Spectral Statistics Beyond Random Matrix Theory
Extracted source text ยท arXiv cond-mat/9503141
%\documentstyle[aps,floats,epsf]{revtex}
%\documentstyle[aps,twocolumn,floats,epsf]{revtex}
\documentstyle[aps,tighten,preprint,epsf]{revtex}
%Last updated on March 27, 1995
%Remove in final version
\def\ret{\nonumber \\ & &\qquad}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\newcommand\inp{-\!\!\!\!\!\int}%PRINCIPAL VALUE
\newlength{\len}
\def\intp#1#2{{\renewcommand{\arraystretch}{0.4}%
\settowidth\len{$\displaystyle\int$}%
\begin{array}{c}%
\makebox[0pt]{\hspace{ \len}$\scriptstyle{#2}$}\\
\mbox{${\displaystyle{-\!\!\!\!\!\!\int}}$}\\%
\makebox[0pt]{\hspace{- \len}$\scriptstyle{#1}$}
\end{array}%
}\! }
%\intp defines the PrinVal integ with up and bott limits in displaymath
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\draft
\begin{document}
\title{Spectral Statistics Beyond Random Matrix Theory.
%Energy Level Correlations in Diffusive Chaotic Systems:
%Deviations from Universality.
}
\author{A.~V.~Andreev and B.~L.~Altshuler}
\address{
NECI, 4 Independence Way, Princeton, NJ\ 08540, USA,\\
Department of Physics, Massachusetts Institute of Technology, 77
Massachusetts Avenue, Cambridge, MA\ 02139, USA
}
%Department of Physics, Massachusetts Institute of Technology, 77
%Massachusetts Avenue, Cambridge, Massachusetts 02139 \\
\maketitle
%\medskip\\
%\parbox{14cm}{\rm
\begin{abstract}
Using a nonperturbative approach we examine the large frequency
asymptotics of the two-point level density correlator
in weakly disordered metallic grains.
This allows us to study the behavior of the two-level
structure factor close to the Heisenberg time. We find that the singularities
(present for random matrix ensembles) are washed out in a grain with
a finite conductance. The results are nonuniversal (they depend on the
shape of the grain and on its conductance), though they suggest
a generalization for any system with finite Heisenberg time.
%To our knowledge this is the
%first attempt of such an analysis. We also develop some new formalism
%in the supersymmetric nonlinear $\sigma$-model for disordered conductors.
%\smallskip\\
\end{abstract}
\pacs{PACS numbers:71.30.+h, 05.60.+w, 72.15.Rn}
\narrowtext
%\section{Introduction}
A great variety of physical systems are known to exhibit quantum
chaos. The common examples are atomic nuclei, Rydberg atoms in a
strong magnetic field, electrons in
disordered metals, etc \cite{LesHouches89}.
Chaotic behavior manifests itself in the energy level
statistics. It was a remarkable discovery of Wigner and Dyson,
that these statistics in a particular system can be approximated
by those of an ensemble of random matrices (RM). Here we consider
deviations from the RM theory taking an ensemble of weakly disordered
metallic grains with a finite conductance $g$ as an example.
The results seem to be extendible to general chaotic systems.
There are two characteristic energy scales associated with
a particular system: a classical one $E_c$ and a quantum one.
The quantum energy scale is the mean level spacing $\Delta$.
In a chaotic billiard, for example, $E_c$ is set by the
frequency of the shortest periodic orbit. Well developed chaotic
behavior can take place only if $E_c\gg \Delta$.
In a disordered
metallic grain the classical energy is the Thouless energy $E_c=D/L^2$,
where $D$ is the diffusion constant, and $L$ is the system size.
For a weakly
disordered grain the two scales are separated by
the dimensionless conductance
$g=E_c/\Delta \gg 1$ \cite{Thouless}.
For frequencies $\omega \ll E_c$ the behavior of
the system becomes universal (independent of particular parameters
of the system ).
In this regime in the zeroth approximation
the level statistics depend only
on the symmetry of the system and are
described by one of the RM ensembles:
unitary, orthogonal or symplectic \cite{Mehta91}.
One of the conventional statistical spectral characteristics is the
two-point level density correlator
\begin{equation}
K(\omega ,x)=
\langle \rho (\epsilon +\omega, \hat{H}+x\delta\hat{H} )
\rho (\epsilon , \hat{H})\rangle -\Delta^{-2},
\label{Kparam}
%\tag{Kparam}
\end{equation}
\noindent where
$\hat{H}$ is the Hamiltonian of the system,
$\delta\hat{H}$ is a perturbation, $x$ is the dimensionless
perturbation strength and
$\rho (\epsilon,\hat{H}+x\delta\hat{H})={\rm Tr}
\delta(\epsilon -\hat{H}-x\delta\hat{H})$
is the $x$-dependent density of states at
energy $\epsilon$. It is convenient to introduce
the dimensionless frequency $s= \omega/\Delta$ and the
dimensionless two-level correlator $R(s,x)=\Delta^2 K(\omega ,x)$.
Dyson \cite{Dyson} determined $R(s,x=0)$ for RM.
For example, $ R(s ,o) $ in the unitary case equals to
%
\begin{equation}
R(s ,0)
%= \delta(s)-
%\frac{\sin^2 (\pi s)}{(\pi s)^2}
= \delta(s)-
\frac{1-\cos(2 \pi s)}{2(\pi s)^2}
\label{Runivers}
%\tag{Runivers}
\end{equation}
%
and is plotted in the insert in Fig.~\ref{fig:1}.
Perhaps the most striking signature of the Wigner-Dyson
statistics is
the rigidity of the energy spectrum \cite{Haake}.
Among the major consequences of this phenomenon are:
a) the probability to find two levels separated by $\omega \ll \Delta$
vanishes as $\omega \to 0$;
b) the level number variance in an energy strip of width $N\Delta$
is proportional to $\ln N$ rather than $N$; and
c) oscillations in the
correlator $R(s , 0)$ in Eq.~(\ref{Runivers}) decay only algebraically.
In the two level structure factor \cite{Prigodin94}
$S(\tau, x)=\int^{\infty}_{-\infty} ds \exp(i\tau s) R(s , x)$
the reduced fluctuations of the level number manifest themselves
in the vanishing of $S(\tau, 0)$ at
$\tau=0$, and the algebraic decay of the oscillations in $R(s ,0)$
leads to the singularity in $S(\tau, 0)$ at the
Heisenberg time $\tau=2\pi$.
In the unitary case, e.~g.
\[ S(\tau, 0)={\rm min}\{|\tau|/(2\pi), 1\}
%\left\{
%\begin{array}{lc}
%|\tau|/(2\pi), & |\tau|\leq 2\pi \\
%1,& |\tau| > 2\pi
%\end{array}
%\right.
.\]
\noindent At $\tau \ll 2\pi $ this Dyson result
was obtained by Berry \cite{Berry85} for a generic chaotic
system by use of a
semiclassical approximation.
To the best of our knowledge nobody succeeded in analyzing the
behavior of $S(\tau, 0)$ around $\tau=2\pi$ using this formalism.
Wigner-Dyson statistics become exact in the limit $g=E_c/\Delta \to \infty$.
We consider corrections to these statistics for finite $g$.
One of the better understood systems in this respect is a
weakly disordered metallic grain.
For frequencies much smaller than $E_c$ the statistics are close
to universal ones, the corrections being small as $(s/g)^2$
\cite{Kravtsov94}. At $s\gg 1$ the monotonic part
of $R(s,x)$ can be
obtained perturbatively \cite{Altshuler85}
%
\begin{equation}
\label{Rpert}
% R(s,x)=
R_{{\rm p}}(s,x)=
\Re\sum_{\mu}
\frac{1}{\alpha \pi^2(-is+ x^2 + \epsilon _\mu)^2},
%g \pi^2 \vec{n}^2
%\tag{Rpert}
\end{equation}
\noindent where $\epsilon_\mu$ are eigenvalues (in units of $\Delta$)
of the
diffusion equation in the grain, $\alpha=2$ for the
unitary ensemble and $\alpha=1$ for the orthogonal and symplectic
ensembles \cite{footnote}.
At this point we can define
%
\begin{equation}
\label{gdef}
E_c=\epsilon_1 \Delta/\pi^2, {\qquad} g=\epsilon_1/\pi^2,
%\tag{gdef}
\end{equation}
%
where $\epsilon_1$ is the smallest nonzero eigenvalue.
Perturbation theory allows one to
determine $S(\tau, 0)$ at small times $\tau \ll 1$.
Since the oscillatory part of $R(s,x)$ is non-analytic in $1/s$
it can not be obtained perturbatively.
In this Letter we
obtain the leading $s \gg 1$ asymptotics of $R(s , x)$
retaining the oscillatory terms
and monitor how the singularity in $S(\tau, 0)$
at the Heisenberg time is modified by the finite conductance $g$.
We make use of the nonterturbative approach
\cite{Efetov83} that is valid for arbitrary relation between
$s$ and $g$.
The oscillatory part $R_{\rm osc}(s,x)\equiv R(s,x)-R_{\rm p}(s,x)$
for the unitary (${\rm u}$), orthogonal (${\rm o}$) and
symplectic (${\rm s}$) cases equals to
%For a cubic sample in dimension $d$ with
%closed boundary conditions we obtain
\begin{mathletters}
\label{allresults}
%\tag{allresults}
\begin{eqnarray}
R_{\rm osc}^{\rm u}(s, x)&=&
\frac{\cos(2\pi s)}{2\pi^2y^2}P(s,x), \label{resultu}
%\tag{resultu}
\\
R_{\rm osc}^{\rm o}(s, x)&=&
-\frac{\cos(2\pi s)}{2\pi^4y^4}P^2(s,x), \label{resulto}
%\tag{resulto}
\\
R_{\rm osc}^{\rm s}(s, x)&=&
\frac{\cos(\pi s)}{2y}P(s,x)
-\frac{\cos(2\pi s)}{2\pi^4y^2}P^2(s,x),
\label{results}
%\tag{results}
\end{eqnarray}
\end{mathletters}
%
where $y^2=s^2+x^4$, and $P(s,x)$ is
the spectral determinant of the diffusion operator
%
\begin{equation}
P(s,x)= \prod_{\mu, \epsilon_\mu
\neq 0}\left[ \left(\frac{s}{\epsilon_\mu}\right)^2
+\left(1+\frac{x^2}
{\epsilon_\mu}\right)^2\right]^{-1}.
\label{product}
%\tag{product}
\end{equation}
\noindent Note that Eq.~(\ref{Rpert}) expresses
$R_p(s,x)$ through the Green function of this
operator. Thus, regardless of the spectrum $\epsilon_\mu$,
$R_p(s,x)$ and $R_{\rm osc}(s, x)$ are related:
\begin{equation}
\label{relation}
R_{{\rm p}}(s,x)=
\Re\frac{1}{\alpha \pi^2(-is+ x^2)^2}-
\frac{1}{2 \alpha \pi^2}\frac{\partial^2 \ln[P(s,x)]}{\partial s^2}.
\end{equation}
It follows from Eq.~(\ref{product}) that $R_{\rm osc}(s, x)$
together with $P(s,x)$ decays exponentially at $s\gg g$.
As a result, the singularity in $S(\tau, 0)$ at the Heisenberg time
is washed out:
$S(\tau, 0)$ becomes {\it analytic} around $\tau=2\pi$.
The scale of smoothening of the singularity is
$1/E_c$ ( see Fig.~\ref{fig:1}).
At $1\ll s\ll g$ the sum of Eq.~(\ref{allresults}) and Eq.~(\ref{Rpert})
gives the leading high frequency asymptotics
of the universal result, for $s\gg g$ it coincides with the perturbative
result $R_{{\rm P}}(s,x)$ of Ref.~\cite{Altshuler85}.
In a closed ${\rm d}$-dimensional cubic sample
(diffusion equation with Dirichlet boundary conditions)
$\epsilon_{\mu}=g\pi^2 \vec{n}^2$, where $\vec{n}=(n_1,\ldots, n_d)$
and $n_i$ are non-negative integers.
For $s\gg g$ and $d<4$
we obtain the asymptotics
$P(s,0)\to \exp\{-\pi(s/\pi g)^{d/2}
/[\Gamma (d/2)d \sin(\pi d/4)]\} $.
At $1\ll s\ll g$ we obtain
\begin{equation}
R(s, 0) = -\frac{\sin^2(\pi s)}{(\pi s)^2}+
\frac{\sin^2(\pi s)}{\pi^2 g^2} \sum_{\vec{n}}
\frac{1}{(\pi^2 \vec{n}^2)^2},
\label{kkravtsov}
%\tag{kkravtsov}
\end{equation}
\noindent
This result was shown in Ref.~\cite{Kravtsov94} to be valid
even for $s < 1$.
Thus, it is natural to assume that for the unitary ensemble the
sum of Eq.~(\ref{resultu}) and Eq.~(\ref{Rpert}) gives
the correct $g \to \infty$ asymptotics
at arbitrary frequency. This is related to the
absence of higher order corrections to the leading term of
the perturbation theory $S(\tau, 0)\propto \tau$
in the unitary case.
%\section{$\sigma$-model for disordered conductors at high frequency.}
Now we sketch the derivation of our results.
Consider a quantum particle
moving in a random potential $V(\vec{r})$.
The perturbation acting on the system is a
change in the potential $\delta V(\vec{r})$.
Both $V(\vec{r})$ and $\delta V(\vec{r})$ are taken to be white noise
random potentials with variances
$\langle V(\vec{r})V(\vec{r'})\rangle =\delta (\vec{r}-\vec{r'})
/2\pi \nu \tau $
and $\langle \delta V(\vec{r})\delta V(\vec{r}')\rangle =
x^2 \Delta \delta (\vec{r}-\vec{r'})/(4\pi \nu )$,
$\Delta \tau\ll 1$, $\langle V(\vec{r}) \delta V(\vec{r}')\rangle =0$,
where $\langle \rangle $ denotes ensemble averaging and $\nu$ is
the density of states per unit volume.
The dimensionless perturbation strength $x^2$
is assumed to be of order unity.
We use the supersymmetric nonlinear $\sigma$-model
introduced by Efetov \cite{Efetov83}, and follow his notations everywhere.
One can show that for the system under consideration the $\sigma$-model
expression for $K(\omega ,x)$ is given by
\begin{mathletters}
\label{allfunc}
%\tag{allfunc}
\begin{eqnarray}
K(\omega, x) & = &\left. -\frac{1}{\pi^2}\Re
\frac{\partial ^2 }{\partial J ^2}\int {\cal D} Q
\exp(-F_J)\right|_{J=0}
\label{kfunc},
%\tag{kfunc}
\\
F_J&=&\frac{\pi \nu}{8} \int d\vec{r}
{\rm STr}\left\{
D(\nabla Q)^2 +2i \omega \Lambda Q +i J \Lambda k Q-
\frac{x^2\Delta}{2} (\Lambda Q)^2 \right\} .
\label{Zfunc}
%\tag{Zfunc}
\end{eqnarray}
\end{mathletters}
\noindent The $8\times 8$ supermatrix $Q(\vec{r})$ obeys the nonlinear
constraint $Q(\vec{r})^2=1$ and takes on its values on a certain
symmetric space ${\bf H}={\bf G}/{\bf K}$, where ${\bf G}$ and ${\bf K}$
are groups \cite{Weidenmuller85}. For example, for the unitary ensemble
${\bf H}={\bf SU(1,1/2)}/{\bf SU(1/1)\otimes SU(1/1)}$ \cite{Zirnbauer86}.
The integration measure for $Q$
in the functional integral Eq.~(\ref{kfunc}) is the invariant
measure on ${\bf H}$.
The hierarchy of blocks of supermatrices is as follows:
advanced-retarded (A-R) blocks, fermion-boson (F-B) blocks, and
blocks corresponding to time-reversal.
$\Lambda={\rm diag}\{ 1, 1, 1, 1, -1, -1, -1, -1\}$ is the
matrix breaking the symmetry in the advanced-retarded (A-R ) space,
$k={\rm diag}\{1,1,-1,-1, 1,1,-1,-1\}$ is the symmetry breaking matrix
in the Fermion-Boson ( F-B ) space.
The large frequency asymptotics of
$K(\omega, x)$ can be obtained from Eq.~(\ref{kfunc})
by use of the stationary phase method.
The conventional perturbation theory corresponds to integrating over the
small fluctuations of $Q$ around $\Lambda$ \cite{Efetov83},
\begin{equation}
Q=\Lambda
(1+iP)(1-iP)^{-1}, P= \left(\begin{array}{cc}
0 & B \\
\bar{B} & 0
\end{array}\right),
\label{parametrization}
%\tag{parametrization}
\end{equation}
\noindent where the matrix $P$ describes these small fluctuations.
$Q=\Lambda$
is not the only stationary point on ${\bf H}$.
This fact to the best of our knowledge was not appreciated in the
literature.
The existence of other stationary points makes the basis for our
main results.
It is possible to parameterize fluctuations around a point $Q_0$
in the form $Q=Q_0(1+iP_0)(1-iP_0)^{-1}$. Expanding the Free Energy
$F_J$ in Eq.~(\ref{Zfunc}) in $P_0$ we would obtain the stationarity
condition $\partial F_J/\partial P_0 =0$.
This route however
is inconvenient because the parametrization of $P_0$ will depend on
$Q_0$.
Instead we perform a global coordinate transformation on ${\bf H}$
that maps $Q_0$ to $\Lambda$,
$Q_0\rightarrow T_0^{-1} Q_0 T_0=\Lambda$.
We note that the matrices $\Lambda$ and $-\Lambda k$
belong to ${\bf H}$, and the corresponding terms
in Eq.~(\ref{Zfunc}) can be viewed as symmetry
breaking sources.
This transformation changes the sources, but allows us to
keep the parametrization of Eq.~(\ref{parametrization})
and preserves the invariant measure. Introducing the notation
$Q_{\Lambda}=T_0^{-1}\Lambda T_0$ and
$Q_{\Lambda k}=T_0^{-1}\Lambda k T_0$ we rewrite Eq.~(\ref{allfunc}) as
\begin{eqnarray}
K(\omega, x) & = & \frac{\nu ^2}{64}\Re
\int {\cal D} Q
\int d\vec{r} \int d\vec{r}' {\rm STr}\Big(Q_{\Lambda k}
Q(\vec{r})\Big) {\rm Str}\Big(Q_{\Lambda k} Q(\vec{r}')\Big)
\exp\left(-F(Q_{\Lambda})\right),
\nonumber \\
F(Q_{\Lambda}) & = & \frac{\pi \nu}{8} \int d\vec{r}
{\rm STr}\left\{
D(\nabla Q)^2 +2i \omega Q_{\Lambda} Q - \frac{x^2\Delta}{2} (
Q_{\Lambda} Q)^2 \right\} .
%\tag{kfuncsource}
\label{kfuncsource}
\end{eqnarray}
The stationarity condition
%\begin{equation}
% \label{stationpt}
$ \left. \partial F(Q_{\Lambda})/\partial P\right|_{P=0}=0 $
% \tag{stationpt}
%\end{equation}
%
%\noindent
implies that all the elements
of $Q_{\Lambda}$ in the AR and RA blocks should vanish (this can be
seen from Eq.~(\ref{parametrization})).
Here we discuss in detail only the calculation for the unitary ensemble.
The calculation for the other cases proceeds
analogously, and we just point out the important
differences from the unitary case.
%\subsection{unitary ensemble.}
%\label{sec:unitary}
Consider now the unitary case.
The $4\times 4$ supermatrices $B$ and $\bar{B}$
in Eq.~(\ref{parametrization}) are given by
\[ B= \left(\begin{array}{cccc}
a & 0 & i\sigma _1 & 0\\
0 & a^* & 0 & -i\sigma _1^* \\
\sigma _2^* & 0 & ib & 0 \\
0 & \sigma _2 & 0 & ib^*
\end{array}\right),
\bar{B}=\left(\begin{array}{cccc}
a^* & 0 & -\sigma _2 & 0\\
0 & a & 0 & \sigma _2^* \\
-i\sigma _1^* & 0 & ib^* & 0 \\
0 & -i \sigma _1 & 0 & ib
\end{array}\right),
\]
\noindent where $a$, $b$ and their conjugates are ordinary variables,
and $\sigma_1$, $\sigma_2$ and their conjugates are grassmann variables.
The only matrix besides $\Lambda$
that satisfies the stationarity condition is
$Q_{\Lambda}=
\tilde{\Lambda}=-k \Lambda$.
In this case $Q_{ \Lambda k}= -\Lambda$.
All other matrices from ${\bf H}$
contain nonzero elements in the AR and RA blocks.
Both stationary points contribute substantially to
$K(\omega, x)$.
Consider the contribution of
$Q_{\Lambda}=\tilde{\Lambda}$ to $K(\omega, x)$ first.
We substitute $Q_{\Lambda}=-k\Lambda$ and $Q_{ \Lambda k}= -\Lambda$
into Eq.~(\ref{kfuncsource}) and expand
both the Free Energy $F(Q_\Lambda)$ and the pre-exponent to the second order
in $B$ and $\bar{B}$.
Expanding $B(\vec{r})$ in the eigenfunctions of the diffusion
operator $\phi_{\mu}(\vec{r})$ as
$B(\vec{r})=\sum_{\mu}\phi_{\mu}(\vec{r})B_{\mu}$, and
introducing ${\cal E}_\mu=is+\epsilon_\mu +x^2$
we arrive at the following expression for the dimensionless
density-density correlator:
\begin{eqnarray}
R^{\rm u}_{\rm osc}(s, x) & = &
\Re \int {\cal D} B_{\mu}
\Big(\sum_{\mu}[|a_{\mu}|^2+ |b_{\mu}|^2 -
\sigma_{1, \mu}^* \sigma
_{1, \mu}+\sigma_{2, \mu}^*\sigma_{2, \mu}]\Big)^2
\nonumber \\
& \times & \exp( - 2\pi \{-is +\sum_{\mu}[
{\cal E}_\mu|a_{\mu}|^2 +
{\cal E}^*_\mu|b_{\mu}|^2
- \Re{\cal E}_\mu (\sigma_{1, \mu}^* \sigma
_{1, \mu}-\sigma_{2, \mu}^*\sigma_{2, \mu})
] \}).
%\tag{kpt}
\label{kpt}
\end{eqnarray}
\noindent We have to keep the perturbation
strength $ x^2 $ finite to avoid the divergence of
the integral over $a_0$ caused by the
presence of the infinitesimal imaginary part in $\omega$.
For the non parametric case we should take the
$x^2\to 0$ limit after the integral
in Eq.~(\ref{kpt}) is evaluated.
Since the Free Energy in Eq.~(\ref{kpt})
contains no Grassmann variables in the
zero mode they have to come from the pre-exponent. Therefore out of the
whole square of the sum in the pre-exponent only the
terms containing all four zero mode Grassmann variables
contribute. In these terms the prefactor does not contain
any variables from non-zero modes. Thus,
the evaluation of the Gaussian integrals over non-zero modes
yields the superdeterminant of the quadratic form in the
exponent.
Supersymmetry around $\tilde{\Lambda}$ is broken by $s$,
therefore this superdeterminant differs from unity and is given by
$P(s,x)$ of Eq.~(\ref{product}).
The correctly ordered integration measure for Grassmann variables is
$\prod_{\mu} d\sigma_{1\mu} d\sigma_{1\mu}^*
d\sigma_{2\mu}^* d\sigma_{2\mu}$.
Evaluating the integral we arrive at Eq.~(\ref{resultu}).
In quasi-1D for closed boundary conditions and $x=0$
the spectral determinant $P(s,0)$ can be evaluated exactly,
and from Eq.~(\ref{resultu}) we obtain
\begin{equation}
R^{\rm u, osc}_{1D}(s,0) = \frac{s}{2g\pi^2s^2}
\frac{\cos( 2
\pi s) }{\sinh ^2 \left(\sqrt{\frac{s}{2g}}\right)+\sin ^2
\left(\sqrt{\frac{s}{2g}}\right)}.
%\tag{k1D}
\label{k1D}
\end{equation}
For $Q_{\Lambda}=\Lambda$ the same procedure as used above
leads to Eq.~(\ref{Rpert}), which coincides with
the result of Ref.~\cite{Altshuler85}.
%\subsection{Structure Factor $S(\tau,0)$ at the Heisenberg time.}
The behavior of $S(\tau,0)$ at $\tau=0$ and $\tau=2\pi$
is associated respectively with $R_{\rm p}(s,0)$ (Eq.~(\ref{Rpert}))
and $R^{\rm u}_{\rm osc}(s,0)$ (Eq.~(\ref{resultu})).
In other words the singularity
at the Heisenberg time is determined by the contribution
to $R(s,0)$ from ${\tilde{\Lambda}}$.
It is clear that the cusp in $S(\tau, 0)$ at $\tau =2\pi $
will be rounded off because $R^{\rm u}_{\rm osc}(s,0)$
decays exponentially at large $s$. The scale of the smoothening
is of order $1/g$.
The Fourier transform of Eq.~(\ref{k1D}) ( see Fig.~\ref{fig:1} ) is
%
\begin{equation}
S^{\rm u}_{1D}(2\pi +t, 0)_{\tilde{\Lambda}} =
\sum_{n=1}^{\infty}\frac{(-1)^n \exp(-\pi^2 n^2 g |t|)}{\pi^2 g
n \sinh(\pi n)}-\frac{|t|}{4\pi} .
\label{ktildefourier}
%\tag{ktildefourier}
\end{equation}
\noindent Even though $S^{\rm u}_{1D}(2\pi +t, 0)_{\tilde{\Lambda}} $
appears to be a function of $|t|$, it is regular at $t=0$.
We can also estimate
$S^{\rm u}(2\pi , 0)_{\tilde{\Lambda}}$
in any dimension. It is proportional to $1/g$ of Eq.~(\ref{gdef})
and is given by
\[ S^{\rm u}(2\pi , 0)_{\tilde{\Lambda}}
= \frac{1}{4 \pi^4 g} \int_{-\infty+i\eta}^{\infty+i\eta} \frac{d z}{ z^2}
\prod _{\mu \epsilon_\mu \neq 0}\left( 1+\left[\frac{z\epsilon_1}
{\epsilon_\mu}\right]^2\right)^{-1} .\]
%\noindent where the integration contour is displaced to the upper half
%plane by an infinitesimal amount.
%\subsection{Orthogonal Ensemble}
%\label{sec:orth}
Consider now T-invariant systems.
For the orthogonal ensemble there are still only two stationary points
on ${\bf H}$: $\Lambda$ and $\tilde{\Lambda}= -k \Lambda$.
To determine the contribution of the $\tilde{\Lambda}$-point
we use the formula Eq.~(\ref{kfuncsource})
with $Q_{\Lambda}=\tilde{\Lambda}$ and $Q_{k \Lambda }=-\Lambda$
and Efetov's parametrization for the perturbation theory \cite{Efetov83}.
The calculations are analogous to those for the
unitary ensemble and lead to Eq.~(\ref{resulto}).
The contribution of $Q_\Lambda=\Lambda$ gives Eq.~(\ref{Rpert}).
At $\tau=2\pi$ the third derivative of $S(\tau,0)$ for the orthogonal
ensemble has a jump. This
singularity also disappears at finite $g$.
%\subsection{symplectic ensemble}
%\label{sec:symplectic}
In the symplectic case there are three types of stationary
points which correspond to singularities in the
structure factor $S(\tau,0)$ at $\tau=0, \pi , 2\pi$ \cite{Mehta91}.
The $\tau=2\pi$ singularity corresponds to $Q_{\Lambda}=\tilde{\Lambda}=
-k\Lambda$, and
its contribution to $R(s, x)$,
given by the second term in Eq.~(\ref{results}), is
exactly the same as $R^{\rm o}_{\rm osc}(s,x)$.
The stationary point $Q_{\Lambda}=\Lambda$ corresponds
to the $\tau=0$ singularity in $S(\tau,0)$ and leads to Eq.~(\ref{Rpert}).
The $\tau=\pi$ singularity corresponds to a degenerate manifold
of matrices $Q_{\Lambda}$ on ${\bf H}$
$Q_\Lambda={\rm diag}(\tau_{\vec{m}},\openone_2, -\tau_{\vec{m}},
-\openone_2)$,
$ Q_{k\Lambda} =-k Q_\Lambda $,
where ${\openone_2}$ is a $2\times 2$ unit matrix,
$\tau_{\vec{m}}=m_x \tau_{x}+m_y\tau_{y}$,
$\vec{m}^2=1$ and $\tau_{x,y}$ are Pauli matrices in the time-reversal
block.
%
%\begin{equation}
% \label{lambdatilone}
% Q_{\Lambda} =\tilde{\Lambda}_{1}= \left(\begin{array}{cccc}
%\sigma_x & 0 & 0 & 0 \\
%0 & 1 & 0 & 0 \\
%0 & 0 & -\sigma_x & 0 \\
%0 & 0 & 0 & -1 \\
%\end{array}\right),\qquad
% Q_{k\Lambda} =k \tilde{\Lambda}_{1},
%\tag{lambdatilone}
%\end{equation}
%
%\noindent where $\sigma_x$ is the Pauli matrix in the space corresponding
%to time-reversal conjugation. {\bf
%In Efetov's parametrization \cite{Efetov83} $\tilde{\Lambda}_{1}$
%corresponds to
%$\theta_1=\theta_2=\pi/2$ and all other variables being equal to zero.
%The calculations proceed in the same manner as before.
%The contribution of these stationary points to $R(s, x)$ is
%given by Eq.~(\ref{results}). }
The calculation proceeds as before and leads to the first term in
Eq.~(\ref{results}).
In quasi-1D we can obtain the leading contribution to the structure factor
$S(\tau, 0)$ around $\tau=\pi$
%
\begin{equation}
\label{formfsympl}
S^{\rm s}(t+\pi,0)=\int_0^{\infty}
\frac{-4\sin^2(g|t|z)dz}{\sinh^2\sqrt{z}+\sin^2\sqrt{z}}+\ln(1.9 g)+ O(1/g).
%\tag{formfsympl}
\end{equation}
\noindent
The result is plotted in Fig.~\ref{fig:2}. In all dimensions the
logarithmic divergence
in the zero mode result is now cut off by finite $g$, and
$S^{\rm s}(\pi,0)\propto \ln g$.
%\section{Discussion of the results.}
In conclusion we mention several points about our results.
1) Equation (\ref{allresults}) describes the deviation of
the level statistics of a weakly
disordered chaotic grain from the universal ones.
This deviation is controlled by the diffusion operator. This
operator is purely classical.
It seems plausible that the nonuniversal part of spectral statistics
of any chaotic system can be expressed through a spectral determinant
of some classical system-specific operator. If so, the relation
Eq.~(\ref{relation}) should be universally correct!
%
%every chaotic system can be
%described by a certain operator whose spectrum will appear
%in the spectral determinant $P(s,x)$ of Eq.~(\ref{product}).
%The nature of this operator for ballistic systems is
%unclear at present.
%It would be interesting to see whether this operator can be
%described by means of the semiclassical approximation ({\bf something like
%Gutzwiller approach}).
%
%
2) The formalism used here should be
applicable even to the systems
weakly coupled to the outside world (say through tunnel contacts).
As long as the level broadening $\Gamma$ ($\omega=\Re\omega +i\Gamma$)
is smaller than $\Delta x^2$
the integration over the zero mode variables
in Eq.~(\ref{kpt}) is convergent. The integral over the other modes
is always convergent provided $\Gamma < E_c$.
Thus, the presence of a perturbation can effectively ``close'' a
weakly coupled system. Under these conditions Eq.~(\ref{allresults})
remains valid after the substitution $\cos(2\pi s)\to
\exp(-2\pi\Gamma/\Delta)\cos(2\pi s)$ and $x^2\to x^2-\Gamma/\Delta$.
3)
The classification of physical systems into the three universality
classes (unitary, orthogonal and symplectic) is, of course, an
oversimplification. In practice there is always
a time scale which determines the crossover from one
ensemble to another. For example if a system is subjected to
a magnetic field for very short times it
will still effectively remain orthogonal.
On the other hand, the long time behavior will be unitary.
The characteristic time is
set by the strength of the magnetic field.
For a disordered metallic grain in a magnetic field this
characteristic time is given by
$l^2_{H}/D$. For frequencies larger than $D/l^2_{H}$
the system effectively becomes orthogonal.
This implies that even if we neglect the spatially nonuniform
fluctuations of the $Q$-matrix the cusp in $S(\tau,0)$ at $\tau=2\pi$
will be washed out on the scale of $\Delta l_H^2/D$
( although there will still remain a jump in the third derivative of
$ S(\tau, 0)$ ). For the system to behave as unitary for frequencies
of order $E_c$ the magnetic length $l_H$ has to be
shorter than the size of the
system. Spin-orbit interaction that causes the orthogonal-to-symplectic
crossover can be considered in a similar way.
4) The rounding off of the singularity in $S(2\pi,0)$ is also
present in the random matrix model with preferred basis \cite{Pichard93}
\cite{Shapiro94}. Note that our results differ from those in
Ref.~\cite{Shapiro94} substantially. This means that finite $g$ is not
equivalent to finite temperature for the corresponding
Calogero-Sutherland model \cite{Simons93}.
We are grateful to D.~E.~Khmel'nitskii, B.~D.~Simons and N.~Taniguchi
for numerous discussions
throughout the course of this work.
\begin{references}
\bibitem{LesHouches89} Chaos in Quantum Physics,
eds., M.~J.~Jianonni, A.~Voros and J.~Zinn-Justin, Les Houches,
Session LII 1989 (North-Holland, Amsterdam, 1991)
\bibitem{Thouless} D.~J.~Thouless, Phys. Rep. {\bf 13}, 93 (1974)
\bibitem{Mehta91} M.~L.~Mehta, Random Matrices, (Academic Press,
New York, 1991)
\bibitem{Dyson} F.~J.~Dyson, J. Math. Phys. {\bf 3}, 140, 157, 166
(1962)
\bibitem{Haake}F.~Haake, Quantum Signatures of Chaos (Springer, Berlin,
1991)
\bibitem{Prigodin94} The function $S(\tau,0)$, for example, describes
the phenomenon of ``quantum echo'' in mesoscopic quantum dots.
See V.~N.~Prigodin et al, Phys. Rev. Lett. {\bf 72}, 546 (1994)
\bibitem{Berry85}M.~V.~Berry, Proc. Roy. Soc. London {\bf A 400},
229 (1985)
\bibitem{Kravtsov94}V.~E.~Kravtsov, A.~D.~Mirlin, Sov. Phys. JETP Lett.,
{\bf 60}, 656 (1994). [Pis'ma ZhETF, {\bf 60}, 645 (1994) ]
\bibitem{Altshuler85}B.~L.~Altshuler, B.~I.~Shklovskii, JETP
{\bf 64}, 127 (1986).
\bibitem{footnote}This definition of $\alpha$ is related to the fact that
$\Delta$ for the unitary and orthogonal cases
is the mean level spacing per spin polarization, and for the
symplectic case it is one-half of that.
\bibitem{Efetov83} K.~B.~Efetov Adv. Phys. {\bf 32}, 53 (1983).
\bibitem{Weidenmuller85} J.~J.~Verbaarscot, H.~A.~Weidenmuller,
M.~R.~Zirnbauer, Phys. Rep. {\bf 129}, 367 (1985)
\bibitem{Zirnbauer86} M.~R.~Zirnbauer, Nucl. Phys.{\bf B [FS] 265},
375 (1986)
\bibitem{Pichard93} J.~-L.~Pichard, B.~Shapiro J. Phys. May(1994)
\bibitem{Shapiro94}M.~Moshe, H.~Neunberger, B.~Shapiro,
preprint RU-94-28, Technion PH-12-94 (1994)
\bibitem{Simons93} B.~D.~Simons, P.~A.~Lee and B.~L.~Altshuler,
Nucl. Phys.{\bf B 409}, 487 (1993)
%\bibitem{Wegner79}F.~?.~Wegner
%{\bf 64}, 127 (1986).
\end{references}
\begin{figure}
\begin{center}
\epsfxsize=12cm \epsfbox{splot12.eps}
\end{center}
\caption{Structure factor
in quasi-1D case for unitary symmetry (solid line)
and the universal structure factor (dashed line).
Insert: the two level correlator
as a function of level separation.}
\label{fig:1}
\end{figure}
\begin{figure}
\begin{center}
\epsfxsize=15cm \epsfbox{symplcom4.eps}
\end{center}
\caption{The structure factor for the
symplectic case in quasi-1D (solid line) and
the universal result (dashed line).
Insert: the universal stucture factor.}
\label{fig:2}
\end{figure}
\end{document}