Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Short Range Plasma

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Eur. Phys. J.

B 19, 121132 (2001)

THE EUROPEAN
PHYSICAL JOURNAL B
EDP Sciences
c Societ`

a Italiana di Fisica
Springer-Verlag 2001

Short-range plasma model for intermediate spectral statistics


E. Bogomolny1,a , U. Gerland2 , and C. Schmit1
1
2

Laboratoire de Physique Theorique et Mod`eles Statistiquesb , Universite Paris-Sud, 91405 Orsay Cedex, France
Phys. Dep., UCSD, La Jolla, CA 92093-0319, USA
Received 13 September 2000
Abstract. We propose a plasma model for spectral statistics displaying level repulsion without long-range
spectral rigidity, i.e. statistics intermediate between random matrix and Poisson statistics similar to the
ones found numerically at the critical point of the Anderson metal-insulator transition in disordered systems
and in certain dynamical systems. The model emerges from Dysons one-dimensional gas corresponding to
the eigenvalue distribution of the classical random matrix ensembles by restricting the logarithmic pair
interaction to a finite number k of nearest neighbors. We calculate analytically the spacing distributions
and the two-level statistics. In particular we show that the number variance has the asymptotic form
2 (L) L for large L and the nearest-neighbor distribution decreases exponentially when s ,
P (s) exp(s) with = 1/ = k + 1, where is the inverse temperature of the gas ( = 1, 2
and 4 for the orthogonal, unitary and symplectic symmetry class respectively). In the simplest case of
k = = 1, the model leads to the so-called Semi-Poisson statistics characterized by particular simple
correlation functions e.g. P (s) = 4s exp(2s). Furthermore we investigate the spectral statistics of several
pseudointegrable quantum billiards numerically and compare them to the Semi-Poisson statistics.
PACS. 05.45.-a Nonlinear dynamics and nonlinear dynamical systems 03.65.Sq Semiclassical theories
and applications 05.40.-a Fluctuation phenomena, random processes, noise, and Brownian motion

1 Introduction

Since the advent of random matrix theory there has been


considerable interest in the statistical analysis of spectra
[14]. Two diametrically opposed statistical distributions
have been found to be of universal relevance: the Poisson
distribution, i.e. completely uncorrelated levels, and the
Wigner-Dyson distributions of random matrix theory. The
prominent features of these distributions are conveniently
characterized with the help of spectral observables such
as the nearest-neighbor spacing distribution P (s) and the
number variance 2 (L) (the variance of the number of
levels in an energy interval of length L in the unfolded
spectrum). The former stresses the correlations on a short
scale, while the latter measures the stiffness of the spectrum, i.e. long-range spectral correlations. For the standard random matrix ensembles (orthogonal, unitary and
symplectic symmetry labeled by = 1, 2 and 4 respectively) the spacing distribution is approximately given by
the Wigner surmise and the number variance increases

e-mail: bogomol@ipno.in2p3.fr
b
Unite de Recherche de lUniversite Paris XI et du CNRS
(UMR 8626)

only logarithmically with large L,


P (s) = a s exp(c s2 )
2
2 (L)
log(L)
(L )
2

(1)

(a and c are determined by normalization and by requiring the mean level spacing to be one). Thus the WignerDyson distributions are marked by level repulsion and
long-range spectral rigidity. In contrast, for the Poisson
distribution one has neither level repulsion nor spectral
rigidity,
P (s) = exp(s)
2 (L) = L .

(2)

In the present article we devise a discrete set of statistical


distributions with properties intermediate between those
of the Poisson and Wigner-Dyson distributions, namely
level repulsion
P (s) s

(s 0)

(3)

paired with an exponential decay of the nearest-neighbor


spacing distribution,
P (s) exp(s)

(s ) ,

(4)

122

The European Physical Journal B

and a linear asymptotic of the number variance,


2 (L) L

(L ) .

(5)

Our distributions arise as the Gibbs distribution of a onedimensional gas of interacting classical particles. It was
noted already by Dyson that the distribution of eigenvalues Ej of random matrices is equivalent to the distribution
of particles in a one-dimensional gas with inverse temperature and the repulsive two-body interaction potential
V (x) = log |x|, since the joint probability distribution
can be written in the form
1 W (E1 ,... ,EN )
PN (E1 , . . . , EN ) = ZN
e
,

(6)

where the normalization constant ZN is the canonical partition function of the gas and
X
X
W (E1 , . . . , EN ) =
U (Ei ) +
V (|Ei Ej |) (7)
i

A second motivation which led us to consider the shortrange plasma model is the following. In the context of
quantum chaos the statistical analysis of spectra plays
a major role. In the semiclassical limit the quantum energy spectra of systems with integrable classical dynamics
generally display Poisson statistics [8], while the WignerDyson distribution is associated with fully chaotic classical dynamics [9]. But there are systems which are neither
integrable nor chaotic. In [10] we examined numerically
a few such systems and found that their spectral statistics exhibit all properties (35) typical for intermediate
statistics. In particular, the simple expression

i<j

is the total potential energy. The one-body potential U


serves to confine the Ej to a finite stretch of the E-axis,
e.g. U (E) E 2 for the Gaussian ensembles. It determines
the mean particle (level) density, but it is irrelevant for
the statistical correlations on the scale of the mean level
spacing (random matrix universality) as long as U is steep
enough to actually confine the levels [5].
Our plasma model of intermediate spectral statistics
is defined by the same equations (6) and (7), except that
we restrict the second sum on the right hand side of equation (7) to a finite number k of nearest neighbors. It is this
screening of the Coulomb interaction between the levels,
which is the essential ingredient leading to the properties (35).
Our first motivation stems from the physics of disordered electronic systems. A 3-dimensional disordered
sample undergoes a phase transition between an insulating and a metallic phase as a function of the disorder strength (Anderson metal-insulator transition). In
the insulating phase the electron eigenfunctions are localized and since non-overlapping eigenfunctions are uncorrelated, the eigenenergies are Poisson distributed. In
contrast, in the metallic phase the eigenfunctions extend
homogeneously over the whole sample and overlap
strongly which leads to a Wigner-Dyson distribution of the
energy levels. Exactly at the transition point the electron
eigenfunctions are extended, but strongly inhomogeneous
(multifractal). This leads to intermediate spectral statistics, which are believed to be universal, i.e. independent
of the microscopic details of the disordered system [6].
Fyodorov and Mirlin [7] calculated the overlap of two
critical eigenstates with slightly different energy and found
that the overlap is one if the energy separation, s, between
the eigenfunctions is of the order of mean level spacing,
while for larger s it decays as a certain power of s. In
contrast, for Wigner-Dyson statistics the overlap is one
for all values of s. Hence one may conclude that at the
critical point of the Anderson metal-insulator transition
only eigenvalues which are separated by at most a few
level spacings interact strongly.

P (s) = 4s e2s

(8)

is an excellent fit to the spacing distribution of a subclass of systems considered. We further noticed that the
short-range plasma model with nearest-neighbor interaction (k = 1) leads precisely to this distribution (and we
checked that other correlation functions are also well described by this model).
Another interesting class of plasma models with
screened Coulomb interaction is the Gaudin model
[11,12] defined by equations (6) and (7) with the two-body
interaction potential
1
x2
V (x) = log 2

2
a + x2

(9)

When a it reduces to the random matrix models,


while in the limit a 0 it yields Poisson statistics. For
= 2 the model is solvable and all correlation functions
can be written in a closed form [11,12]. They obey all three
conditions (35) characteristic for intermediate statistics
with the following constants (for = 2)
1
(1 e ),

(10)

and
=

Z
0

e 1

ln(1 + t)
dt.
t

(11)

Here = 2a and is the mean density of the levels.


In general, in any plasma model with short-range interaction the nearest-neighbor spacing distribution will decay exponentially at large distances. It is the otherwise
unusual property of long-range interaction in the standard random matrix models (which manifests itself in the
summation over all pairs of particles in Eq. (7)) that is
responsible for the exp(cs2 ) decay of P (s) at large s.
The main drawback of the Gaudin model is that no analytical solution is known for 6= 2. In contrast, our model
can be solved for arbitrary (and arbitrary potential).
We stress that our one-dimensional gas model is meant
as a toy model which deserves interest, because it provides a natural discrete interpolation between Poisson and
Wigner-Dyson statistics, it is analytically solvable, and
it is physically motivated. It does not pretend to furnish
an exact description of the critical statistics of a physical

E. Bogomolny et al.: Short-range plasma model for intermediate spectral statistics

model (e.g. the Anderson model in the MIT point). Despite significant theoretical progress (see [7,17,2426] and
references therein), the precise form of the critical distribution is not yet known though there exist arguments
([2426]) that the main difference between the standard
random matrix ensembles and the intermediate statistics
is in the form of the correlation kernel, K(x, y), which determines the correlation functions in the random matrix
theory [2]. In the standard ensembles
K(x, y) =

sin (x y)

(x y)

(12)

123

N-1
2

L/2

Fig. 1. N particles on a circle of circumference L. The positive spacings between consecutive particles measured along the
circle are denoted by j with j = 1 . . . N.

For intermediate statistics it has been argue that


K(x, y)

a sin (x y)
,
sinh a(x y)

(13)

where a is a parameter. When a 0 the standard random


ensembles are recoved. Non-zero value of a characterizes
the intermediate statistics. The detailed discussion of this
type of critical behavior is beyond the scope of this paper.
The paper is organized as follows. In the next section
we present the calculation of the spacing distributions and
the number variance in the one-dimensional plasma, first
for k = 1, then for k = 2, and finally for arbitrary k.
We compare the resulting distributions with our numerical results for the pseudointegrable billiards in Section 3.
Finally, in the discussion of Section 4 we compare the
short-range plasma model to other existing interpolations
between Poisson and Wigner-Dyson statistics.

2 The model
Instead of a linear one-dimensional gas as in equations (6)
and (7), where the levels are confined by a one-body
potential U (E), we choose a circular geometry. This is
convenient, since on a circle the levels are automatically
confined and it is unnecessary to introduce an external
potential. The mean level density is then constant and
unfolding is trivial. At the same time the correlations in
the unfolded level sequence are the same as in the linear geometry (in the limit of a large number of levels),
just as the Gaussian ensembles of random matrix theory
are locally equivalent to Dysons circular ensembles. The
method of calculation that we apply in this section goes
back to G
ursey [13] for the case of nearest-neighbor interaction, and to van Hove [14] for the general case (see also
Ref. [15]).
We consider N particles (representing energy levels)
on a circle of circumference L. We denote the positive
spacings between neighboring particles by j with j =
1 . . . N (see Fig. 1). Hence,
N
X

V () = log || as in random matrix theory), but we let


each particle interact only with its k nearest neighbors to
the left and to the right (k = 1, 2, 3, . . . ).
The canonical partition function of this one-dimensional gas is
Z
ZN (L, ) =

j=1

For convenience we use a periodic index, i.e. j+N := j .


We introduce an interaction between the particles via a repulsive two-body potential V () (eventually we will choose

d1 . . .
0

N


X
dN L
i
i=1


exp

N
X


W (j , . . . , j+k1 ) , (14)

j=1

where
W (j , . . . , j+k1 ) = V (j ) + V (j + j+1 )
+ . . . + V (j + . . . + j+k1 ) . (15)
W (j , . . . , j+k1 ) includes the interaction energy of the
particle which is located at the left of j with its k nearest
right neighbors only to avoid double counting. Since for
the application to spectral statistics is not a free variable, but takes only the values 1, 2, and 4, we suppress
the -dependence of the partition function in the following. For brevity of notation we define


f () := exp V () ,
(16)
which for the case of V () = ln || becomes
f () = || .

(17)

The joint probability distribution of n consecutive spacings then takes the form
1
p(1 , . . . , n ) =
ZN (L)

N
Y

Z
dn+1 . . .

j = L .

N


X
dN L
i
i=1

f (j )f (j + j+1 ) f (j + . . . + j+k1 ) . (18)

j=1

In the following the variable s denotes distances measured in units of the mean spacing = L/N . Our aim is
to calculate correlation functions such as the nth nearestneighbor spacing distribution P (n, s) (the distribution of

124

The European Physical Journal B

the distance s between two particles that have n particles between them) and the two-point correlation function
R2 (s) (the probability of finding any two particles at distance s). The latter is related to the former by summation
over n,

R2 (s) =

P (n, s) .

The large N limit of the partition function can now be


calculated by performing the Laplace inversion in saddle
point approximation. We have
c+i
Z

1
ZN (L) =
2i

(19)

dt gN (t) eLt
ci
c+i
Z

n=0

1
=
2i

Using equation (18) the nth nearest-neighbor spacing distribution can be expressed as
Z
Z
n+1

1 X 
P (n, s) = d1 . . . dn+1 s
i
i=1
0

p(1 , . . . , n+1 ) .

(20)

dt eN (t+log g(t))
ci

h
iN
g(c) eLc ,

where = L/N is the mean level spacing and c is determined from the saddle point equation

The number variance can be obtained from the two-point


correlation function with help of the relation

ZL
2 (L) = L 2

ds (L s)(1 R2 (s)) .

(21)

A convenient way to calculate the asymptotic behavior of


the number variance for large L is to consider the asymptotic expansion of the Laplace transform of the two-point
correlation function,
Z
(22)
g2 (t) = ds R2 (s) ets ,

p(1 , . . . , n ) =

n
Y
j=1

1
+ 0 + 1 t + O(t2 )
t
then we get from equation (21)

(t 0)

2 (L) = L + C + O(L1 )

(L ) ,

(23)

(24)

where = 1 + 20 and C = 21 (for the determination


of the constant term we have assumed that R2 (s) falls off
faster than 1/s2 , which is true for the cases considered in
this article).

Z
d1 . . .

dN
0

i=1

dn p(1 , . . . , n ) = p(1 , . . . , n1 ) .
0

With help of equations (20) and (29) we find for the


Laplace transformation of the nth nearest-neighbor spacing distribution

 Z
=
0

dx f (x) etx

N

g(n, t) :=

ts

ds e

h
iN
=: g(t)
.

n+1
.

(30)

Using the relation (19) we calculate the Laplace transform


of the two-point correlation function, equation (22),
g2 (t) =

g(c + t/)

g(c) g(c + t/)

(31)

The small-t asymptotic behavior is


g2 (t) =

(26)

g(c + t/)
P (n, s) =
g(c)

j=1

By Laplace transformation with respect to L we obtain


Z
gN (t) := dL ZN (L) etL

(29)

N
N
Y

X
L
i
f (j ) . (25)

1 cj
f (j ) .
e
g(c)

Note that the factor g(c)1 assures the proper normalization,

2.1 Nearest-neighbor interaction (k = 1)


We begin with the simplest case, where the interaction is
restricted to nearest neighbors, so that the expression for
the partition function equation (14) simplifies to

(28)

Using equation (27) and assuming n  N we obtain

for small t. If

ZN (L) =

g 0 (c)
=0.
g(c)

The expression for the joint probability distribution


of n consecutive spacings equation (18) reduces in the
present case to
Pn
n
ZN n (L i=1 i ) Y
f (j ) .
p(1 , . . . , n ) =
ZN (L)
j=1

g2 (t) =

(27)

1
1 g 00 (c)
1
+
t
2 g 0 (c)
+

3 g 00 (c)2 2 g 0 (c) g (3) (c) t


+ O(t2 ) , (32)
12 g 0 (c)2
2

E. Bogomolny et al.: Short-range plasma model for intermediate spectral statistics

where we have employed the saddle point equation (28).


At this point we specialize to the random matrix interaction potential, i.e. we substitute equation (17). Its
Laplace transform reads
( + 1)
g(t) =
t+1

(33)

125

Table 1. The spacing distributions P (s) and P (1, s) for the


one-dimensional gas model with nearest-neighbor interaction.
(k = 1)

=1

=2

2s

=4

2 3s

P (s)
4s e
27/2 s e
3125/24 s4 e5s
3 2s
5 3s
P (1, s) 8/3 s e
243/40 s e
1 953 125/72 576 s9 e5s

and the saddle point equation (28) yields


+1

c=

(34)

By Laplace inversion of equation (30) we then obtain the


nth nearest-neighbor spacing distribution
P (n, s) =

( + 1)(n+1)(+1) +n(+1) (+1)s


s
e
. (35)
(( + 1)(n + 1))

In particular, the nearest-neighbor spacing distribution for


= 1, 2, 4 is (see also Table 1)
P (s) = 4s e2s
27 2 3s
P (s) =
s e
2
3125 4 5s
P (s) =
s e
24

( = 1) ,
( = 2) ,
( = 4) .

(36)

The Laplace transform of the two-point correlation function equation (31) becomes
g2 (t) =

1
(1 +

t
+1
+1 )

(37)

from which R2 (s) can be computed by Laplace inversion.


R2 (s) = 1 exp(4s)


R2 (s) = 1 cos(3 3 s/2)




+ 3 sin(3 3 s/2) e9s/2


R2 (s) = e5s

4
X



exp 5s e2ik/5 + 2ik/5

plasma model with k = = 1 the Semi-Poisson statistics. This name was originally motivated from the fact
that if one takes an ordered Poisson distributed sequence
{xn }, the nearest-neighbor spacing distribution of the new
sequence {yn } with yn = (xn + xn+1 )/2 coincides with
equation (8) (of course the other correlation functions of
the sequence {yn } are different from those of the plasma
model).
It is interesting to note that again starting from the
sequence {xn } and dropping every second level one obtains a new sequence with precisely the same statistical
distribution as the plasma model with k = = 1 [16].
More generally, retaining only every (p + 1)th level of the
sequence {xn } leads to the same statistical distribution as
the plasma model with k = 1 and = p [16].
2.2 Interaction between nearest and next-to-nearest
neighbors (k = 2)
Next we consider the one-dimensional gas, where the interaction is restricted to nearest and next-to-nearest neighbors. In this case the partition function takes the form
Z
ZN (L) =
0

i=1

( = 2)

N
Y

f (j )f (j + j+1 ) .

( = 4).

As in the previous section we first compute the Laplace


transform of the partition function,
Z
gN (t) :=

dL ZN (L) etL

Z
=

Z
d1 . . .

C=

N


X
dN L
i

j=1

With help of equations (23) and (24) and the small t expansion of equation (37) we find the large L behavior of
the number variance 2 (L) L + C with
1
,
+1

d1 . . .

( = 1)

k=0

( + 2)

6(1 + )2

The spacing distribution and the two-point correlation


function of the simplest model with k = = 1 are very
close to the corresponding distributions of a number of
different dynamical systems, whose spectral statistics can
be calculated only numerically (see Sects. 3 and 4). In
lack of a deeper understanding of the spectral statistics
of these systems, the plasma model is valuable in that
it provides simple fitting distributions to these numerical
results. We propose to call the statistics derived from the

dN

N
Y

etj f (j )f (j + j+1 ) .

j=1

To evaluate this integral we introduce the transfer


operator [14]
p
p
0
K(, 0 ) = f () et/2 f ( + 0 ) f ( 0 ) et /2 , (38)
so that gN (t) can be rewritten as
gN (t) = tr K N
Z
Z
d1 . . . dN K(1 , 2 )K(2 , 3 ) . . .
0

. . . K(N 1 , N )K(N , 1 ). (39)

126

The European Physical Journal B

The operator K(, 0 ) is real symmetric and therefore admits the eigenbasis expansion
X
K(, 0 ) =
j j () j ( 0 )
(40)
j

with real eigenvalues j and eigenfunctions j (),


Z

d 0 K(, 0 ) j ( 0 ) = j j () .

To calculate the asymptotic behavior of the number variance we first take the Laplace transform of the nth
nearest-neighbor spacing distribution equation (20) which
yields
X  j (c + t/) n
g(n, t) =
0 (c)
j
 Z
2
t/2

d 0 (c; ) j (c + t/; ) e
. (48)

(41)

The eigenfunctions are normalized according to

Hence, by equation (19), the Laplace transform of the twolevel correlation function takes the form

Z
d j ()j0 () = j,j0

g2 (t) =

and we choose the ordering of the eigenvalues by decreasing magnitude (0 > 1 > . . . ). Consequently, in the large
N limit the Laplace transform of the partition function,
equation (39), reduces to
h
iN
,
(42)
gN (t) = 0 (t)
where we have explicitly indicated the t-dependence of
the eigenvalue. Again we perform the Laplace inversion in
saddle point approximation, which results in t being fixed
to t = c,
h
iN
ZN (L) 0 (c) eLc ,
(43)

X
j

 Z

00 (c)
=0.
0 (c)

(44)

Next we calculate the joint probability distribution of n


consecutive spacings, equation (18). It turns out to be
convenient to write the eigenfunctions in the form
p
(t; ) = f () et/2 (t; )
(45)
(again the dependence on t is explicitly indicated). We
express the Laplace transform of the integral on the right
hand side of equation (18) in terms of the transfer operator K, use the eigenfunction expansion equation (40),
perform the Laplace inversion in saddle point approximation (the saddle point equation is identical with Eq. (44)),
and arrive at
n
1
p(1 , . . . , n ) =
0 (c; 1 ) 0 (c; n ) ec i=1 i
0 (c)n1
n1
Y
f (n )
f (j )f (j + j+1 ) .
(46)

j=1

With the help of equation (46) we can now calculate all


correlation functions. The simplest is the nearest-neighbor
spacing distribution
h
i2
P (s) = 0 (c; s) .
(47)

d 0 (c; ) j (c + t/; ) e

t/2

2
. (49)

For small t this becomes


g2 (t) =

1
1 000 (c)

1 + O(t) ,
t
2 00 (c)

(50)

where we have used the saddle point equation (44) and


1

Z
d 0 (c, )2 = 1 ,
0

where c is determined from the saddle point equation


+

0 (c)
0 (c) j (c + t/)

which follows from equation (47)Rand the definition of the


mean level spacing (implying ds sP (s) = 1).
We now specialize to the random matrix interaction
equation (17) and restrict ourselves to the case of integer. The integral equation (41) with the eigenfunctions in
the form of equation (45) then reads
Z

d 0 et 0 ( 0 + ) j (t; 0 ) = j (t)j (t; ) .

(51)

It is clear from equation (51) that the t dependence factorizes and


j (t) = t21 j (1) ,
j (t; ) = t(+1)/2 (1; t) .

(52)

This permits us to determine c from the saddle point equation (44), which yields
c=

2 + 1

(53)

Using equations (23, 24, 50) we then find the asymptotic


behavior of the number variance,
2 (L) L ,

2 + 1

E. Bogomolny et al.: Short-range plasma model for intermediate spectral statistics

127

Table 2. Same as Table 1, but with nearest- and next-to-nearest-neighbor interaction.


(k = 2)
P (s) = s e(2+1)s

P cs
j

P (1, s) = s3+1 e(2+1)s

P ds
j

=1

=2

=4

c0 = 2.4773
c1 = 6.0681
c2 = 3.7159

c0 = 6.2603
c1 = 24.9958
c2 = 41.1181
c3 = 32.2768
c4 = 10.4386
d0 = 5.9771
d1 = 11.9326
d2 = 9.5151
d3 = 3.3018
d4 = 0.4746

c0 = 35.6018 , c1 = 252.5852
c2 = 826.3343 , c3 = 1630.9839
c4 = 2127.3725 , c5 = 1880.5911
c6 = 1103.6323 , c7 = 395.1819
c8 = 66.6078
d0 = 29.4495 , d1 = 104.4681
d2 = 169.5742 , d3 = 164.1739
d4 = 103.3668 , d5 = 43.2920
d6 = 11.9204 , d7 = 2.0002
d8 = 0.1586

d0 = 2.5054
d1 = 3.068
d2 = 0.7516

In order to determine the spacing distributions we actually need to solve the integral equation (51). It is straightforward to see that it has + 1 solutions each being a
polynomial of degree ,

(t; ) =

where x = (x1 , . . . , xk1 ) and x0 is defined likewise. One


then easily verifies that
gN (t) = tr K N
Z
Z
dx1 . . . dxN K(x1 , x2 )
K(x2 , x3 ) . . . K(xN , x1 ) .

aj j .

(54)

j=0

The coefficients aj are easily obtained by substituting


equation (54) into equation (51). The spacing distributions then follow from equations (46) and (20). Table 2
shows the explicit expression for the nearest and next-tonearest-neighbor spacing distribution for = 1, 2, and 4.

It is clear from the definition (56) that the transfer operator obeys the symmetry relation
K(x, x0 ) = K(x0 , xT ) ,
T

(57)

where we use the notation xT = (xk1 , . . . , x1 ). The presence of this symmetry permits the eigenbasis expansion
X
T
K(x, x0 ) =
j j (x) j (x0 ) ,
(58)
j

with the eigenvalues j and eigenfunctions j (x) obeying


Z
dx0 K(x, x0 ) j (x0 ) = j j (x)
(59)

2.3 General case


Finally we extend the calculation of the previous section
to an interaction between an arbitrary number k of neighboring particles. In this general case the Laplace transform
of the partition function (14) takes the form
Z
gN (t) =

Z
d1 . . .

dN
0

N
Y

etj

j=1

f (j )f (j + j+1 ) f (j + . . . + j+k1 ) . (55)


Again we seek to express gN (t) as the trace of the N th
power of a transfer operator. To this end we define [14]
K(x, x0 ) = (x2 x01 ) (x3 x02 ) (xk1 x0k2 )
p
p
p
etx1 /2 f (x1 ) f (x1 + x2 ) f (x1 + . . . + xk1 )
q
f (x1 + . . . + xk1 + x0k1 ) f (x01 + . . . + x0k1 )
q
q
0
f (x0k2 + x0k1 ) f (x0k1 ) etxk1 /2 ,
(56)

and the normalization


Z
dx j (x) j0 (xT ) = jj0 .

(60)

The above considerations show that in the limit of large


N the Laplace transform of the partition function, equation (55), can again be expressed as
h
iN
(61)
gN (t) = 0 (t) ,
where 0 is the largest eigenvalue of the transfer operator
(56) which parametrically depends on t. Hence the form of
the partition function and the saddle point equation will
also be the same as in the last section, see equations (43)
and (44).
In the following it will be useful to introduce the functions j (t; x) by writing j (t; x) in the form (again we
indicate the t-dependence explicitly)
p
j (t; x) = j (t; x) R(t; x) ,
(62)

128

The European Physical Journal B

where

which determines the asymptotic behavior of the number


variance, see equation (24). For n < k 2 we get

k1

X 
R(t; x) = exp t
xj

j=1

k2
Y kj1
Y
j=0

g (n, t) =
f (xi + . . . + xi+j ) .

...

j (t; 2 , . . . , k ) = j (t) j (t; 1 , . . . , k1 ) ,

(64)

which in conjunction with the normalization condition


dx R(t; x) j (t; x) j0 (t; xT ) = jj0

(65)

determines j (t; x) and j (t).


For the calculation of the joint probability distribution
of n consecutive spacings, equation (18), we follow the
same steps as in the last section, except that we have to
distinguish two cases, namely n < k 1 and n k 1. In
the former case,
Z
dk1

P
n

p( 1 , . . . , n ) =
, nk+2 ) e
 k1

nj
YY
f (i , . . . , i+j ) 0 (c; 1 , . . . , k1 ) . (67)
i=1

j=0 i=1

Equations (66) and (67) in conjunction with equation (20)


allow the calculation of the spacing distributions. For example, the nearest-neighbor spacing distribution becomes

dk1 0 (c; 1 , . . . , k1 ) 0 (c; k1 , . . . , 1 ) ,

i=1


 2
i
t
0 (c; 1 , . . . , k1 ) j c+ ; k1 , . . . , 1 .

k1

g(n, t) =

n=k2

1
1 000 (c)

(k 1) + O(t) ,
t
2 00 (c)

where
we have used the saddle point equation (44) and
R
ds sP (s) = 1 with P (s) given by equation (68). Since
the k 2 remaining terms each contribute 1 + O(t), we
finally get
g2 (t) =

1
1 000 (c)

1 + O(t) .
t
2 00 (c)

(69)

Note that this expression depends on k only by the largest


eigenvalue of the corresponding transfer operator and it is
identical with equation (50).
At this point we again specialize to the random matrix interaction equation (17) with integer . The integral
equation (64) which determines j (t) and j (t; x) then
reads
Z

dk k (k + k1 ) . . . (k + . . . + 1 ) etk

j (t; 2 , . . . , k ) = j (t) j (t; 1 , . . . , k1 ) . (70)

Z
dk1 (s j /)

d1 . . .

P (s) =

where c is determined from the saddle point equation (44),


while in the latter case,

dn+2 . . .

Summing over n and expanding asymptotically for small


t we obtain

0 (c; 1 , . . . , k1 ) 0 (c; k1 , . . . , 1 ) , (66)

n+k1
0 (c; n , . . .
0

Z
X  j (c + t/) nk+2 Z
d1 . . . dk1
g(n, t) =
0 (c)
j
t
2

i=1

whose behavior for small t is simply g(n, t) = 1 + O(t) due


to the normalization equation (60). For n k 2 we find

dn+1 . . .

dn+1 e

n+1

dk etk f (k ) f (k + k1 ) . . . f (k + . . . + 1 )

p(1 , . . . , n ) =

i=1

d1 . . .
0

(63)

From equation (59) we obtain the equation


Z

0 (c; 1 , . . . , k1 ) 0 (c; k1 , . . . , 1 ), (68)


where j may take any value from 1 to k 1. In order to determine the asymptotic behavior of the number variance,
we follow the same procedure as in the last section, i.e. we
calculate the Laplace transform g(n, t) of the nth nearestneighbor spacing distribution, sum this over n to obtain
the Laplace transform of the two-point correlation function, see equation (19), and find its small-t asymptotic,

The t-dependence of the eigenvalues and eigenfunctions


factorizes and
j (t) = tk1 j (1) ,
j (t; x) = t(k1)(k+2)/4 (1; tx) .

(71)

This permits us to determine c from the saddle point equation (44), which yields
c=

k + 1

(72)

E. Bogomolny et al.: Short-range plasma model for intermediate spectral statistics

Using equations (23, 24, 69) we then find the asymptotic


behavior of the number variance,
2 (L) L ,

1
k + 1

(L  1) .

(73)

In critical statistics this quantity is related with a certain


(multi)fractal dimension [17]
=

,
2d

(74)

where d is the dimensionality of the system and = d


D2 and D2 is one of the multifractional exponent defined
through the behavior of the mean inverse participation
ratio with the size of the system, L,
Z

dd x|n (x)|4 LD2 .
(75)
In our model geometrical interpretation of non-zero values
of (73) remains unclear.
In order to find the spacing distributions explicitly we
need to determine 0 and 0 from equation (70) and then
use equation (66) or (67) together with equation (20). The
solutions of equation (70) have the form
(t; 1 , . . . , k1 ) =
X
3
X
i1 =0 i2 =0

(k1)k/2

...

k1
ai1 i2 ...ik1 1i1 2i2 . . . k1
. (76)

129

where 2 (L) as above is the number variance.


Therefore the probability that there is no levels inside
this interval (i.e. the nearest-neighbor distribution) is


L2
.
(81)
P0 (L) exp 2
2 (L)
Assuming that, when L , 2 (L) L one gets
equation (79).
But it is clear that the assumption that Pn (L) has the
Gaussian form (80) even for small n is an oversimplification and cannot be strictly valid in general. In our shortrange plasma model Pn (L) is like the one for the Poissonian process which gives equation (78). In the Gaudin
model with = 2 there is no simple relation between
and (see Eqs. (10) and (11)) though for small one
obtains equation (79).
This difference between the Gaudin model when 0
and the relation (78) derived in our short range plasma
model is probably related with the different truncation of
the interaction between two levels. In the former model
the distance between two levels is important while in the
later one only the number of levels between the two given
levels is taken into account.
Numerical calculations of the Anderson model at the
MIT point and certain analytical arguments [25,26] based
on the kernel (13) are in the favor of the relation (79) at
least in the limit of small . The short range plasma model
demonstrates that there exist (mathematical) models of
intermediate statistics which do not obey this relation.

ik1 =0

The coefficients ai1 i2 ...ik1 must be determined numerically by substituting equation (76) into equation (70). In
Tables 1, 2, and 3 the explicit form of P (s) and P (1, s)
are shown for k = 1, 2, and 3 respectively, where in each
case the distributions for = 1, 2, and 4 are given.
These calculations become more and more tedious as
k or increase. However it is straightforward to find the
large s asymptotic of the spacing distributions,
P (n, s) es ,

= k + 1

(s  1) .

(77)

Note that in our model the asymptotics of the spacing distributions and the number variance are related as follows,
=

(78)

For critical distribution of of the Anderson model at MIT


point it is often assumed [18] that
=

(79)

To derive this relation it was assumed [18] that the probability to find n levels in an interval which contains in
average L levels has the Gaussian form


(n L)2
Pn (L) exp
,
(80)
2 2 (L)

3 Spectral statistics of pseudointegrable


billiards
Pseudointegrable systems, as introduced by Richens and
Berry [19], are dynamical systems which are neither integrable nor chaotic. The difference between integrable and
pseudointegrable systems is that for the former the phase
space is foliated into 2-dimensional surfaces with genus
g = 1 (i.e. tori) while for the latter it is foliated into surfaces of higher genus. The simplest example of pseudointegrable systems is plane polygonal billiards whose angles
are all rational multiples of . For these models the genus
of the corresponding surface is given by
g =1+

N X mk 1
,
2
nk

(82)

where mk /nk are the interior vertex angles of the polygon, N is the least common multiple of the integers nk ,
and the sum is taken over all vertices of the polygon. The
polygons with g = 1 (e.g. rectangle, equilateral triangle)
are integrable, whereas those with g 2 are pseudointegrable. In references [20,21] the spectral statistics of a
number of pseudointegrable polygonal billiards have been
investigated numerically (and in the last reference even experimentally using a microcavity) and it was found that
they display level repulsion but deviate from random matrix theory.

130

The European Physical Journal B


Table 3. Same as Table 1, but with three interacting neighbors.

P (s) = s e(3+1)s

P cs
j

P (1, s) = s3+1 e(3+1)s

P ds
j

=1

=2

=4

c0 = 2.1342
c1 = 7.8196
c2 = 11.6945
c3 = 9.0743
c4 = 3.6855
c5 = 0.6259

c0 = 4.9711
c1 = 31.6156
c2 = 93.9581
c3 = 171.4838
c4 = 212.8904
c5 = 188.1012
c6 = 120.0046
c7 = 54.6739
c8 = 17.0418
c9 = 3.2829
c10 = 0.2968
d0 = 1.6110
d1 = 8.8782
d2 = 23.0847
d3 = 37.4284
d4 = 42.0839
d5 = 34.4593
d6 = 20.9804
d7 = 9.5327
d8 = 3.2010
d9 = 0.7747
d10 = 0.1284
d11 = 0.0131
d12 = 6.2 104

c0 = 24.0711 , c1 = 282.9984
c2 = 1615.27 , c3 = 5952.78
c4 = 15890.11 , c5 = 32666.46
c6 = 53678.64 , c7 = 72222.20
c8 = 80836.91 , c9 = 76044.48
c10 = 60480.73 , c11 = 40759.87
c12 = 23247.21 , c13 = 11165.36
c14 = 4473.73 , c15 = 1472.98
c16 = 389.3105 , c17 = 79.6241
c18 = 11.8577 , c19 = 1.1466
c20 = 0.0541
d0 = 2.9196 , d1 = 29.5956
d2 = 146.1401 , d3 = 467.9231
d4 = 1090.9150 , d5 = 1971.2442
d6 = 2868.9524 , d7 = 3449.8348
d8 = 3488.0347 , d9 = 3001.6809
d10 = 2216.98 , d11 = 1412.97
d12 = 779.3162 , d13 = 372.3224
d14 = 153.8623 , d15 = 54.8050
d16 = 16.7526 , d17 = 4.3271
d18 = 0.94224 , d19 = 0.1687
d20 = 0.0241 , d21 = 2.76 103
d22 = 2.5 104 , d23 = 1.9 104
d24 = 6.2 107

d0
d1
d2
d3
d4
d5
d6

= 1.1340
= 3.6334
= 5.0189
= 3.7112
= 1.5037
= 0.3100
= 0.0259

In [10] a number of different models had been considered numerically which clearly demonstrate the existence
of intermediate statistics. In this section we present a more
detailed analysis of the data.
We consider the sequence of pseudointegrable billiards (with Dirichlet boundary conditions) in the shape
of the right triangle with one angle equal /n, where
n = 5, 7, 8, 9, . . . , 30 (the triangles with n = 3, 4, 6 are
integrable) and using the boundary integral method we
have calculated the first 20 00030 000 levels for all these
triangles (for some triangles even up to 80 000 levels).
Figure 2 shows the difference between the cumulative
nearest-neighbor spacing distribution
Z s
N (s) =
dx P (x)
(83)

0.05

0.00

dN(s)

(k = 3)

0.05

(calculated from the levels 500020 000) for all triangles


and the Semi-Poisson prediction
N1 (s) = 1 (2s + 1) exp(2s)

(84)

which is obtained from our model with = k = 1 (see


Eq. (36)).
Roughly four close groups of curves are observed. For
the first group, consisting of the curves corresponding to

GOE
0.10

Fig. 2. Difference between the cumulative spacing distribution


for the pseudointegrable rational right triangles with n 30
(calculated from the levels 500020 000) and the Semi-Poisson
curve. See explanation in the text.

E. Bogomolny et al.: Short-range plasma model for intermediate spectral statistics

131

0.10
0.005

0.05

0.00

0.05
0.005

0.10

0.01

Fig. 3. The same as in Figure 2 but for n = 12. The thin


line corresponds to 134 000 levels and the thick one includes
34 00168 000 levels.

triangles with n = 5, 8, 10, 12, this difference is less than


one percent and consequently the spectral statistics of
these billiards is quite well described by the Semi-Poisson
model. The remaining three groups correspond to the triangles with n = 7, 14, 18, with n = 9, 16, 20, 24, 30, and
the rest.
The first conclusion from this and many others figures is that spectral statistics of pseudointegrable systems
is not universal and depends on the billiard angles. The
grouping of the triangles which emerges from Figure 2 does
not agree with the classification of the billiards according
to their geometrical genus g. In first approximation the
spectral statistics of the above triangular billiards are reasonably well described by a quantity q which we proposed
to call arithmetical genus
(
g(n), n odd
q(n) =
(85)
(n)/2, n even.
Hence q is equal to the geometrical genus g for n odd
but for even n it is given by half the Euler function (n)
(equal to the number of integers not exceeding and relatively prime to n). The first group now correspond to the
triangles with q = 2, the second to q = 3, the third to
q = 4, and the last to q > 4. It seems that for q > 4 the
spacing distribution does not change noticeably any more,
but the resulting distribution may differs from the WignerDyson spacing distribution. We stress that this classification is only an approximative one and serves mostly for
the crude arrangement of the spectral statistics of different
triangles.
In Figure 3 we present the evolution of the cumulative nearest-neighbor distribution for the triangular billiard with n = 12 with increasing energy. It is clearly
seen that for higher energy the spacing distribution moves

0.15

Fig. 4. The difference between the cumulative next-to-nearest


spacing distribution for n = 12 and the Semi-Poisson prediction. Two curves are the same as in Figure 3. The dotted line
corresponds to GOE.

closer to the Semi-Poisson result (though a limiting distribution may deviates from the Semi-Poisson one).
In Figure 4 the difference between the cumulative nextto-nearest distribution for the same triangle and the SemiPoisson prediction (see Tab. 1) is plotted. Once more we
observed that this distribution is close to the Semi-Poisson
result and that, with increasing energy, the agreement is
better.
The main conclusion of this section is that for certain
pseudointegrable systems the short-range spectral statistics is very close to the Semi-Poisson statistics (though
theoretical explication of this fact is not yet clear).

4 Summary and discussion


In order to model intermediate spectral statistics, we considered a one-dimensional gas of energy levels interacting
via a logarithmic pair potential, whose action is restricted
to a small number of nearest neighbors. As shown in Section 2, its correlation functions can be calculated analytically, so the model deserves interest as a simple reference point for comparisons with numerics. In Section 3
we performed a comparison with the spectral statistics of
a number of pseudointegrable billiards and demonstrated
that the plasma model with nearest-neighbor interaction
(the Semi-Poisson model) gives an excellent phenomenological description of the short-range spectral observables
of certain pseudointegrable systems. Unfortunately a full
theoretical understanding of the spectral statistics of pseudointegrable systems is still lacking.
Of course there are many ways to interpolate between Wigner-Dyson and Poisson statistics. For the spacing distribution a well-known interpolation is the Brody

132

The European Physical Journal B

distribution [22], which shows fractional level repulsion.


Since for intermediate spectral statistics P (s) s for
small spacings, the Brody distribution cannot give an adequate description. A very natural way to construct interpolating ensembles is to take a weighted average between
the ensemble of diagonal random matrices and one of the
standard random matrix ensembles, e.g. for = 1
H = HPoisson + HGOE

(86)

(see e.g. Ref. [23]). However, as we checked numerically, for


no value of the resulting distribution is close to the distribution (8), which we have found to be an excellent fit to
the spacing distribution of the pseudointegrable triangles
with q = 2 and also to that of several other systems [10].
Hence this interpolation do not seem to be suitable for the
description of intermediate statistics.
All data suggest that a necessary requirement of theoretical description of intermediate statistics is the screening of two-body potential. Its precise form is not yet
known (but see [2426]) and may depend on the problem
considered.

The authors is greatly indebted to A. Pandey who had investigated the short-range plasma model in the framework of band
random matrices and whose unpublished notes were useful to
check our calculations.
It is a pleasure to thank O. Bohigas, D. Delande, A. Mirlin
and G. Montambaux for useful discussions. UG acknowledges
a HSP2-fellowship of the Deutscher Akademischer Austauschdienst, and thanks the IPN Orsay for kind hospitality during
an extended stay.

References
1. C.E. Porter, Statistical theories of spectra: fluctuations
(Academic Press, New York, 1965).

2. M.L. Mehta, Random matrices, 2nd ed. (Academic Press,


New York, 1991).
3. O. Bohigas in Chaos and Quantum Physics, Proceedings
of the Les Houches Summer School of Theoretical Physics,
1989, edited by M.-J. Giannoni, A. Voros, J. Zinn-Justin
(Elsevier, New York, 1991).
4. T. Guhr, A. M
uller-Groeling, H.A. Weidenm
uller, Phys.
Rep. 299, 190 (1998).
5. E. Bogomolny, O. Bohigas, M.P. Pato, Phys. Rev. E 55,
6707 (1997).
6. B.I. Shklovskii et al., Phys. Rev. B 47, 11487 (1993).
7. Y.V. Fyodorov, A.D. Mirlin, Phys. Rev. B 55, R16001
(1997).
8. M.V. Berry, M. Tabor, Proc. R. Soc. London A 356, 375
(1977).
9. O. Bohigas, M.-J. Giannoni, C. Schmit, Phys. Rev. Lett.
52, 1 (1984); ibid. J. Phys. Lett. 45, L1015 (1984).
10. E. Bogomolny, U. Gerland, C. Schmit, Phys. Rev. E 59,
R1315 (1999).
11. M. Gaudin, Nucl. Phys. 85, 545 (1966).
12. P.J. Forrester, Phys. Lett. A 173, 355 (1993).
13. F. G
ursey, Proc. Cambridge Phil. Soc. 46, 182 (1950).
14. L. van Hove, Physica 16, 137 (1950).
15. E.H. Lieb, D.C. Mattis, Mathematical Physics in One Dimension (Academic Press, New York, 1966).
16. H. Hern
andez-Salda
na, J. Flores, T.H. Seligman, Phys.
Rev. E 60, 449 (1999).
17. J.T. Chalker, V.E. Kravtsov, I.V. Lerner, JETP Lett., 64,
386 (1996).
18. B. Altshuler, I.Kh. Zharekeshev, S.A. Kotochigova, B.I.
Shklovskii, Sov. Phys. JETP, 67, 625 (1988).
19. P.J. Richens, M.V. Berry, Physica D 2, 495 (1981).
20. A. Shudo, Y. Shimizu, Phys. Rev. E 47, 54 (1993).

21. A. Shudo, Y. Shimizu, P. Seba,


J. Stein, H.J. St
ockmann,
K. Zyczkowski, Phys. Rev. E 49, 3748 (1994).
22. T.A. Brody, Lett. Nuov. Cim. 7, 482 (1973).
23. T. Guhr, H.A. Weidenm
uller, Ann. Phys. 193, 472 (1989).
24. V.E. Kravtsov, I.V. Lerner, J. Phys. A 28, 3623 (1995).
25. S.M. Nishigaki, Phys. Rev. E 59, 2853 (1999).
26. V.E. Kravtsov, A.M. Tsvelik, Phys. Rev. B 62, 9888
(2000), cond-mat/0002120.

You might also like