Persistence of Manifolds in Nonequilibrium Critical Dynamics
Extracted source text ยท arXiv cond-mat/0302537
%\documentstyle[aps,prl,multicol,psfig]{revtex}
\documentstyle[prl,aps,epsf,multicol]{revtex}
\begin{document}
\title{\bf Persistence of Manifolds in Nonequilibrium Critical Dynamics}
\author{Satya N. Majumdar$^1$ and Alan J. Bray$^{1,2}$}
\address{$^1$Laboratoire de Physique Quantique (UMR C5626 du CNRS),
Universit'e Paul Sabatier, 31062 Toulouse Cedex, France \\
$^2$Department of Physics and Astronomy, University of
Manchester, Manchester M13 9PL, UK}
\date{\today}
\maketitle
\begin{abstract}
We study the persistence $P(t)$ of the magnetization of a $d'$
dimensional manifold (i.e.\ the probability that the manifold
magnetization does not flip up to time $t$, starting from a random
initial condition) in a $d$-dimensional spin system at its critical
point. We show analytically that there are three distinct late time
decay forms for $P(t)$ : exponential, stretched exponential and power
law, depending on a single parameter $\zeta=(D-2+\eta)/z$ where
$D=d-d'$ and $\eta$, $z$ are standard critical exponents. In
particular, our theory predicts that the persistence of a line
magnetization decays as a power law in the $d=2$ Ising model at its
critical point. For the $d=3$ critical Ising model, the persistence of
the plane magnetization decays as a power law, while that of a line
magnetization decays as a stretched exponential. Numerical results
are consistent with these analytical predictions.
\noindent
\medskip\noindent {PACS numbers: 05.70.Ln, 05.50.+q, 05.70.Jk}
\end{abstract}
\begin{multicols}{2}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%\section{Introduction}
Following a rapid quench of a spin system from infinite temperature to
zero temperature, domains of competing ground states form and grow
with time\cite{Bray}. In the standard nonconserved dynamics a given
spin inside the sample flips only when a domain wall passes through
it, which happens rather rarely. As a result, persistence, i.e.\ the
probability that the spin remains unflipped up to time $t$ decays
slowly as a power law, $P(t)\sim t^{-\theta}$ at late
times\cite{Review}. In contrast, if the spin system is quenched to
its critical temperature $T_c$, any given spin fluctuates rapidly due
to the finite temperature and the persistence of a single spin has an
exponential tail at late times. However, for a quench to $T_c$, the
persistence of the total magnetization of the sample (as opposed to
that of a single spin) again decays as a power law, $P(t)\sim
t^{-\theta_g}$, where the exponent $\theta_g$ has been argued to be a
new nonequilibrium critical exponent\cite{MBCS}. The global
persistence has since been studied in a wide variety of
systems\cite{globalp}.
In a $d$-dimensional sample, a single spin is a $0$-dimensional
manifold. On the other hand, the global magnetization corresponds to
summing over all the spins of the full sample which is a
$d$-dimensional manifold. To interpolate between these two limits, it
is natural to study the persistence of the magnetization of a
$d'$-dimensional manifold with $0 \leq d'\leq d$. In the two limits
$d'=0$ and $d'=d$, the persistence has very different asymptotic
decay, respectively exponential and power law. The following question
then naturally arises: if one tunes the manifold dimension $d'$ from
$d'=0$ to $d'=d$, how does the asymptotic behavior of persistence
changes from an exponential ($d'=0$) to a power law decay ($d'=d$)?
Does this change occur abruptly at some intermediate value of $d'$, or
is there an intermediate regime of $d'$ where the behavior is neither
exponential nor power law, but something in between? In this Letter we
address this interesting issue and we show analytically that indeed
there is a novel intermediate regime of $d'$ where the persistence has
a stretched exponential tail. Our main results can be summarized in
terms of a single number $\zeta= (D-2+\eta)/z$, where $D=d-d'$ is the
co-dimension of the manifold and $\eta$ and $z$ are the standard
critical exponents: $z$ is the dynamical exponent that describes the
temporal growth of the correlation length $\xi(t)\sim t^{1/z}$ and
$\eta$ describes the power law decay of the equilibrium spin-spin
correlation function at $T_c$, $\langle \phi (0)\phi(r)\rangle \sim
r^{-(d-2+\eta)}$ for large separation $r$. Depending on the value of
$\zeta$, the persistence of the magnetization of a $d'$-dimensional
manifold at $T_c$ has the asymptotic behavior,
\begin{eqnarray}
P(t) &\sim & t^{-\theta(d',d)}, \quad\quad\quad\quad\, \zeta<0
\nonumber \\ &\sim & \exp\left(-a_1 t^{\zeta}\right), \,\,\,\,
0\leq \zeta\leq 1 \nonumber \\ & \sim & \exp\left(-b_1 t\right),
\quad\quad\quad \zeta>1 ,
\label{asymp1}
\end{eqnarray}
where the exponent $\theta(d',d)$ depends on $d'$ and $d$ and $a_1$,
$b_1$ are constants. Strictly speaking, in the intermediate regime
$0\leq \zeta\leq 1$ we will show that, $\exp\left(-a_2 t^{\zeta}\ln
t\right)\leq P(t) \leq \exp\left(-a_1 t^{\zeta}\right)$ for large
$t$. We will derive the results in Eq.\ (\ref{asymp1}) within the
mean-field theory (valid for $d>4$), in the $n\to \infty$ limit of the
$O(n)$ model, followed by a general scaling theory.
Two specific applications of our general results are as follows: (i)
Consider the persistence of the line magnetization ($d'=1$) in the
$d=2$ Ising model at $T_c$. Using $d'=1$, $d=2$, $\eta=1/4$ and
$z\approx 2.172$, one gets $\zeta=(d-d'-2+\eta)/z\approx -0.3453<0$.
Hence Eq.\ (\ref{asymp1}) predicts a power law decay for the
persistence of the line magnetization; (ii) For the $d=3$ Ising model,
using $\eta\approx 0.032$ and $z\approx 2$, one finds that while for
the plane magnetization ($d'=2$), $\zeta\approx -0.484 <0$, for the
line magnetization ($d'=1$), $0<\zeta\approx 0.016<1$. Our Eq.\
(\ref{asymp1}) then predicts a power law decay for the persistence of
the plane magnetization, but a stretched exponential decay for the line
magnetization. Numerical simulations for small samples in $2$ and $3$
dimensions are consistent with these analytical predictions.
Our starting point is the standard Langevin equation for the vector
order parameter $\vec \phi=(\phi_1,\ldots, \phi_n)$,
\begin{equation}
\partial_t\phi_i=\nabla^2 \phi_i -r\phi_i -(u/n)\phi_i \sum_{j=1}^n
\phi_j^2 +\eta_i,
\label{Lange1}
\end{equation}
where ${\vec \eta} ({\bf x}, t)$ is a Gaussian white noise with mean
zero and correlator $\langle \eta_i({\bf x},t)\eta_j({\bf
x'},t')\rangle=2\delta_{i,j}\delta^d({\bf x}-{\bf
x'})\delta(t-t')$. The magnetization of a $d'$-dimensional manifold
will be defined by the vector field, $\psi_i (x_{d'+1},\ldots, x_d,t)
=\int \phi_i(\vec x ,t)\prod_{i=1}^{d'}{ {dx_i}\over {\sqrt L} }$,
obtained by integrating the order parameter over the $d'$
directions. Here $L$ denotes the length of the sample along any
direction. For a vector order parameter, we define the persistence
$P(t)$ of the manifold to be the probability that any given component
of the manifold magnetization, say $\psi_1$, does not change sign up
to time $t$. Since all the components of the spin are equivalent,
henceforth we will drop the subscript $i$ of $\psi_i$ for convenience.
Note that the magnetization $\psi$ is a field over the remaining
$D=d-d'$ dimensional space whose coordinates $(x_{d'+1},\ldots, x_d)$
will be relabelled by the vector $\vec r= (r_1,r_2,\ldots r_D)$ for
convenience. The observation that allows us to make analytical
predictions for the persistence of the magnetization $\psi(\vec r,t)$
is that it is a Gaussian variable at all finite times. This follows
simply from the fact that $\psi(\vec r, t)$ is a sum of $L^{d'}$
random variables which are correlated but only over a {\em finite}
correlation length $\xi(t)\sim t^{1/z}$. Thus in the thermodynamic
limit when $t^{1/z}\ll L$, the central limit theorem asserts that
$\psi(\vec r ,t)$ is a Gaussian field. Hence its persistence is
determined by the autocorrelation function $C(t_1,t_2)=\langle
\psi(\vec r,t_1)\psi(\vec r,t_2)\rangle$\cite{Review}.
We start with the mean-field theory, valid for $d\geq 4$, where we set
$u=0$, and also $r=0$ (at the critical point) in
Eq.\ (\ref{Lange1}). Next we integrate the Langevin equation over the
$d'$ space directions and then solve the resulting linear equation in
the Fourier space. Defining ${\tilde \psi}(\vec k, t)=\int \psi(\vec
r, t)e^{i {\vec k}\cdot{\vec r}}d^Dr$, it is easy to compute the
two-time correlation function,
\begin{eqnarray}
\langle {\tilde \psi}(\vec k, t_1) {\tilde \psi}(-\vec k, t_2)\rangle
&=& \Delta (\vec k) e^{-k^2(t_1+t_2)}\nonumber \\ &+& {1\over
{k^2}}\left(e^{-k^2|t_1-t_2|}-e^{-k^2(t_1+t_2)}\right),
\label{corrk1}
\end{eqnarray}
where $\Delta(\vec k)=\langle {\tilde \psi}(\vec k, 0) {\tilde
\psi}(-\vec k, 0)\rangle$ is taken to be a constant (for uncorrelated
initial condition). At late times, the initial condition dependent
term becomes negligible and it is sufficient to retain only the
second term on the right-hand side of Eq.\ (\ref{corrk1}). The
autocorrelation function is then obtained by integrating over the
$\vec k$ space, $C(t_1,t_2)=\int d^Dk\,e^{-k^2a^2} \langle {\tilde
\psi}(\vec k, t_1) {\tilde \psi}(-\vec k, t_2)\rangle$ where we have
introduced a soft ultraviolet cut-off $a$. One finds
\begin{equation}
C(t_1,t_2) = \beta A\left[(t_1+t_2+a^2)^{2\beta}
-(|t_1-t_2|+a^2)^{2\beta}\right],
\label{corr2}
\end{equation}
where $2\beta=(2-D)/2$ and $A$ is an unimportant constant.
Consider first the case $D<2$, i.e., $\beta>0$. In this case there is
no need for the ultraviolet cut-off since $\langle \psi^2(\vec r,
t)\rangle =C(t,t)$ does not diverge at any finite time even if one
puts $a=0$. Setting $a=0$ in Eq.\ (\ref{corr2}) we find that the
correlation function $C(t_1,t_2)$ is still non-stationary. However,
one can render it stationary by employing a well known
trick\cite{Review}, where one introduces a normalized Gaussian
process, $X=\psi/\sqrt{\langle \psi^2\rangle}$ which, when observed in
the logarithmic time $T=\ln t$ becomes a stationary Gaussian variable
with the correlator, $C(T)=\langle X(0)X(T)\rangle ={\cosh
(T/2)}^{2\beta}-|\sinh(T/2)|^{2\beta}$. Interestingly, the same
Gaussian correlator also appears in the context of the persistence
of the Edwards-Wilkinson type rough interfaces\cite{interface}.
It is well known\cite{Slepian,Review} that
for such a stationary Gaussian correlator (decaying exponentially for
large $T$), the persistence of the process also decays exponentially
for large $T$, $P(T)\sim \exp(-\theta T)$. In terms of the real time
$t=e^T$, this indicates a power law decay, $P(t)\sim t^{-\theta}$,
with persistence exponent $\theta$. For this particular
correlator, the exponent $\theta$ has been studied in great detail in
the context of the interface problem\cite{interface} and the exponent
is known to depend continuously on the roughness parameter
$\beta=(2-D)/4 >0$.
For $D>2$, on the other hand, one needs the ultraviolet cut-off $a$
explicitly in order to keep $\langle \psi^2(\vec r, t)\rangle=C(t,t)$
finite. In that case, the appropriate scaling limit is $t_1, t_2\to
\infty$, but keeping their difference $|t_1-t_2|$ fixed. In this
limit, Eq.\ (\ref{corr2}) reduces to a stationary correlator in the
original time variable, $ C(t_1,t_2)\sim (|t_1-t_2|+a^2)^{-(D-2)/2}$
that decays as a power law for large $|t_1-t_2|$. To calculate the
persistence of Gaussian stationary processes with an algebraically
decaying correlator is nontrivial. However, there exists a powerful
theorem due to Newell and Rosenblatt\cite{NR}, which states that if
the stationary correlator decays as $C(t)\sim t^{-\alpha}$ with
$\alpha>0$ for large time difference $t=|t_1-t_2|$, then the
persistence $P(t)$ (probability of no zero crossing between $t_1$ and
$t_2$) of such a process has the following asymptotic behaviors: (i)
$P(t)\sim \exp(-K_1 t)$ if $\alpha>1$ and (ii) $\exp(-K_2
t^{\alpha}\ln t) \leq P(t)\leq \exp(-K_3 t^{\alpha})$ if $0< \alpha
<1$, where the $K_i$'s are constants. Applying this theorem to our
problem, we find that $P(t)\sim \exp(-K_1 t)$ for $D>4$ and $
\exp(-K_2 t^{(D-2)/2}\ln t) \leq P(t)\leq \exp(-K_3 t^{(D-2)/2})$ for
$2< D <4$. In the borderline case $D=4$, there will be an additional
logarithmic correction.
Combining these results for $D<2$ and $D>2$ and noting that $z=2$ and
$\eta=0$ within the mean-field theory, we find that the explicit exact
results for the mean-field theory derived above are just the special
cases of the general result in Eq.\ (\ref{asymp1}) provided one uses
the mean-field value $\zeta=(D-2)/2$ in Eq.\ (\ref{asymp1}).
The mean-field theory is valid for $d\geq 4$. In order to access the
physically relevant dimensions $d\leq 4$, we now consider another
solvable limit, namely the $n\to \infty$ limit where Eq.\
(\ref{Lange1}) becomes
\begin{equation}
\partial_t\phi_i=\nabla^2 \phi_i -[r+S(t)]\phi_i +\eta_i,
\label{Lange2}
\end{equation}
where $S(t)=u\langle \phi_i^2\rangle$ has to be determined
self-consistently. The critical point corresponds to
$r+S(\infty)=0$. This self-consistent determination of $S(t)$ can be
done using standard techniques\cite{MBCS} and one finds that at late
times, $S(t)\to S(\infty)- (4-d)/{4t}$ for $2< d \leq 4$. Substituting
this result into Eq.\ (\ref{Lange2}), summing over the $d'$
directions, solving the resulting equation in the Fourier space and
finally integrating over the $\vec k$ space (as in the mean-field
theory), we finally arrive at the following autocorrelation function
for the manifold magnetization $\psi(\vec r,t)$ in $2< d \leq 4$,
\begin{equation}
C(t_1,t_2)= A_1 (t_1t_2)^{(4-d)/4}\int_0^{t_1} {{dt'\,{t'}^{(d-4)/2}
}\over {(t_1+t_2-2t'+a^2)^{D/2}} },
\label{auto1}
\end{equation}
where $t_1\leq t_2$, $A_1$ is a constant and $a$ represents the soft
ultraviolet cut-off as before.
For $D<2$, as in the mean-field theory, one can set the cut-off $a=0$
and the resulting nonstationary correlator in Eq.\ (\ref{auto1}) can be
reduced to a stationary correlator for the normalized process
$X=\psi/\sqrt{\langle\psi^2\rangle}$ in the logarithmic time $T=\ln t$,
\begin{equation}
A(T)=\left[\cosh(T/2)\right]^{\mu-D/2}
{{B\left[\mu,2\beta,2/(1+e^T)\right]}\over{B\left[\mu,2\beta\right]}},
\label{auto2}
\end{equation}
where $\mu=(d-2)/2$, $2\beta=(2-D)/2$, $B[m,n]$ is the standard Beta
function and $B[m,n,x]=\int_0^x dy y^{m-1}(1-y)^{n-1}$. Since the
stationary correlator in Eq.\ (\ref{auto2}) decays exponentially for
large $T$, $A(T)\sim \exp[-(d+D-2)T/4]$, one concludes
\cite{Slepian,Review} that the corresponding persistence also
decays exponentially for large $T$, $P(T)\sim \exp(-\theta T)$ and
hence as a power law in the original $t=e^T$ variable, $P(t)\sim
t^{-\theta}$. Determining the exponent $\theta$ analytically is still
a challenging task. However one can make progress in the limit of
small co-dimension $D\to 0$. Note that for $D=0$, i.e.\ for the global
persistence, Eq.\ (\ref{auto2}) becomes a pure exponential $A(T)=
\exp[-(d-2)T/4]$ for all $T$, indicating that the process is
Markovian\cite{MBCS}. One then finds $P(T)\sim \exp(-\theta_0 T)$ with
$\theta_0=(d-2)/4$\cite{MBCS}. For $D$ nonzero but small, one can
expand the correlator in Eq.\ (\ref{auto2}) around the Markov process
($D=0$) and then use a perturbation theory result\cite{perturb} to
calculate $\theta$ to first order in $D$. We get $\theta= \theta_0
+ D \theta_0^2 I_d/\pi +O(D^2)$ for all $2<d\leq 4$, where $I_d$ is a
complicated $d$-dependent integral. For special values of $d$, this
integral simplifies\cite{unpub}. For example, for $d=4$, we get,
$\theta= 1/2 + (2\sqrt 2-1)D/4 +O(D^2)$ and for $d=3$, $\theta= 1/4 +
0.183615\dots D +O(D^2)$.
For $D>2$, it is evident from Eq.\ (\ref{auto1}) that one needs to keep
a nonzero cut-off $a$ in order that the integral does not diverge at
$t_1=t_2$. In this case, for large $t_1$, the dominant contribution to
the integral comes from the regime $t'\to t_1$. It is easy to see that
in the limit when $t_1, t_2$ are both large with their difference
$|t_1-t_2|$ fixed, the autocorrelator in Eq.\ (\ref{auto1}) reduces to
a stationary one, $C(t_1,t_2) \approx B_1(|t_1-t_2|+a^2)^{-(D-2)/2}$,
where $B_1$ is an unimportant constant. One can then invoke the
Newell-Rosenblatt theorem\cite{NR} once again to conclude that for
large $t$, the persistence $P(t)\sim \exp(-\kappa_1 t)$ if $D>4$ and
$\exp(-\kappa_2 t^{(D-2)/2}\ln t) \leq P(t)\leq \exp(-\kappa_3
t^{(D-2)/2})$ for $2<D<4$, where the $\kappa_i$'s are constants.
Combining these results we thus find that the $n\to \infty$
results for the persistence of manifold magnetization in $2<d\leq 4$
are also compatible with our general results in Eq.\ (\ref{asymp1}) on
noting that $\zeta=(D-2)/2$ since $\eta=0$ and $z=2$ within the large
$n$ limit.
Taking hints from the two solvable cases above we now construct a
general scaling theory valid for all $d\geq 2$. The two-point
correlation function of the order parameter, at the critical point,
has the generic scaling form, $\langle \phi({\bf 0}, t_1)\phi({\bf x},
t_2)\rangle \sim x^{-(d-2+\eta)}F(xt_1^{-1/z}, t_2/t_1)$ for large
distance $x$ and large times $t_1$, $t_2$, where $\eta$ and $z$ are
the standard critical exponents defined in the introduction. This
indicates that in the Fourier space, $\langle {\tilde \phi}({\bf K},
t_1){\tilde \phi}(-{\bf K}, t_2)\rangle \sim K^{-(2-\eta)}G(K
t_1^{1/z}, t_2/t_1)$, where ${\bf K}$ is a $d$-dimensional vector
conjugate to $\bf x$. The manifold magnetization $\psi$ is obtained
by summing the order parameter $\phi$ over $d'$ directions. This is
equivalent to putting $K_i=0$ along the $i=1,\dots, d'$
directions. One then obtains the scaling behavior of the two point
correlator of the manifold magnetization, $\langle {\tilde \psi}(\vec
k, t_1){\tilde \psi}(-\vec k, t_2)\rangle \sim
k^{-(2-\eta)}g(kt_1^{1/z}, t_2/t_1)$, where $\vec k$ is now a $D=d-d'$
dimensional vector. For example, within the mean-field theory, the
scaling function $g(x,y)=\exp[-x^2|1-y|]-\exp[-x^2(1+y)]$, as evident
from Eq.\ (\ref{corrk1}) after dropping the $\Delta$ dependent term at
late times. The autocorrelation function $C(t_1,t_2)=\langle\psi(\vec
r, t_1)\psi(\vec r, t_2)\rangle$ is then obtained by integrating over
$\vec k$,
\begin{equation}
C(t_1,t_2)=\int {{d^Dk}\over {k^{2-\eta}}}\,g(kt_1^{1/z}, t_2/t_1)
e^{-k^2a^2},
\label{auto3}
\end{equation}
where $a$ is the soft ultraviolet cut-off as before.
Consider first the case when $D-2+\eta<0$. One can then set the
cut-off $a=0$ (since the integral in Eq.\ (\ref{auto3}) is convergent
at the upper limit), and one obtains $C(t_1,t_2)\sim
t_1^{-(D-2+\eta)/z}f(t_2/t_1)$ in the limit $t_1, t_2 \to \infty$ with
$t_2/t_1$ arbitrary. Note that the function $f(x)\sim
x^{-\lambda_c/z}$ for large $x$ such that $C(t_1,t_2)\sim
t_2^{-\lambda_c/z}$ for $t_2 \gg t_1$ where $\lambda_c$ is the standard
autocorrelation exponent\cite{Huse,JSS,Bray}. This nonstationary
Gaussian correlator can then be reduced, as before, to a stationary
one for the normalized variable $X=\psi/\sqrt{\langle \psi^2\rangle}$
in the logarithmic time, $T=\ln t$ and one gets, $A(T)=\langle
X(0)X(T)\rangle = \exp[(D-2+\eta)T/2z]f(e^T)/f(1)$. Since $A(T)\sim
\exp\left[-{(\lambda_c -(D-2+\eta)/2)}T/z\right]$ for large $T$, it
follows, as before, that the persistence $P(T)\sim \exp(-\theta T)$
for large $T$. This means that the persistence decays as a power law
in the original time variable $t=e^T$, $P(t)\sim t^{-\theta}$ for
large $t$.
In the complementary case $D-2+\eta>0$, the integral in Eq.\ (\ref{auto3})
is, for $t_1=t_2$, divergent near the upper limit without the cut-off.
Hence one needs to keep $a$ nonzero and then the appropriate scaling
limit is obtained by taking $t_1, t_2$ both large keeping their
difference $|t_1-t_2|$ fixed but arbitrary. Then one can replace the
scaling function $g(kt_1^{1/z}, t_2/t_1)$ in Eq.\ (\ref{auto3}) by
another function $g_1(k|t_1-t_2|^{1/z})$ of a single scaling
variable, as in the two previous solvable cases. Performing the
integral, one then finds $C(t_1,t_2)\sim |t_1-t_2|^{-(D-2+\eta)/z}$
for $|t_1-t_2| \gg a^2$. This correlator is stationary and decays as
a power law. Invoking the Newell-Rosenblatt theorem once more,
we find that $P(t)$ decays exponentially for $(D-2+\eta)/z >1$ and
as a stretched exponential for $0 < (D-2+\eta)/z <1$. Combining this
with the result for $D-2+\eta <0$ outlined in the previous paragraph,
gives our general result in Eq.\ (\ref{asymp1}) on defining $\zeta=
(D-2+\eta)/z$.
\begin{figure}
\narrowtext\centerline{\epsfxsize\columnwidth \epsfbox{pt2p.eps}}
\caption{The three outer curves represent the persistence of the line
magnetization in $d=2$ critical Ising model as a function of time for
system sizes $L=63$, $95$ and $127$. The two curves in the middle
represent the persistence of the plane magnetization in $d=3$ critical
Ising model for system sizes $L=15$ and $31$. For the same system
sizes, the two leftmost curves represent the persistence of the line
magnetization in $d=3$.}
\end{figure}
\vspace{-0.2cm}
To test the analytical predictions (i) and (ii), we have done
preliminary Monte Carlo simulations for $d=2$ and $d=3$ Ising model
for small size lattices with periodic boundary conditions. The results
are summarized in Fig.\ 1. The systems were evolved using heat-bath
Monte Carlo dynamics at their bulk critical couplings: $K_c=[\ln
(1+{\sqrt 2})]/2$ in $d=2$ and $K_c \approx 0.221656$ in $d=3$. For
$d=2$, the data for the persistence of the line magnetization for
system sizes $L=63$, $95$ and $127$ (in each case the data was
averaged over $1000$ samples) shows a power law decay $P(t)\sim
t^{-\theta}$ for $t \ll L^{z}$ and crosses over to a faster decay for
$t\gg L^z$. A fit to the linear part of the data on the log-log plot
gives an estimate $\theta\approx 0.72$. Similarly the persistence for
the plane magnetization in $d=3$ shows a power law decay, $P(t)\sim
t^{-\theta}$ with $\theta \approx 0.88$, estimated from small lattice
sizes $L=15$ and $31$. The estimates of these exponents are only rough
and may shift a little with larger lattices. The persistence of the
line magnetization in $d=3$, in contrast, has a much faster decay.
Note that our theory predicts a decay, $P(t)\sim \exp(-a_1t^{\zeta})$
where $\zeta\approx 0.016$. Such a small stretching exponent is
difficult to determine from the small size lattices due to the strong
finite-size effects at late times. All we can say is that the present
data for the line magnetization in $d=3$ is consistent with a decay of
persistence that is faster than a power law but slower than an
exponential. More extensive simulations and a somewhat sophisticated
finite size scaling analysis are required to pin down the precise
value of the stretching exponent\cite{unpub}.
In summary, we have shown that the persistence of submanifolds in a
critical system decays with time in a manner that depends on the
dimensions $d$ and $d'$ of the system and manifold respectively. The
crossover between the power-law decay observed for the `global'
persistence ($d'=d$) and the exponential decay of the `local'
persistence ($d'=0$), occurs via an intermediate regime of stretched
exponential decay with stretched exponent $\zeta=(d-d'-2+\eta)/z$, for
$0 \le \zeta \le 1$. This latter behavior is predicted for the line
magnetization of the $d=3$ critical Ising model. Numerical simulations
are consistent with the analytical predictions.
\vspace*{-0.3cm}
\begin{references}
\vspace*{-1.2cm}
\bibitem{Bray} A.J.\ Bray, Adv.\ Phys.\ {\bf 43}, 357 (1994).
\bibitem{Review} For a review, see S.N.\ Majumdar, Curr.\ Sci.\ {\bf 77},
370 (1999).
\bibitem{MBCS} S.N.\ Majumdar, A.J.\ Bray, S.J.\ Cornell, and C. Sire,
Phys.\ Rev.\ Lett.\ {\bf 77}, 3704 (1996).
\bibitem{globalp} N. Menyhard and G. Odor, J.\ Phys.\ A: Math.\ Gen.\
{\bf 30}, 8515 (1997); L. Schuelke and B. Zheng, Phys.\ Lett.\ A {\bf 233},
93 (1997); H. Hinrichsen and H. Koduvely, Eur.\ Phys.\ J.\ B {\bf 5}, 257
(1998); K. Oerding and F. van Wijland, J.\ Phys.\ A: Math.\ Gen.\ {\bf 31},
7011 (1998); S. Cueille and C. Sire, Eur.\ Phys.\ J.\ B {\bf 7}, 111 (1999);
E. Albano and M.A.\ Munoz, Phys.\ Rev.\ E {\bf 63}, 031104 (2001);
S. Lubeck and A. Misra, Eur.\ Phys.\ J.\ B {\bf 26}, 75 (2002);
M. Saharay and P. Sen, Physica A {\bf 318}, 243 (2003); R. da Silva, N.A.\
Alves, and J.R.D.\ de Felicio, cond-mat/0302166.
\bibitem{Slepian} D. Slepian, Bell Syst.\ Tech.\ J.\ {\bf 41}, 463 (1962).
\bibitem{interface} J. Krug et. al., Phys.\ Rev.\ E {\bf 56}, 2702 (1997).
\bibitem{NR} G.F.\ Newell and M. Rosenblatt, Ann.\ Math.\ Stat.\ {\bf 33},
1306 (1962).
\bibitem{perturb} S.N.\ Majumdar and C. Sire, Phys.\ Rev.\ Lett.\ {\bf 77},
1420 (1996); K. Oerding, S.J.\ Cornell, and A.J.\ Bray, Phys.\ Rev.\ E
{\bf 56}, R25 (1997).
\bibitem{unpub} Details will be published elsewhere.
\bibitem{Huse} D.A.\ Huse, Phys.\ Rev.\ B {\bf 40}, 304 (1989).
\bibitem{JSS} H.K.\ Janssen, B.\ Schaub, and B.\ Schmittman, Z.\ Phys.\
B {\bf 73}, 539 (1989).
\end{references}
\end{multicols}
\end{document}