown in \fig{TwoParticleCutFigure} --- it cuts the amplitude into two lower-loop four-point amplitudes. Using the factorization \eref{FourPointFactorization}, it will be possible to write down the cut in terms of lower-loop universal factors, thus straightforwardly determining a class of integrals in the amplitude that are detectable in this cut. The result of this exercise
will be valid in $D$ dimensions, whenever the lower-loop universal factors are valid in $D$ dimensions.
We evaluate the generic two-particle color-dressed cut depicted in \fig{TwoParticleCutFigure}. It cuts the $(L+L'+1)$-loop
amplitude ${\cal A}_4^{(L+L'+1)}(1,2,3,4)$ into the two four-point amplitudes ${\cal A}_4^{(L)}(-l_1,1,2,l_2)$ and ${\cal A}_4^{(L')}(-l_2,3,4,l_1)$ of loop orders $L$ and $L'$,
\begin{equation}
{\cal A}_4^{(L+L'+1)}(1,2,3,4)\Big|_{\rm 2\hbox{-}cut}
\hskip-2mm = i^2 \, \sum_{{{\cal N}=4} \atop {\rm states}}
{\cal A}_4^{(L)}(-l_1,1,2,l_2) \, {\cal A}_4^{(L')}(-l_2,3,4,l_1) \,,
\label{GenLoopTwoParticleCut}
\end{equation}
where the state sum is over the particles with momenta $l_1$ and $l_2$.
Using the factorization~(\ref{FourPointFactorization}) and the state-independence of $U^{(L)}$, we can immediately rewrite the cut as follows:
\begin{eqnarray}
{\cal K}(1,2,3,4) \, U^{(L+L'+1)}(1,2,3,4)\Big|_{\rm 2\hbox{-}cut}
\hskip-2.5mm &=& i^2
\, U^{(L)}(-l_1,1,2,l_2) \,
U^{(L')}(-l_2,3,4,l_1) \nonumber\\
&& \hskip-1mm \times
\sum_{{{\cal N}=4} \atop {\rm states}}
{\cal K}(-l_1,1,2,l_2) \, {\cal K}(-l_2,3,4,l_1) \,. \nonumber\\
\label{GenLoopTwoParticleCutRewrite}
\end{eqnarray}
To evaluate this, we use the sewing relation between two four-point color-ordered {${\cal N}=4$~sYM}\ trees~\cite{BRY,BDDPR},
\begin{equation}
\sum_{{{\cal N}=4} \atop \rm states}
\hskip-1mm A_4^{(0)}(-l_1,1,2,l_2) A_4^{(0)}(-l_2,3,4,l_1)
=\hskip-1mm - i s_{12} s_{23} \, A_4^{(0)}(1,2,3,4) {1\over s_{2l_2} s_{4l_1}} \,.
\label{SewingRelation}
\end{equation}
This sewing relation is valid in any dimension $D$ and for any external states in the ${{\cal N}=4}$ multiplet. A straightforward way to confirm \eqn{SewingRelation} is to work in $D=10$ and evaluate the sum over states in components, using the fact that in $D=10$ {${\cal N}=4$~sYM}\ is equivalent to an ${{\cal N}=1}$ theory composed of a gluon and a gluino. By dimensional reduction the sewing relation~(\ref{SewingRelation}) then holds in any dimension $D\le 10$. Recently, this equation has also been verified directly in six dimensions using an on-shell superspace~\cite{SixDimSusy}.
Substituting in the relation $A_4^{(0)}={\cal K}/(s_{12}s_{23})$ given in \eref{FourPointFactorization}, we find that \eref{SewingRelation} is equivalent to
\begin{eqnarray}
\sum_{{{\cal N}=4} \atop {\rm states}}
{\cal K}(-l_1,1,2,l_2) \, {\cal K}(-l_2,3,4,l_1)&=& -i\, s_{12}^2 \, {\cal K}(1,2,3,4)\,.
\label{KSewingRelation}
\end{eqnarray}
\begin{figure}[tb]
\centerline{\epsfxsize 5.8 truein \epsfbox{TwoParticleCutConstruct.eps}}
\caption[a]{\small Sample contributions to the full color-dressed
two-particle cut for $L=2$ and $L'=1$. The diagrams on the
right-hand side show some of the terms in
$U^{(4)}(1,2,3,4)$ that are constructed from these
cuts, using $U^{(2)}$ and $U^{(1)}$. The explicit
color factors, as well as factors of $i$, have been omitted.}
\label{TwoParticleCutConstructFigure}
\end{figure}
Applying \eqn{KSewingRelation} to \eqn{GenLoopTwoParticleCutRewrite}, we find the key equation for building all contributions from two-particle cuts directly in terms of the $U$s:
\begin{equation}
U^{(L+L'+1)}(1,2,3,4)\Big|_{\rm 2\hbox{-}cut}
= i \, s^2_{12}\, U^{(L)}(-l_1,1,2,l_2)
\, U^{(L')}(-l_2,3,4,l_1) \,.
\label{TwoParticleUSewing}
\end{equation}
\Eqn{TwoParticleUSewing} is rather powerful. No complicated calculations remain in order to obtain all contributions visible in two-particle cuts; they are given simply by taking the product of lower-loop results. The color-dressed $U^{(L+L'+1)}$ is given immediately as a sum over products of individual integrals residing inside the $U^{(L)}$ and $U^{(L')}$ factors, up to terms that vanish because of the on-shell conditions, $l_1^2 = l_2^2 = 0$.
\Fig{TwoParticleCutConstructFigure} illustrates diagrammatically some of the terms generated by \eqn{TwoParticleUSewing} for the case $L=2$ and $L'=1$. For simplicity, we draw only the planar contributions of $U^{(2)}(-l_1,1,2,l_2)$, omitting the color factor $\tilde{f}^{abc}$s as well as all factors of $i$ (which can be restored at one's convenience). The denominator factors in $U^{(2)}$ and $U^{(1)}$ correspond to the un-cut propagators that are graphically visible on the left- and right-hand sides of the dashed line, respectively. Similarly, the numerator factor for each parent graph on the right-hand side of \fig{TwoParticleCutConstructFigure} is given by forming the product of the numerator factors for the one-loop box diagram and two-loop double box (taking into account the proper permutation of legs), and then multiplying by two powers of $s_{12}$.
\begin{figure}[t]
\centerline{\epsfxsize 3.3 truein \epsfbox{rungrule.eps}}
\caption[a]{\small The rung rule~\cite{BRY,BDDPR}
for generating a higher-loop contribution by inserting vertical rung
between two lines inside a loop diagram of {${\cal N}=4$~sYM}\ amplitudes. The dimension of the extra propagator is compensated by the extra factor $i(p+q)^2$ in the numerator. This rule reproduces the numerators of planar diagrams generated by iterated two-particle cuts.}
\label{rungruleFigure}
\end{figure}
The simple formula~\eqn{TwoParticleUSewing} for the two-particle cuts provides a rather useful tool for generating many higher-loop contributions by recycling lower-loop ones. Indeed, it is the mechanism behind the many of the diagrams given by the well-known ``rung rule''~\cite{BRY,BDDPR}, shown in \fig{rungruleFigure}. Unlike the two-particle cut formula, the rung rule is a heuristic rule that should be applied with some care; it can potentially suggest non-contributing terms when the two-particle cut does not apply~\cite{BCDKS}. That warning aside, for planar {${\cal N}=4$~sYM}, the rung rule is rather powerful. For example, it is quite interesting that the rung rule generates terms that are invariant under the dual conformal symmetry of {${\cal N}=4$~sYM}~\cite{BCDKS,FiveLoop}, presaging modern awareness of this symmetry.
\subsection{Box cuts of multiparticle {${\cal N}=4$~sYM}\ amplitudes}
\label{BoxCutSection}
The two-particle cut has a particularly nice generalization that also relies on the simple structure of the four-point amplitude of {${\cal N}=4$~sYM}.
This cut can be thought of as a (generalized) four-particle cut that isolates a four-point subamplitude, which in the simplest nontrivial case is a box subdiagram; thus, the cut was named ``box cut'' in~\cite{Neq44np}. A crude version of this cut appeared already~\cite{FiveLoop} as a ``box-substitution rule''. It allowed the
construction of $L$-loop contributions with a box subgraph, starting from $(L-1)$-loop contributions with a contact interaction, as illustrated in \fig{BoxSubRuleFigure}. Similar rules were discussed in conjunction with the leading singularity method in~\cite{CachazoSkinner}. Unlike the two-particle cut discussed in the previous section, the box cut is applicable to multiparticle amplitudes. Now we will review the derivation of the $D$-dimensional box cut.
\begin{figure}[t]
\centerline{\epsfxsize 4.7 truein \epsfbox{BoxSubRule.eps}}
\caption[a]{\small The box-substitution rule~\cite{FiveLoop}
for generating a higher-loop contribution by inserting a one-loop
four-point box subintegral into a four-point vertex. In
this example, we substitute a box into the central four-point vertex
in the four-loop ``window'' diagram. The result is a five-loop
integral that cannot be obtained from two-particle cuts or the rung rule. }
\label{BoxSubRuleFigure}
\end{figure}
Consider the generalized cut of an $L$-loop $n$-point amplitude,
\begin{equation}
{\cal A}_n^{(L)}\Big|_{\rm box~cut} \equiv
\sum_{{{\cal N}=4} \atop \rm states}
{\cal A}_{(1)}\cdots {\cal A}_{4, (i)}^{(L')} \cdots {\cal A}_{(m)} \,,
\label{BoxCut}
\end{equation}
that is composed of a generic set of color-dressed amplitude factors, except for the $i^{\rm th}$ such factor, which we take to be a color-dressed $L'$-loop four-point subamplitude, ${\cal A}_{4, (i)}^{(L')}$. Example of such box cuts are given in \fig{GeneralBoxCutFigure}.
\begin{figure}[t]
\centerline{\epsfxsize 4.5 truein \epsfbox{GeneralBoxCut.eps}}
\caption[a]{\small Two examples of multiparticle multiloop ``box
cuts''. They reduce to lower-loop cuts by replacing the red
four-point subamplitudes by four-point color-ordered trees,
multiplied by known numerator and denominator factors.
This property allows these cuts to be computed easily in $D$
dimensions, once the amplitudes with fewer loops are
known. White holes represent loops and the darker (red)
subamplitudes mark four-point amplitudes amenable to reduction.}
\label{GeneralBoxCutFigure}
\end{figure}
As mentioned, the $L'$-loop four-point subamplitude of {${\cal N}=4$~sYM}\ is special because of the factorization property
(\ref{FourPointFactorization}). Labeling the cut momenta by $l_1,l_2,l_3,l_4$, we have
\begin{equation}
{\cal A}_{4,(i)}^{(L')}(l_1,l_2,l_3,l_4)
= A_{4,(i)}^{(0)}(l_1,l_2,l_3,l_4)
\, (l_1+l_2)^2 (l_2 + l_3)^2
\, U^{(L')}(l_1,l_2,l_3,l_4) \,, \,
\label{FourPointFactorizationBoxCut}
\end{equation}
where, as in the previous section, we use ${\cal A}_4$ to represent the color-dressed amplitude, and only the color-ordered tree amplitude factor $A_{4,(i)}^{(0)}$ depends on the states crossing the cuts. Therefore, we can pull the factor $U^{(L')}$ out of the sum over states in \eqn{BoxCut}, leading to a simpler expression in the summand:
\begin{equation}
{\cal A}_n^{(L)}\Big|_{\rm box \atop cut}\hskip-1mm =(l_1+l_2)^2 (l_2 + l_3)^2
U^{(L')}(l_1,l_2,l_3,l_4)
\, \sum_{{{\cal N}=4} \atop \rm states}
{\cal A}_{(1)}\cdots A_{4, (i)}^{(0)} \cdots {\cal A}_{(m)} \,.\,
\label{BoxCutTrue}
\end{equation}
The state-sum is identical to a lower-loop cut, that of the $(L-L')$-loop amplitude, but utilizing the color-ordered contribution to the $i^{\rm th}$ tree. This fact immediately gives a simple relation between the $L$-loop box cut and contributions to the reduced $(L-L')$-loop cut under the same cut conditions.
We can formally write down an equation relating the cut of an $L$-loop amplitude to a cut of a lower-loop one as,
\begin{equation}
{\cal A}_n^{(L)}\Big|_{\rm box~cut}
= (l_1+l_2)^2 (l_2+l_3)^2
\,U^{(L')}(l_1,l_2,l_3,l_4) \,
\tilde{{\cal A}}_n^{(L-L')}\Big|_{\rm cut} \,.
\label{BoxCutReduction}
\end{equation}
We introduced the reduced cut $\tilde{{\cal A}}$ notation to emphasize that the state-sum in \eqn{BoxCutTrue} is exactly a $(L-L')$ loop unitarity cut which is color-dressed with $\tilde f^{abc}$ everywhere, except for the four-point color-ordered tree amplitude whose associated color factors are accounted for in the $L'$-loop universal factor $U^{(L')}$.
\begin{figure}[t]
\centerline{\epsfxsize 5.5 truein \epsfbox{boxcut2.eps}}
\caption[a]{\small Practical application of the box cut to determine the
numerator $N$ for a four-loop integral. The (red) dashed cut conditions around the
upper left box in this diagram allow us to replace it by the product
of a reduced cut diagram, some kinematic factors and a box
integral. The reduced cut diagram is then expanded into two
three-loop ``tennis-court'' diagrams, corresponding to the two
allowed channels of the marked four-point vertex. The relevant
kinematic pieces of the tennis-court diagrams, {\it i.e.} the
numerators and the spurious propagators, are extracted from the
known three-loop contribution. Assembling all the kinematic
factors gives the result for $N$ (with the overall factor ${\cal K}$ suppressed), which is free of spurious
propagators.}
\label{boxcutFigure}
\end{figure}
\Fig{boxcutFigure} gives an example illustrating that the box cut is very a practical tool to work with. Using a back-of-the-envelope calculation, one can determine the numerator polynomial of a four-loop integral~\cite{BCDKS,Neq44np}, using the known three-loop four-point amplitude~\cite{BRY,BDS, CompactThree}. Although this particular example is planar, it is just as simple to use the box cut for non-planar contributions. For example, inserting the box into the four-point vertex in \fig{boxcutFigure} in a non-planar fashion generates non-planar four-loop integrals.
The box cut is an extremely efficient way to obtain contributions to parent graphs that contain a lower-loop four-point subgraph. As mentioned earlier, the $L'=1$ box cut is closely related to the box-substitution rule depicted in~\fig{BoxSubRuleFigure}. The box cut also generates contributions that are consistent with the two-particle cut and/or the rung rule~\cite{BRY}.
From the above covariant derivation, it follows that box cuts are valid in any dimension, if both the reduced cut $\tilde{{\cal A}}_n^{(L-L')}\big|_{\rm cut}$ and the four-point universal factor $U^{(L')}$ are known in $D$ dimensions. As a practical matter, the universal factors entering the lower-loop amplitudes should already be known in $D$ dimensions, prior to attempting the higher-loop calculation in $D$ dimensions.
\section{Duality between color and kinematic factors}
\label{BCJ}
In modern S-matrix schemes, one successfully constructs multiparticle and multiloop amplitudes using minimal input from the theory as defined by the Lagrangian. Rather than employing the nuts and bolts of the action, one relies heavily on quantum field theoretic consistency. For example, a wealth of tree-level amplitudes for interesting theories can be fully constructed through BCFW~\cite{BCFW} using only the on-shell three-point amplitude as input. Similarly, as discussed in \sect{MaximalCutsSection} and in greater detail in \cite{ZviYutinReview}, the unitarity method allows the construction of all loop level contributions using as input only lower-loop subamplitudes. A natural question is whether one can further minimize the theory-specific information needed to construct a full color-dressed scattering amplitude. The conjectured duality between color and kinematic factors~\cite{BCJ, BCJloops} described in this section suggests such an approach.
The duality relies upon the cubic-graph representation of gauge theory amplitudes, introduced earlier \eref{cubicRep}, which we write again as
\begin{equation}
{\cal A}_n^{(L)}= i^L g^{n-2 +2L}\sum_{i\in\Gamma_3} \, \int \frac{d^{LD}\ell}{(2\pi)^{LD}} \frac{1}{S_i} \, \frac{N_i C_i}{p_{i_1}^2p_{i_2}^2p_{i_3}^2\cdots p_{i_{m}}^2}\,.
\label{LoopBCJ}
\end{equation}
This time we have made the dependence on the coupling constant explicit, and pulled out an overall phase relative to \eref{cubicRep}. We note that in general the $N_i$ cannot be uniquely specified; if calculated by Feynman diagrams, they are gauge choice dependent. If calculated by other means, they are allowed to transform under a generalized gauge transformation~\cite{BCJ, BCJloops,LagrangianSquare}, which corresponds to the ability to move contact terms between different graphs, while leaving the amplitude invariant. This freedom allows for many different representations of the same amplitude, some of which make special properties of the theory manifest.
One very special property is the duality between color and kinematics. This duality asserts the existence of representations \eref{LoopBCJ} where the kinematic numerators $N_i$ obey the same general algebraic properties as the color factors $C_i$. Specifically, the $N_i$ obey Jacobi relations and have antisymmetric cubic vertices analogous to the color factors, schematically:
\begin{eqnarray}
N_i+N_j+N_k=0~~ &\Leftrightarrow~~ C_i+C_j+C_k=0&\,,~~~~~~~~{\rm (Jacobi~identity)} \label{DefiningJacobi} \\
~~~~~~~~~~N_{i}\rightarrow -N_{i}~ &\Leftrightarrow~~C_{i} \rightarrow -C_{i}\,& ~~~\mbox{(vertex-flip~antisymmetry)} \nonumber
\end{eqnarray}
where the first line signifies the Jacobi identity valid for specific triplets of graphs in the amplitude, and the second line represents the action of flipping the ordering of a cubic vertex in a graph. In addition, it is often convenient at loop-level to impose the self-symmetries or graph automorphisms of the $N_i$, similar to the self-symmetries obeyed by the color factors $C_i$.
The duality between color and kinematics was first proposed by Bern and the present authors at tree-level~\cite{BCJ}, and later conjectured to hold to all loop orders~\cite{BCJloops} in generic gauge theories. It generalizes a four-point property of tree-amplitudes observed in cuts of multi-loop-amplitudes~\cite{BCJ}, and also noticed in studies of zeros of scattering amplitudes in the 1980s~\cite{ZhuGHL}.
Many of the the interesting and surprising consequences of this duality will not be treated in detail, as they are outside the scope of this review. Nevertheless, in~\sect{TLGSection} we will briefly comment on some of these consequences, as related to tree amplitudes, Lagrangian manifestation, and constructing gravity amplitudes. For the main body of this discussion, we will instead focus on how the duality can be used in the explicit construction of multiloop gauge theory amplitudes. We will see that the constraints imposed by the duality allow us to drastically reduce the necessary input from unitarity cuts. (As long as the duality remains a conjecture, it is of course necessary to verify any obtained amplitude on a set of spanning cuts.) We may loosely quantify the amount of theory-specific information required for the construction of an amplitude by counting the number of necessary maximal and near-maximal cuts. Alternatively, one may count the graphs used to represent the full amplitude and compare this number to the very small number of graph numerators not completely fixed by the kinematic Jacobi relations~\eref{DefiningJacobi}. We call the latter ones ``master graphs''; from these, all other contributing graph numerators can be expressed using linear relations. In \sect{ThreeLoopExampleSection}, the four-point three-loop {${\cal N}=4$~sYM}\ amplitude, comprised of 12 graphs, is constructed using a single master graph. The numerator of this single master graph, after imposing all symmetries and duality relations, is completely fixed by a single maximal cut, illustrating the dramatic simplification involved.
Nontrivial evidence supporting the duality conjecture at loop level is found by studying the four-point amplitude of the {${\cal N}=4$~sYM}\ theory at one~\cite{GSB}, two~\cite{BRY}, three~\cite{BCJloops}, and four loops~\cite{fourLoopBCJ}, and for the five-point amplitude through three loops~\cite{FivePointBCJ}. Evidence that the duality holds even in nonsupersymmetric theories is provided~\cite{BCJloops} by the two-loop four-gluon identical-helicity pure Yang-Mills amplitude~\cite{BDKtwofourqcd}.
We will elaborate on the duality using graph operators, but first it will be instructive to look at an venerable example from the literature where the duality \eref{DefiningJacobi} is known to hold.
\begin{figure}
\centerline{\epsfxsize 5 truein \epsfbox{pentatriangle.eps}} \caption{ A Jacobi identity between three cubic four-point two-loop topologies. These three graphs have the same connectivity except for the shaded (pink) edge. Every Jacobi identity and kinematic identity between numerator factors can be graphically seen as a relation local to a set of three edges. For the {${\cal N}=4$~sYM}~amplitude, only graphs ${(\rm P)}$ and ${(\rm NP)}$ contribute~\cite{BRY}. } \label{twoGraphs}
\end{figure}
\subsection{Two-loop example}
An instructive nontrivial example can be found using the four-point two-loop ${\cal N}=4$ sYM amplitude considered previously in \sect{MaximalCutsSection}. As we have already shown how to construct the planar part using other methods, here we merely verify that the full amplitude, given in the literature{~\cite{BRY}}, satisfies the color--kinematics duality. In the subsequent discussion we will show how such amplitudes can be directly constructed using the duality.
The two-loop amplitude is
\begin{equation}
{\cal A}^{(2)}_{4}=-g^6 \Bigl(N^{(\rm P)}C_{1234}^{(\rm P)}I^{(\rm P)}(s,t) +N^{(\rm NP)}C_{1234}^{(\rm NP)} I^{(\rm NP)}(s,t)~+{\rm perms}\Bigr)\,,
\label{TwoLoopAmpEqn}
\end{equation}
where the two scalar integrals contributing are drawn in~\fig{twoGraphs}, and the
``perms'' instructs us to sum over all distinct permutations of external labels. For convenience we have pulled the kinematic (and color) factors outside the integral, as they do not depend on the loop momenta. The numerators $N^{(\rm P)}$ and $N^{(\rm NP)}$ are given by
\begin{equation}
N^{(\rm P)}=N^{(\rm NP)}= {\cal K} \, (k_1+k_2)^2 \,,
\label{Ndef}
\end{equation}
where ${\cal K}$ is the state-dependent prefactor introduced in \eref{Kprefactor}.
The color factors $C^{(\rm P)}$ and $C^{(\rm NP)}$, obtained by \eref{ColorDef},
are related through the Jacobi relation shown in \fig{twoGraphs}, to a penta-triangle graph $C^{(\rm PT)}$,
\begin{equation}
C^{(\rm P)}=C^{(\rm NP)}+ C^{(\rm PT)}\,.
\end{equation}
If \eref{TwoLoopAmpEqn} satisfies the color--kinematics~duality, we should expect a similar relation for the kinematic factors:
\begin{equation}
N^{(\rm P)}=N^{(\rm NP)}+ N^{(\rm PT)}\,.
\label{NJacobi}
\end{equation}
This indeed holds. The amplitude \eref{TwoLoopAmpEqn} does not involve any penta-triangle integral; therefore, we set $N^{(\rm PT)}=0$, and thus \eref{NJacobi} is consistent with the result in \eref{Ndef}.
\subsection{Graph operators and the color--kinematics~duality}
In order to phrase the duality precisely, we will now introduce graph operators. This is not an exercise in pedantry. Indeed, precisely defining the duality in terms of graph operators allows for natural automation---especially useful when considering higher-loop contributions that may involve many graph topologies.
The color Jacobi identity directly relates graphs that share entirely the same connectivity except for one edge, {\it e.g.} the shaded (pink) edge in \fig{twoGraphs}. The language we will use to specify this relation will involve operators that take a graph and a particular edge of that graph, and rearrange that edge's connectivity so as to get another graph. We can specify any internal edge $e$ of a cubic graph in terms of the four edges that touch its bounding nodes. If the order of edges around one node is $(a,b,e)$, and the order of edges around the other node is $(-e,c,d)$, we can specify the edge $e$ as $e=\{(a,b),(c,d)\}$. With this terminology we are ready to consider some mapping $\hat{t}$ to take a graph~${\cal G}$ and one of its edges $e$, to another graph ${\cal G}_{t(e)}$:
\begin{equation}
\hat{t}\big ({\cal G},\, e
\big ) = {\cal G}_{t(e)}\,,
\end{equation}
where ${\cal G}_{t(e)}$ has the same connectivity of ${\cal G}$ except the nodes $(a,b,e)$ and $(-e,c,d)$ have been replaced with $(d,a,t)$ and $(-t,b,c)$, thereby defining the edge $t=\{(d,a),(b,c)\}$. We will also use
\begin{equation}
\hat{u}\big ({\cal G},\, e) = {\cal G}_{u(e)}\,,
~~~~~~~
\hat{s} \big ({\cal G},\, e \big )= {\cal G}\,,
\end{equation}
where ${\cal G}_{u(e)}$ has the same connectivity of ${\cal G}$ except nodes $(a,b,e)$ and $(-e,c,d)$ have been replaced with $(d,b,u)$ and $(-u,c,a)$, with edge $u=\{(c,a),(d,b)\}$. The $\hat{s}$ operator simply returns the original graph unchanged independent of the edge specified. These operations are depicted visually in \fig{GraphOperators}.
\begin{figure}
\centerline{\epsfxsize 4.8 truein \epsfbox{graphOp2.eps}} \caption{Three graph operators taking $($graph ${\cal G}$ , edge $e$$)$ to graphs ${\cal G}_{t_e}, {\cal G}_{u_e}$ and ${\cal G}$. All three graphs have the same connectivity, except for the edge touching edges ($a,b,c,d$).} \label{GraphOperators}
\end{figure}
We are now ready to concisely specify the color Jacobi identity associated with every edge $e$ of every graph $\cal G$,
\begin{eqnarray}
0&= C \left[ \hat{s} \big ( {\cal G}, e \big )\right]-C \left[ \hat{t} \big ( {\cal G}, e \big )\right] -C \left[ \hat{u} \big ( {\cal G}, e \big )\right] \nonumber \\
&=C[{\cal G}]- C[ {\cal G}_{t(e)}] -C[ {\cal G}_{u(e)}] \,,
\label{colorJacobi}
\end{eqnarray}
where $C[{\cal G}]$ gives the color factor arrived at by dressing every node of graph ${\cal G}$ with the $\tilde f^{abc}$ structure constants. The color--kinematics duality is satisfied in a representation where for all edges $e$ of all graphs $\cal G$,
\begin{equation}
\label{numIdentity}
0= N \left[ \hat{s} \big ( {\cal G}, e \big )\right]-N \left[ \hat{t} \big ( {\cal G}, e \big )\right] -N \left[ \hat{u} \big ( {\cal G}, e \big )\right] \,,
\end{equation}
where $N[{\cal G}]$ gives the kinematic numerator factor associated with the graph ${\cal G}$.
Additionally, we may use this formalism to write down some relations that originate from the antisymmetry of the cubic vertices
\begin{eqnarray}
C[{\cal G}_{\{(d,a),(b,c)\}}]&=&-C[{\cal G}_{\{(a,d),(b,c)\}}]\,,\nonumber \\
N[{\cal G}_{\{(d,a),(b,c)\}}]&=&-N[{\cal G}_{\{(a,d),(b,c)\}}]\,.
\label{antisym}
\end{eqnarray}
In the following section, we will see that the set of constraints on the numerator factors \eref{numIdentity} and \eref{antisym} can be understood as a set of functional equations that must hold for the duality to be satisfied.
\subsection{Three-loop example}
\label{ThreeLoopExampleSection}
\begin{figure}[t]
\centerline{\epsfxsize 4.5 truein \epsfbox{3LoopDiagrams2.eps}} \caption{Three-loop four-point cubic graphs considered in the main text. The external momenta is outgoing and the shaded (pink) edges mark the application of kinematic Jacobi relations used in \eref{BCJjacobi}. Note that only graphs (a)--(l) contribute to the {${\cal N}=4$~sYM}~amplitude where the duality between color and kinematics is made manifest.}
\label{threeLoop14}
\end{figure}
In this section, we reexamine the four-point three-loop {${\cal N}=4$~sYM}\ amplitude using the duality between color and kinematics~\cite{BCJloops}. This amplitude was originally given in \cite{superfinite,CompactThree} in terms of nine cubic diagrams. For this exercise, we start by considering a larger set of 25 graphs, which are related to any of the original nine diagrams by a single application of a kinematic Jacobi relation. However, eleven of these diagrams contain triangle subgraphs, which the no-triangle property of {${\cal N}=4$~sYM}~\cite{MHVoneloop} suggests will not contribute. After removing those with one-loop triangle subgraphs we have the 14 graphs depicted in \fig{threeLoop14}. We will see that this set of diagrams is sufficiently large to admit a manifest representation of the duality.
Now we will introduce the kinematic Jacobi relations that the numerators of each diagram must satisfy. Each numerator depends on three independent external momenta and (at most) three independent loop momenta. While in general a graph's numerator factor will depend on the states of all external particles, the four-point amplitude of {${\cal N}=4$~sYM}\ is special in that all of its state dependence can be captured in the overall factor ${\cal K}$ described in \sect{RecursiveCutsSection}. Since this factor is common to all terms and the kinematic Jacobi relations are homogeneous, this factor plays no role in the equations. With this in mind, we suppress all state dependence but make the dependence on momenta explicit:
\begin{equation}
N^{(x)}=N^{(x)}(k_1,k_2,k_3,l_5,l_6,l_7)\,,
\end{equation}
where the canonical labeling for each graph $x\in \{\rm a,b,c,d,e,f,g,h,i,j,k,l,m,n\}$ is given in \fig{threeLoop14}.
For every internal edge and for every graph there is one functional equation relating the numerator of that graph to the numerator of two other graphs through the kinematic Jacobi relations. The $14$ graphs of \fig{threeLoop14} each have ten propagators, giving 140 relations in total (before accounting for possible overcounts). One can write down many more equations using the graph automorphisms of each numerator -- ensuring that numerators are invariant under equivalent ways of labeling symmetric graphs. However, this large set of equations contains much redundancy; the number of independent functional relations is quite small. We will not elaborate on how to methodically reduce this system; instead, we observe that the following 13 kinematic Jacobi relations, together with all automorphism relations, solve all constraints imposed by the color--kinematics duality:
%
\begin{equation}
\begin{array}{lc}
N^{(\rm a)}= N^{(\rm b)}(k_1,k_2,k_3,l_5,l_6,l_7)\,, & (J_{\rm a })\\
N^{(\rm b)}= N^{(\rm d)}(k_1,k_2,k_3,l_5,l_6,l_7)\,, & (J_{\rm b })\\
N^{(\rm c)}= N^{(\rm a)}(k_1,k_2,k_3,l_5,l_6,l_7)\,, & (J_{\rm c })\\
N^{(\rm d)}= N^{(\rm h)}(k_3,k_1,k_2, l_7, l_6, k_{1,3}-l_5+l_6-l_7)&\\
~~~~~~~~~\null+N^{(\rm h)}(k_3, k_2, k_1,l_7,l_6, k_{2,3}+l_5-l_7)\,,& (J_{\rm d })\\
N^{(\rm f)}= N^{(\rm e)}(k_1,k_2,k_3,l_5,l_6,l_7)\,, & (J_{\rm f })\\
N^{(\rm g)}= N^{(\rm e)}(k_1,k_2,k_3,l_5,l_6,l_7)\,, & (J_{\rm g })\\
N^{(\rm h)}= - N^{(\rm g)}(k_1, k_2, k_3, l_5, l_6, k_{1,2} - l_5 - l_7)&\\
~~~~~~~~~\null-N^{(\rm i)}(k_4, k_3, k_2, l_6-l_5, l_5 - l_6 + l_7-k_{1 ,2}, l_6)\,, & (J_{\rm h })\\
N^{(\rm i)}\,= N^{(\rm e)}(k_1,k_2,k_3,l_5,l_7,l_6)&\\
~~~~~~~~~\null-N^{(\rm e)}(k_3, k_2, k_1, -k_4 - l_5 - l_6, - l_6 - l_7, l_6)\,, & (J_{\rm i })\\
N^{(\rm j)}\,= N^{(\rm e)}(k_1,k_2,k_3,l_5,l_6,l_7)-N^{(\rm e)}(k_2,k_1,k_3,l_5,l_6,l_7)\,,& (J_{\rm j })\\
N^{(\rm k)}= N^{(\rm f)}(k_1,k_2,k_3,l_5,l_6,l_7)-N^{(\rm f)}(k_2,k_1,k_3,l_5,l_6,l_7)\,, & (J_{\rm k })\\
N^{(\rm l)}\,= N^{(\rm g)}(k_1,k_2,k_3,l_5,l_6,l_7)-N^{(\rm g)}(k_2,k_1,k_3,l_5,l_6,l_7)\,, & (J_{\rm l })\\
N^{(\rm m)}= 0\,, & (J_{\rm m })\\
N^{(\rm n)}\,=0\,, & (J_{\rm n})
\label{BCJjacobi}
\end{array}
\end{equation}
where $k_4=-(k_1+k_2+k_3)$, $k_{i,j}=k_i+k_j$, and for notational simplicity we have suppressed the arguments of the numerators on the left-hand-sides; these all have the canonical argument $(k_1,k_2,k_3,l_5,l_6,l_7)$. The 13 relations are labeled by $(J_x)$ which correspond to Jacobi relations specified by the marked edges in \fig{threeLoop14}. In the above relations we have immediately set to zero any diagrams that do not feature in \fig{threeLoop14}. This explains why some relations in \eref{BCJjacobi} involve fewer than three terms.
Analyzing the result of \eref{BCJjacobi}, we immediately see that the color--kinematics~duality has constrained $N^{(\rm m)}$ and $N^{(\rm n)}$ to vanish, given that we insist on the absence of graphs with one-loop triangles.
We see that $N^{(\rm f)},N^{(\rm g)},N^{(\rm i)},N^{(\rm j)}$ are directly expressed as linear combinations of the function $N^{(\rm e)}$. Furthermore, $N^{(\rm k)},N^{(\rm l)},N^{(\rm h)}$ are expressed as linear combinations of the functions $N^{(\rm f)},N^{(\rm g)},N^{(\rm i)},N^{(\rm j)}$, and thus they can also be solved in terms of $N^{(\rm e)}$. Finally we note that $N^{(\rm a)}=N^{(\rm b)}=N^{(\rm c)}=N^{(\rm d)}$ and that $N^{(\rm d)}$ can be expressed as a linear combination of $N^{(\rm h)}$. So each of these numerators can also be solved in terms of $N^{(\rm e)}$. Therefore, every numerator $N^{(x)}$ is entirely expressible in terms of the planar diagram numerator $N^{(\rm e)}$.
Although the solution to \eref{BCJjacobi} is quite constraining, it does not enforce all of the 140 Jacobi-like relations we began with, nor does it enforce all automorphism relations. As mentioned, we will obtain the full solution by enforcing automorphic invariance of each graph numerator. Using the partial solution of \eref{BCJjacobi}, these relations can be translated to self-constraints on $N^{(\rm e)}$. For example, the \fig{threeLoop14}(e) diagram is invariant under reflections up $\leftrightarrow$ down; thus, $N^{(\rm e)}$ must satisfy
\begin{equation}
N^{(\rm e)}(k_1,k_2,k_3,l_5,l_6,l_7) =N^{(\rm e)}(k_2,k_1,k_4,k_{1,2}-l_5,l_7,l_6)\,.
\end{equation}
Similarly, \fig{threeLoop14}(j) is invariant under $1\leftrightarrow2$. However, just like color factors, numerators are odd under the reordering of a cubic vertex; therefore, we have
\begin{equation}
N^{(\rm j)}(k_1,k_2,k_3,l_5,l_6,l_7) =-N^{(\rm j)}(k_2,k_1,k_3,l_5,l_6,l_7) \,.
\end{equation}
In fact, the reader may check that this equation is already enforced by \eref{BCJjacobi}; thus, this particular automorphism relation does not give rise to any new constraints on $N^{(\rm e)}$.
Beyond these two examples, every diagram but \fig{threeLoop14}(g) is constrained by automorphism relations, which through \eref{BCJjacobi} translates to possible further constraints on $N^{(\rm e)}$.
One way to efficiently solve the remaining kinematic Jacobi and automorphism relations is to write down an ansatz for $N^{(\rm e)}$. As mentioned, we assume that the state-dependent prefactor is ${\cal K}\equiv s_{12} s_{23} A_4^{\rm tree}(1,2,3,4)$. We write the remaining factors as a local degree-two polynomial in 17 momentum invariants:
\begin{equation}
N^{(\rm e)}_{\rm ansatz}={\cal K} \, P(s_{12},s_{23},k_i\cdot l_j,l_i\cdot l_j)\,,
\end{equation}
where $k_i=k_1,k_2,k_3$ and $l_i=l_5,l_6,l_7$, and the degree of the polynomial is given by simple dimensional counting. Although we could use various facts about the power counting of {${\cal N}=4$~sYM}\ to simplify the ansatz, we proceed with the most general homogeneous degree-two polynomial of 17 variables, which gives $17 \times(17+1)/2=153$ parameters to be determined.
Using such an ansatz, we find that only the following four automorphism relations are required in order for all automorphic relations to be satisfied:
\begin{eqnarray}
N^{(\rm c)}(k_1,k_2,k_3,l_5,l_6,l_7)&=&N^{(\rm c)}(k_1,k_2,k_4,l_5,l_5-l_6,k_{3,4}-l_7)\nonumber \,, \\
N^{(\rm f)}(k_1,k_2,k_3,l_5,l_6,l_7)&=&N^{(\rm f)}(k_2,k_1,k_4,k_{1,2}-l_5,k_4-l_5-l_7,k_4+l_5-l_6)\nonumber \,, \\
N^{(\rm h)}(k_1,k_2,k_3,l_5,l_6,l_7)&=&N^{(\rm h)}(k_1,k_4,k_3,k_1 -l_5,l_7,l_6)\nonumber\,, \\
N^{(\rm h)}(k_1,k_2,k_3,l_5,l_6,l_7)&=&N^{(\rm h)}(k_2,k_1,k_4,k_{1,2}-l_5-l_7,-l_6,l_7) \,.
\label{automorph}
\end{eqnarray}
Enforcing \eref{BCJjacobi} and \eref{automorph}, the ansatz is constrained from 153 parameters to only three free parameters, which we label $\alpha, \beta, \gamma$, giving
\begin{eqnarray}
N^{(\rm e)}_{\rm sol}=&{\cal K} \, \Big[ \alpha\, s_{12}s_{23} +
\beta \, ( s_{12}^2 + s_{23}^2) + \gamma \, \Big(s_{12}^2+
2 (s_{12}- s_{23})\, k_1 \cdot l_5 \nonumber \\& \hskip 3.5cm
+2 (2 s_{12} +s_{23})\, k_2 \cdot l_5 + 6 s_{12} \, k_3 \cdot l_5 \Big) \Big]\,.
\label{BCJsolution}
\end{eqnarray}
After checking that this also solves the original 140 Jacobi relations and all automorphisms, we conclude that \eref{BCJsolution} is a threefold family of potential solutions that makes the duality between color and kinematics manifest.
We still need to ensure that these solutions actually describe the ${\cal N}=4$ theory. One such source of theory-specific data can be a generalized unitarity cut. We consider the maximal cut of graph (e), using the known expression in the literature~\cite{BRY},
\begin{equation}
N^{(\rm e)}\Bigl|_{\rm cut}=2 \, {\cal K} \, s_{12}\,k_4 \cdot l_5 \,.
\label{MCut3L}
\end{equation}
Then, we compare to the maximal cut of the solution \eref{BCJsolution},
\begin{equation}
N^{(\rm e)}_{\rm sol}\Bigl|_{\rm cut}={\cal K} \, \Big[ \alpha\, s_{12}s_{23} +
\beta \, ( s_{12}^2 + s_{23}^2) + \gamma \, (
s_{12}s_{23} - 6 s_{12} \, k_4 \cdot l_5 ) \Big]\,,
\end{equation}
where we have used the relations $k_1\cdot l_5=0$, $ 2 k_2\cdot l_5=s_{12}$, and $ 2 k_3\cdot l_5=-s_{12}-2 k_4\cdot l_5$, valid on the maximal cut.
Comparing \eref{BCJsolution} and \eref{MCut3L}, we find that $\alpha=1/3$,
$\beta=0$, $\gamma=-1/3$, giving the fully solved numerator
\begin{eqnarray}
N^{(\rm e)}&=&\frac{1}{3}{\cal K} \, \Bigl( s_{12}(s_{23}-s_{12})
-2 (s_{12}- s_{23})\, k_1 \cdot l_5 \nonumber \\&&
~~~~~~~ -2 (2 s_{12} +s_{23})\, k_2 \cdot l_5 - 6 s_{12} \, k_3 \cdot l_5 \Bigr)\,.
\label{FullBCJsolution}
\end{eqnarray}
All 12 nonvanishing numerators can now be obtained from \eref{FullBCJsolution}, using \eref{BCJjacobi}. They are explicitly given in~\cite{BCJloops}. Together, they represent a duality-satisfying expression of the three-point four-loop {${\cal N}=4$~sYM}~amplitude. This was verified on a spanning set of cuts in~\cite{BCJloops}.
\subsection{Tree level, Lagrangians, and gravity}
\label{TLGSection}
While this review focuses on multiloop methods for gauge theory amplitudes, we should mention a few other interesting results intimately related to the kinematics--color duality.
Consider the formula \eref{LoopBCJ} in the tree-level limit $L=0$. This gives the cubic representation of tree amplitudes introduced in~\cite{BCJ}:
\begin{equation}
{\cal A}_n^{\rm tree}= g^{n-2}\sum_{i\in\gamma_3} \, \frac{n_i c_i}{p_{i_1}^2p_{i_2}^2p_{i_3}^2\cdots p_{i_{m}}^2}\,,
\end{equation}
where the sum runs over all cubic tree-level graphs $\gamma_3$ (including all distinct permutations of external legs), which all have unit symmetry factors. For notational uniformity with~\cite{BCJ} we use $n_i=N_i^{\rm tree}$ and $c_i=C_i^{\rm tree}$.
As shown in~\cite{BCJ}, assuming the duality between tree-level color and kinematics leads to novel relations between color ordered tree amplitudes. These decompose any color-ordered partial amplitude into a sum of $(n-3)!$ basis amplitudes, schematically,
\begin{equation}
A_n^{{\rm tree}}( \sigma(1),\sigma(2),\ldots, \sigma(n))= \sum_{\rho \in S_{n-3}} K^{(\sigma)}_{\rho} A_n^{{\rm tree}}(1,2,3,\rho(4),\ldots, \rho(n))\,,
\label{amplituderelations}
\end{equation}
where $\sigma$ and $\rho$ are permutations, and $K^{(\sigma)}_{\rho}$ are some universal (state-independent) non-local kinematic coefficients. We will not give these factors here; they can be deduced using the results in~\cite{BCJ}.
Similar relations between partial amplitudes were subsequently found in string theory~\cite{Monodromy} using monodromy of integration on the worldsheet. In the field-theory limit, these string relations provided a first proof of \eref{amplituderelations}. Using BCFW recursion, the same relations were later derived from field theory~\cite{amplituderelationProof}. In \cite{noncommutative}, analogous relations was shown to exist in non-commutative Yang-Mills theory.
As for the duality itself, phrased in terms of color and kinematic numerator factors and corresponding Jacobi relations~\eref{DefiningJacobi}, less is known. An investigation in this direction is \cite{TyeDoubleCopy} where the framework of heterotic string theory is used to shed light on the duality. In addition, there are known explicit solutions for tree-level duality-satisfying numerators $n_i$ to all multiplicity~\cite{KiermaierTalk, BBDSVsolution}. While these solutions demonstrate the existence of numerators satisfying the duality, they take a particularly non-local and asymmetric form (under relabeling of external states). The existence of local or symmetric forms of duality-satisfying tree-level numerators to all mulitplicity remains an open question.
In~\cite{LagrangianSquare}, it was shown that the duality between color and kinematics could be made manifest at the Lagrangian level of Yang-Mills theory, by introducing higher degree interactions at the five-point and six-point levels, schematically:
\begin{equation}
{\cal L}_{\rm YM}=\frac{1}{4g^2}F^{\mu \nu}F_{\mu \nu}+ {\cal L}_5+ {\cal L}_6 + \ldots \, ,
\end{equation}
and at the same time introducing auxiliary fields so to make all interactions cubic. Remarkably the corrections ${\cal L}_5+ {\cal L}_6 + \ldots$ vanish by the usual gauge-group Jacobi identity, and do not alter the theory in any way other than modifying the relative values of individual Feynman diagrams.
Finally, we remark that one can easily construct gravity amplitudes using the duality; gravity numerators are double copies of gauge theory kinematic factors~\cite{BCJ,BCJloops}. Given an $n$-point $L$-loop gauge theory amplitude with duality satisfying numerators $N_i$, a gravity amplitude takes the form
\begin{equation}
{\cal M}_n^{(L)}= i^{L+1} \left(\frac{\kappa}{2}\right)^{n-2 +2L}\sum_{i\in\Gamma_3} \, \int \frac{d^{LD}\ell}{(2\pi)^{LD}} \frac{1}{S_i} \, \frac{N_i \widetilde N_i}{p_{i_1}^2p_{i_2}^2p_{i_3}^2\cdots p_{i_{m}}^2}\,,
\label{GravBCJ}
\end{equation}
where $\tilde{N}_i$ are a set of kinematic numerators for the amplitude of a possibly different gauge theory (which need not explicitly satisfy the duality), and $\kappa$ is the gravitational coupling constant. This formula is the generalization of the analogous tree-level ($L=0$) formula~\cite{BCJ}. Assuming the color--kinematics duality, the double copy property was proven~\cite{LagrangianSquare} recursively in $n$ for the explicit case of maximal supersymmetry
$[{{\cal N}=4}] \otimes [{{\cal N}=4}] \rightarrow [{{\cal N}=8}]$ and similarly for pure Yang-Mills theory going to pure gravity plus axion and dilaton. At the classical level, \eref{GravBCJ} encodes (and arguably explains) the field-theory forms of the string-theoretic Kawai-Lewellen-Tye (KLT) tree-level relations~\cite{KLT} between gravity scattering amplitudes and gauge theory amplitudes. Intriguingly, the double copy form of gravity \eref{GravBCJ} appears to generalize seamlessly to the quantum level, to any loop order~\cite{BCJloops}.
We close this section by referring the reader to the literature for further application of the duality, the new tree-level amplitude relations, and the gravity double copy property~\cite{BCJOther,Neq44np,BBDSVsolution,Mafra}; see also reviews~\cite{BCJreview}.
\section{Conclusions}
In this review we have covered some recent methods for constructing compact integral representations of multiloop scattering amplitudes, applicable to planar and non-planar gauge theory in general, and to {${\cal N}=4$~sYM}\ theory in particular, valid in four dimensions and in $D>4$.
The maximal cut method is a convenient scheme in the generalized unitary framework that allows the construction of amplitudes in any massless theory starting from the cuts with the most on-shell propagators. The method has a natural hierarchy which determines the order in which unitary cuts should be evaluated in order to capture missing contributions of an initial amplitude ansatz. Maximal cuts are followed by next-to-maximal cuts and thereafter next-to-next-to-maximal cuts and so on, until the full amplitude has been determined. These maximal and near-maximal cuts can be evaluated in four dimensions and in $D>4$, for planar as well as non-planar cuts. A close relative to the maximal cut method is the leading singularity method~\cite{CachazoSkinner,CachazoLeadingSingularityAndCalcs,Spradlin3loop}, which focuses on the {${\cal N}=4$~sYM}\ theory in four dimensions.
Complementary to the maximal cut method, we have reviewed two particularly simple $D$-dimensional multiloop unitarity cuts in the {${\cal N}=4$~sYM}\ theory: the ordinary two-particle cut of the four-point amplitude, and the box cut. These two types of cuts have a recursive structure that can be implemented using simple pictorial calculations.
Although the {${\cal N}=4$~sYM}\ theory naturally lives in four dimensions, the $D$-dimensional cuts and amplitudes are important for several reasons.
Since {${\cal N}=4$~sYM}\ amplitudes exhibit infrared divergences at loop level they need to be regulated, either using dimensional regularization, or using mass regulators~\cite{MassiveRegulatorProgress}. The former method naturally requires $D$-dimensional amplitudes, and the latter can in many cases be rephrased in terms of $D$-dimensional amplitudes, where the higher-dimensional momenta effectively act as masses from a four-dimensional perspective. Beyond this, there is, of course, the inherent theoretical interest in studying higher-dimensional theories.
Using the duality between color and kinematic factors we considered how to compute multiloop integrands in a very efficient manner. We discussed the three-loop four-point amplitude of {${\cal N}=4$~sYM}\ in some detail, showing how the duality can be made manifest using 12 nonvanishing cubic three-loop graphs, all determined from only one master graph. The duality, which is conjectured to hold for any multiparticle and multiloop amplitude in {${\cal N}=4$~sYM}, as well as less-supersymmetric theories (including pure Yang-Mills), hints of some deep field-theoretic origin, which once understood will explain the suggestive new ways to organize and calculate scattering amplitudes. Similarly, the connection to gravity amplitudes as double copies of gauge theories, similar to the string-theoretic KLT relations, hints of an underlying unifying framework that includes a particularly novel notion of graviton compositeness. From this perspective a connection to string theory is not surprising.
\ack
We are grateful to Zvi Bern, Camille Boucher-Veronneau, Johannes Br\"odel, Tristan Dennen, Lance Dixon, Daniel Freedman, Yu-tin Huang, Harald Ita, Renata Kallosh, David Kosower, and Radu Roiban for collaboration on these very topics, valuable comments on this manuscript, and/or many stimulating discussions on related subjects. JJMC gratefully acknowledges the Stanford Institute for Theoretical Physics for financial support. HJ's research is supported by the European Research Council under Advanced Investigator Grant ERC-AdG-228301.
The figures were generated using Jaxodraw~\cite{Jaxo1and2}, based on Axodraw~\cite{Axo}.
\section*{References}
|
\section{Introduction}
The supersolid (SS) phase that
simultaneously exhibits a superfluid (SF) phase
and a density-wave order
has been studied, since it was first
proposed theoretically \cite{Andreev,Chester,Leggett}.
Although its existence is still controversial,
a non-classical moment of inertia of solid ${}^4$He
in Vycor glass that suggests its existence
was recently reported \cite{Kim}.
SS phases may also exist in optical lattices.
Cold atoms in optical lattices are new experimental
systems in which theoretically modeled Hamiltonians
can be more distinctly simulated experimentally with no disorders.
For instance, an optical lattice system, as described
by a Bose--Hubbard model \cite{Fisher1,Jaksch}, distinctly showed
the phase transition between the SF phase and
the Mott insulator (MI) phase \cite{Greiner}.
Moreover, recent observations of the
long-range dipole--dipole interaction in
${}^{52}$Cr atoms \cite{Cr} may lead to the realization of the SS phase.
The simplest lattice model with a long-range interaction
is the extended Bose--Hubbard model with a nearest-neighbor interaction:
\begin{eqnarray}
H&\ = \ &H_{kin}+H_{int}^U+H_{int}^{V},\nonumber\\
H_{kin}&\ = \ &-t\sum_{\langle i,j \rangle}
(a_{i}^\dagger a_{j}^{}
+ a_{j}^\dagger a_{i}), \nonumber\\
H_{int}^U&\ = \ &\frac{U}{2}\sum_i n_i(n_i-1),\nonumber\\
H_{int}^V&\ = \ &V\sum_{\langle i,j \rangle} n_in_j. \label{Hamiltonian}
\end{eqnarray}
Here $t$, $U$, and $V$ denote the transfer integral
between nearest-neighbor sites,
the repulsive intra-site interaction,
and the repulsive inter-site interaction between nearest-neighbor sites,
respectively. Furthermore,
$a_i$ $(a_i^\dagger)$ is the annihilation (creation) operator
at the site $i$.
In this study, we assume that the lattice is bipartite and consists of
sublattices $A$ and $B$ and that the number of
nearest-neighbor sites is $z$.
The extended Bose--Hubbard model that prohibits (allows)
multiple boson occupations
at one site is called the hard-core (soft-core)-extended Hubbard model.
Several analytical and numerical methods exist
for studying the ground state of
the extended Bose--Hubbard model.
The strong-coupling perturbation theory \cite{Iskin}
has been applied to obtain the phase boundary
between the SF and non--SF phases.
At least in the absence of the nearest-neighbor interaction
\cite{Freericks1,Freericks2},
the phase boundary determined by this theory agrees perfectly with that
determined by quantum Monte Carlo (QMC) simulations
\cite{Scalettar1,Niyaz1,Batrouni1,Otterlo1,Capogrosso1,Capogrosso2}.
However, this theory can neither distinguish between the SF and SS phases
nor describe discontinuous transitions even if they exist.
In contrast, the mean field (MF) approximation
and equivalent Gutzwiller approximations
\cite{Rokhsar,Krauth}
can distinguish between the SF and SS phases and
describe discontinuous transitions \cite{Kimura1,Kimura2,Krutitsky},
although their validity is limited for high dimensions.
For the hard-core model at half filling ($N\ = \ 1/2$), both the MF approximation \cite{Matsuda,Liu,Scalettar2} for the mapped XXZ model
and two-dimensional (2D) QMC studies \cite{Scalettar2,Batrouni2}
identified the first-order SF--solid transition and no SS phase.
The MF approximation \cite{Matsuda}
also showed that the transition from
the SF phase to the checker-board solid phase is of the first order,
and at the transition point $\mu\ = \ \mu^*$,
the SS phase; solid phase; and PS phase in which
the SS phase and the solid phase coexistx are all degenerate.
On the other hand, later QMC studies \cite{Batrouni3,Hebert1}
showed that the SS phase is unstable against PS away from half filling.
For the soft-core model, both
Gutzwiller approximations \cite{Otterlo2,Scarola1}
and QMC simulations \cite{Scalettar2,Batrouni2}
have shown that checker-board SS phases can be stable against the PS.
However, the details of the two types of study differ.
The Gutzwiller studies showed stable SS phases
for both broad filling and half filling, but
the PS issue is not a concern in these cases.
In contrast, the
2D QMC studies did not show the SS phase at half filling but identified
SS phases above half filling; however,
the system energy as a function of boson density
was found to be concave below half filling, which indicates
a possible PS instability \cite{Scalettar2}.
A later 2D QMC study \cite{Sengupta1}
showed more distinctly that the SS phase
is unstable against PS below half filling
by exhibiting the negative compressibility $\kappa\ = \ dN/d\mu<0$.
This situation is somewhat similar in one dimension (1D).
Both QMC \cite{Hebert2,Batrouni4} and density-matrix renormalization group
\cite{Kuhner1,Kuhner2,Mishra1} studies have been performed in 1D.
The SS phase was found only above half filling
as in the 2D case but PS did not occur \cite{Batrouni4,Mishra1}.
In a three-dimensional (3D) cubic lattice,
Yamamoto and coworkers \cite{Yamamoto1}
found some region of the parameter sets where the SS phase
is stable below half filling in QMC calculations,
which is consistent with a MF phase diagram on the $N$--$t/U$ plane
for $V\ = \ U/6$.
Hence, it is interesting to study the 3D system further.
On the other hand, other previous calculations
assumed a grand canonical ensemble and
did not directly include the possibility that
the system can be separated into two phases.
As a result, the phases comprising
the separate phases were not directly shown.
However, if we start with a canonical
ensemble and explicitly include the possibility
that the system can be separated into phases,
we can precisely describe the phase diagram with a fixed
boson number for the entire system.
Such calculations might also be easily considered using the
grand canonical ensemble if we explicitly assume that
the total free energy of the system
is $E_{\rm tot}\ = \ \gamma E_{\rm SS}+(1-\gamma) E_{\rm sol}$.
Here $\gamma$ is the ratio of the areas of the SS phases,
and we neglect the surface energy between the SS and solid phases.
However, if we minimize the free energy at the chemical potential $\mu$,
we can obtain only $\gamma\ = \ 0$ or 1, but not
an intermediate value $0< \gamma <1$
except for $\mu\ = \ \mu^*$ at which the SS and solid phases
are degenerate [$E_{\rm SS}(\mu^*)\ = \ E_{\rm sol}(\mu^*)$].
That is, the solid and SS phases cannot be in the ground
state at the same time,
and a separate phase including two states cannot be obtained
for $\mu\ \neq\ \mu^*$.
To obtain a separate phase,
we must tune the chemical potential from $\mu$ to $\mu^*$;
furthermore, we must try to find a $\gamma$
that satisfies the boson number condition
$\gamma N_{\rm SS}+(1-\gamma)N_{\rm sol}\ = \ N$.
These calculations become more complicated when the number of possible phases
is large (SF, SS, MI, and several solids).
This issue, in which PS can be obtained only
on the phase boundary point(s) or curve(s) in the phase diagram
but not for a finite region,
generally occurs when we employ intensive
variable(s) such as chemical potential
in thermodynamics.
The most popular case is the gas--liquid transition,
where PS occurs only on the
phase boundary curve in the pressure--volume plane for a given temperature.
However, if we employ particle number,
volume, and internal energy as thermodynamic variables,
which are extensive quantities,
we can obtain a finite region of PSs on the
particle density--energy density plane.
In this study, we study the ground-state properties of the
extended Bose--Hubbard model on a bipartite lattice
based on the Gutzwiller approximation
in a canonical ensemble; hence, we do not have to
tune the chemical potential of PS.
We study a wide parameter regime and
present three main figures for the phase diagram on the
particle density--nearest-neighbor interaction plane.
Our phase diagrams for small, intermediate, and large transfer integrals
differ significantly.
The other main topic is the solid order parameter
$\delta n\ = \ |N_A-N_B|$ [$N_{A(B)}$: the boson density on the
$A(B)$ sublattice], which is also of interest
and also depends strongly on the transfer integral.
We employed the linear programming method \cite{numerical}
to minimize the total energy for a particular boson number.
Following this method,
we can simultaneously determine the PS region,
the phases that comprise the separate phase,
and the ratio of each phase to the entire system.
This paper is organized as follows, Section II
introduces our calculation method, and
Secs. III and IV present the hard-core and soft-core
Bose--Hubbard models, respectively.
Section V describes the effect of an
improved calculation on the energy
of the solid and MI phases.
Section VI discusses the conclusions based on our results.
Finally, the appendices explain the details of several perturbative
calculations, the results of which are compared with the numerical
results from the Gutzwiller approximation.
\section{CALCULATION METHOD}
We employ the following Gutzwiller approximation
for the SS and SF phases: $\Psi\equiv \prod_i\Phi_i$. Here,
$\Phi_i$ is a variational wave function at the site $i$.
We assume a bipartite lattice with sublattices $A$ and $B$
and a checker-board symmetry for the SS phase.
The variational wave function
is assumed to be $\Phi_i\ = \ \Phi_{A(B)}$
if the site $i$ belongs to the $A(B)$ sublattice.
$\Phi_{A(B)}$ is written as a linear combination of states $|n\rangle$
with $n$ bosons as $\Phi_{A(B)}\ = \ \sum_{n}c_{A(B)n}|n\rangle$.
The variational parameters $c_{A(B)n}$ are
determined so as to minimize the energy expectation value.
We use Powell's method \cite{numerical}
to numerically optimize the variational parameters.
If the result of the optimization shows
$c_{An}\ \ne\ c_{Bn}$ and
$c_{A(B)n}\ \ne\ \delta_{i(j)n}$,
the phase is SS; however, if
$c_{An}\ = \ c_{Bn}\ \ne\ \delta_{in}$, the phase is SF
($i$ and $j$ are non-negative and positive integers,
respectively).
The first inequality for the SS phase
corresponds to the existence of a finite density difference
between sublattices $A$ and $B$ $\delta n\ = \ |N_{A}-N_{B}|$
(i.e., the checker-board density order).
Here $N_{A(B)}\ = \ \sum_n n|c_{A(B)n}|^2$ is the expectation value
of the boson density at the $A(B)$ sublattice.
The second inequality implies that
the phase is not solid. For the soft-core model,
in principle, the sum of $n$ must be taken from $n\ = \ 0$
to $\infty$; however,
we take the sum from $n\ = \ 0$ to $n\ = \ 9$,
which is sufficient for our calculation when $N\le 1$.
For the hard-core model, by definition,
we consider only the two states $|n\ = \ 0\rangle$
and $|n\ = \ 1\rangle$.
The energy expectation value per site is calculated
by the Gutzwiller variational wave function
as
\begin{eqnarray}
E&\ = \ &\langle H\rangle\nonumber\\
&\ = \ &\langle H_{\rm kin}\rangle+\langle H_{\rm int}^U\rangle
+\langle H_{\rm int}^V\rangle\nonumber\\
&\ = \ &-zt \prod_{i\ = \ A,B}\sum_n \sqrt{n+1} c_{in} c_{i(n+1)}\nonumber\\
&&+\frac{U}{4}
\sum_{i\ = \ A,B}\sum_n n(n-1)|c_{in}|^2
+\frac{zV}{2}N_A N_B.
\end{eqnarray}
Furthermore, we assume that the energy per site of the MI phase
with $N$ bosons per site is
\begin{eqnarray}
E_{\rm MI}\ = \ \frac{U}{2}N(N-1)+\frac{zV}{2}N^2\ = \ \frac{zV}{2}
\ \ \ {\rm for}\ N\ = \ 1\label{EMI}
\end{eqnarray}
and that of the solid phase with
$N_A$ ($N_B$) bosons per site on the $A$($B$) sublattice is
\begin{eqnarray}
E_{\rm S}\ = \ \frac{U}{4}\sum_{i\ = \ A,B}N_i(N_i-1)+\frac{zV}{2}N_AN_B, \label{ES}
\end{eqnarray}
where $N_{A(B)}$ is a non-negative integer.
In particular, for
the solid phase when $N_A\ = \ 1$ and $N_B\ = \ 0$ (the S$_1$ phase),
$E_{{\rm S}_1}\ = \ 0$;
for the solid phase when $N_A\ = \ 2$ and $N_B\ = \ 0$ (the S$_2$ phase),
$E_{{\rm S}_2}\ = \ U/2$.
These energies correspond to those obtained from the Gutzwiller
variational wave function with
$c_{An}\ = \ c_{Bn}\ = \ \delta_{nN}$ for the MI phase
and $c_{An}\ = \ \delta_{nN_A}$ and $c_{Bn}\ = \ \delta_{nN_B}$ ($N_A\ \neq\ N_B$)
for the solid phase.
Because we
neglect the surface energy between different phases
by assuming the thermodynamic limit,
the system's total energy per site is
\begin{eqnarray}
E_{\rm tot}\ = \ \sum_{i} \gamma_i E_i
\end{eqnarray}
for the boson number condition
\begin{eqnarray}
N_{\rm tot}\ = \ \sum_{i} \gamma_i N_i.
\end{eqnarray}
Here $\{E_i\}$ and $\{N_i\}$ (where $i\ = \ {\rm SF}$, SS, MI, and solids)
are the energies and boson number densities,
respectively, of all possible phases and $\{\gamma_i\}$
represent the of area
or volume ratio of the phase $i$ in the entire system.
Following the linear programming method \cite{numerical}, we can
minimize $E_{\rm tot}$ as a function of $\{\gamma_i\}$ \cite{note0}.
Only one $\gamma_i$ value (corresponding to the
uniform phase) or two $\gamma_i$ values (corresponding to PS)
are automatically chosen to be nonzero as a result of minimizing $E_{\rm tot}$
because only one additional condition, the boson number condition, exists.
For simplicity, however, we neglect the possibility
of the separated phase that splits into
the SF and SS phases, which is highly unlikely. Hereafter, we represent
the separated phase consisting of phases X and Y as the PS(X + Y) phase.
\section{HARD-CORE MODEL}
In the hard-core-extended Hubbard model,
we prohibit multiple occupations at a site
and define
the Gutzwiller variational wave function as
\begin{eqnarray}
\Phi_{A(B)}\ = \ c_{A(B)0}|0\rangle+c_{A(B)1}|1\rangle. \label{hardcore}
\end{eqnarray}
At half filling,
$|c_{A(B)0}|^2+|c_{A(B)1}|^2\ = \ 1$
according to the normalization condition of the wave function,
and $|c_{A1}|^2+|c_{B1}|^2\ = \ 1$ according to the
boson number condition. If we set $x\ = \ c_{A0}$, then
$c_{A1}\ = \ c_{B0}\ = \ \sqrt{1-x^2}$ and $c_{B1}\ = \ x$.
Hence, the system energy per site is
\begin{eqnarray}
E&\ = \ &-zt c_{A0}c_{A1}c_{B0}c_{B1}
+\frac{zV}{2}|c_{A1}|^2|c_{B1}|^2 \nonumber\\
&\ = \ &\Big(zt-\frac{zV}{2}\Big)\Big[\Big(x^2-\frac{1}{2}\Big)^2-\frac{1}{4}\Big].
\end{eqnarray}
If $V>2t$, then $x^2\ = \ 0$ or $1$ and the phase is S$_1$;
however, if $V<2t$, then $x^2\ = \ \frac{1}{2}$ and the
phase is SF. Hence, the SF--S$_1$ phase transition occurs at $V\ = \ 2t$
and is discontinuous.
This result agrees with those of previous studies \cite{Bruder1,Nozieres1}.
Away from half filling,
we can obtain the SF--SS phase boundary by a perturbative calculation
setting $c_{A(B)n}\ = \ c_n+\delta c_{A(B)n}$,
where $c_n$ is the optimized value when
we assume the SF phase ($c_{An}\ = \ c_{Bn}$) and
$\delta c_{A(B)n}$ represents infinitesimal quantities that
describe the possible SS instability (see Appendix A.1 for details).
The result is
\begin{eqnarray}
V_{\rm C}\ = \ t\ \frac{N^2+(1-N)^2}{N(1-N)}. \label{perturbation1}
\end{eqnarray}
That is, the energy of the SS phase is lower (higher)
than that of the SF phase for $V>V_{\rm C}$ ($V<V_{\rm C}$).
$V_{\rm C}^{\rm SS}$ is invariant under $N\leftrightarrow 1-N$
because of the particle--hole symmetry of the hard-core model.
We also obtain the phase boundary between
the SF phase and the PS(SF + S$_1$) phase, which
consists of the SF and the S$_1$ phases
by another perturbative calculation, assuming that the
ratio of the S$_1$ phase to the entire system is infinitesimal
(see Appendix A.2 for details).
The result agrees with Eq. \ref{perturbation1}.
To determine the ground state for $V>V_{\rm C}$,
we numerically compared the energy of the SS and
PS(SF + S$_1$) phases by using the Gutzwiller variational
wave function as explained in the Introduction.
We found that the PS(SF + S$_1$) and SS phases are
degenerate for $V>V_{\rm C}$ within the numerical error.
The phase diagram is shown in Fig. 1.
The numerical phase boundary agrees with Eq. \ref{perturbation1}
within the numerical error and the phase transitions
are continuous as the perturbative calculation assumed.
These results are consistent with those in Ref. \cite{Matsuda} which showed
that the SF--S$_1$ transition is of the first order
and the SS, S$_1$, and PS(SF + S$_1$) phases are all degenerate
at the transition point $\mu\ = \ \mu^*$.
We also confirmed that both $\gamma_{\rm S_1}$
(the ratio of the S$_1$ phase) in the PS(SF + S$_1$) phase
and $\delta n\ = \ |N_A-N_B|$ in the SS phase
are finite for $V>V_{\rm C}$ [Figs. 2(a) and 2(b)].
As $N$ approaches $1/2$,
$\gamma_{\rm S_1}$ increases rapidly as a function of $(V-V_{\rm C})/V_{\rm C}$,
because at half filling, no SS phase occurs and
the SF--S$_1$ phase transition is discontinuous.
\begin{figure}
\includegraphics[height= 2.5in]{fig1.eps}
\caption{
Phase diagram of the hard-core model.
In the PS(SF + S$_1$)/SS phase, the SS and
PS(SF + S$_1$) phases are degenerate.
The two-dot-dashed line at half filling indicates the solid phase;
the dashed line at unit filling indicates the MI phase.
}
\label{fig1}
\end{figure}
\begin{figure}
\subfigure[]{
\includegraphics[height=2.5in]{fig2a.eps}
\label{fig2a}}
\subfigure[]{
\includegraphics[height=2.5in]{fig2b.eps}
\label{fig2b}}
\caption{
(a) $V$ dependence of
$\gamma_{\rm S_1}$ [the ratio of the S$_1$ phase
in the PS(SF + S$_1$) phase] for $zt\ = \ 1$.
(b) $\delta n\ = \ |N_A-N_B|$
in the SS phase for $zt\ = \ 1$.
$V_{\rm C}$ is the critical value for the
SF--PS(SF + S$_1$) or SF--SS transition at each value of $N$
($zV_{\rm C}\ = \ 4.250$, 2.762, 2.167, 2.002 for
$|N-1/2|\ = \ 0.3$, 0.2, 0.1, 0.01, respectively).
}
\label{fig2}
\end{figure}
\section{SOFT-CORE MODEL}
In this section, we study the soft-core-extended Hubbard model.
We employ the Gutzwiller variational wave function and
optimize its variational parameters numerically.
Hereafter, we call this the full numerical calculation(s);
when we do not refer to the calculation method,
the result was obtained by the full numerical calculations.
We also perform perturbative calculations that
limit the Hilbert space of the Gutzwiller variational
wave function and include partial numerical calculations.
We compare these perturbative calculations with
the full numerical calculations.
In subsection A, we examine
the SS phase between the SF
and solid phases at half filling.
In subsection B, we examine the case away from half filling,
which is the main part of this paper.
Here we show that both the phase diagram on the $zV/U-N$ plane
and $\delta n\ = \ |N_A-N_B|$ change
qualitatively from small $zt/U$ to large $zt/U$.
\subsection{HALF FILLING}
In this subsection, we primarily examine the SS phase at half filling.
We compare the critical value of $V$
for the SS--S$_1$ transition with that of the SF--SS transition
because if the former is larger than the latter, we can
obtain the SS phase between the SF and S$_1$ phases.
By a perturbative calculation with
an infinitesimal SS component added to the S$_1$ phase,
we obtain the critical value of $V$ for the SS--S$_1$ transition
by a power-law expansion of $zt/U$:
\begin{eqnarray}
zV_{\rm C}^{\rm SS-S_1}\ = \ 2zt+2\frac{z^2t^2}{U}+2\frac{z^3t^3}{U^2}+O\Big(\frac{z^4t^4}{U^3}\Big) \label{eq3}
\end{eqnarray}
(see Appendix B.1 for details).
We can also obtain the critical
value of $V$ for the SF--SS transition
\begin{eqnarray}
zV_{\rm C}^{\rm SF-SS}\ = \ 2zt+2\frac{z^2t^2}{U}+O\Big(\frac{z^4t^4}{U^3}\Big),
\label{eq4}
\end{eqnarray}
by another perturbative calculation, in
which an infinitesimal $\delta n\ = \ |N_A-N_B|$ is added to the SF phase
(see Appendix B.2 for details).
Because $V_{\rm C}^{\rm SF-SS}<V_{\rm C}^{\rm SS-S_1}$,
the SS phase is possible for
intermediate $V$ satisfying $V_{\rm C}^{\rm SF-SS}<V<V_{\rm C}^{\rm SS-S_1}$.
We verified these results numerically.
Figure 3 shows the phase boundaries obtained
by the above-mentioned perturbative calculations
(dashed and dot-dashed curves) and
the full numerical calculations (solid curves).
A stable SS phase appears between the SF
and S$_1$ phases. Both the SS--S$_1$ and SF--SS phase boundaries
determined by perturbative calculations
agree well with those determined by the full numerical calculations
(especially at small $zt/U$ as expected).
Figure 4 shows the interaction dependence of $\delta n\ = \ |N_A-N_B|$.
Here, $zV_{\rm C}\ = \ zV_{\rm C}^{\rm SF-SS}$ is the critical value
for the SF--SS transition at each value of $zt/U$.
A finite $\delta n$ less than unity
shows the density order of the SS phase,
whereas $\delta n\ = \ 0(1)$ corresponds to the SF(S$_1$) phase.
We found that $\delta n$ continuously becomes finite at
the SF--SS phase transition. It
changes more rapidly for smaller $zt/U$.
This demonstrates that
the SS phase disappears and the SF-S$_1$ discontinuous transition occurs
in the hard core limit $zt/U\rightarrow 0$.
On the other hand, 2D QMC calculations at half filling \cite{Scalettar2,Batrouni2}
showed that the SF phase directly transitions to the S$_1$ phase
and no SS phase appears.
However, $zt/U$ might not be sufficiently large to allow the SS phase
to be easily found there, and a QMC simulation (as a matter of course,
not only 2D but also 3D) for a large $zt/U$
has a possibility for finding the SS phase.
\begin{figure}
\includegraphics[height= 2.5in]{fig3.eps}
\caption{
Phase diagram of the soft-core model at half filling.
Solid curves represent numerical SF--SS and SS--S$_1$ phase boundaries.
Dashed (dot-dashed) curve represents the SF--SS
(SS--S$_1$) phase boundary obtained
by a perturbative calculation using Eq. \ref{eq4} (Eq. \ref{eq3}).
}
\label{fig3}
\end{figure}
\begin{figure}
\includegraphics[height= 2.5in]{fig4.eps}
\caption{
$V$ dependence of the density difference
between sublattices $A$ and $B$ $\delta n\ = \ |N_A-N_B|$ in the SS phase
of the soft-core model at half filling.
$V_c$ is the critical value for the
SF--SS transition at each value of $zt/U$
($zV_{\rm C}/U\ = \ 0.219$, 0.469, 0.730 for
$zt/U\ = \ 0.1$, 0.2, 0.3, respectively).
}
\label{fig4}
\end{figure}
\subsection{AWAY FROM HALF FILLING}
In this section,
for comparison with the results of the full numerical calculations,
we calculate the results of the following three perturbative
equations for the phase boundary (see Appendices A, B.2, and B.3 for details).
Perturbation 1 is Eq. \ref{perturbation1}, which
assumes a limited Hilbert space with $|N\ = \ 0\rangle$ and $N\ = \ 1\rangle$
and is the same as that used in Sec. III;
perturbation 2 is Eq. \ref{perturbation2}, which yields
the SF--SS phase boundary; and
perturbation 3 is Eq. \ref{perturbation3}, which yields
the SF--PS(SF + S$_1$) or SF--PS(SF + S$_2$) phase boundary.
For perturbations 2 and 3,
we employ a limited Hilbert space with $|N\ = \ i\rangle$, where
$i\ = \ 0$, 1, and 2.
Hence, results obtained by
perturbations 2 and 3 are expected to be better than those obtained
by perturbation 1.
However, perturbations 2 and 3 require
a numerical calculation to
minimize the SF energy in the limited Hilbert space.
We begin by analyzing the phase diagram for a
small transfer integral $zt/U$.
Figure 5(a) shows the phase diagram for $zt/U\ = \ 0.03$,
which resembles that of the hard-core model.
Namely, the upper bound of $zV/U$ for the SF phase
has a minimum value near half filling,
which is similar to that of the hard-core model
with particle-hole symmetry.
PS occurs for large $zV/U$
below half filling, and
the separated phase is the PS(SF + S$_1$) phase, as in the hard-core model.
However, the SS phase appears above half filling.
These properties agree with those obtained in
a 2D QMC study with $zt/U\ = \ 0.08$ \cite{Sengupta1}.
Figure 5(a) plots the result of perturbation 1 (dot-dashed curve).
It agrees almost perfectly with the full numerical calculation
(solid curve) except at large $N$ where the SF--SS phase
boundary disappears because the PS(SS + S$_2$) phase appears there
[see also Fig. 5(b), which expands Fig. 5(a) around $zV/U\ = \ 1$].
Both perturbation 2 for the SF--SS boundary curve and
perturbation 3 for the SF--PS(SF + S$_1$) boundary curve
yield almost the same results as perturbation 1.
Therefore, they also agree almost exactly
with the full numerical calculations except at large $N$.
Figure 5(b) shows that PSs also appear above half filling
but the region is very small compared to that obtained by
2D QMC \cite{Sengupta1}.
The appearance of the PS(SS + S$_2$) phase and the intricate structure of the
phase diagram are non-trivial;
however, the two SS phases that sandwich
the PS(SS + S$_2$) phase have different characteristics:
one having a smaller zV/U resembles the SF phase and
the other having a larger zV/U resembles the S$_2$ phase,
as we will see below
through $\delta n\ = \ |N_{\rm A}-N_{\rm B}|$.
Namely, SF, SS, S$_2$, and the PS comprising
these phases are almost
degenerate for $zV/U\sim 1$ above half filling.
\begin{figure}
\subfigure[]{
\includegraphics[height=2.5in]{fig5a.eps}
\label{fig5a}}
\subfigure[]{
\includegraphics[height=2.5in]{fig5b.eps}
\label{fig5b}}
\caption{
Phase diagram of the soft-core model for $zt/U\ = \ 0.03$.
(a) Entire phase diagram.
Solid curve represents the SF--PS(SF + S$_1$) or SF--SS phase boundary.
Dot-dashed curve for the SF--SS phase boundary
obtained by the perturbative calculation in Eq. \ref{perturbation1}
cannot be distinguished from the solid curve
except at large $N$, where the SF--SS transition curve disappears.
(b) Expansion of (a) around $zV/U\ = \ 1$ and $N\ = \ 1$.
Solid (dashed) curves for the phase boundaries
show continuous (discontinuous) phase transitions.
In both (a) and (b), the two-dot-dashed lines
at half filling and unit filling
indicate the solid phases (S$_1$ and S$_2$, respectively),
and the long-dashed line at unit filling
indicates the MI phase.
}
\label{fig5}
\end{figure}
Figure 6 confirms that the SS phase overcomes the PS(SF + S$_1$) phase
above half filling because the critical value of the nearest-neighbor
interaction $V_{\rm C}$ for the SF--SS transition
is always smaller than that for the virtual SF--PS(SF + S$_1$) transition.
Here, the virtual SF--PS(SF + S$_1$) transition was obtained
by setting the variational parameter of the
Gutzwiller variational wave function to $c_{An}\ = \ c_{Bn}$.
The difference between the curves for the SF--SS transition and
the virtual SF--PS(SF + S$_1$) transition
disappears in the hard-core limit of $zt/U\rightarrow 0$,
where the SS and PS(SF + S$_1$) phases are degenerate,
as discussed in Sec. III.
\begin{figure}
\includegraphics[height=2.5in]{fig6.eps}
\caption{
Critical nearest-neighbor interaction $V_{\rm C}$
for the SF--SS transition (solid curves)
and the virtual SF--PS(SF + S$_1$) transition (dashed curves)
above half filling.
Note that both solid and dashed curves appear
for $N\ = \ 0.6$, although they are very close.
}
\label{fig6}
\end{figure}
Figure 7(a) shows the $V$ dependence of $\delta n$ obtained
by the full numerical calculation.
Figure 7(b) shows an expansion of the region around $1.001\le zV/U\le 1.002$.
Note that $\delta n$ has a discontinuity at $zV_{\rm C}/U\simeq 1.001$,
because the discontinuous SS--PS(SS + S$_2$) transition occurs there.
Interestingly, the SS phase exhibits a small $\delta n$ for $zV/U\leq 1$,
whereas it exhibits a large $\delta n\simeq 2N$ ($N_A\simeq 2N$ and
$N_B\simeq 0$) for $zV/U\ge 1.002$
after a rapid increase in $\delta n$ in a narrow PS(SS + S$_2$) region
($1.001\leq zV/U\leq 1.002$), where the curves of $\delta n$
for different $N$ are almost indistinguishable
[see Fig. 7(b)].
Because $\delta n\simeq 2N$,
$\delta n$ is larger for larger $N$ at $zV/U\ge 1.002$.
In contrast, $\delta n$ is smaller for larger $N$ for $zV/U\le 1$,
demonstrating that the critical value of $zV/U$ for the SF--SS transition
is an increasing function of $N$ above half filling.
The fact that $\delta n\simeq 2N$ for $zV/U\gtrsim 1$ indicates that
the SS phase is similar to the S$_2$ phase, in which
$N_{\rm A}\ = \ 2$ and $N_{\rm B}\ = \ 0$. The fact that
$\delta n\simeq 0$ for $zV/U\lesssim 1$ indicates that
the SS phase resembles the SF phase.
Note that $zV/U\ = \ 1$ is also the phase transition point
for $t\ = \ 0$ at unit filling: the MI phase (i.e., no-density-order phase)
for $zV/U<1$ becomes
the S$_2$ phase (density order phase) for $zV/U>1$. That is,
even for a finite but small $t$
and $N<1$, the characteristics of the ground state
change rapidly near $zV/U\simeq 1$.
\begin{figure}
\subfigure[]{
\includegraphics[height=2.5in]{fig7a.eps}
\label{fig7a}}
\subfigure[]{
\includegraphics[height=2.5in]{fig7b.eps}
\label{fig7b}}
\caption{
$V$ dependence of
$\delta n\ = \ |N_A-N_B|$ in the SS phase
of the soft-core model for $zt/U\ = \ 0.03$.
$\delta n$ is larger at smaller $N$ for $zt/U
\le 1.000$,
discontinuous at $zV_{\rm C}/U\simeq 1.001$,
and smaller at smaller $N$ for $zt/U\ge 1.002$.
(b) Expansion around the PS(SS + S) phase
($1.001\leq zV_{\rm C}/U\leq 1.002$).
In the PS(SS + S) phase, $\delta n$ is a sharply
increasing function of $zV/U$, and the curves of $\delta n$
for different $N$ cannot be distinguished.
Thus, the curved arrows describe the increase in $\delta n$ at
each value of $N$ when we enlarge $zV/U$.
}
\end{figure}
Next, we study the case of an intermediate $zt/U$ value of 0.3.
Figures 8(a) and 8(b) show the phase diagram;
the PS region above half filling
exists, but again is very small.
The phase diagram clearly departs from that
of the hard-core model with the particle-hole symmetry.
Interestingly, for an intermediate $zV/U$, the SS phase
overcomes the PS even below half filling. This agrees with
a recent 3D QMC simulation with $zt/U\ = \ 0.33$ \cite{Yamamoto1}.
Figure 9 shows $\delta n$ ($\gamma_{\rm S_1}$)
below half filling in the SS (PS(SF + S$_1$)) phase.
The SS--PS(SF + S$_1$) transition is discontinuous:
for $N\ = \ 0.3$, $0.35$, and $0.4$, $\delta n$ ($\gamma_{\rm S_1}$)
is finite (zero)
in the SS phase and discontinuously becomes zero (finite)
in the PS(SF + S$_1$) phase.
For $N\ = \ 0.25$, $\delta n$ is zero in the entire figure
because the SS phase does not exist.
Returning to Fig. 8(a),
we see that the results from
perturbative calculations (dot-dashed curves)
agree well with those from the full numerical
calculations (solid curves).
Assuming the S$_1$ phase for the solid, perturbation 3
is in almost perfect agreement
with the SF--PS(SF + S$_1$) phase boundary below half filling.
Perturbation 2 is also in excellent agreement with
the SF--SS phase boundary except for
the region of large $N\sim 1$ where the SF--SS transition
becomes discontinuous, and thus
cannot be described by the perturbative calculation.
Figure 8(b) shows an enlargement of Fig. 8(a) around $zV/U\ = \ 1.1$ and $N\sim 1$,
where a small PS(SF + S$_2$) phase appears.
The curve of perturbation 3 in Fig. 8(b), which is calculated
assuming the S$_2$ phase for the solid,
seems to be far from the results obtained by
the full numerical calculations; however,
the difference is in fact approximately 1$\%$ at most.
\begin{figure}
\subfigure[]{
\includegraphics[height=2.5in]{fig8a.eps}
\label{fig8a}}
\subfigure[]{
\includegraphics[height=2.5in]{fig8b.eps}
\label{fig8b}}
\caption{
Phase diagrams of the soft-core model for $zt/U\ = \ 0.3$.
(a) shows Entire phase diagram.
Solid curve shows the continuous SF--PS(SF + S$_1$) and SF--SS transitions.
Dot-dashed curves
are the results of the perturbative calculations
Eq. \ref{perturbation1} (perturbation 1),
Eq. \ref{perturbation2} (perturbation 2), and Eq. \ref{perturbation3} (perturbation 3) for the SF--PS(SF + S$_1$) transition.
Two-dot-dashed lines at half filling and
unit filling represent the solid phases (S$_1$ and S$_2$, respectively).
(b) Expansion of the region around $zV/U\ = \ 1$ and $N\ = \ 1$.
Solid curve shows the continuous
SF--PS(SF + S$_2$) phase transition; dashed curve
shows the discontinuous SF--SS and PS(SF + S$_2$)--SS phase transitions.
Dot-dashed curve shows Eq. \ref{perturbation3} (perturbation 3)
for the SF--PS(SF + S$_2$) transition.
}
\label{fig8}
\end{figure}
\begin{figure}
\includegraphics[height=2.5in]{fig9.eps}
\caption{
$V$ dependences of
$\delta n\ = \ |N_A-N_B|$ (solid curves) and $\gamma_{\rm S_1}$ (dashed curves)
in the soft-core model for $zt/U\ = \ 0.3$ below half-filling.
}
\label{fig9}
\end{figure}
Figure 10 shows the $V$ dependence of
$\delta n$ above half filling for $zt/U\ = \ 0.3$.
As in Fig. 7 for $zt/U\ = \ 0.03$,
$\delta n\simeq 2N$ is larger for larger $N$ at $zV/U>1.1$, whereas
$\delta n$ is smaller for larger $N$ for $zV/U<1.1$.
Hence, as in the case of small $zt/U$,
the characteristics of the SS change drastically from SF-like
to solid (S$_2$)-like around $zV/U\sim 1$ when we increase $zV/U$.
Note that in contrast to the case of $zt/U\ = \ 0.03$,
$\delta n$ shows no discontinuity except at $N\ = \ 0.9$
because the SF--SS transition is continuous.
As a result, the two $\delta n$ curves for two different $N$ values
intersect smoothly around $zV/U\simeq 1.1$ except at $N\ = \ 0.9$.
For $N\ = \ 0.9$, $\delta n$ is zero for $zt/U<1.105$
(because the phase is SF, at which $\delta n\ = \ 0$) and
discontinuously becomes finite for $zt/U>1.105$,
because the SF--SS transition is discontinuous there.
\begin{figure}
\includegraphics[height=2.5in]{fig10.eps}
\caption{
$V$ dependence of
$\delta n\ = \ |N_A-N_B|$ of the soft-core model for $zt/U\ = \ 0.3$ above half filling.
}
\label{fig10}
\end{figure}
Finally, we examine a large $zt/U$($\ = \ 1$).
Figure 11 shows the phase diagram.
Only the SF and SS phases exist, and there are no PSs.
The SF--SS transition is continuous throughout the figure.
The critical $zV/U$ value between these phases is a decreasing function
of the boson density $N$. This may be explained as follows.
For large $zt/U$, the ratio $t$ to $V$ is the only important factor
determining the phase (SF or SS, not the PS),
because the lattice does not play an important role except
at half or unit filling, and $N$ affects only
the SF--SS phase boundary ($V$
affects the phase more strongly for larger $N$).
As a result, no rapid change in the ground state properties
occurs around $zV/U\ = \ 1$ in contrast to the case of small or intermediate $zt/U$.
The dot-dashed curve of perturbation 1
does not agree with the full numerical calculation (solid curve)
because perturbation 1 exhibits
the particle-hole symmetry of the hard-core model, which
is distinctly broken here. However, the
dot-dashed curve of perturbation 2 is in excellent agreement
with the solid curve except at large $N$($\sim 1$).
\begin{figure}
\includegraphics[height=2.5in]{fig11.eps}
\caption{
Phase diagram of the soft-core model for $zt/U\ = \ 1$.
Solid curve represents the SF--SS phase boundary at which
the phase transition is continuous.
Dot-dashed curves
show the results of perturbation 1 (Eq. \ref{perturbation1})
and 2 (Eq. \ref{perturbation2}).
Two-dot-dashed lines at half filling and
unit filling show the S$_1$ and S$_2$ phases, respectively.
}
\label{fig11}
\end{figure}
Figure 12 shows the $V$ dependence of
$\delta n$ in the SS phase of the soft-core model for $zt/U\ = \ 1$.
$\delta n$ is larger for larger $N$ at the same $zV/U$
because the critical value of $zV/U$ for the SF--SS transition
is smaller for larger $N$ (Fig. 11).
Unlike the cases of $zt/U\ = \ 0.03$ (Fig. 6) or $zt/U\ = \ 0.3$ (Fig. 10),
$\delta n$ is a smooth increasing function
of $N$ and $zV/U$, and the
two curves of $\delta n$ for two different $N$ values
no longer intersect. Note that
$\delta n$ does not change rapidly
around $zV/U\sim 1$ as expected from the phase diagram (Fig. 11).
\begin{figure}
\includegraphics[height=2.5in]{fig12.eps}
\caption{
$V$ dependence of
$\delta n\ = \ |N_A-N_B|$ in the soft-core model at $N\ = \ 0.3$, 0.5, 0,7, and 0.9 for $zt/U\ = \ 1$.
}
\label{fig12}
\end{figure}
\section{Effect of improved calculation
on the energy of the solid and MI phases}
The energy of the solid and MI phases employed in the previous sections
may be significantly higher than the exact energy.
Hence, in this section,
we improve the calculated energies in these phases
by employing the perturbation theory \cite{Iskin} up to the order of
$t^2/U$ or $t^2/V$
as
\begin{eqnarray}
E_{\rm MI}&\ = \ &\frac{zV}{2}-\frac{2zt^2}{U-V},\\
E_{\rm S_1}&\ = \ &-\frac{zt^2}{2(z-1)V},\\
E_{\rm S_2}&\ = \ &\frac{U}{2}-\frac{zt^2}{(2z-1)V-U}.
\end{eqnarray}
We can obtain the phase diagram by employing these
improved energies.
Hereafter, we assume $z\ = \ 6$ for the 3D cubic lattice.
However, note that this improvement may be excessively favorable
for the solid and MI phases and unfavorable for the SF and SS phases.
Therefore, a correction on the order of $t^2/U$ or $t^2/V$ should also be
applied to the energies of the SF and SS phases
close to the SF (or SS)--solid (or MI) phase boundary.
In addition, the denominators of the equations describing
$E_{\rm MI}$, $E_{{\rm S}_1}$, and $E_{{\rm S}_2}$
diverge for $V\ = \ U$, $V\ = \ 0$ and $(2z-1)V\ = \ U$, respectively.
Hence, we exclude the S$_1$ phase around $V\ = \ 0$ and
the S$_2$ phase around $zV/U\ = \ z/(2z-1)\ = \ 6/11$.
We also neglect the PS into two solids (S$_1$ and S$_2$)
which is unlikely but indeed appears in this approximation.
Figures 13 and 14 show the phase diagrams
for $zt/U\ = \ 0.03$ and
$zt/U\ = \ 0.3$, respectively.
(The phase diagram for $zt/U\ = \ 1$ is the same as that in Fig. 11.)
In both figures, the regions of separated phases become large,
as expected. For $zt/U\ = \ 0.03$, the PS
into the SF and MI phases [PS(SF + MI) phase] appears. Furthermore,
for $zt/U\ = \ 0.3$, the SS phase disappears below half filling
because of existence of the PS(SF + S$_1$) phase.
These results suggest that the solid and MI energies
have somewhat over-improved.
On the other hand, the regions of PS into
the SS and solid phases become large above half filling,
which resembles the results of a 2D QMC study by Sengupta and coworkers \cite{Sengupta1}.
\begin{figure}
\includegraphics[height=2.5in]{fig13.eps}
\caption{
Phase diagram of the soft-core model
for $zt/U\ = \ 0.03$.
Energies of the solid and MI phases are improved (see text).
}
\label{fig13}
\end{figure}
\begin{figure}
\includegraphics[height=2.5in]{fig14.eps}
\caption{
Phase diagram of the soft-core model
for $zt/U\ = \ 0.3$. Energies of the solid
and MI phases are improved (see text).
}
\label{fig14}
\end{figure}
\section{Conclusion}
In this work, we studied the hard-core and soft-core-extended
Hubbard models by the Gutzwiller variational
wave function. We adopted a canonical ensemble and
a linear programming method to include PSs more directly
in our calculations.
In the hard-core model away from half filling,
we showed that the PS(SF + S$_1$) and SS phases are degenerate
above a critical value of the nearest-neighbor interaction $V$,
which is consistent with a previous MF study \cite{Matsuda}.
Unlike the hard-core model,
the soft-core model at half filling has
a possible SS phase between the solid and SF phases
and all the phase transitions are continuous.
Away from half filling, the phase diagram
depends drastically on the transfer integral $t$.
For small $zt/U$, the shape of the SF
region is similar to that of the hard-core model.
The PS(SF + S$_1$) phase appears below half filling
and the SS phase appears only above half filling,
as in the 2D QMC studies.
For intermediate $zt/U$, the SS phase appears close to the
PS below half filling
as in the 3D QMC study.
The phase diagram becomes simpler for large $zt/U$, where
only the continuous SF--SS phase transition appears,
and the critical value of $zV/U$ at the phase boundary
is a smooth decreasing function of $N$.
The nearest-neighbor interaction
dependence of $\delta n\ = \ |N_A-N_B|$,
which shows the density wave order of the SS phase,
is also interesting.
For a small $zt/U$ of 0.03, $\delta n$ is a discontinuous
function of $zV/U$; furthermore,
the SS phase has a small $\delta n$ for small $zV/U$($\lesssim 1$)
and large $\delta n$($\simeq 2N$)
($N_A\simeq 2N$ and $N_B\simeq 0$) for large $zV/U$($\gtrsim 1$).
In addition,
$\delta n$ is larger for smaller (larger) $N$
for $zV/U\lesssim 1$ ($zV/U\gtrsim 1$).
For an intermediate $zt/U$ of 0.3,
the behavior of $\delta n$ is similar to that for a small n
$zt/U\ = \ 0.03$.
In detail, however, unlike the case of $zt/U\ = \ 0.03$,
the two curves of $\delta n$ for two different $N$ values continuously
intersect around $zV/U\simeq 1.1$ because $\delta n$
is a continuous increasing function of $zV/U$ except at large $N$.
For a large $zt/U\ = \ 1$, $\delta n$ is a smooth increasing
function of $N$ and $zV/U$, and the
two curves of $\delta n$ for two different $N$ values
no longer intersect.
Throughout this paper,
we found that our perturbative calculations
determined the phase boundary curves very well
except for large $N$($\sim 1$).
We also studied the effects of the improved perturbative calculation
on the energy of the solid and MI phases. The improvement enlarges the
region of PS into the SS and solid phases above half filling,
which resembles the results of the 2D QMC study. However, the improvement is
excessively favorable for the solid and MI phases,
resulting in an unusual enlargement of the PSs. Therefore,
the energies of the SF and SS phases should also be improved in future work.
Because the Gutzwiller approximation is not precise,
exact numerical calculations such as QMC simulations are needed
to check our results.
Although some details (such as the very complicated
phase diagrams for $N\sim 1$ and small $zt$)
might be artifacts of our approximation,
we believe that the important results in
the phase diagrams and $\delta n$ values in the SS phase
are worth studying in detail.
For instance, the transfer integral dependence of the
entire phase diagram seems to remain an open question
not only in the 3D case but also in the 2D case,
especially for the large $zt/U$ regime, in which
an SS phase below or at half filling might exist.
\begin{acknowledgments}
I sincerely thank D. Yamamoto and I. Danshita
for fruitful discussions.
\end{acknowledgments}
|
\section{Introduction}
The Perona-Malik equation
\begin{equation}
u_{t}=\mathrm{div}\left(\frac{\nabla u}{1+|\nabla
u|^{2}}\right)
\quad\quad
(x,t)\in\Omega\times(0,+\infty)
\label{eqn:PM}
\end{equation}
(where $\Omega\subseteq\mathbb{R}^{n}$ is a bounded open set) is arguably the
most celebrated example of forward-backward diffusion process. It was
introduced by P.~Perona and J.~Malik~\cite{PM} in the context of image
denoising. It is formally the gradient-flow of the functional
\begin{equation}
PM(u):=\frac{1}{2}\int_{\Omega}\log\left(1+|\nabla
u(x)|^{2}\right)\,dx.
\label{defn:PM-funct}
\end{equation}
The forward-backward nature of (\ref{eqn:PM}) depends on the
convex-concave behavior of the integrand in~(\ref{defn:PM-funct}).
Equation (\ref{eqn:PM}) has generated a considerable literature (see
\cite{amann,BFG,BF,BNP,BNP3,BNPT1,Z2,CLMC,nes,E1,E2,E3,GG,GG-tams,GG-cpdes,GG-ifb,G-PMEntire,
guidotti,gl,Z3,Z1}), focussed both on numerical and on analytical
aspects. The big challenge is to reconcile the empirical and
practical efficacy of the method, supported by several numerical
observations, with the expected analytical ill-posedness of a backward
parabolic equation. A satisfactory theory would represent a solution
to the Perona-Malik paradox, as named after~\cite{kich}, but it still
seems to be out of reach.
In a recent paper, P.\ Guidotti~\cite{patrick-DM} introduced a mild
regularization of (\ref{eqn:PM}). He considered the family of
functionals
$$ PM_{\delta}(u):=\frac{1}{2}\int_{\Omega}\log\left(1+|\nabla
u(x)|^{2}\right)\,dx+ \delta
\int_{\Omega}|\nabla u(x)|^{2}\,dx,$$
where $\delta>0$ is a parameter. The integrand remains nonconvex, at
least in the interesting cases where $\delta$ is small, but now it is
convex-concave-convex, and it grows quadratically at infinity. This
is enough to guarantee that the corresponding gradient-flow equation,
which simply reads as
\begin{equation}
u_{t}=\mathrm{div}\left(\frac{\nabla u}{1+|\nabla
u|^{2}}\right)+\delta\,\Delta u
\quad\quad
(x,t)\in\Omega\times(0,+\infty),
\label{eqn:PM-delta}
\end{equation}
has a unique global-in-time solution in the sense of Young measures,
according to the theory developed in \cite{KP,demoulini}.
Several qualitative properties of these solutions are reported in
\cite{patrick-DM}. In particular, it seems that the well-known
staircasing effect of the original Perona-Malik equation is now
replaced by a ``ramping'' effect, namely the tendency of solutions to
alternate flat plateaus and bounded growth regions in a piecewise
fashion (see all figures in~\cite{patrick-DM}).
In \cite{patrick-DM} it is also observed that the limit of
approximated solutions as $\delta\to 0^{+}$ is the trivial stationary
solution frozen in the initial condition, and ``thus the only way to
produce a meaningful limit would involve a rescaling of time in the
process''.
In this paper we follow this path. We take a family of solutions
$v_{\delta}(t)$ of the regularized model (with initial conditions
$v_{\delta}(0)$ converging to some $u_{0}$), and then we speed up the
evolution by considering the family of functions
\begin{equation}
u_{\ep}(t):=v_{4^{-1}\varepsilon^2|\log\varepsilon|^{}}\left(\frac{t}{\varepsilon|\log\varepsilon|}\right)
\quad\quad
t\geq 0.
\label{defn:uep}
\end{equation}
This happens to be the right rescaling factor, in the sense that
$u_{\ep}(t)$ uniformly converges to a nontrivial limit $u(t)$ as $\varepsilon\to
0^{+}$. Moreover $u(t)$ turns out to be the solution of the total
variation flow (see~\cite{TV-nonno,TV-esistenza})
\begin{equation}
u_{t}=\mathrm{div}\left(\frac{\nabla u}{|\nabla u|}\right)
\quad\quad
(x,t)\in\Omega\times(0,+\infty),
\label{eqn:TVF}
\end{equation}
with initial condition $u_{0}$. Any slower time rescaling leads in
the limit to the stationary solution $u(t)\equiv u_{0}$, and any
faster time rescaling produces a limit which for every $t>0$
coincides with the constant function equal to the average of $u_{0}$
in $\Omega$.
All qualitative properties of the total variation flow
(see~\cite{TV-property}) are consistent with the numerical experiments
presented in~\cite{patrick-DM}. Therefore, our convergence result
provides a rigorous justification of all these properties.
The proof of the convergence result involves three main steps. First
of all, we interpret $u_{\ep}(t)$ as a gradient-flow. As it comes
from~\cite{patrick-DM}, $u_{\ep}(t)$ is the solution in the sense of
Young measures of the forward-backward equation associated to the
formal gradient-flow of the nonconvex energy
\begin{equation}
E_{\ep}(u):=\int_{\Omega}
\varphi_{\ep}\left(\strut|\nabla u(x)|\right)\,dx,
\label{defn:Eep}
\end{equation}
where
$$\varphi_{\ep}(\sigma):=\frac{1}{2\varepsilon|\log\varepsilon|}
\log\left(1+\sigma^{2}\right)+\frac \varepsilon 4\,\sigma^{2}.$$
The key point is that $u_{\ep}(t)$ is also the gradient-flow of
the relaxed (convex) energy
\begin{equation}
E_{\ep}^{**}(u):=\int_{\Omega}\varphi_{\ep}^{**}\left(\strut|\nabla u(x)|\right)\,dx,
\label{defn:Fep}
\end{equation}
where $\varphi_{\ep}^{**}$ is the convexification of $\varphi_{\ep}$.
In other words, as a result of the first step we can forget about
Young measures and forward-backward equations, and think of $u_{\ep}(t)$
as the solution of a degenerate forward parabolic equation, or better
as the gradient-flow of a convex (although not strictly convex)
functional. We stress that this is a general fact. Solutions
provided by the theory developed in~\cite{KP,demoulini} always
coincide with gradient-flows of the corresponding convexified (or
relaxed) energies. Quite surprisingly, this has never been observed
in the literature up to our knowledge.
In the second step we compute the Gamma-limit of the energies
(Theorem~\ref{thm:Gamma-conv}), and we discover that
\begin{equation}
\Gamma\,\mbox{--}\!\lim_{\varepsilon\to 0^{+}}E_{\ep}(u)=
\Gamma\,\mbox{--}\!\lim_{\varepsilon\to 0^{+}}E_{\ep}^{**}(u)=
TV(u),
\label{th:Gamma-lim}
\end{equation}
where $TV(u)$ denotes the total variation of $u$. Since the total
variation flow is the gradient-flow of $TV(u)$, our convergence result
is now equivalent to say that the limit of gradient-flows is the
gradient-flow of the Gamma-limit.
This is the content of the third step, where we deduce it from a
general result (Theorem~\ref{thm:main-cmp}) which we state and prove
in the abstract setting of \emph{maximal slope curves} in metric
spaces (see Section~\ref{sec:CMP} for further details and references).
We believe that the scope of the general result goes far beyond this
simple application.
We conclude by discussing how this problem relates to recent
investigations about the ``slow time'' behavior of approximations of
the Perona-Malik equation. The starting point is the observation that
the evolution, despite of the different approximation methods, seems to
exhibit always three different time scales, named ``fast time'',
``standard time'', and ``slow time'' according to~\cite{BF}.
In a fast time of order $o(1)$ solutions develop microstructures in
the concave regime (staircasing). In a time scale of order $O(1)$
(standard time) solutions behave as expected in the original model,
with a smoothing effect in the concave regime, and sharpening of
regions where the gradient is large. At the ``end'' of standard time,
solutions have an almost piecewise constant structure, and this is consistent
with the intuitive idea that piecewise constant functions are
stationary points of $PM(u)$. On the other hand, only constant
functions (and not piecewise constant functions) are stationary points
of the usual approximating models. As a consequence, approximating
solutions exhibit a transition from a piecewise constant structure to
a constant value (equal to the average of $u_{0}$). The transition
turns out to be very slow because there is almost no energy left.
This gives rise to the so called ``slow time'' motion, in which the
plateaus of the piecewise constant function move in the vertical
direction, with jump points which remain fixed in space. Although the
existence of this phase seems to be independent of the approximation
method, the law of the vertical motion does depend on it.
All previous results on this problem are limited to the one
dimensional case. Let $u$ be a piecewise constant function defined in
an interval, and let $S_{u}$ be the (finite or countable) set of its
jump points. Let $J_{x}$ denote the jump height in a point $x\in
S_{u}$, and let us consider the following energies
$$H_{\alpha}(u):=\sum_{x\in S_{u}}|J_{x}|^{\alpha} \quad\mbox{(with
$\alpha\in(0,1]$)},
\quad\quad
H_{0}(u):=\sum_{x\in S_{u}}\log|J_{x}|.$$
In the case of a fourth order regularization of (\ref{eqn:PM}),
corresponding to adding a vanishing second order term to
(\ref{defn:PM-funct}), G.~Bellettini and A.~Fusco~\cite{BF}
conjectured that the vertical motion is governed by the gradient-flow
of $H_{1/2}(u)$. They supported their conjecture by proving the
corresponding Gamma-limit result for the energies. The missing step
is a rigorous proof that also in that case the limit of gradient-flows
is the gradient-flow of the limit. Unfortunately
Theorem~\ref{thm:main-cmp} does not apply to their functionals.
In the case of the semidiscrete scheme (see~\cite{BNP3,nes}) the
vertical motion is governed by the gradient-flow of $H_{0}(u)$, which
in a certain sense represents a limit case. The proof given
in~\cite{nes} exploits a variant of Theorem~\ref{thm:main-cmp},
complicated by the fact that $H_{0}(u)$ is not bounded from below.
What we show in this paper is that in the model proposed
in~\cite{patrick-DM} the slow-time vertical motion is governed by the
gradient-flow of $H_{1}(u)$. This is the opposite limit case, and
phenomena are completely different. The good news is that the limit
energy is convex. This simplifies the analysis, which here can be
carried out in any space dimension, and delivers a well known limit
problem. On the other hand, the relaxation of $H_{1}(u)$ is $TV(u)$,
hence it is finite in the whole space of bounded variation functions.
As a consequence, in this case the slow time motion is not limited to
piecewise constant functions. This is hardly surprising after
reminding that in this model the motion in standard time is trivial.
This paper is organized as follows. In Section~\ref{sec:statements}
we fix notations and we state our convergence results.
Section~\ref{sec:CMP} is devoted to limits of maximal slope curves in
metric spaces. In Section~\ref{sec:proofs} we prove our main result.
\setcounter{equation}{0}
\section{Notations and statements}\label{sec:statements}
Let $n$ be a positive integer, and let $\Omega\subseteq\mathbb{R}^{n}$ be an
open set, which for simplicity we assume to be bounded and an
extension domain (see Definition~3.20 in~\cite{AFP}, satisfied by all
bounded open sets with Lipschitz boundary). The more general ambient
space we consider is $L^{2}(\Omega)$. We write
$\|u\|_{L^{p}(\Omega)}$, or simply $\|u\|_{p}$, to denote the $p$-norm
(with $p\in[1,+\infty]$) of a function $u\in L^{p}(\Omega)$. All the
energies we consider, and in particular $E_{\ep}(u)$ and $E_{\ep}^{**}(u)$, are
always thought as defined in the whole space $L^{2}(\Omega)$ by
setting them equal to $+\infty$ outside their natural domain. We
write $BV(\Omega)$ to denote the space of all functions $u\in
L^{2}(\Omega)$ with finite total variation $TV(u)$. Once again, we
think of $TV(u)$ as defined for every $u\in L^{2}(\Omega)$, with
$TV(u)<+\infty$ if and only if $u\in BV(\Omega)$.
For every $\delta>0$ we consider equation (\ref{eqn:PM-delta}), with
Neumann boundary conditions, and an initial datum. It has been shown
in~\cite{patrick-DM} that solutions $v_{\delta}(t)$ exist in a suitable weak
sense. For every $\varepsilon\in(0,1)$, we define $u_{\ep}(t)$ by rescaling
$v_{\delta}(t)$ according to (\ref{defn:uep}). It turns out that
$u_{\ep}(t)$ is a solution in the same weak sense of equation
\begin{equation}
{u_{\ep}}_{t}=\frac{1}{\varepsilon|\log\varepsilon|}\,\mathrm{div}\left(
\frac{\nabla u_{\ep}}{1+|\nabla
u_{\ep}|^{2}}\right)+\frac \varepsilon 4\, \Delta u_{\ep}
\quad\quad
(x,t)\in\Omega\times(0,+\infty),
\label{eqn:PM-ep}
\end{equation}
with Neumann boundary conditions
\begin{equation}
\frac{\partialu_{\ep}}{\partial n}(x,t)=0
\quad\quad
(x,t)\in\partial\Omega\times(0,+\infty),
\label{eqn:PM-NBC}
\end{equation}
and initial condition
\begin{equation}
u_{\ep}(x,0)=u_{0\ep}(x)
\quad\quad
x\in\Omega.
\label{eqn:uep-data}
\end{equation}
In the following result we collect properties of $u_{\ep}(t)$.
\begin{thmbibl}[Properties of rescaled approximating solutions]\label{thm:uep}
Let $n$ be a positive integer, let $\Omega\subseteq\mathbb{R}^{n}$ be a
bounded extension domain, let $\varepsilon\in (0,1)$, and let $u_{0\ep}\in
L^{2}(\Omega)$. Let $E_{\ep}^{**}$ be the functional defined in
(\ref{defn:Fep}).
Then the following properties hold true.
\begin{enumerate}
\renewcommand{\labelenumi}{(\arabic{enumi})}
\item
\label{stat:yms} \emph{(Weak Young measure solution and
regularity)} There exists a unique function $u_{\ep}(t)$ and a
(not necessarily unique) gradient Young measure $\nu_{\varepsilon}$ in
$\Omega\times[0,+\infty)$ such that the pair
$(u_{\ep},\nu_{\varepsilon})$ is a weak Young measure valued solution of
problem (\ref{eqn:PM-ep}) through (\ref{eqn:uep-data}) in the
sense of~\cite{KP,demoulini}.
Moreover, we have that
$$u_{\ep}\in C^{0}\left([0,+\infty);L^{2}(\Omega)\right)\cap
C^{1}\left((0,+\infty);L^{2}(\Omega)\right),$$
and for every $t>0$ we have that $u_{\ep}(t)$ is regular enough
so that the right-hand side of (\ref{eqn:fd-pde}) lies in
$L^{2}(\Omega)$.
\item \emph{(Degenerate forward parabolic equation)} The
function $u_{\ep}$ of statement~(\ref{stat:yms}) is the unique
solution in $\Omega\times(0,+\infty)$ of the partial
differential equation
\begin{equation}
{u_{\ep}}_{t}=-\nablaE_{\ep}^{**}(u_{\ep})=
\mathrm{div}\left[(\varphi_{\ep}^{**})'
\left(\strut|\nablau_{\ep}|\right)
\frac{\nablau_{\ep}}{|\nablau_{\ep}|}\right],
\label{eqn:fd-pde}
\end{equation}
with Neumann boundary conditions (\ref{eqn:PM-NBC}), and
initial condition (\ref{eqn:uep-data}).
\item \emph{(Gradient-flow integral inequality)} The
function $u_{\ep}$ of statement~(\ref{stat:yms})
is the unique function satisfying (\ref{eqn:uep-data}) and
the inequality
\begin{equation}
E_{\ep}^{**}(u_{\ep}(s))-E_{\ep}^{**}(u_{\ep}(t))\geq
\frac{1}{2}\int_{s}^{t}\|u_{\ep}'(\tau)\|_{2}^{2}\,d\tau+
\frac{1}{2}\int_{s}^{t}\|\nablaE_{\ep}^{**}(u_{\ep}(\tau))\|_{2}^{2}\,d\tau
\label{hp:uep-gf}
\end{equation}
for every $0\leq s\leq t$.
\item \label{stat:Linfty}\emph{($L^{p}$ estimate)}
If $u_{0\ep}\in L^{p}(\Omega)$ for some $p\in[1,+\infty]$, then
$u_{\ep}(t)\in L^{p}(\Omega)$ for every $t\geq 0$, and the function
$t\to\|u_{\ep}(t)\|_{L^{p}(\Omega)}$ is nonincreasing.
\end{enumerate}
\end{thmbibl}
Theorem~\ref{thm:uep} above shows that the function $u_{\ep}(t)$ can be
characterized in at least three different ways, either as the solution
in the sense of Young measures of a forward-backward equation, or as
the solution of a forward degenerate parabolic equation (for example
in the sense of~\cite{brezis}), or as a maximal slope curve
(gradient-flow inequalities). What we need in this paper is only the
last one. Solutions generate a contraction semigroup in
$L^{2}(\Omega)$.
As far as we know this equivalence has never been stated explicitly.
Nevertheless, it follows from some general facts which nowadays are
quite well known, and which we now recall briefly.
First of all, the three approaches lead to a \emph{unique} solution.
Uniqueness follows in the first case from the strong requirements
imposed on the structure of the corresponding Young measure, in the
second case from the contraction property of the semigroup generated
by a forward parabolic equation, in the third case from the convexity
of the energy $E_{\ep}^{**}(u)$.
Secondly, in all three approaches the solution is usually obtained (or
at least it can be obtained) as the limit of approximated solutions
constructed via an iterated minimization process, known as minimizing
movement (see \cite{dg-min}). As already observed
in~\cite{KP,demoulini,patrick-DM}, the minimization procedure gives
the same result when applied to $E_{\ep}(u)$ or $E_{\ep}^{**}(u)$.
In conclusion, all three approaches define a unique solution through
an analogous procedure, hence the solution is the same.
In this paper we are interested in the behavior of $u_{\ep}(t)$ as
$\varepsilon\to 0^{+}$. Following the gradient-flow approach, the first thing
to do is understanding the limit behavior of the energies. This is
the content of next result.
\begin{thm}[Gamma-convergence and compactness]\label{thm:Gamma-conv}
Let $n$ be a positive integer, and let $\Omega\subseteq\mathbb{R}^{n}$ be
a bounded extension domain. For every $\varepsilon>0$, let
$E_{\ep}:L^{2}(\Omega)\to[0,+\infty]$ and
$E_{\ep}^{**}:L^{2}(\Omega)\to[0,+\infty]$ be defined as in
(\ref{defn:Eep}) and (\ref{defn:Fep}), respectively, if $u\in
H^{1}(\Omega)$, and $+\infty$ otherwise.
Then we have the following conclusions.
\begin{enumerate}
\renewcommand{\labelenumi}{(\arabic{enumi})}
\item \emph{(Gamma-convergence)} We have that
(\ref{th:Gamma-lim}) holds true with respect to the topology
of $L^{2}(\Omega)$.
\item \emph{(Compactness)} Let
$\{u_{\ep}\}_{\varepsilon\in(0,1)}\subseteq L^{2}(\Omega)$ be a family
of functions such that
\begin{equation}
\sup_{\varepsilon\in(0,1)}\left\{\strut\|u_{\ep}\|_{\infty}+
E_{\ep}^{**}(u_{\ep})\right\}<+\infty.
\label{hp:cpt}
\end{equation}
Then the family $\{u_{\ep}\}$ is relatively compact in
$L^{2}(\Omega)$.
\end{enumerate}
\end{thm}
The gradient-flow of the limit functional $TV(u)$ is the so called
\emph{total variation flow}, and the corresponding partial
differential equation is (\ref{eqn:TVF}). The right-hand side of
(\ref{eqn:TVF}) needs to be interpreted in a suitable weak sense when
the gradient vanishes, and this happens in large regions because
solutions tend to develop flat plateaus. A general existence and
uniqueness result was proved by F.\ Andreu, C.\ Ballester, V.\
Caselles, and J.\ M.\ Maz\'{o}n~\cite{TV-esistenza} (see
also~\cite{TV-property}) using the theory of accretive operators in
Banach spaces. In~\cite{TV-esistenza} the operator in the right-hand
side of~(\ref{eqn:TVF}) is interpreted as the limit of the
$p$-Laplacian as $p\to 1^{+}$.
In this paper we limit ourselves to initial data in $L^{2}(\Omega)$,
in which case existence of a unique solution is provided also by the
theory of maximal monotone operators~\cite{brezis}, as explained
in~\cite{TV-nonno}. In this context the right-hand side
of~(\ref{eqn:TVF}) is the subdifferential of the convex functional
$TV(u)$. As in the case of approximating problems, this formulation
is equivalent to the gradient-flow integral inequality
$$TV(u(s))-TV(u(t))\geq
\frac{1}{2}\int_{s}^{t}\|u'(\tau)\|_{2}^{2}\,d\tau+
\frac{1}{2}\int_{s}^{t}\|\nabla TV(u(\tau))\|_{2}^{2}\,d\tau$$
for every $0\leq s\leq t$, where $\|\nabla TV(u(\tau))\|_{2}$ is the
minimal norm of an element in the subdifferential of the functional
$TV$ in the point $u(\tau)$, which in turn coincides with the slope
of $TV$ in $u(\tau)$ as defined in Section~\ref{sec:CMP} in an
abstract metric setting. This is the characterization of the total
variation flow which we need in this paper.
Our main result is the convergence of $u_{\ep}(t)$ to the solution of the
total variation flow with the same boundary conditions, and the limit
initial datum. We point out that we do not assume initial data to be
a recovery sequence, and we do not ask their energy to be bounded.
\begin{thm}[Global-in-time convergence]\label{thm:main}
Let $n$ be a positive integer, let $\Omega\subseteq\mathbb{R}^{n}$ be a
bounded extension domain, let $u_{0}\in L^{2}(\Omega)$, and let
$\{u_{0\ep}\}_{\varepsilon\in(0,1)}\subseteq L^{2}(\Omega)$ be a family of
functions such that
\begin{equation}
\lim_{\varepsilon\to 0^{+}}u_{0\ep}=u_{0}
\quad\quad
\mbox{in }L^{2}(\Omega).
\label{hp:data-conv}
\end{equation}
For every $\varepsilon\in(0,1)$, let $u_{\ep}$ be the solution of the
rescaled approximating problem with initial condition $u_{0\ep}$, in
the sense of Theorem~\ref{thm:uep}. Let $u(t)$ be the solution of
the total variation flow with Neumann boundary conditions and
initial datum $u_{0}$.
Then we have that $u_{\ep}(t)\to u(t)$ in
$C^{0}\left([0,+\infty);L^{2}(\Omega)\right)$, namely
\begin{equation}
\lim_{\varepsilon\to 0^{+}}\sup_{t\geq 0}
\|u_{\ep}(t)-u(t)\|_{L^{2}(\Omega)}=0.
\label{th:main}
\end{equation}
\end{thm}
We conclude this section with a heuristic argument which justifies
the rescaling leading to (\ref{defn:Eep}).
Let us consider the one dimensional case where $\Omega=(-1,1)$, let
$J>0$, and let $v\in BV((-1,1))$ be the piecewise constant function
equal to $-J/2$ in $(-1,0)$, and equal to $J/2$ in $(0,1)$. Let
$\eta>0$, and let $h\in H^{1}(\mathbb{R})$ be a function such that
$h(x)= J/2$ for every $x\geq \eta$, $h(x)= -J/2$ for every $x\leq
-\eta$, and $h'(x)> 0$ for almost every $x\in (-\eta, \eta)$.
For every $\varepsilon>0$, let $v_{\ep}(x)$ be the approximation of $v$ defined as $v_{\ep}(x):=
h(x/\varepsilon)$ for every $x\in (-1,1)$.
For every $\varepsilon \leq \eta^{-1}$, plugging $v_{\ep}$ into (\ref{defn:Eep}),
with a variable change we obtain that
$$E_{\ep}(v_{\ep})=\frac{1}{2|\log\varepsilon|}\int_{-\eta}^{\eta}
\log\left(1+\frac{1}{\varepsilon^{2}}\left[h'(x)\right]^{2}\right)dx+
\frac 1 4 \int_{-\eta}^{\eta}\left[h'(x)\right]^{2}dx.$$
Letting $\varepsilon\to 0^{+}$ we find that
\begin{equation}
\liminf_{\varepsilon\to 0^{+}}E_{\ep}(v_{\ep})\geq
2\eta+\frac 1 4 \int_{-\eta}^{\eta}\left[h'(x)\right]^{2}dx.
\label{heu:liminf}
\end{equation}
In order to estimate the right-hand side, we first minimize with
respect to $h$, and we discover that the optimal choice is the
function $h(x)$ defined as $J(2\eta)^{-1}x$ for every
$x\in(-\eta,\eta)$. Then we compute the integral, and finally we
apply the inequality between arithmetic and geometric mean to deduce
that
\begin{equation}
2\eta+\frac 1 4 \int_{-\eta}^{\eta}\left[h'(x)\right]^{2}dx\geq
2\eta+\frac{J^{2}}{8\eta}\geq J.
\label{heu:AM-GM}
\end{equation}
Estimates (\ref{heu:liminf}) and (\ref{heu:AM-GM}) suggest that the
cost of a jump could be the jump height, which leads to conjecture
that the Gamma-limit is $TV(u)$ for a general $u$ in any dimension.
\setcounter{equation}{0}
\section{Passing to the limit in maximal slope curves}\label{sec:CMP}
The abstract theory of gradient-flows in metric spaces was introduced
in~\cite{dgmt}, and then developed by the same authors and
collaborators in a series of papers (see~\cite{DMT,MST-SNS} and the
references quoted therein). For a modern presentation we refer
to~\cite{AGS}. Here we just recall some basic definitions.
Let $(X,d)$ be a metric space, let
$\overline{\mathbb{R}}:=\mathbb{R}\cup\{-\infty,+\infty\}$ be the extended real line, and
let $F:X\to\overline{\mathbb{R}}$ be any function. The (descending) \emph{slope}
$|\nabla F|(x)$ of $F$ in $x$ is defined to be $+\infty$ if
$F(x)\not\in\mathbb{R}$, and otherwise
$$|\nabla F|(x):=\limsup_{y\to x}\frac{\max\{F(x)-F(y),0\}}{d(x,y)}
\in[0,+\infty].$$
For every $T>0$, the space $AC^{2}\left([0,T];X\right)$ is the set of
all functions $v:[0,T]\to X$ for which there exists $g\in
L^{2}((0,T))$ such that
\begin{equation}
d(v(t),v(s))\leq\int_{s}^{t}g(\tau)\,d\tau
\quad\quad
\forall\, 0\leq s\leq t\leq T.
\label{defn:metric-deriv}
\end{equation}
It can be seen that there exists a smallest function $g(t)$
satisfying (\ref{defn:metric-deriv}). This function is called the
\emph{metric derivative} of $v$, and it is denoted by $|v'|(t)$.
A \emph{maximal slope curve} for $F$ in $[0,T]$ is a triple
$(u,\psi,E)$ where
\begin{itemize}
\item $u\in AC^{2}\left([0,T];X\right)$,
\item $\psi:[0,T]\to\mathbb{R}$ is a nonincreasing function
such that for every $0\leq s\leq t\leq T$ we have that
\begin{equation}
\psi(s)-\psi(t)\geq \frac 1 2
\int_s^t |u'|^2(\tau)\,d\tau+
\frac 1 2 \int_s^t \left|\nabla F \right|^2
(u(\tau))\,d\tau,
\label{eqn:cmp-lim-psi}
\end{equation}
\item $E\subseteq [0,T]$ is a set with Lebesgue measure
equal to 0 such that
\begin{equation}
\psi(t)=F(u(t))
\quad\quad
\forall t\in[0,T]\setminus E.
\label{eqn:psi-ae}
\end{equation}
\end{itemize}
To be more precise, the second integral in the right-hand side of
(\ref{eqn:cmp-lim-psi}) should be an upper integral, since at this level of
generality there is no reason for the function $t\to|\nabla F|(u(t))$
to be measurable. On the other hand, it can be easily proved that
actually it is always true that $|u'|(t)=|\nabla F|(u(t))$ for almost
every $t\in[0,T]$, which implies the required measurability.
When $F$ is a $C^{1}$ function in a Hilbert space $X$, this weak
formulation is equivalent to the classical one, namely to asking that
$u'(t)=-\nabla F(u(t))$ for every $t\in[0,T]$.
Besides generality, the advantage of this weak formulation is that
inequalities and integrals are more stable than equalities and
derivatives. It follows that maximal slope curves exist under general
assumptions on $F$ (see Theorem 2.3.1 in~\cite{AGS}), and are quite
stable when passing to the limit, both with respect to initial
conditions, and with respect to functionals. Results in this
direction are contained in~\cite{DMT} and~\cite{ss-cpam} in a Hilbert
setting, and in~\cite{ss-09} in a metric setting, but assuming that
initial data are a recovery sequence.
Here we state a quite general result, used in a special case also
in~\cite{nes}.
\begin{thm}[Limits of maximal slope curves]\label{thm:main-cmp}
Let $X$ be a metric space, let $F: X \to\overline{\mathbb{R}}$ be a
function, and let $F_n: X \to\overline{\mathbb{R}}$ be a sequence of
functions.
Let us assume that for every $x\in X$, and every pair of sequences
$\{n_{k}\}\subseteq\mathbb{N}$ and $\{x_{k}\}\subseteq X$, we have the
implication
\begin{equation}
\fbox{$\begin{array}{r}
n_{k}\to +\infty,\ x_{k}\to x \\
\noalign{\vspace{1ex}}
\displaystyle{\sup_{k \in \mathbb{N}}
\left\{|F_{n_{k}}(x_{k})|+|\nabla F_{n_{k}}
|(x_{k})\strut\right\} < +\infty}
\end{array}$}
\Longrightarrow
\fbox{$\begin{array}{l}
\displaystyle{\lim_{k\to +\infty}F_{n_{k}}(x_{k})=F(x).} \\
\noalign{\vspace{1ex}}
\displaystyle{\liminf_{k\to +\infty}|\nabla
F_{n_{k}}|(x_{k}) \geq |\nabla F|(x).}
\end{array}$}
\label{hp:slope-energy}
\end{equation}
Let $T>0$, and for every $n\in \mathbb{N}$ let $(u_n,\psi_{n},E_{n})$ be
a maximal slope curve for $F_n$ in $[0,T]$. Let us assume that
\begin{equation}
\sup_{n\in\mathbb{N}}\;\max_{t\in[0,T]}|\psi_{n}(t)|
< +\infty,
\label{hp:sup-fn}
\end{equation}
and that $\{u_{n}\}$ has a pointwise limit, namely there
exists $u:[0,T]\to X$ such that
$$\lim_{n\to +\infty}u_{n}(t)=u(t)
\quad\quad
\forall t\in [0,T].$$
Then there exist $\psi$ and $E$ such that $(u,\psi,E)$ is a
maximal slope curve for $F$.
\end{thm}
\paragraph{\textmd{\emph{Proof}}}
From the definition of maximal slope curve, for every
$n\in\mathbb{N}$ we have that
\begin{equation}
\psi_n(s)-\psi_n(t)\geq \frac 1 2 \int_s^t
|u_n'|^2(\tau)\,d\tau+ \frac 1 2 \int_s^t \left|\nabla F_n \right|^2
(u_n(\tau))\,d\tau
\label{eqn:cmpn}
\end{equation}
for every $0\leq s\leq t\leq T$, and
\begin{equation}
\psi_{n}(t)=F_{n}(u_{n}(t))
\quad\quad
\forall t\in[0,T]\setminus E_{n}.
\label{eqn:psi-ae-n}
\end{equation}
The functions $\psi_n(t)$ are nonincreasing, and equi-bounded because
of (\ref{hp:sup-fn}). Thus the usual compactness result for monotone
functions (known as Helly's Lemma, see for
example~\cite[Lemma~3.3.3]{AGS}) implies the existence of a
nonincreasing function $\psi:[0,T]\to [0,+\infty)$ such that (up to
subsequences, not relabeled)
\begin{equation}
\lim_{n\to +\infty}\psi_n(t)=\psi(t)
\quad\quad
\forall t\in[0,T].
\label{defn:psi}
\end{equation}
Let us consider now the right-hand side of
(\ref{eqn:cmpn}).
Setting $s=0$ and $t=T$, and using once more assumption
(\ref{hp:sup-fn}), we obtain that
\begin{equation}
\sup_{n\in\mathbb{N}}\int_{0}^{T}|u_n'|^2(\tau)\,d\tau<+\infty,
\label{est:t1-deriv}
\end{equation}
\begin{equation}
\sup_{n\in\mathbb{N}}\int_{0}^{T}\left|\nabla F_n \right|^2
(u_n(\tau))\,d\tau<+\infty.
\label{est:t1-slope}
\end{equation}
From (\ref{est:t1-deriv}) we easily deduce that $u\in
AC^{2}\left([0,T];X\right)$, and
\begin{equation}
\liminf_{n\to +\infty}\int_s^t |u_n'|^2(\tau)\,d\tau\geq
\int_s^t |u'|^2(\tau)\,d\tau
\quad\quad
\forall\,0\leq s\leq t\leq T.
\label{est:liminf-deriv}
\end{equation}
From (\ref{est:t1-slope}) and Fatou's Lemma we obtain that
$$\int_{0}^{T_{}} \left(\liminf_{n\to +\infty}\left|\nabla F_n
\right|^2 (u_n(\tau))\right)\,d\tau\leq \liminf_{n\to
+\infty}\int_{0}^{T_{}} \left|\nabla F_n \right|^2
(u_n(\tau))\,d\tau<+\infty,$$
hence there exists a set $E' \subseteq[0,T]$, with Lebesgue
measure equal to 0, such that
\begin{equation}
\liminf_{n\to +\infty}\left|\nabla F_n \right|
(u_n(t))<+\infty
\quad\quad
\forall t\in[0,T]\setminus E'.
\label{est:slliminf}
\end{equation}
Let us introduce the set
$$E := E' \cup \left( \bigcup_{n\in\mathbb{N}} E_n\right),$$
which has clearly Lebesgue measure equal to 0. Let us consider any
$t\in[0,T]\setminus E$. Since $t\not\in E_{n}$, from
(\ref{eqn:psi-ae-n}) and (\ref{hp:sup-fn}) we have that
\begin{equation}
\sup_{n\in\mathbb{N}}\left|F_{n}(u_{n}(t)) \right| =
\sup_{n\in\mathbb{N}}|\psi_{n}(t)|<+\infty.
\label{est:bounded-energy}
\end{equation}
Moreover, due to (\ref{est:slliminf}) there exists a ($t$-dependent)
sequence $n_{k}\to +\infty$ such that
\begin{equation}
\lim_{k\to +\infty}|\nabla F_{n_{k}}|(u_{n_{k}}(t))=
\liminf_{n\to +\infty}|\nabla F_{n}|(u_{n}(t))<+\infty.
\label{est:bounded-slope}
\end{equation}
Thanks to (\ref{est:bounded-energy}) and (\ref{est:bounded-slope}),
sequences $\{n_{k}\}$ and $\{u_{n_{k}}(t)\}$ satisfy the assumptions
in the left-hand side of (\ref{hp:slope-energy}). If follows that for
every $t\in[0,T]\setminus E$ we have that
$$\psi(t)=\lim_{n\to +\infty}\psi_{n}(t)=\lim_{k\to +\infty}\psi_{n_k}(t)=
\lim_{k\to +\infty}F_{n_{k}}(u_{n_{k}}(t))=
F(u(t)),$$
which proves (\ref{eqn:psi-ae}), and
$$|\nabla F|(u(t))\leq \liminf_{k\to +\infty}|\nabla
F_{n_{k}}|(u_{n_{k}}(t))=\!
\lim_{k\to +\infty}|\nabla F_{n_{k}}|(u_{n_{k}}(t))=
\liminf_{n\to +\infty}|\nabla
F_{n}|(u_{n}(t)),$$
so that one more application of Fatou's Lemma gives that
\begin{equation}
\int_s^t \left|\nabla F \right|^2 (u(\tau))\,d\tau\leq
\liminf_{n\to +\infty}\int_s^t\left|\nabla F_n \right|^2
(u_n(\tau))\,d\tau
\quad\quad
\forall\,0\leq s\leq t\leq T.
\label{est:liminf-slope}
\end{equation}
We can now take the $\liminf$ of both sides of (\ref{eqn:cmpn}).
Thanks to (\ref{defn:psi}), (\ref{est:liminf-deriv}), and
(\ref{est:liminf-slope}) we obtain that (\ref{eqn:cmp-lim-psi}) holds
true. This completes the proof that $(u,\psi,E)$ is a maximal slope
curve for $F$.{\penalty 10000\mbox{$\quad\Box$}}
\medskip
Thanks to Theorem~\ref{thm:main-cmp} above, any convergence result for
maximal slope curves is reduced to verifying three assumptions. The
first one is the existence of a pointwise limit, namely a compactness
result. The second one is estimate (\ref{hp:sup-fn}), which in
general follows from suitable assumptions on the sequence of initial
data and some boundedness from below of the functionals. The third and
more important assumption is (\ref{hp:slope-energy}). In the last
part of this section we show that (\ref{hp:slope-energy}) follows from
Gamma-convergence in a class of functionals which contains all convex
functionals in Banach spaces.
\begin{defn}[Slope Cone Property]\label{defn:scp}
\begin{em}
Let $(X,d)$ be a metric space. A function $F: X \to\overline{\mathbb{R}}$
satisfies the \emph{Slope Cone Property} if
$$F(y) \geq F(x) - |\nabla F |(x)\cdot d(x,y)$$
for every $y\in X$ and every $x\in X$ such that $F(x)\in\mathbb{R}$
and $|\nabla F |(x)<+\infty$.
\end{em}
\end{defn}
\begin{rmk}\label{rmk:convex->scp}
\begin{em}
Let $X$ be a Banach space. Then every \emph{convex} function
$F: X \to [0, +\infty]$ fulfils the Slope Cone Property. If
in addition $F$ is lower semicontinuous, then its slope
$|\nabla F|(x)$ is lower semicontinuous.
On the other hand, also in Banach spaces there do exist
nonconvex functions satisfying the Slope Cone Property. An
example is $F(x)=-|x|$ in $\mathbb{R}$.
\end{em}
\end{rmk}
\begin{prop}\label{prop:scp}
Let $X$ be a metric space, and let $F_n: X \to\overline{\mathbb{R}}$ be a sequence of
functions satisfying the Slope Cone Property. Let us assume that
there exists
\begin{equation}
F(x):=\Gamma\,\mbox{--}\!\lim_{n\to +\infty} F_n(x).
\label{hp:gamma-conv}
\end{equation}
Then the sequence $\{F_{n}\}$ satisfies assumption
(\ref{hp:slope-energy}) of Theorem~\ref{thm:main-cmp}.
\end{prop}
\paragraph{\textmd{\emph{Proof}}}
Let $n_{k}\to +\infty$ and $x_{k}\to x$ be two sequences as in
assumption (\ref{hp:slope-energy}). Let $M$ be the supremum in the
left-hand side of (\ref{hp:slope-energy}). Let $z_{n}\to x$ be a
recovery sequence for $x$, namely a sequence such that
$F_{n}(z_{n})\to F(x)$.
From the Slope Cone Property, and the uniform bound on slopes, it
follows that
$$F_{n_{k}}(z_{n_{k}}) \geq F_{n_{k}}(x_{k}) - |\nabla F_{n_{k}}
|(x_{k})\cdot d(x_{k},z_{n_{k}}) \geq F_{n_{k}}(x_{k}) - M \,
d(x_{k},z_{n_{k}}).$$
Taking the $\limsup$ of both sides we obtain that
$$F(x) = \limsup_{k\to+\infty} F_{n_{k}}(z_{n_{k}}) \geq \limsup_{k\to+\infty}
F_{n_{k}}(x_k).$$
The opposite inequality with the $\liminf$ follows from assumption
(\ref{hp:gamma-conv}). This proves the first limit in the right-hand
side of (\ref{hp:slope-energy}).
Let us prove now the $\liminf$ inequality for slopes. Let $L$ denote
the $\liminf$ in the left-hand side of (\ref{hp:slope-energy}). Let
us take any $y\in X$, and let $y_{n}\to y$ be a corresponding recovery
sequence. Due to the Slope Cone Property we have that
$$F_{n_{k}}(y_{n_{k}})\geq F_{n_{k}}(x_{k})-
|\nabla F_{n_{k}}|(x_{k})\cdot d(x_{k},y_{n_{k}}).$$
We already proved that the first term in the right-hand side tends to
$F(x)$. Therefore, taking the $\limsup$ of both sides we obtain that
$F(y)\geq F(x)-L\,d(x,y)$. Since $y$ is arbitrary, this easily
implies that $L\geq|\nabla F|(x)$, which completes the proof.{\penalty 10000\mbox{$\quad\Box$}}
\medskip
We conclude by mentioning a straightforward extension of
Definition~\ref{defn:scp} and Proposition~\ref{prop:scp} in the same
spirit of \cite{DMT, MST-SNS}.
\begin{rmk}
\begin{em}
One can weaken Definition~\ref{defn:scp} by asking that
$F:X\to\mathbb{R}\cup\{+\infty\}$
(so we exclude the value $-\infty$),
and there exists a continuous function
$\Phi:X^{2}\times\mathbb{R}^{3}\to\mathbb{R}$ such that $\Phi(x,x,u,v,w)=0$
for every $(x,u,v,w)\in X\times\mathbb{R}^{3}$, and
$$F(y) \geq F(x) - |\nabla F |(x)\cdot d(x,y)-\Phi\left(\strut
x,y,F(x),F(y),|\nabla F|(x)\right)\cdot d(x,y)$$
for every $(x,y)\in X^{2}$ such that $F(x)\in\mathbb{R}$,
$F(y)\in\mathbb{R}$, and $|\nabla F|(x)\in\mathbb{R}$.
We call this property ``$\Phi$ Slope Cone Property''. In
a Banach space it is fulfilled, for example, by the sum of a
$C^{1}$ function and a convex function.
It can be easily proved that Proposition~\ref{prop:scp} holds
true also if all functions $F_{n}$ satisfy the $\Phi$ Slope
Cone Property with respect to the same function $\Phi$.
\end{em}
\end{rmk}
\setcounter{equation}{0}
\section{Proofs}\label{sec:proofs}
\subsection{Gamma-convergence and compactness}
The first equality in (\ref{th:Gamma-lim}) is a general property of
Gamma-convergence. So we can concentrate on the second one.
A standard approach, suggested by the heuristic argument at the end of
Section~\ref{sec:statements}, involves a blow up argument in order to
reduce the Gamma-liminf inequality to minimizing the right-hand side
of (\ref{heu:liminf}) with respect to $\eta$ and $h$, and a proof of
the Gamma-limsup inequality via the density of piecewise constant
functions with smooth level sets, for which a recovery sequence can be
constructed by adapting the optimal profile in the direction
orthogonal to level sets.
In both cases we follow a different and more elementary approach,
which exploits that our functionals depend only on the gradient.
\paragraph{\emph{\textmd{Gamma-liminf inequality}}}
We claim that for every $a\in(0,1)$ and $b\in(0,1)$ there exists
$\varepsilon_{1}\in(0,1)$ such that
\begin{equation}
\varphi_{\ep}^{**}(\sigma)\geq a|\sigma|-b
\quad\quad
\forall\sigma\in\mathbb{R},\ \forall\varepsilon\in(0,\varepsilon_{1}).
\label{th:phiep>}
\end{equation}
If we prove this claim, then for every $u\in L^{2}(\Omega)$ and every
$\varepsilon\in(0,\varepsilon_{1})$ we have that
$$E_{\ep}^{**}(u)=\int_\Omega \varphi_{\ep}^{**}\left(\strut|\nabla u(x)|\right)\,dx \geq
\int_\Omega
\left(\strut a|\nabla u(x)|-b\right)dx = a\, TV(u)-b|\Omega|.$$
Since the functional $TV(u)$ is lower
semicontinuous, this proves that
$$\Gamma\,\mbox{--}\liminf_{\varepsilon \to 0^{+}}E_{\ep}^{**}(u) \geq a\,TV(u)-b|\Omega|.$$
Letting $a\to 1^{-}$ and $b\to 0^{+}$, we obtain the required
inequality.
So we are left to prove (\ref{th:phiep>}). Since the right-hand side
is convex, it is enough to prove (\ref{th:phiep>}) with $\varphi_{\ep}$
instead of $\varphi_{\ep}^{**}$. Moreover, without loss of generality we can
assume that $\sigma\geq 0$. Now we distinguish four cases.
If $\sigma\in[0,b]$, then we have that
$$\varphi_{\ep}(\sigma)\geq 0\geq ab-b\geq a \sigma-b.$$
If $\sigma\in[b,(e^{2}-1)^{1/2}]$, then we have that
$$\varphi_{\ep}(\sigma)-a\sigma\geq\frac{1}{2\varepsilon|\log\varepsilon|}\log(1+b^{2})-
a\sqrt{e^{2}-1}.$$
The right-hand side tends to $+\infty$ as $\varepsilon\to 0^{+}$, so it is
greater than $-b$ when $\varepsilon$ is small enough.
If $\sigma\in[(e^{2}-1)^{1/2},(\varepsilon|\log\varepsilon|)^{-1}]$, then we have that
$$\varphi_{\ep}(\sigma)-a\sigma\geq\frac{1}{2\varepsilon|\log\varepsilon|}\log e^{2}-
\frac{a}{\varepsilon|\log\varepsilon|}=
\frac{1-a}{\varepsilon|\log\varepsilon|},$$
so that the conclusion follows as in the previous case.
Finally, when $\sigma\geq(\varepsilon|\log\varepsilon|)^{-1}$ we apply the inequality
between arithmetic mean and geometric mean, and we obtain that
$$\varphi_{\ep}(\sigma)\geq \frac{\log\sigma}{\varepsilon|\log\varepsilon|}+
\frac{\varepsilon}{4}\,\sigma^{2}\geq\sigma\cdot
\left(\frac{\log\sigma}{|\log\varepsilon|}\right)^{1/2}
\geq\sigma\cdot\left\{\frac{1}{|\log\varepsilon|}\log\left(
\frac{1}{\varepsilon|\log\varepsilon|}\right)\right\}^{1/2}.$$
It is not difficult to see that the coefficient of $\sigma$ tends to
1 as $\varepsilon\to 0^{+}$, hence it is greater than $a$ when $\varepsilon$ is small
enough (in this point it is essential that $a<1$).
This completes the proof of (\ref{th:phiep>}), hence also of
the Gamma-liminf inequality.
\paragraph{\emph{\textmd{Gamma-limsup inequality}}}
By a classical density argument it is enough to find a recovery
sequence for all functions $u\in C^{1}(\Omega)$ whose gradient is
bounded in $\Omega$. To this end, it is enough to show that for any
such function we have that
\begin{equation}
\limsup_{\varepsilon\to 0^{+}}\int_{\Omega}
\varphi_{\ep}^{**}\left(\strut|\nabla u(x)|\right)\,dx
\leq TV(u).
\label{th:gamma-limsup}
\end{equation}
In turn, (\ref{th:gamma-limsup}) is proved if we show that for every
$M>0$ we have that
$$\varphi_{\ep}^{**}(\sigma)\leq a_{\varepsilon}|\sigma|
\quad\quad
\forall\sigma\in[-M,M]$$
for a suitable coefficient $a_{\varepsilon}$ which tends to $1$ as $\varepsilon\to
0^{+}$. Let us assume, without loss of generality, that
$\sigma\in[0,M]$. Since $\varphi_{\ep}^{**}$ is the convexification of $\varphi_{\ep}$,
we can estimate $\varphi_{\ep}^{**}(\sigma)$ from above with the linear function
interpolating the values of $\varphi_{\ep}$ in $0$ and $2\varepsilon^{-1}$. As soon
as $M\geq 2\varepsilon^{-1}$ we obtain that
$$\varphi_{\ep}^{**}(\sigma)\leq\frac \varepsilon
2\varphi_{\ep}\left(\frac 2 \varepsilon\right)\cdot\sigma \leq \frac 1 2
\left(\frac {\log(1+4\varepsilon^{-2})}{2|\log \varepsilon|}+ 1\right)\cdot\sigma
\quad\quad
\forall\sigma\in [0,M].$$
As required, the coefficient of $\sigma$ in the right-hand side tends
to $1$. This completes the proof of (\ref{th:gamma-limsup}).
\paragraph{\emph{\textmd{Compactness}}}
From (\ref{th:phiep>}) with $a=b=1/2$ we obtain that
$$E_{\ep}^{**}(u_{\ep}) \geq \frac 1 2 \int_\Omega
\left(\strut|\nablau_{\ep}(x)| - 1\right)dx = \frac 1 2 \,TV(u_{\ep}) - \frac 1
2|\Omega|.$$
This estimate and assumption (\ref{hp:cpt}) yield a uniform bound on
the $L^{\infty}$-norm and on the total variation of $u_{\ep}$. Thanks to
well known embedding theorems in $BV(\Omega)$ (see~\cite{AFP}, this is
the point where we need $\Omega$ to be an extension domain), this
implies that the family $\{u_{\ep}\}$ is relatively compact in
$L^{p}(\Omega)$ for every $p<+\infty$, and in particular in
$L^{2}(\Omega)$.{\penalty 10000\mbox{$\quad\Box$}}
\subsection{Convergence of approximating solutions}
In the first three paragraphs we prove the result with the further
assumption that
\begin{equation}
\sup_{\varepsilon\in(0,1)}\left\{\strut\|u_{0\ep}\|_{\infty}+
E_{\ep}^{**}(u_{0\ep})\right\}<+\infty.
\label{hp:uzep}
\end{equation}
Then in the last paragraph we prove it for general data.
\paragraph{\emph{\textmd{Compactness on bounded time intervals}}}
We show that, for every $T>0$, the family $\{u_{\ep}\}$ is relatively
compact in $C^{0}\left([0,T];L^{2}(\Omega)\right)$. This
follows from Ascoli's theorem provided that we show that solutions are
$1/2$-H\"{o}lder continuous with equi-bounded H\"{o}lder constants,
and that for every fixed $t\geq 0$ the family $\{u_{\ep}(t)\}\subseteq
L^{2}(\Omega)$ is relatively compact.
From (\ref{hp:uep-gf}) and H\"{o}lder's inequality we have
that
\begin{eqnarray*}
\|u_{\ep}(t)-u_{\ep}(s)\|_{2}\ \leq\
\int_{s}^{t}\|u_{\ep}'(\tau)\|_{2}\,d\tau & \leq & |t-s|^{1/2}\left\{
\int_{0}^{t}\|u_{\ep}'(\tau)\|_{2}^{2}\,d\tau\right\}^{1/2} \\
& \leq & |t-s|^{1/2}\left\{2E_{\ep}^{**}(u_{0\ep})\right\}^{1/2},
\end{eqnarray*}
so that the uniform bound on H\"{o}lder constants follows from
assumption (\ref{hp:uzep}).
Moreover, from (\ref{hp:uep-gf}) and statement~(\ref{stat:Linfty}) of
Theorem~\ref{thm:uep} we have also that the functions
$t\toE_{\ep}^{**}(u_{\ep}(t))$ and $t\to\|u_{\ep}(t)\|_{\infty}$ are nonincreasing,
hence
\begin{equation}
\|u_{\ep}(t)\|_{\infty}+ E_{\ep}^{**}(u_{\ep}(t))\leq
\|u_{0\ep}\|_{\infty}+ E_{\ep}^{**}(u_{0\ep})
\quad\quad
\forall t\geq 0.
\label{est:t-fix}
\end{equation}
Thanks to assumption~(\ref{hp:uzep}), the right-hand side is bounded
independently of $\varepsilon$. Therefore the compactness result in
statement~(2) of Theorem~\ref{thm:Gamma-conv} implies that the family
$\{u_{\ep}(t)\}$ is relatively compact in $L^{2}(\Omega)$ for every fixed
$t\geq 0$.
\paragraph{\emph{\textmd{Characterization of the limit}}}
We prove that, for every $T>0$, any limit point in the interval
$[0,T]$ of the family $\{u_{\ep}\}$ of approximating solutions is the
solution $u(t)$ of the total variation flow in the same interval.
Since the solution of the limit problem is unique, this is enough to
prove the convergence of the whole family.
Let $\varepsilon_{n}\to 0^{+}$ be any sequence such that $u_{\varepsilon_{n}}(t)$
uniformly converges to some $v(t)$ in $[0,T]$. By
(\ref{hp:data-conv}) we have that $v(0)=u_{0}$. So it is enough to
show that $v(t)$ is a maximal slope curve for the functional $TV(u)$
in $[0,T]$.
To this end, we apply Theorem~\ref{thm:main-cmp} to the sequence of
functionals $\{E^{**}_{\varepsilon_{n}}(u)\}$. Indeed they are convex
functionals, and they Gamma-converge to $TV(u)$ because of
Theorem~\ref{thm:Gamma-conv}. Thus Remark~\ref{rmk:convex->scp} and
Proposition~\ref{prop:scp} prove that
assumption~(\ref{hp:slope-energy}) of Theorem~\ref{thm:main-cmp} is
satisfied. Also (\ref{hp:sup-fn}) holds true because of
(\ref{est:t-fix}), and the fact that the functionals are nonnegative.
Since all the assumptions are satisfied, Theorem~\ref{thm:main-cmp}
implies that $v(t)$ is a maximal slope curve for the functional
$TV(u)$.
\paragraph{\emph{\textmd{Uniform convergence for all positive times}}}
It remains to prove that the convergence is global-in-time, as stated
in (\ref{th:main}). This follows from two general facts.
The first one is that $u(t)$ tends, as $t\to +\infty$, to the constant
function $u_{\infty}$ equal to the average of $u_{0}$ in $\Omega$
(see~\cite{TV-property}).
The second fact is that $u_{\ep}(t)-u_{\infty}$ is the solution of
the approximating problem with
initial datum $u_{0\varepsilon}-u_{\infty}$, hence
statement~(\ref{stat:Linfty}) of Theorem~\ref{thm:uep} (with $p=2$)
implies that
$t\to\|u_\varepsilon(t)-u_{\infty}\|_{2}$ is a nonincreasing function.
Therefore for every
$T>0$ we have that
\begin{eqnarray*}
\sup_{t\geq T}\|u_{\ep}(t)-u(t)\|_{2} & \leq &
\sup_{t\geq T}\|u_{\ep}(t)-u_{\infty}\|_{2}+
\sup_{t\geq T}\|u(t)-u_{\infty}\|_{2}\\
& = & \|u_{\ep}(T)-u_{\infty}\|_{2}+
\sup_{t\geq T}\|u(t)-u_{\infty}\|_{2} \\
& \leq & \|u_{\ep}(T)-u(T)\|_{2}+
2\sup_{t\geq T}\|u(t)-u_{\infty}\|_{2},
\end{eqnarray*}
hence
\begin{eqnarray*}
\sup_{t\geq 0}\|u_{\ep}(t)-u(t)\|_{2}& = & \max\Bigl\{
\sup_{t\in[0,T]}\|u_{\ep}(t)-u(t)\|_{2},\; \sup_{t\geq
T}\|u_{\ep}(t)-u(t)\|_{2}\Bigr\} \\
& \leq & \sup_{t\in[0,T]}\|u_{\ep}(t)-u(t)\|_{2}+
2\sup_{t\geq T}\|u(t)-u_{\infty}\|_{2}.
\end{eqnarray*}
Letting $\varepsilon\to 0^{+}$, the first term tends to 0 because of the
convergence result in $[0,T]$. Letting $T\to +\infty$, the second
term tends to 0 because $u(t)\to u_{\infty}$. This proves
(\ref{th:main}) provided that initial data satisfy (\ref{hp:uzep}).
\paragraph{\emph{\textmd{Convergence for general data}}}
Let now $u_{0\ep}\to u_{0}$ be any family satisfying
(\ref{hp:data-conv}). Let us choose a sequence $\{u_{0n}\}\subseteq
L^{\infty}(\Omega)\cap W^{1,\infty}(\Omega)$ with $u_{0n}\to u_{0}$. For
every $n\in\mathbb{N}$, let $u_{\varepsilon,n}(t)$ be the solution of the approximating
problem with $u_{\varepsilon,n}(0)=u_{0n}$, and let $u_{n}(t)$ be the solution of
the limit problem with $u_{n}(0)=u_{0n}$. For every $n\in\mathbb{N}$ we
already know that $u_{\varepsilon,n}\to u_{n}$ in
$C^{0}\left([0,+\infty);L^{2}(\Omega)\right)$, because in this case
the sequence of initial data does not depend on $\varepsilon$ and satisfies
(\ref{hp:uzep}).
Since both the approximating problems and the limit problem generate a
contraction semigroup in $L^{2}(\Omega)$, we have that
\begin{eqnarray*}
\|u_{\ep}(t)-u(t)\|_{2} & \leq &
\|u_{\ep}(t)-u_{\varepsilon,n}(t)\|_{2}+\|u_{\varepsilon,n}(t)-u_{n}(t)\|_{2}+
\|u_{n}(t)-u(t)\|_{2} \\
& \leq & \|u_{0\ep}-u_{0n}\|_{2}+\|u_{\varepsilon,n}(t)-u_{n}(t)\|_{2}+
\|u_{0n}-u_{0}\|_{2}.
\end{eqnarray*}
Taking the supremum over all $t\geq 0$, and letting $\varepsilon\to 0^{+}$,
we obtain that
$$\limsup_{\varepsilon\to 0^{+}}\;\sup_{t\geq 0}\|u_{\ep}(t)-u(t)\|_{2}\leq
2\|u_{0n}-u_{0}\|_{2}.$$
Letting $n\to +\infty$, we finally obtain (\ref{th:main}) for general
data.{\penalty 10000\mbox{$\quad\Box$}}
\subsubsection*{\centering Acknowledgments}
We would like to thank Patrick Guidotti for sending us a preliminary
version of reference~\cite{patrick-DM}.
|
\section{Introduction}
\bigskip In this article we consider self-similar solutions to Smoluchowski's
mean-field model for coagulation \cite{Drake,Smolu} that is given by the
equation
\begin{equation}%
\begin{split}
\partial_{t}f\left( \xi,t\right) & =\frac{1}{2}\int_{0}^{\xi}K\left(
\xi-\eta,\eta\right) f\left( \xi-\eta,t\right) f\left( \eta,t\right) d\eta\\
& \quad-\int_{0}^{\infty}K\left( \xi,\eta\right) f\left( \xi,t\right)
f\left( \eta,t\right) d\eta=:Q\left[ f\right]\,, \label{S1E1}%
\end{split}
\end{equation}
where $f(\xi,t)$ denotes the number density of clusters of size $\xi$ at time
$t$. The kernel $K(\xi,\eta)$ describes the rate of coalescence of clusters of
size $\xi$ and $\eta$ and subsumes all the microscopic aspects of the
coagulation process. It is well-known, cf. for example the review article
\cite{LM1} and references therein, that if the kernel grows at most linearly,
then the initial value problem for (\ref{S1E1}) is well-posed for initial data
with finite mass and the mass is conserved for all times. For homogeneous
kernels it is furthermore expected that solutions converge to a self-similar
form for large times. In fact, for the kernels $K(\xi,\eta) \equiv1$ and
$K(\xi,\eta)=\xi+\eta$, the constant and additive kernel respectively, this
issue has by now been completely solved. It has been established
\cite{Bertoin1,MP1} that besides explicitly known exponentially decaying
self-similar solutions also solutions with algebraic decay exist. In
\cite{MP1} their domains of attraction could also be completely characterized.
However, not much is known about self-similar solutions for other kernels than
these solvable ones. Existence of fast-decaying self-similar solutions has been
established for a large class of kernels \cite{FL1,EMR}, and local properties
of such solutions have been investigated in \cite{CM,EM,FL2,NV10}, but it is
not known whether they are unique in the class of solutions with finite mass.
It is also not clear, not even on the formal level, whether self-similar
solutions with algebraic decay exist.
In this article we will consider mass-conserving self-similar solutions to
coagulation equations with the following product kernel of homogeneity
$2\lambda\in(0,1)$:
\begin{equation}
K\left( \xi,\eta\right) =\left( \xi\eta\right) ^{\lambda}%
\ ,\\ \ \ 0<\lambda<\frac{1}{2}\,. \label{S1E2}%
\end{equation}
Existence of a solution for this kernel
has been established in \cite{FL1}. The purpose of this
paper is to give an asymptotic description of such
a solution
in the regime
$\lambda\to0$.
Let us describe briefly what is known for self-similar solutions for kernels
as in (\ref{S1E2}) and compare it to results for the so-called sum
kernels. To fix ideas, we restrict ourselves to kernels of the form
\begin{equation}
\label{S1E2b}K(\xi,\eta) = \xi^{\alpha}\eta^{\beta}+\xi^{\beta}\eta^{\alpha},
\qquad\alpha+\beta= 2 \lambda, \; 0\leq\alpha\leq\beta,
\end{equation}
even though most of the results that we mention also apply to more general
kernels with the same growth behavior as the ones in (\ref{S1E2b}). We will
need in particular to distinguish between the case $\alpha=0$, the sum
kernel, and the case $\alpha>0$, the product kernel.
It has long been predicted \cite{Le1,DE1}
that self-similar solutions for kernels as in (\ref{S1E2b}) exhibit a singular
power-law behavior of the form $x^{-\tau}$ with $\tau<1+2\lambda$ in the case
$\alpha=0$, and $\tau = 1+2\lambda$ for the case $\alpha>0$. This has been
rigorously proved for the case $\alpha=0$ in \cite{FL2} and for the case
$\alpha>0$ in \cite{EM,NV10}. As has been pointed out in \cite{FL2}, the next
order behavior for small clusters in the case $\alpha=0$ can then easily be
established and is as predicted by the physicists. In the case $\alpha>0$,
however, the next order behavior has not been known \cite{DE1}, and only
numerical simulations suggested that it is oscillatory \cite{FilL1,Lee}.
Our goal in this paper is to describe formally how to construct
mass-\linebreak conserving self-similar solutions for kernels as in
(\ref{S1E2}) in the limit $\lambda\to0$ and by this also describe their
asymptotic behavior on the whole positive real line. While we believe that
such self-similar solutions are unique, our approach only gives the
construction of one such solution. In forthcoming work \cite{NV11} we will
also show how to make this construction rigorous. We will see that in the
limit $\lambda\to0$ the oscillatory character of the solutions becomes very
explicit and can be interpreted in terms of simple ODEs coupled with explicitly
solvable equations.
In the limit $\lambda\rightarrow0^{+}$ the product kernel becomes close
to the sum kernel $x^{2\lambda}+y^{2\lambda}$.
As described above the power law
for the sum kernel is different from the one for the product
kernel. This is due to the fact that, thinking in terms of coagulating
particles, the physical behaviour of these particles is different in these two
cases. Indeed, the power law obtained for the sum case shows that
particles with small $x$ coalesce mostly with particles that are much bigger
than themselves. On the contrary, in the case of the product kernel
small particles interact mostly with the ones having a comparable size. The
point $\lambda=0$ can be thought as a bifurcation point where both kind of
behaviours take place simultaneously.
\section{Preliminaries and Overview}
\subsection{Equation for self-similar solutions}
We are now going to derive the equations that are solved by self-similar
solutions. It is known that the coagulation equation (\ref{S1E1}) can be
written in the following conservative form that makes the
conservation of the number of monomers $\int_{0}^{\infty}\xi f\left( \xi,t\right)
d\xi$ transparent.
\begin{equation}
\partial_{t}\left( \xi f\left( \xi,t\right) \right) +\partial_{\xi}\left(
\int_{0}^{\xi}d\eta\int_{\xi-\eta}^{\infty}d\rho K\left( \eta,\rho\right) \eta f\left(
\eta,t\right) f\left( \rho,t\right) \right) =0\,. \label{S2E1}%
\end{equation}
Using the self-similar variables
\begin{equation}
f\left( \xi,t\right) =t^{-\frac{2}{1-2\lambda}}g\left( x\right)
\ ,\ \ \ x=\frac{\xi}{t^{\frac{1}{1-2\lambda}}}\,, \label{S1E3}%
\end{equation}
as well as the form of the kernel (\ref{S1E2}), equation (\ref{S2E1}) becomes
\[
-\partial_{x}\left( \frac{x^{2}g}{\left( 1{-}2\lambda\right) }\right)
+\partial_{x}\left( \int_{0}^{x}dy\int_{x-y}^{\infty}dzK\left( y,z\right)
yg\left( y\right) g\left( z\right) \right) =0\,.
\]
Integrating this equation, assuming decay of the solutions as $x\rightarrow
\infty$ and imposing absence of particle fluxes, we obtain
\begin{equation}
\frac{x^{2}g\left( x\right) }{\left( 1{-}2\lambda\right) }=\int_{0}%
^{x}dy\int_{x-y}^{\infty}dzK\left( y,z\right) yg\left( y\right) g\left(
z\right)\,. \label{S2E2}%
\end{equation}
The balance of the terms on both sides of (\ref{S2E2}) suggests the following
power law behaviour for $g$ near the origin,
\begin{equation}
g\left( x\right) \sim\frac{H_{\lambda}}{\left( 1{-}2\lambda\right) }\frac
{1}{x^{1+2\lambda}}\ \ \text{as\ \ }x\rightarrow0^{+}\ ,\ \ \ H_{\lambda
}=\frac{\lambda}{B\left( 1{-}\lambda,1{-}\lambda\right) } \label{S2E3}%
\end{equation}
where $B\left( \cdot,\cdot\right) $ is the classical Beta function
\cite{AS}. For further reference we notice that
\begin{equation}
\label{S2E3b}B\left( 1{-}\lambda,1{-}\lambda\right) \sim1+2\lambda+ o(\lambda) \qquad
\mbox{ as } \lambda \to 0.
\end{equation}
In order to remove the power law behaviour we introduce the new function
\begin{equation}
h\left( x\right) =\frac{\left( 1{-}2\lambda\right) }{H_{\lambda}%
}x^{1+2\lambda}g\left( x\right)\,. \label{S2E4}%
\end{equation}
Then $h$ solves
\begin{equation}
h\left( x\right) =H_{\lambda}x^{2\lambda-1}\int_{0}^{x}\,dy\,y^{-2\lambda}h\left(
y\right) \int_{x-y}^{\infty}dzK\left( y,z\right) z^{-\left(
1+2\lambda\right) }h\left( z\right) \,. \label{S2E5}%
\end{equation}
In the rest of the paper we will study solutions of (\ref{S2E5}) in the limit
$\lambda\to0$ that satisfy $h(x) \to0$ as $x \to\infty$. Notice that
(\ref{S2E5}) has the explicit constant solution $h\left( x\right) =1$ that
corresponds to the well-known solution with infinite mass
$g\left( x\right) =\frac{H_{\lambda}}{\left(
1-2\lambda\right) }\frac{1}{x^{1+2\lambda}}$ in the formulation (\ref{S2E2}).
\subsection{Reformulation as a Volterra integro-differential equation}
\label{ShootingProblem}
A main idea in our approach is to reformulate the problem of finding solutions
of (\ref{S2E5}) as a shooting problem for a Volterra integro-differential
equation. Indeed, we can rewrite (\ref{S2E5}) as
\begin{align}
h\left( x\right) & =\frac{H_{\lambda}}{\lambda\left( 1{-}\lambda\right)
}u\left( x\right) v\left( x\right) +H_{\lambda}x^{2\lambda-1}\int_{0}%
^{x}dy\int_{x-y}^{x}dzy^{-\lambda}z^{-\left( \lambda+1\right) }h\left(
y\right) h\left( z\right)\,, \label{S3E4}\\
u\left( x\right) & =\left( 1{-}\lambda\right) x^{\lambda-1}\int_{0}%
^{x} dy y^{-\lambda}h\left( y\right) \ ,\ \ \ \ v\left( x\right) =\lambda
x^{\lambda}\int_{x}^{\infty} dz z^{-\left( \lambda+1\right) }h\left( z\right)
\,. \label{S3E5}%
\end{align}
Differentiating (\ref{S3E5}) we obtain the equations
\begin{equation}
xu_{x}=-\left( 1-\lambda\right) u+\left( 1-\lambda\right)
h\ ,\ \ \ \ xv_{x}=\lambda v-\lambda h\,. \label{S3E7}%
\end{equation}
Thus the solution of (\ref{S2E5}) solves (\ref{S3E7}) with $h$ as in
(\ref{S3E4}), that has the advantage of being a Volterra integro-differential
equation. However, the problem (\ref{S3E4}), (\ref{S3E7}) admits solutions
that in general do not decay as $x\rightarrow\infty$ and therefore they do not
provide solutions of (\ref{S2E5}).
Therefore, among all the solutions of (\ref{S3E4}),
(\ref{S3E7}) we must select those satisfying
\begin{equation}
v\left( x\right) \rightarrow0\ ,\ \ \ \ h\left( x\right) \rightarrow
0\ ,\ \ \ \ u\left( x\right) \rightarrow0\ \ \text{ as \ \ }x\rightarrow
\infty\,.\label{S3E7a}%
\end{equation}
\subsection{Overview of different regimes}
We now sketch the asymptotics of the solution that we are going to construct. It
passes, roughly speaking, through three stages that are described in detail in Sections
\ref{S.zero}, \ref{CombRegion} and \ref{S.asymptotics} respectively.
Section \ref{S.zero} discusses the behavior of the solution near $x =0$. As
pointed out before, near the origin the solution is oscillatory and behaves as
\begin{equation}
\label{oscillations}h\left( x\right) \sim1+Cx^{\beta\left( \lambda\right)
}\cos\left( \alpha\left( \lambda\right) \log\left( x\right)
+\varphi\right) \ \ \text{as\ \ }x\rightarrow0^{+}
\end{equation}
where $\beta(\lambda) \sim\lambda/2$ and $\alpha(\lambda) \sim\sqrt{\lambda}$
as $\lambda\to0$. Our analysis consists of constructing a solution by starting
with $h$ as in (\ref{oscillations}) and using $K=Ce^{i\varphi}$ as a shooting
parameter. However, since equation (\ref{S2E5}) is invariant under the
rescaling $x \mapsto a x$ for $a>0$, we can restrict the range of $K$ to
$[1,\exp{2 \pi(\beta/\alpha)})$ (see also the comment in Section
\ref{Ss.variables}). In this regime near the origin we can approximate
(\ref{S3E4}) and (\ref{S3E5}) by a nonlinear ODE (see (\ref{S5E6})), that is
a
perturbation of a simple ODE system. For the perturbed system we
can use an adiabatic approximation to compute the increase of an associated
energy $E$ along trajectories. As we discuss in Section \ref{Ss.validity} this
first ODE approximation is valid as long as $E \ll1/\lambda$. When $E
\sim1/\lambda$ we enter a new regime that is described in Section
\ref{CombRegion}. In this regime $h$ develops peaks of height $1/\lambda$ and
width of order one that are separated by wide regions in which $h$ is small.
The regime where $h$ is small is again described by an ODE
(cf. Section \ref{Ss.ode}), while the peaks are described by an
integro-differential equation (see Section \ref{Ss.integro}). The analysis of
these respective regimes is done in Sections \ref{Ss.intanalysis} and
\ref{Ss.odeanalysis}, while their coupling is described in
Section \ref{asymptPeaks}. In Section \ref{S.shooting} we will then show by a
continuity argument that there exists a shooting parameter such that the
corresponding solution $h$ converges to zero as $x \to\infty$. In the third
regime, that is discussed in Section \ref{S.asymptotics},
$h$ then decays exponentially fast to zero.
\section{The behaviour as $x \to0$}
\label{S.zero}
\subsection{Oscillations}
It has been observed in \cite{FilL1,Lee} that the self-similar solutions
associated with (\ref{S1E1}) exhibit oscillations for small values of $x.$ This
oscillatory behaviour can be seen as follows. If we make the ansatz
$h(x)=1+\alpha x^{\mu} + \cdots $ with some $\mu\in\mathbb{C}$,
plug this into (\ref{S2E5}) and keep only the leading order terms, we obtain
that $\mu$ must satisfy
\[%
\begin{split}
\frac{1}{H_{\lambda}}& =\int_{0}^{1} dy y^{-2\lambda}\int_{1-y}^{\infty}dzK\left(
y,z\right) z^{-\left( 1+2\lambda\right) +\mu}
\\
&
+\int_{0}^{1} dy y^{-2\lambda+\mu}\int_{1-y}^{\infty}dzK\left(
y,z\right) z^{-\left( 1+2\lambda\right) }\,.%
\end{split}
\]
After some computations, we find that this is equivalent to
\begin{equation}
\Psi_{\lambda}\left( \mu\right) : =\left( \lambda{-}\mu\right) B\left(
1{-}\lambda,1{-}\lambda\right) -\left( 2\lambda{-}\mu\right) B\left(
1{-}\lambda,1{-}\lambda{+}\mu\right) =0 \,. \label{S3E3}%
\end{equation}
One can show that $\Psi_{\lambda}$ has two complex conjugate roots
$\mu_{\lambda}^{\pm}=\pm\alpha\left( \lambda\right) i+\beta\left(
\lambda\right) $ with positive real part and nonzero imaginary part. Indeed, in the limit
$\lambda \to 0$ this can be seen easily as the dominating terms in (\ref{S3E3}) are,
assuming $|\mu| \gg \lambda$,
\[
\Psi_{\lambda}\left( \mu\right) \sim (\lambda {-} \mu) (1{+}2\lambda) - (2\lambda{-}\mu)(1
{-}\mu{+}2\lambda + \mu^2)
\]
and thus the roots of $\Psi_{\lambda}$ satisfy
\begin{equation}
\mu_{\lambda}^{\pm}=\frac{\lambda}{2}\pm\sqrt{\lambda}i+O\left(
\lambda^{\frac{3}{2}}\right) \ \ \text{as\ \ }\lambda\rightarrow0^{+}.
\label{D1E2}%
\end{equation}
Consequently we expect the following asymptotics for the
solutions of (\ref{S2E5}):%
\begin{equation}
h\left( x\right) \sim1+Cx^{\beta\left( \lambda\right) }\cos\left(
\alpha\left( \lambda\right) \log\left( x\right) +\varphi\right)
\ \ \text{as\ \ }x\rightarrow0^{+} \label{S2E6}%
\end{equation}
for suitable real constants $C$ and $\varphi.$
Notice that the asymptotics (\ref{S2E6}) indicate that two consecutive local
maxima of $h$, called $x_{1}$ and $ x_{2}$, satisfy
$
\frac{x_{1}}{x_{2}}\approx\exp\left( \frac{2\pi}{\sqrt{\lambda}}\right)$
for small $\lambda.$ In comparison, to double the amplitude of (\ref{S2E6}) we
must multiply $x$ by a number of order $\exp\left( \frac{2\log\left(
2\right) }{\lambda}\right) $. Thus, the behaviour (\ref{S2E6}) is
essentially oscillatory, with a growth in the amplitude that takes place on a
much larger scale.
\subsection{A new set of variables}
\label{Ss.variables}
For our forthcoming analysis it will be more convenient to use the following variables:
\begin{align}
x & =e^{X} ,\ \ \ y=e^{Y} ,\ \ \ z=e^{Z}\,,\label{S3E8}\\
h\left( x\right) & =H\left( X\right) ,\ \ \ \ u\left( x\right)
=U\left( X\right), \ \ \ \ v\left( x\right) =V\left( X\right)\,.
\label{S3E9}%
\end{align}
Then (\ref{S3E4}), (\ref{S3E7}) becomes
\begin{align}
H & =\frac{H_{\lambda}}{\lambda\left( 1{-}\lambda\right)}
\,U V +I\left[ H\right] \,,
\label{S3E10}\\
\frac{dU}{dX} & =-\left( 1{-}\lambda\right) U +\left(
1{-}\lambda\right) H\,, \qquad
\frac{dV}{dX} =\lambda V -\lambda H\,,
\label{S3E12}%
\end{align}
where
\begin{equation}
I\left[ H\right] \left( X\right) =H_{\lambda}\int_{-\infty}^{X}%
dYe^{\left( 1-\lambda\right) \left( Y-X\right) }H\left( Y\right)
\int_{\log\left( e^{X}-e^{Y}\right) }^{X}dZe^{-\lambda\left( Z-X\right)
}H\left( Z\right)\,. \label{S3E13}%
\end{equation}
The system (\ref{S3E10})-(\ref{S3E13}) has two explicit solutions, namely
$(H,U,V)$ \linebreak
$ \equiv(1,1,1)$ and $(H,U,V) \equiv(0,0,0)$. Our goal is to construct
a solution of (\ref{S3E10})-(\ref{S3E13}) for small values of $\lambda$ with
the property
\begin{align}
\lim_{X\rightarrow-\infty}\left( H\left( X\right) ,U\left( X\right)
,V\left( X\right) \right) & =\left( 1,1,1\right)\,, \label{S4Ea}\\
\lim_{X\rightarrow\infty}\left( H\left( X\right) ,U\left( X\right)
,V\left( X\right) \right) & =\left( 0,0,0\right) \,. \label{S4Eb}%
\end{align}
We will see that a solution that satisfies (\ref{S4Eb}) decays exponentially fast
to zero. This in particular implies that the corresponding self-similar solution
$g$ has finite mass.
Equations
(\ref{S3E8}), (\ref{S3E9}) and (\ref{S2E6}) yield the following asymptotics of
the solutions as $X\rightarrow-\infty$:
\begin{equation}
\left( H\left( X\right) ,U\left( X\right) ,V\left( X\right) \right)
\sim\left( 1,1,1\right) +\operatorname{Re}\left( K\left( 1,\frac{\left(
1{-}\lambda\right) }{\left( 1{-}\lambda\right) +\mu^{+}},\frac{-\lambda
}{\left( \mu^{+}{-}\lambda\right) }\right) e^{\mu^{+}X}\right)\,,
\label{S4E0}%
\end{equation}
where $\mu^{+}$ is as in (\ref{D1E2}) and $K=Ce^{i\varphi}%
\in\mathbb{C}.$
The complex number $K$ determines the solution of (\ref{S3E10}%
)-(\ref{S3E13}) uniquely. However, we notice that (\ref{S3E10})-(\ref{S3E13}) is
invariant under translations $X\rightarrow X+a$
and thus all the complex
numbers in the spiral $\mathcal{S}_{\lambda,K}=\left\{ Ke^{\mu^{+}a}%
:a\in\mathbb{R}\right\} \subset\mathbb{C}$\ yield the same solution up to
translations. Thus we can identify the solutions of
(\ref{S3E10})-(\ref{S3E13}) satisfying (\ref{S4Ea}), up to translations, with
the set of positive real numbers contained between two consecutive
intersections of $\mathcal{S}_{\lambda}$ with the real axis. In particular,
given $K=1$ we have, since $\mu_{\lambda}^{+}=\alpha\left( \lambda\right)
i+\beta\left( \lambda\right)$, that the next consecutive point in
$\mathbb{R}^{+}\cap\mathcal{S}_{\lambda,1}$ is $\exp\left( \frac{2\pi
\beta\left( \lambda\right) }{\alpha\left( \lambda\right) }\right) .$
Therefore, there is a one-to-one correspondence between the points in the
interval $\left[ 1,\exp\left( \frac{2\pi\beta\left( \lambda\right)
}{\alpha\left( \lambda\right) }\right) \right) $ and the solutions of
(\ref{S3E10})-(\ref{S3E13}) satisfying (\ref{S4Ea}).
The main result of this paper is that we show, using
asymptotic arguments, that for small $\lambda$ a value of
$K\in\left[ 1,\exp\left( \frac{2\pi\beta\left(
\lambda\right) }{\alpha\left( \lambda\right) }\right) \right) $ exists such
that (\ref{S4Eb}) holds.
\begin{figure}[ht!]
\centering{%
\includegraphics[width=.46\textwidth]{u}
\includegraphics[width=.46\textwidth]{phase}
}%
\caption{Oscillatory behavior of $U$ (left) and phase plane for $U,V$ (right).}
\label{figure1}
\end{figure}
\subsection{The ODE regime\label{ODELolkaVolterra}}
The key idea in computing the asymptotics of the solutions of (\ref{S3E10}%
)-(\ref{S3E13}) is to obtain suitable approximations of the operator $I\left[
H\right] \left( X\right) $ in (\ref{S3E13}) for small $\lambda$. The
formula for $H_{\lambda}$ (cf. (\ref{S2E3})), combined with (\ref{S2E3b}) and
(\ref{S3E13}), suggests the approximation
\begin{equation}
I\left[ H\right] \left( X\right) =\lambda\int_{-\infty}^{X}e^{\left(
Y-X\right) }H\left( Y\right) dY\int_{\log\left( e^{X}-e^{Y}\right) }%
^{X}H\left( Z\right) dZ\,. \label{E1}%
\end{equation}
We will discuss the consistency of this assumption in Section \ref{S.consistency}.
The operator $I\left[ H\right] \left( X\right) $ can be further approximated
if $Y/X$ changes significantly faster than $H(Y)/H(X)$, as we saw is the case for
$X \to -\infty$. Then it is natural to approximate $H(Y)$ by $H(X)$ and we obtain
\ignore{
If the characteristic length in which
$\left( H\left( Y\right) {-} 1\right) $ has relevant variations compared
with itself for $Y$ of order $X$ is much larger than one, it would be natural
to approximate $H\left( Y\right) $ by $H\left( X\right) .$
}
\[
I\left[ H\right] \left( X\right) \sim\lambda H\left( X\right)
\int_{-\infty}^{X}e^{\left( Y-X\right) }dY\int_{\log\left( e^{X}%
-e^{Y}\right) }^{X}H\left( Z\right) dZ\ \
\]
and changing the order of the integrals
\begin{equation}
I\left[ H\right] \left( X\right) \sim\lambda H\left( X\right)
\int_{-\infty}^{X}H\left( Z\right) e^{Z-X}dZ\,. \label{S5E1}%
\end{equation}
From (\ref{S3E12}) we have
\[
U\left( X\right) =\left( 1{-}\lambda\right) \int_{-\infty}^{X}H\left(
Z\right) e^{\left( 1{-}\lambda\right) \left( Z-X\right) }dZ \sim
\int_{-\infty}^{X}H\left( Z\right) e^{\left( Z-X\right) }dZ
\]
which together with (\ref{S5E1}) leads to
\begin{equation}
I\left[ H\right] \sim\lambda H U
\,. \label{S5E2}%
\end{equation}
Using (\ref{S5E2}) as well as (\ref{S2E3}) and (\ref{S2E3b}) we obtain the
following approximation of (\ref{S3E10}) up to order $\lambda$:
\begin{equation}
H =U V +\lambda U V \left( U -1\right) \label{S5E3a}%
\end{equation}
and plugging this approximation into (\ref{S3E12}), (\ref{S3E13}) we obtain%
\begin{align}
\frac{dU}{dX} & =-\left( 1{-}\lambda\right) U + \left( 1{-}\lambda\right)
U V +\lambda\left( 1{-}\lambda\right) U V \left( U -1\right)\,,
\label{S5E3}\\
\frac{dV}{dX} & =\lambda V -\lambda U V -\lambda^{2}U V \left( U -1\right)\,.
\label{S5E4}%
\end{align}
As (\ref{D1E2}) and (\ref{S4E0})
imply that the changes of $\left( V{-}1\right) $ compared to those of
$\left( U{-}1\right) $ are of order $\sqrt{\lambda}$, we make the
change of variables
\begin{equation}
V=1+\sqrt{\lambda}\omega\ \ ,\ \ \xi=\sqrt{\lambda}X\,, \label{S5E5}%
\end{equation}
that transform (\ref{S5E3}), (\ref{S5E4}) up to order $\sqrt{\lambda}$ into
\begin{equation}
\frac{dU}{d\xi}=U\omega+\sqrt{\lambda}U\left( U-1\right) \ ,\ \ \ \ \frac
{d\omega}{d\xi}=1-U+\sqrt{\lambda}\omega\left( 1-U\right) \,. \label{S5E6}%
\end{equation}
\subsection{Analysis of the ODE (\ref{S5E6})\label{ODE}}
Let us denote by $\left( U_{\lambda},\omega_{\lambda}\right) $ any solution
of (\ref{S5E6}). We first notice that the asymptotics of the solutions of
(\ref{S5E6}) as $\left( U_{\lambda},\omega_{\lambda}\right) \rightarrow
\left( 1,0\right) $ agree with those obtained in (\ref{S4E0}). Indeed, in
the limit $\lambda\rightarrow0,$ we can compute the asymptotics of $\left(
U,\omega\right) $ as $\xi\rightarrow-\infty$ using (\ref{S4E0}), (\ref{D1E2})
and (\ref{S5E5}):
\begin{equation}
\left( (U_{\lambda}{-}1)\left( \xi\right) ,\omega_{\lambda}\left( \xi\right)
\right) \sim\operatorname{Re}\left( K\left(
1,i\right) e^{\left( i+O\left(
\sqrt{\lambda}\right) \right) \xi}\right) \ \ \text{as}\ \ \xi\rightarrow
-\infty\,.\label{S6E2}%
\end{equation}
Linearizing (\ref{S5E6}) near $\left( U,\omega\right) =\left( 1,0\right) $
we obtain the same asymptotics for the functions $\left( U_{\lambda}\left(
\xi\right) ,\omega_{\lambda}\left( \xi\right) \right) $ as $\xi
\rightarrow-\infty.$ This gives the desired matching between the solutions of
the linearization of (\ref{S3E10})-(\ref{S3E13}) around $\left( H,U,V\right)
=\left( 1,1,1\right) $ and the solutions of the approximated problem
(\ref{S5E6}).
The approximation (\ref{S6E2}) is valid as long as $\left\vert \left(
U_{\lambda}-1\left( \xi\right) ,\omega_{\lambda}\left( \xi\right) \right)
\right\vert $ is small. However, a detailed analysis of the nonlinear problem
(\ref{S5E6}) is needed if \linebreak$\left\vert \left( U_{\lambda}-1\left(
\xi\right) ,\omega_{\lambda}\left( \xi\right) \right) \right\vert $
becomes of order one. In order to describe the solutions of (\ref{S5E6}) in
this regime, we notice that this equation is a perturbation of the
equation
\begin{equation}
\frac{dU_{0}}{d\xi}=U_{0}\omega_{0}\ ,\ \ \ \ \frac{d\omega_{0}}{d\xi
}=1-U_{0} \label{S6E3}\,.
\end{equation}
This equation can be explicitly integrated, since the following quantity is
conserved along trajectories:
\begin{equation}
E=-\log\left( U_{0}\right) +\left( U_{0}-1\right) +\frac{\omega_{0}^{2}%
}{2} \,. \label{S6E4}%
\end{equation}
The solutions of (\ref{S6E3}) are periodic, but this behavior is not
compatible with the exponential growth obtained in (\ref{S6E2}). This growth
is due to the increase in $E$ produced by the terms of order $\sqrt{\lambda} $
in (\ref{S5E6}). We now compute this change of energy for values of $\left(
U,\omega\right) $ of order one. Using (\ref{S5E6}) and (\ref{S6E4}) we
obtain
\begin{equation}
\frac{dE}{d\xi}=\sqrt{\lambda}\left[ \left( U_{\lambda}-1\right)
^{2}+\omega_{\lambda}^{2}\left( 1-U_{\lambda}\right) \right] \label{S6E5}\,.
\end{equation}
Since the solutions of (\ref{S5E6}) are close to those of (\ref{S6E3}) during
finite time intervals, we can adiabatically compute the change in $E.$ More
precisely, the solutions of (\ref{S6E3}) satisfying (\ref{S6E4}) are periodic
with a period $T\left( E\right) $ given by
\begin{equation}
T\left( E\right) =\sqrt{2}\int_{U_{-}\left( E,0\right) }^{U_{+}\left(
E,0\right) }\frac{d\eta}{\eta\sqrt{E+\log\left( \eta\right) -\left(
\eta-1\right) }} \label{S6E5a}%
\end{equation}
where the functions $U_{-}\left( E,\omega\right) \leq U_{+}\left(
E,\omega\right) $ are defined as the roots of the equation
\begin{equation}
\log\left( U\right) -\left( U-1\right) =\frac{\omega^{2}}{2}-E
\label{S6E6}%
\end{equation}
for any given value of $\omega$ and $E\geq\frac{\omega^{2}}{2}.$
Integrating (\ref{S6E5}) and using the adiabatic approximation, we can then
approximate the change of $E$ during an interval of length $T\left( E\right)
$ as
\begin{equation}
\frac{E\left( T\left( E\right) \right) -E\left( 0\right) }{\sqrt
{\lambda}}=\int_{0}^{T\left( E\right) }\left[ \left( U_{0}\left(
\xi\right) -1\right) ^{2}+\omega_{0}^{2}\left( \xi\right) \left(
1-U_{0}\left( \xi\right) \right) \right] d\xi=\Phi\left( E\right)
\label{S6E7}%
\end{equation}
where $\left( U_{0}\left( \xi\right) ,\omega_{0}\left( \xi\right)
\right) $ is a solution of (\ref{S6E3}) satisfying (\ref{S6E4}).
It turns out that the function $\Phi\left( E\right) $ is positive for any
$E>0.$ This follows, using
the second
equation in (\ref{S6E3}), via
\[
\Phi\left( E\right) =\int_{-\sqrt{2E}}^{\sqrt{2E}}\left[ \left(
1-U_{-}\left( E,\omega\right) \right) +\omega^{2}\right] d\omega
-\int_{-\sqrt{2E}}^{\sqrt{2E}}\left[ \left( 1-U_{+}\left( E,\omega\right)
\right) +\omega^{2}\right] d\omega\,,
\]
whence
\begin{equation}
\Phi\left( E\right) =2\int_{0}^{\sqrt{2E}}\left( U_{+}\left(
E,\omega\right) -U_{-}\left( E,\omega\right) \right) d\omega>0\,.
\label{S6E8}%
\end{equation}
We can compute the asymptotics of (\ref{S6E8}) as $E\rightarrow\infty$ using
(\ref{S6E6}). Notice that the leading contribution to the integral is given by
$U_{+}\left( E,\omega\right) ,$ since $U_{-}\left( E,\omega\right) \leq1.$
Then
\begin{equation}
\Phi\left( E\right) \sim2\int_{0}^{\sqrt{2E}}\left( E-\frac{\omega^{2}}%
{2}\right) d\omega=\frac{4\sqrt{2}}{3}E^{\frac{3}{2}}\ \ \text{as\ \ }%
E\rightarrow\infty\,.\label{S6E8a}%
\end{equation}
\subsection{Range of validity of the ODE regime}
\label{Ss.validity}
We have obtained that, as long as the approximation
(\ref{S5E6}) is valid, we can approximate the functions $\left( U_{\lambda
}\left( \xi\right) ,\omega_{\lambda}\left( \xi\right) \right) $ by
$\left( U_{0}\left( \xi\right) ,\omega_{0}\left( \xi\right) \right) $
while $E$ is computed via the iterative recursion (\ref{S6E7}).
Due to (\ref{S6E8a}) these values of $E$ increase an amount
of order $\sqrt{\lambda}\left( 1+E\right) ^{\frac{3}{2}}$ in each period of
size $T\left( E\right) .$ We now proceed to discuss the range of validity of
the different approximations that have been used to derive (\ref{S6E7}) and
(\ref{S5E6}).
To examine the range
of values of $E$ for which the adiabatic
approximation is valid, we need to show that the value of $E\left(
\xi\right) $ remains close to $E\left( 0\right) $ for $0\leq\xi\leq
T\left( E\right) .$ Due to (\ref{S6E6}) and (\ref{S6E8}) the
adiabatic approximation is valid as long as
\begin{equation}
\sup_{0\leq\xi\leq T\left( E\right) }\left\vert E\left( \xi\right)
-E\left( 0\right) \right\vert \ll E\left( 0\right)\,. \label{S6E9}%
\end{equation}
Let us remark that (\ref{S6E9}) does not follow immediately from the
inequality $\left\vert E\left( T\left( E\left( 0\right) \right) \right)
-E\left( 0\right) \right\vert \ll E\left( 0\right) $ because the right-hand
side of (\ref{S6E5}) does not have a sign and therefore it is not clear that
\begin{equation}
\sup_{0\leq\xi\leq T\left( E\right) }\left\vert E\left( \xi\right)
-E\left( 0\right) \right\vert \leq C\left\vert E\left( T\left( E\left(
0\right) \right) \right) -E\left( 0\right) \right\vert \label{S6E10}%
\end{equation}
for some $C>0.$ A priori it is not possible to rule out the possibility of big
changes of $E\left( \xi\right) $ along each cycle balanced in such a way
that the overall change of $E$ along the cycle is just $\Phi\left( E\right)
.$ However, as shown in Appendix \ref{A.energy}, it turns out that
(\ref{S6E10}) holds true.
Using (\ref{S6E7}), (\ref{S6E8a}) and (\ref{S6E10}) it follows that the
adiabatic approximation is valid as long as $\sqrt{\lambda}\left( 1+E\right)
^{\frac{3}{2}}\ll \left( 1+E\right) $ and this is satisfied as long as
\begin{equation}
\lambda\left( 1+E\right) \ll 1 \label{S6E11}\,.
\end{equation}
We now study the range of validity of the approximation (\ref{S5E2}) that is
the main ingredient in deriving the approximate problem (\ref{S5E6}). The main
assumption used in the derivation of (\ref{S5E2}) is that the characteristic
length scale for which $\left( H\left( X\right) -1\right) $ has changes
comparable to itself is much larger than one.
Due to (\ref{S5E3a}) it follows that as long as $V$ remains close to $1$ the
characteristic length scale for
$\left( H-1\right) $ is the same as for $\left( U-1\right) .$ Notice that
$\left( V-1\right) $ is small if $\sqrt{\lambda}\omega\ll 1.$
Since
we have $\frac 1 2 \omega^2 \leq E$, the condition $\sqrt{\lambda} \omega \ll 1$
holds if (\ref{S6E11}) is true, and so
the characteristic
length scales for $\left( H-1\right) $ and $\left( U-1\right) $ are the
same under the assumption (\ref{S6E11}). For $\left\vert \left(
U,\omega\right) \right\vert $ of order one, the evolution of (\ref{S6E3})
takes place in the length scale $\xi,$ or equivalently for changes in $X$ of
order $\frac{1}{\sqrt{\lambda}}\gg 1.$ Therefore, the condition in (\ref{S6E11})
that allows us to
obtain (\ref{S5E2}) is immediately satisfied and we can restrict our analysis
to the case $\left\vert \left( U,\omega\right) \right\vert \gg 1. $ The first
equation in (\ref{S6E3}) combined with (\ref{S5E5}) implies that
\begin{equation}
\left\vert \frac{d\left( \log\left( U\right) \right) }{dX}\right\vert
=\sqrt{\lambda}\left\vert \omega\right\vert \label{S6E12}\,.
\end{equation}
The left-hand side measures the relative variations in $U$ with respect to
changes in $X.$ The characteristic length scale that describes the changes in
this quantity is large as long as $\sqrt{\lambda}\left\vert \omega\right\vert
\ll 1$ which is again true if (\ref{S6E11}) holds.
Therefore, if the condition (\ref{S6E11}) is satisfied, we can apply
simultaneously all the approximations that lead to (\ref{S5E6}), the adiabatic
approximation yielding (\ref{S6E7}) as well as the condition $\left\vert
V-1\right\vert \ll 1.$ However, the three assumptions break down simultaneously
if $E$\ becomes of order $\frac{1}{\lambda}$. Since (\ref{S6E7}) and
(\ref{S6E8a}) imply that $E$ increases to arbitrarily large values, the
failure of (\ref{S6E11}) happens for every solution of (\ref{S3E10}%
)-(\ref{S3E13}) satisfying (\ref{S4E0}).
\section{ The intermediate regime}
\label{CombRegion}
In order to describe the solutions of (\ref{S3E10})-(\ref{S3E13}) when
$\lambda E\approx1$ we introduce a new group of variables. It turns out to be
convenient to split the regime into the one where $U$ (or
$H$) is of order one or smaller, and the one where $U$ (or $H$) is large.
Notice that (\ref{S6E6}) implies that for a given value of $E\sim\frac
{1}{\lambda}$ the minimum of $U$ scales as $e^{-1/\lambda}$, and the
maximum scales as $1/\lambda$. We remark that the
forthcoming analysis will show that as long as $V$ remains of order one and
is not small the
values of $H$ and $U$ have the same order of magnitude. This assumption will
be made implicitly in all the remaining computations and will be justified by the
self-consistency of the derived asymptotics.
\subsection{The integro-differential equation regime}
\label{Ss.integro}
We begin by studying the case in which $U$ (and $H$) are large. Since we are
interested in the case $E\approx\frac{1}{\lambda}$, equation (\ref{S6E6}) suggests
the rescaling $U\approx\frac{1}{\lambda},\ \omega\approx\frac{1}{\sqrt
{\lambda}}.$ Then (\ref{S6E12}) yields a characteristic length scale for $X$
of order one. This suggests introducing the new set of variables
\begin{equation}
U\left( X\right) =\frac{1}{\lambda}\mathcal{U}\left( X\right)
, \ \ \ V\left( X\right) =\mathcal{V}\left( X\right) ,\ \ \ H\left(
X\right) =\frac{1}{\lambda}\mathcal{H}\left( X\right) \,.
\ \label{S7E1}%
\end{equation}
Plugging
(\ref{S7E1}) into (\ref{S3E10})-(\ref{S3E13}) we obtain
\begin{align}
\mathcal{H} & =\frac{H_{\lambda}}{\lambda\left(
1{-}\lambda\right) }\mathcal{U} \mathcal{V}
+\frac{1}{\lambda}I\left[ \mathcal{H}\right]\,, \label{S7E4}\\
\frac{d\mathcal{U}}{dX} & =-\left( 1{-}\lambda\right) \mathcal{U}
+\left( 1{-}\lambda\right) \mathcal{H}\,,
\, \qquad \frac{d\mathcal{V}}{dX} =\lambda\mathcal{V}
-\mathcal{H}\,, \label{S7E6}%
\end{align}
where the operator $I\left[ \mathcal{H}\right] $ is as in (\ref{S3E13}).
Using the approximation (\ref{E1}) that will be seen to be still valid in this
region we obtain
\[
\frac{1}{\lambda}I\left[ \mathcal{H}\right] \left( X\right) =\left(
1+O\left( \lambda\right) \right) \int_{-\infty}^{X}e^{\left( Y-X\right)
}\mathcal{H}\left( Y\right) dY\int_{\log\left( e^{X}-e^{Y}\right) }%
^{X}\mathcal{H}\left( Z\right) dZ\,.
\]
Taking the limit $\lambda\rightarrow0$ in (\ref{S7E4})-(\ref{S7E6}) we obtain
\begin{align}
\mathcal{H}\left( X\right) & =\mathcal{U}\left( X\right) \mathcal{V}%
\left( X\right) +\int_{-\infty}^{X}e^{\left( Y-X\right) }\mathcal{H}%
\left( Y\right) dY\int_{\log\left( e^{X}-e^{Y}\right) }^{X}\mathcal{H}%
\left( Z\right)\,, dZ\label{S7E7}\\
\frac{d\mathcal{U}}{dX} & =-\mathcal{U} +\mathcal{H}
\ ,\ \ \ \frac{d\mathcal{V}}{dX}=-\mathcal{H}\,.
\label{S7E9}%
\end{align}
\subsection{The ODE regime}
\label{Ss.ode}
If $\lambda E\approx1$ but $U$ (and $H$) remains bounded we can
approximate (\ref{S3E10})-(\ref{S3E13}) as follows.
Keeping just the leading terms in
(\ref{S3E10}), (\ref{S3E13}) we arrive at the approximation $H=UV$.
Plugging this approximation
into (\ref{S3E12}) we obtain
\begin{equation}
\frac{dU}{dX}=-U +U V
\ \ ,\ \ \ \ \frac{dV}{dX}=\lambda\left( V -U
V \right) \,. \label{S7E10}%
\end{equation}
\subsection{Explicit solution of (\ref{S7E7}), (\ref{S7E9})}
\label{Ss.intanalysis}
We are going to compute the solutions of (\ref{S7E7}), (\ref{S7E9}) explicitly using
Laplace transforms. In order to bring (\ref{S7E7}), (\ref{S7E9}) into the form
of a convolution equation we introduce the set of variables
\begin{equation}
x=e^{X}\ ,\ \ \mathcal{H}\left( X\right) =\bar{h}\left( x\right)
\ ,\ \ \ \mathcal{U}\left( X\right) =\bar{u}\left( x\right)
\ ,\ \ \mathcal{V}\left( X\right) =\bar{v}\left( x\right) \label{S8E0}%
\end{equation}
that transforms (\ref{S7E7}), (\ref{S7E9}) into
\begin{align}
\bar{h}\left( x\right) & =\bar{u}\left( x\right) \bar{v}\left(
x\right) +\frac{1}{x}\int_{0}^{x}\bar{h}\left( y\right) dy\int_{x-y}%
^{x}\frac{\bar{h}\left( z\right) }{z}dz\,,\label{S8E1}\\
x\bar{u}_{x} & =-\bar{u}+\bar{h}\ \ \ ,\ \ \ x\bar{v}_{x}=-\bar
{h}\,.
\label{S8E2}%
\end{align}
Notice that this system of equations can be obtained from (\ref{S3E4}),
(\ref{S3E7}) by means of the rescaling $h=\frac{\bar{h}}{\lambda}%
,\ u=\frac{\bar{u}}{\lambda},\ v=\bar{v}$ and taking the limit $\lambda
\rightarrow0^{+}.$ We can compute explicitly a family of solutions of
(\ref{S8E1}), (\ref{S8E2}) that will be used to describe the solutions of
(\ref{S3E10})-(\ref{S3E13}) in the limit $\lambda\rightarrow0^{+}.$
\begin{theorem}
\label{solInt}
For any given $a \in (0,2)$ and $\kappa>0$ there exists a solution
to (\ref{S8E1}), (\ref{S8E2}) that satifies
\begin{equation}
\bar h(0)=\bar h(\infty)= \bar u(0)= \bar u(\infty)=0\,, \ \
\lim_{x\rightarrow0^{+}}\bar{v}\left( x\right) =1+a , \ \ \lim
_{x\rightarrow\infty}\bar{v}\left( x\right) =1-a\,. \label{F7E7}%
\end{equation}
It is given by
\begin{equation}
\bar{v}\left( x\right) =\frac{1}{2\pi i}\int_{\gamma}\left[ 1+a\left(
\frac{\zeta^{a}-\kappa x^{a}}{\zeta^{a}+\kappa x^{a}}\right) \right]
\frac{e^{\zeta}}{\zeta}d\zeta\label{F7E5}%
\end{equation}
where $\gamma=\left\{ \gamma\left( s\right) :s\in\mathbb{R}\right\} $ is a
contour in the complex plane contained in the domain $\left\{ \arg\left(
\zeta\right) \in\left( -\theta_{0},\theta_{0}\right) \right\} $ satisfying
$\arg\left( \gamma\left( s\right) \right) \rightarrow-\theta_{0}$ as
$s\rightarrow-\infty$ and $\arg\left( \gamma\left( s\right) \right)
\rightarrow\theta_{0}$ as $s\rightarrow\infty$ with $\frac{\pi}{2}<\theta
_{0}<\min\left\{ \frac{\pi}{a},\pi\right\} .$
If $0<a<1$ we have $\bar{v}_{x}\left( x\right) <0,\ \bar{h}\left( x\right)
>0 $ and $\ \bar{u}\left( x\right) >0$ for any $x\in\mathbb{R}^{+}.$
The following asymptotics hold:
\begin{align}
\bar{v}_{x}\left( x\right) & \sim-\frac{2a\kappa}{\Gamma\left( a\right)
}\frac{1}{x^{1-a}}\ \ \ \ \text{as\ \ }x\rightarrow0^{+}\ \ ,\ \ a\in\left(
0,2\right) \label{G1E1}\\
\bar{v}_{x}\left( x\right) & \sim\frac{2a}{\kappa\Gamma\left( -a\right)
}\frac{1}{x^{a+1}}\ \ \ \ \text{as\ \ }x\rightarrow\infty\ \ \ ,\ \ a\in
\left( 0,1\right) \cup\left( 1,2\right) \nonumber
\end{align}%
\begin{align}
\bar{h}\left( x\right) & \sim\frac{2a\kappa}{\Gamma\left( a\right)
}x^{a}\ \ \ \ \text{as\ \ }x\rightarrow0^{+}\ \ \ \ ,\ \ a\in\left(
0,2\right) \label{G1E2}\\
\bar{h}\left( x\right) & \sim-\frac{2a}{\kappa\Gamma\left( -a\right)
}\frac{1}{x^{a}}\ \ \ \ \text{as\ \ }x\rightarrow\infty\ \ \ ,\ \ a\in\left(
0,1\right) \cup\left( 1,2\right) \nonumber
\end{align}%
\begin{align}
\bar{u}\left( x\right) & \sim\frac{2a\kappa}{\Gamma\left( a\right)
\left( a+1\right) }x^{a}\ \ \ \ \text{as\ \ }x\rightarrow0^{+}%
\ \ \ \ ,\ \ a\in\left( 0,2\right) \label{G1E3}\\
\bar{u}\left( x\right) & \sim-\frac{2a}{\kappa\Gamma\left( -a\right)
\left( 1-a\right) }\frac{1}{x^{a}}\ \ \ \ \text{as\ \ }x\rightarrow
\infty\ \ \ ,\ \ a\in\left( 0,1\right) \nonumber
\end{align}
If $a=1:$%
\begin{equation}
\bar{v}\left( x\right) =2e^{-\kappa x}\ \ ,\ \ \bar{h}\left( x\right)
=2\left( \kappa x\right) e^{-\kappa x}\ \ ,\ \ \bar{u}\left( x\right)
=\frac{2}{\kappa x}\left( 1-\left( 1+\kappa x\right) e^{-\kappa x}\right)
\label{G1E4}%
\end{equation}
\end{theorem}
\begin{remark}
The solutions with different values of $\kappa$ are essentially the same up to
rescaling. Notice also that (\ref{F7E7}) implies that $\bar{v}\left(
x\right) $ becomes negative for $x$ sufficiently large if $a>1.$
\end{remark}
\begin{remark}
Some comments about the choice of the range of values of $a$ are in order. We
need the restriction $0<a<2$ to avoid the singular points $\left(
-\kappa\right) ^{\frac{1}{a}}$ crossing the imaginary axis. It would also be
possible to include the value $a=1$ in the second formula of (\ref{G1E1}) and
(\ref{G1E2}) since the right-hand side vanishes. However this
would result in asymptotic formulas like $f\sim0 $ that do not have a precise
meaning. Finally the constraint $a<1$ in the second formula of (\ref{G1E3}) is
strictly needed, since the asymptotics of $\bar{u}\left( x\right) $ for
$a>1$ is proportional to $\frac{1}{x}.$
\end{remark}
\begin{remark}
It is interesting to note that the solutions described in Theorem
\ref{solInt} for $0<a<1 $ are just the self-similar solutions with
algebraic decay that have been obtained for the coagulation equation with
constant kernel in \cite{MP1}. This correspondence can be seen as
follows. The solutions we obtained in Theorem \ref{solInt} satisfy
\[
\bar{v}\left( x\right) =\left( 1-a\right) +\int_{x}^{\infty}\bar{g}\left(
z\right) dz\ \ ,\ \ \bar{u}\left( x\right) =\frac{1}{x}\int_{0}^{x}y\bar
{g}\left( y\right) dy
\]
with $\bar{g}\left( x\right) =\frac{\bar{h}\left( x\right) }{x}.$ Using
this formula to eliminate $\bar{u},\ \bar{v}$ from (\ref{S8E1}) we obtain
\begin{equation}
x^{2}\bar{g}\left( x\right) =\left( 1-a\right) \int_{0}^{x}y\bar{g}\left(
y\right) dy+\int_{0}^{x}y\bar{g}\left( y\right) dy\int_{x-y}^{\infty}%
\bar{g}\left( z\right) dz \,.\label{A1}%
\end{equation}
Differentiating (\ref{A1}) we obtain the equation that is satisfied by the
self-similar solutions of
\begin{equation}
\label{A2}\partial_{\tau}F=Q\left[ F\right] -\frac{\left( 1-a\right)
}{\tau}F
\end{equation}
having the form $f\left( \xi,\tau\right) =\frac{1}{\tau^{2}}\bar{g}\left(
\frac{\xi}{\tau}\right) $ with $Q\left[ \cdot\right] $ as in (\ref{S1E1}%
)$.$ Equation (\ref{A2}) can be transformed into (\ref{S1E1}) using the change
of variables
$F\left( \xi,\tau\right) =\left( \tau\right) ^{-\left(
1-a\right) }f\left( \xi,t\right) $ and $t=\frac{\tau^{a}}{a}$ and
correspondingly the solutions obtained in Theorem \ref{solInt} are transformed
into those obtained in \cite{MP1} if $0<a\leq1.$
\end{remark}
\begin{proof}
Integrating the first equation in (\ref{S8E2}) we arrive at
\begin{equation}
\bar{u}\left( x\right) =\frac{1}{x}\int_{0}^{x}\bar{h}\left( z\right) dz\,.
\label{S8E2a}%
\end{equation}
Using the second equation in (\ref{S8E2}) we obtain
\begin{align}
\int_{x-y}^{x}\frac{\bar{h}\left( z\right) }{z}dz & =-\bar{v}\left(
x\right) +\bar{v}\left( x-y\right)\,, \label{S8E3}\\
\int_{0}^{x}\bar{h}\left( z\right) dz & =-\int_{0}^{x}z\bar{v}_{z}\left(
z\right) =-x\bar{v}\left( x\right) +\int_{0}^{x}\bar{v}\left( z\right) dz\,,
\label{S8E4}%
\end{align}
assuming that all the integrals involved are convergent. Therefore, using
(\ref{S8E2a}) and (\ref{S8E4})
we find
\begin{equation}
\bar{u}\left( x\right) =-\bar{v}\left( x\right) +\frac{1}{x}\int_{0}%
^{x}\bar{v}\left( z\right) dz \,.\label{S8E5}%
\end{equation}
Eliminating $\bar{h}$ and $\bar{u}$ from (\ref{S8E1}), the second equation in
(\ref{S8E2}), as well as (\ref{S8E2a}), (\ref{S8E3}), (\ref{S8E4}),
implies
\[%
\begin{split}
-x\bar{v}_{x}\left( x\right) & =-\left( \bar{v}\left( x\right) \right)
^{2}+\frac{\bar{v}\left( x\right) }{x}\int_{0}^{x}\bar{v}\left( z\right)
dz\\
& \quad-\frac{1}{x}\int_{0}^{x}y\bar{v}_{y}\left( y\right) \bar{v}\left(
x-y\right) dy+\frac{\bar{v}\left( x\right) }{x}\int_{0}^{x}y\bar{v}%
_{y}\left( y\right) dy.
\end{split}
\]
Integrating by parts in the last equation we obtain
\begin{equation}
x^{2}\bar{v}_{x}\left( x\right) =\int_{0}^{x}y\bar{v}_{y}\left( y\right)
\bar{v}\left( x-y\right) dy \,.\label{S8E6}%
\end{equation}
Equation (\ref{S8E6}) is a convolution equation. In order to solve it we
introduce the Laplace transform of $\bar{v}\left( x\right) $
\begin{equation}
w\left( \zeta\right) =\int_{0}^{\infty}\bar{v}\left( x\right) e^{-\zeta
x}dx\,. \label{S8E7}%
\end{equation}
Then (\ref{S8E6}) becomes
$\frac{d^{2}}{d\zeta^{2}}\big( \zeta w \big) +w
\frac{d}{d\zeta}\big( \zeta w \big) =0$.
This equation can be integrated explicitly. Its only solution that does not
have singularities and takes real values along the line $\zeta\in
\mathbb{R}^{+}$ is
\begin{equation}
w\left( \zeta\right) =\frac{1}{\zeta}\left[ 1+a\left( \frac{\zeta
^{a}-\kappa}{\zeta^{a}+\kappa}\right) \right] \,. \label{S8E8}%
\end{equation}
Inverting the Laplace transform we obtain%
\begin{equation}
\bar{v}\left( x\right) =\frac{1}{2\pi i}\int_{\gamma}\left[ 1+a\left(
\frac{\zeta^{a}-\kappa}{\zeta^{a}+\kappa}\right) \right] \frac{e^{\zeta x}%
}{\zeta}d\zeta\,. \label{S8E9}%
\end{equation}
The change of variables $\zeta x=\hat{\zeta}$ transforms this formula into
(\ref{F7E5}). The choice of the contour of integration ensures the exponential
convergence of the integral as well as the absence of singularities of the
integrand along the contour of integration. Taking the limits $x\rightarrow
0^{+}$ and $x\rightarrow\infty$ we obtain (\ref{F7E7}).
The negativity of $\bar{v}_{x}$ for $0<a<1$ can be proved by differentiating
(\ref{S8E9}) and deforming the contour of integration to a double half-line
following $\mathbb{R}^{-}$ in a positive and negative direction with
$\arg\left( \zeta\right) =-\pi$ and $\arg\left( \zeta\right) =\pi$
respectively. Then
\[
\bar{v}_{x}\left( x\right) =-\frac{2a\kappa\sin\left( a\pi\right) }{\pi
}\int_{0}^{\infty}\frac{e^{-rx}r^{a}dr}{\left\vert \kappa+e^{a\pi i}%
r^{a}\right\vert ^{2}}\ \ ,\ \ 0<a<1\,.
\]
The asymptotics in (\ref{G1E1}) can be obtained rewriting (\ref{S8E9}) as
\[
\bar{v}\left( x\right) =\left( 1+a\right) -\frac{\kappa a}{\pi i}%
\int_{\gamma}\left( \frac{1}{\zeta^{a}+\kappa}\right) \frac{e^{\zeta x}%
}{\zeta}d\zeta
\]
and%
\[
\bar{v}\left( x\right) =\left( 1-a\right) +\frac{a}{\pi i}\int_{\gamma
}\left( \frac{\zeta^{a}}{\zeta^{a}+\kappa}\right) \frac{e^{\zeta x}}{\zeta
}d\zeta\,.
\]
Differentiating these formulas we obtain
\[
\bar{v}_{x}\left( x\right) =-\frac{\kappa a}{\pi i}\int_{\gamma}%
\frac{e^{\zeta x}}{\zeta^{a}+\kappa}d\zeta\ \ ,\ \ \bar{v}_{x}\left(
x\right) =\frac{a}{\pi i}\int_{\gamma}\left( \frac{\zeta^{a}}{\zeta
^{a}+\kappa}\right) e^{\zeta x}d\zeta
\]
and using the change of variables $\zeta x=\hat{\zeta}$ in both integrals it
follows that
\[
\bar{v}_{x}\left( x\right) =-\frac{\kappa a}{\pi i}\frac{1}{x^{1-a}}%
\int_{\gamma}\frac{e^{\zeta}}{\zeta^{a}+\kappa x^{a}}d\zeta\ \ ,\ \ \bar
{v}_{x}\left( x\right) =\frac{a}{\pi i}\frac{1}{x}\int_{\gamma}\left(
\frac{\zeta^{a}}{\zeta^{a}+\kappa x^{a}}\right) e^{\zeta}d\zeta\,.
\]
Taking the limit $x\rightarrow0^{+}$ in the first formula and $x\rightarrow
\infty$ in the second and using the fact that
\[
\frac{1}{\pi i}\int_{\gamma}\frac{e^{\zeta}}{\zeta^{a}}d\zeta=\frac{2}%
{\Gamma\left( a\right) }\ \ \ \ \ ,\ \ \ \ \frac{1}{\pi i}\int_{\gamma}%
\zeta^{a}e^{\zeta}d\zeta=\frac{2}{\Gamma\left( -a\right) }
\]
we obtain (\ref{G1E1}). Using then (\ref{S8E2}) we obtain (\ref{G1E2}),
(\ref{G1E3}).
Formula (\ref{G1E4}) follows by integrating by residues.
\end{proof}
We can reformulate the results in Theorem \ref{solInt} in terms of the
original functions $\left( \mathcal{H}\left( X\right) ,\mathcal{U}\left(
X\right) ,\mathcal{V}\left( X\right) \right) .$ In the following we choose $\kappa=1$
as a suitable normalization.
\begin{theorem}
\label{ThIntAs}For any $0<a<2$ there exists a solution of (\ref{S7E7}),
(\ref{S7E9}) satisfying $\mathcal{H}(-\infty)=\mathcal{H}(\infty)=
\mathcal{U}(-\infty)=\mathcal{U}(\infty)=0$. It is given by
\begin{equation}
\mathcal{V}_{a}\left( X\right) -1=\frac{a}{2\pi i}\int_{\gamma}\left(
\frac{\zeta^{a}-e^{aX}}{\zeta^{a}+e^{aX}}\right) \frac{e^{\zeta}}{\zeta
}d\zeta\label{T1E1}%
\end{equation}%
\begin{equation}
\mathcal{H}_{a}\left( X\right) =-\frac{d\mathcal{V}_{a}\left( X\right)
}{dX}\ \ ,\ \ \mathcal{U}_{a}\left( X\right) =-\int_{-\infty}^{X}e^{\left(
Z-X\right) }\frac{d\mathcal{V}_{a}\left( Z\right) }{dZ}dZ \label{T1E2}%
\end{equation}%
\begin{equation}
\lim_{X\rightarrow-\infty}\mathcal{V}_{a}\left( X\right)
=1+a\ \ \ ,\ \ \ \lim_{X\rightarrow\infty}\mathcal{V}_{a}\left( X\right)
=1-a\,. \label{T1E3}%
\end{equation}
If $0<a<1$ we have $\frac{d\mathcal{V}_{a}\left( X\right) }{dX}%
<0,\ \mathcal{H}_{a}\left( X\right) >0,\ \mathcal{U}_{a}\left( X\right)
>0$ for any $X\in\mathbb{R}$. We have the asymptotics
\begin{align}
\frac{d\mathcal{V}_{a}\left( X\right) }{dX} & \sim-\frac{2a}{\Gamma\left(
a\right) }e^{aX}\ \ \text{as\ \ }X\rightarrow-\infty,\ \ a\in\left(
0,2\right) \label{T1E4}\\
\frac{d\mathcal{V}_{a}\left( X\right) }{dX} & \sim\frac{2a}{\Gamma\left(
-a\right) }e^{-aX}\ \ \text{as\ \ }X\rightarrow\infty\ \ \ ,\ \ a\in\left(
0,1\right) \cup\left( 1,2\right) \nonumber
\end{align}%
\begin{align}
\mathcal{H}_{a}\left( X\right) & \sim\frac{2a}{\Gamma\left( a\right)
}e^{aX}\ \ \ \ \text{as\ \ }X\rightarrow-\infty\ \ \ \ ,\ \ a\in\left(
0,2\right) \label{T1E5}\\
\mathcal{H}_{a}\left( X\right) & \sim-\frac{2a}{\Gamma\left( -a\right)
}e^{-aX}\ \ \ \ \text{as\ \ }X\rightarrow\infty\ \ \ ,\ \ a\in\left(
0,1\right) \cup\left( 1,2\right) \nonumber
\end{align}%
\begin{align}
\mathcal{U}_{a}\left( X\right) & \sim\frac{2a}{\Gamma\left( a\right)
\left( a+1\right) }e^{aX}\ \ \ \ \text{as\ \ }X\rightarrow-\infty
\ \ \ \ ,\ \ a\in\left( 0,2\right) \label{T1E6}\\
\mathcal{U}_{a}\left( X\right) & \sim-\frac{2a}{\Gamma\left( -a\right)
\left( 1-a\right) }e^{-aX}\ \ \ \ \text{as\ \ }X\rightarrow\infty
\ \ \ ,\ \ a\in\left( 0,1\right) \nonumber
\end{align}
If $a=1:$%
\begin{equation}
\mathcal{V}_{a}\left( X\right) =2e^{-e^{X}} ,\, \mathcal{H}_{a}\left(
X\right) =2 e^{X} e^{-e^{X}} ,\, \mathcal{U}_{a}\left( X\right)
=2e^{-X}\left( 1-\left( 1+\kappa e^{X}\right) e^{-e^{X}}\right)
\label{T1E7}%
\end{equation}
\end{theorem}
\subsection{Solution of (\ref{S7E10})}
\label{Ss.odeanalysis}
Equation (\ref{S7E10}) that approximates the behaviour of the solutions of
(\ref{S3E10})-(\ref{S3E13}) if $\lambda E\approx1$ and $U$ is of order one,
can be solved explicitly since
\begin{equation}
\hat{E}=\lambda\left( \log\left( U\right) -U\right) -V+\log\left(
V\right) \label{S9E1}%
\end{equation}
is preserved along trajectories.
The solutions of (\ref{S7E10}) will be used to match the solutions of
(\ref{S7E7}), (\ref{S7E9}) obtained in the previous Section for $1\ll U\ll \frac
{1}{\lambda}.$ For this range of values $V$ is almost constant if
$\lambda\rightarrow0.$ Moreover, since we are away from the ODE regime
described in Section \ref{ODE} we can assume that $\lambda E$ in (\ref{S6E6})
is of order one and therefore $\hat{E}$ is of order one,
whence also $\left\vert V-1\right\vert $ is of order one.
Suppose that we consider a matching region (to be made precise later) where
$\hat{E}$ is of order one, $V$ is almost constant and satisfies $V<1$.
Then
the first equation in (\ref{S7E10}) implies that $U$ decreases
exponentially. More preciesely, let us assume that $U=1$ for $X=X^{-}.$ Then, the asymptotics
of $U,\ V$ are
\begin{equation}
V\sim\left( 1-a^{-}\right) \ \ ,\ \ \ \ U\sim\exp\left( -a_{-}\left(
X-X^{-}\right) \right) \label{S9E2}%
\end{equation}
for some $a^{-}\in\left( 0,1\right) $.
The asymptotics (\ref{S9E2}) will be shown to match with suitable solutions
among the ones described in Theorem \ref{solInt}. Moreover, notice that
(\ref{S7E10}) implies that these asymptotics are valid as long as $U$ remains
larger than some small number $\varepsilon_{0}$ that we can assume to be of
order one, although small. The exponential decay of $U$ in (\ref{S9E2})
implies that the transition between $U=1$ and $U=\varepsilon_{0}$ takes place
on a length scale $\left( X-X^{-}\right) $ of order one.
We now describe the solution when $U$ is less than $\varepsilon_0$. For this
purpose
let us define by $a_{-}$ and $a_{+}$ the two roots of
the equation $\hat{E}=-V+\log\left(
V\right) $. Then,
in the limit $\lambda\rightarrow0^{+}$, the curves given by
(\ref{S9E1}) in the plane $\left( U,V\right) $
are approximately
the two horizontal lines $\left\{ V=1-a_{-}\right\} $ and $\left\{
V=1+a_{+}\right\}$
connected by a curve that is approximately vertical. Along this connecting curve
$V$ changes an amount of order one and
$\lambda\log\left( U\right) $ continues to be of order one since $\hat E$ is
preserved.
When $U$ is less than $\varepsilon_{0}$ we can approximate the second
equation in (\ref{S7E10}) as $\frac{dV}{dX}=\lambda\left( 1+O\left(
\varepsilon_{0}\right) \right) V.$ Then $V\left( X\right) $ increases
exponentially. The first equation in (\ref{S7E10})
indicates that for the
solutions of (\ref{S7E10}) under consideration, $U$ remains small while $V$
changes from $\left( 1-a_{-}\right) $ to$\ \left( 1+a_{+}\right) .$ Then
we can use the approximation
\begin{equation}
V=\left( 1-a_{-}\right) \exp\left( \lambda\left( 1+O\left( \varepsilon
_{0}\right) \right) \left( X-X^{-}-\ell_{trans}^{-}\right) \right)
\label{S9E3}%
\end{equation}
where $\ell_{trans}^{-}$ is the range of $X$ that it takes for $U$ to decrease
from $U=1$ to $U=\varepsilon_{0}.$ We recall that $\ell_{trans}^{-}$ is of
order one.
Notice that the structure of the curve (\ref{S9E1}) implies that eventually
$U$ becomes again of order $\varepsilon_{0}$ with $V$ close to $\left(
1+a_{+}\right) .$ It then follows from (\ref{S7E10}) that for small
$\lambda,$ $U$ reaches again the value $1$ at $X^{+}$ after
an additional length
$\ell_{trans}^{+}$ of order one. Notice that, using (\ref{S9E3}), we have
\begin{equation}
V=\left( 1-a_{-}\right) \exp\left( \lambda\left( 1+O\left( \varepsilon
_{0}\right) \right) \left( X^{+}-X^{-}-\ell_{trans}^{+}-\ell_{trans}%
^{-}\right) \right) \label{S9E4}%
\end{equation}
and $U$ has the behaviour, for $\varepsilon_{0}\leq U\ll
\frac{1}{\lambda}$,
\begin{equation}
U\sim\exp\left( a_{+}\left( X-X^{+}\right) \right)\,. \label{S9E5}%
\end{equation}
We need to estimate the relation between $a_{-}$ and $a_{+}.$ From
(\ref{S9E1}) we deduce that to leading order
\begin{equation}
\log\left( 1-a_{-}\right) +a_{-}=\log\left( 1+a_{+}\right) -a_{+}\,.
\label{S9E6}%
\end{equation}
On the other hand (\ref{S9E4}) allows us to obtain an approximation for the
transition length $\left( X^{+}-X^{-}\right) .$ Indeed, we have
\[
\left( 1+a_{+}\right) =\left( 1-a_{-}\right) \exp\left( \lambda\left(
1+O\left( \varepsilon_{0}\right) \right) \left( X^{+}-X^{-}-\ell
_{trans}^{+}-\ell_{trans}^{-}\right) \right)
\]
whence, since $\ell_{trans}^{+},\ \ell_{trans}^{-}$ are of order one and
$\varepsilon_{0}$ can be made arbitrarily small
\begin{equation}
\left( X^{+}-X^{-}\right) \sim\frac{1}{\lambda}\log\left( \frac{1+a_{+}%
}{1-a_{-}}\right) \label{S9E7}%
\end{equation}
as $\lambda\rightarrow0^{+}.$
We have described the transition of the solution $\left( U,V,H\right) $ of
(\ref{S3E10})-(\ref{S3E13}) for small $\lambda$ and $U\ll \frac{1}{\lambda}.$
The following result will play a crucial role in describing the solutions,
because it will show that the amplitude of the oscillations increases with
$X.$
\begin{lemma}
\label{L1}Suppose that $a_{-}>0,$ $a_{+}>0$ satisfy (\ref{S9E6}). Then
$a_{+}>a_{-}$.
\end{lemma}
\begin{proof}
This follows from the inequality $\log\left( 1+x\right) -x>\log\left(
1-x\right) +x$ that is just a consequence of
$\log\left( 1+x\right) -\log\left( 1-x\right) =\int_{0}^{x}\frac
{2dr}{\left( 1-r^{2}\right) }>2x$.
\end{proof}
\subsection{Coupling of the regimes}
\label{asymptPeaks}
\subsubsection{Description of the iterative procedure\label{iteration}}
We now describe the behaviour of the solutions of (\ref{S3E10})-(\ref{S3E13})
for $\lambda\rightarrow0$ and $\lambda E\approx1,$ with $E$ as in
(\ref{S6E6}). It turns out that these solutions can be described for this
range of values by a sequence of intervals where the solutions can be
approximated alternately by solutions of (\ref{S7E7}), (\ref{S7E9})
or by solutions of (\ref{S7E10}) with both types of regions connected with
suitable matching regions.
More precisely, suppose that $X_{n}^{+}$ is a point where
$U\left( X_{n}^{+}\right) =1\ ,\ \ V\left( X_{n}^{+}\right) =1+a_{n}$.
Notice that at such a point we can expect to be able to use the approximation
(\ref{S7E10}). This will be seen matching the ODE-Integrodifferential equation
regime with the ODE regime described in Section
\ref{ODELolkaVolterra}. We will assume for the moment that this approximation
is valid. Then $H$ is also of order one and we can approximate it to the
leading order as $H=UV$ (cf. Section \ref{Ss.ode}). We can also use the
approximation (\ref{S9E5}) with $a_{+}=a_{n}$ and approximate $V$ by a
constant. Then we obtain for $\left\vert X-X_{n}^{+}\right\vert $ of order one that
\begin{equation}
U\sim\exp\left( a_{n}\left( X-X_{n}^{+}\right) \right) \ \ ,\ \ V\sim
1+a_{n}\ \ ,\ \ H\sim\left( 1+a_{n}\right) \exp\left( a_{n}\left(
X-X_{n}^{+}\right) \right)\,. \label{T2E1}%
\end{equation}
We can now match the asymptotics (\ref{T2E1}) with those obtained for the
solutions of (\ref{S7E7}), (\ref{S7E9}). We use (\ref{S7E1}) to obtain the
matching condition
\begin{equation}%
\begin{split}
\mathcal{U}\left( X\right) \sim\lambda\exp\left( a_{n}\left( X-X_{n}%
^{+}\right) \right) \ & ,\ \ \mathcal{V}\left( X\right) \sim\left(
1+a_{n}\right)\,, \\
\mathcal{H}\left( X\right) & \sim\lambda\left( 1+a_{n}\right) \exp\left(
a_{n}\left( X-X_{n}^{+}\right) \right)\,.
\end{split}
\label{T2E2}%
\end{equation}
This behaviour must be matched with the one for the solutions of (\ref{S7E7}),
(\ref{S7E9}) obtained in Theorem \ref{ThIntAs}. More precisely we will match
these behaviours with the ones of
\[
\Phi_{n}\left( X\right) =\left( \mathcal{U}_{a_n}\left( X-X_{n}^{0}\right)
,\mathcal{V}_{a_n}\left( X-X_{n}^{0}\right) ,\mathcal{H}_{a_n}\left(
X-X_{n}^{0}\right) \right)
\]
for a suitable choice of
$a_n$ and $ X_{n}^{0}.$ The asymptotics of $\mathcal{V}\left( X\right) $ combined
with the one in (\ref{T1E3}) yield $a=a_{n},$ since in the matching region we
expect to have $\left( X-X_{n}^{0}\right) \rightarrow-\infty.$ On the other
hand (\ref{T1E5}), (\ref{T1E6}) give the choice
\begin{equation}
X_{n}^{0}=X_{n}^{+}+\frac{1}{a_n}\log\left( \frac{2a_{n}}{\left( 1+a_{n}\right)
\Gamma\left( a_{n}\right) }\frac{1}{\lambda}\right)\,. \label{T2E3}%
\end{equation}
With these choices of $a_{n}$ and $ X_{n}^{0}$ we obtain, using (\ref{T1E3}),
(\ref{T1E5}), (\ref{T1E6}) and (\ref{T2E2}), a matching to the first order
between (\ref{T2E1}) and $\Phi_{n}\left( X\right) $ in the common region of
validity where
$\left( X-X_{n}^{+}\right) \gg 1$ and $ \left( X-X_{n}^{0}\right) \ll 1$.
We can then use the approximation $\Phi_{n}\left( X\right) $ to describe the
solutions of (\ref{S3E10})-(\ref{S3E13}) for small $\lambda$ and $\left(
X-X_{n}^{0}\right) $ of order one. This approximation breaks down for
$\left( X-X_{n}^{0}\right) $ sufficiently large. Indeed, (\ref{T1E5}) and
(\ref{T1E6}) imply that $\mathcal{U}_{a}\left( X-X_{n}^{0}\right) $ and
$\mathcal{H}_{a}\left( X-X_{n}^{0}\right) $ converge exponentially to zero
as $\left( X-X_{n}^{0}\right) \rightarrow\infty.$ However, the approximation
(\ref{S7E7}), (\ref{S7E9}) is only valid if $\left\vert \left( \mathcal{U}%
\left( X\right) ,\mathcal{H}\left( X\right) \right) \right\vert
\gg \lambda.$ In the region where $\left\vert \left( \mathcal{U}\left(
X\right) ,\mathcal{H}\left( X\right) \right) \right\vert $ becomes of
order $\lambda$ we must use again the approximation (\ref{S7E10}). In
particular this region can be described using the analysis in Section
\ref{Ss.odeanalysis}. The asymptotics of the solutions for $U$ of order one is
as in (\ref{S9E2}) for suitable choices of $a_{-},\ X^{-}.$ More precisely,
since this matching is made for $\left( X-X_{n}^{0}\right) \gg 1$ we can use
the asymptotics (\ref{T1E3}), (\ref{T1E5}), (\ref{T1E6}) for $X\rightarrow
\infty.$ We then obtain, using also the rescaling (\ref{S7E1}), the matching
conditions
\[%
\begin{split}
U\left( X\right) \sim-\frac{2a_{n}}{\Gamma\left( -a_{n}\right) \left(
1-a_{n}\right) \lambda}e^{-a_{n}\left( X-X_{n}^{0}\right) }\ , &
\ \ \ V\left( X\right) \sim\left( 1-a_{n}\right) \\
H\left( X\right) \sim-\frac{2a_{n}}{\Gamma\left( -a_{n}\right)
\lambda}e^{-a_{n}\left( X-X_{n}^{0}\right) } & \qquad \mbox{ for }
\left( X-X_{n}^{0}\right) \gg 1,\ \ U\gg 1\,.%
\end{split}
\]
We now match these formulas with (\ref{S9E2}) (combined with the approximation
$H=UV$). To this end we must choose
\begin{equation}
a_{-}=a_{n}\ ,\ \ \ X_{n}^{-}=X_{n}^{0}+\frac{1}{a_n}\log\left( -\frac{2a_{n}%
}{\left( 1-a_{n}\right) \Gamma\left( -a_{n}\right) }\frac{1}{\lambda
}\right) \,. \label{T2E4}%
\end{equation}
We remark that all this analysis will be meaningful only if $0<a_{n}<1.$
Therefore $\Gamma\left( -a_{n}\right) <0.$
The region where $U\leq1,\ H\leq1$ can then be described using the analysis in
Section \ref{Ss.odeanalysis}. The conclusion of this analysis is that $\left(
U,V\right) $ moves close to the point $\left( 1,\left( 1+a_{n+1}\right)
\right) $ for some suitable $a_{n+1}$ in a characteristic length given by
(\ref{S9E7}). More precisely if we define (cf. (\ref{S9E6}))
\begin{equation}
\log\left( 1-a_{n}\right) +a_{n}=\log\left( 1+a_{n+1}\right)
-a_{n+1}\ \ ,\ \ a_{n}>0\ \ ,\ \ a_{n+1}>0 \label{T2E5}%
\end{equation}
\begin{equation}
X_{n+1}^{+}=X_{n}^{-}+\frac{1}{\lambda}\log\left( \frac{1+a_{n+1}}{1-a_{n}%
}\right) \label{T2E6}%
\end{equation}
we can approximate $U,\ V$ by means of (\ref{T2E1}) with $X_{n}^{+}$ replaced
by $X_{n+1}^{+}.$
Combining (\ref{T2E3}), (\ref{T2E4}) and (\ref{T2E6}) we obtain to the leading
order%
\begin{equation}
X_{n+1}^{+}-X_{n}^{+}\sim\frac{1}{\lambda}\log\left( \frac{1+a_{n+1}}%
{1-a_{n}}\right)\,. \label{T2E6a}%
\end{equation}
\begin{figure}[ht!]
\centering{%
\includegraphics[width=.46\textwidth]{u2}
\includegraphics[width=.46\textwidth]{phase2}
}%
\caption{Intermediate regime: $U$ (left) and phase plane for $U,V$ (right)}
\label{figure2}
\end{figure}
\subsubsection{Matching with the ODE regime}
We first remark that equation (\ref{S5E6}) agrees
with the approximating equations (\ref{S7E7}), (\ref{S7E9}) and (\ref{S7E10})
if $E\gg 1,\ E\ll \frac{1}{\lambda}$ in their respective regimes of validity.
Indeed, we can rewrite (\ref{S5E6}) using (\ref{S5E5}) as
\begin{equation}
\frac{dU}{dX}=U\left( V-1\right) +\lambda U\left( U-1\right)
\ \ \ ,\ \ \ \frac{dV}{dX}=\lambda\left( 1-U\right) V \label{T2E7}%
\end{equation}
and these equations yield the same behaviour as (\ref{S7E10}) if $U$ is bounded.
On the other hand, in order to approximate the solutions of (\ref{S7E7}),
(\ref{S7E9}) for $E\gg 1,\ E\ll \frac{1}{\lambda}$ we use the fact that for small
$a,$ the solutions of (\ref{S7E7}), (\ref{S7E9}) in Theorem \ref{ThIntAs} have
the characteristic length scale $\frac{1}{a}.$ Then, for $a$ small, which
corresponds to the matching region indicated above, we have a characteristic
length very large compared with one. Then, as in (\ref{S5E2}) we can approximate $I\left[
\mathcal{H}\right] \sim \lambda \mathcal{H} \mathcal{U}$.
Since $E\ll \frac{1}{\lambda}$ we have $\mathcal{U}\left( X\right) \ll 1$
and we obtain from (\ref{S7E7}) the approximation
$\mathcal{H}\left( X\right) =\mathcal{U}\left( X\right) \mathcal{V}\left(
X\right) +\left( \mathcal{U}\left( X\right) \right) ^{2}\mathcal{V}%
\left( X\right)$
and plugging this formula into (\ref{S7E9}) and using the rescaling
(\ref{S7E1}) we obtain
\[
\frac{dU}{dX}=U\left( V-1\right) +\lambda\left( U\right) ^{2}%
V\ \ ,\ \ \frac{dV}{dX}=-\lambda UV-\lambda^{2}\left( U\right) ^{2}V
\]
that agrees with (\ref{T2E7}) in the range of values $U\gg 1,\ U\ll \frac
{1}{\lambda}.$
The previous computations show that the equations used in both transition regimes
agree in the intermediate matching regime. Actually it is possible to check
that the main characteristics of the computed solution also agree, as could
be expected. Notice that the length scale $\left( X_{n+1}^{+}-X_{n}%
^{+}\right) $ can be computed using the
function $T\left( E\right) $ in (\ref{S6E5a}) for
$E\gg 1,\ E\ll \frac{1}{\lambda}$:
\begin{equation}
T\left( E\right) \sim2\sqrt{2}\sqrt{E}\,. \ \label{T3E1}%
\end{equation}
In the computation of this integral, we have split the region of integration
into the two subsets $\left( U_{-}\left( E,0\right) ,\varepsilon_{0}\right)
$ and $\left( \varepsilon_{0},U_{+}\left( E,0\right) \right) $ where
$\varepsilon_{0}>0$ is a small but fixed number. It turns out that the
contribution of the second integral is much smaller than the one due to the
first integrand, exactly in the same manner that the contributions in $\left(
X_{n+1}^{+}-X_{n}^{+}\right) $ computed in Subsection \ref{iteration} are due
mostly to the region with small values of $U.$ Using the rescaling
(\ref{S5E5}) we then obtain the approximation
\begin{equation}
\left( X_{n+1}^{+}-X_{n}^{+}\right) \sim\frac{2\sqrt{2}\sqrt{E}}%
{\sqrt{\lambda}}\,. \label{T3E2}%
\end{equation}
On the other hand, we can compute the same length using (\ref{T2E6a}). Taking
into account that in the matching region $a_{n},\ a_{n+1}$ are small, we
obtain from (\ref{T2E5}) that
\begin{equation}
a_{n+1}\sim a_{n}+\frac{2}{3}\left( a_{n}\right) ^{2}\,. \label{T3E3}%
\end{equation}
Using (\ref{T2E6a}) we then obtain to
leading order that
\begin{equation}
X_{n+1}^{+}-X_{n}^{+}\sim\frac{\left( a_{n}+a_{n+1}\right) }{\lambda}%
\sim\frac{2a_{n}}{\lambda} \,.\label{T3E4}%
\end{equation}
Next, equation (\ref{S6E6}) and the definition of $a_{n}$ by means of $U=1,$ imply
\begin{equation}
a_{n}=\sqrt{2}\sqrt{E}\sqrt{\lambda} \label{T3E4a}%
\end{equation}
and plugging (\ref{T3E4a}) into (\ref{T3E4}) we obtain the sought-for matching with
(\ref{T3E2}).
Finally we obtain the matching for the formula of the change of energy in each
iteration. We have obtained using the ODE approximation that the
change of energy in each cycle is given by (\ref{S6E7}) where in the matching
region $1\ll E\ll \frac{1}{\lambda}$ we can use the approximation (\ref{S6E8a}).
Then
\begin{equation}
E_{n+1}-E_{n}\sim\frac{4\sqrt{2}}{3}E_{n}^{\frac{3}{2}}\,. \label{T3E5}%
\end{equation}
On the other hand, using again $a_{n}=\sqrt{2}\sqrt{E}\sqrt{\lambda}$ and
plugging it in (\ref{T3E3}) we obtain again (\ref{T3E5}) and this yields the
desired matching.
\section{Shooting argument}
\label{S.shooting}
The structure of the solutions of (\ref{S3E10})-(\ref{S3E13}) described in
Section \ref{asymptPeaks} by means of alternate approximate solutions of
(\ref{S7E7})-(\ref{S7E9}) and (\ref{S7E10}) yields several sequences of
numbers $\left\{ a_{n}\right\} ,\ \left\{ X_{n}^{+}\right\} ,\ \left\{
X_{n}^{-}\right\} ,\ \left\{ X_{n}^{0}\right\} $ that measure the amplitude
of the oscillations, the positions of the points where $U=1,\ V>1,\ U=1,\ V<1$
and the position of the peaks where $U$ and $H$ are of order $\frac{1}%
{\lambda}$ respectively. It is relevant to notice that the sequence $\left\{
a_{n}\right\} $ is increasing due to Lemma \ref{L1}. The values of $a_{n}$
approach zero in the matching region (see the approximation (\ref{T3E4a}) for
$1\ll E\ll \frac{1}{\lambda}$), but they become of order one for $E\approx\frac
{1}{\lambda}$ with increases in each iteration of order one. Eventually, this
sequence of points becomes larger than or equal to one.
A detailed study of the solutions of (\ref{S3E10})-(\ref{S3E13}) shows that
there are several possibilities. If $a_{n}$ becomes strictly larger than one,
Theorem \ref{ThIntAs} shows that during the peak regime $V$
becomes strictly negative. In particular $V$ reaches the value $V=0$ at some
finite $X_{\ast}.$ If $V$ remains positive for larger values of $X$, something
that could happen if $a_{n}=1,$ it would be possible to approximate
(\ref{S3E10})-(\ref{S3E13}) by means of (\ref{S7E10}). Therefore $V$ would
increase, and the value of $a_{n}$ would be increased to a new value
$a_{n+1}>1$ at the next iteration. The only possibility of not having such a
behaviour would be with $V\rightarrow0$ as $X\rightarrow\infty.$ Such
approximation to zero cannot be described by the approximate equation
(\ref{S7E10}), because in such a regime the integral terms in (\ref{S3E10}%
)-(\ref{S3E13}) must be taken into account in full detail.
We now recall that for any $K\in\mathbb{C}$ there exists a unique solution of
(\ref{S3E10})-(\ref{S3E13}) satisfying (\ref{S4E0}). The discussion in Section
\ref{Ss.variables} shows that all the solutions of (\ref{S3E10})-(\ref{S3E13})
with such a behaviour can be obtained, up to rescaling, by choosing $K$ in the
interval $\left[ 1,\exp\left( \frac{2\pi\beta\left( \lambda\right)
}{\alpha\left( \lambda\right) }\right) \right) .$ A continuity argument
taking into account the behaviour of the trajectories shows that there exists
a value of $K=K_{\ast}$ in this interval for which the corresponding solution
of (\ref{S3E10})-(\ref{S3E13}) satisfies $\lim_{X\rightarrow\infty}V\left(
X\right) =0.$ This also implies $\lim_{X\rightarrow\infty}H\left( X\right)
=0$ and $\lim_{X\rightarrow\infty}U\left( X\right) =0$ using (\ref{S3E10}),
(\ref{S3E12}). More precisely, the inversion of (\ref{S3E10}) allows us to write
$H\left( X\right) $ in terms of $UV.$ Then $\lim_{X\rightarrow\infty
}H\left( X\right) =0$ and (\ref{S3E12}) yields $\lim_{X\rightarrow\infty
}U\left( X\right) =0.$
\section{Asymptotics as $X\rightarrow\infty$}
\label{S.asymptotics}
We finally describe the asymptotics of the solution of (\ref{S3E10}%
)-(\ref{S3E13}) satisfying (\ref{S4E0}) with $K=K_{\ast}$ described in the
previous Section. We have seen that $
\lim_{X\rightarrow\infty}U\left( X\right) =\lim_{X\rightarrow\infty}V\left(
X\right) =\lim_{X\rightarrow\infty}H\left( X\right) =0\,.$
To leading order the behaviour of this trajectory is described by the
solution of (\ref{S7E7}), (\ref{S7E9}) given in Theorem \ref{ThIntAs} with
$a=1.$ The asymptotics
\begin{equation}
H\left( X\right) \sim\frac{2e^{\left( X-X_{n}^{0}\right) }e^{-e^{\left(
X-X_{n}^{0}\right) }}}{\lambda} \label{T4E1}%
\end{equation}
is valid as long as $H\gg 1,\ \left(
X-X^0_{n}\right) \gg 1$,
where, by assumption, to the leading order $a_{n}=1.$ It is convenient to
reformulate (\ref{T4E1}) in the original variables (cf. (\ref{S3E8}),
(\ref{S3E9}))
\begin{equation}
h\left( x\right) \sim\frac{2}{\lambda}\left( \frac{x}{x_{n}}\right)
\exp\left( -\left( \frac{x}{x_{n}}\right) \right) \label{T4E2}%
\end{equation}
that is valid for $\frac{x}{x_{n}}\gg 1,\ h\gg 1.$
We can match (\ref{T4E2}) with the following exponential asymptotics for the
solutions of (\ref{S2E5}) as $x\rightarrow\infty$:
\begin{equation}
h\left( x\right) \sim\frac{2}{\lambda}
\left( \kappa x\right) \exp\left( -\kappa x\right)
\label{T4E3}%
\end{equation}
The asymptotics (\ref{T4E2}), (\ref{T4E3}) match for small $\lambda$ if
$\kappa=\frac{1}{x_{n}}.$ This gives the desired exponential decay for the
obtained solution.
\section{Self-consistency of the approximations of
$I\left[ H\right]$}
\label{S.consistency}
We now show that the solution we obtained is self-consistent with
the approximations made for the integral operator $I\left[ H\right] \left(
X\right) $ defined in (\ref{S3E13}). We have made two main approximations.
The first one is (\ref{E1}), that approximates $I\left[ H\right] $ by an
operator with a simple dependence on $\lambda.$ The second approximation is
(\ref{S5E2}) and it allows us to replace the integral operator by a much simpler
local operator.
Concerning the validity of (\ref{E1}) we remark that its precise meaning is that the error
made in the approximation is smaller than the right-hand side. In order to
check its validity we distinguish between the regions where $H$ is of order
$\frac{1}{\lambda}$ and the regions where $H$ is smaller than that quantity.
Notice that for any region we can expect the errors made in the approximation
to be of order
\begin{align*}
R_{1} & =\lambda^{2}\int_{-\infty}^{X}dYe^{\left( Y-X\right) }\left(
X-Y\right) H\left( Y\right) \int_{\log\left( e^{X}-e^{Y}\right) }%
^{X}dZH\left( Z\right) \,, \\
R_{2} & =\lambda^{2}\int_{-\infty}^{X}dYe^{\left( Y-X\right) }H\left(
Y\right) \int_{\log\left( e^{X}-e^{Y}\right) }^{X}dZ\left( X-Z\right)
H\left( Z\right)
\end{align*}
and we can expect these two terms to be very small compared with the
right-hand side of (\ref{E1}). Indeed, in the case of $R_{1}$, the
contribution due to the region $\lambda\left( X-Y\right) \leq\varepsilon
_{0}$ is small compared with the term in (\ref{E1}) if $\varepsilon_{0}$ is
small. On the other hand, if $\lambda\left( X-Y\right) >\varepsilon_{0}$ we
can use the smallness of the exponential factor $e^{\left( Y-X\right) }$ to
obtain estimates for the corresponding terms in $R_{1}$ as $C\lambda
^{2}\left\Vert H\right\Vert _{\infty}^{2}e^{-\frac{\varepsilon_{0}}{\lambda}}$
and since $\left\Vert H\right\Vert _{\infty}\leq C/\lambda$ it then follows
that this contribution is exponentially small. Concerning $R_{2}$ we can argue
similarly for $\lambda\left( X-Z\right) $ smaller than $\varepsilon_{0}$. If
$\lambda\left( X-Z\right) >\varepsilon_{0}$, since $Z\geq X+\log\left(
1-e^{Y-X}\right) $ it follows that the size of the region of integration can
be bounded by $Ce^{-\frac{\varepsilon_{0}}{\lambda}}$ and therefore, it gives
also a very small contribution.
Concerning the approximations of $I\left[ H\right] $ by means of local terms
that have been made in the derivation of the ODE approximations (\ref{S5E6})
or (\ref{S7E10}) two main ingredients are required. In the derivation of
(\ref{S5E6}) we have used just the fact that $H\left( Y\right) $ has
significant changes over distances much longer than one. Since the order of
magnitude of $H$ is roughly the same for the range of values described by
means of (\ref{S5E6}) no special care is required to estimate the values of
$H\left( Y\right) $ with $X-Y\gg 1.$ On the other hand, in the case of the
approximation (\ref{S7E10}) we use the fact that $I\left[ H\right] $ is a
corrective term that is completely ignored in (\ref{S7E10}). Notice that
$H\left( Y\right) $ takes values for $Y<X$ much larger than the ones of
$H\left( X\right) $ in the region where the approximation (\ref{S7E10}) is
used. However, due to the exponential factor $e^{\left( Y-X\right) }$ as
well as the exponential decay of $H\left( Y\right) $ in the region described
by the integro-differential equation (cf. Subsection \ref{iteration}) the
corresponding contribution of such values of $H\left( Y\right) $ is
negligible.
|
\section{Introduction}
Boundary value problems governed by discrete equations have received some
attention lately by both variational and topological approach. The
variational techniques applied for discrete problems include, among others,
the mountain pass methodology, the linking theorem, the Morse theory, the
three critical point, compare with \cite{agrawal}, \cite{caiYu}, \cit
{AppMathLett}, \cite{TianZeng}, \cite{zhangcheng}, \cite{nonzero}. Moreover,
the fixed point approach is in fact much more prolific in the case of
discrete problem and covers the techniques already applied for continuous
problems, see for example \cite{FPAgrawal}, \cite{FPYangPing}, with both
list of references far from being exhaustive. \bigskip
While in the literature mainly the problem of the existence of solutions and
their multiplicity is considered, we are going to go a bit further and
investigate also the dependence on a functional parameter $u$ for the
following discrete boundary value problem which is a saddle -point type
system. Let $D>0$ be fixed. The problem which we consider reads
\begin{equation}
\left\{
\begin{array}{ll}
\Delta ^{2}x(k-1)=F_{x}(k,x(k),y(k),u(k)),\bigskip & \\
\Delta ^{2}y(k-1)=-F_{y}(k,x(k),y(k),u(k)),\bigskip & \\
x(0)=x(T+1)=y(0)=y(T+1)=0, &
\end{array
\right. \label{zad}
\end{equation
where $F:[1,T]\times \mathbb{R}\times \mathbb{R}\times \left[ -D,D\right]
\rightarrow \mathbb{R}$ is a continuous function differentiable with respect
to the second and the third variable,
\begin{equation*}
u\in L_{D}=\{u\in C([1,T],\mathbb{R}):||u||_{C}\leq D\},
\end{equation*
where $||u||_{C}$ denotes the classical maximum norm $||u||_{C}=\max_{k\in
\lbrack 1,T]}|u(k)|$ and $[a,b]$ for $a<b$, $a,b\in
\mathbb{Z}
$ denotes a discrete interval $\{a,a+1,...,b\}$. By a solution to (\ref{zad
) we mean a function $x:[0,T+1]\rightarrow \mathbb{R}$\ which satisfies the
given equation and the associated boundary conditions.\bigskip
Such type of a difference equation as (\ref{zad}) may arise from evaluating
the Dirichlet boundary value proble
\begin{equation*}
\begin{array}{l}
\frac{d^{2}}{dt^{2}}x=G_{x}\left( t,x,y,u\right) \text{,}\bigskip \text{
\frac{d^{2}}{dt^{2}}y=-G_{y}\left( t,x,y,u\right) \text{,} \\
0<t<1\text{, }x\left( 0\right) =x(1)=0,y\left( 0\right) =y(1)=
\end{array
\end{equation*
where $G:\left[ 0,1\right] \times \mathbb{R}\times \mathbb{R}\times \mathbb{
}\rightarrow \mathbb{R}$ is continuous and subject to some growth
conditions. Such a continuous problem subject to a functional parameter has
been considered in \cite{JS}. \ \bigskip
The question whether the system depends continuously on a parameter is vital
in context of the applications, where the measurements are known with some
accuracy. This question is even more important when \ the solution to the
problem under consideration is not unique as is the case of the present
note. In the boundary value problems for differential equations there are
some results towards the dependence of a solution on a functional parameter,
see \cite{LedzewiczWalczak}, \cite{JS} with references therein. This is not
the case with discrete equations where we have only some results which use
the critical point theory, see \cite{w}. The approach of this note is
different from this of \cite{w} since it does not relay on coercivity
arguments but on a min-max inequality due to Ky Fan, see \cite{nirenberg}.
In our approach we use some ideas developed in \cite{JS} suitable modified
due to the finite dimensionality of the space under consideration.
The following results will be used in the sequel, see \cite{nirenberg}.
\begin{theorem}[Fan's Min--Max Theorem]
\label{kYfanTh}Let $X$ and $Y$ be Hausdorff topological vector spaces,
A\subset X$ and $B\subset Y$ be convex sets, and $J:A\times B\rightarrow
\mathbb{R}$ be a function which satisfies the following conditions:
\begin{itemize}
\item[(i)] for each $y\in B$, the functional $x\rightarrow J(x,y)\rightarrow
\mathbb{R}$ convex and lower semi-continuous on $A$;
\item[(ii)] for each $x\in A$, the functional $y\rightarrow
J(x,y)\rightarrow \mathbb{R}$ is concave and upper semi-continuous on $B$;
\item[(iii)] for some $x_0\in A$ and some $\delta_0<\inf_{x\in A}\sup_{y\in
B}J(x,y)$, the set $\{y\in B:J(x_0,y)\}$ is compact.
\end{itemize}
Then
\begin{equation*}
\sup_{y}\inf_{x}J(x,y)=\inf_{x}\sup_{y}J(x,y).
\end{equation*}
\end{theorem}
\begin{definition}
Let $(X,\tau )$ be a Hausdorff topological space and let
(A_{n})_{n=1}^{\infty }$ be a sequence of nonempty subsets of $X$. The set
of accumulation points of sequences $(a_{n})_{n=1}^{\infty }$ with $a_{n}\in
A_{n}$ for $n=1,2,3,...$ is called the upper limit of $(A_{n})_{n=1}^{\infty
}$ and denoted by $\limsup A_{n}$.
\end{definition}
\section{Variational framework for problem (\protect\ref{zad})}
Solutions to (\ref{zad}) will be investigated in the space
\begin{equation*}
H=\{x:[0,T+1]\rightarrow \mathbb{R}:x(0)=x(T+1)=0\}
\end{equation*}
considered with the norm
\begin{equation*}
||x||=\left( \sum_{k=1}^{T+1}|\Delta x(k-1)|^{2}\right) ^{1/2}.
\end{equation*
Then $(H,||\cdot ||)$ becomes a Hilbert space. For any $m\geq 2$ let $c_{m}$
be the smallest positive constant such that
\begin{equation*}
\sum_{k=1}^{T}|x(k)|^{m}\leq c_{m}\cdot \sum_{k=1}^{T+1}|\Delta x(k-1)|^{m}
\end{equation*
for any $x\in H$; see \cite[Lemma 1]{MRT}. \bigskip
Since the approach of present note is a variational one we investigate the
action functional $J_{u}:H\times H\rightarrow R$, corresponding to problem
\ref{zad}). For a fixed parameter $u\in L_{D}$, $J_{u}$ is of the form
\begin{equation*}
J_{u}(x,y)=\sum_{k=1}^{T+1}\frac{|\Delta x(k-1)|^{2}}{2}-\frac{|\Delta
y(k-1)|^{2}}{2}+\sum_{k=1}^{T}F(k,x(k),y(k),u(k)).
\end{equation*}
We assume that $F$ has the following properties:\bigskip
\begin{itemize}
\item[H1] $F:[1,T]\times \mathbb{R}\times \mathbb{R}\times \mathbb{R
\rightarrow \mathbb{R}$ is a continuous function which is differentiable
with respect to the second and the third variable; $F_{x},F_{y}:[1,T]\times
\mathbb{R}\times \mathbb{R}\times \mathbb{R}\rightarrow \mathbb{R}$ are
continuous functions. \bigskip
\item[H2] For any fixed $y\in H$ there are a constant $\beta _{1},$ a
function $\gamma _{1}:\left[ 1,T\right] \rightarrow \mathbb{R}$ and a
constant $\alpha _{1}<1/(2c_{2})$ such that
\begin{equation*}
F(k,x,y(k),u)\geq -\alpha _{1}|x|^{2}+\beta _{1}x+\gamma _{1}(k)
\end{equation*
for all $x\in \mathbb{R}$, all $u\in \mathbb{R}$, $\left\vert u\right\vert
\leq D$ and all $k\in \left[ 1,T\right] .$\bigskip
\item[H3] For any fixed $x\in H$ there are a constant $\beta _{2},$ a
function $\gamma _{2}:\left[ 1,T\right] \rightarrow \mathbb{R}$ and a
constant $\alpha _{2}<1/(2c_{2})$ such that
\begin{equation*}
F(k,x(k),y,u)\leq \alpha _{2}|y|^{2}+\beta _{2}y+\gamma _{2}(k)
\end{equation*
for all $y\in \mathbb{R}$, all $u\in \mathbb{R}$, $\left\vert u\right\vert
\leq D$ and all $k\in \left[ 1,T\right] .$\bigskip
\item[H4] Functional $x\rightarrow J_{u}(x,y)$ is convex for all $y\in H$,
u\in L_{D}$.\bigskip
\item[H5] Functional $y\rightarrow J_{u}(x,y)$ is concave for all $x\in H$,
u\in L_{D}$.\bigskip
\end{itemize}
We observe that with any fixed $u\in L_{D}$ functional $J_{u}$ is
continuous. With the aid of Theorem \ref{kYfanTh} we are able to find saddle
points for functional $J_{u}$. Since $J_{u}$ is differentiable in the sense
of G\^{a}teaux, it is apparent that such points are the critical points to
J_{u}$. Since in turn critical points to $J_{u}$ constitute solutions to
\ref{zad}), we arrive at existence result once we get the existence of
saddle points. Moreover, since the spaces in which we work are finite
dimensional one, there is no need to distinguish between the weak and the
strong solutions.
\section{Existence of saddle point solutions}
\begin{theorem}[Existence of saddle points]
\label{ExistSaddlePoints}Assume that conditions H1-H2 hold. Let $u\in L_{D}$
be fixed. Then it follows that\newline
(A) There is a saddle point $(x_{u},y_{u})$ for the functional $J$;\newline
(B) There are balls $B_{1}=\{x:||x||\leq r_{1}\}$ and $B_{2}=\{y:||y||\leq
r_{2}\}$ such that $(x_{u},y_{u})\in B_{1}\times B_{2}$;\newline
(C) The set of all saddle points of $J_{u}$ is compact.
\end{theorem}
\begin{proof}
For fixed $y\in H$ using H2 we obtai
\begin{equation*}
\begin{array}{l}
J_{u}(x,y)\geq \sum_{k=1}^{T+1}\left( \frac{|\Delta x(k-1)|^{2}}{2}-\frac
|\Delta y(k-1)|^{2}}{2}-\alpha _{1}|x(k)|^{2}+\beta _{1}x(k)+\gamma
_{1}\right) \geq \bigskip \\
\left( \frac{1}{2}-c_{2}\alpha _{1}\right) ||x||^{2}+\tilde{\beta}_{1}||x||
\tilde{\gamma}_{1}
\end{array
\end{equation*}
where $\tilde{\beta}_{1}>0$ depends only on $\beta _{1}$ (note that $||x||$
and $\sum_{k=1}^{T+1}|x(k)|$ are equivalent norms, since $H$ is
finite-dimensional) and $\tilde{\gamma}_{1}>0$ depends only on $\gamma _{1}$
and $y$. Since $\frac{1}{2}-c_{2}\alpha _{1}>0$, the functional $J_{u}(x,y)$
is coercive on $H$. By H1 and H4 it is continuous and convex for each $u$.
Put
\begin{equation*}
J_{u}^{-}(y)=\min_{x}J_{u}(x,y).
\end{equation*
By H5 the functional $J_{u}^{-}$ is concave. By H3 we obtain that
\begin{equation}
\begin{array}{l}
J_{u}^{-}(y)\leq J_{u}(0,y)\leq \bigskip \\
\sum_{k=1}^{T+1}\left( -\frac{|\Delta y(k-1)|^{2}}{2}+\alpha
_{2}|y(k)|^{2}+\beta _{2}y(k)+\gamma _{2}\right) \leq \bigskip \\
\left( -\frac{1}{2}+c_{2}\alpha _{1}\right) ||y||^{2}+\tilde{\beta}_{2}||y||
\tilde{\gamma}_{2}
\end{array}
\label{eq1}
\end{equation
where $\tilde{\beta}_{2}>0$ depends only on $\beta _{2}$ and $\tilde{\gamma
_{2}>0$ depends only on $\gamma _{2}$. Since the constant $-\frac{1}{2
+c_{2}\alpha _{1}$ is negative, then $J_{u}^{-}$ is anti-coercive. Hence it
attains its supremum at some point $y_{u}$. By H2 we hav
\begin{equation*}
\begin{array}{l}
J_{u}^{-}(y_{u})\geq J_{u}^{-}(0)=\min_{x}J_{u}(x,0)\geq \bigskip \\
\min_{x}\left( \left( \frac{1}{2}-c_{2}\alpha _{1}\right) ||x||^{2}
\overline{\beta }_{1}||x||+\gamma _{1}\right) =\gamma _{1}
\end{array
\end{equation*}
Since $J_{u}^{-}$ is anti-coercive, there is $r_{2}>0$ such that
J_{u}^{-}(y)<\gamma _{1}$ for every $||y||>r_{2}$. Since $J_{u}^{-}$ is
continuous the set $\{y:J_{u}^{-}(y)\geq \gamma _{1}\}$ is compact and is
contained in $B_{2}$. Hence each $y_{u}$ is in $B_{2}$.\bigskip
Analogously one can show that there is $x_{u}$ with
\begin{equation*}
J_{u}^{+}(x_{u})=\min_{x}J_{u}^{+}=\min_{x}\max_{y}J_{u}(x,y).
\end{equation*
Furthermore, there is a ball $B_{1}$ with $x_{u}\in B_{1}$ for each such
x_{u}$.
We have already showed that for each $x$ there exists $\max_{y}J_{u}(x,y)$.
Hence for some $\delta _{0}$ we have
\begin{equation*}
\delta _{0}<\min_{x}J_{u}(x,0)\leq \min_{x}\max_{y}J_{u}(x,y).
\end{equation*
By (\ref{eq1}) we obtain
\begin{equation*}
\{y:J_{u}(0,y)\geq \delta _{0}\}\subset \{y:(-1/2+c_{2}\alpha _{1})||y||^{2}
\tilde{\beta}_{2}||y||+\tilde{\gamma}_{2}\geq \delta _{0}\}.
\end{equation*
Since the set of right hand of inclusion is compact, so is the set
\{y:J_{u}(0,y)\geq \delta _{0}\}$. Thus, the assumptions H4 and H5 and Fan's
minimax Theorem \ref{kYfanTh}, give the existence of a saddle point of
J_{u} $. Moreover the set of all saddle points of $J_{u}$ is compact.
\end{proof}
\begin{theorem}[Existence of saddle point solutions]
\label{ExistSaddlePointSol}Assume that conditions H1-H5 hold. Let $u\in
L_{D} $ be fixed. Then it follows that there exists is at least one saddle
point $(x_{u},y_{u})\in H\times H$ for the functional $J_{u}$ which solves
\ref{zad}).
\end{theorem}
\begin{proof}
By Theorem \ref{ExistSaddlePoints} there is at least one saddle point
(x_{u},y_{u})$ for the functional $J_{u}$. Since $J_{u}$ is a G\^{a}teaux
differentiable functional we see that $J_{u}^{^{\prime }}(x_{u},y_{u})=0$
and therefore $(x_{u},y_{u})$ solves (\ref{zad}).
\end{proof}
In order to obtain existence results we do not need to impose conditions
H2-H5 uniformly in $u$. This is not the case when one is interested in the
dependence on parameters, when assumptions must be placed uniformly with
respect to $u$. Indeed, let us consider a following problem
\begin{equation}
\left\{
\begin{array}{l}
\Delta ^{2}x(k-1)=F_{x}(k,x(k),y(k)),\bigskip \\
\Delta ^{2}y(k-1)=-F_{y}(k,x(k),y(k)),\bigskip \\
x(0)=x(T+1)=y(0)=y(T+1)=0
\end{array
\right. \label{zad2}
\end{equation
where $F:[1,T]\times \mathbb{R}\times \mathbb{R}\rightarrow \mathbb{R}$ is a
continuous function which is differentiable with respect to the second and
the third variable. The action functional $J:H\times H\rightarrow R$,
corresponding to problem (\ref{zad2}) is
\begin{equation*}
J(x,y)=\sum_{k=1}^{T+1}\frac{|\Delta x(k-1)|^{2}}{2}-\frac{|\Delta
y(k-1)|^{2}}{2}+\sum_{k=1}^{T}F(k,x(k),y(k)).
\end{equation*
We assume that\bigskip
\begin{itemize}
\item[H6] $F:[1,T]\times \mathbb{R}\times \mathbb{R}\rightarrow \mathbb{R}$
is a continuous function which is differentiable with respect to the second
and the third variable; $F_{x},F_{y}:[1,T]\times \mathbb{R}\times \mathbb{R
\rightarrow \mathbb{R}$ are continuous functions.\bigskip
\item[H7] For any fixed $y\in H$ there are a constant $\beta _{1},$ a
function $\gamma _{1}:\left[ 1,T\right] \rightarrow \mathbb{R}$ and a
constant $\alpha _{1}<1/(2c_{2})$ such that
\begin{equation*}
F(k,x,y(k),u)\geq -\alpha _{1}|x|^{2}+\beta _{1}x+\gamma _{1}(k)
\end{equation*
for all $x\in \mathbb{R}$ and all $k\in \left[ 1,T\right] .$\bigskip
\item[H8] For any fixed $x\in H$ there are a constant $\beta _{2},$ a
function $\gamma _{2}:\left[ 1,T\right] \rightarrow \mathbb{R}$ and a
constant $\alpha _{2}<1/(2c_{2})$ such that
\begin{equation*}
F(k,x(k),y,u)\leq \alpha _{2}|y|^{2}+\beta _{2}y+\gamma _{2}(k)
\end{equation*
for all $y\in \mathbb{R}$ and all $k\in \left[ 1,T\right] .$\bigskip
\item[H9] Functional $x\rightarrow J_{u}(x,y)$ is convex for any $y\in H
.\bigskip
\item[H10] Functional $y\rightarrow J_{u}(x,y)$ is concave for any $x\in H
.\bigskip
\end{itemize}
Then we have
\begin{corollary}
Assume that conditions H6-H10 hold. Then it follows that there exists is at
least one saddle point $(x,y)\in H\times H$ for the functional $J$ which
solves (\ref{zad}).
\end{corollary}
\section{Continuous dependence on parameters}
Now we are interested of the behavior of the sequence of saddle points which
correspond to a sequence of parameters. Dependence on parameters in
investigated through the convergence of the sequence of action functionals
corresponding the sequence of parameters - this approach has already been
applied with some success for the continuous and also the discrete problems,
see \cite{w}, \cite{LedzewiczWalczak}. Let $(u_{n})_{n=1}^{\infty }\subset
L_{D}$ be a sequence of parameters. We put $J_{n}=J_{u_{n}}$ and let
\begin{equation*}
V_{n}=\{(\overline{x},\overline{y}):J_{n}(\overline{x},\overline{y
)=\max_{y}\min_{x}J_{n}(x,y)\}\subset B_{1}\times B_{2}
\end{equation*
be the set of all saddle points of $J_{n}$. Due to Theorem \re
{ExistSaddlePoints} $V_{n}\neq \emptyset $ for all $n=1,2,...$ .
\begin{theorem}
\label{condepd}Assume that conditions H1-H5 hold. Let $(u_{n})_{n=1}^{\infty
}\subset L_{D}$ be a \ convergent sequence of parameters and
u_{n}\rightarrow u_{0}\in L_{D}$ as $n\rightarrow \infty $. Then $\emptyset
\neq \limsup_{n\rightarrow \infty }V_{n}\subset V_{0}$.
\end{theorem}
\begin{proof}
At first we observe by continuity of $F$ that $J_{n}$ tends to $J_{0}$
uniformly on $B_{1}\times B_{2}$, where $B_{1},$ $B_{2}$ are defined in
Theorem \ref{ExistSaddlePoints}. We will prove that $\emptyset \neq \limsup
V_{n}\subset V_{0}$. Let $a_{n}=\max_{y}\min_{x}J_{n}(x,y)$ and let
\varepsilon >0$. Since $J_{n}$ tends uniformly to $J_{0}$, then
J_{n}(x,y)\leq J_{0}(x,y)+\varepsilon $ for each $(x,y)\in B_{1}\times B_{2}$
and every $n\geq n_{0}$ for some $n_{0}$. Then
\begin{equation*}
\min_{x}J_{n}(x,y)\leq \min_{x}J_{0}(x,y)+\varepsilon ,
\end{equation*
\begin{equation*}
\max_{y}\min_{x}J_{n}(x,y)\leq \max_{y}\min_{x}J_{0}(x,y)+\varepsilon .
\end{equation*
Hence $a_{k}-a_{0}\leq \varepsilon $. Similarly one can show that
a_{k}-a_{0}\geq -\varepsilon $. Therefore $a_{k}\rightarrow a_{0}$.\bigskip
Let $(x_{n},y_{n})\in V_{n}$ for $n=1,2,...$. Since
\begin{equation*}
\{(x_{n},y_{n})\}_{n=1}^{\infty }\subset B_{1}\times B_{2}
\end{equation*
we may assume that $(x_{n},y_{n})\rightarrow (x_{0},y_{0})$. In particular
\limsup V_{n}\neq \emptyset $. Suppose now that $(x_{0},y_{0})\notin V_{0}$.
Let $(\overline{x},\overline{y})\in V_{0}$. Then $J_{0}(\overline{x}
\overline{y})\neq J_{0}(x_{0},y_{0})$. Consider the case
\begin{equation*}
J_{0}(\overline{x},\overline{y})-J_{0}(x_{0},y_{0})=\eta <0.
\end{equation*
The
\begin{equation*}
\begin{array}{l}
a_{n}-a_{0}=J_{n}(x_{n},y_{n})-J_{0}(x_{0},y_{0})=\bigskip \\
\min_{x}J_{n}(x,y_{n})-J_{0}(x_{0},y_{0})\leq \bigskip \\
\leq J_{n}(\overline{x},y_{n})-J_{0}(x_{0},y_{0})=\bigskip \\
J_{n}(\overline{x},y_{n})-J_{0}(\overline{x},y_{n})+J_{0}(\overline{x
,y_{n})-J_{0}(\overline{x},\overline{y})+J_{0}(\overline{x},\overline{y
)-J_{0}(x_{0},y_{0})
\end{array
\end{equation*}
Sinc
\begin{equation*}
J_{0}(\overline{x},\overline{y})=\max_{y}J_{0}(\overline{x},y)\geq J_{0}
\overline{x},y_{n}),
\end{equation*
then
\begin{equation*}
\limsup_{n}J_{0}(\overline{x},y_{n})-J_{0}(\overline{x},\overline{y})\leq 0.
\end{equation*
By the continuity of $F$ we obtain that $J_{n}(\overline{x
,y_{n})\rightarrow J_{0}(\overline{x},y_{n})$. Therefore
\begin{equation*}
\limsup_{n\rightarrow \infty }(a_{n}-a_{0})<\eta .
\end{equation*
A contradiction. Similarly, a contradiction can be obtained when $\eta >0$.
\end{proof}
Theorem \ref{condepd} combined with Theorem \ref{ExistSaddlePointSol} yield
the following main result of our note
\begin{theorem}
Assume H1-H5. For any fixed $u\in L_{D}$ there exists at least one solution
y\in V_{u}$ to problem (\ref{zad}). Let $\{u_{n}\}\subset L_{D}$ be a
convergent sequence of parameters, where $\underset{n\rightarrow \infty }
\lim }u_{n}=u_{0}\in L_{D}$. For any sequence $\{\left( x_{n},y_{n}\right) \}
$ of solutions $\left( x_{n},y_{n}\right) \in V_{n}$ to the problem (\re
{zad}) corresponding to $u_{n}$, there exist a subsequence $\{\left(
x_{n_{i}},y_{n_{i}}\right) \}\subset H\times H$ and an element $\left(
x_{0},y_{0}\right) \subset H\times H$ such that $\underset{i\rightarrow
\infty }{\lim }x_{n_{i}}=x_{0}$, $\underset{i\rightarrow \infty }{\lim
y_{n_{i}}=y_{0}$ and $J_{0}(x_{0},y_{0})=\max_{y}\min_{x}J_{0}(x,y)$.
Moreover $x_{0},y_{0}\in V_{0}$, i.e. the pair $\left( x_{0},y_{0}\right) $
satisfies (\ref{zad}) with $u=u_{0}$, namely
\begin{equation*}
\left\{
\begin{array}{l}
\Delta ^{2}x_{0}(k-1)=F_{x}(k,x_{0}(k),y_{0}(k),u_{0}(k)),\bigskip \\
\Delta ^{2}y_{0}(k-1)=-F_{y}(k,x_{0}(k),y_{0}(k),u_{0}(k)),\bigskip \\
x_{0}(0)=x_{0}(T+1)=y_{0}(0)=y_{0}(T+1)=0
\end{array
\right.
\end{equation*}
\end{theorem}
|
\section{Introduction}
Some might argue that quantum mechanics became a universal theory of matter when, at the last Einstein-Bohr debate at the Solvay Conference in 1930, Einstein proposed a weight measurement to observe unobtrusively a particle decay in order to contradict the Heisenberg energy-time uncertainty relation \cite{1.}. After nearly being defeated by Einstein in the debate, Bohr surprisingly countered with a general relativistic gravitational argument. Henceforth, the relation between gravity and quantum mechanics was to become an important question in fundamental physics.
It could also have been questioned whether the relations between energy and frequency and between momentum and wave vector, introduced for matter waves six years earlier by de Broglie \cite{1.}, were rigorously valid in a general curved spacetime. This question can be shown to be equivalent to the question of whether sufficiently small wave packets travel along classical paths consistent with the de Broglie relations. Recall that a wave packet is a wave whose amplitude, frequency, and wave vector vary slowly over a region of spacetime comparable to a period or wave length. (E.g., for an electron traveling at half the speed of light, the wave length is approximately 5 $\times$ 10${}^{-}$${}^{12}$ m, and a typical wave packet has dimensions 10${}^{-}$${}^{6}$ m \cite{2.}.) The de Broglie relations can be observed for such wave packets.
The well known WKB approximation is commonly used to derive wave packet approximations in quantum mechanics (e.g., \cite{Audretsch1981A}). The WKB approximation is based on taking the limit as a physical constant, namely Planck's constant $\hbar $, approaches zero.
\footnote{\
More generally, a common conception is that classical physics emerges from quantum mechanics in the limit as Planck's constant $\hbar $ approaches zero. However, the limit $\hbar \rightarrow 0$ ``is not well defined mathematically unless one specifies what quantities are to be held constant during the limiting process" \cite{Ballentine2001}. It is interesting to note that the most classical behaving Gaussian wave functions, the coherent states of the ordinary harmonic oscillator, whose expected position and momentum obey classical equations by Ehrenfest's theorem, do not resemble wave packets in the limit $\hbar \rightarrow 0$ \cite{Park1990}.
}
However, the spin connection in the Dirac equation of a curved spacetime has no effect in the first WKB approximation (i.e., the one retaining only the zero order term in $\hbar $) \cite{Audretsch1981A}. The assumption that the spin connection can be neglected, as it would be in a first WKB approximation, is unnecessary
and is too strong for many applications in a curved spacetime, or even in a Minkowski spacetime with arbitrary coordinates. Note that throughout this paper, except for the brief description of a post Newtonian approximation in Section \ref{WhithamMethod}, we may set both the speed of light $c$ and Planck's constant $\hbar $ equal to one.
The Whitham approximation \cite{12.}, which we adopt in this paper, places no restriction on Planck's constant. To implement the Whitham approximation and to show that it leads to propagation along classical paths, we will first show in Section \ref{Canonical Form} that any Dirac equation in a curved spacetime can be transformed into an equivalent canonical form known in the literature as the ``local representation" \cite{3.}, \cite{4.}. In general, transformation to equivalent canonical form is a necessary step to simplify a Dirac equation so that propagation along classical paths can be derived. It will be evident in Theorem 1 of Section \ref{Canonical Form} that Planck's constant $\hbar $ does not appear in the transformations mapping Dirac equations to their equivalent canonical forms \cite{3.}, \cite{4.}. Previously, these canonical forms, or ``local representations" as they are called in the literature, have only been discussed in the special case of orthogonal coordinates \cite{3.}, \cite{4.}, \cite{5.}.
Then in Section \ref{WhithamMethod}, with each Dirac equation transformed into equivalent canonical form, we apply Whitham's Lagrangian method \cite{12.} to derive wave packets in general curved spacetimes. Whitham's method preserves the symmetries of the Lagrangian, and in particular, the Whitham wave packet equations conserve the probability current. We also show that generalized de Broglie relations, as well as COW and Sagnac type terms \cite{10.}, emerge from the Whitham equations. It will become clear in Sections 2 and 3 that for every Dirac equation transformed into equivalent canonical form, the generalized de Broglie relations have no other meaning than the fact that sufficiently small wave packets propagate along classical paths in a background of gravitational and electromagnetic fields. This is what is observed in experiments
\footnote{ Note, however, that the electron's magnetic moment, predicted by the Dirac equation, is not contained in the wave packet approximation. Indeed, quoting from Ref. \cite{13.}: ``The uncertainty principle, together with the Lorentz force, prevents spin-up and spin-down electrons from being separated by a macroscopic field of the Stern-Gerlach type.'' In practice, wave packet splitting in Stern-Gerlach experiments is only observed using neutral atoms or molecules, which are undisturbed by the Lorentz force \cite{Ballentine2001}. In Section \ref{WhithamMethod}, the wave packet approximation is expressed by neglecting in the Lagrangian the variation in the amplitude of the wave function as compared to the variation of its phase. This leads to wave packet equations which do not involve the electron's magnetic moment.
}
and therefore more physically precise than the statement often made that particles with a given energy and momentum possess a frequency and wave vector given by the de Broglie relations. In fact, the generalized de Broglie relations are a \textit{direct consequence} of Whitham's method applied to each Dirac equation transformed into equivalent canonical form, and not just a physical \textit{hypothesis} as introduced by Einstein and de Broglie, and by many quantum mechanics textbooks.
In the WKB approximation of the standard Dirac equation in a curved spacetime, classical trajectories are derived from the Gordon decomposition of the probability current $J^\mu = J^\mu _c + J^\mu _s$ into a convection current $J^\mu _c$ and a spin current $J^\mu _s$ \cite{Audretsch1981A}. In Section \ref{WhithamMethod}, we also prove the existence of the Gordon decomposition for Dirac equations transformed into canonical form. We further show that in the Whitham approximation, the spin current $J^\mu _s$ vanishes, which explains the negligible effect of spin on wave packet solutions, independent of the size of Planck's constant $\hbar $.
It is also clear that the canonical forms (or ``local representations'') of the Dirac equations, while not unique, are the preferred representations to understand certain phenomena associated with the Dirac equations in a curved spacetime, particularly, the emergence of classical physics and its conservation laws in a quantum world.
Section \ref{Classical-Quantum} concludes this paper with a discussion of the classical-quantum correspondence in a curved spacetime based on both Lagrangian and Hamiltonian formulations of the Whitham equations. In this section we also include results from a previous analysis of the classical-quantum correspondence \cite{15.}, which can be applied to the canonical forms of Dirac equations in a curved spacetime considered in this paper.
\section{New Representations of the Dirac Equation in a Curved Spacetime and their Equivalent Canonical Forms}\label{Canonical Form}
Shortly after Dirac discovered his celebrated four component wave equation:
\be
\gamma ^{\mu } \partial _{\mu } \Psi = -\frac{imc}{\hbar } \Psi,
\label{GrindEQ__1_}
\ee
together with its conserved probability current:
\be J^{\mu } = c \Psi ^{+} A\gamma ^{\mu } \Psi,
\label{GrindEQ__2_}
\ee
his equation was studied in its widest representations for a Minkowski spacetime \cite{6.}, \cite{16.}.
In Eq. \eqref{GrindEQ__1_}, the Dirac field $ \Psi $ is a four component complex function of spacetime coordinates $ x^{\mu } $, $ \mu = 0 , 1 , 2, 3 $, whose partial derivatives with respect to $ x^{\mu } $ are denoted as $ \partial _{\mu } \Psi $. The Dirac gamma matrices $ \gamma ^{\mu } $, acting on $ \Psi $, satisfy the anticommutation formula:
\be
\gamma ^{\mu } \gamma ^{\nu } + \gamma ^{\nu } \gamma ^{\mu } = 2 \eta ^{\mu \nu } {\bf 1}_{{\bf 4}}, \label{GrindEQ__3_}
\ee
where $ \eta ^{\mu \nu } $ is the inverse of the Minkowski metric tensor $ \eta _{\mu \nu } $, and $ {\bf 1}_{{\bf 4}} $ denotes the identity matrix acting on the Dirac field $ \Psi $. The mass, the speed of light, and Planck's constant are denoted by $ m $, $ c $, and $ \hbar $, respectively. Repeated indices are summed.
In Eq. \eqref{GrindEQ__2_}, $ \Psi ^{+} $ denotes the complex conjugate transpose (Hermitian conjugate) of the Dirac field $ \Psi $, and $ A $ is a hermitizing matrix for the Dirac gamma matrices $ \gamma ^{\mu } $. That is, \cite{6.}, \cite{16.}:
\noindent
\be\begin{array}{l} {A^{+} = A,} \\ {} \\ {\gamma ^{\mu +} = A \gamma ^{\mu } A^{-1} }, \end{array} \label{GrindEQ__4_}\ee
\noindent where $ \gamma ^{\mu +} $ and $ A^{+} $ denote the Hermitian conjugates of the matrices $ \gamma ^{\mu } $ and $ A $, respectively. The hermitizing matrix $ A $ is uniquely determined by the matrices $ \gamma ^{\mu } $ up to a nonzero real scalar multiple \cite{6.}.
For a Minkowski spacetime, assuming that $ \left( \gamma ^{\mu } , A \right) $ are chosen to be constant matrices satisfying Eqs. \eqref{GrindEQ__3_} and \eqref{GrindEQ__4_}, every solution of Eq. \eqref{GrindEQ__1_} satisfies the Klein-Gordon equation, and the probability current $ J^{\mu } $ in Eq. \eqref{GrindEQ__2_} is then also conserved. That is, Eqs. \eqref{GrindEQ__3_} and \eqref{GrindEQ__4_} are the only conditions that the constant matrices $ \left( \gamma ^{\mu } , A \right) $ need satisfy.
The ``coefficient matrices'' $ \left( \gamma ^{\mu } , A \right) $ in Eqs. \eqref{GrindEQ__1_} and \eqref{GrindEQ__2_} are far from unique. Given a Dirac field $ \Psi $, and any set of constant coefficient matrices $ \left( \gamma ^{\mu } , A \right) $ satisfying Eqs. \eqref{GrindEQ__3_} and \eqref{GrindEQ__4_}, they may be transformed by any constant complex $ 4\times 4 $ matrix $ S $ as follows:
\noindent
\be\begin{array}{l} {\widetilde{\Psi }\quad =\quad S^{-1} \Psi, } \\ {} \\ {\widetilde{\gamma }^{\mu } = S^{-1} \gamma ^{\mu } S,} \\ {} \\ {\widetilde{A}\quad =\quad S^{+} A S.} \end{array}\label{GrindEQ__5_}
\ee
\noindent Such a transformation $ S $ is called a ``similarity transformation'' or a ``spin-base transformation'', the latter referring to simply a change of basis for the four components of the Dirac field $ \Psi $. It is straightforward to see that Eqs. \eqref{GrindEQ__1_} $-$ \eqref{GrindEQ__4_} are invariant under all similarity transformations $ S $ by Eq. \eqref{GrindEQ__5_}. Thus, in the widest sense no restriction to a smaller group of transformations was deemed necessary in the early development of the Dirac equation \cite{6.}, \cite{16.}.
\noindent
\noindent Nevertheless, further choices were necessary when the Poincar\'e group of coordinate transformations of the Dirac equation was considered. At least three possibilities have been considered for Poincar\'e coordinate transformations in a Minkowski spacetime as follows:
\noindent
\bi
\item A) The Dirac field $ \Psi $ transforms as a quadruplet of complex scalars under a coordinate transformation. \cite{17.}, \cite{18.}, \cite{19.}, \cite{20.}\\
\item B) The Dirac field $ \Psi $ transforms as a complex four-vector $ \Psi ^{\mu } $ under a coordinate transformation. \cite{21.}\\
\item C) The Dirac field $ \Psi $ transforms as a quadruplet of complex scalars under a coordinate transformation, which is then followed by a similarity transformation. (The combined transformation leaves the constant coefficient matrices $ \left(\gamma ^{\mu } , A\right) $ invariant.) \cite{6.}\\
\ei
\noindent For a Minkowski spacetime with Poincar\'e coordinate transformations, all three possibilities may be considered. However, with general coordinate transformations, as required for a curved spacetime, only the first two possibilities (A) and (B) exist. Since in a general spacetime, the possibility (C) does not exist, it was replaced by the possibility (A) in what has become the standard Dirac equation, which was proposed independently by Weyl \cite{17.} and by Fock \cite{18.}, hereafter called the ``Dirac-Fock-Weyl'' (DFW) equation. See Refs. \cite{19.} and \cite{20.}. Possibility (B) was investigated recently \cite{21.}, which became the impetus for a more general study of Dirac equations \cite{7.}, \cite{8.}, \cite{22.}.
General Dirac fields of type (A) will be said to belong to the Quadruplet Representation of the Dirac theory (or QRD theory). General Dirac fields of type (B) will be said to belong to the Tensor Representation of the Dirac theory (or TRD theory). It was recently shown that in an open neighborhood of each spacetime point, every TRD equation is in fact equivalent to a QRD equation and vice versa \cite{7.}. Since TRD equations are locally equivalent to QRD equations, we will only consider QRD equations in this paper. Note that there are QRD equations in a curved spacetime that are not locally equivalent to any DFW equation \cite{7.}.
In a further evolution of the Dirac equation, which applies also to a Minkowski spacetime, the partial derivatives in the Dirac equation \eqref{GrindEQ__1_} were replaced by covariant derivatives $ D_{\mu } = \partial _{\mu } + \Gamma _{\mu } $ where $ \Gamma _{\mu } $ are four $ 4\times 4 $ complex matrices
\footnote{\
In the case of a Majorana representation of the Dirac field, the coefficient matrices $ \left( \gamma ^{\mu } , A \right) $ are pure imaginary, and the Dirac equation is real. In this case the spin connection matrices $ \Gamma _{\mu } $ are real.
}
, called ``spin connection matrices'', acting on the four components of the Dirac field $ \Psi $. At the same time, the Dirac equation was generalized by substituting a general metric $ g_{\mu \nu } $ of Lorentz signature and determinant $g$ for the Minkowski metric $ \eta _{\mu \nu } $ in the anticommutation formula \eqref{GrindEQ__3_} for the Dirac gamma matrices:
\noindent
\be\gamma ^{\mu } \gamma ^{\nu } + \gamma ^{\nu } \gamma ^{\mu } = 2 g^{\mu \nu } {\bf 1}_{{\bf 4}}. \label{GrindEQ__7_}
\ee
For the results in this paper, mild restrictions must be placed on the metric components $ g_{\mu \nu } $, namely that $ g_{00} > 0 $ and the components $ g_{jk} $ for $ j, k = 1, 2, 3 $ form a negative definite $ 3\times 3 $ matrix. Even though these mild conditions hold for almost all spacetime metrics $ g_{\mu \nu } $ of interest, the G\"odel metric is a notable exception \cite{23.}, \cite{24.}.
For this generalization, the coefficient matrices $ \left( \gamma ^{\mu } , A \right) $ defining the Dirac equation are augmented to become ``coefficient fields'' $ \left( \gamma ^{\mu } , A , \Gamma _{\mu } \right) $, which now may vary with the spacetime point. Then, in order that the Dirac equation \eqref{GrindEQ__1_} be invariant, the transformation equations \eqref{GrindEQ__5_} are augmented with the transformation of the spin connection matrices $ \Gamma _{\mu } $ under ``local similarity transformations'' $ S $ (i.e., similarity transformations $ S $ that may also vary with the spacetime point
\footnote{
Because $ S$ depends on the spacetime point, Schl\"utert, Wietschorke, and Greiner call $ S $ a ``local transformation'' \cite{3.}, \cite{4.}. }
) as follows: \cite{20.}
\noindent
\be\begin{array}{l} { \widetilde{\Gamma }_{\mu } = S^{-1} \Gamma _{\mu } S + S^{-1} \partial _{\mu } S} \\ {} \\ {\qquad =S^{-1} \left( \partial _{\mu } +\Gamma _{\mu } \right) S.} \end{array}
\label{GrindEQ__9_}
\ee
Indeed, for this type of transformation, we have the covariant derivatives transforming as: $\widetilde{D}_{\mu }= S^{-1} \circ D_{\mu } \circ S$. Transformations of the kind given by Eqs. \eqref{GrindEQ__5_} and \eqref{GrindEQ__9_} we will call local similarity transformations ``of the first kind''.
Local similarity transformations ``of the second kind'' are defined by setting:
\be
\widetilde{\Gamma }_{\mu } = \Gamma _{\mu }. \label{GrindEQ__11_}
\ee
For this second kind of transformation, given by Eqs. \eqref{GrindEQ__5_} and \eqref{GrindEQ__11_}, we have the covariant derivatives transforming as: $ \widetilde{D}_{\mu } = D_{\mu } $.
Two Dirac equations will be said to be ``equivalent'' or ``classically equivalent'' if there exists a local similarity transformation $ \Psi \to S^{-1} \Psi $ of any kind that transforms the solutions of one Dirac equation into the solutions of the other.
\footnote{
The notion of equivalence here is somewhat different than the notion used in Ref. \cite{25.}, where equivalence was applied only to classify the coefficient fields $ \left( \gamma ^{\mu } , A \right) $ without requiring the existence of a map $ \Psi \to S^{-1} \Psi $ between the solutions of two Dirac equations.
}
From Eqs. \eqref{GrindEQ__2_} and \eqref{GrindEQ__5_}, the conserved probability currents for two equivalent Dirac equations are equal. Hence in any spacetime, scattering experiments will give the same results regardless of the representation of a given Dirac equation. However, in a first quantized theory, which is our concern in this paper, a local similarity transformation $ S $ may not intertwine with the quantum mechanical operators. In that case, the operators corresponding to a given observable generally will not have the same eigenvalues before and after the transformation $S$. This has been shown for the Hamiltonian (or energy) operator in previous work \cite{25.}. Thus, two Dirac equations that are equivalent as partial differential equations via a local similarity transformation $S$, need not be equivalent with respect to all quantum mechanical operators \cite{25.}, \cite{26.}, \cite{27.}. \\
The standard ``Dirac Lagrangian'' applies to the DFW equation \cite{19.} and has to be extended to include the coefficient field $A$ in the exact place of the constant hermitizing matrix valid for DFW \cite{25.}:
\be\begin{array}{l} {L= L\left(\Psi, \partial _{\mu } \Psi , x^{\mu } \right)} \\ {} \\ {= \sqrt{ -g}\ \frac{i\hbar c }{2}\, {\rm [} \Psi ^{+} A\gamma ^{\mu } \left( D_{\mu } \Psi \right) -\left( D_{\mu } \Psi \right)^{+} A\gamma ^{\mu } \Psi + \frac{ 2 mc }{\hbar }\, i\, \Psi ^{+} A\Psi {\rm ]}.} \end{array}\label{GrindEQ__8_}
\ee
Note that the Lagrangian (\ref{GrindEQ__8_}) is the local expression of a global Dirac Lagrangian based on a general Dirac operator ${\not}\mathcal{D}$ acting on the cross-sections of a spinor bundle ${\sf E}$ over the spacetime.
\footnote{\
A globally defined generalized Dirac Lagrangian has the form:
\be\label{Lagrangian-intrinsic
L=\ \frac{i}{2}
\left [(\psi,{\not}\mathcal{D}\psi)-({\not}\mathcal{D}\psi,\psi)+2im(\psi ,\psi) \right].
\ee
where ${\not}\mathcal{D}$ is a Dirac operator acting on the cross-sections $\psi $ of a spinor bundle ${\sf E}$ over the spacetime, and $(\ ,\ )$ denotes a hermitizing metric on the fibers of ${\sf E}$. See Ref. \cite{7.}, Sect. 2.1 and the references therein for the precise definitions. Once any coordinate chart of the spacetime and any frame field on the spinor bundle have been chosen, one gets the local expression of the global generalized Lagrangian as Eq. (\ref{GrindEQ__8_}). In particular, in Eq. (\ref{GrindEQ__8_}) and in the rest of this paper, $\Psi $ is the column vector made with the components of $\psi $ in the chosen frame field on ${\sf E}$. See Ref. \cite{7.}, Sect. 2.2.
}
Thus, the Lagrangian (\ref{GrindEQ__8_}) gives a generalized formulation of the Dirac theory for a general Dirac operator ${\not}\mathcal{D}$ on a curved spacetime. We will see in Theorem 1 that the equivalent canonical forms of DFW equations require such a generalization.
The Euler-Lagrange equations for the Lagrangian \eqref{GrindEQ__8_} give the following generalized Dirac equation \cite{7.}, \cite{8.}:
\noindent
\be\gamma ^{\mu } D_{\mu } \Psi + \frac{1}{2} A^{-1} D_{\mu } \left(\, A\gamma ^{\mu } \right) \Psi = -\frac{imc}{\hbar } \Psi,
\label{GrindEQ__14_}\ee
\noindent where
\be\begin{array}{l} {D_{\mu } \Psi \quad \equiv \quad \partial _{\mu } \Psi + \Gamma _{\mu } \Psi } \\ {} \\ {D_{\mu } \gamma ^{\nu } \quad \equiv \quad \nabla _{\mu } \gamma ^{\nu } + \Gamma _{\mu } \gamma ^{\nu } - \gamma ^{\nu } \Gamma _{\mu } } \\ {} \\ {D_{\mu } A\quad \equiv \quad \partial _{\mu } A - \Gamma _{\mu }^{+} A - A \Gamma _{\mu }, } \end{array}
\label{GrindEQ__15_}\ee
\noindent and we define the Levi-Civita covariant derivatives \textbf{$ \nabla _{\mu } $ }acting on the Dirac field $ \Psi $ and the coefficient fields $ \left( \gamma ^{\mu } , A \right) $ as follows:
\noindent
\be\begin{array}{l} {\nabla _{\mu } \Psi \quad \equiv \quad \partial _{\mu } \Psi } \\ {} \\ {\nabla _{\mu } \gamma ^{\nu } \quad \equiv \quad \partial _{\mu } \gamma ^{\nu } + \left\{^\nu _{\rho\, \mu } \right\} \gamma ^{\rho } } \\ {} \\ {\nabla _{\mu } A\quad \equiv \quad \partial _{\mu } A} \end{array}
\label{GrindEQ__16_}\ee
\noindent where $\left\{^\nu _{\rho\, \mu } \right\}$ are the Christoffel symbols belonging to the Levi-Civita connection. The covariant derivatives $ D_{\mu } $ extend to $\Psi ^{+}$ by the formula $D_{\mu } \Psi ^{+}= \left( D_{\mu } \Psi \right)^{+} $, and similarly $\nabla _{\mu } \Psi ^{+}= \left(\nabla _{\mu } \Psi \right)^{+} $. As usual, covariant derivatives extend to products of fields via Leibniz's rule for differentiating products.
If $D_{\mu } \left(A\gamma ^{\mu } \right)= 0$, then the generalized Dirac equation \eqref{GrindEQ__14_} reduces to normal form:
\noindent
\be
\gamma ^{\mu } D_{\mu } \Psi = -\frac{imc}{\hbar } \Psi.
\label{GrindEQ__17_}
\ee
\noindent Normal Dirac equations generalize the DFW property that the coefficient fields $ \left( \gamma ^{\mu } , A \right) $ be covariantly constant: $ D_{\mu } \gamma ^{\nu } = 0 $ and $ D_{\mu } A = 0 $. One can show that the weaker normality condition $ D_{\mu } \left( A\gamma ^{\mu } \right) = 0 $ is preserved under all local similarity transformations of the first kind \cite{7.}. The normality condition is also preserved under all coordinate transformations.
Thus we have several invariant classes of Dirac equations. First, we have the class of Dirac equations for which the coefficient fields $ \left( \gamma ^{\mu } , A \right) $ are covariantly constant $-$ that is, $ D_{\mu } \gamma ^{\nu } = 0 $ and $ D_{\mu } A = 0 $. This class contains the DFW equations as a proper subset. Second, we have the class of Dirac equations for which $ D_{\mu } \left( A\gamma ^{\mu } \right) = 0 $. This is the class of normal Dirac equations \eqref{GrindEQ__17_} which contains the first class as a proper subset. Finally, we have the class of generalized Dirac equations \eqref{GrindEQ__14_} which contains the other two classes. Each of these classes is invariant under all coordinate transformations and also under all local similarity transformations of the first kind \cite{7.}.
A non-invariant class of Dirac equations that we will discuss in this paper is the class of the QRD--0 equations, in which the contracted spin connection matrix $ \Gamma \equiv \gamma ^{\mu } \Gamma _{\mu } = 0 $. See Ref. \cite{7.}, Sect. 3.2.1. Equations in the QRD--0 class (with $ \Gamma = 0$) are generalized Dirac equations which may be written as follows:
\be
\gamma ^{\mu } \partial _{\mu } \Psi + \frac{1}{2} A^{-1} \nabla _{\mu } \left(\, A\gamma ^{\mu } \right)\Psi =-\frac{imc}{\hbar } \Psi,
\label{GrindEQ__18_}
\ee
with the Levi-Civita covariant derivatives $ \nabla _{\mu } $ acting on the coefficient fields $ \left( \gamma ^{\mu } , A \right) $ as previously defined in Eq. \eqref{GrindEQ__16_}. If $ \nabla _{\mu } \left( A\gamma ^{\mu } \right) = 0 $, then the QRD--0 equation \eqref{GrindEQ__18_} reduces to the simpler normal form:
\noindent
\be\gamma ^{\mu } \partial _{\mu } \Psi = -\frac{imc}{\hbar } \Psi .
\label{GrindEQ__19_}\ee
\noindent Note that a QRD--0 equation which is both normal as in Eq. \eqref{GrindEQ__19_} and equivalent to a DFW equation is called a ``local representation'' of the DFW equation by other authors \cite{3.}, \cite{4.}, \cite{5.}. Finding a ``local representation'' for a DFW equation often simplifies the process of deriving its solutions. Previously, this ``local representation'' of the DFW equation in a curved spacetime in the form of Eq. \eqref{GrindEQ__19_} has only been discussed in the special case of orthogonal coordinates \cite{3.}, \cite{4.}, \cite{5.}.
Given any normal Dirac equation \eqref{GrindEQ__17_}, one can construct an equivalent normal QRD--0 equation \eqref{GrindEQ__19_} in terms of a basis of solutions of the massless equation associated with \eqref{GrindEQ__17_}:
\be\gamma ^{\mu } D_{\mu } \Psi = \gamma ^{\mu } \left( \partial _{\mu } +\Gamma _{\mu } \right) \Psi = 0.
\label{GrindEQ__20_}\ee
This is useful when solutions to the massless equation \eqref{GrindEQ__20_} are known, for which there are many examples in general relativity, including all diagonal metrics and G\"odel type metrics \cite{29.} $-$ \cite{33.}. Indeed, consider a local similarity transformation $ S $ which takes the Dirac field $ \Psi $ to $ \widetilde{\Psi } $ and the coefficient fields $ \left( \gamma ^{\mu } , A , \Gamma _{\mu } \right) $ to $ \left( \widetilde{\gamma }^{\mu } , \widetilde{A} , \widetilde{\Gamma }_{\mu } \right) $, according to Eqs. \eqref{GrindEQ__5_} and \eqref{GrindEQ__9_}. Recall that the contracted spin connection matrix $ \widetilde{\Gamma }\equiv \widetilde{\gamma }^\mu \widetilde{\Gamma }_\mu= 0$ if the transformed equation is a QRD--0 equation. Then from Eqs. \eqref{GrindEQ__5_} and \eqref{GrindEQ__9_}, we get:
\noindent
\be\gamma ^{\mu } \left( \partial _{\mu } +\Gamma _{\mu } \right) S= 0 \label{GrindEQ__21_},\ee
\noindent whereby any four linearly independent solutions to the massless equation \eqref{GrindEQ__20_} can be used to form the columns of the matrix-valued field $ S $. In this case, an equivalent normal QRD--0 equation can be explicitly and globally constructed.
\noindent \textbf{}
\noindent We can show that any generalized Dirac equation \eqref{GrindEQ__14_}, with very minor conditions imposed on the spacetime metric $ g_{\mu \nu } $, is equivalent to a normal QRD--0 equation \eqref{GrindEQ__19_}, in an open neighborhood of each spacetime point, by applying local similarity transformations of the first and second kind to the Dirac field $ \Psi $. Previously, this so called ``local representation'' of the DFW equation in a curved spacetime has only been discussed in the special case of orthogonal coordinates \cite{3.}, \cite{4.}, \cite{5.}. Here we generalize it to essentially all Dirac equations:\\
\noindent \textbf{THEOREM 1.} Let $ {\bf U} $ be any open subset of a spacetime on which local coordinates are defined. Suppose that the metric components $ g_{\mu \nu } $ for $ \mu , \nu = 0, 1, 2, 3 $ in $ {\bf U} $ satisfy $ g_{00} > 0 $ and the components $ g_{jk} $ for $ j, k = 1, 2, 3 $ form a negative definite $ 3\times 3 $ matrix. Then, for any choice of smooth coefficient fields $ \left( \gamma ^{\mu } , A \right) $ and any choice of covariant derivatives $ D_{\mu } = \partial _{\mu } + \Gamma _{\mu } $ acting on smooth Dirac fields $ \Psi $ defined on $ {\bf U} $, there exists a smooth local similarity transformation $ S $ of the first kind, composed with a smooth local similarity transformation $ T $ of the second kind, taking $ \Psi \to \left( T\circ S \right)^{-1} \Psi $, which transforms the generalized Dirac equation with smooth coefficient fields $ \left( \gamma ^{\mu } , A , \Gamma _{\mu } \right) $ into an equivalent normal QRD--0 equation, in an open neighborhood of each point $ X_{0} \in {\bf U} $.\\
\noindent Thus, we may regard the normal QRD--0 equations \eqref{GrindEQ__19_} as canonical forms for all generalized Dirac equations \eqref{GrindEQ__14_}, in open neighborhoods of each spacetime point.
The proof of Theorem 1 relies heavily on the theory of linear hyperbolic partial differential equations \cite{7.}, \cite{34.}, \cite{35.}. Since it is not particularly enlightening beyond the explicit construction given above for the normal case, we will postpone writing out the full proof of Theorem 1 until the Appendix.
Note that every DFW equation has the normal form \eqref{GrindEQ__17_}. By Theorem 1, every DFW equation is equivalent to a normal QRD--0 equation (called a ``local representation'' of the DFW equation by Schl\"uter, Wietschorke, and Greiner) which generally is not a DFW equation \cite{3.}, \cite{4.}, \cite{5.}.
Conversely, we can show that not every normal QRD--0 equation is equivalent to a DFW equation, so that DFW equations are in fact equivalent to only a proper subset of the possible normal QRD--0 equations. \\
\noindent {\bf EXAMPLE.} Consider the flat metric $ g_{\mu \nu } $ on $ {\bf R}^{4} $ whose line element $ ds $ in rotating cylindrical coordinates $ \left( t , r , \phi , z \right) $ has the form:
\be
ds^{2} = (c \,dt)^{2} - dr^{2} - r^{2} \left( d\phi + \omega \, dt \right)^{2} - dz^{2} .
\label{GrindEQ__22_}
\ee
Using the orthonormal tetrad \cite{28.} as indicated by the parsing of the metric in Eq. (\ref{GrindEQ__22_}), the DFW equation is given by:
\be
\gamma ^{\mu } D_{\mu } \Psi = \gamma ^{\mu } \partial _{\mu } \Psi + \frac{ 1 }{2r} \gamma ^{ 1} \Psi = -i\,\frac{mc}{\hbar }\Psi .
\label{GrindEQ__29_}
\ee
\noindent A simple local similarity transformation of the first kind:
\be
S^{-1} : \Psi \mapsto \Psi '= \sqrt{ r }\, \Psi
\label{GrindEQ__30_}
\ee
which is independent of Planck's constant, time independent, and also is independent of the rotation rate $\omega $, transforms the DFW equation \eqref{GrindEQ__29_} into a Dirac equation of the canonical form \eqref{GrindEQ__19_}.
Note that the canonical Dirac equation \eqref{GrindEQ__19_} does not have the Mashhoon term \cite{9.}, \cite{10.}, whose presence or absence thus depends on the chosen representation of the Dirac field. See also Ryder \cite{11.}.
\section{Whitham's Lagrangian Method --- The Main Theorem}\label{WhithamMethod}
\noindent
\noindent In this section we apply Whitham's Lagrangian method \cite{12.} to derive wave packets for Dirac equations in general curved spacetimes. Whitham's method preserves the symmetries of the Lagrangian, and in particular, the Whitham wave packet equations conserve the probability current. We show in this section that generalized de Broglie relations, as well as COW and Sagnac type terms \cite{10.}, emerge from the Whitham equations after transforming each Dirac equation into an equivalent canonical form. Thus, the normal QRD--0 representations (or canonical forms) of the Dirac equations are the preferred representations to express the generalized de Broglie relations in a curved spacetime.
It is noteworthy that the Whitham approximation places no restriction on Planck's constant $\hbar $. Indeed, the transformation which takes the Dirac equation to its canonical form, is independent of Planck's constant $\hbar $. This is obvious in the example above, as seen in Eq. (\ref{GrindEQ__30_}). In fact, this independence is a general fact that follows from Eqs. (\ref{GrindEQ__74_}) and (\ref{GrindEQ__76_}) used in the proof of Theorem 1 in the Appendix. Thus, the Whitham approximation is not equivalent to the WKB approximation, since the latter does not require any transformation of variables.
\footnote{\
As stated in Section \ref{Canonical Form}, two Dirac equations that are classically equivalent, need not be equivalent with respect to their quantum mechanical energy-momentum operators \cite{25.}, \cite{26.}, \cite{27.}. Clearly, the Whitham approximation also distinguishes them. A striking example is a Dirac equation with a Mashhoon term \cite{10.}, \cite{9.}, \cite{11.}. Applying the Whitham approximation directly to a Dirac equation with a Mashhoon term, without first transforming the Dirac field $\Psi $, does not produce wave packet motion along classical paths.
}
Including both gravitational and electromagnetic external fields, the generalized Lagrangian \eqref{GrindEQ__8_} for the Dirac equation can be written as follows:
\be\begin{array}{l} {L= L \left( \Psi , \partial _{\mu } \Psi , x^{\mu } \right)\quad } \\ {} \\ {\quad =\quad \sqrt{ -g }\, \frac{ i\hbar c }{2}\, {\rm [} \Psi ^{+} A\gamma ^{\mu } \left( D_{\mu } \Psi \right) - \left( D_{\mu } \Psi \right)^{+} A\gamma ^{\mu } \Psi + \frac{ 2 mc }{\hbar } \, i \Psi ^{+} A\Psi {\rm ]},} \end{array}
\label{GrindEQ__33_}\ee
with the covariant derivatives $ D_{\mu } $ defined by:
\be
D_{\mu } = \partial _{\mu } + \Gamma _{\mu } + \frac{ie}{\hbar c} V_{\mu },
\label{GrindEQ__34_}
\ee
where $ V_{\mu } $ are electromagnetic gauge potentials and $ e $ is the electric charge. We can display the interaction terms of the Lagrangian \eqref{GrindEQ__33_} more explicitly by expressing the Lagrangian \eqref{GrindEQ__33_} as a sum of a free and an interaction part as follows:
\be
\begin{array}{l} {L= \sqrt{ -g }\, \frac{ i\hbar c }{2}\, {\rm [} \Psi ^{+} A\gamma ^{\mu } \left( \partial _{\mu } \Psi \right) - \left( \partial _{\mu } \Psi \right)^{ +} A\gamma ^{\mu } \Psi + \frac{ 2 mc }{\hbar } i \Psi ^{+} A\Psi {\rm ]}} \\ {} \\ {\quad \quad \quad \quad \quad \quad \quad \quad \quad + \sqrt{ -g } \, {\rm [} \frac{ i\hbar c }{2} \Psi ^{+} \left( A\Gamma - \Gamma ^{+} A \right) \Psi - \frac{e}{c} J^{\mu } V_{\mu } {\rm ]},} \end{array}
\label{GrindEQ__35_}
\ee
where $ J^{\mu } \equiv c \Psi ^{+} A\gamma ^{\mu } \Psi $ is the probability current, and $ \Gamma \equiv \gamma ^{\mu } \Gamma _{\mu } $ is the contracted spin connection matrix, and also noting that since $ A $ is a Hermitizing matrix for the Dirac matrices $ \gamma ^{\mu } $, we have from Eq. \eqref{GrindEQ__4_}:
\noindent
\be\Gamma ^{+} A= \Gamma _{\mu } ^{+} \gamma ^{\mu +} A= \Gamma _{\mu } ^{+} A\gamma ^{\mu }.
\label{GrindEQ__36_}
\ee
Since we can first transform any Dirac equation into a normal QRD-0 equation (or canonical form) as stated in Theorem 1 of Section \ref{Canonical Form}, we can transform the fields in this Lagrangian so that $ \nabla _{\mu } \left( A\gamma ^{\mu } \right) = 0 $ and $ \Gamma \equiv \gamma ^{\mu } \Gamma _{\mu } = 0 $. Thus, instead of the Lagrangian \eqref{GrindEQ__33_}, we may substitute an equivalent canonical Dirac Lagrangian:
\noindent
\be\begin{array}{l} {L= \sqrt{ -g } \, \frac{ i\hbar c }{2}\, {\rm [} \Psi ^{+} A\gamma ^{\mu } \left( \partial _{\mu } \Psi \right) - \left( \partial _{\mu } \Psi \right)^{ +} A\gamma ^{\mu } \Psi + \frac{ 2 mc }{\hbar } i \Psi ^{+} A\Psi {\rm ]}} \\ {} \\ {\quad \quad \quad \quad \quad \quad \quad \quad \quad - \sqrt{ -g } \, \frac{e}{c} \, J^{\mu } V_{\mu },} \end{array}
\label{GrindEQ__37_}
\ee
Clearly, the normal QRD-0 equation \eqref{GrindEQ__19_} is derived from the Euler-Lagrange equations of the Lagrangian \eqref{GrindEQ__37_} by setting the electromagnetic gauge potentials $ V_{\mu } $ equal to zero and using the normality condition: $ \nabla _{\mu } \left( A\gamma ^{\mu } \right) = 0 $.
For Whitham's method, we set $ \Psi = \chi e^{i\theta } $ where $ \chi = \chi \left(X\right) $ is also a complex wave function with four components, and $ \theta = \theta \left(X\right) $ is a real phase at each point $ X $ in the spacetime. Whitham's method assumes that $ \chi $ is slowly changing compared to the rapidly changing phase $ \theta $, so that we may obtain approximate wave packet solutions to the Dirac equations by neglecting $ \partial _{\mu } \chi $ with respect to $\left( \partial _{\mu } \theta \right) \chi $ in the Lagrangian. Substituting the wave function $ \Psi = \chi e^{i\theta } $ into the Lagrangian \eqref{GrindEQ__37_} and using this approximation, we get:
\noindent
\be L\quad =\quad c \sqrt{ -g } \left[ \left( -\hbar {\kern 1pt} \partial _{\mu } \theta - \frac{e}{c} V_{\mu } \right) \chi ^{+} A\gamma ^{\mu } \chi - mc \chi ^{+} A \chi \right].
\label{GrindEQ__38_}
\ee
In Whitham's method, this Lagrangian governs the wave packet motion. Clearly the Lagrangian (\ref{GrindEQ__38_}) is invariant under the global gauge symmetry $\theta \rightarrow \theta + \tau $, where $\tau $ is a real constant. This leads by Noether's theorem to the conservation of a current. In Section \ref{Current} we will derive explicitly the conservation of the probability current for this Lagrangian.
Thus, our goal in this section is to derive the Euler-Lagrange equations for the Lagrangian (\ref{GrindEQ__38_}), which by change of field variables leads to the following main theorem of this paper.
\noindent \textbf{}
\noindent \textbf{THEOREM 2.} Define a four-vector field $ u^{\mu } $ and a scalar field $ J $, related to the amplitude $ \chi $ and phase $ \theta $ of the wave function $ \Psi = \chi e^{i\theta } $ as follows:
\noindent
\be\begin{array}{l} {u_{\mu } \equiv -\frac{\hbar }{mc} \partial _{\mu } \theta - \frac{e}{ mc^{2} } V_{\mu }, } \\ {} \\ {u^{\mu } \equiv g^{\mu \nu } u_{\nu },} \\ {} \\ {J\equiv c \chi ^{+} A \chi .} \end{array}
\label{GrindEQ__39_} \ee
\noindent
\noindent i) Then the Euler-Lagrange equations for the Whitham Lagrangian \eqref{GrindEQ__38_} imply the following equations for the fields $ u^{\mu } $ and $ J $ in a curved spacetime:
\noindent
\be\begin{array}{l} {g_{\mu \nu } u^{\mu } u^{\nu } = 1,} \\ {} \\ {\nabla _{\mu } \left( Ju^{\mu } \right)= 0,} \\ {} \\ {u_{\mu } = g_{\mu \nu } u^{\nu } ,} \\ {} \\ {\nabla _{\mu } u_{\nu } - \nabla _{\nu } u_{\mu } = -\frac{e}{ mc^{2} } F_{\mu \nu }, } \end{array}
\label{GrindEQ__40_} \ee
\noindent where $ F_{\mu \nu } \equiv \nabla _{\mu } V_{\nu } - \nabla _{\nu } V_{\mu } $ is the electromagnetic field tensor.
\noindent
\noindent ii) The four-vector field $ u^{\mu } $ is a unit velocity field, such that $ J^{\mu } = Ju^{\mu } $ is a conserved probability current.
\noindent
\noindent iii) The integral curves $ x^{\mu } \left(s\right) $ of the four-vector field $ u^{\mu } $, parameterized by arc-length $ s $, are given by the classical equations:
\noindent
\bea
& \frac{\displaystyle dx^{\mu }} {\displaystyle ds} = u^{\mu } , \nonumber \\ & \nonumber \\ & \frac{\displaystyle du^{\mu } }{\displaystyle ds} + \left\{^{\ \mu} _{\nu \, \rho} \right\} u^{\nu } u^{\rho } = \frac{\displaystyle e}{\displaystyle mc^{2} } F^{\mu } _{\ \ \nu } u^{\nu },
\label{GrindEQ__41_} \eea
\noindent along which the scalar field $ J $ satisfies:
\noindent
\be\frac{dJ}{ds} = -J \nabla _{\mu} u^{\mu }.
\label{GrindEQ__42_}\ee
\noindent
\noindent The proof and interpretation of Theorem 2 will occupy the rest of Section 3.
\subsection{Euler-Lagrange Equations for the Wave Packet Lagrangian}
\noindent
\noindent The Euler-Lagrange equations for the amplitude $ \chi = \chi \left(X\right) $ and phase $ \theta = \theta \left(X\right) $ can be derived from the wave packet Lagrangian $ L $ in Eq. \eqref{GrindEQ__38_} as follows. First we have from Eq. \eqref{GrindEQ__38_}:
\noindent
\be\begin{array}{l} {\frac{\partial L}{\partial \chi ^{+} } = \quad c \sqrt{ -g } \left[ \left( -\hbar \, \partial _{\mu } \theta - \frac{e}{c} V_{\mu } \right) A\gamma ^{\mu } \chi - mc A \chi \right],} \\ {} \\ {\frac{\partial L}{\partial \left(\partial _{\mu } \theta \right)} = -\hbar c \sqrt{ -g } \chi ^{+} A\gamma ^{\mu } \chi.} \end{array}
\label{GrindEQ__43_}\ee
\noindent Then, since no derivatives of $ \chi ^{+} $ and only derivatives of $ \theta $ occur in the Lagrangian \eqref{GrindEQ__38_}, we set equal to zero, using Eq. \eqref{GrindEQ__43_}, the following expressions:
\noindent
\be\begin{array}{l} {0= \frac{\delta L}{ \delta \chi ^{+} } =\frac{\partial L}{ \partial \chi ^{+} } } { = c \sqrt{ -g } \left[ \left(-\hbar \partial _{\mu } \theta - \frac{e}{c} V_{\mu } \right) A\gamma ^{\mu } \chi - mc A \chi \right],} \\ {} \\ 0=\frac{\delta L}{\delta \theta } = \partial _{\mu } \left(\frac{\partial L}{\partial \left(\partial _{\mu } \theta \right)} \right)= \partial _{\mu } \left( -\hbar c \sqrt{ -g } \chi ^{+} A\gamma ^{\mu } \chi \right), \end{array} \ee
\noindent which then gives the following Euler-Lagrange equations:
\noindent
\be\begin{array}{l} {\left( -\hbar \, \partial _{\mu } \theta - \frac{e}{c} V_{\mu } \right) A\gamma ^{\mu } \chi } {= mc A\chi}, \\ {} \\
\partial _{\mu } \left( c \sqrt{ -g } \, \chi ^{+} A\gamma ^{\mu } \chi \right) = 0.\end{array} \label{GrindEQ__45_}\ee
\noindent For the wave function $ \Psi = \chi e^{i\theta } $, with phase $ \theta $, the wave covector is $ K_{\mu } \equiv \partial _{\mu } \theta $. Thus $ \omega \equiv -K_{0} = -\partial _{0} \theta $ is the angular frequency of the wave. We define a four-vector field $ p^{\mu } $ as follows:
\noindent
\be p_{\mu } \equiv -\hbar\, \partial _{\mu } \theta = -\hbar K_{\mu }. \label{GrindEQ__46_}
\ee
\noindent It will be shown in Section \ref{Classical-Quantum} that $ p_{\mu } =-P_{\mu } $, where $ P_{\mu } $ are canonical momentum variables.
Eq. \eqref{GrindEQ__46_} expresses the generalized de Broglie relations $ P_{\mu } = \hbar K_{\mu } $ between the canonical momentum variables $ P_{\mu } $ and the wave covector $ K_{\mu } $. We also define a four-vector velocity field $ u^{\mu } $ from the usual classical equation with kinetic and potential terms as follows:
\noindent
\be p^{\mu } = mc u^{\mu } + \frac{e}{c} V^{\mu }.
\label{GrindEQ__47_}\ee
\noindent From Eq. \eqref{GrindEQ__46_} we have $ \partial _{\mu } p_{\nu } = \partial _{\nu } p_{\mu } $. Substituting $ \Psi = \chi e^{i\theta } $ into Eq. \eqref{GrindEQ__2_} we have $ J^{\mu } = c \chi ^{+} A\gamma ^{\mu } \chi $. We denote $ \gamma \left( u \right) \equiv u_{\mu } \gamma ^{\mu } $. Then, Eqs. \eqref{GrindEQ__45_} become:
\bea
\gamma \left( u \right) \chi & = & \chi \label{gamma chi = chi},
\\ \label{D_mu J^mu = 0}
\partial _{\mu } \left( \sqrt{ -g } J^{\mu } \right) & = & 0,
\\ \label{rot p = 0}
\partial _{\mu } p_{\nu } - \partial _{\nu } p_{\mu } & = & 0.
\eea
\noindent
\noindent The first equation \eqref{gamma chi = chi} is an algebraic eigenvalue equation. The second equation \eqref{D_mu J^mu = 0} can be written as the covariant conservation of the probability current, $ \nabla _{\mu } J^{\mu } = 0 $, where $ \nabla _{\mu } $ denotes the Levi-Civita covariant derivatives. Since $ p_{\mu } = -\hbar {\kern 1pt} \partial _{\mu } \theta $, the third equation \eqref{rot p = 0} expresses the equality of mixed partial derivatives of $ \theta $. Furthermore, since the left-hand side of the third equation \eqref{rot p = 0} is an antisymmetric tensor, the partial derivatives $ \partial _{\mu } $ can be replaced by Levi-Civita covariant derivatives $ \nabla _{\mu } $. Thus, Eqs. \eqref{gamma chi = chi}--\eqref{rot p = 0} become the following covariant equations:
\bea\label{gamma chi = chi-bis}
\gamma \left( u \right) \chi & = & \chi, \\
\label{nabla_mu J^mu = 0}
\nabla _{\mu } J^{\mu } & = & 0, \\
\label{rot p = 0-bis}
\nabla _{\mu } p_{\nu } - \nabla _{\nu } p_{\mu } & = & 0.
\eea
We will show below that these equations, taken together, reduce to a set of quasi-linear partial differential equations describing a scalar density $ J = c \chi ^{+} A\chi $ and the four-vector velocity field $ u^{\mu } $, whose integral curves are classical relativistic trajectories. We will further show that certain of these equations give rise to initial conditions, and of the rest, only four equations are independent. First, let us derive a dispersion relation from the algebraic equation \eqref{gamma chi = chi-bis}.
\subsection{Dispersion Relation}
From the algebraic equation \eqref{gamma chi = chi-bis}, and the anticommutation relation of Dirac gamma matrices \eqref{GrindEQ__7_}, we have:
\be\left( g^{\mu \nu } u_{\mu } u_{\nu } \right) \chi = \gamma \left( u \right)^{ 2} \chi = \gamma \left( u \right) \chi = \chi .
\label{GrindEQ__50_}\ee
Equation \eqref{GrindEQ__50_} implies that at any spacetime point where the wave function $ \chi $ is not zero, the four-vector velocity field $ u^{\mu } $ satisfies $ u^{\mu } u_{\mu } = 1 $. From Eq. \eqref{GrindEQ__47_}, this gives the dispersion relation:
\be g^{\mu \nu } \left( p_{\mu } - \frac{e}{c} V_{\mu } \right) \left( p_{\nu } - \frac{e}{c} V_{\nu } \right) - m^{2} c^{2} =0.
\label{GrindEQ__51_}\ee
Since $ p_{0} = \varepsilon /c $, where $ \varepsilon $ is the energy and $ P_{j} = -p_{j} $ for $ j = 1 , 2 , 3 $ are momentum variables, Eq. \eqref{GrindEQ__51_} is a quadratic equation for the energy $ \varepsilon $.
Let us consider the dispersion relation \eqref{GrindEQ__51_} in the absence of the electromagnetic potentials $ V_{\mu } $. We have:
\be
g^{\mu \nu } p_{\mu } p_{\nu } = g^{00} \left( \frac{\varepsilon }{c} \right)^{2} + 2g^{0j} p_{j} \left( \frac{\varepsilon }{c} \right) + g^{jk} p_{j} p_{k} = m^{2} c^{2}. \label{GrindEQ__52_}
\ee
That is,
\be\varepsilon = \frac{c g^{0j} p_{j} \pm c \sqrt{ \left( g^{0j} p_{j} \right)^{ 2} - g^{00} \left( g^{jk} p_{j} p_{k} -m^{2} c^{2} \right) } }{g^{00} }.
\label{GrindEQ__53_}
\ee
\noindent Choosing positive energy $ \varepsilon > 0 $ and setting:
\be g^{00} = 1 - \frac{2\phi }{c^{2} } ,\qquad g^{0j} = \frac{\phi ^{j} }{c} ,\qquad g^{jk} = -\delta ^{jk} - \frac{2\phi ^{jk} }{c^{2} },
\label{GrindEQ__54_}\ee
\noindent where $ \phi $, $ \phi ^{j} $, and $ \phi ^{jk} $ are gravitational potentials, and $ \delta ^{jk} $ is the Kronecker delta, equal to one if $ j = k $ and equal to zero otherwise, then we have for a non-relativistic approximation, i.e., taking the limit of $ \varepsilon - mc^{2} $ as the speed of light $ c $ goes to infinity in Eq. \eqref{GrindEQ__53_}:
\be\begin{array}{l} {\varepsilon - mc^{2} = \frac{cg^{0j} p_{j} + \sqrt{ g^{00} m^{2} c^{4} + \left( cg^{0j} p_{j} \right)^{ 2} - c^{2} g^{00} g^{jk} p_{j} p_{k} } }{g^{00} } - mc^{2} } \\ {} \\ {\quad \quad \quad \quad =\quad \frac{cg^{0j} p_{j} + \sqrt{g^{00} } mc^{2} \sqrt{ 1 + \frac{\left( cg^{0j} p_{j} \right)^{ 2} }{g^{00} m^{2} c^{4} } - \frac{g^{jk} p_{j} p_{k} }{m^{2} c^{2} } } }{g^{00} } - mc^{2} } \\ {} \\ {\quad \quad \quad \quad \approx \quad \frac{cg^{0j} p_{j} + \sqrt{g^{00} } mc^{2} \left( 1 + \frac{\left( cg^{0j} p_{j} \right)^{ 2} }{2g^{00} m^{2} c^{4} } - \frac{g^{jk} p_{j} p_{k} }{2m^{2} c^{2} } \right)}{g^{00} } - mc^{2} } \\ {} \\ {\quad \quad \quad = c\frac{g^{0j} }{g^{00} } p_{j} + \frac{1}{\sqrt{g^{00} } } \left( mc^{2} + \frac{\left( cg^{0j} p_{j} \right)^{ 2} }{2g^{00} mc^{2} } - \frac{g^{jk} p_{j} p_{k} }{2m} \right) - mc^{2} } \\ {} \\ {\quad \quad \quad \quad \approx \quad \frac{1}{2m} \delta ^{jk} p_{j} p_{k} + m\phi + \phi ^{j} p_{j} }. \end{array}
\label{GrindEQ__55_}\ee
\noindent That is, the non-relativistic approximation gives the energy as follows;
\noindent
\be\varepsilon - mc^{2} \quad \approx \quad \frac{1}{2m} \delta ^{jk} p_{j} p_{k} + m\phi + \phi ^{j} p_{j} .
\label{GrindEQ__56_}\ee
\noindent It is straightforward to identify the three energy terms on the right hand side of Eq. \eqref{GrindEQ__56_} as the kinetic energy, a COW potential energy, and a Sagnac potential energy, respectively.
\subsection{Probability Current and Classical Trajectories}\label{Current}
\noindent
\noindent Recall that for Whitham's method, we set the wave function $ \Psi = \chi e^{i\theta } $ where $ \chi = \chi \left(X\right) $ is also a wave function and $ \theta = \theta \left(X\right) $ is a real phase at each point $ X $ in the spacetime. Then the probability current $ J^{\mu } $ and scalar field $J$ are given by:
\noindent
\bea
J^{\mu } \equiv c \Psi ^{+} A\gamma ^{\mu } \Psi = c \chi ^{+} A\gamma ^{\mu } \chi, \nonumber \\
J \equiv c \Psi ^{+} A\Psi = c \chi ^{+} A\chi.
\label{GrindEQ__57_}
\eea
\noindent From Eq. (\ref{gamma chi = chi-bis}), $\chi $ is a solution of the equation $\gamma \left( u \right) \chi = \chi$, where $u^\mu $ is a unit four-vector field satisfying $u^\mu u_\mu =1$ and where $\gamma (u)\equiv u_\mu\gamma ^\mu $. From the anticommutation relation of the Dirac gamma matrices in Eq. (\ref{GrindEQ__3_}), we have:
\be\label{u vs gamma(u)}
u^\mu =g^{\mu \nu }u_\nu =\frac{1}{2}\left(\gamma ^\mu\gamma ^\nu+\gamma ^\nu\gamma ^\mu \right )u_\nu =\frac{1}{2}\left [\gamma ^\mu\gamma (u)+\gamma (u)\gamma ^\mu \right ].
\ee
Moreover, using again the definition $\gamma (u)\equiv u_\mu\gamma ^\mu $, it follows easily from the properties of the hermitizing matrix $A$ [Eq. (\ref{GrindEQ__4_})] that $\chi ^+ A \gamma (u)=\left[ \gamma (u)\chi \right]^+A$. We get thus from Eqs. (\ref{gamma chi = chi-bis}), (\ref{GrindEQ__57_}) and (\ref{u vs gamma(u)}):
\bea
J u^\mu & = & \frac{c}{2}\,\chi ^+A \left[\gamma ^\mu\gamma (u)+\gamma (u)\gamma ^\mu \right ]\chi \nonumber \\
& = & \frac{c}{2}\,\chi ^+A\gamma ^\mu \left[\gamma (u)\chi \right ] + \frac{c}{2}\, \left[\gamma (u)\chi \right ]^+ A\gamma ^\mu\chi \nonumber \\
& = & \frac{c}{2}\,\chi ^+ A\gamma ^\mu\chi + \frac{c}{2}\,\chi ^+A\gamma ^\mu\chi \nonumber \\
& = & J^\mu .
\eea
That is, $J^\mu =J u^\mu $. Using Eqs. \eqref{GrindEQ__47_} and \eqref{GrindEQ__51_} together with this result, Eqs. \eqref{gamma chi = chi-bis}--\eqref{rot p = 0-bis} can be written as:
\bea\label{u normed}
g_{\mu \nu } u^{\mu } u^{\nu } & = & 1,\\
\label{u bemol}
u_{\mu } & = & g_{\mu \nu } u^{\nu }, \\
\label{nabla Ju=0}
\nabla _{\mu } \left( Ju^{\mu } \right) & = & 0,\\
\label{rot u = C F}
\nabla _{\mu } u_{\nu } - \nabla _{\nu } u_{\mu } & = & -\frac{e}{ mc^{2} } F_{\mu \nu },
\eea
where $ F_{\mu \nu } \equiv \nabla _{\mu } V_{\nu } - \nabla _{\nu } V_{\mu } $ is the electromagnetic field tensor. Multiply by $ u^{\nu } $ and contract the index $ \nu $ on both sides of Eq. \eqref{rot u = C F}. Then, using Eq. \eqref{u normed} to set $ u^{\nu } \left( \nabla _{\mu } u_{\nu } \right) = 0 $, and finally raising the index $ \mu $, we get:
\noindent
\be\left( u^{\nu } \nabla _{\nu } \right) u^{\mu } = \frac{e}{ mc^{2} } F^{\mu } _{\ \ \nu } u^{\nu }.
\label{GrindEQ__59_}\ee
\noindent Consider the integral curves $ x^{\mu } \left(s\right) $ of the four-vector velocity field $ u^{\mu } \left(X\right) $. Since the four-vector velocity field consists of unit vectors by Eq. \eqref{u normed}, the integral curves $ x^{\mu } \left(s\right) $ are parameterized by arc-length $ s $. That is, from Eqs. \eqref{u normed} and \eqref{GrindEQ__59_}:
\noindent
\bea
& \frac{\displaystyle dx^{\mu }} {\displaystyle ds} = u^{\mu } , \nonumber \\ & \nonumber \\ & \frac{\displaystyle du^{\mu } }{\displaystyle ds} + \left\{^{\ \mu} _{\nu \, \rho} \right\} u^{\nu } u^{\rho } = \frac{\displaystyle e}{\displaystyle mc^{2} } F^{\mu } _{\ \ \nu } u^{\nu }.
\label{GrindEQ__60_} \eea
\noindent Note that Eqs. \eqref{GrindEQ__60_} are precisely the classical relativistic equations of a particle of mass $ m $ and electric charge $ e $ in a gravitational field $ g_{\mu \nu } $ in the presence of an electromagnetic field $ F_{\mu \nu } $. Thus, the integral curves of the four-vector velocity field $ u^{\mu } \left(X\right) $, describing the motion of wave packets, coincide with the trajectories of classical relativistic particles. Note that in the absence of the electromagnetic field $ F_{\mu \nu } $, the classical trajectories \eqref{GrindEQ__60_} are geodesics of the spacetime.
Finally, we note from Eq. (\ref{nabla Ju=0}) that along the integral curves of $ u^{\mu } \left(X\right) $ we have:
\be\frac{dJ}{ds} = -J \nabla _{\mu } u^{\mu }.
\label{GrindEQ__61_}
\ee
This completes the proof of Theorem 2.
Note that the wave packet equations \eqref{GrindEQ__40_} describe a certain congruence of classical trajectories \eqref{GrindEQ__41_} together with a scalar density $ J $ that on each classical trajectory in the congruence satisfies Eq. \eqref{GrindEQ__42_}. This congruence satisfies certain initial conditions on a spatial submanifold $ {\bf M} $ discussed in the next subsection.
\subsection{Mathematical Structure of the Wave Packet Equations}
Eq. \eqref{u normed}--\eqref{rot u = C F} can be written as follows:
\bea\label{u normed-bis}
g_{\mu \nu } u^{\mu } u^{\nu } & = & 1,\\
\label{u bemol-bis}
u_{\mu } & = & g_{\mu \nu } u^{\nu }, \\
\label{nabla Ju=0-2}
\frac{\partial }{ \partial x^{\mu } } \left( \sqrt{ -g } Ju^{\mu } \right) & = & 0,\\
\label{rot u = C F-0j}
\frac{\partial u_{0} }{\partial x^{j} } - \frac{\partial u_{j} }{\partial x^{0} } & = & \frac{e}{ mc^{2} } F_{0j} \qquad (j = 1 , 2 , 3), \\
\label{rot u = C F-jk}
\frac{\partial u_{j} }{\partial x^{k} } - \frac{\partial u_{k} }{\partial x^{j} } & = & \frac{e}{ mc^{2} } F_{jk}\qquad (j, k = 1 , 2 , 3).
\eea
Eqs. \eqref{u normed-bis} and \eqref{u bemol-bis} allow us to solve algebraically for $ u^{0} $, $ u_{0} $, and $ u_{j} $ in terms of $ u^{j} $ where $ j = 1 , 2 , 3 $. We will show below that Eq. \eqref{rot u = C F-jk} gives merely a set of initial conditions. Thus, we are left with only four real quasi-linear partial differential equations contained in Eqs. \eqref{nabla Ju=0-2} and \eqref{rot u = C F-0j}, for the four real fields $ u^{j} $ and $ J $. Indeed, from Eq. \eqref{GrindEQ__47_}, we have that \eqref{rot u = C F-0j} and \eqref{rot u = C F-jk} are equivalent to:
\bea\label{rot p 0j = 0}
\frac{\partial p_{j} }{\partial x^{0} } & =& \frac{\partial p_{0} }{\partial x^{j} }, \\
\label{rot p jk = 0}
\frac{\partial p_{j} }{\partial x^{k} } & = & \frac{\partial p_{k} }{\partial x^{j} }.
\eea
It follows from Eq. \eqref{rot p 0j = 0} that Eq. \eqref{rot p jk = 0} remains true for all time if and only if it is true at an initial time. That is because from Eq. \eqref{rot p 0j = 0}, we derive:
\be\frac{\partial }{\partial x^{0} } \left(\frac{\partial p_{j} }{\partial x^{k} } \right) = \frac{\partial ^{2} p_{0} }{\partial x^{j} \partial x^{k} } = \frac{\partial }{\partial x^{0} } \left(\frac{\partial p_{k} }{\partial x^{j} } \right).
\label{GrindEQ__65_}\ee
\noindent It follows that Eq. \eqref{rot u = C F-jk} gives a set of initial conditions. Thus, provided that the initial conditions \eqref{rot u = C F-jk} are satisfied at any initial time --- i.e., on a spatial submanifold $ {\bf M} $--- the wave packet equations \eqref{u normed-bis}--\eqref{rot u = C F-jk} give rise to well-defined solutions. As previously stated these solutions comprise a congruence of classical trajectories together with a scalar density.
\subsection{Gordon Decomposition for Dirac Equations in Canonical Form}
In the WKB approximation of the DFW equation in a curved spacetime, classical trajectories are derived from the Gordon decomposition of the probability current $J^\mu = J^\mu _c + J^\mu _s$ into a convection current $J^\mu _c$ and a spin current $J^\mu _s$ \cite{Audretsch1981A}. In this subsection, we will prove the existence of the Gordon decomposition for all normal Dirac equations, noting that both DFW and canonical equations are normal. We will further show that in the Whitham approximation, the spin current $J^\mu _s$ vanishes, which explains the negligible effect of spin on wave packet solutions. More specifically, we will show that wave packet solutions of the form $\Psi =\chi e^{i\theta }$, where $\chi $ is slowly changing compared to a rapidly changing phase $\theta $, can only exist if the spin current $J^\mu_s$ is negligible. As discussed above, this definition of wave packet is independent of the size of Planck's constant $\hbar $. Indeed, no assumption will be made in this subsection regarding the size of Planck's constant $\hbar $ or the speed of light $c$, both of which we will set equal to one, $\hbar =c=1$.
For a normal Dirac equation, define the probability current $J^\mu$, the spin current $J^\mu_s$, and the convection current $J^\mu_c$ as follows:
\bea
J^{\mu } & \equiv & \Psi ^{+} A\gamma ^{\mu } \Psi, \nonumber \\
J^{\mu }_s & \equiv & \frac{1}{2m} \left[\Psi ^{+} A\sigma ^{\mu \nu }(D_\nu \Psi) + (D_\nu \Psi)^+A\sigma ^{\mu \nu }\Psi \right ], \nonumber \\
J^{\mu }_c & \equiv & \frac{i}{2m} g^{\mu \nu }\left[\Psi ^{+} A(D_\nu \Psi) - (D_\nu \Psi)^+A\Psi \right ],
\label{Gordon}
\eea
where $\sigma ^{\mu \nu }\equiv \frac{i}{2} \left (\gamma ^\mu \gamma ^\nu -\gamma ^\nu \gamma ^\mu \right )$ are the Dirac spin matrices. Substituting the normal Dirac equation (\ref{GrindEQ__17_}) written in the following form:
\be\label{Dirac Psi=}
\Psi = \frac{i}{m}\gamma ^\mu D_\mu \Psi
\ee
into the formula for $J^\mu$ in Eq. (\ref{Gordon}), using the gamma matrix formula $\gamma ^\mu \gamma ^\nu= g^{\mu \nu }-i\sigma ^{\mu \nu }$, and noting the asymmetry $\sigma ^{\mu \nu }=-\sigma ^{\nu \mu }$, we get from Eqs. (\ref{Gordon}) and (\ref{Dirac Psi=}):
\bea
J^{\mu } & \equiv & \Psi ^{+} A\gamma ^{\mu } \Psi = \frac{i}{2m} \left[\Psi ^{+} A\gamma ^\mu \gamma ^\nu(D_\nu \Psi) - (D_\nu \Psi)^+A\gamma ^\nu \gamma ^\mu\Psi \right ]\nonumber \\
& = & \frac{i}{2m} \left[\Psi ^{+} A\left(g^{\mu \nu }-i\sigma ^{\mu \nu }\right)(D_\nu \Psi) - (D_\nu \Psi)^+A\left(g^{\mu \nu }+i\sigma ^{\mu \nu }\right)\Psi \right ] \nonumber \\
& = & J^\mu_c + J^\mu_s.
\label{Gordon2}
\eea
That is, $J^\mu = J^\mu_c + J^\mu_s$, which proves the Gordon decomposition for normal Dirac equations.
Note that the probability current $J^\mu = J^\mu_c + J^\mu_s$ is covariantly conserved. Consequently, the probability density current $\sqrt{-g} J^\mu$ is conserved. In general, the currents $J^\mu_c$ and $J^\mu_s$ are not separately covariantly conserved, unless the coefficient fields $(\gamma ^\mu ,A)$ are covariantly constant. However, it is worthy to note that for Dirac equations transformed into canonical form, we may replace the covariant derivatives $D_\mu$, including electromagnetic field potentials $V_\mu $ as in Eq. (\ref{GrindEQ__34_}), with the Levi-Civita covariant derivatives $\nabla _\mu +ieV_\mu $.
\footnote{\
Note that the probability current $J^\mu =\Psi ^+A\gamma ^\mu \Psi $ is invariant under local similarity transformations $S$ defined in Eq. (\ref{GrindEQ__5_}), so that when transforming a Dirac equation into canonical form, it is only $J^\mu _s$ and $J^\mu _c$ which change their form.
}
Then the spin current $J^\mu_s$ and the convection current $J^\mu_c$ become:
\bea
J^{\mu }_s & = & \frac{1}{2m} \left[\Psi ^{+} A\sigma ^{\mu \nu }(\partial _\nu \Psi) + (\partial _\nu \Psi)^+A\sigma ^{\mu \nu }\Psi \right ], \nonumber \\
J^{\mu }_c & = & \frac{i}{2m} g^{\mu \nu }\left[\Psi ^{+} A(\partial _\nu +ieV_\nu) \Psi - \left((\partial _\nu +ieV_\nu ) \Psi\right )^+A\Psi \right ].
\label{Gordon3}
\eea
Note that the convection current $J^{\mu }_c$ in Eq. (\ref{Gordon3}) closely resembles the spin zero current of the Klein-Gordon equation in the presence of both gravitational and electromagnetic external fields \cite{28.}, whereas the spin motion resides in the spin current $J^{\mu }_s$ \cite{Audretsch1981A}. Recall that we are using Whitham's approximation, for which we set $\Psi =\chi e^{i\theta }$ and neglect $\partial _\mu \chi $ with respect to $(\partial _\mu \theta ) \chi $ [see before Eq. (\ref{GrindEQ__38_})]. Using the definitions in Eq. (\ref{GrindEQ__39_}), this gives us
\be\label{Whitham to D Psi}
(\partial _\nu +ieV_\nu) \Psi \approx i (\partial _\nu \theta +eV_\nu) \Psi=-imu_\nu \Psi .
\ee
Then we use the definitions in Eq. (\ref{GrindEQ__39_}) together with Eqs. (\ref{Gordon3}) and (\ref{Whitham to D Psi}) to obtain the following formulas for the spin current and the convection current:
\bea
J^{\mu }_s & \approx & -\frac{i}{2} \left[\Psi ^{+} A\sigma ^{\mu \nu }u_\nu \Psi - \Psi^+A\sigma ^{\mu \nu }u_\nu \Psi \right ]=0, \nonumber \\
J^{\mu }_c & \approx & \frac{1}{2} g^{\mu \nu }u_\nu \left[\Psi ^{+} A\Psi + \Psi ^{+} A\Psi \right ]=Ju^\mu .
\label{Gordon4}
\eea
Thus, in the Whitham approximation, the spin current $J^{\mu }_s$ vanishes and the convection current $J^\mu_c = Ju^\mu$. Therefore, the probability current $J^\mu$ equals the convection current $J^\mu_c$.
\section{The Classical-Quantum Correspondence}\label{Classical-Quantum}
We proved in Theorem 2 that the solutions of the Whitham approximation to the Dirac equation, Eq. \eqref{GrindEQ__40_}, consist of a four-velocity vector field $ u^{\mu }$ whose integral curves are classical trajectories, and a scalar field $ J $ representing a conserved particle density $-$ see Eqs. \eqref{GrindEQ__41_} and \eqref{GrindEQ__42_}. From Whitham's approximation of the Dirac equation, we have derived the dispersion relation \eqref{GrindEQ__51_}, the motion of wave packets along classical trajectories \eqref{GrindEQ__41_}, conservation of the probability current \eqref{GrindEQ__42_}, as well as the generalized de Broglie relations $ P_{\mu } = \hbar K_{\mu } $ in Eqs. \eqref{GrindEQ__46_} and \eqref{GrindEQ__47_}.
We will conclude this paper by summarizing results from a previous analysis of the ``classical-quantum correspondence'' based on the dispersion relation \eqref{GrindEQ__51_} alone, which can be applied to the Dirac equation \cite{15.}. We will interpret the fact that the integral curves of the four-velocity field $ u^{\mu } $ are classical trajectories as a ``geometrical optics limit'' of the Dirac equation. However, as we have seen, unlike the WKB limit, this ``geometrical optics limit'' is one that places no restriction on Planck's constant. Whitham's Lagrangian method, applied to the Dirac Lagrangian (\ref{GrindEQ__33_}) transformed into equivalent canonical form (\ref{GrindEQ__37_}), has also many advantages over using the dispersion relation \eqref{GrindEQ__51_} alone as the starting point for a wave packet approximation \cite{12.}. Whitham's method preserves the conservation laws inherent in the starting Lagrangian \eqref{GrindEQ__37_}, and in particular, the Whitham wave packet equations conserve the probability current $ J^{\mu } = Ju^{\mu } $.
To any linear partial differential equation for scalar wave functions $ \Psi $, of the form:
\noindent
\be a\left( X \right) \Psi + \sum _{n = 1}^{d} a^{^{\mu _{_{1} } \cdots \cdots \mu _{n} } } \left( X \right) \frac{\partial ^{n} \Psi }{\partial x^{\mu _{ 1} } ...... \partial x^{\mu _{ n} } } = 0\label{GrindEQ__66_}\ee
\noindent (summing over coordinate indices $ \mu _{r} = 0, 1, 2, 3 $ for $ r = 1, 2, 3, \cdots , n $ and over the index $ n = 1, 2, 3, \cdots , d $) where the coefficient fields $ a\left( X \right) $ and $ a^{^{\mu _{_{1} } \cdots \cdots \mu _{n} } } \left( X \right) $ depend on the spacetime point $ X $$-$ one may associate its dispersion polynomial $ \Pi _{X} \left( K \right)$. That is to say, a polynomial function of covector fields $ K_{\mu } $ at each fixed spacetime point $ X $ is given by:
\be
\Pi _{X} \left( K \right)= a\left( X \right) + \sum _{n = 1}^{d} i^{n} a^{^{\mu _{{1} } \cdots \cdots \mu _{ n} } } \left( X \right) K_{\mu _1 } ...... K_{\mu _n}.
\label{GrindEQ__67_}
\ee
The dispersion relation is thereby obtained from the polynomial equation $ \Pi _{X} \left( K \right)= 0 $ at each fixed spacetime point $ X $ by solving for the time component $ K_{0} $. Applications of this one-to-one correspondence are discussed in Ref. \cite{15.}.
This applies also if the wave functions $ \Psi $ have $ m $ components and the coefficient fields $ a\left( X \right) $ and $ a^{^{\mu _{_{1} } \cdots \cdots \mu _{n} } } \left( X \right) $ are $ m\times m $ matrices, as is the case for the Dirac equation \cite{21.}, \cite{36.}. Note that in the matrix case, the dispersion relation is obtained from the scalar polynomial equation $ \det \Pi _{X} \left( K \right)= 0 $. Consider a dispersion polynomial \eqref{GrindEQ__67_} where $ a\left( X \right) $ and $ a^{^{\mu _{_{1} } \cdots \cdots \mu _{n} } } \left( X \right) $ are $ m\times m $ matrices. By solving $ \det \Pi _{X} \left( K \right)= 0 $ for the component $ K_{0} $ we get a dispersion relation:
\noindent
\be \omega = W\left( {\bf k} , {\bf x} {\bf ,} t \right)
\label{GrindEQ__68_}\ee
\noindent expressing the angular frequency $ \omega \equiv -c\,K_{0} $ as a function of the spatial wave covector $ {\bf k} = \left( {\it K}_{{\rm 1}} , K_{{\rm 2}} , K_{{\rm 3}} \right) $, the spatial coordinates $ {\bf x} = \left( x^{1} , x^{2} , x^{3} \right) $, and time $ t $, together with the auxiliary equations \cite{12.}, \cite{15.}:
\noindent
\bea \frac{\displaystyle \partial K_{j} }{\displaystyle \partial t } + \frac{\displaystyle \partial \omega }{\displaystyle \partial x^{j} } = 0, \nonumber \\ \nonumber \\ \frac{\displaystyle \partial K_{j} }{\displaystyle \partial x^{k} } - \frac{\displaystyle \partial K_{k} }{\displaystyle \partial x^{j} } = 0,
\label{GrindEQ__69_}\eea
\noindent for $ j, k = 1 , 2 , 3 $. In general, none or multiple such dispersion relations \eqref{GrindEQ__68_} can be derived as distinct real roots of the polynomial equation $ \det \Pi _{X} \left( \omega , {\bf k} \right)= 0 $ when solving for $ \omega $. Assuming at least one real root, let us choose one of them to be $ \omega = W\left( {\bf k} , {\bf x} {\bf ,} {\rm t} \right) $. Then from Eqs. \eqref{GrindEQ__68_} and \eqref{GrindEQ__69_} one derives the Hamiltonian system as in Ref. \cite{15.}, Sect. 2.2:
\be
\frac{d K_j}{d t}= -\frac{\partial W}{\partial x^j},\qquad \frac{d x^j}{d t}= \frac{\partial W}{\partial K_j}.
\label{GrindEQ__70_}\ee
\noindent Noting in Eq. \eqref{GrindEQ__68_} that $ W \equiv -c\,K_{0} $, and recalling the generalized de Broglie relations $ P_{\mu } \equiv -p_{\mu } = \hbar K_{\mu } $ derived from wave packet motion in Eqs. \eqref{GrindEQ__46_} and \eqref{GrindEQ__47_}, we see that Eq. \eqref{GrindEQ__70_} leads us to define the Hamiltonian $ H \equiv \hbar W $ and the momentum variables $ P_{j} \equiv \hbar K_{j} $ for $ j = 1 , 2 , 3 $, whereby a system of classical point particle trajectories emerges as follows: Indeed, solving the dispersion equation \eqref{GrindEQ__51_} for the energy $ \varepsilon = H\left( {\bf p} , {\bf x} {\bf ,} {\rm t} \right) $ where $ {\bf p} = \left( P_{{\rm 1}} , P_{{\rm 2}} , P_{{\rm 3}} \right) $ is equivalent to solving it for the angular frequency $ \omega = W\left( {\bf k} , {\bf x} , t \right) $ as in Eq. \eqref{GrindEQ__68_}. Then Eq. \eqref{GrindEQ__70_} is equivalent to:
\noindent
\be\frac{d P_j}{d t}= -\frac{\partial H}{\partial x^j},\qquad \frac{d x^j}{d t}= \frac{\partial H}{\partial P_j}.
\label{GrindEQ__71_}\ee
\noindent One can show that the trajectories associated with the Hamiltonian $ H $, that is, the solution trajectories of the Hamiltonian equations \eqref{GrindEQ__71_}, are identical to the solution trajectories of the Euler-Lagrange equations deduced from the well known Lagrangian for classical point particles in a background of electromagnetic and gravitational fields, which is given by \cite{37.}, \cite{38.}:
\noindent
\be\ell = -mc \sqrt{ g_{\mu \nu } \frac{dx^{\mu } }{d\xi } \frac{dx^{\nu } }{d\xi } } - \frac{e}{c} V_{\mu } \frac{dx^{\mu } }{d\xi }, \label{GrindEQ__72_}\ee
\noindent where $ \xi $ is an arbitrary parameter for the classical trajectory $ x^{\mu } \left( \xi \right) $. To show this one first applies an inverse Legendre transformation
\footnote{ The Legendre transformation is its own inverse \cite{38.}, pages 563-565. Thus, an inverse Legendre transformation is also a Legendre transformation.
}
to the Hamiltonian $ H = H\left( {\bf p} , {\bf x} {\bf ,} {\rm t} \right) $ to obtain a traditional Lagrangian $ L \left( {\bf x} , \frac{d{\bf x}}{dt} , t \right) $ and then, as in Ref. \cite{38.}, pages 267-271, one generalizes the trajectory parameter to be an arbitrary parameter $ \xi $, instead of the coordinate time $ t $. It is straightforward to check that the equations for the classical trajectories \eqref{GrindEQ__41_} are the Euler$-$Lagrange equations for the Lagrangian \eqref{GrindEQ__72_}. Thus, the dispersion relation \eqref{GrindEQ__68_} and the auxiliary equations \eqref{GrindEQ__69_} give rise to the classical point particle trajectory equations \eqref{GrindEQ__41_}, as is also the case for the integral curves of the Whitham equations \eqref{GrindEQ__40_}.
Similar to the previous analysis of the classical-quantum correspondence (\cite{15.}, Sect. 2.3), it is in the ``geometrical optics limit'' that the solutions of each Dirac equation transformed into equivalent canonical form satisfy the dispersion equation \eqref{GrindEQ__51_}. Indeed, Whitham's approximation: $ \partial _{\mu } \chi \ll \left( \partial _{\mu } \theta \right) \chi $, which we applied in Section \ref{WhithamMethod} to the Dirac Lagrangian \eqref{GrindEQ__37_}, is one way of defining precisely this limit. However, the classical Hamiltonian equations \eqref{GrindEQ__71_}, which are based solely on the dispersion relation, give only part of the Whitham equations \eqref{GrindEQ__40_}. In addition to providing equations \eqref{GrindEQ__41_} equivalent to the Hamiltonian equations \eqref{GrindEQ__71_}, the Whitham equations \eqref{GrindEQ__40_} preserve the symmetries of the Dirac Lagrangian \eqref{GrindEQ__37_}, and provide for the conservation of the probability current $ J^{\mu } = Ju^{\mu } $, which is a property inherited from the exact Dirac equations in a curved spacetime.
|
\section{Introduction}
The Wilsonian perspective is a powerful guiding principle in
constructing theories with the given field content and symmetries.
It tells that one should include in the action all terms that can be
constructed from the fields and are compatible with the symmetries
of the theory. In the context of first order gravity we have to do
with two fields, tetrad $e^a$ and connection $\omega^{ab}$, and two symmetries, local Lorentz invariance and spacetime diffeomorphisms. If we implement
the diffeomorphism invariance, assuming that the action of gravity
is written as a four form polynomial constructed from the tetrad and the connection, the
list of possible terms turns out to be rather short and includes
\begin{itemize}
\item Palatini Lagrangian
\begin{equation}\label{1}
\mathcal{L}_P= R^{ab}\wedge e^{c}\wedge e^{d}\,\epsilon_{abcd}\, ,
\end{equation}
\item Cosmological term
\begin{equation}\label{2}
\mathcal{L}_{C}=e^{a}\wedge e^{b}\wedge e^{c}\wedge e^{d}\, \epsilon_{abcd}\, ,
\end{equation}
\item Holst term \cite{Holst:1995pc}
\begin{equation}\label{3}
H_4 = R^{ab}\wedge e_a \wedge e_b\, ,
\end{equation}
\item Pontryagin, Euler and Nieh-Yan topological terms
\begin{eqnarray}
P_4 &=& R^{ab}\wedge R_{ab}\, ,\nonumber\\
\label{4}E_4 &=& R^{ab}\wedge R^{cd} \,\epsilon_{abcd}\, ,\\
NY_4&=& T^a \wedge T_a - R^{ab}\wedge e_a \wedge e_b\, ,\nonumber
\end{eqnarray}
\end{itemize}
where $R^{ab}$ is the curvature of $\omega^{ab}$ and $T^a$ is torsion.
Each of these terms comes with its own coupling constant. One could
ask if there is an additional principle that could be used to reduce
the number of independent parameters of the theory. As it turns out,
this can be achieved in the framework of the formulation of gravity
as a constrained BF theory.
This approach has its roots in MacDowell-Mansouri
\cite{MacDowell:1977jt}, \cite{Stelle:1979aj} and Plebanski
\cite{Plebanski:1977zz,Capovilla:1991qb,Capovilla:1991kx} theories
and
was developed in the series of papers \cite{Starodubtsev:2003xq,Smolin:2003qu,Freidel:2005ak,Wise:2006sm}. In
this formulation we have the anti-de Sitter
algebra $so(2,3)$-valued\footnote{The de Sitter case $so(1,4)$ can
be constructed analogously. Here we use the anti-de Sitter algebra
because it leads to the asymptotically anti-de Sitter spacetimes.}
connection $A^{IJ}$, with $I,J=0, \ldots, 4$, which can be decomposed
into Lorentz connection $\omega^{ab}$ and the tetrad (soldering)
one-form $e^a$ ($a,b = 0, \dots,3$) as follows
\begin{equation}\label{5}
A^{ab} = \omega^{ab}\,,\qquad A^{a4}=\frac{1}{\ell}\, e^a\,.
\end{equation}
Here $\ell$ is a length scale necessary for dimensional reason since
the tetrad is dimensionless. As we will see this scale is naturally
associated with the cosmological constant. The components of the
curvature of connection $A^{IJ}$ are related to the curvature of
Lorentz connection $\omega$
\begin{equation}\label{6}
F^{ab}(A)= R^{ab}(\omega)+ \frac1{\ell^2}\, e^a\wedge e^b
\end{equation}
and the torsion
\begin{equation}\label{7}
F^{a4} =\frac1{\ell}\left(d e^{a} +
\omega^a{}_b\wedge e^b \right)=
\frac1{\ell}\, T^a\, .
\end{equation}
With the help of the second ingredient, the $so(2,3)$ Lie algebra
valued two-from field $B^{IJ}$ one can write down the action of the
theory as follows
\begin{equation}\label{8}
16\pi \, S(A,B)= \int F^{IJ}\wedge B_{IJ} -
\frac{\beta}{2} B^{IJ}\wedge B_{IJ} - \frac{\alpha}{4}\epsilon^{IJKL4} B_{IJ}\wedge
B_{KL}
\end{equation}
After solving $B$ field equations we find
\begin{equation}\label{9}
B^{a4} = \frac1\beta\, F^{a4},
\quad B^{ab}
=\frac{1}{2(\alpha^2+\beta^2)}( \beta \delta^{ab}_{cd}-\alpha \epsilon^{ab}{}_{cd})\, F^{cd}\, .
\end{equation}
Before substituting this result back to the action (\ref{8}) let us
provide the expressions for dimensionless coupling constants
$\alpha$ and $\beta$ and the scale $\ell$ in terms of the physical
coupling constants, Newton's constant $G$, a negative cosmological
constant $\Lambda$, and the Immirzi parameter $\gamma$
\cite{Immirzi:1996di}
\begin{equation}\label{10}
\alpha =
\frac{G\Lambda}{3}\frac{1}{(1+\gamma^2)}, \quad \beta =
\frac{G\Lambda}{3}\frac{\gamma}{(1+\gamma^2)} , \quad
\gamma=\frac{\beta}{\alpha}\, , \quad \Lambda = -\frac{3}{\ell^2}\,
.
\end{equation}
Substituting (\ref{9}) and (\ref{10}) to the action (\ref{8}) gives
\begin{eqnarray}
32\pi &G&\, S=\int\left(
R^{ab}\wedge e^{c}\wedge e^{d}
+\frac{1}{2\ell^2} e^{a}\wedge e^{b}\wedge e^{c}\wedge
e^{d}\right) \epsilon_{abcd}\nonumber\\
& &\qquad+\frac{2}{\gamma}\int R^{ab}\wedge e_{a}\wedge e_{b}
\label{11}\\
&&\qquad+\frac{\ell^2}{2}\,
E_4-\ell^2\gamma\, P_4+\frac{2(\gamma^2+1)}{\gamma}\, NY_4\, .\nonumber
\end{eqnarray}
The first line in (\ref{11}) is the standard first order form of the
general relativity action with the cosmological constant. The third
line contains the combination of topological invariants and
therefore can be written as a total derivative (see
\cite{Durka:2010zx}). The middle term is called the Holst term
\cite{Holst:1995pc}. Although it is not a total derivative, by the virtue of the Bianchi identity, it does not influence equations of motion when
torsion vanishes.
The first order action above can be also written down in a compact
form
\begin{equation}\label{our}
S(\omega,e)=\frac{1}{16\pi} \int_{M}\left( \frac{1}{4}M^{abcd} F_{ab}\wedge F_{cd}-\frac{1}{\beta\ell^2} \,T^a \wedge T_a\right)\,
\end{equation}
with
\begin{equation}\label{M}
M^{ab}{}_{cd}=\frac{\alpha}{(\alpha^2+\beta^2)}( \gamma\, \delta^{ab}_{cd}
-\epsilon^{ab}_{\;\;cd}) \equiv -\frac{\ell^2}{G}( \gamma\, \delta^{ab}_{cd}-\epsilon^{ab}_{\;\;cd})\,.
\end{equation}
The field equations following from (\ref{our}) read
\begin{eqnarray}
K^{abcd}\, F_{ab}\wedge e_c&=&0 \label{12}\\
\frac{1}{\ell^2}{\cal D}^\omega \left(K^{abcd}(e_a \wedge
e_b)\right)&=&0\label{13}
\end{eqnarray}
where the operator $K$ has the form
\begin{equation}\label{14}
K^{ab}{}_{cd}\equiv -\frac{\ell^2}{G}\left(\frac{1}{\gamma}\,
\delta^{ab}{}_{cd}+\epsilon^{ab}{}_{cd}\right)
\end{equation}
and we have introduced AdS curvature $F$
\begin{equation}\label{15}
F^{ab}=(R^{ab}+\frac{1}{\ell^2}e^a\wedge e^b)\,.
\end{equation}
Later we will make use of the fact that this curvature vanishes for anti-de
Sitter spacetime.
It follows from equation (\ref{13}) that
torsion $T^a={\cal D}^\omega e^a$ vanishes (one has to assume that
$\gamma^2 \neq -1$) and thus the field equations (\ref{14}) are
Einstein equations with a negative cosmological constant in the first
order form.\newline
\indent The Immirzi parameter, being the coupling constant associated with
the Holst term, is a mysterious beast. It was first introduced by
Barbero \cite{Barbero:1994ap} in the context of Ashtekar variables,
parametrizing a family of canonical transformations on the gravity
phase space and inequivalent quantizations. It was soon realized
that $\gamma$ is explicitly present in the Loop Quantum Gravity
formula for area spectrum \cite{Immirzi:1996di}, \cite{Rovelli:1994ge}. As a consequence, Immirzi parameter is also
present in the formula for black hole entropy calculated by counting LQG
microstates of isolated horizon \cite{Ashtekar:2000eq,Domagala:2004jt,Meissner:2004ju,Engle:2009vc,Engle:2010kt};
for the recent discussion of the results see \cite{Agullo:2010zz}. On the other
hand, as it was said above, the inclusion of the Holst term does not lead to any modifications of classical field equations of gravity and therefore is seemingly completely irrelevant classically.
However, it is well known that black hole entropy can be computed
\cite{Wald:1993nt}, \cite{Iyer:1995kg} in a class of diffeomorphism
invariant theories as a Noether charge associated with a timelike
Killing vector with a vanishing norm at the horizon.
A natural question arises: if we calculate the black hole entropy
following the Wald and Iyer recipe in the theory of gravity with
Holst term, will we reproduce the Loop Quantum Gravity result? This
is the main problem we would like to address in this paper.
The BF formulation of gravity is a very convenient starting point in this context. First, it naturally leads to the emergence of the Holst term. Second,
the analysis of the boundary terms is particularly simple in this case.
As we will see below, in this formulation the problem of notorious
counterterms, that usually have to be added to the action in order
to make it differentiable and finite, is automatically taken care
of. Last, but not least, the calculation of Noether charges in this
formulation is much simpler than in the case of the standard first order gravity.
The plan of this paper is as follows. In the next section we will
show that in a black hole, asymptotically anti-de Sitter spacetime,
the action (\ref{8}) is differentiable. This remarkable fact can be
understood in the complementary first order gravity formulation as
being due to the presence of the topological invariants with the right coefficients.
In Sec.\ III, returning to the constrained BF theory, we will construct the Noether
charges following the construction of Wald and Iyer \cite{Wald:1993nt}, \cite{Iyer:1995kg}.
Next, in Sec.\ IV we make use of these
expressions to calculate entropy of Schwarzschild--AdS black hole.
The final section will be devoted to discussion and conclusions.
\section{Boundaries and differentiability}
When spacetime has boundaries we must make sure that the action (\ref{8}) is
differentiable and the variational principle is well
defined\footnote{Usually one also assumes that the action should be
finite for physically reasonable asymptotic conditions for the
fields at infinity, so as to make the path integral meaningful. We
will not investigate this issue in details here.}. The
differentiability of the action means essentially that the values of
the fields and the form of variations are chosen in such a way that
the boundary contribution to the variation of the action vanishes.
Investigating this we will see how powerful is the BF formulation
outlined in the previous section. In what follows we will restrict
ourselves to the black hole spacetimes with the anti-de Sitter
asymptotic; therefore we will have to do with a manifold with the
boundary at infinity, where the gravitational field satisfies
$F^{ab} =0$ (cf.\ \ref{15}), and the inner black hole boundary,
where we assume that the variation of connection vanishes
$\delta\omega^{ab}=0$. This latter condition is imposed because
fixing connection at the horizon means fixing the black hole
temperature, and therefore this boundary condition is essentially
equivalent to imposing the zeroth law of black hole mechanics.
Consider the variation of the action (\ref{8}) keeping only the
terms that contribute to the boundary integral
$$
16\pi \,\delta S(A,B)= \int_M \delta F^{IJ}\wedge B_{IJ} + \ldots =
\int_M d \delta A^{IJ}\wedge B_{IJ} + \ldots
$$
\begin{equation}\label{16}
= \int_{\partial M}\delta A^{IJ}\wedge B_{IJ}+\mbox{ bulk
terms,}
\end{equation}
with $\partial M=(\mathbb{R}\times \partial\Sigma_{\infty}) \cup (\mathbb{R}\times \partial\Sigma_{H})$.
There are two contributions to the integral at infinity,
proportional to $B_{a4}$ and $B_{ab}$. The first vanishes because
$B_{a4}$ is proportional to torsion which vanishes by the field
equations, and the second is zero because $B_{ab} \sim F_{ab}$ which
vanishes by the virtue of asymptotic condition.
Similarly, at the black hole horizon the term $\delta A^{a4}\wedge
B_{a4}$ is proportional to torsion and therefore zero, while the
term $\delta A^{ab}\wedge B_{ab}$ vanishes because we choose the
boundary condition $\delta\omega^{ab}=0$ there, as discussed above.
Therefore, remarkably, we find that the BF action is differentiable
without any need of adding counterterms.
To understand how this result comes about let us notice that the
action (\ref{8}) written in the components has the form\footnote{The prefactor $\gamma$ results from combining the original Holst term with torsionless part of Nieh-Yan term.}
$$
\left[(\mbox{Palatini}+\Lambda)+\ell^2 \mbox{Euler}\right]
- \gamma \left[\mbox{Holst}+\ell^2 \mbox{Pontryagin}\right]
$$
It can be checked that these are exactly the combinations needed to
cancel out the boundary terms at infinity resulting from varying the
Palatini and Holst actions. To see this consider the first
combination above. Take an arbitrary variation of the Palatini
action, to wit
\begin{equation}
\delta (\mbox{Palatini}+\Lambda)=\int_{M}(f.e.)_{a}\delta
e^a+(f.e.)_{ab}\,\delta \omega ^{ab}+\int_{\partial M}\Theta ,
\end{equation}
where $(f.e.)$ denote field (Einstein and torsion) equations, while
\begin{equation}\label{theta}
\Theta =\frac{1}{32\pi G}\epsilon_{abcd}\,\delta \omega ^{ab}\wedge
e^{c}\wedge e^{d}\,.
\end{equation}
Let us now turn to the Euler term. As it is well known
\begin{equation}
E_4 = 32\pi\, \chi(M) + 2\int_{\partial M} \widetilde{CS}_3\, ,
\end{equation}
where $\chi(M)$ is the Euler characteristics of the manifold $M$ and
$\widetilde{CS}_3$ is the Chern--Simons three-form for the Lorentz gauge
algebra. The Euler characteristics is a fixed number and its
variation vanishes; the variation of Chern--Simons form is
\begin{equation}\label{CS}
\delta\, \widetilde{CS}_3=\epsilon_{abcd}\,\delta \omega ^{ab}\wedge R^{cd}\,.
\end{equation}
It can be now checked directly that the terms (\ref{theta}) and
(\ref{CS}) are being combined to give $\delta\omega ^{ab}\wedge
F^{cd}\, \epsilon_{abcd}$, which is zero by the virtue of the
asymptotic condition at infinity, and by boundary condition at the
horizon. The Holst term and the Pontryagin counterterm can be
analyzed similarly.
\section{Noether charges and entropy}
Now knowing that the action (\ref{8}) is differentiable we can turn
to the discussion of the Noether charges associated with its
symmetries. In our derivation below we will follow the procedure
proposed in the papers \cite{Wald:1993nt} and \cite{Iyer:1995kg}.
Let us start with an arbitrary variation of the action (\ref{8})
$$ \begin{aligned}
\displaystyle 16\pi\,\delta S= \int \Big(&\delta B^{IJ}
\wedge (F_{IJ}-\beta B_{IJ}-\frac{\alpha}{2}B^{KL} \,\epsilon_{IJKL4})+\\
&+\delta A_{IJ} \wedge ({\cal D}^A B^{IJ} ) +d( B^{IJ} \wedge \delta
A_{IJ})\Big)\,.
\end{aligned}$$
The expressions proportional to the variations of $B^{IJ}$ and
$A^{IJ}$ in the bulk are field equations, while the last term is the total
derivative of the 3-form symplectic potential:
\begin{equation}\label{17}
\Theta= B^{IJ} \wedge \delta A_{IJ}\,.
\end{equation}
For an arbitrary diffeomorphism generated by a smooth vector field
$\xi^\mu$, one can derive the conserved Noether current 3-form $J$
given by
\begin{equation}\label{18}
J[\xi] = \Theta[\phi, L_\xi \phi] - I_\xi {\cal L},
\qquad J[\xi] =B^{IJ} \wedge L_\xi A_{IJ}-I_\xi {\cal L}
\end{equation}
where ${\cal L}$ is the Lagrangian, $L_\xi$ denotes the Lie derivative in
the direction $\xi$ and contraction $I_\xi$ (acting on a $p$-form $\alpha$) is
defined to be
$$
I_\xi \alpha_p=\frac{1}{(p-1)!}\xi^\mu\,
\alpha_{\mu\nu^1...\nu^{p-1}}dx^{\nu^1}\wedge\cdots\wedge
dx^{\nu^{p-1}}\,.
$$
By direct calculation we find
$$ \begin{aligned}
16\pi\, J[\xi] &= \left(F_{IJ}-\beta B_{IJ}-\frac{\alpha}{2}B^{KL} \,\epsilon_{IJKL4}\right)\wedge I_\xi B_{IJ}\\
&+I_\xi A_{IJ}\wedge\left({\cal D}^A B^{IJ}\right) +d\left(B^{IJ}\wedge
I_\xi A_{IJ}\right)\,.
\end{aligned}$$
When field equations are satisfied this current is an exact
differential of a two form and thus we can write down the associated
charge to be
\begin{equation}\label{19}
Q[\xi]=\frac1{16\pi}\int_{\partial \Sigma} B^{IJ}\, I_\xi A_{IJ}
\end{equation}
which, after substituting the solution of the $B$ field equations
takes the form
\begin{equation}\label{20}
Q=\frac1{16\pi}\int_{\partial \Sigma} \left(\frac{1}{2} M^{ab}{}_{cd}\, F_{ab}\, I_\xi \omega^{cd}-\frac{2}{\beta\ell^2}\,
T_a \, I_\xi e^a\right)\,,
\end{equation}
where $\partial\Sigma$ is a spatial section of the manifold.
One can check that the expression for the Noether charge (\ref{20})
agrees with the one that can be obtained from the first order action
(\ref{our}), as it should. It is also worth noticing that the
Noether charge can be expressed compactly as
$$
Q=\frac{1}{16\pi}\int_{\partial \Sigma} \frac{\delta {\cal L}}{\delta F^{IJ}}\;I_\xi A^{IJ}\,.
$$
Turning back to the formula (\ref{20}) and taking torsion \mbox{$T^a=0$} we
can express the charge in the final form
\begin{equation}\label{21}
Q[\xi]=\frac{\ell^2}{32\pi G}\int_{\partial \Sigma} I_\xi \omega_{ab}
\left(\epsilon^{ab}_{\;\;\;cd}F^{cd}_{jk}-2\gamma F^{ab}_{jk}\right)dx^j\wedge dx^k\,.
\end{equation}
This generalizes the result of \cite{Wald:1993nt},
\cite{Iyer:1995kg}, \cite{Aros:1999kt}, \cite{Aros:1999id}, and
\cite{Olea:2005gb} to the case of first order gravity with Immirzi
parameter.
Having the general expression for the charge, we can now turn to
finding the formula for the entropy. According to \cite{Wald:1993nt} and
\cite{Iyer:1995kg} the black hole entropy $S$ is proportional to the
value of the Noether charge (\ref{21}) calculated at the black hole
horizon and associated with a timelike Killing vector
$\partial/\partial t$, which vanishes at the horizon $\partial\Sigma_H$
\begin{equation}\label{22}
\left. Q\left(\frac\partial{\partial t}\right)\right|_{\partial\Sigma_H} = \frac{\kappa}{2\pi}\,
\mbox{Entropy}\, ,
\end{equation}
where $\kappa$ is the surface gravity. The question we would like to
address here is how the presence of the Immirzi parameter influence
the resulting expression for entropy. In this paper we will
investigate only the case of AdS--Schwarzschild black hole, leaving
another examples of the asymptotically anti-de Sitter black hole
spacetimes to the forthcoming publication.\newline
To calculate the value of the Noether charges (\ref{21}) for the
Schwarzschild--AdS spacetime let us first fix the metric to be
\begin{equation}\label{23}
ds^2=-f(r)^2 dt^2 +f(r)^{-2} dr^2+r^2(d\theta^2+\sin^2 \theta d\varphi^2)
\end{equation}
with
\begin{equation}\label{24}
f(r)^2=(1-\frac{2GM}{r}+\frac{r^2}{\ell^2})\,.
\end{equation}
It can be checked that for the case of the metric (\ref{23}) the
surface gravity $\kappa$ defined by the equation
\begin{equation}
I_\xi \omega^{ab} \xi_b=\kappa\xi^a
\end{equation}
is given by
\begin{equation}\label{sg}
\kappa=\omega^{01}_t \Big|_{r_H}= \left(\frac{1}{2} \frac{\partial f(r)^2}{\partial r}
\right)\Big|_{r_H} \qquad\qquad T=\frac{\kappa}{2\pi}\,.
\end{equation}
The charge associated with the timelike Killing vector
$\xi \equiv\partial/\partial t$ equals
$$
Q[\xi]=\frac{4\ell^2}{32\pi G}\int_{\partial \Sigma} \omega^{01}_{t}\left(\epsilon_{0123}F^{23}_{jk}-\gamma F^{}_{jk\,01}\right)dx^j\wedge dx^k
=$$\begin{equation}\label{25} \frac{4\ell^2}{32\pi G}\int_{\partial \Sigma}
\left(\frac{1}{2} \frac{\partial f(r)^2}{\partial r}
\right)\left(1+\frac{r^2}{\ell^2}-f(r)^2\right)\sin \theta d\theta
\wedge d\varphi\,.
\end{equation}
Notice that this expression does not depend of the Immirzi
parameter; the $\gamma$-dependent terms in (\ref{21}) have just
dropped out.
The value of this charge calculated at the boundary at infinity
gives
\begin{equation}\label{26}
Q [\xi ]{}_\infty=\lim_{r\rightarrow\infty}\frac{1}{4\pi }
\int_{\partial \Sigma} \left( M+\frac{\ell^2 GM^2}{r^3}\right)
\sin \theta d\theta \wedge d\varphi = M
\end{equation}
as it should be \cite{Aros:1999kt}, \cite{Aros:1999id}.\newline
The charge calculated at the Schwarzschild--AdS black hole horizon
equals
\begin{eqnarray}
\label{27}
Q[\xi]_H&=&\frac{\kappa \,\ell^2}{8\pi G}\left(1+\frac{r_H^2}{\ell^2}\right)\int_{\partial \Sigma}
\sin \theta d\theta \wedge d\varphi\nonumber\\
&=&\frac{\kappa}{2\pi}\frac{4\pi(r_H^2 +\ell^2)}{4G}\, ,
\end{eqnarray}
where $\kappa$ is the surface gravity defined by eq.\ (\ref{sg}).
The horizon radius $r_H$ is the largest real solution of the third
order equation
$$
r^3/\ell^2+r-2GM=0\, ,
$$
which allows us to rewrite the expression (\ref{27}) as
\begin{equation}\label{28}
Q[\xi]_{H}= \kappa \;\frac{M\ell^2}{r_H}\,.
\end{equation}
From (\ref{27}) it's straightforward to see that the black hole entropy yields the form
\begin{equation}\label{30}
\textsc{S}=\frac{A}{4G}+\frac{4\pi\ell^2}{4G}\, .
\end{equation}
The first term is the standard Bekenstein--Hawking area law, while
the second is just a constant, which does not alter the
first law of thermodynamics. In the
first order formalism its presence can be regarded as price that has
to be paid for the regularization at infinity and the presence of
the Euler term in the action. The appearance of the additive
constant in the expression for black hole has been discussed in the
context of Lovelock and Gauss-Bonet gravity theories e.g.\ in
\cite{Krishnan:2009tj}, \cite{Sarkar:2010xp}. It is worth mentioning that our model avoids the problem of the negative entropy \cite{Clunan:2004tb}.
In the context of the BF construction presented above this constant
could be understood as an indication that the vacuum of the
constrained BF theory (being the maximally symmetric spacetime with
$SO(2,3)$ symmetry) carries some entropy. A deeper origin of this
entropy remains still to be understood.
\section{Discussion: what about the Immirzi parameter?}
In the precedent section we calculated the entropy of the
Schwarzschild--AdS black hole by making use of the Wald-Iyer
prescription. Interestingly, the resulting expression (\ref{30})
does not contain any trace of the Immirzi parameter, in spite of the
fact this parameter was present in the Lagrangian of the dynamical
theory that we started with.
Before turning to the discussion of this intriguing result let us
try to trace the reason for the Immirzi parameter disappearance. Let
us consider, for simplicity, the action without the regularizing
Euler and Pontryagin terms, using just the Palatini and Holst
actions. In the case of axisymmetric stationary spacetime the charge
associated with the Killing vector $\partial_\chi$ being either
$\partial_t$ or $\partial_\varphi$ (related to the mass and angular
momentum at infinity) reads
\begin{eqnarray}
Q[\partial_\chi]&=&\frac{1}{32\pi G}\int_{\partial\Sigma}\omega^{ab}_\chi
\big(\epsilon_{abcd}(e^c_\theta e^d_\varphi-e^c_\varphi e^d_\theta)\nonumber\\
&-&\frac{2\gamma}{32\pi G}\int_{\partial\Sigma}\omega^{ab}_\chi
\left(e_{\theta\,a}e_{\varphi\,b}-e_{\varphi\,a}e_{\theta\,b})\right)\,.
\end{eqnarray}
Using the definition of the connection
$\omega^{ab}_\mu=e^{\nu\,a}\nabla_\mu
e^b_\nu=e^{\nu\,a}\left(\partial_\mu
e^b_\nu-\Gamma^\lambda_{\;\mu\nu}e^{b}_\lambda \right)$ we can
drastically simplify this formula and for $\chi=t$ we have
\begin{equation}
Q[\partial_t]=\frac{1}{16\pi G}\int_{\partial\Sigma} \left(\epsilon^{\mu\nu}_{\quad\theta\varphi}\Gamma_{\mu t\nu}-\gamma\,(\Gamma_{\theta t\varphi}-\Gamma_{\varphi t\theta})\right)\,.
\end{equation}
Therefore Immirzi parameter might be present in the expression for
black black hole thermodynamics if
\begin{equation}
\gamma \int_{\partial \Sigma}\partial_\theta g_{t\varphi}\neq 0\,.
\end{equation}
Thus we expect that such a contribution proportional to Immirzi
parameter can be present in Taub--NUT--AdS spacetime, and it is
going to be proportional to Taub--NUT 'mass', and not to the
Schwarzschild one. We will present the detailed discussion of
several black hole asymptotically AdS spacetimes in the forthcoming
paper.
Let us now return to the problem if and how our expression for the
entropy (\ref{30}) can be reconciled with the Loop Quantum Gravity
calculation \cite{Ashtekar:2000eq}, \cite{Domagala:2004jt},
\cite{Meissner:2004ju}, \cite{Agullo:2010zz} according to which the black hole entropy computed by counting the black hole horizon microstates equals
\begin{equation}\label{31}
\textsc{S}_{LQG}=\frac{\gamma_M}{\gamma}\frac{A}{4G}
\end{equation}
where $\gamma_M$ is a parameter, whose numerical value is between
$0.2$ and $0.3$, accompanied by higher order corrections; see
\cite{Agullo:2010zz} for detailed discussion. It is not hard to
understand why $\gamma$ should be explicitly present in this
formula. Indeed the Immirzi parameter defines the size of the
quantum of the area, and therefore it must show up in the state
counting for black hole horizon. It would have been for some quite
unnatural cancelations to make it disappear from the entropy formula
in the semiclassical limit of Loop Quantum Gravity. Yet the
expression for entropy presented above (\ref{30}), which holds in
the semiclassical theory, whose quantum counterpart LQG is supposed
to be, shows no trace of $\gamma$.
A possible way to resolve this dilemma, as suggested in
\cite{Jacobson:2007uj}, is to notice that the entropy in (\ref{31})
was calculated using microscopic quantities, while in eq.\
(\ref{30}) with the help of those of effective low energy ones. It
follows that there might be highly nontrivial relations between the
area $A$ and and Newton's constant $G$ of (\ref{31}) and those of
(\ref{30}), so that, when the relations between them are properly
understood, and the renormalization effects are taken into account,
the two expression may turn out to be completely equivalent.
Another possible way out was proposed recently in
\cite{Perez:2010pq}. In this paper it was observed that there exist
an additional ambiguity parameter associated with the construction
of the $SU(2)$ Chern--Simons theory that describes the microscopic
degrees of freedom of the isolated horizon. This parameter is of the
similar nature as the Immirzi one, and one can adjust the two in
such a way, so as to make the final expression for the black hole
entropy having the standard Bekenstein--Hawking form.
In both cases it remains to be understood in details how the
proposed mechanisms work. This question is related to the notorious
problem of the semiclassical limit of Loop Quantum Gravity, and it
seems that without controlling this limit one cannot make any
definite conclusions.
\section*{ACKNOWLEDGEMENTS}
We thank A.\ Perez for discussion and bringing the paper
\cite{Perez:2010pq} to our attention. The work of J.\
Kowalski-Glikman was supported in part by grants 182/N-QGG/2008/0,
and the work of R.\ Durka was supported by the National Science
Centre grant N202 112740.
|
\section{Introduction} \label{sec:introduction}
The detection of neutrinos from a core-collapse supernova (SN) is the key to understanding SNe and neutrino properties. The fact that SNe emit all six flavors of neutrinos and antineutrinos, with average energies that depend on their different cross sections, offers a rich potential to reveal the detailed properties of the proto-neutron star. And the fact that these extreme conditions can lead to strong neutrino mixing may amplify the effects of neutrino properties too subtle to be seen in the laboratory. However, to fully separate and identify the astrophysical and particle physics effects, all the flavors must be detected and their spectra measured. While \mbox{SN 1987A} provided evidence of the expected strong neutrino emission from a core-collapse supernova~\cite{Hirata:1987hu, Bionta:1987qt}, only $\sim 20$ $\bar{\nu}_e$ events were detected, allowing only modest precision in the spectrum measurement and no clear evidence of what mixture of the initial $\bar{\nu}_e$ and $\bar{\nu}_{\mu}/\bar{\nu}_{\tau}$ led to the received $\bar{\nu}_e$~\cite{Jegerlehner:1996kx, Lunardini:2004bj, Yuksel:2007mn, Pagliaroli:2008ur}.
It will be challenging to experimentally measure and theoretically interpret supernova neutrino data. The prospects for success depend crucially on the values of the emission parameters and especially on the differences between flavors, {\it e.g., } their average energies. Compared to a decade ago, it is now widely thought that the average energies and the differences between flavors are both less~\cite{Fischer:2009af, Huedepohl:2009wh, Brandt:2010xa} (see~\cite{Janka:2006fh, Cardall-Talk-2010} for reviews), which reduces the expected numbers of events and the effects of neutrino mixing, making the challenges greater. Further, there is now an appreciation that neutrino mixing is much more complicated due to neutrino-neutrino interactions~\mbox{\cite{Duan:2006an, Hannestad:2006nj, Raffelt:2007cb, EstebanPretel:2007ec, Fogli:2007bk, Dasgupta:2007ws, EstebanPretel:2008ni, Dasgupta:2009mg, Chakraborty:2009ej, Dasgupta:2010ae, Duan:2010bf, Choubey:2010up, Mirizzi:2010uz, Galais:2011jh} (see \cite{Duan:2010bg} for a review)}, making the importance of high-statistics measurements with flavor and energy dependence even greater.
It is likely that a Milky Way supernova will occur in the coming decades, and it is exceedingly important that we are prepared to capture the most detailed data possible. Present detection capabilities are much better than for SN 1987A, and the likely supernova distance much closer, so a large detected yield is expected. The $\bar{\nu}_e$ spectrum will be measured in Super-Kamiokande and other detectors with $\sim 10^4$ events. The $\nu_e$ spectra could be measured well in one of the proposed liquid Argon detectors. For these flavors, there are charged current detection channels with large cross sections and good spectral fidelity.
The other flavors, $\nu_{\mu}$, $\nu_{\tau}$, $\bar{\nu}_{\mu}$, and $\bar{\nu}_{\tau}$, collectively called $\nu_x$ and assumed to be similar, while harder to measure, are especially important because they constitute the bulk of the emission and drive neutrino mixing effects. Because the charged current channels are energetically forbidden, detection depends on the smaller neutral current cross sections; this is partially compensated by the correspondingly higher emitted average energy and the four flavors. The most serious difficulty is separating these events from other channels and measuring their spectrum without having charged leptons in the final states.
The larger the average energies of all flavors and their spectral differences, the easier it is to measure $\nu_x$ spectra. Ideally, the charged current channels would show two distinct spectral components, due to mixing. However, on the basis of SN 1987A data and contemporary supernova simulations, this seems unlikely. Some neutral current channels, {\it e.g., } those on $^{12}$C and $^{16}$O, lead to the emission of gamma rays with energies characteristic of the nuclei and not the incident neutrinos~\cite{Kolbe:1992xu, Langanke:1995he}. The yields are low for what now appear to be reasonable average energies, and sensitivity to the assumed average energy is degenerate with uncertainties in the total flux and the nuclear cross sections. Neutral current neutrino-electron elastic scattering does yield spectral information, but it is very difficult to isolate these events from charged current events in this and other channels.
As a solution to these problems, Beacom, Farr, and Vogel (BFV) pointed out that neutrino-proton elastic scattering ($\nu + p \rightarrow \nu + p$) in scintillator detectors could give a large yield of separable $\nu_x$ events with spectral information, provided that the low-energy detector backgrounds are low enough~\cite{Beacom:2002hs}. Fairly optimistic assumptions about the properties of then-future detectors and the $\nu_x$ average energy were required. This detection reaction mainly probes high neutrino energies, so it is important to ask if it is viable with the known capabilities of present detectors and the lower average energies assumed today.
We show that this technique is indeed viable with realistic inputs and in fact is essential to adequately understand supernova emission and neutrino properties, especially in light of collective oscillation effects~\cite{Duan:2010bg}. We provide new detailed calculations with contemporary inputs for several detectors: the presently-running Borexino and KamLAND, the near-term SNO+, and the much larger proposed LENA. These results are needed so that the experiments have triggers in place that will ensure that this data is not missed, and to show how to interpret the signals. Most importantly, we develop a new method to directly invert the measured proton spectrum for the unknown $\nu_x$ spectrum, allowing one to go beyond the thermal spectra assumed by BFV.
The outline for this paper is as follows. In Sec.~\ref{sec:inverseSN}, we discuss why learning about SN neutrino spectra is a complicated problem, and how measuring $\nu$--$p$ elastic scattering events addresses the issue. In Sec.~\ref{sec:inputs}, we review the general framework of $\nu$--$p$ elastic scattering at scintillators. In Sec.~\ref{sec:signals}, we present the expected signals, and, in Sec.~\ref{sec:inversion}, our prescription to reconstruct the $\nu_x$ spectrum at Earth. We discuss, in Sec.~\ref{sec:LENA}, possible improvements if a larger detector is built. We discuss phenomenological implications in Sec.~\ref{sec:implications} and conclude in Sec.~\ref{sec:summary}.
\section{Collective Neutrino Mixing} \label{sec:inverseSN}
Ideally, one would like to determine the SN neutrino emission parameters, study SN explosion dynamics, and determine neutrino parameters from a measurement of SN neutrino spectra. Different aspects of the uncertain SN astrophysics and neutrino oscillations are coupled due to the effects of neutrino mixing in dense matter. These effects can be strong, have a complicated phenomenology, and are still not completely understood. All this makes the problem of reconstructing the parameters of SN emission and neutrino mixing a highly challenging problem.
While neutrino parameters may eventually be determined from cosmology~ \cite{Carbone:2010ik} or oscillation experiments~\cite{Bandyopadhyay:2007kx}, there is no other way to probe SN neutrino emission or explosion parameters in detail. One needs measurements of the received neutrino fluxes, which, when interpreted with guidance from state-of-the-art SN simulations~\cite{Janka:2006fh, Cardall-Talk-2010}, could reveal important aspects of SN physics.
This inverse SN neutrino problem was already difficult in simple neutrino mixing scenarios~\cite{Dighe:1999bi}. However, there was at least one major simplification. The flavor evolution did not depend on the initial spectra themselves, {\it i.e., } the Hamiltonian was independent of the neutrino fluxes. As a consequence, observing the $\nu_x$ spectra wasn't necessary for determining the flavor evolution of $\nu_e$ or $\bar{\nu}_e$.
However, it has since been realized that SN neutrinos are also subject to so-called ``collective effects,'' due to neutrino-neutrino forward scattering~\cite{Duan:2006an, Hannestad:2006nj, Raffelt:2007cb, EstebanPretel:2007ec, Fogli:2007bk, Dasgupta:2007ws, EstebanPretel:2008ni, Dasgupta:2009mg, Chakraborty:2009ej, Dasgupta:2010ae, Duan:2010bf, Choubey:2010up, Mirizzi:2010uz, Galais:2011jh}. This is unavoidable near the SN core, and can lead to large effects for almost the entire duration of the SN burst. In the presence of these collective effects, the flavor histories of all species get coupled to each other along their trajectory. The evolution depends explicitly on the flavor-dependent fluxes of the neutrinos. Therefore, to calculate the flavor evolution of any neutrino flavor in a SN, one requires knowledge of initial conditions for all others -- a situation fundamentally different from the older paradigm. However, if one knows the final state of all species, one can choose possible initial conditions for the neutrinos, evolve them forward, and check if they reproduce the final spectra. This allows one to interpret SN neutrinos. Of course, this bootstrap is meaningful only if one has a complete characterization of the final state. Without that, there will be strong degeneracies -- a large suite of initial conditions can produce the same final spectra for appropriate choice of neutrino and SN emission parameters, and no firm inference is possible.
In this light, it becomes crucial that all SN neutrino flavors be observed at Earth. This can be achieved through a detection of $\nu$--$p$ elastic scattering events in addition to the charged current measurements. Detecting all flavors would break degeneracies and allow determination of the primary neutrino spectra. This would also allow robust analysis of model independent signatures of SN neutrinos, {\it e.g., } to determine neutrino mixing parameters or probe shockwave dynamics~~\cite{Dasgupta:2008my,Gava:2009pj}. It is with this motivation that we study the detection of $\nu$--$p$ elastic scattering events from SN neutrinos.
\section{General framework} \label{sec:inputs}
A core-collapse SN provides a neutrino fluence, {\it i.e., } time-integrated flux, $dF/dE$ at Earth, spread over an energy range \mbox{$E\sim(5-50)$~MeV}. These neutrinos interact with $N_p$ free protons in the detector through neutral current elastic scattering. The differential cross section $d\sigma/dT$ is broad and slightly forward peaked, leading to proton recoils with kinetic energies \mbox{$T\lesssim5$~MeV}. Each recoiling proton, as it is brought to rest, deposits energy in the scintillator. These protons are slow, so the scintillated light is quenched, {\it i.e., } the number of photons scintillated corresponds to a lower effective proton recoil $T'\lesssim2$~MeV. One observes the effective proton spectrum $dN/dT'$, and the objective is to extract the neutrino fluence $dF/dE$.
The observed event spectrum due to SN neutrinos is
\be
\frac{dN}{dT'}=\frac{N_p}{dT'/dT}\int_{E_{\rm min}}^{\infty} dE\, \frac{dF}{dE}\frac{d\sigma}{dT}(E)\;.
\label{eq:mastereq}
\end{equation}
A neutrino of energy $E$ can produce a proton recoil energy between $0$ and $T_{\rm max}=2E^2/m_p$, where $m_p$ is the proton mass. In other words, a minimum neutrino energy $E_{\rm min}=\sqrt{m_p T/2}$ is needed to produce a recoil energy $T$. The recoiling proton is unbound from its atom and molecule, so that its energy loss in the medium is dominated by collisions with electrons.
There are three ingredients needed to calculate the time-integrated SN neutrino event spectrum due to $\nu$-$p$ elastic scattering at a scintillator detector: \emph{(i)} the neutrino fluence $dF/dE$ over the signal duration $\Delta t$, \emph{(ii)} the cross section $d\sigma/dT$, and \emph{(iii)} detector specific information, {\it e.g., } the number of target protons $N_{p}$, quenching function $T'(T)$, and energy resolution. Additionally, the relevant backgrounds over the duration of the burst are needed to estimate signal significance.
\subsection{SN neutrino fluence}
A SN emits a total energy ${\cal E}\approx3\times10^{53}$~erg over a burst of $\Delta t\approx10\,\,$s in neutrinos of all six flavors. The neutrino flavors $\nu_{\mu}$, $\nu_{\tau}$, and their antiparticles, have similar interactions and thus similar average energies and fluences. Therefore, the total energy is divided as ${\cal E}={\cal E}_{\nu_e}+{\cal E}_{\bar{\nu}_e}+4{\cal E}_{\nu_x}$. The various flavors have different average energies -- lowest for $\nu_e$ and highest for $\nu_x$. The fluence in each flavor is distributed in energy according to a normalized spectrum $d\varphi_{\alpha}/dE$. A SN at a distance $d$ from Earth thus provides a neutrino fluence
\be
\frac{dF}{dE}=\sum_{\alpha} \frac{dF_{\alpha}}{dE}=\frac{1}{4\pi d^2}\sum_{\alpha}\frac{{\cal E}_{\alpha}}{\langle E_\alpha\rangle}\frac{d\varphi_\alpha}{dE}\;,
\end{equation}
where $dF_{\alpha}/dE$ is the fluence in each flavor. For a neutral current process, the initial flavor distribution is not important: only the sum of all active-flavor neutrinos is relevant. Assuming no active-sterile mixing, one can ignore subsequent neutrino oscillation effects for calculating this proton recoil signal.
The parameters of the neutrino fluences are not known accurately, and the objective is to measure them. For concreteness, we choose as a nominal spectrum \mbox{$d\varphi_\alpha/dE=(128/3)\,(E^3/\langle E_\alpha\rangle^4)\,{\rm exp}(-4E/\langle E_\alpha\rangle)$}. This is one variant of the Keil parametrization for the spectrum~\cite{Keil:2002in}, and is normalized as $\int dE\,d\varphi_\alpha/dE=1$. The fluence in each flavor is then
\be
\frac{dF_\alpha}{dE}=\frac{2.35\times10^{13}}{{\rm cm}^{2}\,{\rm MeV}}\cdot\frac{{\cal E}_{\alpha}}{d^2}\frac{E^3}{\langle E_\alpha\rangle^5}\,{\rm exp}\left({-\frac{4E}{\langle E_\alpha\rangle}}\right)\;.
\label{eq:fluence}
\end{equation}
In the last expression, ${\cal E}_{\alpha}$ is in $10^{52}$ erg, $d$ is in $10$~kpc, and energies are in MeV. For the numerical evaluations, we take a representative supernova at the Galactic center region at $d=10$~kpc, with a total energy output of $3\times10^{53}$ erg equipartitioned in all neutrino flavors, {\it i.e., } ${\cal E}_{\alpha}=5\times10^{52}$ erg for each of the 6 flavors. Further, we choose $\langle E\rangle$ to be $12$~MeV for $\nu_e$, $15$~MeV for $\bar{\nu}_e$, and $18$~MeV for the 4 other flavors represented by $\nu_x$, respectively. We show, in Sec.~{\ref{sec:signals}}, how the proposed measurement is highly sensitive to SN emission parameters.
\subsection{Detection cross section}
The yield from elastic scattering on protons, {\it i.e., }
\be
\nu + p\rightarrow \nu +p\quad\quad(\rm any\,flavor)\,.
\end{equation}
is comparable to that of inverse-beta reactions ($\bar{\nu}_e+p\rightarrow e^{+}+n$). The total cross section is about a factor of four smaller, but this neutral current channel couples to all active flavors of $\nu$ and $\bar\nu$, as opposed to only $\bar{\nu}_e$. The crucial difference is that the elastic scattering events are at low quenched energies~$T'\lesssim2$~MeV, and one needs a significantly low threshold to detect these events~\cite{Beacom:2002hs}.
\begin{table*}[!t]\centering
\caption{Detector properties, {\it i.e., } number of free proton targets ($N_p$), Birks constant ($k_B$), and energy resolution ($\Delta T'/T'$), and signal yields above $0.2$~MeV for the large scintillator detectors considered here. Note that the Birks formula for quenching in the KamLAND detector includes the quadratic correction, while others do not.}
\setlength{\extrarowheight}{4pt}
\begin{ruledtabular}
\begin{spacing}{1.1}
\begin{tabular}{cccccccc}
Detector&~Mass\;&Chemical composition&$N_{p}$&$k_{B}$&$\Delta T'/T'$&Signal Yield~\\
&[kton]&(rounded to nearest $\%$)&$~\left[10^{31}\right]\;$&~[cm/MeV]\;&($T'$ in MeV)&~($T'>$0.2 MeV)\;\vspace{0.1cm}\\\hline
Borexino&0.278&${\rm C}_{9}{\rm H}_{12}$&$1.7$&$0.010$&$4.5\%/\sqrt{T'}$&$27$\\
~KamLAND\;&0.697&~${\rm C}_{12}{\rm H}_{26}$($80\%$v/v)+${\rm C}_{9}{\rm H}_{12}$($20\%$v/v)\;&$5.9$&$0.0100$&$6.9\%/\sqrt{T'}$&$66$\\
SNO+&0.800&${\rm C}_{6}{\rm H}_{5}{\rm C}_{12}{\rm H}_{25}$&$5.9$&$0.0073$&$~5.0\%/\sqrt{T'}$\;&$111$\vspace{0.1cm}\\
\end{tabular}
\end{spacing}
\end{ruledtabular}
\label{Tab:det-vals}
\end{table*}
The differential cross section $d\sigma/dT$ for a neutrino of energy $E$
to produce a proton recoil of kinetic energy $T$, to zeroth order in $E/m_{p}$, is given by~\cite{Weinberg:1972tu,Beacom:2002hs}
\bea
\frac{d\sigma}{dT}&=&\frac{G_F^2 m_p}{\pi}\left[\left(1-\frac{m_p T}{2E^2}\right)c_v^2+ \left(1+\frac{m_p T}{2E^2}\right)c_a^2\right]\nonumber\\
&=&\frac{4.83\times10^{-42}\,{\rm cm}^{2}}{\rm MeV}\cdot\left(1 + 466\,\frac{T}{E^2}\right)\;,
\end{eqnarray}
where $T$ and $E$ are in MeV and we have used $m_p=938$ MeV, $c_v=0.04$, and $c_a=1.27/2$~\cite{Beacom:2002hs}. The recoil kinetic energy is minimum, $T=0$, for a grazing collision, and maximum, $T_{\rm max}$, when the neutrino momentum is reversed. The cross section rises linearly by a factor of $\sim2$ over this allowed range of recoil energies $T=(0-T_{\rm max})$, {\it i.e., } higher recoil energies are preferred. Note that the recoil direction cannot be measured, due to the isotropic emission of scintillation light, so one can determine the energy of the neutrino using the recoil energy only on a statistical basis. The cross section for antineutrinos is slightly different, but in practice this difference is negligible at SN neutrino energies~\cite{Beacom:2002hs}. These differences due to weak magnetism corrections almost cancel between neutrinos and antineutrinos, and become only important when average neutrino energies are $\gtrsim 30$~MeV~\cite{Beacom:2002hs}, so we ignore them in our analysis.
\subsection{Detector response}
We estimate the number of free proton targets using the fiducial mass and composition as
\bea
N_p&=&N_{A}M\sum_{i}\frac{w_i\,f_{i}}{m_i}\;,\nonumber\\
&=&6.02\times10^{32}\cdot M\sum_{i}\frac{w_i f_i}{m_{i}}\;,
\end{eqnarray}
where the fiducial detector mass $M$ (in ktons) is composed of different components $i$ with weight fractions $w_i$, molecular weights $m_i$ (in a.m.u), each contributing $f_i$ free protons per molecule, and $N_A$ is the Avogadro number.
A proton with recoil energy $T$ loses energy by repeatedly colliding with electrons in the detector material. The rate of energy loss is predicted to be approximately $\propto1/T$ by Bethe theory~\cite{Bethe:1930ku}. The energy loss rate is about \mbox{$(10^3$--$10^2)$~MeV/cm} for the considered range of recoils with $T=(1-5)$~MeV, which is much more than the typical $2$~MeV/cm for a relativistic electron in a carbon target. Also, protons have velocities in the range \mbox{$\beta=(0.03$--$0.07)$}, which at the lower end are comparable to atomic electron velocities and the Bethe approximation is no longer valid (see~\cite{Nakamura:2010zzi} for details). These subtleties are accounted for by using accurate numerical tables for $\langle dT/dx\rangle$ taken from the PSTAR tables at {\tt{http://physics.nist.gov}}. Similar data is also available at {\tt{http://srim.org}}.
The energy loss on hydrogen targets is significantly larger (almost a factor of 2) than that on carbon, so we account for the composition of each detector and add the $\langle dT/dx\rangle$ for protons on carbon and hydrogen targets in the ratio of their weights in the detector.
While all the recoil energy is deposited in the detector, only part of it leads to scintillation light. This ``quenching'' is an important effect because the proton is slow. The quenching function $T'(T)$ maps a recoil kinetic energy of the proton $T$ to an electron-equivalent quenched kinetic energy $T'$ as
\be
T'(T)=\int_{0}^{T}\frac{dT}{1+k_B \langle dT/dx\rangle}\,,
\end{equation}
where $k_B$ is Birks constant~\cite{Birks:1951}. This parametrization for the quenching function is approximate. An additional term in the denominator, $C\langle dT/dx\rangle^2$, is sometimes included to achieve a better fit to data~\cite{Chou:1952}. For a typical Birks constant of $0.01$~cm/MeV, there is no quenching when $\langle dT/dx\rangle \ll100$~MeV/cm. Whereas for $\langle dT/dx\rangle\gg100$~MeV/cm, the light yield is quenched.
Although the relationship between $T$ and $T'$ is nonlinear, it is one-to-one and can be calibrated accurately. For typical parameter values noted above, the highest recoil energies $T\sim5$~MeV are quenched by a factor $T/T'\sim2$, whereas at $T\sim1$~MeV, the quenching factor is $\sim5$. Clearly, quenching affects the recoil spectrum and the number of signal events above a given threshold. See BFV~\cite{Beacom:2002hs} for details. Of the detectors we consider, KamLAND has published its quenching factor including the quadratic correction, while Borexino and SNO+ quote their Birks constant, or equivalent.
The energy resolution of the detector depends on the number of detected photoelectrons per unit energy, {\it i.e., } $dn_{\rm pe}/dT'$. Assuming $dn_{\rm pe}/dT'$ to be almost constant in the relevant regime, one gets $\langle dn_{\rm pe}/dT'\rangle T'$ hits at energy $T'$. In the limit of large number of photoelectrons, their Poisson fluctuations are well approximated by a Gaussian, which leads to an energy resolution
\be
\Delta T'/T'=1/\sqrt{\langle dn_{\rm pe}/dT'\rangle T'}\;.
\end{equation}
The yield is $\langle dn_{\rm pe}/dT' \rangle \approx$ few hundred detected photoelectrons/MeV, leading to a resolution $\Delta T'/T'$ better than $10\%$ in the relevant energy range above $0.2$~MeV. We simulate the energy resolution by smearing the signal locally at each energy with a Gaussian of width given by the energy resolution.
\subsection{Backgrounds}
The backgrounds at a scintillator detector can arise from either steady detector backgrounds, or a variety of other charged/neutral current signals due to the SN itself.
Cosmic ray induced backgrounds are very small over the relevant times~\cite{Beacom:2002hs}, and the most important steady backgrounds come from radioactivities in the scintillator and surroundings. Of these, the most obvious is $\beta$-decays of $^{14}$C that produce a high rate of electrons below $0.2$~MeV. This is a common background at all carbon-based scintillators, and sets a threshold, below which the signal is almost completely background dominated. Pulse shape discrimination may be used to reject this background and lower the threshold, probing lower energy neutrinos and greatly enhancing the yield, but we do not assume that here.
Above $0.2$~MeV, most background events are due to $\alpha$-decays of $^{210}$Po in the energy range $T'=(0.2$--$0.5)$~MeV. In their common quest to detect solar neutrinos, all the detectors we consider have purified their scintillator to similar levels, and the $^{210}$Po rates are similar and manageably small. This background has been measured, including its spectral shape, and can be subtracted statistically or by pulse shape discrimination.
Charged current signals from SN neutrinos that could be important backgrounds at these energies are either small, or can be tagged and subtracted~\cite{Beacom:2002hs}. Other neutral current channels, {\it i.e., } reactions on $^{12}$C, have a high energy threshold~\cite{Beacom:2002hs} and are disfavored, relative to BFV~\cite{Beacom:2002hs}, in the light of low average neutrino energies that are preferred by state-of-the-art simulations~\cite{Janka:2006fh, Cardall-Talk-2010}. Further study of the inelastic neutral current channels are needed~\cite{Beacom:2002hs}. Elastic scattering events on carbon nuclei are heavily quenched and and unobservable at present detectors~\cite{Beacom:2002hs}. Finally, elastic scattering on electrons has a small rate and produces larger recoil energies.
Consequently, we can use an almost universal description of the detector backgrounds for the SN burst signal -- a sharp threshold at $T'\sim0.2$~MeV, and a well understood reducible background above threshold. This was hoped for in BFV, and it is remarkable that it has been achieved in working experiments.
\section{Detected signals} \label{sec:signals}
We consider the large scintillator detectors, {\it i.e., } Borexino and KamLAND, which are already available now, and SNO+, which should be operational shortly. For each of these detectors, we need to know three relevant quantities: \emph{(i)} the number of free proton targets $N_p$, \emph{(ii)} Birks constant $k_B$, and \emph{(iii)} the energy resolution $\Delta T'/T'$. The values of these detector parameters are summarized in Table~\ref{Tab:det-vals}.
\begin{figure}[!thpb]
\includegraphics[width=1.0\columnwidth]{SNnus-Borexino.eps}
\caption{Galactic SN neutrino-proton elastic scattering event spectrum at Borexino. The dashed vertical line at $0.2$~MeV shows the threshold used.}
\label{fig:1}
\end{figure}
\subsection{Signals}
With these detectors, the expected neutrino signals in the $\nu$--$p$ elastic scattering channel are shown in Fig.~\ref{fig:1} for Borexino, in Fig.~\ref{fig:2} for KamLAND, and in Fig.~\ref{fig:3} for SNO+. We find that all three detectors can detect a significant number of events, {\it i.e., } $\sim100/$kton, with a modest number of background events, above $0.2$~MeV. We have checked that our results agree in detail with BFV~\cite{Beacom:2002hs} when identical inputs are used. Note that most of the signal in the $\nu-p$ elastic scattering channel comes from the hotter $\nu_x$ flavors. This is partly due to their average energies being higher than those of $\nu_e$ and $\bar{\nu}_e$, and partly because there are $4$ flavors that contribute to the signal. Also note, as an interesting aside, that the total yield with zero threshold is about $\sim500/$kton, similar to the $\bar{\nu}_e$ yield from inverse beta decays at water Cherenkov detectors. In the remainder of this section, we justify our choices for the detector parameters.
\begin{figure}[!thbp]
\includegraphics[width=1.0\columnwidth]{SNnus-Kamland.eps}
\caption{Galactic SN neutrino-proton elastic scattering event spectrum at KamLAND.}
\label{fig:2}
\end{figure}
\subsubsection*{Borexino} A SN signal will be localized in time and allow use of $0.278$ kton as the fiducial mass~\cite{Cadonati:2000kq, Miramonti:2003hw} instead of the smaller volume used for solar neutrinos. The detector material is pseudocumene (${\rm C}_{9}{\rm H}_{12}$)~\cite{Bellini:2010hy}, with a specific density $0.875$. The number of free proton targets is thus $1.7\times10^{31}$. We could not find a direct published reference to the Birks constant for the Borexino scintillator. However, we determine $k_B=0.010$~cm/MeV using the available quenching data on protons at Borexino~\cite{Bellini:2009jr}, {\it i.e., } protons of energy $8.3$~MeV and $4.6$~MeV get quenched to $4.1$~MeV and $1.86$~MeV, respectively and $\alpha$ particles of $5.4$~MeV are quenched by a factor of 13. Note that this is lower than the $k_B$ found for electrons in Borexino~\cite{Wagner-thesis} because low energy electrons are quenched more than protons or ions of the same kinetic energy~\cite{Robinson:1954}. About 500 photoelectrons/MeV are detected at Borexino~\cite{Arpesella:2008mt}, which sets its energy resolution to be $\Delta T'/T'=4.5\%/\sqrt{T'}$~(with $T'$ in MeV). Finally, the backgrounds at Borexino at energies below $0.2$~MeV, due to $^{14}$C, make events at those energies unusable. Above that threshold, the background is negligible. It is comprised of the entire event yield in $\sim 10$~s without a SN burst~\cite{Arpesella:2008mt}, and there are $\sim2$ background events, mostly from the $^{210}$Po peak.
\subsubsection*{KamLAND} The Kamland fiducial volume is assumed to be $0.697$~kton~\cite{Araki:2005qa} for a SN burst, corresponding to the inner $5.5$~m sphere of the detector. The detector material is a mixture of $80\%$~(by volume) of dodecane (${\rm C}_{12}{\rm H}_{26}$) with $20\%$ pseudocumene (${\rm C}_{9}{\rm H}_{12}$). The number of free proton targets is thus $5.9\times10^{31}$. From the observed quenching of protons with recoil energies \mbox{$(1-10)$~MeV}, the quenching factor was reported~\cite{Yoshida:2010zz} to be\footnote{Ref.~\cite{Yoshida:2010zz} reports $k_{B}$ in units of g/cm$^2$/MeV which we have converted to units of cm/MeV using known specific densities for dodecane ($0.750$) and pseudocumene ($0.875$).} \mbox{$k_{B}=(0.0100\pm 0.0002)$~cm/MeV} and \mbox{$C=(2.73\pm0.08)\times 10^{-5}$~(cm/MeV)$^2$}. The energy resolution at KamLAND is determined by a photoelectron yield of $210$/MeV, giving $\Delta T'/T'=6.9\%/\sqrt{T'}$~\cite{Piepke-Grant}. Note that these are marginally different from the values in Ref.~\cite{Yoshida:2010zz}, for the pre-purification scintillator. Again, the backgrounds at KamLAND at energies below $0.2$~MeV, due to $^{14}$C, make events at those energies unusable. Above that threshold, the background is comprised of the entire event yield in the absence of a SN burst~\cite{Grant:2010}, wherein one expects $\sim 31$ background events~\cite{Piepke-Grant}, mostly from the $^{210}$Po peak. The background rates are known very well, and the expected fluctuations in the background are relatively small ($\sim 5$ events), so we expect that these events can be statistically subtracted.
\begin{figure}[!t]
\includegraphics[width=1.0\columnwidth]{SNnus-SNO.eps}
\caption{Galactic SN neutrino-proton elastic scattering event spectrum at SNO+.}
\label{fig:3}
\end{figure}
\begin{figure}[!thpb]
\includegraphics[width=1.0\columnwidth]{SNr.eps}
\caption{Sensitivity (defined as fractional change in event yield relative to benchmark SN values) to SN emission parameters with KamLAND as the detector.}
\label{fig:4}
\end{figure}
\subsubsection*{SNO+} The SNO+ fiducial volume is taken to be $0.8$~kton. The detector material is linear alkyl benzene (${\rm C}_{6}{\rm H}_{5}{\rm C}_{n}{\rm H}_{2n+1}$). The alkyl group typically has a size $n=(10-16)$, and we assume $n=12$ for definiteness and a specific density of $0.86$. The number of free proton targets is thus $5.9\times10^{31}$. In recent laboratory tests of the detector material, the Birks constant for the scintillator in SNO+ is reported to be about $0.0073$~cm/MeV~\cite{Tolich:2009nnn}. The energy resolution at SNO+ is expected to be $5.0\%$ at 1~MeV and $3.5\%$ at 3.4~MeV respectively~\cite{Maneira:2010now}, from which we estimate an energy resolution $\Delta T'/T'=5.0\%/\sqrt{T'}$. These values could change in the full detector. A measurement of backgrounds at SNO+ is not available yet, but SNO+ has solar neutrino physics goals that are similar to Borexino and KamLAND, so we expect that the region above $0.2$~MeV to have similar backgrounds.
\subsection{Sensitivity and Robustness} \label{sec:results}
We now investigate the expected signal dependence on different choices of SN neutrino fluence parameters and detector characteristics. To quantify this sensitivity, we define $S$ as the fractional change in the event yield for non-benchmark values for the SN or detector parameters.
In Fig.~\ref{fig:4}, we show the signal sensitivity to the average energy of $\nu_x$, overall shape, and total energy. For these calculations we have assumed the detector properties to be that of KamLAND. Using the SN model described in Sec.~\ref{sec:inputs} as the benchmark, we show how the signal varies for \emph{(i)} a $\nu_x$ average energy $\langle E_{\nu_x} \rangle=21$~MeV or $16$~MeV, \emph{(ii)} a Maxwell-Boltzmann spectrum with the same average energy $\langle E_{\nu_x}\rangle=18$~MeV, and \emph{(iii)} a higher or lower $\nu_x$ total energy, ${\cal E}_{\nu_x}=1.3$ or $0.7$~times the benchmark value. We can see that the expected signal depends strongly on the average energy or the spectral shape. This is because the signal is mainly from the ``tail'' of the distribution, which is exponentially sensitive to $\langle E \rangle$. The dependence on ${\cal E}_{\nu_x}$, which sets the overall normalization, is linear, as expected. The yield varies by a factor of few in the range of SN emission values commonly predicted from SN theory~\cite{Janka:2006fh, Cardall-Talk-2010}.
\begin{figure}[!thpb]
\includegraphics[width=1.0\columnwidth]{Detr.eps}
\caption{Sensitivity (defined as fractional change in event yield relative to benchmark values for the KamLAND detector) to detector parameters for the benchmark SN emission.}
\label{fig:5}
\end{figure}
In Fig.~\ref{fig:5}, we show the dependence on relevant detector characteristics. We choose \emph{(i)} a higher and lower quenching, {\it i.e., } $k_B=0.0102$~cm/MeV with $C=2.81\times10^{-5}$~(cm/MeV)$^2$, and $k_B=0.0098$~cm/MeV with $C=2.65\times10^{-5}$~(cm/MeV)$^2$, respectively, {\it i.e., } the reported range of uncertainty in $k_B$ for the KamLAND scintillator~\cite{Yoshida:2010zz}, \emph{(ii)} a worsening of the KamLAND energy resolution to $\Delta T'/T'=10\%/\sqrt{T'}$. We find that these uncertainties leads to very small effects on the predicted event yield. The signal varies by only a few percent in the range of uncertainties of the detector parameters.
We checked the dependence on energy loss rate (not shown) by using a $1/T$ fit to the Bethe theory prediction, instead of accurate numerical tables. Such a naive choice of $\langle dT/dx\rangle$ leads to an erroneously larger yield ($\sim 50\%$ more). It is therefore important that accurate values for $\langle dT/dx\rangle$ be used.
It is thus clear that in the range of detector uncertainties the signal changes by few~$\%$, while the statistical uncertainties with $\sim100$ detected events are about~$10\%$. The expected signal variation between allowed SN emission models is much larger, with changes of up to a factor of two. Therefore this signal can be expected to identify specific emission scenarios.
\section{Extracting the neutrino fluence} \label{sec:inversion}
We have shown that the quenched recoil spectrum of protons due to neutrinos from a SN burst can be reliably measured at scintillator detectors. The measurement is almost completely background-free above $0.2$~MeV, the detector properties are well-calibrated, and the energy resolution doesn't play a crucial role. The task at hand is to invert this signal into a measurement of the neutrino fluence as a function of energy.
One could choose a parametrization for the spectrum and fit for those parameters using data. Two obviously interesting quantities are the average energy $\langle E_{\nu_x} \rangle$ and the total energy ${\cal E}_{\nu_x}$. In BFV, fitting for those parameters, with about $\sim 100$ signal events, led to an expected precision of about $1/\sqrt{100}\sim 10\%$, though the parameters are correlated~\cite{Beacom:2002hs}.
We show here that it is possible to take a more general approach. One can extract the spectrum non-parametrically, and the extracted spectrum has high fidelity to the true spectrum, even for realistic choices of detector characteristics.
The problem of reconstructing the neutrino spectrum from the proton recoil spectrum would be quite simple if there was a known one-to-one relationship between the recoil energy and the neutrino energy. In that case, as for inverse-beta events, one could have simply scaled the observed number of events by the cross section and detector size, and translated the observed energies $T'$ to neutrino energies using their known relationship.
This does not work for elastic scattering events. The differential cross section is broad and slightly forward peaked~\cite{Beacom:2002hs}, but the recoil energy $T$ is uniquely related to the minimum neutrino energy $E_{\rm min}=\sqrt{m_p T/2}$ that is capable of producing that recoil. Furthermore, $T$ is related in a unique way to the quenched recoil $T'$. So, having measured $T'$, we can uniquely determine $T$ using the known quenching function $T'(T)$, and then $T$ leads to its corresponding $E_{\rm min}$. This allows us to use the recoil data as a function of quenched recoil energies to extract the neutrino fluence as a function of neutrino energy $E$.
\subsection{Inversion recipe}\vspace{0.2cm}
Suppose the quenched recoil data is in a range of energies which we divide into $N_{\rm bin}$ bins. We denote the value of $T'$ at the midpoint of each bin by $T'_i$ and the width of the bin by $\Delta T'_i$, with $i$ going from $1$ to $N_{\rm bin}$. Each $T'$ is uniquely related to some recoil energy $T=T(T')$ using the inverse of the quenching function. Each of these recoil energies $T$ in turn are related to a minimum neutrino energy $E_{j}=\sqrt{m_pT(T'_j)/2}$.
The data is simply the number of observed events $N_i$ in the $i^{\rm th}$ bin
\be
N_i=\Delta T'_i \left(\frac{dN}{dT'}\right)_{T'_i}\,.
\end{equation}
Using the expression for $dN/dT'$, this can be written as
\be
N_i=\sum_{j=1}^{N_{\rm bin}}K_{ij}F_j\,,
\label{eq:Meq}
\end{equation}
where
\bea
F_{j}&=&(dF/dE)_{E_j}\Delta E_j\,,\label{eq:M2eq}\\
K_{ij}&=&\Bigg\lbrace
\begin{array}{l}
N_{p}\Delta T'_i \left(\frac{dT}{dT'}\right)_{T'_i}\left(\frac{d\sigma}{dT}\right)_{T'_i,E_j}{\rm\quad for~}\,i\leq j\nonumber\\
0\,\hspace{0.455\linewidth}{\rm~for}~i>j\end{array}
\,.\nonumber
\end{eqnarray}
Note the upper-triangular form of the matrix $K_{ij}$. This is because only neutrinos with energy more than $\sqrt{m_pT/2}$ are able to produce a proton recoil energy $T$. We illustrate the form of Eq.~(\ref{eq:Meq}) explicitly for $N_{\rm bin}=3$,
\be
\left(\begin{array}{c}
N_1\\[2.0ex]
N_2\\[2.0ex]
N_3
\end{array}\right)
=
\left(\begin{array}{ccc}
K_{11}&K_{12}&K_{13}\\[2.0ex]
0&K_{22}&K_{23}\\[2.0ex]
0&0&K_{33}
\end{array}\right)
\left(\begin{array}{c}
F_1\\[2.0ex]
F_2\\[2.0ex]
F_3
\end{array}\right)\,.
\end{equation}
It is easy to see how the above equation generalizes to any number of bins.
Now the extraction of the neutrino spectrum is simply an inversion of this matrix $K$. That is, we write
\be
F_j = \sum_{i=1}^{N_{\rm bin}}{(K^{-1})}_{ji}\,N_i\,,
\label{eq:Fsoln}
\end{equation}
and find the neutrino fluence at the detector $dF/dE=F_j/\Delta E_j$ in the energy bin around $E=E_j$. The inversion can be done trivially using back-substition, as the matrix $K_{ij}$ is upper triangular. That is, we can start with determining the fluence in the highest energy bin and proceed to lower ones. Again, with $N_{\rm bin}=3$ as an example, we have
\bea
F_{3}&=&N_{3}/K_{33}\;,\\
F_{2}&=&\left(N_{2}-F_{3}K_{23}\right)/K_{22}\;,\nonumber\\
F_{1}&=&\left(N_{1}-F_{2}K_{12}-F_{3}K_{13}\right)/K_{11}\;.\nonumber
\end{eqnarray}
It is again easy to see how the above formula generalizes to any number of bins.
The procedure is unchanged even in the presence of known backgrounds $B_i$ in each bin, except that, after the $F_i$ are determined, one would subtract the expected backgrounds. This subtraction procedure works to an accuracy $\approx \sqrt{B_i}/(F_i+B_i)$, {\it i.e., } as long as the backgrounds are small.
There is freedom in the choice of number and width of bins. The optimal number of bins depends on a compromise between finer sampling or lower noise in each bin, subject to the constraint that bins are wider than the energy resolution. Excessively narrow bins have too few events and lead to spurious features from noisy data, whereas excessively wide bins lead to a breakdown of Eq.~(\ref{eq:M2eq}), which uses a linear interpolation within each bin. We find that choosing the bins such that the events are almost equally distributed between them, leads to an accurate reconstruction.
\begin{figure}[!thpb]
\includegraphics[width=1.0\columnwidth]{Mock.eps}
\caption{Mock data for KamLAND generated from the benchmark scenario. Note the logarithmic scale on the abscissa and the unequally-sized bins.}
\label{fig:6}
\end{figure}
\begin{figure}[!thpb]
\includegraphics[width=1.0\columnwidth]{mash.eps}
\caption{Reconstructed neutrino fluence from the mock data for KamLAND (upper panel) and residuals from the benchmark model (lower panel). Note the unequally sized bins for the data. The fluences are to be compared to Eq.~(\ref{eq:fluence}), accounting for $\langle E_{\nu_x}\rangle$ and ${\cal E}_{\nu_x}$.}
\label{fig:7}
\end{figure}
Finally, a short remark about the stability of our prescription. Formally, this inversion problem is known as a Volterra integral equation of the first kind. Volterra equations are relatively stable, because the kernel $K_{ij}$ is upper triangular. There is, however, literature on how to numerically stabilize the matrix inversions, if the kernel is even mildly singular~\cite{CraigBrown:1986}. As a safeguard against unphysical artifacts in the reconstructed spectrum, due to noisy data and possible instability of the inversion procedure, the inversion procedure may need to be regulated. The regulated solution is more stable, and smoothed over expected statistical fluctuations. For our purposes, we find that we don't need a regulator when we bin the data as above.
\subsection{Numerical example}
To demonstrate the viability of this procedure we generate mock data at the KamLAND detector, using our benchmark SN model. The data ($66$ signal events) falls in the range $T'=(0.2$--$5.9)$~MeV, which we divide into $8$ bins with comparable numbers of events in each bin. The lower end of the observable range in $T'$ is set by the large $^{14}$C background below $0.2$~MeV. The upper end is where the neutrino fluxes become negligible. The chosen bin widths are significantly larger than the energy resolution and thus no significant correlation is expected between bins.
These data, without statistical noise, are shown in Fig.~\ref{fig:6}, with the number of events in each bin mentioned alongside. Expected one-sigma errors in the reconstruction due to Poisson fluctuations are plotted as error bars. We have already subtracted the modest backgrounds. This figure has a logarithmic scale on the abscissa, and unequally-sized bins. On a linear scale, all the bins would have comparable areas.
We apply our inversion procedure to the mock data and recover the total fluence $dF/dE$. We assume that the $\bar{\nu}_e$ contribution will be known accurately using the data from the inverse-beta channel, while $\nu_e$ events are negligible. Thus, we subtract the subleading contribution expected from $\bar{\nu}_e$, and divide by $4$ to find the $\nu_x$ fluence in each flavor -- $\nu_{\mu}$, $\nu_{\tau}$, $\bar{\nu}_{\mu}$, and $\bar{\nu}_{\tau}$.
The range of available recoil energies maps to neutrino energies in the range $E=(25$--$73)$~MeV. The lowest available energy is set by backgrounds, and is close to the typical average neutrino energies expected. The higher end goes well into the tail. The reconstructed $\nu_x$ fluence in this energy range is shown in the top panel of Fig.~\ref{fig:7} using the points for discrete energy bins, with their one-sigma fluctuations shown with error bars. Although we have shown here a reconstruction for mock data without statistical fluctuations, we have separately checked that a reconstruction for noisy data typically remains within the shown error bars. In the inset, we show the region of the neutrino spectrum that is being probed. Note that the peak energies are somewhat lower than the average energies, due to the asymmetric spectral shapes. We also plot some alternate fluences -- one with an average energy energy $\langle E_{\nu_x}\rangle=21$~MeV, one with a Maxwell-Boltzmann spectrum with $\langle E_{\nu_x}\rangle=18$~MeV, and one with a higher total energy ${\cal E}_{\nu_x}$ being $1.3$ times the benchmark value, to test the reconstruction.
The reconstructed spectrum shows that only the higher part of the $\nu_x$ spectrum is probed, because $\nu_x$ energies lower than $\sim 25$~MeV are swamped by background. This is clearly a drawback of this channel. Notwithstanding this, it must be emphasized that for determining non-thermal features, the tail is in fact the most interesting region of the spectrum~\cite{Langanke:2007ua}. The spectrum near the peak does not depend as strongly on average energy, and would not have significantly improved the ability to distinguish between spectral parameters. Also, signatures of flavor mixing in the signals, due to charged current events, appear at the higher energy tails where the flavor spectra of $\nu_e$ and $\bar{\nu}_e$ are most different from that of $\nu_x$. Thus it is quite important to be sensitive to the tail of the SN neutrino energy spectrum.
The main advantage of this more general reconstruction is that we obtain a \emph{direct} measurement of the $\nu_x$ flavor spectra in the higher energy range. This data is obtained without any fitting or parametrization of the primary neutrino spectra, and can be used as an independent measurement to compare with the high energy $\nu_e$ or $\bar{\nu}_e$ spectra measured in charged current interactions, to see if mixing occurs. Also, this provides a model independent comparison with SN simulations. Any departure for thermal spectra, or any unexpected features, can be seen directly.
Note that knowledge of the distance to the supernova is needed to correctly normalize the neutrino fluence. However, it is not crucial for determining the average energy, or for discovering nonthermal features. On the other hand, for an accurate measurement of the total energy, or for comparing with other flavors, the distance must be determined by other means.
In the lower panel, we plot the residuals from the true spectrum. With about $10$ events in each bin, we expect $1/\sqrt{10}\sim30\%$ uncertainties in each bin. Our bins are chosen to be wide enough to be almost uncorrelated. On the other hand, when a parametric expression for the spectrum is being tested, the statistical power in different bins can combine and lead to smaller uncertainties of about $1/\sqrt{100}\sim10\%$. These are the capabilities of a detector like KamLAND. If one combines the results from multiple detectors, {\it e.g., } KamLAND, Borexino, and SNO+, the uncertainties are $30\%$ smaller.
The events in the tail have a lot of statistical power to distinguish between models. In particular, the final data point that is far out in the tail seems to be very powerful, but it is also subject to larger systematic errors. This is simply because the highest energy event decides the bin-width of the last bin, and large Poisson fluctuations there can lead to an wrong estimate of the average flux in that bin. With actual data, the binning will need to make optimal use of the available signal. We find that, even after omitting the final bin, the alternate scenarios are disfavored with~$\Delta\chi^2>3.5$.
\begin{figure}[!thbp]
\includegraphics[width=1.0\columnwidth]{SNnus-LENA.eps}
\caption{Galactic SN neutrino-proton elastic scattering event spectrum at LENA. No backgrounds have been shown, but for $T'>0.2$~MeV, the backgrounds are expected to be small.}
\label{fig:8}
\includegraphics[width=1.0\columnwidth]{mash-LENA-minbias.eps}
\caption{Reconstructed neutrino fluence from mock data generated for LENA (upper panel) and residuals from the benchmark model (lower panel). Note the unequally sized bins. The fluences are to be compared to Eq.~(\ref{eq:fluence}), accounting for $\langle E_{\nu_x}\rangle$ and ${\cal E}_{\nu_x}$. Note that, compared to Fig.~\ref{fig:7}, much smaller differences in spectral parameters can be distinguished. In the same context, note that the range shown for residuals is much smaller.}
\label{fig:9}
\end{figure}
Note that the reconstructed spectrum is not identical to the true spectrum. This is due to biases in our discretization. We evaluated the cross section in each bin at the midpoint. Given that the differential cross section has a term that varies as $1/E^2$ with $E$, it means that we assumed a larger than effective cross section. On the other hand, the flux itself is falling exponentially in $E$. Our prescription to calculate all quantities in a bin at the mid-point has a bias for the largest sized bins. For the SN signal, we find that the bias in our scheme is within statistical uncertainty, as long as the binning is not too crude. The bias can be reduced by considering a properly weighted scheme, as opposed to the mid-point scheme we have employed.
\section{Capabilities of a larger detector} \label{sec:LENA}
In this section we discuss how the potential of this channel increases if a larger scintillation detector is built. To illustrate the potential, we choose LENA, a proposal for a $(50$--$100)$ kton scintillator detector as a possibility~\cite{Winter-thesis, Wurm:2010ny}. The number of free proton targets is expected to be $3.3\times10^{33}$, which corresponds to a fiducial mass of $44$~kton~\cite{Wurm-PC}. We assume the Birks constant for the scintillator to be $0.010$~cm/MeV, in the ballpark of other scintillators. The energy resolution is expected to be worse than smaller experiments, due to more absorption in scintillator of the scintillated light, and we assume $\Delta T'/T'=10\%/\sqrt{T'}$. These values are likely to be different in the future detector, but we believe that our choices are conservative. Backgrounds at LENA are not available yet, but LENA has solar neutrino physics goals that are similar to Borexino , KamLAND, and SNO+, so we expect that the region above $0.2$~MeV to be similarly background free for a SN burst. With these assumptions, and our benchmark SN model, the expected event rate at LENA is shown in Fig.~{\ref{fig:8}}.
We find approximately $5250$ events, almost comparable to the yield of $\bar{\nu}_e$ at Super-K for a similar SN burst. With such a large number of events, we expect this data will also allow for a time-dependent analysis. This may reveal important time dependent variations in the $\nu_x$ emission properties. These possibilities need to be investigated in more detail, but we do not pursue them here.
Applying our reconstruction procedure to this data from a LENA-type experiment will lead to extraordinary results. One could opt for finer binning in energy, as long as allowed by energy resolution, or simply much smaller statistical uncertainties in each bin. We choose the former option and, in Fig.~{\ref{fig:9}}, show a reconstruction with 16 bins. With about $70$ times larger statistics than KamLAND, we get a precision of $\lesssim 5\%$ in each bin. Thus we expect spectral and luminosity parameters to be determined to a precision of few $\%$. Although biases begin to become important, they remain manageable. The statistical power is significantly better than what can be achieved at present detectors, and would give strong constraints on acceptable SN neutrino spectra, once systematic uncertainties are under control.
\section{What can we learn?}
\label{sec:implications}
The most important message is that the $\nu$--$p$ elastic scattering channel is potentially accessible at already available scintillator detectors, and the triggers required to gather the relevant data must be put in place. Missing out on this crucial piece of data would be a huge loss to SN neutrino phenomenology.
The primary neutrino physics result from this measurement would be a direct measurement of the $\nu_x$ spectrum. This information complements the flavor information available at water Cherenkov and liquid Argon detectors. This would allow for easy comparison between these detectors and one would be able to identify the flavor exchanges in both the ``disappearance'' and ``appearance'' channel, providing a complete picture.
A revealing aspect would be the relationship between the observed $\nu_e$, $\bar{\nu}_e$, and $\nu_x$ spectra. First, one would check if there are non-thermal features in the observed spectra. If so, one would ask if the non-thermal features appear at identical energies for different flavors. An answer in the affirmative would reveal that flavor conversions are at the heart of this observed non-thermality. This would reveal the pattern of flavor exchanges over the observed energy range, shedding light on the unknown neutrino parameters, {\it i.e., } $\rm{sign}(\Delta m_{\rm atm}^2)$ and $\sin^2\tht{13}$. Additionally, this allows one to test the equipartition hypothesis and compare with SN simulations.
This data would for the first time allow us to empirically test the claim that almost all the SN energy goes in neutrinos. Only when we have detected all flavors can we find the energy output in neutrinos and compare that to the binding energy released by the star. Determination of the binding energy will also be a useful diagnostic for the proto-neutron star mass and radius, which relates to the neutron star mass and radius~\cite{Horowitz:2000xj,Page:2006ud,Lattimer:2006xb}. The measured $\nu_x$ spectrum is also a probe of nucleosynthesis~\cite{Woosley:1988ip, Fetter:2002xx, Heger:2003mm, Yoshida:2005uy, Austin:2011gw}.
Any oscillation into sterile neutrinos would be most readily observable using this neutral current channel. This will set stringent bounds on their masses and mixing, through constraints on total energy loss and a flavor composition in different detection channels.
It is important that all three experiments -- Borexino, KamLAND, and SNO+, be running and actively looking for these events. The most obvious reason is to increase statistics. However, having more than one detection also allows for useful cross-calibration to rule out any unexpected backgrounds. Additionally, a multi-detector search makes it unlikely that this important signal would be missed owing to down-time for a specific experiment.
Even the smallest of the considered experiments, Borexino, is expected to see $\sim27$ events, and is capable of measuring the total energy at about $20\%$. Combining all three available detectors gives $\sim204$ events above threshold, with $\lsim 10\%$ precision on total energy. Average energies that are significantly larger ($\lsim 10\%$) than our benchmark value would be easily distinguished or constrained.
The proposed large liquid scintillator detector LENA could be extremely useful for this channel. At present, its detector specifications are not yet certain. However, with conservative estimates on the specifications, we find that the prospects are promising. One can expect $\sim5250$ events, and few \% level measurements of SN $\nu_x$ emission parameters. This is almost comparable to what Super-K can do for $\bar{\nu}_e$. Clearly, such a significant detection in multiple flavors will allow for meaningful comparison of the data from these two detectors.
\section{Summary and Outlook}
\label{sec:summary}
We have presented an updated calculation of the SN neutrino signal due to elastic scattering on protons at scintillator detectors. Using more realistic assumptions on SN emission and detector properties than were used in BFV, we find that the signal is observable at available detectors. Additionally, we have demonstrated a simple procedure to numerically reconstruct the $\nu_x$ spectrum from the data, which will allow for new analyses and probe different aspects of SN physics and astrophysics.
As a note for the future, we must remark that detectors with lower quenching will lead to more promising results. Another avenue for drastic improvement is if the threshold can be lowered below the $^{14}$C background. This may be possible through pulse shape discrimination~\cite{Bellini:2009jr}. That allows us to extend our range to lower neutrino energies, increases the overall yield drastically, and may allow reconstructing the SN spectral peak. Together, one could hope for significantly better results than we have outlined for already achieved experiments.
In conclusion, we exhort the experimentalists working in the Borexino, KamLAND and SNO+ collaborations, and on proposed detectors like LENA, to seriously consider the physics impact of this channel and be adequately prepared to acquire this data from a future Galactic SN signal.
\section*{Acknowledgements}
We thank Bruce Berger, Frank Calaprice, Mark Chen, Christiano Galbiati, Greg Keefer, Ranjan Laha and especially Alvaro Chavarria, Brian Fujikawa, Chris Grant, Andreas Piepke, Alex Wright, and Michael Wurm for helpful discussions and comments on the manuscript. J.F.B was supported by NSF CAREER Grant PHY-0547102.
|
\section{Introduction}
\label{SEC:Intro}
In biomechanics the application of electromyography is especially a field rife with various signal processing methods. One reason for this is the complexity of the EMG signal which requires processing if one wants to proceed beyond the simple level of on-off interpretation. A number of standard processing tools can be found in EMG softwares by commercial vendors. Researchers are usually also interested in new, or modified processing methods, which may take long to implement in commercial softwares if it happens at all. In this case one can choose ''prototyping'' tools like \textsc{LabView, Mathcad, Matlab, Octave, SciLab}, and so on. Over the the years we have for instance frequently used \textsc{Labwindows CVI} (a C-based graphical programming tool) and \textsc{Mathcad} for data acquisition, visualization and analysis. One common problem in data intensive projects is that the data and analysis documents get scattered among computers and data disks. Our solution is to transfer all data to a single database (in our case based on \textsc{MySql}). Subsets of data can then be selected by \texttt{sql} searches and exported as \texttt{csv}-tables or similar. The salient point is that all the analyses can be gathered in one single script file using the R-tool \citep{R2010}. Running this script generates all the graphs, figures and statistical reports that one may need for the project. This script is easily modified when needed and can be shared among collaborators and other researchers. We believe that R provides a useful platform for defining and using various signal analyzing algorithms in EMG studies, which may stimulate further refinements and developments through collaborative efforts. As noted by \citep[Preface]{Everitt2010}:
\begin{quote}
(...) R has started to become the main computing engine for statistical research (...) For reproducible piece of research, the original observations, all data processing steps, the statistical analysis as well as the scientific report form a unit and all need to be available for inspection, reproduction and modification by the readers.
\end{quote}
With this paper and the supplementary material we hope to give a demonstration of the potential uses of R in biosignal processing.
\section{What is R?}
R is an interpreted functional programming language for statistical computing and data visualization created by Ross Ihaka and Robert Gentleman (University of Auckland, New Zealand). R is part of the GNU project and since 1997 it is developed by the \emph{R Development Core Team}. The version R 1.0.0 was released on 29 February 2000. The official project website is \url{www.r-project.org} where one can naturally download the software (binaries, or source code) and manuals \citep{Venables2009}. R is an implementation of the S language developed at the AT\&T Bell Laboratories around 1975 (by Rick Becker, John Chambers and Allan Wilks), and is influenced by the \textsc{Lisp} dialect \textsc{Scheme} created at the MIT AI Lab, also around 1975 (by Guy L Steele and Gerald Jay Sussman). A large part of the software is written in the R language itself, but core functions are written in C and \textsc{Fortran}. As an interpreted language R may be slower than say pure C-programs. Klemens, who advocates C as general scientific programming tool, gives an example with running the Fisher test 5 million times: the speed relation came out as about 1:30 in favour of C \citep{Klemens2009}. Thus in speed-critical cases, such as where we have large simulation samples, it may be necessary to revert to compiled C programs. However, in many other cases the versatility of R will outweigh the possible gains from optimizations for speed. The R tool is continually extended by packages contributed by an active community; there are presently more than 2000 packages available, see \url{http://cran.r-project.org/web/packages/}.
Downloading and installing the R software on your computer should require nothing more than basic computer skills. A useful companion is the R scripting tool \textsc{NppToR} (\url{http://sourceforge.net/projects/npptor/}) which is an adds-on to the editing tool \textsc{Notepad++} which can be downloaded from \url{http://notepad-plus.sourceforge.net/uk/site.htm}. Since documentation about R is easily found on the web we will move on to describe what we can do with the software.
\section{Importing data}
The first thing we want to do is to get our data into the workspace. R can access various databases directly but the most typical situation is where we have, say, EMG data exported on a text format. We use R as console program, so the command for reading a simple ASCI table of data from a file into variable \verb|M| using the \verb|read.table()| function is:
\begin{verbatim}
M <- read.table("C:/.../myData.asc", header = FALSE)
\end{verbatim}
where the first argument contains the path to your data file. If the columns are separated for instance by \verb|";"| (instead of space) we have to add to the arguments \verb|sep = ";"|. Note also that the assignment operator in R is an arrow \verb|<-| and not \verb|=|. We have assumed that the data table had no header. In this case column names can added by the command,
\begin{verbatim}
names(M) <- c("colName1", ..., "colNameN")
\end{verbatim}
if there are $N$ columns. You can then refer, for instance, to the first column as \verb|M$colName1|. The other way to extract the first column is as \verb|M[,1]|. One can inspect the content of the table read into \verb|M| by the command \verb|edit(M)|. Finally you can quit R by the command \verb|q()|. Instead of typing and running the commands at the console in a sequence, you can collect them in a script file (an ordinary text file) with the extension R, named say \verb|myScript.R|, and run it by the command \verb|source("C:/.../myScript.R")| where the argument contains the path to your script file.
\section{Analysing data}
\subsection{Plotting data}
The next thing you want to do is to inspect your data. For a quick plot of the data in the example above use the command,
\begin{verbatim}
plot(M$colName1, type = "l")
lines(M$colName2, type = "l")
\end{verbatim}
The second line adds the graph of the second column to the same plot. The \verb|plot| function has various arguments for controlling colours, titles, scales, and so on. Information about the function \verb|plot| is obtained by the command \verb|?plot|. An interesting feature is that with a few lines of code it is for example possible to plot histograms/graphs of hundreds of variables and print them to a pdf document for quick browsing on the computer.
After these preliminaries we can move on to the processing of data. In this connection we must explain how we define new functions in R.
\subsection{Basic signal processing -- some examples}
\subsubsection{User defined functions}
In most programming tasks it us convenient to able to define your own functions. We give a simple example how it works in R. Let us say that we want to sum the elements of vectors using a function \verb|sumVec(V)| which returns the sum and mean of a numeric vector \verb|V|. This can be defined in R by:
\begin{lstlisting}
sumVec <- function(V, start = 1){
n <- length(V)
sm <- 0
for(i in start:n){
sm <- sm + V[i]
}
mn <- sm/(n - start + 1)
return(list(sum = sm, mean = mn))
}
\end{lstlisting}
We can call this as \verb|result <- sumVec(V)| which puts the sum and mean of the elements of the vector \verb|V| into to a variable \verb|result|. Here we have also demonstrated the very useful \verb|list| structure in R. The sum (\verb|sm|) and mean (\verb|mn|) are put into a list with respective (arbitrary) names \verb|sum| and \verb|mean|. The variables are then referenced as \verb|result$sum| and \verb|result$mean| after the call
\begin{lstlisting}
result <- sumVec(V).
\end{lstlisting}
It is an advantage to have a function return a \verb|list| since it is then easy to modify the function by adding new variables to the output \verb|list| without affecting previous uses of the function. The function example also illustrates an other aspect of R function; that is, we may have arguments with default values, like \verb|start| as in the example. If the argument is not listed, as in
\begin{lstlisting}
result <- sumVec(V),
\end{lstlisting}
it will use the default value (\verb|start = 1|). In many functions we have set as default \verb|plot = FALSE| for a variable \verb|plot|, which means that the results will not plotted unless one adds \verb|plot = TRUE| to the arguments.
Functions can also have other functions as arguments, as for example in the case of \verb|EMG_spec| below. This example also illustrates the similarity with the common C and \textsc{Java} syntax. User defined functions can be collected in a separate file \verb|myFunctions.R| which can be made active (loaded into the workspace) by calling it using
\begin{verbatim}
source("C:/.../myFunctions.R")
\end{verbatim}
This corresponds to the \verb|include| statement of header files and source files in C programming. The EMG processing functions to be described below are collected in the file \verb|EMGfuns.R| in the supplementary material to this paper. The hash-symbol \verb|#| is used for comment lines in the R scripts.
\subsubsection{Simulated data}
Is useful to have access to various test data, and as example we have implemented a a standard algorithm \citep[pp.70-71]{Hermens1999} in \verb|EMG_sim| which generates simulated EMG data of desired length. This function returns a \verb|list| where the data is contained in the component \verb|sim|. For instance one can probe the spectrum function using the test data as follows:
\begin{lstlisting}
mysim <- EMG_sim(3000)
myspec <- EMG_spec(mysim$sim, plot = TRUE)
\end{lstlisting}
\subsubsection{Rectification, RMS and turns}
Rectifying raw data means simply taking the absolute value of the elements. For EMG data it may be useful to first subtract any offset (bias) from the raw data. Thus the rectification of \verb|V| could be written as \verb|abs(V - mean(V))| and the \verb|EMG_rect| function in \verb|EMGfuns.R| becomes very simple. It is noteworthy that effect of the rectification of EMG is similar to the rectification of AM radio waves whose purpose is to enhance the low frequency components which encode the voice signals. For EMG the low frequency ''voice'' part corresponds to the encoded force \citep{Borg2007}.
A bit less trivial from the programming point of view is the \verb|EMG_rms| function which computes the Random Mean Square of the EMG and can basically be represented as
\begin{equation}
\sqrt{ \frac{1}{\Delta T} \int_{t - \Delta T/2}^{t + \Delta T/2} EMG(u)^2 du}.
\end{equation}
To write these sorts of filtering or enveloping functions it is convenient to use the built-in \verb|filter(V, filtc, ...)| function. This takes input data \verb|V| and outputs a vector with elements
\begin{equation}
y[i] = \mathrm{filtc}[1] \cdot V[i+o] + \cdots + \mathrm{filtc}[p] \cdot V[i+o-(p-1)],
\end{equation}
where \verb|filtc[i]| are the filter coefficients. The offset $o$ depends on the argument \verb|sides| such that \verb|sides = 2| corresponds to zero lag with $o = (p-1)/2$ if $p$ is odd. For the moving average one uses \verb|method = "convolution"|. The function \verb|EMG_rms| has an argument \verb|DT| which determines the size in milliseconds of the moving window over which one calculates the RMS. This function also illustrates a special feature of R: the use of the \verb|'...'| argument.
\begin{lstlisting}
EMG_rms <- function(V, sampFreq = 1000, DT = 250, plot = FALSE, ...)
{ #part of the function declaration
rectV <- sqrt(filter(rectV^2, filter1, sides = 2,
method = "convolution", ...))
#rest of the function declaration
}
\end{lstlisting}
In this case it means that we can pass arguments to the \verb|filter| function employed by \verb|EMG_rms|. For instance, \verb|filter| has an argument called \verb|circular|, and via \verb|EMG_rms(..., circular = TRUE)|, we can pass a value (\verb|TRUE| in this case) to this argument of the function \verb|filter|.
A version of the classical method of counting turns \citep{Willison1963,Willison1964} is implemented by the function \verb|EMG_turns| and it also uses a moving window \verb|DT| over which one sums the number of turns. The implementation \verb|EMG_wturns| is maybe closer to the original idea by Willison but the practical difference between the two seems small. The turns functions return a structure
\begin{verbatim}
list(turns.ps = turns_per_sec, turns.where = turns).
\end{verbatim}
The variable \verb|turns.where| contains the time indexes where the turns are counted. (This is also shown in the plot of the function.) This data may be of interest when, for instance, one wants to calculate an entropy metric for the signal.
\subsubsection{Time-frequency domain}
Part of the inspection of the EMG signal is to study its frequency properties. This is typically performed by calculating the power spectrum of the data. For this purpose one subdivides the original times series into blocks of some time length $\Delta T$, then calculates the power spectrum for these and take their mean as the final power spectrum. For the subdivision one normally use a 50 \% overlap which further suppresses the variance of the final spectrum estimate \citep{Press2002}. This method is implemented in the function \verb|EMG_spec|. It uses a default windowing of the data by the filter \verb|filtWelch|. The window size $\Delta T$ is given by the argument \verb|DT| in milliseconds. The nominal frequency resolution is then given by $\Delta f = 1/ \Delta T $. The function outputs a \verb|list| with the power density estimate in \verb|psd|, which also contains mean (MNF) and median frequency (MDF) in \verb|meanf| and \verb|medianf|. The time-frequency methods naturally rely on the \emph{Fast Fourier Transformation} (FFT) which in R is called by \verb|fft|. Using the \emph{Short Time Fourier Transformation} (STFT), which applies the FFT to subintervals of the time series, we can, for instance, calculate how the mean frequency varies with time. This is implemented by the function \verb|EMG_stft_f|. One use of mean/median frequency is for the study of muscle fatigue as a function of time, which is often associated with a decrease in mean frequency \citep{Lindstrom1977}.
Using the \verb|fft| transform we can filter EMG signals by suppressing the higher frequency components. The method employed in \citep{Borg2007} is here implemented by the function \verb|EMG_bw0|. It first calculates the Average Rectified Value, then applies \verb|fft| which gives the Fourier coefficients $c(f_k)$. These are multiplied by a filter factor,
\[
c(f_k) \mapsto {\tilde c}(f_k) = \frac{c(f_k)}{1 + \left(\frac{f_k}{f_c}\right)^n},
\]
where $f_c$ is the low pass cut-off frequency of the filter and $n$ is the order of the filter. Finally we obtain the filtered signal by applying the inverse \verb|fft| to ${\tilde c}(f_k)$. This is basically a zero-lag version of the Butterworth filter. With $n = 4$ and $f_c = 1$ Hz we obtained quite a good correspondence between gastrocnemius EMG and the muscle force as expressed by the anterior-posterior COP (center of pressure) during quiet standing. We have also implemented the filter corresponding to a second order critically damped system,
\[
c(f_k) \mapsto {\tilde c}(f_k) = \frac{c(f_k)}{\left( 1 + i\frac{f_k}{f_c}\right)^2},
\]
which is one of the basic models for the EMG-to-force transfer functio7n \citep{Soechting1975}. Note that the function \verb|EMG_crit2| too rectifies the EMG before filtering.
For comparison of EMG vs EMG, or (filtered) EMG vs force etc, the correlation methods are essential. Given two time series $x$ and $y$ we may define a correlation function $c_{xy}(t)$ by,
\[
c_{xy}(t) = \frac{\int_0^T {\tilde x}(u) {\tilde y}(u + t) du}{\sqrt{\int_0^T {\tilde x}(u)^2 du} \sqrt{\int_0^T {\tilde y}(u)^2 du}},
\]
where ${\tilde x}$ is $x$ with the mean value $\bar x$ subtracted, ${\tilde x}(t) = x(t) - \bar x$, etc. In the discrete version this is implemented by \verb|EMG_corr| employing \verb|fft| methods. In the frequency domain a coherence function is defined by,
\[
\label{EQ:Coherency}
\text{coh}_{xy}(f) = \frac{\langle \hat{x}^\star(f) \hat{y}(f)\rangle}{\sqrt{\langle |\hat{x}(f)|^2\rangle} \sqrt{\langle |\hat{y}(f)|^2\rangle}},
\]
where $\langle \cdots \rangle$ denotes statistical averaging. This is estimated by \verb|EMG_coh| by dividing the time series into time slices of size \verb|DT| and calculating the Fourier coefficients for these slices, and finally take the averages over the blocks. \verb|EMG_corr| and \verb|EMG_coh| can be used to investigate time lags and phase shifts between signals. In the case that we have a linear relationship, $y = H \star x + n$, with a transfer function $H$ (and uncorrelated ''noise'' $n$), we would get
\[
\mathrm{coh}_{xy}(f) = \frac{\hat H(f)}{|\hat H(f)|} = e^{i \phi(f)},
\]
where $\phi(f)$ is the phase function of the transfer function. A related time shift $\tau$ can then be obtained from $2 \pi \tau = d\phi(f)/df$.
\subsubsection{Frequency band analysis}
As is well known from musical transcription, it is convenient to study sound by how the power (intensity) is distributed over the frequency bands (pitch) as a function of time. This is useful also in basic signal analysis. The idea is to decompose signals using filter banks. Wavelets can be considered as a special realization of the idea of filter banks \citep{Vetterli1995}. An interesting hybrid method for ''intensity analysis'' of EMG has been proposed by \citep{Tscharner2000}. The idea is to divide the frequency band of interest, say one from 10 Hz to 200 Hz, into subbands centered on frequencies $f_c^{(j)}$ ($j = 1, \cdots , J$) such that the relative bandwidth $BW = \Delta f(j)/f_c^{(j)}$ scales as $1/\sqrt{f_c^{(j)}}$ over the frequency band. Here $\Delta f(j)$ is the frequency resolution of the ''mother wavelet'' $\psi$ at the center frequency $f_c^{(j)}$. In this way one can provide a distribution of signal power among the frequency bands. The power in the frequency band $j$ at time $t$ is given by $\left|c_j\right|^2$ where,
\[
c_{j}(t) = \int \bar{\psi}_j (u - t) x(u) du,
\]
and ${\psi}_j$ is the wavelet centered at $f_c^{(j)}$. This differs from the recipe in \citep{Tscharner2000} but that is mainly because we use here the full complex coefficient. We present here a modification \citep{Borg2003} which is based on the Morlet function which in frequency space is given as,
\[
\hat{\psi}(f_c, \alpha, f) = \exp\left( - \frac{2 \pi^2}{\alpha f_c} (f - f_c)^2 \right).
\]
For the center frequencies we select, following von Tscharner,
\[
f_c^{(j)} = \frac{1}{s}(q+j)^2 \quad (j = 0, \cdots, J-1),
\]
determined by the parameters $s$ (''scale'') and $q$. The implementation is given by \verb|EMG_morvt| which is again based on using the \verb|fft| transformation. This function returns a \verb|list| where \verb|powc| refers to the matrix of the $c^{(j)}[i]$ coefficients, \verb|freqc| to the array with the center frequencies, and \verb|freqm| contains an estimate of the instantaneous mean frequency calculated as the average of the center frequencies weighted with the power coefficients $\left|c_j\right|^2$.
\subsubsection{Multi resolution analysis, MRA}
In the above examples we have relied on the basic libraries that belong the the default setup of the R system. In the next example we will take advantage of a library that provides functions for discrete wavelet analysis. There is for instance a package appropriately named \verb|wavelets| (by Erich Aldrich). In order to install it one enters the command
\begin{verbatim}
install.packages("wavelets")
\end{verbatim}
which will look up a depository and ask you to download the package. When successfully installed it can be loaded by the command
\begin{verbatim}
library(wavelets)
\end{verbatim}
The command \verb|library()| with empty argument will show the packages installed on your system. Information about the package \verb|wavelets| can be obtained by the command
\begin{verbatim}
help(package = "wavelets")
\end{verbatim}
or \verb|??wavelets|. In multi resolution analysis (MRA) we repeatedly apply low- and high-pass filters to a discrete time series which thus can be decomposed into fine and coarse grained parts. The simplest example is the Haar filter. If $x = (x_1, x_2, \cdots)$ then Haar low pass and high pass filters produce the series,
$a = (a_1, a_2, \cdots)$ and If $b = (b_1, b_2, \cdots)$, with
\begin{align*}
&a_i = \frac{x_{2i} + x_{2i-1}}{\sqrt{2}}, \\
&b_i = \frac{x_{2i} - x_{2i-1}}{\sqrt{2}}. \\
\end{align*}
The averaging procedure produces a coarse grained sum version $a$, while $b$ contains the detail. Symbolically the decomposition can be written $x = (a|b)$. This procedure can be repeated taking $a$ as an input for the decomposition procedure. In this fashion we obtain
\[
x = (a^J|b^J|b^{J-1}| \cdots |b^1),
\]
for a decomposition of order $J$. The $k$:th level detail coefficients $b^k$ represent information about the changes in the times series on a time scale proportional to $2^k$.
The function \verb|EMGx_mra()| is a wrapper for \verb|mra| in the \emph{wavelet} package. A new feature here is that the function \verb|mra| returns a \verb|class| object with \verb|slot|s whose names can be accessed by the function \verb|slotNames|. For instance the detail coefficients have the name \verb|D| and the sum coefficients the name \verb|S|. If
\begin{verbatim}
res <- mra(X)
\end{verbatim}
then the vectors with the coefficients are accessed as
\begin{verbatim}
res@D[[j]], and res@S[[j]],
\end{verbatim}
for the level $j$. The original data can be obtained as a sum of the decomposition,
\begin{verbatim}
X = res@D[[1]] + res@D[[2]] + ... + res@D[[J]] + res@S[[J]].
\end{verbatim}
Thus \verb|res@D[[j]]| reflects the signal content on a time scale of the order $2^j \cdot f_s^{-1}$ where $f_s$ is the sampling rate.
\subsubsection{Batch processing}
As the number of data files grow it is important to be able to process them in one row. This kind of batch processing can be simply implemented in R. We will assume that have a set of data files \verb|name1.asc|, ... , \verb|nameN.asc|. One can collect these paths of these files into \verb|filelist.asc| and write a R-script which opens each of these files for processing. One thing to remember is that the file paths must be on the Unix format using \verb|/| (or \verb|\\| ) instead of \verb|\|. The files can also be selected interactively by using the \verb|tk_choose.files()| function,
\begin{lstlisting}
library(tcltk) # load the tcltk package
Filters <- matrix(c("EMG data", ".asc", "All files", "*"),
2, 2, byrow = TRUE)
if(interactive()) filelist <- tk_choose.files(filter = Filters)
\end{lstlisting}
This will open the ''Select files'' dialogue and put the selected files into the \verb|filelist| variable (with file paths on the Unix format). The following snippet is a simple example which opens the files in the \verb|filelist| and plots the first column to a pdf-file, and writes the standard deviation to a text-file.
\begin{lstlisting}
# filelist -- contains paths to asci files with EMG data
outputpdf <- "C:/EMGanalysis.pdf" # output graphs to this file
outputtxt <- "C:/EMGanalysis.txt" # output text/numbers to this file
pdf(outputpdf) # starts the pdf driver and opens the output pdf-file
fp <- file(outputtxt, "w") # opens text-file for writing
n <- length(filelist)
for(i in 1:n){
EMG <- read.table(filelist[i], header = FALSE)
title <- paste("Data from ", filelist[i])
# this one goes to the text file -->
cat("Standard deviation = ", sd(EMG$V1),
" for data EMG$V1 in file ",
filelist[i], "\n", file = fp)
# this one goes to the pdf file -->
plot(EMG$V1, main = title, type = "l")
}
close(fp) # closes the text file
dev.off() # closes the pdf driver
\end{lstlisting}
It illustrates how one reads the data, opens a file for writing, where the writing to the file is performed using the function \verb|cat|. (It computes the standard deviation of the time series using the function \verb|sd| and writes it to the file.) This example is easily generalized to more complicated processings.
\subsection{Statistics}
R is by definition a statistics software package whence all the well-known, and many less well-known, statistical procedures are implemented. Important sub\-topics are descriptive statistics, statistical testing, and modeling data. Since our emphasis here is on signal processing we will not go into the statistical methods. At the very basic level we have, for instance, \verb|hist(X)| which computes and plots a histogram of numerical data \verb|X|, while \verb|plot(ecdf(X))| first calculates the empirical cumulative distribution function (ecdf) and the plots the result. The \emph{Student test} is performed by \verb|t.test| and, for instance, \verb|t.test(X, mu = 2)| computes the $p$-value for the mean to differ from 2, and the 95\% confidence interval for the mean of \verb|X|.
For an introduction to statistical analysis using R we recommend \cite{Everitt2010} which is provided with a R-package \verb|HSAUR2| that contains the codes and the data sets.
\section{Conclusions}
We have given a brief introduction to some basic features of the R software used as a tool for analyzing and displaying EMG data, and biosignal data in general. Using R it is easy to document the exact procedures employed in analyzing the data so that it can be replicated by other researchers. A next level would be to develop a dedicated \textsc{Remg} package with tools covering various aspects of EMG and related kinesiological data and signals (MMG, ECG, etc). Such a package could be supplied with a representative set of data for testing and demonstrating the analysis methods. Finally we would also like to emphasize the usefulness of R in teaching basic data processing and visualization methods to biomechanics students.
\begin{comment}
\section*{Conflict of interest statement}
The author acknowledges that he does not have any financial or
personal relationships with other people or organizations that
would inappropriately influence the results of this study.
\end{comment}
\section*{Acknowledgments}
The author thanks Maria Finell for gathering the EMG- and balance data used in the examples (supplementary material). He is also indebted to W Jeffrey Armstrong for exchanges on the ''intensity analysis'' which has resulted in an update of the \verb|EMGfuns.R| file.
\section*{Supplementary material}
The file \verb|data.bal| contains quiet standing balance COP-data in ASCI format. First column contains COP X, second column contains COP Y, both in millimeters. The third column contains total vertical force (Newton). The columns are \verb|tab|-separated (\verb|\t|). Sampling rate is 100 Hz. The file \verb|data.emg| contains the EMG-data in ASCI format. The columns contain data from the muscles Tibialis anterior (right), Lateral Gastrocnemius (right), Medial Gastrocnemious (right), Tibialis anterior (left), Lateral Gastrocnemius (left), Medial Gastrocnemious (left), all sampled at 1000 Hz. The file \verb|emg_analysis.R| is the R-script which demonstrates a few basic analyzing methods with the EMG and the balance data. The script either produces a \verb|pdf| report (\verb|result.pdf|), or shows the results on the console, depending on the setting of the variable \verb|report| (\verb|TRUE| or \verb|FALSE|). The file \verb|EMGfiles.R| contains a script which demonstrates how to set up batch processing. The file \verb|EMGfuns.R| contains the basic R-scripts (functions) for analyzing EMG which are employed by \verb|emg_analysis.R|. The files are by default assumed to reside in the archive \verb|C:/EMGR|. The files can be downloaded as \url{http://terra.chydenius.fi/~frborg/emg/EMGR.zip}, and also from the arXiv site as supplementary material.
|
\section{Introduction}
Dynamics in strong (color) electromagnetic fields
has been an interesting subject in theoretical physics
\cite{Heisenberg, Schwinger, Schwinger3, Suganuma, Tanji, Iwazaki, Hidaka}.
Recently, this subject is being paid attention also
in high-energy heavy-ion collision experiments.
At the so-called Glasma stage~\cite{Glasma} just after the collision,
longitudinal color-electric and color-magnetic fields are expected to be
produced in the context of the color glass condensate
of order $1$--$2$GeV in RHIC and $5$GeV in LHC.
In the peripheral collision, a strong magnetic field of order $100$ MeV
would be also induced.
Such strong fields will play important role of particle productions
and thermalization to the quark gluon plasma.
In this paper, we concentrate on how the strong fields decay into particles.
For this purpose, we consider the Schwinger mechanism for quarks
in the coexistence of color-electric and color-magnetic fields.
We will point out that, in the case of massless quarks,
the vacuum immediately decays
in general homogeneous (color) electromagnetic fields.
\section{Vacuum decay in color electromagnetic field}
To simplify the situation of particle creations, we consider
{\it covariantly constant} color electromagnetic fields~\cite{Suganuma} of
$[D_\mu, F_{\alpha \beta}]=0$, i.e.,
$[D_\mu,\bm{E}]=[D_\mu,\bm{B}]=\bm{0}$,
where $D_\mu \equiv \partial_\mu-ig A_\mu$ is the covariant derivative
with the gauge field $A_\mu$.
We here set $g>0$.
The color electric and magnetic fields are defined as
$\bm{E}^i=F^{i0}$ and $\bm{B}^i=-\epsilon^{ijk}F_{jk}/2$
with $F_{\mu\nu}\equiv i[D_{\mu},D_{\nu}]/g$.
This is a gauge-covariant generalization of constant fields in QED,
$\partial_{\mu} \bm{E}=\partial_{\mu} \bm{B}=\bm{0}$,
to the non-Abelian fields, and the translational invariance
leads to this condition.
For the covariantly constant fields, all the components of
$\bm{E}$ and $\bm{B}$ can be diagonalized to be constant matrices
in color space by a gauge transformation.
Without loss of generality,
one can set $\bm{E}=(0,0,E)$ and $\bm{B}=(0,0,B)$
in a suitable Lorentz frame.
Here, $B$ and $E$ can be written as Lorentz scalar defined by
$B, E \equiv [({\cal F}^2+{\cal G}^2)^{1/2} \pm {\cal F}]^{1/2}$
\cite{Suganuma}
with ${\cal F} \equiv F_{\mu\nu}F^{\mu\nu}/4=(\bm{B}^2-\bm{E}^2)/2$
and ${\cal G} \equiv F_{\mu\nu}\tilde F^{\mu\nu}/4=\bm{B}\cdot\bm{E}$.
Like the vacuum decay in QED~\cite{Heisenberg,Schwinger},
we consider the vacuum persistency probability,
\begin{equation}
|\langle\Omega_\mathrm{out}|\Omega_\mathrm{in}\rangle|^2=\exp(-VTw) ,
\end{equation}
where $V$ and $T$ are infinite space volume and time length.
$|\Omega_\mathrm{in}\rangle$ and $|\Omega_\mathrm{out}\rangle$ are
the in-vacuum and the out-vacuum defined
at $t=-T/2$ and $t=T/2$, respectively.
$w$ is called ``pair-creation rate'', and
denotes magnitude of the vacuum decay per unit space-time volume.
The analytic formula of $w$ for the quark-pair creation
with the quark mass $m$
in the covariantly constant color fields is obtained as \cite{Suganuma,Hidaka}
\begin{equation}
\begin{split}
w
&=\sum_{N=1}^\infty \mathrm{tr} \left\{\frac{g^2 EB}{4\pi^2}
\frac{1}{N}\coth(\pi N BE^{-1}) e^{-N\pi m^2/gE} \right\} \\
&=\sum_{s_z=\pm 1/2}\sum_{n=0}^\infty \mathrm{tr}
\left\{ \,\frac{g^2 EB}{4\pi^2}
\ln\frac{1}{1-\exp{(-\pi m_T^2(n,s_z)/gE})}\right\},
\end{split}
\label{eq:wqcd}
\end{equation}
where $m_T(n,s_z)\equiv \{2gB(n +1/2\mp s_z)+m^2\}^{1/2}$
is the transverse energy, $n$ labels the Landau level,
and $s_z=\pm1/2$ denotes the spin component.
The trace is taken over the indices of color and flavor.
This is the non-Abelian extension of the formula for QED.
(On the quark mass $m$, chiral symmetry is restored by
strong (color) electric field \cite{Suganuma}.)
The lowest Landau level (LLL) with $n=0$ and $s_z=1/2$ gives
$m_T^2(n,s_z)=m^2$, which is independent of $B$, while
$m_T^2(n,s_z)$ of higher levels includes $O(gB)$ term.
Hence, the LLL state gives the dominant contribution for $gB\gg m^2$.
The LLL contribution for small $m$ is
\begin{equation}
w=\mathrm{tr}\left\{ \, \frac{g^2 EB}{4\pi^2}
\ln \frac{1}{1-\exp(-\pi m^2/gE)} \right\}
\simeq \mathrm{tr}
\left\{\,\frac{g^2EB}{4\pi^2}\ln\frac{gE}{\pi m^2} \right\}.
\end{equation}
As $m\to0$, $w$ diverges as $w \propto -\ln m\to \infty$.
Note that $w$ diverge, only when the (color) magnetic field coexists, and
the Landau quantization causes the divergence of $w$~\cite{Hidaka}.
\section{Effective theory in a strong magnetic field}
When the (color) magnetic field is strong enough,
the LLL dominates and hence the LLL projected theory is applicable,
because of large $m_T(n,s_z)$ for higher Landau levels.
As a remarkable merit of the LLL projected theory,
we can take into account the back reaction of created particle pairs.
In this section, we investigate the LLL projected theory
in QED or Abelianized QCD,
since the fermion-pair creation formalism in the covariantly constant fields
in QCD has Abelian nature and is similar to the Schwinger mechanism in QED.
The wave function of the fermion is projected to the LLL state, and
the LLL wave function in a symmetric gauge of ${\bf A} =(-yB/2,xB/2,0)$ is
\begin{equation}
\phi_l(x,y)=\sqrt{\frac{eB}{2\pi l!}}
\left(\frac{eB}{2}\right)^{\frac{l}{2}}(x+iy)^l
\exp\left(-\frac{eB}{4}(x^2+y^2)\right),
\end{equation}
where $l (\geq 0)$ denotes the angular momentum in $z$ direction for the LLL,
and the energy is degenerate on $l$.
The projected wave function can be written as
\begin{equation}
\psi(x,y,z,t)=
\begin{pmatrix}
\sum_l \phi_l(x,y)\varphi_l(t,z)\\
0
\end{pmatrix} ,
\end{equation}
in a suitable representation,
where $\varphi_l(t,z)$ is the two component Dirac field in 1+1 dimensions.
Then the fermion action of QED in $3+1$ dimensions reduces to
that of non-Abelian gauge theory in $1+1$ dimensions:
\begin{equation}
S=\int d^4x \bar{\psi}(x)i\gamma^\mu D_\mu\psi(x)
\simeq
\sum_{l,l'}\int dtdz \bar{\varphi}_{l'}(t,z)i\tilde{\gamma}^\mu \tilde{D}^{l'l}_\mu\varphi_{l}(t,z),
\label{eq:action}
\end{equation}
where $\tilde{\gamma}^{~t}$ and $\tilde{\gamma}^{~z}$ are the gamma matrices
in $1+1$ dimensions and $\tilde{\gamma}^{~x}=\tilde{\gamma}^{~y}=0$.
The covariant derivative is defined by $\tilde{D}^{l'l}_\mu
=\delta^{l'l}\partial_\mu-ie \tilde{A}^{l'l}_\mu$ with
\begin{equation}
\tilde{A}^{l'l}_\mu(t,z)=\int dxdy\phi_{l'}^*(x,y)\phi_l(x,y) A_\mu(x,y,z,t).
\label{eq:gaugeField}
\end{equation}
$\tilde{A}^{l'l}_\mu(t,z)$ turns out the gauge field
in $U({\cal N})$ gauge theory with
${\cal N} \equiv eBV_\perp/(2\pi)(=+\infty)$,
since $\tilde{A}^{l'l}_\mu(t,z)$ is an Hermite matrix,
$\tilde{A}^{*l'l}_\mu(t,z)=\tilde{A}^{ll'}_\mu(t,z)$,
and the indices $l$ and $l'$ run from $0$ to ${\cal N}-1$. $V_\perp$
is the volume of the transverse directions.
While the original gauge field lives in 3+1 dimensions
as Abelian gauge field,
the gauge field in the effective fermion action
is reduced to non-Abelian gauge field in 1+1 dimensions.
The gauge field can be decomposed to $U(1)$ and $SU({\cal N})$ parts,
which are $\tilde{A}_\mu(t,z) \equiv
\sum_l\tilde{A}^{ll}_\mu(t,z)/{\cal N}$, and $A^{ll'}_\mu(t,z)
\equiv \tilde{A}^{ll'}_\mu(t,z)-\delta^{ll'}\tilde{A}_\mu(t,z)$, respectively.
For the homogeneous system on the transverse directions, $x$ and $y$,
$A_t$ and $A_z$ can be taken to be independent of $x$ and $y$ in a gauge,
so that the non-Abelian part vanishes.
Then, the effective fermion action in Eq.~(\ref{eq:action}) becomes
\begin{equation}
S\simeq {\cal N}\int dtdz
\bar{\varphi}(t,z)i\tilde{\gamma}^\mu \tilde{D}_\mu\varphi(t,z) ,
\label{eq:action2}
\end{equation}
where $\varphi_l(t,z)$ is rewritten as $\varphi(t,z)$.
This action is nothing but that in 1+1 QED, i.e., the Schwinger model.
The exact solution of the effective action for the fermion is known as
\begin{figure}[tb]
\centering
\includegraphics*[width=0.42\linewidth]{electricField.eps}
\centering
\includegraphics*[width=0.42\linewidth]{numberDensity.eps}
\caption{
The electric field $E(t,z)$ (left) and
the number density $N(t,z)$ of created fermions (right)
at $z=0$ plotted against time $t$.
Both values are normalized by that at maximum values.}
\label{fig:dynamics}
\end{figure}
\begin{equation}
\Gamma(A_\mu)= -\frac{m_\gamma^2V_\perp}{2}\int dtdz \tilde{A}_{\mu}(t,z)
\left(g_\parallel^{\mu\nu}-\frac{\partial_\parallel^\mu\partial_\parallel^\nu}{\partial_{\parallel}^2}\right)\tilde{A}_\nu(t,z),
\label{eq:effectiveAction}
\end{equation}
where $\parallel$ denotes for longitudinal directions, $t$ and $z$.
$m_\gamma$ is the effective photon mass,
$m_\gamma^2\equiv e^3B/(2\pi^2)$.
Equation~(\ref{eq:effectiveAction}) is manifestly gauge invariant.
In the Lorenz gauge, it reduces to the result by \cite{Schwinger3}.
The mass $m_\gamma$ is induced by the axial anomaly, whose effect is called
``dynamical Higgs effect.''
From Eq.~(\ref{eq:effectiveAction}), the fermion and axial currents read
\begin{equation}
j^\mu(x)=\frac{\delta\Gamma(A)}{\delta (eA_\mu)}=-\frac{e^{2}B}{2\pi^{2}}\tilde{A}^\mu(t,z), \qquad
j_5^\mu(x)=-\epsilon^{\mu\nu}j_\nu(x),
\end{equation}
where we choose the Lorenz gauge, $\partial_\mu \tilde{A}^\mu=0$.
The divergence of the axial current leads to
the axial anomaly in $1+1$ dimensions
except for the overall factor $eB/(2\pi)$:
$\partial_\mu j_5^\mu(x)=e^2B\epsilon^{\mu\nu}\partial_{\mu}\tilde{A}_{\nu}(t,z)/(2\pi^2)
=e^2BE/(2\pi^2).$
This relation is nothing but axial anomaly in $3+1$ dimensions.
Since the effective action (\ref{eq:effectiveAction})
has a quadratic form in $\tilde{A}_{\mu}$,
the equation of motion for the photon can be solved.
For example, the electric field of $z$ direction is
$E=E_0\cos(\omega t-k_zz )$,
with $\omega=(k_z^2+m_\gamma^2)^{1/2}$. The currents satisfy
$ej^\mu=-\epsilon^{\mu\nu}\partial_\nu E$ and
$ej_5^\mu=\partial^\mu E$.
We show in Fig.~\ref{fig:dynamics}
the electric field $E$ and the particle number density $N$
as functions of time $t$, for the spatially homogeneous case, $k_z=0$.
The electric field oscillates with a frequency $\omega$.
In this case, $j^{~t}=0$, but $j^{~t}_5\neq0$.
The number density of created fermions is $|j_5^{~t}|$.
These results agree with the previous works~\cite{Tanji,Iwazaki}.
Since $m_\gamma^2$ is the only scale in this theory,
the time scale of the particle decay is $\sim 1/m_\gamma$.
The generalization to non-Abelian theories is straightforward.
The fermion determinant becomes the Wess-Zumino-Witten action.
\section{Summary and Concluding Remarks}
In this paper, we have studied the vacuum decay
in homogeneous color electromagnetic fields.
When the fermion is massless,
the vacuum persistency probability per unit space-time volume becomes zero,
and hence $w$ diverges. This divergence originates from
spectral discretization of transverse directions and the lowest Landau level.
With the LLL projection,
we have analytically calculated the effective action in this situation.
\bibliographystyle{aipproc}
|
\section{Introduction}
\label{sec:intro}
\subsection{Background}
\label{subsec:RWDRE}
\emph{Random walk in random environment} (RWRE) has been an active area of research for
more than three decades. Informally, RWRE's are random walks in discrete or continuous
space-time whose transition kernels or transition rates are not fixed but are random
themselves, constituting a random environment. Typically, the law of the random environment
is taken to be translation invariant. Once a realization of the random environment is
fixed, we say that the law of the random walk is \emph{quenched}. Under the quenched law,
the random walk is Markovian but not translation invariant. It is also interesting to
consider the quenched law averaged over the law of the random environment, which is called
the \emph{annealed law}. Under the annealed law, the random walk is not Markovian but
translation invariant. For an overview on RWRE, we refer the reader to Zeitouni~\cite{Ze1,Ze2},
Sznitman~\cite{Sz1,Sz2}, and references therein.
In the past decade, several models have been considered in which the random environment
itself evolves in time. These are referred to as \emph{random walk in dynamic random
environment} (RWDRE). By viewing time as an additional spatial dimension, RWDRE can be
seen as a special case of RWRE, and as such it inherits the difficulties present in RWRE
in dimensions two or higher. However, RWDRE can be harder than RWRE because it is an interpolation
between RWRE and homogeneous random walk, which arise as limits when the dynamics is slow,
respectively, fast. For a list of mathematical papers dealing with RWDRE, we refer the
reader to Avena, den Hollander and Redig~\cite{AvdHoRe2}. Most of the literature on RWDRE
is restricted to situations in which the space-time correlations of the random environment
are either absent or rapidly decaying.
One paper in which a milder space-time mixing property is considered is Avena, den Hollander
and Redig~\cite{AvdHoRe1}, where a law of large numbers (LLN) is derived for a class of
one-dimensional RWDRE's in which the role of the random environment is taken by an
\emph{interacting particle system} (IPS) with configuration space
\be{}
\Omega:=\{0,1\}^{\mathbb{Z}}.
\end{equation}
\vspace{-0.3cm}
\begin{figure}[htbp]
\begin{center}
\setlength{\unitlength}{0.4cm}
\begin{picture}(12,2)(-6,1)
\put(-7.8,0){$x-1$}
\put(-4.7,0){$x$}
\put(-3.4,0){$x+1$}
\put(-4.7,1.2){$0$}
\put(-4.3,2){\vector(1,0){1.6}}
\put(-4.7,2){\vector(-1,0){1.6}}
\put(-3.6,2.5){$\beta$}
\put(-5.9,2.5){$\alpha$}
\put(0.7,0){$x-1$}
\put(3.8,0){$x$}
\put(5.1,0){$x+1$}
\put(3.8,1.2){$1$}
\put(4.2,2){\vector(1,0){1.6}}
\put(3.8,2){\vector(-1,0){1.6}}
\put(4.8,2.5){$\alpha$}
\put(2.6,2.5){$\beta$}
\end{picture}
\end{center}
\vspace{0.2cm}
\caption{Jump rates of the $(\alpha,\beta)$-walk on top of a hole ($=0$), respectively,
a particle ($=1$).}
\label{fig-jumprates}
\end{figure}
\noindent
In their paper, the random walk starts at $0$ and has transition rates as in Fig.~\ref{fig-jumprates}:
on a \emph{hole} (i.e., on a $0$) the random walk has rate $\alpha$ to jump one unit to the left and rate $\beta$
to jump one unit to the right, while on a \emph{particle} (i.e., on a $1$) the rates are
reversed (w.l.o.g.\ it may be assumed that $0<\beta<\alpha<\infty$, so that the random
walk has a drift to the left on holes and a drift to the right on particles).
Hereafter, we will refer to this model as the $(\alpha,\beta)$-model.
The LLN is proved under the assumption that the IPS satisfies a space-time mixing property called
\emph{cone-mixing} (see Fig.~\ref{fig-cone}), which means that the states inside a space-time cone are almost
independent of the states in a space plane far below this cone. The proof uses a
regeneration scheme originally developed by Comets and Zeitouni~\cite{CoZe} for RWRE
and adapted to deal with RWDRE. This proof can be easily extended to $\mathbb{Z}^d$, $d \geq 2$,
with the appropriate corresponding notion of cone-mixing.
\vspace{0.5cm}
\begin{figure}[htbp]
\begin{center}
\setlength{\unitlength}{0.3cm}
\begin{picture}(12,10)(-8,1)
\put(-6,0){\line(12,0){12}}
\put(0,0){\line(0,10){10}}
{\thicklines
\qbezier(0,4)(3,7)(6,10)
\qbezier(0,4)(-3,7)(-6,10)
}
\put(-11,0){\vector(1,0){4}}
\put(-5,7){\vector(1,0){4}}
\put(0,4){\circle*{.25}}
\put(-13.3,6.8){\small\mbox{space-time cone}}
\put(-1,10.7){\small\mbox{time}}
\put(7,-.3){\small\mbox{space}}
\put(-17.2,-.2){\small\mbox{space plane}}
\end{picture}
\end{center}
\vspace{0.5cm}
\caption{Cone-mixing property: asymptotic independence of states inside a space-time
cone from states inside a space plane.}
\label{fig-cone}
\end{figure}
\subsection{Elliptic vs.\ non-elliptic}
\label{subsec:ellnonell}
The original motivation for the present paper was to study the $(\alpha,\beta)$-model in
the limit as $\alpha\to\infty$ and $\beta\downarrow 0$. In this limit, which we will refer
to as the $(\infty,0)$-model, the walk is almost a deterministic functional of the IPS;
in particular, it is non-elliptic. The challenge was to find a way to deal with the \emph{lack
of ellipticity}. As we will see in Section~\ref{sec:model}, our set-up will be rather general
and will include the $(\alpha,\beta)$-model, the $(\infty,0)$-model, as well as various
other models. Examples of papers that deal with non-elliptic (actually, deterministic)
RW(D)RE's are Madras~\cite{Ma1} and Matic~\cite{Mat1}, where a recurrence vs.\ transience
criterion, respectively, a large deviation principle are derived.
In the RW(D)RE literature, ellipticity assumptions play an important role. In the static case,
RWRE in $\mathbb{Z}^d$, $d \geq 1$, is called \emph{elliptic} when, almost surely w.r.t.\ the
random environment, all the rates are \emph{finite} and there is a basis $\{e_i\}_{1 \leq i
\leq d}$ of $\mathbb{Z}^d$ such that the rate to go from $x$ to $x+e_i$ is \emph{positive}
for $1 \leq i \leq d$. It is called \emph{uniformly elliptic} when these rates are \emph{bounded away from infinity},
respectively, \emph{bounded away from zero}. In \cite{CoZe}, in order to take advantage
of the mixing property assumed on the random environment, it is important to have uniform ellipticity not necessarily in all directions, but in at least one direction in which the random walk is transient. One way to state this ``uniform directional ellipticity'' in a way that encompasses also the dynamic setting is to require the existence of a deterministic time $T>0$ and a vector $e \in \mathbb{Z}^d$ such that the quenched probability for the random walk to
displace itself along $e$ during time $T$ is uniformly positive for almost every realization of the random environment.
This is satisfied by the $(\alpha,\beta)$-model for $e=0$ and any $T>0$. This model is also transient
(indeed, non-nestling) in the time direction, which enables the use of the cone-mixing property of \cite{AvdHoRe1}.
In the case of the $(\infty,0)$-model, however, there are in general no such $T$ and $e$.
For example, when the random environment is a spin-flip system with bounded flip rates,
any fixed space-time position has positive probability of being unreachable by the random walk.
For all such models, the approach in \cite{AvdHoRe1} fails.
In the present paper, in order to deal with the possible lack of ellipticity we require a
different space-time mixing property for the dynamic random environment, which we call
\emph{conditional cone-mixing}. Moreover, as in \cite{CoZe} and \cite{AvdHoRe1}, we must
require the random walk to have a tendency to stay inside space-time cones. Under these
assumptions, we are able to set up a regeneration scheme and prove a LLN. Our result
includes the LLN for the $(\alpha,\beta)$-model in \cite{AvdHoRe1}, the $(\infty,0)$-model
for at least two subclasses of IPS's that we will exhibit, as well as models that are
intermediate, in the sense that they are neither uniformly elliptic in any direction,
nor deterministic as the $(\infty,0)$-model.
\subsection{Outline}
\label{subsec:outline}
The rest of the paper is organized as follows. In Section~\ref{sec:motiv} we discuss,
still informally, the $(\infty,0)$-model and the regeneration strategy. This section
serves as a motivation for the formal definition in Section~\ref{sec:model} of the
class of models we are after, which is based on three \textit{structural assumptions}.
Section~\ref{sec:results} contains the statement of our LLN under four \textit{hypotheses},
and a description of two classes of one-dimensional IPS's that satisfy these hypotheses
for the $(\infty,0)$-model, namely, spin-flip systems with bounded flip rates
that either are in Liggett's $M < \epsilon$ regime, or have finite range
and a small enough ratio of maximal/minimal flip rates.
Section \ref{sec:prep} contains preparation material,
given in a general context, that is used in the proof of the LLN given in
Section~\ref{sec:proofmainthm}. In Section~\ref{sec:proofthm2} we verify our hypotheses
for the two classes of IPS's described in Section~\ref{sec:results}. We also obtain a
criterion to determine the sign of the speed in the LLN, via a comparison with independent
spin-flip systems. Finally, in Section~\ref{sec:oex}, we discuss how to adapt the proofs
in Section \ref{sec:proofthm2} to other models, namely, generalizations of the $(\alpha,
\beta)$-model and the $(\infty,0)$-model, and mixtures thereof.
We also give an example where our hypotheses fail. The examples in our paper are all one-dimensional, even though our LLN is valid
in $\mathbb{Z}^d$, $d\geq 1$.
\section{Motivation}
\label{sec:motiv}
\subsection{The $(\infty,0)$-model}
\label{subsec:infty0}
Let
\be{IPSdef}
\xi := (\xi_t)_{t \geq 0} \quad \mbox{ with } \quad \xi_t:= \big(\xi_t(x)\big)_{x\in\mathbb{Z}}
\end{equation}
be a c\`adl\`ag Markov process on $\Omega$. We will interpret $\xi$ by saying that at time $t$ site $x$ contains either
a \emph{hole} ($\xi_t(x)=0$) or a \emph{particle} ($\xi_t(x)=1$).
Typical examples are interacting particle systems on $\Omega$, such as independent spin-flips and simple exclusion.
Suppose that we run the $(\alpha,\beta)$-model on $\xi$ with
$0 < \beta \ll 1 \ll \alpha < \infty$. Then the behavior of the random walk is as follows.
Suppose that $\xi_0(0)=1$ and that the walk starts at $0$. The walk rapidly moves to the
first hole on its right, typically before any of the particles it encounters manages to flip
to a hole. When it arrives at the hole, the walk starts to rapidly jump back and forth between
the hole and the particle to the left of the hole: we say that it sits in a \emph{trap}. If
$\xi_0(0)=0$ instead, then the walk rapidly moves to the first particle on its left, where
it starts to rapidly jump back and forth in a trap. In both cases, before moving away from
the trap, the walk typically waits until one or both of the sites in the trap flip. If only
one site flips, then the walk typically moves in the direction of the flip until it hits a
next trap, etc. If both sites flip simultaneously, then the probability for the walk to sit
at either of these sites is close to $\tfrac12$, and hence it leaves the trap in a direction
that is close to being determined by an independent fair coin.
The limiting dynamics when $\alpha\to\infty$ and $\beta\downarrow 0$ can be obtained from
the above description by removing the words ``rapidly, ``typically'' and ``close to''.
Except for the extra Bernoulli($\tfrac12$) random variables needed to decide in which
direction to go to when both sites in a trap flip simultaneously, the walk up to time $t$
is a deterministic functional of $(\xi_s)_{0 \leq s \leq t}$. In particular, if
$\xi$ changes only by single-site flips, then apart from the first jump
the walk is completely deterministic. Since the walk spends all of its time in traps where it
jumps back and forth between a hole and a particle, we may imagine that it lives on the
edges of $\mathbb{Z}$. We implement this observation by associating with each edge its left-most
site, i.e., we say that the walk is at $x$ when we actually mean that it is jumping back
and forth between $x$ and $x+1$.
\begin{figure}[hbtp]
\vspace{0.5cm}
\begin{center}
\setlength{\unitlength}{0.3cm}
\begin{picture}(20,10)(0,0)
\put(0,0){\line(1,0){22}}
\put(0,11){\line(1,0){22}}
\put(2,0){\line(0,1){4.8}}
\put(2,6.5){\line(0,1){1.6}}
\put(2,10){\line(0,1){1}}
\put(5,4){\line(0,1){5.8}}
\put(8,0){\line(0,1){1.5}}
\put(8,3){\line(0,1){4}}
\put(8,8.5){\line(0,1){1}}
\put(11,3.5){\line(0,1){2.2}}
\put(11,7.8){\line(0,1){3.2}}
\put(14,1.5){\line(0,1){2.2}}
\put(14,7){\line(0,1){3.2}}
\put(17,0){\line(0,1){2.8}}
\put(17,5.6){\line(0,1){5.4}}
\put(20,3.8){\line(0,1){3}}
\put(20,9.5){\line(0,1){1.6}}
{\thicklines
\qbezier[9](11.2,0.2)(9.7,0.2)(8.2,0.2)
\qbezier[4](8.2,0.2)(8.2,0.85)(8.2,1.5)
\qbezier[18](8.2,1.5)(5.2,1.5)(2.2,1.5)
\qbezier[7](2.2,1.5)(2.2,2.8)(2.2,4.1)
\qbezier[27](2.2,4.1)(6.7,4.1)(11.2,4.1)
\qbezier[5](11.2,4.1)(11.2,4.8)(11.2,5.7)
\qbezier[9](11.2,5.7)(9.7,5.7)(8.2,5.7)
\qbezier[4](8.2,5.7)(8.2,6.3)(8.2,7)
\qbezier[9](8.2,7)(6.7,7)(5.2,7)
\qbezier[5](5.2,7)(5.2,7.8)(5.2,8.6)
\qbezier[36](5.2,8.6)(11.2,8.6)(17.2,8.6)
\qbezier[3](17.2,8.6)(17.2,9.1)(17.2,9.6)
\qbezier[9](17.2,9.6)(18.7,9.6)(20.2,9.6)
\qbezier[4](20.2,9.6)(20.2,10.3)(20.2,11)
}
\put(10.7,-1.2){$0$}\put(-1.2,10.7){$t$}
\put(-1.2,-.3){$0$}
\put(20.3,11.6){$W_t$}
\put(11.2,0){\circle*{.35}}
\put(20.2,11){\circle*{.35}}
\put(23,-0.3){$\mathbb{Z}$}
\end{picture}
\end{center}
\caption{\small The vertical lines represent the presence of particles. The dotted line is the path of the $(\infty,0)$-walk.} \label{Inftyzero}
\end{figure}
Let
\be{walkdef}
W:=(W_t)_{t\geq 0}
\end{equation}
denote the random walk path. By the description above, $W$ is c\`adl\`ag and
\be{eq:form0}
W_t \text{ is a function of } \big((\xi_s)_{0 \leq s \leq t} , Y\big),
\end{equation}
where $Y$ is a sequence of i.i.d.\ Bernoulli($\tfrac12$) random variables independent of $\xi$. Note that $W$ also
has the following three properties:
\begin{itemize}
\item[(1)]
For any fixed time $s$, the increment $W_{s+t}-W_s$ is found by applying the same function in \eqref{eq:form0} to the environment
shifted in space and time by $(W_s,s)$ and an independent copy of $Y$; in particular, the pair $(W_t, \xi_t)$ is Markovian.
\item[(2)]
Given that $W$ stays inside a space-time cone until time $t$, $(W_s)_{0\leq s \leq t}$ is a
functional only of $Y$ and of the states in $\xi$ up to time $t$ inside a slightly larger cone, obtained by by adding all neighboring sites to the right.
\item[(3)]
Each jump of the path follows the same mechanism as the first jump, i.e.,
$W_t - W_{t-}$ is computed using the same rules as those for $W_0$ but applied to the environment shifted in space and time by $(W_{t-},t)$.
\end{itemize}
\noindent
The reason for emphasizing these properties will become clearer in Section~\ref{subsec:regen}.
\subsection{Regeneration}
\label{subsec:regen}
The cone-mixing property that is assumed in \cite{AvdHoRe1} to prove the LLN for the
$(\alpha,\beta)$-model can be loosely described as the requirement that all the states
of the IPS inside a space-time cone opening upwards depend weakly on the states inside
a space plane far below the tip (recall Fig.~\ref{fig-cone}). Let us give a rough idea
of how this property can lead to \emph{regeneration}. Consider the event that the walk
stands still for a long time. Since the jump times of the walk are independent of the
IPS, so is this event. During this pause, the environment around the walk is allowed
to mix, which by the cone-mixing property means that by the end of the pause all the
states inside a cone with a tip at the space-time position of the walk are almost
independent of the past of the walk. If thereafter the walk stays confined to the cone,
then its future increments will be almost independent of its past, and so we get an
approximate regeneration. Since in the $(\alpha,\beta)$-model there is a uniformly
positive probability for the walk to stay inside a space-time cone with a large
enough inclination, we see that this regeneration strategy can indeed be made to work.
\begin{figure}
\begin{picture}(1,0.6)
\put(70,0){\includegraphics[height= 120pt,width=280pt]{reg_fig2.pdf}}
\end{picture}
\put(75,0){\vector(1,0){280}}
\put(75,0){\vector(0,1){120}}
\put(62,127){\small\mbox{time}}
\put(365,-5){\small\mbox{space}}
\put(182,33){\small\mbox{pause}}
\put(210,33){\mbox{$\Big\{$}}
\put(60,45){$\tau$}
\caption{\small Regeneration at time $\tau$.} \label{fig-reg}
\end{figure}
For the actual proof of the LLN in \cite{AvdHoRe1}, cone-mixing must be more carefully
defined. For technical reasons, there must be some uniformity in the decay of correlations
between events in the space-time cone and in the space plane. This uniformity holds, for
instance, for any spin-flip system in the $M<\epsilon$ regime (Liggett~\cite{Li}, Section
I.3), but not for the exclusion process or the supercritical contact process. Therefore the
approach outlined above works for the first IPS, but not for the other two.
There are three properties of the $(\alpha,\beta)$-model that make the above heuristics
plausible. First, to be able to apply the cone-mixing property relative to the space-time
position of the walk, it is important that the pair (IPS,walk) is Markovian and that the
law of the environment as seen from the walk at any time is comparable to the initial law.
Second, there is a uniformly positive probability for the walk to stand still for a long
time and afterwards stay inside a space-time cone. Third, once the walk stays inside a space-time
cone, its increments depend on the IPS only through the states inside that cone. Let us
compare these observations with what happens in the $(\infty,0)$-model. Property (1) from
Section~\ref{subsec:infty0} gives us the Markov property, while property (2) gives us the
measurability inside cones. As we will see, when the environment is translation-invariant,
property (3) implies absolute continuity of the law of the environment as seen from the walk at any positive time with respect to its
counterpart at time zero. Therefore, as long as we can make sure that the walk has a tendency
to stay inside space-time cones (which is reasonable when we are looking for a LLN), the
main difference is that the event of standing still for a long time is not independent of
the environment, but rather is a \emph{deterministic} functional of the environment.
Consequently, it is not at all clear whether cone-mixing is enough to allow for regeneration.
On the other hand, the event of standing still is local, since it only depends on the states
of the two neighboring sites of the trap where the walk is pausing. For many IPS's, the
observation of a local event will not affect the weak dependence between states that are
far away in space-time. Hence, if such IPS's are cone-mixing, then states inside a space-time
cone remain almost independent of the initial configuration even when we condition on seeing
a trap for a long time.
Thus, under suitable assumptions, the event ``standing still for a long time'' is a
candidate to induce regeneration. In the $(\alpha,\beta)$-model this event does not
depend on the environment whereas in the $(\infty,0)$-model it is a deterministic
functional of the environment. If we put the $(\alpha,\beta)$-model in the form (\ref{eq:form0})
by taking for $Y$ two independent Poisson processes with rates $\alpha$ and $\beta$, then
we can restate the previous sentence by saying that in the $(\alpha,\beta)$-model the
regeneration-inducing event depends only on $Y$, while in the $(\infty,0)$-model it depends
only on $\xi$. We may therefore imagine that, also for other models of the type
\eqref{eq:form0} and that share properties (1)--(3), it will be possible to find more
general regeneration-inducing events that depend on both $\xi$ and $Y$ in a non-trivial
manner. This motivates our setup in Section~\ref{sec:model}.
\section{Model setting}
\label{sec:model}
So far we have mostly been discussing RWDRE driven by an IPS. However, there are convenient
constructions of IPS's on richer state spaces (such as graphical representations)
that can facilitate the construction of the regeneration-inducing events mentioned in
Section~\ref{subsec:regen}. We will therefore allow for more general Markov processes
to represent the dynamic random environment $\xi$. Notation is set up in
Section~\ref{subsec:notation}. Section~\ref{subsec:mainass} contains the three
structural assumptions that define the class of models we will consider.
\subsection{Notation and setup}
\label{subsec:notation}
Let $\mathbb{N} = \{1,2,\ldots \}$ be the set of natural numbers, and $\mathbb{N}_0 := \mathbb{N} \cup \{0\}$.
Let $E$ be a Polish space and $\xi:=(\xi_t)_{t \geq 0}$ a Markov process
with state space $E^{\mathbb{Z}^d}$ where $d \in \mathbb{N}$. Let $Y:=(Y_n)_{n \in \mathbb{N}}$ be an i.i.d.\
sequence of random elements independent of $\xi$. For $I \subset [0,\infty)$,
abbreviate $\xi_{I}:= (\xi_u)_{u \in I}$, and analogously for $Y$. The joint law
of $\xi$ and $Y$ when $\xi_0 = \eta \in E^{\mathbb{Z}^d}$ will be denoted by
$\mathbb{P}_{\eta}$. For $n \in \mathbb{N}$, put $\mathscr{Y}_n:=\sigma(Y_{[1,n]})$. Let $\mathscr{F}_0 := \sigma(\xi_0)$ and,
for $t>0$, $\mathscr{F}_t := \sigma(\xi_{[0,t]}) \vee \mathscr{Y}_{\lceil t \rceil}$.
For $t \ge 0$ and $x \in \mathbb{Z}^d$, let $\theta_t$ and $\theta_x$ be the
time-shift and space-shift operators given by
\be{}
\theta_t (\xi,Y) := \big((\xi_{t+s})_{s \geq 0},
(Y_{\lfloor t\rfloor + n})_{n\in\mathbb{N}}\big),
\qquad
\theta_x (\xi,Y) := \big((\theta_x \xi_t)_{t \geq 0},
(Y_n)_{n\in\mathbb{N}}\big),
\end{equation}
where $\theta_x \xi_t(y) = \xi_t(x+y)$.
In the sequel, whether $\theta$ is a
time-shift or a space-shift operator will always be clear from the index.
We assume that $\xi$ is translation-invariant, i.e., $\theta_x\xi$ has under $\mathbb{P}_\eta$
the same distribution as $\xi$ under $\mathbb{P}_{\theta_x\eta}$. We also assume
the existence of a (not necessarily unique) translation-invariant equilibrium distribution
$\mu$ for $\xi$, and write $\mathbb{P}_{\mu}(\cdot):=\int \mu(d\eta)\,\mathbb{P}_{\eta}(\cdot)$
to denote the joint law of $\xi$ and $Y$ when $\xi_0$ is drawn from $\mu$.
The random walk will be denoted by $W=(W_t)_{t\ge0}$, and we will write $\bar\xi:=(\bar\xi_t)_{t\geq 0}$
to denote the \emph{environment process as seen from $W$}, i.e., $\bar\xi_t := \theta_{W_t}\xi_t$.
Let $\bar\mu_t$ denote the law of $\bar \xi_t$ under $\mathbb{P}_{\mu}$.
We abbreviate $\bar\mu := \bar\mu_0$. Note that $\bar \mu = \mu$ when $\mathbb{P}_{\mu}(W_0 = 0)=1$.
For $m>0$ and $R\in\mathbb{N}_0$, define the $R$-enlarged $m$-cone by
\be{}
\begin{aligned}
C_R(m) &:= \big\{(x,t)\in\mathbb{Z}^d\times [0,\infty)\colon\,\|x\| \leq mt + R\big\},
\end{aligned}
\end{equation}
where $\| \cdot \|$ is the $L^1$ norm. Let $\mathscr{C}_{R,t}(m)$ be
the $\sigma$-algebras generated by the states of $\xi$ up to time $t$ inside $C_R(m)$.
\subsection{Structural assumptions}
\label{subsec:mainass}
We will assume that $W$ is random translation of a random walk starting at $0$.
More precisely, we assume that $Z = (Z_t)_{t \ge 0}$ is a c\`adl\`ag $\mathscr{F}$-adapted $\mathbb{Z}^d$-valued process
with $Z_0=0$ $\mathbb{P}_{\bar\mu}$-a.s. such that
\begin{equation}
\label{assumpWZ}
W_t = W_0 + \theta_{W_0}Z_t \quad \forall \; t \ge 0.
\end{equation}
We also assume that $W_0 \in \mathbb{Z}^d$ and depends on $\xi$ and $Y$ only through $\xi_0$, i.e.,
\be{assumpX0}
\mathbb{P}_{\mu}(W_0=x \mid \mathscr{F}_{\infty}) = \mathbb{P}_{\mu}(W_0=x \mid \xi_0) \; \text{ a.s. }
\forall \; x \in \mathbb{Z}^d.
\end{equation}
Under these assumptions, $(W_t-W_0)_{t\ge0}$ has under $\mathbb{P}_{\mu}$ the same distribution as $Z$ under $\mathbb{P}_{\bar \mu}$.
In what follows we make \emph{three structural assumptions} on $Z$:
\begin{itemize}
\item[(A1)] {\bf (Additivity)}\\
For all $n \in \mathbb{N}$,
\begin{equation}
\label{eqA1}
\begin{array}{rcl}
(Z_{t+n}-Z_n)_{t \ge 0} = \theta_{Z_n}\theta_n Z \quad \mathbb{P}_{\bar \mu} \text{-a.s.}
\end{array}
\end{equation}
\item[(A2)] {\bf (Locality)}\\
For $m>0$, let $\mathcal{D}_m := \{\|Z_t\| \le mt \,\,\forall\,
t\geq 0\}$. Then there exists $R \in \mathbb{N}_0$ such that, $\forall$ $m >0$, both $\mathcal{D}_m$ and $(1_{\mathcal{D}_m}Z_t)_{t \ge 0}$ are measurable w.r.t.\
$\mathscr{C}_{R,\infty}(m) \vee \mathscr{Y}_{\infty}$.
\item[(A3)]{\bf (Homogeneity of jumps)}\\
For all $n\in\mathbb{N}$ and $x \in \mathbb{Z}^d$,
\begin{equation}
\mathbb{P}_{\bar \mu}\left(Z_n-Z_{n-} = x \mid \xi_{[0,n]},
Z_{[0,n)}\right) = \mathbb{P}_{\theta_{Z_{n-}}\xi_n}\big(W_0 = x\big)
\qquad \mathbb{P}_{\bar \mu} \text{-a.s.}
\end{equation}
\end{itemize}
These properties are analogues of properties (1)--(3) of the $(\infty,0)$-model mentioned in
Section~\ref{subsec:infty0}, with the difference that we only require them to hold at integer times;
this will be enough as our proof relies on integer-valued regeneration times.
We also assume the `extra randomness' $Y$ to be split independently among time intervals of length $1$;
for example, in the case of the $(\infty,0)$-model,
each $Y_n$ would \emph{not} be a Bernoulli($\frac12$) random variable but a whole \emph{sequence} of such variables instead.
This is discussed in detail in Section \ref{sec:formaldef}.
\noindent
Another remark: assumption (A3) might seem strange since many random walk models have no deterministic jumps,
which is indeed the case for the examples described in Section \ref{sec:results}.
Note however that, in this case, (A3) severely restricts $W_0$, implying $W_0=0$ a.s.\ when $\xi$ is started from $\theta_{Z_{n-}}\xi_n$.
Furthermore, our main theorem (Theorem \ref{mainthm} below) is not restricted to this situation and includes also cases with deterministic jumps. For example,
one could modify the $(\infty,0)$-walk to jump exactly at integer times. Additional examples with deterministic jumps are described in item 4 of Section~\ref{sec:oex}.
The relevance of assumption (A3) is in showing that the law of the environment as seen by the RW after any jump is
absolutely continuous w.r.t.\ the law after the first jump; this is done in Lemma \ref{lem1} below.
\section{Main results}
\label{sec:results}
Theorems~\ref{mainthm} and \ref{thm:cond} below are the main results of our paper.
Theorem~\ref{mainthm} in Section~\ref{subsec:LLN} is our LLN. Theorem \ref{thm:cond}
in Section~\ref{subsec:examp} verifies the hypotheses in this LLN for the $(\infty,0)$-model
in two classes of one-dimensional IPS's. For these classes some more information is
available, namely, convergence in $L^p$, $p\geq 1$, and a criterion to determine the
sign of the speed.
\subsection{Law of large numbers}
\label{subsec:LLN}
In order to develop a regeneration scheme for a random walk subject to assumptions
(A1)--(A3) based on the heuristics discussed in Section~\ref{subsec:regen}, we need
suitable regeneration-inducing events. In the \emph{four hypotheses} stated below,
these events appear as a sequence $(\Gamma_L)_{L\in \mathbb{N}}$ such that,
for a certain fixed $m \in (0,\infty)$ and $R$ as in (A2),
$\Gamma_L\in\mathscr{C}_{R,L}(m)\vee\mathscr{Y}_L$ for all $L \in \mathbb{N}$.
\begin{itemize}
\item[(H1)] {\bf (Determinacy)}\\
On $\Gamma_L$, $Z_t = 0$ for all $t \in [0,L]$ $\mathbb{P}_{\bar\mu}$-a.s.
\item[(H2)] {\bf (Non-degeneracy)}\\
For $L$ large enough, there exists a $\gamma_L>0$ such that $\mathbb{P}_{\eta}(\Gamma_L)
\geq\gamma_L$ for $\bar\mu$-a.e.\ $\eta$.
\item[(H3)] {\bf (Cone constraints)}\\
Let $\mathcal{S}:=\inf\{t > 0\colon\,\|Z_t\| > mt\}$.
Then there exist $a \in (1, \infty)$, $\kappa_L \in (0,1]$ and $\psi_L \in [0,\infty)$
such that, for $L$ large enough and $\bar \mu$-a.e.\ $\eta$,
\be{}
\begin{array}{lll}
&\text{(1)} &\mathbb{P}_{\eta}(\theta_L \mathcal{S} = \infty \mid \Gamma_L )
\geq \kappa_L,\\[0.2cm]
&\text{(2)} &\mathbb{E}_{\eta} \left[1_{\{\theta_L \mathcal{S} < \infty\}}
\left(\theta_L\mathcal{S}\right)^a \mid \Gamma_L \right] \leq \psi_L^a .
\end{array}
\end{equation}
\item[(H4)] {\bf (Conditional cone-mixing)}\\
There exists a sequence of non-negative numbers $(\Phi_L)_{L \in \mathbb{N}}$
satisfying $\lim_{L\to\infty}\kappa_L^{-1}\Phi_L=0$ such that, for $L$ large enough and for $\bar\mu$-a.e. $\eta$, \be{condconemix}
\left \lvert \mathbb{E}_{\eta} \left(\theta_L f \mid \Gamma_L \right)
- \mathbb{E}_{\bar \mu}(\theta_L f \mid \Gamma_L )\right\rvert
\leq \Phi_L\,\| f\|_{\infty} \qquad \forall\,f\in\mathscr{C}_{R,\infty}(m), f \geq 0.
\end{equation}
\end{itemize}
We are now ready to state our LLN.
\bt{mainthm}
Under assumptions {\rm (A1)--(A3)} and hypotheses {\rm (H1)--(H4)}, there exists a
$w\in\mathbb{R}^d$ such that
\be{}
\lim_{t\to\infty} t^{-1}\,W_t = w \qquad \mathbb{P}_{\mu}-a.s.
\end{equation}
\end{theorem}
\noindent
\textbf{Remark 1:} Hypothesis (H4) above without the conditioning on $\Gamma_L$ in
\eqref{condconemix} and with constant $\kappa_L$ is the same as the cone-mixing condition used in Avena, den Hollander
and Redig \cite{AvdHoRe1}. There, $W_0=0$ $\mathbb{P}_{\mu}$-a.s., so that $\bar \mu = \mu$.
\noindent
\textbf{Remark 2:} Theorem \ref{mainthm} provides no information about the value of $w$, not
even its sign when $d=1$. Understanding the dependence of $w$ on model parameters is in general
a highly non-trivial problem.
\subsection{Examples}
\label{subsec:examp}
We next describe two classes of one-dimensional IPS's for which the $(\infty,0)$-model satisfies
hypotheses (H1)--(H4). Further details will be given in Section~\ref{sec:proofthm2}. In both
classes, $\xi$ is a spin-flip system in $\Omega=\{0,1\}^{\mathbb{Z}}$ with bounded and translation-invariant
single-site flip rates. We may assume that the flip rates at the origin are of the form
\be{eq1}
c(\eta) = \left\{\begin{array}{ll}
c_0 + \lambda_0 p_0(\eta) & \text{ if } \; \eta(0)=1,\\
c_1 + \lambda_1 p_1(\eta) & \text{ if } \; \eta(0)=0,\\
\end{array}\right.
\qquad \eta \in \Omega,
\end{equation}
for some $c_i, \lambda_i \geq 0$ and $p_i\colon\,\Omega\to [0,1]$, $i=0,1$.
\medskip\noindent
{\bf Example 1:} $c(\cdot)$ is in the $M<\epsilon$ regime (see Liggett~\cite{Li}, Section I.3).
\medskip\noindent
{\bf Example 2:} $p(\cdot)$ has finite range and $(\lambda_0 + \lambda_1)/(c_0+c_1)<\lambda_c$,
where $\lambda_c$ is the critical infection rate of the one-dimensional contact process with the same range.
\bt{thm:cond}
Consider the $(\infty,0)$-model. Suppose that $\xi$ is a spin-flip system with flip rates
given by {\rm (\ref{eq1})}. Then for Examples $1$ and $2$ there exist a version of $\xi$
and events $\Gamma_L\in \mathscr{C}_{R,L}(m)\vee\mathscr{Y}_L$, $L\in\mathbb{N}$,
satisfying hypotheses {\rm (H1)--(H4)}. Furthermore, the convergence in Theorem {\rm \ref{mainthm}}
holds also in $L^p$ for all $p \geq 1$, and
\be{signofw}
\begin{array}{ll}
w \ge \frac{c_0 + \lambda_0}{c_1 + c_0 + \lambda_0}(c_1-c_0 - \lambda_0)
& \text{ if } c_1 \ge c_0 + \lambda_0, \\[0.2cm]
w \le - \frac{c_1 + \lambda_1}{c_0 + c_1 + \lambda_1}(c_0-c_1 - \lambda_1)
& \text{ if } c_0 \ge c_1 + \lambda_1. \\
\end{array}
\end{equation}
\end{theorem}
For independent spin-flip systems (i.e., when $\lambda_0 = \lambda_1 = 0$), \eqref{signofw} shows
that $w$ is positive, zero or negative when the density $c_1/(c_0+c_1)$ is,
respectively, larger than, equal to or smaller than $\tfrac12$. The criterion
for other $\xi$ is obtained by comparison with independent spin-flip systems.
We expect hypotheses (H1)--(H4) to hold for a very large class of IPS's and walks. For each
choice of IPS and walk, the \emph{verification} of hypotheses (H1)--(H4) constitutes a
\emph{separate problem}. Typically, (H1)--(H2) are immediate, (H3) requires some work,
while (H4) is hard.
Additional models will be discussed in Section~\ref{sec:oex}. We will consider generalizations
of the $(\alpha,\beta)$-model and the $(\infty,0)$-model, namely, \emph{internal noise} models and
\emph{pattern} models, as well as mixtures of them.
The verification of (H1)--(H4) will be analogous to the two examples discussed above and
will not be carried out in detail.
This concludes the motivation and the statement of our main results. The remainder of the
paper will be devoted to the proofs of Theorems~\ref{mainthm} and \ref{thm:cond}, with the
exception of Section~\ref{sec:oex}, which contains additional examples and remarks.
\section{Preparation}
\label{sec:prep}
The aim of this section is to prove two propositions (Propositions~\ref{prop:avrt} and
\ref{prop:adfun} below) that will be needed in Section~\ref{sec:proofmainthm} to prove
the LLN. In Section~\ref{sec:alln} we deal with approximate laws
of large numbers for general discrete- or continuous-time random walks in $\mathbb{R}^d$.
In Section~\ref{sec:adfun} we specialize to additive functionals of a Markov chain
whose transition kernel satisfies a certain absolute-continuity property.
\subsection{Approximate law of large numbers}
\label{sec:alln}
This section contains two fundamental facts that are the basis of our proof of the LLN.
They deal with the notion of an approximate law of large numbers.
\begin{definition}\label{def:av}
Let $W=(W_t)_{t \geq 0}$ be a random process in $\mathbb{R}^d$ with $t\in\mathbb{N}_0$ or
$t\in [0,\infty)$. For $\varepsilon \geq 0$ and $v \in \mathbb{R}^d$, we say that $W$
has an $\varepsilon$-approximate asymptotic velocity $v$, written $W \in AV(\varepsilon,v)$,
if
\begin{equation}
\limsup_{t \to \infty} \left \| \frac{W_t}{t} - v \right \| \leq \varepsilon \quad
\text{ a.s.}
\end{equation}
\end{definition}
\noindent
We take $\|\cdot\|$ to be the $L_1$-norm. A simple
observation is that $W$ a.s.\ has an asymptotic velocity if and only if for every
$\varepsilon > 0$ there exists a $v_{\varepsilon} \in \mathbb{R}^d$ such that $W
\in AV(\varepsilon, v_{\varepsilon})$. In this case $\lim_{\varepsilon \downarrow 0}
v_{\varepsilon}$ exists and is equal to the asymptotic velocity.
\subsubsection{First key proposition: skeleton approximate velocity}
\label{sec:skelav}
The following proposition gives conditions under which an approximate velocity for the process
observed along a random sequence of times implies an approximate velocity for the full process.
\begin{proposition}
\label{prop:avrt}
Let $W$ be as in Definition~{\rm \ref{def:av}}. Set $\tau_0 := 0$, let $(\tau_k)_{k \in \mathbb{N}}$ be
an increasing sequence of random times in $(0,\infty)$ (or $\mathbb{N}$) with
$\lim_{k \to \infty}\tau_k = \infty$ a.s.\ and put $X_k := (W_{\tau_k},\tau_k) \in
\mathbb{R}^{d+1}$, $k \in \mathbb{N}_0$. Suppose that the following hold:\\
(i) There exists an $m > 0$ such that
\begin{equation}
\limsup_{k \to \infty} \sup_{s \in (\tau_k,\tau_{k+1}]}
\left \| \frac{ W_s - W_{\tau_k}}{s-\tau_k}\right \| \leq m \; \text{ a.s.}
\end{equation}
(ii) There exist $v \in \mathbb{R}^d$, $u>0$ and $\varepsilon \ge 0$ such that $X \in
AV(\varepsilon, (v,u))$.\\
Then $W \in AV((3m+1)\varepsilon/u, v/u)$.
\end{proposition}
\begin{proof}
First, let us check that \emph{(i)} implies
\begin{equation}
\label{sublin}
\limsup_{t \to \infty}\frac{\|W_t\|}{t} \le m \;\; \text{ a.s.}
\end{equation}
Suppose that
\begin{equation}
\label{ineq2}
\limsup_{k \to \infty} \sup_{s > \tau_k}
\left\|\frac{W_s-W_{\tau_k}}{s-\tau_k}\right\| \leq m \quad \text{a.s.}
\end{equation}
Since, for every $k$ and $t > \tau_k$,
\begin{equation}
\left\|\frac{W_t}{t}\right\| \leq \frac{\|W_{\tau_k}\|}{t}
+\left\|\frac{W_t-W_{\tau_k}}{t-\tau_k}\right\|
\left|1-\frac{\tau_k}{t}\right| \leq \frac{\|W_{\tau_k}\|}{t}
+\sup_{s>\tau_k}\left\|\frac{W_s-W_{\tau_k}}{s-\tau_k}\right\|
\left|1-\frac{\tau_k}{t}\right|,
\end{equation}
(\ref{sublin}) follows from (\ref{ineq2}) by letting $t\to\infty$ followed by $k\to\infty$.
To check (\ref{ineq2}), define, for $k \in \mathbb{N}_0$ and $l \in \mathbb{N}$,
\begin{equation}
m(k,l) := \sup_{s \in (\tau_k, \tau_{k+l}]}
\left\|\frac{W_s - W_{\tau_k}}{s-\tau_k}\right\| \; \text{ and } \;
m(k,\infty) := \sup_{s > \tau_k}\left\|\frac{W_s - W_{\tau_k}}{s-\tau_k}\right\| = \lim_{l \to \infty}m(k,l).
\end{equation}
Using the fact that $(x_1+x_2)/(y_1+y_2)\leq (x_1/y_1)\vee(x_2/y_2)$ for all $x_1, x_2
\in \mathbb{R}$ and $y_1, y_2 > 0$, we can prove by induction that
\begin{equation}
\label{ineq3}
m(k,l) \leq \max\{m(k,1), \dots, m(k+l-1,1)\}, \qquad l\in\mathbb{N}.
\end{equation}
Fix $\varepsilon>0$. By \emph{(i)}, a.s.\ there exists a $k_{\varepsilon}$ such that $m(k,1) \leq
m+\varepsilon$ for $k>k_{\varepsilon}$. By (\ref{ineq3}), the same is true for $m(k,l)$ for
all $l \in\mathbb{N}$, and therefore also for $m(k,\infty)$.
Since $\varepsilon$ is arbitrary, (\ref{ineq2}) follows.
Let us now proceed with the proof of the proposition. Assumption $(ii)$ implies that, a.s.,
\begin{equation}
\label{eq:rt1}
\limsup_{k \to \infty}\left\| \frac{W_{\tau_k}}{k} - v \right\|
\leq \varepsilon \;\;\; \text{ and } \;\;\;
\limsup_{k \to \infty}\left| \frac{\tau_k}{k} - u \right| \leq \varepsilon.
\end{equation}
For $t \ge 0$, let $k_t$ be the (random) non-negative
integer such that
\begin{equation}
\label{eq:rt2}
\tau_{k_t} \le t < \tau_{k_t+1}.
\end{equation}
Since $\tau_1 < \infty$ a.s., $k_t>0$ for large enough $t$. From (\ref{eq:rt1}) and (\ref{eq:rt2}) we deduce that
\begin{equation}
\label{eq:rt3}
\limsup_{t \to \infty}\left| \frac{t}{k_t} - u \right| \leq
\varepsilon \quad \text{ and so } \quad
\limsup_{t \to \infty}\left| \frac{t}{k_t} - \frac{\tau_{k_t}}{k_t} \right| \le 2\varepsilon.
\end{equation}
For $t$ large enough we may write
\begin{align}
\label{ineq:avrt1}
\left\| \frac{uW_t}{t}-v \right\| \leq
& \frac{\|W_t\|}{t}\left|u - \frac{t}{k_t} \right|
+ \left\|\frac{W_t - W_{\tau_{k_t}}}{k_t}\right\|
+ \left\| \frac{W_{\tau_{k_t}}}{k_t}-v \right\|\nonumber \\
\leq & \frac{\|W_t\|}{t}\left|u - \frac{t}{k_t} \right|
+ \sup_{s \in (\tau_{k_t},\tau_{k_t+1}]}\left\|\frac
{W_s - W_{\tau_{k_t}}}{s-\tau_{k_t}} \right\|
\left| \frac{t - \tau_{k_t}}{k_t}\right| + \left\| \frac{W_{\tau_{k_t}}}{k_t}-v \right\|,
\end{align}
from which we obtain the conclusion by taking the limsup as $t\to\infty$ in (\ref{ineq:avrt1}),
using (i), (\ref{sublin}), (\ref{eq:rt1}) and (\ref{eq:rt3}), and then dividing by $u$.
\end{proof}
\subsubsection{Conditions for the skeleton to have an approximate velocity}
\label{sec:condskelav}
The following lemma states sufficient conditions for a discrete-time process to have
an approximate velocity. It will be used in the proof of Proposition \ref{prop:adfun} below.
\bl{lemma:aplln}
Let $X=(X_k)_{k\in\mathbb{N}_ 0}$ be a sequence of random vectors in $\mathbb{R}^d$ with joint
law $P$ such that $P(X_0 =0)=1$. Suppose that there exist a probability measure $Q$ on $\mathbb{R}^d$ and numbers
$\phi \in [0,1)$, $a>1$, $K>0$ with $\int_{\mathbb{R}^d} \|x\|^a\,Q(dx) \leq K^a$, such
that, $P$-a.s. for all $k\in\mathbb{N}_0$,\\
(i) $\left \lvert P(X_{k+1}-X_k\in A \mid X_0,\dots, X_k) - Q(A) \right \rvert \leq \phi$
for all $A$ measurable;\\
(ii) $E[\|X_{k+1}-X_k\|^a|X_0,\dots,X_k] \leq K^a$.\\
Then
\be{eq:genaplln}
\limsup_{n \to \infty} \left\lVert \frac{X_n}{n} - v \right \rVert
\leq 2K \phi^{(a-1)/a} \qquad P\text{-a.s.,}
\end{equation}
where $v=\int_{\mathbb{R}^d} x\,Q(dx)$. In other words, $X \in AV(2K\phi^{(a-1)/a},v)$.
\end{lemma}
\begin{proof}
The proof is an adaptation of the proof of Lemma 3.13 in \cite{CoZe}; we include it here
for completeness. With regular conditional probabilities, we can, using $(i)$, couple $P$
and $Q^{\otimes\mathbb{N}_0}$ according to a standard splitting representation (see e.g.\
Berbee~\cite{Be}). More precisely, on an enlarged probability space we can construct random
variables
\be{}
(\Delta_k,V_k,R_k)_{k\in\mathbb{N}}
\end{equation}
such that
\begin{itemize}
\item[(1)]
$(\Delta_k)_{k\in\mathbb{N}}$ is an i.i.d.\ sequence of Bernoulli($\phi$) random variables.
\item[(2)]
$(V_k)_{k\in\mathbb{N}}$ is an i.i.d.\ sequence of random vectors with law $Q$.
\item[(3)]
$(\Delta_l)_{l \geq k}$ is independent of $(\Delta_l,V_l,R_l)_{0\leq l<k}$, $R_k$.
\item[(4)]
Setting $\hat{X}_0:=0$ and, for $k \in \mathbb{N}_0$, $\hat{X}_{k+1}-\hat{X}_k:= (1-\Delta_k)V_k +\Delta_k R_k$, then $\hat{X}$ is equal in distribution to $X$.
\item[(5)]
Setting $\mathcal{G}_k := \sigma(\Delta_l,V_l,R_l\colon\,0\leq l \leq k)$, then $E[ f(V_k) \mid \mathcal{G}_{k-1}]$ is measurable w.r.t.\
$\sigma(\hat{X}_l\colon\,0\leq l\leq k-1)$ for any Borel nonnegative function $f$.
\end{itemize}
Using (4), we may write
\be{eq5}
\frac{X_n}{n} \; \,{\buildrel d \over =}\, \; \frac{\hat{X}_n}{n} = \frac{1}{n} \sum_{k=1}^n V_k
- \frac{1}{n} \sum_{k=1}^n \Delta_k V_k
+ \frac{1}{n}\sum_{k=1}^n \Delta_k R_k.
\end{equation}
As $n\to\infty$, the first term on the r.h.s.\ converges a.s.\ to $v$
by the LLN for i.i.d.\ random variables. By H\"older's inequality, the norm of the second term
is at most
\be{eq2}
\left(\frac{1}{n} \sum_{k=1}^n \Delta_k \right)^{(a-1)/a}
\left(\frac{1}{n} \sum_{k=1}^n \|V_k\|^a\right)^{1/a},
\end{equation}
which, by (1) and (2), converges a.s.\ as $n\to\infty$ to
\be{}
\phi^{(a-1)/a} \left(\int_{\mathbb{R}^d} \|x\|^a\,Q(dx)\right)^{1/a}
\leq K \phi^{(a-1)/a}.
\end{equation}
To control the third term, put $R^*_k := E[R_k\mid\mathcal{G}_{k-1}]$. Since $\|\Delta_k R_k\| \leq \|\hat{X}_{k+1}-\hat{X}_k\|$,
using (1), (3), (4), (5) and (ii), we get
\be{eq3}
\phi E[\|R_k\|^a \mid \mathcal{G}_{k-1}]
= E[\Delta_k \|R_k\|^a \mid \mathcal{G}_{k-1}]
\leq E[\|\hat{X}_{k+1}-\hat{X}_k\|^{a}\mid\mathcal{G}_{k-1}] \leq K^a.
\end{equation}
Combining \eqref{eq3} with Jensen's inequality, we obtain
\be{eq4}
\|R^*_k\| \leq E\big[\|R_k\|^{a} \mid \mathcal{G}_{k-1}\big]^{1/a}
\leq \frac{K}{\phi^{1/a}},
\end{equation}
so that
\be{eq6}
\left\|\frac{1}{n} \sum_{k=1}^n \Delta_k R^*_k \right\|
\leq \frac{K}{\phi^{1/a}}
\left(\frac{1}{n} \sum_{k=1}^n \Delta_k \right) \xrightarrow[n \to \infty]{} K \phi^{(a-1)/a}.
\end{equation}
Now fix $y \in \mathbb{R}^d$ and put
\be{}
M^y_n := \sum_{k=1}^n \frac{\Delta_k}{k} \langle R_k-R^*_k, y \rangle.
\end{equation}
where $\langle \cdot, \cdot \rangle$ denotes the usual inner product. Then $(M^y_n)_{n\in\mathbb{N}_0}$ is
a $(\mathcal{G}_n)_{n\in\mathbb{N}_0}$-martingale whose quadratic variation is
\be{}
\langle M^y \rangle_n = \sum_{k=1}^{n} \frac{\Delta_k}{k^2}\langle R_k - R^*_k,y \rangle^2.
\end{equation}
By the Burkholder-Gundy inequality and (\ref{eq3}--\ref{eq4}), we have
\be{}
\begin{aligned}
E\left[\sup_{n\in\mathbb{N}}|M^y_n|^{a \wedge 2}\right]
&\leq C\,E \left[\langle M^y \rangle_{\infty}^{(a \wedge 2)/2}\right]\\
&\leq C\,E \left[\sum_{k=1}^{\infty} \frac{\Delta_k}{k^{a \wedge 2}}
\left| \langle R_k-R^*_k,y \rangle \right |^{a \wedge 2}\right] \leq C\,\|y\|^{a \wedge 2}
K^{a \wedge 2},
\end{aligned}
\end{equation}
where $C$ is a positive constant that may change after each inequality. This implies that $M^y_n$ is uniformly integrable
and therefore converges a.s.\ as $n\to\infty$. Kronecker's lemma then gives
\be{}
\lim_{n\to\infty} \frac{1}{n} \sum_{k=1}^n \Delta_k \langle R_k-R^*_k, y \rangle
= 0 \, \text{ a.s.}
\end{equation}
Since $y$ is arbitrary, this in turn implies that
\be{eq7}
\lim_{n\to\infty} \frac{1}{n} \sum_{k=1}^n \Delta_k(R_k-R^*_k) = 0 \; \text{ a.s.}
\end{equation}
Therefore, by \eqref{eq6} and \eqref{eq7}, the limsup of the norm of the last term in the r.h.s.\ of \eqref{eq5} is also bounded by $K\phi^{(a-1)/a}$, which finishes the proof.
\end{proof}
\subsection{Additive functionals of a discrete-time Markov chain}
\label{sec:adfun}
\subsubsection{Notation}
Let $\mathcal{X} = (\mathcal{X}_n)_{n \in \mathbb{N}_0}$ be a time-homogeneous Markov chain in the canonical space equipped
with the time-shift operators $(\theta_n)_{n \in \mathbb{N}_0}$. For $n \ge 1$, put $\mathcal{F}_n :=
\sigma(\mathcal{X}_{[1,n]})$ (note that $\mathcal{X}_0 \notin \mathcal{F}_{\infty}$) and let $P_\chi$ denote the law of $(\mathcal{X}_n)_{n
\in \mathbb{N}_0}$ when $\mathcal{X}_0 = \chi$. Fix an initial measure $\nu$ and suppose that, for any nonnegative $f \in \mathcal{F}_{\infty}$,
\begin{equation}
\label{sup:abscont}
P_\nu(E_{\mathcal{X}_n}[f] \in \cdot) \ll P_\nu(E_{\mathcal{X}_0}[f] \in \cdot),
\end{equation}
where $P_\nu := \int \nu(d\chi) P_\chi$.
Let $Z = (Z_n)_{n \in \mathbb{N}_0}$ be a $\mathbb{Z}^d$-valued $\mathcal{F}$-adapted process that is an \emph{additive functional} of $(\mathcal{X}_n)_{n \in \mathbb{N}}$, i.e.,
$Z_0=0$ and, for any $k \in \mathbb{N}_0$,
\begin{equation}\label{defadditivefunc}
(Z_{k+n}-Z_{k})_{n \in \mathbb{N}_0} = \theta_k Z \qquad P_\nu \text{-a.s.}
\end{equation}
We are interested in finding random times $(\tau_k)_{k \in \mathbb{N}_0}$ such
that $(Z_{\tau_k},\tau_k)_{k \in \mathbb{N}_0}$ satisfies the hypotheses of
Lemma~\ref{lemma:aplln}. In the Markovian setting it makes sense to look for
$\tau_k$ of the form
\begin{equation}
\label{eq:recrt}
\tau_0 = 0, \qquad \tau_{k+1} = \tau_k + \theta_{\tau_k}\tau, \quad k\in\mathbb{N}_0,
\end{equation}
where $\tau$ is a random time.
Condition (i) of Lemma~\ref{lemma:aplln} is a ``decoupling condition''. It states
that the law of an increment of the process depends weakly on the previous increments. Such a
condition can be enforced by the occurrence of a ``decoupling event'' under which
the increments of $(Z_{\tau_k},\tau_k)_{k \in \mathbb{N}_0}$ lose dependence. In this
setting, $\tau$ is a time at which the decoupling event is observed.
\subsubsection{Second key proposition: approximate regeneration times}
Proposition~\ref{prop:adfun} below is a consequence of Lemma~\ref{lemma:aplln} and
is the main result of this section. It will be used together with Proposition~\ref{prop:avrt}
to prove the LLN in Section~\ref{sec:proofmainthm}. It gives a way to construct $\tau$
when the decoupling event can be detected by ``probing the future'' with a stopping time.
For a random variable $\mathcal{T}$ taking values in $\mathbb{N}_0 \cup \{\infty\}$, we define the
\emph{image} of $\mathcal{T}$ by $\mathcal{I}_{\mathcal{T}}:= \{n \in \mathbb{N}\colon\,
P_{\nu}(\mathcal{T} = n)>0\}$, and its closure under addition by $\bar{\mathcal{I}}_{\mathcal{T}}
:= \{n \in \mathbb{N}\colon\,\exists\,\,l \in \mathbb{N},\,i_1,\ldots,i_l \in
\mathcal{I}_{\mathcal{T}}\colon\, n = i_1 + \dots + i_l\}$.
Note that $\mathcal{I}_{\mathcal{T}} = \emptyset$ if and only if $\mathcal{T} \in \{0,\infty\}$ a.s..
\begin{proposition}
\label{prop:adfun}
Let $\mathcal{T}$ be a stopping time for the filtration $\mathcal{F}$
taking values in $\mathbb{N}\cup\{\infty\}$. Put $D := \{\mathcal{T} = \infty\}$ and suppose that
the following properties hold:\\
(i) For every $n \in \bar{\mathcal{I}}_{\mathcal{T}}$ there exists a $D_n \in \mathcal{F}_n$
such that
\begin{equation*}
D \cap \theta_n D = D_n \cap \theta_n D \quad \; P_{\nu} \text{-a.s.}
\end{equation*}
(ii) There exist numbers $\rho \in (0,1]$, $a>1$, $C>0$, $m>0$ and $\phi \in [0,1)$ such that, $P_{\nu}$-a.s.,
\begin{equation*}
\begin{array}{rl}
\text{(a)} & P_{\mathcal{X}_0}\left(D \right) \geq \rho, \\
\text{(b)} & E_{\mathcal{X}_0}\left[1_{\{\mathcal{T} < \infty\}} \mathcal{T}^a \right] \leq C^a, \\
\text{(c)} & \text{On } D \text{, } \|Z_t\| \leq mt \text{ for all } t \in \mathbb{N}_0,\\
\text{(d)} & \left| E_{\mathcal{X}_0}\left[f(Z, (\theta_n\mathcal{T})_{n\in \bar{\mathcal{I}}_{\mathcal{T}}}) \mid D\right] - E_{\nu}\left[f(Z, (\theta_n\mathcal{T})_{n\in \bar{\mathcal{I}}_{\mathcal{T}}}) \mid D\right] \right| \le \phi \|f\|_{\infty} \; \forall f \ge 0 \text{ measurable.}
\end{array}
\end{equation*}
Then there exists a random time $\tau \in \mathcal{F}_{\infty}$ taking values in $\mathbb{N}$
such that, setting $\tau_k$ as in {\rm (\ref{eq:recrt})} and $X_k:= (Z_{\tau_k}, \tau_k)$,
then $X \in AV(\varepsilon, (v,u))$ where $(v,u)=E_{\nu}\left[(Z_{\tau},\tau) \mid
D\right]$, $u>0$ and $\varepsilon = 12(m+1)u\phi^{(a-1)/a}$.
\end{proposition}
\subsubsection{Two further propositions}
In order to prove Proposition~\ref{prop:adfun}, we will need two further propositions (Propositions \ref{prop:adfun1} and \ref{prop:stoptimes} below).
\begin{proposition}
\label{prop:adfun1}
Let $\tau$ be a random time measurable w.r.t.\ $\mathcal{F}_{\infty}$ taking values in
$\mathbb{N}$. Put $\tau_k$ as in {\rm (\ref{eq:recrt})} and $X_k := (Z_{\tau_k},\tau_k)$.
Suppose that there exists an event $D \in \mathcal{F}_{\infty}$ such that the following
hold $P_{\nu}$-a.s.:\\
(i) For $n \in \mathcal{I}_{\tau}$, there exist events $H_n$ and $D_n \in \mathcal{F}_n$
such that
\begin{equation}
\begin{array}{rl}
\text{(a)} & \{\tau = n\} = H_n \cap \theta_n D, \\
\text{(b)} & D \cap \theta_n D = D_n \cap \theta_n D.
\end{array}
\end{equation}
(ii) There exist $\phi \in [0,1)$, $K>0$ and $a>1$ such that that, on $\{P_{\mathcal{X}_0}(D)>0\}$,
\begin{equation}
\begin{array}{rl}
\text{(a)} & E_{\mathcal{X}_0}[\|X_1\|^a | D] \le K^a,\\
\text{(b)} &\left | P_{\mathcal{X}_0}\left(X_1 \in A | D \right)
- P_{\nu}\left(X_1 \in A | D \right) \right | \leq \phi
\quad \forall\,A \text{ measurable.}\\
\end{array}
\end{equation}
Then $X \in AV\big(\varepsilon, (v,u)\big)$, where $\varepsilon= 2K\phi^{(a-1)/a}$ and $(v,u)
:= E_{\nu}[X_1|D]$.
\end{proposition}
\begin{proof}
Since $\tau < \infty$, by (i)(a) and (\ref{sup:abscont}) we must have $P_{\nu}(D)>0$.
Let $\mathcal{F}_{\tau_k}$ be the $\sigma$-algebra of the events $B \in \mathcal{F}_{\infty}$
such that, for all $n \in \mathbb{N}$, there exists
$B_n \in \mathcal{F}_n$ with $B \cap \{\tau_k = n\} = B_n \cap \{\tau_k = n\}$. We will
show that, $P_{\nu}$-a.s., for all $k\in\mathbb{N}$,
\begin{equation}
\label{eq:af1}
E_{\nu}\left[ \| \theta_{\tau_k}X_1 \|^a | \mathcal{F}_{\tau_k}\right] \le K^a
\end{equation}
and
\begin{equation}
\label{eq:af2}
\left|P_{\nu}\left(\theta_{\tau_k}X_1 \in A | \mathcal{F}_{\tau_k}\right)
- P_{\nu}(X_1 \in A | D) \right| \leq \phi \quad \forall\,A \text{ measurable.}
\end{equation}
Then, setting $Q(\cdot):= P_{\nu}(X_1 \in \cdot | D)$ and noting that $\theta_{\tau_k}X_1 = X_{k+1} - X_k$
and $X_j \in \mathcal{F}_{\tau_k}$ for all $0\leq j \leq k$, we will be able to conclude since
(\ref{eq:af1}--\ref{eq:af2}) and (ii)(a) imply that the conditions of Lemma~\ref{lemma:aplln} are all satisfied.
To prove (\ref{eq:af1}--\ref{eq:af2}), first note that, using (i), one can verify by induction
that (i)(a) holds also for $\tau_k$, i.e., for every $n \in \mathcal{I}_{\tau_k}$ there
exists $H_{k,n} \in \mathcal{F}_n$ such that
\begin{equation}
\{\tau_k = n \} = H_{k,n} \cap \theta_n D \qquad P_\nu \text{-a.s.}
\end{equation}
Take $B \in \mathcal{F}_{\tau_k}$ and a measurable nonnegative function $f$, and write
\begin{align}
\label{eq:af3}
E_{\nu}\left[ 1_B \theta_{\tau_k}f(X_1)\right]
& = \sum_{n \in \mathcal{I}_{\tau_k}}E_{\nu}
\left[ 1_{B \cap \{\tau_k = n\}}\theta_n f(X_1)\right]
= \sum_{n \in \mathcal{I}_{\tau_k}}
E_{\nu}\left[1_{B_n \cap H_{k,n}}
\theta_n\big(1_D f(X_1)\big) \right] \nonumber \\
& = \sum_{n \in \mathcal{I}_{\tau_k}} E_{\nu}\left[1_{B_n \cap H_{k,n}}
P_{\mathcal{X}_n}(D)E_{\mathcal{X}_n}\left[f(X_1) | D \right]\right].
\end{align}
Noting that $P_{\nu}(B) = \sum_{n \in \mathcal{I}_{\tau_k}} E_{\nu}\left[1_{B_n \cap H_{k,n}} P_{\mathcal{X}_n}(D) \right]$,
obtain (\ref{eq:af1}) by taking $f(x) = \|x\|^a$ and using (ii)(a) together
with (\ref{sup:abscont}). For (\ref{eq:af2}), choose $f = 1_A$, subtract $P_{\nu}(B)
E_{\nu}\left[ f(X_1) | D \right]$ from (\ref{eq:af3}) and use (ii)(b).
\end{proof}
\noindent
\begin{proposition}
\label{prop:stoptimes}
Let $\mathcal{T}$ be a stopping time as in Proposition \ref{prop:adfun}
and suppose that conditions (ii)(a) and (ii)(b) of that proposition are satisfied.
Define a sequence of stopping times $(T_k)_{k \in \mathbb{N}_0}$ as follows.
Put $T_0 = 0$ and, for $k \in \mathbb{N}_0$,
\begin{equation}
\label{recursionT_k}
T_{k+1} := \left\{
\begin{array}{ll}
\infty & \;\;\;\;\;\; \text{ if } T_k = \infty \\
T_k + \theta_{T_k}\mathcal{T} & \;\;\;\;\;\; \text{ otherwise.}
\end{array}\right.
\end{equation}
Put
\begin{equation}
N := \inf\{ k \in \mathbb{N}_0\colon\, T_k < \infty \text{ and } T_{k+1} = \infty\}.
\end{equation}
Then $N < \infty$ a.s.\ and there exists a constant $\varkappa = \varkappa(a,\rho) \in (0,\infty)$ such that,
$P_\nu$-a.s.,
\begin{equation}
\label{conclem}
E_{\mathcal{X}_0}\left[T_N^a \right] \le (\varkappa C)^a.
\end{equation}
Furthermore, $\mathcal{I}_{T_N} \subset \bar{\mathcal{I}}_{\mathcal{T}}$.
\end{proposition}
\begin{proof}
First, let us check that
\begin{equation}
\label{nfin}
P_{\mathcal{X}_0}(N \geq n) \le (1-\rho)^n.
\end{equation}
Indeed, $N \geq n$ if and only if $T_n < \infty$, so that, for $k \in \mathbb{N}_0$,
\begin{equation}
\label{nfinn}
P_{\mathcal{X}_0}(T_{k+1} < \infty)
= E_{\mathcal{X}_0}\left[ 1_{\{T_k < \infty\}}
P_{\mathcal{X}_{T_k}}(\mathcal{T} < \infty)\right] \le (1-\rho)P_{\mathcal{X}_0}(T_k < \infty),
\end{equation}
where we use (ii)(a) and the fact that (\ref{sup:abscont}) holds also with a stopping time in place of $n$.
Clearly, (\ref{nfin}) follows from (\ref{nfinn}) by induction. In particular, $N < \infty$ a.s.
From \eqref{recursionT_k} we see that, for $0\leq k \leq n$,
\begin{equation}\label{recursion2}
T_n = T_k + \theta_{T_k}T_{n-k} \; \text{ on } \; \{T_k < \infty \}.
\end{equation}
Using (ii)(a) and (b), with the help of (\ref{sup:abscont}) again, we can a.s.\ estimate,
for $0\leq k < n$,
\begin{align}
\label{estim}
E_{\mathcal{X}_0}\left[1_{\{T_n < \infty \}} \left|T_{k+1}-T_k \right|^a\right]
& = E_{\mathcal{X}_0}\left[1_{\{T_{k+1} < \infty\}} \left|T_{k+1}-T_k \right|^a
P_{\mathcal{X}_{T_{k+1}}}(T_{n-k-1}<\infty)\right] \nonumber \\
& \le (1-\rho)^{n-k-1} E_{\mathcal{X}_0}
\left[1_{\{T_k < \infty, \theta_{T_k}\mathcal{T} < \infty\}}
\theta_{T_k}\mathcal{T}^a\right] \nonumber \\
& = (1-\rho)^{n-k-1} E_{\mathcal{X}_0}\left[1_{\{T_k < \infty\}}
E_{\mathcal{X}_{T_k}}\left[1_{\{\mathcal{T} < \infty\}}
\mathcal{T}^a \right] \right] \nonumber \\
& \le (1-\rho)^{n-k-1} C^a P_{\mathcal{X}_0}(T_k < \infty) \nonumber \\
& \le (1-\rho)^{n-1} C^a.
\end{align}
Now write
\begin{equation}
T_N = \sum_{k=0}^{N-1} T_{k+1}-T_k.
\end{equation}
By Jensen's inequality,
\begin{equation}
T_N^a \le N^{a-1}\sum_{k=0}^{N-1} \left| T_{k+1}-T_k \right|^a
\end{equation}
so that, by \eqref{estim},
\begin{align}
E_{\mathcal{X}_0}\left[ T_N^a \right] \le \sum_{n=1}^{\infty} n^{a-1}
\sum_{k=0}^{n-1}E_{\mathcal{X}_0}\left[ 1_{\{N=n\}}
\left|T_{k+1}-T_k \right|^a \right] \le C^a \sum_{n=1}^{\infty} n^a (1-\rho)^{n-1} \; \text{ a.s.}
\end{align}
and (\ref{conclem}) follows by taking $\varkappa = \left( \sum_{n=1}^{\infty} n^a
(1-\rho)^{n-1} \right)^{1/a}$.
As for the claim that $\mathcal{I}_{T_N} \subset \bar{\mathcal{I}}_{\mathcal{T}}$,
write, for $n \in \mathbb{N}$,
\begin{equation}
\{T_N = n\} = \sum_{k=1}^{\infty} \{T_k=n, N=k\}
\end{equation}
to see that $\mathcal{I}_{T_N} \subset \bigcup_{k=1}^{\infty} \mathcal{I}_{T_k}$. Using
(\ref{recursionT_k}), we can verify by induction that, for each $k \in \mathbb{N}$, $\mathcal{I}_{T_k}
\subset \{n \in \mathbb{N} \colon\,\exists\,i_1,\ldots,i_k \in \mathcal{I}_{\mathcal{T}}
\colon\,n = i_1 + \dots + i_k \} \subset \bar{\mathcal{I}}_{\mathcal{T}}$, and the claim
follows.
\end{proof}
\subsubsection{Proof of Proposition \ref{prop:adfun}}
We can now combine Propositions \ref{prop:adfun1} and \ref{prop:stoptimes} to prove
Proposition~\ref{prop:adfun}.
\begin{proof}
In the following we will refer to the hypotheses of Proposition~\ref{prop:adfun1}
with the prefix P. For example, P(i)(a) denotes hypothesis (i)(a) in that proposition.
The hypotheses in Proposition~\ref{prop:adfun} will be referred to without a prefix. Since
the hypotheses of Proposition~\ref{prop:stoptimes} are a subset of those of
Proposition~\ref{prop:adfun}, the conclusions of the former are valid.
We will show that, if $\tau := t_0 + \theta_{t_0}T_N$ for a suitable $t_0 \in \mathbb{N}$, then $\tau$
satisfies the hypotheses of Proposition~\ref{prop:adfun1} for a suitable $K$. There are
two cases. If $\mathcal{I}_{\mathcal{T}} = \emptyset$, then $T_N \equiv 0$. Choosing
$t_0 = 1$, we basically fall in the context of Lemma~\ref{lemma:aplln}. P(i)(a) and
P(i)(b) are trivial, (ii)(c) implies that P(ii)(a) holds with $K = (m+1)$, while P(ii)(b)
follows immediately from (ii)(d). Therefore, we may suppose that $\mathcal{I}_{\mathcal{T}} \neq \emptyset$
and put $\iota := \min \mathcal{I}_{\mathcal{T}} \in \mathbb{N}$.
Let $\hat C := 1 \vee (\varkappa C)$ and $t_0 := \iota \lceil \hat C \rho^{-1/a} \rceil$.
We will show that $\tau$ satisfies the hypotheses of Proposition~\ref{prop:adfun1}
with $K = 6\iota(m+1)\hat C\rho^{-1/a}$.
\medskip\noindent
\underline{P(i)(a)}:
First we show that this property is true for $T_N$. Indeed,
\begin{align}
\{T_N = n\} & = \sum_{k \in \mathbb{N}_0}\{N=k, T_k = n\}
= \sum_{k \in \mathbb{N}_0}\{T_k=n, \theta_n \mathcal{T} = \infty\} \\
& = \theta_n D \cap \left( \bigcup_{k \in \mathbb{N}_0} \{T_k=n\} \right),
\end{align}
and $\hat H_n := \bigcup_{k \in \mathbb{N}_0} \{T_k=n\} \in \mathcal{F}_n$ since the $T_k$'s
are all stopping times. Now we observe that $\{ \tau = n\} = \theta_{t_0}\{T_N = n-t_0\}$,
so we can take $H_n := \emptyset$ if $n<t_0$ and $H_n := \theta_{t_0} \hat H_{n-t_0}$
otherwise.
\medskip\noindent
\underline{P(i)(b)}:
By $(i)$, it suffices to show that $\mathcal{I}_{\tau} \subset \bar{\mathcal{I}}_{\mathcal{T}}$.
Since $t_0 \in \bar{\mathcal{I}}_{\mathcal{T}}$ (as an integer multiple of $\iota$), this
follows from the definition of $\tau$ and the last conclusion of Proposition \ref{prop:stoptimes}.
\medskip\noindent
\underline{P(ii)(a)}:
By (ii)(c), $\|X_1\|^a = \left(\|Z_{\tau}\|+\tau\right)^a \le \left((m+1)\tau\right)^a$ on
$D$. Therefore, we just need to show that
\begin{equation}
\label{mombound}
E_{\mathcal{X}_0}\left[\tau^a |D\right] \le (6 \iota \hat C)^a/\rho.
\end{equation}
Now, $\tau^a \le 2^{a-1}\left(t_0^a + \theta_{t_0}T_N^a\right)$ and, by
Proposition~\ref{prop:stoptimes} and (\ref{sup:abscont}),
\begin{equation}
E_{\mathcal{X}_0}\left[ \theta_{t_0} T_N^a\right]
= E_{\mathcal{X}_0}\left[ E_{\mathcal{X}_{t_0}}\left[T_N^a\right]\right] \le \hat C^a.
\end{equation}
Using (ii)(a), we obtain
\begin{equation}
E_{\mathcal{X}_0}\left[ \theta_{t_0} T_N^a\right | D] \le \hat C^a / \rho.
\end{equation}
Since $t_0 \le 2 \iota \hat C \rho^{-1/a}$ and $\iota \ge 1$, (\ref{mombound}) follows.
\medskip\noindent
\underline{P(ii)(b)}:
Let $S = (S_n)_{n\in \bar{\mathcal{I}}_{\mathcal{T}}}$ with $S_n := \theta_n \mathcal{T}$. By (ii)(d), it is enough
to show that $X_1 = (Z_{\tau}, \tau) \in \sigma(Z,S)$ a.s.. Since $Z_{\tau} = \sum_{n=0}^{\infty}
1_{\{\tau = n\}}Z_n \in \sigma(Z, \tau)$, it suffices to show that $\tau \in \sigma(S)$ a.s..
Using the definition of the $T_k$'s, we verify by induction that each $T_k$ is a.s.\ measurable in
$\sigma(S)$. Since $N \in \sigma((T_k)_{k \in \mathbb{N}_0})$, both $N$ and $T_N$ are also a.s.\
in $\sigma(S)$. Therefore, a.s. $\tau \in \sigma(\theta_{t_0}S) \subset \sigma(S)$.
\medskip
With all hypotheses verified, Proposition~\ref{prop:adfun1} implies that $X \in AV(\hat
\varepsilon, (v,u))$, where $(v,u) = E_{\nu}[X_1|D]$ and $\hat \varepsilon = 2K\phi^{(a-1)/a}$.
To conclude, observe that $u = E_{\nu}[\tau|D] \ge t_0 \ge \iota \hat C \rho^{-1/a} > 0$,
so that $K = 6(m+1) \iota \hat C \rho^{-1/a} \le 6(m+1)u$. Therefore, $\hat \varepsilon
\le \varepsilon$ and the proposition follows. In the case $\mathcal{I}_{\mathcal{T}} =
\emptyset$, we conclude similarly since $u=1$ and $K=(m+1)$.
\end{proof}
\section{Proof of Theorem~\ref{mainthm}}
\label{sec:proofmainthm}
In this section we show how to put the model defined in Section~\ref{sec:model} in the
context of Section~\ref{sec:prep}, and we prove the LLN using Propositions~\ref{prop:avrt}
and \ref{prop:adfun}.
\subsection{Two further lemmas}
Before we start, we first derive two lemmas (Lemmas~\ref{lem1} and \ref{lem2} below) that
will be needed in Section~\ref{subsec:proofmainthm}. The first lemma relates the laws of
the environment as seen from $W_n$ and from $W_0$. The second lemma is an extension of the
conditional cone-mixing property for functions that depend also on $Y$.
\bl{lem1}
$\bar{\mu}_n\ll \bar\mu$ for all $n \in \mathbb{N}$.
\end{lemma}
\begin{proof}
For $t\geq 0$, let $\bar\mu_{t-}$ denote the law of $\theta_{W_{t-}}\xi_t$ under
$\mathbb{P}_{\mu}$. First we will show that $\bar\mu_{t-} \ll \mu$. This is a consequence
of the fact that $\mu$ is translation-invariant equilibrium, and remains true if we replace
$W_{t-}$ by any random variable taking values in $\mathbb{Z}^d$. Indeed, if $\mu(A)=0$ then
$\mathbb{P}_{\mu}(\theta_x \xi_t \in A)=0$ for every $x \in \mathbb{Z}^d$, so
\be{}
\begin{aligned}
\bar{\mu}_{t-}(A) = \mathbb{P}_{\mu}(\theta_{W_{t-}}\xi_t \in A)
= \sum_{x \in \mathbb{Z}^d} \mathbb{P}_{\mu}(W_{t-}=x, \theta_x\xi_t \in A) = 0.
\end{aligned}
\end{equation}
Now take $n \in \mathbb{N}$ and let $g_n := \frac{d\bar\mu_{n-}}{d \mu}$. For any measurable
$f\geq 0$,
\begin{align}
\mathbb{E}_{\mu}\left[f(\theta_{W_n}\xi_n)\right] = \mathbb{E}_{\bar \mu}\left[f(\theta_{Z_n}\xi_n)\right]
= & \sum_{x \in \mathbb{Z}^d} \mathbb{E}_{\bar \mu}
\left[1_{\{Z_n-Z_{n-}=x\}}f(\theta_x\theta_{Z_{n-}}\xi_n)\right] \nonumber \\
= & \sum_{x \in \mathbb{Z}^d} \mathbb{E}_{\bar \mu}
\left[\mathbb{P}_{\theta_{Z_{n-}}\xi_n}(W_0=x)f(\theta_x\theta_{Z_{n-}}\xi_n)\right] \nonumber \\
= & \sum_{x \in \mathbb{Z}^d} \mathbb{E}_{\mu}
\left[\mathbb{P}_{\theta_{W_{n-}}\xi_n}(W_0=x)f(\theta_x\theta_{W_{n-}}\xi_n)\right] \nonumber \\
= & \sum_{x \in \mathbb{Z}^d} \mathbb{E}_{\mu}
\left[g_n(\xi_0)\mathbb{P}_{\xi_0}(W_0=x)f(\theta_x\xi_0)\right] \nonumber \\
= & \sum_{x \in \mathbb{Z}^d} \mathbb{E}_{\mu}
\left[g_n(\xi_0)1_{\{W_0=x\}} f(\theta_x\xi_0)\right] \nonumber\\
= & \; \mathbb{E}_{\mu} \left[ g_n(\xi_0) f(\theta_{W_0} \xi_0)\right]
\end{align}
where, for the second equality, we use (A3).
\end{proof}
\bl{lem2}
For $L$ large enough and for all nonnegative $f\in\mathscr{C}_{R,\infty}(m) \vee \mathscr{Y}_{\infty}$,
\be{eq:lem2}
\left\lvert\mathbb{E}_{\eta} \left[\theta_L f \mid \Gamma_L\right]
- \mathbb{E}_{\bar \mu}[\theta_Lf\mid\Gamma_L ]\right\rvert
\leq \Phi_L \|f\|_{\infty} \quad \text{for } \;\bar{\mu} \text{-a.e. } \eta.
\end{equation}
\end{lemma}
\begin{proof}
Put $f_y(\eta) = f(\eta,y)$ and abbreviate $Y^{(L)} = (Y_k)_{k > L}$. Then $\theta_L f
= \theta_L f_{Y^{(L)}}$. Since $\Gamma_L$ depends on $Y$ only through $(Y_k)_{k \le L}$, we have
\be{}
\mathbb{E}_{\eta}[\theta_Lf\,1_{\Gamma_L} \mid Y^{(L)}]
= \mathbb{E}_{\eta}\big[\theta_Lf_{(\cdot)}\,1_{\Gamma_L}\big] \circ (Y^{(L)}),
\end{equation}
and \eqref{eq:lem2} follows from (H4) applied to $f_y$.
\end{proof}
\subsection{Proof of Theorem~\ref{mainthm}}
\label{subsec:proofmainthm}
\begin{proof}
Extend $\xi$ and $Z$ for times $t \in [-1,0]$ by taking them constant in this interval,
and let $Y_0$ be a copy of $Y_1$ independent of $\mathscr{F}_{\infty}$. Put
\begin{equation}\label{defmkvchain}
\begin{array}{rcl}
\mathcal{X}_0 & := & \left( \xi_{[-1,0]},Z_{[-1,0]},Y_0\right), \\
\mathcal{X}_{n+1} & := & \left(\theta_{Z_n}\xi_{[n,n+1]},(Z_{t+n}-Z_n)_{0 \le t \le 1}, Y_{n+1}\right), \; n \in \mathbb{N}_0.
\end{array}
\end{equation}
Then $(\mathcal{X}_n)_{n \in \mathbb{N}_0}$ is a time-homogeneous Markov chain;
to avoid confusion, we will denote its time-shift operator by $\bar\theta_n$.
Note that $\mathcal{F}_n = \mathscr{F}_n$ $\forall$ $n \in \mathbb{N}\cup\{\infty\}$ and that, for functions $f \in \mathcal{F}_{\infty}$,
$\bar \theta_n f = \theta_{Z_n} \theta_n f$ $\forall$ $n \in \mathbb{N}_0$.
Fix $L \in \mathbb{N}$ large enough and put
\begin{equation}\label{defcurlyTL}
\mathcal{T}_L := L + 1_{\Gamma_L} \lceil \theta_L \mathcal{S} \rceil.
\end{equation}
By (\ref{eqA1}) and since $\Gamma_L \in \mathscr{F}_L$ and $Z$ is $\mathscr{F}$-adapted,
$\mathcal{T}_L$ is an $\mathcal{F}$-stopping time and
$(Z_n)_{n \in \mathbb{N}_0}$ is an additive functional of $(\mathcal{X}_n)_{n \in \mathbb{N}}$ as in Section \ref{sec:adfun}.
Next, we will verify (\ref{sup:abscont}) for $\mathcal{X}$ and the hypotheses of Proposition~\ref{prop:adfun} for
$Z$ and $\mathcal{T}_L$ under $\mathbb{P}_{\bar \mu}$. These hypotheses will be referred
to with the prefix P. The notation here is consistent in the sense that parameters in
Section~\ref{sec:model} are named according to their role in Section~\ref{sec:prep}; the
presence/absence of a subscript $L$ indicates whether the parameter depends on $L$ or not.
\medskip\noindent
\underline{(\ref{sup:abscont})}:
Noting that, for nonnegative $f \in \mathcal{F}_{\infty}$ and $n \in \mathbb{N}_0$,
\begin{equation}\label{relMCwithmodel}
E_{\mathcal{X}_n} \left[ f \right] = \mathbb{E}_{\theta_{Z_n}\xi_n}\left[ f \right] \qquad \mathbb{P}_{\bar \mu} \text{-a.s.},
\end{equation}
this follows from Lemma \ref{lem1} and (\ref{assumpWZ}--\ref{assumpX0}).
\medskip\noindent
\underline{P(i)}:
We will find $D_n$ for $n\ge L$. This is enough, since both $\mathcal{I}_{\mathcal{T}_L}$
and $\bar{ \mathcal{I}}_{\mathcal{T}_L}$ are subsets of $[L,\infty) \cap \mathbb{N}$. Using (A1) and (H1), we
may write
\begin{equation}
\begin{array}{rcl}
D & = & \Gamma_L \cap \{\|Z_{t+L}\|\le mt \; \forall \; t \ge 0\}, \\
\bar\theta_n D & = & \bar \theta_n \Gamma_L \cap \{\|Z_{t+n+L}-Z_n\|\le mt \; \forall \; t \ge 0\}.\\
\end{array}
\end{equation}
Intersecting the two above events, we get
\begin{equation}
D\cap\bar\theta_n D = \Gamma_L \cap \{\|Z_t\| \le mt
\; \forall \; t \in [0,n]\} \cap \bar \theta_n D,
\end{equation}
i.e., P(i) holds with $D_n :=\Gamma_L \cap \{\|Z_t\| \le mt \; \forall \;
t \in [0,n]\} \in \mathcal{F}_n$ for $n \ge L$.
For the remaining items, note that, by \eqref{relMCwithmodel},
the distribution of $(Z,\mathcal{T}_L)$ under $P_{\mathcal{X}_0}$ is $\mathbb{P}_{\bar \mu}$-a.s. the same as under $\mathbb{P}_{\xi_0}$.
\medskip\noindent
\underline{P(ii)(a)}:
Since $\{\mathcal{T}_L = \infty\} = \{\theta_L\mathcal{S} = \infty\} \cap \Gamma_L$, we get
from (H2) and (H3)(1) that, $\mathbb{P}_{\bar \mu}$-a.s.,
\begin{equation}
\mathbb{P}_{\xi_0}\left(\mathcal{T}_L = \infty \right)
= \mathbb{P}_{\xi_0}\left( \theta_L \mathcal{S} = \infty \mid \Gamma_L\right)\mathbb{P}_{\xi_0}(\Gamma_L)\ge \kappa_L \gamma_L > 0,
\end{equation}
so that we can take $\rho_L := \kappa_L \gamma_L$.
\medskip\noindent
\underline{P$(ii)(b)$}:
By the definition of $\mathcal{T}_L$, we have
\begin{align}
\mathcal{T}_L^a 1_{\{\mathcal{T}_L < \infty\}}
& = L^a 1_{\Gamma_L^c} + \left(L+ \lceil \theta_L \mathcal{S} \rceil \right)^a
1_{\Gamma_L \cap \{\theta_L \mathcal{S} < \infty\}}\nonumber \\
& \le L^a 1_{\Gamma_L^c} + \left(L+1+\theta_L\mathcal{S} \right)^a
1_{\Gamma_L \cap \{\theta_L \mathcal{S} < \infty\}}\nonumber \\
& \le 2^{a-1}(L+1)^a + 2^{a-1} \left((\theta_L \mathcal{S})^a
1_{ \{\theta_L \mathcal{S}<\infty\}}\right) 1_{\Gamma_L} .
\end{align}
Therefore, by (H3)(2), we get
\begin{align}
\mathbb{E}_{\xi_0}\left[ \mathcal{T}_L^a 1_{\{\mathcal{T}_L<\infty\}}\right]
\le 2^a((L+1)^a+(1\vee\psi_L)^a) \le \left[2(L+1+1\vee\psi_L)\right]^a \quad \mathbb{P}_{\bar \mu} \text{-a.s.},
\end{align}
so that we can take $C_L := 2(L+1+1\vee\psi_L)$.
\medskip\noindent
\underline{P(ii)(c)}:
This follows from (H1) and the definition of $\mathcal{S}$.
\medskip\noindent
\underline{P(ii)(d)}:
First note that, for any $n \in \bar{\mathcal{I}}_{\mathcal{T}_L}$, $\bar\theta_n \mathcal{T}_L \in \sigma(Z, \bar\theta_n\Gamma_L)$.
Since $n \ge L$, on $\{\mathcal{T}_L = \infty\} = \Gamma_L \cap \{\theta_L \mathcal{S}=\infty\}$,
$Z$, $\bar\theta_n \Gamma_L$ and $\{\theta_L \mathcal{S} = \infty\}$ are all measurable in $\theta_L(\mathscr{C}_{R,\infty}(m) \vee \mathscr{Y}_{\infty})$;
this follows from (A2), (H1) and the assumptions on $\Gamma_L$.
Noting that, for any two probability measures $\nu_1$, $\nu_2$ and an event $A$,
\begin{equation}\label{compcondTV}
\| \nu_1(\cdot \mid A) - \nu_2(\cdot \mid A) \|_{TV} \le 2 \frac{\| \nu_1 - \nu_2 \|_{TV}}{\nu_1(A) \vee \nu_2(A)}
\end{equation}
where $\| \cdot \|_{TV}$ stands for total variation distance, we see that
P(ii)(d) follows from Lemma \ref{lem2} and (H3)(1) with $\phi_L := 2 \Phi_L/\kappa_L \to 0$ as $L \to \infty$ by (H4).
\medskip
Thus, for large enough $L$, we can conclude by Proposition~\ref{prop:adfun} that there
exists a sequence of times $(\tau_k)_{k \in \mathbb{N}_0}$ with $\lim_{k\to\infty}
\tau_k = \infty$ a.s.\ such that $(Z_{\tau_k},\tau_k)_{k \in \mathbb{N}_0} \in
AV(\varepsilon_L, (v_L,u_L))$, where
\begin{equation}
\label{vLuL}
\begin{array}{rcl}
v_L & = & \mathbb{E}_{\bar\mu}[Z_{\tau_1}|D], \\
u_L & = & \mathbb{E}_{\bar\mu}[\tau_1|D] > 0, \\
\varepsilon_L & = & 12(m+1)u_L \phi_L^{(a-1)/a}.
\end{array}
\end{equation}
From (\ref{vLuL}) and P(ii)(c), Proposition~\ref{prop:avrt} implies that $Z \in
AV(\delta_L, w_L)$, where
\begin{equation}
\begin{array}{rcl}
w_L & = & v_L/u_L, \\
\delta_L & = & (3m+1)12(m+1)\phi_L^{(a-1)/a}.
\end{array}
\end{equation}
By (H4), $\lim_{L\to\infty}\delta_L = 0$. As was observed after Definition~\ref{def:av},
this implies that $w := \lim_{L \to \infty} w_L$ exists and that $\lim_{t \to \infty}t^{-1}
Z_t = w$ $\mathbb{P}_{\bar\mu}$-a.s., which, by (\ref{assumpWZ}--\ref{assumpX0}), implies the same for
$W$, $\mathbb{P}_{\mu}$-a.s.
\end{proof}
We have at this point finished the proof of our LLN. In the following sections, we will
look at examples that satisfy (H1)--(H4). Section~\ref{sec:proofthm2} is devoted to the
$(\infty,0)$-model for two classes of one-dimensional spin-flip systems. In Section~\ref{sec:oex}
we discuss three additional models where the hypotheses are satisfied, and one where
they are not.
\section{Proof of Theorem \ref{thm:cond}}
\label{sec:proofthm2}
We begin with a proper definition of the $(\infty,0)$-model in Section~\ref{sec:formaldef},
where we identify $Z$ and $W_0$ of Section~\ref{subsec:mainass}.
In Section~\ref{sec:sfsbr}, we first define suitable versions of spin-flip systems with bounded
rates. After checking assumptions (A1)--(A3), we define events
$\Gamma_L$ satisfying (H1) and (H2) for which we then verify (H3). We also derive uniform integrability properties of
$t^{-1}W_t$ which are the key for convergence in $L^p$ once we have the LLN. In
Sections~\ref{sec:mlessep} and \ref{sec:subcritdep}, we specialize to particular constructions
in order to prove (H4), which is the hardest of the four hypotheses. Section~\ref{sec:signspeed}
is devoted to proving a criterion for positive or negative speed.
\subsection{Definition of the model}
\label{sec:formaldef}
Assume that $\xi$ is a c\`adl\`ag process with state space $E:=\{0,1\}^{\mathbb{Z}}$. We will
define the walk $W$ in several steps, and a monotonicity property will follow.
\subsubsection{Identification of $Z$ and $W_0$}
\label{identZW0}
First, let $Tr^+=Tr^+(\eta)$ and $Tr^-=Tr^-(\eta)$ denote the locations of the closest
traps to the right and to the left of the origin in the configuration $\eta \in E$, i.e.,
\begin{equation}
\label{deftraps}
\begin{array}{rcl}
Tr^+(\eta) & := & \inf \{x \in \mathbb{N}_0\colon\,\eta(x) = 1, \eta(x+1)=0\},\\
Tr^-(\eta) & := & \sup \{x \in -\mathbb{N}_0\colon\,\eta(x) = 1, \eta(x+1)=0\},
\end{array}
\end{equation}
with the convention that $\inf\emptyset = \infty$ and $\sup\emptyset = - \infty$. For $i,j
\in \{0,1\}$, abbreviate $\langle i,j \rangle := \{\eta \in E \colon\, \eta(0) = i, \eta(1) = j\}$.
Let $\bar E := \langle 1,0 \rangle$, i.e., the set of all the configurations with a trap
at the origin.
Next, we define the functional $J$ that gives the jumps in $W$. For $b \in \{0,1\}$ and
$\eta \in E$, let
\begin{equation}
\label{defJ}
J(\eta,b) := Tr^+\left( 1_{\langle 1,1 \rangle}
+ b1_{\langle 0,1 \rangle} \right) + Tr^-\left( 1_{\langle 0,0 \rangle}
+ (1-b)1_{\langle 0,1 \rangle} \right),
\end{equation}
i.e., $J$ is equal to either the left or the right trap, depending on the configuration around the
origin. In the case of an inverted trap ($\langle 0,1 \rangle$), the
direction of the jump is decided by the value of $b$. Observe that $J=Tr^+=Tr^-=0$ when $\eta \in
\bar E$, independently of the value of $b$.
Let $b_0$ be a Bernoulli($\frac12$) random variable independent of $\xi$ and set
\begin{equation}\label{defW0}
W_0 = X_0 := J(\xi_0,b_0).
\end{equation}
Now let $(b_{n,k})_{n,k \in \mathbb{N}}$ be a double-indexed i.i.d.\ sequence
of Bernoulli($\frac12$) r.v.'s independent of $(\xi,b_0)$.
Put $\tau_0 := 0$ and, for $k \geq 0$,
\begin{equation}
\label{defjumps}
\begin{array}{rcll}
\tau_{k+1} & := & \left\{ \begin{array}{l} \infty \\
\inf\left\{t> \tau_k \colon\, \left(\xi_t(X_{k}),\xi_t(X_{k}+1)\right) \neq (1,0)\right\}
\end{array}\right. & \begin{array}{l} \text{if } |X_k| = \infty,\\ \text{otherwise,} \end{array} \\
X_{k+1} & := & \left\{ \begin{array}{l} X_k \\
X_k + J\left(\theta_{X_k}\xi_{\tau_k},b_{\lceil \tau_{k+1} \rceil, k+1}\right)\\
\end{array}\right. & \begin{array}{l} \text{if } \tau_{k+1} = \infty,\\ \text{otherwise.} \end{array}
\end{array}
\end{equation}
Since $\xi$ is c\`adl\`ag, for any $k \in \mathbb{N}_0$ we either have $\tau_{k} = \infty$ or $\tau_{k+1} > \tau_{k}$. We define
$(W_t)_{t\geq 0}$ as the path that jumps $X_{k+1} - X_k$ at time $\tau_{k+1}$ and is constant
between jumps, i.e.,
\begin{equation}
\label{defW}
W_t := \sum_{k=0}^{\infty}1_{\{\tau_k \le t<\tau_{k+1}\}} X_k.
\end{equation}
With this definition, it is clear that $W_t$ is c\`adl\`ag and, by (\ref{defW0}--\ref{defjumps}),
\begin{equation}\label{Wadditive}
W_{n+t}-W_n = \theta_{W_n} \theta_n W_t \;\; \text{on } \{W_n < \infty\} \;\; \forall \; n \in \mathbb{N}_0, t \ge 0.
\end{equation}
Therefore, defining $Z$ by
\begin{equation}\label{defZ}
Z_t := 1_{\{ \xi_0 \in \bar E \}} W_t, \qquad t \ge 0,
\end{equation}
we get $W_t = W_0 + \theta_{W_0}Z_t$ on $\{W_0 < \infty\}$
since, in this case, $\theta_{W_0}\xi_0 \in \bar E$, and $W_0=0$ on $\bar E$.
\subsubsection{Monotonicity}
\label{subsubsec:monotonicity}
The following monotonicity property will be helpful in checking (H3). In order to state it,
we first endow both $E$ and $D([0,\infty),E)$ with the usual partial ordering, i.e., for $\eta_1,
\eta_2 \in E$, $\eta_1 \leq \eta_2$ means that $\eta_1(x)\leq\eta_2(x)$ for all $x \in \mathbb{Z}$, while,
for $\xi^{(1)}, \xi^{(2)} \in D([0,\infty),E)$, $\xi^{(1)} \le \xi^{(2)}$ means that $\xi^{(1)}_t
\leq \xi^{(2)}_t$ for all $t \geq 0$.
\begin{lemma}
\label{monot}
Fix a realization of $b_0$ and $(b_{n,k})_{n,k \in \mathbb{N}}$. If $\xi^{(1)} \le \xi^{(2)}$, then
\begin{equation}\label{eqmonot}
W_t\left(\xi^{(1)},b_0,(b_{n,k})_{n,k \in \mathbb{N}}\right) \leq W_t\left(\xi^{(2)}, b_0, (b_{n,k})_{n,k \in \mathbb{N}}\right)
\end{equation}
for all $t \ge 0$.
\end{lemma}
\begin{proof}
This is a straightforward consequence of the definition. We need only to understand what
happens when the two walks separate and, at such moments, the second walk is always to the right of
the first.
\end{proof}
\subsection{Spin-flip systems with bounded flip rates}
\label{sec:sfsbr}
\subsubsection{Dynamic random environment}
\label{subsubsec:DRE}
From now on we will take $\xi$ to be a single-site spin-flip system with translation-invariant
and bounded flip rates. We may assume that the rates at the origin are of the form
\begin{equation}
\label{rates}
c(\eta) = \left\{ \begin{array}{ccl}
c_0 + \lambda_0 p_0(\eta) & \text{ when } & \eta(0) = 1, \\
c_1 + \lambda_1 p_1(\eta) & \text{ when } & \eta(0) = 0,
\end{array}
\right.
\end{equation}
where $c_i,\lambda_i > 0$ and $p_i \in [0,1]$. We assume the existence conditions of Liggett~\cite{Li},
Chapter I, which in our setting amounts to the additional requirement that $c(\cdot)$ has finite triple norm.
This is automatically satisfied in the $M<\epsilon$ regime or when $c(\cdot)$ has finite range.
From (\ref{rates}), we see that the IPS is stochastically dominated by the system $\xi^+$ with rates
\begin{equation}
\label{upperisf}
c^+(\eta) = \left\{ \begin{array}{ccl}
c_0 & \text{ when } & \eta(0) = 1, \\
c_1 + \lambda_1 & \text{ when } &\eta(0) = 0,
\end{array}\right.
\end{equation}
while it stochastically dominates the system $\xi^-$ with rates
\begin{equation}
\label{lowerisf}
c^-(\eta) = \left\{ \begin{array}{ccl}
c_0 + \lambda_0& \text{ when } & \eta(0) = 1, \\
c_1 & \text{ when } & \eta(0) = 0.
\end{array}
\right.
\end{equation}
These are the rates of two independent spin-flip systems with respective densities
$\rho^+ := (c_1+\lambda_1)/\lambda^+$ and $\rho^- := c_1/\lambda^-$ where $\lambda^+ := c_0+c_1+\lambda_1$ and $\lambda^- := c_0 + \lambda_0+c_1$.
Consequently, any equilibrium for $\xi$ is stochastically dominated by $\nu_{\rho^+}$ and dominates $\nu_{\rho^-}$,
where $\nu_\rho$ is a Bernoulli product measure with density $\rho$.
We will take as the dynamic random environment the triple $\Xi := (\xi^-,\xi, \xi^+)$
starting from the same initial configuration and coupled together via the basic
(or Vasershtein) coupling, which implements
the stochastic ordering as an a.s.\ partial ordering. More precisely, $\Xi$ is the IPS with
state space $E^3$ whose rates are translation invariant and at the origin are given
schematically by (the configuration of the middle coordinate is $\eta$),
\begin{equation}\
\label{ratescoupl}
\begin{array}{ccl}
(000) & \to & \left\{\begin{array}{cl}
(111) & c_1, \\
(011) & c(\eta) - c_1, \\
(001) & c_1 + \lambda_1 - c(\eta), \\
\end{array}\right.\\
(001) & \to & \left\{\begin{array}{cl}
(111) & c_1,\\
(011) & c(\eta) - c_1, \\
(000) & c_0,\\
\end{array}\right. \\
(011) & \to & \left\{\begin{array}{cl}
(111) & c_1, \\
(000) & c_0, \\
(001) & c(\eta) - c_0, \\
\end{array}\right.\\
(111) & \to & \left\{\begin{array}{cl}
(000) & c_0, \\
(001) & c(\eta) - c_0, \\
(011) & c_0 + \lambda_0 - c(\eta). \\
\end{array}\right. \\
\end{array}
\end{equation}
\subsubsection{Verification of (A1)--(A3)}
\label{checkassumptions}
Under our assumptions, $\lim_{k \to \infty}\tau_k = \infty$ and $X_0 < \infty$ $\mathbb{P}_\mu$-a.s., as $\xi$ has bounded flip rates per site and $\mu$ dominates and is dominated by non-trivial product measures. By induction, $X_k < \infty$ a.s.\ for every $k \in \mathbb{N}$ as well, since the law of $\theta_{X_{k-1}}\xi_{\tau_k}$ is absolutely continuous w.r.t.\ $\mu$, which can be verified by approximating $\tau_k$ from above by times taking values in a countable set. Therefore, $W_t$ is finite for all $t \ge 0$.
Set $Y_n := (b_{n,k})_{k \in \mathbb{N}}$. Then $Z$ is $\mathscr{F}$-adapted as it is independent of $b_0$.
(A1) follows by \eqref{Wadditive} and \eqref{defZ}, and (A3) follows either from the recursive construction \eqref{defjumps}
or by noting that $Z$ has no deterministic jumps and $\theta_{Z_n} \theta_n W_0 = 0$.
To verify (A2), note that $\{J=x\}$ depends on $\eta$ only through $(\eta(y))_{ y \in \{0 \wedge x, \ldots, 0 \vee x + 1\}}$
so we may take $R=1$.
\subsubsection{Definition of $\Gamma_L$ and verification of (H1)--(H3)}
\label{subsubsec:defGammaL}
Using $\Xi$, we can define the events $\Gamma_L$ by
\begin{equation}
\Gamma_L := \big\{\xi_t^{\pm}(x)=\xi_0^{\pm}(x) \;
\forall \; t \in [0,L] \text{, } x = 0,1\big\}.
\end{equation}
Then $\Gamma_L \in \mathscr{C}_{1,L}(m)$ for any $m > 0$.
When $\xi_0^{\pm} \in \bar E$, $\Gamma_L$ implies that there is a trap at the origin between times $0$ and $L$; since $\bar \mu(\bar E)=1$, (H1) holds. The probability of $\Gamma_L$ is positive
and depends on $\Xi_0$ only through the states at $0$ and $1$,
so (H2) is also satisfied.
In order to verify (H3), we will take advantage of Lemma~\ref{monot} and the stochastic domination in
$\Xi$ to define two auxiliary processes $H^{\pm} = (H^{\pm}_t)_{t\ge0}$ which we can control and which will bound $Z$.
This will also allow us to deduce uniform integrability properties.
In the following we will suppose that $\xi^{{\pm}}_0 \in \bar E$. Let $G_0 = U_0 := 0$ and, for $k \geq 0$,
\begin{equation}\label{defUG+}
\begin{array}{rcl}
U_{k+1} & := & \inf \left\{t>U_k \colon\, \xi^{+}_t(G_k+1)=1 \right\}, \\
G_{k+1} & := & G_k + Tr^{+}\Big(\theta_{G_k}\xi^{+}_{U_{k+1}}\Big) \\
\end{array}
\end{equation}
and put
\begin{equation}\label{defH+}
H^{+}_t := \sum_{k=0}^{\infty} 1_{\{U_k \le t < U_{k+1}\}}G_{k+1}.
\end{equation}
Define $H^-$ analogously, using $Tr^-$ and $\xi^-$ instead and switching $1$'s to $0$'s in \eqref{defUG+}.
Then $H^+$ ($H^-$) is the process that, observing $\xi^+$ ($\xi^-$) , waits to the left of a hole (on a particle) until it flips to a
particle (hole), and then jumps to the right (left) to the next trap.
Therefore, by Lemma~\ref{monot} and the definition of $Z$, $H^-_t \le Z_t \le H^+_t$ $\forall$ $t \ge 0$.
Note that $H^+$ depends only on $(\xi^+(x))_{x \ge 1}$, and analogously for $H^-$.
In the following, we will write $\mathbb{Z}_{\le x} : = \mathbb{Z} \cap (-\infty,x]$ and analogously for $\mathbb{Z}_{\ge x}$.
\begin{lemma}
\label{sbisf}
Fix $\rho_* \in (0,\rho^-]$ and $\rho^* \in [\rho^+,1)$.
There exist $m,a,\psi_*$ $\in (0,\infty)$ and $\kappa_* \in (0,1)$, depending on $\rho_*$, $\rho^*$ and $\lambda^{\pm}$, such that, for any probability measure $\bar \nu$ on $\bar E$ that stochastically dominates $\nu_{\rho_*}$ on $\mathbb{Z}_{\le-1}$ and is dominated by $\nu_{\rho^*}$ on $\mathbb{Z}_{\ge 2}$,
\begin{equation}\label{sbisfeq1}
\text{(a)} \;\; \sup_{t \ge 1}\mathbb{E}_{\bar \nu}\left[e^{a \left( t^{-1}|H^{\pm}_t| \right)} \right] \le \psi_*
\end{equation}
and, setting
\begin{equation}\label{defSpm}
\mathcal{S}^{\pm}:= \inf\{t > 0 \colon\, |H^{\pm}_t| > mt\}, \qquad \widehat{\mathcal{S}}^{\pm}:= \sup\{t > 0 \colon\, |H^{\pm}_t| > mt\},
\end{equation}
then
\begin{equation}\label{sbisfeq3}
\begin{array}{rl}
\text{(b)} & \mathbb{P}_{\bar\nu}\left(\mathcal{S}^{\pm} = \infty \right) \ge \kappa_*,\\
\text{(c)} & \mathbb{E}_{\bar\nu}\left[ e^{a \widehat{\mathcal{S}}^{\pm}} \right] \le \psi_*.
\end{array}
\end{equation}
\end{lemma}
Before proving this lemma, let us see how it leads to (H3).
We will show that there exist $m,a,\psi \in (0,\infty)$ and $\kappa \in (0,1)$
such that, for all $L \ge 1$ and $\eta \in \bar E$,
\begin{equation}\label{verifH31}
\mathbb{P}_{\eta} \left( \theta_L\mathcal{S}=\infty \mid \Gamma_L \right) \ge \kappa
\end{equation}
and
\begin{equation}\label{verifH32}
\mathbb{E}_{\eta} \left[ e^{a (\theta_L \mathcal{S}) } 1_{\{\theta_L\mathcal{S}<\infty \}} \mid \Gamma_L \right] \le \psi,
\end{equation}
which clearly imply (H3).
Let us verify \eqref{verifH31}.
First note that $\theta_L \mathcal{S} \ge \theta_L (\mathcal{S}^+ \wedge \mathcal{S}^-)$,
and that the latter is nonincreasing in $(\eta(x))_{x \ge 2}$ and nondecreasing in $(\eta(x))_{x \le -1}$.
Therefore we may assume that $\eta = \eta_{01}$ which is the configuration in $\bar E$
with all $0$'s on $\mathbb{Z}_{\le-1}$ and all $1$'s on $\mathbb{Z}_{\ge 2}$. In this case,
$\xi^-_L$ is distributed as $\nu_{\rho^L_0}$ on $\mathbb{Z}_{\le-1}$ and $\xi^+_L$ as $\nu_{\rho^L_1}$ on $\mathbb{Z}_{\ge 2}$,
where $\rho_0^L = \rho^-(1-e^{-\lambda^- L})$ and $\rho_1^L = \rho^+ + e^{-\lambda^+ L}(1-\rho^+)$.
Furthermore, on $\Gamma_L$, $\xi^{\pm}_L \in \bar E$.
Let now $m,a,\psi_*$ and $\kappa^*$ as in Lemma~\ref{sbisf} for $\rho_*:=\rho_0^1$ and $\rho^* := \rho_1^1$, and let $\bar\nu_L$ be the distribution of $\bar \eta_L \in \bar E$ given by $\xi^-_L$ on $\mathbb{Z}_{\le-1}$ and $\xi^+_L$ on $\mathbb{Z}_{\ge 2}$.
Noting that $\bar \eta_L$ is independent of $\Gamma_L$ and that $\mathcal{S}^+$ and $\mathcal{S}^-$ are independent,
we use the previous observations, the Markov property and Lemma~\ref{sbisf}(b) to write
\begin{align}\label{compverifH31}
\mathbb{P}_{\eta}\left(\theta_L \mathcal{S} = \infty \mid \Gamma_L\right)
& \ge \mathbb{P}_{\eta_{01}}\left(\theta_L(\mathcal{S}^+ \wedge \mathcal{S}^-) = \infty \mid \Gamma_L\right) \nonumber \\
& = \mathbb{E}_{\eta_{01}}\left[ 1_{\Gamma_L} \mathbb{P}_{\bar \eta_L}\left(\mathcal{S}^+ \wedge \mathcal{S}^- = \infty\right)\right]\mathbb{P}_{\eta_{01}}\left(\Gamma_L\right)^{-1} \nonumber \\
& = \mathbb{P}_{\bar\nu_L}\left(\mathcal{S}^+ = \infty\right)\mathbb{P}_{\bar\nu_L}\left(\mathcal{S}^- = \infty\right) \ge \kappa_*^2 \in (0,1),
\end{align}
and we may take $\kappa := \kappa_*^2$.
For \eqref{verifH32}, note now that, when finite, $\theta_L\mathcal{S} < \theta_L(\widehat{\mathcal{S}}^+ \vee \widehat{\mathcal{S}}^-)$
and the latter is is nondecreasing in $(\eta(x))_{x \ge 2}$ and nonincreasing in $(\eta(x))_{x \le -1}$.
Therefore we may again assume $\eta= \eta_{01}$ and write, using Lemma~\ref{sbisf}(c),
\begin{align}\label{compverifH32}
\mathbb{E}_{\eta}\left[\theta_L \left(e^{a \mathcal{S}}1_{\{\mathcal{S}=\infty\}}\right) \mid \Gamma_L\right]
& \le \mathbb{E}_{\eta_{01}}\left[ \theta_L e^{a \left( \widehat{\mathcal{S}}^+ + \widehat{\mathcal{S}}^- \right) } \mid \Gamma_L\right] \nonumber \\
& = \mathbb{E}_{\bar\nu_L}\left[e^{a \widehat{\mathcal{S}}^+} \right] \mathbb{E}_{\bar\nu_L}\left[e^{a \widehat{\mathcal{S}}^-} \right] \le \psi_*^2 \in (0,\infty),
\end{align}
and we can take $\psi := \psi^2_*$. All that is left to do is to prove Lemma~\ref{sbisf}.
\begin{proof}[Proof of Lemma~\ref{sbisf}]
By symmetry, it is enough to prove (a)--(c) for $H^+$.
Since $H^+$, $\mathcal{S}^+$ and $\widehat{\mathcal{S}}^+$ are monotone, we may assume that
$\xi^+$ has rates $\lambda^+ \rho^*$ to flip from holes to particles
and $\lambda^+ (1-\rho^*)$ from particles to holes and starts from $\nu_{\rho^*}$, which is the equilibrium measure.
In this case, the increments $G_{k+1}-G_k$ are i.i.d.\ Geom($1-\rho^*$),
and $U_{k+1}-U_k$ are i.i.d.\ Exp($\lambda^+ \rho^*$), independent from $(G_k)_{k \in \mathbb{N}_0}$.
Therefore, $H^+$ is a c\`adl\`ag L\'evy process and $H^+_1$ has an exponential moment, so (a) promptly follows.
Moreover, $H^+$ satisfies a large deviation estimate of the type
\begin{equation}\label{pr:sbisfeq1}
\mathbb{P}_{\nu_{\rho^*}}\left(\exists \; s>t \text{ such that } H^+_s > ms\right)
\leq K_1 e^{-K_2 t} \; \text{ for all } t > 0,
\end{equation}
where $m$, $K_1$ and $K_2$ are functions of ($\rho^*$, $\lambda^+$), which proves (c).
In particular, $\widehat{\mathcal{S}}^+ < \infty$ a.s., which implies that $\mathbb{P}_{\nu_{\rho^*}}(H^+_s \le m(s+n^*) \; \forall \; s \ge 0) \ge \frac12$ for some $n^*$ large enough;
then
\begin{align}\label{pr:sbisfeq2}
\mathbb{P}_{\nu_{\rho^*}}\left(\mathcal{S}^+ = \infty \right)
& \ge \mathbb{P}_{\nu_{\rho^*}}\left( H^+_{n^*} =0, H^+_{n^*+s} - H^+_{n^*} \le m(s+n^*) \; \forall \; s \ge 0 \right) \nonumber\\
& = \mathbb{P}_{\nu_{\rho^*}}\left( H^+_{n^*} =0 \right)\mathbb{P}_{\nu_{\rho^*}}\left(H^+_s \le m(s+n^*) \; \forall \; s \ge 0 \right) =: \kappa_* >0,
\end{align}
proving (b).
\end{proof}
\subsubsection{Uniform integrability}
\label{subsubsec:ui}
The following corollary implies that, for systems given by (\ref{rates}), $(t^{-1}|W_t|^p)_{t \geq 1}$
is uniformly integrable for any $p \geq 1 $, so that, whenever we have a LLN, the convergence holds
also in $L^p$.
\begin{corollary}
\label{cor:ui}
Let $\xi$ be a spin-flip system with rates as in \eqref{rates}, starting from equilibrium. Then
$(t^{-1}W_t)_{t \geq 1}$ is bounded in $L^p$ for all $p \ge 1$.
\end{corollary}
\begin{proof}
The claim for $Z$ under $\mathbb{P}_{\bar \mu}$ follows from Lemma~\ref{sbisf}(a)
by noting that $\bar \mu$ stochastically dominates $\nu_{\rho^-}$ on $\mathbb{Z}_{\le-1}$ and is dominated by $\nu_{\rho^+}$ on $\mathbb{Z}_{\ge 2}$; this can be verified noting that $W_0\ge0$ corresponds to finding particles to the left of $W_0$, and $W_0\le0$ to holes to its right. The same for $W$ follows from (\ref{assumpWZ}--\ref{assumpX0}) since $W_0$ has exponential moments under $\mathbb{P}_{\mu}$.
\end{proof}
We still need to verify (H4). This will be done in Sections~\ref{sec:mlessep} and
\ref{sec:subcritdep} below.
As $\kappa$ in \eqref{verifH31} could be taken independently of $L$ for (H3),
we only need $\lim_{L \to \infty}\Phi_L = 0$ in (H4).
\subsection{Example 1: $M < \epsilon$}
\label{sec:mlessep}
We recall the definition of $M$ and $\epsilon$ for a translation-invariant spin-flip system:
\begin{eqnarray}
\label{defm}
M &:=& \sum_{x \ne 0} \sup_{\eta} \left| c(\eta^x) - c(\eta) \right|,\\
\label{defep}
\epsilon &:=& \inf_{\eta} \left\{ c(\eta)+c(\eta^0) \right \},
\end{eqnarray}
where $\eta^x$ is the configuration obtained from $\eta$ by flipping the $x$-coordinate.
\subsubsection{Mixing for $\xi$}
\label{subsubsec:mixxi}
If $\xi$ is in the $M<\epsilon$ regime, then there is exponential decay of space-time correlations
(see Liggett~\cite{Li}, Section I.3). In fact, if $\xi$, $\xi'$ are two copies starting
from initial configurations $\eta$, $\eta'$ and coupled according to the Vasershtein coupling,
then, as was shown in Maes and Shlosman~\cite{MaSh}, the following estimate holds uniformly
in $x \in \mathbb{Z}$ and in the initial configurations:
\begin{equation}
\label{expest1}
\mathbb{P}_{\eta, \eta'}\left( \xi_t(x) \ne {\xi'}_t(x)\right)
\leq e^{-(\epsilon-M)t}.
\end{equation}
Since the system has uniformly bounded flip rates, it follows that there exist constants
$K_1, K_2 \in (0,\infty)$, independent of $x \in \mathbb{Z}$ and of the initial configurations, such that
\begin{equation}
\label{expest2}
\mathbb{P}_{\eta, \eta'}\left( \exists \; s> t \text{ s.t. } \xi_s(x)
\neq {\xi'}_s(x)\right) \le K_1 e^{-K_2t}.
\end{equation}
For $A \subset \mathbb{Z} \times \mathbb{R}_+$ measurable, let $\text{Discr}(A)$ be the event in
which there is a discrepancy between $\xi$ and $\xi'$ in $A$, i.e., $\text{Discr}(A)
:= \{\exists \; (x,t) \in A \colon\, \xi_t(x) \ne \xi'_t(x)\}$. Recall the definition of $C_R(m)$
in Section~\ref{subsec:notation}, and let $C_{R,t}(m) := C_R(m) \cap \mathbb{Z} \times [0,t]$. From (\ref{expest2}) we
deduce that, for any fixed $m>0$ and $R \in \mathbb{N}_0$, there exist (possibly different) constants
$K_1, K_2 \in (0,\infty)$ such that
\begin{equation}
\label{expestrestcone}
\mathbb{P}_{\eta,\eta'}(\text{Discr}(C_R(m) \setminus C_{R,t}(m))) \le K_1 e^{-K_2t}.
\end{equation}
\subsubsection{Mixing for $\Xi$}
\label{subsubsec:mixXi}
Bounds of the same type as \eqref{expest1}--\eqref{expestrestcone} hold for $\xi^{\pm}$,
since $M = 0$ and $\epsilon > 0$ for independent spin-flips. Therefore, in order to have
such bounds for the triple $\Xi$, we need only couple a pair $\Xi$, $\Xi'$ in such a
way that each coordinate is coupled with its primed counterpart by the Vasershtein coupling.
A set of coupling rates for $\Xi$, $\Xi'$ that accomplishes this goal is given
in \eqref{ratesbigcoupling}, in Appendix \ref{appendix}.
Redefining $\text{Discr}(A):= \{\exists \; (x,t) \in A \colon\, \Xi_t(x)
\neq \Xi'_t(x)\}$, by the previous results we see that (\ref{expestrestcone}) still
holds for this coupling, with possibly different constants. As a consequence, we get the following lemma.
\begin{lemma}
\label{conecont}
Define $d(\eta,\eta'):= \sum_{x \in \mathbb{Z}}1_{\{\eta(x) \neq \eta'(x)\}}2^{-|x|-1}$. For
any $m>0$ and $R \in \mathbb{N}_0$,
\begin{equation}
\label{eq:conecont}
\lim_{d(\Xi_0, \Xi'_0) \to 0}
\mathbb{P}_{\Xi_0,\Xi'_0}\big( \emph{Discr}(C_R(m)) \big) = 0.
\end{equation}
\end{lemma}
\begin{proof}
For any $t>0$, we may split $\text{Discr}(C_R(m)) = \text{Discr}(C_{R,t}(m)) \cup \text{Discr}
(C_R(m)\setminus C_{R,t}(m))$, so that
\begin{equation}
\label{estsplit}
\mathbb{P}_{\eta,\eta'}\big( \text{Discr}(C_R(m)) \big)
\leq \mathbb{P}_{\eta,\eta'}\big( \text{Discr}(C_{R,t}(m))
+ \mathbb{P}_{\eta,\eta'}\big( \text{Discr}(C_R(m) \setminus C_{R,t}(m)) \big).
\end{equation}
Fix $\varepsilon > 0$. By (\ref{expestrestcone}), for $t$ large enough the second term in
(\ref{estsplit}) is smaller than $\varepsilon$ uniformly in $\eta$, $\eta'$. For this
fixed $t$, the first term goes to zero as $d(\eta,\eta') \to 0$, since $C_{R,t}(m)$ is
contained in a finite space-time box and the coupling in (\ref{ratesbigcoupling}) is Feller
with uniformly bounded total flip rates per site. (Note that the metric $d$ generates the
product topology, under which the configuration space is compact.) Therefore $\limsup_{d(\eta,
\eta') \to 0} \mathbb{P}_{\eta, \eta '} \left( \text{Discr}(C_R(m)) \right) \leq
\varepsilon$. Since $\varepsilon$ is arbitrary, (\ref{eq:conecont}) follows.
\end{proof}
\subsubsection{Conditional mixing}
\label{subsubsec:condmix}
Next, we define an auxiliary process $\bar \Xi$ that, for each $L$, has the law of $\Xi$
conditioned on $\Gamma_L$ up to time $L$. We restrict to initial configurations $\eta \in
\bar E$. In this case, $\bar \Xi$ is a process on $\left(\{0,1\}^{\mathbb{Z} \setminus \{0,1\}}\right)^3$
with rates that are equal to those of $\Xi$, evaluated with a trap at the origin. More precisely,
for $\bar \eta \in \{0,1\}^{\mathbb{Z}\setminus\{0,1\}}$, denote by $(\bar \eta)_{1,0}$ the configuration
in $\{0,1\}^\mathbb{Z}$ that is equal to $\bar \eta$ in $\mathbb{Z}\setminus\{0,1\}$ and has a trap at the origin.
Then set $\bar C_x(\bar \eta) := C_x((\bar\eta)_{1,0})$, where $\bar C_x$ are the rates of
$\bar \Xi$ and $C_x$ the rates of $\Xi$ at a site $x \in \mathbb{Z}$. Observe that the latter depend
only on the middle configuration $\eta$, and not on $\eta^{\pm}$. These rates give the correct
law for $\bar \Xi$ because $\Xi$ conditioned on $\Gamma_L$ is Markovian up to time $L$. Indeed,
the probability of $\Gamma_L$ does not depend on $\eta$ (for $\eta \in \bar E$) and, for $s < L$,
$\Gamma_L = \Gamma_s \cap \theta_s \Gamma_{L-s}$. Thus, the rates follow by uniqueness. Observe
that they are no longer translation-invariant.
Two copies of the process $\bar \Xi$ can be coupled analogously to $\Xi$ by restricting the rates in (\ref{ratesbigcoupling}) to $\bar E$.
Since each coordinate of $\bar \Xi$ has similar properties as the
corresponding coordinate in $\Xi$ (i.e., $\bar \xi^{\pm}$ are independent spin-flip systems and
$\bar \xi$ is the in $M<\epsilon$ regime), it satisfies an estimate of the type
\begin{equation}
\label{expestauxproc}
\mathbb{\bar P}_{\eta, \eta'} \left( \text{Discr}([-t,t]\times \{t\}) \right)
\leq K_1 e^{-K_2t} \quad \forall \;\; \eta, \eta' \in \bar E,
\end{equation}
for appropriate constants $K_1,K_2 \in (0,\infty)$. From this estimate we see that $d(\bar \Xi_t,\bar \Xi'_t)
\to 0$ in probability as $t \to \infty$, uniformly in the initial configurations. By
Lemma~\ref{conecont}, this is also true for $\mathbb{P}_{(\bar \Xi_t)_{1,0}, (\bar \Xi '_t)_{1,0}}
(\text{Discr}(C_R(m)))$. Since the latter is bounded, the convergence holds in $L_1$ as well, uniformly
in $\eta, \eta'$.
\subsubsection{Proof of (H4)}
\label{proofH4ex1}
Let $f$ be a bounded function measurable in $\mathscr{C}_{R,\infty}(m)$ and estimate
\begin{align}
\label{h4mlessep}
&\left| \mathbb{E}_{\eta}\left[\theta_L f \mid \Gamma_L \right]
- \mathbb{E}_{\eta'}\left[\theta_L f \mid \Gamma_L \right]\right|
\leq 2 \|f\|_{\infty}\mathbb{P}_{\eta,\eta'}\big(\theta_L\text{Discr}(C_R(m))
\mid \Gamma_L \big) \nonumber \\
&\qquad \leq 2 \|f\|_{\infty} \sup_{\eta,\eta'}\mathbb{\bar E}_{\eta,\eta'}
\left[\mathbb{P}_{(\bar \Xi_L)_{1,0}, (\bar \Xi '_L)_{1,0}}
\left(\text{Discr}(C_R(m))\right)\right],
\end{align}
where $\mathbb{\bar E}$ denotes expectation under the (coupled) law of $\bar \Xi$. Therefore (H4)
follows with
\be{}
\Phi_L:= 2 \sup_{\eta,\eta'}\mathbb{\bar E}_{\eta,\eta'}
\left[\mathbb{P}_{(\bar \Xi_L)_{1,0}, (\bar \Xi '_L)_{1,0}}\left(\text{Discr}
(C_R(m))\right)\right],
\end{equation}
which converges to zero as $L \to \infty$ by the previous discussion.
This is enough since $\kappa_L$ could be taken constant in the verification of (H3)(1), as we saw in \eqref{verifH31}.
\subsection{Example 2: subcritical dependence spread}
\label{sec:subcritdep}
In this section, we suppose that the rates $c(\eta)$ have a finite range of dependence
$r \in \mathbb{N}_0$. In this case, the system can be constructed via a graphical representation as
follows.
\subsubsection{Graphical representation}
\label{subsubsec:graphrep}
For each $x \in \mathbb{Z}$, let $I^j_t(x)$ and $\Lambda^j_t(x)$ be independent Poisson processes
with rates $c_j$ and $\lambda_j$ respectively, where $j=0,1$. At each event of
$I^{j}_t(x)$, put a $j$-cross on the corresponding space-time point. At each event of
$\Lambda^j(x)$, put two $j$-arrows pointing at $x$, one from each side, extending over the
whole range of dependence. Start with an arbitrary initial configuration $\xi_0 \in \{0,1\}^{\mathbb{Z}}$.
Then obtain the subsequent states $\xi_t(x)$ from $\xi_0$ and the Poisson processes by, at
each $j$-cross, choosing the next state at site $x$ to be $j$ and, at at each $j$-arrow pair,
choosing the next state to be $j$ if an independent Bernoulli($p_j(\theta_x\xi_s)$) trial
succeeds, where $s$ is the time of the $j$-arrow event. This algorithm is well defined since,
because of the finite range, up to each fixed positive time it can a.s.\ be performed locally.
Any collection of processes with the same range and with rates of the form (\ref{rates}) with
$c_i$, $\lambda_i$ fixed ($i=0,1$) can be coupled together via this representation by fixing
additionally for each site $x$ a sequence $(U_n(x))_{n \in \mathbb{N}}$ of independent Uniform$[0,1]$
random variables to evaluate the Bernoulli trials at $j$-arrow events. In particular, $\xi^{\pm}$
can be coupled together with $\xi$ in the graphical representation by noting that, for $\xi^-$,
$p_0 \equiv 1$ and $p_1 \equiv 0$ and the opposite is true for $\xi^+$. For example, $\xi^-$
is the process obtained by ignoring all $1$-arrows and using all $0$-arrows. This gives the
same coupling as the one given by the rates (\ref{ratescoupl}). In particular, we see that
in this setting the events $\Gamma_L$ are given by (when $\xi_0 \in \bar E$)
\begin{equation}
\label{defgammalfiniterange}
\Gamma_L := \big\{I^0_L(0)=\Lambda^0_L(0)=I^1_L(1)=\Lambda^1_L(1)=0\big\}.
\end{equation}
\subsubsection{Coupling with a contact process}
\label{subsubsec:contproc}
We will couple $\Xi$ with a contact process $\zeta = (\zeta_t)_{t\geq 0}$ in the following way.
We keep all Poisson events and start with a configuration $\zeta_0 \in \{i,h\}^\mathbb{Z}$, where $i$
stands for ``infected" and $h$ for ``healthy". We then interpret every cross as a recovery,
and every arrow pair as infection transmission from any infected site within a neighborhood
of radius $r$ to the site the arrows point to. This gives rise to a `threshold contact process'
(TCP), i.e., a process with transitions at a site $x$ given by
\begin{equation}
\begin{array}{rcl}
i \rightarrow h & \text{ with rate } & c_0+c_1, \\
h \rightarrow i & \text{ with rate } & (\lambda_0 + \lambda_1)
1_{\{ \exists \text{ infected site within range $r$ of $x$} \}}.
\end{array}
\end{equation}
In the graphical representation for $\xi$, we can interpret crosses as moments of memory loss
and arrows as propagation of influence from the neighbors. Therefore, looking at the pair
$(\Xi_t(x),\zeta_t(x))$, we can interpret the second coordinate being healthy as the first
coordinate being independent of the initial configuration.
\begin{proposition}
\label{prop:couplingcp1}
Let $\underline i$ represent the configuration with all sites infected, and let $\Xi_0$,
$\Xi'_0 \in E^3$. Couple $\Xi$, $\Xi'$ and $\zeta$ by fixing a realization of
all crosses, arrows and uniform random variables, where $\Xi$ and $\Xi'$ are obtained from
the respective initial configurations and $\zeta$ is started from $\underline i$. Then a.s.\
$\Xi_t(x)=\Xi'_t(x)$ for all $t > 0$ and $x \in \mathbb{Z}$ such that $\zeta_t(x) = h$.
\end{proposition}
\begin{proof}
Fix $t>0$ and $x \in \mathbb{Z}$. With all Poisson and Uniform random variables fixed, an algorithm to find
the state at $(x,t)$, simultaneously for any collection of systems of type (\ref{rates}) with
fixed $c_i, \lambda_i$ and finite range $r$ from their respective initial configurations runs as follows. Find
the first Poisson event before $t$ at site $x$. If it is a $j$-cross, then the state is
$j$. If it is a $j$-arrow, then to decide the state we must evaluate $p_j$ and, therefore,
we must first take note of the states at this time at each site within range $r$ of $x$,
including $x$ itself. In order to do so, we restart the algorithm for each of these sites.
This process ends when time $0$ or a cross is reached along every possible path from $(x,t)$
to $\mathbb{Z}\times\{0\}$ that uses arrows (transversed in the direction opposite to which they
point) and vertical lines. In particular, if along each of these paths time $0$ is never
reached, then the state at $(x,t)$ does not change when we change the initial configuration.
On the other hand, time $0$ is not reached if and only if every path ends in a cross, which is
exactly the description of the event $\{\zeta_t(x)=h\}$.
\end{proof}
\subsubsection{Cone-mixing in the subcritical regime}
\label{subsubsec:subcritreg}
The process $(\zeta_t)_{t \geq 0}$ is stochastically dominated by a standard (linear) contact
process (LCP) with the same range and rates. Therefore, if the LCP is subcritical, i.e., if
$\lambda := (\lambda_0+\lambda_1)/(c_0+c_1) < \lambda_c$ where $\lambda_c$ is the critical
parameter for the corresponding LCP, then the TCP will die out as well.
Moreover, we have the following lemma:
\begin{lemma}
\label{lemma:subcritcp}
Let $A_t$ be the set of infected sites at time $t$. If $\lambda < \lambda_c$,
then there exist positive constants $K_1, K_2, K_3, K_4$ such that
\begin{equation}
\mathbb{P}_{\underline i}\big(\exists\,s>t \colon\, A_s \cap [-K_1 e^{K_2s}, K_1 e^{K_2s}] \neq \emptyset \big)
\leq K_3 e^{-K_4t}.
\end{equation}
\end{lemma}
\begin{proof} This is a straightforward consequence of Proposition 1.1 in Aizenman-Jung \cite{AiJu07}.
\end{proof}
According to Lemma~\ref{lemma:subcritcp}, the infection disappears exponentially fast around the
origin. For $r=1$, a proof can be found in Liggett~\cite{Li}, Chapter VI,
but it relies strongly on the nearest-neighbor nature of the interaction.
Let us now prove cone-mixing for $\xi$ when the rates are subcritical.
Pick a cone $C_t$ with any inclination and tip at time $t$,
and let $\mathcal{H}_t := \{\text{all sites inside } C_t \text{ are healthy}\}$. This event is independent of $\xi_0$
and, because of Lemma~\ref{lemma:subcritcp}, has large probability if $t$ is large.
Furthermore, by Proposition \ref{prop:couplingcp1}, on $\mathcal{H}_t$ the states of $\xi$ in $C_t$ are equal
to a random variable that is independent $\xi_0$, which implies the cone-mixing property.
\subsubsection{Proof of (H4)}
\label{subsubsec:proofH4ex2}
In order to prove the \emph{conditional} cone-mixing property, we couple the
conditioned process to a conditioned contact process as follows. First, let
\begin{equation}
\tilde \Gamma_L :=\left\{I_L^{j}(i)=\Lambda_L^{j}(i)=0 \colon\, j, i \in \{0,1\}\right\}.
\end{equation}
\begin{proposition}
\label{prop:couplingcp2}
Let $\hat i$ represent the configuration with all sites infected except for $\{0,1\}$, which are
healthy. Let $\Xi_0$, $\Xi'_0 \in \bar{E}^3$. Couple $\Xi$, $\Xi'$ conditioned on
$\Gamma_L$ and $\zeta$ conditioned on $\tilde \Gamma_L$ by fixing a realization of all crosses,
arrows and uniform random variables as in Proposition~\ref{prop:couplingcp1} and starting, respectively,
from $\Xi_0$, $\Xi'_0$ and $\hat i$, but, for $\Xi$ and $\Xi'$, remove the Poisson
events that characterize $\Gamma_L$ and, for $\zeta$, remove all Poisson events up to time $L$
at sites $0$ and $1$, which characterizes $\tilde \Gamma_L$. Then a.s.\ $\Xi_t(x)=\Xi'_t(x)$
for all $t > 0$ and $x \in \mathbb{Z}$ such that $\zeta_t(x) = h$.
\end{proposition}
\begin{proof}
On $\Gamma_L$, the states at sites $0$ and $1$ are fixed for time $[0,L]$. Therefore, in order
to determine the state at $(x,t)$, we need not extend paths that touch $\{0,1\}\times[0,L]$:
when every path from $(x,t)$ either ends in a cross or touches $\{0,1\}\times[0,L]$, the state
at $(x,t)$ does not change when the initial configuration is changed in $\mathbb{Z} \setminus \{0,1\}$.
But this is precisely the characterization of $\{\eta_t(x) = h\}$ on $\tilde \Gamma_L$ when
started from $\hat i$.
\end{proof}
The proof of (H4) is finished by noting that $(\eta_t)_{t\geq 0}$ starting from $\hat i$ and
conditioned on $\tilde \Gamma_L$ is stochastically dominated by $(\eta_t)_{t\ge0}$ starting
from $\underline i$. Therefore, by Lemma~\ref{lemma:subcritcp}, the ``dependence infection"
still dies out exponentially fast, and we conclude as for the unconditioned cone-mixing.
\subsection{The sign of the speed}
\label{sec:signspeed}
For independent spin flips, we are able to characterize with the help of a coupling argument
the regimes in which the speed is positive, zero or negative. By the stochastic domination
described in Section~\ref{sec:sfsbr}, this gives us a criterion for positive (or negative)
speed in the two classes addressed in Sections~\ref{sec:mlessep} and \ref{sec:subcritdep} above.
\subsubsection{Lipschitz property of the speed for independent spin-flip systems}
Let $\xi$ be an independent spin-flip system with rates $d_0$ and $d_1$ to flip to holes and
particles, respectively. Since it fits both classes of IPS considered in Sections \ref{sec:mlessep} and \ref{sec:subcritdep}, by Theorem~\ref{mainthm} there exists a $w(d_0, d_1) \in \mathbb{R}$
that is the a.s.\ speed of the $(\infty,0)$-walk in this environment. This speed has the following
local Lipschitz property.
\begin{lemma}
\label{signspeedisf}
Let $d_0, d_1, \delta > 0$. Then
\begin{equation}
w(d_0, d_1+\delta) - w(d_0, d_1 ) \geq \frac{d_0}{d_0 + d_1}\delta.
\end{equation}
\end{lemma}
\begin{proof}
Our proof strategy is based on the proof of Theorem 2.24, Chapter VI in \cite{Li}.
Construct $\xi$ from a graphical representation by taking,
for each site $x \in \mathbb{Z}$, two independent Poisson processes $N^i(x)$ with rates
$d_i$, $i=0,1$, with each event of $N^i$ representing a flip to state $i$. For a fixed $\delta >0$,
a second system $\xi^\delta$ with rates $d_0$ and $d_1 + \delta$ can be coupled to $\xi$ by starting
from a configuration $\xi_0^\delta \ge \xi_0$ and adding to each site $x$ an independent Poisson process
$N^\delta(x)$ with rate $\delta$, whose events also represent flips to particles, but only for
$\xi^\delta$. Let us denote by $W$ and $W^\delta$ the walks in these respective environments.
Under this coupling, $\xi \leq \xi^\delta$, so, by monotonicity, $W_t \le W_t^{\delta}$
for all $t \ge 0$ as well. We aim to prove that
\begin{equation}
\label{lips1}
\mathbb{E}_{\mu^{\delta}}\left[W^\delta_t\right]
-\mathbb{E}_{\mu}\left[W_t\right] \geq \frac{d_0}{d_0+d_1}\delta t,
\end{equation}
where $\mu$ and $\mu^{\delta}$ are the equilibria of the respective systems. From this the
conclusion will follow after dividing by $t$ and letting $t \to \infty$.
Define a third walk $W^*$ that is allowed to use one and only one event
of $N^\delta$. More precisely, let $S$ be the first time when there is an event of $N^\delta$ at $W_S+1$.
Take $W^*$ equal to $W$ on $[0,S)$ and, for times $\ge S$, let $W^*$ evolve by the same rules as $W$ but adding a particle at $W_S+1$ at time $S$, and using no more $N^\delta$ events. By construction, we have $W_t \le W_t^* \leq W_t^\delta$ $\forall$ $t \ge 0$.
Let $\eta_1 := \theta_{W_S}\xi_S \in \bar E$ and $\eta_2 := (\eta_1)^1$ be the configurations
around $W_S$ and $W^*_{S-}$, respectively. Then
\begin{align}
\label{lips2}
\mathbb{E}_{\mu^{\delta}}\left[W^\delta_t\right] -\mathbb{E}_{\mu}\left[W_t\right]
&\geq \mathbb{E}_{\mu}\left[W_t^*-W_t, S \le t\right]
\geq \mathbb{E}_{\mu}\left[W_t^*-W_t, S \le t, \eta_1(2) =0\right] \nonumber \\
&= \mathbb{E}_{\mu}\left[\mathbb{E}_{\eta_1,\eta_2}
\left[ W^2_{t-S}-W^1_{t-S}\right], \eta_1(2)=0, S \le t \right],
\end{align}
where $W^i$, $i=1,2$ are copies of $W$ starting from $\eta_i$ and coupled via the graphical
representation. We claim that, if $\eta_1(2) = 0$,
\begin{equation}\label{lips7}
\mathbb{E}_{\eta_1,\eta_2}\left[W^2_s - W^1_s \right] \ge 1 \quad \forall \; s \ge 0.
\end{equation}
Indeed, we will argue that the difference $W^2_s-W^1_s$ can only decrease
when we flip all states of $\eta_1,\eta_2$ on $\mathbb{Z}_{\le -1}$ to particles
and on $\mathbb{Z}_{\ge 2}$ to holes; but after
doing these operations, we find that $W^2$ has the same distribution as $W^1+1$, which gives \eqref{lips7}.
It is enough to consider a single $x > 2$. Let $\tau := \inf\{t > 0 \colon\, N^0_t(x) + N^1_t(x) > 0 \} \wedge s$,
and put $T := \inf\{t > 0 \colon\, W^1_t = x-1\}$. There are two cases: either $T > \tau$ or not.
In the first case, $W^1_s$ remains constant if we set $\eta_{1,2}(x)=0$, while $W^2_s$ does not increase.
In the second case, if $\eta_{1,2}(x)=0$, then $W^1_{T} = W^2_{T}$;
but then they must remain equal thereafter since, for them to meet, the state at site $1$ must have flipped,
and therefore they see the same configuration in the environment at time $T$.
Hence, in this case, $W^2_s-W^1_s=0$ which is the minimum value, and our claim follows.
From \eqref{lips2} and \eqref{lips7} we get
\begin{equation}
\label{lips3}
\mathbb{E}_{\mu^{\delta}}\left[W^\delta_t\right] -\mathbb{E}_{\mu}\left[W_t\right]
\geq \mathbb{P}_{\mu}\left(\eta_1(2)=0, S \le t \right).
\end{equation}
Consider the event $\{\eta_1(2)=0\} $. There are two possible situations: either at time $S$
the site $W_S +2$ was not yet visited by $W$, in which case $\eta_1(2)$ is still in equilibrium,
or it was. In the latter case, let $s$
be the time of the last visit to this site before $S$. By geometrical constraints, at time
$s$ only a hole could have been observed at this site, so the probability that its state at
time $S$ is a hole is larger than at equilibrium, which is $d_0/(d_0+d_1)$. In other words,
\begin{equation}
\label{lips4}
\mathbb{P}_{\mu}\left(\eta_1(2)=0 \mid S, W_{[0,S]}\right) \geq \frac{d_0}{d_0+d_1},
\end{equation}
which, together with (\ref{lips3}) and the fact that $S$ has distribution Exp($\delta$),
gives us
\begin{equation}
\label{lips5}
\mathbb{E}_{\mu^{\delta}}\left[W^\delta_t\right] -\mathbb{E}_{\mu}\left[W_t\right]
\geq \frac{d_0}{d_0+d_1}\left(1-e^{\delta t}\right).
\end{equation}
Since $\delta$ is arbitrary, we may repeat the argument for systems with rates $d_1+ (k/n)
\delta$, $n \in \mathbb{N}$ and $k =0,1,\ldots,n$, to obtain
\begin{align}
\label{lips5*}
\mathbb{E}_{\mu^{\delta}}\left[W^\delta_t\right] -\mathbb{E}_{\mu}\left[W_t\right]
& \geq \frac{d_0}{d_0+d_1}n\left(1-e^{\delta t /n}\right),
\end{align}
and we get (\ref{lips1}) by letting $n \to \infty$.
\end{proof}
\subsubsection{Sign of the speed}
\label{subsubsec:signspeed}
If $d_0 = d_1$, then $w = 0$, since by symmetry $W_t=-W_t$ in distribution. Hence we can summarize:
\begin{corollary}
\label{summarysignisf}
For an independent spin-flip system with rates $d_0$ and $d_1$,
\begin{equation}
\label{eqsumsignisf}
\begin{array}{ll}
w \ge \frac{d_0}{d_0+d_1}\left(d_1 - d_0\right) & \text{ if } d_1 > d_0,\\[0.2cm]
w = 0 & \text{ if } d_1 = d_0 ,\\[0.2cm]
w \le -\frac{d_1}{d_0+d_1}\left(d_0 - d_1\right) & \text{ if } d_1 < d_0.
\end{array}
\end{equation}
\end{corollary}
\noindent
Applying this result to the systems $\xi^{\pm}$ of Section~\ref{sec:sfsbr}, we obtain the following.
\begin{proposition}\label{critsignspeed}
Let $W$ be the random walk for the $(\infty,0)$-model in a spin-flip system
with rates given by \eqref{rates}. Then, $\mathbb{P}_{\mu}$-a.s.,
\begin{equation}
\begin{array}{ll}
\liminf_{t \to \infty} t^{-1}W_t \geq \frac{c_0+\lambda_0}{c_1+c_0+\lambda_0}
\left(c_1 - c_0 - \lambda_0 \right) & \text{ if } c_1 \ge c_0 + \lambda_0, \\[0.2cm]
\limsup_{t \to \infty} t^{-1}W_t \leq -\frac{c_1+\lambda_1}{c_0+c_1+\lambda_1}\left(c_0 - c_1 - \lambda_1 \right)
& \text{ if } c_0 \ge c_1 + \lambda_1.
\end{array}
\end{equation}
\end{proposition}
This concludes the proof of Theorem~\ref{thm:cond} and the discussion of our two classes
of IPS's for the $(\infty,0)$-model. In Section~\ref{sec:oex} we give additional examples
and discuss some limitations of our setting.
\section{Other examples}
\label{sec:oex}
We describe here three types of examples that satisfy our hypotheses:
generalizations of the $(\alpha,\beta)$-model and of the $(\infty,0)$-model, and mixed models.
We also discuss an example that is beyond the reach of our setting.
\medskip\noindent
{\bf 1.\ Internal noise models.}
For $x \in \mathbb{Z}\setminus\{0\}$ and $\eta \in E$, let $\pi_x(\eta)$ be functions with a finite range
of dependence $R$. These are the rates to jump $x$ from the position of the walk. Let $\pi_x
:= \sup_{\eta}\pi_x(\eta)$ and suppose that, for some $u>0$,
\begin{equation}
\label{expmomboundrates}
\sum_{x \in \mathbb{Z}\setminus\{0\}}e^{u|x|}\pi_x < \infty.
\end{equation}
This implies that also
\begin{equation}
\label{deftotrateinm}
\Pi := \sum_{x \in \mathbb{Z}\setminus\{0\}}\pi_x < \infty.
\end{equation}
The walk starts at the origin, and waits an independent Exponential($\Pi$) time $\tau$ until it
jumps to $x$ with probability $\pi_x(\xi_\tau)/\Pi$. These probabilities do not necessarily sum
up to one, so the walk may well stay at the origin. The subsequent jumps are obtained analogously,
with $\xi_\tau$ substituted by the environment around the walk at the time of the attempted jump.
It is clear that (A1)--(A3) hold. The walk has a bounded probability of standing still independently
of the environment, and its jumps have an exponential tail. We take
\begin{equation}
\label{defgammainm}
\Gamma_L := \{\tau > L\}.
\end{equation}
By defining an auxiliary walk $(H_t)_{t \geq 0}$ that also tries to jump at time $\tau$, but only
to sites $x>0$ with probability $\pi_x/\Pi$, we see that $W_t \leq H_t$ and that $H_t$ has
properties analogous to the process defined in the proof of Lemma~\ref{sbisf}. Therefore,
(H1)--(H3) are always satisfied for this model. Since $\Gamma_L$ is independent of $\xi$,
(H4) is the (unconditional) cone-mixing property. Observe that $W_0=0$, so that $\bar \mu = \mu$.
Therefore the LLN for this model holds in both examples discussed in Section \ref{sec:proofthm2}, and
also for the IPS's for which cone-mixing was shown in Avena, den Hollander and Redig~\cite{AvdHoRe1}.
The $(\alpha,\beta)$-model is an internal noise model with $R=0$ (the rates depend only on the
state of the site where the walker is) and $\pi_x(\eta)=0$, except for $x = \pm1$, for which
$\pi_{1}(1) = \alpha = \pi_{-1}(0)$ and $\pi_1(0)= \beta = \pi_{-1}(1)$.
\medskip\noindent
{\bf 2.\ Pattern models.}
Take $\aleph$ to be a finite sequence of $0$'s and $1$'s, which we call a \emph{pattern}, and
let $R$ be the length of this sequence. Take the environment $\xi$ to be of the same type used
to define the $(\infty,0)$-walk. Let $q:\{0,1\}^R\setminus\{\aleph\} \to [0,1]$. The pattern
walk is defined similarly as the $(\infty,0)$-walk, with the trap being substituted by the
pattern, and a Bernoulli($q$) random variable being used to decide whether the walk jumps to
the right or to the left. More precisely, let $\vartheta = (\xi_0(0),\ldots,\xi_0(R-1))$.
If $\vartheta = \aleph$, then we set $W_0 = 0$, otherwise we sample $b_0$ as an independent Bernoulli($q(\vartheta$)) trial.
If $b_0=1$, then $W_0$ is set to be the starting position
of the first occurrence of $\aleph$ in $\xi_0$ to the right of the origin, while if $b_0=0$,
then the first occurrence of $\aleph$ to the left of the origin is taken instead. Then the
walk waits at this position until the configuration of one of the $R$ states to its right
changes, at which time the procedure to find the jump is repeated with the environment as
seen from $W_0$. Subsequent jumps are obtained analogously. The $(\infty,0)$-model is a
pattern model with $\aleph := (1,0)$, $q(1,1) := 1$, $q(0,0):=0$ and $q(0,1):=1/2$.
For spin-flip systems given by (\ref{rates}), the pattern walk is defined and finite
for all times, no matter what $\aleph$ is, the reasoning being exactly the same as for the
$(\infty,0)$-walk. Also, it may be analogously defined so as to satisfy assumptions (A1)--(A3). Defining the events $\Gamma_L$ as
\begin{equation}
\label{deceventspattern}
\Gamma_L := \big\{\xi_s^{\pm}(j)=\xi_0^{\pm}(j)
\; \forall \; s \in [0,L] \text{ and } j \in \{0,\ldots,R-1\}\big\},
\end{equation}
we may indeed, by completely analogous arguments, reobtain all the results of
Section~\ref{sec:proofthm2}, so that hypotheses (H1)--(H4) hold and, therefore,
the LLN as well.
\medskip\noindent
{\bf 3. Pattern models with extra jumps.}
Examples of models that fall into our setting and for which the events $\Gamma_L$ depend
non-trivially both on $\xi$ and $Y$ can be constructed by taking a pattern model and adding noise
in the form of non-zero jump rates while sitting on the pattern.
More precisely, add to $Y$ an independent Poisson process $N$ with positive rate
and let $W$ jump also at events of $N$ but with the same jump mechanism, i.e.,
choosing the sign of the jump according to the result of a Bernoulli($q$) random variable, and the displacement using the pattern.
Taking $\Gamma_L := \Gamma^{\aleph}_L \cap \{N_L=0\}$, where $\Gamma^{\aleph}_L$ is the corresponding event for the pattern model,
we may check that, for the two examples of dynamic random environments
considered in Theorem~\ref{thm:cond}, (A1)--(A3) and (H1)--(H4) are all verified.
\medskip\noindent
{\bf 4. Mixtures of pattern and internal noise models.}
Another class of models with nontrivial dependence structure for the renegeration-inducing events can be constructed as follows.
Let $W^0$ be an internal noise model and $W^1$ a pattern model (with or without extra jumps) on the the same random environment $\xi$
and let $Y^i$, $i \in \{0,1\}$, be the corresponding random elements associated to each model.
Let $X = (X)_{n \in \mathbb{N}}$ be a sequence of i.i.d.\ Bernoulli($p$) random variables independent of all the rest, where $p \in (0,1)$.
Then the mixture is the model for which the dynamics associated to $i \in \{0,1\}$ are applied in the time interval $[n-1,n)$ when
$X_n = i$. Note that this model will have deterministic jumps.
Letting $Y := \left(Y^0,Y^1, X \right)$ where $Y^i$ is the corresponding
random element associated to the model $i$,
it is easily checked that this model indeed falls into our setting.
Choosing
\begin{equation}
\label{defgammamix}
\Gamma_L := \Gamma_L^1 \cap \{X_k = 1, k = 1,\ldots,L\}
\end{equation}
where $\Gamma^1_L$ is the corresponding event for the pattern model,
it is not hard to verify, using the results of Section~\ref{sec:proofthm2}, that, for the two classes of random environments considered in
Theorem~\ref{thm:cond}, the mixed model satisfies (A1)--(A3) and (H1)--(H4).
\medskip\noindent
{\bf An open example.}
We will close with an example of a model that does not satisfy the hypotheses of our
LLN (in dynamic random environments given by spin-flip systems). When $\xi(0)=j$, let
$C^j$ be the cluster of $j$'s around the origin. Define jump rates for the walk as follows:
\begin{equation}
\label{ratescounterex}
\begin{array}{lcl}
\pi_1(\eta) & = & \left\{\begin{array}{lll}
|C^1| & \text{ if } & \eta(0)=1,\\
|C^0|^{-1} & \text{ if } & \eta(0)=0,
\end{array}\right.\\
\pi_{-1}(\eta) & = & \left\{ \begin{array}{lll}
|C^0| & \text{ if } & \eta(0)=0,\\
|C^1|^{-1} & \text{ if } & \eta(0)=1.
\end{array} \right.
\end{array}
\end{equation}
Even though this looks like a fairly natural model,
it does not satisfy (A2). It also won't satisfy (H1) and (H2) together for any reasonable random environment,
which is actually the hardest obstacle.
The problem is that, while we are able to transport
a.s.\ properties of the equilibrium measure to the measure of the environment as seen from the walk,
we cannot control the distortion in events
of positive measure. Thus, even if $\Gamma_L$ has positive probability at time zero, there
is no a priori guarantee that it will have an appreciable probability
from the point of view of the walk at later times.
Because of this, we cannot implement our regeneration strategy, and our proof of the LLN breaks down.
\newpage
|
\section{Introduction}
\label{sec:intro}
Cosmic accelerating expansion is the most tantalizing problem in modern
cosmology and physics. Within the framework of Einstein's general
relativity (GR), the cosmic acceleration requires that
roughly 70\% of total energy of the present-day universe is in the form
of unknown, mysterious energy component having negative pressure, dubbed
as dark energy. An alternative explanation is the so-called
modified gravity scenario, where the cosmic acceleration is conjectured
as a result of
breakdown of Einstein's gravity on cosmological scales. There are
growing attempts in the community trying to develop a consistent model
of modified gravity that can explain the cosmic acceleration on
cosmological scales, yet recovering GR on small scales
such as solar system scales, without the need of dark energy \citep[e.g. see][for a
review]{JainKhoury:10}.
There are various methods capable of addressing the nature of the cosmic
acceleration: type-Ia supernovae, cluster experiments,
galaxy clustering, and weak gravitational lensing. These methods are
sensitive to cosmic expansion and structure formation histories, in
a complementary way, over different length scales and/or different ranges
of redshifts. In particular, an essential approach to discriminate the
dark energy and modified gravity scenarios is exploring both the cosmic
expansion history and the growth rate of structure formation by
combining more than two different methods above
\citep{DETF:06,EsoFundamental:06,JainZhang:08,Guziketal:10}.
In this paper we focus on cosmological observables derivable from a
wide-field galaxy redshift survey. A robust method feasible with a
galaxy redshift survey is the baryon acoustic oscillation (BAO)
experiment, which allows us to infer the angular diameter distance as
well as the Hubble expansion rate from the measured pattern of galaxy
clustering \citep{Eisenstein:05,Cole:05,BlakeetalBAO:11}.
There are many ongoing and planned galaxy redshift surveys aimed at
achieving the BAO experiments at higher precisions: the Baryon
Oscillation Spectroscopic Survey
(BOSS)\footnote{http://cosmology.lbl.gov/BOSS/}, the BigBOSS project
\citep{BigBoSS}, the HETDEX survey\footnote{http://hetdex.org/}, and the
Subaru Prime Focus Spectrograph (PFS)
project\footnote{http://sumire.ipmu.jp/en/}.
Adding the redshift distortion measurement
can further improve the cosmological power of a galaxy
redshift survey
\citep{Peacocketal:01,Guzzoetal:08,Whiteetal:10,Yamamotoetal:10,Blakeetal:11,songsabiu:11}.
In real space
galaxy clustering is statistically isotropic in a statistically
homogeneous and isotropic universe. However, in redshift space the
line-of-sight component of galaxies' peculiar velocities induces an
angular anisotropic modulation in the clustering pattern. In a
structure formation scenario the peculiar velocities of galaxies are
caused by gravitational attracting force in large-scale structure and
the gravitational field can be inferred from the observed galaxy
distribution or directly probed by weak gravitational lensing.
More precisely there are two kinds of redshift distortion effects.
One is caused by large-scale coherent velocities or bulk motions of
halos which are associated with large-scale structure of large length
scales, $\lower.5ex\hbox{$\; \buildrel > \over \sim \;$} 1~{\rm Mpc}$. This large-scale redshift distortion in
the linear regime is
called Kaiser effect \citep{Kaiser:87}. This effect amplifies clustering
amplitudes of galaxies in redshift space. It is now becoming recognized
that, even at length scales of $100~h^{-1}{\rm Mpc}$ relevant for
BAO experiments, the Kaiser effect ceases to be
accurate, and nonlinearity effects needs to be included for a level of
precision ongoing/upcoming survey can achieve. Encouragingly, however,
the refined, accurate modeling has been developed based on perturbation
theory and/or simulations \citep[][and see references
therein]{Scoccimarro04,Matsubara:08,Taruyaetal:09,Taruyaetal:10}. In
this sense the Kaiser effect contains a cleaner cosmological
information. Hence, if Einstein GR is a priori assumed, adding the
large-scale velocity information to the BAO constraints or more
generally the density clustering information allows us to
significantly improve geometrical constraints
\citep[e.g.][]{AP,MatsubaraSuto:96,Ballingeretal:96} as well as
cosmological parameter estimation
\citep[e.g.][]{Eisensteinetal:99,Takadaetal:06,Takada:06,Saitoetal:08}.
Probably more interestingly, if the density and velocity power spectra
can be reconstructed from the measured redshift-space clustering of
galaxies without assuming any gravity theory, we can now open up a
window of exploring properties of gravity on cosmological scales in a
model-independent way, by comparing the reconstructed density and
velocity power spectra, because the density and velocity fields are
related via gravity theory
\citep{Linder:05,Zhangetal:07,Zhangetal:08,Guzzoetal:08,Wang:08,Yamamotoetal:08,
Whiteetal:09,PercivalWhite:09,
SimpsonPeacock:10,Song:10,SongKayo:10,Yamamotoetal:10,Reyesetal:10,Shapiroetal:10}. For
example, Einstein gravity or a concordance $\Lambda$CDM model gives us
specific predictions on how these two spectra are related to each other:
the two spectra have a constant overall offset in the amplitudes in the
linear regime. Hence, if any scale-dependent differences in the
amplitudes are found from data, it is a signature of failure of Einstein
gravity.
However, a viable reconstruction method needs to be not much influenced
by uncertainties arising from small-scale, nonlinear redshift distortion
effect due to
internal virial motions
of galaxies within halos, the so-called Fingers-of-God (FoG) effect
\citep[e.g. see][]{Jackson:72,Peacock:book99,Hamilton:98,Scoccimarro04}.
This effect causes a significant suppression
in redshift-space clustering amplitudes along the line-of-sight
direction.
How does the small-scale velocity field affect the BAO-scale clustering
of galaxies? Here is a rough estimate on the physics. Recall that virial
velocity dispersion for massive halos of $10^{15}M_\odot$ can have
velocities of a few $ 10^3$~km s$^{-1}$. This causes a redshift
modulation given as $\Delta z\simeq v_{\parallel} \simeq 10^{-2}$ (in
units of speed-of-light $c=1$), which in turn causes an apparent
displacement in the position space as $\Delta r_\parallel=\Delta
z/H(z)\simeq 30~h^{-1}{\rm Mpc}$. This corresponds to Fourier modes of
$k=2\pi/\lambda\simeq 0.2~h{\rm Mpc}^{-1}$, which are indeed relevant
for the BAO scales. Thus the real-space galaxy clustering within halos
at scales smaller than a few Mpc blows up to large scales up to $\sim
$50Mpc in redshift space. Since the FoG effect arises from highly
nonlinear regime and is affected by baryonic and astrophysical effects,
it is still very challenging to have a sufficiently accurate model
needed for precision cosmology. In fact, the FoG effect is one of the
major systematic errors in galaxy clustering observables.
Hence the purpose of this paper is to develop a method that allows us to
unbiasedly reconstruct the real-space density and velocity power spectra of
large length scales from the measured redshift-space clustering of
galaxies, removing the contamination of FoG effect \citep[also see][for a similar study]{SongKayo:10}. This can be done by
developing a maximum likelihood based method of reconstructing band
powers of the real-space power spectra
at each wavenumber bins, including marginalization over
uncertainties in parameters to model the FoG effect. In this method the
real-space power spectra at each wavenumber bins are estimated such that
the likelihood of redshift-space power spectrum is maximized, assuming
that the original density perturbation field is a Gaussian field. The
{\em real-space} velocity power spectra on large scales include only the
information on the large-scale redshift distortion effect, because the
spectra arise from the density and velocity fields at physical scales
corresponding to the wavenumbers without the FoG effect contamination. Our
method is analogous to the cosmic microwave background (CMB) power
spectrum reconstruction \citep{Verde03}.
By using N-body simulations of 70 realizations and the halo
catalogs, we will carefully test the method by studying whether or not
the reconstructed real-space power spectra can recover the input spectra
in the simulations.
The structure of this paper is as follows. In
\S~\ref{sec:redshift_distortion} we review how the redshift distortion
effect due to peculiar velocities causes an angular modulation in
redshift-space power spectrum, after briefly describing how the peculiar
velocity field is related to metric scalar perturbations. In
\S~\ref{sec:method} we develop a maximum likelihood method of
reconstructing the real-space density and velocity power spectra from
the redshift-space power spectrum. After describing the N-body
simulations and halo catalogs in \S~\ref{sec:dm_catalog}, we will show
in \S~\ref{sec:results} the main results of this paper; by applying the
method to the N-body simulations and halo catalogs, we assess accuracies
of reconstructing the real-space power spectra with the
method. \S~\ref{sec:summary} is devoted to summary and discussion.
\section{Preliminaries}\label{sec:redshift_distortion}
\subsection{Metric perturbations}
\label{sec:metric}
In the Newtonian gauge the perturbed Friedmann-Robertson-Walker metric
that has scalar perturbations can be fully specified by the form of
\begin{equation}
ds^2=-(1+2\Psi)dt^2+a^2(1-2\Phi)d\bmf{x}^2,
\label{eq:metric}
\end{equation}
where $a(t)$ is the expansion scale factor. Note that we here assumed a
flat universe for simplicity. The metric form (\ref{eq:metric}) is
fully general for any metric theory of gravity, as long as the vector
and tensor perturbations are negligible. Redshift $z$ is the most
important observable in astronomy, and it is given as $1+z=1/a(t_e)$,
where $a(t_e)$ is the scale factor at the epoch when an object of
interest, e.g. galaxy, emitted the photon to be observed by an
observer. We use the convention $a(t_0)=1$ at present. $\Psi$
corresponds to the Newtonian potential that describes the acceleration
of particles, while $\Phi$ denotes the curvature perturbation.
The expansion history of the universe is specified by the function of
$a(t)$ or the Hubble function $H(t)=\dot{a}/a$, where $\dot{}$ denotes
the derivative with respect to time $t$. Given gravity theory, the time
evolution of $a(t)$ or $H(t)$ is specified once the energy content of
the universe is specified, as in the case of Einstein gravity.
\subsection{The case of Einstein gravity}
Although the rest of this paper does not assume any theory of gravity,
it would be instructive to discuss the case of Einstein gravity. This
subsection also gives a background motivation of our work.
Theory of gravity relates the metric perturbations in
Eq.~(\ref{eq:metric}) to matter variables. In the matter dominated
era, if assuming the Einstein gravity, the Einstein equations yield, for
example, the Poisson equation on sub-horizon scales, which relates the
metric perturbation $\Phi$ to the density perturbation field of total
matter as
\begin{equation}
-k^2\Phi=4\pi G a^2\delta.
\end{equation}
Note that the Poisson equation here is given in the Fourier space,
yielding the factor $k^2$ on the l.h.s.
The matter distribution can be inferred from galaxy surveys or weak
lensing surveys.
In a case that the anisotropic energy stress is negligible as in a
cold dark matter (CDM) dominated structure formation
model, the two metric perturbations are equivalent to each other
on sub-horizon
scales:
\begin{equation}
\Psi\simeq\Phi.
\end{equation}
Thus the two metric perturbations have only one degree of freedom, which
corresponds to the density field $\delta$ in the matter sector.
The geodesic equation for a test particle is given by
\begin{equation}
\frac{dp^\alpha}{d\lambda}+\Gamma^{\alpha}_{\mu\nu}p^\mu p^\nu=0,
\end{equation}
where $\Gamma$ is the Christophel symbols. Let's consider a test
particle which only slowly move with respect to the comoving coordinates
i.e. a non-relativistic particle. Dark matter and galaxies are such
particles. The equation of motion for
such a test particle is given in the linear regime as
\begin{equation}
\frac{d\bmf{v}}{dt}+H\bmf{v}=-\frac{1}{a}\nabla\Psi,
\end{equation}
where $\bmf{v}$ is the comoving peculiar velocity defined as
$\bmf{v}\equiv d\bmf{x}/dt$.
Thus the velocity field follows the gravitational potential. For this
reason the peculiar velocity field of galaxies is expected as a powerful
tool for probing the gravitational potential field.
Another important observable is gravitational lensing. Solving the
geodesic equation for a photon, which is a relativistic particle, leads
the lensing deflection angle to be given as
\begin{equation}
\bmf{\alpha}=\int\!d\chi~W_{\rm GL}(\chi)\nabla_\perp
\left(\Psi+\Phi\right),
\end{equation}
where $W_{\rm GL}(\chi)$ is the lensing geometrical kernel that depends
on the background metric quantity, i.e the scale factor
\cite[e.g.][]{Guziketal:10}.
Thus lensing depends on a combination of the two metric perturbations,
$\Psi+\Phi$.
Therefore combining different observables such as the galaxy
distribution, the peculiar velocity and weak lensing in principle allow
to test the consistency relation $\Psi=\Phi$ or more generally explore
properties of gravity on cosmological scales
\citep[e.g.][]{JainZhang:08}.
However in this paper we address
we can use the measured clustering features of galaxies in redshift space
to reconstruct the power spectrum of the peculiar velocity field
$\bmf{v}$, independently from the density power spectrum. Hence, our
method allows us to use the redshift-space clustering to test gravity
theory on cosmological scales, by comparing the reconstructed power
spectra of density and velocity fields.
\subsection{Redshift-space power spectrum}\label{sec:FOG_theory}
What we can measure from a spectroscopic survey of galaxies is angular
positions and redshifts of the galaxies. However, the observed redshift
of a given galaxy, $\hat{z}$, is modulated from the true redshift, $z$,
due to its peculiar velocity as well as the metric perturbations
-- the so-called redshift-space
distortion. According to the metric theory of gravity, the observed
redshift is given \citep[e.g., see][]{Sasaki:87} as
\begin{eqnarray}
1+\hat{z}&\simeq& (1+z)\left\{
1+
\left[\Phi+
v_z\right]^{\rm e}_{\rm o}
\right\}\nonumber\\
&\simeq &(1+z)\left[1+\left.v_\chi\right|^{\rm e}_{\rm o}\right],
\end{eqnarray}
where $\Phi$ is
the gravitational potential perturbation (Eq.~[\ref{eq:metric}]), $v_z$
denotes the line-of-sight component of the comoving peculiar velocity of
tracer considered, and the notation $\left.\cdots\right|^{\rm e}_{\rm
o}$ denotes the difference between quantities at observer's and galaxy's
positions. In the second line on the r.h.s. of the equation above, we
assumed that the effect of peculiar velocity $v_\chi$, which is the
order of $10^{-3}$ (corresponding to $300$km s$^{-1}$) for the
large-scale coherent peculiar velocities, is much larger than the
potential amplitude $\Phi\sim O(10^{-5})$ for $\Lambda$CDM-like
cosmologies. In other words, while we will later focus on the density
perturbations of matter or galaxies in large-scale structure, the effect
of the metric perturbation $\Phi$ is safely negligible compared to the
density perturbations on relevant length scales. In addition the
perturbation contributions at an observer's position ($O$) only
contribute to the monopole offset (e.g. a shift in the overall
normalization of galaxy number density at a given redshift), therefore
we can ignore these contributions in the following.
Via the redshift-distance relation $\chi(z)$, the apparent radial
distance to a galaxy at redshift $z$, $\hat{\chi}(z)$, is modulated from
the true position $\chi(z)$ due to the peculiar velocity as
\begin{eqnarray}
\hat{\chi}&\equiv& \chi(\hat{z})\nonumber\\
&\simeq& \chi(z)+(1+z)\frac{d\chi}{dz}v_\chi\nonumber\\
&=& \chi(z)+\frac{(1+z)}{H(z)}v_\chi\nonumber\\
&=& \chi(z)+u_\chi,
\label{eq:u_chi}
\end{eqnarray}
where $u_\chi$ is the normalized peculiar velocity field defined as
$u_\chi\equiv (1+z)v_\chi/H(z)$.
The mass conservation, or the number conservation of galaxies, tells
that the density perturbation in redshift space, $\delta_s$, is related
to the real-space density perturbation as
\begin{eqnarray}
1+\delta_s&=&(1+\delta)
\left(1+\frac{\partial u_\chi}{\partial \chi}\right)^{-1}\nonumber\\
&\simeq& 1+\delta - \frac{\partial u_\chi}{\partial \chi}+
O\!\left(\delta u, u^2\right),
\label{eq:delta_s}
\end{eqnarray}
where in the second equality of the equation above we have used the
Taylor expansion of $(1+\partial u/\partial \chi)^{-1}$, and we have
ignored the higher-order terms of the perturbations (see below for
further discussion).
Exactly
speaking the Jacobian transformation above breaks down when the particle
motions have shell crossing or multi-streamings at a single spatial
position, which can occur in the nonlinear stage such as a region within
a virialized halo. In other words, the equation above is valid only at
large length scales greater than a size of halos, which is validated on
scales $k\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 0.3~h^{-1}{\rm Mpc}$ we are interested in.
Fourier-transforming the equation above yields
\begin{equation}
\tilde{\delta}_s(\bmf{k})
\simeq \tilde{\delta}(\bmf{k}) + \mu^2 \tilde{\theta}(\bmf{k})
+O(\delta \theta, \theta^2),
\label{eq:delta_s_app}
\end{equation}
where $\mu$ is the cosine between the wavevector
$\bmf{k}$ and the line-of-sight direction.
Here we have assumed that the peculiar velocity is irrotational,
therefore is given in terms of the scalar velocity potential, and the
quantity, $\tilde{\theta}$, denotes the Fourier-transformed coefficient
of the divergence of peculiar velocity field, $\theta\equiv
-\nabla\cdot\bmf{u}$.
Also notice that in the equation above we have employed a distant
observer approximation and ignored the curvature of the sky, where one
axis of the coordinate system can be chosen to be along the
line-of-sight direction.
We again ignored the higher-order terms of the
perturbations such
as $O(\delta \theta, \theta^2)$.
Thus the redshift distortion induces angle-dependent
modulations, given by $\mu^{2n}$ ($n=1,2,\cdots$), in the redshift-space density field.
Motivated by the discussion above and the previous works
\citep{Kaiser:87,Scoccimarro04}, we {\em assume} that the redshift-space
power spectrum of galaxies (dark matter or halos)
is given by the
following functional form:
\begin{eqnarray}
&&\hspace{-1em}\langle\tilde{\delta}_s(\bmf{k})
\tilde{\delta}^\ast_s(\bmf{k}')
\rangle\equiv (2\pi)^3P_{\delta\delta}^{s}(k,\mu)\delta_D^3(\bmf{k}-\bmf{k}')\nonumber\\
&&\hspace{-1em}\rightarrow
P_{\delta\delta}^{s}(k,\mu)=
\left[P_{\delta\delta}(k)+2\mu^2
P_{\delta\theta}(k)+\mu^4
P_{\theta\theta}(k)\right]
F(k,\mu),\nonumber\\
\label{eq:Pgg}
\end{eqnarray}
where $P_{\delta\delta}$ and $P_{\theta\theta}$ are the power spectra of
density perturbation and velocity divergence and $P_{\delta\theta}$ is
the cross power spectrum:
\begin{eqnarray}
\langle\tilde{\delta}(\bmf{k})
\tilde{\delta}^\ast(\bmf{k}')\rangle&\equiv&
(2\pi)^3P_{\delta\delta}(k)\delta^3_D(\bmf{k}-\bmf{k}'),\nonumber \\
\langle\tilde{\delta}(\bmf{k})
\tilde{\theta}^\ast(\bmf{k}')\rangle&\equiv &
(2\pi)^3P_{\delta\theta}(k)\delta^3_D(\bmf{k}-\bmf{k}'),\nonumber \\
\langle\tilde{\theta}(\bmf{k})
\tilde{\theta}^\ast(\bmf{k}')\rangle&\equiv &
(2\pi)^3P_{\theta\theta}(k)\delta^3_D(\bmf{k}-\bmf{k}').
\label{eq:ps_def}
\end{eqnarray}
The form of Eq.~(\ref{eq:Pgg}) is often assumed in the literature
\citep[e.g.][and references
therein]{Hamilton:98,Taruyaetal:09}. However, as can be found from
Eqs.~(\ref{eq:delta_s_app}), (\ref{eq:Pgg}) and (\ref{eq:ps_def}), we
ignored the contributions of higher-order perturbations to the
redshift-space power spectrum. In fact we will discuss later that the
higher-order terms of Eq.~(\ref{eq:delta_s_app})
can be important for the redshift-space power
spectrum, especially for massive halos. To be more precise, if we
recall that the velocity perturbation is smaller than the density
perturbation at relevant low redshifts,
the leading-order correction to the Kaiser formula is
found in \cite{Taruyaetal:10} to be
\begin{equation}
\delta P_s(k,\mu)\leftarrow
k_\parallel\left\langle\tilde{\delta}(\bmf{k}')
\int\!\frac{d^3\bmf{q}}{(2\pi)^3}\frac{q_\parallel}{q^2}\tilde{\theta}(\bmf{q})
\tilde{\delta}(\bmf{k}-\bmf{q})
\right
\rangle.
\label{eq:higher_Ps}
\end{equation}
In Appendix~\ref{sec:higher_Kaiser} we derive the correction terms for
dark matter and halos based on the perturbation theory, and will use the
results for the following discussion.
Meanwhile we will assume Eq.~(\ref{eq:Pgg}) for simplicity.
The function in the square bracket on the r.h.s. of Eq.~(\ref{eq:Pgg})
denotes the Kaiser formula for redshift-space power spectrum, which is
valid only at large length scales in the linear regime. The function
$F(k,\mu)$ was introduced so as to take into account the nonlinear
distortion effect, the so-called Fingers-of-God (FoG) effect, which
causes a smearing of redshift-space clustering due to random virial
motions of dark matter particles or galaxies within halos. Thus the
assumption we employed in Eq.~(\ref{eq:Pgg}) is the FoG redshift
distortion and the Kaiser formula are separable functions in the
redshift-space power spectrum. This does not necessarily hold, although
an empirical model based on the halo model gives such a functional form
of redshift-space power spectrum \citep[][also see Hikage et al. in
preparation]{White:01,Seljak:01}. Hence the validity of
Eq.~(\ref{eq:Pgg}) needs to be further tested in combination with
simulations. The recent study done in \cite{Taruyaetal:09} gives a
possible verification on this treatment, where it was shown that the
form (\ref{eq:Pgg}) can well reproduce the simulation results in the
weakly nonlinear regime down to $k\simeq 0.2~ h{\rm Mpc}^{-1}$ if an
appropriate function $F(k,\mu)$ is employed.
A theoretical understanding of the FoG effect is still lacking
due to complicated physics involved in the nonlinear clustering
regime. In this paper we rather employ an empirical approach: we will
consider the following functional forms of $F(k,\mu)$ in order to study
how the results change with the different FoG models:
\begin{equation}
F(k,\mu)=
\left\{
\begin{array}{l}
\exp[-\sigma^2k^2\mu^2],\\
{\displaystyle \frac{1}{1+\sigma^2k^2\mu^2}},\\
{\displaystyle 1-\sigma^2k^2\mu^2+\frac{1}{2}\tau^4 k^4\mu^4}.\\
\end{array}
\right.
\label{eq:FOG}
\end{equation}
All the models have a limit of $F\rightarrow 1$ when $k\rightarrow 0$.
The first and second forms correspond to the Gaussian and Lorentzian FoG
models that are sometimes employed in the literature
\citep[e.g. see][for a review]{Hamilton:98}. We will treat $\sigma$
appearing in the forms as a free parameter in the following analyses,
motivated by the results in \cite{Taruyaetal:09}. The third form can be
considered as a more general form, in analogy to the Taylor expansion of
the FoG function in terms of $k\mu$, and this includes the Gaussian and
Lorentzian models in the range $k\mu\ll 1$. Similarly we will treat
$\sigma$ and $\tau$ as free parameters in the model fitting. We will
refer to these models as Gaussian, Lorentzian, Taylor-$\sigma$
and Taylor-($\sigma+\tau$) models, respectively.
Besides the assumed form of redshift-space power spectrum and the
FoG function (see Eqs.~[\ref{eq:Pgg}] and [\ref{eq:FOG}]), in the
following we will explore a model-independent reconstruction of the
density and velocity power spectra $P_{\delta\delta}$,
$P_{\delta\theta}$ and $P_{\theta\theta}$ at each $k$ bins, from the
measured galaxy distribution in redshift space.
More exactly speaking,
since the reconstructed power spectra are not necessarily same as the
density and velocity spectra,
our method recovers the real-space power spectra that are proportional
to $\mu^{2n}$ ($n=0,1,2$) in the redshift-space power spectrum, being
marginalized over uncertainties of the FoG effect.
Then we will assess the
performance of this reconstruction method by comparing the reconstructed
spectra with the spectra directly measured from simulations. This
reconstruction problem is not a linear problem, because the FoG function
is non-linearly coupled with the density and velocity spectra. Hence,
the reconstructed band powers at different $k$-bins become correlated
with each other even if the underlying density and velocity fields are
Gaussian.
Finally we remark on the Einstein gravity case, where the two metric
perturbations are equivalent: $\Psi=\Phi$ as discussed in
\S~\ref{sec:metric}. In this case the density and velocity power spectra
are related to each other in the linear regime
as
$P_{\delta\theta}\simeq \beta P_{\delta\delta}$ and
$P_{\theta\theta}\simeq \beta^2P_{\delta\delta}$, where $\beta=(1/b)d\ln
D/d\ln a$ with $D$ and $b$ being the linear growth rate and the linear
bias parameter, respectively.
Note that the possible nonlinear correction terms (see
Eq.~[\ref{eq:higher_Ps}]) can be also accurately computed based on
perturbation theory and/or simulations for a given cosmological model
\citep[e.g.][]{Taruyaetal:10,Jenningsetal:11}.
Thus, if the Einstein gravity is {\em a
priori} assumed, measuring the density and velocity power spectra helps
to break parameter degeneracies, especially the degeneracy between the
galaxy bias and the power spectrum amplitudes, which in turn helps to
significantly improve parameter constraints \citep[e.g.][]{Takadaetal:06}.
\section{A maximum likelihood reconstruction method of
redshift-space power spectra}
\label{sec:method}
In this section we develop a method for reconstructing the real-space
power spectra from the galaxy distribution in redshift space, based on
a maximum likelihood method.
We start with assuming that the mass density fluctuation field
$\delta_m(\bmf{x})$ in redshift space obeys the Gaussian likelihood
function:
\begin{eqnarray}
&&{\cal L}[\delta_s\!(\bmf{x})]
\propto \frac{1}{\sqrt{\rm det(\bmf{C})}}
\int\!\frac{d^3\bmf{x}_i}{V_s}
\int\!\frac{d^3\bmf{x}_j}{V_s}\nonumber\\
&&\hspace{5em}\times
\exp\left[
-\frac{1}{2}\delta_s(\bmf{x}_i)
(\bmf{C}^{-1})_{ij}\delta_s(\bmf{x}_j)
\right],
\label{eq:like_real}
\end{eqnarray}
where $V_s$ is the survey volume, $\bmf{C}(\bmf{x}_i-\bmf{x}_j)$ is
defined as $\bmf{C}(\bmf{x}_i-\bmf{x}_j)\equiv
\langle\delta_s(\bmf{x}_i)\delta_s(\bmf{x}_j) \rangle$, the two-point
correlation function between the density fields $\delta_s(\bmf{x}_i)$
and $\delta_s(\bmf{x}_j)$ in redshift space, and $\bmf{C}^{-1}$ is its
inverse matrix.
Analogously to the likelihood of the CMB temperature power spectrum
\citep[e.g.,][]{Verde03}, by converting the likelihood function to
Fourier space, we can derive the log-likelihood function for the
redshift-space power spectrum (see Appendix~\ref{sec:like} for the
detailed derivation; also see \cite{Percival:2005yq} for the similar
discussion):
\begin{equation}
-2\ln{\cal L}=\sum_{k_i, \mu_a}N(k_i,\mu_a)\left[ \frac{\hat{P^s}(k_i,\mu_a)}{P^s(k_i,\mu_a)} +\ln \frac{P^s(k_i,\mu_a)}{\hat{P^s}(k_i,\mu_a)}-1
\right], \label{eq:2dmethod}
\end{equation}
where $\hat{P}_s(k_i,\mu_a)$ is the power spectrum estimated at the bin
($k_i, \mu_a$):
\begin{equation}
\hat{P}^s(k_i,\mu_a)\equiv \frac{1}{N(k_i,\mu_a)}\sum_{\bmf{k}\in
(k_i,\mu_a)} \left|
\tilde{\delta}_s(\bmf{k})
\right|^2.
\label{eq:measuredPk}
\end{equation}
The quantity $N(k_i,\mu_a)$ is the number of independent Fourier modes
confined within the bin ($k_i, \mu_a$):
$
N(k_i,\mu_a)\equiv \sum_{\bmf{k}\in (k_i,\mu_a)} 1.
$
If a surveyed volume has a cubic geometry with side length $L$,
i.e. $V_s=L^3$, the fundamental mode to discriminate different Fourier
modes has the length given by $k_f=2\pi/L$. Hence the number of
independent Fourier modes, for the bin $(k_i,\mu_a)$, is approximately
given as $N(k_i,\mu_a)\approx 2\pi k_i^2\Delta k\Delta\mu/(2\pi/L)^3$
in the limit $k_i\gg k_f$, where $\Delta k$ and $\Delta \mu$ are the bin
widths. In the equation above, we ignored observational effects due to
survey geometry and masking of the surveyed region for simplicity. For
actual data we need to include these effects.
In Eq.~(\ref{eq:2dmethod}) $P^s(k_i,\mu_a)$ is the underlying true
redshift-space power spectrum at the bin $(k_i,\mu_a)$. We assume the
form given by Eq.~(\ref{eq:Pgg}) for $P^s(k_i,\mu_a)$, which is given by
the model power spectra $P_{\delta\delta}(k)$, $P_{\delta\theta}(k)$ and
$P_{\theta\theta}(k)$ and the parameters to model the FoG effect (see
Eq.~[\ref{eq:FOG}]). Hence, given the measured redshift-space power
spectrum $\hat{P}^s(k_i,\mu_a)$
(Eq.~[\ref{eq:measuredPk}]), we can estimate the best-fit
power spectra $P_{\delta\delta}(k)$, $P_{\delta\theta}(k)$, and
$P_{\theta\theta}(k)$ at each $k_i$-bin, including marginalization
over the band
powers at different $k$-bins and the FoG effect parameters, in such a
way that the log-likelihood (\ref{eq:2dmethod}) is maximized. This is
the maximum likelihood method for reconstructing the real-space
spectra.
We will demonstrate how the method above allows a reconstruction of the
real-space power spectra using simulations. To do this,
we will use the Markov-Chain Monte Carlo (MCMC) sampling method
\citep[e.g.][]{cosmomc}, more specifically Metropolis-Hasting algorithm in
our work.
The chain convergence is diagnosed by using the criteria given in
\citet{dunkley_mcmc}. The free parameters are: the band powers at each
$k$
bin, $P_{\delta\delta}(k_i)$, $P_{\delta\theta}(k_i)$ and
$P_{\theta\theta}(k_i)$, and the parameters to model the FoG effect given by
Eq.~(\ref{eq:FOG}), where $k_i$ denotes the $i$-th wavenumber bin and
the index $i$ runs over the number of bins. If we employ $N_{\rm bin}$
for the bin number over $k_{\rm min}\le k\le k_{\rm max}$, the total
number of model parameters are $3\times N_{\rm bin}$ plus the number of
the FoG parameters, 1 or 2 ($\sigma$ or $\sigma$ and $\tau$,
respectively), depending on which FoG model to use.
In the MCMC parameter search, we adopted
the following priors on model parameters:
$P_{\delta\delta}>P_{\delta\theta}>P_{\theta\theta}>0$ and the FoG function $0<F(k,\mu)\le 1$.
Note that the latter prior, $0<F(k,\mu)\le 1$ is automatically satisfied by the
Gaussian and Lorentzian FoG models in Eq.~(\ref{eq:FOG}).
Another assumption we employed in the log-likelihood function is the Gaussian
field assumption. This Gaussian assumption breaks down in the weakly
nonlinear regime, indeed over a range of wavenumbers relevant for the
method above. However, \cite{Takahashietal:11} showed that, using 5000
N-body simulation realizations, the non-Gaussianity of the density field
does not cause any large impact on parameter estimation in the weakly
nonlinear regime. Hence we do not think that the non-Gaussianity
affects the following results.
\section{N-body simulations and halo catalogs}
\label{sec:dm_catalog}
To test the performance of the power spectrum reconstruction method
described in the preceding section, we will
implement a hypothetical
experiment: we will apply the method to mock data from N-body
simulations, and then compare the reconstructed spectra with the input
spectra directly measured from simulations.
In this section we describe
some details of N-body simulations and the halo catalogs we will use in
the following sections.
\subsection{N-body Simulations}
The N-body simulations are generated by running the GADGET
\citep{Springel:05} assuming a flat universe; the matter density
$\Omega_{\rm m}=0.238$, the baryon content $\Omega_{\rm b}=0.041$, the
Hubble constant $H_0=73.2 {\rm km s^{-1} Mpc^{-1}}$, the spectral index
$n_s=0.958$,
and the amplitude of
the linear power spectrum $\sigma_8=0.76$ \citep{Spergel:2007fk}. The
transfer function is calculated by the CAMB \citep{Lewis:1999bs}. We
include $512^{3}$ N-body particles in a box of
1${ h^{-3}{\rm
Gpc}}^3$ volume. We started the simulations from the initial redshift $z=30$,
and set the initial conditions of N-body particles using the Zel'dovich
approximation. In this paper we use the outputs of $z=0$ and 1.
Our initial redshift may not be sufficiently early to accurately
compute the nonlinear clustering of N-body particles as discussed in
e.g. \citet[][]{Crocce:2006uq}.
However, the main purpose of this paper is to study
whether the reconstruction method of the power spectra $P_{\delta\delta}$ and $P_{\delta\theta}$
can reproduce the spectra directly measured from simulations, so
the accuracy of N-body simulations is not our concern.
We will use 70 realizations in order to reduce the statistical scatters.
Using these N-body simulations, we also construct halo catalogs by
adopting the friend-of-friend (FOF) method with the linking length of
$b=0.2$ ($20\%$ of the mean separation). The minimum number of member
particles is set to 20, which corresponds to the mass threshold of halos
$9.8\times10^{12}M_\odot$ for both the $z=0$ and $1$ outputs, and
the resulting number density of halos is $\bar{n}\simeq 3.8\times
10^{-4}~h^3{\rm Mpc}^{-3}$, which is comparable to the number density of
SDSS luminous red galaxies (LRGs) targeted for the ongoing BOSS survey
\citep{Whiteetal:10}.
We will use these
halo catalogs to compute the halo power spectrum, and then address
whether the power spectrum reconstruction method can also work for the
halo power spectrum.
\subsection{Power Spectrum Measurement from Simulations}\label{sec:ps_measurement}
From the simulation data above, we measure
the redshift-space power spectrum $P^s(k_i,\mu_a)$ in the
two-dimensional $(k_i,\mu_a)$ bins,
as well as the real-space spectra; the density-density power spectrum
$P_{\delta\delta}(k)$, the density-velocity power spectrum
$P_{\delta\theta}(k)$, and the velocity-velocity power spectrum
$P_{\theta\theta}(k)$ (see Eq.~[\ref{eq:ps_def}]).
In the following we describe how we measure these spectra from the
N-body simulations and the halo catalogs.
\subsubsection{Dark matter spectra}
\begin{figure*}
\begin{center}
\includegraphics[width=15.0cm,angle=0]{P2d_dm.eps}
\end{center}
\caption{Power spectra measured from N-body simulations at redshifts
$z=0$ ({\em right panel}) and $z=1$ ({\em left}), respectively. The
color scales show the redshift-space power spectrum amplitudes as a
function of $k_\perp$ and $k_{\parallel}$, where $k_{\perp}$ and
$k_{\parallel}$ are wavelengths perpendicular and parallel to the
line-of-sight direction (which is taken as the $z$-axis direction in
simulations). Shown is the mean power spectrum among the spectra of
70 realizations, each of which has a volume of $1~[h^{-1}{\rm
Gpc}]^3$. The anisotropic modulations of the band powers are due to the
redshift distortion effect due to the peculiar motions of N-body
particles (see text for the details). The spectra for $z=0$ shows a
stronger FoG effect: a stronger squashed feature of the iso-contours
along the $k_{\parallel}$ direction. For comparison, the solid
contours show the real-space spectra, which have isotropic
contours. The contours are stepped by $\Delta \log P(k) = 0.11$. } \label{fig:2dps_dm}
\end{figure*}
First let us discuss the spectra measured from the N-body simulations.
For redshift-space power spectrum, the distribution of N-body
particles is mapped into the redshift-space distribution taking into
account the modulation of their positions due to the redshift
distortion, where the line-of-sight direction is simply taken to be in
the $z$-axis direction in each simulation.
Note that we here adopted the distant observer approximation for simplicity.
The power spectrum of N-body particle distribution is measured using the
fast Fourier transform method (FFT). In doing this we first used the
``Cloud-in-Cell'' (CIC) interpolation method for assigning N-body
particles to the $512^{3}$ uniformly-distributed grids in order to
construct the grid-based density field. Then we implemented the FFT
method on the density field to obtain the Fourier-transformed
coefficients,
$\tilde{\delta}_s(\mathbf k)$.
Using Eq.~(\ref{eq:measuredPk}), we
estimate, in each simulation realization, the redshift-space power
spectrum, $\hat{P}^s(k_i,\mu_a)$, from the Fourier coefficients of the
density field. To reduce the statistical scatters, we will use the
averaged power spectrum of 70 realizations, and infer the 1$\sigma$
statistical errors from the scatters among the 70 realizations, which
correspond to the sampling variance for a volume of $1~h^{-3}{\rm Gpc}^3$.
Fig.~\ref{fig:2dps_dm} shows the redshift-space power spectrum (color
scales) for dark matter (N-body particles), measured from the
simulations of $z=0$ and 1 outputs. The redshift-space power spectrum
is compared with the real-space density power spectrum (contours). The
figure clearly shows redshift-space distortion effects. The Kaiser
effect due to large-scale bulk motions increases the redshift-space
power spectrum amplitudes along the line-of-sight direction or
equivalently $k_\parallel$, stretching the iso-contours towards the
larger $k_\parallel$. On the other hands, the FoG effect squashes the
iso-contours towards the smaller $k_\parallel$. Comparing the left- and
right-panels clarifies that the FoG effect is stronger at lower redshifts.
The real-space power spectra $P_{\delta\delta}(k), P_{\delta\theta}(k)$ and $P_{\theta\theta}(k)$ are
estimated from simulations as follows. The power spectrum $P_{\delta\delta}(k)$ is
just similar to the redshift-space spectrum as described above, but
skipping a step to compute redshift modulation due to the peculiar
velocities. For $P_{\delta\theta}(k)$ and $P_{\theta\theta}(k)$, we first assign the velocity
components of each N-body particles to the $512^3$ uniformly-distributed
grids based on the CIC method, and then use the FFT method to generate
the Fourier coefficients of the velocity fields,
$\tilde{v}_i(\bmf{k})$. The velocity-divergence field is computed at
each Fourier grid as $\tilde{\theta}(\bmf{k})\propto
\bmf{k}\cdot\tilde{\bmf{v}}(\bmf{k})$. If a larger number of grids than
$512^3$ (i.e. the smaller-size grid) is used, some grids may not contain
any N-body particle, which causes an ill-behaved spectrum $P_{\delta\theta}(k)$ at
small $k$ bins. On the other hand, if we use a smaller number of
grids than $512^3$, the CIC interpolation causes a smoothing of the
velocity power spectrum amplitudes at large $k$ bins we are interested
in, as carefully studied in \cite{Pueblas:2009qy}. Hence we checked that
the 512$^3$ grids are rather close to an optimal choice of the grid
number in order to avoid these artificial effects over a range of scales
we are interested in. We again use the averaged power spectra, $P_{\delta\delta},
P_{\delta\theta}$ and $P_{\theta\theta}$ from 70 realizations to reduce the statistical
scatters.
\begin{figure}
\begin{center}
\includegraphics[width=9.0cm,angle=0]{./Ps_z=0vs1.eps}
\end{center}
\caption{The density-density ($P_{\delta\delta}$), density-velocity ($P_{\delta\theta}$) and
velocity-velocity ($P_{\theta\theta}$) power spectra at $z=0$ and $1$, respectively,
for N-body simulation particles. Similarly to the previous plot, shown
is the mean spectra of 70 realizations (see text for the details). The
statistical scatters around the mean spectra are sufficiently small, so
we do not show the scatters here (in other words, the average spectra
are well-converged).
}
\label{fig:ps_dm}
\end{figure}
\begin{figure}
\includegraphics[width=8.0cm,angle=0]{./P2d_Halo.eps}\\
\includegraphics[width=8.0cm,angle=0]{./Ps_z=0vs1_Halo.eps}
\caption{ {\em Top panel}: The redshift-space power spectra of halos,
similarly to Fig.~\ref{fig:2dps_dm}. We used halo catalogs
containing halos with masses greater than $\approx
10^{13}h^{-1}M_\odot$, roughly corresponding to halos hosting LRGs.
Slightly jaggy contours in the plot are due to the smaller number of
halos compared to the case of N-body particles.
Compared to Fig.~\ref{fig:2dps_dm}, the halo spectra shows a less FoG
effect than in N-body particles, because halos have only bulk-velocity
contributions in large-scale structure.
{\em Lower panel}: Similarly to
Fig.~\ref{fig:ps_dm}, the real-space power spectra
for the halo
distribution. For the density-density power spectrum $P_{\delta\delta}$, the shot
noise contamination $P_{\rm sn}=1/\bar{n}_{\rm halo}$ is subtracted.
Note that the density power spectra (solid curves) have
greater amplitudes at $z=1$ than at $z=0$ because of the greater halo
biases, where we used the same mass threshold of $M_{\rm min}\approx
10^{13}h^{-1}M_\odot$ for both the two redshift outputs. For the
velocity-related spectra we used the velocity field defined from N-body
particles, instead of the velocity field of halos, because we found the
difficulty of defining the velocity field for halos that are too
sparsely sampled (see text for the details). Hence the velocity-velocity
spectra $P_{\theta\theta}$ shown here are same as those in
Fig.~\ref{fig:ps_dm}. \label{fig:ps_halo}}
\end{figure}
Fig.~\ref{fig:ps_dm} shows the real-space spectra of dark
matter (N-body particles): $P_{\delta\delta}(k)$, $P_{\delta\theta}(k)$ and $P_{\theta\theta}(k)$, for the two
redshift outputs of $z=0$ and 1.
One can clearly find the relation $P_{\delta\delta}>P_{\delta\theta}>P_{\theta\theta}$,
and therefore the approximation given by Eq.~(\ref{eq:delta_s_app}) is
considered valid.
\subsubsection{Halo spectra}
\label{sec:halo_spectra}
Now let us move on to discussion on the the power spectra measured from
the halo catalogs.
First we need to define the spatial position and the velocity for each
halo in the simulation.
We use the center-of-mass position, computed from N-body particles
contained within each halo, as the spatial position of the halo, while
we assign the mean of member N-body particles' velocities to the
velocity of the halo. Then, to get the density field for the
discrete halo distribution in each realization, we adopt the
Nearest-Grid-Point (NGP) method to assign the density field in $512^3$
uniformly-distributed grids, after the redshift modulation due to the
halo velocity field in redshift space are taken into account. Similarly
to the cases for N-body particles, we computed the density power spectra
in redshift- and real-space, $P^{s}(k,\mu)$ and
$P_{\delta\delta}(k)$.
On the other hand, however, the velocity related power spectra for the
halo distribution, $P_{\delta\theta}$ and $P_{\theta\theta}$, require
some caution, because it is not straightforward to define the
continuously-varying velocity field from the halo distribution that has
a much smaller number density (typically $\sim 10^{-4}~[h^{-1}{\rm
Mpc}]^{-3}$) than that of N-body particles. We tried several
interpolation methods such as the CIC and the Delaunay triangulation
interpolation method for which we used the publicly available code from
the Computational Geometrical Algorithms Library (CGAL:
http://www.cgal.org/). However, we could not find a reliable result for
the velocity power spectra at scales of interest in such a way that the
power spectra obtained become insensitive to the interpolation
method. Hence, instead of pursuing a more appropriate method to obtain
the halo velocity field, we use the grid-based velocity field of N-body
particles, {\em assuming} no velocity bias between the halo and N-body
particle (dark matter) distribution.
Thus we computed the power spectra, $P_{\delta\theta}$ and
$P_{\theta\theta}$, combining the halo density field and the N-body particle
velocity field.
We will again use the mean halo power spectra from
70 realizations.
Fig.~\ref{fig:ps_halo} shows the redshift-space and real-space spectra
for the halo distribution, measured from the 70 realizations according
to the method we described above. Compared to Fig.~\ref{fig:2dps_dm},
the redshift-space power spectrum of halos shows almost no FoG effect,
because the halo spectrum does not include contributions from the virial
motions of particles within each halo, and rather includes only the
contribution from the bulk motion of each halo.
\section{Results}
\label{sec:results}
\subsection{Reconstruction of matter power spectra}
\label{sec:dm_reconstruction}
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm,angle=0]{./RecPk_dm.eps}
\end{center}
\caption{The symbols show the reconstructed power spectra,
$P_{\delta\delta}(k)$ and $P_{\delta\theta}(k)$, based on the maximum
likelihood method (see Eq.~[\ref{eq:2dmethod}] and
\S~\ref{sec:method}), for dark matter.
The different symbols are the results assuming the
different models of FoG effect: $F(k,\mu)$ (Eq.~[\ref{eq:FOG}]) in the
model redshift-space power spectrum (Eq.~[\ref{eq:Pgg}]). The triangle
symbols are the results ignoring the FoG effect $F=1$; the square shows
the results assuming the Gaussian $F(k,\mu)$, which has a single
parameter; the diamond and plus symbols show the results assuming the
Taylor expansion forms for $F(k,\mu)$ up to different orders of $(k\mu)
$, which are characterized by one ($\sigma$) and two ($\sigma, \tau$)
free parameters, respectively. The reconstructed band powers at each $k$
bins include marginalization over uncertainties in reconstructing band
powers of $P_{\delta\delta}$, $P_{\delta\theta}$ and $P_{\theta\theta}$
at different $k$ bins as well as the FoG parameter(s). The error bars
around the symbols denote statistical uncertainties in the
reconstruction for the volume $1~h^{-3}{\rm Gpc}^3$, computed from the
MCMC-based posterior distributions.
Note that the
band powers at different $k$ bins are correlated. For comparison, the
dotted curves show the power spectra, $P_{\delta\delta}(k)$ and $P_{\delta\theta}(k)$,
directly measured from the
simulations. For $P_{\delta\delta}(k)$ the dotted curve and the symbols are
almost perfectly overlaid. } \label{fig:re_dm}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=8.cm,angle=0]{./dRecPk_dm.eps}
\end{center}
\caption{Using the results in Fig.~\ref{fig:re_dm}, the plot shows
fractional differences between the reconstructed power spectra and the
spectra measured from the simulations for $P_{\delta\delta}$ (upper
panel) and $P_{\delta\theta}$ (lower), respectively. To be more
precise, $\Delta P/P\equiv \left[ P({\rm reconst.})-P({\rm
input})\right]/P({\rm input})$, where $P_{\rm input}$ and $P_{\rm
reconst.}$ are the input and reconstructed power spectra,
respectively.
The different
symbols are as in the previous plot. } \label{fig:re_diff_dm}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=5.7cm,angle=0]{./slice_RecPk_dm_k1.eps}
\includegraphics[width=5.7cm,angle=0]{./slice_RecPk_dm_k5.eps}
\includegraphics[width=5.7cm,angle=0]{./slice_RecPk_dm_k18.eps}
\end{center}
\caption{Comparing the best-fit redshift-space power spectrum, based on
the maximum likelihood method, with the redshift-space spectrum
directly measured from simulations. The solid histograms show slices
of the redshift-space power spectrum amplitudes as a function of $\mu$,
for a fixed $k$: $k=0.08$, $0.16$ and
$0.29~ h{\rm Mpc}^{-1}$ from the left to right panels,
respectively. Note $\mu$ denotes the cosine angle between the
wavevector $\bmf{k}$ and the line-of-sight direction: $\mu\equiv
\cos(\hat{\bmf{k}\cdot}\hat{\bmf{k}}_\parallel)$. The best-fit spectra,
denoted by the different symbols, are computed by inserting the
best-fit band powers of $P_{\delta\delta},P_{\delta\theta}$ and $P_{\theta\theta}$ at each $k$ bin and the
best-fit FoG parameters into Eq.~(\ref{eq:Pgg}), which are shown in
Fig.~\ref{fig:re_dm}. For comparison, the dashed curves in each panel
show the spectra computed by inserting the simulation-measured spectra
$P_{\delta\delta}, P_{\delta\theta}$ and $P_{\theta\theta}$ in
Eq.~(\ref{eq:Pgg}), but ignoring the FoG effect, i.e. setting
$F(k,\mu)=1$. Hence the differences between the dashed curves and, for
example, the histograms are due to the FoG effect. The reconstructed
power spectra obtained using the Taylor ($\sigma+\tau$) FoG model
(denoted by the cross symbols) are found to well reproduce the
redshift-space power spectrum.
} \label{fig:2d_slices_z=0}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=4.1cm,angle=0]{./pdf_pdd.eps}
\includegraphics[width=4.1cm,angle=0]{./pdf_pdv.eps}
\includegraphics[width=4.1cm,angle=0]{./pdf_pvv.eps}
\includegraphics[width=4.1cm,angle=0]{./pdf_sigma.eps}
\end{center}
\caption{
The posterior distributions of the band powers $P_{\delta\delta}$, $P_{\delta\theta}$ and $P_{\theta\theta}$
at $k=0.16h^{-1}{\rm Mpc}$ and the FoG parameter $\sigma$, for the power spectrum
reconstruction of $z=0$ using the Gaussian FoG model as a demonstration
example. The histograms are computed from MCMC chains. The solid curve
in each panel represents a Gaussian distribution with the mean and
variance given by the MCMC chains.
The distribution of $P_{\theta\theta}$ includes a range of $P_{\theta\theta}=0$, meaning that
the band power is not well constrained.
}
\label{fig:exmp_pxx}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=5.8cm,angle=0]{./pdf2d_pdv-pdd_bin5.eps}
\includegraphics[width=5.8cm,angle=0]{./pdf2d_pvv-pdv_bin5.eps}
\includegraphics[width=5.8cm,angle=0]{./pdf2d_sigma-pdv_bin5.eps}
\end{center}
\caption{The color scales represent the marginalized 2d probability
distribution between the band powers ($P_{\delta\delta}$, $P_{\delta\theta}$, $P_{\theta\theta}$) and
$\sigma$ as in the previous figure.
The two contours levels represent the confidence levels of
68\% and 95\%,
respectively.
}
\label{fig:exmp_contour}
\end{figure}
We first assess performance of the maximum likelihood reconstruction
method developed in \S~\ref{sec:method} for matter spectra, by using
N-body simulations.
We stress here again that the real-space density and velocity spectra
used to compare with the reconstructed power spectra, shown in figures
of this and following sections, are the spectra directly measured from
the simulations, and therefore include nonlinearity effects arising from
nonlinear clustering in structure formation.
To apply our method to N-body simulations, we need to compute the
likelihood function, given by Eq.~(\ref{eq:2dmethod}), for the
redshift-space density field. More precisely, in
Eq.~(\ref{eq:2dmethod}), we need to specify a survey volume $V_s$, which
determines the statistical uncertainties, and need to compute the
redshift-space power spectrum $P_s(k_i,\mu_a)$ at each $k$- and
$\mu$-bins from the simulations. In the following we assume
$V_s=1h^{-3}{\rm Gpc}^3$ and use the spectrum $P_s(k_i,\mu_a)$ averaged
from 70 realizations to reduce the statistical scatters for illustrative
purpose. For the $k$-binning, we mainly use 19
wavenumber bins over $0.034\le k\le 0.3$~$h$Mpc$^{-1}$. We determined
the bin widths of $k$ and $\mu$ such that the area $\Delta k \times
k\Delta \mu$ in the two-dimensional Fourier space of $(k,\mu)$ is kept
about constant.
Therefore, instead of using the constant bin width,
$\Delta k$ is taken to be
large at small $k$ and gradually become smaller at larger $k$.
Since we use the bin width $\Delta \mu =0.067$, we use $\Delta k\simeq
0.02~h$Mpc$^{-1}$
at small $k$ bins,
while we use $\Delta k=0.01$ around $k\simeq 0.2~{\rm Mpc}^{-1}$.
We will
show below the reconstruction results for the density-density and
density-velocity power spectra, $P_{\delta\delta}$ and
$P_{\delta\theta}$, but not for the velocity-velocity spectrum
$P_{\theta\theta}$, because the reconstruction of $P_{\theta\theta}$ is
very noisy due to the smaller amplitudes, i.e. the small
signal-to-noise ratios, compared to $P_{\delta\delta}$ and
$P_{\delta\theta}$ \citep[also see][for the similar discussion]{Tegmarketal:04}.
Fig.~\ref{fig:re_dm} shows the results when using the simulation outputs
at $z=0$, for dark matter (N-body particle) distribution.
The top-dotted curve shows the input density-density power
spectrum, $P_{\delta\delta}(k)$, directly measured from N-body
simulations (the average of 70 realizations). The three symbols around
the curve, although almost perfectly overlapped with each other, show
the reconstructed power spectra assuming different FoG models
(Eq.~[\ref{eq:FOG}]). Note that we here show the results for the
Gaussian and Taylor-type FoG models, and do not and will not show the
results for the Lorentzian FoG model for illustrative purpose. The
results for the Lorentzian FoG model is very similar to the results of
the Gaussian and Taylor ($\sigma$) models; that is, we have found that
all the results for
one-parameter FoG models are similar.
The error bars around the symbols, although again
overlapped, show $1\sigma$ statistical uncertainties in the band power
reconstruction at each $k$-bin, including marginalization over
uncertainties in the reconstructed band powers at different $k$ bins and
for
different spectra $(P_{\delta\delta}, P_{\delta\theta},
P_{\theta\theta}$) as well as the FoG effect parameters.
Encouragingly our method can well recover $P_{\delta\delta}(k)$, rather
irrespectively of the assumed FoG model.
To be more precise the upper panel of Fig.~\ref{fig:re_diff_dm} shows
fractional differences between the input and reconstructed spectra:
$\Delta P/P\equiv \left[ P({\rm reconst.})-P({\rm input})\right]/P({\rm
sim.})$, where $P({\rm reconst.})$ and $P({\rm input})$ are the
reconstructed spectrum and the spectrum directly measured from
simulations, respectively. Our method recovers the input power spectrum
within the statistical errors, achieving a few percent accuracy up to
$k\simeq 0.2h$Mpc$^{-1}$. If we employ the Taylor-type FoG model
including the orders up to $(k\mu)^4$ (hereafter Taylor ``$\sigma+\tau$
model'') in Eq.~(\ref{eq:FOG}), our method can recover $P_{\delta\delta
}$ up to $k\simeq 0.3h$Mpc$^{-1}$, which is well in the nonlinear
regime.
The lower curves with different symbols in Fig.~\ref{fig:re_dm} show the
reconstruction results for the density-velocity power spectrum,
$P_{\delta\theta}(k)$, assuming different FoG models
(Eq.~[\ref{eq:FOG}]). The reconstruction of $P_{\delta\theta}$ is noisier
than in $P_{\delta \delta}$ due to the lower signal-to-noise ratios
\citep{Tegmarketal:04}. Also the reconstruction is sensitive to which
FoG model is assumed, reflecting that the redshift-space power spectrum
is affected by the FoG effect over a range of wavenumbers we
consider. Fig.~\ref{fig:re_dm} shows that the reconstructed
$P_{\delta\theta}$ is in closest agreement with the input spectrum, if
using the Taylor-($\sigma+\tau$) FoG model that is given by two free
parameters and has more degrees of freedom to describe a scale-dependent
FoG effect than other models (that respectively has only one free
parameter).
The lower panel of Fig.~\ref{fig:re_diff_dm} shows fractional
differences between the input and reconstructed spectra for
$P_{\delta\theta}$. Combined with the result for $P_{\delta\delta}$ shown in the
upper panel, one can notice that, although $P_{\delta\delta}$ and $P_{\delta\theta}$ are
unbiasedly recovered regardless of the FoG models at small $k$, the
results are substantially different at large $k$ depending on which FoG
model to use. The FoG redshift distortion increasingly affects the
power spectrum with increasing $k$.
As a result, the reconstructed band
powers at different $k$ bins are correlated
with each other via the FoG effect being marginalized over,
and therefore the correlations need to be properly taken into account
(see below).
Fig.~\ref{fig:re_diff_dm} shows that, among the different FoG
models, the performance of the Taylor ($\sigma+\tau$) model is of
promise; it can unbiasedly recover $P_{\delta\delta}$ over all the scales as well as
$P_{\delta\theta}$ within the statistical uncertainties up to $k\simeq 0.25~h{\rm
Mpc}^{-1}$, implying that the FoG model can nicely fit the strong FoG
effect in simulations.
To have more insights on the results in Figs.~\ref{fig:re_dm} and
\ref{fig:re_diff_dm}, Fig.~\ref{fig:2d_slices_z=0} show slices of the
redshift-space power spectrum amplitudes, $P_s(k,\mu)$, as a function of
the azimuthal angle $\mu$, for a fixed radius $k$: $k=0.08$, $0.16$ and
$0.29$~$h{\rm Mpc}^{-1}$ from the left to right panels, respectively.
The histograms in each panel show the band powers measured from
simulations, which are compared with the best-fit power spectra
(different symbols) obtained in Fig.~\ref{fig:re_dm}. The best-fit
spectra are computed by inserting the best-fit parameters (band powers
at each $k$ bins and the FoG parameters) into Eq.~(\ref{eq:Pgg}). As in
Fig.~\ref{fig:re_dm}, the different symbols are computed for the
different FoG models. For comparison, we also plot the spectra, by
dashed curves, which are computed by inserting the directly- measured
$P_{\delta\delta}$, $P_{\delta\theta}$ and $P_{\theta\theta}$ in
Eq.~(\ref{eq:Pgg}), without FoG term, i.e. $F(k,\mu)=1$. Hence, the
difference between the dashed curve and the solid histogram clarifies
how strongly the FoG affects the redshift-space power spectrum at each
$k$ bin.
The two extreme
cases are very distinctive. At small $k$ ($k=0.08h^{-1}{\rm Mpc}$) the FoG
effect is very small, and all the reconstructed power spectra can well
match the input spectra
independently of the FoG models.
One the other hand, at the largest $k$
($k=0.29h^{-1}{\rm Mpc}$), one can clearly see that the FoG effect is so
strong that none of the Taylor ($\sigma$) or Gaussian models (also or
Lorentzian model)
can fit the $\mu$-dependence of redshift-space power spectrum.
It
is essential to add an additional parameter in the FoG model, like
Taylor ($\sigma+\tau$) model, to
reproduce the simulation results.
The middle panel shows the result at the
intermediate scale ($k=0.16h^{-1}{\rm Mpc}$), where the FoG
effect is mild and the FoG models of one parameter work to
a good approximation.
Fig.~\ref{fig:exmp_pxx} shows the posterior, marginalized
distributions of the band
powers, $P_{\delta\delta}$, $P_{\delta\theta}$, and $P_{\theta\theta}$ at $k=0.16~ h^{-1}{\rm Mpc}$ and the $\sigma$
parameter of the Gaussian FoG model,
which are obtained from the MCMC
chains.
Here we show the reconstruction results
assuming the Gaussian
FoG model, which well works at the scale of $k=0.16h^{-1}{\rm Mpc}$ as shown in
Fig.~\ref{fig:2d_slices_z=0}.
The figure shows that, while the distribution of $P_{\delta\delta}$ looks Gaussian,
the distributions of $P_{\delta\theta}$, $P_{\theta\theta}$ and $\sigma$ show skewed,
non-Gaussian distributions. In particular, the distribution of $P_{\theta\theta}$
has a wide distribution and includes a region around $P_{\theta\theta}=0$, showing
no constraint on the band power of $P_{\theta\theta}$. Given these results, we
conclude that it is very difficult to reliably reconstruct $P_{\theta\theta}$ based on the
method developed in this paper, at least for a survey with survey volume
$\sim 1$Gpc$^3$.
The origin of the skewed distributions in Fig.~\ref{fig:exmp_pxx} is
explored in Fig.~\ref{fig:exmp_contour}, which shows the posterior
distributions in a two-parameter sub-space between the parameters in
Fig.~\ref{fig:exmp_pxx}. The figure clearly shows that the different
parameters are correlated with each other after the nonlinear
reconstruction. In particular, the $\sigma$ parameter of Gaussian
FoG model, shown here as an example, shows a strong correlation with the
band power $P_{\delta\theta}$. Thus the band powers of different spectra
($P_{\delta\delta},P_{\delta\theta},P_{\theta\theta}$) at different $k$ bins are correlated with each other,
and the correlation needs to be properly taken into account for the power
spectrum reconstruction.
\begin{figure}
\centering
\includegraphics[width=8.8cm,angle=0]{./RecPk_comp-nk.eps}
\caption{ {\em Upper panel}: Sensitivity of the reconstructed power
spectra $P_{\delta \delta}$ and $P_{\delta\theta}$ at $z=0$ to the
maximum wavenumber $k_{\rm max}$, where the redshift-space power
spectrum information up to $k_{\rm max}$ is used for the power
spectrum reconstruction. The triangle and square symbols show the
results for $k_{\rm max}=0.16$ and $0.30~h^{-1}{\rm Mpc}$, respectively,
assuming the Gaussian FoG effect as a working example. From
comparison, the star symbols show the results assuming $k_{\rm
max}=0.16~h^{-1}{\rm Mpc}$ and the Taylor ($\sigma$) FoG model. {\em Middle
and lower panels}: The fractional differences between the input and
reconstructed power spectra as in Fig.~\ref{fig:re_diff_dm}.
\label{fig:2d_nk_cmp}
}
\end{figure}
Given such strong correlations between the band powers, how sensitive is
the reconstruction of the power spectrum to a choice of the maximum
wavenumber $k_{\rm max}$? Fig.~\ref{fig:2d_nk_cmp} studies this
question. With increasing $k_{\rm max}$, the redshift-space power
spectrum to use for the reconstruction is more affected by the FoG
effect, and in turn the reconstructed $P_{\delta\delta}$ and $P_{\delta\theta}$ become
increasingly affected by the FoG effect after marginalization. The
figure shows the reconstructed $P_{\delta\delta}$ and $P_{\delta\theta}$ obtained when including
the redshift-space power spectrum information up to $k_{\rm max}=0.16$
and $0.3~h{\rm Mpc}^{-1}$, respectively. For the triangle and square
symbols, we assumed the Gaussian FoG model for comparison. The results
for $P_{\delta\delta}$ agree and are only slightly different at scales around
$k_{\rm max}=0.16~h{\rm Mpc}^{-1}$, while the results for $P_{\delta\theta}$ are
systematically different. Since the Gaussian FoG model cannot well
describe the FoG effect seen in simulations as implied in
Fig.~\ref{fig:re_dm}, the inaccuracy of the Gaussian FoG model causes a
systematic underestimation in the band powers of $P_{\delta\theta}$ if including the
higher-$k$ modes. However, the two results for $P_{\delta\delta}$ and $P_{\delta\theta}$ agree
over an overlapping range of $k$, up to $k=0.16~h{\rm Mpc}^{-1}$, within
the error bars. For comparison, we also show the results obtained
assuming $k_{\rm max}=0.16~h{\rm Mpc}^{-1}$ and the Taylor ($\sigma$)
FoG model (see Eq.~[\ref{eq:FOG}]). The Taylor ($\sigma$) model is found
to give a less biased reconstruction of $P_{\delta\theta}$, implying the importance
of the assumed FoG model even around $k\simeq 0.16~h{\rm Mpc}^{-1}$.
\begin{table}
\begin{center}
\begin{tabular}{|l|l|l|}
\hline
FoG model & $k_{\rm max}=0.3h^{-1}{\rm Mpc}$ & $k_{\rm max}=0.16h^{-1}{\rm Mpc}$ \\
\hline
Gaussian ($\sigma$) & $155^{+24.5}_{-30.5}$ & $215^{+118}_{-137}$ \\
Taylor ($\sigma$) & $105^{+23.1}_{34.4}$ & $170^{+96.2}_{-109}$ \\
Taylor ($\sigma+\tau$) & ($287^{+17.9}_{-17.0}$, $356^{+73.1}_{-76.2}$) & ---\\
\hline
\end{tabular}
\caption{The best-fit FoG parameters
assuming different $k_{\rm max}$ for the results of $z=0$ simulations.
The units of the numbers
shown here are $h{\rm Mpc}^{-1}$.
For the Taylor
($\sigma+\tau$) model the parameters $\sigma$ and $\tau$ are not well
constrained if using the redshift-space spectrum information up to
$k_{\rm max}=0.16~h^{-1}{\rm Mpc}$, and therefor are not shown here.}
\label{tb:FoG_002}
\end{center}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=8.cm,angle=0]{./pdf2d_tau-sigma.eps}
\end{center}
\caption{The contours represents the marginalized 2d probability
distribution in the parameter space ($\sigma, \tau$) of the Taylor
($\sigma+\tau$) FoG model (for the results shown in
Figs.~\ref{fig:re_dm} and \ref{fig:re_diff_dm}).
The two contours levels represent the
confidence levels 68\% and 95\%, respectively. }
\label{fig:sigma_tau_002}
\end{figure}
Table~\ref{tb:FoG_002} summarizes the best-fit FoG parameters and the
marginalized confidence ranges that are obtained for the reconstructions
at $z=0$ assuming different $k_{\rm max}$: $k_{\rm max}=0.3$ and
$0.16h^{-1}{\rm Mpc}$, respectively. For the Taylor ($\sigma+\tau$) FoG model,
the error of $\tau$ parameter is smaller than that of $\sigma$, because
the $\tau$ parameter has a stronger dependence on $k\mu$ as $
(\tau k\mu)^4$, than the $\sigma$-term does. Fig.~\ref{fig:sigma_tau_002}
shows the 2D posterior distribution in the ($\sigma$,$\tau$) sub-space.
\begin{figure*}
\begin{center}
\begin{minipage}[c]{1.00\textwidth}
\centering
\includegraphics[width=8.0cm,angle=0]{./RecPk_dm_z=1.eps}
\includegraphics[width=7.6cm,angle=0]{./dRecPk_dm_z=1.eps}\\
\includegraphics[width=5cm,angle=0]{./slice_RecPk_dm_k1_z=1.eps}
\includegraphics[width=5cm,angle=0]{./slice_RecPk_dm_k5_z=1.eps}
\includegraphics[width=5cm,angle=0]{./slice_RecPk_dm_k18_z=1.eps}
\end{minipage}
\caption{Same as in Figs.~\ref{fig:re_dm}, \ref{fig:re_diff_dm} and
\ref{fig:2d_slices_z=0}, but for redshift $z=1$.
\label{fig:2d_all_methods_001}}
\end{center}
\end{figure*}
\begin{table}
\begin{center}
\begin{tabular}{|l|l|l|}
\hline
FoG model & $k_{\rm max}=0.3h^{-1}{\rm Mpc}$ & $k_{\rm max}=0.16h^{-1}{\rm Mpc}$ \\
\hline
Gaussian ($\sigma$) & $17.4^{+18.5}_{-12.2}$ & $82.6^{+82.4}_{-57.6}$ \\
Taylor ($\sigma$) & $18.6^{+18.8}_{-12.8}$ & $82.4^{+76.4}_{-57.1}$ \\
Taylor ($\sigma+\tau$) & ($240^{+18.4}_{-20.2}$, $336^{+12.6}_{-14.4}$) & ---\\
\hline
\end{tabular}
\caption{The best fit FoG parameters assuming different $k_{\rm
max}$ for $z=1$, as in Table~\ref{tb:FoG_002}. }
\label{tb:FoG_001}
\end{center}
\end{table}
Nonlinearities in matter clustering are less significant at higher
redshifts. Hence the likelihood reconstruction of power spectrum we are
studying may work better for higher
redshifts. Fig.~\ref{fig:2d_all_methods_001} shows the results using
simulations at $z=1$, similarly to Figs.~\ref{fig:re_dm},
\ref{fig:re_diff_dm}, and \ref{fig:2d_slices_z=0}. Note that we assumed
$k_{\rm max}=0.3~h{\rm Mpc}^{-1}$ as done in the $z=0$ reconstruction.
In fact the FoG effect is smaller at $z=1$ than $z=0$, e.g. as seen from
the bottom-right panel of Fig.~\ref{fig:2d_all_methods_001} compared to
Fig.~\ref{fig:2d_slices_z=0}. However, the Taylor ($\sigma+\tau$) FoG
model seems still needed in order to better recover the simulation
spectra. The bottom-right panel clearly shows that, although the FoG effect is
smaller around $\mu=0$ compared to the $z=0$ result
(Fig.~\ref{fig:2d_slices_z=0}), the Taylor ($\sigma+\tau$) model better
captures the simulation results around $\mu=\pm 1$, showing a stronger
dependence of $\mu$ than the Gaussian or Taylor ($\sigma$)
(or Lorentzian) models do.
This implies that it is important to properly include the
scale-dependent FoG effect for the power spectrum reconstruction, at
least up to $z\simeq 1$ we have studied. As given in
Table~\ref{tb:FoG_001}, a non-zero $\tau$ parameter is favored to
capture the FoG effect seen in simulations. Fig.~\ref{fig:sigma_tau_001}
shows the 2D posterior distribution in the ($\sigma$,$\tau$) sub-space,
displaying a strong correlation between the two parameters.
\begin{figure}
\begin{center}
\includegraphics[width=8.5cm,angle=0]{./pdf2d_tau-sigma_z=1.eps}
\end{center}
\caption{The posterior distribution in the ($\sigma,\tau$) parameter
space for the Taylor ($\sigma+\tau$) FoG model at $z=1$, as in
Fig.~\ref{fig:exmp_contour}.
}
\label{fig:sigma_tau_001}
\end{figure}
\subsection{Reconstruction of halo power spectra}
\label{sec:results_halos}
Now let us move on to the reconstruction of halo power spectra, which
are more relevant for a galaxy survey, using the halo catalogs
constructed from 70 simulation realizations (see \S~\ref{sec:dm_catalog}
for details). The halo clustering in redshift space is least affected by
the FoG effect, because the halos are treated as points and the redshift
distortion effect on halo clustering is caused only by their bulk
motions in large-scale structure, not by the internal virial motion
within one halo. Therefore we can naively expect a more accurate
reconstruction of the density and velocity power spectra for halos based
on the maximum likelihood method we have developed in this
paper. However, unfortunately, this is not that simple as shown below.
For halo power spectrum we need to take into account the effect of shot
noise arising from an imperfect sampling of the density fluctuation
field
due to the finite number of halos. In this case the maximum
likelihood for power spectrum reconstruction needs to be modified as
\begin{eqnarray}
-2\ln{\cal L}&=&\sum_{k_i, \mu_a}N(k_i,\mu_a)\left[
\frac{\hat{P^s}(k_i,\mu_a)-P_{\rm sn}}{P^s(k_i,\mu_a)}\right.\nonumber\\
&&\hspace{2em}
+\left.\ln \frac{P^s(k_i,\mu_a)}{\hat{P^s}(k_i,\mu_a)-P_{\rm sn}}-1
\right], \label{eq:2dmethod_halo}
\end{eqnarray}
where $P_{\rm sn}=1/\bar{n}$ is the shot noise contamination, and
$\bar{n}$ is the mean number density of halos. In the following we
simply assume that the shot noise is not a free parameter and
given by the mean number density of halos we use for the power spectrum
reconstruction: $P_{\rm sn}=1/\bar{n}$ \citep[see][for a promising
method to further suppress the shot noise
contamination]{Seljaketal:09}. The shot noise expression is not accurate
for the actually measured power spectrum \citep[e.g.][]{Smithetal:07},
but the residual shot noise, even if exists, primarily contaminates to
the spectrum that is proportional to $\mu^0$, i.e. the density-density
spectrum $P_{\delta\delta}$.
However the main obstacle we have faced is that we cannot reliably
measure the velocity field of halos (see \S~\ref{sec:halo_spectra} for
details) and cannot therefore have the velocity-related power spectra,
which are needed to assess the performance of our reconstruction method
by comparing with the reconstructed spectra $P_{\delta\delta}$ and $P_{\delta\theta}$. Rather
we decided to use the dark matter (N-body particles) velocity field
instead of estimating the halo velocity field, assuming that the halo
bulk-velocity field is unbiased from the matter velocity field, which is
often assumed in the literature.
Hence, before going to the halo spectrum reconstruction, we make a
simple test to study whether or not the power spectrum reconstruction is
affected by the shot noise contamination. This test can be done
by applying our reconstruction method to the catalogs with reduced
N-body particles.
To be more precise we randomly select N-body particles from
each simulation realization ($z=0$) until the number density of
particles selected becomes the same to the density of halo catalogs,
$\bar{n}\simeq 3.8\times10^{-4} h^{3}{\rm
Mpc}^{-3}$.
Then by using the reduced N-body particles in each
simulation we compute the redshift-space power spectrum taking into
account redshift modulation due to the velocity field of each
particle. These procedures preserve the underlying spectra of $P_{\delta\delta}$ and
$P_{\delta\theta}$. Thus we can compare the spectra with the spectra reconstructed
by applying the maximum likelihood method to the redshift-space spectrum
of reduced N-body particles, where the shot noise contamination is
subtracted from the measured spectrum according to
Eq.~(\ref{eq:2dmethod_halo}).
Fig.~\ref{fig:2d_all_methods_002_reduce} shows the reconstruction
results (symbols in each panels) for the catalogs of reduced N-body
particles. The directly measured spectra in the left panel are similar
to the curves in Fig.~\ref{fig:re_dm}, although we found a small
difference in the directly measured $P_{\delta\delta}(k)$ at $k\lower.5ex\hbox{$\; \buildrel > \over \sim \;$} 0.2~h{\rm
Mpc}^{-1}$ due to the residual shot noise effect
\citep{Smithetal:07}. Fig.~\ref{fig:2d_all_methods_002_reduce}
clearly shows that, even in the presence
of shot noise, our reconstruction method recovers the power spectra to a
similar precision to the results in Figs.~\ref{fig:re_dm},
\ref{fig:re_diff_dm} and \ref{fig:2d_slices_z=0}. Hence we conclude that
the shot noise is not a serious source of systematics for our method.
\begin{figure*}
\begin{center}
\begin{minipage}[c]{1.00\textwidth}
\centering
\includegraphics[width=5.8cm,angle=0]{./RecPk_reduced-dm.eps}
\includegraphics[width=5.8cm,angle=0]{./dRecPk_reduced-dm.eps}
\includegraphics[width=5.8cm,angle=0]{./slice_RecPk_reduced-dm_k18.eps}
\end{minipage}
\caption{Reconstruction results of the power spectra for the reduced
N-body particle distribution as in Fig.~\ref{fig:2d_all_methods_001},
where a smaller number of N-body particles are randomly selected in each
simulation realization in such a way that the mean number density of the
resulting particles becomes comparable to that of halo catalogs we will
use below (see \S~\ref{sec:dm_catalog}): $\bar{n}=3.8\times
10^{-4}~h^{3}{\rm Mpc}^{-3}$.
In this case the shot noise term arising from
a finite number of the sampled N-body particles affects the
redshift-space power spectrum measurement.
We applied the power spectrum reconstruction method
(Eq.~[\ref{eq:2dmethod}]) to the
redshift-space power spectrum after simply subtracting the expected shot noise
term $1/\bar{n}$ from the measured spectrum.
The dotted curves in the left and right panels and the spectrum in
denominator in the middle panel are the input power spectrum, which are
the same as the spectra of original N-body particles in
Figs.~\ref{fig:re_dm}, \ref{fig:re_diff_dm} and \ref{fig:2d_slices_z=0}.
It is found that our method nicely recovers the input power spectra
even in the presence of shot noise contamination.
\label{fig:2d_all_methods_002_reduce}}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\begin{minipage}[c]{1.00\textwidth}
\centering
\includegraphics[width=8.0cm,angle=0]{./RecPk_halo.eps}
\includegraphics[width=7.6cm,angle=0]{./dRecPk_halo.eps}
\includegraphics[width=5cm,angle=0]{./slice_RecPk_halo_k1.eps}
\includegraphics[width=5cm,angle=0]{./slice_RecPk_halo_k5.eps}
\includegraphics[width=5cm,angle=0]{./slice_RecPk_halo_k18.eps}
\end{minipage}
\caption{Same as in Figs.~\ref{fig:re_dm}, \ref{fig:re_diff_dm} and
\ref{fig:2d_slices_z=0}, but for halo spectra at $z=0$. As described in
\S~\ref{sec:halo_spectra}, the input density-velocity spectra $P_{\delta\theta}$,
which are used to compare with the reconstructed spectra, we used the
spectra between the halo density field and the N-body particle velocity
field, because the halo velocity field is hard to construct due to a
coarse sampling of the velocity field. As can be clearly seen from the
lower three panels (especially two lower-right panels), the measured
redshift-space power spectra show greater amplitudes
than the redshift-space
spectrum inferred from the simulation, without the FoG effect
$F=1$ (see text for discussion). The dashed curves in the upper-left and
-right panels show the results obtained by adding the perturbation
theory prediction of the nonlinearity correction term
(Eq.~[\ref{eq:higher_Ps_halo_text}]) to the directly measured
$P_{\delta\theta}(k)$. In the predictions, we assumed the halo bias parameters
$b_1=1.6$ and $b_2=0$, where the linear bias parameter is estimated by
comparing the density power spectra ($P_{\delta\delta}$) of dark matter and halos
at small $k$.
\label{fig:2d_all_methods_halo002}}
\end{center}
\end{figure*}
\begin{figure}
\includegraphics[width=8.7cm,angle=0]{./dRecPdv_compare-halobias.eps}
\caption{Ratios of the reconstructed spectra of $\mu^2$
to the directly measured $P_{\delta\theta}(k)$
(as in
Fig.~\ref{fig:2d_all_methods_halo002}), for different halo catalogs.
The two results are slightly shifted in a horizontal direction for
illustrative purpose. The two different halo catalogs are defined from
halos with masses greater than $9.8\times 10^{12}~h^{-1}M_\odot$
(triangles) and $1.86\times 10^{13}~h^{-1}M_\odot$ (squares), which
contain at least 20 and 38 N-body particles as members,
respectively. For more massive halos, the reconstructed spectrum of
$\mu^2$ shows greater amplitudes. The solid and dashed curves show the
results including the nonlinearity correction term to $P_{\delta\theta}(k)$ as done
in Fig.~\ref{fig:2d_all_methods_halo002}: $P_{\delta\theta}(k)+\delta
P_{\mu^2}(k)$. To compute the nonlinearity corrections,
we assumed $b_1=1.6$
and 1.84 for the less and more massive halo catalogs, respectively, but
assumed $b_2=0$ for both the catalogs.
\label{fig:halo_biastest}}
\end{figure}
Now we move to the reconstruction of halo power spectra.
Fig~\ref{fig:2d_all_methods_halo002} shows the results for halo catalogs
at $z=0$.
First of all, the reconstruction
can successfully recover the density power spectrum $P_{\delta\delta}(k)$ over a
range of wavenumbers we consider, as a result of properly correcting for
the shot noise and the redshift distortion.
The accurate reconstruction of $P_{\delta\delta}$ is relevant for the BAO
experiments, and the results imply that our method may allow us to
further use the broad-band shape of $P_{\delta\delta}(k)$ to improve cosmological
constraints. However, in contrast to the results for N-body particles
shown in Figs.~\ref{fig:re_dm} and \ref{fig:2d_all_methods_002_reduce},
the reconstruction fails to recover the density-velocity power spectrum
$P_{\delta\theta}(k)$. The reconstructed $P_{\delta\theta}$ for halos gives higher amplitudes
than the directly measured power spectrum, irrespective of the different
FoG models.
In fact, as explicitly shown
in the right panel, the measured redshift-space power spectrum shows
{\em greater} amplitudes than
predicted by the Kaiser formula
(Eq.~[\ref{eq:Pgg}] with no FoG effect, i.e. $F=1$). The enhancement in
the power spectrum amplitudes is opposite to the FoG effect, which
always suppresses the amplitudes.
We argue below that the results in Fig.~\ref{fig:2d_all_methods_halo002}
can be understood by the nonlinearity effect on the redshift-space power
spectrum. As we briefly discussed around Eq.~(\ref{eq:higher_Ps}), the
nonlinear clustering causes a correction to the Kaiser formula of
redshift-space power spectrum \citep[also see][for a more extensive
discussion]{Scoccimarro04,Taruyaetal:10}. Assuming that the density
perturbation is greater than the velocity field, which can be even more
validated for highly biased halos with $b>1$, the leading-order
correction term is found to arise from the
cross-bispectrum $\langle
\tilde{\delta}\tilde{\delta}\tilde{\theta}\rangle$ (see
Eq.~[\ref{eq:higher_Ps}]):
\begin{equation}
\delta P_s(k,\mu)\leftarrow
k_\parallel\left\langle\tilde{\delta}(\bmf{k}')
\int\!\frac{d^3\bmf{q}}{(2\pi)^3}\frac{q_\parallel}{q^2}\tilde{\theta}(\bmf{q})
\tilde{\delta}(\bmf{k}-\bmf{q})
\right
\rangle.
\end{equation}
We tried to measure this correction term from the simulations, but could
not obtain the reliable results as the bispectrum measurement is very
noisy. Instead we here use the perturbation theory prediction assuming a
$\Lambda$CDM cosmology (or equivalently Einstein gravity). In
Appendix~\ref{sec:higher_Kaiser} we explicitly derive the leading-order
correction term given as a function of the linear mass power
spectrum. We find that the leading-order correction only contributes to
the redshift-space power spectrum at the power of $\mu^2$:
\begin{equation}
P_{\rm halo}^s(k,\mu)=P_{\delta\delta}(k)+2\mu^2\left[P_{\delta\theta}(k) + \delta P_{\mu^2}(k)\right]
+\mu^4P_{\theta\theta}(k),
\label{eq:phh}
\end{equation}
where $\delta P_{\mu^2}(k)$ is the correction term. Including the halo
bias parameters, the correction term is expressed as
\begin{eqnarray}
&&\hspace{-2em}\delta P_{\mu^2}(k) =\frac{fb_1^2k^3}{(2\pi)^2}\left[
\int_0^\infty\!dr
\int_{-1}^{1}dx~ x\right.
\nonumber\\
&&\hspace{0em}
\times \left\{ \frac{r^3}{7}(-1+7rx-6x^2)
P_{\delta\delta}^L(k)
\right.\nonumber\\
&&\hspace{1em}
\left.
+\frac{1}{7}(7x+3r-10rx^2)P_{\delta\delta}^L(kr)
\right\}
\frac{P^L_{\delta\delta}(k\sqrt{1+r^2-2rx})}{(1+r^2-2rx)}
\nonumber\\
&&\hspace{1em}
\left.
-P_{\rm \delta\delta}^L(k)\int_0^\infty\!dr
\frac{2}{3}(1+r^2)P_{\delta\delta}^L(kr)
\right]\nonumber\\
&&+
\frac{fb_1b_2k^3}{(2\pi)^2}\int_0^\infty\!dr\int_{-1}^{1}
\!dx~ xr P_{\delta\delta}^L(kr)P_{\delta\delta}^L(k\sqrt{1+r^2-2rx}),
\nonumber\\
\label{eq:higher_Ps_halo_text}
\end{eqnarray}
where $f\equiv d\ln D/d\ln a$ ($D$ is the growth rate) and $b_1$ and
$b_2$ are the linear and nonlinear bias parameters. Note that we here
assumed that the velocity field of halos is unbiased with respect to the
velocity field of dark matter.
Eq.~(\ref{eq:higher_Ps_halo_text}) clearly shows that the nonlinearity
correction term depends on halo bias parameters. The first term depends
on the linear bias parameter as $\propto b_1^2$. Compared to the
density-velocity power spectrum $P_{\delta\theta}(k)$, which depends on $b_1$ as
$P_{\delta\theta} \propto b_1$, the nonlinearity correction term can be more
important for more biased halos, or equivalently more massive halos.
\cite{Taruyaetal:10} derived more comprehensive equations for dark
matter including other nonlinear terms which arise from the
cross-bispectra such as
$\langle\tilde{\delta}\tilde{\theta}\tilde{\theta} \rangle$. The other
terms are found to have the contributions of $\mu^{2n}$ ($n=1,2,\cdots
4$), but depend on halo bias as $\propto b_1$. For highly biased halos
with $b_1>1$ as halos we are studying, the term given by
Eq.~(\ref{eq:higher_Ps_halo_text}) has most dominant
contribution. However, the perturbation theory is known to be less
accurate for lower redshifts such as $z=0$, due to the stronger
nonlinear clustering effects. Hence the results shown here still need to
be more carefully studied.
The dashed curves in the upper-left and -right panels of
Fig.~\ref{fig:2d_all_methods_halo002} show the results where we added
the nonlinearity correction term to $P_{\delta\theta}(k)$ measured from the
simulation assuming the Taylor $(\sigma+\tau)$ FoG model:
$P_{\delta\theta}(k)+\delta P_{\mu^2}(k)$. To obtain the theory prediction of the
nonlinearity correction term, we assumed
$b_1=1.6$ and $b_2=0$, where $b_1$ is estimated by comparing the density
power spectra $P_{\delta\delta}$ for dark matter and halos at small $k$. More
exactly we computed the correction power spectrum of $\mu^2$ using the
full expression in \cite{Taruyaetal:10}, also taking into account the
halo bias dependences on the different terms that are either
proportional to $b_1$ or $b_1^2$. Eq.~(\ref{eq:higher_Ps_halo_text})
gives the similar shape, but about 10\% higher amplitudes than the
curves in Fig.~\ref{fig:2d_all_methods_halo002}. Interestingly, the
nonlinearity correction term increases, rather than suppresses, the
amplitudes of real-space power spectrum that is proportional to $\mu^2$
in the redshift-space power spectrum. The enhancement increases with
increasing $k$. We should also emphasize that such a nice agreement
including the nonlinearity correction can be found only if using the
Taylor-($\sigma+\tau$) FoG model. Given the fact that the redshift-space
power spectrum of halos is least affected by the FoG effect, our results
imply that the Taylor ($\sigma+\tau$) model has more degrees of freedom
than other one-parameter FoG models and can effectively capture
higher-order contributions of $\mu^{2n}$ ($n=1,2,\cdots$) that arise
from nonlinearity effects as studied in \cite{Taruyaetal:10}.
We also studied the halo power spectrum reconstruction using the halo
catalogs constructed from $z=1$ simulations. We similarly found that the
reconstructed power spectra of $\mu^2$ show greater amplitudes than
expected from the measured $P_{\delta\theta}(k)$. The nonlinearity correction term
is found to similarly explain the reconstructed power spectrum, in
slightly less agreement, if using the reconstructed power spectrum
obtained from the Taylor $(\sigma+\tau)$ FoG model. We
have also found a subtle contamination of the residual shot noise for
the $z=1$ results, and
therefore we here show the results for $z=0$ for illustrative clarity.
To obtain more insights on the halo power spectrum results, in
Fig.~\ref{fig:halo_biastest} we study the reconstruction results for
$P_{\delta\theta}$ using different halo catalogs where halos are selected with
different mass thresholds.
To be more precise, we made the new catalogs by employing higher mass
threshold, $1.86\times10^{13}h^{-1}M_{\sun}$ (including more than 38
N-body member particles), rather than the threshold
$9.8\times10^{12}h^{-1}M_{\sun}$ (20 particles) we have so far
used. Note that the number density for more massive halos is
$\bar{n}\simeq 1.9\times 10^{-4}~h^3{\rm Mpc}^{-3}$, compared to
$3.8\times 10^{-4}~h^3{\rm Mpc}^{-3}$ for our fiducial halo
catalogs. The estimated bias parameter is $b_1=1.84$ compared to
$b_1=1.6$. Fig.~\ref{fig:halo_biastest} shows the reconstruction power
spectrum of $\mu^2$ for
the different halo catalogs, compared to the density-velocity
spectrum.
The figure shows that the reconstructed spectrum for more massive halos
has higher amplitudes than for less massive halos. The solid and dotted
curves show the predictions obtained by adding the nonlinearity
correction term (Eq.~[\ref{eq:higher_Ps_halo_text}]) to the directly
measured $P_{\delta\theta}(k)$, where we used in the computation the linear bias
parameters above and assumed $b_2=0$ for simplicity.
The nonlinearity correction, which depends on the halo bias,
fairly well reproduces the reconstruction
results. In summary such nonlinearity correction terms need to be
included when interpreting the reconstructed power spectra for halos, or
more generally galaxies. We again emphasize that our method reconstructs
band powers of the real-space power spectrum, which are proportional to $\mu^2$ in the
redshift-space power spectrum, rather than the band powers of $P_{\delta\theta}$
alone.
Finally we comment on the impact of nonlinearity correction on the
reconstructions results for dark matter (N-body particles), which we
showed in the preceding section. For the dark matter spectrum, which
has $b=1$ by definition, the nonlinearity correction
(Eq.~[\ref{eq:higher_Ps_halo_text}]) is smaller compared to the
results of halos. However, using the perturbation theory predictions, we
found that the nonlinearity correction is not negligible. Including the
nonlinearity correction improves agreement with the input power
spectrum over a range of wavenumbers up to $ k \simeq 0.2~h$Mpc$^{-1}$
for the
Taylor-($\sigma+\tau$) FoG results shown in the middle panel of
Fig.~\ref{fig:re_diff_dm}. However, the nonlinearity correction
increases the disagreement at the larger $k$. In summary we conclude
that our reconstruction method can well recover the real-space power
spectrum, which is proportional to $\mu^2$ in the redshift-space power
spectrum, up to $k\simeq 0.2~h$Mpc$^{-1}$ for both dark matter and
halos, {\em if} including the nonlinearity corrections.
\section{Summary and Discussion}
\label{sec:summary}
In this paper we have developed a maximum likelihood based method of
reconstructing the real-space power spectra of density and velocity
fields, from the two-dimensional, redshift-space clustering of dark
matter and halos (supposedly galaxies). This method is developed in
analogy with the CMB power spectrum reconstruction method
\citep{Verde03}.
By assuming the form of redshift-space power spectrum given by
Eq.~(\ref{eq:2dmethod}), we developed a method of reconstructing the
band powers of $P_{\delta\delta}$ and $P_{\delta\theta}$ at each $k$ bins, being marginalized
over uncertainties in the band powers at different $k$ bins and the
parameters to model the FoG effect, in such a way that the
likelihood of the redshift-space power spectrum measured becomes
maximized. One assumption we have employed for the method is the
functional form of redshift-space power spectrum
(Eq.~[\ref{eq:2dmethod}]), where the Kaiser formula and the FoG effect
is given by multiplicative functions. In fact this form is expected
based on the halo model picture \citep[][also see Hikage et al. in
preparation]{White:01,Seljak:01}.
The real-space power spectra, especially at such large length scales
($k\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 0.3~h^{-1}{\rm Mpc}$), contains cleaner cosmological information in the
linear or quasi-nonlinear regimes, and are relatively easier to develop
a sufficiently accurate model by using a suit of
simulations and/or refined perturbation theory.
Furthermore, by measuring the velocity-related power
spectra in a model-independent way,
we
can open up a new window of testing gravity on cosmological scales. That
is, we can address whether or not the velocity field inferred is
consistent with the gravity field inferred from the density field,
because the density and velocity fields are related to each other via
gravity theory.
We have carefully tested our method by comparing the reconstructed
real-space power spectra with the spectra directly measured from
simulations of 70 realizations, for dark matter (N-body particles) as
well as halos. For matter power spectra (i.e. N-body particles), we
showed our method nicely recovers the power spectra $P_{\delta\delta}$ and $P_{\delta\theta}$
over a range of scales $k\lower.5ex\hbox{$\; \buildrel < \over \sim \;$} 0.3~h^{-1}{\rm Mpc}$ and at redshifts $z=0$ and
$z=1$ (see Figs.~\ref{fig:re_dm}, \ref{fig:re_diff_dm} and
\ref{fig:2d_all_methods_001}), to accuracies within the statistical
errors, {\em if} we use the Taylor ($\sigma+\tau$) FoG model (see
Eq.~[\ref{eq:FOG}]), which has more degrees of freedom (2 parameters)
than the other models, Gaussian, Lorentzian and single-parameter Taylor
FoG models. Our results imply that the FoG effect seen in simulations
has a complex scale-dependence, and is important to take into account
the scale dependence in order to obtain an unbiased reconstruction of the
band powers of $P_{\delta\delta}$ and $P_{\delta\theta}$ at scales down to
$k\simeq0.3h^{-1}{\rm Mpc}$. In other words, the FoG effect affects the
redshift-space power spectrum over a wide range of wavenumbers. Hence an
inaccurate modeling of the FoG effect causes a biased estimate of the
power spectra.
It is also worth noting that the reconstruction causes
correlations between the band powers of different power spectra and at
different $k$-bins and the FoG model parameters (see
Figs.~\ref{fig:exmp_contour} and \ref{fig:2d_all_methods_001}).
For the halo power spectrum, we showed that our method again nicely
recovers the density power spectrum $P_{\delta\delta}$ over a wide range of
wavelengths, up to $k\simeq 0.3~h{\rm Mpc}^{-1}$
(see Fig.~\ref{fig:2d_all_methods_halo002}). Such an
accurate reconstruction of $P_{\delta\delta}$ is very promising, because the shape
and amplitude information of $P_{\delta\delta}$ are sensitive to cosmological
parameters such as the tilt and running index of the primordial power
spectrum and neutrino masses
\citep[e.g.][]{Takadaetal:06,Saitoetal:09,Saitoetal:10}. On a
measurement side, the halo power spectrum can be estimated from actual
galaxy redshift survey, e.g. based on the method developed in
\cite{Reidetal:10} where galaxy pairs with small spatial separations are
clipped out. Although the halo power spectrum is supposed to be less
contaminated by the FoG effect, it is very important to minimize the
residual FoG contamination in order to extract unbiased cosmological
information from the measured halo power spectrum. For example, since
the FoG effect causes a suppression in the power spectrum amplitude, a
residual FoG contamination would cause a bias in neutrino mass
constraints because the main effect of massive neutrinos is also the
suppression on power spectrum amplitude (Hikage et al. in preparation).
Our method can give a robust way of measuring the density power
spectrum, minimizing the FoG contamination.
For the halo velocity power spectrum $P_{\delta\theta}$, we found some
difficulty. First of all, we could not reliably reconstruct the velocity
field of halos from simulations, due to too sparse sampling of halos'
velocities. Hence we instead used the velocity field of dark matter
(N-body particles) assuming that the large-scale bulk motions of halos
are unbiased from the velocities of dark matter, which has been often
assumed in the literature. Note that the {\em real-space} velocity field
we considered here contains only the large-scale information at $k\lower.5ex\hbox{$\; \buildrel < \over \sim \;$}
0.3~h{\rm Mpc}^{-1}$, and therefore is not affected by any virial
motions within halos. As a result, we found that the reconstructed power
spectrum of $\mu^2$ systematically differs from $P_{\delta\theta}(k)$ directly
measured from the simulations
(Fig.~\ref{fig:2d_all_methods_halo002}). In fact, the measured
redshift-space halo power spectrum shows greater amplitudes than the
spectrum inferred from the Kaiser formula of redshift-space spectrum,
without the FoG effect that causes a suppression in the redshift-space
power spectrum amplitudes.
Therefore we argued that the halo power spectrum is affected by the
nonlinearity effect.
In Appendix~\ref{sec:higher_Kaiser}, assuming a $\Lambda$CDM cosmology
or Einstein gravity, we derived the leading-order nonlinearity correction
to the Kaiser formula of redshift-space power spectrum, which arises
from the cross-bispectrum
between the density and velocity
perturbations. We meant by the leading-order term that the term appears
to have
the largest contribution, assuming that the density perturbation is
greater than the velocity field at length scales of interest.
We found
that the leading-order contribution is proportional to $\mu^2$ and has
greater amplitudes for more biased halos, i.e. more massive halos. We
showed that adding the perturbation theory prediction to the simulation
$P_{\delta\theta}(k)$ better matches the reconstructed power spectrum (see
Fig.~\ref{fig:2d_all_methods_halo002}). We also found that, by using the
different halo catalogs defined with different mass thresholds,
the halo bias
dependence of the nonlinearity correction is seen in the
reconstructed power spectra of $\mu^2$
(see
Fig.~\ref{fig:halo_biastest}).
Hence a more appropriate statement for our maximum likelihood method is
that the method can recover the real-space power spectra, which are
proportional to $\mu^0$ and $\mu^2$, respectively, in the measured
redshift-space power spectrum, including marginalization over
uncertainties in the FoG effect. In other words the reconstructed
spectrum of $\mu^2$ is not necessarily the same as the density-velocity
power spectrum $P_{\delta\theta}(k)$, which we have used in our comparison. The
nonlinearity effect on the real-space power spectrum of $\mu^2$ needs to
be included if we want to use the reconstructed power spectrum to
constrain cosmological parameters as well as to test
gravity theory. On the other hand, we
found that the power spectrum of $\mu^4$ is very noisy to reconstruct,
for the ranges of wavenumbers and redshifts we have considered in this
paper.
Recently \cite{Taruyaetal:10} studied the redshift-space power spectrum
including nonlinearity effects, based on the extended perturbation
theory. They found that the nonlinear correction terms including the
higher-order terms of $O(\theta^2)$ have the contributions that are
proportional to $\mu^{2n}$ ($n=1,2,\dots, 4$) in the redshift-space
power spectrum. The comparison of the theoretical prediction with the
reconstructed power spectrum based on our method is very interesting,
and will be studied elsewhere.
One encouraging result is our method can unbiasedly recover the
real-space density power spectrum $P_{\delta\delta}(k)$ even in the presence of
redshift distortion effect. Given this result our method may offer even
a new means of obtaining geometrical constraints on the Hubble expansion
rate and the angular diameter distances beyond the usual BAO
constraints. We have assumed throughout this paper that the underlying
cosmology is known. However, this is obviously not true for an actual
observation. In reality, we have to assume a reference cosmological
model to perform the clustering analysis of galaxies, and the assumed
cosmology generally differs from the underlying true cosmology. An
imperfect cosmological model causes additional angular anisotropies in
the measured redshift-space power spectrum -- the so-called cosmological
distortion. In terms of Eq.~(\ref{eq:2dmethod}) an incorrect cosmology
leads some power of the density power spectrum $P_{\delta\delta}(k)$ of $\mu^0$ to
leak into the power spectrum with powers higher than $\mu^2$ in the
redshift-space power spectrum. Contrary, if we seek the reconstructed
$P_{\delta\delta}$ of maximum amplitudes with varying reference cosmological models,
we may be able to obtain the cosmological constraints \citep[also
see][for a similar discussion]{Padmanabhan:08}. This method looks
similar to the Alcock-Paczynski test
\citep{AP,MatsubaraSuto:96,Ballingeretal:96}, but our method may have
practical advantages: our method allows us to measure the real-space
power spectrum of $\mu^0$ in a model-independent way as well as to
derive cosmological constraints being marginalized over uncertainties in
the FoG effect. The feasibility of this method is our future work and
will be presented elsewhere.
Our reconstruction method is done in the two-dimensional Fourier space
of $(k,\mu)$. In practice one may want to exclude the Fourier modes
around $\mu\simeq \pm 1$, which are more affected by the FoG effect. In
our method it is straightforward to include a masking of the modes
around $\mu\pm 1$; that is, the real-space power spectra are
reconstructed by using the Fourier modes in redshift space, excluding
the modes around $\mu\pm 1$. We have tried several masking methods of
$\mu$,
but
could not find any significant differences from the results shown in
this paper.
In this paper we have ignored some observational effects for
simplicity. For example, to apply our method to actual data, we need to
include effects such as survey window function and the curvature of the
sky. These effects have been well studied \citep[e.g.][]{Tegmarketal:04},
and would be rather straightforward to include, although a further
careful study needs to be done.
\section*{Acknowledgments}
We thank E. Komatsu, B. Jain, T. Nishimichi, N. Padmanabhan, R. Sheth,
D. Spergel and A. Taruya for useful discussion and valuable comments.
This work is in part supported in part by JSPS Core-to-Core Program
``International Research Network for Dark Energy'', by Grant-in-Aid for
Scientific Research from the JSPS Promotion of Science, by Grant-in-Aid
for Scientific Research on Priority Areas No. 467 ``Probing the Dark
Energy through an Extremely Wide \& Deep Survey with Subaru Telescope'',
by World Premier International Research Center Initiative (WPI
Initiative), MEXT, Japan, and by the FIRST program ``Subaru Measurements
of Images and Redshifts (SuMRe)''.
|
\section{Introduction}
Over the past two decades, the Lyman-$\alpha$ (Ly$\alpha$) forest in the line-of-sight
to distant quasars has emerged as one of the
most important probes of the high-redshift ($ z > 2$) universe
\citep[see, e.g.,][]{croft+98,mcd+00,croft+02,zald+03,mcd+05}.
This has been underpinned by theoretical advances
that established the Ly$\alpha$\ forest as arising from neutral hydrogen
embedded in a warm
photo-ionized inter-galactic medium (IGM), tracing the density fluctuations due
to gravitational instability in hierarchical clustering cosmological models
\citep[see, e.g.,][]{cen+94,mirald+96,dave+99,theuns+98} .
In recent years, the Ly$\alpha$\ forest is increasingly being used
to gain a more detailed understanding of the inter-galactic medium.
In particular, the thermal history of the IGM holds
the key to understanding hydrogen and \ion{He}{2}\ reionization at
$z > 6$ and $z \sim 3$, respectively.
For an underlying density distribution, $\Delta(x)=\rho(x) / \bar{\rho}$,
the astrophysics of the IGM mediates
the optical depth of the Ly$\alpha$\ forest.
In the standard photoionization equilibrium model of the Ly$\alpha$\ forest \citep{gp65},
the thermal properties of the IGM
are usually described in terms of 3 main parameters: the ionization rate of the
photoionizing background radiation, $\Gamma$, the temperature of the gas at
mean density, $T_0$, and the temperature-density
relationship\footnote{The
temperature-density relationship of the IGM is often called the `equation of
state'.
This is technically incorrect as the equation of state of the IGM is that of a perfect gas,
thus in this Letter we do not use this term.}
approximated as a power-law, $T = T_0 \Delta^{\gamma - 1}$.
Various authors have placed constraints on the
background ionization rate, $\Gamma$, of the IGM
using the effective optical depth, $\tau_\mathrm{eff}$, of
the Ly$\alpha$\ forest \citep{bolton+05,fg+08b} and the
quasar proximity effect \citep{scott+00,dall+09}.
Constraints on $T_0$ have been made using detailed
line-profile fitting of individual Ly$\alpha$\ absorption lines
from high-resolution spectra \citep[see, e.g.,][]{schaye+00,becker+11}.
Meanwhile, the probability
distribution function (PDF) of the transmitted Ly$\alpha$\ forest flux \citep{jen+ost91}
has been used to place constraints
on the temperature-density relation\ of the IGM.
Using this technique, \citet{becker+07}, \citet{bolton+08}, and \citet{viel+09}
have recently found evidence of a highly inverted
temperature-density relation, $\gamma \sim 0.5$, in the Ly$\alpha$\ forest at $z \sim 2-3$.
While some theories \citep[e.g.,][]{furl+oh08} predict a mildly-inverted
temperature-density relation ($\gamma \approx 0.8$)
as a consequence of inside-out \ion{He}{2}\ reionization,
the amount of energy that needs to be injected into the IGM to obtain $\gamma \sim 0.5$
is inconsistent with the observational constraints on heating sources that those redshifts \citep{mcquinn+09}.
Although various systematics such as continuum fitting
errors, noise and metal line contamination can bias the interpretation of the flux PDF,
high-resolution ($R \equiv \lambda / \Delta \lambda \sim 10^4$) and high signal-to-noise
($\ensuremath{\mathrm{S/N}} \gtrsim 50$ per pixel) Ly$\alpha$\ forest spectra can ameliorate these
effects \citep[see, e.g.,][]{kim+07}.
With such data, metal lines in the Ly$\alpha$\ forest region can be directly identified and removed, while
the low pixel noise allows precise determinations of the continuum from
the peaks of the observed Ly$\alpha$\ transmission, with random errors as low as $1-2 \%$.
However, depending on the underlying properties of the IGM,
a uniform Gunn-Peterson absorption component\footnote{While this
term may be reminiscent of the obsolete picture
of the IGM as a two-phase medium consisting of
dense and cool Ly$\alpha$\ clouds embedded in a hot, tenuous inter-cloud medium
\citep{sargent+80}, in this paper we merely use this term to refer generically to
a uniform or large-scale low-absorption component in the Ly$\alpha$\ forest transmission field.}
could exist in
the Ly$\alpha$\ forest even at relatively low redshifts ($z \sim 3$).
In such a case, a continuum fitted to the transmission peaks of the observed Ly$\alpha$\ forest flux
could underestimate the continuum, since few of the peaks would
reach the true quasar continuum.
This \emph{systematic} continuum bias could exist even in high-\ensuremath{\mathrm{S/N}}\ spectra which have small
\emph{random} errors in the pixel fluxes.
Workers in this field are aware of this possibility:
\citet{bolton+08} tested this by fitting their flux PDFs with the continuum
raised by 1.5\% and 5\%, but found that these provided worse fits to their data.
\citet{becker+07} and \citet{viel+09} made the
continuum level a free parameter in their likelihood analysis, although neither
found a significant continuum bias.
However, none of the aforementioned studies tried to
directly estimate the amount of continuum bias at low-redshift ($z \lesssim 3-4$),
nor is it well-understood
how such a bias could affect measurements of the IGM temperature-density relation.
In a paper measuring the effective optical depth $\tau_{\mathrm{eff}}$ of the Ly$\alpha$\ forest,
\citet{fg+08} did attempt to constrain the amount of bias in their continuum fits
by hand-fitting mock spectra derived from numerical simulations.
They found that even at $z = 3$, continua fitted to the peaks of the transmission
underestimate the level of the
continuum by 5\%, although they assumed an extreme value of $\gamma = 1.6$ in their
mock spectra.
While this was the temperature-density relation calculated for a relaxed IGM following
hydrogen reionization \citep{hui+gned97},
it is not expected to be valid during the epoch of \ion{He}{2}\ reionization at
$z \lesssim 3$ \citep{furl+oh08, mcquinn+09}.
Since $\tau_{\mathrm{eff}} = - \ln( \langle F \rangle)$, a systematic bias in the estimated continuum level
leads to errors in $\tau_{\mathrm{eff}}$ at approximately the same level
(for example, a 2\% underestimate in the continuum would lead to roughly a 2\% overestimate
in $\tau_{\mathrm{eff}}$).
In contrast, such continuum errors would bias the estimated temperature-density relation
in a more complicated manner, which is not well-understood.
This short paper is intended to investigate this bias in a simple, easily-reproducible,
manner.
In Section~\ref{sec:simple} we first make a simple back-of-the-envelope calculation which
indicates that systematic biases in the continuum fitting of even a few percent
could cause large errors in the estimated temperature-density relation.
Section~\ref{sec:analysis} then introduces a simple semi-analytic model of the
Ly$\alpha$\ forest flux PDF to quantify this error, followed by a
discussion (Section~\ref{sec:conclusion}) of these biases
and some methods to correct for them.
\section{Simple Estimates} \label{sec:simple}
In this section we make an order-of-magnitude
calculation of the bias
in the IGM temperature-density relation that could potentially
arise from systematic errors in the placement of the quasar continuum level.
We first define the fractional continuum error,
\begin{equation}\label{fc_def}
f_c = \frac{\ensuremath{C_{\mathrm{est}}} - \ensuremath{C_{\mathrm{true}}}}{\ensuremath{C_{\mathrm{true}}}},
\end{equation}
where \ensuremath{C_{\mathrm{true}}}\ is the true quasar continuum in the Ly$\alpha$\ forest, while \ensuremath{C_{\mathrm{est}}}\ is the
estimated continuum.
The flux transmission through the IGM $F$ is related to the intervening
optical depth $\tau$ by $F = \exp(-\tau)$, while $\tau$ is in turn
related to the underlying matter distribution,
$\Delta(x) \equiv \rho(x) / \bar{\rho}$.
For this calculation we use the fluctuating Gunn-Peterson approximation
\citep[FGPA, see, e.g.,][]{croft+98},
\begin{equation} \label{eq:fgpa}
\tau = \tau_0 \ensuremath{\Delta}^{2-\alpha},
\end{equation}
where $\ensuremath{\Delta} \equiv \ensuremath{\Delta}(x)$, $\alpha = 0.7(\gamma-1)$,
and $\tau_0$ is a factor which includes astrophysics such as the
Ly$\alpha$\ recombination coefficient and the photoionizing UV background $\Gamma$.
Here we assume $\tau_0$ to be constant, so that the only unknown parameter
of the IGM is the exponent $\alpha$ in the equation of state.
We rearrange Equation~\ref{eq:fgpa} to
\begin{equation}
\alpha = \frac{\ln \tau_0 - \ln \tau}{\ln \ensuremath{\Delta}} + 2.
\end{equation}
The derivative of $\alpha$ with respect to $\tau$ is then
\begin{equation}
\frac{d\alpha}{d\tau} = - \frac{1}{\tau \ln \ensuremath{\Delta}}.
\end{equation}
This allows us to estimate the error
in the derived equation of state,
$\delta \alpha \equiv \alpha^{\mathrm{est}} -\alpha^{\mathrm{true}} $
(the superscripts
`est' and `true' refer to the estimated
and true underlying quantities respectively),
arising from a systematic error in the measured optical depth
$\delta \tau $ for a given matter overdensity $\ensuremath{\Delta}$:
\begin{eqnarray}\label{eq:dalpha}
\delta \alpha &\sim& -\frac{1}{\ln \ensuremath{\Delta}} \frac{\delta \tau}{\tau} \nonumber \\
&\sim& \frac{\delta \tau}{\tau}.
\end{eqnarray}
In the second line, we have approximated $-\ln \ensuremath{\Delta} $ to be of order unity,
since the low-absorption regions of the IGM are in the underdense
($\Delta < 1$) parts of the universe.
Now consider the effects of a misunderestimated quasar continuum level
\ensuremath{C_{\mathrm{est}}}, where \ensuremath{C_{\mathrm{true}}}\ is the true continuum level.
\citet{fg+08} have shown that the true optical depth \ensuremath{\tau^{\mathrm{true}}}\ is related to the measured
optical depth corresponding to the underestimated continuum, \ensuremath{\tau^{\mathrm{est}}}\ by
\begin{eqnarray} \label{eq:dtau}
\delta \tau &=& \ensuremath{\tau^{\mathrm{est}}} - \ensuremath{\tau^{\mathrm{true}}} \nonumber \\
&=& \ln\left(1 + \frac{\ensuremath{C_{\mathrm{est}}} - \ensuremath{C_{\mathrm{true}}} }{\ensuremath{C_{\mathrm{true}}}}\right) \nonumber \\
&\simeq& f_c,
\end{eqnarray}
where in the last line we have substituted in Equation~\ref{fc_def} and assumed that it
is small ($|f_c| \lesssim 0.1$).
Inserting Equation~\ref{eq:dtau} into Equation~\ref{eq:dalpha}, and
using the fact that in the low-absorption regions
$F = \exp(-\tau) \simeq 1 - \tau$, we get
\begin{equation}
\delta \alpha \sim \frac{f_c}{1 - F}.
\end{equation}
As the strongest constraints on the temperature-density relation\ of the IGM come from the
low-absorption/high-transmission ($1 - F \sim \%$) pixels of the
Ly$\alpha$\ forest, systematic errors in the continuum estimate of just
a few percent can cause large ($\delta \gamma \sim \delta \alpha \sim 1$)
errors in the estimated value of $\gamma$.
\section{Mock Analysis}\label{sec:analysis}
In this section, we generate theoretical flux PDFs of the Ly$\alpha$\ forest
using a simple semi-analytic model, and quantitatively study the errors introduced into
the estimated temperature-density relation\ caused by systematic
errors in the Ly$\alpha$\ forest continuum.
\subsection{Toy Flux PDFs}
\begin{figure}
\epsscale{1}
\plotone{pdfs_1.6_0.5_rev.ps}
\caption{\label{fig:rawpdfs}
Unbinned flux PDFs generated from Eq.~\ref{eq:fpdf}, with
$\gamma=0.5$ (thin red dashed line) and $\gamma=1.6$
(thin red solid line). The flux PDFs are normalized
to $\langle F \rangle = 0.70$.
The thick black lines denote
the PDFs after being smoothed by a Gaussian kernel corresponding
to $\ensuremath{\mathrm{S/N}}=50$ in our simplified noise model.
}
\end{figure}
\begin{figure}
\epsscale{1.}
\plotone{plot_pdfbin_rev.ps}
\caption{\label{fig:binnedpdfs}
(Upper panel) Binned flux PDFs generated from Eq.~\ref{eq:fpdf} with
different temperature-density
relations, $\gamma=0.5$ (green dashed line), $\gamma=1.0$ (blue dotted line),
$\gamma=1.2$ (black solid line), and $\gamma=1.6$ (red dot-dashed line).
The lower panel shows the ratio of the various flux PDFs
with respect to the $\gamma = 1.0$ model. The relative differences
are similar to those presented in \citet{bolton+08}.
}
\end{figure}
We first generate theoretical flux PDFs by using as a starting point the semi-analytic density
PDF from \citet{mhr00} (hereafter MHR00):
\begin{equation} \label{eq:pmhr}
p(\Delta) d\Delta = A \exp{\left[\frac{-(\Delta^{-2/3} - C_0)^2}{2 (2\delta_0 /3)^2}
\right]} \Delta^{-\beta} d\Delta,
\end{equation}
where $\Delta$ is the matter density, while $A$, $\beta$, $C_0$, and
$\delta_0$ are redshift-dependent parameters interpolated from the values
published in MHR00.
At $z=3$, the corresponding values are $A = 0.558$, $\beta=2.35$,
$C_0 = 0.599$, and $\delta_0 = 1.90$.
The corresponding Ly$\alpha$\ forest flux PDF is then derived by substituting in $\Delta(\tau)$ using the
FGPA (Equation~\ref{eq:fgpa}) and
then the relation $F = \exp{(-\tau)}$:
\begin{eqnarray} \label{eq:fpdf}
p(F) dF &=& \frac{dF}{F} \frac{A}{\tau_0 (2 - \alpha)}
\left( \frac{-\ln{F}}{\tau_0} \right)^{\frac{\alpha-1-\beta}{2 - \alpha}}
\nonumber \\
& & \times \exp{\left\{ \frac{-\left[(-\ln{F}/\tau_0)^{-\frac{2}{3(2-\alpha)}}
- C_0 \right]^2}{2 (2\delta_0/3)^2} \right\}}.
\end{eqnarray}
This is then normalized such that $\int_0^1 p(F) dF = 1$.
The parameters in Equation~\ref{eq:fpdf}
that characterize the IGM are $\alpha$ and $\tau_0$.
$\tau_0$ is a function that depends on the background photoionization rate
in the IGM and the temperature at mean density, but for the purposes
of this paper we adjust it to yield a fixed value of mean optical
depth $\langle F \rangle (z) = \exp{(-\tau_{\mathrm{eff}}(z))}$
for a given redshift.
Thus, at fixed redshift the only free parameter in this model is
$\gamma = 1 + \alpha/0.7$ which parametrizes
the temperature density equation-of-state.
Note that this model does not correctly account for thermal broadening
of the Ly$\alpha$\ forest lines, nor peculiar velocities.
However, \citet{bolton+08} have shown from hydrodynamical simulations
that the flux PDF is not sensitive to $T_0$, which affects the
Ly$\alpha$\ forest primarily through thermal broadening and by
changing the Jean's smoothing scale.
Peculiar velocities need to be included to obtain the correct flux PDF shape,
but they arise primarily from gravitational collapse of large-scale
structure. Thus, we do not expect the omission of peculiar velocities
to seriously affect the relative behavior of the flux PDF with changes to the
temperature-density relation.
We show the flux PDFs calculated for
2 extreme values, $\gamma=0.5$ and $\gamma=1.6$
in Figure~\ref{fig:rawpdfs}.
$\gamma=0.5$ corresponds to a highly inverted temperature-density
relation, which had been detected by \citep{becker+07, bolton+08}.
While $\gamma=1.6$ is a value calculated by \citet{hui+gned97} for a
post-reionization relaxed IGM, which should be valid in the epoch between
the end of hydrogen reionization and prior to \ion{He}{2}\ reionization,
$ 3 \lesssim z \lesssim 6$.
Figure~\ref{fig:rawpdfs} illustrates the effects of changing $\gamma$:
at fixed temperature at mean density $T_0$,
lowering $\gamma$ increases the temperature in the underdense
($\ensuremath{\Delta} < 1$) regions of the IGM.
This decreases the hydrogen recombination rate, therefore
reducing the \ion{H}{1} optical depth and increasing the transmission
in those regions. The differences are apparent in
the $F \gtrsim 0.5$ regions of the PDF, where the peak is shifted
to higher $F$.
The probability of having pixels with close to
100\% transmission decreases as $\gamma$ goes up;
at $\gamma=1.6$, the probability drops to zero by $F \simeq 0.98$
(in our model) thus
leading to Gunn-Peterson absorption at the 2\% level.
In this case, the lack of transmission peaks reaching the true
continuum level is likely to lead to an underestimate
of the continuum, potentially biasing the estimate of $\gamma$.
Note that the moderate-absorption regions of
the PDF ($0.2 \lesssim F \lesssim 0.6$)
do not vary much with $\gamma$. This agrees with \citet{bolton+08},
who found that using only these moderate-absorption pixels
significantly weakens constraints on $\gamma$.
The effects of finite signal-to-noise are introduced by smoothing
the flux PDF from Equation~\ref{eq:fpdf} with a Gaussian kernel.
We use a constant smoothing length of
$\sigma_{sm} = 1/50$ to simulate an average $\ensuremath{\mathrm{S/N}}= 50$ per
pixel, a typical value for high-resolution and high-$\ensuremath{\mathrm{S/N}}$ Ly$\alpha$\ forest
spectra.
This is an acceptable simplification, as Table~3 in \citet{mcd+00}
shows that the typical pixel noise in real data is
roughly constant across the different flux bins.
Note that the presence of noise scatters some pixels to
$F \gtrsim 1$, changing the shape of the flux PDFs
(black curves in Figure~\ref{fig:rawpdfs}).
We then bin the flux PDFs in the fashion of \citet{mcd+00} and
\citet{kim+07}: the `noisy' flux PDF is divided into 21
bins with size $\Delta F = 0.05$ in the range $0 < F < 1$.
The `noisy' portions of the flux PDF
with $F < 0$ and $F > 1$ are transferred to the $F=0$ and $F = 1$ bins,
respectively.
In Figure~\ref{fig:binnedpdfs}, we show the binned PDFs for different
temperature-density relations. The overall trend of the unbinned PDFs
are similar to that in the unbinned case:
the high-transmission peak of the PDF
shifts to larger $F$ as $\gamma$ is decreased. In addition, the
number of pixels in the $F=1.0$ bin increases with decreasing $\gamma$.
The overall shape of the flux PDFs are in broad agreement with other
theoretical flux PDFs published in the literature, including
\citet{mcd+00}, \citet{bolton+08}, and \citet{white+10}.
However, note that we are unable to reproduce the exact shapes of the
flux PDFs as seen in the \citet{kim+07} data,
nor from hydrodynamical simulations \citep[e.g.,][]{bolton+08}:
our PDFs tend to peak at lower values of $F$ for a given value of $\gamma$,
while the $\gamma=0.5$
PDF here does not have the same shape in the high-transmission end obtained
from the \citet{bolton+08} hydrodynamical simulations,
in which the PDF peaks in the $F=1.0$ bin.
The MHR-FGPA model also appears to underpredict the number of
low-transmission pixels ($F\approx 0$) in comparison with the
\citet{bolton+08} simulations.
In the lower panel of Figure~\ref{fig:binnedpdfs}, we plot the ratio of
the flux PDFs with respect to that with $\gamma = 1.2$.
These appear
similar to the analogous plots shown in the lower panels of Figures~2
and 4 in \citet{bolton+08}.
The main differences are that the flux PDFs in the hydrodynamical
simulations pivot at $F\approx 0.1$, while those in our model
pivot at a higher value of $F\approx 0.3$.
Nevertheless, the \emph{relative} behavior of our model PDFs with
respect to $\gamma$ appear similar to those in the \citet{bolton+08}
hydrodynamical simulations.
Since the primary effect of systematic continuum errors is to
rescale the flux PDF along the abscissa,
the fact that our toy model correctly reproduces the relative
changes in the PDF with respect to $\gamma$ means that it can be used
to study continuum errors, even if it should not be used to make direct
comparisons with data.
\subsection{Effect of Continuum Errors}
\begin{figure}
\epsscale{1.}
\plotone{plot_conterr_rev.ps}
\caption{\label{fig:conterr}
The effect of systematic continuum errors on the Ly$\alpha$\
forest flux PDF.
The top panel shows the unbinned flux PDFs generated with
$\gamma = 1.2$ and $\ensuremath{\mathrm{S/N}} = 50$, with no continuum error (black line),
2\% overestimated continuum (dotted line), and 2\% underestimated
continuum(dashed line). The lower panel shows the corresponding
flux PDFs in $\Delta F=0.05$ bins. 2\% errors in the continuum
determination can drastically change the shape of the binned flux PDF.
}
\end{figure}
\begin{figure}
\epsscale{1.}
\plotone{plot_meanflux_rev.ps}
\caption{\label{fig:meanflux}
The effect of changing $\tau_0$ (or equivalently $\ensuremath{\langle F \rangle}$)
on the flux PDF. The top panel shows the unbinned flux PDFs generated with
$\gamma = 1.2$ and $\ensuremath{\mathrm{S/N}} = 50$, with $\tau_0$ adjusted to match the effective
mean-flux of the PDFs with the corresponding line-styles in Figure~\ref{fig:conterr}.
The lower panel shows the corresponding
flux PDFs in $\Delta F=0.05$ bins. $\tau_0$ changes the flux PDF
differently from the temperature-density relation or continuum errors. No
continuum errors have been applied in these PDFs.
}
\end{figure}
In this section, we study the effect of systematic continuum biases
on the value of $\gamma$ measured from the mock flux PDFs
described above,
and the corresponding errors on this estimate.
This mock analysis is carried out at a fixed redshift of $z=3$,
an epoch at which Gunn-Peterson absorption is usually assumed to
be negligible,
but could account for as much as 4-5\% of the observed flux
\citep{gial+92,fg+08}.
All mock PDFs are set to a mean flux of
$\langle F \rangle(z=3)=0.70$ \citep{meik+white04}, and the model parameters
from Equation~\ref{eq:pmhr} are set to $z=3$.
Systematic continuum errors are introduced into the flux PDFs
by multiplying the flux scales with
the factor $1 + f_c$ prior to binning --- positive values of $f_c$ denote
an overestimate with respect to the true continuum, and vice versa.
The dotted and dashed curves in Figure~\ref{fig:conterr} illustrate
the effects of these errors on a flux PDF with $\gamma=1.2$.
Clearly, just $2\%$ systematic errors in the continuum estimation can
dramatically change the shape of the binned flux PDFs:
with a 2\% underestimate of the continuum, the number of unabsorbed
pixels ($F=1.0$) increases by a factor of two.
Note that the continuum errors also change the measured mean-flux
$\langle F \rangle$. For the cases with $f_c = [-0.02, 0, 0.02]$, the
effective mean-fluxes are $\ensuremath{\langle F \rangle} = [0.723, 0.699, 0.669]$, respectively.
The scatter in these mean-flux values are at the same
level as the current observational uncertainty of several percent
at $z \approx 3$.
To investigate if the mean-flux might be degenerate
with $\gamma$ and $f_c$, in Figure~\ref{fig:meanflux} we show $\gamma=1.2$ flux PDFs with
$\tau_0$ adjusted to give $\ensuremath{\langle F \rangle}$ values corresponding to the curves in
Figure~\ref{fig:conterr}, without continuum errors included.
The corresponding values used to generate these PDFs
are $\tau_0 = [1.003, 1.185, 1.451]$ for $\ensuremath{\langle F \rangle} = [0.723, 0.699, 0.669]$,
respectively.
We see that a smaller $\tau_0$ or higher $\ensuremath{\langle F \rangle}$ can shift the high-transmission peak
of the flux PDF to larger $F$, analogous to lowering $\gamma$.
However,
the relative ratio of high-absorption/low-transmission pixels ($F < 0.85$),
with respect to the low-absorption/high-transmission end ($F > 0.85$)
behaves differently with changes in $\tau_0$ compared to when
$\gamma$ or $f_c$ is varied (c.f. Fig.~\ref{fig:binnedpdfs}
and Fig.~\ref{fig:conterr}).
This suggests that the degeneracy of $\tau_0$ with respect to
$\gamma$ and $f_c$ can be broken in an analysis of the full flux PDF,
but nevertheless $\tau_0$ needs to be taken into account.
To quantify the effects on the estimated temperature-density relation,
the flux PDFs with different combinations of $[\gamma, f_c]$ are
then compared with mock data PDFs with similar noise properties to
the \citet{kim+07} flux PDFs, which is the best existing data set.
However, the bins of the flux PDF are highly correlated,
thus one needs a covariance
matrix in order to generate random samples with the correct covariances.
As our semi-analytic model does not actually result in simulated
Ly$\alpha$\ forest sightlines, it is problematic to estimate the necessary covariances.
Instead, we use the same covariance matrices described in
\S4.1 of \citet{bolton+08}:
the covariance matrix (kindly provided by James Bolton) derived from
their $z=2.94,\gamma=0.44$ simulations are used to estimate
the correlation coefficients between different flux bins.
This is then used in conjunction with the errors
(i.e. diagonal covariance terms) in the \citet{kim+07} data to estimate the
cross-terms of the final covariance matrix.
We then carry out Cholesky decomposition on this covariance matrix to generate
random correlated errors, which are then applied to the model flux PDFs.
This process assumes that all the model covariances are the same as
those in \citet{bolton+08},
but this should be reasonable considering the
approximate nature of our analysis.
We can then quantify the effects of continuum bias by sampling
the reduced Chi-squared between each mock data PDF, and different
model curves
with various combinations of $\gamma$ and $f_c$. Since the bins are
highly correlated, the expression for the Chi-squared is:
\begin{equation}
\chi^2 =
[d_i - p_i^{\mathrm{model}}]^T C_{ij}^{-1} [d_i - p_i^{\mathrm{model}}]
\end{equation}
where $d_i$ are the bins in the mock data PDFs,
$p_i^{\mathrm{model}}$ is the model flux PDF with some combination of
$\gamma$, $f_c$ and $\tau_0$, while
$C_{ij}$ is the covariance matrix discussed above.
The confidence level of the various parameter combinations can then be
estimated from $\Delta \chi^2 \equiv \chi^2 - \chi^2_{min}$.
In our analysis, we regard the Ly$\alpha$\ forest mean-flux, $\ensuremath{\langle F \rangle}$,
(or equivalently within our model, $\tau_0$)
as a nuisance parameter to be marginalized over.
At each point in the $[\gamma,f_c]$ model parameter space, we
marginalize over the mean-flux in the likelihood,
$\mathcal{L} \equiv [ (2\pi)^N \det{C_{ij}} ]^{-1} \exp\{-(1/2) \chi^2 \}$.
If we assume the error in the mean-flux is Gaussian, we can then
marginalize over the possible mean-flux values $\ensuremath{\langle F \rangle}'$:
\begin{eqnarray}
\mathcal{L}(\gamma,f_c|\ensuremath{\langle F \rangle}) &\propto&
\int^{\infty}_{\infty}
\mathcal{L}(\gamma, f_c, \ensuremath{\langle F \rangle}' ) \nonumber \\
& & \exp \left(- \frac{(\ensuremath{\langle F \rangle}' - \ensuremath{\langle F \rangle})^2} {\sigma_{\ensuremath{\langle F \rangle}}^2} \right)
d\ensuremath{\langle F \rangle}',
\end{eqnarray}
where we have used $\ensuremath{\langle F \rangle} = 0.70$ and $\sigma_{\ensuremath{\langle F \rangle}} = 0.02$ for $z=3$.
Note that we have ignored normalization factors that will be canceled out when we evaluate
$\Delta \chi^2$, and have evaluated the integral using
7-point Gauss-Hermite quadrature.
\begin{figure}
\epsscale{1.}
\plotone{chisq2d_g1.0_rev2.ps}
\caption{\label{fig:chisq}
Likelihood plots for different combinations of $\gamma$ and $f_c$,
computed with respect to a mock flux PDFs generated
with $\gamma=1.0$.
The contours enclose the 68\% and 95\% confidence intervals,
while the color scale denotes $\Delta \chi^2 \equiv \chi^2 - \chi^2_{min}$.
The square and cross denote the extreme values of $\gamma$
plotted in Figure~\ref{fig:binpdf} which
constitute acceptable $\sim 1\sigma$ fits to the $\ensuremath{\gamma^{\mathrm{true}}}=1.0$ mock flux PDF
(also plotted in Figure~\ref{fig:binpdf}) once continuum errors are considered.
The white horizontal arrow denotes the 68 \% confidence interval for
$\gamma$ when the continuum is perfectly known, i.e. $f_c = 0.$.
}
\end{figure}
\begin{figure}
\epsscale{1.}
\plotone{plotbin_g1.0_rev2.ps}
\caption{\label{fig:binpdf}
Flux PDFs corresponding to Figure~\ref{fig:chisq}.
The error bars denote the mock $\gamma=1.0$ flux PDF
with simulated scatter,
against which the likelihoods in Figure~\ref{fig:chisq} were
calculated.
The dashed- and dotted lines denote model PDFs with
$[\gamma=0.75,f_c=-0.007] $ and $[\gamma=1.2,f_c=0.006]$, which constitute
acceptable ($\Delta \chi^2 \approx 2$, 68\% confidence level)
fits to the mock $\ensuremath{\gamma^{\mathrm{true}}}=1.0$ data points.
Their respective positions in $[\gamma,f_c]$ space are shown as
the square and cross in Fig~\ref{fig:chisq}.
Note that the ordinate axis is plotted logarithmically in this figure.}
\end{figure}
Figure~\ref{fig:chisq} shows the likelihood plot for
fits to a mock data PDF generated with $\ensuremath{\gamma^{\mathrm{true}}}=1.0$,
computed with respect to different
combinations of $\gamma$ and $f_c$.
The contours indicates the 68\% and 95\% confidence intervals
($\Delta \chi^2 = 2.30$ and $\Delta \chi^2 = 6.17$, respectively).
The error ellipse shows a significant degeneracy
between $\gamma$ and $f_c$.
For reference, we also show the 68\% confidence level ($\Delta \chi^2 = 1$)
in the case where the continuum is not allowed to vary (horizontal
white arrows).
The range of $\gamma$ which falls within the 68\% confidence
interval increases significantly to $\sigma_{\gamma} = 0.2-0.3$ once the continuum level is allowed to vary.
Without continuum errors, the
error\footnote{We use the terms `$1\sigma$ errors' and `68\% confidence intervals'
interchangeably but the latter is the quantity we have really measured.
These two terms would be identical in the case of Gaussian errors.} in the estimated $\gamma$ is $\sigma_{\gamma} \approx 0.05$.
The square and cross symbols in Figure~\ref{fig:chisq}
denote two extreme points
in the $[\gamma, f_c]$ parameter space
that fall within the
68\% confidence interval of the underlying
flux PDF with $\ensuremath{\gamma^{\mathrm{true}}} = 1.0$.
The binned flux PDFs for these two combinations of $[\gamma, f_c]$
are juxtaposed
with the underlying PDF in Figure~\ref{fig:binpdf}.
The similarity of these PDFs with significantly different
temperature-density relations ($\gamma = 0.75-1.2$) illustrates
the effect of continuum errors on the measurement.
Figure~\ref{fig:eos_errs} shows the average
estimated value of $\gamma$ as a function of the `true' underlying
temperature-density power-law $\ensuremath{\gamma^{\mathrm{true}}}$ and its associated errors,
averaged over $40$ Gaussian
realizations for each value of $\ensuremath{\gamma^{\mathrm{true}}}$.
The dotted lines denote the average 68\% confidence intervals if systematic continuum errors
are considered, while the shaded area show the $1\sigma$
intervals if
the continuum is known perfectly.
We see that continuum errors cause
a significant increase in the error on $\gamma$:
at higher temperature-density slopes ($\gamma \sim 1.5$) the errors increase
from $\sigma_{\gamma} \approx 0.05$ to $\sigma_{\gamma}^\mathrm{cont} \approx 0.15$;
at flatter or inverted slopes, the errors are generally larger but still increase
from $\sigma_{\gamma} \approx 0.1$ to $\sigma_{\gamma}^\mathrm{cont} \approx 0.2$.
In general, the errors increase by a factor of $\sim 2$ once we consider continuum
errors.
\begin{figure}
\epsscale{1.}
\plotone{ploterrs_deltachisq_margmf.ps}
\caption{\label{fig:eos_errs}
The $1\sigma$ upper- and lower-limits in the value of $\gamma$ estimated from
mock flux PDFs generated with some underlying value of \ensuremath{\gamma^{\mathrm{true}}}, when continuum
biases are taken into account (dotted lines) and if the continuum is perfectly known (shaded area
enclosed by solid lines). Continuum uncertainties roughly double the error on $\gamma$.
These values are averaged over $40$ mock data realizations for each value of $\ensuremath{\gamma^{\mathrm{true}}}$.
}
\end{figure}
\section{Discussion \& Conclusion} \label{sec:conclusion}
\begin{figure}[t]
\epsscale{1.}
\plotone{fluxplot_cen.ps}
\caption{\label{fig:cen_spec}
A Ly$\alpha$\ forest sightline from the \citet{cen+chisari10} hydrodynamical simulations extracted at $z = 3.0$
and smoothed to a resolution of $6.7\, \mathrm{km\, s^{-1}}$.
The dashed horizontal line indicates the $F = 0.95$ flux level.
Even though no pixel noise is present, an attempt to fit the continuum to this mock spectrum
will likely result in a continuum bias of $f_c \approx -0.05$. Courtesy of R. Cen.
}
\end{figure}
In this paper we have quantified the effect
of systematic continuum errors on the temperature-density relation,
$\gamma$, estimated from the Ly$\alpha$\ forest flux PDF, using a simple toy model
which correctly reproduces the relative changes in the flux PDF with
respect to $\gamma$.
We found that small systematic errors of just $\sim 1-2\%$ in the overall continuum level
can bias the estimated $\gamma$
to smaller values.
In the absence of continuum errors, the $1\sigma$ errors on the estimated value of $\gamma$
is $\sigma_{\gamma} \approx 0.1$, but with continuum errors this interval is increased
to $\sigma_{\gamma} \approx 0.2$.
While we have used a simple semi-analytic model
to calculate the flux PDFs
(although note that this is the same model used by \citet{becker+07} to
analyze their spectra), the relative scaling of the resultant flux
PDF with $\gamma$ is reasonably accurate and thus
the biases discussed here are qualitatively valid. We have also marginalized over
the mean-flux in our analysis.
In addition, in the simple mock analysis presented here,
both the data realizations and `theory' flux PDFs are
derived from exactly the same model.
In a real data analysis, uncertainties in the underlying
physical model (e.g.\ gas temperature, UV ionizing background, $\sigma_8$,
Jeans' smoothing scale etc.),
and other observational uncertainties such as metal contamination
must increase the error in $\gamma$ beyond those presented here.
Another point that should be emphasized is that prior to
the biases in $\gamma$ considered in this paper,
systematic continuum errors are not a symmetric effect. Overestimates
of the continuum ($f_c > 1$) are less likely with high-quality data
--- an observer is unlikely to place the continuum level
significantly above the observed transmission peaks,
whereas low-level Gunn-Peterson absorption
can degrade the transmission peaks and lead to
underestimates of the continuum ($f_c < 0$).
There is thus an additional bias towards smaller values of $\gamma$
due the higher probability of
underestimating the continuum rather than overestimating it.
In the MHR00 model considered here, the maximum amount of possible
Gunn-Peterson absorption is in fact fairly limited at $z = 3$: about 2\%
with $\gamma = 1.6$ (Figure~\ref{fig:rawpdfs}).
Other models could provide even more Gunn-Peterson absorption
at these redshifts.
An example is shown in Figure~\ref{fig:cen_spec}, which plots a $z=3$
Ly$\alpha$\ forest spectrum extracted from the detailed hydrodynamical simulations
described in \citet{cen+chisari10}. If this were actual data, an observer
would probably underestimate the spectrum by 5\% even
in the absence of noise.
Recent studies of the flux PDF that claimed a highly inverted temperature-density relation\
\citep{becker+07,bolton+08,viel+09} have
considered the possibility of a systematic continuum bias.
However, \citet{becker+07} used the same semi-analytic MHR00 model for the flux PDF
used in this paper, which may not accurately capture the details of
the Ly$\alpha$\ forest flux field at a sufficiently level for data analysis.
\citet{bolton+08} and \citet{viel+09} both analyzed the same data set from \citet{kim+07}:
\citet{bolton+08} checked for continuum errors by comparing their
calculating likelihoods after rescaling their continua by 1.5\% and 5\%, while
\citet{viel+09} marginalized the continuum in their analysis ---
both concluded that continuum errors were not significant and that
the inverted temperature-density relation\ was favored.
However, it is also interesting to note that
\citet{viel+09} arrived at a best-fit continuum error suggesting an \emph{over}estimated
continuum of 1\% in the \citet{kim+07} data
(c.f.\ the random errors in the continuum fitting, $\sigma_{\mathrm{fit}} = 1-2 \%$).
In other words, they favor a continuum which is \emph{lower} than that
which was actually fitted to the \citet{kim+07} spectra.
In the context of our analysis, this would be consistent with an underestimation of
$\gamma$.
We also note that \citet{viel+09} had fitted for a redshift-independent continuum error
when analyzing the \citet{kim+07} flux PDFs.
Since this data set is dominated by lower-redshift ($z< 2.5$)
spectra (which should be less affected by low-level Gunn-Peterson absorption),
it is possible that the true continuum error cannot be fully accounted for,
using a redshift-independent approach.
Considering the controversial nature of the claims of a highly-inverted
IGM temperature-density relation, we feel a more direct approach towards dealing with continuum bias
is required that had been done previously.
There are some possibilities:
\citet{gial+92} had extrapolated a power-law continuum from redwards of the
quasar Ly$\alpha$\ emission line to estimate the amount of uniform GP
absorption in the peaks of the Ly$\alpha$\ forest.
Existing data sets will allow much stronger constraints to be placed using
this method, although it requires the assumption that the mean quasar continuum
slope does not change with redshift.
Alternatively, when comparing simulations with data, the simulated
sightlines need to be processed through the same continuum-fitting method
as the observed data, i.e. through `forward-modeling'.
\citet{fg+08} estimated continuum biases by
fitting simulated mock spectra, and comparing these fits
with the underlying continuum. However, they were attempting to
measure the optical evolution of the Ly$\alpha$\ forest, and used
simulations with a fixed value of $\gamma=1.6$.
For this method to account for continuum biases in
studies exploring a large parameter space, one would need to hand-fit
large sets of mock spectra (ideally including realistic quasar continua)
covering the explored parameter space.
This would be time-consuming, but in principle would be a robust
method to account for systematic errors in the continuum fits,
especially if automated continuum-fitting methods are used \citep{dall+09}.
Another avenue for improvement to use larger Ly$\alpha$\ forest data sets,
and hence reduce the errors in the flux PDF and other statistics.
The \citet{kim+07} sample (18 quasars) represent a
significant increase in data size in comparison with \citet{mcd+00} (8 quasars),
but considerably more high-resolution, high-\ensuremath{\mathrm{S/N}}\, spectra than these currently exist.
Increasing the sample size would clearly reduce the errors in the measured flux PDF,
which would limit the scope of continuum errors to bias the measured
temperature-density relation.
At time of writing, the measurement of the inverted temperature-density relation
has been carried out by the same group of authors \citep{bolton+08,viel+09}
analyzing the same flux PDF data \citep{kim+07}.
Alternative and independent analyses are urgently required in order to verify this phenomenon,
which would have important implications on IGM science as well as energetic
sources in the high-redshift universe.
\acknowledgements{
The author thanks David Spergel, Renyue Cen and Xavier Prochaska for useful
discussions and comments. He is also grateful to James Bolton for
providing covariance matrices from his simulations, and to the anonymous
referee for useful criticisms which have improved the paper.
}
|
\section{Introduction}\label{S:intro}
Modern imaging surveys such as the Cosmic Evolution Survey (COSMOS) \citep{Scoville07}) and the Great Observatories Origins Deep Survey (GOODS) \citep{Giavalisco04} contain a wealth of galaxy information in the form of stamp images and multiwavelength photometry across the electromagnetic spectrum. Spectroscopy adds to preexisting survey data through, among other information, precise redshift determination and firm galaxy classification based on spectral features. While spectral analysis can be automated in some cases, human examination is often necessary. This is especially true when the spectroscopic targets are faint or unusual sources for which automated redshift determination is not reliable.
Imaging and photometric data from a multiwavelength survey can play a vital role in deciphering difficult spectra. For example, a high-redshift galaxy is expected to have a strong photometric break (the ``Lyman break'') at rest-frame $\lambda$1216~\AA, with essentially no flux short of the rest-frame $\lambda$912~\AA\ Lyman limit. The presence of this spectral break in the SED can suggest an initial redshift guess and validate a high-redshift assignment based on weak spectral features. Conversely, photometric detections blueward of the Lyman limit can preclude a high-redshift assignment. As another example, passive galaxies typically show a peak in the SED near rest-frame 1.6~$\mu$m and a spectral break near rest-frame $\lambda$4000~\AA, while their spectra reveal little more than faint absorption features. Again, the photometric characteristics can be used to inform the redshift assignment. Stamp images of a source can also be very helpful in resolving puzzling spectra, as they reveal the morphology and spatial extent of a source as well as possible sources of photometric or spectroscopic contamination. High-resolution imaging can thus help to distinguish stars from galaxies and reveal nearby or coincident systems giving rise to anamolous spectra.
With the large numbers of spectra taken in survey fields, it is necessary to have a means of rapid spectral analysis. SpecPro was developed to analyze faint spectra of high-redshift galaxies and active galactic nuclei (AGN) in the COSMOS field, where the available photometric data effectively provides low-resolution spectra covering 0.1 to 8~$\mu$m. The spectra, taken with the Keck II DEIMOS multislit spectrometer \citep{Faber03}, often show few strong features. By incorporating the SED and postage stamp images from COSMOS in the SpecPro interface we have significantly improved our ability to find redshifts and understand our galaxy sample. In addition to facilitating the analysis of faint sources, the cross-correlation capability makes redshift determination for bright sources extremely fast.
Since SpecPro was originally designed to examine spectra taken with a multislit spectrometer, data structures are organized around the idea of a slit mask. Data directories contain files for sources on a mask, the files within the mask directory being differentiated by slit number. Aside from this general structure, we have made the input formats quite flexible in order to accommodate data from other instruments/surveys. In addition to DEIMOS, the code has to date been used successfully with data from the IMACS multislit spectrometer on Magellan \citep{Dressler06} as well as the FMOS fiber spectrometer and FOCAS multisplit spectrometer, both on Subaru \citep{Kashikawa02, Kimura10}. Any spectroscopic data can be viewed with SpecPro after conforming to the simple formats outlined in Section 3.
The structure of this paper is as follows. In Section 2 we provide an overview of the interface and discuss the motivation behind its design. In Section 3 we discuss the inputs to the program. Section 4 concludes with a summary and a brief discussion of possible further development.
The code with installation instructions as well as example data, an overview of data formats, and a brief tutorial are available at http://specpro.caltech.edu. \newpage
\section{SpecPro Overview}
\begin{figure*}[ht]
\centering
\includegraphics[scale=0.36]{f1}
\caption{The SpecPro widget interface. The object shown is a star-forming galaxy at redshift 0.89. Stamp images of the source are displayed on the left side of the interface, and the 1-D spectrum is displayed in the upper right window. On the 1-D plot, vertical lines mark the positions of prominent emission features, and a spectral template is overlaid in blue. Under the 1-D spectrum is the corresponding 2-D spectrum, binned to fit the display. The same emission lines plotted over the 1-D spectrum are indicated in the 2-D spectrum, and the atmospheric absorption regions are marked with horizontal red lines above and below the spectrum. Emission lines ([OII], H$\gamma$, H$\beta$, [OIII]) are easily visible in the binned 2-D image. The window on the bottom right shows the SED of the source from the ultraviolet to the mid-infrared. The AB magnitude (increasing downwards) is plotted against filter wavelength. The SED data points indicate filter FWHM with horizontal bars and photometric errors with vertical bars. The observed-frame positions of important photometric indicators, including the Lyman limit at $\lambda$912~\AA, Ly$\alpha$ at $\lambda$1216~\AA, the 4000~\AA\ break, and 1.6~$\mu$m are indicated with vertical dashed lines.}
\label{Fi:plotone}
\end{figure*}
The SpecPro interface is shown in Figures~\ref{Fi:plotone} and \ref{Fi:plottwo}. Figure~\ref{Fi:plotone} shows the full interface, while Figure~\ref{Fi:plottwo} shows a smaller version that can be invoked for use on smaller monitors. A galaxy with strong emission lines from the COSMOS survey is displayed.
Surrounding the data displays are widgets allowing the user to perform various tasks, such as navigate through spectra in the mask directory, adjust the current redshift guess, bin and smooth the 1-D spectrum, overlay galaxy templates, plot the redshifted positions of common emission and absorption features, change stamp image and 2-D spectrum contrast, and save results to an external file. In the following sections we outline the key capabilities of the interface, emphasizing how they assist in spectral analysis.
It is worth emphasizing that the SpecPro interface does not require all of the various pieces of information (2-D spectrum, stamp images, SED) to run. It can be used with a subset of this information and the windows for which there is no data are left blank. Therefore the interface can be used to view and analyze spectra for which the data associated with a multiwavelength survey is absent, or to pre-select targets for spectroscopy using just photometric and imaging data.
\begin{figure*}[ht]
\centering
\includegraphics[scale=0.36]{f2}
\caption{The SpecPro widget interface, small version.}
\label{Fi:plottwo}
\end{figure*}
\subsection{Stamp images}
Available stamp images of the source are displayed on the left side of the interface in 100x100 pixel windows. Under the final window is a scroll list containing the names of overflow images. The user can click entries in this list to display them in the final stamp window, allowing an arbitrary number of stamp images to be stored and viewed with the interface. The images are displayed in the order in which they are stored in the structure array, beginning with the upper left stamp window and going down.
The displayed images are signal-to-noise (S/N) plots generated by dividing the image flux by the average error in the image. This maximizes the dynamic range in order to display both bright sources and faint counterparts blueward of a spectral break. By viewing the stamps the user can quickly determine the source of a serendipitous spectrum and whether photometric points in the source SED correspond to real detections or artifacts.
Selecting the ``Show extraction" button (located underneath the 2-D plot window) causes the scale in arcseconds to be displayed on the stamps. In addition, if the slit information (RA, DEC, length, width, and angle) is provided, it is plotted over the stamp images, as in Figures~\ref{Fi:plotone} and \ref{Fi:plottwo}.
\subsection{Spectrum Display}
The one-dimensional spectrum is displayed in the upper right window of the display. The spectrum is shown in white, with overlaid spectral templates shown in blue. Regions of significant atmospheric absorption are indicated with horizontal blue lines. The residual error, which can help determine whether a weak feature is real or an artifact due to sky subtraction, can be overlaid on the plot by clicking the ``Show sky'' button on the second row down from the 1-D display.
The 1-D spectrum can be binned and smoothed to bring out features in low S/N spectra. Binning combines pixels, reducing the resolution of the spectrum to increase the S/N, while smoothing performs an inverse variance weighted average in the vicinity of each pixel to reduce the noise. In Figure~\ref{Fi:plotone} the spectrum has been binned and smoothed, clearly revealing prominent emission features.
The 2-D spectrum is displayed beneath the 1-D spectrum, binned, if necessary, to fit the display. The 2-D spectrum can be helpful to: (1)~verify the fidelity of the 1-D spectrum, (2)~determine whether features observed in the 1-D spectrum are real or artifacts, (3)~identify serendipitous objects, and (4)~examine the spatial morphology of emission features. As described in Section 2.3, the user has the ability to reextract the 1-D spectrum from the 2-D, which can be very useful in cases in which there was a problem with 1-D extraction or the 1-D spectrum of a serendipitous source is desired.
The contrast of the 2-D spectrum can be adjusted with a slider widget. Next to the contrast slider is a droplist widget labeled ``Max sigma:". This is set to 10 by default, which indicates that contrast scaling has an upper limit of 10$\sigma$, so that pixels 10$\sigma$ from the mean are rescaled to the maximum pixel value of 255. Lowering this value can help in cases in which the spectrum of interest is being drowned out by the light of bright serendipitous sources.
The positions (observed-frame) of emission and absorption features of different galaxy types can be displayed over both the 1-D and the 2-D spectrum by selecting the buttons above the 2-D display. Typical lines from star-forming galaxies, passive galaxies, quasistellar objects (QSOs), Seyferts, and high-redshift galaxies can be selected. Redshifts can be found manually by overlaying the appropriate lines on the spectrum and adjusting the redshift until they line up with observed spectral features. These lines can also be used to identify absorption and emission features in the various spectral templates.
Both the 1-D and the 2-D spectrum can be zoomed on by clicking, dragging and releasing around the region of interest. In the case of the 2-D, this action causes the zoomed region to be displayed in its unbinned state in a separate window. This can be helpful, for instance, to determine whether an emission feature is [OII] or Ly$\alpha$, which can often be deduced from the morphology of the emission line.
\subsection{Reextraction of 1-D spectrum}
It is often necessary to reextract the 1-D spectrum from the 2-D, particularly in order to recover serendipitously observed sources that fall in the slit. We have provided a method of easily reextracting the 1-D spectrum from the 2-D, adapting the routine extract1d.pro from the DEIMOS DEEP2 pipeline code \citep{Marinoni01} for this purpose. To reextract, the user places the cursor over the 2-D spectrum at the vertical position desired, then clicks, drags horizontally (by any amount, in either direction) and releases. This triggers the reextraction, and when it is complete the 1-D plot is updated with the newly extracted spectrum. A dialog box lets the user save the extracted 1-D spectrum in .fits format.
\subsection{SED}
The source SED is displayed in the bottom right window of the interface. The information for this plot comes from the photometry file, described in Section 3.5. The plot extends from 0.1~to~10~$\mu$m and displays photometric AB magnitudes with errors. The effective widths of the photometric bands are indicated with horizontal bars and errors in the magnitude estimates are indicated with vertical bars. Observed-frame positions of important SED features are indicated with vertical blue lines. These features include the Lyman limit at 912~\AA, Ly$\alpha$ at 1216~\AA, the 4000~\AA~break, and the typical SED peak of evolved galaxies at 1.6~$\mu$m. The wavelength coverage of the spectrum is indicated with vertical green lines.
\subsection{Cross-Correlation}
Automated redshift determination by convolution against spectral templates is a powerful and time-saving capability. We have adapted cross-correlation routines originally written for the SDSS spectral reduction package for use in SpecPro, with a library of spectral templates available to correlate against (see Table~1). The templates span a broad range of galaxy types. We have also included some stellar templates to help identify high-redshift selected sources that are actually stars.
The user is expected to be able to choose an appropriate template based on the spectrum and other information about the source. When one of the galaxy templates from the template library is selected, the six best redshift matches are computed and the best solution is displayed. The droplist ``Auto-z solution:" under the 1-D spectrum contains all of the solutions for examination. Clicking an alternate solution will switch to the new redshift guess and move the template accordingly. Because the spectral template is overplotted on the 1-D spectrum, the accuracy of the result can be easily checked. The template can be scaled via the ``Template scaling:'' button to provide a better visual match to the spectrum.
In some cases cross-correlation fails to find the solution, even though the correct answer is apparent to the user. In this case the redshift can be adjusted manually to find the answer, but finding the exact solution can be time consuming. Therefore we added an ``Auto center" button that cross-correlates against the currently selected galaxy template in a small region (z~$\pm$~0.05) around the current redshift guess.
In some very difficult cases cross-correlation will fail to find the correct redshift, even using auto-centering. In these unusual instances the user must resort to adjusting the redshift manually to the correct answer.
\begin{deluxetable*}{ll}
\tabletypesize{\scriptsize}
\tablecaption{Summary of the spectral templates available for cross-correlation in SpecPro.}\label{Ta:templates}
\tablehead{\colhead{\bf Template} & \colhead{\bf Description}}
\startdata
VVDS LBG & Lyman-break galaxy template from the VIRMOS-VLT Deep Survey \citep{Lefevre05} \\
VVDS Elliptical & Passive galaxy template, strong absorption, no emission \\
VVDS S0 & S0 classification template, weak emission, absorption \\
VVDS Early Spiral & Stronger emission, less absorption \\
VVDS Spiral & Typical spiral spectrum, with strong emission features \\
VVDS Starburst & Extremely strong nebular emission features \\
SDSS Quasar & Typical broad-line quasar spectrum from the Sloan Digital Sky Survey \citep{Schneider10} \\
Red Galaxy & Passive galaxy template from PEGASE spectral evolution model \citep{Fioc97} \\
Green Galaxy & Early spiral/spiral template from PEGASE \\
Blue Galaxy & Spiral/Starburst template from PEGASE \\
LBG Shapley & Lyman-break galaxy template from Shapley 2003 \citep{Shapley03} \\
SDSS LoBAL & Low-ionization broad absorption line (BAL) quasar template \citep{Reichard03} \\
SDSS HiBAL & High-ionization BAL quasar template \citep{Reichard03} \\
A0-M6 & Stellar templates from the Pickles catalog \citep{Pickles98} \\
\enddata
\end{deluxetable*}
\subsection{Output}
The interface allows the user to rapidly record results to a formatted file. To the left of the SED plot window are four output fields with which the user can save the redshift, a confidence associated with it, their initials, and notes on the spectrum. The output is saved in ascii format and contains the mask name, source name, slit number, RA, and DEC as well as the redshift result. Results are always appended to the end of the selected output file and multiple entries for a single source can be added. Once an output file has been selected, it is the default output location for the rest of the SpecPro session.
The ``Save Spec1D" button saves the displayed 1-D spectrum (wavelength and flux at the current zoom, bin, and smooth level) to an ascii file for plotting in other programs. In addition, images in .tif format of any of the data windows can be saved by double-clicking on them.
\section{Data Formats}
We have attempted to make the required formats simple and general so that data from any instrument/survey can be easily adapted for use with SpecPro. While the files are expected to be arranged by slit number, any sequential numbering scheme can be used to differentiate sources. None of the inputs accepted by SpecPro are actually required for the program to run; if one is missing, the corresponding field in the interface is left blank. The program accepts the following five files$\colon$~\begin{itemize}
\item 1-D spectrum file (fits)
\item 2-D spectrum file (fits)
\item Stamp image file (fits)
\item Photometry file (ascii)
\item Information file (ascii)
\end{itemize}
These are briefly described in the rest of this section. More detailed descriptions are given on the website, and example data is provided with the program download.
\subsection{1-D Spectrum}
The 1-D spectrum is a stored as a structure with three fields$\colon$ \emph{flux}, \emph{ivar}, and \emph{lambda}. Each field holds a one-dimensional double array. The \emph{flux} field is the spectrum, the \emph{ivar} field is the inverse variance of the flux, and \emph{lambda} is the wavelength at each pixel, in angstroms. Note that the units of flux are arbitrary, but values proportional to F$_\nu$ are preferable because this matches the units of the spectral templates. The structure must be saved in FITS format.
\subsection{2-D Spectrum}
The 2-D spectrum is also stored in a structure with the three fields \emph{flux}, \emph{ivar}, and \emph{lambda}. However, each field here is a two-dimensional double array, including the \emph{lambda} field, which contains the wavelength solution for each pixel in the 2-D spectrum. The 2-D spectrum is stored in FITS format.
\subsection{Stamp Images}
Stamp images of the source are stored in a single FITS file, with the data stored as an array of structures. Each element in the array contains a stamp image and ancillary information, with the size of the array determined by how many stamps are available. The stamp structures must all be of the same form, with fields as follows$\colon$
\begin{itemize}
\item \emph{name} - string giving name of the image, e.g. ``U band"
\item \emph{flux} - 100x100 double array
\item \emph{ivar} - 100x100 double array
\item \emph{RA} - image center in degrees, double
\item \emph{DEC} - image center in degrees, double
\item \emph{pixscale} - arcseconds per pixel, double
\end{itemize}
The stamps should be oriented with north up. If the inverse variance for the image is not available then it should either be approximated in some way or set to all zeros.
\subsection{Information File}
The information file contains ancillary information about a source displayed by the interface and/or used in some of its functionality. This information includes source RA and DEC, slit information, photometric redshift estimate, and other quantities (see the website for a detailed description). If the user only has a subset of the information they can omit the missing values, or set them to zero.
\subsection{Photometry}
The photometry file is a space delimited file with five columns: filter name, filter central wavelength, effective filter bandwidth, AB magnitude, and AB magnitude error. Both the filter central wavelength and effective filter bandwith should be expressed in angstroms. Non-detections are indicated with an AB magnitude of -99, with the AB magnitude error set to the lower bound on the AB magnitude established by the non-detection. Any photometric point can be included, but the program's SED plot extends from 0.1 to 10.0~$\mu$m.
\section{Summary}
SpecPro is an interactive IDL-based interface particularly suited to viewing astronomical spectra in the context of multiwavelength surveys. In addition to 1-D and 2-D spectra, the interface can display stamp images and the spectral energy distribution of the source. Having all of this information in one display makes it much easier to identify the nature and redshift of faint spectral targets.
We have made an effort to make the data formats required by the program both simple and general. The program could potentially incorporate wrappers for various instrument's data or allow more flexible inputs; however, the effort required to achieve this flexibility is probably too high considering the relative ease with which data can be made to conform with the formats outlined in Section 3.
One possibility for future development is to incorporate tools for more detailed spectroscopic analysis (line-fitting, determination of equivalent widths, etc.) from within the interface. This was not the our intention when designing SpecPro, but the addition of such functionality would make it a much more general tool for spectral analysis. Another related possibility would be to allow the user to interactively fit the source SED from within the interface, in order to determine galaxy stellar mass, age, and other parameters. However, at present the interface is optimized for spectral classification and redshift determination and we have no concrete plans for further development.
We have found the program valuable in our own analysis of DEIMOS spectra of targets in the COSMOS survey field. In addition to making the identification of redshifts very fast, it has also allowed us to efficiently characterize our sample and identify particularly interesting galaxies for follow-up study. We feel that SpecPro will prove useful to others working with spectroscopic samples as well.
\acknowledgments
The authors would like to thank Dr. Yuko Kakazu, Dr. Hai Fu, Dr. Lin Yan, and Dr. Nick Scoville from the California Institute of Technology for their helpful comments and suggestions during the development of the software. In addition, we would like to acknowledge and thank Dr. Mara Salvato of the Max Planck Institute for Plasma Physics and Dr. Francesca Civano of the Harvard-Smithsonian Center for Astronomy for providing useful feedback on early versions of the code. We also thank Dr. Bahram Mobasher of the University of California, Riverside for carefully reading a draft of this paper and making suggestions that significantly improved its content.
SpecPro makes use of IDL code written and maintained by others. We thank Wayne Landsman at NASA/GSFC for his work maintaining the IDL Astronomy User's Library. We also thank the authors of code we have used for our implementation of automated cross-correlation, including David Schlegel, Doug Finkbeiner, Michael Cooper, and John Johnson. We thank Craig Markwardt, whose MPFIT least-squares fitting package \citep{Markwardt09} is integral to the cross-correlatation functionality.
Finally, we thank an anonymous referee for carefully reading this manuscript and providing very constructive feedback both in terms of content and with regard to the distribution of the software on the web.
This work was supported in part by a visiting graduate student fellowship at the Caltech Infrared Processing and Analysis Center (IPAC).
\bibliographystyle{plainnat}
|
\section{Introduction}
\par{H. Goguen \cite{healf:thesis,healf:TLCA99YY} has developed a method called \emph{typed operational semantics} (TOS for short) to prove meta-theoretic properties of type theories, including strong normalisation, Church-Rosser and subject reduction. In this paper, using the TOS approach, we study the meta-theoretic properties of a type system with dependent record types. }
\par{A record type is a type of labelled tuples called records. A dependent record type (DRT) is a type of records whose fields may have types that depend on the values of earlier fields. Dependent records have been studied previously for various different type systems \cite{Harper-Lillibridge93, bet-tar:subtyping98, Pollack:records02, ctp:semantic-records05}, with applications to the study of module mechanisms for both programming and proof languages. Recently, in the context of studying manifest fields of module types, the second author has proposed a formulation of dependent record types \cite{luo:TYPES08}, for type theories with canonical objects such as \ML's type theory, and shown in \cite{luo:MLPA09} that, in some applications, dependent record types are more useful than $\Sigma$-types (dependent types of tuples without labels). }
Studying the meta-theory of dependent record types, the contributions of the current paper are two-fold. First of all, the meta-theory of dependent record types has not been well-studied. This work makes a positive contribution, showing that our formulation of dependent record types has the good meta-theoretic properties such as strong normalisation. Secondly, the type theory we study has record \emph{types} as studied in \cite{Pollack:records02,luo:TYPES08}, rather than record \emph{kinds} as in \cite{bet-tar:subtyping98,ctp:semantic-records05}. Since types have a much more sophisticated structure than kinds, the meta-theory for dependent record types is expected to be much more difficult than that for dependent record kinds as found in, {e.g.}, \cite{ctp:semantic-records05}. We shall study the meta-theory by taking the TOS approach, which is shown to be robust enough to deal with dependent record types. In particular, we study the \emph{intensional} DRTs, that is, the dependent record types without the so-called weakly extensional rules (these rules are considered in \cite{luo:TYPES08}). The typed operational semantics for intensional DRTs is developed and shown to be sound and complete and, based on this, it is proved that the intensional DRTs have good meta-theoretic properties, including strong normalisation, Church-Rosser and subject reduction.
The paper is arranged as follows. The type system IDRT for intensional DRTs is described in Section~\ref{sec:LFwithDRT}. In Section~\ref{sec:TOS}, after introducing the basic idea of TOS, we define the TOS for dependent record types. The properties of the TOS are studied in Section~\ref{sec:metaTOS} and the meta-theoretic properties of IDRT in Section~\ref{sec:metaDRT}. Discussions of related work and future work are given in the conclusion.
\section{Dependent Record Types}
\label{sec:LFwithDRT}
A dependent record type is a type of labelled tuples whose fields may have types that depend on the values of earlier fields. For instance, if $Nat$ and $Vect(n)$ are the types of natural numbers and vectors of length $n$, respectively, $\record{n\colon Nat,\ v\colon Vect(n)}$ is the dependent record type with objects (called {\sl records}) such as $\record{n=2,\ v=[5,6]}$, where dependency is respected: the vector $[5,6]$ must be of type $Vect(2)$.
Formally, in our study, dependent record types are formulated as an extension of the logical framework that we describe briefly first.
\paragraph{Logical Framework.} LF \cite{luo:book94} is the typed version of Martin-L\"{o}f's logical framework \cite{NPS:book}. It is itself a type system that serves as a meta-language to specify type theories such as \ML's intensional type theory \cite{NPS:book} and the Unifying Theory of dependent Types (UTT) \cite{luo:book94}. Here, we give only a brief introduction, fixing the notations to be used in the paper. (For details of, for example, how inductive types like $Nat$, $\Pi$-types and $\Sigma$-types can be specified in the logical framework, see Part III of \cite{NPS:book} or Chapter 9 of \cite{luo:book94}.)
In LF, the syntactical entities \emph{contexts}, \emph{kinds} and \emph{terms} are of the following forms:
\begin{eqnarray*}
Contexts \ & \ \Gamma & ::= \ \ \ () \ \ | \ \ \Gamma, x \colon A \\
LF\ Kinds \ & \ K & ::= \ \ \ Type \ \ | \ \ El(A) \ \ | \ \ (x:K)K' \\
LF\ Terms \ & \ M & ::= \ \ \ x \ \ | \ \ [x:K]M \ \ | \ \ M(M')
\end{eqnarray*}
The types in LF are called \emph{kinds}, including:
\begin{itemize}
\item $Type$ -- the kind representing the collection of all types ($A$ is a type if $A \colon Type$);
\item $El(A)$ -- the kind of objects of type $A$ (we often omit $El$); and
\item $(x:K)K'$ (or simply $(K)K'$ when $x \notin FV (K')$) -- the kind of dependent functional operations.
\end{itemize}
The judgement forms in LF include, for example,
\begin{itemize}
\item $\Gamma\ts k\colon K$, which asserts that $k$ is an object of kind $K$; and
\item $\Gamma\ts k=k'\colon K$, which asserts that $k$ and $k'$ are (computationally) equal objects of kind $K$.
\end{itemize}
The inference rules of LF to define the typing relation and the computational equality are given in Appendix~\ref{app:LF-rules}. In particular, $\beta\eta$-equal objects are computationally equal. For instance, an abstraction $[x:K]M$ can be applied to form $([x:K]M)(a)$ that is computationally equal to $[a/x]M$.
\begin{notation}
We shall use $\equiv$ to denote the syntactical identity (up to $\alpha$-conversion).
\end{notation}
\paragraph{Dependent Record Types.}
We now give a formal presentation of the system IDRT of intensional dependent record types, which is an extension of LF. The syntax of this type system is given as follows, where $\mathcal{L}$ is an (infinite) set of labels, $l\in \mathcal{L}$ and $L\subset\mathcal{L}$ is finite:
\begin{eqnarray*}
Kinds\ of\ Record\ Types \ & K_R & ::= \ \ \ RType \ | \ RType[L] \\
Record\ Types \ & \ R & ::= \ \ \ \langle \rangle \ | \ \langle R, \ l \colon A \rangle \\
Records \ & \ r & ::= \ \ \ \langle \rangle \ | \ \langle r,\ l=a \colon A \rangle
\end{eqnarray*}
The inference rules of IDRT consist of the rules for LF (Appendix~\ref{app:LF-rules}) and the additional rules in Figure~\ref{DRT-rules}. Here are some informal explanations.
\begin{figure}[top]
\framebox[5.8in][l]{
\begin{minipage}{\linewidth}
\ \\
\ \ \ \emph{Kinds of record types}
$$
\frac{\Gamma \: valid}{\Gamma \vdash RType \: kind}\: \: \: \:
\frac{\Gamma \: valid \: \: \:
}{\Gamma \vdash RType[L] \: kind}
$$
$$
\frac{\Gamma \vdash R \colon RType[L] \: \: \: L \subseteq L'
}{\Gamma \vdash R \colon RType[L']} \: \: \: \:
\frac{\Gamma \vdash R \colon RType[L]}{\Gamma \vdash R \colon RType} \: \: \: \:
\frac{\Gamma \vdash R \colon RType}{\Gamma \vdash R \colon Type}
$$
\ \ \ \emph{Formation rules}
$$
\frac{\Gamma \: valid}{\Gamma \vdash \langle \rangle \colon RType[\emptyset]}
\: \: \: \:
\frac{\Gamma \vdash R \colon RType[L] \: \: \: \Gamma \vdash A \colon (R)Type \: \: \: l \notin L}{\Gamma \vdash \langle R, \ l \colon A \rangle \colon RType[L \cup \{l\}]}
$$
\ \ \ \emph{Introduction rules}
$$
\frac{\Gamma \: valid}{\Gamma \vdash \langle \rangle \colon \langle \rangle}
\: \: \: \:
\frac{\Gamma \vdash \langle R, \ l\colon A \rangle \colon RType \: \: \: \Gamma \vdash r \colon R \: \: \: \Gamma \vdash a \colon A(r)}{\Gamma \vdash \langle r, \ l=a \colon A \rangle \colon \langle R, \ l \colon A \rangle}
$$
\ \ \ \emph{Elimination rules}
$$
\frac{\Gamma \vdash r \colon \langle R, \ l \colon A \rangle}{\Gamma \vdash [r]\colon R}
\: \: \: \:
\frac{\Gamma \vdash r \colon \langle R, \ l \colon A \rangle}{\Gamma \vdash r.l \colon A([r])}
$$
$$
\frac{\Gamma \vdash r \colon \langle R, \ l \colon A \rangle \: \: \: \Gamma \vdash [r].l' \colon B \: \: \: l \neq l'}{\Gamma \vdash r.l' \colon B}
$$
\ \ \ \emph{Computation rules}
$$
\frac{\Gamma \vdash \langle r, \ l = a \colon A \rangle \colon \langle R, \ l \colon A \rangle}{\Gamma \vdash [\langle r, \ l = a \colon A \rangle] = r \colon R}
\: \: \: \:
\frac{\Gamma \vdash \langle r, \ l = a \colon A \rangle \colon \langle R, \ l \colon A \rangle}{\Gamma \vdash \langle r, \ l = a \colon A \rangle . l = a \colon A(r)}
$$
$$
\frac{\Gamma \vdash r \colon \langle R, \ l \colon A \rangle \: \: \: \Gamma \vdash [r].l' \colon B \: \: \: l \neq l'}{\Gamma \vdash r.l' = [r].l' \colon B}
$$
\ \ \ \emph{Congruence rules for record types}
$$
\frac{\Gamma \: valid}{\Gamma \vdash \langle \rangle = \langle \rangle \colon RType[\emptyset]}
\: \: \: \:
\frac{\Gamma \vdash R = R' \colon RType[L] \: \: \: \Gamma \vdash A = A' \colon (R)Type \; \: \: l \notin L}{\Gamma \vdash \langle R, \ l \colon A \rangle = \langle R', \ l \colon A' \rangle \colon RType[L \cup \{l\}]}
$$
\ \ \ \emph{Congruence rules for records}
$$
\frac{\Gamma \: valid}{\Gamma \vdash \langle \rangle = \langle \rangle \colon \langle \rangle}
\: \: \: \:
\frac{\begin{array}{c} \Gamma \vdash R \colon RType[L] \: \: \: l \notin L \\ \Gamma \vdash r = r' \colon R \: \: \: \Gamma \vdash a = a' \colon A(r) \: \: \: \Gamma \vdash A = A' \colon (R)Type \end{array}}{\Gamma \vdash \langle r, \ l = a \colon A \rangle = \langle r', \ l = a' \colon A' \rangle \colon \langle R, \ l \colon A \rangle}
$$
$$
\frac{\Gamma \vdash r = r' \colon \langle R, l \colon A \rangle}{\Gamma \vdash [r] = [r'] \colon R}
\: \: \: \:
\frac{\Gamma \vdash r = r' \colon \langle R, l \colon A \rangle}{\Gamma \vdash r.l = r'.l \colon A([r])}
$$
\\
\end{minipage}
}
\caption{Inference Rules of IDRT} \label{DRT-rules}
\end{figure}
\begin{itemize}
\item We add new kinds $RType$ and $RType[L]$ of record types. Intuitively, $RType[L]$ is the kind of the record types whose (top-level) labels are all in $L$, a finite set of labels. Naturally, if $L\subseteq L'$, every record type in $RType[L]$ is also in $RType[L']$. The kind $RType$ is the kind of all record types and could conceptually be understood as `$RType[\mathcal{L}]$'. Finally, every record type is also a type. These are formally reflected in the rules for the kinds of record types in Figure~\ref{DRT-rules}.
\item Record types are types of the form $\record{}$ or $\record{R,\ l\colon A}$. Intuitively, a record type is of the form $\record{l_1\colon A_1,\ ...,\ l_n\colon A_n}$,\footnote{We overload the $\record{\ ...\ }$ notation for records and their types. It is always possible to distinguish between the two.} where each $l_i\colon A_i$ is a \emph{field} labelled by $l$. An object of this record type is a labelled tuple $\record{l_1 = a_1\colon A_1,\ ..., \ l_n = a_n\colon A_n}$, where $a_i$ is of the type of the corresponding field. \selfcomment{The kind of the field types are dependent kinds from \emph{previous} record types to $Type$.} Note that, formally, each $A_i$ in the record type is not a type, but a family of types; this is how dependency is incorporated -- we have dependent record types.
Notation-wise, we shall adopt the following notational conventions: for record types, we write $\langle l_1 \colon A_1,\ ...,\ l_n \colon A_n \rangle$ for $\langle \langle \langle \rangle,\ l_1 \colon A_1 \rangle,\ ...,\ l_n \colon A_n \rangle$ and often use label occurrences/non-occurrences to show dependency/non-dependency respectively.
For instance, we write $\langle n \colon Nat, \ v \colon Vect(n) \rangle$ for $\langle \langle \langle \rangle,\ n \colon NAT \rangle, \ v \colon [x:\langle n \colon NAT \rangle] Vect(x.n) \rangle$ where $NAT \equiv [\_\ :\langle \rangle]Nat$, and $\langle R, \ l \colon Vect(2) \rangle$ for $\langle R, \ l \colon [\_\ :R]Vect(2) \rangle$.
\item There are two operations on records: \emph{restriction} (or first projection) $[r]$ that removes the last component of record $r$ and \emph{field selection} $r.l$ that selects the value of the field labelled by $l$. For instance, intuitively, for the record $r \equiv \record{l_1 = a_1\colon A_1,\ l_2 = a_2\colon A_2, \ l_3 = a_3\colon A_3}$ of type $\record{l_1\colon A_1,\ l_2\colon A_2,\ l_3\colon A_3}$, we have $[r] = \record{l_1\colon A_1,\ l_2\colon A_2}$ and $r.l_2 = [r].l_2 = a_2$. These are formally reflected in the introduction, elimination and computation rules in Figure~\ref{DRT-rules}.
\item The congruence rules for record types and records in Figure~\ref{DRT-rules} propagate the computational equality through the term structure. Also, we do not include the weakly extensional equality rules as considered in \cite{luo:TYPES08}. Therefore, we call the system the type system for intensional DRTs.
\end{itemize}
We shall adopt the following terminology: the terms of the form $\record{r,\ l=a\colon A}$ will be called \emph{pair-records}. (For example, we shall use this terminology in specifying the TOS-rules for record types in Figure~\ref{TOS-DRT-rules} in Section~\ref{sec:TOS-IDRT}.)
\paragraph{Record types v.s. record kinds.} It is worth pointing out that our type system contains dependent record \emph{types} (as studied by Pollack \cite{Pollack:records02}, Luo \cite{luo:TYPES08,luo:MLPA09} and the current paper), rather than dependent record \emph{kinds} (as studied by Betarte and Tasistro \cite{bet-tar:subtyping98} and Coquand, Pollack and Takeyama \cite{ctp:semantic-records05}\footnote{\emph{Types} in the terminology of \ML's type theory are what we call \emph{kinds} in this paper. Therefore, the so-called record types in \cite{bet-tar:subtyping98} and \cite{ctp:semantic-records05} are really record kinds.}). We would like to distinguish these two notions clearly: in a type theory with inductive types, types include those such as $Nat$ of natural numbers and $\Sigma$-types of dependent pairs, while the examples of kinds include, for example, the kind $Type$ of all types. They exist at two completely different levels and have rather different structures and properties.
In general, types have a much more sophisticated and richer structure than kinds. For instance, it is easy to show that a kind is of the form either $Type$ or $(x:K)K'$, but types are not ({e.g.}, a type may be of the form $f(a)$). To appreciate the difference, let us consider the issue of ensuring label distinctness. If one considers only record kinds, it is easy to guarantee that the labels in the same record kind are distinct because of the limited syntactic forms of kinds (see, for example, \cite{ctp:semantic-records05}). However, this is not easy at all for record types (think, for example, how one ensures that a label does not occur in a type of the form $f(a)$). In our case, we have to introduce the kinds $RType[L]$ to ensure that it is the case that the (top-level) labels in the same record type are distinct. In other words, intuitively, $l\not= l'$ for any record type $\record{...,\ l\colon A,\ ...,\ l'\colon A',\ ...}$. This is guaranteed by means of the side condition $\l\not\in L$ of the second formation rule in Figure~\ref{DRT-rules}.
That a type system with record types is more powerful than one with only record kinds can be understood from another angle when one wants to introduce universes of record types. It is possible to introduce type universes for dependent record types, as shown in \cite{luo:MLPA09}; this, however, cannot be done for record kinds. Therefore, record types are more useful than record kinds (for example, in representing module types in data refinement \cite{luo:MLPA09}).
Since types have a more sophisticated structure than kinds, it is more difficult to study the meta-theoretic properties of a system with record types, as compared with a meta-theoretic study of record kinds. As we show in this paper, the approach of using typed operational semantics can be used in this endeavour.
\section{Typed Operational Semantics for Dependent Record Types}
\label{sec:TOS}
The typed operational semantics (TOS for short) is a proof-theoretic method to prove the meta-theoretic properties of type theories. It was developed by H. Goguen in his PhD thesis \cite{healf:thesis}, where he studied the meta-theory of UTT and proved that UTT has the nice properties such as Church-Rosser, Subject Reduction and Strong Normalisation.
In this paper, the TOS approach is applied to study the meta-theory of dependent record types. After a brief informal introduction of the approach, we develop the typed operational semantics for the system IDRT of intensional DRTs and show that it has the soundness and completeness properties. The meta-theoretic properties of dependent record types are studied in the next section.
\subsection{The TOS Approach}
\label{sec:introTOS}
For a type theory, its typed operational semantics captures its computational behaviour, usually given by its (untyped) reduction relation. For example, in TOS, the following judgement
\[ \Gamma\models M\to N\to P\colon A \]
informally asserts that, among other things, $N$ and $P$ are the weak-head normal form and the normal form of the term $M$, respectively.\footnote{Formally, the reduction relation and the TOS are related to each other by means of the `adequacy theorems' such as Lemmas~\ref{AUR} and~\ref{ANF} for IDRT in Section~\ref{subsec: adeq}.} For the logical framework LF, for example, its corresponding TOS has been studied \cite{healf:TLCA99YY} and its inference rules are given in Appendix~\ref{app:LF-TOSrules}. Since many meta-theoretic properties of a type theory are concerned with its computational behaviour, it is not a surprise that TOS provides an effective approach to the meta-theory of type theories.\footnote{It is worth noting that, although it is useful to study the meta-theory for many type theories, the TOS approach would not be suitable for non-normalising type theories. See \cite{healf:thesis} for discussions.}
The TOS and its corresponding type theory are related to each other by means of the soundness and completeness theorems. Using the judgement $\Gamma\models M\colon A$ to abbreviate `$\Gamma\models M\to N\to P\colon A$ for some $N$ and $P$', we can state the soundness and completeness properties as follows:
\begin{itemize}
\item Soundness: $\Gamma\ts M\colon A$ implies $\Gamma\models M\colon A'$ (for $A'$ that is the `normal form' of $A$).
\item Completeness: $\Gamma\models M\colon A$ implies $\Gamma\ts M\colon A$.
\end{itemize}
Based on soundness and completeness, we can prove many meta-theoretic properties of the type theory. For example, it can be shown that, if $\Gamma\models M\colon A'$, then $M$ is strongly normalisable. Therefore, strong normalisation, the property that every well-typed term is strongly normalisable, can be proved by means of such a fact together with the soundness property, as pictured as follows:
\begin{picture}(100, 80)(-60, -50)
\put(30, 10){\makebox(0,0){$\Gamma \ts M \colon A$}}
\put(120, -30){\makebox(0,0){$\Gamma \models M \colon A'$}}
\put(210, 10){\makebox(0,0){$M$ is SN}}
\put(40, -30){\line(0, 1){35}}
\put(40, -30){\vector(1, 0){50}}
\put(150, -30){\line(1, 0){60}}
\put(210, -30){\vector(0, 1){35}}
\put(60, 10){\dashbox{.5}(120, 0)[t]{}
\put(180, 10){\vector(1, 0){3}}
\put(120, 14){\makebox(0,0){\emph{\begin{small}Strong Normalisation\end{small}}}}
\put(15, -15){\makebox(0,0){\emph{\begin{small}Soundness\end{small}}}}
\put(245, -15){\makebox(0,0){\emph{\begin{small}SN for TOS\end{small}}}}
\end{picture}
\noindent As shown in this paper, for dependent record types, the SN property for the corresponding TOS is proven in Theorem~\ref{SN-TOS}. Then, by the Soundness Theorem (Theorem~\ref{Soundness-TOS-LF}), we can show that strong normalisation for IDRT (Corollary~\ref{SN-LFDRT}).
Note that, to implement such ideas is not a simple matter: it requires one to prove:
\begin{itemize}
\item that the TOS is `adequate' {w.r.t.}\ the (untyped) reduction relation,
\item that the TOS is sound and complete {w.r.t.}\ the original type theory, and
\item that the TOS satisfies some specific meta-theoretic properties ({e.g.}, strong normalisation).
\end{itemize}
Then, one can transfer the results to the original type theory to show that it has nice meta-theoretic properties. This is what we shall do for IDRT, the type theory with dependent record types.
\subsection{TOS for Dependent Record Types}
\label{sec:TOS-IDRT}
The typed operational semantics for dependent record types is described in this section. The judgement forms in a TOS are given in Figure~\ref{TOS-judgements},
\begin{figure}[top]
\begin{minipage}{\linewidth}
\begin{eqnarray*}
Basic\ forms: \ \ \ & \ \ \ \ \ \ \ \ Abbreviated\ forms: \\
\models \Gamma \rightarrow \Delta \ \ \ & \ \ \ \ \ \ \ \ \Gamma \models\ ok \\
\Gamma \models A \rightarrow B \ \ \ & \ \ \ \ \ \ \ \ \Gamma \models M \rightarrow_{w} N \colon A \\
\Gamma \models M \rightarrow N \rightarrow P \colon A \ \ \ & \ \ \ \ \ \ \ \ \Gamma \models M \rightarrow_{n} P \colon A \\
\ \ \ & \ \ \ \ \ \ \ \ \Gamma \models M \colon A
\end{eqnarray*}
\end{minipage}
\caption{Judgement Forms in Typed Operational Semantics} \label{TOS-judgements}
\end{figure}
three of which are the basic forms of judgements whose informal meanings are:
\begin{itemize}
\item $\models \Gamma \rightarrow \Delta$: the context $\Gamma$ has context $\Delta$ as its normal form;
\item $\Gamma \models A \rightarrow B$: the kind $A$ is well-formed in context $\Gamma$ and has normal form $B$; and
\item $\Gamma \models M \rightarrow N \rightarrow P \colon A$: the terms $M$, $N$, $P$ are well-formed in context $\Gamma$ of kind $A$ and $M$ has weak-head normal form $N$ and normal form $P$.
\end{itemize}
From these basic judgements, one can define other forms of judgements, including the following:
\begin{itemize}
\item $\Gamma \models\ ok$ stands for `$\models \Gamma \rightarrow \Delta$ for some $\Delta$';
\item $\Gamma \models M \rightarrow_{w} N \colon A$ stands for `$\Gamma \models M \rightarrow N \rightarrow P \colon A$ for some $P$';
\item $\Gamma \models M \rightarrow_{n} P \colon A$ stands for `$\Gamma \models M \rightarrow N \rightarrow P \colon A$ for some $N$'; and
\item $\Gamma \models M \colon A$ stands for `$\Gamma \models M \rightarrow N \rightarrow P \colon A$ for some $N$ and $P$'.
\end{itemize}
The typed operational semantics for the type system IDRT of intensional DRTs is the extension of that for LF (Appendix~\ref{app:LF-TOSrules}) with the inference rules given in Figure~\ref{TOS-DRT-rules}.
\begin{figure}[top]
\framebox[6.2in][l]{
\begin{minipage}{\linewidth}
\ \\
\ \ \ \emph{Record Kinds}
$$
\frac{\Gamma \models ok}{\Gamma \models RType \rightarrow RType} \ RTYPE \: \: \: \: \: \:
\frac{\Gamma \models ok \: \: \:
}{\Gamma \models RType[L] \rightarrow RType[L]} \ RTYPE[L]
$$
\ \ \ \emph{Record Types}
$$
\frac{\Gamma \models ok}{\Gamma \models \langle \rangle \rightarrow \langle \rangle \rightarrow \langle \rangle \colon RType[\emptyset]} \ \ EMP_{RCDT}
$$
$$
\frac{\Gamma \models R \rightarrow_n P \colon RType[L] \: \: \: \: \Gamma \models A \rightarrow_n B \colon (P)Type \: \: \: \: l \notin L}{\Gamma \models \langle R, l \colon A \rangle \rightarrow \langle R, l \colon A \rangle \rightarrow \langle P, l \colon B \rangle \colon RType[L \cup \{l\}]} \ \ RCDT
$$
\ \ \ \emph{Pair-records}
$$
\frac{\Gamma \models ok}{\Gamma \models \langle \rangle \rightarrow \langle \rangle \rightarrow \langle \rangle \colon \langle \rangle} \ \ EMP_{RCD}
$$
$$
\frac{\begin{array}{c} \Gamma \models \langle R, l \colon A \rangle \rightarrow_n \langle P, l \colon B \rangle \colon RType \: \: \: \: \Gamma \models r \rightarrow_n p \colon P \\ \Gamma \models A(r) \rightarrow_n C \colon Type \: \: \: \Gamma \models a \rightarrow_n b \colon C \: \: \: \:
\end{array}}{\Gamma \models \langle r, l=a \colon A \rangle \rightarrow \langle r, l=a \colon A \rangle \rightarrow \langle p, l=b \colon B \rangle \colon \langle P, l \colon B \rangle} \ RCD
$$
\ \ \ \emph{Restrictions}
$$
\frac{\Gamma \models r \rightarrow q \rightarrow p \colon \langle P, l \colon B \rangle \: \: \: \: p,\ q\ not\ pair\texttt{-}records}{\Gamma \models [r] \rightarrow [q] \rightarrow [p] \colon P} \ \ BASE_{RESTR}
$$
$$
\frac{\Gamma \models r \rightarrow_w \langle p, l = b \colon A \rangle \colon \langle P, l \colon B \rangle \: \: \: \: \Gamma \models p \rightarrow s \rightarrow t \colon P \selfcomment{\: \: \: \: \Gamma \models P \rightarrow_n C \colon Q}}{\Gamma \models [r] \rightarrow s \rightarrow t \colon P} \ RESTR
$$
\ \ \ \emph{Selections}
$$
\frac{\begin{array}{c} \Gamma \models r \rightarrow q \rightarrow p \colon \langle P, l \colon B \rangle \: \: \: \: p,\ q\ not\ pair\texttt{-}records \\ \Gamma \models B([r]) \rightarrow_n C \colon Type \end{array}}{\Gamma \models r.l \rightarrow q.l \rightarrow p.l \colon C} \ BASE_{FLDSEL}
$$
$$
\frac{\Gamma \models r \rightarrow_w \langle p,\ l = b \colon A \rangle \colon \langle P,\ l \colon B \rangle \: \: \: \: \Gamma \models b \rightarrow c \rightarrow d \colon C \: \: \: \: \Gamma \models A(p) \rightarrow_n C \colon Type}{\Gamma \models r.l \rightarrow c \rightarrow d \colon C} \ FLDSEL
$$
$$
\frac{\begin{array}{c} \Gamma \models r \rightarrow_n s \colon \langle P, l \colon B \rangle \: \: \: \: \Gamma \models [r].l' \rightarrow c \rightarrow d \colon C \: \: \: \: \: l \neq l' \end{array}}{\Gamma \models r.l' \rightarrow c \rightarrow d \colon C}\ FLDSL'
$$
\\
\end{minipage}
\caption{Inference Rules of Typed Operational Semantics for IDRT} \label{TOS-DRT-rules}
\end{figure}
Most of the rules are self-explanatory. We only mention that, besides using the abbreviated forms of judgement (see above) in the rules, we also use the terminology of `pair-record' as introduced in Section~\ref{sec:LFwithDRT}. For example, in $(BASE_{RESTR})$, we require that $p$ or $q$ be not a pair-record, for otherwise, for instance, $[p]$ could be a redex and would not be in normal form.
\section{Properties of TOS for Dependent Record Types}
\label{sec:metaTOS}
We shall study the properties of the TOS for IDRT, as presented above in Section~\ref{sec:TOS-IDRT}. These include those properties {w.r.t.}\ the relationship with IDRT (soundness and completeness) and those {w.r.t.}\ the reduction relation.
\subsection{Basic Structural Properties}
\label{subsec:struc}
The typed operational semantics satisfy some basic properties as stated in the following lemma, which can all be proved by induction on the TOS-derivations.\footnote{Some of the lemmas ({e.g.}, the strengthening lemma) can only be proved by proving a stronger statement by induction on derivations. We omit the details here.}
\begin{lemma} \label{Stru-prop}\
\begin{enumerate}
\item (Context Validity) Any derivation of $\Gamma_0, \Gamma_1 \models J$ has a sub-derivation of $\Gamma_0 \models ok$.
\item (Variables) Let $dom\{\Gamma\}$ be the set of variables declared in context $\Gamma$ and $FV(M)$ the set of free variables occurring in term $M$.
\begin{enumerate}
\item If $\models \Gamma \rightarrow \Delta$, then $dom\{\Delta\} = dom\{\Gamma\}$.
\item If $\Gamma \models A \rightarrow B$, then $FV(A) \cup FV(B) \subseteq dom\{\Gamma\}$.
\item If $\Gamma \models M \rightarrow N \rightarrow P \colon A$, then $FV(M) \cup FV(N) \cup FV(P) \cup FV(A)\subseteq dom\{\Gamma\}$.
\end{enumerate}
\selfcomment
\item (Renaming) Suppose $\gamma$ is a renaming from $\Delta$ to $\Gamma$. If $\Gamma \models J$, then $\Delta \models J[\gamma]$ (where $J[\gamma]$ is the $\gamma$-renaming from $J$, i.e.\ $A[\gamma] \rightarrow B[\gamma]$ or $M[\gamma] \rightarrow N[\gamma] \rightarrow P[\gamma] \colon B[\gamma]$ when $J$ appears as the form $A \rightarrow B$ or $M \rightarrow N \rightarrow P \colon B$ respectively).
\item (Weakening) If $\Gamma \models J$ and $\Gamma, \Delta\models ok$, then $\Gamma,\Delta \models J$.
\item (Strengthening) If $\Gamma_0, z:C, \Gamma_1 \models J$ and $z\not\in FV(\Gamma_1)\cup FV(J)$, then $\Gamma_0, \Gamma_1 \models J$.
\item (Determinacy) \label{UNF}
\begin{itemize}
\item If $\models \Gamma \rightarrow \Delta$ and $\models \Gamma \rightarrow \Phi$, then $\Delta \equiv \Phi$.
\item If $\Gamma \models A \rightarrow B$ and $\Gamma \models A \rightarrow C$, then $B \equiv C$.
\item If $\Gamma \models M \rightarrow N \rightarrow P \colon B$ and $\Gamma \models M \rightarrow Q \rightarrow R \colon C$, then $N \equiv Q$, $P \equiv R$ and $B \equiv C$.
\end{itemize}
\end{enumerate}
\end{lemma}
\ \\{\bf Remark}\ \ The above Lemma~\ref{Stru-prop}(\ref{UNF}) of `Determinacy' says that the TOS-normal forms are unique. Of course, in order to show that the normal form of a well-typed term (under the usual reduction relation) is unique, one has to prove that the TOS-reductions are adequate. This is what we do in the following subsection.
\subsection{Adequacy {w.r.t.}\ the Untyped Reduction}
\label{subsec: adeq}
We shall show in this section that the notions of computation captured in TOS are adequate {w.r.t.}\ the usual (untyped) reduction relation, which is defined in the following definition.
\begin{defn}[Untyped Reduction for IDRT] \label{untyped-red}
The (untyped) one-step reduction over terms, notation $\to$, is the compatible closure\footnote{The compatible closure of a relation $R$ over terms propagates R to all of the terms. We omit its formal definition here; see \cite{healf:thesis,Fen10} for formal details.} of the relation given by the following rules:
\begin{eqnarray*}
(\beta) \ \ \ \ \ \ \ \ \ \ ([x:A]M)N & \rightarrow & [N/x]M \\
(\eta) \ \ \ \ \ \ \ \ \ \ \ [x:A]M(x) & \rightarrow &M\ \ \ \ \ \ \ \ \ \ \ \ \ (x \notin FV(M)) \\
(\pi_1) \ \ \ \ [\langle r, l=a \colon A \rangle] & \rightarrow & r \\
(\pi_2) \ \ \ \ \langle r, l=a \colon A \rangle .l & \rightarrow & a \\
(\pi_2') \ \ \ \langle r, l=a \colon A \rangle.l' & \rightarrow & r.l' \ \ \ \ \ \ \ \ \ \ \ \ \ (l \neq l')
\end{eqnarray*}
We write $\rightarrow^+$ and $\rightarrow^*$ for the corresponding transitive closure and reflexive and transitive closure, respectively.
A term of the form on the left of an arrow is called a \emph{redex}. For example, a $\pi_2$-redex is a term of the form $\langle r, l=a \colon A \rangle .l$.
\end{defn}
\selfcomment{
\begin{definition}[Compatible Closure] \label{comp-clos}
Let $R$ be a relation on terms. The \emph{compatible closure} of $R$, notation $\triangleright_R$, is the least relation satisfying the following rules:
$$
\frac{M\ R\ N}{M\ \triangleright_R\ N} (R-Inc) \ \ \
\frac{A_1\ \triangleright_R\ B_1}{(x:A_1)A_2 \ \triangleright_R\ (x:B_1)A_2} (\Pi-L) \ \ \
\frac{A_2\ \triangleright_R\ B_2}{(x:A_1)A_2 \ \triangleright_R\ (x:A_1)B_2} (\Pi-R)
$$
$$
\frac{M\ \triangleright_R\ N}{El(M)\ \triangleright_R\ El(N)} (El) \ \ \
\frac{R_1\ \triangleright_R\ R_2}{\langle R_1, l \colon A\rangle\ \triangleright_R\ \langle R_2, l \colon A \rangle} (RCD_{fst}) \ \ \
\frac{A_1\ \triangleright_R\ B_1}{\langle R, l \colon A_1\rangle\ \triangleright_R\ \langle R, l \colon B_1 \rangle} (RCD_{snd}) \ \ \
$$
$$
\frac{A_1\ \triangleright_R\ B_1}{[x:A_1]M_0 \ \triangleright_R\ [x:B_1]M_0} (\lambda-L) \ \ \
\frac{M\ \triangleright_R\ N}{[x:A]M\ \triangleright_R\ [x:A]N} (\xi) \ \ \
\frac{A_1\ \triangleright_R\ B_1}{\langle r, l=a \colon A_1\rangle\ \triangleright_R\ \langle r, l=a \colon B_1\rangle} (rcd)
$$
$$
\frac{r_1\ \triangleright_R\ r_2}{\langle r_1, l=a \colon A\rangle\ \triangleright_R\ \langle r_2, l=a \colon A\rangle} (rcd_{fst}) \ \ \
\frac{a_1\ \triangleright_R\ a_2}{\langle r, l=a_1 \colon A\rangle\ \triangleright_R\ \langle r, l=a_2 \colon A\rangle} (rcd_{snd})
$$
$$
\frac{M\ \triangleright_R\ N}{M(P)\ \triangleright_R\ N(P)} (App-L) \ \ \
\frac{M\ \triangleright_R\ N}{P(M)\ \triangleright_R\ P(N)} (App-R) \ \ \
\frac{M\ \triangleright_R\ N}{M.l\ \triangleright_R\ N.l} (Sel) \ \ \
\frac{M\ \triangleright_R\ N}{[M]\ \triangleright_R\ [N]} (Res) \ \ \
$$
\\
\end{definition}
\begin{definition}[Untyped Reduction $\rightarrow^*$] \label{un-red}
We introduce the one-step reduction rules over untyped terms in IDRT:
the untyped reduction (or just reduction) $\rightarrow$ is the compatible closure of all the following rules in Figure \ref{Untyped-reduction}; we write $\rightarrow^+$ for the transitive closure of reduction and $\rightarrow^*$ for the reflexive, transitive closure of reduction. We also write $\rightarrow_{R}$ for the compatible closure of the last three rules that operate only on the record terms, and $\rightarrow_{R}^+$ and $\rightarrow_{R}^*$ accordingly. Similarly $\rightarrow_{\beta R}$ is the compatible closure of $\rightarrow_\beta$ and $\rightarrow_R$, and $\rightarrow_{\beta R}^+$ and $\rightarrow_{\beta R}^*$ accordingly. \\
\begin{figure}[here]
\framebox[5.8in][c]{
\begin{minipage}{\linewidth}
\begin{eqnarray*}
(\beta) \ \ \ \ & ([x:A]M)N \rightarrow_\beta \ [N/x]M &\\
(\eta) \ \ \ \ & [x:A]M(x) \rightarrow_\eta \ M & (x \notin FV(M)) \\
(restriction) \ \ \ \ & [\langle r, l=a \colon A \rangle] \rightarrow_{RESTR} \ r &\\
(field-selection) \ \ \ \ & \langle r, l=a \colon A \rangle .l \rightarrow_{FLDSEL} \ a &\\
(diff.\ field-sele.) \ \ \ \ & \langle r, l=a \colon A \rangle.l' \rightarrow_{FLDSL'} \ r.l' & (l \neq l') \\
\end{eqnarray*}
\end{minipage}
} \caption{One-step Reduction of Untyped Terms in IDRT} \label{Untyped-reduction}
\end{figure}
(\textbf{Redex}) Let $R$ be a relation, a term $M$ is an $R$-redex if there is some $N$ such that $M\ R\ N$. A term $M$ in IDRT is a redex if $M$ is a $\rightarrow$-redex. \\
\end{definition}
\begin{defn}[Weak-Head Normal Forms and Normal Forms] \label{WHN-N}\
\begin{itemize}
\item A term $M$ is \emph{weak-head normal} if
\begin{itemize}
\item $M\equiv x$ is a variable;
\item $M\equiv [x:K]k$;
\item $M\equiv f(a)$, where $f$ is weak-head normal and not an abstraction;
\item $M\equiv\record{}$;
\item $M\equiv \record{r,\ l=a\colon A}$; or
\item $M\equiv [r]$ or $M\equiv r.l$, where $r$ is weak-head normal and not a pair-record.
\end{itemize}
\item A term $M$ is \emph{normal} if
\begin{itemize}
\item $M\equiv x$ is a variable;
\item $M\equiv [x:K]k$, which is not an $\eta$-redex, and $K$ and $k$ are normal;
\item $M\equiv f(a)$ and $f$ and $a$ are normal and $f$ not an abstraction;
\item $M\equiv \record{}$;
\item $M\equiv \record{r,\ l=a\colon A}$ and $r$, $a$ and $A$ are normal.
\item $M\equiv [r]$ or $M\equiv r.l$, where $r$ is normal and not a pair-record.
\end{itemize}
\end{itemize}
The notions of weak-head normal forms and normal forms are lifted to record types, kinds and contexts in the usual way.
\end{defn}
This case was in our original proof of a DRT system with the WER rules, it was of interest because the weakly extensional rules are $\eta$-like rules that cause problems, such as strong normalization fails for untyped raw terms. For reason of discussion we keep this case still here.
The following lemmas show that the notion of computation captured in TOS is adequate {w.r.t.}\ the untyped reduction and the associated notions of normal forms.
\begin{lemma}[Adequacy of TOS {w.r.t.}\ Untyped Reduction] \label{AUR}\
\begin{itemize}
\item If $\Gamma \models A \rightarrow C$ then there exists $B$ such that $A \rightarrow_{\beta R}^* B \rightarrow_\eta^* C$.
\item If $\Gamma \models M \rightarrow N \rightarrow P \colon A$, then there exists $N'$ such that $M \rightarrow_{\beta R}^* N \rightarrow_{\beta R}^* N' \rightarrow_\eta^* P$.
\end{itemize}
\end{lemma}
\textbf{Proof.} By induction on derivations.
\begin{lemma}[Adequacy of TOS {w.r.t.}\ Normal Forms and WHNFs] \label{ANF}\
\begin{itemize}
\item If $\models \Gamma \rightarrow \Delta$, then $\Delta$ is normal.
\item If $\Gamma \models A \rightarrow B$, then $B$ is normal.
\item If $\Gamma \models M \rightarrow N \rightarrow P \colon A$, then $N$ is weak-head normal and $P$ and $A$ are normal.
\end{itemize}
\end{lemma}
\textbf{Proof.} By induction on derivations.
\selfcomment
\begin{lemma}[Determinacy (Unique Normal Form)] \label{UNF}
\ \\
(1) If $\models \Gamma \rightarrow \Delta$, $\models \Gamma \rightarrow \Phi$, then $\Delta \equiv \Phi$; \\
(2) If $\Gamma \models A \rightarrow B$, $\Gamma \models A \rightarrow C$, then $B \equiv C$; \\
(3) If $\Gamma \models M \rightarrow N \rightarrow P \colon B$, $\Gamma \models M \rightarrow Q \rightarrow R \colon C$, then $N \equiv Q$, $P \equiv R$, $B \equiv C$.
\end{lemma}
\textbf{Proof.} By simultaneous induction on derivations.
\subsubsection{Soundness and Completeness}
\label{subsec:comp-sound}
The TOS we have studied is sound and complete {w.r.t.}\ the type system IDRT of dependent record types. In the informal introduction to TOS in Section~\ref{sec:introTOS}, we have over-simplified the situation. In fact, what we shall do is to show that completeness holds for a simpler system $IDRT^-$ (with judgements of the form $\Gamma\ts^- J$), which is obtained from IDRT by removing the seven substitution rules in Appendix~\ref{app:LF-rules}. Therefore, the soundness and completeness may be pictured as follows:
\begin{picture}(100, 65)(-115, 0)
\put(50, 50){\makebox(0,0){$\vdash$}}
\put(120, 50){\makebox(0,0){$\vdash^-$}}
\put(82, 10){\makebox(0,0){$\models$}}
\put(60, 40){\oval(20, 60)[bl]}
\put(105, 40){\oval(20, 60)[br]}
\put(82, 50){\oval(48, 10)[t]}
\put(24, 30){\makebox(0,0){\emph{\begin{small}Soundness\end{small}}}}
\put(150, 30){\makebox(0,0){\emph{\begin{small}Completeness\end{small}}}}
\put(82, 48){\makebox(0,0){\emph{\begin{small}$\supset$\end{small}}}}
\put(60, 10){\vector(1, 0){2}}
\put(115, 40){\vector(0, 1){2}}
\put(58, 50){\vector(-1, -2){2}}
\end{picture}
\begin{theorem}[Completeness of TOS {w.r.t.}\ $IDRT^-$] \label{compl}\
\begin{itemize}
\item If $\Gamma \models ok$ then $\vdash^- \Gamma \ valid$.
\item If $\Gamma \models A \rightarrow B$ then $\Gamma \vdash^- A \ kind$ and $\Gamma \vdash^- A=B$.
\item If $\Gamma \models M \rightarrow N \rightarrow P \colon A$ then $\Gamma \vdash^- M \colon A$, $\Gamma \vdash^- M=N \colon A$, $\Gamma \vdash^- M=P \colon A$ and $\Gamma \vdash^- A=A$.
\end{itemize}
\end{theorem}
\textbf{Proof.} By simultaneous induction on derivations and examining each case of the $TOS$ inference rules.
\begin{cor}[Completeness of TOS {w.r.t.}\ IDRT] \label{completeness}\
\begin{itemize}
\item If $\Gamma \models ok$ then $\vdash \Gamma \ valid$.
\item If $\Gamma \models A \rightarrow B$ then $\Gamma \vdash A \ kind$ and $\Gamma \vdash A=B$.
\item If $\Gamma \models M \rightarrow N \rightarrow P \colon A$ then $\Gamma \vdash M \colon A$, $\Gamma \vdash M=N \colon A$, $\Gamma \vdash M=P \colon A$ and $\Gamma \vdash A=A$.
\end{itemize}
\end{cor}
\textbf{Proof.} By Theorem~\ref{compl} and the inclusion of $IDRT^-$ in IDRT.
\ \\
The Soundness Theorem is harder to prove. We have to consider all the inference rules of IDRT including the structural rules. In the following, we only consider some selected cases. The detailed proof can be found in \cite{Fen10}.
\begin{theorem}[Soundness of TOS {w.r.t.}\ IDRT] \label{Soundness-TOS-LF}\
\begin{itemize}
\item If $\Gamma \vdash ok$, then there exists $\Delta$ such that $\models \Gamma \rightarrow \Delta$.
\item If $\Gamma \vdash A\ kind$, then there exists $B$ such that $\Gamma \models A \rightarrow B$.
\item If $\Gamma \vdash A=B$ then there exists $C$ such that $\Gamma \models A \rightarrow C$ and $\Gamma \models B \rightarrow C$.
\item If $\Gamma \vdash M \colon A$ then there exist $P$, $B$ such that $\Gamma \models A \rightarrow B$ and $\Gamma \models M \rightarrow_n P \colon B$.
\item If $\Gamma \vdash M=N \colon A$, then there exist $P$, $B$ such that $\Gamma \models A \rightarrow B$, and $\Gamma \models M \rightarrow_n P \colon B$, $\Gamma \models N \rightarrow_n P \colon B$.
\end{itemize}
\end{theorem}
\textbf{Proof.} By induction on derivations. For the cases of LF-rules, see \cite{healf:TLCA99YY}. We consider the following two cases about record types.
\begin{itemize}
\item The second introduction rule in Figure~\ref{DRT-rules}:
\[ \frac{\Gamma \vdash \langle R, \ l\colon A \rangle \colon RType \: \: \: \Gamma \vdash r \colon R \: \: \: \Gamma \vdash a \colon A(r)}{\Gamma \vdash \langle r, \ l=a \colon A \rangle \colon \langle R, \ l \colon A \rangle}
\]
By induction hypothesis, the following hold:
\begin{enumerate}
\item \label{(*)}
$\Gamma \models \langle R, l \colon A \rangle \rightarrow_n \langle P, l \colon B \rangle \colon RType$ for some $P$ and $B$,
\item $\Gamma \models r \rightarrow_n p \colon P'$ and $\Gamma \models R \rightarrow_n P' \colon RType[L]$ for some $p$, $P'$ and $L$, and
\item \label{(**)}
$\Gamma \models a \rightarrow_n b \colon C$ and $\Gamma \models A(r) \rightarrow C$ for some $b$ and $C$.
\end{enumerate}
By Lemma~\ref{Stru-prop}(\ref{UNF}) (Determinacy) and inversion of the rule $(RCDT)$ in Figure~\ref{TOS-DRT-rules}, $P \equiv P'$. Therefore, by rule $(RCD)$ in Figure~\ref{TOS-DRT-rules}, $\Gamma \models \langle r, l=a \colon A \rangle \rightarrow_n \langle p, l=b \colon B \rangle \colon \langle P, l \colon B \rangle$.
\item The third elimination rule in Figure~\ref{DRT-rules}:
\[ \frac{\Gamma \vdash r \colon \langle R, \ l \colon A \rangle \: \: \: \Gamma \vdash [r].l' \colon B \: \: \: l \neq l'}{\Gamma \vdash r.l' \colon B}
\]
By induction hypothesis, the following hold:
\begin{enumerate}
\item $\Gamma \models r \rightarrow_n s \colon \langle P, l \colon B \rangle$ and $\Gamma \models \langle R, l \colon A \rangle \rightarrow \langle P, l \colon B \rangle$ for some $s$, $P$ and $B$, and
\item $\Gamma \models [r].l' \rightarrow_n c \colon C$ and $\Gamma \models B \rightarrow C$ for some $c$ and $C$.
\end{enumerate}
Since $l \neq l'$, by $(FLDSL')$ in Figure~\ref{TOS-DRT-rules}, we have $\Gamma \models r.l' \rightarrow_n c \colon C$.
\end{itemize}
\selfcomment{
We shall examine each case of $LF$ and $DRT$ inference rules
and consider all the structural rules (selected proof segment): \\
-- Kinding rules. We show for $RType$: By induction hypothesis $\Gamma \models ok$, using $RTYPE_{TOS}$ we have $\Gamma \models RType \rightarrow RType$. \\
-- Formation rules for $RType[\_]$. We show for the second one: By induction hypothesis there exist $P, B'$ such that $\Gamma \models RType[L] \rightarrow B'$ (so by inversion $B' \equiv RType[L]$), and $\Gamma \models R \rightarrow_n P \colon RType[L]$. Also by induction hypothesis there exist $B$ such that $\Gamma \models A \rightarrow B \colon (P)Type$. Together with the condition $l \notin L$, using $RCDT_{TOS}$ we have $\Gamma \models \langle R, l \colon A \rangle \rightarrow_n \langle P, l \colon B \rangle \colon RType[L \cup \{l\}]$. \\
-- Introduction rules. We show for the second one: By induction hypothesis there exist $P, B$ such that $\Gamma \models \langle R, l \colon A \rangle \rightarrow_n \langle P, l \colon B \rangle \colon RType$ $^{(*)}$; also that there exist $s, P'$ such that $\Gamma \models r \rightarrow_n s \colon P'$ and $\Gamma \models R \rightarrow_n P' \colon RType[L]$ for some $L$. By inversion of $RCDT_{TOS}$ and by Determinacy we know $P \equiv P'$. Also by induction hypothesis there exist $b, B'$ such that $\Gamma \models a \rightarrow_n b \colon B'$ and $\Gamma \models A(r) \rightarrow B'$ $^{(**)}$. Here, consider two subcases when we construct the new normal form for the small record $\langle r, l=a \colon A \rangle$: \\
\indent
First subcase, $\langle s, l=b \colon B \rangle$ is not an $\eta$-redex. By inversion on $^{(*)}$, we have $\Gamma \models A \rightarrow_n B \colon (P)Type$ and $\Gamma \models r \rightarrow_n s \colon P$. Use the $TOS-LF$ rule for application, we have $\Gamma \models A(r) \rightarrow_n B(s) \colon Type$. Use the $TOS-LF$ rule for $El$, we have $\Gamma \models El(A(r)) \rightarrow El(B(s))$ $^{(***)}$. Compare $^{(**)}$ and $^{(***)}$, by Determinacy we have $El(B(s)) \equiv El(B')$. By inversion on $LF$'s $El$ rule we have $\Gamma \vdash B(s) = B' \colon Type$. Apply induction hypothesis again and by the fact that $B'$ is normal (Adequacy for normal forms, Lemma \ref{ANF}), we have $\Gamma \models B(s) \rightarrow_n B' \colon Type$. Finally, with all these consitions got, we apply $RCD_{TOS}$ rule and get $\Gamma \models \langle r, l=a \colon A \rangle \rightarrow_n \langle s, l=b \colon B \rangle \colon \langle P, l \colon B \rangle$. \\
\indent
Second subcase, $\langle s, l=b \colon B \rangle$ is an $\eta$-redex. This means there exists $r'$ such that $[r'] \equiv s$ and $r'.l \equiv b$. Similarly to the above subcase, by inversion on $BASE_{RESTR}$ and $BASE_{FLDSEL}$ rules of $TOS-DRT$, we have $\Gamma \models r' \rightarrow_n p \colon \langle P, l \colon B \rangle$ for some $p$. \\
-- Elimination rules. We show for the third one (the first two could be easily deduced from inverting the $BASE_{TOS}$ rules): By induction hypothesis there exist $p, b, P, B$ such that $\Gamma \models r \rightarrow_n \langle p, l = b \colon B \rangle \colon \langle P, l \colon B \rangle$ and $\Gamma \models \langle R, l \colon A \rangle \rightarrow \langle P, l \colon B \rangle$; also that there exist $c$ such that $\Gamma \models [r].l' \rightarrow_n c \colon C$ and $\Gamma \models B \rightarrow C$. There are two subcases here, apply an inversion either on the rule $FLDSEL_{TOS}$ or on the rule $BASE_{TOS}$, we will have $\Gamma \models B(P) \rightarrow C \colon Type$. Together with the condition $l \neq l'$, using $FLDSL'_{TOS}$ we have $\Gamma \models r.l' \rightarrow_n c \colon C$. \\
-- Computation rules. We show for the third one: From the induction hypothesis and by inversion on the rule $RESTR_{TOS}$ and using the rule $FLDSL'_{TOS}$ it will be easily derived that $\Gamma \models r.l' \rightarrow_n s \colon C$ and $\Gamma \models [r].l' \rightarrow_n s \colon C$ and $\Gamma \models B \rightarrow C$ for some $s, C$.
-- Congruence rules. We show for the one for $RType[L \cup \{l\}]$: Using induction hypothesis and the rule $RCDT_{TOS}$, also by Determinacy, we will derive that $\Gamma \models \langle R, l \colon A \rangle \rightarrow \langle P, l \colon B \rangle \colon RType[L \cup \{l\}]$, $\Gamma \models \langle R', l \colon A' \rangle \rightarrow \langle P, l \colon B \rangle \colon RType[L \cup \{l\}]$ for some $P, B$ and $\Gamma \models RType[L \cup \{l\}] \rightarrow RType[L \cup \{l\}]$.
}%\quad\hfill\mbox{$\Box$} \\
\subsection{Strong Normalisation in TOS}
\label{sec:PSR-SN}
The strong normalisation property of the TOS says that, if a term $M$ is well-typed in the TOS ({i.e.}, $\Gamma\models M\colon A$), then $M$ is strongly normalisable. This result, together with the soundness theorem, will then be enough to show that the original type theory has the property of strong normalisation.
The strong normalisation property of the TOS by introducing a notion of \emph{parallel reduction} and showing that it has the so-called \emph{Parallel Subject Reduction} property.
\begin{definition}[Parallel Reduction] \label{Pal-Red}
Parallel reduction $\Rightarrow$ is defined as the least relation closed under the rules in Figure~\ref{Parallel-reduction}, and is extended in an obvious way to kinds and contexts.
\begin{figure}[here]
\framebox[5.8in][c]{
\begin{minipage}{\linewidth}
\ \\
$$
VAR\ \frac{}{x \Rightarrow x} \ \ \
\lambda\ \frac{A \Rightarrow A' \ \ M \Rightarrow M'}{[x:A]M \Rightarrow [x:A']M'} \ \ \
APP\ \frac{M \Rightarrow M' \ \ N \Rightarrow N'}{M(N) \Rightarrow M'(N')}
$$
$$
\beta\ \frac{M \Rightarrow M' \ \ N \Rightarrow N'}{([x:A]M)(N) \Rightarrow [N'/x]M'} \ \ \
\eta\ \frac{M_0 \Rightarrow M'(x) \ \ x \notin FV(M')}{[x:A_1]M_0 \Rightarrow M'}
$$
$$
RCDT_{EMP}\ \frac{}{\langle \rangle \Rightarrow \langle \rangle} \ \ \ \ \
RCDT\ \frac{R \Rightarrow R' \ \ A \Rightarrow A'}{\langle R, l \colon A \rangle \Rightarrow \langle R', l \colon A' \rangle}
$$
$$
RCD_{EMP}\ \frac{}{\langle \rangle \Rightarrow \langle \rangle} \ \ \ \
RCD\ \frac{r \Rightarrow r' \ \ a \Rightarrow a' \ \ A \Rightarrow A'}{\langle r, l = a \colon A \rangle \Rightarrow \langle r', l = a' \colon A' \rangle}
$$
$$
BASE_{RESTR}\ \frac{r \Rightarrow r'}{[r] \Rightarrow [r']}\ \ \ \
BASE_{FLDSEL}\ \frac{r \Rightarrow r'}{r.l \Rightarrow r'.l}
$$
$$
RESTR\ \frac{r \Rightarrow r'}{[\langle r, l=a \colon A \rangle] \Rightarrow r'} \ \ \ \
FLDSEL\ \frac{a \Rightarrow a'}{\langle r, l=a \colon A \rangle.l \Rightarrow a'}
$$
$$
FLDSL'\ \frac{r \Rightarrow r' \ \ l \neq l'}{\langle r, l=a \colon A \rangle.l' \Rightarrow r'.l' } \ \ \ \
$$
\\
\end{minipage}
\caption{Parallel Reduction for IDRT} \label{Parallel-reduction}
\end{figure}
\end{definition}
\ \\{\bf Remark}\ \ Parallel reduction has some simple properties. First, $M\To M$ for all M. Furthermore, if $M\to N$ then $M\To N$, and if $M\To N$ then $M\to^\ast N$. Finally, if $M\To M'$ and $N\To N'$ then $[N/x]M \To [N'/x]M'$.
\selfcomment{
\begin{lemma}[Parallel Reduction $\Rightarrow$ $\subset$ Untyped Reduction $\rightarrow^*$] \label{rel-reductions}
If $M \Rightarrow N$ then $M \rightarrow^* N$.
\end{lemma}
\textbf{Proof.} By structural induction on the terms. From Definition \ref{Parallel-reduction} we know that no $\rightarrow^*$-redex was created by parallel reduction $\Rightarrow$. \footnote{As consequence of Lemma \ref{rel-reductions}, we have $\rightarrow \ \ \subset \ \ \Rightarrow \ \ \subset \ \ \rightarrow^* \ \ \subset \ \ \Rightarrow^*$, Church-Rosser holds for both $\rightarrow^*$ and $\Rightarrow^*$. } }%\quad\hfill\mbox{$\Box$} \\
\begin{lemma}[Parallel Subject Reduction] \label{PSR}\
\begin{enumerate}
\item If $\models \Gamma \rightarrow \Delta$ and $\Gamma \Rightarrow \Gamma'$, then $\models \Gamma' \rightarrow \Delta$.
\item If $\Gamma \models A \rightarrow B$, $\Gamma \Rightarrow \Gamma'$ and $A \Rightarrow A'$, then $\Gamma' \models A' \rightarrow B$.
\item If $\Gamma \models M \rightarrow N \rightarrow P \colon A$, $\Gamma \Rightarrow \Gamma'$ and $M \Rightarrow M'$, then there exist $N'$ and $N''$ such that $N \Rightarrow N'$, $\Gamma' \models M' \rightarrow N'' \rightarrow P \colon A$ and $\Gamma' \models N' \rightarrow N'' \rightarrow P \colon A$.
\end{enumerate}
\end{lemma}
\textbf{Proof.} By simultaneous induction on derivations. The detailed proof can be found in \cite{Fen10}.
\begin{lemma}[Subject Reduction] \label{SR}
If $\Gamma \models M \rightarrow N \rightarrow P \colon A$, $\Gamma \Rightarrow \Gamma'$ and $M \Rightarrow M'$, then there exists $N'$ such that $\Gamma' \models M' \rightarrow N' \rightarrow P \colon A$ and $N \rightarrow^* N'$.
\end{lemma}
\textbf{Proof.} By simultaneous induction on derivations, using Lemma \ref{PSR}.
\ \\
The proof of strong normalisation of the TOS (Theorem~\ref{SN-TOS}) uses the following lemma.
\begin{lemma}\label{NOT-ABS-PAIR}\
\begin{itemize}
\item If $M$ is weak-head normal and not an abstraction and $M \rightarrow^* N$, then $N$ is weak-head normal and not an abstraction.
\item If $M$ is weak-head normal and not a pair record and $M \rightarrow^* N$, then $N$ is weak-head normal and not a pair record.
\end{itemize}
\end{lemma}
\textbf{Proof.} By induction on length of reduction for untyped terms.
\begin{theorem}[Strong Normalisation of TOS] \label{SN-TOS}\
\begin{enumerate}
\item If $\Gamma \models A \rightarrow B$ then $A$ is strongly normalisable.
\item If $\Gamma \models M \rightarrow N \rightarrow P \colon A$ then $M$ is strongly normalisable.
\end{enumerate}
\end{theorem}
\textbf{Proof.} By simultaneous induction on derivations. The full proof can be found in \cite{Fen10}. We shall consider one of the most difficult cases -- when the last rule used is $(BASE_{RESTR})$ in Figure~\ref{TOS-DRT-rules}:
$$
\frac{\Gamma \models r \rightarrow q \rightarrow p \colon \langle P,\ l \colon B \rangle \: \: \: \: p,\ q\ not\ pair\texttt{-}records}{\Gamma \models [r] \rightarrow [q] \rightarrow [p] \colon P}
$$
By \IH, $r$ is \SN. It suffices for us to show that $[r]$ is \SN\ if $\Gamma\models r \rightarrow_w q \colon \langle P, l \colon B \rangle$ such that $q$ is not a pair-record. We do this by induction on the maximal length of reductions starting from $r$.
Assume that $r\to r_1$. \selfcomment{As $q$ is not a pair-record and $r_1$ is arbitrary, it is enough to show that $[r_1]$ is \SN. Since $r\to r_1$,} We then have $r\To r_1$, which implies by Lemma~\ref{PSR} (Parallel Subject Reduction) that there exist $r'$ and $r''$ such that
\[
\Gamma \models r_1 \rightarrow_w r'' \colon \langle P, l \colon B\rangle, \ \ \
\Gamma \models r' \rightarrow_w r'' \colon \langle P, l \colon B\rangle, \ \ \ \textrm{and}\ \ \
q \Rightarrow r'.
\]
We therefore have that $q$ is weak-head normal (by Lemma~\ref{ANF}) and that $q \rightarrow^* r'$ (see the remark above). From these and Lemma~\ref{NOT-ABS-PAIR}, we have
\begin{itemize}
\item[(*)] $r'$ is weak-head normal and not a pair-record.
\end{itemize}
Furthermore, by Lemma~\ref{AUR}, we have
\begin{itemize}
\item[(**)] $r' \rightarrow^* r''$.
\end{itemize}
From $(*)$ and $(**)$, $r''$ is not a pair-record by Lemma~\ref{NOT-ABS-PAIR}. Therefore, since $\Gamma \models r_1 \rightarrow_w r'' \colon \langle P, l \colon B\rangle$, we conclude by \IH\ that $[r_1]$ is \SN.
\selfcomment{
\par{And eventually, we arrive at the following main Corollary, which is strong normalisation of our original system of dependent record types IDRT: }
\begin{corollary}[Strong Normalisation of IDRT] \label{SN-LFDRT}
\ \\
(1) If $\Gamma\ valid$ then $\Gamma$ is strongly normalisable; \\
(2) If $\Gamma \vdash A \ kind$ then $A$ is strongly normalisable; \\
(3) If $\Gamma \vdash A = B$ then both $A$ and $B$ are strongly normalisable to some $C$; \\
(4) If $\Gamma \vdash M \colon A$, then $M$ is strongly normalisable to some $P$, $A$ is strongly normalisable to some $B$ such that $P \colon B$ in $\Gamma$; \\
(5) If $\Gamma \vdash M = N \colon A$, then both $M$ and $N$ are strongly normalisable to some $P$, $A$ is strongly normalisable to some $B$ such that $P \colon B$ in $\Gamma$.
\end{corollary}
\textbf{Proof.} Derived from Theorem \ref{Soundness-TOS-LF}
and Theorem \ref{SN-TOS}. \\
\par{\noindent Other metatheory such as the Church-Rosser property, uniqueness of normal form, etc. could be derived in a similar way as strong normalisation and subject reduction are done. }
\begin{corollary}[Church-Rosser Property of IDRT] \label{CR-LFDRT}
\ \\
If $\Gamma \vdash M = N \colon A$, then $\exists P$ such that $M \rightarrow^* P$ and $N \rightarrow^* P$.
\end{corollary}
\textbf{Proof.} By Soundness (Theorem \ref{Soundness-TOS-LF}), there exist $P, B$ such that $\Gamma \models M \rightarrow_n P \colon B$ and $\Gamma \models N \rightarrow_n P \colon B$. By Adequacy (Lemma \ref{AUR}) we prove that $M \rightarrow^* P$ and $N \rightarrow^* P$. \\
\section{Meta-theoretic Properties of IDRT}
\label{sec:metaDRT}
From the properties of the TOS that have been proved in the last section, the meta-theoretic properties of IDRT, the type theory for dependent record types, can be proved. Here, we give the theorems for Subject Reduction, Church-Rosser and Strong Normalisation. (For further details and other properties, see \cite{Fen10}.)
\begin{thm}[Subject Reduction for IDRT] \label{SR-LFDRT}
If $\Gamma \vdash M \colon A$ and $M \rightarrow N$, then $\Gamma \vdash N \colon A$.
\end{thm}
\textbf{Proof.} First of all, we have $\Gamma \models M \colon A$ (by the Soundness Theorem \ref{Soundness-TOS-LF}) and $M \Rightarrow N$ (since $M \rightarrow N$). Therefore, by Lemma~\ref{SR}, we have $\Gamma \models N \colon A$. Now, by Completeness (Theorem~\ref{completeness}), $\Gamma \vdash N \colon A$.
\begin{thm}[Strong Normalisation for IDRT] \label{SN-LFDRT}\
\begin{enumerate}
\item If $\Gamma\ valid$, then $\Gamma$ is strongly normalisable.
\item If $\Gamma \vdash A \ kind$, then $A$ is strongly normalisable.
\item If $\Gamma \vdash A = B$, then both $A$ and $B$ are strongly normalisable to some $C$.
\item If $\Gamma \vdash M \colon A$, then $M$ and $A$ are strongly normalisable and $\Gamma\ts P \colon B$, where $P$ and $B$ are the normal forms of $M$ and $A$, respectively.
\item If $\Gamma \vdash M = N \colon A$, then both $M$ and $N$ are strongly normalisable to some $P$, $A$ is strongly normalisable to some $B$ such that $\Gamma\ts P \colon B$.
\end{enumerate}
\end{thm}
\textbf{Proof.} By the Soundness Theorem~\ref{Soundness-TOS-LF}
and Theorem \ref{SN-TOS}.
\begin{thm}[Church-Rosser for IDRT] \label{CR-LFDRT}
If $\Gamma \vdash M = N \colon A$, then $M \rightarrow^* P$ and $N \rightarrow^* P$ for some $P$.
\end{thm}
\textbf{Proof.} By Soundness (Theorem \ref{Soundness-TOS-LF}), there exist $P$ and $B$ such that $\Gamma \models M \rightarrow_n P \colon B$ and $\Gamma \models N \rightarrow_n P \colon B$. Then, by Adequacy (Lemma \ref{AUR}), we have $M \rightarrow^* P$ and $N \rightarrow^* P$.
\section{Conclusions}
\label{sec:conclusion}
We have studied the meta-theory of a type theory with dependent record types, by studying its typed operational semantics. As we have mentioned in Section~\ref{sec:LFwithDRT}, dependent record \emph{types} are rather different from dependent record \emph{kinds}, with the former having a much richer structure and being more difficult to study. The meta-theory of dependent record kinds has been studied by Coquand \emph{et. al} \cite{ctp:semantic-records05}, where they have given a proof of termination of type-checking. As far as we know, ours is the first attempt to study the meta-theory of dependent record types formulated in a logical framework \cite{Pollack:records02,luo:TYPES08,luo:MLPA09}. In the light that the TOS-approach has been successfully applied to the meta-theoretic study of type theories with $\eta$-equality \cite{healf:TLCA99YY} and with inductive types \cite{healf:thesis}, the current work can be used to justify the incorporation of dependent record types in a full-scale type theory as implemented in the proof assistants such as Agda and Coq.
The dependent record types studied in this paper are \emph{intensional} in the sense that we do not have the following extensional equality rules \cite{bet-tar:subtyping98,luo:TYPES08}:
$$
\frac{\Gamma \vdash r\colon \langle \rangle}{\Gamma \vdash r = \langle \rangle \colon \langle \rangle}
\: \: \: \: \: \:
\frac{\begin{array}{c} \Gamma \vdash r \colon \langle R, \ l\colon A \rangle \: \: \: \Gamma \vdash r' \colon \langle R, \ l\colon A \rangle \\ \Gamma \vdash [r] = [r']\colon R \: \: \: \Gamma \vdash r.l = r'.l \colon A([r]) \end{array}}{\Gamma \vdash r = r' \colon \langle R, \ l\colon A \rangle}
$$
They basically say that two records are computationally equal if their components are. For instance, from the second rule above, we would have $\record{r,\ l=r.l} = r$ for any $r$ of type $\record{R,\ l\colon A}$. It is unclear whether the TOS-approach as adopted in this paper can be applied to such (weakly) extensional record types. It would be obviously problematic if one considered the reduction relation for the records as follows:
\[ \record{r,\ l=r.l} \to r \]
for, together with the $\eta$-reduction for $\lambda$-terms, the Church-Rosser property would fail to hold. A natural question arises here: would it possible if one takes the TOS-approach by considering a reduction relation that treats $\eta$-long normal forms ({e.g.}, by taking the above reduction in the other direction)? This involves the development of the TOS-approach to incorporate $\eta$-long normal forms and research is needed to see whether it is possible.
\selfcomment{
\par{
This is generally problematic with systems with $\eta$-equality, an example for why Church-Rosser fails could be as: In a system with both $\Pi$-type, $\Sigma$-type and unit type $\textbf{1}$, suppose we allow $\eta$-rules for $\Pi$, $\Sigma$ and for unit types, }
\begin{eqnarray*}
(\eta_\Pi) \ \ \ & [x:A]f(x) \ \rightarrow \ f \colon B \ (if\ x\ \notin\ FV(f)) \\
(\eta_\Sigma) \ \ \ & \langle \pi_1(x), \pi_2(x) \rangle \ \rightarrow \ x \colon A \\
(\eta_\textbf{1}) \ \ \ & x \ \rightarrow \ * \colon \textbf{1}
\end{eqnarray*}
then, a term $[x:A]f(x)$ with $f \colon A \rightarrow \textbf{1}$ could have two reductions by $(\eta_\Pi)$ and by $(\eta_\textbf{1})$ that do not commute: $[x:A]f(x)\ \rightarrow_{\eta_\Pi} \ f$ and $[x:A]f(x)\ \rightarrow_{\eta_\textbf{1}} \ [x:A].*$. Note that $[x:A].*$ is already a canonical form.
\par{Similarly for a term $\langle \pi_1(x), \pi_2(x) \rangle \colon \textbf{1} \times A$, it reduces to both $x$ and to $\langle *, \pi_2(x) \rangle$} under $(\eta_\Sigma)$ and $(\eta_\textbf{1})$ respectively, which do not commute.
This situation happens because there are two types of $\eta$-reductions: one is the structural $\eta$s such as $(\eta_\Pi)$ or $(\eta_\Sigma)$, the other one is the non-structural or ``terminating'' $\eta$s such as $(\eta_\textbf{1})$ (they are called ``terminating'' because the calculations are towards a canonical form). The mixture of these two different types of $\eta$-rules has caused the failure of the Church-Rosser property. But if a system simply has only the structural $\eta$s, Church-Rosser still holds, for example,
\begin{eqnarray*}
\ [x:A]\langle \pi_1(x), \pi_2(x) \rangle & \rightarrow_{\eta_\Pi} \ & \langle \pi_1(x), \pi_2(x) \rangle \ \ \rightarrow_{\eta_\Sigma} \ \ x, \ and \\
\ [x:A]\langle \pi_1(x), \pi_2(x) \rangle & \rightarrow_{\eta_\Sigma} \ & [x:A]x \ \ \rightarrow_{\eta_\Pi} \ \ x \\
\end{eqnarray*}
\par{In our system, the weakly extensional equality rules are similar to the $\eta$-rules, and in particular, if one allows $\eta$-reduction for the DRTs in the definition of untyped reduction (Definition \ref{Untyped-reduction}), our proofs will not work, because the first weakly extensional equality rule is one in the category of non-structural $\eta$s, while the second still belongs to the category of structural $\eta$s. So if this system is extended with non-structural $\eta$s, i.e., the first WER, Church-Rosser fails. However, in our proof, the part related to the weakly extensionality rules in TOS and in other related definitions are supposed to be \emph{disconnected} with other part, which means, if we add a case in the proof for Strong Normalisation that deals with the weakly extensional equality rules given above,
it will only use rules provided by the TOS rule $ETA_{RCD}$ in Figure \ref{TOS-DRT-rules} and the $\eta$-parallel reduction definition in Figure \ref{Parallel-reduction}. }
\par{In another word, if one takes the weakly extensional equality rules out of the system, and remove also the $ETA_{RCD}$ rule in TOS and the other related definitions, such as the ETA for parallel reduction, the proof still go through. If one adds back these rules concerning structural $\eta$-calculation, as a corollary, Church-Rosser works for the typeable terms in TOS. This typeable semantic control ensures that $\eta$s do not cause problem, and it is one of the main features of applying the TOS approach. This approach could be useful for other systems without $\eta$-definitions or non-structural $\eta$-definitions for similar metatheoretical proof. }
\par{In this paper we have introduced a new formulation of dependent record types as an extension of the logical framework LF, and we give a typed operational semantics for this formulation. We have proved some metatheoretical results on this typed semantics, and by its soundness theorem, we have derived the strong normalisation of the dependent record types. Other metatheoretical results such as the Church-Rosser property, subject reduction, etc., could be similarly derived. }
\par{Coquand \emph{et. al} have presented also a metatheoretical study in \cite{ctp:semantic-records05}, on a system of record kinds. They have given a proof of termination of type-checking for the record types; different with their method, we have shown that the typed semantical method to prove strong normalisation that Goguen suggested could also apply on our dependent record types. }
\par{As Goguen's work has established a full metatheory for the UTT and the LF, we here also similarly set up a full metatheory for the DRTs with a same way. Our work has led to an understanding on the following two points: First, the DRTs with intensional manifest fields developed in previous work \cite{luo:TYPES08,luo:MLPA09} own good metatheoretical properties, which would support themselves further as a modelling foundation to module mechanism; Second, this technique used here of the typed operational semantics could be applied to other type systems with structural $\eta$-rules. In conclusion, the methodology of applying TOS has set up a relation between well-formedness in the TOS and well-behaveness in original systems, which is ensured by the typing control of the TOS. }
\par{In the future, we have a few attempts concerning the approach discussed in this paper. Firstly, we would like to extend this technique to wider range of type systems, such as to the Logical Framework with inductive schemata and the dependent record types together with other types such as the unit types.
Also, we are interested to find out if this technique could be applied
to ill-behaving (e.g. non-SN) calculi, or, for not only proving strong normalisation, we would like to also prove weak-head normalisation in related type systems. }
\paragraph{Comparason with other Systems of Dependent Records}
\par{Dependent record types have been previously studied in \cite{MacQ:module86, Harper-Lillibridge93, bet-tar:subtyping98, Pollack:records02, ctp:semantic-records05}, with applications to the study of module mechanisms for both programming and proof languages.
MacQueen has firstly suggested in \cite{MacQ:module86} a module system using dependent $\Sigma$-types as module signatures, which combines the first ideas coming from ML on data abstraction, and from Pebble on dependent Sigma types.
Later, Harper and Lillibridge have put more emphasis on the \emph{control of information flow} between program units, i.e. the idea of sharing, and have proposed in \cite{Harper-Lillibridge93} the weak sum type to implement the ML-style ``sharing by equation'', or called ``sharing by coherence conditions''; but in their formulation, the bounded quantification used is proven to be undecidable, the reason is by the FORGET-rule which loses type information. This has been proved similarly as in Pierce's proof for undecidability of $F_{\leq}$ \cite{Pie94}. }
\par{Betarte and Tasistro have suggested in \cite{bet-tar:subtyping98} a different record system, i.e. of \emph{record kinds},
with type inclusion to represent subtyping, this system is proved to be decidable (the proof is called unicity of right identity); however the record kinds are simpler than record types as a structure.
Pollack presented in \cite{Pollack:records02} a system with dependent record types which is similar to ours however allows repetition of field labels. }
\vspace{0.6cm}
\noindent\textbf{Acknowledgement} The authors would like to thank Robin Adams for discussions on dependent record types and the first author thanks Cody Roux for discussions.
\bibliographystyle{alpha}
|
\section*{Introduction}
What is the lowest geometric genus $\eta(n,d)$ of a reduced,
irreducible curve on a very general hypersurface of degree $d$ in
${\mathbb P}^n$? The case $n=2$ is trivial. For $n=3$ one has
\begin{eqnarray}\label{in0} \qquad \eta(3,d)=0\quad\mbox{if}\quad d\le
4\quad\mbox{while}\quad \eta(3,d)={d-1 \choose 2}-3
\quad\mbox{if}\quad d\ge 5 \,\end{eqnarray} and for any $d\ge 6$ this bound
is achieved by tritangent plane sections, and only by these
\cite{Xu1}. Similarly, $\eta(4,d)=0$ if $d\le 5$, while
$\eta(4,6)\ge 2\,$ \cite{ClR}. More generally, for $n \geq 4$ one
has
\[\,\quad\quad \eta(n,d)=0\quad \forall d\le
2n-3\qquad\mbox{and}\quad \eta(n,d)\ge 1\quad \forall d \geq
2n-2\,
\]
see \cite{Vo1} (in the case $d=2n-3$ see also \cite{Ha, Pac1,
Sh}). Presumably, $\eta(n,d)\to\infty$ as $d\to\infty$, however,
the asymptotic of $\eta(n,d)$ is unknown. One is equally
interested in bounds for the geometric genus or other numerical
invariants of higher dimensional subvarieties in general
hypersurfaces, see e.g.\ \cite{CLR, Ei1, Ei2, Pac2, Vo2, Wa2,
Xu3}.
A projective variety $X$ is {\em algebraically hyperbolic} if it
does not admit a non--constant morphism from an abelian variety.
If there is an algebraically hyperbolic hypersurface of degree $d$
in ${\mathbb P}^n$, then a very general hypersurface of degree $d$ is
algebraically hyperbolic as well. For instance, a very general
surface $X$ of degree $d\ge 5$ in ${\mathbb P}^3$ is algebraically
hyperbolic. Indeed, $X$ does not contain rational or elliptic
curves since $\eta(3,d)\ge 3$ for $d\ge
5$ by \eqref{in0}. This also follows from Proposition \ref {201}
below if $d\ge 6$, while Corollary \ref {agthm} offers a short
proof of Xu's and Voisin's result about non-existence of rational
curves on a very general quintic in ${\mathbb P}^ 3$. Since $X$ is of
general type it cannot be dominated by an abelian variety.
Similarly, a general sextic threefold $X$ in ${\mathbb P}^4$ is
algebraically hyperbolic. Indeed, $X$ does not contain rational or
elliptic curves since $\eta(4,6)\ge 2$. By
\cite[Theorem 1]{Xu3} it also does not contain surfaces with
desingularization of geometric genus at most $2$. Therefore, every map
from an abelian variety to $X$ is constant.
A variety $X$ is \emph{Kobayashi hyperbolic}, or simply
\emph{hyperbolic}, if it does not admit any non--constant entire
curve $\ensuremath{\mathbb{C}}\to X$. Hyperbolicity implies algebraic hyperbolicity,
and it is stable under small deformations.
Given one of the two above hyperbolicity notions,
one can ask what is the lowest degree $d=d(n)$ such that
a very general projective hypersurface in ${\mathbb P}^n$ of degree $d$
possesses this property. For instance, the classical \emph{Kobayashi
problem} suggests that a very general hypersurface of degree $d\ge
2n-1$ in ${\mathbb P}^n$ is hyperbolic.
It is known that, indeed, a very general surface of degree $d\ge
18$ in ${\mathbb P}^3$ is hyperbolic \cite{Pau} (see also \cite{DEG, MQ}).
The existence of hyperbolic surfaces in ${\mathbb P}^3$ of degree $d$ for
all $d\ge 8$ was established with a degeneration argument in
\cite{SZ2} (see the references in \cite{SZ2} for other
constructions), and for $d=6$ in \cite{Du}. In \S \ref
{ssec:d=67} below (see, in particular, Theorem \ref {mthm}) we
give an alternative proof for the case $d=6$, which works also in
the (so far unknown) case $d=7$. The case $d=5$ in the Kobayashi
problem for ${\mathbb P}^3$ remains open.
Our method consist in degenerating a general hypersurface to a
certain special one, following the limits in the degeneration of
entire curves or of algebraic curves or surfaces, according to the
hyperbolicity notion we are dealing with. In this framework the
concept of Brody curves and their limits is very useful, cf.\ e.g.
\cite{SZ1, SZ2, Za1, Za2}. We recall a minimum of basics on this
subject in \S \ref {sec:brody}. Our preferable degenerations here
are to \emph{scrolls}, and we recall their main properties in \S
\ref {sec:scrolls}. In subsection \ref {ssec:ah} we give an easy
proof of non--existence of curves of low genera on very general
surfaces in ${\mathbb P}^ 3$ of a given degree. In \S \ref {sec:higher} we
treat the higher dimensional case. In particular, in Theorem \ref
{thm:pg} we provide a lower bound for the geometric genus of
surfaces contained in very general hypersurfaces of degree $3d\ge
15$ in ${\mathbb P}^ 4$.
By a well-known theorem of Bogomolov \cite{Bo}, on a smooth
surface $S$ of general type with $c^2_1 (S) > c_2(S)$, the curves
of a fixed geometric genus vary in a bounded family. This was
partially extended in \cite{LM} to any smooth surface $S$ of
general type by showing that there are only a finite number of
rational and elliptic curves on $S$ with a fixed number of nodes
and ordinary triple points and no other singularities. In
subsection \ref {ssec:bound} we address the question whether
curves of a given geometric genus have bounded degree on a general
surface of degree $d\ge 5$ in ${\mathbb P}^ 3$. We give an affirmative
answer for all genera $g\le d^ 2+O(d)$.
Finally in subsection \ref {ssec:d=67} we prove the aforementioned
Theorem \ref {mthm}.
\section{Scrolls}\label{sec:scrolls}
\subsection{Generalities on scrolls}\label{ssec:general}
By a \emph{scroll} in ${\mathbb P}^n$ we mean the image $\Sigma=\phi(S)$
of a smooth, proper ${\mathbb P}^1$-bundle $\pi:S\to E$ under a birational
morphism $\phi:S\to \Sigma\hookrightarrow{\mathbb P}^n$ which sends the
\emph{rulings} of $S$ (i.e. the fibres of $\pi$) to projective
lines, called \emph{rulings} of $\Sigma$. The variety $E$ is
called the \emph{base} of the scroll. We will denote by $H$ and
$F$ a hyperplane section and a ruling of $\Sigma$, respectively.
We may abuse notation denoting by $H$ and $F$ also their proper
transforms on $S$.
The induced morphism $\mu: E \to \rm Gr(1,n)$ to the Grassmanian of
lines in ${\mathbb P}^n$ is birational onto its image. Any such morphism
$\mu$ appears in this way, where $\pi:S\to E$ is induced via $\mu$
by the tautological ${\mathbb P}^1$--bundle over the Grassmanian.
Furthermore, $d=\mathop{\rm deg}(\Sigma)$ is equal to the degree of the
subvariety $\mu(E)$ under the Pl\"ucker embedding of the
Grassmanian \cite[12.4]{BCGMB}, \cite[11.4.1]{Do}.
We will suppose form now on that $\phi: S\to \Sigma$ coincides
with the normalization morphism. We denote by $ \operatorname{{\rm br}}(\Sigma)$ the
set of \emph{multibranch points} of $\Sigma$, i.e. the set of
points $x\in \Sigma$ such that $\phi^ {-1}(x)$ consists of more
than one point. If $x\not\in \operatorname{{\rm br}}(\Sigma)$, e.g. $x$ is a smooth
point of $\Sigma$, then there is just one ruling passing through
$x$. Since $\phi$ is finite, there is no point on $\Sigma$ which
belongs to infinitely many rulings.
We let $\Delta_\Sigma=\overline{ \operatorname{{\rm br}}
(\Sigma)}\subseteq\Sigma$ and
$\Delta_S=\phi^{-1}(\Delta_\Sigma)\subseteq S$.
We will assume that the following conditions hold:
\begin{itemize}
\item [(C1)] ~ $\dim(\Sigma)=n-1$;
\item [(C2)] ~ $\Delta_\Sigma$ coincides with $ \operatorname{{\rm Sing}} (\Sigma)$;
\item [(C3)] ~ $\Delta_\Sigma$ and $\Delta_S$
are both irreducible of dimension
$n-2$;
\item [(C4)] ~ a general point $x\in \Delta_\Sigma$
is a normal crossing double point of $\Sigma$. In particular,
$\phi^ {-1}(x)$ has cardinality 2, and $x$ sits on two different
rulings;
\item [(C5)] ~ $\Delta_\Sigma$ contains no ruling, i.e.
$\mu: E \to \rm Gr(1,n)$
is injective.
\end{itemize}
In this situation $\Delta_\Sigma$ and $\Delta_S$ both have natural
scheme structures, and $\Delta_S$ is a reduced divisor on $S$.
Conditions (C1)--(C4) are verified if $S\subseteq{\mathbb P}^{n+k}$ is a
smooth scroll of dimension $n-1$ and $\phi:S\to\Sigma$ is induced
by a general linear projection ${\mathbb P}^{n+k}\dashrightarrow{\mathbb P}^{n}$;
see \cite {Fr}. The last condition (C5) can be easily checked by
induction; we leave the details to the reader.
\begin{lem}\label{199} In the above setting, a general ruling of
$\Sigma$ meets the double locus $\Delta_\Sigma$ in $d-n+1$ points.
In particular, $\Sigma$ is swept out by an $(n-2)$--dimensional
family of $(d-n+1)$--secant lines of $\Delta_\Sigma$.\end{lem}
\begin{proof} By the Ramification Formula \cite[9.3.7(b)]{Fu}
there is a linear equivalence relation on $S$
\begin{eqnarray}\label{f201} \Delta_S \sim
(d-n-1)H-K_S\,.\end{eqnarray} Since $F\cdot H=1$ and $F\cdot K_S=-2$, we have
$F\cdot\Delta_S=d-n+1$. Since $\Delta_S$ is reduced, the general
ruling of $S$ meets $\Delta_S$ in $d-n+1$ distinct points.
The assertions follow because $\phi$ induces an isomorphism of
each ruling of $S$ to its image.
\end{proof}
\subsection{Surface scrolls with ordinary singularities}
We restrict here to the case $n=3$. So $E$ is a smooth curve
of genus $g$ and $S\subseteq{\mathbb P}^{3+k}$ and
$\Sigma\subseteq{\mathbb P}^3$ are surfaces, called \emph{scrolls of
genus} $g$: here $g$ is the \emph{sectional genus}
of the scroll.
\begin{rem}\label{rem:genus}
For an irreducible curve $C$ on $S$ of genus $g'$ such that
$C\cdot F=\nu$, the Riemann--Hurwitz Formula implies the
inequalities $g'\ge \nu(g-1)+1\ge g$. In particular, for $g\ge 1$
the only irreducible curves on $S$ of geometric genus $g'<g$ are
the rulings, and for $g'=g\ge 2$ the curve $C$ is a
\emph{unisecant} i.e., the intersection number $\nu$ of $C$ with
rulings is 1. The same holds on $\Sigma$.
\end{rem}
We say that $\Sigma$ has \emph{ordinary singularities} if, in
addition to conditions (C1)--(C5), the following hold:
\begin{itemize}
\item [(C6)] the singularities of the \emph{double curve} $\Delta_\Sigma$
consist of finitely many triple points, which are also ordinary
triple points of the surface $\Sigma$ (these are locally
analytically isomorphic to the surface singularity $xyz=0$ in
$\mathbb C^ 3$ at the origin);
\item [(C7)] the non--normal crossings singularities of $\Sigma$
are finitely many \emph{pinch points}. These are the points in
$\Delta_\Sigma\setminus \operatorname{{\rm br}}(\Sigma)$, and there is just one ruling
through each of them. A pinch point has just one preimage on $S$,
which, abusing terminology, we will also call a pinch point;
\item [(C8)] the only
singularities of $\Delta_S$ are ordinary double points, three of
them over each triple point of $\Delta_\Sigma$. Furthermore, the
degree two map $\phi: \Delta_S\to \Delta_\Sigma$ is ramified
exactly over the pinch points of $\Sigma$.
\end{itemize}
These are the singularities of a general projection to ${\mathbb P}^3$ of
a smooth surface in ${\mathbb P}^ 4$, or even of a surface in ${\mathbb P}^ 4$
with finitely many \emph{nodes}, i.e. double points with tangent
cone formed by two planes spanning ${\mathbb P}^ 4$.
In this case the curves $\Delta_\Sigma$ and
$\Delta_S$ are irreducible, except for the projection in ${\mathbb P}^ 3$
of the Veronese surface of degree $4$ in ${\mathbb P}^ 5$ (cf.\ \cite
{Fr1, Fr2, MP, Mo}). Note that a
general projection to ${\mathbb P}^4$ of any smooth surface in ${\mathbb P}^{r}$ (with $r>4$)
has only nodes as singularities
and the Veronese surface of degree $4$ in ${\mathbb P}^ 5$ is the only one
whose general projection to ${\mathbb P}^ 4$ is smooth (see \cite {Sev, Zak1}).
The basic invariants of $S$ are
\[c_1^2=K_S^2=8(1-g),\quad c_2=e(S)=4(1-g),\quad\mbox{and}\quad
\chi({\ensuremath{\mathcal{O}}}_S)=\frac{c_1^2+c_2}{12}=1-g\] (see \cite{En}, \cite [Ch.
5, \S 2]{Ha}). The following \emph{projective invariants} are
also important \begin{eqnarray}\label{401}
\begin{aligned}
\delta_\Sigma\quad=&\quad \mathop{\rm deg}(\Delta_\Sigma)\cr
\gamma_\Sigma\quad=&\quad \text{the geometric genus of}\quad
\Delta_\Sigma\cr t_\Sigma\quad=&\quad\text{the number of triple
points of} \quad\Delta_\Sigma\cr p_\Sigma\quad=&\quad\text {the
number of pinch points of} \quad\Sigma\cr \tilde
\gamma_\Sigma\quad=&\quad \text{the geometric genus of}\quad
\Delta_S\cr
\end{aligned}
\end{eqnarray} (in the sequel we suppress the index $\Sigma$ when unnecessary).
For the proof of the following formulas see e.g.\ \cite{Bo},
\cite[\S 11.5]{Do}, \cite[p.\ 176]{En}, \cite{Pi},
\cite[(1)-(10)]{SZ0}, and references therein.
\begin{prop}\label{300} Let $\Sigma$ stands as before for a scroll in
${\mathbb P}^3$ of degree $d$ and genus $g$ with ordinary singularities.
Then the projective invariants of $\Sigma$ are given by the {\rm
Bonnesen's formulas}
\begin{eqnarray}\label{f100}
\delta={d-1\choose 2}-g\,,\end{eqnarray}
\begin{eqnarray}\label{f200}\gamma={d-3\choose 2}+(d-5)g\, ,
\end{eqnarray}
\begin{eqnarray}\label{f300} t={d-2\choose 3}-(d-4)g\,,\end{eqnarray}
\begin{eqnarray}\label{f400} p=2d+4(g-1)\,,\end{eqnarray}
\begin{eqnarray}\label{f500}
\tilde \gamma=2(\gamma+g)+d-3\,.\end{eqnarray}
\end{prop}
\begin{rem}\label{001} Due to (\ref{f300}), for $d\ge 5$ the inequality
$t\ge 0$ reads $g\le\frac{1}{6}(d-2)(d-3)\,$. This implies
\begin{eqnarray}\label{f600} g\le d-4, \quad\text{ if}\quad d=5,6,7\,.\end{eqnarray} In
the sequel we also need the inequality \begin{eqnarray}\label{002} \gamma >
3(g-1)\qquad\text {for all}\quad g\ge 1\quad\text{and}\quad d\ge
5\,.\end{eqnarray} This follows from (\ref{f200}) for $d\ge 8$ and from
(\ref{f200}) and (\ref{f600}) for $d=5,6,7$ (actually, $\gamma
>3g$ for all $g\ge 1$ and $d\ge 5$ except for $g=2$, $d=6$). \end{rem}
\subsection{Surface scrolls with general moduli}
We recall a result from \cite{APS} (cf. also
\cite[Theorem 1.2]{CCFMLincei}).
\begin{thm}\label{thm:scrolls}
Let $g \geq 0$ be an integer and let $k = {\rm min}\{1, g-1\}$. If
$ d \geq 2 g + 3 + k$, then there exists a unique irreducible
component $\mathcal{H}_{d,g}$ of the Hilbert scheme of scrolls of degree
$d$ and sectional genus $g$ in ${\mathbb P}^{r}$, where
$r=d-2g+1$,
such that the general point $[S] \in
\mathcal{H}_{d,g}$ represents a smooth scroll $S$ with $h^ 1(S, \mathcal
O_S(1))=0$, i.e. $S$ is \emph{non--special}. Furthermore
$\mathcal{H}_{d,g}$ dominates the moduli space ${\mathcal M}_g$ of smooth
curves of genus $g$ via the map sending a scroll to its base.
\end{thm}
\begin{rems}\label{rem:3space} (i) Assuming that $ d \geq 2 g + 3 + k$
(as in the above theorem), we have
$r\ge 3$ if $g=0$, $r\ge 4$ if $g=1$, and $r \ge 5$ if $g\ge 2$,
and we can project smooth scrolls $S$ with
$[S] \in \mathcal{H}_{d,g}$ thus obtaining scrolls $\Sigma$ in ${\mathbb P}^ 3$
with ordinary singularities and irreducible double curve.
(ii) The assumption of Theorem \ref{thm:scrolls} gives $d\ge 2g+4$
for $g\ge 2$ and $d\ge 2g+3=5$ for $g=1$. In fact, similar results
hold also for $g\ge 2$ and $d=2g+3$ or $d=2g+2$, while the
corresponding scrolls are no longer smooth.
More precisely, let $g\ge 2$ and $d=2g+3$ (i.e., $r=4$). Then
$\mathcal{H}_{d,g}$ is a component of the Hilbert scheme, whose general
point $[S'] \in \mathcal{H}_{d,g}$ represents a scroll $S'\subseteq {\mathbb P}^
4$ with only nodes as singularities and
with a smooth normalization $S$ such that $h^ 0(S, \mathcal
O_S(1))=5$ and $h^ 1(S, \mathcal O_S(1))=0$
(this can be shown with
the same analysis as in \cite{CCFMLincei}). Once again,
$\mathcal{H}_{d,g}$ dominates the moduli space ${\mathcal M}_g$.
If $g\ge 2$ and $d=2g+2$ (i.e., $r=3$), a similar assertion
holds. However, now $\mathcal{H}_{d,g}$ is no longer a component of the
Hilbert scheme, but a locally closed subset of the projective
space $\mathcal L_d=\vert \mathcal O_{{\mathbb P}^ 3}(d)\vert$ of all
surfaces of degree $d$ in ${\mathbb P}^3$. It is reasonable to expect that
a general point $[\Sigma] \in \mathcal{H}_{d,g}$ represents a scroll
$\Sigma\subseteq {\mathbb P}^ 3$ with ordinary singularities. This would
follow by going deeper into the analysis performed in
\cite{CCFMLincei}, but we do not use this here in the full
generality. We investigate below in more detail various examples
(see especially Example \ref{sextic}).
\end{rems}
\begin{exa}\label{quartic} {\em Elliptic quartic scrolls.} Let $E$ be
a smooth curve of type $(a,b)$ on ${\mathbb P}^1\times{\mathbb P}^1$, identified
with a smooth quadric in ${\mathbb P}^ 3$. The genus of $E$ is
$g=ab-a-b+1$. Consider a
pair of skew lines $R_1, R_2$ in ${\mathbb P}^ 3$. Identifying these lines
with the factors of ${\mathbb P}^1\times{\mathbb P}^1$, we can interpret the
canonical projections of $E$ to the factors as maps $\phi_i: E\to
R_i$, $i=1,2$, of degree $a$ and $b$, respectively. For each $x\in
E$ we consider the line $L_x$ joining the points
$\phi_i(x)$, $i=1,2$. This yields the map $\mu: x\in E\to L_x\in \rm Gr(1,3)$.
Its image is a smooth curve on $\rm Gr(1,3)$
under the Pl\"ucker embedding of the Grassmanian
$\rm Gr(1,3)$ as a quadric in ${\mathbb P}^5$. The associated scroll
\[\Sigma=\Sigma_{a,b}=\bigcup_{x\in E} L_x\]
in ${\mathbb P}^3$ with base $E$ has degree $a+b$. Indeed, it has
singularities of multiplicities $a$ along $R_1$ and $b$ along
$R_2$. So a line $\langle A,B\rangle$, where $A\in R_1$ and $B\in
R_2$, meets $\Sigma$ only in $A$ and $B$.
In particular, for $a=b=2$ we obtain a quartic scroll in ${\mathbb P}^3$
of genus $1$ with two skew double lines, and for $a=3,\, b=2$ a
quintic scroll of genus $2$ with a double line and a triple line.
From now on, we concentrate on an elliptic quartic scroll
$\Sigma=\Sigma_{2,2}$. The preimage $\Delta_S$ of $\Delta_\Sigma$
on $S$ consists of two disjoint copies $E_1,E_2$ of $E$ with
$\phi_i: E_i\to R_i$, $i=1,2$, corresponding to two distinct
$g_2^1$'s on $E$. There are in total 8 pinch points of
$\Sigma$, 4 on each of the lines $R_1, R_2$. These are the branch
points of the maps $\phi_i$, $i=1,2$. If these maps are
sufficiently general, also the pinch points are generically
located along $R_1, R_2$ and the ruling passing through a pinch
point does not contain any other pinch point.
Let us illustrate on this example our degeneration method. Any
smooth elliptic quartic curve is a complete intersection of two
quadrics in ${\mathbb P}^ 3$. Hence it embeds as well to the Grassmanian
$\rm Gr(1,3)$. By virtue of Remark \ref{rem:3space} to Theorem \ref
{thm:scrolls} (the case $r=3$) these curves fill in a unique
irreducible component $\mathcal H_{4,1}$ of the Hilbert scheme of
curves of degree $4$ in $\rm Gr(1,3)$, which dominates the moduli
space $\mathcal M_1$. The component ${\mathcal H_{4,1}}$ contains
all \emph{limit curves}, e.g. all reduced, nodal curves of degree
4 and arithmetic genus 1 spanning a ${\mathbb P}^ 3$. For instance, the
union $E_0$ of two conics $\Gamma_1,\Gamma_2$ meeting
transversally at two distinct points $f_1,f_2$ is such a limit
curve. The curve $E_0$ corresponds to the union $\Sigma_0$ of two
quadrics surfaces $Q_1,Q_2$ in ${\mathbb P}^3$ associated to the conics
$\Gamma_1, \Gamma_2$ on the Grassmanian $\rm Gr(1,3)$. We may assume
these quadrics to be smooth. They intersect along the
quadrilateral $F_1\cup F_2\cup G_1\cup G_2$, where the lines $F_1,
F_2$ correspond to $f_1, f_2$ and belong to the same ruling on
each quadric, and $G_1, G_2$ are distinct lines belonging to the
other ruling. We let $p_{ij}=F_i\cap G_j$, $i,j=1,2$.
The surface $\Sigma_0$ can be seen as a flat limit of surfaces of
type $\Sigma$, since it corresponds to a point in
${\mathcal H_{4,1}}$. The limit of the ruling of
$\Sigma$ is the union of the two rulings of $Q_1$ and $Q_2$
containing $F_1,F_2$. The limits of the double lines $R_1, R_2$
are the lines $G_1,G_2$. The limit of each of the components $E_i$
of the curve $\Delta_S$ on $S$ consists of two copies of $G_i$
glued at $p_{1i}, p_{2,i}$. Each of these points is the limit of
two pinch points of $\Sigma$.
Conversely, when we deform $\Sigma_0$ to $\Sigma$, the two double
lines $F_1$ and $F_2$ of $\Sigma_0$ disappear, because we are
smoothing the two nodes of $E_0$. Each of the points $p_{ij}$
($i,j=1,2$) gives rise to two pinch points generically located
along the double line of $\Sigma$, which deforms $G_j$. \end{exa}
\begin{exa}\label{quintic} {\em Elliptic quintic scrolls.} Consider now
the case where $d=5$ and $ g=1$. By Theorem \ref{thm:scrolls}, a
general point $[S] \in \mathcal{H}_{5,1}$ represents a smooth scroll $S$
in ${\mathbb P}^ 4$, whose general projection $\Sigma$ to ${\mathbb P}^ 3$ has
ordinary singularities. According to Bonnesen's formulas
(\ref{f100})-(\ref{f500}), the double curve $C=\Delta_\Sigma$ is
an irreducible, smooth, elliptic quintic curve, which contains the
$10$ pinch points of $S$. Its preimage $\tilde C=\Delta_S$ is a
smooth, irreducible curve on $S$ of genus $6$. By Lemma \ref
{199}, the rulings of $\Sigma$ are trisecant lines to $C$.
Conversely, for any smooth elliptic quintic curve $C$ in ${\mathbb P}^3$,
the trisecant lines to $C$ sweep out a quintic scroll $\Sigma$,
which is singular exactly along $C$ (cf.\ Berzolari's Formula,
Proposition 1 and Corollary 2 in \cite{MBer}). Such a surface
$\Sigma$ is an elliptic scroll, and by the Riemann--Roch Theorem
it comes as a projection of a surface represented by a point in
$\mathcal{H}_{5,1}$ as above.
Any such scroll $\Sigma$ corresponds to an embedding of an
elliptic quintic curve $E$ in $\rm Gr(1,3)$ via the map $\mu$ as in
\S \ref {ssec:general}. The image of $E$ is a quintic elliptic
normal curve, contained in a hyperplane section of $\rm Gr(1,3)$.
Indeed, any normal, elliptic quintic curve lies on some smooth
quadric in ${\mathbb P}^ 4$, hence on a hyperplane section of
$\rm Gr(1,3)$.
There is another interpretation of these elliptic quintic scrolls.
Let $E$ be an elliptic curve. Consider its symmetric product
$E(2)$, formed by all degree 2 effective divisors on $E$. The
class of the \emph{diagonal} $D=\{2p, p\in E\}$ is divisible by
$2$ in ${\rm Pic}(E(2))$; we denote by $\vartheta$ the class of
its half. One has $K_{E(2)}\sim -\vartheta$.
The Abel--Jacobi map $\alpha: E(2)\to {\rm Pic}^ {(2)}(E)\cong E$
makes $E(2)$ a ${\mathbb P}^ 1$--bundle with base $E$. The rulings are the
$g_2^ 1$'s on $E$. The \emph{coordinate curves} $E_p=\{x+p, x\in
E\}\cong E$ are unisecant curves of the rulings and form a
one--dimensional family parametrized by the point $p$ varying on
$E$. We have $E_p^ 2=1$. If $F_1, F_2$ are rulings, then the
divisor class of the curve $E_p+F_1+F_2$ is very ample on $E(2)$
and maps isomorphically the surface $E(2)$ onto a quintic scroll
$S$ in ${\mathbb P}^ 4$. Each coordinate curve $E_p$ is mapped to a
smooth plane cubic on $S$ which is the residual intersection of
$S$ with a hyperplane containing two rulings. Conversely any
smooth plane cubic on $S$ is a coordinate curve: indeed, it sits
on a 1--dimensional family of hyperplane sections of $S$ and their
residual intersections with $S$ is a pair of lines.
Let as before $\Sigma$ denote the image of $S$ under a general
projection ${\mathbb P}^4\dashrightarrow{\mathbb P}^ 3$. Any coordinate curve on $S$
is isomorphically mapped to a smooth plane cubic and
the images on $\Sigma$ of two distinct
coordinate cubics on $S$ are distinct. This provides a complete,
one--parameter family of smooth plane cubic curves on $\Sigma$
which are the only plane cubics on $\Sigma$.
Let $L$ be the plane containing one of them $\bar E$.
The residual intersection on
$L\cap\Sigma$ must be a union of two rulings, which meet on $C$.
The corresponding rulings on $S$ span a hyperplane
which cuts out on $S$ a coordinate cubic $\tilde E$
plus the two rulings. Hence $\bar E$ is the image of $\tilde
E$ on $\Sigma$.
Let $x\in C$ be a general point and $F_1,F_2$ the two
rulings through $x$. The plane $\pi$ spanned by them cuts
$\Sigma$ in the union of $F_1,F_2$ and a smooth cubic $\bar E$,
which is the projection of a unique coordinate curve. When $x$ varies,
we obtain in this way all projections of coordinate curves.
This shows that $C$ is isomorphic to $E$, since it parametrizes
the family of coordinate curves.
When the center of projection ${\mathbb P}^4\dashrightarrow{\mathbb P}^ 3$
varies we obtain a monodromy action. The following argument shows
that this monodromy is irreducible on appropriately chosen
objects.
The cubic curve $\bar E$ as above does not pass through
$x$, and cuts the ruling $F_i$ in three (generically distinct)
points $p_i, q_{i1},q_{i2}$, $i=1,2$, such that $q_{i1},q_{i2}\in
C$. Indeed, $p, q_{i1},q_{i2}$, $i=1,2$, are the five intersection
points of $\pi$ with $C$. By moving the centre of projection, we
may assume that the pair of rulings $(F_1,F_2)$ corresponds to a
general divisor of a given $g^ 1_2$ on $E$, and that
$q_{11}+q_{12}$ ($q_{21}+q_{22}$, respectively) is a general
divisor in the $g^ 1_2$ cut out on $\bar E$ by the lines through
$p_1$ (through $p_2$, respectively). In conclusion, by moving the
centre of projection the monodromy interchanges the pairs
$q_{11}+q_{12}$ and $q_{21}+q_{22}$ and also interchanges the
points in each pair separately.
\end{exa}
\begin{exa}\label{sextic} {\em Sextic scroll of genus two.}
By the case $g\ge 2,\,r=3$ of Remark \ref {rem:3space}.2,
there exist sextic scrolls $\Sigma$
of genus two in ${\mathbb P}^ 3$.
They correspond to genus 2 curves of degree 6
on the Grassmanian $\rm Gr(1,3)$. In fact,
by the Riemann--Roch Theorem, any smooth curve of genus 2
embeds in ${\mathbb P}^ 4$ as a sextic. This sextic spans ${\mathbb P}^ 4$
and lies on
a smooth quadric in ${\mathbb P}^ 4$, hence on a hyperplane section of
the Grassmanian $\rm Gr(1,3)$.
These curves fill in a unique component $\mathcal H_{6,2}$
of the Hilbert
scheme of curves of degree $6$ and genus 2 in $\rm Gr(1,3)$,
which dominates
$\mathcal M_2$ via the natural map.
As in Example \ref {quartic}.2,
${\mathcal H_{6,2}}$ contains
{limit curves}, and in particular all reduced, nodal curves of
degree 6 and arithmetic genus 2 spanning a ${\mathbb P}^ 4$.
Assuming that a general such scroll $\Sigma$ has ordinary
singularities, Lemma \ref {199} and Proposition \ref {300} say
that the rulings of $\Sigma$ are four--secant lines to the double
curve $C=\Delta_\Sigma$, which is a smooth, irreducible curve in
${\mathbb P}^3$ of degree $8$ and genus $5$, passing through all $16$
pinch points of $\Sigma$. The preimage $\tilde C=\Delta_S$ of $C$
on $S$ is a smooth curve of degree $16$ and of genus $17$.
Let us show that a general sextic scroll $\Sigma$ in ${\mathbb P}^3$ of
genus 2 has ordinary singularities and an irreducible double curve
$C$. Consider a reducible sextic curve $E_0\subseteq {\mathbb P}^ 4$ of
arithmetic genus $2$, which consists of a general smooth elliptic
normal quintic curve $E'$ and a line $D$ meeting $E'$
transversally in two distinct points. Such a curve $E_0$
corresponds to a point in ${\mathcal H_{6,2}}$, hence to a
reducible surface $\Sigma_0$, which is a limit of genus 2 sextic
scrolls $\Sigma$. On the other hand, $\Sigma_0$ is the union of a
general quintic elliptic scroll $\Sigma'$ in ${\mathbb P}^3$ arising from
$E'$ as in Example \ref {quintic} plus a plane $\pi$ through the
two rulings $F_1,F_2$ of $\Sigma_0$, which correspond to the
intersection points of $E'$ and $D$. These rulings meet at a point
$p$ of the double curve $C'$ of $\Sigma'$. The ruling on $\pi$ is
given by the pencil of lines passing through $p$, which
corresponds to the line $D$. The plane $\pi$ cuts out on $\Sigma'$
the union of the rulings $F_1$, $F_2$ and a smooth plane cubic
$\bar E$, as described in Example \ref {quintic}. The
singularities of $\Sigma'$ consist of $C'$, $F_1$, $F_2$, and
$\bar E$.
When we deform $E_0$ to a general smooth sextic $E$ on $\rm Gr(1,3)$,
the scroll $\Sigma_0$ is deformed to an irreducible sextic scroll
$\Sigma$. The double lines $F_1$ and $F_2$ of $\Sigma_0$
disappear, because we are smoothing the two nodes of $E_0$. This
means that the flat limit on $\Sigma_0$ of the singular locus of
$\Sigma$ is the nodal curve $C_0=C'\cup\bar E$ of arithmetic
genus $5$. Hence $\Sigma$ is singular only along a double curve
$C$, which has arithmetic genus $5$. The latter curve is
irreducible. Indeed, otherwise this would be still a union of the
form $C'\cup\bar E$, and so the four-secant lines to $C$ would
sweep out a union of an elliptic scroll and a plane. However, this
is impossible since $\Sigma$ is irreducible and swept out by the
four-secants of the double curve $C=\Delta_\Sigma$.
Since $C_0$ is nodal so is
$C$. We claim that $C$ is actually smooth. Indeed, we may
restrict our family to a general irreducible curve germ in
${{\ensuremath{\mathcal{H}}}_{2,6}}$
through
$\Sigma_0$, and then normalize this germ. In this way we obtain a
family of sextic scrolls over the disc $\ensuremath{\mathbb{D}}$ with a family
${\ensuremath{\mathcal{C}}}\to\ensuremath{\mathbb{D}}$ of double curves. The central fibre of ${\ensuremath{\mathcal{C}}}$ is a
reducible nodal curve $C_0=C'\cup\bar E$ with $4$ nodes. Assuming
that no one of these nodes is smoothed on a general fibre $C$ of
the family, $C$ should also have $4$ nodes. These nodes represent
an \'etale four-sheeted cover over the disc. Now we can normalize
the fibres of the family ${\ensuremath{\mathcal{C}}}$ simultaneously (see e.g.,
\cite{Se}), thus obtaining a smooth family with an irreducible
general fibre and a disconnected cental fibre. The latter
contradicts the \emph{Connectedness Principle} (see \cite [Ch.
III, Ex.\ 11.4, p.\ 281] {har}).
Consequently,
at least one of the four nodes of $C_0$ has to be smoothed
in the deformation to $C$.
But then by the irreducibility of the monodromy
(see the final part of Example \ref {quintic}) all nodes of
$C_0$ have to be smoothed. \end{exa}
\section{Bounding degrees of low genera curves on surfaces}
\subsection{Algebraic hyperbolicity} \label{ssec:ah} Scrolls can be
used to establish algebraic hyperbolicity of very general surfaces
of a given degree $d$ in ${\mathbb P}^3$. For $d\ge 6$ this is done in
Proposition \ref{201} below. In the proof we use the Albanese
inequality (see \cite {Alb, No} (see also \cite[\S
4(b)]{Morr}), which says the following: if a reduced projective
curve $C$ of geometric genus $g$ degenerates into an effective
cycle $C_0=\sum_i m_iC_i$, where $C_i$ is a reduced projective
curve of geometric genus $g_i$, then \begin{eqnarray}\label{alb} g\ge
\sum_{g_i\ge 1} (m_i(g_i-1)+1)\,.\end{eqnarray} In particular, $m_i(g_i-1)\le
g-1$ if $g_i\ge 1$. So $g_i\le g$ for all $i$.
\begin{prop}\label{201} Assume that there exists a scroll $\Sigma$ of
degree $d\ge 5$ and genus $g\ge 1$ in ${\mathbb P}^3$ with ordinary
singularities. Then a very general surface $X$ in ${\mathbb P}^3$ of
degree $d$ does not contain curves of geometric genus $g'<g$.
\end{prop}
\begin{proof} Let $X$ be a very general surface in ${\mathbb P}^3$ of degree
$d$. By the Noether-Lefschetz Theorem, the Picard group of $X$ is
generated by $\mathcal O_X(1)$. Consider the pencil $\{X_t\}_{t\in
{\mathbb P}^ 1}$ generated by $X_0=\Sigma$ and $X_\infty=X$. This gives
rise to a flat family of surfaces $f: \mathcal X\to \mathbb D$
over a disc $\mathbb D$, where the {central fibre} over $0$ is
$X_0$, all fibres $X_t$ with $t\in\ensuremath{\mathbb{D}}\setminus\{0\}$ are smooth
and ${\rm Pic}(X_t)$ is generated by $\mathcal O_{X_t}(1)$ for a
very general such fibre. We claim that a very general surface of
this family does not contain any curve of geometric genus $g'<g$.
We argue by contradiction and assume that this is not the case for
some $g'< g$.
For each positive integer $n$ we may consider the locally closed
subset $\mathcal H_{n,g'}$ of the relative Hilbert scheme of $f:
\mathcal X\setminus X_0\to \mathbb D\setminus \{0\}$, whose points
correspond, for each $t\neq 0$, to the irreducible curves of
geometric genus $g'$ in $\vert \mathcal O_{X_t}(n)\vert$. By our
assumption, there is a component of $\mathcal H_{n,g'}$ which
dominates $\ensuremath{\mathbb{D}}\setminus \{0\}$. Let $\mathcal H$ be the closure of
this component in the relative Hilbert scheme of $f: \mathcal X\to
\mathbb D$. By the properness of the relative Hilbert scheme,
$\mathcal H$ surjects onto $\mathbb D$. Hence there is a curve
$C_0\in \mathcal O_{X_0}(n)$ on $X_0$, which corresponds to a
point in $\mathcal H$. By Albanese's inequality (\ref{alb}), every
component of $C_0$ has geometric genus $g''\le g'<g$. By \eqref
{f200} (for $g\ge 2$) and Example \ref{quintic} (for $g=1$) we
have $\gamma\ge g>g''$, where $\gamma$ stands as before for the
geometric genus of the double curve $\Delta_\Sigma$ of
$X_0=\Sigma$. Hence no component of $C_0$ coincides with
$\Delta_\Sigma$. Now the pull--back $\Gamma$ of $C_0$ on the
normalization $\phi: S\to \Sigma$ belongs to the linear system
$\vert \phi^ *(\mathcal O_\Sigma(n))\vert$ and maps birationally
to $C_0$ by the finite map $\phi$. Since the only curves of genus
smaller than $g$ on $S$ are rulings, $\Gamma$ consists of rulings.
In particular, $\Gamma^ 2=0$. On the other hand, since $\Gamma \in
\vert \phi^ *(\mathcal O_\Sigma(n))\vert=\vert \mathcal
O_S(n)\vert$ we have $\Gamma^2=n^2d>0$, a contradiction.
\end{proof}
In Proposition \ref{201.1} below we slightly strengthen Proposition \ref
{201}, using Proposition \ref{acz} and Corollary \ref{acz2}.
Keeping in mind Example \ref {quintic}, Proposition \ref{201}
provides an alternative quick proof of the following result
originally established by Xu \cite{Xu1} and Voisin \cite{Vo1,
Vo2}.
\begin{cor}\label{agthm} On a very general surface of degree $d\ge 5$ in
${\mathbb P}^3$ there is no rational curve. \end{cor}
\emph{Very general} in Corollary \ref{agthm}
can be replaced by \emph{general}
provided the following question is answered in negative.
\begin{quest}\label{212} {\em Does there exist a sequence of
smooth quintic surfaces $X_n$ in ${\mathbb P}^3$ such that $X_n$ contains
a rational curve of degree $d_n$ and not smaller, with
$d_n\to\infty$?}
\end{quest}
\begin{rem}\label{213} Notice that for any integers $n \ge 3$, $d > 0$
and $0\le \delta \le d^2(n-1)+1$, the linear system $\vert
\mathcal O_S(d)\vert$ on a general K3 surface $S$ of degree $2n-2$
in ${\mathbb P}^ n$ with Picard group generated by $\mathcal O_S(1)$,
contains a $(d^2(n -1)-\delta + 1)$--dimensional family of
irreducible $\delta$--nodal curves, whose geometric genus equals
$d^2(n -1)-\delta + 1$ (see \cite {chen}). So $S$ contains nodal
curves of every geometric genus $g\ge 0$. This applies in
particular to general quartic surfaces in ${\mathbb P}^ 3$. \end{rem}
\subsection{Bounding degrees of curves of low genera on general surfaces in
${\mathbb P}^3$}\label{ssec:bound} In this section we address the
following \emph{boundedness question} (cf.\ \cite{Bo, LM}
and the related discussion in the Introduction):
\begin{quest} \label{quest:bound} {\em Given integers $d\ge 5$ and $g\ge 0$,
does there exist a bound $n_{d,g}$ such that every irreducible
curve of geometric genus $g$ on a very general surface of degree
$d$ in ${\mathbb P}^ 3$ has degree $n\le n_{d,g}$?}
\end{quest}
If $d=4$ the answer is negative (see \cite {chen, Hal} and Remark
\ref{213}). The argument in the proof of Propositions \ref {201}
and \ref {201.1} can be used to give an affirmative answer for
$d\ge 6$ and small enough $g$.
\begin{prop}\label{prop:bound} Suppose there exists a scroll $\Sigma$ of
degree $d\ge 6$ and genus $g\ge 2$ with ordinary singularities.
Then the answer to Question \ref {quest:bound} is affirmative for
all genera $g'<\gamma$, where $\gamma$ is defined in (\ref{401}).
\end{prop}
\begin{proof} We apply the same argument as in the proof of Proposition
\ref {201}. Keeping the notation of this proposition, we let again
$C_0\in \vert \mathcal O_\Sigma(n)\vert$ denote a curve which is a
limit of a flat family of irreducible curves $\{C_t\}_{t\in
\mathbb D-\{0\}}$, $C_t\in \vert \mathcal O_{X_t}(n)\vert$, of
genus $g'$, where $g'\ge g\ge 2$ by Proposition \ref{201}.
Write $C_0=m_1C_1+\ldots+m_hC_h+C'$ as a cycle, where for every
$i=1,\ldots, h$ the curve $C_i$ is irreducible of geometric
genus $g_i\ge 1$ and its transform on $S$ has positive
intersections $n_i$ with the rulings, whereas $C'$ consists of
rulings. Note that $n=\sum_{i=1}^ h m_in_i$. By Albanese's
inequality (\ref{alb}) and our hypothesis $g'<\gamma$, none of the
components of $C_0$ coincides with $\Delta_\Sigma$, and
\[
g'\ge h+ \sum_{i=1}^ h m_i(g_i-1)\,.
\]
The Riemann--Hurwitz formula yields: $g_i-1\ge n_i(g-1)$ for all
$i=1,\ldots, h$, so that $\gamma> g'\ge h + n(g-1)$. This
provides a bound $n<(\gamma-1)/(g-1)$ (we remind that $g\ge
2$).
\end{proof}
\begin{cor}\label{cor:bound} Question \ref {quest:bound} has an
affirmative answer for
\[
\begin{aligned}
&d=6, \quad g\le 5\,,\cr &d\ge 7 \quad {\text even}, \quad
g<(d-4)^2 \,,\cr &d\ge 7 \quad {\text odd}, \quad g<\frac
{(d-3)(2d-9)}2\,\,.\cr
\end{aligned}
\]
\end{cor}
\begin{proof} For $d=6$ we use the sextic scroll of genus 2 as in
Example \ref {sextic}. For $d\ge 7$ even we write $d=2m+4$ and we
consider in ${\mathbb P}^3$ general projections of smooth scrolls of genus
$m$ and degree $d$ in ${\mathbb P}^ 5$ as in Theorem \ref {thm:scrolls}.
For $d\ge 7$ odd we write $d=2m+3$ and we consider general
projections of scrolls of genus $m$ and degree $d$ in ${\mathbb P}^ 4$ as
in Remark \ref {rem:3space}.2. Applying Proposition \ref
{prop:bound} and taking into account \eqref {f200}, the assertion
follows. \end{proof}
\subsection{Families of low degree curves of a given genus
on general surfaces in ${\mathbb P}^3$}
Proposition \ref{acz} below
extends a similar result by Arbarello--Cornalba \cite[Theorem
3.1]{ac2}, \cite {ac1} and Zariski \cite{za}; cf. also Knutsen
\cite[Lemma 4.4]{Kn}.
Let $S$ be a smooth projective surface, Hilb$_1(S)$ the Hilbert
scheme of curves on $S$, and $\mathcal V_g(S)$ the locally closed
subset of Hilb$_1(S)$ formed by irreducible curves of geometric
genus $g$.
\begin {prop} \label{acz} In the setting as before,
for an irreducible component $\mathcal V$ of $\mathcal V_g(S)$ we
let $v=\dim(\mathcal V)$ and $\kappa=K_S\cdot \Gamma$, where a
curve $\Gamma$ on $S$ corresponds to a general point in $\mathcal
V$. Then $v\le \max\{g,g-1-\kappa\}$. Furthermore, if $v>g$ then
$v=g-1-\kappa$, and the general curve $\Gamma$ of $\mathcal V$ has
only nodes as singularities.
\end{prop}
\begin{proof} Let $f: C\to \Gamma$ be the normalization.
The exact sequence
\[
0\to T_C\to f^ *(T_S)\to N_f\to 0
\]
defines the \emph{normal sheaf} $N_f$ to the map $f: C\to S$. It
can be included into an exact sequence
\[
0\to \tau \to N_f \to N'\to 0 \,,
\]
where $\tau$ is the torsion subsheaf of $N_f$ supported at the
points, where the rank of the differential of $f$ drops, and $N'$
is an invertible sheaf. Due to the {\em Horikawa inclusion}
$T_{[\Gamma]}(\mathcal V)\subseteq H^ 0(C,N')$ (see
\cite[(1.3)]{ac2} or \cite[Lemma 1.4]{ac1}) we have $v\le
h^0(C,N')$. By Riemann-Roch, $$h^0(C,N')=\mathop{\rm deg} (N')-g+1+h^1(C,N'),
\quad\text{where}\quad \mathop{\rm deg}(N')\le\mathop{\rm deg}(N_f)=2g-2-\kappa\,.$$ If
$h^1(C,N')=0$ this gives $v\le g-1-\kappa$. Otherwise $N'$ is
special, so $h^0(C,N')\le g$.
In any case, $v\le \max\{g,g-1-\kappa\}$, as stated.
If $v>g$ then $h^ 1(C, N')=0$.
Since $H^1(C,\tau)=0$
this yields $H^ 1(C, N_f)=0$. As in \cite[proof of (1.5) and p.
96]{ac2} this implies $\tau=0$, hence $\Gamma$ is immersed
(i.e., has no cuspidal
singularities). One ends the proof as in \cite[pp. 96--98]{ac2}.
\end{proof}
For $\mathcal L_d=\vert \mathcal O_{{\mathbb P}^ 3}(d)\vert$ we let
\begin{eqnarray}\label{400} N_d=\dim\,(\mathcal L_d)={{d+3}\choose 3}-1\,. \end{eqnarray}
Given a smooth surface $X$ of degree $d$ in ${\mathbb P}^ 3$ and
non--negative integers $n,g$, we let $\mathcal V_{n,g}=\mathcal
V_{n,g}(X)$ denote the locally closed subset of $\mathcal
L_{X,n}=\vert \mathcal O_X(n)\vert$ formed by irreducible curves
on $X$ of geometric genus $g$. We also let
\[
g_{d,n}=\frac {dn(d+n-4)}2+1\,
\]
denote the arithmetic genus of the curves in $\mathcal L_{X,n}$.
Notice that $g_{d,n}=g+\nu$ if a general member of $\mathcal
V_{n,g}$ is nodal with $\nu$ nodes.
\begin{cor}\label{acz2} Let $X$ be a general surface of degree $d\ge 3$
in ${\mathbb P}^ 3$. If $g\ge 0$ and $n\in\{1,2\}$ are such that
$\mathcal V_{n,g}$ is nonempty, then
\[
g_{d,1}-3 \le g\le g_{d,1}
\quad \text {if} \quad n=1 \quad \text {and} \quad g_{d,2}-9\le
g\le g_{d,2} \quad \text {if} \quad n=2\,.
\]
Furthermore, for every irreducible component $\mathcal V$ of $\mathcal
V_{n,g}$, its general curve has exactly $\nu$ nodes as singularities and
its dimension is
\begin{eqnarray}\label{trahtibidoh} 3-\nu=
g-g_{d,1}+3 \quad \text {if} \quad n=1\quad \text {and}\quad
9-\nu= g- g_{d,2}+9
\quad \text {if} \quad n=2\,.
\end{eqnarray}
\end{cor}
\begin{proof} Let us show the assertion in the case $n=2$,
the case $n=1$ being similar. Consider the incidence relation
$I\subseteq\mathcal L_d \times \mathcal L_2$ consisting of all
pairs $(X,Q)$ such that $X$ is smooth and $Q$ and $X$ intersect in an
irreducible curve $C$ of geometric genus $g$. Then $I$ is locally
closed and comes equipped with the natural projections $p: I\to
\mathcal L_d$ and $q: I\to \mathcal L_2$.
Note that if $(X,Q)\in I$ and $C$ is the intersection of $X$ and
$Q$, then we have a family of dimension $\dim(\mathcal L_{d-2})+1$
of pairs $(X',Q)\in I$ such that intersection of $X'$ and $Q$ is
$C$: indeed we can take $X'$ general in the span of $X$ and of all
surfaces of degree $d$ containing $Q$.
By our assumption $p$ is dominant. Let $I'$ be an irreducible
component of $I$ which dominates $\mathcal L_d$ via $p$, so that
$\dim(I')\ge N_d$. We assume that $q(I')$ contains a smooth
quadric $Q$ (the argument is similar otherwise, the details are
left to the reader). Then $I'$ dominates $\mathcal L_2$ via $q$
and we may assume $Q$ to be a general quadric. All components of
$q^ {-1}(Q)$ have dimension $\dim(I')-\dim(\mathcal L_2)$. Any
such component can be identified with a family of surfaces of
degree $d$. By the above discussion, the family of curves
$\mathcal V$ they cut out on $Q$ has dimension
\[v=\dim(I')-\dim(\mathcal L_{d-2})-\dim(\mathcal L_2)-1\,.\]
Moreover, $\mathcal V$ is an irreducible component of $\mathcal
V_{d,g}(Q)$. We have \[v\ge
N_d-N_{d-2}-N_2-1=g_{d,2}+4d-10>g_{d,2}\ge g.\] By Proposition
\ref {acz}, one has $v=g-1+4d$, which yields $g_{d,2}-g\le 9$.
Furthermore, by Proposition \ref{acz} the general curve in
$\mathcal V$ has at most nodes as singularities, which implies
(\ref{trahtibidoh}).
\end{proof}
Corollary \ref {acz2} could be extended to handle also the case $n=3$.
This requires however to analyze a number of cases,
which we avoid here.
Now we can strengthen Proposition \ref{201} as follows.
\begin{prop}\label{201.1} Assume that there exists a scroll
$\Sigma$ of degree $d\ge 5$ and genus $g\ge 1$ in ${\mathbb P}^3$ with ordinary
singularities. Then a
very general surface $X$ of degree $d$ in ${\mathbb P}^3$ does not contain
curves of geometric genus $g'\le 3(g-1)$. \end{prop}
\begin{proof}
By Proposition \ref{201} we may suppose that $d\ge 6$ and $g'\ge
g\ge 2$. We proceed as in the proof of this proposition, using the
same notation. We argue by contradiction and assume that there is
a positive $g'\le 3(g-1)$, a positive integer $n$ and a component
of $\mathcal H_{n,g'}$ which dominates $\mathbb D\setminus \{0\}$.
Consider a curve $C_0\in{\ensuremath{\mathcal{O}}}_{\Sigma}(n)$ as in the proof of
Proposition \ref {201}. As shown in this proof, $C_0$ cannot be
composed of rulings. Hence it contains a component $C_i$ of
geometric genus $g_i>0$, appearing in $C_0$ with multiplicity
$m_i$. By Albanese's inequality \eqref {alb} one has $g'-1\ge
m_i(g_i-1)$. By (\ref{002}) and our assumption $g'\le
3(g-1)<\gamma$, hence $C_i\neq \Delta_\Sigma$. Therefore $C_i$
lifts birationally to the normalization $S$ of $\Sigma$ yielding a
$\nu_i$--secant of the ruling on $S$. Combining the inequalities
above, by Hurwitz Formula (see Remark \ref {rem:genus}) we obtain
$$3(g-1)-1\ge g'-1\ge m_i(g_i-1)\ge \nu_i m_i(g-1)\,.$$
Hence $\nu_i m_i\le 2$ and so the only
possibilities are
\[ \nu_i=m_i=1, \quad \nu_i=1,\, m_i=2,\quad {\rm and}\quad
\nu_i=2,\, m_i=1.\]
In the former case by \eqref {alb} there can be at most two such
components, while in the latter two cases at most one.
We have $n=\sum_i\nu_i m_i$, the sum over all components
$C_i$ of $C_0$ of positive genus.
It follows that $1\le n\le 2$. Then Corollary \ref
{acz2} yields $g'\ge g_{d,1}-3$, since $g_{d,1}-3<g_{d,2}-9$ for
$d\ge 6$. Thus we must have
\begin{equation}\label{eq:ineq}
\frac {(d-1)(d-2)}2-3=g_{d,1}-3\le g'\le 3(g-1)\le \frac
{d(d-5)}2\,,
\end{equation}
the last inequality coming from \eqref {f300}
for $d\ge 6$. But \eqref {eq:ineq}
gives a contradiction.\end{proof}
\section{Bounding geometric genera of divisors on general
$3$-folds in ${\mathbb P}^4$}
\label{sec:higher} A simple way of constructing higher dimensional
scrolls consists in starting with the trivial ${\mathbb P}^1$--bundle
$\pi:S=E\times{\mathbb P}^1\to E$ over a smooth projective variety
$E\subseteq{\mathbb P}^m$ of degree $d$ and dimension $n$. Let
$ \operatorname{{\rm Seg}}_{a,b}$ denote the image of ${\mathbb P}^a\times{\mathbb P}^b$ via the
\emph{Segre embedding}. Then
$$S\hookrightarrow{\mathbb P}^m\times{\mathbb P}^1\stackrel{\simeq}
{\longrightarrow} \operatorname{{\rm Seg}}_{m,1} \hookrightarrow{\mathbb P}^{2m+1}\, $$ yields
an embedding of $S$ as a smooth scroll of dimension $n+1$ and
degree $(n+1)d$ in ${\mathbb P}^{2m+1}$. A general linear projection of
$S$ to ${\mathbb P}^{n+2}$ gives a hypersurface scroll $\Sigma\subseteq
{\mathbb P}^{n+2}$ of degree $(n+1)d$.
Consider, for instance, a surface $E_d$ in ${\mathbb P}^3$ of degree $d$,
which we suppose to be very general. The above construction gives
$$S_d:=E_d\times{\mathbb P}^1\hookrightarrow \operatorname{{\rm Seg}}_{3,1}\hookrightarrow{\mathbb P}^7\,,$$
and $S_d$ is a threefold of degree $3d$ in ${\mathbb P}^7$. A general
linear projection of $S_d$ to ${\mathbb P}^4$ yields a threefold scroll
$\Sigma_d$ of degree $3d$ in ${\mathbb P}^4$. It is swept out by a
two-dimensional family of $(3d-3)$-secant lines to the double
surface $\Delta_\Sigma$ (see Lemma \ref{199}).
The following version of the Albanese inequality follows
immediately from the Semistable Reduction Theorem \cite[\S
1]{Morr} and the Geometric Genus Criterion (see formula (1) on p.\
119 in \cite[\S 6]{Morr} or, in the surface case, formula (8) in
\cite[Ch.\ 5, \S 5]{KK}).
\begin{lem}\label{lem:limit} Let $X$ be a flat limit of a one-parameter
family of smooth, irreducible, projective varieties
of geometric genus $\rho$. Let $X_i$ be irreducible components of
$X_0$ with geometric genera $\rho_i$, $i=1,\ldots,h$. Then
\[ \rho\ge \sum_{i=1}^ h \rho_i\,.\]
\end{lem}
Recall that, in the notation as in (\ref{400}), the geometric
genus $\rho(E_d)$ of a smooth surface $E_d$ of degree $d$ in
${\mathbb P}^3$ is equal to $\rho(E_d)={{d-1}\choose{3}}=N_{d-4}+1$ (see e.g.\
\cite[Ch.\ 4, (5.12.2)]{KK}). The following lower bound on the
geometric genus of the double surface is an analog of (\ref{002})
in the case of surface scrolls.
\begin{lem}\label{lem:limit2} Let $\Sigma_d\subseteq {\mathbb P}^ 4$ be a
threefold scroll of degree $3d$, constructed as before over a very
general surface $E_d$ in ${\mathbb P}^3$ of degree $d\ge 5$ as a base.
Then for the geometric genus $\rho_d$ of the double surface
$\Delta_{\Sigma_d}$ we have a lower bound
\begin{equation}\label{eq:pd}
\quad \forall d\ge 5\,.
\end{equation}
\end{lem}
\begin{proof} Degenerate $E_d$ to $E_{d-1}\cup E_1$, where $E_{d-1}$ and
$E_1$ are general. Then $S_d$ degenerates to the union of
$S_{d-1}$ and $S_1= \operatorname{{\rm Seg}}_{1,2}$, meeting along the Segre image $X$
of $C\times {\mathbb P}^ 1$, where $C=E_1\cap E_{d-1}$. Accordingly,
$\Sigma_d$ degenerates in ${\mathbb P}^4$ to the union of $\Sigma_{d-1}$
and $\Sigma_1$, the latter being a hypersurface of degree 3 with a
double plane. These threefolds intersect along the general
projection $Y$ of $X$, plus another surface $Z$. The limit of the
double locus $\Delta_{\Sigma_d}$ consists of the union of
$\Delta_{\Sigma_{d-1}}$, of the plane $\Delta_{\Sigma_{1}}$, and
of $Z$. The ruling determines a dominant rational map $Z\dasharrow
E_{d-1}$. So there is at least one component $Z'$ of $Z$ with
geometric genus $\rho'\ge\rho(E_{d-1})$.
Now the first inequality in \eqref {eq:pd} follows from Lemma
\ref {lem:limit}. In particular, $\rho_5\ge\rho'\ge 1$. By
induction for every $d\ge 5$ we obtain $$\rho_d\ge\rho(E_{d-1})+
\rho_{d-1}\ge\sum_{k=5}^{d-1}
\rho(E_{k})+\rho_5\ge\sum_{k=0}^{d-2}
{{k}\choose{3}}={{d-1}\choose{4}}\,,$$ as required. \end{proof}
It would be interesting to find the precise value of $\rho_d$.
\begin{thm}\label{thm:pg} Any irreducible surface contained in a
very general hypersurface of degree $3d\ge 15$ in ${\mathbb P}^4$
has geometric genus $\rho\ge \min\{\rho_d, N_{d-4}+1\}$. In
particular,
$\rho\ge \rho(E_{d})$ if $d\ge 8$.
\end{thm}
\begin{proof} The argument is similar to that in the proof of
Proposition \ref {201}, so we will be brief. We let $X_0$ be the
scroll $\Sigma_d$ and $X$ be a general hypersurface in ${\mathbb P}^4$
of degree $3d$. The pencil generated by $X_0$ and $X$ gives rise
as usual to a flat family $f: \mathcal X\to \mathbb D$. Suppose
that the general fibre of this family contains an irreducible
surface $Y$ of geometric genus $\rho< \min\{\rho_d, N_{d-4}+1\}$.
By Lemma \ref {lem:limit2} the limit $Y_0$ of such a surface in
the central fibre does not contain $\Delta_{\Sigma_d}$. By Lemma
\ref {lem:limit} all of its components have geometric genus
$\rho'\le\rho< \min\{\rho_d, N_{d-4}+1\}\le N_{d-4}+1$. Hence they
cannot dominate $E_d$, which has geometric genus $N_{d-4}+1$. Thus
all components of $Y_0$ pull--back to $S_d$ to surfaces with zero
intersection with the ruling. This yields a contradiction as in
the proof of Proposition \ref {201}.
\end{proof}
\begin{rem}\label{rem:bound2} G. Xu gave in \cite[Theorem 2]{Xu1} a
sharp lower bound for the geometric genus of an irreducible
divisor on a very general hypersurface of degree $d\ge n+2$ in
${\mathbb P}^ n$, with $n\ge 4$. Of course Theorem \ref {thm:pg} above is
weaker than Xu's result. However, the method of proof is simple
and it may possibly have further applications. Hence it would be
interesting to extend Theorem \ref {thm:pg} to other degrees
(non-divisible by $3$), as well as to higher dimensions. We wonder
also whether in higher dimensions an analog of Proposition
\ref{prop:bound} holds. For instance, one can suggest by analogy
that on a very general threefold in ${\mathbb P}^4$ of degree $\ge 6$, the
divisors of geometric genera $\rho'<\rho_d$ form bounded families.
\end{rem}
\section{Degeneration to scrolls and Kobayashi hyperbolicity}\label{sec:brody}
\subsection{Limiting Brody curves and Hurwitz Theorem}
Let $V$ be a subvariety of a
hermitian complex manifold.
A {\em Brody curve} in $V$ is a holomorphic map
$f:\ensuremath{\mathbb{C}}\to V$ satisfying
$$\sup_{z\in\ensuremath{\mathbb{C}}} ||df(z)||=||df(0)||=1\,.$$ By \emph{Brody's
reparametrization lemma} (\cite {Bro}), if $V$ is proper and
non--hyperbolic then it contains a Brody curve. Furthermore, from
any sequence of Brody curves in $V$ one can extract a subsequence
converging to a Brody curve, which is called a {\em limiting Brody
curve}.
Assume there is a proper dominant map $\pi:V\to C$ onto a smooth
projective curve $C$. If general fibres $D_c=\pi^{-1}(c)$ ($c\in
C$) are non--hyperbolic, i.e., contain Brody curves, then every
special fibre $D_0:=D_{c_0}$ is non-hyperbolic as well and
contains limiting Brody curves. The Hurwitz Theorem imposes
constrains on limiting Brody curve with respect to the
singularities of $D_{0}$ (cf.\ e.g., \cite[\S 1]{SZ1},
\cite[Theorem 2.1]{Za1}, and \cite[Lemma 1.2]{Za2}). Let
$\Delta_0= \operatorname{{\rm br}} (D_{0})$ be the set of multi--branch points of
$D_{0}$ such that locally the branches of $D_{0}$ are
$\ensuremath{\mathbb{Q}}$--Cartier divisors on $V$, and let $\Delta$ be the Zariski
closure of ${\Delta_0}$. Consider a limit $f:\Omega \to D_{0}$ of
a sequence of holomorphic maps $f_n:\Omega \to
D_{c_n}$, with $c_n\in C\setminus \{c_0\}$ such that $c_n\to c_0$,
where $\Omega\subseteq\ensuremath{\mathbb{C}}$ is a connected domain. Hurwitz' Theorem
says that, if $f(\Omega)\cap\Delta_0\neq\emptyset$, then
$f(\Omega)\subseteq\Delta$. In particular, if $\Delta$ is
hyperbolic then any limiting Brody curve in $D_{0}$ is contained
in $D_{0}\setminus \Delta_0$. Hence if both $\Delta$ and
$D_{0}\setminus \Delta_0$ are hyperbolic then all fibres $D_{c}$
($c\neq c_0$) close enough to $D_{0}$ are hyperbolic as well (cf.\
\cite{Za1}).
\subsection{A hyperbolicity criterion for hypersurfaces in
${\mathbb P}^n$}
Let $X_0$, $X_\infty$ be distinct hypersurfaces in ${\mathbb P}^n$ of
degree $d$. Typically, $X_\infty$ will be a general surface of
degree $d$ meeting $ \operatorname{{\rm Sing}}(X_0)$ in points, where locally $X_0$ is
a union of two smooth branches intersecting transversally.
Consider the associated linear pencil $\{X_t\}_{t\in {\mathbb P}^ 1}$.
Assume that for a general $t\in{\mathbb P}^1$ the hypersurface $X_t$ is
non--hyperbolic. Then there exists a sequence of Brody curves
$\varphi_n:\ensuremath{\mathbb{C}}\to X_{t_n}$ (with respect to the Fubini--Study
metric on ${\mathbb P}^n$), where $t_n\to 0$, converging to a limiting
(non--constant) Brody curve $\varphi_0:\ensuremath{\mathbb{C}}\to X_{0}$.
\begin{prop} \label{pr1} In the above setting, let $B=
X_\infty\cap\overline{ \operatorname{{\rm br}}( X_0)}$. If $\overline{ \operatorname{{\rm br}}( X_0)}$
and $(X_0\setminus \operatorname{{\rm br}} (X_0))\cup B$ are both hyperbolic, then
$X_{t}$, for $t\neq 0$ close enough to $0$, is hyperbolic as
well.
\end{prop}
\begin{proof} By Hurwitz' Theorem and the hypotheses, the image of
$\varphi$ cannot be contained in $\overline{ \operatorname{{\rm br}}( X_0)}$, and it
can meet $\overline{ \operatorname{{\rm br}}( X_0)}$ only at $\left(\overline{ \operatorname{{\rm br}}(
X_0})\setminus \operatorname{{\rm br}}( X_0)\right)\cup B$. But then it is contained
in $(X_0\setminus \operatorname{{\rm br}} (X_0))\cup B$, a contradiction. \end{proof}
\begin{rem}\label{010} Hurwitz' Theorem cannot be applied at points in
$\left(\overline{ \operatorname{{\rm br}}( X_0})\setminus \operatorname{{\rm br}}( X_0)\right)\cup B$, e.g.
at a \emph{pinch point} of $X_0\subseteq {\mathbb P}^3$, where $X_0$ is
locally analytically isomorphic to the surface $x^ 2=yz^ 2$ in
$\mathbb A^ 3=\mathbb A_\ensuremath{\mathbb{C}}^ 3$ at the origin, or at a base
point of the pencil situated on $ \operatorname{{\rm br}}( X_0)$.
Indeed, consider a linear pencil of surfaces given in an affine
chart $\mathbb A^3$ of ${\mathbb P}^3$ as $X_t=\{x^2-y^2z=t\}$. The origin
${\bf 0}\in \mathbb A^3$ is a pinch point of $X_0$ and is not a
base point of the pencil. Consider also the family of entire
curves
$$\phi_\tau: \ensuremath{\mathbb{C}}\to X_{t},\qquad u\longmapsto
(u^2+\tau,u,u^2+2\tau),\quad\mbox{where}\quad t=\tau^2\in\ensuremath{\mathbb{C}}\,.$$
The limiting entire curve $\phi_0(\ensuremath{\mathbb{C}})\subseteq X_0$ passes through
the pinch point ${\bf 0}\in X_0$ and is not contained in the
singular locus $\{x=y=0\}= \operatorname{{\rm br}}( X_0)\cup\{\bf 0\}$ of $X_0$.
\end{rem}
\begin{cor}\label{mcor} In the same setting as before, consider the
normalization $\nu:\bar X_0\to X_0$. Suppose that $\overline{ \operatorname{{\rm br}}
(X_0)}$ is hyperbolic and there is a morphism $\pi:\bar X_0 \to E$
onto a hyperbolic variety $E$ such that for every $x\in E$
\begin{eqnarray}\label{vr} \pi^{-1}(x)\setminus \nu^{-1}({ \operatorname{{\rm br}} (X_0)}\setminus
(X_\infty\cap \operatorname{{\rm br}} (X_0))\end{eqnarray} is hyperbolic. Then any hypersurface
$X_t\neq X_0$ for $t$ close enough to $0$ is hyperbolic.
Consequently, a very general hypersurface of degree $d$ in ${\mathbb P}^n$
is algebraically hyperbolic.
\end{cor}
\begin{proof} We keep the notation introduced before. Suppose that for
$t\in {\mathbb P}^ 1$ general, $X_t$ is not hyperbolic. Let $\phi_0:\ensuremath{\mathbb{C}}\to
X_0$ be a (non--constant) limiting Brody curve. Since its image
cannot be contained in $\overline{ \operatorname{{\rm br}} (X_0)}$, there is a pullback
$\tilde\phi_0:\ensuremath{\mathbb{C}}\to \bar X_0$. Since $E$ is hyperbolic, the
composition $\pi\circ\tilde\phi_0:\ensuremath{\mathbb{C}}\to E$ is constant. Hence
$\tilde\phi_0(\ensuremath{\mathbb{C}})$ is contained in a fibre $\pi^{-1}(x)$ over a
point $x\in E$. Furthermore, it does not meet $\nu^{-1}({ \operatorname{{\rm br}}
(X_0)}\setminus (X_\infty\cap \operatorname{{\rm br}} (X_0))$. Indeed, otherwise
$\phi_0(\ensuremath{\mathbb{C}})$ would meet $ \operatorname{{\rm br}} (X_0)\setminus (X_\infty\cap \operatorname{{\rm br}}
(X_0))$ and, by Hurwitz' Theorem, it would be contained in
$\overline { \operatorname{{\rm br}} (X_0)}$, which is impossible. Then
$\tilde\phi_0(\ensuremath{\mathbb{C}})$ lies in \eqref{vr}, a contradiction.
\end{proof}
\subsection{Applying scrolls to Kobayashi hyperbolicity}\label{ssec:d=67}
\begin{prop}\label{200} We keep the notation as in Subsection
\ref{ssec:general}. Let $\Sigma\subseteq{\mathbb P}^n$ be a hypersurface
scroll with ordinary singularities satisfying conditions
(C1)-(C5). Suppose that: \begin{enumerate}\item[(i)] the base $E$ of $\Sigma$
and its double locus $\Delta_\Sigma$ are both hyperbolic;
\item[(ii)] for a general hypersurface $X$ in ${\mathbb P}^n$ of degree $d=\mathop{\rm deg}(\Sigma)$,
every ruling $F$ of $\Sigma$ meets $ \operatorname{{\rm br}}(\Sigma)$ in at least three
distinct points off $X\cap F$. \end{enumerate} Then a general hypersurface
of degree $d$ in ${\mathbb P}^n$ is hyperbolic. \end{prop}
\begin{proof} The assertion follows by applying
Corollary \ref{mcor} with $X_\infty=X$, $X_0=\Sigma$,
and $\overline{ \operatorname{{\rm br}}(X_0)}=\Delta_\Sigma$.
\end{proof}
Consider a general sextic scroll of genus 2 as introduced in
Example \ref {sextic} and a general septic scroll, also of genus
2, with ordinary singularities in ${\mathbb P}^ 3$. The latter scroll
exists according to Theorem \ref {thm:scrolls} and Remark \ref
{rem:3space}, (ii).
\begin{lem}\label{100} For a general scroll $\Sigma\subseteq {\mathbb P}^ 3$ of
genus 2 and degree either $d=6$ or $d=7$, the following hold:
\begin{itemize}
\item [(i)]~the projection $\pi:\Delta_S\to E$ has only
simple ramifications; in particular $\Delta_S$ meets
every ruling in at least three distinct points;
\item [(ii)]~ no pair of pinch points on $S$ sit on the same ruling;
\item [(iii)]~ the rulings passing through the pinch points on $S$ are not
tangent to $\Delta_S$.
\end{itemize}
\end{lem}
\begin{proof} We first treat the case $d=6$.
The conditions (i)--(iii) are open in $\mathcal H=\overline
{\mathcal H_{6,2}}$. So it suffices to show that there is a
surface in $\mathcal H$ satisfying these conditions. The reducible
surface $\Sigma_0$ in Example \ref {sextic} could be used for
this, once we know that the analogues of (i)--(iii) hold for a
general elliptic quintic scroll. This is in fact the case, but we
do not dwell on this here. We use instead a different degeneration
of a general sextic scroll of genus 2. We keep the notation
introduced in Example \ref {sextic}.
A smooth quadric $\tilde Q$ in ${\mathbb P}^4$ can be viewed as a
hyperplane section of the Grassmanian $\rm Gr(1,3)$ under the
Pl\"ucker embedding of $\rm Gr(1,3)$ in ${\mathbb P}^5$. There exists a curve
$E_0$ of degree $6$ and arithmetic genus $2$ on $\tilde Q$, which
consists of three conics $\Gamma_0, \Gamma_1, \Gamma_2$ such that
$\Gamma_1, \Gamma_2$ are disjoint and intersect both $\Gamma_0$
transversally at two points. Indeed, it is enough to take two
general hyperplanes $H_1,\, H_2$ in ${\mathbb P}^4$ meeting in a plane
$L_0$ and two other general planes $L_i\subseteq H_i$, $i=1,2$,
and let $\Gamma_i=L_i\cap\tilde Q$, $i=0,1,2$.
The surface $\Sigma_0\subseteq{\mathbb P}^3$, which corresponds to the
curve $E_0$, is the union of the three quadrics $Q_0, Q_1, Q_2$,
corresponding to $\Gamma_0, \Gamma_1, \Gamma_2$, respectively. We
may suppose that these quadrics are smooth. The surface $\Sigma_0$
belongs to $\mathcal H$.
One has $Q_0\cap Q_i= F_{ij}\cup G_{ij}$, $i,j=1,2$, where the
lines $F_{ij}$'s correspond to the intersection points of
$\Gamma_0$ with $\Gamma_i$ and belong to the same rulings of $Q_0$
and $Q_i$, and $G_{ij}$ are lines of the other rulings of $Q_0$
and $Q_i$. Furthermore $Q_1\cap Q_2=\varrho$ is a smooth
quartic curve of genus $1$. By taking $Q_0, Q_1, Q_2$
sufficiently general, we may suppose that the lines $F_{ij},
G_{ij}$ are general in their rulings and $\varrho$ is also
general. We denote by $p_{ij;hk}$ the intersection of $F_{ij}$
with $G_{hk}$, where $i,j,h,k\in\{1,2\}$. We note that $\varrho$
meets $Q_0$ at the eight points $p_{ij;3-i,h}$, with $i,j,
h\in\{1,2\}$.
Regard now $\Sigma_0$ as a limit of a general sextic scroll
$\Sigma$ of genus 2. The points of $\Gamma_0\cap\Gamma_i$,
$i=1,2$, are smoothed when deforming $E_0$ to $E$, hence also the
lines $F_{ij}$ are. Therefore the limit of the smooth double curve
$C=\Delta_\Sigma$ is the curve
\[ C_0=\Delta_{\Sigma_0}=\varrho \cup \bigcup _{i,j=1,2} G_{ij} \]
of degree 8 and arithmetic genus 5. The limit on $\Sigma_0$ of the
ruling on $\Sigma$ is the union of the rulings of $Q_0, Q_1, Q_2$
containing the lines $F_{ij}$, $1\le i\le j$. By (\ref{f400})
there are 16 pinch points on $\Sigma$. Similarly as in Example
\ref {quartic}, each of the eight points $p_{ij;ih}$, $i,j,h=1,2$
(not lying on $\varrho$) is the limit of two pinch points of
$\Sigma$. We call them \emph{limit pinch points}.
The smooth normalization $S$ of $\Sigma$ specializes to
a partial normalization
$S_0$ of $\Sigma_0$, ruled over the same nodal base curve $E_0$.
The singular surface $S_0$ consists of three irreducible quadric surfaces
$\tilde Q_0, \tilde Q_1,\tilde Q_2$ glued together along
the common rulings $\tilde F_{ij}$
in the same way as before.
The limit $\tilde C_0=\Delta_{S_0}$ of $\tilde C=\Delta_S$
is a nodal curve of arithmetic genus $17$. It maps to $E_0$
with degree $4$, and
consists of ten components:
\begin{itemize}
\item two copies $\varrho_i\subseteq \tilde Q_i$ of $\varrho$, each is
mapped with degree $2$ to $\Gamma_i$, $i=1,2$;
\item two copies $G_{i;hk}\subseteq \tilde Q_i$ of $G_{hk}$,
with $h,k=1,2$ and $i\in \{0,h\}$, eight curves in total. The
curves $G_{0;hk}$ and $G_{h;hk}$ are glued at two points $p_{h1;hk}$ and
$p_{h2;hk}$. Each of them is also glued to $\varrho_h$ at two
points, $h=1,2$. Hence the curves $\varrho_1$ and $\varrho_2$ meet in
the eight points $p_{ij;3-i,h}$, with $i,j,h=1,2$.
The four disjoint curves $G_{0;hk}$ on $\tilde Q_0$
are all mapped isomorphically
to $\Gamma_0$, whereas for $h=1,2$ the two disjoint curves
$G_{h;hk}$ on $\tilde Q_h$ ($k=1,2$) are mapped
isomorphically to $\Gamma_h$.
\end{itemize}
Therefore, the limit of the $24$ ramification points of the projection
$\pi: \tilde C\to E$ are:
\begin{itemize}
\item [(a)] the ramification points of the degree $2$ covers
$\varrho_i\to \Gamma_i$, $i=1,2$, in total $8$ such points;
\item [(b)] the connecting nodes of $\varrho_h$ with $G_{h;hk}$,
$k,h=1,2$, in total $8$ distinct such points, each counted with
multiplicity two.
\end{itemize}
We call these the \emph{limit ramification points}.
Part (i) follows from this description, our generality assumption,
and the observation that
every limit ramification point of type (b) smooths to two
ramification points on $\tilde C$ lying on different rulings.
As for (ii), the ruling $F_{ij}$ through $p_{ij;ih}$ misses all
limit pinch points other than $p_{ij;i,3-h}$. Consider a partial
deformation of $\Sigma_0$ to the union of a general elliptic
quartic scroll $\Sigma_0'$ and a quadric $Q'_1$ containing two
general rulings. This corresponds to a partial smoothing of $E_0$
to the union of an elliptic quartic curve $E'$, obtained by
smoothing $\Gamma_0+\Gamma_2$, plus a conic $\Gamma'_1$
(specializing to $\Gamma_1$) meeting $E'$ transversally at two
points. In this way $\Sigma'_0$ has two double lines $R_1,R_2$
which respectively specialize to $G_{21}$ and $G_{22}$. For a
fixed index $i\in \{0,1\}$, the two limit pinch points
$p_{2j;2i}$, $j=1,2$, deform to four pinch points of $\Sigma'_0$
on $R_i$, and, as we saw in Example \ref {quartic}, they are
general points on $R_1,R_2$ and are never pairwise on a ruling.
For the proof of (iii) note that, by generality assumptions, the
rulings through the limit ramification points of type (a) do not
contain any of the limit pinch points. In contrast, the rulings
through limit ramification points of type (b) do contain limit
pinch points. However, the same proof as for (ii) and generality
assumptions imply that, in a general deformation of $\Sigma_0$ to
$\Sigma$, this is no longer the case.
The case $d=7$ is similar, hence we will be as brief as possible.
The closure of $\mathcal H_{7,2}$ contains points corresponding to
a surface $\Sigma_0=\Sigma'\cup P$, where $\Sigma'_0$ is a general
sextic scroll of genus 2 and $P$ is a general plane containing a
general ruling $F$. The intersection of $P$ with $\Sigma'$
consists of $F$ plus a plane quintic curve $D$ of
genus $2$, which has four nodes $p_i$, $i=1,\ldots,4$. The
intersection of $C'=\Delta_ {\Sigma'}$ with $P$ consists of the
points $p_i$, $i=1,\ldots, 4$, and four more points $q_i\in F$,
$i=1,\ldots, 4$. The intersection of $D$ with $F$ consists
of the points $q_i$, $i=1,\ldots, 4$, and of a further point $q$
which is smooth on $\Sigma'$, so that $P$ is tangent to $\Sigma'$
at $q$.
The surface $\Sigma_0$ is the limit of a general scroll $\Sigma$
of degree $7$ and genus $2$. If $E$ is the base of $\Sigma$
regarded as a curve in $\rm Gr(1,3)$, this corresponds to $E$
degenerating to $E_0$, which is the union of a general sextic $E'$
of genus $2$ and a line $L$ meeting $E'$ transversally at
one points $f$, which corresponds to $F$. The limit of the ruling
of $\Sigma$ is the ruling of $\Sigma'$ plus the pencil in $P$,
corresponding to $L$, with center a general point of $F$.
The limit of $C=\Delta_\Sigma$ is the curve $C_0=C'\cup D$
of degree $13$. The points $p_i$, $i=1,\ldots, 4$, are limits of
the four triple points of $C$.
The geometric genus of a partial smoothing of $C_0$ at the points $q_i$,
$i=1,\ldots, 4$, is $10$.
All this agrees with \eqref {f200}, \eqref {f300},
and \eqref {f400}.
The usual analysis shows that the limit of the $18$ pinch points
of $\Sigma$ (see \eqref {f500}) are the $16$ pinch points of
$\Sigma'$ plus the point $q$ counted with multiplicity $2$.
The limit $\tilde C_0$ of $\tilde C=\Delta_S$ maps with degree five
to the curve $E_0=E'\cup L$. It consists of:
\begin{itemize}
\item a copy $\tilde C'$ of $\Delta_{S'}$ which maps to $E'$ with
degree four;
\item a copy of the normalization $\bar D$ of $D$, which
maps isomorphically to $E'$ and meets $\tilde C'$
transversally at four points;
\item a copy of $D$
which maps to $L$ with multiplicity five
via the projection induced by the ruling on $P$, and meets
$\tilde C'$ transversally at four points.
\end{itemize}
Hence the limit of the $44$ ramification points of the projection
$\pi: \tilde C\to E$ are
\begin{itemize}
\item the $24$ ramification points of the map $\tilde C' \to E'$;
\item the $12$ ramification points of the map $D \to L$;
\item the $4$ connecting nodes of $\tilde C'$ with $\bar D$,
each counted with multiplicity two.
\end{itemize}
With this in mind the proof proceeds similarly to the case $d=6$.
The details can be left to the reader. \end{proof}
\begin{thm}\label{mthm} For every $d\ge 6$ there exists a hyperbolic
surface in ${\mathbb P}^3$ of degree $d$. Consequently, a very general
surface in ${\mathbb P}^3$ of degree $d\ge 6$ is algebraically hyperbolic.
\end{thm}
\begin{proof} For $d=6,7$ this follows from Corollary \ref{mcor} and
Lemma \ref{100}. For $d\ge 8$ one can consider e.g. a general
deformation of the union of two general cones in ${\mathbb P}^3$ of
degrees $d_1,d_2$, where $d_1+d_2=d$ and $d_i\ge 4$ (see \cite {SZ2}).
\end{proof}
\begin{rem}\label{sz} Consider the union $X_0=X_1\cup X_2$ of projective
cones with distinct vertices in ${\mathbb P}^4$ over two smooth hyperbolic
surfaces in ${\mathbb P}^3$. According to \cite{SZ2}, $X_0$ can be
deformed to a smooth hyperbolic threefold in ${\mathbb P}^4$ of degree
$\mathop{\rm deg} (X_1)+\mathop{\rm deg} (X_2)$. Thus there exist hyperbolic threefolds in
${\mathbb P}^4$ of any given degree $d\ge 12$. Consequently, a very
general threefold in ${\mathbb P}^4$ of degree $d\ge 12$ is algebraically
hyperbolic. \end{rem}
\providecommand{\bysame}{\leavevmode\hboxto3em{\hrulefill}\thinspace}
|
\section{Introduction}
Scalar field models with derivative self interactions have attracted attention in various contexts.
For instance, models of $k$-essence \cite{ArmendarizPicon:2000ah} provide an interesting framework in which to investigate important early \cite{ArmendarizPicon:1999rj} as well as late time \cite{ArmendarizPicon:2000dh} issues of modern cosmology, while models similar to $k$-essence have been proposed in the context of relativistic MOND \cite{Bekenstein:1984tv}. All these models have the characteristic feature that their action depends solely on a scalar field $\pi$ and its first derivative --- clearly then, whatever the Lagrangian, the field equations stay second order.
More recently, scalar models with actions depending on second derivatives of the fields have been considered, mainly inspired by the decoupling limit of the
Dvali-Gabadadze-Porrati (DGP) model and its cosmology \cite{Dvali:2000hr,COSMODGP} as well as the resulting modification of the gravitational interaction via the so-called Vainshtein mechanism \cite{Vainshtein:1972sx,Deffayet:2001uk}. Such models range from the ``Galileon'' \cite{Nicolis:2008in}, to ``$k$-Mouflage" \cite{Babichev:2009ee} or ``kinetic gravity braided scalars" \cite{Deffayet:2010qz,Kobayashi:2010cm} and have different defining properties. A feature shared by the former and latter class of models, as well as some $k$-Mouflage models, is that they have an action which depends on second order derivatives of the fields. Hence, it is not {\it a priori} obvious how the field equations can stay second order, a property necessary in order to avoid propagating ghosts or extra degrees of freedom. This, however, can be achieved. The Galileon \cite{Nicolis:2008in}, for example, can be defined as the most general scalar theory which, in flat space-time, has field equations which are uniquely second order in derivatives. We note that in fact, Galileons were introduced rather earlier than \cite{Nicolis:2008in}, by Fairlie {\it et al.~} \cite{Fairlie} though in a different context. As we will also outline in section \ref{EULERHIER}, there the relevant Lagrangians were constructed through the successive application (called ``Euler hierachies") of the Euler-Lagrange operator to an arbitrary initial Lagrangian depending solely on the first derivative of a scalar field (with also the possibility of introducing arbitrary functions of the field first derivatives at intermediate steps).
The curved space-time generalizations of those models are also interesting.
As shown in Ref.~\cite{US}
the simplest covariantization of the original, four dimensional, Galileons led to field equations for the
scalar and its stress-tensor that contained third
derivatives. However, \cite{US} also showed how to eliminate
these higher derivatives by introducing suitable non-minimal,
curvature, couplings. Single scalar Galileons and their non minimal covariantization were further generalized to the multi-field (and $p$-forms) case, as well as to arbitrary dimensions, in Refs.~\cite{Deffayet:2009mn,Deffayet:2010zh,Padilla:2010de,Hinterbichler:2010xn}. More recently, \cite{deRham:2010eu,VanAcoleyen:2011mj} showed how to obtain the Galileons and covariant Galileons from models with extra-dimensions.
Finally, Ref.~\cite{Deffayet:2010qz} pointed out that a family of models which have Lagrangians depending linearly on second derivatives of the fields, but also have second order field equations, have interesting properties when considered on curved space-time, due to an essential mixing between the scalar and the metric dubbed in \cite{Deffayet:2010qz}, kinetic gravity braiding.
The DGP model also generated new interest in massive gravity and its Vainshtein mechanism. This mechanism was first discussed \cite{Vainshtein:1972sx} in a simple non linear extension of the free theory for a massive graviton, the so-called Pauli-Fierz theory, as a way of getting rid of the bothering effects of the scalar polarization of the massive graviton, present for any non-vanishing graviton mass --- what is known as the van Dam-Veltman-Zakharov discontinuity \cite{vanDam:1970vg}. The Vainshtein mechanism, which was recently shown to work even in the simplest theories of massive gravity \cite{WORKVAIN}, can be attributed to the self-interactions of the scalar polarization of the graviton \cite{Deffayet:2001uk,ArkaniHamed:2002sp,Creminelli:2005qk,Deffayet:2005ys}. The latter interactions, which can be studied by taking an appropriate ``decoupling limit" \cite{ArkaniHamed:2002sp,Luty:2003vm}, take the form of derivative scalar self-couplings and the same is in fact true in the DGP model.
For the Vainshtein mechanism to operate, however, there is no need to have field equations that are purely second order. This was shown explicitly in Ref.~\cite{Babichev:2009ee} in particular, which introduced a large family of scalar tensor models called $k$-Mouflage, and which used the Vainshtein mechanism to screen the effect of a scalar field at small distances. Note also that the Vainshtein mechanism also serves as one of the basis of the recently introduced ``classicalization" \cite{CLASSIC}. Finally, we stress that a recent attempt to obtain a massive gravity devoided of the unwanted Boulware Deser ghost \cite{Boulware:1973my} has a decoupling limit sharing crucial properties with some of the above mentioned scalar field models \cite{DRG}.
Hence, it is clear that scalar models which have derivative self interactions, possibly depending on second order derivatives,
have numerous interesting properties. However, to our knowledge, these theories have so far not been extensively classified nor even constructed. It is the purpose of this work to do so. Namely, here, we will construct all theories of a scalar field $\pi$ in $D$ dimension and on flat space-time, which have actions depending on first and twice differentiated $\pi$'s as well as on undifferentiated $\pi$'s (hence without assuming necessarily a shift symmetry) but have field equations which stay of order two and lower. This will be carried out in section
\ref{UNICSECTION} where our main result is first stated and summarized (subsection \ref{SEC:RESULT}) before being proven.
We then show how to non minimally complete those theories in curved space-time, maintaining second order field equations of the scalar as well as for the metric (section \ref{COVAR}). Some examples are then discussed in relation with the Euler hierarchies construction, and we also illustrate our results giving the covariantization of the conformal Galileons (section \ref{EXAMPLES}). An introductory section \ref{GALIL} gathers some useful results.
\section{Galileon, kinetic gravity braiding and some useful results and notations}
\label{GALIL}
In this section we work in flat space only, and introduce the models studied in the remainder of this paper as well as some useful notation and results. We also revisit the Galileon model studied in \cite{Nicolis:2008in}. Throughout we work in $D$ space-time dimensions, with signature $(-,+,+,\ldots)$.
\subsection{Introduction and two useful lemma}
\label{LEM1}
All models we consider depend on a single scalar field $\pi$ whose partial derivatives will be denoted by\footnote{Note that, when considered on curved space-time, $\pi_\mu$ will denote the covariant derivative acting on $\pi$, $\nabla_\mu \pi$, and so on for $\pi_{\mu \nu} ...$, i.e. partial derivatives are just to be replaced by covariant derivatives in the notation (\ref{defpimunu}).}
\begin{equation} \label{defpimunu}
\pi_\mu \equiv \partial_\mu \pi \, , \qquad \pi_{\mu \nu} \equiv \partial_\mu \partial_\nu \pi \, , \qquad
\pi_{\mu \nu \alpha } \equiv \partial_\mu \partial_\nu \partial_\alpha \pi \, \qquad {\rm etc}.
\end{equation}
Since derivatives commute on flat space-time, these tensors are symmetric under interchange of any indices.
The Lagrangians considered take the form
\begin{eqnarray} \label{L0}
\mathcal{L} = \mathcal{T}_{(2n)}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n}\nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} } \pi_{\mu_{\vphantom{()}1} \nu_{\vphantom{()}1}} \ldots \pi_{\mu_{\vphantom{()}n} \nu_{\vphantom{()}n}} \ ,
\end{eqnarray}
where $\mathcal{T}_{(2n)}$ is a $2n$-contravariant tensor function of $\pi$ and $\pi_\alpha$ only
\begin{equation}
\mathcal{T}_{(2n)} = \mathcal{T}_{(2n)}(\pi,\pi_\alpha) \, .
\label{propT}
\end{equation}
Note that the integer $n$ also denotes the number of twice differentiated $\pi$'s appearing in the Lagrangian (\ref{L0}).
Thus $\mathcal{L} = \mathcal{L}(\pi,\pi_\alpha,\pi_{\alpha \beta})$ and the corresponding field equations are ${\cal E}=0$, where
\begin{eqnarray}
{\mathcal{\cal E}} &=&
\left[ \frac{\partial}{\partial \pi} - \partial_\mu \left(\frac{\partial}{\partial \pi_\mu} \right) + \partial_\mu \partial_\nu \left(\frac{\partial }{\partial \pi_{\mu \nu}} \right) \right] {\cal L}
\label{EL}
\\
&\equiv&\hat{\mathcal{\cal E}} {\cal L} \, .
\label{ELHAT}
\end{eqnarray}
For future use, we begin by giving the {\it sufficient} conditions such that these equations are of order 2 or lower (in derivatives) on flat space-time. To do so, first note the following lemma:
\newline
\noindent {\bf Lemma. }{\it The field equation derived from the Lagrangian $\mathcal{L}= {\cal T}_{(2)}^{\mu\nu}\pi_{\mu\nu}$
does not contain any derivative of order higher than 2. }\newline
This is straightforward to verify. The second term in (\ref{EL}) yields one contribution in third derivatives of $\pi$, namely $\pi_{\alpha \beta \mu} \partial \mathcal{T}^{\alpha \beta}/\partial \pi_\mu$. An identical contribution arises from the last term in (\ref{EL}). However, given the relative sign in (\ref{EL}), these terms in third derivatives cancel.
This simple result is in fact at the basis of the model of kinetically braided scalar of Ref. \cite{Deffayet:2010qz}, where $\mathcal{T}^{\mu \nu}_{(2)} = f(\pi,\pi_\alpha) g^{\mu \nu}$ and $g_{\mu \nu}$ is the metric. In fact the above lemma generalizes to curved space-time as shown in appendix \ref{APPA} and used in section \ref{COVAR}.
Now consider any $n > 1$. The second term in (\ref{EL}) again yields contributions to the equation of motion which are third order in derivatives. Those are cancelled by terms coming from the final term in Eq.~(\ref{EL}) by virtue of the above lemma. However, the final term in Eq.~(\ref{EL}) also yields ``dangerous terms'' (by which we mean terms of order three or more in derivatives) which are of the form $\pi_{\mu_{\vphantom{()}k} \mu_{\vphantom{()}l} \nu_{\vphantom{()}l}}$, $\pi_{\nu_{\vphantom{()}k} \mu_{\vphantom{()}l} \nu_{\vphantom{()}l}}$ or $\pi_{\nu_{\vphantom{()}k} \mu_{\vphantom{()}k} \mu_{\vphantom{()}l} \nu_{\vphantom{()}l}}$ where all the indices are contracted with those of $\mathcal{T}_{(2n)}$. Since derivatives commute on flat space-times, we immediately have the following result:
\newline
\noindent {\bf Main lemma.} {\it A sufficient condition for the field equations derived from the Lagrangian (\ref{L0}) to stay of order less or equal to 2 is that the tensor $\mathcal{T}_{(2n)}^{\mu_{\vphantom{()}1} \mu_{\vphantom{()}2}
\ldots \mu_{\vphantom{()}n}\nu_{\vphantom{()}1} \nu_{\vphantom{()}2}
\ldots \nu_{\vphantom{()}n}}$ is totally antisymmetric in its first $n$ indices $\mu_{\vphantom{()}i}$ as well as (separately) in its last $n$ indices $\nu_{\vphantom{()}i}$.}\newline
The main purpose of this paper is to study the converse of this simple result. Before doing so, and for future use, we first revisit the Galileon theory. As we will see, this provides a simple and fundamental example of the type of theory we will discuss.
\subsection{The flat space time Galileon revisited}
The starting point of Galileon models \cite{Nicolis:2008in,US}
are Lagrangians
${\cal L}^{\rm{Gal}}$
of the form (\ref{L0}) with the tensor ${\cal T}_{(2n)}$ satisfying (\ref{propT}) as well as the properties of the main lemma above. Furthermore, in flat space-time, they have equations of motion of order {\it strictly equal} to 2 (that is, they do not contain undifferentiated or once differentiated $\pi$, but only twice differentiated $\pi$).
As we now outline, there exist several possible ways of writing the Galileon action:
each differs from the other by a different choice of tensor $\mathcal{T}_{(2n)}$ and a total derivative. However they lead to the same equations of motion and are hence equivalent.
To begin with, define the $2m$-contravariant tensor ${\mathcal A}_{(2m)}$ by
\begin{equation} \label{DEFAten}
\mathcal{A}_{(2m)}^{\mu_{\vphantom{()}1} \mu_{\vphantom{()}2}
\ldots \mu_{\vphantom{()}m} \nu_{\vphantom{()}1} \nu_{\vphantom{()}2}
\ldots \nu_{\vphantom{()}m}} \equiv
\frac{1}{(D-m)!}\,
\varepsilon^{\mu_{\vphantom{()}1}
\mu_{\vphantom{()}2} \ldots
\mu_{\vphantom{()}m} \sigma_{\vphantom{()}1}\sigma_{\vphantom{()}2}\ldots
\sigma_{\vphantom{()}D-m}}_{\vphantom{\mu_{\vphantom{()}1}}}
\,\varepsilon^{\nu_{\vphantom{()}1} \nu_{\vphantom{()}2} \ldots
\nu_{\vphantom{()}m}}_{\hphantom{\nu_{\vphantom{()}1}
\nu_{\vphantom{()}2} \ldots
\nu_{\vphantom{()}2m}}\sigma_{\vphantom{()}1}
\sigma_{\vphantom{()}2}\ldots \sigma_{\vphantom{()}D-m}}
\end{equation}
where the totally antisymmetric Levi-Civita tensor is given by
\begin{equation} \label{DEFLC}
\varepsilon^{\mu_{\vphantom{()}1} \mu_{\vphantom{()}2} \ldots
\mu_{\vphantom{()}D}} = - \frac{1}{\sqrt{-g}}
\delta^{[\mu_{\vphantom{()}1}}_1 \delta^{\mu_{\vphantom{()}2}}_2
\ldots \delta^{\mu_{\vphantom{()}D}]}_D \,
\end{equation}
with square brackets denoting unnormalized permutations. (The definitions (\ref{DEFAten}) and (\ref{DEFLC}) are also valid in arbitrary curved space-times with metric $g_{\mu \nu}$ and $D\geq m$.)
Thus $\mathcal{A}_{(2m)}$ is antisymmetric
in its first $m$ indices as well as, separately, in its last $m$ indices. Other useful properties of $\mathcal{A}_{(2m)}$ are given in Appendix \ref{PROPA}.
A first possible Lagrangian for the Galileon
is given by \cite{Deffayet:2009mn}
\begin{eqnarray} \label{LGAL1}
{\cal L}^{\rm{Gal},1}_{N} &=&\left( {\mathcal A}_{(2n+2)}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n+1}\nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n+1} } \pi_{\mu_{n+1}} \pi_{\nu_{n+1}} \right) \pi_{\mu_{\vphantom{()}1} \nu_{\vphantom{()}1}} \ldots \pi_{\mu_{\vphantom{()}n} \nu_{\vphantom{()}n}}
\nonumber
\\
&\equiv & \mathcal{T}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n} \nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} }_{(2n),\rm{Gal},1} \pi_{\mu_{\vphantom{()}1} \nu_{\vphantom{()}1}} \ldots \pi_{\mu_{\vphantom{()}n} \nu_{\vphantom{()}n}},
\end{eqnarray}
with
\begin{eqnarray}
\mathcal{T}^{\mu_{\vphantom{()}1}\label{TENSGAL1}
\ldots \mu_{\vphantom{()}n} \nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} }_{(2n),\rm{Gal},1} &\equiv & {\mathcal A}_{(2n+2)}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n+1}\nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n+1} } \pi_{\mu_{n+1}} \pi_{\nu_{n+1}} \, .
\end{eqnarray}
Here and henceforth $N$ indicates the number of $\pi$'s appearing in the Lagrangian of a given Galileon model
so that
\begin{equation}
N= n+2.
\nonumber
\end{equation}
As discussed in \cite{Deffayet:2009mn}, the Lagrangian ${\cal L}^{\rm{Gal},1}_{N}$ also reads:
\begin{eqnarray} \label{nico1}
{\cal L}^{\rm{Gal},1}_{N}&=& - \sum_{\sigma \in S_{n+1}} \epsilon(\sigma)
\bigl[ \pi_{\vphantom{\mu_1}}^{\mu_{\sigma(1)}} \pi_{\mu_1}\bigr]
\bigl[\pi_{\hphantom{\mu_{\sigma(2)}} \mu_2}^{\mu_{\sigma(2)}}
\pi_{\hphantom{\mu_{\sigma(3)}} \mu_3}^{\mu_{\sigma(3)}} \ldots
\pi_{\hphantom{\mu_{\sigma(n+1)}} \mu_{n+1}}^{\mu_{\sigma(n+1)}}\bigr], \nonumber \\&=& \label{nico1bis}
-\sum_{\sigma \in S_{n+1}} \epsilon(\sigma)
g^{\mu_{\sigma(1)}\nu_{\vphantom{()}1}}
g^{\mu_{\sigma(2)}\nu_{\vphantom{()}2}} \ldots
g^{\mu_{\sigma(n+1)}\nu_{\vphantom{()}n+1}}
(\pi_{\nu_1} \pi_{\mu_1})
(\pi_{\nu_2 \mu_2} \pi_{\nu_3 \mu_3}\ldots \pi_{\nu_{n+1} \mu_{n+1}}),
\end{eqnarray}
where $\sigma$ denotes a permutation of signature $\epsilon(\sigma)$ of the permutation group $S_{n+1}$, and in order for the Lagrangian to be non-vanishing,
\begin{equation}
n+1 \leq D \qquad \Longleftrightarrow \qquad N \leq D+1\,.
\label{MAXX}
\end{equation}
This is the original form presented in \cite{Nicolis:2008in}, and the equality of (\ref{LGAL1}) and (\ref{nico1bis}) can be seen \cite{Deffayet:2009mn} using the identity
\begin{equation}
\sum_{\sigma \in S_D} \epsilon(\sigma)
g^{\mu_{\sigma(1)}\nu_{\vphantom{()}1}}
g^{\mu_{\sigma(2)}\nu_{\vphantom{()}2}} \ldots
g^{\mu_{\sigma(D)}\nu_{\vphantom{()}D}} =
- \varepsilon^{\mu_{\vphantom{()}1} \mu_{\vphantom{()}2}
\ldots \mu_{\vphantom{()}D}}\,
\varepsilon^{\nu_{\vphantom{()}1} \nu_{\vphantom{()}2}
\ldots \nu_{\vphantom{()}D}}.
\end{equation}
Using (\ref{EL}), the field equations derived from the Lagrangian ${\cal L}^{{\rm Gal},1}_{N}$ read
\begin{equation}
\mathcal{\cal E} = -N \times {\cal E}_{N}=0 \, ,
\end{equation}
where
\begin{eqnarray}
{\cal E}_{N} &=&-\sum_{\sigma \in S_{n+1}} \epsilon(\sigma)
\prod_{i=1}^{n+1} \pi_{\hphantom{\mu_{\sigma(i)}}
\mu_i}^{\mu_{\sigma(i)}}, \nonumber
\\
&=& {\mathcal A}_{(2n+2)}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n+1}\nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n+1} }
\pi_{\mu_1\nu_1} \pi_{\mu_2\nu_2} \ldots \pi_{\nu_{n+1} \mu_{n+1}}.
\label{eofmG}
\end{eqnarray}
These are only second order, as advertised. Notice that the index $N$ on $\mathcal{\cal E}_N$ indicates that it is the equation of motion coming from ${\cal L}^{\rm{Gal},1}_{N}$ (which contains $N$ factors of $\pi$): thus $\mathcal{\cal E}_N$ contains $N-1$ factors of $\pi$.
The Galileon model with the largest number of fields in $D$ dimensions has $N=D+1$. In this case, ${\cal E}_{D+1}$ is simply proportional to the determinant of the Hessian, the matrix of second derivatives $\pi_{\mu \nu}$. As such, the equation ${\cal E}_{D+1} = 0 $ is known as the Monge-Amp\`ere equation, and it has various interesting properties, in particular in relation to integrability (see e.g.~\cite{Fairlie:1994in}). At the same time
\begin{equation}
{\cal L}^{{\rm Gal},1}_{D+1} \propto \det
\left(\begin{array}{cc}\pi_{\mu \nu} & \pi_{\nu} \\ \pi_{\mu} & 0\end{array}\right) \, ,
\nonumber
\end{equation}
which is the left-hand-side of the Bateman equation \cite{Fairlie:1994in,Bateman}.
Finally,
in $D$ dimensions, the total Galileon Lagrangian is given by a linear combination of Lagrangians ${\cal L}_{N}^{{\rm Gal},1}$ with $N=2, \cdots, D+1$.
In $D=4$ dimensions, these are simply the 4 terms given in \cite{Nicolis:2008in}.
A second possible Lagrangian for the Galileon with, again, $N = n+2$ fields is given by
\begin{eqnarray} \label{LGAL2}
{\cal L}^{\rm{Gal},2}_{N} &=&\left( {\mathcal A}_{(2n)}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n}\nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} } \pi_{\mu_1} \pi_{\lambda} \pi^{\lambda}_{\hphantom{\lambda} \nu_1} \right) \pi_{\mu_{\vphantom{()}2} \nu_{\vphantom{()}2}} \ldots \pi_{\mu_{\vphantom{()}n} \nu_{\vphantom{()}n}},\\
&\equiv & \mathcal{T}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n} \nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} }_{(2n),\rm{Gal},2} \pi_{\mu_{\vphantom{()}1} \nu_{\vphantom{()}1}} \ldots \pi_{\mu_{\vphantom{()}n} \nu_{\vphantom{()}n}},
\label{EXEQ21}
\end{eqnarray}
where (see also section \ref{ANTI})
\begin{eqnarray}
\mathcal{T}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n} \nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} }_{(2n),\rm{Gal},2} = \frac{1}{n}
\,
{\mathcal A}_{(2n)}^{\alpha_{\vphantom{()}1}
\ldots \alpha_{\vphantom{()}n}\nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} }
&& \Big[\left( \pi^{\mu_1} \pi_{\alpha_1}\right) \delta^{\mu_2}_{\; \; \alpha_2} \ldots \delta^{\mu_n}_{\; \; \alpha_n} \nonumber \\
&& + \delta^{\mu_1}_{\; \; \alpha_1} \left( \pi^{\mu_2} \pi_{\alpha_2} \right) \delta^{\mu_3}_{\; \; \alpha_3} \ldots \delta^{\mu_n}_{\; \; \alpha_n} \nonumber \\
&& + \ldots \nonumber \\
&& + \delta^{\mu_1}_{\; \; \alpha_1} \ldots \delta^{\mu_{n-1}}_{\; \; \alpha_{n-1}} \left( \pi^{\mu_n} \pi_{\alpha_n} \right) \Big].
\label{TT2}
\label{TGAL2}
\end{eqnarray}
Finally, the third form of interest is given by
\begin{eqnarray} \label{LGAL3}
{\cal L}^{{\rm Gal},3}_{N} &= & \left( {\mathcal A}_{(2n)}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n}\nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} } \pi_{\lambda} \pi^{\lambda}\right) \pi_{\mu_{\vphantom{()}1} \nu_{\vphantom{()}1}} \ldots \pi_{\mu_{\vphantom{()}n} \nu_{\vphantom{()}n}}
\end{eqnarray}
so that
\begin{eqnarray}
\mathcal{T}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n} \nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} }_{(2n),{\rm Gal},3} &=& X \mathcal{A}_{(2n)}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n} \nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} }
\end{eqnarray}
where\footnote{Note that in the context of $k$-inflation and other models, $X$ is often defined to be $-\pi_\mu \pi^\mu/2$. In order to simplify our equations, the factor of $-1/2$ is not included here.}
\begin{equation}
X\equiv \pi_\mu \pi^\mu \, .
\label{Xdef}
\end{equation}
The three Lagrangians (\ref{LGAL1}), (\ref{LGAL2}) and (\ref{LGAL3}) are in fact all equal up to a total derivative. Indeed, on defining $J^\mu_{N}$ by
\begin{eqnarray} \label{DEFJ}
J^\mu_{N}= X \mathcal{A}_{(2n)}^{\mu \mu_{2}\cdots\mu_{n}\nu_{1}\nu_{2}\cdots\nu_{n}}\pi_{\nu_{1}}\pi_{\mu_{2}\nu_{2}}\cdots\pi_{\mu_{n}\nu_{n}} \, ,
\end{eqnarray}
it follows directly that
\begin{eqnarray}
\mathcal{L}_{N}^{\rm{Gal},2}&=&-\frac{1}{2}\mathcal{L}_{N}^{\rm{Gal},3}+\frac{1}{2} \partial_\mu J^\mu_{N} \, .
\label{REL2}
\end{eqnarray}
Furthermore on using the properties of $\mathcal{A}_{(2n)}$ given in appendix \ref{PROPA}, it follows that
\begin{eqnarray}{\label{3Lid}}
(N-2)\mathcal{L}_{N}^{\rm{Gal},2}=\mathcal{L}_{N}^{\rm{Gal},3}-\mathcal{L}_{N}^{\rm{Gal},1} \, .
\end{eqnarray}
Thus we also have
\begin{eqnarray}
\mathcal{L}_{N}^{\rm{Gal},1}&=&\frac{N}{2}\mathcal{L}_{N}^{\rm{Gal},3}-\frac{N-2}{2} \partial_\mu J^\mu_{N}, \label{REL1}\\
\mathcal{L}_{N}^{\rm{Gal},1}&=&-N \mathcal{L}_{N}^{\rm{Gal},2}+\partial_\mu J^\mu_{N} \, .
\label{REL3}
\end{eqnarray}
From (\ref{REL2}), (\ref{REL1}) and (\ref{REL3}) it therefore follows that the equations of motion of all three Galileon Lagrangians are identical, given by (\ref{eofmG}), and strictly of second order.
Finally, observe from (\ref{eofmG}) and (\ref{LGAL3}) that ${\cal L}^{\rm{Gal},3}_{N}$ can be rewritten as
\begin{equation}
{\cal L}^{\rm{Gal},3}_{N} = X {\cal E}_{N-1}
\label{DEFGAL3}
\end{equation}
where ${\cal E}_{N-1}$ are the equations of motion coming from ${\cal L}^{\rm{Gal}}_{N-1}$ (where we drop the index $1,2,3$). In this form, it is manifest that Galileon models containing a given number $N$ of $\pi$ fields can be obtained from the field equations of the same models with one less field. This property, though implicit in \cite{Nicolis:2008in}, is very well explained by the hierachical construction of \cite{Fairlie} which preceeded by far Ref.~\cite{Nicolis:2008in} and discussed first, as far as we know, what are called here and elsewhere Galileons (see \cite{Fairlie:2011md}). This hierachical construction will be discussed in section \ref{EXAMPLES}.
\subsection{Galileon Lagrangians in terms of cycles}
The three equivalent Galileon Lagrangians presented above all satisfy the {sufficient} conditions of the main lemma of section \ref{LEM1}. In order to study the {necessary} conditions it will be useful to introduce a new notation consisting of the cycles $[i]$ and $\langle i \rangle$.
We define $[i]$ by
\begin{eqnarray} \label{DEFBIS[}
[i] &\equiv & \pi^{\mu_1}_{\hphantom{\mu_1} \mu_2} \pi^{\mu_2}_{\hphantom{\mu_2}{\mu_3}} \pi^{\mu_3}_{\hphantom{\mu_3}\mu_4} \cdots \pi^{\mu_{i}}_{\hphantom{\mu_{i}}\mu_1},
\end{eqnarray}
so that for example
\begin{equation}
[1]=\Box \pi \, , \qquad [2] = \pi^{\alpha}_{\hphantom{\alpha} \beta} \pi^{\beta}_{\hphantom{\beta}\alpha} \, .
\end{equation}
Similarly,
\begin{eqnarray} \label{DEFBIS<}
\left<i\right> &\equiv & \pi_{\mu_1} \pi^{\mu_1}_{\hphantom{\mu_1} \mu_2} \pi^{\mu_2}_{\hphantom{\mu_2}{\mu_3}} \pi^{\mu_3}_{\hphantom{\mu_3}\mu_4} \cdots \pi^{\mu_{i}}_{\hphantom{\mu_{i}}\mu_{i+1}}\pi^{\mu_{i+1}},
\end{eqnarray}
so that
\begin{equation}
\left<1 \right> = \pi_{\alpha} \pi^{\alpha}_{\hphantom{\alpha} \beta} \pi^{\beta}
\end{equation}
Note that $[i]$ contains $i$ factors of $\pi$ as well as $i$ twice-differentiated $\pi$'s, whereas $\left<i\right>$ contains $i+2$ factors of $\pi$, but again $i$ twice-differentiated $\pi$'s.
Using this notation, and in the case of $N = 4$ fields, the three Galileon Lagrangians can be written as
\begin{eqnarray}
{\cal L}_{N=4}^{{\rm Gal},1} &=& - \left(\Box \pi\right)^2 \left(\pi_{\mu}\,\pi^{\mu}\right)
+ 2 \left(\Box \pi\right)\left(\pi_{\mu}\,\pi^{\mu\nu}\,\pi_{\nu}\right) \,
\nonumber
\\
&& + \left(\pi_{\mu\nu}\,\pi^{\mu\nu}\right) \left(\pi_{\rho}\,\pi^{\rho}\right)
-2 \left(\pi_{\mu}\pi^{\mu\nu}\,\pi_{\nu\rho}\,\pi^{\rho}\right) \,
\nonumber
\\
&=& X \left( [2] - [1]^2 \right) + 2 \left( [1]\left<1 \right> - \left<2 \right> \right) \, ,
\nonumber
\\
{\cal L}_{N=4}^{{\rm Gal},2} &=& - \left(\Box \pi\right) \pi_\mu \pi^{\mu \nu} \pi_\nu + \pi_\mu \pi^{\mu \nu} \pi_{\nu \rho} \pi^\rho = \left<2\right> - [1] \left<1\right> \, ,
\nonumber
\\
{\cal L}_{N=4}^{{\rm Gal},3} &=& \left(\pi_\lambda \pi^\lambda\right) \left( \pi_{\mu \nu} \pi^{\mu \nu} - \left(\Box \pi\right)^2 \right)
= X \left( \left[2\right] - \left[1\right]^2\right) \, .
\nonumber
\end{eqnarray}
Furthermore it will be useful to define
\begin{eqnarray}
\left[
\begin{array}{cccc}
p_1 & p_2 & \cdots & p_r \\1 & 2 & \cdots & r
\end{array}\right] &=& [1]^{p_1} [2]^{p_2} \cdots [r]^{p_r} \label{DEF[},
\end{eqnarray}
as well as
\begin{eqnarray}
\left<
\begin{array}{cccc}
q_1 & q_2 & \cdots & q_s \\1 & 2 & \cdots & s
\end{array}\right> &=& \left<1\right>^{q_1} \left<2\right>^{q_2} \cdots \left<s\right>^{q_s}\label{DEF<},
\end{eqnarray}
where the $p_i$ and $q_i$ are positive (or vanishing) integers.\footnote{Note that the right hand sides of equations (\ref{DEF[}) and (\ref{DEF<}) are uniquely specified respectively by the ordered sets $(p_1,p_2,\cdots,p_r)$ as well as $(q_1,q_2,\cdots,q_s)$. Hence our notation on the left hand side of equations (\ref{DEF[}) and (\ref{DEF<}) is a bit redundant, but we feel this will ease the reading of some of the equations below.} Using this notation it follows from (\ref{TT2}) that ${\cal L}_{N}^{{\rm Gal},2}$ can be expressed as a sum of terms $\left[
\begin{array}{cccc}
p_1& p_2 & \cdots & p_r \\1 & 2 & \cdots & r
\end{array}\right]\left<
\begin{array}{ccc}
q_1 & \cdots & q_s \\1 & \cdots&s
\end{array}\right>$
with $\sum_{j=1}^{s} {q_j}=1$. (See above in case $N=4$.)
Thus, for example, the equations of motion coming ${\cal L}_{N=4}$ are
\begin{eqnarray}
0 = {\cal E}_{N=4}&=&
- \left(\Box \pi \right)^3
-2 \left(\pi_{\mu}^{\hphantom{\mu}\nu}\,\pi_{\nu}^{\hphantom{\nu}\rho}\,\pi_{\rho}^{\hphantom{\rho}\mu}\right)
+3 \left(\Box \pi\right) \left(\pi_{\mu\nu}\, \pi^{\mu\nu} \right)
\nonumber
\\
&=& - [1]^3 - 2 [3] + 3 [1][2]
\nonumber
\\
&=& - \left[\begin{array}{c}3 \\1 \end{array} \right] - 2 \left[\begin{array}{c}1 \\3 \end{array} \right] + 3 \left[\begin{array}{cc}1 &1\\1&2 \end{array} \right] \, .
\label{E4a}
\end{eqnarray}
More generally, for any $N$, ${\cal E}_{N}$ can be expressed as a specific linear combination of monomials of the form (\ref{DEF[}):
\begin{eqnarray} \label{COEFC}
{\cal E}_{N} = \sum {\cal C}^{N}_{p_1,\cdots,p_r} \left[
\begin{array}{cccc}
p_1 & p_2 & \cdots & p_r \\1 & 2 & \cdots & r
\end{array}\right],
\end{eqnarray}
where the sum runs over the $r-$uplets $(p_1,\cdots,p_r)$ verifying the constraint
\begin{eqnarray}
N= 1+\sum_{i=1}^{i=r} i \times p_i ,
\end{eqnarray}
and
${\cal C}^{N}_{p_1,\cdots,p_r}$ are real coefficients which do not depend on the dimension of space-time, but only depend on the number of field $N$, explicitly
\begin{eqnarray*}
\mathcal{C}_{p_{1},p_{2},\cdots p_{r}}^{N}
& = &
\left(-1\right)^{N+p_{1}+p_{2}+\cdots + p_{r}}\frac{\left(N-1\right)!}{\left(p_{1}!p_{2}!\cdots p_{r}!\right)1^{p_{1}}2^{p_{2}}\cdots r^{p_{r}}}.
\end{eqnarray*}
\section{Uniqueness theorem in flat space-time}
\label{UNICSECTION}\label{UNIK}
\subsection{The result}
\label{SEC:RESULT}
We are now in the position to study the converse of the main lemma of section \ref{LEM1}. However, before presenting the proof in section \ref{PROOF}, we first state our result.
Namely, in flat space-time, the most general theory satisfying the three conditions
\begin{itemize}
\item[(i)] its Lagrangian contains derivatives of order 2 or less of the scalar field $\pi$;
\item[(ii)] its Lagrangian is polynomial in the second derivatives of $\pi$;
\item[(iii)] the corresponding field equations are of order 2 or lower in derivatives
\end{itemize}
has a Lagrangian which is given by an arbitrary linear combination of the Lagrangians ${\cal L}_{n}\{f\}$ (each containing $n$ of twice differentiated $\pi$) of the form
\begin{eqnarray}
{\cal L}_{n}\{f\}&=& f(\pi,X) \times {\cal L}^{\rm{Gal},3}_{N = n + 2}, \nonumber \\
&=& f(\pi,X) \times \left( X {\mathcal A}_{(2 n)}^{\mu_{\vphantom{()}1}
\ldots \mu_{\vphantom{()}n}\nu_{\vphantom{()}1}
\ldots \nu_{\vphantom{()}n} } \pi_{\mu_{\vphantom{()}1} \nu_{\vphantom{()}1}} \ldots \pi_{\mu_{\vphantom{()}n} \nu_{\vphantom{()}n}}\right). \label{GENL}
\end{eqnarray}
Here $f(\pi,X)$ is an arbitrary scalar function of $\pi$ and $X$, generally different for each $n$, and the braces in ${\cal L}_{n}\{f\}$ denote that ${\cal L}_n$ is a functional of $f$. The equations of motion corresponding to each ${\cal L}_{n}\{f\}$ are\footnote{We use the notation $f_{X} \equiv f_{,X}$, $f_{\pi} \equiv f_{,\pi}$ and so on.}
\begin{eqnarray}
0 & = & 2\left(f+Xf_{X}\right)\mathcal{E}_{N}+4\left(2f_{X}+Xf_{XX}\right)\mathcal{L}_{N+1}^{\text{Gal},2}\nonumber \\
& & +X\left[2Xf_{X\pi}-\left(n-1\right)f_{\pi}\right]\mathcal{E}_{N-1} \nonumber \\
&& -n\left(4Xf_{X\pi}+4f_{\pi}\right)\mathcal{L}_{N}^{\text{Gal},2} - nXf_{\pi\pi} \mathcal{L}_{N-1}^{{\rm Gal},1}.\label{EOFM}\end{eqnarray}
where $N=n+2$. Notice the dependence of these equations on $\pi^{\mu \nu}$ as well as $\pi^\mu$ (when $f$ is non-constant).
\subsection{Proof of uniqueness}
\label{PROOF}
In flat space-time, the only scalar quantities which are polynomial in second
derivatives
of $\pi$ must be constructed from $[i]$ and $\left<i\right>$, defined respectively in Eqs.~(\ref{DEFBIS[}) and (\ref{DEFBIS<}). Recall that these both contain $i$ times a twice differentiated $\pi$.
Hence, the most general scalar theory obeying conditions (i) and (ii) has a Lagrangian which is a linear combination of monomials, each of the form ${\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}$ defined by
\begin{eqnarray} \label{MONO}
{\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s} = f(\pi,X) \times \left[
\begin{array}{cccc}
p_1 & p_2 & \cdots & p_r \\1 & 2 & \cdots & r
\end{array}\right]\left<
\begin{array}{cccc}
q_1 & q_2 & \cdots & q_s \\1 & 2 & \cdots & s
\end{array}\right>
\end{eqnarray}
where $f$ is an arbitrary scalar function of $\pi$ and $X$ (different for each monomial ${\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}$), and the other quantities have been defined in (\ref{DEF[}) and (\ref{DEF<}). Formally then
\begin{equation}
{\cal L} = \sum_{\{p_i\}\{q_j\}} C_{\{p_i\}\{q_j\}} {\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}
\label{SUMC}
\end{equation}
where the sum is over the set of $\{p_i\} =(p_1,\cdots, p_r)$ and $\{q_j\} =(q_1,\cdots, q_s)$, and (the ratio of) the constant coefficients $C_{\{p_i\}\{q_i\}}$ will be determined below.
Note also that the number $N$ of fields, and the number $n$ of twice differentiated $\pi$ which appear in a given product
$\left[
\begin{array}{cccc}
p_1 & p_2 & \cdots & p_r \\1 & 2 & \cdots & r
\end{array}\right]\left<
\begin{array}{cccc}
q_1 & q_2 & \cdots & q_s \\1 & 2 & \cdots & s
\end{array}\right> $ are given respectively by
\begin{eqnarray}
N &=& \left(\sum_{i=1}^{i=r} p_i \times i\right) +\left(\sum_{j=1}^{j=s} q_j \times (j+2)\right),
\label{NPI} \\
n &=& N- 2 \sum_{j=1}^{j=s}q_j.
\label{NDD}
\end{eqnarray}
We now look for the most general theory which obeys condition (iii) as well as (i) and (ii).
\subsubsection{Fourth order derivatives containing $\Box \pi^{\alpha}{}_{\beta}$}
Inspired by the proof given in \cite{Nicolis:2008in}, start from a particular monomial of the form
${\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}$, given in (\ref{MONO}), with specific values of the $p_i$ (for $1\leq i\leq r$) and of the $q_j$ (for $1\leq j\leq s$). If the theory considered obeys condition (iii), {\it all} third and fourth order derivatives must disappear from the field equations. There are many 4th order derivative terms, but first focus on those containing $\Box \pi^{\alpha}{}_{\beta}$ (others will be discussed later). When varying ${\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}$, such terms can appear through
\begin{eqnarray}
\delta_{\pi} \; [i]^{p_i} &\supset& \frac{2 i p_{i}}{i-1} [\Box(i-1)] [i]^{p_i - 1} \; \delta \pi \qquad \qquad (i>1)
\label{WAY1}
\\
\delta_{\pi} \; \langle i \rangle^{q_i} &\supset& 2 q_i \langle \Box(i-1) \rangle \langle i \rangle^{q_i - 1}\; \delta \pi\qquad \qquad (i>1)
\label{WAY2}
\end{eqnarray}
where
\begin{eqnarray}
[\Box(j)] &\equiv & \sum_{k=1}^j \pi^{\mu_1}_{\hphantom{\mu_1}{\mu_2}} \pi^{\mu_2}_{\hphantom{\mu_2}\mu_3} \cdots \pi^{\mu_{k-1}}_{\hphantom{\mu_{k-1}}\mu_k} (\Box \pi^{\mu_{k}}_{\hphantom{\mu_{k}}\mu_{k+1}}) \cdots
\pi^{\mu_{j-1}}_{\hphantom{\mu_{j-1}}\mu_j} \pi^{\mu_j}_{\hphantom{\mu_j} \mu_1}
\\
&=& j\times \pi^{\mu_1}_{\hphantom{\mu_1}{\mu_2}} \pi^{\mu_2}_{\hphantom{\mu_2}\mu_3} \cdots \pi^{\mu_{j-1}}_{\hphantom{\mu_{j-1}}\mu_j} (\Box \pi^{\mu_{j}}_{\hphantom{\mu_{j}}\mu_{1}}) \\
\langle \Box(j) \rangle & \equiv &
\sum_{k=1}^j \pi_{\mu_1} \pi^{\mu_1}_{\hphantom{\mu_1}{\mu_2}} \cdots \pi^{\mu_{k-1}}_{\hphantom{\mu_{k-1}}\mu_k} (\Box \pi^{\mu_{k}}_{\hphantom{\mu_{k}}\mu_{k+1}}) \cdots
\pi^{\mu_{j}}_{\hphantom{\mu_{j}}\mu_{j+1}} \pi^{\mu_{j+1}} \, .
\end{eqnarray}
Now consider the contribution from (\ref{WAY1}). On varying ${\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}$ this yields the term
\begin{eqnarray} \label{4DER1}
\frac{2 i p_i}{i-1} f(\pi,X) \times [1]^{p_1} [2]^{p_2} \cdots
[i-1]^{p_{i-1}}
[\Box(i-1)] [i]^{p_i-1} \cdots [r]^{p_r}\left<
\begin{array}{cccc}
q_1 & q_2 & \cdots & q_s \\1 & 2 & \cdots & s
\end{array}\right>
\end{eqnarray}
in the equations of motion, and it can
only be cancelled if one adds to the Lagrangian a term proportional to
\begin{eqnarray}
f \times \left[
\begin{array}{ccccccccc}
p_1+1 & p_2 & \cdots & p_{i-2}& p_{i-1}+1&p_{i}-1&p_{i+1}&\cdots& p_r \\1 & 2 & \cdots & i-2& i-1 &i&i+1&\cdots& r
\end{array}\right]\left<
\begin{array}{cccc}
q_1 & q_2 & \cdots & q_s \\1 & 2 & \cdots & s
\end{array}\right>.
\label{NOTHING}
\end{eqnarray}
Indeed, variation of the first term $[1]^{p_1+1} \equiv \left(\Box \pi \right)^{p_1+1}$ gives, after integrating by parts and shifting the $\Box$ operator onto one of the cycles $[i-1]$, a term proportional to (\ref{4DER1}). The same term is obtained from varying one of the twice differentiated $\pi$ inside a cycle $[i-1]$ and, on integrating by parts, acting with the derivatives on one $\Box \pi$.
Thus a {\it necessary} condition for the theory considered to obey conditions (i), (ii) and (iii) is that it must contain in its action the specific linear combination
\begin{eqnarray}
& f& \left<
\begin{array}{cccc}
q_1 & q_2 & \cdots & q_s \\1 & 2 & \cdots & s
\end{array}\right> \times
\nonumber
\\
&&\; \; \left\{ \left[
\begin{array}{cccc}
p_1 & p_2 & \cdots & p_{r}\\1 & 2 & \cdots & r
\end{array}\right]
+
\alpha_{[\; ]}
\left[
\begin{array}{ccccccccc}
p_1+1 & p_2 & \cdots & p_{i-2}& p_{i-1}+1&p_{i}-1&p_{i+1}&\cdots& p_r \\1 & 2 & \cdots & i-2& i-1 &i&i+1&\cdots& r
\end{array}\right]\right\}
\label{TERM2}
\end{eqnarray}
where $\alpha_{[ \; ]}$, which is nothing other than the ratio of two of the $C$ coefficients defined in (\ref{SUMC}),
is given by
\begin{equation}
\alpha_{[\; ]} = - \frac{ i p_i}{(p_1+1)(p_{i-1} + 1)(i-1)} \, .
\nonumber
\end{equation}
Notice that $\alpha_{[ \; ]}$ is independent of $f(\pi,X)$.
It is straightforward to check that the two terms in factor of $f$ in (\ref{TERM2}) each have the same number $N$ of fields, and number $n$ of twice differentiated $\pi$, as given in (\ref{NPI}) and (\ref{NDD}).
A similar reasoning can be applied to the ``$\left<\right>$'' piece in (\ref{MONO}). Indeed, using (\ref{WAY2}) it follows that any theory obeying conditions (i), (ii) and (iii), and which has a term in its action given by
${\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}$, must also contain a term
\begin{eqnarray} \label{TERM3}
\alpha_{\langle \rangle} \times f \times \left[
\begin{array}{cccc}
p_1+1 & p_2 & \cdots & p_r \\1 & 2 & \cdots & r
\end{array}\right]\left<
\begin{array}{cccccccc}
q_1 & \cdots & q_{j-1}+1&q_j-1&q_{j+1}&\cdots & q_s \\1 & \cdots & j-1&j&j+1&\cdots&s
\end{array}\right>
\end{eqnarray}
where
\begin{equation}
\alpha_{\langle \rangle} = - \frac{ q_j}{(p_1+1)(q_{j-1} + 1)} \, .
\nonumber
\end{equation}
Once again observe that the above terms (\ref{MONO}) and (\ref{TERM3}) have the same number $N$ of fields, and same number $n$ of twice differentiated $\pi$'s.
We can thus define mappings $F$ and $G$ on the set of monomials (\ref{MONO})
which appear in the Lagrangian of any theory obeying conditions (i), (ii) and (iii), such that $F$ maps any term
${\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}$, i.e.~(\ref{MONO}), to the monomial (\ref{4DER1}) and similarly $G$ maps the term (\ref{MONO}) to (\ref{TERM3}). Then, any two terms related by those mapping (or their inverse $F^{-1}$ and $G^{-1}$) must have coefficients which are equal (up to combinatorial factors) in order to eliminate fourth order derivatives containing $\Box \pi^{\alpha}{}_{\beta}$. These mapping can easily be pictured by a graph whose nodes are labelled by the set $\{p_1,p_2,\cdots,p_r,q_1,q_2,\cdots,q_s\}$ and represent the monomial, and such that two nodes are connected if and only if they are image of each other by the mappings $F$ or $G$ or their inverse (or successive applications of $F$,$G$, $F^{-1}$ or $G^{-1}$).
Now observe that in going from (\ref{MONO}) to (\ref{NOTHING}), the power of the cycle $[i]$ is lowered by 1, whereas the power of the cycles $[i-1]$ and $[1]$ is increased by 1. Similarly, in going from the term (\ref{MONO}) to (\ref{TERM3}) the power of the cycle $\langle j \rangle $ is lowered by 1 whereas that of the cycles $\langle j-1 \rangle $ and $[1]$ increases by 1. Hence, by acting recursively with the mappings $F$ and $G$, starting from the cycles $[r]$ and $\left<s\right>$ which have the {\it largest} length $r$ and $s$ respectively, one ends up with the conclusion that {\it any term ${\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}$ is connected (after several applications of the maps $F$ and $G$) to a term which has all but $p_1$ and $q_1$ vanishing}. That is, a term of the form
\begin{eqnarray}
{\cal L}^p_q
&=& f \times [1]^p \left<1\right>^q
\nonumber
\\
&=& f \times \left(\Box \pi \right)^p \left(\pi^\mu \pi_{\mu \nu} \pi^\nu\right)^q \, ,
\label{TERM5}
\end{eqnarray}
where
\begin{eqnarray}
q &=& \sum_{j=1}^{s} q_j \, ,
\label{Q}
\\
p &=& \sum_{i=1}^{r} (i p_i) + \sum_{j=1}^{s} q_j (j-1) \, .
\label{P}
\end{eqnarray}
Alternatively, on using (\ref{NPI}) and (\ref{NDD}), $p$ and $q$ are determined by the number of fields and twice differentiated $\pi$ in the part of the monomial containing second derivatives, through
\begin{eqnarray}
p &=& \frac{1}{2} \left( 3 n-N\right), \\
q&=& \frac{1}{2}\left(N - n\right).
\label{Q2}
\end{eqnarray}
Since the term (\ref{TERM5}) is the {\it same} for all ${\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}$ which has fixed values of $N$ and $n$, one can conclude that the graph introduced above, and representing monomials with fixed values of $N$ and $n$, is {\it connected}. In other words, it means that any term ${\cal L}^{p_1,p_2,\cdots, p_r}_{q_1,q_2,\cdots,q_s}$ with fixed $N$ and $n$ must appear in the action, with a common function $f$ and fixed coefficients.
Hence, to conclude the first part of our reasoning, we get a family of (possibly trivial but each uniquely determined) theories indexed by the values of $N$ and $n$. More specifically, for a given values of $N$ and $n$ the theory is uniquely specified by the coefficient in front of (\ref{TERM5}) (as well as the function $f$). The coefficients of all the other monomials with the same number of $N$ and $n$ will be proportional to the coefficient of (\ref{TERM5}) and the proportionality factor will be independent of the specific function $f$ (since this factor is uniquely determined by requiring the vanishing of terms with four derivatives $\Box \pi^{\alpha}{}_{\beta}$, and this procedure is blind to the chosen form for $f$).
\subsubsection{Other fourth order derivative terms}
Now consider other fourth order derivative terms which may appear in the equations of motion. In particular, if $q \geq 2$, the field equation derived from (\ref{TERM5}) will contain a fourth derivative term proportional to
\begin{eqnarray}
f \times \left(\Box \pi \right)^p \left(\pi^\lambda \pi^\rho \pi^\sigma \pi^\tau \pi_{\lambda \rho \sigma \tau} \right)\left(\pi^\mu \pi_{\mu \nu} \pi^\nu\right)^{q-2}
\end{eqnarray}
(obtained by varying one of the twice differentiated $\pi$ appearing in one cycle $\left<1\right>$ and integrating by parts on the other cycle $\left<1\right>$). However, it is impossible to cancel this term by the variation of any of the other terms in the Lagrangian (connected to ${\cal L}^{p}_{q}$ through $F$ and $G$). Thus we conclude that there are two possibilities --- the power $q$ appearing in (\ref{TERM5}) must take the value $0$ or $1$, so that on using (\ref{Q2}),
\begin{eqnarray}
q=0 &\Longleftrightarrow& N =n
\label{POS1}
\\
q=1 & \Longleftrightarrow& N =n + 2.
\label{POS2}
\end{eqnarray}
\subsubsection{Second order equations of motion}
First focus on $q=0$, namely one particular theory which obeys (i), (ii), (iii), has a fixed value of $n$, and $N=n$. From (\ref{Q}), it follows that all the $q_j$ must vanish so that the Lagrangian is a sum of monomials made of $\left[
\begin{array}{cccc}
p_1& p_2 & \cdots & p_r \\1 & 2 & \cdots & r
\end{array}\right]$ only. As we have just seen, these have relative coefficients which are fixed and independent of the choice of the function $f$. One can conclude that those coefficients must be the ones, ${\cal C}^{N+1}_{p_1,\cdots,p_r}$, appearing in the expansion of ${\cal E}_{N+1}$ (see Eq.~(\ref{COEFC})).
Indeed, we know that if we consider ${\cal E}_{N+1}$ as an action, this action has a vanishing equation of motion since ${\cal E}_{N+1}$ is a total derivative. Hence it obeys the hypotheses (i), (ii) and (iii) and it has also the correct power, $n = N$, of twice differentiated $\pi$ appearing in the expansion in terms of monomials $\left[
\begin{array}{cccc}
p_1& p_2 & \cdots & p_r \\1 & 2 & \cdots & r
\end{array}\right]$.
Hence, one is led to the conclusion that any theory obeying (i), (ii) and (iii), with a fixed value of $n$ and $N= n$ must have a Lagrangian of the form $g \times \mathcal{E}_{N+1}$, where $g$ is some function of $\pi$ and $X$. Thus, on using (\ref{DEFGAL3}), it can be rewritten as (defining $f\equiv g X^{-1}$)
\begin{eqnarray}
\label{FAM1}
\mathcal{L}_{n}^{(3)}\{f\}& \equiv &f(\pi,X) \times \mathcal{L}^{{\rm Gal},3}_{n+2}
\\
&=& f(\pi,X) \times \mathcal{L}^{{\rm Gal},3}_{N} \, .
\nonumber
\end{eqnarray}
A similar argument applies to the other family of models. Here $q=1=\sum_{j=1}^{s} {q_j}$, and $N=n+2$ for a given value of $n$: these models must have a Lagrangian of the form
\begin{eqnarray} \label{FAM2}
\mathcal{L}_{n}^{(2)}\{f\}&\equiv &f(\pi,X) \times \mathcal{L}^{{\rm Gal},2}_{n+2}, \\
&=&f(\pi,X) \times \mathcal{L}^{{\rm Gal},2}_{N}.
\nonumber
\end{eqnarray}
Indeed $\mathcal{L}^{{\rm Gal},2}_{N}$ obeys hypotheses (i), (ii) and (iii), and, as recalled after equation (\ref{DEF<}), has an expansion in terms of monomials $\left[
\begin{array}{cccc}
p_1& p_2 & \cdots & p_r \\1 & 2 & \cdots & r
\end{array}\right]\left<
\begin{array}{ccc}
q_1 & \cdots & q_s \\1 & \cdots&s
\end{array}\right>$
with $\sum_{j=1}^{j=s} {q_j}=1$.
Finally, it is straightforward to show that the field equations derived from Lagrangians (\ref{FAM1}) and (\ref{FAM2}) are indeed second order (so that third order derivative terms also vanish).
This is in fact a direct consequence of the ``main lemma'' of the previous section, since using the expressions (\ref{LGAL2}) and (\ref{LGAL3}) one can see that Lagrangians (\ref{FAM1}) and (\ref{FAM2}) are of the form (\ref{L0}). In the case of ${\cal L}^{(3)}_{n}$, the explicit equations of motion are given in (\ref{EOFM}).
\subsubsection{Unique family of models satisfying the ``main lemma''}
Finally, we are in a position to show that (\ref{FAM1}) and (\ref{FAM2}) are in fact equivalent up to total derivatives, so that the unique theory satisfying (i), (ii) and (iii) is indeed given by an arbitrary linear combination of the Lagrangians ${\cal L}^{(3)}_{n}\{f\}$
as advertised in section \ref{SEC:RESULT}.
Define $\mathcal{L}_{n}^{(1)}\{f\} \equiv f(\pi,X) {\cal{L}}_{n+2}^{\rm{Gal}, 1}$. Then the identity (\ref{3Lid}) amongst the 3 Galileon models is now generalised to
\begin{equation}
{\label{3Lidgen}}
n\,\mathcal{L}_{n}^{(2)}\{f\}=\mathcal{L}_{n}^{(3)}\{f\}-\mathcal{L}_{n}^{(1)}\{f\} \, ,
\end{equation}
and similarly (\ref{REL2}) becomes
\begin{equation}
{\label{3Libp}}
2{\cal L}_{n}^{(2)}\left\{ f+Xf_{X}\right\} =-{\cal L}_{n-1}^{(1)}\left\{ Xf_{\pi}\right\} -{\cal L}_{n}^{(3)}\left\{ f\right\} +\partial_{\mu}\left(f(\pi,X)J_{(n)}^{\mu}\right),
\end{equation}
$J^\mu_{(n)}\equiv J^\mu_N$ is given in Eq.~(\ref{DEFJ}). Elimination of $\mathcal{L}^{(1)}_{n}$ between these two equations yields a recurrence relation between $\mathcal{L}^{(2)}_{n}$ and $\mathcal{L}^{(3)}_{n}$, namely
\begin{equation}
{\label{L2frec}}
{\cal L}_{n}^{(2)}\left\{ f\right\} =-\left(n-1\right){\cal L}_{n-1}^{(2)}\left\{ \frac{\partial g_{1}}{\partial\pi}\right\} +{\cal L}_{n}^{(3)}\left\{ \frac{g_{1}}{X}\right\} +{\cal L}_{n-1}^{(3)}\left\{ \frac{\partial g_{1}}{\partial\pi}\right\} + \text{tot.~div.}\;,
\end{equation}
with
\begin{equation}
\nonumber
g_{1}\{f\} =-\frac{1}{2}\int_{0}^{X}dYf\left(\pi,Y\right).
\end{equation}
On using (\ref{L2frec}) repeatedly, it follows that (up to a total derivative) ${\cal L}_{n}^{(2)}\{f\}$ can be expressed
as linear combination of ${\cal L}_{i}^{(3)}\{f\}$. Specifically
\begin{eqnarray}{\label{L2rec}}
{\cal L}_{n}^{(2)}\{f\}={\cal L}_{0}^{(3)}\left\{\frac{\partial g_{n,1}}{\partial\pi}\right\}+
\sum_{i=1}^{n-1} {\cal L}_{i}^{(3)}\left\{\frac{g_{n,i}}{X}+\frac{\partial g_{n,i+1}}{\partial\pi}\right\}+{\cal L}_{n}^{(3)}\left\{\frac{g_{n,n}}{X}\right\}+\text{tot.~div.}\;,
\end{eqnarray}
where $\mathcal{L}_0^{(3)}\{f\}=Xf$ for consistency, and
\begin{eqnarray}
g_{n,i}\{f\} &\equiv& \frac{\left(n-1\right)!}{\left(i-1\right)!}g_{n-i+1}\{f\},
\nonumber
\\
{g}_{i}\{f\} &\equiv& - \frac{1}{2^{i}}\left(\frac{\partial}{\partial\pi}\right)^{i-1}
\int_{X_0}^{X}dX_{1}\int_{X_0}^{X_{1}}dX_{2}\cdots\int_{X_0}^{X_{i-1}}dX_{i}\,f\left(\pi,X_{i}\right),
\nonumber
\end{eqnarray}
Thus Eq.~(\ref{L2rec}) shows the equivalence of the $q=0$ and $q=1$ Lagrangians given in (\ref{FAM1}) and (\ref{FAM2}) respectively.
Finally, observe that $\mathcal{L}_{D}^{(3)}\{f\}$ is in fact a linear combination of $\mathcal{L}^{(3)}_k$ with $k=0,\ldots,D-1$. This follows from (\ref{L2rec}) together with the fact that for $n=D$ equation (\ref{3Lidgen}) reduces to $D\,\mathcal{L}_{D}^{(2)}\{f\}=\mathcal{L}_{D}^{(3)}\{f\}$. Thus to conclude, the most general Lagrangian in $D$-dimensions obeying conditions (i), (ii) and (iii) is given by
\begin{equation}
{\cal L} = \sum_{n=0}^{D-1} {\cal L}_{n}\{ f_n \} \, ,
\label{final}
\end{equation}
where $f_n$ are arbitrary functions of $\pi$ and $X$ and
${\cal L}_{n}\{f\} \equiv {\cal L}_{n}^{(3)}\{ f \}$.
\section{Covariantization}
\label{COVAR}
We now turn our discussion from flat space-time to curved space-time.
Let us start from the general Lagrangian ${\cal L}_{n}\{f\}$ in our family (\ref{GENL}), writing it in the form of Eq.~(\ref{L0}) with ${\cal T}_{(2n)}$ defined by
\begin{eqnarray}
\mathcal{T}_{(2n)}&=& \mathcal{T}_{(2n)}\left(\pi,X\right)
\nonumber
\\
&=&f(\pi,X)\times X \mathcal{A}_{(2n)} ,
\end{eqnarray}
and where $n$ is the number of second derivatives of the field in the Lagrangian. On replacing all partial derivatives appearing in this Lagrangian by covariant derivatives, we obtain a minimally covariantized theory. As we will now see, following \cite{Deffayet:2009mn}, the field equations of this covariantized model contain derivatives of order higher than two (varying the action with respect to $\pi$ or with respect to the metric): however, we will also see that there exists a non minimal covariantization removing all such higher order derivatives.
Let us then consider variation with respect to the scalar $\pi$ of the minimally covariantized version of (\ref{GENL}).
It reads
\begin{equation}
\delta\mathcal{L}_{n}\{f\} =\delta \mathcal{T}_{(2n)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\pi_{\mu_{1}\nu_{1}}\cdots\pi_{\mu_{n}\nu_{n}}+n\mathcal{T}_{(2n)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\delta\pi_{\mu_{1}\nu_{1}}\pi_{\mu_{2}\nu_{2}}\cdots\pi_{\mu_{n}\nu_{n}}.
\label{DELC}
\end{equation}
Above, the only ``dangerous terms" (recall that these are terms leading to expressions in the field equations depending on derivatives of order higher than two) come only from the second term on the right hand side. Indeed by virtue of a straightforward generalisation to curved space-time of the first lemma in section \ref{LEM1} (see the proof in appendix \ref{APPA}), dangerous terms coming from the first piece on the right hand side of (\ref{DELC}) are exactly compensated by terms coming from the second piece. After these compensations, first derivatives of the Riemann tensor (i.e.~third derivatives of the metric) remain,
and are therefore troublesome. Indeed we have
\begin{equation}
\delta\mathcal{L}_{n}\{f\}
\sim -\frac{n(n-1)}{4}\mathcal{T}_{(2n)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\pi^{\lambda}R_{\mu_{1}\mu_{2}\nu_{1}\nu_{2};\lambda}\pi_{\mu_{3}\nu_{3}}\cdots\pi_{\mu_{n}\nu_{n}}\delta\pi,
\end{equation}
where here and below we use the same notation as in \cite{Deffayet:2009mn} so that
the symbol $\sim$ means that we write only the dangerous terms, up to total derivatives.
In order for the equations of motion to be second order, it is enough to add a finite number of terms whose variations exactly cancel the third derivatives of the metric. Starting with the above dangerous term, we can add a term proportional to $\left(\int_{X_{0}}^{X}\mathcal{T}_{(2n)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}} \left(\pi,X_1\right) dX_{1}\right)R_{\mu_{1}\mu_{2}\nu_{1}\nu_{2}}\pi_{\mu_{3}\nu_{3}}\cdots\pi_{\mu_{n}\nu_{n}}$ to our Lagrangian. Then we need to add another term in order to compensate for the dangerous terms arising from our correction, and so on. The general term in this series of Lagrangians is
\begin{equation}
\mathcal{L}_{n,p}\{ f\}=\mathcal{P}_{(p)}^{\mu_{1}\mu_{2}\cdots\mu_{n}\nu_{1}\nu_{2}\cdots\nu_{n}}\mathcal{R}_{(p)}\mathcal{S}_{(q\equiv n-2p)}
\end{equation}
where we again follow the notation of \cite{Deffayet:2009mn} and
\begin{eqnarray}
\mathcal{R}_{(p)}&\equiv &\prod_{i=1}^{p}R_{\mu_{2i-1}\mu_{2i}\nu_{2i-1}\nu_{2i}},
\nonumber
\\
\mathcal{S}_{(q\equiv n-2p)}&\equiv&\prod_{i=0}^{q-1}\pi_{\mu_{n-i}\nu_{n-i}},
\nonumber
\end{eqnarray}
and $\mathcal{P}_{(p)}$ is the ($p$ times) repeated integral of $\mathcal{T}_{(2n)}$ with respect to $X$ defined by\footnote{$X_0$ being an arbitrary constant, whose presence is related to the possibility
of adding terms, all vanishing in flat space, that avoid higher derivatives. See e.g. (\ref{L4CGALC})-(\ref{L5CGALC}).}
\begin{eqnarray}
\mathcal{P}_{(p)} \equiv \int_{X_{0}}^{X} dX_1 \int_{X_{0}}^{X_{1}} dX_2 \cdots\int_{X_{0}}^{X_{p-1}} dX_p \; \mathcal{T}_{(2n)}^{\mu_{1}\mu_{2}\cdots\mu_{n}\nu_{1}\nu_{2}\cdots\nu_{n}}\left(\pi,X_1\right).
\nonumber
\end{eqnarray}
Notice that we use the conventions $\mathcal{L}_{n,0}\{f\}=\mathcal{L}_{n}\{f\}$, ${\cal{S}}_{1}=\pi_{\mu_n \nu_n} $ and ${\cal{S}}_{q\leq 0} = 1$. It follows from the lemma of Appendix \ref{APPA} that $\mathcal{L}_{n,p}\{f\}$ does not yield any more dangerous terms for $p \geq \lfloor n/2 \rfloor$ (where $\lfloor n/2 \rfloor$ denotes the integer part of $n/2$). Thus, as we will confirm, only a finite number of terms $\mathcal{L}_{n,p}\{f\}$ is necessary.
In order to make sure that adding these terms to the initial Lagrangian exactly cancels all higher derivatives in the equations of motion, and to determine the coefficients that make this possible, we now compute their variation $\delta\mathcal{L}_{n,p}$ with respect to the field $\pi$, only paying attention to dangerous terms. We get (after suitable integrations by parts)
\begin{eqnarray}
\delta\mathcal{L}_{n,p} &=&\delta \mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p)}\mathcal{S}_{(q)}+(n-2p)\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p)}\mathcal{S}_{(q-1)}\delta\pi_{\mu_{2p+1}\nu_{2p+1}}\nonumber \\
&=&2\mathcal{P}_{(p-1)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\pi^{\lambda}\delta\pi_{\lambda}\mathcal{R}_{(p)}\mathcal{S}_{(q)}+(n-2p)\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p)}\mathcal{S}_{(q-1)}\delta\pi_{\mu_{2p+1}\nu_{2p+1}}\nonumber \\
&\sim& -2p\mathcal{P}_{(p-1)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p-1)}\mathcal{S}_{(q)}\pi^{\lambda}R_{\mu_{2p-1}\mu_{2p}\nu_{2p-1}\nu_{2p};\lambda}\delta\pi\nonumber \\
&& -\frac{(n-2p)(n-2p-1)}{4}\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p)}\mathcal{S}_{(q-2)}\pi^{\lambda}R_{\mu_{2p+1}\mu_{2p+2}\nu_{2p+1}\nu_{2p+2};\lambda}\delta\pi. \nonumber
\end{eqnarray}
Notice that, once again, the third derivatives of $\pi$ coming from the variation of $\mathcal{P}_{(p)}$ are exactly canceled by those coming from the variation of $\mathcal{S}_{(q)}$. The second Bianchi identity {\it ensures} that no other derivatives of the Riemann tensor appear. To summarise, in order that the equations of motion contain no more than second order derivatives, the Lagrangian must be given by the linear combination
\begin{equation}
\mathcal{L}^{\rm {cov}}_{n}\{f\}=\sum_{p=0}^{\lfloor\frac{n}{2}\rfloor}\mathcal{C}_{n,p}\mathcal{L}_{n,p}\{f\},
\label{THEANSWER}
\end{equation}
where the coefficients obey the specific recurrence relation $\mathcal{C}_{n,p+1}=-\frac{(n-2p)(n-2p-1)}{8(p+1)}\mathcal{C}_{n,p}$ for $p\geq 0$ and $\mathcal{C}_{n,0}=1$.
This gives
\begin{eqnarray} \label{COEFCNP} \mathcal{C}_{n,p}=\left(-\frac{1}{8}\right)^{p}\frac{n!}{(n-2p)!p!}.\end{eqnarray}
Note that similar expressions have been obtained in \cite{Deffayet:2009mn}. However, there, the covariantization was derived for the Lagrangians ${\cal L}^{Gal,1}_N$. It was also observed that the covariantized action could be written in different forms using various total derivatives. The same is true here and hence, taking also into account (\ref{REL1}), it is not straightforward to compare the covariantized form (\ref{THEANSWER})-(\ref{COEFCNP}) to the one given in \cite{Deffayet:2009mn}.
Remarkably the Lagrangian $\mathcal{L}^{{\rm cov}}_{n}\{f\}$ also yields second order equations for the metric. Indeed by computing the variation of each term appearing in the above linear combination, but this time with respect to the metric $g_{\mu \nu}$ (denoting the metric variation by $\delta g_{\mu \nu}$), we obtain (given that the tensor $\mathcal{P}_{(p)}$ does not depend on derivatives of the metric)
\begin{equation}
\delta\mathcal{L}_{n,p} \sim p\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p-1)}\delta R_{\mu_{2p-1}\mu_{2p}\nu_{2p-1}\nu_{2p}}\mathcal{S}_{(q)}+(n-2p)\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p)}\mathcal{S}_{(q-1)}\delta\pi_{\mu_{2p+1}\nu_{2p+1}}
\nonumber \end{equation}
with (see equations (31) and (32) in Ref. \cite{Deffayet:2009mn})
\begin{eqnarray}
\delta\pi_{\mu_{2p+1}\nu_{2p+1}}&=&-\frac{1}{2}\pi^{\lambda}(\delta g_{\lambda\mu_{2p+1};\nu_{2p+1}}+\delta g_{\lambda\nu_{2p+1};\mu_{2p+1}}-\delta g_{\mu_{2p+1}\nu_{2p+1};\lambda})
\nonumber
\\
\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\delta R_{\mu_{2p-1}\mu_{2p}\nu_{2p-1}\nu_{2p}}&=&2\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\delta g_{\mu_{2p-1}\nu_{2p};\mu_{2p}\nu_{2p-1}}
\nonumber
\\ && \qquad \qquad \qquad +\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\delta g^{\sigma}{}_{\mu_{2p-1}}R_{\sigma\mu_{2p}\nu_{2p-1}\nu_{2p}} \nonumber.
\end{eqnarray}
This yields
\begin{eqnarray}
\delta\mathcal{L}_{n,p} &\sim& 2p\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p-1)}\mathcal{S}_{(q)}\delta g_{\mu_{2p-1}\nu_{2p};\mu_{2p}\nu_{2p-1}}
\nonumber \\ &&
+\frac{(n-2p)}{2}\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p)}\mathcal{S}_{(q-1)}\pi^{\lambda}\delta g_{\mu_{2p+1}\nu_{2p+1};\lambda}. \nonumber
\end{eqnarray}
After two integrations by parts, we are led to
\begin{eqnarray}
\delta\mathcal{L}_{n,p}&\sim& 4p\mathcal{P}_{(p-1)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p-1)}\mathcal{S}_{(q)}\pi^{\lambda}\pi_{\nu_{2p-1}\mu_{2p}\lambda}\delta g_{\mu_{2p-1}\nu_{2p}}\nonumber \\
&& -\frac{p(n-2p)}{2}\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p-1)}\mathcal{S}_{(q-1)}\pi^{\lambda}R_{\mu_{2p}\mu_{2p+1}\nu_{2p-1}\nu_{2p+1};\lambda}\delta g_{\mu_{2p-1}\nu_{2p}} \nonumber \\
&& -\frac{p(n-2p)}{2}\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p-1)}\mathcal{S}_{(q-1)}\pi^{\lambda}R_{\mu_{2p-1}\mu_{2p}\nu_{2p-1}\nu_{2p};\lambda}\delta g_{\mu_{2p+1}\nu_{2p+1}}\nonumber \\
&& -\frac{(n-2p)(n-2p-1)}{2}\mathcal{P}_{(p)}^{\mu_{1}\cdots\mu_{n}\nu_{1}\cdots\nu_{n}}\mathcal{R}_{(p)}\mathcal{S}_{(q-2)}\pi^{\lambda}\pi_{\lambda\mu_{2p+2}\nu_{2p+2}}\delta g_{\mu_{2p+1}\nu_{2p+1}} \, .
\nonumber
\end{eqnarray}
On relabeling $\mu_{2p-1}\leftrightarrow\mu_{2p+1}$ and $\nu_{2p}\leftrightarrow\nu_{2p+1}$ in the second term, the second and third terms are seen to cancel. Also, by changing $\nu_{2p}\leftrightarrow\nu_{2p-1}$ in the first term, it is clear that all the other dangerous terms vanish if the coefficients $\mathcal{C}_{n,p}$ are given by Eq.~(\ref{COEFCNP}).
To conclude, the generalisation to curved space-time of our family of Lagrangians (\ref{GENL}) which yield second order equations of motion, is given by (\ref{THEANSWER}).
\section{Examples}
\label{EXAMPLES}
Although in section \ref{UNIK} we have completely classified all models of the form (\ref{L0}) (and more generally all models obeying conditions (i), (ii) and (iii) of section \ref{UNIK}), it is of interest to discuss some specific models in which the tensor $\mathcal{T}_{(2n)}$ of Eqs.(\ref{L0}-\ref{propT}) takes a simple form. We will in turn give a straightforward method with which to construct a suitable tensor $\mathcal{T}_{(2n)}$ --- namely satisfying the properties given in the the main lemma of section \ref{LEM1}. We then discuss the hierachical constructions of \cite{Fairlie} (see also \cite{Fairlie:2011md} for a recent summary) in our notation. Finally, we give the covariantization of the conformal Galileon in 4 dimensions (section \ref{CONFGALCOV}).
\subsection{Construction with antisymmetric tensors}
One way to obtain a tensor $\mathcal{T}_{(2n)}$ is to consider a set of twice contravariant tensors $F^{\mu \nu}_{i}$, $i = 1 \ldots n$, (and $n \leq D$) depending on $\pi$ and $\pi^\lambda$ (where $\mu$ and $\nu$ label space-time indices, whereas $i$ labels the tensors). Then to define the $D$-contravariant tensor $\mathpzc{E}_{\{F_i\}}$ by
\begin{eqnarray}
\label{DEFE}
\mathpzc{E}_{\{F_i\}}^{\hphantom{\{F_i\}}\mu_{\vphantom{()}1} \ldots \mu_{\vphantom{()}n} \sigma_{\vphantom{()}1} \ldots \sigma_{\vphantom{()}D-n}} \equiv
\,\varepsilon_{\nu_{\vphantom{()}1} \ldots
\nu_{\vphantom{()}n}}^{\hphantom{nu_{\vphantom{()}1} \ldots
\nu_{\vphantom{()}n}}
\sigma_{\vphantom{()}1} \ldots
\sigma_{\vphantom{()}D-n}} F_{(1}^{\mu_{1}\nu_{1}} F_{2}^{\mu_2 \nu_2} \ldots F_{n)}^{\mu_n\nu_n}
\end{eqnarray}
where the brackets denote symmetrisation over the tensors $F^{\mu \nu}_{i}$. The tensor $\mathpzc{E}_{\{F_i\}}$ is easily seen to be completely antisymmetric over its first $n$ indices as well as (separately) on it last $D-n$ indices.
The special choice
\begin{eqnarray} \label{MET}
F^{\mu \nu}_i = g^{\mu \nu}, \qquad {\rm for \; all} \; i
\end{eqnarray}
gives a tensor $\mathpzc{E}$ equal to the Levi-Civita $\varepsilon$ tensor. Another interesting case arises by choosing, for example, $F_1^{\mu \nu}=\pi^\mu \pi^\nu$, and all the other $F^{\mu \nu}_i$ given by the metric as in (\ref{MET}). In that case the tensor $\mathpzc{E}$ will be denoted by $\tilde{\varepsilon}_{(n)}$
and is given by
\begin{eqnarray}
\tilde{\varepsilon}_{(n)}{}^{\mu_{\vphantom{()}1} \ldots \mu_{\vphantom{()}n} \sigma_{\vphantom{()}1} \ldots \sigma_{\vphantom{()}D-n}} = \frac{1}{n}
\,\varepsilon_{\nu_{\vphantom{()}1} \ldots
\nu_{\vphantom{()}n}}^{\hphantom{nu_{\vphantom{()}1} \ldots
\nu_{\vphantom{()}n}}
\sigma_{\vphantom{()}1} \ldots
\sigma_{\vphantom{()}D-n}} && \left(\pi^{\mu_1} \pi^{\nu_1} g^{\mu_2 \nu_2} \ldots g^{\mu_n \nu_n} \right. \nonumber \\
&& + g^{\mu_1 \nu_1} \pi^{\mu_2} \pi^{\nu_2} g^{\mu_3 \nu_3} \ldots g^{\mu_n \nu_n} \nonumber \\
&& + \ldots \nonumber \\
&& + \left.g^{\mu_1 \nu_1} \ldots g^{\mu_{n-1} \nu_{n-1}} \pi^{\mu_n} \pi^{\nu_n} \right).
\end{eqnarray}
This tensor enters in particular in the form (\ref{LGAL2})-(\ref{TT2}) of the action for the Galileon, as we will see below. We can also construct a tensor $\mathpzc{E}$ given by a linear combination of $\varepsilon$ and $\tilde{\varepsilon}_{(n)}$ by choosing
\begin{eqnarray}
F_i^{\mu \nu} = \frac{\partial^2 F_i}{\partial \pi_\mu \partial \pi_\nu}
\end{eqnarray}
where $F_i$ is some scalar function of $X$ and $\pi$: such $F_i^{
\mu \nu}$ arise in the hierachical construction of \cite{Fairlie} (see section \ref{ANTI}). In this case
\begin{eqnarray}
\frac{\partial^2 F_i}{\partial \pi_\mu \partial \pi_\nu} = 4 \frac{\partial^2 F_i}{\partial X^2} \pi^\mu \pi^\nu+ 2 \frac{\partial F_i}{\partial X} g^{\mu \nu}.
\end{eqnarray}
Note that if there is more than one tensor $F^{\mu \nu}_{i}$ proportional to $\pi^\mu$ (or $\pi^\nu$) in Eq.~(\ref{DEFE}), then the contraction with the Levi-Civita tensor on the right hand side of this equation implies a vanishing $\mathpzc{E}$.
Having at hand the tensors $\mathpzc{E}$,
it is easy to build tensors $\mathcal{T}$ with the required properties: namely simply replace the Levi-Civita $\varepsilon$ tensor on the right hand side of equation (\ref{DEFAten}) by the $\mathpzc{E}$ tensor, and define, in analogy with (\ref{TENSGAL1}) a tensor $\mathcal{T}$ given by
\begin{equation} \label{DEFEbis}
\mathcal{T}_{(2n)}^{\mu_{\vphantom{()}1} \mu_{\vphantom{()}2}
\ldots \mu_{\vphantom{()}n} \nu_{\vphantom{()}1} \nu_{\vphantom{()}2}
\ldots \nu_{\vphantom{()}n}} \equiv f \times \,
\mathpzc{E}_{\{F_i\}}^{\hphantom{\{F_i\}}\mu_{\vphantom{()}1}
\mu_{\vphantom{()}2} \ldots
\mu_{\vphantom{()}n} \sigma_{\vphantom{()}1}\sigma_{\vphantom{()}2}\ldots
\sigma_{\vphantom{()}D-n}}
\,\mathpzc{E}_{\{G_i\} \hphantom{\nu_{\vphantom{()}1} \nu_{\vphantom{()}2} \ldots
\nu_{\vphantom{()}n}} \sigma_{\vphantom{()}1}
\sigma_{\vphantom{()}2}\ldots \sigma_{\vphantom{()}D-n}}^{\hphantom{\{G_i\}}\nu_{\vphantom{()}1} \nu_{\vphantom{()}2} \ldots
\nu_{\vphantom{()}n}}.
\end{equation}
Here the tensors $\mathpzc{E}$ are built with (possibly different) given sets of functions $F_{i}$ and $G_{i}$, and the function $f$ is an arbitrary scalar function of $\pi_\mu$ and $\pi$.
For example, if we choose $f=1$, one of the $\mathpzc{E}$ tensors to be the tensor
$\tilde{\varepsilon}_{(n)}$, and the other to be the Levi-Civita tensor $\epsilon$, then $\mathcal{T}$ is simply the tensor $\mathcal{T}_{(2n),{\rm{Gal}},2}$ of Eq.(\ref{TGAL2}), namely
\begin{equation} \label{DEFEter}
\mathcal{T}_{(2n)}^{\mu_{\vphantom{()}1} \mu_{\vphantom{()}2}
\ldots \mu_{\vphantom{()}n} \nu_{\vphantom{()}1} \nu_{\vphantom{()}2}
\ldots \nu_{\vphantom{()}n}} \equiv
\tilde{\varepsilon}_{(n)}^{\hphantom{(n)}\mu_{\vphantom{()}1}
\mu_{\vphantom{()}2} \ldots
\mu_{\vphantom{()}n} \sigma_{\vphantom{()}1}\sigma_{\vphantom{()}2}\ldots
\sigma_{\vphantom{()}D-n}}
\,\varepsilon_{ \hphantom{\nu_{\vphantom{()}1} \nu_{\vphantom{()}2} \ldots
\nu_{\vphantom{()}n}} \sigma_{\vphantom{()}1}
\sigma_{\vphantom{()}2}\ldots \sigma_{\vphantom{()}D-n}}^{\nu_{\vphantom{()}1} \nu_{\vphantom{()}2} \ldots
\nu_{\vphantom{()}n}}.
\end{equation}
\label{ANTI}
\subsection{Euler hierachies} \label{EULERHIER}
Another way to generate tensor of the form (\ref{DEFE}) is provided by the {\it Euler hierachies} of \cite{Fairlie}, which we now revisit using the notations and results of the present paper.
We start from a set of arbitrary scalar functions $F_\ell = F_\ell(\pi^\mu)$ which depend {\it only} on first derivatives of the scalar field $\pi$, and work in flat $D$-dimensional space-time. Then, denoting $W_0 = 1$, we define the recursion relation
\begin{eqnarray}
W_{\ell+1}=-\hat{\mathcal{E}}F_{\ell+1}W_{\ell},
\end{eqnarray}
where $\hat{\mathcal{E}}$ is the Euler-Lagrange operator defined in Eq.~(\ref{ELHAT}). Hence $W_{\ell}$ is the field equation of a Lagrangian $\mathcal{L}_{\ell}$ defined by
\begin{eqnarray}
\mathcal{L}_{\ell}=F_{\ell}W_{\ell-1},
\end{eqnarray}
and we build that way an ``Euler hierarchy" of equations of motion and Lagrangian for each $\ell$. An interesting aspect of this hierarchy is that one can show that $W_\ell$ is given by
\begin{eqnarray} \label{WlWl}
W_{\ell}=\mathcal{A}_{\mu_{1}\mu_{2}\cdots\mu_{\ell}}{}^{\nu_{1}\nu_{2}\cdots\nu_{\ell}} \frac{\partial^{2} F_{1}}{\partial\pi_{\mu_{1}}\partial\pi_{\rho_{1}}}\frac{\partial^{2} F_{2}}{\partial\pi_{\mu_{2}}\partial\pi_{\rho_{2}}}\cdots \frac{\partial^{2} F_{\ell}}{\partial\pi_{\mu_{\ell}}\partial\pi_{\rho_{\ell}}}\pi_{\rho_{1}\nu_{1}}\pi_{\rho_{2}\nu_{2}}\cdots\pi_{\rho_{\ell}\nu_{\ell}},
\end{eqnarray}
and
\begin{equation}
\mathcal{L}_{\ell}=\mathcal{A}_{\mu_{1}\cdots\mu_{\ell-1}}{}^{\nu_{1}\cdots\nu_{\ell-1}} \frac{\partial^{2} F_{1}}{\partial\pi_{\mu_{1}}\partial\pi_{\rho_{1}}}\cdots \frac{\partial^{2} F_{\ell-1}}{\partial\pi_{\mu_{\ell-1}}\partial\pi_{\rho_{\ell-1}}}F_{\ell}\pi_{\rho_{1}\nu_{1}}\cdots\pi_{\rho_{\ell-1}\nu_{\ell-1}} \, .
\end{equation}
The proof proceeds by induction \cite{Fairlie}, and is also given in appendix \ref{PROOFINDUC}. Notice that $\mathcal{L}_{\ell}$ can be obtained from (\ref{DEFEbis}) by setting $f\equiv \frac{1}{(\ell-1)!}F_{\ell}$ and contracting the tensor $\mathpzc{E}_{\{\frac{\partial^2 F _{k}}{\partial\pi_{\mu}\partial\pi_{\nu}}\}}$ with $\varepsilon$.
The above generic hierarchy stops after {\it at most} $D$ steps (it may stop earlier depending on the properties of the functions $F_k$) and one finds a vanishing $W_{D+1}$. As discussed in \cite{Fairlie}, and as we will see in a simple example below, the last non trivial equation of motion $W_D$ is simply given (whatever the choice made for the functions $F_k$) by the equation of motion of the maximal Galileon in $D$ dimensions, namely the one for which $N=D+1$, see (\ref{MAXX}).
The maximal and non maximal Galileons can also be obtained by choosing $F_{k} = \pi_{\mu}\pi^{\mu}/2$ for all ${k}=0,\ldots,\ell$. In this case, one has
\begin{eqnarray}
W_{\ell}&=&\mathcal{A}_{(2\ell)}^{\mu_{1}\cdots\mu_{\ell}\nu_{1}\cdots\nu_{\ell}}\pi_{\mu_{1}\nu_{1}}\cdots\pi_{\mu_{\ell}\nu_{\ell}} = \mathcal{E}_{\ell+1} \\
\mathcal{L}_{\ell}&=& \frac{1}{2} X W_{\ell-1} =\frac{1}{2}\mathcal{L}_{\ell+1}^{{\rm Gal},3}.
\end{eqnarray}
More generally, when the $F_k$ are (possibly different) functions of $X$, that is $F_k= f_k(\pi^\lambda \pi_\lambda)$, one can show that
\begin{eqnarray}
W_{\ell}&=&\alpha_\ell \mathcal{E}_{\ell+1}-\beta_\ell \mathcal{L}_{\ell+2}^{{\rm Gal},1},
\\
{\cal L}_{\ell} &=& f_\ell(X) \left( \alpha_{\ell-1} \mathcal{E}_{\ell}-\beta_{\ell-1} \mathcal{L}_{\ell+1}^{{\rm Gal},1} \right)
\label{LEULER}
\end{eqnarray}
where $\alpha$ and $\beta$ are given by
\begin{eqnarray}
\alpha_\ell(\pi^{\lambda}\pi_{\lambda})&=& \frac{2^{\ell-1}}{\ell}\sum_{k=1}^{\ell}\left(\left(2f_{k}'+4(\pi^{\lambda}\pi_{\lambda})f_{k}''\right)\prod_{j\neq k}f_{j}'\right)
\nonumber
\\
\beta_\ell(\pi^{\lambda}\pi_{\lambda})&=&\frac{2^{\ell+1}}{\ell}\sum_{k=1}^{\ell}\left(f_{k}''\prod_{j\neq k}f_{j}'\right) \, .
\nonumber
\end{eqnarray}
To conclude, since both ${\cal L}_{\ell+1}^{{\rm Gal},1}$ and $\mathcal{E}_{\ell}$ are directly related to ${\cal L}_{\ell+1}^{{\rm Gal},3}$ (see equations (\ref{REL1}) and (\ref{DEFGAL3}) respectively), the Lagrangian (\ref{LEULER}) which we have constructed using the Euler hierarchies belongs to our family of general Lagrangians given in (\ref{GENL}).
\subsection{Conformal covariant Galileons in 4 dimensions}
\label{CONFGALCOV}
The conformal galileons \cite{Nicolis:2008in} provide an other simple example of theories of the kind obtained in this work (other non trivial examples in flat and curved space-times have been obtained in \cite{VanAcoleyen:2011mj}). The Lagrangians for the conformal Galileons in 4 dimensions and flat space-time have been given explicitly in \cite{deRham:2010eu}. They read, in our notation,
\begin{eqnarray}
\mathcal{L}^{{\rm C.Gal}}_{4} & = & \frac{1}{20}e^{2\pi}X\left[10\left(\left(\square\pi\right)^{2}-\pi_{\mu\nu}\pi^{\mu\nu}\right)+4\left(X\square\pi-\pi^{\mu}\pi^{\nu}\pi_{\mu\nu}\right)+3X^{2}\right], \nonumber \\
\mathcal{L}^{{\rm C.Gal}}_{5}&=&e^{4\pi}X\left[\frac{1}{3}\left(\left(\square\pi\right)^{3}+2\pi_{\mu_{1}}^{\mu_{2}}\pi_{\mu_{2}}^{\mu_{3}}\pi_{\mu_{3}}^{\mu_{1}}-3\square\pi\pi_{\mu\nu}\pi^{\mu\nu}\right)+X\left(\left(\square\pi\right)^{2}-\pi_{\mu\nu}\pi^{\mu\nu}\right)\right.\nonumber \\
&&\left.+\frac{10}{7}X\left(X\square\pi-\pi^{\mu}\pi^{\nu}\pi_{\mu\nu}\right)-\frac{1}{28}X^{3}\right]. \nonumber
\end{eqnarray}
These can be rewritten in terms of $\mathcal{L}_{N}^{\text{Gal},i}$ and $\mathcal{L}_{n}^{\left(i\right)}\left\{ f\right\} $, as
\begin{eqnarray}
\mathcal{L}^{{\rm C.Gal}}_{4} & = & -\frac{1}{2}e^{2\pi}\mathcal{L}_{4}^{\text{Gal},3}-\frac{1}{5}e^{2\pi}X\mathcal{L}_{3}^{{\rm \text{Gal},1}}+\frac{3}{20}e^{2\pi}X^{3}\nonumber, \\
& = & -\frac{1}{2}\mathcal{L}_{2}^{\left(3\right)}\left\{ e^{2\pi}\right\} -\frac{1}{5}\mathcal{L}_{1}^{\left(1\right)}\left\{ e^{2\pi}X\right\} +\frac{3}{20}e^{2\pi}X^{3},\label{LCG4} \\
\mathcal{L}^{{\rm C.Gal}}_{5} & = & -\frac{1}{3}e^{4\pi}\mathcal{L}_{5}^{\text{Gal},3}-e^{4\pi}X\mathcal{L}_{4}^{\text{Gal},3}-\frac{10}{7}e^{4\pi}X^{2}\mathcal{L}_{3}^{{\rm \text{Gal},1}}-\frac{1}{28}e^{4\pi}X^{4},\nonumber \\
& = & -\frac{1}{3}\mathcal{L}_{3}^{\left(3\right)}\left\{ e^{4\pi}\right\} -\mathcal{L}_{2}^{\left(3\right)}\left\{ e^{4\pi}X\right\} -\frac{10}{7}\mathcal{L}_{1}^{\left(3\right)}\left\{ e^{4\pi}X^{2}\right\} -\frac{1}{28}e^{4\pi}X^{4}.\label{LCG5}\end{eqnarray}
As expected, one sees that those Lagrangians are indeed in the family obtained by our proof of uniqueness.
The covariantization of theories (\ref{LCG4}) and (\ref{LCG5}) is straightforward. Only the first term in (\ref{LCG4}) and the first two terms in (\ref{LCG5})
need compensating factors. The relevant covariantization formula are
\begin{eqnarray}
\mathcal{L}_{2}^{\text{cov}}\left\{ f\right\} & = & \mathcal{L}_{2}^{\left(3\right)}\left\{ f\right\} +\frac{1}{2}R\int_{X_{0}}^{X}dY\, f\left(\pi,Y\right)Y,\label{L2_cov}
\\
\mathcal{L}_{3}^{\text{cov}}\left\{ f\right\} & = & \mathcal{L}_{3}^{\left(3\right)}\left\{ f\right\} -3G_{\mu\nu}\pi^{\mu\nu}\int_{X_{0}}^{X}dY\, f\left(\pi,Y\right)Y.\label{L3_cov}
\end{eqnarray}
To summarize, (\ref{LCG4}) must be completed with
\begin{equation} \label{L4CGALC}
\mathcal{L}_{4}^{{\rm C.Gal,c}}=-\frac{1}{8}e^{2\pi}R\left(X^{2}-X_{0}^{2}\right),\end{equation}
and (\ref{LCG5}) must be completed with\begin{equation} \label{L5CGALC}
\mathcal{L}_{5}^{{\rm C.Gal,c}}=\frac{1}{2}e^{4\pi}\left[G_{\mu\nu}\pi^{\mu\nu}\left(X^{2}-X_{0}^{2}\right)-\frac{1}{3}R\left(X^{3}-X_{0}^{3}\right)\right],
\end{equation}
where $X_{0}$ is an arbitrary integration constant that can be taken to vanish.
Note that the coefficients of the terms proportional to $(X^2-X_0^2)$ in (\ref{L4CGALC}) and (\ref{L5CGALC}) can be directly read from the expressions given in \cite{US,Deffayet:2009mn}. However, one should take into account that one has to integrate once $e^{4 \pi} X {\cal T}_{(4),\text{Gal},3}$ to obtain the term proportional to $(X^3-X_0^3)$ in (\ref{L5CGALC}).
\section{Conclusions}
In this work, we have obtained the most general scalar theory which has an action depending on derivatives of order up to second and has second order (and lower) field equations. Those theories were shown to have Lagrangians made by taking the product of an arbitrary function of the scalar field and its first derivatives with a special form of the Galileon Lagrangian, or any linear combinations of those Lagrangians. We have also shown how to covariantize those models, while maintaining the key property that field equations are second order. Finally, we have also discussed the relation between our construction and the Euler hierarchies of Fairlie {\it et al.}. We have shown in particular that the latter construction allows one to obtain all theories which are shift symmetric.
Several questions are left for future work. On the formal side, it would be interesting to see how the above proof can be generalized to the case of $p$-forms and/or multifields. One could also investigate the possibility of having actions which depend on derivatives of order higher than two, and yet give rise to field equations of second order. On the phenomenological side, work is needed to see which subsets of the theories introduced here retain the several interesting aspects recalled in the introduction. Finally, it would be interesting to study the cosmology of, as well as cosmological perturbation theory in, phenomenologically interesting models.
\section*{Acknowledgements}
C.D.~thanks S.~Deser and G.~Esposito-Farese, as well as C.~de~Rham, J.~Khoury and A.~Tolley for discussions. X.G.~thanks N.~Deruelle, D.~Israel, B.~Shou, J.~Troost for interesting discussions and comments. We also thank D.~Langlois for many useful discussions and for his interest at an early stage of this project.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.