Home

Search

Collections

Journals

About

Contact us

My IOPscience

Anosov parameter values for the triple linkage and a physical system with a uniformly chaotic
attractor

This content has been downloaded from IOPscience. Please scroll down to see the full text.
2003 Nonlinearity 16 1499
(http://iopscience.iop.org/0951-7715/16/4/318)
View the table of contents for this issue, or go to the journal homepage for more

Download details:
IP Address: 164.73.224.2
This content was downloaded on 20/03/2015 at 20:49

Please note that terms and conditions apply.

INSTITUTE OF PHYSICS PUBLISHING

NONLINEARITY

Nonlinearity 16 (2003) 14991510

PII: S0951-7715(03)54789-2

Anosov parameter values for the triple linkage and a

physical system with a uniformly chaotic attractor
T J Hunt1 and R S MacKay2
1
2

Open University, Walton Hall, Milton Keynes MK7 6AA, UK

Mathematics Institute, University of Warwick, Coventry CV4 7AL, UK

Received 8 October 2002, in final form 7 May 2003

Published 6 June 2003
Online at stacks.iop.org/Non/16/1499
Recommended by C Liverani
Abstract
First, the free frictionless dynamics of the triple linkage of Thurston and Weeks
is proved to be Anosov on each positive energy level for a non-empty open
set of parameter values. Second, in the presence of weak friction and any
smooth potential, an open set of physically implementable feedback control
laws are found for which the system possesses a uniformly hyperbolic attracting
submanifold.
Mathematics Subject Classification: 37D20, 70E55, 53Z05

1. Introduction
There is a well worked out mathematical theory of uniformly hyperbolic sets (e.g. [KH]).
Their definition and principal properties are recalled in appendix A. The main point is that,
if not just a union of hyperbolic periodic orbits, a locally maximal uniformly hyperbolic set
has well characterized chaotic behaviour. As a convenient shorthand, we call any locally
maximal uniformly hyperbolic set other than a union of hyperbolic periodic orbits uniformly
chaotic.
Although the theory is usually presented in discrete time, continuous time is much
more relevant to physics, so the latter is the context used in this paper. Mathematical
examples of uniformly chaotic sets in continuous time include: (i) suspensions of discrete-time
examples like a horseshoe, Plykin attractor, solenoid map, DA attractor and hyperbolic toral
automorphism, and (ii) the unit tangent bundle for a geodesic flow on a compact Riemannian
surface with everywhere negative Gaussian curvature or higher dimensional manifold with
everywhere negative sectional curvature.
The principal types of chaos of relevance to physics are those which can be observed
without special choice of initial conditions. Thus, physical interest focuses on (i) chaotic
attractors, and (ii) systems with a foliation into compact invariant submanifolds on each
0951-7715/03/041499+12$30.00

2003 IOP Publishing Ltd and LMS Publishing Ltd

Printed in the UK

1499

1500

T J Hunt and R S MacKay

of which the whole dynamics is chaotic. Physicists have been disappointed, however, that
the mathematical theory of uniformly hyperbolic sets never seems to apply to any physical
examples of these two types of chaotic dynamics, and that the mathematical examples cannot
be realized physically (see appendix B for a discussion). Suspensions of horseshoes occur in
many physical systems, of course, but are unstable. Thus, two long-standing questions have
been whether there are (i) physical examples of a Hamiltonian system with Anosov energy
levels (an energy level is the set of points where the Hamiltonian takes a given value), and
(ii) physical examples of a system with a uniformly chaotic attractor. Both questions are given
positive answers in this paper. Of course, physical does not have a definition, but we hope
that most physicists will agree that our examples qualify.
Our examples are based on the triple linkage of [TW]. For its free frictionless motion
we prove existence of an open set of parameter values for which each positive energy level
is Anosov, thus answering question (i). To answer question (ii), we allow weak friction
and construct feedback control laws such that the system has an attracting uniformly chaotic
invariant submanifold near one of the energy levels.
The triple linkage consists of three disks in a plane, free to rotate about pivots fixed in a
triangle, but constrained by three rods connecting one point pi , i {1, 2, 3} of each disk to
a floating pivot x (see figure 1). Of course, it is not really necessary that all lie in one plane;
it is enough that the projections to a plane satisfy the constraints. The free motion at given
positive energy is (up to timescale) the
geodesic flow of the kinetic energy metric. This is just
the metric that assigns length ds = 2K(v) dt to an infinitesimal displacement v dt, where
K(v) is the kinetic energy for a velocity vector v. Thurston and Weeks presented the triple
linkage (with the disks replaced by rods, but this makes no difference) as an example of a
system whose configuration space has interesting topology: for parameters in the regime of
interest to us, it is a surface of genus three (section 2); but they did not address the dynamics.
Surfaces of genus greater than one can support Riemannian metrics of negative Gaussian

Figure 1. The triple linkage.

Anosov parameter values

1501

curvature, and their geodesic flows are Anosov. So we asked ourselves if for some parameter
values the curvature of the kinetic energy metric for the triple linkage might be everywhere
negative.
We began by a numerical investigation of the curvature. To reduce our search we restricted
attention to the D6 -symmetric case of disks of equal moments of inertia I pivoted on the vertices
of an equilateral triangle of radius 1, connected by equal rods of length l2 and mass m, attached
to points of the disks at equal radius l1 from their centres. The rods are not required to have
uniform line density, nor to be straight, nor to end at the pivots, nor even to be one-dimensional,
but the D6 symmetry forces them to have centre of mass on the line through its two pivots.
Then the kinetic energy can be written





1
2
2
p)
=
(1)
p i + 3
p i
1 x + 2 x
K(x,
2
i
i
(the values of the mass parameters j are related to physical parameters in section 2). After
an extensive search [H], parameter values l1 = 7/40, l2 = 41/40, 1 = 11/50, 2 =
3/100, 3 = 23/100 were found for which the curvature was computed to be negative
everywhere (section 3). To make a proof, however, would probably require computer-assisted
estimates.
To obtain a rigorous result we took a cue from the fact that these parameter values have l1
small and l2 very close to 1, and that calculations are much easier if 1 , 2 = 0, to consider a
limiting metric for which we proved the curvature to be negative except at 8 points (section 4).
That is sufficient for the flow to be Anosov.
Using structural stability of Anosov flows, we deduced a parameter regime where the free
motion of the triple linkage is Anosov on each energy level and other properties (section 5).
Theorem 1. Write l2 = 1 + bl1 . Then there exist l0 , m0 , b0 > 0 such that for l1 < l0 ,
m < m0 , b < b0 the free motion of the triple linkage on each positive energy level is Anosov,
mixing and the correlations of Holder continuous functions decay exponentially.
These properties are preserved even if the D6 -symmetry is broken slightly, and for all high
enough energy levels if a potential is added. Furthermore, they imply diffusion in any Abelian
cover.
Even for parameter values where the free motion is not Anosov, a potential can be chosen
to make a given energy level (and all nearby ones) Anosov, provided the configuration space
has genus greater than one (section 6).
To obtain a uniformly chaotic attractor when friction is taken into account, we add torques
chosen to contract strongly onto or near a particular energy level (section 7). By the theory
of normal hyperbolicity, this gives a normally hyperbolic attracting submanifold close to the
energy level. The motion on it is close to the free motion on that energy level and hence
Anosov and mixing.
2. The triple linkage
In this section, we examine the topology of the configuration space of the triple linkage as a
function of its length parameters, and determine the kinetic energy in terms of the length and
mass parameters.
Given l1 , l2 , the region accessible to the central pivot x is the intersection of three annuli, of
inner and outer radii |l2 l1 | and l1 + l2 , centred on the pivots for the three disks. The accessible
region may have more than one component. When the position x is given in this region, there

1502

T J Hunt and R S MacKay

Figure 2. Partition of the plane of parameters R = l1 + l2 , r = l1 l2 into regions where

the configuration space of the triple linkage has a given diffeomorphism type. I, G, M : X3 ;
H, F, E, J, N : X0 ; X : X12 ; Z, AA : X7 ; Q, V , S : 3X0 ; Y : X7 X0 ; W : 6X0 ;
K, O : 4X0 ; U : 7X0 ; T : X3 3X0 . Xg denotes the oriented surface of genus g. The regions
A, B, C, D, L, P , R and all those above the diagonal are unattainable.

are generically two possibilities for the position of each pi . We label the possibilities by
i {} using + for the choice that is clockwise relative to x and for the anticlockwise
choice. The possibilities for each pi merge into one, denoted , on certain pieces of the
boundary of the accessible region, and the 8 patches corresponding to the choices of glue
together along these edges to form a union of compact orientable surfaces. The genus g of
each can be computed using Eulers formula F E + V = 2 2g for the numbers of faces,
edges and vertices in this decomposition into patches.
We computed all the possibilities for the configuration space of the triple linkage and the
results are shown in figure 2. The topology of the configuration spaces of many other linkages
was studied by [Ni].
A necessary condition for a geodesic flow on a surface to be Anosov is that it have a
genus greater than 1 [Kl], so only those cases where the configuration space has a component
of genus greater than 1 are worth considering. A sufficient condition is that the curvature of
the metric be negative everywhere, though one can relax it to be zero at a finite number of
points (cf [Kn]).
Morse proved that, for a surface of genus greater than 1, the geodesic flow of any
metric has an invariant subset with semiconjugacy to the negative curvature case (the proof
was streamlined and generalized in [DM]). The subset could be small, however, and the
semiconjugacy could map most points to a few periodic orbits, so it would be much better to
know that the whole flow is Anosov.
We decided to concentrate on the regions I, G, M where the configuration space is a
surface of genus 3. Figure 3 shows the case l1 = 19/180, l2 = 181/180, represented as a

Anosov parameter values

1503

Figure 3. The configuration space as a subset of T3 for l1 = 19/180, l2 = 181/180.

subset of T3 using the angles i of the three disks.

The mass parameters for the kinetic energy are:
3
(Ic + m|p c|2 ),
l22
2
2 = 2 (Ic + m(x c) (p c)),
l2
1 =

3 =

(2)

I Ic + m|x c|2
+
,
l12
l22

where Ic is the moment of inertia of a rod about its centre of mass c. The masses of the pivots
can be incorporated into the mass parameters too, any part fixed to a disk being incorporated
into its inertia and any part fixed to a rod being incorporated into its inertia. Real pivots,
however, often contain parts that are fixed to neither, e.g. ball races. Then the exact result
depends on their construction. One could calculate their effect for any given construction, but
for simplicity of exposition, we ignore that.
The mass parameters satisfy some restrictions due to positivity of m, I , Ic , |p c|2
and |x c|2 and the relation |(x c) (p c)| = l2 . The complete set of restrictions is
1 , 3 > 0 and 22 < 43 1 3 (see appendix C). Note that although 2 may be negative if the

1504

T J Hunt and R S MacKay

rods extend far enough beyond the pivots, it is positive for a uniform rod joining the pivots
(and for all nearby mass distributions).
3. Numerics on the curvature
On eliminating the pi in favour of x and the i , the kinetic energy of the triple linkage can be
written entirely in terms of x and the i . We chose the origin for x to be at the centre of the
triangle of fixed pivots, and oriented the triangle to have a vertex at (1, 0). By D6 -symmetry,
it is enough to look at the + + + and + + patches. The kinetic energy metric depends
on 5 parameters: l1 , l2 and the mass parameters i , but an overall scale factor on the masses
just scales the kinetic energy, so we can reduce parameter space to four dimensions (in our
numerics we took 1 + 32 + 33 = 1).
The curvature of the metric can be computed by Brioschis formula (e.g. [DC,G]), though
the formulae are immense. Although still long, the calculation simplifies at the symmetry
points, i.e. the fixed points for (non-identity) elements of the D6 action. So we evaluated
the curvature there to eliminate large regions of the four-dimensional parameter space. To
make the curvature negative at the type I fixed points ((x, ) = (1 l1 l2 , 0, , +, ),
(1 l1 l2 , 0, , , +), and their images under three-fold rotation) required most of the
mass to be in 3 and the lengths to be near l2 = 1, l1 = 0. The type II fixed points
((1 (l2 l1 ), 0, , +, ), (1 (l2 l1 ), 0, , , +) and their three-fold rotations) turned out
to pose no further restriction. The type O points, however, (x = 0, = + + + or )
required further restriction to a neighbourhood of the hyperbola l22 = 1 + l12 .
Having hit on a promising parameter set l1 = 7/40, l2 = 41/40, 1 = 11/50, 2 =
3/100, 3 = 23/100, we computed the curvature on the whole of the + + + and + +
patches. This was done with intelligent use of the chain rule in Mathematica. The result is
shown in figure 4. Indeed the curvature appears to be negative everywhere.
4. The limit system
A priori, the regime = (0, 0, 13 ) does not look promising because at the type O
symmetry points the principal curvatures are equal so the Gaussian curvature is a square,

Figure 4. The curvature on the + + + and + + patches of the configuration space for the triple
linkage at parameter values 1 = 11/50, 2 = 3/100, 3 = 23/100, l1 = 7/40, l2 = 41/40. The
darkness indicates the degree of negativity of the curvature.

Anosov parameter values

1505

thus non-negative. It has the advantage, however, that the kinetic energy metric is just the
restriction of a standard Euclidean metric on T3 , which facilitates the calculations.
For i = 1, 2, 3, let i be the angle of the ith disk from the position where the point of
attachment of the connecting rod is closest to the centroid of the equilateral triangle. The
condition that the attachment points be on a circle of radius l2 can be written
l22 (a + b + c)(a + b c)(b + c a)(c + a b) = a 2 b2 c2 ,

(3)

where a, b, c are the distances between them. Writing l2 = 1 + bl1 , we find the configuration
space to first order in l1 has the equation
(cos 1 + cos 2 + cos 3 )
.
(4)
b=
3
This is a (non-empty real analytic) submanifold of T3 for all values of b [1, 1] except
the critical values of the right-hand side which are 1, 13 , 13 , 1. Solid-state physicists will
recognize it as the Fermi surface for independent electrons in a tight binding model of a cubic
lattice with only nearest-neighbour hopping [AM]. Riemannian geometers will recognize the
case b = 0 as the Schwarz P-surface, an example of a minimal surface (which was the basis
for constructions of [BD]).
If the rods have zero mass (m = 0) and the moments of inertia of the disks are scaled
to I =
1, then the kinetic energy metric is simply the restriction of the Euclidean metric

2
3
ds =
i di on T to the configuration surface. The unit normal vector of this first order

approximation is (sin 1 , sin 2 , sin 3 )/Z 1/2 , where Z = i sin2 i . Note that Z = 0 for b not
a critical value. In any open set which is a graph over (1 , 2 ) the projection of its derivative is


2
2
sin 1 sin 2 cos 3
3/2 cos 2 sin 3 + cos 3 sin 2
.
(5)
Z
sin 1 sin 2 cos 3
cos 1 sin2 3 + cos 3 sin2 1
The curvature (the determinant of the above, multiplied by the appropriate projection factor
Z/ sin2 3 ) is
Z 2 (cos 1 cos 2 sin2 3 + cos 2 cos 3 sin2 1 + cos 3 cos 1 sin2 2 ).

(6)

This can be written as

 2
s
9b
(7)
+ 3bp
Z 2 ,
2
2


where p = i cos i and s = i cos2 i = 3 Z. The same result is obtained for open sets
requiring other pairs of coordinates, so it holds everywhere.
If b = 0 then the curvature is s/2Z 2 , which is non-positive (this follows also from
minimality). It is equal to zero only if all cos i are zero, i.e. at the eight points i {/2}.
This is sufficient for its geodesic flow to be Anosov (cf [Kn]).
5. Perturbation of the limit system
Now, we add the higher-order corrections in l1 and allow the rods to have small positive mass.
By structural stability of Anosov flows, the geodesic flow remains Anosov. It is also mixing
because geodesic flows are contact flows and contact Anosov flows on connected manifolds
are mixing [KH] (both topologically and for every Gibbs measure, which includes Liouville
measure for Anosov geodesic flows). The correlations of Holder continuous functions decay
exponentially by [Do]. This proves our theorem.
One can break the D6 -symmetry weakly too (which can introduce rattleback terms in
the kinetic energy), and the theorem still holds. One can even allow slight out-of-plane effects,

1506

T J Hunt and R S MacKay

e.g. if the axes of rotation of the disks are not quite parallel, provided no new degrees of
freedom are introduced.
One consequence of the exponential decay of correlations is that there is diffusion in any
Abelian cover. An obvious Abelian cover is obtained by lifting the configuration space from
T3 to R3 (as is easy to visualize from figure 3). Thus, if one were to measure the total angles
j , j = 1, 2, 3, turned in time t, one should see the vector (t) perform a random walk in R3 ,
when observed on a scale large compared with 2 . Actually, the first Betti number of the unit
tangent bundle to X3 is 7, so this can be extended to a seven-dimensional random walk.
6. Adding a potential
If the linkage is subjected to a potential energy V (but still no friction), then the motion is that
= K(q)
V (q). The energy H = K + V is conserved. For large
of the Lagrangian L(q, q)
energies, the perturbation of the vector field due to V is relatively small in C 1 , so structural
stability implies that all high enough energy levels remain Anosov. They also remain mixing,
because nonintegrability of the invariant foliations of a transitive Anosov flow is preserved
under C 1 -small perturbation [Ch]. It follows by [Do] that exponential decay of correlations is
also preserved.
An alternative strategy for obtaining Anosov energy levels for the triple linkage is to start
from any parameter value such that the configuration space has a component which is a surface
of genus greater than 1 and then add a carefully designed potential, as follows. For E greater
than the maximum of V , the motion on energy level H1 (E) is up to time-reparametrization the
geodesic flow of the Maupertuis (or Jacobi)
This is conformally
metric 2 (E V (q))K(dq).
equivalent to the kinetic energy metric 2K(dq). The conformal factor 2(E V (q)) can
be chosen to make the Maupertuis metric have constant negative curvature [Au]: the equation
 e2 = , where  and are the Laplacian and the curvature of the kinetic energy metric,
has a unique solution on the configuration space and then one chooses V = E 21 e2 . It
follows that the chosen energy level is Anosov (by constant negative curvature), and so are
all the nearby ones (by structural stability). It is not guaranteed that the flow remains mixing,
however, because the time-reparametrization might remove the joint nonintegrability of the
invariant foliations.
7. Including friction
If there is friction (meaning additional forces whose inner product with the velocity is negative
definite) then by LaSalles principle and compactness of the configuration space, the energy
tends to zero as time goes to infinity (or in the case with a potential, to a critical value of the
potential), so there is no chaos.
If the friction depends linearly on velocity and there is no potential, then the flow induces
one on the unit tangent bundle of the configuration space, given by rescaling the rate of time
to obtain speed 1. If the friction is weak then the resulting flow is close to the geodesic flow,
hence still Anosov and mixing. This is a somewhat artificial fix, however, because the rescaled
time is bounded and the result is not robust to the addition of a potential or nonlinearity in the
friction. So we seek alternatives.
The effect of friction can be compensated by suitable driving forces. To make a physically
realizable system, the compensating forces must lie in an open set of suitable ones. So it is not
enough just to design external forces which exactly annihilate the friction forces. Instead, we
suppose the friction is weak and apply feedback forces which contract strongly onto or close

Anosov parameter values

1507

to a particular energy level H 1 (E). We choose E sufficiently above the maximum of the
potential so that H is dominated by the kinetic energy K. Then for small m the desired effect
can be achieved by torques (K E)i on the three disks, with large ( can be allowed to
be state-dependent). This is because with no friction
 nor potential, and m = 0, such external
torques would make the kinetic energy K = (I /2) i i2 evolve by K = 2 K(K E)/I , so
K would converge rapidly to E from all initial conditions except the rest states, and on K 1 (E)
the torques are zero so there is no change to the motion there. The linearized contraction rate
onto K 1 (E) is 2 E/I . If this is chosen larger than the
maximum contraction rate for the flow
tangent to K 1 (E) (in the limit of section 4, this is 2E/I , because it is the product of the
speed and the square root of minus the most negative curvature of the metric, which is easily
shown to be 1 and to occurat = (0, , /2) and its images under symmetry, so it suffices to
choose a little larger than I /2E), then K 1 (E) is normally hyperbolic and hence persists to
a nearby attracting invariant submanifold under all small perturbations [F], such as the addition
of weak friction, a potential small compared to E, small m, and small deviations from the above
feedback law (provided they are autonomous). Because the unperturbed motion on K 1 (E)
is Anosov and Anosov flows are structurally stable, the flow on the resulting submanifold is
Anosov. It is also mixing, because nonintegrability of the invariant foliations is preserved by
C 1 -small perturbation.
This feedback control could be implemented by detectors to measure the rotation rate of
each disk (e.g. car speedometers), a device to compute the torques (which could perhaps be
mechanical, though a microcomputer would be easier), and actuators to apply the torques to
the disks (e.g. induction motors).
Acknowledgment
TJH was supported by an EPSRC studentship.
Appendix A. Properties of uniformly hyperbolic sets
Definition. An invariant subset  for a C 1 vector field x = v(x) on a manifold M (equipped
with a norm | | on its tangent spaces), with flow denoted : R M M, is said to be
uniformly hyperbolic if there are C, A, > 0 and a splitting Ex+ Ex Rv(x) of the tangent
space to M at each point x , such that (i) Ex , t R implies Dt Et (x) and
|Dt |
Ce|t|

for t R ,

and (ii) Ex , R imply | + + + v(x)|  A max(| |, |||v(x)|) (the splitting is

automatically continuous and the second condition is satisfied automatically if  is compact).
The case where the whole manifold M is uniformly hyperbolic is called Anosov.
Any invariant set that is contained in a small enough neighbourhood of a uniformly
hyperbolic set is also uniformly hyperbolic, so without loss of generality one can restrict
attention to locally maximal uniformly hyperbolic sets.
Locally maximal uniformly hyperbolic invariant sets  have the following nice properties
(e.g. [KH]):
1.  is closed.
2. Each chain transitive component of  is locally maximal and topologically transitive (for
a volume-preserving system such as a Hamiltonian system on an energy level, this implies
that each connected component of  is topologically transitive).

1508

T J Hunt and R S MacKay

3. If compact,  has a finite Markov partition, i.e. equivalence of the flow on  (up to a
negligible subset) to a flow on a finite directed graph (symbolic dynamics).
4.  persists under C 1 -small perturbation of v, and the dynamics on its continuation remain
topologically conjugate to the unperturbed dynamics (structural stability).
5. If a small time-dependent perturbation is added to v, each orbit of the perturbed system
that remains in a small enough neighbourhood of  is shadowed by some orbit of the
unperturbed system in .
6. The sets of points whose orbits converge together with the orbit of a given point x  in
forwards (respectively backwards) time form injectively immersed submanifolds, tangent
to Ex and as smooth as the system, which we call the forwards (respectively backwards)
contracting submanifolds of that point. If  contains all its backwards contracting
submanifolds then it attracts a neighbourhood and they foliate the basin of attraction.
In the Anosov case, the whole manifold is foliated by both the forwards and backwards
contracting submanifolds. The foliations are invariant and Holder continuous; in general
they are not smoother, but for Anosov geodesic flows of Riemannian metrics on surfaces
they are C 1 .
7. If  is chain transitive and attracting, it supports a unique invariant probability measure
(called SRB after Sinai, Ruelle and Bowen) describing the asymptotic distribution of
almost all orbits in its basin of attraction. For volume-preserving flows the SRB measure
is the volume (up to normalization). There are formulae relating the SRB measure to
properties of its periodic orbits.
8. There is a simple formula for how the SRB measure changes under perturbation of the
vector field [R].
9. In some circumstances, it can be proved that the SRB measure is mixing, and explicit
bounds on the decay of correlation functions can be obtained. For example, Anosov
geodesic flows on a compact connected manifold are mixing, and if the flow is C 2+ and
the invariant foliations are C 1 (as is true for Anosov geodesic flows in two dimensions)
then the correlations of Holder continuous functions decay at least exponentially [Do].

10. In particular, if the velocity correlation functions Cij (t) = hi (v(t (x)))hj (v(x)) d(x)
for a basis (hi )i=1...I of any subspace L of first de Rham cohomology H 1 (M) are integrable,
one obtains diffusion in the Abelian cover of M corresponding to L. The simplest
sense in which there is diffusion is that i (x)j (x) d(x) 2Dij t as t , where
t

i (x) = 0 hi (v(t (x))) dt and Dij = 0 Cij (t) dt, but diffusion holds in stronger senses
too [DP] (cf [BS] for a periodic Lorentz gas, though this is not a uniformly hyperbolic
system).
Appendix B. Comments on some standard examples of chaos
The chaotic attractors of most dynamical systems of physical origin seem not to be uniformly
chaotic. For example, the Lorenz system fails even if one takes a parameter regime with no
near-tangencies of the invariant manifolds of the saddle, because the attractor contains the
saddle so the splitting in the definition of uniformly hyperbolic set cannot be continuous there,
whereas continuity of the splitting is an automatic consequence of the definition of uniform
hyperbolicity; this may look just a technical point but it causes properties 3, 4, and 5 to fail.
The Sinai billiard fails because the vector field is discontinuous at bounces and, even worse,
the billiard map is discontinuous at grazing incidence. The periodically forced van der Pol
oscillator (which can be included in our autonomous context by adding the equation t = 1)
in chaotic parameter regimes appears to have near-tangencies of the invariant manifolds of a

Anosov parameter values

1509

periodic orbit, and its Poincare map does not appear to have the minimum of three holes that
are necessary to obtain a uniformly hyperbolic attractor [P].
Also the mathematical examples of uniform chaos do not seem to be physically realizable.
Compact manifolds of negative curvature cannot be embedded in R3 so their geodesic flows
have not been regarded as physical. Similarly, the suspensions of the solenoid map and toral
automorphisms live on three-manifolds that are difficult to imagine, so they are regarded
as artificial. The suspension of a Plykin attractor is realizable in a system of three ODEs
(e.g. [H]), but explicit realizations so far look very contrived. Ruelle and Takens (refined with
Newhouse [NRT]) proved that arbitrarily close to a constant vector field on a torus of dimension
N  3 (which could arise by weakly coupling N limit cycle oscillators or by bifurcations from
a periodic orbit) are vector fields containing uniformly chaotic attractors, but we are not aware
of any positive identifications of this possibility in physical systems.
Appendix C. Restrictions on the mass parameters
The restrictions 1 > 0, 3 > 0 and 22 < 43 1 3 on the mass parameters are derived as
follows.
The necessity of the first two conditions is clear from the formulae (2). To show the third
is necessary, let s be the position of the centre of mass of a rod along the line through p and
x, scaled so that s = 0 puts it at p and s = 1 puts it at x. Put = I / l12 and = Ic / l22 . Then


2

2
1
(8)
+ = (ms(1 s))2 =
(3 ).
2
3
So




22
1 3
1
1

=
+ 2 + 3
.
(9)
4
3
3
3
Now
1
1
+ 2 + 3 = m + > 0
= ms 2  0,
and
3
3
so 22 43 1 3 < 0.
To show sufficiency of the conditions, given satisfying the constraints, choose any > 0
such that 1 , 2 , 3 = 3 still satisfy them. Then let


22
1
1  1 2
1 1 3

m=
,
s=
=
+ 2 + 3 ,

+
.
4
3
m
3
m 3
2
This combination of parameters realizes .
References
[AM]
[Au]
[BS]
[BD]
[Ch]
[DP]
[DM]
[DC]

Ashcroft N W and Mermin N D 1976 Solid State Physics (Philadelphia, PA: Saunders)
Aubin T 1998 Some Nonlinear Problems in Riemannian Geometry (Berlin: Springer)
Bunimovich L A and Sinai Ya G 1981 Statistical properties of Lorentz gas with periodic configuration of
scatterers Commun. Math. Phys. 78 47997
Burns K and Donnay V J 1997 Embedded surfaces with ergodic geodesic flows Int. J. Bif. Chaos 7 150927
Chernov N I 1998 Markov approximations and decay of correlations for Anosov flows Ann. Math. 147
269324
Denker M and Phillips W 1984 Approximation by Brownian motion for Gibbs measures and flows under a
function Ergod. Theor. Dyn. Syst. 4 54152
Denvir J and MacKay R S 1998 Consequences of contractible geodesics on surfaces Trans. Am. Math. Soc.
350 455368
Do Carmo M P 1976 Differential Geometry of Curves and Surfaces (Englewood Cliffs, NJ: Prentice-Hall)

1510
[Do]
[F]
[G]
[H]

T J Hunt and R S MacKay

Dolgopyat D 1998 On decay of correlations in Anosov flows Ann. Math. 147 35790
Fenichel N 1971 Persistence and smoothness of invariant manifolds for flows Indiana U. Math. J. 211 193226
Gray A 1993 Modern Differential Geometry of Curves and Surfaces (Boca Raton, FL: CRC)
Hunt T J 2000 Low dimensional dynamics: bifurcations of cantori and realisations of uniform hyperbolicity
PhD Thesis, Cambridge University
[KH] Katok A and Hasselblatt B 1995 Introduction to the Modern Theory of Dynamical Systems (Cambridge:
Cambridge University Press)
[Kl]
Klingenberg W 1974 Riemannian manifolds with geodesic flow of Anosov type Ann. Math. 99 113
[Kn]
Knauf A 1987 Ergodic and topological properties of Coulombic periodic potentials Commun. Math. Phys.
110 89112
[NRT] Newhouse S, Ruelle D and Takens F 1978 Occurrence of strange Axiom A attractors near quasi periodic
flows on Tm , m
3 Commun. Math. Phys. 64 3540
[Ni]
Niemann S H 1977 Geometry and mechanics PhD Thesis (Mathematics) Oxford
[P]
Plykin R V 1974 Sources and sinks of A-diffeomorphisms of surfaces Mat. USSR Sbornik 23 23353
[R]
Ruelle D 1997 Differentiation of SRB states Commun. Math. Phys. 187 22741
[TW] Thurston W P and Weeks J R 1984 The mathematics of three-dimensional manifolds Sci. Am. 251 94106