Quantum chaos in billiards

Arnd Bäcker

Institut für Theoretische Physik, Technische Universität Dresden, D-01062 Dresden, Germany

Abstract
Billiards are an important class of systems showing a large variety of dynamical behaviour ranging
from integrable (i.e. regular) motion, over mixed dynamics to strongly chaotic behaviour. Using
tools developed with Python one can interactively study the complexity of the dynamics. This
dynamical behaviour is directly reflected in properties of the corresponding quantum systems, like
eigenvalue statistics or the structure of eigenfunctions.

Chaotic behaviour in dynamical systems is a well-studied phenomenon. A particularly illustrative

class of systems are so-called billiard systems where a point particle moves freely along straight lines
inside a two-dimensional domain Ω with elastic reflections at the boundary. In such systems it is
therefore exclusively the boundary which determines the dynamical properties. This is illustrated
in Fig. 1 where 50 iterations of one initial point are shown for two billiards, parametrized in polar
coordinates by ρ(ϕ) = 1 + ε cos(ϕ) with ϕ ∈ [0, 2π], for parameters ε = 0 (circular billiard) and
ε = 1 (cardioid billiard) [1]. The circular billiard is an example of an integrable system showing
regular dynamics. The opposite extreme is the cardioid billiard which is fully chaotic In particular
this means that nearby trajectories separate exponentially as a function of time (hyperbolicity)
and that a typical trajectory will fill out the available space in a uniform way (ergodicity).
As the motion inside the billiard is on straight lines it is convenient to use the boundary to
define a Poincaré section,

P := {(s, p) | s ∈ [0, |∂Ω|], p ∈ [−1, 1]} . (1)

Figure 1: Regular dynamics in a billiard with circular shape vs. chaotic dynamics
in the cardioid billiard.
Here s is the arclength along the boundary ∂Ω and p = hv, T (s)i is the projection of the unit
velocity vector v after the reflection on the unit tangent vector T (s) in the point s ∈ ∂Ω. The
Poincaré map P of a point ξ = (s, p) ∈ P is then obtained by considering the ray starting at the
point r(s) ∈ ∂Ω in the direction specified by p and determining the first intersection with the
0 0 0
boundary, leading to the new p point ξ = (s , p ). Explicitly, the velocity in the T , N coordinate
system is given by (p, n = 1 − p2 ), so that in Cartesian coordinates
µ ¶ ³ p p ´
T x Nx 2 2
v = (vx , vy ) = (p, n) = Tx p + Nx 1 − p , Ty p + Ny 1 − p . (2)
T y Ny

Numerically the main task is to find the next intersection for a given starting point on the boundary
and direction, specified by s and p. If the boundary is determined by an implicit equation

F (x, y) = 0 , (3)

the new point r 0 can be determined by solving

F (x + tvx , y + tvy ) = 0 (4)

for t > 0. In the case of the circular billiard one easily obtains an analytical solution leading to
an explicit prescription of the billiard mapping. In general, however, only a numerical solution of
(4) is possible.
For non–convex billiards there are points ξ = (s, p) ∈ P for which there is more than one
solution (apart from t = 0); obviously the one with the smallest t > 0 has to be chosen. The
condition (3) can sometimes be used to remove the t = 0 solution analytically from (4). If F is
a polynomial in x and y this allows to reduce the order of (4) by one. This approach has for

Figure 2: Screenshot of a Python application to interactively explore billiard sys-

tems.

2
example been used for the cardioid billiard leading to a cubic equation for t, see [2] for details.
From the solution t one gets the coordinates (x0 , y 0 ) = (x, y) + tv.
Numerically, one has to find one or several solutions of (4), depending on the type of boundary.
Therefore one typically tries to bracket the zero with the smallest t, e.g. by evaluating F (x+tvx , y+
tvy ) for sufficiently many values of t, and then use scipy.optimize.brentq to determine the zero
as precisely as possible. Special attentention has to be given to glancing motion, e.g. when p is
near ±1 because then t can get very close to 0, or in non-convex billiards, zeros of F (x+tvx , y+tvy )
can become very close to each otherand can therefore be easily missed.
Both for research and teaching it proved extremely useful to interactively explore the dynamics
in billiards using the visualization of trajectories specified by their initial condition in the Poincaré
section using the mouse. For this purpose we developed an application written in Python, using
wxPython [3] for the graphical user-interface and for fast plotting of many points a particular
canvas [4]. Fig. 2 shows a typical screenshot, where the above billiard family is considered for
ε = 0.3. In this case one has a system which shows both regular and irregular motion, depending
on the initial point.
While classical mechanics describes macroscopic objects correctly, at small scales a quantum
mechanical description is necessary. Due to the Heisenberg uncertainty principle it is no longer
possible to specify both position and momentum of a particle at the same time. Instead the
state of a particle is specified by a wave function whose absolute value squared is interpreted as
the probability density. For quantum billiards finding the stationary solutions of the Schrödinger
equation reduces to the determination of eigenvalues and eigenfunctions of the Helmholtz equation

−∆ψn (q) = En ψn (q) , q∈Ω (5)

with (for example) Dirichlet boundary conditions, i.e. ψ³n (q) = 0 for ´ q ∈ ∂Ω. Here ∆ denotes the
∂2 ∂2
Laplace operator, which reads in two dimensions ∆ = ∂q2 + ∂q2 . The interpretation of ψ is
R 1 2

that D |ψ(q)|2 d2 q is the probability of finding the particle inside the domain D ⊂ Ω.
For some simple domains Ω it is possible to solve eq. (5) analytically. For example for
the billiard in a rectangle with sides a and b the (non–normalized) eigenfunctions are given by
ψn1 ,n2 (q) = sin(πn1 q1 /a) sin(πn2 q2 /b) with corresponding eigenvalues En1 ,n2 = π 2 (n21 /a2 + n22 /b2 )
and (n1 , n2 ) ∈ N 2 . For the billiard in a circle the eigenfunctions are given in polar–coordinates
by ψmn (r, ϕ) = Jm (jmn r) exp(imϕ), where jmn is the n–th zero of the Bessel function Jm (x) and
m ∈ Z, n ∈ N. Using scipy.special.jn_zeros one can easily obtain a given number of zeros of
Jm (x) for fixed m.
However, in general no analytical solutions of eq. (5) exist so that numerical methods have
to be used to compute eigenvalues and eigenfunctions. Among the many different possibilities,
the so-called boundary-integal method is widely used, see [5] for a more detailed account and
references. By employing Green’s theorem the two-dimensional problem (5) is transformed into
a one-dimensional integral equation. Discretization leads to a matrix equation for which the
zeros of a determinant as a function of the energy E have to be determined. To detect nearly
degenerate energy levels it turned out to be most efficient [6] to use a singular value decomposition.
Concerning the numerical implementation this is the most time consuming step, which is provided
by scipy.linalg.flapack.zgesdd. Using this approach 2000-10000 (and more if necessary)
eigenvalues and eigenfunctions can easily be computed. In particular, due to the independence of
the computation at different E, this problem can be straight-forwardly parallelized using as many
CPUs as available. Communication between different CPUs is essentially not needed, only the
initial value of E has to transferred to each CPU (e.g. using MPI).

3
 One of the fundamental questions in quantum chaos concerns the implications of the un-
derylying classical dynamical properties on the statistical behaviour of eigenvalues. It has been
conjectured that for fully chaotic systems these are described by the statics of random matrices
obeying appropriate symmetries [7]. For generic integrable systems one expects that the energy
level statistics can be described by a Poissonian random process [8].
The simplest spectral statistics is the level spacing distribution P (s) obtained from the his-
togramm of the spacings
sn := xn+1 − xn , (6)
where xn are rescaled eigenvalues such that their average spacing is 1. Once the eigenvalues are
computed, the determination of the level-spacing distribution can be done as follows
from pylab import *

spacings = x[1:]-x[0:-1] # x: rescaled eigenvalues

hist(dat, normed=True, bins=100)
show()
The resulting distribution has to be compared with the expectation for integrable systems,

PPoisson (s) = e−s . (7)

Since P (s) → 1 for s → 0 this behaviour is called level attraction. The result for the level-spacing
distribution in the case of the Gaussian orthogonal random matrix ensemble (GOE) is in very
good approximation described by the Wigner distribution [9]
π ³ π ´
PGOE (s) ≈ PWigner (s) = s exp − s2 . (8)
2 4
In this case we have P (s) → 0 for s → 0, which is called level repulsion. To illustrate this
behaviour, we show in fig. 3 the level-spacing distribution for the circle and for the cardioid
billiard and observe good agreement with the expected distributions.
Concerning the eigenfunctions of (5) one would expect that the classical dynamics is reflected
by their structure. According to the “semiclassical eigenfunction hypothesis” the eigenstates
should concentrates on those regions which a generic orbit explores in the long–time limit [10–12].

(a) (b)
1.0 1.0
GOE GOE
P (s) exp(−x) P (s) exp(−x)
circle billiard cardioid billiard
0.5 0.5

0.0 0.0
0 1 2 3 s 4 0 1 2 3 s 4

Figure 3: Level–spacing distribution for (a) circle billiard (100 000 eigenvalues) and
(b) cardioid billiard (11 000 eigenvalues). On observes good agreement with the
expected behaviour of a Poissonian random process and of the GOE, respectively

4
 n = 100 n = 1000 n = 1500 n = 2000
Regular billiard

Chaotic billiard

Figure 4: The eigenstates of the integrable circular billiard and the chaotic cardioid
billiard reflect the structure of the corresponding classical dynamics.

For integrable systems the motion is restricted to invariant tori while for ergodic systems the
whole energy surface is filled in a uniform way. For the case of ergodic systems this statement is
actually proven by the quantum ergodicity theorem (see [13] for an introduction and references),
which states that almost all eigenfunctions become equidistributed in the semiclassical limit, e.g.
restricted to position space we have
Z
vol(D)
lim |ψnj (q)|2 d2 q = (9)
j→∞ vol(Ω)
D

for a subsequence {ψnj } ⊂ {ψn } of density one. So for almost all eigenfunctions the probability
of finding a particle in a certain region D of the position space Ω in the semiclassical limit is just
the same as for the classical system.
Fig. 4 illustrates this for the case of the integrable circle billiard and the chaotic cardioid
billiard. One clearly sees that in the former case the probability is restricted to subregions of the
billiard, while for the ergodic case the probability density is uniformly distributed over the full
billiard region (apart from the inevitable fluctuations).
For systems with a mixed phase space the dynamics is more complicated, because both regular
motion and chaotic motion coexist, see Fig. 2. This is also reflected in the structure of the quantum
eigenstates, which are either located in the regular islands, or extend over the chaotic region, see
fig. 5
To conclude let us mention that in addition to the fundamental questions on the quantum
behaviour of systems with classically chaotic dynamics, such systems are also investigated experi-
mentally like for example microwave cavities, optical cavities (microlasers) and mesoscopic devices
(quantum dots).
From the numerical side, our experiences with using Python for research purposes is extremely
positive. When thinking of scientific computing typically Fortran or C/C++ come to ones mind
first for maximum performance. However many tasks involve fairly small amounts of time-critical
code, so that both development time and programm lengths are substantially reduced. On the

5
 1

-1
-4 0 s 4

-1
-4 0 s 4

-1
-4 0 s 4

Figure 5: Eigenstates in billiard with mixed phase either concentrate in the regular
islands (first two lines), or extend over the chaotic region (last line). This is most
clearly seen in the quantum Poincaré Husimi representation displayed in the last
column for each case.

other hand, due to the efficient usage of numerical libraries, no significant speed reduction arises
in our applications.
We would also like to point out that all the illustrations shown here involve Python. For
example for the two-dimensional graphics shown in Figs. 3 and 5 we have used PYX [14] (together
with PyXGraph to simplify plotting). For the three-dimensional visualization of eigenstates MayaVi
[15] is used with subsequent rendering.

Acknowledgements

I would like to thank L. Bittrich for useful comments on the manuscript and L. Bittrich, S. Löck,
N. Hlubek and J.-M. Braun for their contributions to the Iterator shown in Fig. 2.

6
References
[1] M. Robnik: Classical dynamics of a family of billiards with analytic boundaries, J. Phys. A
16 (1983) 3971–3986.

[2] A. Bäcker and H. R. Dullin: Symbolic dynamics and periodic orbits for the cardioid billiard ,
J. Phys. A 30 (1997) 1991–2020.

[3] wxPython: www.wxpython.org.

[4] A. Bäcker and N. Hlubek: PlottingCanvas, see www.physik.tu-dresden.de/~baecker.

[5] A. Bäcker: Numerical aspects of eigenvalue and eigenfunction computations for chaotic quan-
tum system, in: The Mathematical Aspects of Quantum Maps, M. Degli Esposti and S. Graffi
(Eds.), Springer Lecture Notes in Physics 618 (2003) 91–144.

[6] R. Aurich and F. Steiner: Statistical properties of highly excited quantum eigenstates of a
strongly chaotic system, Physica D 64 (1993) 185–214.

[7] O. Bohigas, M.-J. Giannoni and C. Schmit: Characterization of chaotic quantum spectra and
universality of level fluctuation laws, Phys. Rev. Lett. 52 (1984) 1–4.

[8] M. V. Berry and M. Tabor: Level clustering in the regular spectrum, Proc. R. Soc. London
Ser. A 356 (1977) 375–394.

[9] M. L. Mehta: Random Matrices, Academic Press, San Diego, revised and enlarged second
edn., (1991).

[10] M. V. Berry: Regular and irregular semiclassical wavefunctions, J. Phys. A 10 (1977) 2083–
2091.

[11] A. Voros: Semi-classical ergodicity of quantum eigenstates in the Wigner representation, in:
Stochastic Behavior in Classical and Quantum Hamiltonian Systems, no. 93 in Lecture Notes
in Physics, 326–333, Springer-Verlag, Berlin, (1979).

[12] M. V. Berry: Semiclassical mechanics of regular and irregular motion, in: Comportement
Chaotique des Systèmes Déterministes — Chaotic Behaviour of Deterministic Systems (Eds.
G. Iooss, R. H. G. Hellemann and R. Stora), 171–271, North-Holland, Amsterdam, (1983).

[13] A. Bäcker, R. Schubert and P. Stifter: Rate of quantum ergodicity in Euclidean billiards,
Phys. Rev. E 57 (1998) 5425–5447.

[14] PYX: pyx.sourceforge.net.

[15] MayaVi: mayavi.sourceforge.net.