Scaling near Quantum Chaos Border in Interacting Fermi Systems
Extracted source text ยท arXiv cond-mat/0004237
\documentstyle[aps,multicol,epsf]{revtex}
%\documentstyle[aps,prl,preprint]{revtex}
\begin{document}
\preprint{}
\draft
\title{Scaling near Quantum Chaos Border in Interacting Fermi Systems}
\author{Pil Hun Song\cite{adr}}
\address{Center for Theoretical Physics, Seoul National University,
Seoul 151--742, Korea}
\date{\today}
%\ \hspace{5in}{\normalsize SNUTP }\\
\maketitle
\begin{abstract}
The emergence of quantum chaos for interacting Fermi systems is
investigated by numerical calculation of the level spacing distribution
$P(s)$ as a function of interaction strength $U$ and the excitation energy
$\epsilon$ above the Fermi level. As $U$ increases, $P(s)$ undergoes
a transition from Poissonian (nonchaotic) to Wigner-Dyson (chaotic)
statistics and the transition is described by a single scaling parameter
given by $Z = (U \epsilon^{\alpha}-u_0) \epsilon^{1/2\nu}$, where $u_0$
is a constant. While the exponent $\alpha$, which determines the global
change of the chaos border, is indecisive within
a broad range of $0.9 \sim 2.0$, the finiteness of $\nu$, which comes from
the increase of the Fock space size with $\epsilon$, suggests that the
transition becomes sharp as $\epsilon$ increases.
\end{abstract}
\pacs{PACS numbers: 05.45.-a, 05.30.Fk, 73.23.-b}
\begin{multicols}{2}
\narrowtext
Recently, the emergence of quantum chaos due to electron-electron
interaction has attracted much attention. While the subject
has a longer history in nuclear physics, with the progress
of modern nanofabrication techniques, it has entered condensed
matter physics since a quantum dot system, for example, could be
considered as an artificial atom with some physical parameters
under control. Besides being of interest as its own right, the
importance of the subject stems from the fact that it is related to
a failure of the perturbative approach in interacting many-particle systems.
Traditionally, the perturbative method has been one of the standard
tools in theoretical many-particle physics. The emergence of quantum chaos
means a strong mixing of the unperturbed levels, thereby inducing
breakdown of the perturbation series.
A recent theoretical work by Altshuler {\it et al.}\cite{alt} for
the quasiparticle decay in a quantum dot has especially stimulated
many theoretical investigations. In their paper\cite{alt}, the
quasiparticle decay process was mapped to a single particle diffusion
on the Bethe lattice, which is a nonperturbative treatment of
the problem. They concluded that there is a transition to quantum
chaos at a critical excitation energy $\epsilon_c \sim \sqrt{g}\Delta$,
where $g$ is the dimensionless conductance and $\Delta$ is the mean
level spacing between the single-particle levels. However, many
authors afterwards have pointed out that the
mapping to a Bethe lattice in ref.~\cite{alt} is too naive for a proper
description and they obtained different results using other methods.
The ongoing controversy could be summarised through two main
questions. i) what is the relation between $\epsilon$ and
$g$ at the quantum chaos border, and ii) whether the
transition is sharp or not. Regarding the first question, there exist
further different predictions such as $\epsilon_c
\sim g^{2/3}\Delta$\cite{jac,mir} and $\epsilon_c \sim g\Delta/\ln{g}$
\cite{sil}. As for the second question, Jacquod and
Shepelyansky\cite{jac} argued that the transition is smooth since
the coupling between the Fock states is
of nonlocal nature. The authors of ref.~\cite{mej} concluded that
the transition is smooth based on their numerical result for
the local density of states and the participation number. In ref.~\cite{ley},
the same conclusion has been drawn by calculating the inverse participation
ratio (IPR) for higher values of $\epsilon$ by use of the, so called, layer
model. On the other hand, Berkovits and Avishai\cite{ber} suggested a finite
size scaling hypothesis, according to which the transition becomes sharp as
$\epsilon$ increases, which was based on their exact numerical results for
small system size. Silvestrov\cite{sil} also proposed that the transition
is sharp when the effective high-order interaction is taken
into account. In addition, the authors of ref.\cite{ley2} performed
an analysis of their numerical result for the IPR and found that their
results show features consistent with the prediction of ref.~\cite{sil}.
However, their conclusion is not decisive enough concerning the sharpness
of the transition. A reliable numerical test is, therefore, urgently
needed to settle the issue.
The main difficulty of numerical test is due to the fact that one should
consider
the regime of $g \gg 1$, which corresponds to $\epsilon \gg \Delta$. Since
the size of the matrix to be diagonalized rapidly increases with $\epsilon$,
one needs an alternative to the brute force method. The layer model,
introduced by Georgeot and Shepelyansky\cite{geo}, allows one to handle much
larger system size (higher $\epsilon$) at a given computational cost
by truncating out the Slater determinants which contribute little to
a given eigenstate.
In this paper, we calculate the level spacing distribution $P(s)$ for
interacting fermionic systems up to $\epsilon/\Delta = 27$ by use of
the layer model. The change of $P(s)$ from the Poissonian to Wigner-Dyson
statistics
represents the transition from integrability to chaos. While our result
does not allow us to resolve the controversy over the parametric relation
for the quantum chaos border, i.e. question (i), it gives
strong evidence that the transition becomes sharper as $\epsilon$
increases. The finite size scaling (FSS) behaviour can be understood
through a comparison with an infinite dimensional Anderson model.
Let us begin with $n_f$ spinless fermions on $m$ single-particle levels for
which the Hamiltonian ${\cal H}$ is given by ${\cal H}_0 + {\cal H}_1$ with
\begin{equation}
{\cal H}_0 = \sum_i \epsilon_i c^\dagger_i c_i \ \ \mbox{and} \ \
{\cal H}_1 = \sum_{i<j,k<l} U_{ij,kl} c^\dagger_l c^\dagger_k c_i c_j.
\end{equation}
${\cal H}_0$ represents the noninteracting Hamiltonian. A single-particle
level energy $\epsilon_i$ with $i = 1,2,\cdot\cdot\cdot, m$ is randomly
chosen from the interval $[(i-1/2)\Delta,(i+1/2)\Delta]$ with uniform
probability distribution. $U_{ij,kl}$ denotes the random two-body
interaction matrix element and also has a box distribution $[-U/2,U/2]$.
When ${\cal H}$ is interpreted as the Hamiltonian for a quantum dot
system, $U$ is related to $g$ as $U \sim \Delta/g$\cite{alt,bla}.
${\cal H}$ is widely known as the Two-Body Random Interaction
Model (TBRIM) and its full matrix size is given by $N_t = n_f!/[m!(m-n_f)!]$.
When ${\cal H}$ is written in matrix form on the basis of the
eigenstates of ${\cal H}_0$, a diagonal element is the sum of the
energies of occupied single-particle levels, while an off-diagonal element is
nonzero only for two states which are different upto two particle-hole
pairs.
Since $N_t$ increases very rapidly with $n_f$ and $m$, direct
numerical diagonalisation of ${\cal H}$ is limited to rather small
values of $n_f$ and $m$. Due to this difficulty, the layer model
has been introduced by Georgeot and Shepelyansky\cite{geo} and the
details are as follows.
For each eigenstate of ${\cal H}_0$, we define a sequence $\{f_i\}$
such that $f_i$ = 0 when the level $i$ is empty and $f_i$ = 1 when
it is filled with a fermion.
Then ${\cal E}/\Delta = \sum^{m}_{i=1} i f_i$ ranges from $n_f (n_f+1)/2$
to $n_f (m+1) - n_f (n_f+1)/2$, which is approximately equal to the
ground state energy and the highest excitation energy, respectively, in
units of $\Delta$. If we rewrite ${\cal H}$ in
ascending order of ${\cal E}$ and set $U_{ij,kl} = 0$ for $i+j \neq
k+l$, the whole matrix reduces to a block diagonal form where
${\cal E}$ is constant within each block. We call the submatrix
with ${\cal E}/\Delta-n_f (n_f+1)/2 = j$ the $j$-th layer model.
In general, the matrix size of the $j$-th layer model, $N(j)$,
varies depending on $n_f$ and $m$. However, for $1 \ll n_f \ll m$,
$N(j)$ for $j\leq n_f$ is determined solely by $j$ and behaves as\cite{mot}
\begin{equation}
N(j) \sim \exp(\pi \sqrt{2j/3})/j.
\end{equation}
The layer model hereafter is understood in such a sense. An eigenstate
of ${\cal H}$ with energy $E$ can be written as a superposition of the
eigenstates of ${\cal H}_0$, the energies of which lie within the width
$\Gamma$ around $E$. $\Gamma$ is much less than $\Delta$, when $U \ll
\Delta$, i.e. $g \gg 1$. This defines the valid regime of the layer model.
One of the well established criteria for transition from
integrability to chaos is the change of the level spacing distribution
$P(s)$ from the Poissonian $P_p(s) = e^{-s}$ to the Wigner surmise
$P_w(s) = (\pi s/2) e^{-\pi s^2/4}$. To
quantify the proximity of $P(s)$ to either of the two, it is useful
to define $\eta$ in terms of the variance of $s$ as follows;
\begin{equation}
\eta =\frac{\int_0^{\infty} s^2(P(s)-P_{w}(s)) ds}{\int_0^{\infty}
s^2(P_{p}(s)-P_{w}(s)) ds}.
\end{equation}
In this way, $\eta$ is 1 for $P_p(s)$, 0 for $P_w(s)$ and inbetween
for an intermediate distribution.
In our calculation, the Hamiltonian for the layer model for $19 \leq j
\leq 27$ has been constructed and numerical diagonalisation has been
performed over 200-1000 disorder configurations for each parameter set
$(j,g)$. Corresponding matrix size ranges from 490 ($j=19$) to 3010
($j$=27). To exclude the contribution of the tail states near the
edge, 50 $\%$ of eigenvalues around the band center have been used to
construct $P(s)$. If we choose a smaller part of the eigenvalue set,
e.g. 25 $\%$, the result does not change significantly.
The $\eta$'s for $19 \leq j \leq 27$ and for $1/200 \leq U \leq
1/10$\cite{com1} are shown in Fig.~1. The gradual change from $\eta
\simeq 1$ (Poissonian, integrable) to $\eta \simeq 0$ (Wigner-Dyson,
chaotic) with increase
of $U$ is found with all $j$. As $j$ increases, the transition occurs
at a smaller value of $U$, indicating that the interaction becomes
more efficient for mixing the levels as the energy increases.
The global dependence of the chaos border on the excitation energy may
be found by plotting the data of Fig.~1 on a rescaled $x$-axis with
$x = U j^{\alpha}$\cite{com2}. If there exists $\alpha = \alpha_0$ such that
all the data points lie in the same curve, the transition is described
by the relation $\epsilon = U^{-1/\alpha_0}$ and there is a smooth crossover
instead of a sharp phase transition when $\epsilon \rightarrow \infty$.
However there does not exist such a value $\alpha_0$ for
our data. Instead, as shown in Fig.~2(a-c), there is a single crossing
point where the curves meet one another and the slope at the crossing
point becomes larger as $j$ increases. This
suggests that at sufficiently high energy the transition to chaos is a
sharp phase transition. In fact such a FSS feature
with $j$ is found for a broad range of $0.9 \leq \alpha \leq 2$.
When $\alpha$ is tuned from 0.9 to 2, $\eta_c$ ($\eta$ at the crossing
point) decreases monotonically from 0.8 to 0.2. For $\alpha < 0.9$
or $\alpha > 2$, the crossing point cannot be clearly identified.
These results imply that while we cannot determine the exponent
$\alpha$ which governs the quantum chaos border\cite{com3}, the
transition shows a FSS property regardless of the
choice of $\alpha$ as long as $0.2 < \eta_c < 0.8$.
For a further analysis of this FSS behaviour, we assume that
$\eta$ is given by a function $f$ of a single scaling variable
$Z$ such that
\begin{equation}
\eta(\epsilon,U) = f(Z) = f[(U \epsilon^{\alpha}-u_0) \epsilon^{1/2\nu}],
\end{equation}
where $u_0$ is a constant of the order of unity which is independent
of $U$ and $\epsilon$. To understand the introduction of this
scaling parameter $Z$, it would be helpful to compare the layer model
with an Anderson Hamiltonian. One can think of the standard Anderson
Hamiltonian
defined on a graph, which has the same structure as the $j$-th layer
model; there are $N(j)$ vertices and a bond between two vertices exists
if two corresponding Slater determinants of the $j$-th layer
model have nonzero transition amplitude by the two-body interaction.
The size of the graph can then be defined by the average minimum number
of bonds between two vertices. It is given by $\sim \sqrt{j}$, which
is the same as the typical number of electron-hole pairs for the
$j$-th layer model\cite{mir}. The spatial dimension
of the graph in the limit of $j \rightarrow \infty$ is given by
\begin{equation}
d=\lim_{j \rightarrow \infty} \frac{\ln N(j)}{\ln (\sqrt{j})}
\rightarrow \infty.
\end{equation}
We expect that there exists a transition to chaos as we decrease the
disorder strength of the on-site energy for a fixed value of hopping
energy between nearest neighbors. The transition will show a FSS behaviour
with the system size $L \sim j^{1/2}$ and with $\nu =1/2$, the correlation
length exponent of infinite dimensional Anderson model\cite{efe}.
Returning back to the layer model, we expect that a similar FSS property
exists since there also is the competition between the system size and
a correlation length governing the chaotic property of the system.
A difference between the layer models and the thus defined
Anderson models is that the matrix elements are strongly correlated
in the former, while they are completely random in the latter\cite{com4}.
Therefore, $\nu$ of Eq.~(4) is not necessarily equal to that of the
infinite Anderson model, 1/2. In passing, we note that Berkovits and
Avishai\cite{ber} introduced a scaling hypothesis very similar to Eq.~(4)
based on their numerical results for lower energy.
According to Eq.~(4), the slope of the curves $\eta(U)$ for a given
$j$ behaves as $\sim j^{\alpha + 1/(2\nu)}$ at the crossing point and $\nu$
is obtained by fitting this variation of the slopes for a given value
of $\alpha$. Fig.~3 shows the scaling plot obtained in this way with
$\alpha = 1.5$ ($\eta_c = 0.56$) and in this case $\nu$ is given by
0.59. Such a procedure can
be performed in the same way for $0.2 \leq \eta_c \leq 0.8$ ($0.9 \leq \alpha
\leq 2$) and the result for $1/(2\nu)$ is shown in Fig.~4. Taking
account of uncertainties of all data, $1/(2\nu)$ lies between 0.5 and 1.5
($0.3 < \nu < 1$) over the whole range where the crossing point is
identified. Therefore, though $\nu$ also varies in a rather broad
interval, our data evidently excludes the possibility of $\nu$ being
infinity, i.e. the possibility of a smooth crossover. One should note
that this large variation in $\nu$ over a factor of 3 is the result
of the fact that $\alpha$ is indecisive. If one can pin down the range
of $\alpha$ in some other way than our method, or assume one of the
previously proposed values (e.g. 1.5\cite{jac,mir} or 2\cite{alt}) to
be valid, the uncertainty in $\nu$ can be determined to a higher
accuracy from Fig.~4.
Finally, we discuss the sharpness of the transition when $\epsilon$ is
varied for a fixed value of $U$, since in real measurements with
a quantum dot the conductance of the sample $g \sim \Delta/U$ is kept constant
and the bias voltage, which corresponds to $\epsilon$, is varied. We
define the transition as sharp when the ratio of the transition
interval $\delta \epsilon$ to the transition energy $\epsilon_c$ goes
to zero as $\epsilon_c$ increases. From Eq.~(4), $\epsilon_c$ is
given by $(u_0/U)^{1/\alpha}$ and $\delta \epsilon$ is estimated
using the slope of the curve $\eta(\epsilon,U)$ as
\begin{equation}
\delta \epsilon \sim \left| \frac{\partial\eta}{\partial\epsilon}
\right|^{-1}_{\epsilon = \epsilon_c} \sim \ \
\epsilon_c^{1-\frac{1}{2\nu}}.
\end{equation}
Therefore, $(\delta \epsilon/\epsilon)|_{\epsilon = \epsilon_c}
\sim \epsilon_c^{-1/(2\nu)}$
and we find that the transition becomes sharp as long as $\nu$ is
not infinity. Therefore, the result of Fig.~4 shows that the transition
to chaos takes place sharply if one uses a sufficiently clean sample,
i.e. large $g$, for a quantum dot.
In summary, we have numerically investigated the emergence of quantum
chaos in the interacting many-fermion systems using the layer model.
While our result does not allow us to find the global dependence of
the chaos border as a function of the interaction strength, we find that
the transition becomes sharp at sufficiently high energy. This is
also true when one observes the transition to chaos in a quantum dot
as a function of bias voltage.
I am thankful to Jae Dong Noh, Gun Sang Jeon and especially to Dima
L.~Shepelyansky for helpful discussions.
\begin{thebibliography}{99}
\bibitem[*]{adr}
present address: Max-Planck-Institut f\"{u}r Physik komplexer Systeme,
N\"{o}thnitzer Str.~38, D-01187 Dresden, Germany, Email:
phsong@mpipks-dresden.mpg.de
\bibitem{alt}
B.~L.~Altshuler, Yu.~Gefen, A.~Kamenev and L.~S.~Levitov, Phys. Rev.
Lett. {\bf 78}, 2803 (1997).
\bibitem{jac}
P.~Jacquod and D.~L.~Shepelyansky, Phys. Rev. Lett. {\bf 79}, 1837 (1997).
\bibitem{mir}
A.~Mirlin and Y.~Fyodorov, Phys. Rev. B {\bf 56} 13393 (1997).
\bibitem{sil}
P.~G.~Silvestrov, Phys. Rev. E {\bf 58}, 5629 (1998).
\bibitem{mej}
C.~Mejia-Monasterio, J.~Richert, T.~Rupp, and H.~A.~Weidenm{\"u}ller,
Phys. Rev. Lett. {\bf 81} 5189 (1998).
\bibitem{ley}
X.~Leyronas, J.~Tworzydlo, and C.~W.~Beenakker, Phys. Rev. Lett.
{\bf 82}, 4894 (1999).
\bibitem{ber}
R.~Berkovits and Y.~Avishai, Phys. Rev. Lett. {\bf 80}, 568 (1998).
\bibitem{ley2}
X.~Leyronas, P.~G.~Silvestrov and C.~W.~Beenakker, Phys. Rev. Lett.
{\bf 84} 3414 (2000).
\bibitem{geo}
B.~Georgeot and D.~L.~Shepelyansky, Phys. Rev. Lett. {\bf 79} 4365
(1997).
\bibitem{bla}
Ya.~M.~Blanter, Phys. Rev. B {\bf 54}, 12807 (1996).
\bibitem{mot}
A.~Bohr and B.~R.~Mottelson, {\it Nuclear Structure} (Benjamin, New
York, 1969), Vol.~1, p.~284.
\bibitem{com1}
We may set $\Delta=1$ without loss of generality. Under this
condition, the relations between parameters become simpler to yield
$g \sim 1/U$ and $\epsilon=j$, for example.
\bibitem{com2}
We note that within the finite range of $1/200 \leq U \leq 1/10$ ($10 \leq
g \leq 200$), the functional form of $g/\log{g}$, which was suggested
in ref.~\cite{sil}, does not differ seriously from a power form of
$g^{1/\alpha}$.
\bibitem{com3}
We may try a more refined analysis through higher cumulants of $P(s)$
assuming that just at the transition point, $P(s)$ will be given by a
critical form $P_c(s)$ which is insensitive to $U$ and $j$.
Unfortunately, such an approach is not successful within our data
accuracy.
\bibitem{efe}
K.~B.~Efetov, Physica (Amsterdam) {\bf 167A}, 119 (1990).
\bibitem{com4}
For example, the correlation of the diagonal energies for the
layer model decays very slowly with the distance as
$\langle {\cal H}_{ii} {\cal H}_{jj} \rangle \ \sim \ C-d(i,j)$,
where $C$ is a constant of order of $n_f$ and $d(i,j)$ is the distance
between two Slater determinants or two vertices characterized by $i$
and $j$ respectively. In contrast, it is zero unless $i=j$ for the
Anderson models.
\end{thebibliography}
\pagebreak
\noindent
%\vspace{-0.7cm}
\begin{figure}
\centerline{\epsfxsize=8cm \epsfbox{fig1.epsi}}
\vspace{4mm}
\caption{$\eta$ as a function of the interaction strength $U$ for
the layer model with various values of $j$.
}
\end{figure}
\noindent
%\vspace{5.5cm}
\begin{figure}
\centerline{\epsfxsize=6cm \epsfbox{fig2.epsi}}
\vspace{4mm}
\caption{The same data as Fig.~1 with rescaled $x$-axis of $x=U j^{\alpha}$
with $\alpha = 1.2$, 1.5 and 1.8. A partial set of data from Fig.~1 is
shown for clarity.
}
\end{figure}
\noindent
%\vspace{-0.7cm}
\begin{figure}
\centerline{\epsfxsize=8cm \epsfbox{fig3.epsi}}
\vspace{4mm}
\caption{
Scaling plot of the data of Fig.~1 for $\alpha = 1.5$. In
this case $1/(2\nu)$ is given by 0.85, i.e. $\nu = 0.59$.
}
\end{figure}
\noindent
%\vspace{-0.7cm}
\begin{figure}
\centerline{\epsfxsize=8cm \epsfbox{fig4.epsi}}
\vspace{4mm}
\caption{
$1/(2\nu)$ when $\eta_c$ varies from 0.2 to 0.8.
}
\end{figure}
\end{multicols}
\end{document}