Berry’s conjecture and information theory

C. Jarzynski
Theoretical Astrophysics, T-6, MS B288

Los Alamos National Laboratory

Los Alamos, NM 87545

arXiv:chao-dyn/9703014v1 21 Mar 1997

chrisj@t6-serv.lanl.gov

(October 23, 2018)

Abstract

It is shown that, by applying a principle of information theory, one obtains

Berry’s conjecture regarding the high-lying quantal energy eigenstates of clas-

sically chaotic systems.

Typeset using REVTEX

1
 In many problems of physical interest, it is necessary to abandon a search for the exact
solution, and to turn instead to a statistical approach. This involves mentally replacing the
answer which we seek, with an ensemble of possibilities, then adopting the attitude that each

member of the ensemble is an equally likely candidate for the true solution. The choice of
ensemble then becomes centrally important, and here information theory provides a reliable
guiding principle. The principle instructs us to choose the least biased ensemble (the one
which minimizes information content), subject to some relevant constraints. A well-known

illustration arises in classical statistical mechanics: the least biased distribution in phase
space, subject to a fixed normalization and average energy, is the canonical ensemble of Gibbs
[1]. Another example appears in random matrix theory: by minimizing the information
content of an ensemble of matrices, subject to various simple constraints, one obtains the

standard random matrix ensembles [2]. The purpose of this paper is to point out that
Berry’s conjecture [3] regarding the energy eigenstates of chaotic systems, also emerges
naturally from this principle of least bias.
Berry’s conjecture makes two assertions regarding the high-lying energy eigenstates ψE of

quantal systems whose classical counterparts are chaotic and ergodic1 : (1) Such eigenstates
appear to be random Gaussian functions ψ(x) on configuration space, (2) with two-point
correlations given by
! !
s s 1
Z
ψ∗ x− ψ x+ = dp eip·s/h̄ δ[E − H(x, p)] (1)
2 2 Σ

Here, E is the energy of the eigenstate, H(x, p) is the classical Hamiltonian describing
the system, and Σ ≡ dx dp δ[E − H(x, p)]; if the Hamiltonian is time-reversal-invariant,
R R

then ψ(x) is a real random Gaussian function, otherwise ψ(x) is a complex random Gaussian
function. Berry’s conjecture thus uniquely specifies, for a given energy E, an ensemble ME
of wavefunctions ψ(x) (i.e. ME is the Gaussian ensemble with two-point correlations given

1 By “chaotic and ergodic”, we mean that all trajectories, except a set of measure zero, chaotically

and ergodically explore the surface of constant energy in phase space.

2
by Eq.1), and states that an eigenstate ψE at energy E will look as if it were chosen randomly
from this ensemble.
The correlations given by Eq.1 are motivated by considering the Wigner function [4]

corresponding to the eigenstate ψE ,

! !
s s −ip·s/h̄
Z
WE (x, p) ≡ (2πh̄)−D ds ψE∗ x− ψE x + e , (2)
2 2

where D is the dimensionality of the system. For high-lying states ψE , this Wigner func-
tion, after local smoothing in the x variable, is expected to converge to the microcanonical
distribution in phase space [3,5–7]:

1
WEsm (x, p) ≈ δ[E − H(x, p)]. (3)
Σ

By taking the Fourier transform of both sides of Eq.1, and then smoothing locally in the
x-variable2 rather than averaging over the ensemble ME , it is straightforward to show that
the correlations given by Eq.1 produce the desired result, Eq.3.

The assertion that ψE (x) is a Gaussian random function is most easily motivated by
viewing ψE (x), locally, as a superposition of de Broglie waves with random phases [3].
When the number of these waves becomes infinite, the central limit theorem tells us that
ψE (x) will look like a Gaussian random function.
We can interpret Berry’s conjecture as making a specific prediction about the eigenstate

ψE : once we compute ψE (x) by solving the Schrödinger equation, we can subject it to

various tests (see e.g. Ref. [8]), and we will observe that, yes, ψE (x) really behaves as a
Gaussian random function with two-point correlations given by Eq.1. Alternatively, we can
interpret Berry’s conjecture as providing us with the appropriate ensemble of wavefunctions

from which to choose a surrogate for the true eigenstate ψE , if we cannot (or do not care

2 This smoothing is performed on a scale which is large compared with the local correlation length

of ψ(x), but small compared with a classically relevant distance scale (see e.g. equations 3.29 and

3.30 of Ref. [7]). This allows us to replace ensemble averaging with local smoothing.

3
to) actually solve for ψE (x). In this interpretation, ME stands to ψE much as, in classical
statistical mechanics, the canonical ensemble stands to the instantaneous microscopic state
of a system at a given temperature. It is within the context of the second point of view that

we will show that Berry’s conjecture may be “derived” from information theory. Specifically,
we will show that, by applying the principle of least bias, and accepting the correlations given
by Eq.1 as a set of relevant constraints, we are led immediately to a statement of Berry’s
conjecture.

We thus pose the following question. Suppose we have a quantal Hamiltonian Ĥ, whose
classical counterpart H(x, p) is chaotic and ergodic; and suppose we are told that a high-
lying eigenstate of Ĥ — represented by a wave function ψE (x) — exists at energy E. Given
this limited knowledge, how to we go about making a “best guess” for ψE (x)? By a best

guess, we mean not a single wave function, but rather a probability distribution PE [ψ] in
Hilbert space, such that, by sampling randomly from this distribution, we are making a
guess which takes into account our limited knowledge regarding ψE , but is otherwise unbi-
ased. Information theory provides a general prescription for constructing such a distribution.

First, we quantify the information I contained in an arbitrary distribution P [ψ]. Next, we

identify the constraints on P [ψ] imposed by our limited knowledge. Finally, we minimize
I{P [ψ]} subject to these constraints. The resulting distribution PE [ψ] is the least biased
one, consistent with our limited knowledge. Let us now implement this procedure.
Given a probability distribution P [ψ] in Hilbert space, the amount of information I

contained in this distribution is:

Z
I{P [ψ]} = P [ψ] ln P [ψ]. (4)

The integral is over all square-integrable functions ψ(x), where ψ(x) is taken to be real if

the Hamiltonian is time-reversal-invariant, and complex otherwise. [The integral in Eq.4

requires a measure dµ on Hilbert space. We take the usual Euclidean measure of field
theories [9]: a wavefunction is represented by its value at N discrete points in configuration
space, and the set of these values is regarded as a (real or complex) Cartesian vector. Hence,

4
dµ = dψ1 dψ2 · · · dψN , where ψi ≡ ψ(xi ). The limit N → ∞ is finally taken.]
Since we will want to minimize I{P } subject to relevant constraints on the distribution
P [ψ], our next task is to identify those constraints. The first is simply that P ought to be

normalized to unity:
Z
P [ψ] = 1. (5)

The second constraint is that embodied by Eq.1:

1
Z Z
P [ψ] ψ ∗ (x1 )ψ(x2 ) = dp eip·s/h̄ δ[E − H(x, p)], (6)
Σ

where s ≡ x2 − x1 and x ≡ (x1 + x2 )/2. (Both x1 and x2 are assumed to be within

the classically allowed region; outside this region, the wavefunction is taken to be zero.) As
explained briefly above (see also Refs. [3,6,7]), this constraint is motivated by the expectation
that the smoothed Wigner function corresponding to ψE will approximate the microcanonical

distribution in phase space (Eq.3). Note that Eq.6 does not represent a single constraint,
but rather a set of constraints, where each member of the set is specified by (x1 , x2 ).
Finally, we minimize the information I{P [ψ]}, subject to the constraints in Eqs.5 and 6.
We do this in the usual way, by introducing Lagrange multipliers. That is, we define
Z ZZ Z
A{P [ψ]} ≡ I{P [ψ]} + λ P [ψ] + Λ(x1 , x2 ) P [ψ] ψ ∗ (x1 )ψ(x2 )
Z ZZ !
= P [ψ] ln P [ψ] + λ + Λ(x1 , x2 )ψ (x1 )ψ(x2 ) ,
∗
(7)

where λ is the Lagrange multiplier associated with Eq.5, Λ(x1 , x2 ) is the set of multipliers

associated with Eq.6, and is shorthand for dx1 dx2 . For a given distribution P [ψ], the
RR R R

change in A induced by a small variation δP [ψ] is, to first order in δP [ψ],

Z ZZ !
δA = δP ln P + (λ + 1) + Λ(x1 , x2 )ψ (x1 )ψ(x2 ) .
∗
(8)

To minimize A (i.e. to minimize I subject to the constraints imposed by Eqs.5 and 6) we

insist that δA = 0 for all variations δP . From Eq.8, it follows that the distribution PE [ψ]
which accomplishes this minimization has the form:

5
 ZZ
PE [ψ] = N exp − Λ(x1 , x2 )ψ ∗ (x1 )ψ(x2 ). (9)

The constant N is determined by normalization (Eq.5), whereas the two-point correlations

(Eq.6) uniquely determine Λ(x1 , x2 ). (Specifically, the kernel Λ(x1 , x2 ) is just the inverse of
ψ ∗ (x1 )ψ(x2 ), divided by 2 if H is time-reversal-invariant [10].)

Once N and Λ(x1 , x2 ) are determined, Eq.9 completely specifies a probability distri-
bution PE on Hilbert space. By randomly sampling from this distribution, we generate
a random function ψ(x), with two-point correlations ψ ∗ (x1 )ψ(x2 ) which (by construction)
satisfy Eq.1. But is a function sampled from PE a Gaussian random function? The answer

is yes [8,10]. For a random function f (x), let Pn (f1 , · · · , fn ) denote the joint probability
distribution of finding that f (xi ) = fi , i = 1, · · · , n. Then f (x) is Gaussian if Pn is a Gaus-
sian in (f1 , · · · , fn )-space, for any (x1 , x2 , · · · , xn ), n ≥ 1 [11]. A function ψ(x) obtained by
sampling the probability distribution given by Eq.9 satisfies this condition.

We thus arrive at the following conclusion: given the limited knowledge that an eigenstate
of Ĥ exists at energy E, the least biased guess for ψE (x) (by reasonable construction) is a
Gaussian random function, with two-point correlations given by Eq.1. This is just another
way of stating Berry’s conjecture. (Instead of saying that ψE “looks like” a Gaussian random
function, we say that a Gaussian random function is a “best guess” for ψE .) In this sense,

Berry’s conjecture allows for a statistical description of eigenfunctions, by providing us with

the appropriate ensemble ME in Hilbert space to use as a stand-in for the true energy
eigenstate ψE . When the latter is unobtainable3 calculations performed with ME may be
tractable [12], just as the canonical ensemble of ordinary statistical mechanics makes possible

accurate computations without demanding detailed knowledge of the microscopic state of

the system.
The notion that Berry’s conjecture is gainfully viewed as a statistical theory — in analogy

3 Note that the computational effort required to solve Schrödinger’s equation grows exponentially

with the dimensionality N of configuration space.

6
with classical statistical mechanics, or random matrix theory — has been a guiding theme
of this paper. As stressed in the opening paragraph, the first order of business with such
theories is to identify the proper ensemble to use in place of an exact description of the object

of study (be it the microscopic state of a many-body system, or a complicated Hermitian

matrix, or a quantal eigenstate). A common feature of statistical theories is that this
ensemble follows in a natural way from the information theoretic principle of least bias.
The purpose of this paper has simply been to point out that Berry’s conjecture shares this

feature.

ACKNOWLEDGMENTS

The author wishes to acknowledge useful conversations and correspondence with S.Jain,
J.Morehead, and M.Srednicki.

7
 REFERENCES

[1] D.Chandler, Introduction to Modern Statistical Mechanics, section 3.7 (Oxford Uni-
versity Press, New York, 1987). Note that minimizing the information content of the

ensemble is the same as maximizing its entropy.

[2] R.Balian, Il Nuovo Cimento 57B, 183 (1968). The author would like to thank S.Jain

for bringing this reference to his attention.

[3] M.V.Berry, J.Phys.A 10, 2083 (1977).

[4] E.Wigner, Phys.Rev. 40, 749 (1932).

[5] A.I.Shnirelman, Ups.Mat.Nauk 29, 181 (1974).

[6] A.Voros, Ann.Inst.Henri Poincaré A 24, 31 (1976); 26, 343 (1977); in Stochastic Be-
havior in Classical and Quantal Hamiltonian Systems, ed. by G.Casati and J.Ford

(Springer-Verlag, Berlin, 1979).

[7] M.V.Berry, in Les Houches XXXVI, Chaotic Behavior of Deterministic Systems,

G.Iooss, R.H.G.Helleman, and R.Stora, eds. (North-Holland, Amsterdam, 1983).

[8] M.Srednicki and F.Stiernelof, J.Phys.A 29, 5817 (1996), chao-dyn/9603012.

[9] L.S.Brown, Quantum Field Theory, Chapter 1 (Cambridge University Press, 1992).

[10] M.Srednicki, Phys.Rev.E 54, 954 (1996), cond-mat/9512115.

[11] R.Kubo, M.Toda, N.Hashisume, Statistical Physics II: Nonequilibrium Statistical Me-
chanics, section 1.4 (Springer-Verlag, Berlin, 1985).

[12] See e.g. the use of Berry’s conjecture to derive eigenstate thermalization. M.Srednicki,
Phys.Rev.E 50, 888 (1994), cond-mat/9403051.