Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
DYNAMICAL SYSTEMS
To
VLADIMIR IGOREVICH ARNOLD and
STEPHEN SMALE
for their inspirational work
D. K. ARROWSMITH
Lecturer, School of Mathematical Sciences,
Queen Mary & Westfreld College, Unioersity of London
C.M. PLACE
Lecturer (formerly Department of Mathematics, Westfield College, University of London)
UNIVERSIDAD DE: CANTABRIA BIBLIOTECA
00284078
X : ~S /.
v . <11 Yo CAMBRIDGE
UNIVERSITY PRESS
CONTENTS
Published by the Press Syndicate of the University of Cambridge The Pit! Building, Trumpington Street, Cambridge CB2 I RP 40 West 20th Street. New York, NY 100114211, USA
10 Stamford Road, Oakleigh, Melbourne 3166, Australia
Preface
I Dift'eomorphisms aM Rows . l.I Introduction
1.2 Elementary dynamics of diffeomorphisms 1.2.1 Definitions
1.2.2 Diffeomorphisms of the circle 1.3 Flows and differential equations 1.4 Invariant sets
1.5 Conjugacy
1.6 Equivalence of flows
1.7 Poincare maps and suspensions 1.8 Periodic nonautonomous systems
1.9 Hamiltonian flows and Poincare maps Exercises
{;! Cambridge University Press, 1990
First published 1990 Reprinted 1991, 1994
Printed in Great Britain by Bell and Bain Ltd, Glasgow
British Library cataloguing in publication data
Arrowsmith, O.K.
An introduction to dynamical systems. 1. Differentiable dynamical systems
I. Title II. Place, C.M.
514'.7
2 Local properties of flows and diffeomorphisms 2.1 Hyperbolic linear diffeomorphisms and flows 2.2 Hyperbolic nonlinear fixed points
2.2.1 Diffeomorphisms
2.2.2 Flows
2.3 Normal forms for vector fields
2.4 Nonhyperbolic singular points of vector fields 2.5 Normal forms for diffeomorphisms
2.6 Timedependent normal forms
2.7 Centre manifolds
2.8 Blowingup techniques on R2 2.8.1 Polar blowingup
2.8.2 Directional blowingup Exercises
Library of Congress cataloguing in publication data Arrowsmith, D. K.
An introduction to dynamical systems / O.K. Arrowsmith and C.M. Place
p. em.
Bibliography: p. Includes index.
ISBN 0 521 30362 I.  ISBN 0 521 31650 2 (paperback)
I. Differentiable dynamical systems. I. Place, C. M. II. Title.
QA614.8.A77 1990
515:,352i1c20 897191 CIP
Transferred to digital reprinting 2001 Printed in the United States of America
MP
3 Structural stallility, .yperllo!icity aM ~ points 3.1 Structural stability of linear systems
3.2 Local structural stability
3.3 Flows on twodimensional manifolds 3.4 Anosov diffeomorphisms
I 1 5 5 6
11 16 20 28 33 38 42 56
64 64 67 68 69 72 79 83 89 93
102 102 105 lOS
119 120 123 125 132
Contents
3.5 Horseshoe diffeomorphisms 3.5.1 The canonical example
3.5.2 Dynamics on symbol sequences
3.5.3 Symbolic dynamics for the horseshoe diffeomorphism 3.6 Hyperbolic structure and basic sets
3.7 Homoclinic points
3.8 The Melnikov function Exercises
Contents
138 139 147 149 154 164 170 180
6.7 Birkhoff periodic orbits and Hopf bifurcations
6.8 Double invariant circle bifurcations in planar maps Exercises
355 368 379
394 413 417
Hims for exercises ' References
Index
4 Local bifurcations I: planar vector fields and
diffeomorphisms on R 190
4.1 Introduction 190
4.2 Saddlenode and Hopf bifurcations 199
4.2.1 Saddlenode bifurcation 199
4.2.2 Hopf bifurcation 203
4.3 Cusp and generalised Hopf bifurcations 206
4.3.1 Cusp bifurcation 206
4.3.2 Generalised Hopf bifurcations 211
4.4 Diffeomorphisms on If! 215
4.4.1 Dx/(O) = + I: the fold bifurcation 218
4.4.2 Dx/(O) = I: the flip bifurcation 221
4.5 The logistic map 226
Exercises 234
5 Local bifurcations II: diffeomorphisms on R2 245
5.1 Introduction 245
5.2 Arnold's circle map 248
5.3 Irrational rotations 253
5.4 Rational rotations and weak resonance 258
5.5 Vector field approximations 262
5.5.1 Irrational p 262
5.5.2 Rational (l = p/q, q ~ 3 264
5.5.3 Rational fJ = p/q. q = 1,2 268
5.6 Equivariant versal unfoldings for vector field approximations 271
5.6.1 q = 2 272
5.6.2 q = 3 275
5jJq=4 D6
5.6.4 q ~ 5 282
5.7 Unfoldings of rotations and shears 286
Exercises 291
6 Areapreserving maps and their perturbations 302
6.1 Introduction 302
6.2 Rational rotation numbers and Birkhoff periodic points 309
6.2.1 The PoincareBirkhoff Theorem 309
6.2.2 Vector field approximations and island chains 310
6.3 Irrational rotation numbers and the KAM Theorem 319
6.4 The AubryMather Theorem 332
6.4.1 Invariant Cantor sets for homeomorphisms on Sl 332
6.4.2 Twist homeomorphisms and Mather sets 335
6.5 Generic elliptic points 338
6.6 Weakly dissipative systems and Birkhoff attractors 345
1
PREFACE
Diffeomorphisms and flows
In recent years there has been a marked increase of research interest in dynamical systems and a number of excellent postgraduate texts have been published. This book is specifically aimed at the interface between undergraduate and postgraduate studies. It is intended both to stimulate the interest of final year undergraduates and to provide a solid foundation for postgraduates who intend to embark on research in the field. For example, a challenging thirdyear undergraduate course can be constructed by selecting topics from the first four chapters. Indeed, lecture courses taught by one of us (CMP) provided the basis for Chapters I, 2 and 4. On the other hand, Chapter 6 is directed at firstyear postgraduate students. It contains a selection of current research topics that illustrate the interaction between superficially different research problems.
A major feature of the book is its extensive set of exercises; more than 300 in all. These exercises not only illustrate the topics discussed in the text, but also guide the reader in the completion of technical details omitted from the main discussion. Detailed model solutions have been prepared and hints to their construction are provided.
The reader is assumed to have attended courses in analysis and linear algebra to secondyear undergraduate standard. Prior knowledge of dynamical systems is not necessary; however, some familiarity with the qualitative theory of differential equations and Hamiltonian dynamics might be an advantage.
We would like to thank Martin CasdagJi for sharpening our understanding of Birkhoff attractors, David Knowles and Chris Norman for helpful discussions and Carl Murray for steering some awkward diagrams to a laser printer. We are grateful to the Quarterly Journal of Applied Mathematics and SpringerVerlag for allowing us to use diagrams from some of their publications and our thanks go to Sandra Place for her fast and accurate typing of much of the manuscript. One of us (CMP) would like to 'acknowledge the Brayshay Foundation for its financial support throughout this project. Finally, we must both pay tribute to the patience and support of our families during the long, and often difficult, gestation period of the manuscript.
l.l Introduction
A dynamical system is one whose state changes with time (t). Two main types of dynamical system are encountered in applications: those for which the time variable is discrete (r e Z or N) and those for which it is continuous (tER).
Discrete dynamical systems can be presented as the iteration of a function, i.e.
x,+ 1= f(x,),
tEl or N.
(1.1.1)
When t is continuous, the dynamics are usually described by a differential equation
dx =x=X(x). dt
(1.1.2)
In (1.1.1 and 2), x represents the state of the system and takes values in the state or phase space. Sometimes the phase space is Euclidean space or a subset thereof, but it can also be a nonEuclidean structure such as a circle, a sphere, a torus or some other differentiable manifold.
In this chapter we will consider two special cases of the above equations, namely when:
(i) r in (1.1.1) is a diffeomorphism; and
(ii) the solutions of (1.1.2) can be described by a flow with velocity given by
the vector field X.
These two cases have been widely studied and they are fundamental to our understanding of dynamical systems. Smale, in his definitive work (Smale, 1967), pointed out that (i) and (ii) are closely related and our discussion emphasises this connection.
Any description of the theory of (i) and (ii) involves differentiable maps so let us begin by recalling some definitions. Let U be an open subset of R". Then a function g: U + IR is said to be of class C' ifit is rfold continuously differentiable, 1 ~ r ~ 00. Let V be an open subset of IRm and G: U + V. Given coordinates
2
1.1 1 ntroduction
1 Diffeomorphisms and flows
3
(XI' ... , x.) in V and (YI' ... , Y .. ) in V, G may be expressed in terms of component functions gi: V + R, where
Figure 1.1 Examples of differentiable manifolds and some 'patches' of local coordinates. Several open sets based on patches of this kind may be required in order to cover the whole manifold.
I
(a) cylinder 0
f(!
Em
0 ~
~~~
(1.1.3)
;= 1, ... ,m.
The map G is called a C'map if gi is C' for each i = 1, ... , m. G is said to be differentiable if it is a C" map for some 1 ~ r ~ 00 and to be smooth if it is Coo. Maps that are continuous but not differentiable are, conventionally, referred to as COmaps.
Definition 1.1.1 G is said to be a diffeomorphism if it is a bijection and both G and G I are differentiable mappings. G is called a Ckdiffeomorphism if both G and G I are Ckmaps.
Observe that the bijection G: V + V is a diffeomorphism if and only if m = nand the matrix of partial derivatives
[agiJ·
DG(xI' ... , x.) = 
ax] i,j=1
(1.1.4)
is nonsingular at every x E V. Thus G(x, y) = (exp(y), exp(xW with V = R2 and V = {(x, y)lx, y > O} is a diffeomorphism because Det DG(x, y) = exp(x + y) '" 0 for each (x, Y)ER2.
If G satisfies Definition l.l.l with G and G  I continuous, rather than differentiable, maps then G is said to be a homeomorphism. As we shall see, such maps playa central role in the topological theory of flows and diffeomorphisms.
The above definitions are adequate provided phase space is Euclidean, but, as we have already mentioned, the natural setting for dynamics is a differentiable manifold. The important point here is that manifolds have the property that they are 'locally Euclidean' and this allows us to extend the idea of differentiability to functions defined on them. If M is a manifold of dimension n then, for any x EM, there is a neighbourhood W £; M containing x and a homeomorphism h: W + RO which maps W onto a neighbourhood ofh(x)E RO. Since we can define coordinates in V = h(W) £; RO (the coordinate curves of which can be mapped back onto W), we can think ofh as defining local coordinates on the patch W of M (see Figure 1.1).
The pair (V, h) is called a chart and we can use it to give meaning to differentiability on W. Let us assume, for simplicity, that f: W + W, then f induces a map f = h·(·hI: U + U (see Figure 1.2). We say that f is a Ckmap on W iff is a Ckmap on V. This construction allows us to give a definition of a local diffeomorphism on M.
In order to obtain a global description of the manifold, we cover it with a family of open sets, Wa, each with its associated chart (Va' ha) (predictably, the set of all charts is called an atlas). If Wall WII is not empty, then either (Va' ha) or (V II' hll) can be used to provide local coordinates for Wa II WII. This possibility induces overlap maps, hall and hll. between h.(W.1l WI') £; Va and hll(Wall WII) £; V II (see
(c) sphere: polar 0
coordinates
e
(d) sphere: stereographic projection
s~
4
1 Diffeomorphisms and flows
1.2 Elementary dynamics of diffeomorphisms
5
Figure 1.3). If ,!e now consider !: W«n Wp ..... W«n Wp, w_e hav~ two alternative representatives f« = h •. f.h; 1 and fp = hp.f.hi 1 for f. Since f« and fp are determined by dilTerent charts, they might belong to different differentiability classes, so that the class of f would be ambiguous. A manifold is said to be differentiable if all the overlap maps are dilTeomorphisms of the same dilTerentiability class, C' say. Now, from Figure 1.3,
These maps are usually presented to us in local coordinates so that the manifold structure does not appear explicitly.
fp=hp.f.hi1
== (hp.h; 1 Hh .. f.h; ')(h«.hi I)
 I == h«p' f«· h«p .
(1.1.5)
1.2 Elementary dynamics of diffeomorphisms
1.1.1 Definitions
Let M be a dilTerentiable manifold and suppose f: M ..... M is a diffeomorphism. For each x EM, the iteration (1.1.1) generates a sequence, the distinct points of which define the orbit or trajectory of x under f. More precisely, the orbit of x under f is {f"'(x)lm E Z}. For mE Z + , f"' is the composition of f with itself m times. Since f is a diffeomorphism r 1 exists and f m == WI r. Finally, CO == idM, the identity map on M. Typically, the orbit of x is a biinfinite sequence of distinct points of M. However, there are two important exceptions to this state of affairs.
Thus all local representatives of f have the same dilTerentiability class, Ck say, with k ~ r. It is important to note that r is determined entirely by the charts and hence by the structure of M. A manifold with overlap maps of class C· is called a C' manifold.
The discussion presented above is, of course, incomplete. We have only considered maps taking a chart into itself. This is clearly not true in general. Given f: M ..... M, then f: W« 4 Wp and f: W«n W1 ..... Wpn W6• The generalisation of our simple arguments that allows for these omissions is considered in Exercise 1.1.2. Needless to say, the 'message' is unchanged by these manipulations.
A more detailed discussion of differentiable manifolds is not necessary here (the interested reader should consult Arnold (1973) or Chillingworth (1976)). While the ideas outlined above provide valuable background knowledge, we will rarely find ourselves involved with charts, atlases, etc. This is because our concern is the dynamics of maps defined on M given that they are diffeomorphisms or flows.
Definition 1.2.1 A point x* E M is called a fixed point of fiffm(x*) == x* for all mE I.
Definition 1.2.2 A point x* E M is a periodic point of f if f'i(x*) == x*, for some integer q ~ I.
The least value of q satisfying Definition 1.2.2 is called the period of the point x* and the orbit of x*, i.e.
{x*, f(x*), ... , fqI(x*)},
(1.2.1)
is said to be a periodic orbit of period q or a qcycle of f. Clearly, since fq(x*) == x*, it is the sequence {f"'(x*)}~= _~ which is qperiodic. Notice that a fixed point is a periodic point of period one and a periodic point of f with period q is a fixed point of fq. Morever, if x* is a periodic point of period q for r then so are all of the other points in the orbit of x*. For example, if rq(x*) == x* then f(f'i(x*)) == r(x*) == fq(f(x*)) and f(x*) is therefore a periodic point of period q, and so on for f2(x*), ... , fqI(X*).
Fixed and periodic points can be classified according to the behaviour of the orbits of points in their vicinity. The following ideas are due to Liapunov.
Figure 1.2 Commutative diagram illustrating the representation of f defined on an open set W of M in a local shart (U, h).
Figure 1.3 Illustration of the definition of the overlap maps h.p and hpa· Note that hPa == h.;/.
Definition 1.2.3 A fixed point, x*, is said to be stable if, for every neighbourhood N of x*, there is a neighbourhood N' £, N of x* such that if x E N' then f"'(X)E N for all m > O.
W"nWp
y .~
h" (WanWp)hp (WanWp) haP
= hp' h;;1
WanWp
y ~
ha (Wa n Wp)  h/l (Wa n Wp) hpa
=h,,' hpl
Essentially, Definition 1.2.3 says that iterates of points 'near to' a stable fixed point, remain 'near to' itfor mEl + . If a fixed point x* is stable and Lim f"'(x) = x*,
m~co
for all x in some neighbourhood of x*, then the fixed point is said to be asymptotically stable. Trajectories of points near to an asymptotically stable fixed point move toward it as m increases. Fixed points that are stable, but not
6
J DijJeomorphisms and flows
1.2 Elementary dynamics of diffeomorphisms
7
asymptotically stable, are said to be neutrally or marginally stable and those that are not stable in the sense of Definition 1.2.3 are unstable.
others. It is then difficult to recognise fixed or periodic points from the representation of orbits on the circle itself. This is a problem for any map (f) of the circle, whether it is a diffeomorphism or not, and it is solved by considering a lift of f.
The natural setting for introducing the lift of [: SI + SI is when f is a homeomorphism rather than a diffeomorphism and it would be perverse to artificially confine our discussion to the differentiable case. Moreover, by taking f to be a homeomorphism at this point we can better appreciate the consequences of imposing differentiability on f and r 1. Thus, let f: SI + SI be a homeomorphism and suppose there is a continuous function 1: R + R such that
1.2.2 Diffeomorphisms of the circle
The circle (SI) is arguably the simplest nonEuclidean differentiable manifold. It is compact (see Chillingworth, 1976, p. 143) so 'behaviour at infinity' is not a problem; it has no boundary so that dynamics can be studied without the complication of boundary conditions on the functions concerned and it is onedimensional. The dynamics of dilTeomorphisms on the circle therefore provide an ideal opportunity for us to illustrate the definitions given in § 1.2.1.
Some of the simplest examples of dilTeomorphisms on SI are the pure rotations.
They are easily defined in terms of the angular displacement (0) at the centre of the circle relative to a reference radius (see Figure 1.4). In terms of this local coordinate, an anticlockwise rotation by ex may be written as
1t(J(x)) = f(1t(x))
(1.2.5)
(see Figure 1.5), where
(1.2.2)
1t(x) = x mod 1 = O.
ThenJ is called a lift of f: SI + S· onto R.
(1.2.6)
Here we have assumed that 0 is measured in units of 21t. If ex = p/q, p, q E 1. and relatively prime, then
Proposition 1.2.1 Let J be a lift of the orientationpreserving homeomorphism f: SI + S·. Then
R:(O) = (0 + p) mod I = 0
(1.2.3)
J(x+ 1)=J(x)+ I
(1.2.7)
and we conclude (cf. Definition 1.2.2) that every point of the circle is a periodic point of periodq, i.e. the orbit of any point is a qcyc1e (see Figure 1.4). If ex is irrational then
for every x E R.
Proof. Observe that
R:(8) = (8 + mex) mod I ¥ 0,
(1.2.4)
f(1t(x)) = f(1t(x + 1))
(1.2.8)
for any 0 and, in fact, the orbit of any point fills the circle densely (see Exercise 1.2.1 ).
Obviously more general dilTeomorphisms of SI do not simply rotate all points uniformly. Crudely speaking they compress some arcs of the circle and stretch
because 1t(x) = 1t(x + I) by (1.2.6). If we substitute for J1t from (1.2.5), (1.2.8) becomes
1t(f(x)) = 7t(J(x + ]))
(1.2.9)
and it follows that
Figure 1.4 Typical orbit of the pure rotation R. for IX = »t« = 2/5. Observe that the orbit of 0 winds around the circle p = 2 times before returning to 0 on the fifth iteration.
l(x + 1) = ](x)+ k(x),
(1.2.10)
R~ (8)
where k(x) is an integer possibly depending on x. However, since] is continuous, k(x) must be continuous and this is only possible if k(x) == k E 1..
Ra (8)
Figure 1.5 Commutative diagram illustrating the definition of the lift of a circle homeomorphism f. The map It takes infinitely many equivalent points of R onto a single point of S·.
8
1 DifJeomorphisms and flows
1.2 Elementary dynamics oj diffeomorphisms
Suppose k;;:. 2, then l(x) and l(x + I) differ by more than two and] takes the form shown schematically in Figure 1.6(a). Clearly, the points Xo and XI satisfying 1(."(0) = 1 and ](xd = 2 are both less than unity. This means that 1t maps them to distinct points on SI. However, ](xo) and l(xd differ by unity and therefore represent the same point on SI. This contradicts the hypothesis that J is a homeomorphism. Hence k ~ 1.
If k = 0, 1(0) = 1(1) and 1 fails to be injective on (0, I) (see Figure 1.6(b)). Again
this contradicts the fact that J is a homeomorphism.
If k < ° then continuity of 1 can only be maintained if J is orientationreversing in contradiction to hypothesis. Moreover, it is clear that similar arguments would lead to a minus sign in the right hand side of (1.2.7) for orientationreversing f.
Finally, we conclude that k = 1 and (1.2.7) follows. 0
9
but not a diffeomorphism of SI. To obtain the latter, 1 must be a bijection and differentiable for all x e IR. An example of this type is shown in Figure 1.8(b) where
g(x) = x + t + (ij sin 21tx,
(1.2.12)
xe[O,I].
Notice, we have, without loss of generality, taken I(O)e[O, I) in both of the above examples. Observe that, 1t(j{x) + k) = 1t(j{x», for any k E 7L. Thus if l(x) is a lift of J then so is 1t(x) = 1(x) + k, k e 7L. Therefore, unless otherwise stated, we will assume that 1 is the member of this family of lifts satisfying 1(0)e [0, 1).
Figure 1.7 The function 1 shown here cannot be the lift of a homeomorphism [: Sl + SI because it is not injective.
It is important to realise that not every continuous function satisfying (1.2.7) is the lift of some homeomorphism. The function shown in Figure 1.7 is continuous and satisfies (1.2.7) but fails to be the lift of a homeomorphism because it is not injective. Figure 1.7 also highlights the geometrical significance of (1.2.7); namely that the graph of 1 in the interval [k, k + I] is obtained by shifting the graph of 1 in [0, 1] vertically by k units. In this way any continuous function g, defined on [0, I], that is injective, and such that g(l) = g(O) + 1, can be used to construct a lift 1 for some homeomorphism J: SI ..... SI. The function J is given by (1.2.5). A simple example of this construction is given in Figure 1.8(a) where
g(x) = _x2 + 2x +!, (1.2.11)
x E [0. I]. In this case, 1 is a continuous bijection but it is not differentiable at x = 1,2, .... This reflects on the corresponding J which is a homeomorphism,
Figure 1.6 Schematic forms for J when (1.2.10) has (a) k = 2; (b) k =0. In both cases, the hypothesis that f is a homeomorphism is contradicted.
Figure 1.8 The function shown in (a) is the lift of a homeomorphism. but not of a diffeomorphism, of the circle. Lifts of diffeomorphisms are differentiable functions of X, see (b) for example, where J is obtained from (1.2.12).
i
o LL_'L __ ' X
o Xo xII 2
(a)
i
31++
21++
o LL.J.__JL __ L_ X
o Xo XI 2
(b)
(a)
I
3
V
,/
V
/
V 2
o o
2
x
I
(b)
10
1 Diffeomorphisms and flows
1.3 Flows and differential equations
11
How are the fixed or periodic points of f: S' + S' related to the properties of the lift I?
replaced by
or
lq(x*) = x* + 1+ 1.
(1.2.20)
l(x*) = x*
(1.2.13a)
This point of view often has the advantage that 1. p, ... , 1', ... can be presented on the same diagram (see Figure 1.10) without ending up with a confusion of curves in the vicinity of y = x and y = x + 1.
The lift 1 of f: S' + S' not only provides a means of conveniently finding fixed and periodic points, it can also allow us to determine their stability. If (1.2.13a) is satisfied at x*, then the orbits of points near to x* under 1 can be obtained by moving between y = l(x) and y = x as in Figure 1.I1. The fixed point x* is stable (unstable) if
Proposition 1.2.2 Let f: S' + S' be an orientationpreserving homeomorphism and suppose that 1 is the lift of f with 1(0) E [0, 1). Then n(x*) is a fixed poim of f if and only if either
or
l(x*) = x* + I.
(1.2.13b)
ID1(x*)1 < 1
(> 1),
(1.2.21)
f(n(x*)) = n(x*) by (1.2.5) and n(x*) is a fixed point of [.
If ()* = n(x*) is a fixed point of f, i.e. f«(}*) = ()*, then
f(n(x*)) = n(x*) = n(f(x*))
(1.2.15)
(see any first course in Numerical Analysis). The stability of ()* = n(x*) is clearly ~e same as that of x*. When (1.2.13b) is satisfied, we can either replace 1 by f  I, so that x* is then represented by an intersection with y = x, and proceed as above or construct paths for the orbits of 1 by using y = l(x) and y = x + 1. The stability of the fixed point is still given by (1.2.21).
Proof. If f{x*) = x* (or .f(x*) = x* + 1) then
n(f(x*)) = n(x*) (or n(f(x*)) = n(x* + I) = n(x*)).
(1.2.14)
In either case,
(1.2.16)
1.3 Flows and differential equations
The iteration problem (1.1.1) for a diffeomorphism f: M + M given different Xo E M is equivalent to the study of the set offunctions {f"'lmE I}. This set has the property
by (1.2.5). Thus
f(x*) = x* + k, kE I.
Let x* = y* + I, IE lL, y* E [0, I) then (1.2.17) becomes l(y*) + I = y* + 1+ k.
(1.2.17)
Figure 1.9 Examples illustrating why (1.2.18) can only be satisfied if k = 0 or I forthe case when ! is an orientationpreserving homeomorphism. As 13 shows, if k > I then 1 fails to be injective. Notice (1.2.17) has a countable infinity of solutions for each solution to (1.2.18).
Here we have noted that a simple induction on l(x + 1) = f(x) + 1 gives 1(x + I) = l(x) + I. Thus, if (1.2.17) is satisfied for any x*, it must be satisfied for a point y* E [0, 1). Now.1( 1) = 1(0) + 1 and 1 is injective so that 1(0) :!;J(y) < 1(0) + 1 for yE [0, 1). Therefore, (1.2.18) cannot be satisfied unless k = 0 or 1 (see Figure 1.9).
o
y
y=x
(1.2.18)
Proposition (1.2.2) can be used to locate periodic points of f. Suppose that f has
lift 1, i.e. n(l(x)) = f(n(x)), XE IR, then .
n(p(x)) = n(I(f(x))) = !(n(f(x» = f2(n(x)). (1.2.19)
Thus P is a lift of f2.,1t only remains to ensure that P(O)E[O, 1) (i.e. choose the lift P  [p(O)], where [.] denotes the integer part of .), and Proposition 1.2.2 allows us to find the period2 points of f. These arguments obviously extend to points of period q > 2.
An alternative approach is to recognise that if 1q(0)E[I, I + I) then (1.2.13) is
i2(1)=iz(0)+ I 2 iJ(I) = iJ (0)+ I l. (I) = I, (0)+ I
12(0)
iJ(O) 11(0)
2
1.3 Flows and differential equations
13
12
1 Diffeomorphisms and flows
that:
and
increasing t. It can be shown (see Exercise 1.3.2) that there is one and only one trajectory of tp passing through each point x E M. If tp,(x*) = x* for all t E IR then x* is said to be a fixed point of the flow. Fixed points of flows can be stable, asymptoticaIly stable, neutraIly stable or unstable in the sense of Liapunov. Precise definitions are obtained by the transcription f'" H tp, and mE 71. H t ERin Definition 1.2.3 and the comments following it.
The orbit of a fixed point is just the point itself. If x is not a fixed point it is said to be ordinary or regular. The trajectory through an ordinary point gives rise to an oriented curve on M and tp has periodic points if this curve is closed.
(1.3.1)
for each i, j E 71.. It is said to be an action of the group 1. on M or, more precisely, the 71.action generated by f (see ChiJIingworth, 1976). In this section we consider the action of the group IR on M; such IRactions are caIled flows on M.
Definition 1.3.l A flolV on M is a continuously differentiable function tp: IR x M ~ M such that, for each t E R, the restriction tp(t, • ) = tp,( • ) satisfies
(1.3.2a) (IJ.2b)
(a) (b)
Definition 1.3.2 A closed orbit of a flow is a trajectory, y, which is not a fixed point but is such that tpt(x) = x for some x E y and r i O.
Observe that (I.3.Za and b) imply that (tp,)I exists and is given by tp_,. Since tp E C1, it follows (see Exercise 1.3.1) that tp,: M ~ M is a diffeomorphism for each
tER
Let us pursue the analogy with dilTeomorphisms a little further. We define the orbit or trajectory of tp through x to be {tp,(X)ltER} oriented in the sense of
Clearly, if tpt(x) = x the orbit returns to x after time r. If T is the least, positive time for which this occurs, x is a periodic point with period T. It is easily shown (see Exercise 1.3.3) that if a closed orbit has one point with period T, then every point of y is periodic with period T. Thus, T is also called the period of y.
The set of all trajectories of a flow is called its phase portrait. Since each trajectory corresponds geometrically to an oriented curve or point on M, a valuable pictorial representation of the flow is obtained by sketching or plotting typical trajectories. Some examples are shown in Figure l.lZ. Notice that the caption to this figure does not specify tp" instead a differential equation is given. How are flows related
  2 2 1
Figure 1.10 Plots of f'(x) vs x for (a) fIx) = x + x + 2;
(b) j(x) = x +! + fa sin 2xx. Observe that case (a) corresponds to a homeomorphism with a 3cycle but no fixed points or 2cyc1es. On the other hand, case (b) is the lift of a diffeomorphism with 2, 4 and 6cyc1es but no 1, 3 or 5cyc1es.
y=p(x) y=j2(x)
y=/(x)
(a)
Figure 1.11 Graphical illustration of the iteration x.+ 1 = f(x.) showing the stability of xt, xT + 1, .... Note that IDf(x*H < 1 for all these points. The remaining fixed points, xf, xf + 1, ... , satisfy IDf(x*H > 1 and are unstable. Observe that the graphical representation of the iteration can still give the stability of a fixed point x· even when IDf(x*)1 = I.
y=x
y
y=i(x)
y=i(x)
2
(b)
xi
xi+1 xi+2
14
I Diffeomorphisms and flows
J.3 Flows and differential equations
15
to differential equations? We define the velocity or vector field, X, of a flow fI' by
Thus, ~(t) is a solution of x = X(x) and, since fl'o = idM, ~(O) = fl'o(xo) = xo, as required. 0
X( ) _ dfl'. ()I _ L· {fI'(e, x)  fI'(O, X)}
x  x ,=0 1m
dt ,0 e
for each x E M. Geometrically, {fI'.(x)lt E R} defines a curve on M passing through x. The vector X(x) is directed along the tangent to this curve at x and has magnitude equal to the speed of description of the curve under the parametrisation by t. It is important to realise that, in contrast to vector fields defined on Rn, X(x)¢M. For each XE M, the set, TMx, of all vectors tangent to M at x is called the tangent space to M at x and X(X)E TMx. Figure 1.13 illustrates TMx for a typical point XES2. If M is an ndimensional manifold, then TMx is isomorphic to R" for all x EM. Each element of TM x corresponds to an equivalence class of curves on M having the same tangent vector at x (see Chillingworth, 1976, p. 164).
Proposition 1.3.1 fI'.(xo) is the solution of x = X(x) which passes through Xo at t = 0.
Proof. Let ~(t) = fI'.(xo). Then
e(t) = Lim{W + e)  W)}
,0 e
= Lim {fI't fI',(xo)  fI',(xo)}
,0 e
= Lim {fI'(e, ~(t))  fI'(O, ~(t))}
£+0 e
= X(~(t)).
(1.3.3)
Notice that ifX(x*) = 0 thenql,(x*) = x* is the solution oU = X(x) passing through x*. Moreover, if fI'.(x*) = x* for all t then (1.3.3) implies X(x*) = O. We conclude, therefore, that x* is a fixed point of fI'. if and only if X(x*) = O. Such points are referred to as singular points of the vector field X.
Proposition 1.3.1 means that every flow on M corresponds to an autonomous differential equation. Unfortunately the converse is not true. This is because there are autonomous differential equations with solutions that cannot be extended indefinitely in t. For example, x = x2 has general solution
!(CW1,
x(t)= 0,
(C'WI,
tE(OO,C); tER;
tE (C', (0),
(1.3.5)
C, C' E R. Only the trivial solution has domain IR. In such cases, local flows can still be defined. For example, the function
q>,(xo) = 1 Xo , (1.3.6)
xot
provides a local flow for x=x2. When xo>O, (1.3.5) implies tE(OO,XOI) in (1.3.6). It is easy to verify that q>, satisfies (1.3.2) provided t, sand t + s all belong to ( 00, Xo I). The same function q>, can be used when Xo < ° provided t is restricted
(1.3.4)
Figure 1.13 Illustration of the tangent space, TM., to the sphere 82 at x. Let the circles a and b define the latitude and longitude of x. If a and b are tangent to a and b, respectively, at x then 1M. = 8p{a, b}. Observe that b, c and d are all curves on the sphere having tangent vector h.
Figure 1.12 Some examples of I?hase portraits of flows: (a) 6 = z, i = sin 6; (b) 0 =sin 6,cP = 0; (c)O= 0(0  (3x/4»(0 x), q, = 6(x  0).
(a)
8
(b)
(c)
r
16
J DijJeomorphisms and flows
to the interval (xo 1, 00). Equation (1.3.6) obviously provides the trivial solution when Xo = O. This local flow is sufficient to characterise the solutions of .\: = x2 in a neighbourhood of the origin of the t, xplane. For example, for IXol < e, (1.3.6) certainly gives the solutions to x = x2 for t E (  s I, &  I). Flows of this type are frequently used implicitly when local properties are discussed (e.g. the saddlenode singularity in Example 2.7.4).
With the above proviso in mind, differential equations, vector fields and flows merely provide alternative ways of presenting the same dynamics. These alternatives have arisen for historical reasons; applications frequently lead to differential equations; local analysis is usually presented in terms of vector fields; and global analysis uses the language of flows. We hope the reader will become familiar with all three possibilities.
1.4 Invariant sets
17
The set of nonwandering points for f (tp) is called the nonwandering set, Q(f) (Q(tp)). It is easy to see that fixed points and periodic orbits lie in Q (see Exercises 1.4.2 and 1.4.3), however, points exhibiting milder forms of recurrence are also present. For example, consider an irrational rotation of the circle, SI. No point of the circle is periodic, but the orbit of any point x ultimately approaches x arbitrarily closely. Thus, every point of Sl is a nonwandering point and Q = SI.
The structure of Q will be examined more closely in § 3.6, but we can recognise some important subsets of it by formalising the idea that fixed points and closed' orbits frequently attract or repel the trajectories of phase points not contained in them.
Definition 1.4.2 A point y E M is said to be an {IX limit point of the w
trajectory of f (tp) through x if there is a sequence mi (ti) ~ { 00 such that +00
The set of all {IX limit points of x is known as the {IX limit set of x, denoted
w w
by {Lo(X). These sets are invariant under f (tp). Let z=f"'(y), mEl (z=tp,(y), L.,(x)
t e R), where y satisfies Definition 1.4.2. Then Lim (""+'"(x) = z (Lim tp,,+,(x) = z)
t.4 Invariant sets
Sometimes the orbit of a point under f or tp remains within a particular region of phase space for all mEl or t E It A set I\. £; M is said to be invariant under the diffeomorphism f (or flow tp) if {"'(X)EI\. (tp,(X)EI\.) for each xEI\. and all mEl (t E IR). We write
Lim (""(x) = y (Lim fP,,(x) = y).
j ..... eo iIX)
C"'(I\.) £; I\.
for all mEl
(1.4.1 )
or
tp,(I\.) £; I\.
for all tER
(1.4.2)
Invariant sets are said to be positipely (negatively) invariant if the orbits of their elements remain within them for mEl+ (Z") or t~O (t:::;;O).
Clearly, the orbit of any point is an example of an invariant set. It follows therefore that fixed points, cycles and closed orbits are all invariant sets. However, they are rather special in two main ways.
(i) They are minimal in the sense that they do not have any proper subsets that are themselves invariant. For example, the circle CC is an invariant set for both of the flows shown in Figure 1.14. In contrast to the flow shown in (a), the circle CC in (b) has proper subsets, PI' P2, fl and f2' that are invariant under the flow.
(ii) They exhibit periodicity. This is particularly important for applications where such sets frequently correspond to observable phenomena.
More subtle forms of recurrence than periodicity can occur in dynamical systems and the following definitions allow us to describe them.
ts ec
i+ 00
so that z and y belong to the same limit set of x.
Notice that IX and colimits sets are subsets of Q for any x. Recall if y ¢ Q then
Figure 1.14 The circle rc is an invariant set for both of the flows shown. However, in (a) rc has no proper subsets that are themselves invariant; while in (b) rc is the disjoint union of the invariant sets PI' Pz, r" rz.
Definition 1.4.1 A point x is a nonwandering point for the diffeomorphism f (or flow fP) if, given any neighbourhood W of x, there exists some m > 0 (t > to > 0) for which r(W)n W (cp,(W)n W) is not empty.
(a)
II
(b)
18
1 Diffeomorphisms and flows
1.4 Invariant sets
19
there exists a neighbourhood V3Y such that f"'(V)n V is empty for all m > O. However, yE Lco(x) implies f"'i(X)E V for i ~ N, say, and hence there is z = fN(X)E V such that f"'iN(Z)E V for i> N. Thus f"'(V)n V cannot be empty for all m and y must lie in n.
However, for Ixi > I, the Lim fI',,(x) does not exist for any sequence tj such that
ti +  00 as i + 00 and therefore La(x) is empty.
o
Example 1.4.1 Find L.(x) and L",(x) for (a) x = 0; (b) x # 0, when fI' is the flow on 1R2 induced by
Example 1.4.2 Let the flow fI' have the phase portrait shown in Figure 1.16. What are L",(x) and La(x) for x E A, B, C respectively? What feature do all three wlimit sets have in common?
;=r(lr), where (r, 8) are plane polar coordinates.
(1.4.3)
Solution. Sequences {tj}i=o can be constructed as in Example 1.4.1 to show that xEA:L.(x)={P1}; L",(x)=oA
xeB: L.(x) = {P2}; L.,(x) = oB
xEC:L.(x)=empty set; L",(x)=oAuiJB.
Let rA and rB be the trajectories of the flow which form the separatrices of the saddle point Po. Observe that
Solution. fI' has a unique, attracting closed orbit y given by r(t) = 1, with period T = 2n:, and an unstable fixed point at the origin (see Figure 1.15).
(a) x = 0
Note fI',(O) = 0 for all t therefore
Lo,(O) = La(O) = {O}.
(1.4.6)
(1.4.4)
and it follows that all three colimit sets are unions of fixed points and the trajectories joining them. 0
(b) x #0
Let y = (cos 80, sin 80)Ey and let ti be the sequence of t > 0 at which the orbit of
x crosses the radial line from 0 through y. Then Lim fI',,(x) = y and y is an wlimit
Example 1.4.2 illustrates an important theorem concerning the global properties of planar flows.
point ofx. This argument is valid for any y Ey and any x # O. Therefore, L",(x) = y for any x #0.
A similar argument allows us to show that
Theorem 1.4.1 (PoincareBendixson) A nonempty, compact limit set of a flow on the plane, which contains no fixed point, is a closed orbit.
Ixl< I, Ixl= 1.
(1.4.5)
This theorem states that the types of limit sets illustrated in Examples 1.4.1 and 1.4.2 are the only compact ones that can occur in flows on the plane. It is one of the few theorems which gives the existence of a global feature of a phase portrait.
Figure 1.15 Phase portrait for the flow of (1.4.3).
Figure 1.16 Phase portrait of the flow required for Example 1.4.2. The points PO.I.2 are fixed points. The open sets A, B have boundaries IJA, IJB, respectively. C is the complement of the closure of Au B.
r
C
aA
20
1 Diffeomorphisms and flows
Definition 1.4.3 A limit cycle is a closed orbit y such that either» £; L.,(x)ory £; L.(x) for some x¢y.
1.5 Conjugacy
21
Figure 1.17 Diagram illustrating conjugacy of: (a) diffeomorphisms; (b) flows. Note that (b) is valid for all leR and (1.5.1) implies that h(f'"(x)= g"'(h(x») for all mel.
Theorem 1.4.1 has the important corollary that a nonempty, compact set A which is positively or negatively invariant contains either a limit cycle or a fixed point. This result can be useful in demonstrating the existence of limit cycles (Arrowsmith & Place, 1982, pp. 14751).
1.5 Conjugacy
We now turn to the equivalence relations which allow us to recognise when two diffeomorphisms or two flows exhibit the 'same' behaviour. These equivalence relations lie at the heart of topological or qualitative theory.
Definition 1.5.1 Two diffeomorphisms f, g: M .... M are said to be topologically (or COo) conjugate if there is a homeomorphism, h: M .... M, such that
h·f=g.h.
(1.5.1)
Topological conjugacy of two flows tp" 1/1,: M .... M is defined in the same way with (1.5.1) replaced by b·tp, =I/I,·h for all tER.
Definition 1.5.1 means that h takes each orbit of r (or tp,) onto an orbit of g (1/1,) preserving the parameter m (t), i.e.
("'(x) ~ g'"(h(x»,
for each mEl,
(1.5.2)
or
tp,(x) ~ I/I,(h(x»,
for each t E lit
(1.5.3)
The significance of (1.5.2 and 1.5.3) is illustrated in Figure 1.17. Notice, by uniqueness of the trajectories of each flow, a given trajectory of tp, is mapped onto one and only one trajectory of 1/1, and vice versa.
h
h
(a)
h
(b)
Figure 1.18 Typical graph of a diffeomorphism f: R .... R for which ~(x) >? for some x E R. The fixed points of f are given by the mtersections of the curve y = f(x) and the straight line y = x.
Example 1.5.1 Let f: IR .... IR be a diffeomorphism with Df(x) > 0 for some x E IR. Given that the differential equation x = f(x)  x defines a flow tp,: IR .... IR, show that f is topologically conjugate to tpl'
Solution. If f is a diffeomorphism it is either an increasing or a decreasing function (differentiability of r 1 means that Df can never become zero). Since, Df(x) > 0 for some x, it follows I?f(x) > 0 for all x and f is an increasing function (see Figure 1.18). It follows that f can have any number of fixed points (including zero). Such points, xr, i = 1, 2, ... , are given by xr = f(xn and clearly coincide with the singular points of the vector field f(x)  x.
Let Xo be any point of the open interval (xt, xt+ 1)' The orbit of xo under both
y
x
22
1 Diffeomorphisms and flows
1.5 Conjugacy
23
f and <PI is confined to this interval and has the same orientation for both maps (N.B. sign(x.+ I  x.) = sign(f(x.)  x.) = signtf], nell·
Let xo, Yo e (xr, xr+ I) and consider the orbit of xo under f and the orbit of Yo under <PI' Let P, = f"(xo) and Q. = <P.(yo), lIel. Observe (see Figure 1.19) that
It follows that h: [xr, xr+ 1] + [xr, xr+ d defined by
for
y=xr
ye[Q., Q.+I]'
(1.5.8)
and
(1.5.4)
is a homeomorphism.
Finally, it is easy to verify that h exhibits the conjugacy of f and <P I' If x e [x]', xr+ I] then x e [Q., Q. + I] for some nand
ne I; are orderpreserving diffeomorphisms. Moreover, if xe[Po,PI] then f"(x)e[P., r., .I and similarly with x, P and f replaced by y, Q, <PI'
Our aim is to construct a homeomorphism on [xr, xr+ I] taking orbits of <PI onto orbits of I, preserving the parameter nel. To this end, let ho: [Qo, QI] + [Po, PI] be a homeomorphism, for example we might take
h '<pdx) = h.+ I '<PI(X)
= f"+ l·ho·<p_._I·<PI(x)
= {h.(x) =f·h(x)
(1.5.9) o
as required.
(1.5.5)
(1.5.6)
It is important to note that Example 1.5.1 highlights a special property of some increasing diffeomorphisms of the line. Not all diffeomorphisms on IR are topologically conjugate to the timeone map of some flow. For example. if f is a decreasing diffeomorphism on IR, the orbits of f oscillate about its fixed point (see Figure 1.20). Such behaviour is impossible for the timeone map of any flow on It
If h, in Definition 1.5.1, is a Cidiffeomorphism with k ~ I, rather than a homeomorphism then rand g (or fi', and t{I,) are said to be Ciconjugate. This kind of conjugacy is far more restrictive than topological conjugacy. For example, the real valued functions f(x) = 2x and g(x) = 8x, x e R, are topologically conjugate but they are not Ciconjugate for any k ~ I (see Exercise 1.5.1). Ciconjugacy of
Now, for ye [Q., Q.+ I], define
Clearly, h.: [Q., Q.+ I] + [P., r., I] and, what is more,
(1.5.7)
Figure 1.19 Orbits of the points Xo and Yo under f and <PI' respectively. It is convenient to define P; = !"(xo) and Q. = <P.(Yo), for nElL.
Figure 1.20 Graphical derivation of a typical orbit of a decreasing diffeomorphism f: R + R in the neighbourhood of its fixed point. The orbit clearly oscillates from one side of the fixed point to the other.
y
hll
x
'PI .... 1_....1~
y=x;~
y=x
24
I DifJeomorphisms and flows
1.5 Conjugacy
25
qI, and !{I, corresponds to there being a ktimes differentiable change of coordinates, h, which transforms the differential equation, x = X(x), of qI, into that, y = Y(y) say, of "". Recall ~conjugacy of qI, and t/I, means that there is a function hE Ck such that h(qI,(x)) = ",,(h(x)). Differentiate this equation with respect to t and evaluate at t = 0, to obtain
Figure 1.21 Various representations of the flowbox' containing the ordinary point Xo: (a) in the original coordinates; (b) in local coordinates at Xo and (c) using local coordinates defined by the flow lines.
v
{Dh(qI,(x)) dql, (X)}I = dt/l, (h(X))1
dt ,=0 dt ,=0
(1.5.10)
or
R"
Dh(x)X(x) = Y(h(x)),
(1.5.11)
since qlo = idM• Now consider the change of coordinates y = h(x) applied to x = X(x). With the aid of (1.5.11), we find
y = Dh(x)x = Dh(x)X(x) = Y(h(x)) = Y(y)
(1.5.12)
as required. Thus, when h exhibits the conjugacy of qI and "', the derivative map, Dh, transforms the vector field X(x) into Y(y) with y = h(x).
An important example of C'<coajugacy of flows occurs in the qualitative study of local phase portraits in the neighbourhood of an ordinary point. Let Xo be an ordinary point of the flow qI: R x R" > R" of the vector field X: jR" > jR".
Xo
For convenience, we will assume that H has normal X(xo) in the following discussion. Observe, (see Figure 1.21) that there is a neighbourhood, V, of x, such that any point XE V can be written as x = qI.(y), where YES. In other words, we can use the trajectories of the flow to define new coordinates on V.
These new coordinates are best related to local coordinates at xo, therefore, let x f+ x  Xo so that Xo is at the origin of both sets of coordinates. Now suppose we choose a basis in R" which has X(O) as its first vector. Then the first coordinate of every point YES is zero and S defines a neighbourhood, S, of the origin in R" t (see Figure 1.21 (b)). Each point of S can be specified by ~ E jR"  t and every point x of V can be written as
(b)
Definition 1.5.2 A local (cross) section at Xo is an open set, S, containing xo, in a hyperplane H £; R" which is transverse to X(xo).
X(O)
Xo =0
t/I,(u, ~) = (t + u, n
(1.5.14)
(u, ~)
x = qI.«O, ~)) = qI(u, (0, m = h(u, ~).
(1.5.13)
By definition of qI, h: R" + Rn is a C1function. What is more, hiS is the identity and Duh(O) = X(O), QY (1.3.3). Thus Det Dh(O) l' 0 so that h1 exists and is C1 by the Inverse Function Theorem. In the new coordinates, the trajectories of the flow are simply lines of constant ~ (see Figure 1.21 (c)), i.e.
u
(c)
Xo =0
(0, ~)
26
1 Diffeomorphisms and flows
1.5 Conjugacy
27
To show that "', and cp, are conjugate, observe that (1.5.14) implies h(",,(u, m = h(t + u, ~).
As Figure 1.22 shows
(1.5.15)
(1.5.19)
h(",,(u, ~)) = cp,(h(u, ~))
(1.5.17)
is a lift of the pure rotation Ri8) = (8 + y) mod 1. Thus R~(x) = x + ny and p(Ry) = y, i.e. the rotation number of R; is simply y itself. A rational rotation, R., (X = sl« E 0, cannot be topologically conjugate to an irrational rotation, Rp, p e R\ O. We saw in § 1.2.2 that the orbit of any point 8 under R. was periodic with period q, i.e. R:(8) = 8, while R';(8) # 8 for any 8e [0, 21l) or mE l.. Clearly, any map taking an orbit of RfJ onto an orbit of R. would fail to be injective. Therefore, the pure rotations with rational rotation number are topologically distinct from (i.e. not COconjugate to) those with irrational rotation number. Now, the pure rotations are diffeomorphisms on Sl and p(f) is defined for any diffeomorphism f: SI + s'. To what extent, therefore, can the above result for pure rotations be carried over to general diffeomorphisms on Sl1
However, (1.5.13) gives
h(t + u, ~) = cp(t + u, (0, ~)) = cp,(cp(u, (0, ~»
(1.5.16)
by (1.3.2). Thus
and " is Clconjugate to cp,. The arguments presented above essentially constitute a proof of the 'Flowbox' Theorem.
Theorem 1.5.1 (Flowbox) Let Xo be an ordinary point of the flow cpo Then in every sufficiently small neighbourhood of Xo, cp is Clconjugate to the flow ,(t, x) = x + tel' where el is a unit vector parallel to the xlaxis.
Proposition 1.5.1 A diffeomorphism f: Sl + SI has periodic points if and only if its rotation number, p(f), is rational.
The above examples emphasise that in order to prove two flows or diffeomorphisms conjugate, we must construct an appropriate map satisfying (1.5.1). It is often a great deal easier to recognise when no such map exists. For example, consider two flows: cp, with an isolated fixed point and "', with no fixed points at all. The fixed point is a trajectory of cp, and, therefore, if cp, and " are topologically conjugate, there is a homeomorphism which takes a trajectory of" onto the fixed point. However, every trajectory of" contains more than one point and can only have a single point image under a noninjective map. This contradiction proves that cp, and I{I, are not topologically conjugate. This result has an obvious extension: a necessary condition for two flows to be COconjugate is that they have the same number of fixed points. Here an easily recognisable property of the flows (namely, the number of fixed points) allows us to conclude that they are not conjugate.
Another, perhaps less trivial example of this approach, is afforded by diffeomorphisms on the circle. Let us begin by considering pure rotations.
A property that distinguishes rational and irrational rotations is their rotation number. This quantity can be defined for any homeomorphism f: Sl + s'.
Proof, If f has a periodic point then, given a lift, 1. of f, there exists x* e R such that
lq(x*) = x* + p,
(1.5.20)
for some integers p and q. It follows that l·q(x*) = x* + np, and therefore l"'l(x*)  x· np p
(1.5.21 )
nq nq q
Hence p(f) is rational.
To prove the converse, suppose f has no periodic points then,
lq(x) #x+ p,
(1.5.22)
or
lq(x)x#p,
(1.5.23)
for any integers p, q and any xe R. Since g.(x) = lq(x)  x satisfies gq(x + I) = gq(x),
Definition 1.5.3 The rotation number, p(f), of a homeomorphism f: SI + Sl is given by
Figure 1.22 Commutative diagram illustrating the connection between the pure rotation Ry and its lift Ry•
p(f) = (Lim l"(x)  ~) mod 1,
a+cc n
(1.5.18)
Rr x~x+a
where I is a lift of f.
As our notation suggests, it can be shown that p(f) is independent of the point x occurring in (1.5.18). A proof of this fact can be found in Nitecki (1971, pp. 334).
Rr
6=xmod 1'Rr(6)=(6+a)mod I
= (x+a) mod I
,
28
1 Diffeomorphisms and flows
1.6 Equivalence of flows
29
for each x, (1.5.23) means that there exists e > 0 such that either
for all x;
(1.5.24)
the orbits of the latter are parametrised by a continuous variable t. This allows us some additional freedom in the mapping of orbits onto orbits.
(1.5.25)
Definition 1.6.1 Two flows, fP, and "'" are said to be topologically (or Co) equivalent if there is a homeomorphism, h, taking orbits of fP, onto those of "'" preserving their orientation.
or
for all x.
Suppose (1.5.24) holds, thenlq(x) < x + p  s, for all x, and therefore l·q(x) = lq(P·I)q(x» <p.Ilq(x) + p  e
= lq(P·2Iq(x)) + p  e < ... < x + n(p  e). (1.5.26)
Similarly, when (1.5.25) is valid
Since equivalence only demands that orientation be preserved, we allow h(fP,(x» = "'.,(lI(Y)' with Y = h(x), where t, is an increasing function of t for every y (see Figure 1.24). This relaxation of the requirement that the parameter t be preserved, provides more satisfactory equivalence classes for flows. For example, the planar differential
Figure 1.23 (a) Sketch of 1'(x). Observe that, since 1'(1) = 1'(0) + 1, if a fixed point, x~, occurs then there must be at least one further fixed poi~t x]. Moreover, if 4 is stable then xT must be unstable. (b) Example of f3(X) for a circle diffeomorphism with a stable 3cycle. Note that an unstable 3cycle must also occur. (e) Illustration of periodic points of f on the circle for the lift shown in (b).
l·q(x) > x + n(p + e).
(1.5.27)
Thus, Lim [l"(x)  x]/nq is either greater than (p + e)/q or less than (p  e)/q,
for any integers p and q, and so p(f):I (p/q) mod I.
o
Typically, circle diffeomorphisms with rational rotation number, p/qEQ, have an even number of periodq cycles. A sketch of lq(x) (see Figure 1.23) not only reveals why the number of cycles is even, but it also shows that the stable and unstable points alternate around the circle.
The following result shows that circle diffeomorphisms with irrational rotation number can behave like irrational rotations.
Theorem 1.5.2 (Denjoy) If an orientationpreserving diffeomorphism f: Sl .... SI is of class C2 and p(f) = P E IR\ C, then it is topologically conjugate to the pure rotation Rp.
For a proof of Denjoy's Theorem the interested reader should consult Arnold (\983, pp. 1056) or Nitecki (1971, pp.459). This important result means that every orbit of f is dense in the circle provided f E C2 and p(f) is irrational. If f f C2, then more complicated phenomena, such as invariant Cantor sets, can occur (see §6.4.l and Nitecki, 1971).
(a)
x
(b)
x
o x~ xi
1.6 Equivalence of flows
Topological conjugacy is arguably the natural equivalence relation for maps. A homeomorphism h is used to take successive points in the orbit of one map, f, onto those of another map, g. Given that the aim is to capture the fact that the orbits of f and g behave in a similar way, continuity of h and its inverse is the least we should demand. Moreover, since the orbits of a map are sequences of discrete points, it is hard to envisage anything more sensible than mapping orbits onto orbits in the manner described above. However, this is not the case for flows.
From this point of view, the important difference between maps and flows is that
(e)
30
1 Diffeomorphisms and .flows
J.6 Equivalence of flows
31
r=tr(lr),
where O'(y) = iy(O) is a positive scale factor altering the magnitude but not the direction of Y(y).
equations
r= r(lr),
(1.6.1) (1.6.2)
Example 1.6.1 Show that the vector fields Jx and Jox, with
where (T, B) are polar coordinates, have similar phase portraits. Both have an attractive closed orbit y with r(t) == I and an unstable focus at the origin. However, the closed orbit has periodZz in (1.6.1) and periodIt in (1.6.2). Thus, if h: y + y preserves the parameter t, it must fail to be a bijection. Thus (1.6.1) and (1.6.2) are not topologically conjugate, but they are topologically equivalent. Observe that the time rescaling tf+ 2t transforms (1.6.2) into (1.6.1).
If Definition 1.6.1 is satisfied with hECl, k~ 1, then the stronger relationship between fI' and", can be emphasised by saying that they are Ckequivalent. If two flows fI' and", are Ckequivalent (k ~ 0) then their vectors fields X(x) and Y(y) are also said to be Ckequivalent. This terminology is frequently used because flows are often described implicitly in terms of their vector fields. For example, in applications one is often provided with a model differential equation but no explicit form for its solutions.
When k ~ I there is a Ckdiffeomorphism, h, such that
where IX, P > 0, are topologically equivalent.
and
I)
1 •
(1.6.5)
Solution. The differential equations x = Jx and x = Jox are easily solved using plane polar coordinates. We find x = Jx gives r = IXT, 8 = fJ with solutions
r(t) = ro exp(at),
(1.6.6)
The equation x = Jox becomes R = R, e = 1 and its solutions are
(1.6.7)
Since /3 > 0, the flows defined by (1.6.6) and (1.6.8) are topologically equivalent with h equal to the identity. In other words, they have identical trajectories and differ only in the speed at which they are described.
Elimination of t from (1.6.7) and (1.6.8) gives
!_ = (!_)f1/«,
R 0 = (B  Bo) + 00, (1.6.9)
o ro
Equation (1.6.9) defines a map taking the trajectory of (1.6.8) through (ro, Bo) onto the trajectory of (1.6.7) through (Ro, 00) (see Figure 1.25). For r, ro > 0, this map is 1 :1, continuous and preserves orientation (indeed it preserves t itself);
Figure 1.25 Illustration of the effect of the map (1.6.9) on the orbit of (1.6.8) through (ro• 00), The result is the orbit of (1.6.7) passing through (Ro.eo)·
R(t) = Ro exp(t), If we let t 1+ /3t in (1.6.6), we obtain
r(t) = ro exp(<xt//3).
Dh(x)X(x) = O'(h(x»Y(h(x»
(1.6.3)
(cL (1.5.11), where 0': R" ..... IR takes only positive values corresponding to the reparametrisation of the time. Recall the vector field of ""y(l) is given by
dt/l'y(l) (y)1 = (iy(t) ~ (y»)1
dt 1=0 dry 1=0
= O'(y)Y(y)
(1.6.4)
Figure 1.24 Topological equivalence requires trajectories to be mapped onto trajectories preserving their orientation rather than t itself. Thus. t,(t) is an increasing function of t that is continuously parametrised by y and satisfies f,(O) = O. For example. when ty(t) takes the form shown in (a). h relates 1I',(x) and ';,,(I)(h(x» as indicated in (b).
/ /
/
/
/
t,(/2) /,
/ :
r,(tl) /. !
/: :
(1.6.9)
",(h(x»
(a)
(b)
/ / /
/
/
/
/
/
/
/
/
/
(1.6.8)
32
1 Diffeomorphisms and flows
1.7 Poincare maps and suspensions
33
however, it involves four parameters. In fact, (1.6.9) represents a family of maps of the plane onto itself. We require a single homeomorphism taking each trajectory of (1.6.8) onto an orbit of (1.6.7) and, therefore, we must choose values for the parameters.
Observe, every trajectory of (1.6.7) and (1.6.8) crosses the unit circle once and only once. Let us choose to map the orbit of (1.6.8) that crosses the unit circle at angular coordinate 00 onto the orbit of (1.6.7) that crosses the unit circle with angular coordinate 00 = 00. The map h obtained in this way is given by setting ro = Ro = 1 and 00 = 00 in (1.6.9), i.e.
0=0,
(1.6.10)
that the flows of Ax and Jx are linearly conjugate. Examples 1.6.1 and 1.6.2 show that all such vector fields are topologically equivalent to the vector field x.
The complete classification of linear vector fields on 1R2 is summarised in Figure 1.26. Each point of the (Tr A, Det A)plane represents a similarity class of real, 2 x 2 matrices. The striking feature is that the vast majority of points in Figure 1.26 correspond to vector fields of stable, unstable or saddle type. Such linear vector fields are said to be hyperbolic (see § 2.1) and Figure 1.26 suggests that hyperbolic behaviour is 'typical' for linear vector fields on R2. The point to note is that, without a suitable equivalence relation, the idea of what is typical has no meaning. We will return to the question of typical or generic properties of flows and diffeomorphisms in § 3.1.
with r > 0, 0 ~ 0 < 2n. Thus, if we define h(O) = 0, we have constructed a homeomorphism which exhibits the topological equivalence of (1.6.7) and (1.6.8). Since we have already established the equivalence of (1.6.6) and (1.6.8), we finally conclude that Jx and Jox are topologically equivalent. 0
Solution. Let h be given by
h(x) = h(r cos 0, r sin 0)
= {r cos(O In r), r sin(O In r), (0,0),
r>O r=O.
(1.6.11)
1.7 Poincare maps and suspensions
We have already noted that the flow map If',: M + M is a diffeomorphism for each fixed t. Thus, one way of obtaining a diffeomorphism from a flow is to take its timet map, If't: M + M, r > O. Clearly, the orbits of flit are constrained to follow the trajectories of the flow because {1f'~(x)lmEZ} = {If'mt(x)lmEZ}!; {If',(X)ltER}. This means that the dynamics of If't are strongly influenced by the flow If' and they are not typical of those of dilTeomorphisms on M. It is perhaps worth stressing that, while the orbits of x under If't, and If'n' tl # f2' behave in a similar way for any x E M (because both are subsets of the same trajectory of If'), the two maps are not necessarily of the same topological type. For example, suppose If', has a closed orbit)' of period Tand that tl = exT, exE 0, whilst t2 = PT, PE IR\O.1t follows that If't. has an invariant circle)' consisting entirely of periodic points (of period q if ex = p/q). The same closed curve y is invariant for If'" but the orbit of any point XEy under If'" fills out y densely. Therefore, If'" cannot be topologically conjugate
Example 1.6.2 Use the map r' = r, 0' = 0 In r (r > 0) to demonstrate that the vector fields Jox, where Jo is given in (1.6.5), and x are topologically equivalent.
The map h: R2 + 1R2 is continuous and has continuous inverse, r = r', 0 = 0' + In r', r' > O. Since h(O) = 0, h takes the fixed point trajectory of the flow of Jox onto that of x. For x # 0, h is dilTerentiable so we can check its effect on the flow by transforming the differential equation x = Jox or, in polar coordinates, f = r, e = I. We find
f'=;=r=r'
and
. . i .
0' = 0   = 0  I = 0, r
(1.6.12)
Figure 1.26 Topological types of all linear vector fields on the plane. Each point in the (Tr A, Det A)plane corresponds to an equivalence class of linear vector fields. Details of the derivation of this diagram are given in Chapter 2 of Arrowsmith & Place (1982). The differential equation i = x has Tr A = 2, Det A = 1 and is therefore unstable.
which is just the polar form of x = x. Of course, h is not differentiable at the origin so that (1.6.11) is only a homeomorphism of the plane. Hence Jox and x are topologically equivalent. 0
When two flows are topologically equivalent we say they are ofthe same topological type. The results obtained in Examples 1.6.1 and 1.6.2 play an important role in the classification, up to topological type, of all linear vector fields on 1R2 (see Arrowsmith & Place, 1982, p. 58). The matrix J in (1.6.5) is the real Jordan form of any 2 x 2 real matrix, A, with complex eigenvalues ex ± iP, ex > 0, i.e. there is a real nonsingular matrix M such that M1 AM = J. It follows (see Exercise 1.5.6)
DetA
STABLE 1 UNSTABLE
i
"""·""·roon simple 
SADDLES
I
TrA
,
34
1 Diffeomorphisms and flows
1.7 Poincare maps and suspensions
35
to tp", i.e. the maps are of different topological type. We will have cause to return to timet maps of flows in Chapter 5.
Another, more significant, way of obtaining a diffeomorphism from a flow is to construct its Poincare map. Let tp be a flow on M with vector field X and suppose that L is a codimension one submanifold of M satisfying:
(i) every orbit of tp meets L for arbitrarily large positive and negative times;
(ii) if x E L then X(x) is not tangent to L.
Then L is said to be a global (cross) section of the flow. Let YEL and r(y) be the least, positive time for which tpt(y)(Y) E L.
of flows in one higher dimension. For example, the Poincare map P(x) in (1.7.5) has a fixed point at x = 1 (observe x* = P(x*) implies (I  x*)(1  a) = 0, which is only satisfied for x* = 1). Furthermore, if x ~ 1, then P(x) § x so that x = 1 is an attracting fixed point. This fixed point in P dearly corresponds to the stable limit cycle in the phase portrait of tp (see Figure 1.15).
Another example is afforded by the flow on the torus, T2, defined by
O=a,
(1.7.7)
where (J and tp are as shown in Figure 1.27. The equations (1.7.7)have solutions
and
(1.7.8)
r(t) = ro/{ro + (1  ro) exp( tn,
(1.7.4)
reduced mod 2n, so {qJ first returns to {qJo when {t = tip, where Pt", = 2n. Thus if
o (Jo t = t8 at8 = 2n
a/p=p/q, p,qEI+ and relatively prime, then qt",=pto and the orbit through «(Jo, qJo) returns to this point after q revolutions around the torus in the qJsense and p revolutions in the (Jsense. It follows that if cx and P are rationally related then every point of T2 is a periodic point of the flow, i.e. every point lies on a closed orbit. If on the other hand a and P are not rationally related then the orbit through «(Jo, qJo) never returns to that point although it approaches it arbitrarily closely.
A global section of the torus is obtained by taking qJ = qJo, a constant, when L is a circle, SI, with coordinate (J. Since the orbit of the flow first returns to qJ = qJo after time t", = 2n/p and (J = cxt + (Jo, we conclude that the Poincare map, P: Sl + SI, is a rotation by 2ncx/p. The properties of pure rotations (see § 1.2.2) obviously reflect the behaviour of the flow described above.
There are flows for which there is no global section (see Exercise 1.7.2). Therefore, it is not true to say that every flow corresponds to a diffeomorphism by taking Poincare maps. However, the converse is true, i.e. every diffeomorphism f is the Poincare map of a flow  called the suspension of f. This is a very important
Definition 1.7.1 The Poincare (or first return) map for L is defined to be
P(Y)=tpt(y)(Y), YEL
(1.7.1)
Example 1.7.1 Obtain the Poincare map, P, of the flow defined by
;=r(lr);
(J = 1,
r>O,
(1.7.2)
where (r, 8) are plane polar coordinates, taking L to be the halfline 8 = O. How does P change if L is taken to be the halfline (J = Oo?
Solution. The phase portrait of (1.7.2) is shown in Figure 1.15. L is the positive xaxis in the plane and (1.7.1) can be written
x>O,
(1.7.3)
where tpt(r, (J) E R2 is the flow of (1.7.2), (')x denotes the xcomponent of " and r(x) is the time taken for a phase point at XEL to make one complete revolution about the origin. Since 8 = I, t(x) = 2n.
The radial equation, r = r( 1  r), has solution
with r(O) = ro, so that
P(x)=x/{x+(Ix)a},
x>O,
(1.7.5)
Figure 1.27 Diagram showing how the coordinates, 0 and cp, used in (1.7.7) are defined.
where a = exp(  2n) < I.
If L is taken to be the halfline 0 = 00, then (1.7.3) is replaced by
P(r)= (tp,(,)(r, (0»"
(1.7.6)
where r(r) = 2n and (.), denotes the radial component of '. We, therefore, conclude that P takes the form q.7.5) with x replaced by r, the radial distance along 0 = 00,
D
By construction P: L + L is a diffeomorphism and dim L = dim M  I. In contrast to timer maps we, therefore, expect these diffeomorphisms to reflect the properties
,
36
J DijJeomorphisms and flows
J.7 Poincare maps and suspensions
37
observation. It means that any result that can be proved for diffeomorphisms should have a counterpart for flows in one higher dimension (see Smale, 1967). The following explicit definition is given on p. 59 of Arnold & Avez, 1968.
An alternative way of viewing Definition 1.7.2 is to think of linking (x, 1) and (f(x), 0) by a smooth 'fibre' of unit length along which the suspension is considered to flow. This must be done for each x EM. Since f is a diffeomorphism, if y E M is close to x then f(y) is close to f(x) and the fibres of the identification lie close to each other. If we were to take a finite sample of these fibres we should obtain something resembling unit length of a, possibly twisted, multicored electrical flex.
Obviously, this procedure does not define the precise shape of the identifying fibre or, in other words, it does not uniquely determine the suspended flow. What is important is that the component of the flow in the new dimension is never zero. It then follows that all admissible shapes of the identifying fibres give rise to topologically equivalent suspended flows. The flow given in (1.7.9) is a particular
Definition 1.7.2 The flow
",,(x, 6) = (fI' +8)(X), t + 6  [t + 6]),
(1.7.9)
where x EM, 0 E [0, I] and [ .] denotes the integer part of " defined on a compact manifold M by identification of (x, 1) and (f(x),O) in the topological product M x [0, 1], is called the suspension of the dijJeomorphism f: M + M.
It is easy to verify that ",,(x, 6) in (1.7.9) formally satisfies the requirements of Definition 1.3.1. Geometrically, (1.7.9) corresponds to considering the product M x [0, 1] and taking a unit vector field in the [0, l]direction. Now imagine identifying the lend and the Oend in such a way that (x, 1) is attached to (f(x), 0) for each x E M (see Figure 1.28).
It must be pointed out that the manifold M is not always M x Sl as Figure 1.28 suggests. M x S' is obtained if f is continuously deformable, through diffeomorphisms, to the identity. For example, if we let M = Sl and f be a rotation then if is the torus T2 = s' X st. However, if f is a reflection in a diameter of the circle then M must be a Klein bottle to achieve the identification of (x, I) and (f(x), 0). Another, perhaps simpler example is to let M = (0, 1) and f be reflection in x = 1. As Figure 1.29 shows, M is a Mobius strip.
Figure 1.29 The suspension of the diffeomorphism f: (0, 1) .... (0,1) given by reflection in x =! is defined on a Mobius band. The twist in the manifold on which the suspension is defined arises because (x, 1) must be identified with (f(xl, 0).
Figure 1.28 Schematic illustration of the construction of the suspension of a diffeomorphism r that is continuously deformable into the identity: (a) before; (b) after; identification of [x, 1) and (f(x),O).
(x, 0)
(a)
A
A
B
(b)
A
o
B
r
38
I Diffeomorphisms and flows
1.8 Periodic nonautonomous systems
39
representative of this equivalence class which clearly exhibits the connection with the diffeomorphism f. When looked at from this point of view it is easier to understand how the nature of f (whether or not it is deformable to idM) affects the resultant manifold on which the suspension is defined.
Figure 1.30 (a) Schematic representation, in the extended phase space, M x R, of some possible solutions of the nonautonomous system (1.8.1), (1.8.2). (b) Corresponding solutions of the autonomous equation (1.8.3), (1.8.4) on M ~ SI.to = M x {O} is a global section for the now of (1.8.3), (1.8.4) and this allows us to define the Poincare map Po: to ..... t8.
1.8 Periodic nonautonomous systems
An important application of the ideas developed in § 1.7 is in the analysis of differential equations of the form
x = X(x, 0),
(1.8.3)
periodT
x = X(x, t),
xEM,
(1.8.1)
where
X(x, t + T) = X(x, t),
(1.8.2)
for all tER. The transformation tHtly, X(x, t)HYX(X, 'It), with j » T12n, allows (1.8.1) to be written as the autonomous system
defined on M x IR, where
X(x, 0 + 2n) = X(x, 0)
(1.8.4)
I+T
1+2T
for all 0 E R (see Exercise 1.8.1). It is then convenient to identify 0 + 2nm, mE 1., with 0 to obtain a differential equation on M x SI, where 0 is the circular coordinate. This procedure is illustrated in Figure 1.30 where some possible solutions of (1.8.1 and 2) are shown. Observe that the solutions are not necessarily periodic (see Exercise 1.8.2). However, it is easily verified that if ~(t) is a solution of (1.8.1 and 2) then so is W + T) (see Exercise 1.8.2). i.e. advancing a solution by one period of the vector field also gives a solution.
Figure 1.30 helps us to associate this 'period advance map' of the nonautonomous system with the Poincare map P 8: ~8 + ~8 of (1.8.3) defined on the global section, ~8 = M x {O} of M x SI. It is worth noting that P 8 and P 8" O:f 0' are topologically conjugate (see Exercise 1.8.4). Thus, in discussing topological properties it is sufficient to consider P = Po. Conversely, we can associate the solutions of the nonautonomous system with the suspension, on Mx SI, of the Poincare map, P, which is itself a diffeomorphism on ~o = M x {OJ.
There is a complete correspondence between the properties of the Poincare map, P, and those of its suspension. For example, P has a fixed point x· if and only if its suspension has a closed orbit of period 2n, i.e. if and only if the nonautonomous system has a periodic solution of period T, Figure 1.30 shows a 2cycle of P along with the corresponding solution of (1.8.1,2) with period 2T. Furthermore, a periodic solution of (1.8.1,2) is stable (asymptotically stable), in the sense of Liapunov, if and only if the associated periodic point of P is stable (asymptotically stable). The following example shows how this last result can be applied.
(a)
2m
(}=  mod 2n T
(b)
,
40
J DiffeomorplJisms and flows
1.8 Periodic nonautonomous systems
41
Example 1.8.1 Find the period advance map for the nonautonomous system
x + [w(t)fx = 0,
(1.8.5)
To show that Det P = I, note that (1.8.8) implies that d
 «(jI(t, 0)) = A(t)(jI(t, 0), dt
(1.8.11 )
where w(t) = w(t + T), t E It Obtain the Poincare map P and show that Det P = 1. Hence, deduce that the null solution of (l.8.5) is stable (in the sense of Liapunov) if ITr PI < 2 and unstable if ITr PI > 2.
Solution. The secondorder equation (1.8.5) can be written in the firstorder form
since Q(t) = A(t)Q(t). It follows (see Exercise 1.8.6) that, if W(t) = Det«(jI(t, 0)), then W(t) = Tr(A(t))W(t) = 0 for (1.8.6). Hence W(t) = W(O) = Det«(jI(O, 0» = 1 and, in
particular,
x = A(t)x, where x = (XI' X2)T = (x, .X)T E R2 and
(1.8.6)
W(T) = Det«(jI(T, 0)) = Det P = I.
(1.8.12)
( 0 01).
A(t)=
 [w(t)] 2
The solutions of (1.8.6) form a twodimensional vector space (see Exercise 1.8.3). The solution ~(t) satisfying ~(to)=xo can be written in the form
(I.8.7)
The null solution of (I.8.6) corresponds to the fixed point of P at the origin.
The stability type of the null solution is the same as that of the fixed point and the latter is determined by the eigenvalues, )'1,2' of P. Since Det P = 1, the characteristic equation of P is A 2  (Tr P)A + 1 = 0 and
(1.8.13)
(1.8.8)
If ITr PI < 2 then (Tr p)2 < 4 and the eigenvalues are complex with AI = A! = exp(ip) (since AIA2 = 1), where tan P=[4 (Tr p)2] I/z{fr P. Let u +ir, u, vER2, be the eigenvector of P with eigenvalue AI' Then the matrix K = (v: u) is such that
KI PK = (cos P sin P), (1.8.14)
sin P cos P
i.e. P is conjugate to a rotation about x = O. It follows that the orbit of x " 0 under P lies on an ellipse and, consequently, the fixed point at x = 0 is stable in the sense of Liapunov (see Figure 1.31).
If ITr PI > 2, then A1.Z are real With AI = ). (IAI > 1) Az = A I. In this case, there is a nonsingular K such that
KIPK = D = (~ A ~} (1.8.15)
where the columns of Q(t) form a basis for the solution space of (1.8.6). Q(t) is called a fundamelltal matrix for the problem (see Jordan & Smith, 1977) while (jI(t, to) is known as the state trallsitiol1 matrix (see Barnett, 1975). Now observe that
W + T) = Q(t + T)QI(tO)XO = Q(t + T)QI(t)Q(t)QI(tO)XO = (jI(t + T, t)~(t).
Thus (jI(t + T, r): L, .... L,+T (see notation in Figure 1.30) is the period advance map at t. Clearly, if ~ and" are solutions of (1.8.6) then
(jI(t + T,t)(aW) + b,,(t»= a(jl(t + T,t)~(t) + b(jl(t + T,t),,(t), a,b~R, and the period advance map is linear for any t. Moreover, (jI(t + T, t) = Q(t + T)QI(t)
= Q(t + T)[QI(T)Q(T)QI(O)Q(O)]QI(t)
= (jI(t + T, T)(jI(T, O)(jI(O, t).
(1.8.9)
Figure 1.31 When ITr PI < 2, the orbits of points x 7" 0 under P lie on ellipses as shown. Observe that, for any xeN', P"xeN for all mel. Thus, x = 0 is stable in the sense of Liapunov (see Definition 1.2.3).
It can be shown that:
(i) (jI(t + T, to + T) = (jI(t, to);
(ii) (jI(t, Or I = (jI(O, t);
(see Exercise 1.8.4) so that (1.8.9) can be written in the form (jI(t + T, t)(jI(t, 0) = (jI(t, O)(jI(T, 0).
Xl
(1.8.10)
This result shows that (jI(t + T, t) and (jI(T,O) are topologically (indeed linearly) conjugate and consequently, for the qualitative behaviour of the solutions of (1.8.6), we can focus attention on (jI(T, 0). Expressed in terms of 6, (jI(T, 0) = Po = P: RZ .... 1R2, the Poincare map for (1.8.6).
,
1.9 Hamiltonian flows and Poincare maps
43
42
1 DifJeomorphisms and flows
Here the orbits of P lie on hyperbolae and, as Figure 1.32 shows, the x = 0 is an unstable fixed point. 0
H = H(q, p) is the Hamiltonian for the system and the equations (1.9.1) are known as Hamilton's equations. The state of the system at time t is specified by
Example 1.8.1 suggests that periodic perturbations of the frequency, (0, of a harmonic oscillator can destabilise the equilibrium point with x = O. This is essentially what a child on a swing achieves by appropriate movements of weight, in order to build up the amplitude of the oscillations of the swing. A simple example illustrating how this instability can be achieved is given in Arnold (1973, pp. 2056). This phenomenon is known as parametric resollance.
X(t)=(q(t)). p(t)
The con[l{]uration, q(t), of the system is given by the n generalised coordinates ql(t) and p(t) consists of the n conjugate generalised momenta PI(t). A system with n degrees of freedom is often called an nF system.
In general, ql and Pi change with t but H does not. Observe
for all t, by (1.9.1). Thus, H is a conserved quantity or a constant of the motion. Alternatively, (1.9.1) is an autonomous system of differential equations which defines a Hamiltonian flow, fI'~: U _. R2". Equation (1.9.2) means that H is constant on the trajectories of fI'~ i.e. H is a first integral for (1.9.1) (see Arrowsmith & Place, 1982, pp. 1016).
In general, Hamiltonian flows occur on differentiable manifolds and Definition 1.9.1 is valid for each chart (U., h.). Thus (see Figure 1.33) H.: U. _. R gives rise' to a vector field XH., via (1.9.1), for each IX. Moreover, when W.n W/l (IX # P) is nonempty, the two sets of local coordinates on U. and U (J are related by the
1.9 Hamiltonian flows and Poincare maps
Another application of Poincare maps, that is of current research interest, lies in the study of noninteqrable, conservative Hamiltonian systems. While the reader will no doubt have encountered the integrable case in a Classical Mechanics course, it will be useful to review the basic ideas emphasising the connection with flows.
Definition 1.9.1 Let U be all open subset of R2• and H: U _. IR be a twice conunuousl» differentiable function. The system of differential equations x = XH(x), XJI: U t R2• given by
i= 1, ... , n,
(1.9.1)
(1:9.2)
where x = (qt, ... , q., Pt, ... , p.)T is said to be a conservative Hamiltonian system with udegrees of freedom.
Figure 1.33 Illustration of the way in which a Hamiltonian function defined on a manifold M gives rise to Hamiltonians, H. and H~ on the charts (U., h.) and (U" h,), respectively.
H:MtR
Figure 1.32 For ;. > 1 the hyperbolae XIXZ = C, C # 0, are invariant curves for the map Dx, where D is given by (1.8.15). The origin is a hyperbolic saddle point and therefore for every N' £:: N, there exists x EN' for which P''x ¢ N, for some mE;r . Hence the saddle point is unstable in the sense of Liapunov.
UaR Ha
=H·h;.I
R
yW'''z
UfJ Hp R
=H.hpl
44
1 DijJeomorphisms and flows
1.9 Hamiltonian flows and Poincare maps
45
overlap map h.1I (see Figure 1.3). Thus. if x = (ql' ., ., q •• PI"'" p.)T in U. and y = (QI' ... , Q., PI' ... , p.)T in U /l represent the same point on M, then
Clearly, (1.9.5) is satisfied if and only if the overlap map h.p is such that
[Dxh.p(xj]YSDxh.p(x) = S, (1.9.13)
(1.9.3)
and
(1.9.4)
Definition 1.9.2 A dijJeomorphism h: U + RZ., U C;;; !Rz., is said to be symplectic if
Of course. we require that the vector fields XH• and XHII give rise to the same dynamics on the overlap between two charts and this imposes constraints on the manifold itself. To make the dynamics on W. and Wp agree on W.n Wp, we demand that
[Dh(xWSDh(x) = S
for all x E RZ., with S = (~ ~) where I is the n x n identity matrix.
(1.9.14)
D.h.,(x)XH.(X) = XHII(h.p(x)) (see (1.5.12)). Now, differentiation of (1.9.3) gives
D.H.(x) = DyHp(h.p(x»D.h./1(x).
Equation (1.9.6) looks more familiar in component form, i.e.
(aH. aHa en, aH.)
aql"'" aq.' apl" .. , apR
(1.9.5)
(1.9.6)
A differentiable manifold for which all the overlap maps satisfy (1.9.13) is said to be a symplectic manifold. The theory of symplectic manifolds provides a coordinate free approach to Hamiltonian mechanics (Abraham & Marsden, 1978; Arnold, 1968).
It is important to realise that (1.9.13) is sufficient to ensure that the form (1.9.1) of Hamilton's equations is valid on both U. and U /1 (see (1.9.8,9». The arguments involved in obtaining (1.9.13) are not confined to overlap maps. Consider the effect of a coordinate transformation, h, on a Hamiltonian system defined on !R20• If we demand that the equations of motions of the new coordinates be derived from the transformed Hamiltonian by applying (1.9.1), then we can conclude, by precisely the same steps as we have used above, that h must be symplectic. However, preservation of Hamilton's equations in this sense is the property that defines canonical transformations in Classical Mechanics. Thus symplectic and canonical transformations are one and the same thing.
A property that distinguishes a Hamiltonian flow, If'~, from other flows of even dimension is that If'~ preserves volumes of phase space.
ap. sr,
aql apo
. (1.9.7)
Furthermore, (1.9.1) implies
(1.9.8)
(1.9.9)
Theorem 1.9.1 (Liouville) Let If', be the flow induced by x = X(x) and net) be the volume of the image, 1f',(D), of any region D of its phase space. If div X == 0, then If', preserves volume, i.e. net) = nCO) for all t.
and
with S = (~ ~) and I equal to the n x n unit matrix. Operating from the left with [D.h./1(X)]TS in (1.9.5) gives
[D.h./1(xWSD.h.p(x)XH.(x) = [D.h.p(xWSXH,(h.p(x)), (1.9.10)
= [D.h.p(xW[DyHp(h./1(x»Y = [D.H.(xW, (1.9.11)
To illustrate the ideas behind the proof of Theorem 1.9.1 we will assume that D and 1f',(D) both lie in the same chart. Since If', is a diffeomorphism for each t, we can regard it as a change of coordinates in phase space. With notation in Figure 1.34,
n(t)= f dq'"" .• dp~.
9',(D)
(1.9.15)
by (1.9.9) and (1.9.6), respectively. Finally, [D.h.p(xWSD.h.p(x)XH.(x) = SXH.(x),
(1.9.12)
Since x' = If',(x), this can be written as
net) = L Det(DIf',(x)W·x,
(1.9.16)
by (1.9.8).
46
J Diffeomorphisms and flows
1.9 Hamiltonian flows and Poincare maps
47
tp,(x) = x + tX(x) + 0(t2),
The above arguments do not depend on the initial time being zero and (1.9.22) can be generalised to
whered2·x=dq!> ... ,dpn' Now,
(1.9.17)
Dtp,(x) = I + IDX(x) + 0(12).
(1.9.18)
0(1) = f div X(xW·x. (1.9.23)
DI')
Clearly, if div X(x) == 0 then Ott) == 0 and Q(I) = Q(O) for all t.
Let us apply Theorem 1.9.1 to a Hamiltonian flow tp~. The vector field X is given by (1.9.1) and
and therefore
Thus
dQI . . f (Det(Dtp,(X»  l)d2'
 =Q(O) = LIm x,
dr '=0 ,0 D t
since Q(O) = JD dq), ... , dp •. However, observe that if DX(x) has eigenvalues '1j(x) then
(1.9.19)
(1.9.22)
divX(x)= t !_ (aH)_!_(aH) ==0.
j=) aql apj apj aqj
Hence tp~ preserves phase space volumes. This result highlights, in a geometrical way, the very special nature of Hamiltonian flows. In general, even dimensional flows may expand volumes in some parts of phase space and contract them in others. Clearly, (1.9.24) imposes a global restriction on tp~. The volumepreserving nature of Hamiltonian flows is also reflected in the nature of the transformations that relate them to one another. It can be shown (see Arnold, 1968, p. 222 and Exercise 1.9.5) that (1.9.14) implies Det(Dh(x» == 1 so that symplectic transformations preserve volumes of phase space. However, it is perhaps worth noting that Det(Dh(x)) == 1 only implies h is symplectic when h: il2 t R2 (see Exercises 1.9.5 and 1.9.6).
It is reasonable to consider to what extent Hamilton's equations can be simplified by symplectic transformations. Let h: (q, p) + (Q, P) and ii(Q, P) = H(h)(Q, P». In particular, the transformed equations will be simpler if the new Hamiltonian is independent of one of the generalised coordinates. For example, suppose ii does not depend on Q., then
(1.9.24)
Det(Dtp,(x» = Det(I + IDX(x) + 0(t2)), = n (1 + t'1j(x) + 0(t2»,
j
= I + t L '1j(x) + 0(t2),
j
= 1 + t Tr DX(x) + 0(12).
(1.9.20)
Of course,
Tr DX(x) = div X(x) and substitution in (1.9.19) gives
0(0) = L div X(xW"x.
(1.9.21)
. aii
P ==0
n aQ.
(1.9.25)
Figure 1.34 The now map !p, takes D at time zero (see (a» to fI',(D) at time t (see (b». Since !p, is a diffeomorphism, this transformation can be regarded as a change of coordinates from (ql' ... , q., PI' ... , P.) = x T to (qj, ... , q~, p'(> ... , p~) = X'T.
and
Pn(t) = p.(O) = In
(1.9.26)
p p'
0 "
D
q
(a) (b) The constant value I. can be thought of as a parameter. For a given value of I., H now depends on only (n  I) pairs of conjugate variables; the number of degrees of freedom has been reduced by one and the order of Hamilton's equations has decreased by two.
Ideally, one would like ii to be independent of all Qj, i = I, ... , n. Then
Pj(t) = Pj(O) = I,
(1.9.27)
and
(1.9.28)
r
48
1 Dijfoomorphisms and .flows
1.9 Hamiltonian flows and Poincare maps
49
i = I, ... , n. Notice Qi depends only on the parameters I" ... , I. and is therefore independent of t. Thus (1.9.28) can be trivially integrated to give
and (1.9.30) has solutions of the form
( 1.9.29)
Pi = wjAj sin(wit + tli),
(1.9.32)
tljE R, j = l, 2. The aim is to construct the Poincare map in such a way that one pair of conjugate variables (q2, P2, say) are removed. Thus we argue that by restricting to the Hamiltonian shell H(q, p) = ho > 0 we can express P2 in terms of q" p, and Q2' Since Q2 is periodic with period 21t/W2' the orbit of a phase point in the plane q2 = 0 returns to q2 = 0 after time 27[/w2 (see Figure 1.35). Therefore, the Poincare map P defined on the section q2 = 0 is given by
j = I, ... , n. Kj E IR. Systems for which such a reduction is possible are said to be integrable and the system defined by (1.9.27 and 28) is referred to as their normal form. The variables (Q, P) displaying this form are called actionangle variables; the Pi (or 1j) being the 'actions' and the Qi being the 'angles' (or cyclic variables). The latter name arises because (1.9.27 and 28) is the polar form of a simple harmonic oscillator with radial coordinate I, and angular coordinate Qj.
Traditional courses in Classical Mechanics focus attention on the integrable case. For example, IF systems with analytic H, linear equations of motion (i.e. normal modes), nonlinear systems that are separable into IF systems are commonly discussed. However, these systems are not typical. In general, Hamiltonian systems are nonintegrable and they can exhibit much more exotic dynamics. To illustrate this we must consider systems with at least two degrees of freedom and Poincare maps playa key role in making such problems manageable.
A system with twodegrees of freedom has a fourdimensional phase space and it is, therefore, not feasible to picture its flow directly. Since the system is conservative, (generically) its trajectories lie in threedimensional submanifolds or 'shells' on which the Hamiltonian H(q, p) is constant. Thus, by choosing a particular value for H(q, p) we can reduce the dimensionality of the problem by one. Now, we are frequently interested in systems exhibiting some kind of recurrence. For example, nonintegrable perturbations of an integrable system or the behaviour of a nonintegrable system in the neighbourhood of a closed orbit. In such cases, we can reduce our problem to one in two dimensions by constructing an appropriate Poincare map. Of course, we have lost some detail of the dynamics in this process. After all we are only sampling the orbit periodically. However. the interesting point is that sufficient information is retained to show that the dynamics of 2F, conservative systems can be very complicated. Moreover, since this information is in twodimensions it is quite easy to present and appreciate in graphical form.
To show how the Poincare map is constructed, let us first examine an integrable case, where solutions can be written down explicitly. Consider the biharmonic oscillator
=( cos 2.7[/1 + ~, sin 27[Jl) (pq,),
(1.9.33)
 w, sin 21tJl cos 21tJl '
with Jl = wt/w2• Clearly, P represents a rotation for which the ellipses
pf + wfq: = C, (1.9.34)
with 0 < C < 2ho, are invariant curves. These closed invariant curves correspond to invariant tori in the flow on the H = ho shell.
The important thing to notice about (1.9.33) is that Det P = 1. This means (see Exercise 1.9.5) that the Poincare map, constructed in the manner described above, is areapreserving. That this is also the case when the system is nonintegrable follows from the PoincareCanan invariant (Arnold, 1968, pp. 23340 or Arnold & Avez, 1968, pp. 2302). A derivation of this invariant for 2F systems requires
Figure 1.35 The Poincare map defined on the section q2 = O. It is clear from (1.9.32) that q2 returns to zero periodically with period 21[/W2'
q, =p" q2 = P2,
(1.9.30)
The Hamiltonian H(q, p) is given by
2
H(q"q2;P"P2)=! L (pf+wfqf)
i= 1
(1.9.31)
ql =0
r
1.9 Hamiltonian flows and Poincare maps
51
50
1 Diffeomorphisms and flows
a knowledge of differential forms, however, for IF systems it can be obtained in the familiar notation of vector analysis. Consider the extended phase space for a 1F system with coordinates (q,p,t). Let v=(p,O,H), then curi'l= (iJH/iJp, iJH/iJq, 1) is the vector field of the Hamiltonian H in extended phase space (see (1.9.1». Now apply Stokes Theorem to the tubular region shown in Figure 1.36(a). Here the sides of the tube consist of flow lines of curl 'I. Dissecting
the tube as shown in Figure 1.36(b), we observe that
f curlv·dS=O= f vdr> f vdr,
s y, Y2
where dr = (dq, dp, dt). Thus
f p dq  H dt = f p dq  H dt
y, Y2
(1.9.36)
(1.9.35)
and it follows that Jy p dq  H dt is invariant under the flow.
With the aid of differential two forms (see Arnold, 1968, pp. 2346), we can obtain Stokes' Theorem in five dimensions and derive the corresponding result for 2F systems; namely
f PI dql + P2 dq2  H dt = f PI dql + P2 dq2  H dt,
11 '/1
Figure 1.36 (a) Tubular region to which Stoke's Theorem is applied for 1_ F systems. The vector field curl v is tangent to the surface at every point of the tube so that curl ,,·dS:= O. The closed curvesv, and)'2 are obtained by taking sections transverse to the tube of flow flines. (b) Dissection of the tube shown in (a) used to obtain (1.9.35).
q
(a)
(1.9.37)
Figure 1.38 Some typical orbits of the Henon map (1.9.40) for cos IX = 0.8. Two fixed points can be seen: one elliptic (see § 6.5) and one saddlelike. What appear to be closed curves are each the orbit of a single point, i.e. the orbit is confined to what is topologically an invariant circle. For small numbers of iterations of (1.9.40) individual points of these orbits can be distinguished moving around the origin (cr. Exercise 1.9.9). As the number of iterations increases, the plotted points merge into what looks like a closed curve. Individual orbit points are more apparent in the vicinity of the saddle point. (After Henon, 1969.)
1.0 r,,,..
(b)
Figure 1.37 If)'2 is given by r = r(u), then r(u) = (ql(u), 0, PI(U), P2(U), t(u)), where P2(U) is determined by H(q, p) = ho. The curve, 12' obtained by projecting )'2 onto t = 0 (see (a», is the image of }II under the Poincare map P (see (b».
PI
(a)
0.5
y 0
0.5
1.0 ' __ L 1 ' __ ___J
 1.0 0.5 0 0.5 1.0
(b)
x
52
1 DijJeomorphisms and flows
r
53
1.9 Hamiltonian flows and Poincare maps
where 1'1 and 1'2 are closed curves bounding a tube ofthe flow in the fivedimensional extended phase space with coordinates (ql' qz, PI' pz, t). Now let YI consist entirely of points such that H = ho, qz = t = O. Suppose we follow the lines of the flow fI'~ until we return to qz = O. Although H remains at ho and qz returns to zero, the
y 0
0.5
time required to reach qz = 0 will, in general, be different for different points of 1'1' Thus, in extended phase space, points of the image, I'z, of 1'1 do not all have the same t coordinate. Let us put 1'1 and I'z defined in this way into (1.9.37). Since Hand qz are constant on both curves, we have
Figure 1.39 A selection of plotted orbits of (1.9.40) for cos rx = 0.4. Orbits of points near to the saddle point become highly irregular. Successive iterates still move around the fixed point at (0,0) but they are no longer confined to a closed curve. Instead they appear to spread over a twodimensional region in an erratic manner. Eventually, they are pulled away along the unstable manifold of the saddle and, left to themselves, will cause an overflow error in the computer doing the plotting. On the other hand, orbits of points near the origin still appear to be confined to invariant circles. Between these extremes, a new feature called an island chain can be seen. The 'islands' themselves are formed around the points of an elliptic periodic orbit, here of period six. The 'straits' between successive islands contain a hyperbolic periodic orbit also of period six. The orbits of points near to the elliptic periodic points move from island to island, returning to an invariant circle surrounding the initial elliptic point at every sixth iteration. Some information to help the reader to observe island chains is given in Exercise 1.9.9. (After Henon, 1969.)
(1.9.38)
for i = 1, 2, so that (1.9.37) becomes
(1.9.39)
where 1z is the projection of I'z onto t = O. Now 1z is the image of 1'1 under the
Figure 1.40 Analogous plots to those shown in Figure 1.39 but with cos rx = 0.24. Observe that a fivefold island chain is the dominant feature here. In fact, (see § 6.5) island chains of all periods occur but only a few are easily visible. The orbits looking like separatrices of the hyperbolic periodic points are deceptive (see Figure 1.41). (After Henon, 1969.)
I.o.,.,r.....,
1.0 ryyr~
".:= I.:
0.5
0.5
0.5
y 0
1.0 ' ..L ..1.. t' _._J
1.0
1.0 L._'~_::':7'
_ 1.0 0.5 0 0.5 1.0
0.5
o
0.5
1.0
x
x
54
1 Diffeomorphisms and flows
1.9 Hamiltonian flows and Poincare maps
55
XI+ I = XI cos IX  YI sin IX + x~ sin IX, YI + I = XI sin IX + YI cos IX  x~ cos IX,
as the Poincare map of a Hamiltonian system. Instead it represents the most general quadratic planar map that is areapreserving and has a pure rotation for its linear part. Some striking features of the dynamics of (1.9.40) are illustrated in Figures 1.381.42 but the reader cannot do better than to consult Herron's excellent review (1983) for more details. Figures 1.381.42 show invariant circles, islands chains, chaotic orbits and their repetition on all scales. All this leads to a picture of immense complexity that is by no means fully understood. We will return to such matters in Chapters 3 and 6.
Poincare map P (see Figure 1.37). Hence P is an areapreserving map on the section H(q, p) = ho and q2 = 0 in the phase space of the system.
Numerical experiment has shown that Poincare maps constructed in the manner described above exhibit complicated dynamics (Henon, 1983, pp. 8495; Lichtenberg & Lieberman, 1982). This complexity is a feature of areapreserving maps of the plane and it is typified by the quadratic mapping of Henon: namely
(1.9.40)
where IX is a real parameter and t E 1. (see Henon, 1969). This map is not constructed
Figure 1.41 Two orbits of (1.9.40) for cos IX = 0.22. The first is the orbit of a point near an island centre giving invariant circles around the five elliptic periodic points. In the present context, it serves only to indicate the position of the islands. The remaining points are all generated by iterating a single initial point. Once again, the iterates spread out, in a stochastic manner, over a twodimensional region in the neighbourhood of what appeared to be separatrices in Figure lAO. (After Henon, 1969.)
Figure 1.42 The result of magnifying a detail of Figure 1040 containing one of the hyperbolic periodic points. A twodimensional orbit like that shown in Figure 1041 is apparent. However, not only are more island chains visible around the fixed point (0,0), but also analogous islands can be seen around the adjacent elliptic periodic points. As we shall see, if these islands were again magnified, then we should find more twodimensional orbits and more island chains and so on. Thus the complexity of the map is repeated on all scales. (After Henon, 1969.)
1.0 .1..1..
0.5
y 0
0.5 r
0.200 ~"""~~~ .. .;:,~_r_o(...".·~.,..., ,,'r:~~ .jtL..,,;~"'if...,.~.~;.,.' ':~"'r' ....... ...,
.
.. 1 ." ,~~ ,.,. ...  ". ._  '" I,
i> V~:,./ _. ,
QIOO ~~~~~ __ _L __ ~ ~ ~ ~
0.175
, ,
y
0.150
0.125
.~ .
1.0 L ___.1 l.1 .1..1 ___'
 1.0 0.5 0 0.5
1.0
0.525
0.550
0.575
0.600
x
x
56
1 Diffeomorphisms and flows
r
,
1.1.1
Exercises
1.1 Introduction
Let M be the unit circle in the complex plane. Explain how the map n: R > SI given by xHexp(ix) can be used to define a set of charts on the circle SI. Define explicitly two charts which form an atlas on SI. What is the differentiability of the overlap maps for this atlas?
Let M be a (,,manifold. Show that iff: M > M is a C'map, k,:;; r, and the points Xo and f(xo) are in overlapping charts (U., h.), (U 7' hy) and (Up, hpj, (U" h,) respectively, then the map
1.1.2
(EI.I)
is ct at h.(xo) if and only if
1.1.3
h,.f.h; 1 (EU)
is C' at hy!xo). What does this imply about the differentiability of r at xo?
(a) Find an atlas of the torus T2 = {(x mod I, y mod 1)l(x, Y)ER2} containing four charts by using the local diffeomorphism 11: H2 > T2 given by (x, y)H (x mod I, ymod I).
(b) Use stereographic projection on the unit sphere to obtain an atlas consisting of two charts. Construct the overlap map between these two charts.
1.2.1
1.2 Elementary dynamics of diffeomorphisms 1.2.2 Diffeomorphisms of the circle
Show that the set of points {mIX mod limE zj, IXE R\O, is dense in the interval [0, I] by showing that:
(i) mlX#m'lXmod I ifm#m';
(ii) there exist two points 8 + mIX, 8 + m'lX in any given interval of length 2n/k on the unit circle (consider k + I points of the form mIX, mE I);
(iii) consecutive points of (m  m'p, 2(m  m'p, 3(m  m'p, ... are less than 2n/k
apart;
(iv) any sneighbourhood contains points of the sequence in (iii).
Which of the following maps g: [0, I] > R can be used to construct a lift 1: II > R of an orientationpreserving homeomorphism f of the circle SI.
(a) g(x)=x2,
(b) g(z) = x2  2x, (EU)
(c) g(x) = 2X1  x.
Describe the lift J in each case.
Consider the homeomorphism f: SI > SI, SI = {exp(2nix)IO':;; x < I}, given by exp(2nix)1+ exp(  2nix) (reflection in the xaxis), What are t~e fixed points a~d period2 points of f? Check your answers by finding a lift f: R > II of f With respect to the covering map n: xHexp(2nix) of SI by R and then investigating the intersection Of the graph of J with the lines y = x + n, n E I. Carry out the same investigation for J2.
Find the periodic points of period 2 of the circle diffeomorphism f: SI > SI with lift J(x) = x + 0.5 + 0.1 sin 2nx. Find the stability of the period2 periodic orbits. Prove that there are neither fixed points nor periodic orbits of period 3?
1.2.2
1.2.3
1.2.4
Exercises
57
1.2.5
Consider an orientationreversing homeomorphism f: SI > SI. Show that any lift 1: R > R of f will have the property that J is strictly decreasing and J(x + I) = fix)  I. Show that f always has two fixed points. Is P orientationreversing or preserving?
1.3.1
1.3 Flows and differential equations
Use (1.3.2) to show that (rp,rl=IP, for each tER. Hence show that it follows from Definition 1.3.1 that 11',: M > M is a diffeomorphism for every real t.
Let II' be a now on the manifold M and suppose that the orbits {1P,(xo)} and {1P,(xl)} intersect. Prove that the orbits coincide.
Let y be a closed orbit of the now II' on the manifold M and suppose there exists T> 0 and Xo E l' such that IPT(XO) = Xo. Prove that qlT(X) = x, for every x E y.
Locate two closed orbits YI and l'1 and positive periods TI and T1 for the now of
i = r(r  I)(r  2); 9 = r2. (EI.4)
1.3.2
1.3.3
1.3.4
Verify that
lP(t, x) = x exp(t)/[x exp(t)  x + I]
(EI.5)
1.3.5
is a now on [0, I] and find its associated vector field. Why is it not a now on R1 Find the differential equations associated with the following (local) flows: x
(a) 1P,(x) (I _ 2x2t)t on R; (E. 1.6)
(b) qI,(X,y)=(xexp(t),_Y_) on R2. (EI.7)
Iyt
Convince yourself that, for any interval of t of the form (a, a), there is a neighbourhood of origin in which (EI.6) and (E 1.7), respectively, correctly describe the solutions of the differential equations you have obtained.
1.4.1
1.4 Invariant sets
Find the minimal closed invariant sets for (a) irrational and (b) rational rotations of the circle. What is the most general closed invariant set in both cases?
Prove that the fixed and periodic points of a diffeomorphism f on a manifold M lie in its nonwandering set Q. Show that Q is (a) closed and (b) invariant under f.
Prove that the nonwandering set of a now qI on a manifold M is (a) closed and (b) invariant. Show that a closed orbit of II' is a subset ofthe nonwandering set.
Describe the behaviour of the diffeomorphism f(x) = ax, aE R' when
(i) a< I; (ii) a=I;
(iii) 1 < a < 0; (EI.8)
(iv) O<a< I;
(v) a= I;
(vi) 1 <a.
Give L.(x) and L.,(x), x # 0, for each case.
Sketch examples of phase portraits on the plane with a nonempty, compact limit set containing (a) one; (b) two; (c) three; (d) four fixed points. Suggest how differential equations with flows of these types might be constructed.
1.4.2
1.4.3
1.4.4
1.4.5
by:
(i) J(x)=x+J;
(ii) J(X)=Xl+~, O,,;x< I, f(x+ I)=f(x)+ I;
(iii) fix) = x + 1+! sin 2ltx;
2lt
by locating fixed or periodic points of f.
58
r
,
!
J Diffeomorphisms and flows
1.4.6
Use the PoincareBendixson Theorem to show that the Van der Pol oscillator
E<O,
(EI.9)
has at least one stable limit cycle for sufficiently small values of E.
1.5 Conjugacy
1.5. t Show that:
(i) fix) = 2x andy(x) = 8x are topologically conjugate on R but notdifferentiably conjugate;
(ii) f(x)=2x and g(x}= 2x are not topologically conjugate by showing conjugacy preserves orientation of a map;
(iii) fix) = 2x and g(x) =!x are not topologically conjugate by investigating the nature of the fixed point at the origin in the two cases.
1.5.2 Prove that f(I: x 1+ Xl" + 1, II E I'll, is a topological conjugacy of the diffeomorphisms fix) = 2x and g(y) = 220+ 1 yon R. Why is there no differentiable conjugacy when 11>0.
1.5.3 Let f, g: n" + R" be Clconjugate (k ~ I) diffeomorphisms by h: n" .... Wand suppose flO) = O. Prove that the Jacobian matrices of f at 0 and g at h(O} are similar. What does this imply about the eigenvalues of Df(O) and Dg(h(O»?
Show that fix) = ax and g(x) = px are not Clconjugate for a." p. When are f and 9 not C°conjugate?
1.5.4 Let f, 9 be diffeomorphisms on R given by
Exercises
59
(EI.I5)
• 1.6 Equivalence of flows
1.6.1 Show that flows
exp(Jt), J = (~ ~) and exp(Jot), Jo = (~ ~)
are topologically equivalent. Hence prove that all linear flows, exp(At), for which A has pure imaginary eigenvalues are topologically equivalent to exp(Jot). Why are these flows not all topologically conjugate?
1.6.2 It is often easy to recognise why two flows are not topologically equivalent by noting key features in their phase portraits. Describe, for each diagram in Figure El.l, a distinguishing feature preserved by topological equivalence which is not shared by the others. Explain your answers. Why is the invariant circle in (b) not a limit cycle?
Show that the topological types of a saddle and node are different by considering the separatrices of the saddle.
Consider the system :i = X.(x), a. E R,
on A2 and the flow on T2 induced by the map n: A2 .... T2 where nIx, y) = (x mod I, y mod l ]. Show that the phase portraits of the system (EI.I6) for a. (a) rational and (b) irrational are not topologically equivalent.
Let qI and If" be topologically conjugate flows on the manifold M and", and "" be topologically conjugate flows on the manifold N. Prove that the product flow
fIx) =X + sin x
(EUO)
and
g(x) = x + hl(x),
(EUI)
where
1.5.5
Xl x5 l Xll+l
hk(x)=x++.·· +(1) ,
31 51 (2k + 1)1
kEZ+. Prove that f and 9 are not topologically conjugate for any k.
Find the number of period2 points of the diffeomorphisms f and 9 of the circle S 1 , where the lifts are:
(EI.I2)
1.5.6
J(x) = x + 0.5 + O. t sin 21lx; g(x) = x + OJ + 0.1 sin 2ltx.
Show that f and 9 are not topologically conjugate.
Prove that the two linear systems :i = Ax, Y = By, x, Y E R", have flows which are linearly conjugate if and only if the matrices A and B are similar.
Let tp, '" be flows on R" and suppose that tp has a fixed point at the origin. If qI and", are C"conjugate, k » 0, by h: R" + Rn, show that the vector fields X and Y of II' and", are such that the matrices DX(O) and DY(h(O» are similar. Compare your answer with that of Exercise 1.5.6 and comment on why the converse of the above is not true in the nonlinear case.
1.6.5
(EU3)
(EI.I4)
1.6.3
1.6.4
1.5.7
1.5.8
Find the rotation number of the circle homeomorphism .r with lift 1: R .... R given
Figure El.1
(b)
(e)
(d)
x=l,
y=a.
(E1.I6)
60
J DifJeomorphisms and flows
1.6.6
rp x '" is topologically conjugate to rp' x "" on M x N. Show that this result does not extend to topological equivalence of flows by considering flows on M = N = SI.
The topological types of linear flow, exp(At), given in Figure 1.26 can be subdivided into algebraic types. Sketch phase portraits for exp(AI) when:
(a) [Tr Ar > 4 Det A (nodes);
(b) [Tr Af = 4 Det A (improper nodes); and (c) [Tr A]2 < 4 Det A, Tr A¥O (foci).
Modify Figure 1.26 to show these algebraic types on the (Tr A, Det A)plane.
1.1.1
1.7 Poincare maps and suspensions
Show that a map f: R + R of the form f(x) = lXX, IX e R, has a fixed point at X = 0 and that it is stable or unstable according as IIXI < 1 or I!XI > 1 respectively.
Consider the timeone map CPI of the flow of x = x  x2• Deduce that CPI has fixed points at x = 0 and x = I. Obtain the linear approximations to cP I at these points and deduce their stability. Sketch the behaviour of the map CPl' Compare your results with the phase portrait of the flow.
Show that flows with fixed points do not satisfy the requirements for the existence of a global cross section. In what sense can the system f = r  rJ, iJ = 1, be said to have a global section, S={(r,8l1r~0,8=0}, on which the Poincare map determines the dynamics?
Let y be a periodic orbit of a flow rp. Suppose S 1> S2 are distinct local sections for y (see Definition 1.5.2) at Xt>x2ey such that S2=rpr.(SI)' Let PI,P2 be the corresponding Poincare maps on S I and S2' Prove that for suitable neighbourhoods of XI and X2, there is a CIconjugacy between maps PI and P2.
Suspend the diffeomorphism f: [0,1] + [0, 1] described by fIx) = !(x + x2) to obtain a flow on a surface. What is the surface? Describe the behaviour of the flow.
Draw diagrams to illustrate the suspended flows of the diffeomorphisms
(a) f: 1+ 1,1=[ 1,1], x ix+ !xJ,
(b) f: SI+Si, exp(2nix) exp(2nix).
1.7.2
1.7.3
1.7.4
1.7.5
(EI.I7)
1.7.6
Show that the flow rp,(x, y) = «x + I) mod 1, y + !Xt) on the cylinder (x mod I, y)l(x, y)e R2} is the suspension of the diffeomorphism f: R .... R given by fry) = y +IX.
1.8.1
1.8 Periodic nonautonomous systems
Let x = XIx, r), (x, t)e R" x R, be a periodic differential equation with XIx, t) = XIx, t + T) for some T> O. Show that the transformations I' = (2n/T)1 and X'(x, I') = (T/2n)X(x, t) give a 2nperiodic system dx/dt' = X'(x, I').
Prove that iF X = W) is a solution of x = XIx, t), with XIx, t + T) = XIx, t) and (x, t)eR" x R, then x = W + T) is also a solution. Show that such systems can have nonperiodic solutions by considering x = (1 + sin I)X with (x, l)eR x R.
(a) Show that the set of solutions ofx = A(t)x, (x, t)e If' x R, form a vector space. (b) Let ~i(I), i = I, ... , n + I be a set of solutions. Show that there exist
1.8.2
1.8.3
Exercises
61
"+1
lXI' ... , !X .. 1> not all zero, such that L !Xi~i(O) = O. Use uniqueness of solution 1=1
1.8.4
"+1
to show that L !Xi~i(I).= O.
1=1
(c) Show that the vector space of solutions is ndimensional.
Define the state transition matrix rp(t, to) of the system x = A(t)x, (x, t)e JR" x R, A(t + T) = A(t). Prove that
(i) rp(1 + T, 10 + T) = rp(t, (0),
(ii) fjI(t, Or 1= rp(O, t). (EI.I8)
Use these results to show that
rp(1 + T, t) = rp(O, tr IfjI(T, O)rp(O, t).
(E1.19)
What does this imply about the family of Poincare maps PD.: :E9• > :Eoo' 80 e [0, 2n] of the system
x=A(8)x,
(EI.20)
1.8.5
where A(8 + 2n) = A(8), and 1:,_ = {(x, 80)lxeR2}?
The state transition matrix is useful where A is independent of I or aperiodic.
(a) Verify that ~I(I) = (eXP(AII)) and ~2(1) = ( 0 ) are linearly independent
o exp(A21)
I· f' (AI 0)
so uuons 0 x = Ax, A = 0 A2' Hence find rp(l, (0) for this system.
(b) Verify that ~I(t)=( 2 ) and ~l(I)=(eXp(I») are a basis for the solution
exp(l) 1
space ofi = A(I)x with A(I) = (1 2 exp( I)). Hence construct rp(l, 0)
exptr) 1
and find x(t) given that x(O) = G)'
Let rp(t, to) be the state transition matrix of the system x = A(I)x and define WIt) = Det(rp(t,O». Prove
(i) rp(t + h, 0) = rp(t, 0) + hA(t)rp(t, 0) + O(h2); (E1.21)
(ii) WIt + h) = Det(1 + hA(I) + O(h2))W(I)
1.8.6
"
= n (1 + hAi(l) + O(h2))W(I),
i=1
(EI.22)
1.8.7
where Ai, i = I, ... , n are the eigenvalues of A; (iii) WIt) = Tr(A(I))W(I).
Find solutions for the system
(EI.23)
(~I) = (c~s I +sin I)(XI)
X2 sin I cos I X2
(E1.24)
1.8.8
and hence, or otherwise, obtain the period advance map rp(2n, 0) = P. Calculate the eigenvalues of P and determine the stability of the null solution of (E1.24).
If rp(t, (0) is the state transition matrix of the system
i=A(I)X,
(E1.25)
62
I Diffeomorphisms and flows
show that
x = A(t)x + B(t)
(Et.26)
has solution
x(t) = tp(t, to)( Xo + L tp(to, r)B(r) dr)
(E1.27)
when x = Xo at t = to.
Find the solution of
i=(O 1)x+(sint)
1 0 cost
(EI.28)
when x(O) = Xo'
1.9.1
1.9 Hamiltonian flows and Poincare maps Find the phase portraits of the flows in RZ with Hamiltonian:
(a) H(xhxZ)=xi +x]: (b) H(x., xz) = x~ + x~;
(c) H(X.,xz)=xi+x1xi.
Sketch the phase portraits for Hamilton's equations when H(x., Xz)= xi + x~ + /lX2 and
(i) 11<0; (ii) 11=0; (iii) 11>0.
(E1.29)
1.9.2
Prove that the fixed point of a planar flow with a quadratic Hamiltonian is generically either a centre or a hyperbolic saddle.
Show that, for II> 0, the Hamiltonian vector field given by H = _ W2 + r4 + r5 cos 50, where (r,O) are plane polar coordinates, has 11 fixed points
consisting of 6 centres and 5 saddles. Show that the separatrices of the saddles form a chain of 5 islands around the origin such that each island contains a centre.
1.9.5 Prove that a change of variables on R2 is symplectic if and only if it is orientation and areapreserving. Show that:
(i) the change from Cartesian to plane polar coordinates, (x, Y)1+ (r, 0), is not
symplectic;
(ii) the transformation (x, Y)1+ (r, 0), where 't = r2/2, is symplectic.
Illustrate the fact that symplectic transformations preserve the form of Hamilton's equations by applying the transformations in (i) and (ii) to the system.
x = Y  y(XZ + l); y = x + x(XZ + l). (EI.30)
1.9.3
1.9.4
1.9.6
Let he L(R4) be given by h(x) = Px with
P=(~ :).
where A, S, C, Dare 2 x 2 matrices. Show that h is symplectic if and only if
DTABTC=I, DTB=UTD and CTA = ATe, (E1.32)
(EI.31)
where I is the 2 x 2 unit matrix. Hence construct a counterexample to show that Det(Dh(x) == 1 does not imply h is symplectic.
Exercises
63
1.9.7
Show that the VolterraLotka equations x=(aby)x,
(EU3)
y= (c Jx)y
do not form a Hamiltonian system. Show that the change of variable x = exp(q), Y = exp(p) allows (E 1.33) to be written as a Hamiltonian system with Hamiltonian
H(q, p) = up  b exp(p) + cq  J exp(q).
(E1.34)
1.9.8
Why can it be concluded (without calculation) that the transformation (q, p)1+ (x, y) is not symplectic?
(a) Find actionangle variables for the Hamiltonian system
x=wy,
y= wx
(E1.35)
by using 1= XZ + yZ.
(b) Consider the pendulum equations
x=y,
y=sinx
(E1.36)
and show that the transformation '1': (x, Y)1+ (1, 0), where 1= (yZ/2)  cos x and 0 is the polar angle, has Jacobian 1 + 0(/). Show that the transformed vector field is
9= 1 + 0(111).
(E1.37)
1.9.9
In view of Exercise 1.9.5 what do these results imply about 'I'?
Write a computer program to plot orbits of the Henon areapreserving map (1.9.40).
(a) For. cos IX = 0.8, plot several orbits with initial points (x,O) for x e (0, !].
Compare your results with Figure 1.38.
(b) For cos a =0.4, plot orbits with initial points (x,O) for xe(t, 1]. Observe a sixfold island chain and chaotic orbits.
(E2.42) (E2.43)
3
118
2 Local properties of flows and diJfeomorphisms
2.8.6 Use blowingup techniques to determine the topological types of the singularities at the origin in the following systems:
(a) i=y+x3, (b) X=X2_y2,
y=x3; y= 2xy.
Structural stability, hyperbolicity, and homoclinic points
Explain why the technique fails to determine the topological type of the singularity at [x, y) = 0 in the system x = y+ x3, Y = _x3•
In applications we require our mathematical models to be robust. By this we mean that their qualitative properties should not change significantly when the model is subjected to small, allowable perturbations. If we wish to make these ideas more precise, then we must have some class of perturbations in mind and some way of deciding when they are small. From a theoretical stand point, this means that we regard our model as a member of some chosen space, g, of dynamical systems to which we attach an appropriate metric. It is then possible to give meaning to the idea that a perturbation is 'close' to the original model. A dynamical system whose topological properties are shared (in a sense that must be properly defined) by all sufficiently close neighbouring systems is said to be structurally stable.
Structural stability, like hyperbolicity (see §§2.l and 2.2), is a property of individual dynamical systems and we can ask if this property is, in some sense, typical of the elements of the space f/. The subset of all structurally stable systems is open. This follows directly from the definition of structural stability itself. Clearly, every structurally stable system lies in an open set, each element of which is also structurally stable. Thus the subset of structurally stable systems is a union of open sets and is therefore open itself. It follows that a structurally stable system cannot be approximated arbitrarily closely by structurally unstable systems. However, in some cases the subset of structurally stable systems can be shown to be dense in the space f/. This means that every structurally unstable system can be approximated arbitrarily closely by structurally stable systems. In such cases, we say that structural stability is a generic property of f/. In general, a property is said to be generic if it is shared by a residual subset of the space of systems involved. Such a subset is a countable intersection of open dense sets. We will, however, encounter only open dense sets in this chapter.
In §3.1, we show that a linear flow is structurally stable if and only if it is hyperbolic and that hyperbolicity of flow is a generic property of linear transformations on R". Thus the structurally stable linear flows are characterised by a single, hyperbolic fixed point at the origin and structural stability of flow is a generic property of linear transformations.
120
3 Structural stability, hyperbolicity and homoclinic points
3./ Structural stability of linear systems
121
Studies of flows on twodimensional compact manifolds (see § 3.3) show that the structurally stable systems have nonwandering sets which are characterised by hyperbolic fixed points and closed orbits, together with the global requirement that no saddle connections occur. Structural stability is, once again, a generic property of these flows.
The above findings led to two conjectures about flows on manifolds of dimension n~2:
Definition 3.1.1 A linear flow, exp(At): Rn 4 Rn, (or diffeomorphism, A) is said to be structurally stable in L(Rn) if there is an eneighbourhoocl of A, N,(A) £ L(Rn), such that, for every BE Nt(A), exp(Bt) (or B) is topologically equivalent (conjugate) to exp(At) (or A).
The following result shows that, in this linear case, the structurally stable systems can be completely characterised.
(i) (ii)
structural stability is a generic property of such flows;
the structurally stable flows are characterised in the same way as flows on twomanifolds.
Neither of these conjectures is correct. A counterexample to the first was given by Smale (1%6). We do not discuss this example here; the interested reader can consult Arnold & Avez, 1968, pp. 19fr.200. In §§3.4 and 3.5 we describe two types of diffeomorphism on twomanifolds that are structurally stable but cannot be characterised in the manner described in (ii). Remember (see § 1.7) these diffeomorphisms correspond to flows in three dimensions by suspension. The Anosov automorphisms of the twotorus (see § 3.4) and the horseshoe diffeomorphism of the sphere (see §3.5) have complicated nonwandering sets and to characterise them we must extend our notion of hyperbolic sets beyond fixed points and closed orbits (see §3.6). The systems referred to in (ii) are now known as MorseSmale systems (see Niteeki, 1971).
The horseshoe diffeomorphism also plays a central role in our understanding of the complex dynamics described in §1.9. In particular, it can be shown that it has orbits that behave in a random or chaotic way (see §§ 3.5.2 and 3). In §§ 3.6 and 3.7 we explain how this behaviour is related to the 'twodimensional, chaotic orbits' associated with hyperbolic fixed or periodic points of some planar maps (see Figures 1.3942). The occurrence of homoclinic (or heteroclinic) points is the central feature of this discussion. They arise when the stable and unstable manifolds of a hyperbolic point (or points) intersect transversely and it can be shown that the map must then contain embedded horseshoes.
Proposition 3.1.1 A linear flow or diffeomorphism on Rn is structurally stable in L(Rn) if and only if it is hyperbolic.
Proof. A linear flow exp(At) is hyperbolic if all the eigenvalues of the matrix A have nonzero real parts (see Definition 2.1.2). The eigenvalues of any sclose matrix B differfrom those of A by terms 0(£) (see Exercise 3.1.1). Thus, by making £ sufficiently small we can ensure that the eigenvalues of B are near enough to those of A for their real parts to be nonzero. Moreover, A and B will then have the same number, n,(n.), of eigenvalues with negative (positive) real parts. Theorem 2.1.2 then implies that exp(At) and exp(Bt) are both equivalent to the flow of i =  x, y = y, where x ERn, and y E Rnu. Thus A is structurally stable.
Conversely, suppose the flow exp(At) is not hyperbolic. Then A has at least one eigenvalue with zero real part. However, B = A + EI is hyperbolic for almost all e I 0 and can be made arbitrarily close to A by taking s sufficiently small. Thus, the nonhyperbolic flow exp(At) is not structurally stable. Hence, if a linear flow is structurally stable it must be hyperbolic (see Figure 3.1).
The proof of Proposition 3.1.1 for diffeomorphisms follows similar lines; it is
considered in Exercise 3.1.2. 0
3.1 Structural stability of linear systems
Let L(R") be the set of real linear transformations of R" to itself. Define the norm
"
of an n x n matrix A ='[aiJ] to be IIAII = L la,jl. An eneighbourhood of A is given
I.J=!
by Nt(A) = {BEL(RO)IIIBAll c s}. Each BEN.(A)is said to be eclose to A. We are now able to give a formal definition of structural stability for linear flows and diffeomorphisms on n°.
It is important to note that Definition 3.1.1 specifies a space of systems (L(Rn» to which the perturbations must belong. Whether or not a given system is structurally stable depends on the choice of this space. For example, let CL(R2) £ L(R2) be the subspace of linear transformations with pure imaginary, nonzero eigenvalues. If A E CL(R2), it is structurally stable in CL(R2) but structurally unstable in L(R2). Clearly, if BE CL(R2) and is eclose to A, then B has pure imaginary eigenvalues that are numerically close to those of A. Thus, the flows exp(At) and exp(Bt) are both of centre type and therefore topologically equivalent (see Exercise 1.6.1). Hence A is structurally stable in CL(Rl). Of course, A E CL(R2) is not structurally stable in L(R2) (see Figure 3.2). This follows from Proposition 3.1.1 because A is not hyperbolic.
The relationship between hyperbolic and structurally stable flows in L(Rn) allows us to show that 'structural stability of flow' is a generic property of linear transformations. Let SF(R") £ L(R") denote the set oflinear transformations which give rise to structurally stable flows on R".
(b)
122
3 Structural stability, hyperbolicity and homoclinic points
3.2 Local structural stability
123
Proposition 3.1.2 The set SF(IR") is open and dense in L(R").
Proof. As we have already noted in the opening remarks to this chapter, SF(IR") must be open. More precisely, by Definition 3.1.1, each AESF(IR") lies in an eneighbourhood, N.(A), of topologically equivalent flows. Since N.(A) is open, each of its elements must also be structurally stable, i.e. N.(A) s;; SF(IR"). Hence SF(R") is open because it is a union of open sets.
It follows from Proposition 3.1.1 that SF(R") = HF(R")(s;; L(R")) the set oflinear transformations which give rise to hyperbolic flows on R". However, HF(R") is a dense subset of L(IR"). To show this, suppose A ¢ HF(R"). Then there exist linear transformations, B, arbitrarily close to A for which exp(Bt) is hyperbolic (e.g.
Figure 3.1 Some examples of phase portraits of structurally stable «a), (b» and structurally unstable «e), (d» linear flows on R3: (a) .(I>.(z = «±iP,« > 0, .(3 < 0; (b).(10 .(Z,.(3 < 0; (e).(, = 0, .(Z'.(3 < 0; (d) .(, > 0, ).Z,).3 = tip.
(a)
(e)
B = A + sl, with e small). Thus, every element of L(R") is arbitrarily close to points of H F(R") and H F(IR") = SF(R") is a dense subset of L(R"). Hence, structural stability of flow is a generic property of L(IR"). 0
The analogous result for hyperbolic, linear ditTeomorphisms is dealt with in Exercise 3.1.2.
3.2 Local structural stability (Hirsch & Smale, 1974, Chapter 16)
An important outcome of the discussion in § 3.1 is that hyperbolic fixed points of linear flows and ditTeomorphisms persist under sufficiently small linear perturbations. This result can be extended to the local flows or ditTeomorphisms defined in the neighbourhood of hyperbolic fixed points of nonlinear systems. Of course, we must choose an appropriate class of perturbations and specify when they are small.
Let U be an open subset of R" and Vec'(U) be the set of all CIvector fields defined on U. The magnitude of each vector field XEVecl(U) is taken to be its Clnorm, IIXIII, where
IIXIII = sup{ t IXi(x)1 + t IOXi(X)I}·
xeV i= I il= lOX)
(3.2.1)
(b)
Thus, ifX(x) = (XI(x), ... , X"(xW then IIXIII is 'small' when Xi(X) and oXi(x)/iJx), i, j = 1, ... , n, are 'small' for all x E U. We now define
N.(X) = {Y EVecl(U)IIIX  YIII < s}
to be an eneighbourhood of X in Vecl(U). The use of the CInorm is often emphasised by saying that Y E N.(X) is eCrclose to (or is an eCIperturbation of) X. This highlights the fact that the components of X and Y and their first derivatives are close throughout U.
We can now make our opening remarks more precise.
Figure 3.2 The flow in (a) is an allowed perturbation of the centre, (b), in L(Rz) but it is not an element of CL(Rz). For a sufficiently weak spiral, (a) can be made arbitrarily close to (b) in L(Rz).
, . , .
X2,'
,
r~+~bfq;T~+xl
(a)
(d)
(3.2.2)
3.3 Flows on twodimensional manifolds
125
124
3 Structural stability, hyperbolicity and homoclinic points
Proposition 3.2.1 Let XEVecl(U) have a hyperbolic singularity at x = x*. Then there exists a neighbourhood Vof x* in U and a neighbourhood N of X in Vec'(U) such that each YEN has a unique hyperbolic singular point y*E V. Moreover, the linearised flow, exp(DY(y*)t) has stable and unstable eigenspaces of the same dimension as exp(DX(x*)t).
small C'perturbation of X gives rise to a flow with a hyperbolic closed orbit of the same topological type as y.
fIJ,l U~.  exp(DX(x*)t)  exp(DY(y*)t)  "I U~"
3.3 Flows on twodimensional manifolds
The characterisation of structurally stable flows on twomanifolds provides a good illustration of the technical complications that can arise when we try to extend the local discussion of § 3.2 to include global phenomena. Recall that the results obtained in §3.2 depended on the existence of sufficiently small neighbourhoods of the unperturbed and perturbed fixed points on which topological equivalence could be established. These neighbourhoods were sufficiently small subsets of the open set U on which X was defined and therefore we were not involved with the behaviour of X on the boundary, au, of U. Indeed, as far as the local discussion was concerned we had not considered whether or not X had a natural extension to au. Moreover, since we were involved with perturbations differing from X only on sufficiently small neighbourhoods of the singular point x*, the question of whether IIX  YII, was finite for all X, Y in Vec'(U) was not relevant. Complications of this kind cannot be ignored if precise statements are to be made about global structural stability and they give rise to a number of technical conditions in the resulting theorems.
Let us begin by considering vector fields on R2. We can ensure that IIX  YII, is defined for all X, Y by restricting the discussion to compact subsets of R2 . Therefore, let D2 = {xER211xl ~ I} and let aD2 denote its boundary. To ensure that the vector fields involved are well defined on aD2, we will assume they are defined on an open set U containing D2 and then take their restriction to the unit disc. Let Vecl(D2) be the set of all C'vector fields, X, defined in this way equipped with the C'norm
Notice the perturbed fixed point y* does not, in general, coincide with x*. However, given any () > 0, N can be chosen such that Iy·  x·1 < () for all YEN. Furthermore, by using Proposition 3.2.1 in conjunction with Hartman's Theorem (Theorem 2.2.3) we can deduce that there are neighbourhoods of x" and s: on which X and Yare topologically equivalent. Proposition 3.2.1 gives equality of dimension of the stable eigenspaces of the linearised flows exp(DY(y*)t) and exp(DX(x*)t). It follows that these linear flows are topologically equivalent (see Theorem 2.1.2). The Hartman Theorem states that there is a neighbourhood of x* (y*) on which the flow of i = X(x) (y = Y(y» is topologically conjugate to exp(DX(x*)t) (exp(DY(y*)t». Thus the local, COequivalence of the flows of X and Y follows from
where fIJ, and " are the flows of X and Y respectively. U~. and U~. t;: U denote the neighbourhoods on which Hartman's Theorem is valid. There is, therefore, a sense in which fIJ,: U > R· is structurally stable in a neighbourhood of x*. Namely . that for every pair (Y, y*) with YEN there is a neighbourhood U~. t;: U of y. such that "IU;. is COequivalent to fIJ,IU~ •. Thus, one might say that fIJ, is locally structurally stable at x*. Alternatively we can observe that (3.2.2) means that the topological type of the fixed point (see opening paragraph of Chapter 2) is preserved under all sufficiently small C'perturbations and say that the type of fixed point is structurally stable.
Hyperbolic fixed points of diffeomorphisms also persist under sufficiently small CIperturbations. Let Diff'(U) be the set of C'diffeomorphisms f: U(t;: !R") > n" with the CInorm. Then the following result parallels Proposition 3.2.1 for flows.
(3.3.1 )
Proposition 3.2.2 Let x* be a hyperbolic fixed point of the diffeomorphism f: U + !R". Then there is a neighbourhood Vof x* in U and a neighbourhood N of f in Diff'(U) such that every g E N has a unique hyperbolic fixed point y* E V of the same topological type as x*.
Definition 3.3.1 A vector field X in Vecl(D2) is said to be structurally stable if there exists a neighbourhood, N, of X in Vec'(D2) such that the flow of every Y in N is topologically equivalent to that of X on D2.
Proposition 3.2.2 allows us to show that hyperbolic closed orbits of flows on n" are structurally stable. Let the flow fIJ,: U > R", with vector field XE Vecl(U), have a hyperbolic closed orbit y. Then, its Poincare map P defined on a local section Sat X*EY belongs to Dill"(S). Moreover, x* is a hyperbolic fixed point of P. For sufficiently small e, every eClperturbation of P has a hyperbolic fixed point of the same topological type as x*, by Proposition 3.2.2. Hence, every sufficiently
The above precautions are not enough to focus attention on the structural instabilities occurring in the interior of D2. Unfortunately, instabilities associated with the behaviour of the vector fields at the boundary of the disc can still occur. In the absence of further constraints, vector fields that are tangent to aD2 are still present in Vecl(D2). Topologically distinct C'perturbations of such vector fields are illustrated in Figure 3.3. These unwanted structural instabilities can be excluded by confining the discussion to those vector fields that are transverse to aD2. Clearly, a vector field that satisfies this requirement is a member of one of two disjoint
(a)
(b)
3.3 Flows on twodimensional manifolds
127
126
3 Structural stability, hyperbolicity and homoclinic points
subsets of Vecl(DZ): it either points into or out of D2 at every point of aDz. Let Veci~(D2) be the set of vector fields defined on the disc DZ in the manner described above and such that X(x) points into DZ for every x in aDz. Then the following theorems generalise Propositions 3.1.1 and 3.1.2 for linear vector fields.
with the Clnorm. This norm is defined by imposing the C'<norm on each of the charts of a finite atlas for M. Then, Peixoto's result may be stated as follows.
Theorem 3.3.3 (Peixoto) A vector field in Vecl(M) is structurally stable if and only if its flow satisfies:
Theorem 3.3.1 (Peixoto) Let X belong to Vecf.(DZ). Then X is structurally stable if and only if its flow satisfies:
(i) all fixed points are hyperbolic;
(ii) all closed orbits are hyperbolic;
(iii) there are no orbits connecting saddle points.
(i) all fixed points are hyperbolic;
(ii) all closed orbits are hyperbolic;
(iii) there are no orbits connecting saddle points;
(iv) the nonwandering set consists only of fixed points and periodic orbits.
Moreover, if M is orientable the set of structurally stable Crxector fields forms an open dense subset of Vecl(M).
Notice that items (i) and (ii) in Theorem 3.3. t simply ensure local structural stability of the fixed points and closed orbits in the flow of X. It is really only item (iii) that involves a global property of the flow.
Obviously, parallel results could be stated for those vector fields that point out of D2 at every point of aDZ. Indeed, the flows of vector fields in Vecf.(DZ) and Vec!u.(DZ) are in 1:1 correspondence by time reversal.
By stereographic projection (see Figure 3.4(a», the unit disc DZ is homeomorphic to a closed cap, C2, based on the south pole of the sphere SZ. The boundary condition on aDZ can then be replaced by considering flows on SZ with a single repelling fixed point in S2\CZ• This is conveniently placed at the north pole (see Figure 3.4(b)). Provided that this additional fixed point is hyperbolic, a structurally stable vector field on D1 is also structurally stable on SZ. This follows from the work of Peixoto (1962) who extended Theorems 3.3.1 and 3.3.2 to vector fields on twodimensional, compact manifolds. Let M be a twodimensional, compact manifold without boundary and let Vecl(M) be the set of CIvector fields on M
Here orientable simply means that two distinct sides of M can be recognised. The sphere, torus, pretzel, etc. are all examples of orientable manifolds.
The statement of Theorem 3.3.3 contains an additional condition (iv) that does not appear in Theorem 3.3.1. Its role can be illustrated as follows. Consider the irrational flow on the torus TZ. This flow satisfies items (i)(iii) vacuously. There ~re n~ fixed points, closed orbits or saddle connections. However, it fails to satisfy Item (IV) because the nonwandering set is the whole of T1. Therefore, the irrational flow is not structurally stable on T2. Clearly, there exist f.Clc1ose rational flows for which every orbit is closed. Some examples of structurally stable flows on Sl and T2 are shown in Figure 3.5.
~t should perhaps be noted that, since M is compact, flows on it can only have finitely many fixed and periodic points if they are all hyperbolic. By the Hartman
Theorem 3.3.2 The subset of vector fields in Vecf.(D2) that are structurally stable is open and dense in Vecl.(D1).
Figure 3.4 (a) The stereographic projection from the north pole N of the s~here. ~he circle rt projects onto the boundary of D1. (b) A vector field ID a neighbourhood of N which cuts rt transversely.
N
Figure 3.3 The vector field in Vec1(D2) with a tangency at xoEoD2 in (a) is not structurally stable as the perturbation (b) shows. In (b), an orbit leaves D2 with increasing time.
(a)
(b)
128
3 Structural stability, hyperbolicity and homoclinic points
Y= y+xy
Example 3.3.1 Show that the vector field, X, of the differential equation
(3.3.2)
(a)
JVl
~O+ __ +___ (b)
1(\('
3.3 Flows on twodimensional manifolds
Theorem, the flow in the vicinity of a hyperbolic fixed point is topologically conjugate to that of its linearisation. The latter has an isolated fixed point at the origin, hence hyperbolic points must be isolated. This means that fixed points cannot accumulate at a hyperbolic fixed point. Consequently, only a finite number of fixed points can occur on a compact manifold if they are all hyperbolic.
We might try to extend the discussion of structural stability to noncompact sets such as the whole plane. Of course, a finite CInorm can no longer be guaranteed but an obvious approach is to consider restrictions to compact subsets. Clearly, if a vector field is structurally unstable on any compact subset of the plane then it can be deemed structurally unstable on the whole plane. This approach does have practical merit since in applications we are rarely involved with variables passing to infinity, rather they become very large but finite. The following example illustrates this idea.
129
there are points x in the boundary of D for which X(x) points both into and out of D. Therefore, we are (technically) not able to apply Theorem 3.3.1. However, consider the oneparameter family of systems
Y =  y + xy + p(2x  x2).
(3.3.3)
The vector field of (3.3.3) can be made &CIc1ose to that of (3.3.2) on any compact subset D of the type described above by taking p sufficiently small. The phase portrait for (3.3.3) with p > 0 is shown in Figure 3.6(b). The points (0,0) and (2,0) are saddle points for all real u. However, for nonzero p, y#O on the xaxis between these points. Moreover, the stable separatrix at (2, 0) is tangent to y = JPx. Therefore, for p > 0 there is no saddle connection between the fixed points. Hence, the flows for It = 0 and It> 0 are topologically distinct. The vector field in (3.3.2) is consequently structurally unstable on every compact subset of the plane containing the saddle connection. 0
It is perhaps worth noting that, in view of the role played by the boundary condition on aD2 in Theorem 3.3.1, we can say that a vector field that fails to satisfy the conditions (iHiii) will certainly be structurally unstable independently of the boundary condition on aD2. Of course, the converse is not true unless the vector field is transverse to the boundary.
Example 3.3.1 illustrates a useful way of giving meaning to structural instability on the plane. Some other examples are shown in Figure 3.7. However, we must not be misled into believing that if a vector field is structurally stable on arbitrarily large compact subsets of the plane then it is structurally stable on the whole plane. The following is a counterexample to this erroneous conjecture.
is not structurally stable on any compact subset of the plane with the line segment joining the singular points of X in its interior.
Solution. The system (3.3.2) has saddle points at x* = (0,0) and y* = (2,0). On the xaxis, y = 0 and so there is an orbit connecting these hyperbolic fixed points. The phase portrait is shown in Figure 3.6(a). Let D be any compact set containing the common separatrix of x* and y*. Notice that the stable manifold of x* is the line x = 0 while the line x = 2 is the unstable manifold for y*. This means that
Figure 3.5 Examples of structurally stable phase portraits on the sphere and torus. (a) 0 = sin 0, q, = 0; (b) 0 = sin 28, rP = I; (e) Ii = 1, q, = sin tp; (d) 0 = sin 20, cp = 1.
Example 3.3.2 Show that there are arbitrarily large compact subsets of the plane on which the system
v

6
(a)
(b)
(e)
(d)
x=x,
y = sin(ny) exp( _ y2)
(3.3.4)
y
Figure 3.6 Phase portraits for (a) (3.3,2), (b) (3.3.3) with II> O.
y
x
(3.3.5)
(a)
(b)
130
3 Structural stability, hyperbolicity and homoclinic points
131
3.3 Flows on twodimensional manifolds
is structurally stable. Verify that the topological type of (3.3.4) on the whole plane is changed by the addition of the perturbation (0, Jl) to (x, y), however small the value of Jl I 0.
Solution. Figure 3.8(a) shows y as a function of y. It follows that (3.3.4) has fixed points at (j,y) = (0, p), pEl. These fixed points are alternately stable nodes and saddle points as shown in Figure 3.8(b). Suppose D is a compact subset of R2 whose boundary intersects the jaxis at (0, y,), (0, y.) with y, < y. and y" y. I P for any integer p. If D is such that y(y,) > 0 and Y(y.) < 0 then (3.3.4) is structurally stable on D by Theorem 3.3.1. Clearly, arbitrarily large D can be constructed in this way.
Consider the family of vector fields defined by i = Xp(x), where
x = x, y = sin(ny) exp(  i) + Jl.
Figure 3.7 Some examples of structurally unstable phase portraits on R2. In each case the structural instability is apparent in the restriction of the flow to a compact subset of the plane.
(a)
(e)
The vector field of (3.3.4) is Xo and
(3.3.6)
Thus Xp can be made eC'cIose to Xo on the whole plane. However, Xp, Jl I 0, has only finitely many fixed points (see the broken curve on Figure 3.8(a)). Hence, the flow of (3.3.5) is topologically distinct from that of (3.3.4) for any nonzero value of p. 0
Notice that the topological type of the flow of Xo in Example 3.3.2 changes when Jl departs from zero, even though it has only hyperbolic fixed points and no saddle connections. Since the plane is not compact, infinitely many hyperbolic fixed points can occur. If we wish to maintain contact with Theorem 3.3.1, we must specify a boundary condition at infinity. However, if we require that all vector fields point inward on the boundaries of all sufficiently large discs, we are led back to Theorem 3.3.3 via stereographic projection. In other words, we can obtain structurally stable flows on R2 from structurally stable flows on S2, however, they will have only finitely many hyperbolic fixed points and their behaviour at infinity will correspond to having a hyperbolic fixed point at the north pole of the sphere (see Figure 3.9).
Figure 3.8 Phase portrait of (3.3.4) in relation to plot of y as a function of y. All zeroes of yare simple. The fixed point (x*, .v*) is a saddle if d ytd yl,' > 0 and a node if d ytd .vI" < O. The broken curve in (a) is a plot of y versus y for (3.3.5). The decay of the amplitude of the oscillations in y means that there are no fixed points for exp( i) < 1111. i.e. for y2>ln(llll)·
y
(b)
(d)
132
3 Structural stability, hyperbolicity and homoclinic points
DetA=±1.
3.4 Anosov difJeomorphisms
133
all of which are hyperbolic. Moreover, the structurally stable difJeomorphisms form an open dense subset of DilTl(SI).
Recall (see Proposition 1.5.1) that if f has periodic points then its rotation number p(f) is rational. Therefore, the structurally stable dilTeomorphisms on SI have rational rotation number (the converse of this statement is not true (see Exercise 3.4.1)). Iff is structurally stable with rotation number p(f) = p/q, in lowest terms then its dynamics are very simple. It has an even number of periodq cycles with stable and unstable periodic points alternating around the circle (see Figure 1.23).
It was hoped that generalisations of the behaviour described in Theorems 3.3.3 a~d 3.~.1 would not only characterise structurally stable systems in higher dimensions but would also prove to be generic. To this end, 'MorseSmale' vector fields and dilTeomorphisms were defined (Chillingworth, 1976, p. 231 and Nitecki, 1971, p. 88_l. Unfortu.nate~y, it was found that, while such systems were structurally s~able, .thelr properties did not characterise structurally stable systems in higher d~menslOns .. In particular, it was shown that there were structurally stable diffeomorphisms on manifolds of dimension n ~ 2 (corresponding to vector fields on m~nifo~ds ~f dimension n + I ~ 3, by suspension) whose nonwandering sets con tamed infinitely many periodic points. The Anosov difJeomorphisms of the torus T", are a subset of DiITI(T") that exhibit this behaviour. We can describe a dilTeomorphism f on T" in terms of a 'lift' in much the same way as we did for dilTeomorphisms of the circle (see § 1.2.2). In this case, the lift f is a dilTeomorphism on RO which satisfies
3.4 Anosov diffeomorphisms
The relationship between flows on nmanifolds with global sections and dilTeomorphisms on manifolds of dimension n I (via Poincare maps) was discussed in § 1.7. It is not surprising, therefore, that there is an analogous result to Theorem 3.3.3 for dilTeomorphisms on a compact Imanifold without boundary. Topologically, there is only one such connected manifold; namely the circle, SI.
Let DiITI(SI) be the space of orientationpreserving CIdilTeomorphisms on SI with the CInorm. Then Peixoto's theorem for dilTeomorphisms on the circle can be stated as follows.
Theorem 3.4.1 (Peixoto) A difJeomorphism f E DilTl(SI) is structurally stable if and only if its nonwandering set consists of finitely many fixed points or periodic orbits
Figure 3.9 Some examples or structurally stable phase portraits on RZ derived from flows on S2. All fixed points and periodic orbits are assumed hyperbolic. To ensure stability at infinity the corresponding vector fields on S2 have an unstable hyperbolic fixed point at the north pole (cr. Figure 3.4).
f(n(x)) = n(f(x», for each x E R", where n: R" + T" is given by
n(x) = n«xl> ... , xo)T) = (XI mod 1, ... , x. mod l)T = (91)'' .,9o)T =8
(see Figure 3.10). If kEI·, then (3.4.1) implies
n(f(x + k) = f(n(x + k] = f(n(x)) = n(f(x)), for all x E R". Continuity of f then gives,
f(x + k) = f(x) + I(k),
(a)
(b)
(3.4.1 )
(3.4.2)
(3.4.3)
(3.4.4)
where l(k)E I" (see (1.2.10) et sequel. All lifts of dilTeomorphisms on T" must satisfy (3.4.4).
We will begin by describing a special subset of the Anosov dilTeomorphisms t~at are know~ as ~he An~sov automorphisms. The lift of an Anosov automorphism f IS a hyperbolic, hnear diffeomorphism, A: RO + RO, which is such that:
A: I" + In;
(c)
(d)
(3.4.Sa) (3.4.5b)
(3.4.6)
Thus, the Anosov diffeomorphisms on T", n ~ 2, are examples of structurally stable diffeomorphisms on a compact manifold whose nonwandering set contains infinitely many points. What is more, the dynamics on 0 is very complicated involving infinitely many periodic orbits densely distributed over the torus. In fact, every periodic point is hyperbolic (see Exercise 3.4.2). Since TO is compact, there can only be finitely many such points with a given period q on the torus. However, periodic points of infinitely many periods can be shown to occur (see Exercise 3.4.3) making up the infinite set predicted by Proposition 3.4.1. To gain some insight into how all this complexity arises let us consider the following well quoted example (see Arnold, 1983; Arnold & Avez, 1968).
134
3 Structural stability, h}'perbolicity and homoclinic points
135
3.4 Anosov diffeomorphisms
Together (3.4.5a and b) ensure that A I: IO + Z" and it follows that both A and A I satisfy (3.4.4). Moreover, given that f(n(x)) = n(Ax) and g(n(x» = n(A 'x), then
Proof. Let IJ be a periodic point of f, then f4(0) = 0, for some positive integer q. Suppose XER" satisfies n(x) = 0, then
n(x) = 0 = r(lJ) = f'(n(x)) = n(A'x).
(3.4.8)
and
f(g(n(x))) = f(n(A 'x» = n(AA 'x) = n(x)
This means that
g(f(n(x))) = g(n(Ax» = n(A , Ax) = n(x).
(3.4.7)
(A' I)x=m,
(3.4.9)
Thus, r': T" + yo exists and has lift A I. Finally, observe that f(x) and r' (x) are differentiable because Ax and A 'x are obviously so and n is a local diffeomorphism. Hence f is a diffeomorphism on T".
Hyperbolic, linear diffeomorphisms on R" are often referred to as automorphisms because they are isomorphisms of the group R" with itself. The diffeomorphism f defined above is called an automorphism to distinguish it from other Anosov diffeomorphisms whose lift is not linear. A general definition of Anosov diffeomorphisms can be found in Arnold (1973, p. 126) or Nitecki (1971, p. 103) but, thanks to the following result due to Manning (1974), it will be sufficient for us to consider only the automorphisms here.
where x = (XI" .. , xo)T and m = (m" ... , molT, mjEI, i = I, ... , n. Since A is a hyperbolic matrix, (A q _ W' exists and (3.4.9) has solution
x = (A4_ W'm. (3.4.10)
Now, A q _ I is an integer matrix and therefore (A q _ If' has rational elements. Hence x has rational coordinates.
Conversely, if IJE T" has representative x = «p\O)/r), ... , (p~O)/rW, where pIO), rEI with r ;60, then, for any kEI,
Atx= (pt, ... , p;kJ,
for some integers p\k), ... , p~t). However, there are at most r· points on T" that can be represented in this way and, therefore, there is a q > 0 such that n(A'x)= n(x). 0
(3.4.11 )
Theorem 3.4.2 Every Anosov diffeomorphism f oj T", such that O(f) = T", is topologically conjugate to some Anosov automorphism oj T".
Coupled with Theorem 3.4.2, the following result shows that such Anosov diffeomorphisms have complicated nonwandering sets.
Proposition 3.4.1 A point IJ E T" is a periodic point oj the Anosov automorphism f: T" + T" if and only if IJ = n(x), where x E R" has rational coordinates.
Proposition 3.4.1 not only implies that f has infinitely many periodic points; it also shows that the periodic points are dense in the torus. All these points lie in the nonwandering set 0 of f and, since 0 is closed (see Exercise 1.4.2), we conclude that 0= T",
The final piece of the argument against MorseSmale systems was provided by Mather (1967).
Figure 3.10 (0) Angular coordinates on Tl measured in units of 21t. (b) The map 1t identifies each point (XI' x2)e R2 with a point (8" 82)e {(x" xI)1 0,,;;; XI < 1,0";;; Xl < I} which in turn defines a unique point on T2.
(0)
Theorem 3.4.3 (Mather) The Anosov diffeomorphisms on T" are structurally stable in Diff'(TO).
(b)
136
3 Structural stability, hyperbolicity and homoclinic points
Let A: R2 ..... R2 be given by
A=C ~).
It is easily verified that A satisfies (3.4.S). The behaviour of A on R2 is simple: it has a saddle point at x = 0 with stable and unstable eigenspaces given by the
straight lines
(3.4.12)
(1 51/2)
y=  x
2
v=(~ 51/2)x
. 2 '
(3.4.13)
and
respectively. The complexity arises when it is mapped down onto T2. Forward iterations of A on R2 have the effect of contracting and expanding along the two perpendicular directions in (3.4.13) as shown in Figure 3.11(0). The unit square BI = {(x, y)IO ~ x < I, 0 ~ Y < I} is mapped onto thinner and thinner parallelograms (Figure 3.1I(b) and (e)). The slopes of the longer diagonal of these parallelograms are rational but they approach the irrational HI + 51/2) for large numbers of iterations. When the points in these images of B I under A are identified with points on T2 (see Figure 3.I1(b) and (e)), it can be seen that repeated application of f has the effect of distributing any subset of T2 more and more evenly over the whole torus. An alternative way of seeing this is to recognise that, for any x E R2, ANx can be made arbitrarily close to the line y = }(I + S1/2)X by taking N to be sufficiently large. Since W + 51/2) is irrational, this straight line represents a curve, W·, that winds densely around the torus (see Figure 3.12).
The stable eigenspace of the saddle point in Figure 3.11(0) also corresponds to a densely wound curve, W', on T2. The key to the complexity of the dynamics of f lies in the fact that these stable and unstable manifolds intersect in a dense set of transverse homoclinic points (see Figure 3.12). A Itomoclinie point is one that lies in both the stable and the unstable manifolds of a fixed or periodic point. Such points are said to be transverse if they arise from a transverse, rather than a tangential, intersection of the manifolds. Observe that if ot is a homoclinic point, i.e. ot E W'n W·, then f(8~)E W'n W· because s' E W'··. Hencef(8t) is a homoclinic point. Thus the dynamics of these homoclinic points is confined to the dense set
of intersection points of W' and W·.
The reader must not confuse these homoclinic points with the periodic points of f. Recall, by Proposition 3.4.1, periodic points of r have representatives XE R2 with rational coordinates. However, the stable and unstable eigenspaces of A have irrational slope (see (3.4.13)) and, apart from x = 0, no point on them has rational coordinates. We shall.see later (see §3.7) that the occurrence of transverse homoclinic points is indicative of complicated dynamical behaviour.
It is also worth noting that A4 is a lift of f4 for any q. Thus if r has a periodic point 8* of period q, then its stable and unstable manifolds are densely wound curves on T2 parallel to W' and W·, respectively. This is because, for each x*
3.4 Anosov di.ffeomorphisms
137
.Figure 3.11 JIIustration of the toral automorphism given by (3.4.12): (a) linear saddle point of A = G ~} (b) image A(Bd of the unit square BI under A; (e) image A2(BI). The shading in (b) and (e) indicates how n(A(Bd) defines f: r2 .... ;r2 and n(A2(Bd) gives f2: r2 + t,
(a)
o
2 3
(c)
o
I 2
(b)
4
5
138
3 Structural stability, hyperbolicity and homoclinic points
satisfying 8* = n(x*), we can write
A'(x) = A'(x*) + A'(x  x*).
(3.4.14)
Thus, the stable/unstable eigendirections at x· are given by translates of the stable/unstable eigenspaces of A at the origin. It follows that f'l also has a dense set of transverse homoclinic points for each q. Thus, we can expect the complexity arising from homoclinic points in I, to also occur in f'. This is reminiscent of the 'complexity on all scales' observed in Figure 1.42.
3.5 Horseshoe diffeomorphisms
This is another example of a class of diffeomorphisms which are structurally stable and have a complicated nonwandering set supporting infinitely many periodic orbits. These diffeomorphisms are particularly important because the complexity
Figure 3.12 Segments of the stable and unstable manifolds of the fixed point, 11:(0), of f on T2. Segments AB and A' B' of EU and E' for A have been mapped onto the torus using It(x) = (x mod I, y mod I). All intersections of W' and WU, except at the origin, are transverse homoclinic points. Note WU(WS) always has +ve(ve) slope on 81.
y 2
B
/ I I
/ /
I
A I
3.5 Horseshoe diffeomorphisms
139
exhibited by them can be shown to occur in any map that has transverse homoclinic points (see § 3.7).
3.5.1 The canonical example
Consider a diffeomorphism f:Q_,,~2, where Q={(x,y)lIxl,IYI~I}, that is constructed in the following way. Each point (x, Y)EQ is first mapped to (5x, y/5) and Q is mapped onto the rectangular region R = {(x, Y)lIxl ~ 5, Iyl ~ liS}. This region can be divided into fifths by the lines [x] = I and 3. The map f is completed by bending the central fifth of the rectangle and placing the resulting horseshoeshaped region on the plane in such a way that its second and fourth fifths intersect with Q in Qo and QI as shown in Figure 3.13(a). Observe that, if Po, PI denote the preimages of Qo, QJ> respectively (i.e. f(Pi) = Qi' i = 0, I; see Figure 3.13(b)), then f] Pi is linear for i == 0, I (see Exercise 3.5.1).
We can show that fhas a complicated invariant set by considering the sequence of subsets of Q defined inductively by
(3.5.1)
nEI+, with Q(I)=QOVQI. It is not difficult to see that Q(2)==f(QovQtlnQ consists of four horizontal strips lying inside QOVQI (see Figure 3.14(a)). Moreover, it is apparent that Q(I) ::> Q(2) ::> .•• ::> Q(') ::> ••• and Q(') consists of 2" horizontal strips (see Figure 3.14(b)). If we consider the intersections of the sets Q(') with the jaxis then the relationship between the resulting subintervals for successive values of n is easily recognised as a prescription for the construction of
a Cantor set. It follows that the intersection n Q(') is a Cartesian product of an
neZ +
interval in x with a Cantor set of yvalues, In a similar way, iterations of the inverse of f can be used to obtain an analogous set of vertical strips. Some care is needed here because f11Q is only defined on the subset Q(I) == QOVQI of Q. To avoid this difficulty, we take
Q(O)=POvPI =rI(Q(I») (3.5.2)
and define
(3.5.3)
for nEI+. The latter equality in (3.5.3) follows because Q((nI» is a subset of the two vertical stripsQ(O) = Po v PI for all n E I +. Thus, for each n E I +, the intersection of Q((n I) with Q(I) = Qo v Q I is the same as its intersection with the whole horseshoe f(Q). The map rI stretches Q((·I)nQ(I) linearly by a factor of five in the jdirection, contracts it by five in the xdirection and replaces it on the square as shown in Figure 3.tS(a,b) for n = 1,2, respectively. The sets Q(O), Q(I), Q(2), ••• then have 2,4,8, ... vertical strips, respectively (see Figure
3.15(c)) and n Q(n) is the Cartesian product of an interval in y with a Cantor
.eN
>< ,
I I I
I I I I I
lt l
A:
cc: I I L J
r~__,~
 ....
I
.... ,....
I
Lt' Q:l I
\\' \\t
,\"
\ ,~,
'<:~ I I I I I I
(3.5.4)
3.5 Horseshoe diffeomorphisms
143
142
3 Structural stability, hyperbolicity and homoclinic points
A= n Q(0)
o.l
Together (i) and (ii) imply f(x)e f(Q10)flQ = Qlo+ I). Thus, if xeQ(0), for all ne Z then f(x)eQI.+I) for all neZ. Hence f(x)eA. ' Similar arguments (see Exercise 3.5.2), with the roles of (3.5.1) and (3.5.3)
reversed, show that A is invariant under f" I and, therefore, f(1\) = A. 0
set of xvalues. If we now define
then A is a Cartesian product of two Cantor sets which is itself a Cantor set.
Proposition 3.5.1 The set A = n Q(o) is invariant under f and f" I .
o.l
The map f, as we have defined it up to now, is not a diffeomorphism of the square
Proof. Let xeA, then xeQ(n) for all neZ. Now if xeQ(n), neN, then (3.5.3) implies that f(x)e Q({ni))flf(Q) £ Q((nI)). IfxeQ(n), ne Z+, then observe that:
Co'Do'
(i) (ii)
f(x)ef(Q(n); and f(x)ef(Q(O)£Q, since xeQ(O).
(x, y) I (lx, 5y)
replace on Q
Figure 3.15 (0) Illustration of QIO) = £I(QII)) = Pou PI' Notice £1 only acts in Q(I)=QoUQI' (b) Illustration of QII)=rI(QIO)nQII). The shaded squares in the unprimed part of the diagram represent QIO) n Q(I). QII) consists of the shaded strips in the portion of the diagram labelled with two primes. (c) The vertical strips Q'O), Q'I), Q'Z), defined by (3.5.3) are shown. Observe that Q(O)=>QII)::>QIZ), .... The set QIn) consists of 2"+ 1 disjoint vertical strips each of width 2/510+ I).
Co' Do'
I I I I
I I I I
I I 1 1
I I I I
I 1 1 I
I I I I
I I I I
I 1 I I
I I 1 1
I I I I
1 I I I
I I I 1
I I I I
I I I I (0)
144
3 Struaural stability, hyperbolicity and homoclinic points
Q «((Q) i. Q). Moreover, it does not have an obvious connection with diffeomorphisms on compact manifolds without boundary. However, a diffeomorphism, g: S2 ..... S2 can be constructed such that f is the restriction of g to a subset of the sphere. The first step in the construction is to extend the map (to a capped square Q' as indicated in Figure 3.16. The extension I' is constructed in such a way that ('IF has a unique, attracting, hyperbolic fixed point. This means that once a point is mapped into F its orbit subsequently remains in F. The map f" can, in turn, be extended to a closed disc D2 of suitably large radius. The extension g': D2 ..... D2 is taken to be such that g'(D2) takes the form shown in Figure 3.17. The diffeomorphism g: S2 ..... S2 is finally obtained by identifying the disc, D2, on R2 with a cap, C2, on the sphere (by stereographic projection, see §3.3) and adding a unique, repelling, hyperbolic fixed point in S2\C2•
Since g is a global diffeomorphism on S2, both g and gI are defined for all points of the sphere. However, its construction ensures that g coincides with f or
Figure 3.16 (a) The capped square Q' = GuQuF; (b) the extension f': Q' + Q' is such that G' = 1'(G) and F' = rtF) are both subsets of F.
Q Q
I I
I I
I I
I I
I I {1
I I
I I
I I
L : Po M PI: R
I I
I I QI
I I
I I
I I
I I
I I
I I
(a) (b) Figure 3.17 (a) The disc D2 containing Q'; (b) image of D2 under g' (shaded) with g'IQ' = 1'.
(a)
(b)
3.5 Horseshoe di.fJeomorphisms
145
f" when restricted appropriately. Here we have taken the notational liberty or not distinguishing between these restrictions on the sphere and their representatives on 1R2 via stereographic projection. This distinction does not play a significant role in our discussion and, once noted, should not lead to any confusion. The unstable fixed point in S2/C2 means that the ordinary points of this set move towards C2 under g. As we have shown in Figure l17(b), gl(D2\Q') is again a contraction and, therefore, g essentially delivers points to Q'. On Q', g behaves in the same way as I'. We already know that f' has an invariant Cantor set A arising from its restriction, (, to Q, but what of the points S2\1\ = A <? The following proposition provides part of the answer to this question. It states that those points
of Q' that do not lie on the infinite set of vertical line segments, n Q(",
are eventually swept into F.
neN
Proposition 3.5.2 The orbits under g of points in Q\( n Q(n,) ultimately approach
neN
the stable fixed point of g in F.
Proof. Figure 3.16 shows what happens to the various parts of Q under a single application of glQ' = I'. The left (L) and right (R) fifths of Q are mapped, together with G and F, into F. Since F contains a unique, attracting fixed point, the orbit of any point in F approaches this point asymptotically. Points in the midfifth, M, of Q suffer the same fate after one more iteration. Such points are mapped into G by (' and into F by f'2. Only points in Po u P I remain in Q (in fact in QouQd after one application off'. In other words, points in Q\Q(O, enter F after at most two iterations of g.
Let us focus attention on the partition of the square Q provided by Q(I" rather than Q(O,. Observe that points in Q(O\(Q(O'nQ(I,) are mapped into L, M, Rafter one iteration of glQ = (see Figure 3.l5(b» and thence into F after two or three iterations. Thus we conclude that all points in (Q\Q(O)u(Q(O\(Q(O)nQ(l))= Q\(Q(O'nQ(l,) enter F after at most three applications of g. Similarly, if we consider the partition of Q provided by Q(2), we conclude that points in Q(1)\Q(2'=(Q(0InQ(I')\(Q(0'nQ(I'nQ(21) have images under g in
2
Q(O\(Q(O'nQ(1'). Therefore, if xeQ\ n Q(n, then gk(x)eF for k » 4. Thus, we
n=O
conclude inductively that all points in Q\ n Q(" ultimately enter F. 0
neN
In view of the construction of A, it is clear that A£; n Q(.'. Moreover, it is not
neN
difficult to show that n Q(nl is invariant under g (see Exercise 3.5.3). Bearing
.eN
MH
3.5 Horseshoe difJeomorphisms
147
146
3 Structural stability, hyperbolicity and homoclinic points
in mind that n Q(.) is a set of straight line segments parallel to the jaxis and
.eN
that g involves a contraction along that direction, it is not surprising that points
in (n Q(·)\A have orbits that approach A asymptotically (see Exercise 3.5.3).
uN
It must be emphasised that these orbits are not confined to a single vertical line
subset of n Q(.) .
• eN
Let us now tum to the dynamics of points in S2\Q'. Unlike r1, g1 is defined
for all x E Q (see Figure 3.18). Of course, g 11 Q(1) == r 1 is as illustrated in Figure 3.15 but gI(Q\O'I)£S2\Q'==Q'0. This means that points in Q\Q(1) are the images under g of points lying outside Q'. We have already discussed the fate of such images under forward iterations of g. Since An(Q\Q(1)== 0, points in
(Q\Q(I)n( n Q(.) have orbits approaching A. asymptotically; while those in
we are able to extend the invariant set n Q(.) onto the whole sphere. The
.eN
resulting set of points is called the inset, inCA) of A, i.e.
inCA) = {x ES21g"(x) .... A as n .... co].
(3.5.5)
Similar arguments to those presented above, with g 1 replacing g, lead to analogous conclusions about the infinite set of horizontal lines n Q(.) (see
.. «r+
Exercise 3.5.3). It follows that there is a set of points, out(A) £ GvQv F, the outset of A, whose orbits approach A under reverse iterations of g, i.e.
(3.5.6)
.eN
The role of the set A in the dynamics of g is clearly analogous to that of a saddle point in simpler diffeomorphisms. The inset and outset of A. generalise the stable and unstable manifolds of the saddle. We will return to sets possessing this more general hyperbolic structure in §3.6 but now we must consider the dynamics of g on A .
(Q\Q(I)\ n Q(.) have orbits entering F. However, under reverse iterations of g
• eN
1 • hat ooi .. Q\Q(I)
Figure 3.18 Illustration of g  showing t at points starting In =
TuMHuB are mapped out of Q' by gI. Observe that S2\Q' contains a unique, stable, hyperbolic fixed point of gI so that the orbits of these points under g I do not return to Q'.
T
3.5.2 Dynamics on symbol sequences
Let :E be the set of all biinfinite sequences of the binary symbols {O, I}, i.e. I= {e]e: 1 .... {O,l}}. Thee1ements, a, of I are called symbol sequences and they are defined by specifying u(n)=u.E{O,l} for each nE1. We will write a = {un}:'=  co = { ... a _ 2U _I Uo· U 1 U2 ••• }. Our aim is to study the dynamics of the map a:: I .... I defined by
Q
(3.5.7)
n E 1. This is known as a leftshift on I because it corresponds to moving the binary point one symbol to the left.
B
(a)
Proposition 3.5.3 The left shift a:: :E .... I has periodic orbits of all periods as well as aperiodic orbits.
A point u* E:E is periodic if
a:4(U*) = u*, (3.5.8)
q E 1 ". If q is the least, positive integer for which (3.5.8) is satisfied then u* is said ~o be of period q. It is not difficult to see that (3.5.8) will be satisfied if and only If u: = u:+q, for all n E 1. It is then easy to find periodic points of a: with any given period, q. The required sequence, u*, is generated by repetition of a block of symbols oflength q that is itself not composed of repetitions of any of its subblocks. For example, the point
(h)
U* = { ... 1101011110111011010·11110 ... }
(3.5.9)
(3.5.13)
3.5 Horseshoe diffeomorphisms
149
148
3 Structural stability, hyperbolicity and homoc1inic points
a= { .. 1 .. 110 ... 11110111011010·1011011101111 .. , .},
(3.5.11 )
the following ordered lists of symbol blocks:
(i) all blocks of length I, i.e, {OJ, {I};
(ii) all blocks of length 2, i.e, {O, OJ, {O, I}, {I, OJ, {I, I};
(ii) all blocks oflength 3; and so on.
All possible blocks of all lengths are included in {u n}.:"=O; Un' nE Z+, can be chosen arbitrarily. The orbit of U under 11. contains {l1.m(U)lmE N}. Now, by construction U contains any given symbol block oflength N in its left hand half. After sufficiently many applications of 11. this block will be centrally placed about the binary point. Since N is arbitrary, any element of ~ can be approximated arbitrarily closely by some point on the orbit of a under 11..
In view of the rather special construction used above to obtain a sequence a whose orbit under 11. is dense in ~, the reader may feel that such sequences are in some sense rare or atypical. This is not the case. In fact, most binary biinfinite sequences contain any prescribed block of symbols (see Hardy, 1979) and therefore have a dense orbit under 11.. The particular example chosen above is carefully ordered purely to make the argument more convincing.
Having established some properties of the left shift 11.: ~ ... ~, we must reveal our motive for examining the dynamics of this map: namely to obtain a symbolic description of the dynamics of the horseshoe diffeomorphism on A. Before doing this, it is worth noting that the validity of Propositions 3.5.35 does not depend upon the binary nature of the sequences in ~. Similar results can be derived for sequences of msymbols, {O, I, ... , m  I} say (see Exercise 3.5.5). Binary symbol sequences allow us to deal with the horseshoe map of § 3.5.1. However, there are more sophisticated maps of this type (see Exercise 3.6.5) whose 'symbolic dynamics' involve sequences of m symbols with m > 2.
has period14, while
u* = { ... 101011010 1011010·101 ... }
(3.5.10)
satisfies 11.14(u*)=u* but has period7 because 11.7(u*)=u*, also. It is equally straightforward to show that 11. has aperiodic orbits. For instance,
4 2 I
which contains symbol blocks of the type shown for all n E Z +, is such that there is no q E 1. + such that a.4(u) = a.
Proposition 3.5.4 There is a topology in which the periodic points of 11. are dense ill~.
There is a natural way of defining how close two symbol sequences are to one another. Given two sequences in ~, we can obtain the length of the largest symbol block, centred on the binary point, on which they agree. The larger the size of this block the closer the two sequences are deemed to be. We are then able to define the limit of a sequence of elements in ~. A sequence {u(m)}:=o c;; ~ is said to tend to UE~ as m ... 00, if, given N E1.+, there exists M E1.+ such that u~n) = a; for (N I)~n ~N, when m> M. Clearly, if u(m) ... U as m ... 00 then u(m) and a agree on increasingly large central blocks. For example, the sequence u(m) defined by
{I  (m  1) ~ II ~ m, m E 1. +
u(m) = (3.5.12)
o 0 otherwise,
converges, as m ... 00, to the sequence a with Uo= 1 for aIlIlE1..
With the above definition of convergence, periodic points of 11. are dense in ~.
This follows because, given any a E~, there is a sequence of periodic sequences {u(m)}:=o which tends to a as m'" 00. Each sequence u(m) is simply taken to be periodic with period 2m and such that u~m) = Uo for  (m  1) ~ n ~ m. As an example, let u be the aperiodic sequence (3.5.11), for which
u(1) = ... 0101 if.1 01 01. .. ,
(j(2) = ... 1010 10·10 1010 ... ,
(j(3) = ... 0TIffi)f 01Q.i0f 0TIffi)f ... ,
(j(4) = ... 10101011 1010·101110101011. ..
3.5.1 Symbolic dynamics for the horseshoe diffeomorphism
In this section we show that the restriction of the horseshoe diffeomorphism to the invariant set A is topologically conjugate to the left shift IX on ~. The key idea is that the points of A can be 'coded' as biinfinite sequences of {O, I}.
Recall that A = n Q(n), where Q(n), n E 7L +, is the disjoint union of 2nhorizontal
o.l
strips on the square Q, while Q(n), nEN, is the union of 2n+1 similar vertical strips. As Figures 3.14(b) and 3.l5(c) illustrate these sets of strips are 'nested', i.e. Q(1)::l Q(2)::l ••• ::l Q(n)::l .•• and Q(O)::l Q(I)::l ••• ::l Q(O)::l .. '. Thus, NN) =
N
n Q(o)=Q((NI))nQ(N) is the disjoint union of 22N squares of side 2/5N n= (NI)
(see Figure 3.19). Clearly, as N 00, the size of the squares tends to zero, their
number becomes infinite and NN) A. The coding ofthe points of A follows from
the fact that each square of NN) can be uniquely represented by a symbol block, U(N) = {u (NI)' .. (jo ·ul ... (jN}, a; E {O, I}, of length 2N.
Any given strip in Q(n) can be allocated either 0 or 1 in the following way.
and so on.
Proposition 3.5.5 The left shift 11.: ~ ... ~ has a dense orbit on ~.
To justify Proposition 3.5.5 we must show that IX has an orbit on ~ that approaches every point of ~ arbitrarily closely. Let (jE ~ be such that o n for II E N is given by
(3.5.14)
(c)
150
3 Structural stability, hyperbolicity and homoclinic points
151
Consider the vertical strips Po and P I' Observe that Q(I) = g(Po)ug(PI) = QOUQI
where QonQI = 0 (see Figure 3.13). Furthermore, Q(2) S; g2(PO)ug1(P tl
with g2(Po)ng2(Ptl = 0 (see Figure 3.14(a)). In general, for nE L", Q(n) s; g"(Po)ug"(P tl
and g"(Po)ng"(Ptl is always empty because PonPI = 0 and g is a diffeomorphism. Thus a horizontal strip of Q(") lies either in g"(Po) or g"(P I)' We ~ll~te the symbol 0 to a strip of (t") if it is a subset of g"(P 0) and the symb?l 1 If It. lies in g"(P1) (see Figure 3.20). Obviously, these symbols alone do not provide a umque description of each horizontal strip in Q(n) for n ~ 2, however, they can be used to obtain one. For example, two strips of Q(l) have been allocated the symbol 0 but they are distinguished by the fact that one lies in g(Po) (i.e. strip 0 of Q(1)) and the other lies in g(P I) (i.e. strip 1 of Q(I)). Hence the strips in Q(2) can be
Figure 3.19 Illustration of n~= (NI) Q(n) for (a) N = 1; (b) N = 2. The square regions defined by (3.5.17) with .,.(2) given by {11·01} and {10 ·11}
are indicated.
(a)
3.5 Horseshoe diffeomorphisms
(3.5.15)
uniquely labelled by giving two symbols: the first specifying a strip in Q(1) so that the second uniquely determines a strip in Q(2) (see Figure 3.20). Similarly, the strips of Q(3) can be uniquely labelled by starting from the unique labelling of the strips in Q(2) and appending the symbols allocated to Q(3). It follows that the strips
fQ(") . I . b
o are umque y specified y a set of n of the symbols {O, I}. Similar arguments
can be carried through for Q(  0) by considering the images of Po and Plunder powers of gI (see Exercise 3.5.8). A vertical strip of Q("), n E N, is allocated the symbol i if it is a subset of g°(Pi), i = 0, 1. For n E r: , unique labels for the strips of Q(o) are obtained by appending these allocated symbols to those of the strips in Q((nl)} (see Figure 3.21). Notice we have appended symbols on the left so that the order in the strip label matches that of the negative integers. Finally, we
(3.5.16)
Figure 3.20 Coding of strips in Q(n) for (a) n = 1; (b) n = 2; (c) n = 3. The symbol allocated to each strip is shown on the left and the unique code for the strip is given on the right.
o ~ g(Po)
I ~ g(f\)
(a)
Q(I)g Q(O) EffiI Q(O)nQ(I) •
v
(b)
(b)
Q(21. Q(I).
.~ 1/

/ ~
, 1/
I K I o o I
I o o I
o I I ~ g(Po) n g2(PI) n g3(PI) 010
000
001
101 ~ g(P1) n g2(Po) n g3(PI) 100
110
III
nQ(nl.
11=1
Proposition 3.5.6 implies that the complexity exhibited by the orbits of points in :E under 0( (see §3.5.2) also occurs in the orbits of points of II. under g. Thus gill. has infinitely many periodic points, its periodic points are dense in II. and it has orbits that are themselves dense subsets of 11..
Another feature of the dynamics of 0(: :E + E, that has important repercussions for gill., is that there are points in :E whose orbit under 0( is aperiodic. Since O«(u) and a are not, in general, close in :E, these orbits wander throughout r in an apparently disorganised way. Similarly, their counterpart in the orbits of gill. move around II. by hopping from point to point in a random or chaotic way. Indeed,
152
3 Structural stability, hyperbolicity and homoclinic points
153
• . AIN) E h can construct the symbol blocks representing the squares occurnng In 1\ • ac
. I . fQIINI" ith one
such square is the intersection of one of the veruca stops 0 WI
of the horizontal strips of Q(N). If the vertical strip has label U INI)' ... , U  h 110 and the horizontal strip has label 111, .•• , UN' the symbol block representing the square is taken to be UIN) = {11INI)' ••• , U I' Uo' U h .•. ,11 N}' Thus, for example {II.OI} and {IO.Il}, respectively, represent the top right hand a~d botto.m left hand squares in the illustration of 11.'2) given in Figure 119(b). It IS not difficult to show (see Exercise 3.5.11) that the square represented by the symbol block UIN)
Figure 3.21 Coding for strips of Q(.) for n =0, 1,2: (a) Q(OI; (b) Q(I'; (e) Q(21. Unique labels for the strips are given above and allocated symbols below. Notice that, to match the negative integers, symbols are appended to the left rather than to the right.
Po
I
o
~
I
o
(a)
3.5 Horseshoe diffeomorphisms
is given by
(3.5.17)
10 0 I
(b)
In the limit N + 00, the above construction assigns a unique, biinfinite binary sequence with each point of 11.. Moreover, (3.5.17) allows any such sequence to be converted to a unique point of 11.. We have therefore constructed a bijection h: r +11..
Proposition 3.5.6 The bijection h: :E + II. defined above is a homeomorphism that exhibits the topological conjugacy of g: II. + II. and O(::E + E.
Proof. The nested nature of vertical and horizontal strips defining II. means that sequences that are close, in the sense that they agree over large central blocks, map under h to points of II. that are geometrically close together. Similarly if two points of II. are geometrically close, the symbol sequences agree over a large central block because it is only for N sufficiently large that such points are distinguished in NN). Thus h is a homeomorphism.
Let UEr and
(15.18)
Then
g(h(U))=g(X)=g(LQ g.(p"J]nQ). =[n g'+I(p"J]nQ,
.el
=[n g'(P,,"_,l]nQ,
.el
g2(p, )ng'(Po)nP" ...
100\ III
101
1001 1001
(c)
= h(O«(u».
(3.5.19) o
Therefore h exhibits the conjugacy of g and 0(.
154
3 Structural stability, hyperbolicity and homoclinic points
invariant sets like A are often referred to as chaotic sets (see § 3.6) because of the presence of such orbits.
While the dynamics of glA is very complicated, we must not forget that the dynamics of glAC suggests that A is, in some sense, hyperbolic. In the following section, we consider how such sets fit into a general theoretical framework.
3.6 Hyperbolic structure and basic sets
Let us return to the hyperbolic nature of the invariant set A of the horseshoe diffeomorphism g: S2 + S2. In fact, A is said to have a hyperbolic structure or to be a hyperbolic set for g. Our aim in this section is to explain this statement and to introduce an important theorem about diffeomorphisms whose nonwandering set, n, has a hyperbolic structure.
It is helpful to review our previous encounters with hyperbolicity (see § 2.1 and 2.2). The striking feature is that, thus far, we have only had to consider hyperbolic fixed points. Nontrivial hyperbolic sets such as a hyperbolic periodic orbit, or a normally hyperbolic invariant circle, are defined in terms of a hyperbolic fixed point of a related map (fq or §' in §2.2). We are then involved with the local behaviour of a map at a fixed point in a Euclidean or Banach space. In such cases, the hyperbolic nature of the fixed point is given in terms of the eigenvalues of the derivative map (Dr or O§'). It is not possible to use this approach to characterise the hyperbolicity ofthe invariant set, A, of the horseshoe diffeomorphism. However, it is useful to consider why such an approach fails. There are two problem areas.
(i) The horseshoe diffeomorphism is defined on a manifold (the sphere) and not a Euclidean space. This means that the generalisation of the derivative map to this situation must be considered.
(ii) The complexity of A is such that it is not possible to formulate the problem in terms of a fixed point of some related map. For example, A contains aperiodic orbits which do not correspond to a fixed point of g', for any q E 7/_+ • Thus, having introduced the appropriate generalisation of the derivative map, our definition of hyperbolicity must allow for the fact that x and g(x) are different points in A.
Let us begin by considering how the results of § 2.2 can be applied to a diffeomorphism f: M + M when M is an ndimensional, differentiable manifold that is not a subset of R" (see Figure 3.22). The derivative map, Of(x*): RO + RO, used to discuss the hyperbolic fixed point, x*, of f: RO + RO in §2.2, is replaced by the tangent map TIx': TMx' + TMx" where TMx' is the tangent space to M at x*. Recall (see § 1.3) that T M x can be defined, for any x EM, in terms of equivalence classes of curves on M with the same tangent vector at x. To see the connection with the behaviour of f near x*, let ,,(t), with t el £; R, OEI and ,,(0) = x*, be a parametrised curve on M passing through x*. To find the tangent vector at x*, we need to differentiate ,,(t) with respect to t and (see § 1.1) this can only be done by using a local chart, (Va' ha) say, containing x*. The local representatives, C, ij and ~, of f, " and f.", respectively, in (Va' ha) (or, more
3.6 Hyperbolic structure and basic sets
155
concisely, arepresentatives) are given by
r .. = h .. f·h;l, They satisfy the equation
(3.6.1 )
(c':';,)a(t) = l .. (ij .. (t))
which, provided M is a CImanifold, can be differentiated to give (f;MO) = Of .. (ija(O))ifiO)
(3.6.2)
(3.6.3)
at t = O. The vectors (f .. ':'".)(O) and if .. (O) are IXrepresentatives of elements of TM x' in (V., h.). Strictly speaking, they lie in the tangent space to V. at i: = q.(O) but, as TVi: is a replica of RO, this distinction is not always apparent. The derivative map Dl.(i:) is the local representative of the tangent map Tfx" As the opening remarks to § 2.2 suggest, x* EM is said to be a hyperbolic fixed point of f: M + M if i: is a hyperbolic fixed point of f .. in the sense of Definition 2.2.1, i.e. if ol.(i:) has no eigenvalue with unit modulus. What is more, if we assign a metric to TV i' th:_n (see Exercise 2.1.2) hyperbolicity of i: corresponds to imposing bounds o~ I Df .. (i:)"vl for all v in the stable and unstable eigenspaces of Of.(i:l, Of course, it is only by making such an assignment that we can define a metric on TM" •.
Figure 3.22 For a hyperbolic periodic orbit, the Invariant Manifold Theorem gives the existence of stable and unstable manifolds in each chart for £4. The charting homeomorphisms allow them to be transferred to M (shown here as S2). E~;u are tangent to the images of W~'fu on any local chart.
156
3 Structural stability, hyperbolicity and homoclinic points
Indeed, if x* lies in the overlap of two charts (U., h.) and (U /1' hpj, then we are only allowed to choose metrics on TUx: and TUx; that, for all v E TM .', satisfy Iv.l. = 1'/11/1' where v. (v/1) is the IX (p)representative of v (see Figure 3.23). The common value defines IIvll •• for any v in TM ••. Compatible metrics, such that IIvll. is positive definite, defined at all points x of all overlaps of an atlas provides a Riemannian structure for M (see Exercise 3.6.2). If M is equipped with a Riemannian structure then we can express the hyperbolic nature of x* in a coordinatefree way by requiring that:
(i) TMx' = E!. ~ ~" where E,) is the stable (unstable) eigenspace of Tfx'; (ii) there exist c, C> 0 and 0 < Jl < 1 such that, for every nE T",
II(Tfx.)"(v)II •• < CJl"llvllx' for all vEE~., II (Tfx' )"(v)lIx' > cJl"lIvllx' for all vEE: •.
(3.6.4) (3.6.5)
With this in mind, let us consider an alternative definition of a hyperbolic periodic orbit.
Let f: M ..... M have a qperiodic orbit, Nq,= {xo, Xl>"" xqd. Of course, each point x, = fl(xo) is a fixed point of £4, but we will resist the temptation to use this to test for the hyperbolicity of Nq,. Instead, let us use the approach discussed above. The new feature in this case is that Tfx, maps TMx, to TMnxd, where f(x,) # x,, However, the Riemann structure on M allows us to deal with this change because it provides a norm, 1111 x' for every TMx, X E M. We must therefore
Figure 3.23 Schematic representation of the definition of a norm, 11'llx" on T M •• in terms of compatible norms 1·1. and 1'1, on TV i: and TV iJ' respectively.
3.6 Hyperbolic structure and basic sets
157
generalise (3.6.1 H3.6.5) as follows. Let x, and f(x,) belong to charts (U., h.) and (U 11' hpj, respectively, then
'"
(f·,,) = hp' (f.,,)
(3.6.6)
with
(3.6.7)
Differentiating with respect to t and setting t = 0 gives
vp = (f;)(O) = Dr.p(ij.(O));J.(O) = Dr.,,(i,)v •.
(3.6.8)
"
Now, V.=;J.(O) is a IXrepresentative of vETM., whilst vp=f.,,(O) is a p
representative of Tf.,(V)E TM,(x,,' Thus, the tangent map Tf .. : TM., ..... TM'(Xd and the familiar eigenspace decomposition of TMx, can no longer be used. Instead we require that, for each x,, there exist subspaces E!, and E:, such that TM., = E!, ~ E:, and Tf.,(£!:,U) = E1(~". Of course, the existence of such a decomposition is assured if x, is a hyperbolic fixed point of fq• Finally, it is important to remember that the appropriate norms must be used in the generalisations of (3.6.4) and (3.6.5), i.e.
II (Tf.,fvll,"(." < CJl"lIvllx, for all v E E~" II(Tfx,)"vll,"(x,) > cJl"llvllx, for all vEE:,.
(3.6.9) (3.6.10)
The above discussion motivates the following definition of hyperbolic structure for more general invariant sets.
Definition 3.6.1 An invariant set A is said to be hyperbolic for f (or to have a hyperbolic structure) if for each xEA the tangent space TMx splits into two linear subspaces E!, E: such that:
(i) Tfx(E!'U) = E~i~);
(ii) (3.6.9) and (3.6.10), with x, ...... x, are satisfied for all positive integers n;
(iii) the subspaces E!, E: depend continuously on X E A.
Item (iii) is trivially satisfied if A is a periodic orbit, since the points xEA are isolated. However, it is an important technical restriction for invariant sets containing a dense orbit or a dense subset of periodic orbits.
It is not difficult to accept that the invariant set A = n Q(") is a hyperbolic set
oeZ
for the horseshoe diffeomorphism g. Observe (see § 3.5.1) that the set of vertical
line segments, n Q(.', on the square Q give rise to curves on S2 analogous to
"eN
the stable manifold of a periodic orbit. Similarly, the horizontal line segments,
n Q("), lead to the analogue of the unstable manifold of the periodic orbit. At
nel+
each point X E A we can identify tangents to these curves to obtain E! and E:. Moreover, this splitting into E! and E: depends continuously on x, because, for
158
3 Structural stability, hyperbolicity and homoclinic points
3.6 Hyperbolic structure and basic sets
159
(01 unstable; O2 stable, say) but can A be decomposed into a number of basic sets? Propositions 3.5.5 and 3.5.6 show that A contains a dense orbit of g. This means that further decomposition of A is out of the question and, since it is also closed (it is a Cantor set) and invariant, the only remaining basic set (03) is A itself. Basic sets of this type are referred to as chaotic sets (see § 3.5.3 and Exercise 3.5.11).
The Anosov automorphisms provide another illustration of Theorem 3.6.1 involving a chaotic basic set. Recall that the periodic points of these maps are dense in T" and the nonwandering set is the whole torus. Moreover, in the
twodimensional example with A = G ~), discussed in § 3.4, it is clear that ~. and E:. are given by (1, (1 51/2)/2) and (I, (I + 51/2)/2), respectively, at every periodic point XI (see Exercise 3.4.2). Continuity requires that this be so for each X EO because the periodic points are dense. The splitting of the tangent space is
Figure 3.24 Illustrations of Theorem 3.6.1 where the basic sets are fixed points and periodic orbits. (a) f: SZ + SZ has nonwandering set, 0, consisting of two basic sets, 0, and O2  both fixed points. All wandering points have IXlimit set 0, and wIimit set Oz. (b) The basic sets 0, and O2 are unstable fixed points, 03 is a saddlelike 4cycle and O. is a stable 4cycle. The dynamics of the wandering points are shown schematically on the right. (c) The basic sets are all fixed points in this case: 0, is unstable; Oz is stable and 03, O. are saddlelike. Once again a schematic representation of the dynamics of the wandering points is given.
any x, x' E A, ~ and E~., (or E~ and E~.) are tangent to diffeomorphic images of paralle1line segments on the square Q. Finally, the contraction on ~ and expansion on E~ satisfy (3.6.9) and (3.6.10) with Jl > !and c = C = 1, so that A hasa hyperbolic structure.
In § 1.4, we noted that fixed points and periodic orbits are invariant sets that frequently appear to attract or repel the orbits of points not contained in them. What is more, they are rather special in so far as they have no proper subsets that are themselves invariant. The following theorem for diffeomorphisms whose nonwandering set, n, has a hyperbolic structure, provides the theoretical basis for these observations.
Theorem 3.6.1 Let f: M ~ M be a diffeomorphism on a compact manifold without boundary with a hyperbolic nonwandering set n. If the periodic points of fare dense in n, then n can be written as a disjoint union of finitely many basic sets nj, i.e.
(3.6.11 )
Each nj is closed, invariant and contains a dense orbit of r. Mo,eover, the splitting of n into basic sets is unique and M can be decomposed as a disjoint union
k
M= U in(nj),
i=l
(3.6.12)
where
(3.6.13)
is the inset of OJ.
O2
....~ ....
, If' ,
/,' : ,,\
I I I ,\
\ l : : I (a)
,\ I I I ,\ I I I
" I I / ..... ~~!. ....
0,
Diffeomorphisms with hyperbolic nonwandering set, n, and periodic orbits dense in n are usually referred to as axiomA diffeomorphisms (see Chillingworth, 1976, p. 240; Nitecki, 1971, p. 189). Clearly, any diffeomorphism whose nonwandering set consists of a finite number of fixed points or periodic orbits is axiomA. Moreover, fixed points and periodic orbits are closed, invariant sets that trivially contain a dense orbit, i.e. they are basic sets.
Theorem 3.6.1 does not merely give a decomposition of n. Equation (3.6.12) states that every x E M belongs to the inset (or equivalently, the outset) of one and only one basic set. This means that the wandering points move between the basic sets approaching those that are attracting asymptotically. Some simple examples are illustrated in Figure 3.24.
The horseshoe diffeomorphism, g, on the sphere is a more substantial example.
The nonwandering set, n, of this diffeomorphism consists of the invariant set
11.= n Q(") and the two fixed points; one stable and one unstable. n has a
tiel
hyperbolic structure and Proposition 3.5.4 shows that the periodic points of g are dense in n, so Theorem 3.6.1 applies. Obviously, each fixed point is a basic set
(c)
160
3 Structural stability, hyperbolicity and homoclinic points
therefore trivially continuous, being the same at every point of O. Hyperbolic rates of contraction and expansion follow from the hyperbolicity of A (see Exercise 3.6.4). Hence 0 has a hyperbolic structure. In this case, there is only a single basic set 01 = 0 = T 2 and it follows from Theorem 3.6.1 that the toral automorphism must have an orbit which is dense in the torus.
A further example is the transformation, I, of the solid torus, T = SI X D2, shown in Figure 3.25. The torus is treated rather like a solid rubber ring. It is stretched (with consequent loss of crosssectional area), twisted and folded to lit inside itself. Repeated application of this transformation results in longer and longer tori, wrapped around T increasingly many times. If the disc D2 is a crosssection of T,
then D2 n(OI r(T)) is a Cantor set. Thus, the colimit set of f is locally the product of a Cantor set and a onemanifold. This example has the important property that the chaotic basic set is an attractor.
The set A = n Q(O) in the horseshoe diffeomorphism has only a onedimensional oel
inset. This means that most orbits are not asymptotic to A and this makes A
Figure 3.25 (a) Illustration of a transformation f of the solid torus T
which has an attracting chaotic set. The image, f(T), of T under f is
shown shaded. (b) Intersections of successive images of T under f with
a cross section D2 of the torus. Notice that f I is not defined on the
whole torus, however, only forward iterations are required to observe
the attracting set. The mapping f is sometimes called the 'spinning diffeomorphism' .
(a)
~ ~
•
(b) 3 .6 Hyperbolic structure and basic sets
161
difficult to observe in numerical experiments. In principle, we can lind points whose orbits under g remain in a given neighbourhood of A for an arbitrary number of iterations. However, in practice most plotted orbits spend at most a few iterations near A before vanishing into the sink in F. This is because, with linite computer arithmetic, we are unable to approximate in(A) closely enough to sustain a presence near A in the face of repeated livefold expansion (see Exercise 3.5.4). Thus naive computer experiments involving the orbits of wandering points do not reveal much about the position of A let alone the dynamics on it. Some feeling for the latter aspect of g can be obtained by using symbolic dynamics (see Exercise 3.5.11).
In view of these practical difficulties, it is not surprising that chaotic basic sets are much easier to observe in numerical experiments if they are attracting. Attracting chaotic sets  often referred to as strange attractors  have been observed in a wide variety of computer experiments (see Figures 3.263.30). Detailed documentation ofthis area can be found in Gumowski & Mira, 1980; Heileman, 1980; Lichtenberg & Lieberman, 1982; Sparrow, 1982). Such attracting sets are not fully understood and may not be basic sets in the sense of Theorem 3.6.1. However, they do appear to have the common property that points in them occur on liner and finer scales. For example, in Figure 3.27, the Henon attractor appears to be onedimensional and to consist of a number of segments. Closer examination reveals that each 'segment' consists of several closely spaced curves of similar shape
Figure 3.26 The Duffing altractor (see Guckenheimer & Holmes, 1983, pp, 8291 & 1913). The Duffing equation can be written in the form
x=y,
y=xx3 +&(a cos 0 by),
This system is periodic in 6 and the phase space can be taken as M = R2 X SI. Every surface 0 = constant is a global Poincare section so that the system behaviour is completely described by the Poincare map P",. Numerical approximations to P,.8 appear to have a chaotic attracting set  the Euler approximation is shown in this diagram for to = 0.4 and &b = 0.25. The structure of P e.' is discussed in greater detail in § 3.8.
y
3.6 Hyperbolic structure and basic sets
163
162
3 Structural stability, hyperbolicity and homoc,Iinic points
Figure 3.27 The H enon attractor (see Henon, 1976). The map f producing this attractor is defined by
(x, y)t!. (y  ax2 + I, bx),
where a, be R. The attracting set for a = 1.4 and b = OJ is shown. It arises from the repeated folding and stretching brought about by the action of f. When magnified the attractor is found to consist of many curves, of similar shape to those resolved above, occurring very close together. This 'braided' nature of the attractor appears to be repeated on all scales.
Figure 3.29 The Lorenz attractor (Lorenz, 1963). It must be pointed out that there is a theorem corresponding to Theorem 3.6.1 for flows so that strange attracting sets can also arise in flows that are not the suspension of a diffeomorphism. The Lorenz equations
X= 10(yx),
y=x(28z)y,
i = xy  (8/3)z,
have fixed points at (±6(21/2), ±6(21/2), 27). The system does not have a global section so the projection onto the xzplane is shown. The orbit generated by using the Euler method with step length of 0.005 and initial point (x, y, z) = (0.1, 0, 0) is plotted. The projected orbit switches between revolving about (x, z) = (+6(21/2),27) and (x, z) = (_6(21/2),27) in an apparently random way.
/ .;:::;.::'~~"'/
.... . ~ ~ ... _ ..... .~
_
 ... .._
__________ L +~ __ ~ x
Figure 3.28 Guckenheimer et al. (1977) discussed a Leslie model of a density dependent population with two age classes of size x and y. It is a discretetime model and the dynamics of the two classes are represented by the map
(x, y) .... (r[x + y] exp(  [x + y]/IO), x),
where r is a real parameter. The map appears to exhibit chaotic behaviour for r ~ 17 and a typical orbit for r = 20 is shown here. More numerical details can be found in Guckenheimer et al. (1977), where the origin of the attracting set is discussed in terms of a twisted horseshoe map.
Figure 3.30 The Rossler attractor (Rossler, 1979). This is another threedimensional Dow exhibiting an attracting set with complex dynamics. The system equations are
X= (y+z), y=x+ey, i=! +xzJlz.
A perspective view of an orbit near to the attractor is shown for e = 0.17, ! = 0.4 and Jl = 8.5. It is obtained by using the Euler method with step length 0.005 to approximate the trajectory through (x, y, z) = (1,0,0) and plotting u = x + y, v= Y + z.
y
v
roo . ,
164
3 Structural stability, hyperbolicity and homoclinic points
that are not resolved on the scale of Figure 3.27. Further magnification shows that each of the latter 'curves' has a similar structure, and so on. The attractor is said to have a 'braided' nature which is repeated on all scales. The reader will recall that the chaotic basic set A of the horseshoe diffeomorphism has this property (see Exercise 3.5.9). In fact, a theoretical connection with the basic set of the horseshoe diffeomorphism can be made in some cases: namely when homoclinic points occur.
3.7 HomocIinic points
We have seen that homocIinic points occur in the dynamics of Anosov automorphisms. They also occur in the horseshoe map. Consider for example the fixed point represented by the sequence { ... 1111·1111. ... }. The stable manifold of this point on Q is a vertical line segment and the unstable manifold is a horizontal line segment. The effect of a single iteration of the horseshoe map f of § 3.5.1 is shown in Figure 3.31. Clearly, transverse homocJinic points must occur. Are homoclinic points a feature of chaotic basic sets? The following theorern provides a partial answer to this question.
Let M be a compact twomanifold and Diffl(M) be the set of all Cl diffeomorphisms on M. The elements of a residual subset of Diffl(M) have the property that all their fixed and periodic points are hyperbolic and all intersections of stable and unstable manifolds are transverse. Diffeomorphisms in this subset are usually referred to as KupkaSmale dif.feomorphisms (see Chillingworth, 1976; p. 227; Nitecki, 1971, p. 83).
Figure 3.31 A transverse homoclinic point xt of the fixed point x* = { ... 11·11 ... } of the horseshoe map.
x*
WU(x*)
3.7 Homoclinic points
165
Theorem 3.7.1 (SmaleBirkhoff) Let fEDiffl(M) be KupkaSmale and xt be a transverse homoclinic point of a periodic point x* of f. Then there is a closed subset A of O(f), containing x t, such that:
(i) A is a Cantor set;
(ii) fP(A) = A for some pEl+;
(iii) (P restricted to A is topologically conjugate to a shift on two symbols.
A point x t is a homocIinic point of a periodic point x* of period q if it lies at an intersection (#x*) of the stable and unstable manifolds of the fixed point of rat x*.
The idea behind the proof of Theorem 3.7.1 is illustrated in Figure 3.32. If the stable and unstable manifolds of the hyperbolic saddle point x* intersect at some point x], then they must intersect infinitely many times. Recall (see § 3.4) if x] E WIn W· then f'"(x1)E WIn W· for every mEl. Figure 3.32 illustrates the effect this constraint has on the two manifolds if we attempt to return them directly to x* itself. As the unstable manifold approaches the saddle point the loops between adjacent homocJinic points are stretched parallel to Wioc and squeezed parallel to Wloc. The manifold therefore undergoes oscillations of increasing amplitude and decreasing period. The fate of the stable manifold is similar under reverse iterations resulting in the homoclinic tangle shown in Figure 3.32.
The connection with shifts on twosymbol sequences is apparent if we consider the images of a small 'parallelogram' R, containing xl and with sides parallel to W' and W·, under forward and reverse iterations. For m > 0, the mth iteration of f stretches f< .. ll(R) along w· and contracts it along Ws. Remember, f< .. ll(x1} is a homocJinic point and belongs to r .. ll(R) for every m. Eventually, for some N E r', (N(R), takes the horseshoe shape Rl (see Figure 3.32). For reverse iterations
Figure 3.32 Illustration ofthe homoc1inic tangle occurring at a hyperbolic saddle point. The parallelogram R has images R I = fll(R) and Ro = r II· (R) intersecting in a horseshoe configuration.
166
3 Structural stability, hyperbolicity and homoclinic points
the roles of W' and W· are reversed and, for some N' E I+, (tN')(R) = Ro, where RI and Ro intersect as shown in Figure 3.32. Clearly, if p = N + N', f"(Ro) = RI and we would expect fP to exhibit horseshoelike behaviour, i.e. be conjugate to a left shift on twosymbols. The homoclinic point referred to in Theorem 3.7.1 would in this case be xt =rN'(x1).
Theorem 3.7.1 means that f exhibits all the complexity of the left shift IX::E +:E discussed in § 3.5.2. In particular, in every neighbourhood of a transverse homoclinic point of f, there is a periodic point. By Theorem 3.7.1, the transverse homoclinic point xt E A and fPIA is topologically conjugate to the left shift IX: :E + :E. However, by Proposition 3.5.4 the periodic points of IX are dense in :E. Hence, periodic points of fPIA are dense in A and, therefore, there is a periodic point of f arbitrarily close to x t. Thus there are infinitely many periodic points in any neighbourhood of xt.
It is important to realise that Theorem 3.7.1 employs sufficient conditions to ensure the existence of A. As Smale has pointed out (see Smale, 1963), we might expect a similar result to hold with weaker constraints on r. Figure 3.32 suggests that the key requirement is that the stable and unstable manifolds of a hyperbolic fixed or periodic point intersect transversely. With this in mind, the following example shows that the remarkable phenomena described above really do occur. Let us examine the planar map
XI =X+ YI'
YI=y+kx(xl),
(3.7.1)
numerically, for 0 < k < 4. This map has fixed points at (x, y) = (0, 0) and (1,0) for all values of k. The fixed point at (0, 0) is nonhyperbolic. The linear approximation to (3. 7.1) at (0, 0) is conjugate to an anticlockwise rotation through angle e, where
2 sin e = [k(4  k)]1/2,
2 cos e = (2  k).
(3.7.2)
Linearisation at (1,0) shows that this fixed point is a hyperbolic saddle point with £(1,0) and E('I.o) given by
v = u{  k  [k(4 + k)]1/2}/2 and l' = u{  k + [k(4 + k)]1/2}/2, (3.7.3)
respectively, where (u, v) are local coordinates at (1,0). It is not difficult to then use a microcomputer to plot successive images of, say, one hundred points lying in a small interval of E;\.o) close to (1,0). The result of such a calculation can be quite spectacular (see Figure 3.33(a». Of course, a suitable interval along Eh,o) can be iterated, using the inverse map
X=XI YI'
y = YI  kx(x  I),
(3.7.4)
to complete the homoclinic tangle (see Figure 3.33(b». Some uses of a computer program of this kind are suggested in Exercises 3.7.3 and 3.7.4.
It is important to understand how the contortions of the stable and unstable manifolds influence the orbits of wandering points of (3.7.1). It is tempting to imagine that the latter also undergo wild oscillations but this is not the case. For example, the behaviour of (3.7.1) in the neighbourhood of the saddle point is
3.7 Homoclinic points
167
determined by Hartman's Theorem. Thus, since the eigenvalues of the linearisation at (1,0) are both positive (see Exercise 3.7.3), the orbits of individual points pass the saddle point as shown in Figure 2.1 (e).
It is easily shown that the derivative map of (3.7.1) has positive determinant for all (x, Y)E R2 (see Exercise 3.7.5). A diffeomorphism, f: Ie + Rl, with this property is said to be orientationpreserving (see Chillingworth, 1976, p. 139). A planar closed curve y can be oriented in two ways depending on whether an observer walking along the oriented curve finds the region enclosed by y on his right or lefthand side. When Det(Df(x»>O for all xER1, it can be proved (see Exercise 3.7.6) that the orientation of the image of y under f must be the same as that of y. Now consider the closed region So, with boundary Yo, shown in Figure 3.34(a) and let Yo be oriented according to the sense of description of the unstable manifold. It follows that the image of So under (3.7.1) must be one of the regions
Figure 3.33 (a) Approximation to the unstable manifold of a hyperbolic saddle point of the planar map (3.7.I)at (1, 0) for k = 1.5. (b) Homoclinic tangle for (3.7.1) obtained by adding to (a) an approximation to the stable manifold at (1,0). The latter is obtained by reverse iteration of a small interval of £(1.0) close to (1,0) (see Exercise 3.7.3).
(a)
(b)
168
3 Structural stability, hyperbolicity and homoclinic points
Figure 3.34 (a) Plot of the stable and unstable manifolds of the saddle point at (1,0) of (3.7.1). Since the map is orientationpreserving, t.he image of the manifold loop So must be one of loops Si' i = 1, 2, : .. , "':Ith the same orientation as So. It is not difficult to see that the orientation of the loops Si is opposite to that of So. In fact, for (3.7.1), the image of S, is Si+ I (see Figure 3.34(c» but this is not the case .in general. For example, the image of S, under the square of (3.7.1) IS Sj+2' (b) The images of So under iterates of the inverse of (3.7.1) are the regions Si which wrap further around the fixed point at (0,0) as i increases. (c) Numerical plot of the orbit ofthe point P = (0.64, 0.094) under (3.7.1). It sweeps around (0,0) twice, passing near to the saddle point on each occasion, before arriving in So at the fifteenth iteration. Subsequent iterates are carried away to infinity under the influence of the saddle point. Note that, since the manifold loops become extremely narrow and close together, the number of revolutions of the orbit about (0,0) before expulsion to infinity can depend sensitively on the choice of initial point.
(a)
(b)
(c)
3.7 Homoclinic points
169
Sj, i = 1,2, ... , with the same orientation as So, and not one of Sj' i = 0,1,2, ... , for which the orientation is reversed. Thus points in So are ultimately swept off to infinity under the influence of the saddle point at (1,0). Similarly, points in So are swept around the fixed point at (0,0) and fed back into the vicinity of the saddle point once again. The role of this movement about (0,0) in the dynamics of (3.7.1) is best understood by considering images of So under powers of the inverse map. The preimages of So are a subset of the regions S_j, i = 1,2, ... , shown in Figure 3.34(b). Observe that, as i increases, these regions stretch further around (0,0). Indeed, for each N E 1. +, there is an i(N) such that S_j(N) wraps around (0,0) N times. It follows that there are points in S _ j(N) whose orbit makes N trips around (0,0) before it appears in So and subsequently sweeps out to infinity. It is not difficult to confirm these ideas numerically. An orbit exhibiting this behaviour is shown in Figure 3.34(c).
Similar orbits were shown in Figure 1.39 and 1.40 for the Henon areapreserving map. This is no coincidence. Henon has shown (see Henon, 1969) that every quadratic, areapreserving, planar map, with rotational linear part at the origin, :is conjugate to the form (1.9.40). It is easily verified that the derivative of the map (3.7.1) has unit determinant for all (x, Y)ER2 (see Exercise 3.7.5». Thus (3.7.1) and (1.9.40) must exhibit the same dynamics. For our present purpose, (3.7.1) has the advantage that the saddle point remains at (1,0) for all k, so that E('I,o) and £(1.0) are easily calculated.
In the above discussion, we have assumed that the stable and unstable manifolds that intersect one another come from a single fixed point x*. Recall that Theorem 3.7.1 includes the case where the stable and unstable manifolds involved are associated with a fixed point of fq• Similarly, if x* is a periodic point of period greater than one, then, for example, the unstable manifold of x* may intersect transversely with the stable manifold of f(x*) (see Figure 3.35(a». Once again, the manifolds oscillate wildly because images of homoc1inic points are homoclinic points. Given that the unstable manifold of f(x*) also intersects the stable manifold of x* transversely, then consideration of the images under f of a suitable parallelogram, R, again indicates that some power of f behaves like a horseshoe map (see Figure 3.35(b».
This construction is also relevant to quadratic, areapreserving maps of the plane. Suppose x* has periodq and homoc1inic points arise in the manner described above at each point of the periodic orbit, i.e. in the above argument x* H flmI)(x*) and f(X*)Hf'"'(X*), m' = m mod q, for m = I, ... , q. Then we obtain a chain of homoclinic tangles as shown in Figure 3.36. In this case, the orbit of a point such as P in this figure could sweep around the whole periodic orbit before being fed back into the vicinity of x* at a different point, P'. Because of the massive stretching along the unstable manifold at each periodic point, the position of P' depends sensitively on that of P.
There is evidence of this kind of behaviour in the maps (1.9.40) and (3.7.1). The 'twodimensional' orbits shown in Figures 1.41 and 1.42 are associated with a hyperbolic periodic orbit, they are generated by iterating a single point and their
170
3 Structural stability, hyperbolicity and homoclinic points
extent is similar to that of the expected homoclinic tangles (see Gumowski & Mira, 1980, p. 303). In this situation, there is a good reason (see Figure 6.17) why orbits of this kind do not escape from the influence of the periodic orbit. Therefore, the plotted iterates of a single point appear to fill out the twodimensional region in an apparently random way.
3.8 The Melnikov function
In this section we describe a method for proving that transverse homoc1inic points occur in the Poincare maps of certain types of flow in three dimensions. This
Figure 3.35 (a) Illustration of the unstable manifold of the periodic point x* intersecting the stable manifold of f(x*) transversely at xl and hence at infinitely many other homoclinic points. (b) The parallelogram R is iterated forward to R. and in reverse to R2. The map from R2 to R. is
horseshoelike. •
(a)
xl
R
(b)
3 J3 The M elnikov jUnction
171
method is particularly interesting here because it can be applied to the Duffing equation which appears, numerically, to have a chaotic, attracting set (see Figure 3.26).
Consider the planar differential equation
(3.8.1 )
which has a hyperbolic saddle point at x = 0 and assume there is a homocIinic saddle connection, r, as shown in Figure 3.37. Now consider the product flow in H2 x SI defined by
(3.8.2)
The saddle point of (3.8.1) at x = 0 E H2 becomes a periodic orbit Yo = {(x, O)E H2 x Sllx = 0, OE Sl} of saddle type. Moreover, the unstable manifold of Yo, WV(Yo), intersects the stable manifold, W'(Yo), in the cylindrical surface
Figure 3.36 Chain of homoclinic tangles that can arise on a hyperbolic periodic orbit.
Figure 3.37 Phase portrait for x = fo(x). The origin is a hyperbolic saddle point and r is a homoclinic saddle connection.
(3.8.3)
3.8 The Melnikov jUnction
173
172
3 Structural stability, hyperbolicity and homoclinic points
r x SI £ H2 X SI. This behaviour is nongeneric. In particular, the stable and unstable manifolds of the corresponding fixed point of the Poincare map, Po, of (3.8.2) do not intersect transversely. The Melnikov method applies to small perturbations of (3.8.2) of the form
each curve there will be a unique point of intersection AU.', closest to Xo (see Figure 3.39). Let (xu.S(t; 00, e), r), tEn, be the unique trajectory of (3.8.3) passing through AU" at t = 00, i.e. AU" is the point xU"(Oo; 00, e)E LB' We then define the
timedependent distance function, 0
(3.8.4)
with eE H+ and fl (x, 0) = fl (x, 0 + 2n). For sufficiently small s, it follows from Proposition 3.2.2 that (3.8.3) also has a hyperbolic periodic orbit, '1'., close to Yo· However, the invariant manifolds, WU(Y.) and W'(y,), need not intersect to form a cylinder (see Figure 3.38). The Melnikov function is related to the 'distance' between these two manifolds.
Let Xo E JR2 be a point of the saddle connection r in the unperturbed system (3.8.1). Take a perpendicular section L to the saddle connection at Xo. We use the point Xo and the section L in the 0 = Ooplane, LBo' as follows. Consider the perturbed system and the intersections of '1'" WU(y,) and W'(y,) with L80 This is equivalent to studying the Poincare map P •. Bo: L80 ~ LBo of the flow (3.8.3). P,.B. will have a hyperbolic saddle point, X:'8o' near to x = 0, with stable and unstable manifolds, WU·'(x:".) = Wu.S(Y,)n Ls., which are close to r on LB. (see Figure 3.39).
The distance between WU(y.) and W'(y,) on Ls. is calculated along L. Observe that this distance will, in general, change with 00, since & > 0 implies that the curves WU(x:'8o) and W'(x:'s.) will be Oodependent. Obviously, for the special case e = 0, the distance would be zero for all values of 00,
Of course, the manifolds wu.S(x:'8o) may intersect L many times, however, on
where xo(t) is the homoc1inic trajectory of (3.8.1) with xo(O) = xo. In (3.8.4) the wedge product is defined by a A b=alb2 a2bl where a, bEIR2 have Cartesian coordinates (ai' a2) and (bl, b2), respectively. It follows that ~,(t, 00) is Ifo(xo(t  80))1 times the component of the vector [XU(t; 00, e)  x'(t; 00, e)] perpendicular to fo(xo(t  00)), The latter vector is, of course, tangent to r at xo(t  00), Thus, ~.(Oo, 0o)flfo(xo)1 is the distance between WU(y.) and W'(Y.) measured along L on LB.
We can obtain a useful form for ~.(80, 00) by studying (3.8.4) more closely. Let
XU(t; 00, e) = xo(t  Oo} + exHt, 00) + 0(1:2) (3.8.5)
and
(3.8.6)
where x!, xi are first variations with respect to e. Thus, (see Exercise 3.8.1)
x~·'(t, 00) = Dfo(xo(t  Oo))x~·S(t, 00) + fl(xo(t  00), t). (3.8.7)
Now define
Figure 3.38 The manifolds WU(y.)and W'(,,) for (a)£ = 0 and (b)£ > O.
~:,'(t, 00) = fo(xo(t  00)) A ex~·S(t, 00), so that ~.(t, 00) in (3.8.4) can be written in the form
(3.8.8)
(3.8.9)
Figure 3.39 The intersections of y" WU(y,) and W'(y.) with rOo for £ = 0 and £>0.
(a)
(h)
174
3 Structural stability, hyperbolicity and homoclinic points
We can obtain differential equations for A~ and A:. It can be shown that, since xo(t  (0) = fo(xo(t  (0))'
A:(t, (0) = e[Tr(Dfo(xo(t  Oo)))fo(xo(t  (0)) " xW, (0)
(3.8.10)
The expression (3.8.10) is greatly simplified if fo is a Hamiltonian vector field, as it is for the Duffing equation, for then Tr(Dfo(x)) == 0 (see (1.9.21 and 24» and
A:(t, (0) = efo(xo(t  (0)) " fl (xo(t  (0)' t). (3.8.11)
Integration of (3.8.11) from t =  00 to t = 00 gives
A~(Oo, (0) = e fo", fo(xo(t  (0)) " fl (xo(t  (0), t) dt.
(3.8.12)
Here we have noted that A:(  00, (0) = 0 because xo(  (0) = 0 = fo(O). A similar calculation leads to
A:(Oo, (0) = £ r'" fo(xo(t  (0)) " fl (xo(t  (0)' t) dt.
J80
(3.8.13)
and therefore,
A.(Oo, (0) = e f~", fo(xo(t  (0» " r, (xo(t  (0), t) dt + 0(£2). (3.8.14)
Finally, we define the Melnikov function, M(Oo), by
M(Oo) = f~", fo(xo(t  (0))" fl(xo(t  (0)' t) dt,
(3.8.15)
so that
(3.8.16)
Proposition 3.8.1 If M(Oo) has simple zeroes, then, for sufficiently small £ > 0, W"(x* ) and W·(x· ) intersect transversely for some 00 E [0, 27t). On the other
•• 80 e, 80 10
hand, if M(Oo) is bounded away from zero, then W"(x:'80)rI W'(x:"o) = 0 for al o·
In allowing 00 to vary, we are effectively taking a fixed reference point Xo ~nd section L, perpendicular to fo(xo), in each section 1:80, 00E[0,27t). By taking s sufficiently small, A.(Oo, (0) is an arbitrarily small perturbation of. eM(Oo)' It follows that if eM(Oo) has a simple zero then so does A.(Oo, (0), ThIS ~eans that there is a value, e, of 00 at which A.(Oo, (0) changes sign, corresponding to x"(Oo; 00' e)  x'(Oo; °0, e) reversing its orientation relative to fo(xo)· Clearly x"(e; e, e) =x'(e; e, e) and, therefore, the manifolds WU(x:'s) and W'(x:'s) of the fixed point x:'s of the Poincare map p •. s intersect transversely on L near to
3.8 The Melnikov function
175
xo. Of course, all the Poincare maps P"'o' 00 E [0, 21t), are topologically conjugate (see Exercise (1.7.3)) and, consequently, WU(x:',) and W'(x:',) must intersect transversely for all 00 E [0, 27t) (although, obviously, not always near to Xo (see Figure 3.40). Equally, if M(Oo) is bounded away from zero, then, for sufficiently
Figure 3.40 The manifolds W"(y.) and W'(y,) intersect in a homoclinic trajectory that ultimately approaches ,/, as t ..... 00. When 11.(00, (0) has simple zeroes, this trajectory passes through the section Lx [0, 2n) at least twice. An impression of the nature of the homoclinic trajectory can be gained by recalling that 0 = 2n is to be identified with 0 = O. Thus ~he s~gment BoBI continues as B, B2 and CoC, as c, C2. The trajectory Itself IS U.eZ (B._IB.)u(C._IC.). It follows that corresponding pairs of zeroes occurfor any choice OfxoE r. Moreover, if M(Oo) is bounded away from zero on [0,211), then it is so, independently of the choice of xo, and no homoclinic points occur. For given 00, the stable and unstable manifolds of the fixed point x:',. of the Poincare map P ,.60 are obtained by taking the corresponding section in this figure.
80 = 2n
E8 tE· .... ···~
"
Bj +,
89······:·~
"
(3.8.20)
3.8 The Melnikov funcuon
177
176
3 Structural stability, hyperbolicity and homoclinic points
small e, so is 6.(00, (0), This, in turn, means that transverse homociinic points do not occur on L for any 00 E [0, 2n). As Figure 3.40 shows, this conclusion does not depend on the choice of Xo E r through which L passes. Hence, there are no homoclinic points.
Comparison of (3.8.17) with (3.8.3) gives
fl(x, 0) = (0, a cos 0  by)T,
(3.8.22)
x=y,
y=x _x3 + e(a cos 0 by),
(3.8.17)
which satisfies fl (x, 0) = fl(x, 0 + 2n). It follows (from (3.8.15» that the Melnikov function for the homoclinic orbit r6 is
M(Oo) =  21/2 f_oooo sech(t  (0) tanh(t  Oo)[a cos(t) + 21/2b sech(t  (0) tanh(t  (0)] dt.
(3.8.23)
Example 3.8.1 Show that the Poincare map of the Duffing equation
a, b > 0, has transverse homoclinic points, for sufficiently small values of 6, provided
x=y,
(3.8.19)
The change of variable of integration t 1+ t  00 gives
M(Oo) =  21/2a f:oo sechttr) tanh(t) cos(t + (0) dt
 2b f_~., sech2(t) tanh2(t) dt.
(3.8.24)
a 4 cosh(n/2)
>_._
b 3(21/2)n'
(3.8.18)
Solution. When e = 0, (3.8.17) becomes
so that fo(x) = (y, x  x3r The differential equation i = fo(x) has a hyperbolic saddle point at x = 0 and two further fixed points at x = (± I, Or It is a Hamiltonian system with
H(x, y) = !(l x2 + X24) and the level set of H(x, y) = o consists of two homoclinic orbits, rt, and the saddle point at x = 0 (see Figure 3.41). It can be shown (see Exercise 3.8.2) that the trajectories passing through (x, y) = (± (21/2), 0) at t = 0 are given by
(x± (t), y± (t» = (± (21/2) sech t, + (21/2) sech t tanh t). (3.8.21)
The latter integral is easily evaluated, while the former can be simplified by writing cos(t + (0) = cos(t) cos(Oo)  sin(t) sin(Oo) and noting that
f:oo sech(t) tanh(t) cos(t) dt = 0 (3.8.25)
because the integrand is an odd function of t. Thus,
M(Oo) = 21/2a sin(Oo) f 00 sech(t) tanh(t) sin(t) dt _ 4b.
_00 3
(3.8.26)
The integral occurring in (3.8.26) can be evaluated using the method of residues (see Exercise 3.8.4)) and we finally obtain
Figure 3.41 The phase portrait for the planar system x = fo(x), fo(x) = (y, x  X3)T. Stable and unstable manifolds of saddle point x = o coincide to form a pair of homoclinic orbits r6'. The level set H(x, y) = 0 is rov{O}vr;.
4b (n)
M(Ool = 3 + 21/2na sech "2 sin(Oo)·
(3.8.27)
(2~, 0)
Clearly, if (3.8.18) is satisfied M(Oo) has simple zeroes and, by Proposition 3.8.1, transverse homoclinic points must occur. On the other hand, if the reverse inequality is satisfied, M(Ool is bounded away from zero and Proposition 3.8.1 implies that there are no homociinic points. 0
(2~, 0)
There is one remaining possibility for the system (3.8.17): namely that a 4 cosh(n/2)
b = 3(21/2)n .
(3.8.28)
In this case, M(Oo) has a double zero at 00 = 3n/2. This corresponds to WU(x* )
• * . . •. 31</2
and W (X e, 3,,/2) meeting tangentially rather than transversely. As before, the orbit
of such a homoclinic point under P •. 3./2 consists entirely of tangential intersections
178 3 Structural stability, hyperbolicity and homoclinic points
Figure 3.42 (After Veda, in Guckenheimer & Holmes, 1983, p.192.) Stable and unstable manifolds for the Poincare map of the Duffing equation (3.8.17) with &b = 0.25 and (a) to = 0.11; (b) to = 0.19; (c) to = 0.30. Observe the tangency of the stable and unstable manifolds in
(b).
y
x
2
x
o N ~ I o N
o N I
.." ..... ~ ......
:(;;. )
~\
C?

I
o N I
~o N
o
...:
o N I
180
3.3.2
3 Structural stability, hyperbolicity and homoclinic points
of WU(X:'3x/1) and W'(x:'3xd. Moreover, since p e, so and p •. e• are topologically conjugate for all 80 and 8~, these homoclinic tangencies occur in all p e, Bo' Ueda (see Guckenheimer & Holmes, 1983, p. 192) has computed stable and unstable manifolds for the hyperbolic saddle point of the Poincare map of (3.8.17) and some of his results are reproduced in Figure 3.42. It is not difficult to verify that the value of alb at which homoclinic tangencies occur numerically (see Figure 3.42(b)), is in close agreement with (3.8.28).
The occurrence of homoclinic tangencies has important repercussions the details of which are beyond the scope of this text. Newhouse (1979, 1980) has shown that if such a tangency occurs at XI for fe Diff'(R2), then there is an f eC'close to f for which tangencies also occur stably in a hyperbolic invariant set. This set lies near to the orbit of XI and is known as a wild hyperbolic set. f also has an infinite number of stable periodic orbits  or 'infinitely many sinks'  as the title of Newhouse's original paper had it. We refer the reader to Guckenheimer & Holmes, 1983, pp. 331·40 for a more detailed description of these ideas. However, this kind of behaviour may occur in p e, e for alb near to the critical value (3.8.28).
As we have already noted (see Figure 3.26), numerical approximations to (3.8.17) exhibit a complicated attracting set. Such a set appears even in the Euler approximation and it is then not difficult to verify that alb must exceed a critical value before it appears. This suggests a connection between the attracting set and the occurrence of homoclinic points. Indeed, the careful numerical work of Ueda (in Guckenheimer & Holmes, 1983, p. 90) (see Figure 3.43) has led to the conjecture that the attractor is the closure of the unstable manifold of the saddle point. While this can be justified for ajb less than the value in (3.8.28) (Guckenheimer & Holmes, 1983, p. 91), the situation is more complicated when homoclinic points are present.
Exercises
3.1 Structural stability of linear systems
3.t.l Consider a real, n x n matrix, A, with eigenvalues AI> ••. , A. that are not necessarily distinct. Let B, with eigenvalues 111>"" II", be sclose to A in L(R"). The spectral variation of B with respect to A is defined by
SA(B) = max[min(IAj II)].
} I
(E3.I)
(a) Assume that A can be diagonalised and show that SA(B)<t.
(b) Suppose A cannot be diagonalised and show that SA(B) < (nt)I/"
provided t'< I/n.
(c) Deduce that {II" ... ,II"} + {AI"'" A.} as e +0 for any AeL(R").
3.1.2 Let SD(R") be the subset of structurally stable linear diffeomorphisms in L(R").
Show that a linear diffeomorphism is structurally stable if and only if it is hyperbolic. Hence, or otherwise, show that SD(R") is open and dense in L(R").
Exercises
181
3.1.3
~t S be the subspace of L(R") defined by {(~ ~)IA # 0 or I}. Show that every linear diffeomorphism in S is structurally stable within S but not within L(R2). Consider the subspace, 0(R2), of L(Rl) defined by (AIAT A = I, A e L(R2)}. Show that no element of 0(R2) is structurally stable in L(R2).
3.1.4
3.2.1
3.2 Local structural stability
Le~ the ve;tor field X(x)eVecl(U), U_!;; Rn. and open, have a hyperbolic fixed pomt a~.x = Oe U and sup~se that X(x) I~ an eC'perturbation of X. Verify Proposition 3.2.1 for the special case when X  X is (a) constant· (b) linear (c)
O(!xlt), k ~ 2. ' ,
Find II = II(t) such that each of the following vector fields is eClclose, on U = {(9, r)lr <2}, to;= r(lr), 0= I;
(a) ;=r(1 +IIr), 0= I;
(b) r=r(lr+llr2), 0= I;
(c) 1'= (I +II)r(lr), 0= 1.
Verif~ that the flows (a)(c) all have a hyperbolic periodic orbit near to r = I for sufficiently small values of s,
Show that the nontrivial fixed point x* = (elf, alW of the VolterraLotka vector field
X(x)=«aby)x, (efx)y)T, (E3.2)
a, b, c, f > 0, is nonhyperbolic. Find a first integral for the system x = X(x) and determine the topological type of x*.
Consider vector fields of the form X + Xd on a disc of radius R> Ix*l, where: (a) Xd=(ch, _by)T;
(b) X, = (c5x2, O)T.
Choose c5 in each case such that IIX611, < e. If e is sufficiently small, show that X + Xd has a fixed point y* near to x* for both perturbations but that the topolo~ical type ofy* i.s the same as x* for (a), while it is different for (b). Explain why this result IS consistent with Proposition 3.2.1?
3.2.2
3.2.3
3.3.1
3.3 Flows on twodimensional manifolds
All of the following vector fields are structurally unstable on R2. To which of these .examples does Theorem 3.3.1 apply? Use the theorem, where applicable, to explain the nature of the instability. For the remaining examples construct tC'close s!stems t~ exhibit their structural instability in Vec'(!:iJ), where !:iJ is the closed disc of radius 2 centred on the origin.
(a) ;= r(r1)2, 9= I;
(b) ,,= r(1  r), 0 = sin2(O);
(c) X= 2y(lx2)+xB(x), y=2x(l y2)+ yB(y);
where B(x) =exp{ _x2/(1  x2)} for Ixl < I and =0 for [x] ~ I.
The flows'll,: R2 t R2 of the following systems give rise to flows on the torus T2. = {(~I' 92)10,;;; 9" 92 < I} by taking mod I in both components of'll,. Use PelX?to ~ Theorem to show that these toral flows are structurally unstable.
(a) x = slO(2Itx), y = 0;
(b) x = I, Y = 2.
Illustrate these instabilities by giving topologically distinct systems which are eC'<close for arbitrarily small e.
182
3 Structural stability, hyperbolicity and homoclinic points
3.3.3
Consider the system on the cylinder C = {(II, r)IO ~ 11< 2n, re R}, given by
f = _r_ cos(2nr), 11 = I. (I + r2)
Show that it is structurally stable on every set S. = {(II, r)l n ~ r ~ n}, n e Z+ , but that it is not structurally stable on C.
(E3.3)
3.4.1
3.4 Anosov diffeomorphisms
Use Peixoto's Theorem to show that none of the following dilTeomorphisms Ie DilTl (81) are structurally stable:
(a) 1(11)= (II +a)modl, exeO;
(b) 1(11)= (1I+a)modl,exeA\O;
(c) 1(11) = (II +sin2(2nll»modi ;
(d) 1(11)= (II +}+ 0.1 sin2(2nll»modI.
Show that all of the periodic points of an Anosov automorphism f: TO + TO are hyperbolic. Let 0* be a periodic point of f o~ ~riod ~e 1+. <?i~e ex~ressions for the stable and unstable manifolds of the periodic orbit contammg 0 .
Let the Anosov automorphism f: T2 + T2 have lift A: A2 + IIi given by (3.4.12). Prove that f has periodic points of every prime periodq by showing tha~ there exists xeA2 such that A'xxeZ", where x does not represent a fixed point.
Let f and g be Anosov automorphisms of TO with lifts given by aut~morp~isms A, 8: AO + no. Prove that A and 8 are similar matrices if f and g are dllTe~entIably conjugate. Conversely, show that, if A, B are lifts of Anosov auto~orphlsms f, g and they are similar by a matrix C with integer entries and determinant ± I, then f and g are dilTerentiably conjugate.
Show that the Anosov automorphism f induced by
3.4.2
3.4.3
3.4.4
3.4.5
A = G ~)= A2 + A2
has a saddle point at n(O), where n: R2 + T2 is the map given in ~3.4.~). Find the equation of the separatrices of A at 0 and show that they have l.rratIOnal slope. What are the implications of this on the torus? Show that the point on the torus given byn(xt), where xt = Anosov auto(131/2  1 )/2(131/2), 1/131/2): is a.transv~rse homoclinic point. How does the Anosov automorphism considered m this question dilTer from that given by (3.4.12)?
(E3.4)
3.5.1
3.5 Horseshoe diffeomorphisms 3.5.1 The canonical example
Obtain explicit equations for the horseshoe map f: Q + R2 on its restriction to
r
POuPI cQ (see Figure 3.13). Verify that (X,y)H(fI(X),f2(X,y)).
Let A be the Cantor set of the horseshoe diffeomorphism f as defined in (3.5.4). Complete the proof of Proposition 3.5.1 by showing that A is invariant under r '. Hence show that r(A) = A.
3.5.2
Exercises
183
3.5.3
Let g: 82 + 82 be the globally extended horseshoe dilTeomorphism. Prove that: (a) n Q(o, is invariant under g and n Q(.' is invariant under gI;
"EN ReZ.
(b) n Q(O' (n Q(O') is the inset (outset) of A for g on Q.
"EN "5Z+
3.5.4
Suppose g: Q + R2 is the horseshoe map and take Xo, with coordinates (xo, Yo), to be a point of Q\A. Let d = d(xo, A) = min {lxo  xl} be the horizontal distance (%.1,e,\
from Xo to A. Find the maximum value, N(d), of n such that f"(Xo)eQ.
3.5.2 Dynamics 00 symbol sequences
Let Is be the set ofall biinfinite sequences on the m symbols 8 = {O, I, ... , m  I} and a: Is + Is be the left shift a(I1). = 1101 (see (3.5.7». Prove that:
(a) ex has .periodic orbits of all periods as well as aperiodic orbits;
(b) ~here IS a natural topology on Is for which the periodic points of ex are dense
In Is;
(c) there exist dense orbits of a on I:s.
Let a (fJ) be the left (right) shift on the symbol sequence space I:. Show that: (a) ex is a homeomorphism and ex I = p;
(b) ex is topologically conjugate to p.
~t a: Is + Is be either a left or right shift on Is, the space of symbol sequences WIth 8 = {O, 1, ... , m l}. Show that there are m' sequences a such that cx'(a) =' a. Let m, be the number ofperiodk points ofex and K = {klkel+ &klq}. Prove that
L mt = m' (E3.5)
3.5.5
3.5.6
3.5.7
3.5.8
for m, qel+. For m = 2, use this result to find the number of period12 orbits.
3.5.3 Symbolic dynamics for the horseshoe tliffeomorphism
Let f: Q + H2 be the horseshoe map. Use (3.5.3) to prove that each vertical strip of Q(n" ne N, can be described by a binary sequence of nsymbols.
Let (x, y)eA, the invariant Cantor set of the horseshoe map f: Q + R2. Define {a,}, {b,} j; I as follows.
(i) Consider the nested vertical strips given by n Q(.,.lfxe{PO then al = {I.
o.N PI + I
F ':::.2 {I'f I' {left
or r e ,ai= I x ies to the . of the previous strip in the nesting.
+ I right
(ii) For the nested horizontal strips n Q(o" define bl ={+I if ye{Qo and
.eZ' I QI
bi = r II, i ~ 2, if y lies to the {top of the previous strip in the nesting.
 bottom
Show that each point (x, y) of A can be written in the form
(2 f aj5',2 f bj5i), ai' b, = ± 1.
1=1 '=1
3.5.9
Hence show that the subset of A in the quadrant x, y> 0 is homeomorphic, by the fivefold magnification (x, y)H (5x  2, 5y  2), to the whole of A. What does this imply about the structure of the set A?
184
3 Structural stability, hyperbolicity and homoclinic points
3.5.10
Use the form for x e A given in Exercise 3.5.9 to find the coordinates in Q of the fixed and period2 points of the horseshoe map I: Q + R2.
Let I: Q + H2 be the horseshoe map. In the notation of § 3.5.3, the set II'
A (11') = n Q(.) consists of22N connected components, each one being a square
.=(11'1)
region of side 2/511' lying within Q.
(a) Prove that (3.5.17) uniquely associates a symbol block 11(11') = {11(NI)' •.• ,I1N} with each connected component, K(I1(N) say, of NN). Locate K(I1(N) for the following symbol blocks 11(11'):
3.5.11
(i}{l·l};
(ii) {1I·00};
(iii) {01O·101}.
3.5.12
(b) Let 11(11') = {II (11'1)' ... ' liN} and v(lI') = {V(NI)," ., VII'} be two symbol blocks of length 2N. Explain how to choose a point xeA in K(II(N) su~h that flN(x)e K(v(N). Hence show that there exist points in K(II(N) whose orbit under flN visits every connected component of NNI in any desired order.
(c) What restriction must be imposed on the elements of the sequence 11 if the orbit of the point x = h(l1) e A is to remain in a particular component, K(ulN) say, for k applications of I? What is the maximum number of connected components that can be reached from K(I1(N) in k iterations of I?
(d) What aspect of the dynamics of IIA do the observations (aHc) reflect?
Recall that the horseshoe map I: Q + H1 satisfies (x, y)8 (f1(X), fl(x, y)). Verify this property for I] A by considering the left shift IX: I: + I: that is conjugate to I. Show that II: [ 1, 1] + R has a repelling invariant Cantor set.
The Baker's transformation B: Tl + Tl is defined by
xl=2xmodl, YI=t(2Xxl+y)modl
for (xmodl, ymodl)e Tl (Arnold and Avez, 1968, Appendix 7).
Describe the effect of this transformation on the rectangles Po = [0, !) x [0. I) and PI = n, 1) x [0, 1). Show that every point x e Tl can be written in the form
3.5.13
00
x=h(I1)= n B"(Pa),
,,=  00
where 11 = {11.}:= _., is a biinfinitesequence of {O. I}. Use this result to show that
00
B(x) = n B"(P alaI)'
(E3.6)
.= 00
where IX: I: + I: is the left shift.
Prove that h: I: + T? is not onetoone by finding h(l1tl and h(112)' where 111 = { ·01.0·} and 111 = { '10·0 '}. ( 'i and i . i~dicale indefinite recurre.nce of th~ symbol i to left and right, respectively). What IS the genera~ form of po~nts x e T for which h fails to be injective? Given that these problem pomts can be disregarded (see Arnold & Avez, 1968. p. 125). show that the nonwandering set of B is the
whole of t.
3.6.1
3.6 Hyperbolic structure and basic sets
Let g:Rl+Rl, given by g(x) = (9,(X),g2(X))T, x=(xl.xlf, be a smooth map and y: (  1. 1) + Hl be a smooth curve. Show that the tangents to the curves y
Exercises
185
and g.y at t = 0 are related by the equation
~(g''')1 = Dg(y(O)) dyl
dt ,=0 dtl,=o
where Dg(x) is the matrix of partial derivatives (Og;)l . Illustrate this result
ox} ;.}=I
by finding the image under g(x" x2) = (exp(x, + x2), X,X2)T of the curves: (a) XI = t, X2 = t cos t; (b) XI = I + II, Xz = tan t.
Show that the curves (a) and (b) have a common tangent at x = O. Verify that the image curves under g also have a common tangent that is given by (E3.7).
(E3.7)
3.6.2
Consider the representation of SI shown in Figure E3.I. Find the overlap map hl2 between UI and ti;
The norms 11'11., on TU I., and HI., on TU Z', are compatible if
(E3.8)
3.6.3
(a) Prove that the Euclidean norms on U, and Uz are not compatible.
(b) Find the norm on Uz\Pz which is compatible by hlz to the Euclidean norm on UI• Show that this norm cannot be extended continuously to the whole of Ul.
(c) Verify that, at each point X; of U;
(u. v); = 16uv/(4 + xW.
isa positive definite inner product on TUj.,. Show that the norms II vII; = (v, ». are compatible.
Let M be a differentiable manifold and x* be a fixed point of the diffeomorphism I: M + M. Let h: We M + U be a chart at x· and h(x·) = X·. Show that the eigenvalues of the derivative D(hfh')(x*) are independent of the choice of the chart (U, h) at x*. What is the significance of this result in relation to the problem of defining a hyperbolic fixed point on M?
Show that the Anosov automorphism I: TZ + T1 given by A: RZ + R2. where A = G ~). satisfies the hyperbolicity conditions (3.6.9 and 3.6.10) at each point of n= T2.
3.6.4
Figure E3.l Stereographic projection from (SI\P2) + U, and (SI\p,)+ U2 provides an atlas for the unit circle S' c R2.
y
gives the orientation of (.y.
186
3 Structural stability, hyperbolicity and homoclinic points
Exercises
187
3.6.5 Consider the dilTeomorphisms gl' g2: S2 + S2 defined in Figure E3.2. Outline arguments to show that gj, i = I, 2, has an invariant Cantor set Aj £; Q such that g,IAj is conjugate to a shift on m(i)symbols, where m(l) =.3 ~nd m(2).= 4. W~at are the basic sets of gl and g2? For both maps, draw schematic diagrams illustrating the dynamics of the wandering points. Is g.lAI conjugate to g21A2?
3.6.6 Find which of the following dilTeomorphisms of the torus T2 satisfy the hypotheses of Theorem 3.6.1 and describe their basic sets:
(a) (I (n(x)) = _(Ax), A = (! ~}
(1/2)
(b) (2((X» = _(x + II), 11= 31/2 ;
3.6.7
(c) (3( _(x» = n(fI'l (x)), where fI' I is the timeone map ofthe system (x) = (s~n(27tX»).
y sm(21ty)
Obtain the Henon attractor using a microcomputer to plot the iterates ofthe map
XI = y1.4x2 + I,
YI =O.3x,
(E3.9)
with initial value (1,0). Choose X and Y scales such that the square Q = {(x, y)llxl < 1, IYI < I} fills a large portion of the screen. Observe the braided nature of the attractor by magnification and shifts of the origin.
Figure E3.2 The restrictions of gl and g2 to the capped square Q' are shown in (a) and (b), respectively. On each component of Qlo" i = I, 2, g, is assumed to be linear. The map gl is a contracting diffeomorphism on both Fl and GI with hyperbolic fixed points PIEFi=gl(Ftl and P2EG'I=gl(GI). Observe that GZ=g2(G2)c:F2 and g21F2 is a contracting diffeomorphism with hyperbolic fixed point P 3 E Fz = g2 (F 2)' On S2\Q', both gl and g2 have a single repelling hyperbolic fixed point Po. It may be assumed that Theorem 3.6.1 applies to both gl and g2'
3.7 Homodinic points
3.7.1 Use the explicit form of the horseshoe map I: Q + R2 on POuPI (see Exercise 3.5.1) to locate the fixed points of (and their stable and unstable manifolds. Show that there exist transverse homoclinic points at (x, y) = (t, t) and (lo t). Hence obtain Figure 3.31. Show that any homoclinic point xt of a periodic orbit on A is itself an element of A.
3.7.2 Let f: M + M satisfy the requirements of Theorem 3.7.1, i.e. f is a KupkaSmale diffeomorphism with a transverse homoclinic point associated with one of its
 pI  
periodic points. Show that 11.= U £irA) is a Cantor set such that £(11.) = A.
/=0
,
"
\
\
\
\
\ I I I
fvfj
,
I
3.7.3 Find the eigenvalues of the linearisation of (3.7.1) at the fixed point (x, y) = (1,0) and verify that both are positive for k > O. Show that the eigendireetions are given by y = {k ± [k(k + 4)r/2}(x  1)/2. For k = 1.5 take an interval of approximate length 0.0001 containing 100 points on the appropriate branch of the unstable manifold of the saddle fixed point. Plot 15 iterates of each point under (3.7.1) to obtain a numerical approximation to the unstable manifold at (I, 0). Use the inverse map (3.7.4) to complete the homoclinic tangle shown in Figure 3.33. Modify the program to exhibit the image of each successive iteration of the interval separately. Observe the repeated stretching and folding around the origin.
3.7.4 Use the program developed in Exercise 3.7.3 to study how the extent of the homoclinic tangle depends on k. Plot the tangle for k = 0.4, 0.8, 1.2, 1.6 and 2.0. Comment on your results.
3.7.5 Show that the derivative map of (3.7.1) has unit determinant for all (x, Y)ER2.
Given that there exists a linear conjugacy between (3.7.1) and (1.9.40), find a relation between k and a. Modify the program used in Exercise 1.9.9 to generate orbit plots for (3.7.1) corresponding to Figures 1.38 and 1.39.
3.7.6 Let y(t), tel c: R, define a closed curve, y, in the plane oriented with increasing t. Suppose that I'(s), SEJ c: R, defines a segment ofa planar curve, r, that intersects y transversely. Assume that the point of intersection is given by Xo = y(O) = frO) and that I'(s) lies inside y for s > O. Verify that y(O) " f(0) determines the orientation of y.
Let I: Rl + R2 be a diffeomorphism and show that
Q'
I I
Gr4
I
\
\
\ \ ,
,
"
(a) gl
Q'
(E3.l0)
(C:y)(O) " (C:r)(O) = Det DC(xo)[y(O) " i'(0))
(h) g2
188
3 Structural stability, hyperbolicity and homoclinic points
3.7.7 Let f: Q + R2 be the horseshoe map and x· be a periodic point of period q » I, on the invariant Cantor set A. Use symbolic dynamics to construct a point xt of A which is homoclinic to the periodic orbit containing x· (cf. Figure 3.35). Can the transverse nature of the homoclinic point be detected by the symbolic dynamics?
3.8 The Melnikov function
3.8.1
Consider the solutions of
(E3.H)
given by (3.8.5) and (3.8.6). Obtain (3.8.7) by substitution and comparison of order E terms. Hence deduce (3.8.10).
3.8.2 Show that H(x, y) = !(y2  x2 + tx4) is a Hamiltonian for the system x = y, y = x  Xl and verify that the level set H = 0 consists of a saddle point at x = 0 and two homoclinic orbits r; and ro given by
(x± (t), y± (t)) = (±21/2 sech(t), +21/2 sech(t) tanh(t)).
3.S.3 If the system
(E3.12)
where fl has period 21t/w, has a homoclinic orbit xo(t), for E = 0, then the corresponding Melnikov function is
M(Oo) = r ~ fo(xo(t»" fl (xo(t), t + 00) dt, OoE[O, 21t/w). Use r; obtained in Exercise 3.S.2 to find a Melnikov function for the system
(E3.13)
x=y,
y = x  Xl + E cos(wt).
(ElI4)
3.8.4
Prove that f:~ sech(t) tanh(t) sin (wt) dt = 1tW sech(1twj2) by using contour integration on a rectangle in Cwith vertices at (±R, 0), (± R, i1l')andleuingR + 00. The Melnikov function given in (3.S.lS) can also be used to indicate the separation of stable and unstable manifolds of two different saddle points in a Hamiltonian system (the socalled heteroclinic case). Show that the saddle connections, f~" between the fixed points (11',0) and (1t, 0) of the system
3.S.5
x=y,
y= sin(x)+E(aby)
(E3.1S)
with e = 0, are given by
• (xo(t), yo(t)) = (± 2 arctan(sinh(t)), ± 2 sech(t».
Calculate the Melnikov function for (E3.IS) along these orbits and show that
M(Oo) = ± 2a1t  Sb,
for r,j, ro, respectively. Explain why M(Oo) is constant.
(E3.16)
Exercises
189
3.S.6 The SinoGordon equation
x=y, y=sin(x)+e(acos(wt)by), (ElI7)
a, b > 0, with e = 0, has saddle connection orbits, fer, between fixed points at (±1t, 0). ~alcul~te the Melnikov function for (E3.17) along rer and show that it can be written In the form
M(Oo) = _!. {+ 2a1l'w cos(wOo) }
h Sb .
w cos (1twj2)
(E3.1S)
Describe the regions of the (a, b)plane for which transverse heteroclinic points occur.
,
HINTS FOR EXERCISES
1.1.1
Chapter t
WI = {exp(ix)la < x < b}, W2 = {exp(ix)lc < x < d} such that WI u W2 = SI, (b  a), (d  c) < 2n. C"'overlap maps.
h,.r.h.I = (h,.h~I)(h~.r.h; I)(hy·h.I). Composition of two Clmaps is C and overlap maps h,.h.i I and h,h; I are Cl since r?3 k. Differentiability of f is independent of charts.
(a) Pick open subsets A, B, C, D of H2 such that {n(A), nIB), n(C), n(D)} is an open covering of 1'2 and restrictions of n to A, B, C, D are homeomorphisms. (b) WI = S2\N, N the north pole; Wl = S2\S, S the south pole, hl(h2) is the
stereographic projection from N(S) poles. Overlap map (r, 11') ...... (4/r, ip),r * 0.
Arnold, 1973, pp. 1635.
(a) yes; (b) no; (c) circle map not homeomorphism. Fixed points x = 0, !; all other points period2. Plot y = PIx).
J an orientationreversing homeomorphism on Ii! implies it is strictly decreasing. J(x + I) = j(x)  I as in proof of Proposition 1.2.1. Fixed points of f only at intersection of y = J(x) with y = x and y = x + 1 (see Proposition 1.2.2).
Definition 1.3.1 implies rp, is CI for all t e R. 11',1 = rp _,.
X(x)=xx2•
(a) x = x3; (b)X = x, Y= y2.
Minimal: (a) SI; (b) {x, R.1q(x), ... , R~i;, \'<)} , xe SI.
General: (a) SI; (b) S = UuRpl,(U)u .. · R~/~ I(U), U £; Sl, closed.
(a) Show n° is open.
Separatrices connecting n = 1, 2, 3, 4 saddle points enclosing unstable focus. Consider Hamiltonian system with desired saddle connection and introduce dissipation in the region bounded by the separatrices, e.g. x =  2 y( I  x2) + IL,<B(x), Y = 2x(1  y2) + JlyB(y), II> 0,
{exp[ xl/(I x2)], [x] < I B(x) =
o Ixl~1.
L1.2
LI.3
U.I 1_2.2 1.2.3 1_2.4 1.2.5
1.3.1 1.3.4 1.3.5 1.4.1
1.4.2 1.4.5
Hints for exercises
395
1.4.6 Polar coordinates in the x, xplane give
dr/dO = f;r sirr' 0(1  r2 cos! 0) + 0(£2); rIO)  r(2n) = f.nr( 1~) + 0(£2).
Construct positively invariant set containing no fixed points.
1.5.1 (i) Similar construction to Example 1.5.1 for topological conjugacy with reference
intervals [1,2] for f and [1,8] for g.
1.5.4 Conjugacy preserves fixed points.
1.5.5 Use Proposition 1.2.2. Plot y = PIx). Conjugacy preserves periodic points.
1.5.6 Use (1.5.11).
1.5.7 1.5.8
Differentiate (1.5.11) and set x = O.
If x* periodic with periodq, p(f) = (~~~(J.q(x*)  x*)fnq )mod 1.
Consider separatrix of the saddle which is of opposite stability to that of the node. Consider the lifted flow tP,(x, y) = (x + t, Y + IXt) on R2. Periodic orbits are given by tPT(X, y)=(x+m, y+n), 1'*0, m, neZ.
Recall that (rp x ~),(x, y) = (1I',(x), ~,(y)), xe M, yeN. Consider the lifted flows iP,(x) = x + I, iP;(x) = x + 21/2I,Ifr,(x) = Ifr;(x) = iP,(x).
Arrowsmith & Place, 1982, §2.3.
ipl(X)= xe/[xex+ I].
Fixed point x* must lie in :E but X(x*) = O. Show that P2(rp,.(x))=II"o(PI(x)), xeSI'
Cylinder. Two limit cycles: stable x = 0; unstable x = 1. (a) Mobius band; (b) Klein bottle.
x(t) = C exp(t  cos t).
IfQ(I)isa fundamental matrix so isQ(1 + 1') and Q(t) = Q(I + T)QI(lo + 1')Q(to)· All POo are conjugate.
(eXP(,l.lt) 0)
(a) 11'(1,10) = 1 ' t = I  to;
o exp("2r)
(b) X(I)=(4eXp(t)). 2exp(t)1
Use polar coordinates. Null solution is stable.
1.6.3 1.6.4
1.6.5
1.6.6 1.7.1 1.7.2 1.7.3
1.7.4 1.7.5 1.8.2
1.8.4
1.8.5
1.8.7
1.8.8
X(I)=(:~:: :~:tt)xo+(_~nJ
Sketch level curves of H(x I' Xl)'
The generic case has nonzero eigenvalues.
Hamilton's equations in plane polars are r = ,.1 oHliJO, Ii = =r I oHlor. Examine extrema of H as a function of r for various fixed values of O.
1.9.1 1.9.3 1.9.4
1.9.5
. a(r, 0) 2 .. a(t,O) (1) = ; (11) = I.
a(x, y) r a(x, y)
2.2.1
396
Hints for exercises
1
i
2.2.5
1.9.6
A=G ~),B=C=O'D=G ~),aand/orb#O.
. iJH . oH ox oy
If x =  and Y =   then  = 
oy ax ax a.v·
'I' is symplectic to order 111.
1.9.7
1.9.8
2.1.1
Chapter 2
If the Jordan form of L is not diagonal examine the powers of blocks of the form ).1 + N. Nji = aj.iI' Observe (Nk)ij = aj•ik• I ~ k ~ n  I. N" = O.
(a) ~ = max{j)'II ..... IAtI}. (b) Pick N > 3 such that ~ = N"NIAI < I.
2.1.2
2.1.3
(3 + 51/2)
(i) AIEu: UH 2 u, orientationpreserving expansion;
(351/2)
AlE': IJH 2 v, orientationpreserving contraction;
2.1.4
(ii) AIEu: UH (I + 21/2)U. orientationpreserving expansion; AlE': IIH (I  21/2)V. orientationreversing contraction.
(! t 0)
Real Jordan form of A is ! ! 0 . AlE' is a rotational contraction.
o 0 2
2.1.6
x=Ax is linearly conjugate to y=Ay. A = [AjaiiJ!.i=I' Show that Yj=AiYi is topologically conjugate to ii = sign(Aj)zj. i = 1, 2, 3. and use Exercise 1.6.5.
Use Theorem 2.1.2.
dim E' + dim EU = n and restrictions to E' and EU may be orientationpreserving or reversing.
2.1.7 2.1.8
(a) or(O, 0) = G
~). saddletype with reflection; 2
(b) Df(O.O)=(~ ~} expansion with reflection.
2.2.2 DIPI(O)=eXp(DX(O))=exp(~~) has eigenvalues cosh(ll±sinh(l). Show that W~'U(O) = W.~U(O). where IP is the flow of x = y. Y = x  x2• Obtain W~(O)() W~(O) from a first integral.
2.2.3 If r(x) = y then fql U and f'ljV are conjugate by r.
2.2.4 If YEW'W(X*)) then Limf"4+k(y)=f'+k(x*). k=O .... ,q1. No. construct
• 11+00
counterexample:
(a) for periodI, consider IP, in Figure 1.16, let f = IPI and observe ye A ¢ W;,U(Po) but Po c L,.(y);
(b) for periodq > 1. construct periodic orbit in similar manner to Exercise 2.2.5.
Hints for exercises
397
2.2.6
Vector field is symmetric under clockwise rotation by n/2. A fixed point x* of IPI' with topological type given by DIP I (x*) = exp(DX(x*)), becomes a periodic point off=IP,·R_./2·
Let XI = IP,o(xo) and define So = IPro(SI)' Use flow box coordinates to prove Po: So .... So and Po: So .... So are C'conjugate and result then follows from Exercise 1.7.3.
Introduce cylindrical polar coordinates and recognise closed orbit for r = (x~ + xn'/2 = 1, Z = X3 = O. The Poincare map '1'2. defined on the plane II = constant has a fixed point at [r, z) = (1.0). Hyperbolicity follows from DIP2.(1, 0) = exp(DX(I, 0)) and Hartman's Theorem.
(i) Solve quadratic for y and expand square root.
(ii) Use (i) to obtain YI and substitute into expansion for (I + YI)I.
(a y2 + a Y v + a y2)
Let h2(y) = I I 2 1. 2 ) 2 ,write down LAh2(y) and show that a, and
b,y~ + b2YIYz + b)y~
b.; i = I, 2. 3, can be chosen such that LAh,(y) = X,(y). Find al = a, = a) = bl =
b2 = I, b) =0.
Use resonance condition to show that (x;) is the only resonant term.
Use resonance condition and qAI + P)" = O.
(AIXI + eXi)
Use resonance condition. Normal form when AI = mA" m> 2. .
).,X,
2.2.7
2.3.1
2.3.4
2.3.5
2.3.6
2.3.7
2.3.8
Matrix representing LA is triangular with repeated eigenvalue J.. Since AI = Al = A, A ... i = ). for all m, i.
Use basis {(;,} ... , CT~2m}"" (~).ed)..··. (XT~) .... (~)}for
H'.
(I) ad  be # 0; (2) ad  be = O. a + d # 0; (3) ad  be = a + d = 0, a2 + b' + c2 + d' i' O. cod(S,) = 0; cod(S2) = I; cod(S) = 2. Linear vector fields satisfying (2.4.1) and (2.4.3) have codimension I and 2, respectively.
2.3.9
2.4.1
2.4.3
e
(i) A =t(b + f), B = e, C = 0, D = e/2, E = j, F =0; IX = a+ 2' P =d.
(ii) A =!(b + n, B= c, C =0, D = a. E = j, F=O; y =d. a = e+2a.
2.4.4 2.4.5
e=b, d=2a.
Consider the types of Jordan block which give rise to nonhyperbolic linear systems. Show that each type of block satisfies a resonance condition for all r;': 2.
Observe that c > 0 implies fix) > «)  x for x sufficiently small and positive (negative).
Use (2.5.8) for complex form with n = 2. Observe that f ... 1 = 0 for m, = 0 implies no idependent terms arise. Alternatively. use (2.5.8) with II = 1 and a single (complex) variable z (see Exercise 2.5.2). Note ).q+ 1 =)..
a = exp(3ilX)/[1  exp(4ilX)]. Note exp(4ilX) #1 for IX #2np/q. q = I, 2. 3,4.
2.5.2
2.5.3
2.5.4
398
Hints for exercises
2.6.1 M=(~ ~)haSeigenVaIUes±ibut(~ ~)=exPC~2 ~/2}
2.6.2 Note:
(i) if AB = BA then exp(A + B) = exp(A) exp(B); (ii) N"=O.
2.6.3 Let S'MS = J, find In J from Exercises 2.6.1 and 2. exp(ln J) = J implies L =S In JS1•
Alternative implies state transition matrix (itself a particular fundamental matrix) tp(t, 0) = U(t) exp(CI), make change of variable x = U(t)y and show that y = Cy. Thus alternative implies Theorem 2.6.1 with A = C and B(I) = U(t).
Theorem 2.6.1 implies i = A(I)x has solutions x(t) = B(t) exp(At)yo = '1'(1, O)xo = Q(t)QI(O)XO for any fundamental matrix Q(t). Let Yo = Q I (O)xo to obtain the statement given in the question and Yo = Xo for 8(t) = tp(t, 0) exp( At).
2.6.4 If J( is a Jordan block corresponding to a complex eigenvalue ).2 of M2 then the complex linear transformation that reduces J( to the real Jordan form JR, transforms In Jc into a real matrix, i.e. J" has a real logarithm.
IC p = tp(h, 0) then p2 = fP(4n, 0) = exp(4nA), for real A, by the first part of the question. Show that 8(t) = tp(t, 0) exp( At) is 4nperiodic in t.
2.6.5 Q(t)= (x+: x_), tp(r, 0) = Q(r)Q(Orl,
IP(2n, 0) = (COSh 2n sinh 2n) = exp{2n(0 Ol)}.
sinh 2n cosh 2n I
2.6.8 2.6.9
(a) zW, zlz]"; (b) zjzf, zlz]", z4exp(2it).
For Aj = 0, i = 1,2, (2.6.14) implies resonance only if v = 0, i.e. all timedependent
y2
terms can be removed. Let x = y +  sin(2t) and find a = !. b =  c. 4
2.7.1
AIE'giVenby(a)G ~}(b)(~ ~}(C)(~ ~).(b)giVeSUnboundedmotion. Decompose R" into the direct sum E' EB E' EB EU and consider restrictions of exp(At) to ES and EU•
C"', unstable.
No, origin is hyperbolic node. Maximum differentiability given by [b/a].
2.7.2
2.7.3 2.7.4
2.7.5
co
a2] == 0, a3] == 0, a4] = 41, as] == 0, a6] = 2(6]+ 1)(1  <iY+ I); L ajjll] converges
for i = 4 if II < i and for i = 6 if II <~. t= 0
For C * 0 centre manifold is nonanalytic.
2.7.6
co
2.7.7 Assume centre manifold given by y = L ajxl and show that ao = al = 0;
1=0
a21 = (k  1 )!, all + I =,0, k ~ I. Hence Y has zero radius of convergence.
2.7.8 Assume centre manifold E' of the form y = ao + alx + a2x2 + O(Xl). Show that ao = a I = 0, a 2 = I. Consider restriction of system to E'.
2.7.9 Assume centre manifold of the form x = Co + elY + e21 + O(l) and show that CO=CI =0, C2= 0.
Hints for exercises
399
2.7.10 r = 1, linearise; r> 1, assume EC of the form y = a2x2 + O(x3).
2.7.11 M=(I I), Y= IXX2+0(Xl).
° I
2.8.1 Polar blowingup gives: (i) saddles at (} = 0, n, i » ( <) ° for r > 0 and (} = 0 (x): (ii) (} = 0 unstable node, (} = n stable node.
2.8.2 Singularities on r = 0 circle are: (a) (} = 0 unstable node, (} = n/4 saddle, (} = n/2 unstable node, (} = n stable node, (} = 5n/4 saddle, (} = 3n/2 stable node; (b) (} = 0 unstable node, () = n/4, n/2 saddles, (} = n stable node, (} = 5n/4, 3n/2 saddles.
2.8.3 Repeated blowingup along positive jaxis gives further saddlenode singularities.
2.8.4 Division by lull and Ivll is necessary to prevent orientation reversal.
2.8.5 Positive xblowup, unstable node; negative xblowup, stable node. (cf. Exercise 2.8.1 with a= I, b=2.)
2.8.6 (a) Do polar blowup, investigate resulting singularities at () = n/2, 3n/2 with
further polar blowups. Obtain nonhyperbolic saddle.
(b) Polar blowup gives six hyperbolic singularities. Obtain 'monkey'saddle.
Note that the unfoldings of the vector fields considered in this question appear in Section 5.6 (see (5.6.2) (q = 2) and (5.6.14) (q = 3)). The reader may like to confirm that the underlying singularity for q = 5 (see (5.6.34) and (5.6.35)) is a focus, while q = 4 (see (5.6.21)) admits a variety of singularity types.
Chapter 3
3.1.1
Recall:
(i) the spectral radius, p(A), of A is the maximum of the absolute values of the eigenvalues of A; (ii) the spectral norm, utA), of A is the positive square root of the
largest eigenvalue of AlA; (iii)p(A) <>; utA); (iv)u(A) <>; IIAII, where IIAII = L lalA.
ij
(a) Let M'AM=D, D=[AI.5I]]. Consider det(MIBMIlI), with II an eigenvalue of 8 that is not equal to AI for any i = I, ... , n, and show that u(D;ICI)~I, where Dp=D1l1 and CI=M'CM. Observe
max [1,1.1  111 I] ~ C  1 implies min [li'l  Ill] <>; c.
I
{ 0 '*;+1 .
(b) Let M I AM = D + T, T = [tij]' tl] = ~., Proceed as In (a) and
10rO ) = r + I
nI "
note that (I+D;'T)"'=I+ L (ltD;lT\ u(T) = I, L xl<nx' for
1=1 1=1
3.1.2
x>!.
(c) Observe similar results follow for S.(A).
Hyperbolic implies structurally stable: use Exercise 3.1.1. Structural stability implies hyperbolic: observe a Jordan block J associated with an eigenvalue A of absolute value unity satisfies Wxl = Ixl for all kE 1.+ if x lies in the eigenspace of A. The block (I  .5)J, .5 > 0, can be made arbitrarily close to J but Wxl+ 0 as k + co for all x, Density: note that, if A is nonhyperbolic with eigenvalues Ajo then A +.51 has eigenvalues )', + fJ.
A structurally stable in S: let A, BES be sclose and apply Theorem 2.1.3 in the
3.1.3
400
Hints for exercises
Hints for exercises
401
subspaces on which the restriction of A and B is not the identity. Hence construct a conjugacy for A and B. A is not structurally stable in L(1R2): consider
A=(I +r. ~).
o I.
Show that A E O(R2) is a rotation so that every circle, centre x = 0, is invariant under A. Observe H = (I  r.)A,r. > 0, has no invariant circles. Prove that conjugacy preserves invariant circles.
(a) Use the Implicit Function Theorem and Exercise 3.1.1 (b) Use Exercise 3.1.1.
(c) DX(O) = DX(O).
3.2.2 (a) I'll < r./4; (b) I'll < 6/32; (c) I'll < t/l3. Use Dq>,.(ro) = exp(21tDX,(ro)) (see Exercise 2.2.2), where X, =; and X,(ro) = O.
3.2.3 (a) 1"1 < t/(2 + R); (b) 1"1 < r./(R2 + 2R). The fixed point x· is not hyperbolic.
Restriction of A to its stable manifold is orientationreversing while for (3.4.12) it is orientationpreserving.
3.5.1
I] Po: (x, y)r> (  5x  2,  y/5 + 2/5); flP I: (x, y)r> (5x  2, y/S  2/5).
(a) Note glQ = f; use (3.5.3) and (3.5.1).
(b) Explicit form of f given in Exercise 3.5.1 shows xcomponent of fIx, y),
(x, y)E PouP., is independent of y. Let x E n Q(.) and x' eA have the same
3.1.4
j=O
3.5.3
3.2.1
xcoordinate, show that I{"(x) {"(x'~ ~ 0 as n ~ OC). Eliminate xeQ\ n Q(.).
..N
Similarly, for xe n Q(O) using fI in place of r.
IIeZ .~
3.3.1 Theorem 3.3.1 is:
(a) applicable. nonhyperbolic closed orbit;
(b) applicable, nonhyperbolic fixed point at [x, y) = (I. 0); (c) not applicable, use perturbation (bB(y), 0), be R.
3.3.2 (a) Theorem 3.3.3(i) fails; (b) Theorem 3.3.3(iii) fails.
x = sin(21tx),y = s sin(21ty) } . I {(a). forlbl < s/(1 + 21t)
IS I:C close to
X = I, Y = 2+ b sin[21t(J'  2x)] (b) for Ibl < &/(1 + 61t).
3.3.3 Apply Theorem 3.3.1 to So. The eC'close perturbation given by ; = f. + [r cos(21tr)/(1 + r2)], 6 = 1 has no limit cycles for Irl > I/r..
3.4.1 Use appropriate lifts to examine fixed and periodic points of f.
(a) infinite number of periodic points implies nonwandering set not finite.
(b) no fixed points but every orbit is dense therefore nonwandering set does not consist of fixed and periodic points.
(c) four fixed points on SI but none are hyperbolic. (d) nonhyperbolic period2 points.
3.4.2 Recall that 71: 1R2 + T2 is a local diffeomorphism and differentiate 7I(fQ(x» = fQ(7I(x))
with respect to x to show that Tf'(7I(x)) (see § 3.6) and Dfq(x) are conjugate. Df!(x) = A', for all x, and A' is hyperbolic. W,·"(t'(7I(x·» = 7IWx· + E"O), where ,I
£,(EU) are the stable (unstable) eigenspaces of A. W"" = U W··U(t'(7I(x·)).
3.5.4 N(d) = [In(2/d)/ln 5].
3.5.5 Cf Propositions 3.5.35.
3.5.6 (b) Consider h: l: ~ r defined by h(I1). = 11(n_I)'
3.5.7 Observe that if a'(I1) = 11 then 11 is periodic with periods' where q'lq. 335 period12 orbits.
3.5.9 A is repeated within itself on all scales.
3.5.10 Fixed points: (1/3, 1/3), (1/2, 1/2).
Period2 points: (4/13,6/13), ( 6/13, 4/13).
I
3.5.11 (a) Verify for n g0(p.) and use induction. Exercise 3.5.9 gives:
.=0
(i) square of side 2/5, centre (x, y) = (2/5,  2/5);
(ii) square of side 2/52• centre (x, y) = (8/25. 12/2S); (iii) square of side 2W. centre (x, y) = (38/125. 38/125).
(b) 11 = { ... V_(N_II.···. VN. 'I(NI).···. '10·'11'···. 'IN'" .}.
(e) Central block of 2N symbols must be preserved for k shifts of binary point to the left, i.e.
3.4.3 Observe that 7I(x·), where x· = (A' If Ip. pe Z2. is a periodic point of f of period at most q. Verify that. when A is given by (3.4.12).IDet(A' 1)1. q ~ 2. is an integer greater than unity. hence show that (A' I)  I has at least one element that belongs to Q2\Z2. When q is prime deduce that f has a periodq point.
3.4.4 Differentiable conjugacy of f. g by h implies (h(f(x))) = (g(h(x))) + k where k e Zn and : denotes a lift of '. Differentiate with respect to x and set x = O. Conditions on C mean it is a Ii(t of a diffeomorphism. h. say. on TO. CAx = Hex implies h(f(x)) = g(h(x»; take projection 71 and show h(f(O» = g(h(O». where 0 = 7I(x).
3.4.5 y = (( I ± 131/2)/6 lx, Irrational slope implies stable and unstable manifolds of fixed point at x = 0 wind densely around the torus without closing. Homoclinic point is given by the intersection of y=(1 + 131/2)x/6 and y=(I13112)(xI)/6.
11={ ... i •... ,i.i, ... ,i. i, ...• i •... },i={O, I}.
~..__.. ..__..
k N N
Maximum number of blocks that can be reached in kiterations is 21. (d) 'Chaotic' motion.
3.5.12 Note that if (x, y)e A is given by h(I1). then x is determined by the part of 11 lying to the left of the binary point. If A I denotes the invariant Cantor set of fl> then show that Ifdx)  f( (x')1 = Six  x'i for any x e A to x' ¢ A., i.e. fl is locally repelling at each point of AI'
3.5.13 Show that 111 and 112 both represent the point H. 0). The map h fails to be injective at points corresponding to the dyadic fractions (ml/2°', m2/2°'), m;, n;e Z + , i = 1,2 (see Arnold & Avez. 1968. p. 125). If these points are disregarded. the symbolic dynamics can be used to show that the periodic orbits are dense in T2.
3.6.1 y(O)=(I, W; (g:y)(O) = (2,W.
3.6.2 Overlap map hI2(xl)=4/xl• (b) For vETU2,,(=R), IIvlIX2=4Ivl/x~ on U2\P2• where 1·1 is the Euclidean norm of '.
4()2
Hints for exercises
3.6.3 Let (U" hj), i = 1,2, be overlapping charts and assume that h12 = hzh!l: U .... U
. Cl Sh h h IbI I Z IS : ow t at D( I I )(hl(x)) and D(hzlbi I )(hz(x» are similar. If f(x·) = x·
the eigenvalues of the tangent map Tfx' can be unambiguously defined to be those of i.ts local representati.ves. This means that a fixed point on M is hyperbolic if all Its local representatives are hyperbolic in the sense of Definition 2.2.1.
3.6.4 Char~ TI with n and define II Tf;(vo)11 = IDf"(x)vxl, V,E TTj and v. = Tn;; IV,. A has eigenvalues ).1 = (3 + 51/1)/2, ;'z = (3  51/1)/2. Take P = 1)'111 = l;'zl, c = 2 and c=!.
3.6.6 (a) fl is an Anosov automorphism; Theorem 3.6.1 applies; TI connected implies
there is only one basic set 01 = T2.
(b) f2 has no periodic points but 0 = TI (note f~'(x, y) = (x, y + 2n(31/1)) mod 1, nEZ); Theorem 3.6.1 is not satisfied.
(c) Four hyperbolic fixed points, Ph ... ,P4; Theorem 3.6.1 applies; 0= {Pl>PI,P3,P4}'
3.7.2 Recall: a Cantor set is a closed, uncountable set with empty interior such that
every point is an accumulation point.
3.7.5 k = 4 sinZ(a/2).
3.7.6 Note (f: f)(0) points into the image of l' under f and Au A Av = Det(A)(u A v).
3.7.7 x·=h(u*), u·= ( ... :UI91: 'UI91:uI91: ... ). Let xt =h(ut) where
ut = ( ... :UI91:uI91:UI'0, ... ,0:UI91: ... ),
4.1.5
k
wher~ uld = a IU2 and u I is a subblock of Ul91 containing q  k symbols. Show that a"(u ) .... u· and aft9(ut) .... a1(u·) as n .... 00. No.
3.8.1 AUAv+uAAv=(TrA)uAv.
3.8.3 M(lIo) = 21/2ltW sin(wllo) sech(ltw/2).
3.8.5 3.8.6
(E3.2) is an autonomous system.
Transverse heteroclinic points for bla < iltw sech(ltw/2).
Chapter 4
4.1.1 (a) Terms of order r ~ 3 are not removed as in (4.1.14).
(b) Linear terms are not removed when JlI '" O. Transformed system is not an unfolding of y = i.
4.1.2 (a) Take X(p, x) nonversal and V(v, x) versal. Show X  Y but y,., X. For
example X(p, x) = Xl, Y(v, y) = v + yl.
(b) (i) X  Y: h(Po, Pl> x) = X  PI/2, q>(po, Ptl = /10  pil4;
Y  X: h(vo, x) = y, q>(Yo) = (vo, 0).
(ii) Let x = x*(Po, Ptl be fixed point of x = X(Po, Jll> x). Then X  Y: hlJIo,JII: x) = x  X*(,IO' Ptl, !P(Po,Jltl=(PI3x·(Po,Jltlz, 3x*(Po,Jltl);
. VI (2V~ V:)
Y X.h(vo, Vl>Y)=y, !p(vo, vI)= +I'OYh vo+vl .
3 27 3
4.1.3 Let x=ay, then y= Y(II,Y), Y(O,y)=ayz, becomes X=X(II,x)=aY(II,aIx)so
Hints for exercises
~03
that X(O, x) = xZ. Note for a < 0, x = ay is an order reversing homeomorphism of R. Use versality of x = I' + XZ and transform back.
XZ + Jlzx+ Jl.. .
(a) q(p, x) = .   IS not contmuous at (,I, x) = (0, 0) for any choice
JlX3 + Xl + IlzX + Jl
of q(O,O).
(b) Comparison of coefficients yields bo(Jl) = I or ho(Jl) = Jlz, Latter implies 5 I (P) not defined at II = 0 and therefore certainly not smooth on neighbourhood of Jl=O. Hence bo(lI) = I, q(II,X)=(1 +11Xrl, 51(11)=0, 50(1I)=Jl·
(c) Let q(Jl, x) = _!!Vl, ~),So(Jl) = 50(Jl), whereq(II, x) and 50(,1) are given in (E4.7). (I + Jlx)
(a) Take G(p, x) = Xl in Mather Division Theorem. Set p = 0, differentiate k times and conclude Q(O, 0) = Ijq(O, 0) = g(O).
(b) Let X(P,x) be any unfolding of x= _Xl and take F=X in (E4.14). Since Q(O, 0) = I, X(p, x] is equivalent to family induced by (E4.15) with !p(p) = (50(p),···, SlI(P)f
) VkI  E
(I Let y=x m ( 4.15).
k
(ii) For k odd, right hand side of (E4.15) has at least one real zero.
(a) (iv), (v); isoclines arc tangent to each other but neither is tangent to either coordinate axis.
(b) (iv) and (v); (v).
If y = o then 2yxl x  (2y  2ll) =0. Take y #Oandexaminethediscriminant of this quadratic equation. Equation (E4.19) arises in connection with the averaged forced Van der Pol oscillator (see Arrowsmith & Place, 1984).
(a) Observe distance, d(p, x), between x = 0 and y = 0 isoclines satisfies d(O, x) = Xl and use Malgrange Preparation Theorem;
(b) (Y_X3+JlX,y)T;
(c) (y  X4 + IIX3, }y, (y  x· + 311Xl  2'12, yl, for hyperbolic points isoclines must intersect transversely  three such intersections are not possible for d(O,x)=x4.
4.2.4 Show that "'(x) has a unique maximum on (0, I] at x =!.
4.1.6
4.2.1
4.2.2
4.2.3
4.2.5
(a) P = 0 isocline is gi~en by an increasing function of H bounded by l' with slope yiP at H = O. The H = 0 isocline is independent of /1 and ,. A single fixed point arises at tangential intersection of if = 0 and P = 0 isoclines.
(b) Find where the nontrivial Ii = 0 and p = 0 isoclines meet in a single point.
More details of both of these models appear in Arrowsmith & Place (1984).
4.2.6
[a] Evaluate LAxj'xi'ej, i = I, 2, where e, is the ith column of Im+l and A is the coefficient matrix of the linear part of the extended vector field in (4.2.16). Use (2.3.7).
(b) Consider LAXIJlkeZ and LAx2l1kel'
(c) Apply Theorem 2.7.2 to the 2·jet of the transformed extended system to obtain equivalent system with XI = ).XI' In absence of terms of order three and higher, Xz depends only on xz and p. Complete the square on Xz to obtain (4.2.2)
. (Q(p))2
With az =' 0, b2 = C23' V = P(Jl)  , where 4c23
Q(p) = f (bW  c~z alj)IIJ'
j=1 A
..
Pip) = L a2jllj + OOpI2),
l= 1
412
Hints for exercises
6.8.2
(a) ~s~ repeat~ root of (6.8.5) in (6.8.8) and expand in powers of VI' (b) Use the Implicit Function Theorem on (6.8.8). Substitute r = r(Ul, v)(I + (1) in (6.8.4).
Introduce metric distance between two circles as maximum radial displacement. Show that associated functional equation which maps circles using N. is effectively the second component of (6.8.12).
Use induction on k and show that INl(z) + exp(21[ipk/q)clz,III = INl(z)II up to order Izl' ..
Express in polar coordinates, take logarithms and separate real and imaginary parts. Introduce local coordinate rL and consider Taylor expansion.
Use Theorem 5.4.2 and (5.4.14). Resonance tongue with tip at (0, VI)' Show that (VI' VI) near (0, VI) ~nd [Re A, 1m A) near (cos(2np'/q'), sin(21[p'/q')) are related by a local diffeomorphism (cf Figures 5.5 and 5.6).
Use generalisation of (6.8.5) and (6.8.8).
Abraham, R. & Marsden, 1. (1978) Foundations of Mechanics, Benjamin/Cummings. Arnold, V. I. (1963) Proof of A. N. Kolmogorov's theorem on the preservation of quasiperiodic motions under small perturbations of the Hamiltonian, Russ. Math. Surv., 18(5),936.
Arnold, V. I. (1973) Ordinary Differential Equations, MIT Press.
Arnold, V. I. (1978) Mathematical Methods of Classical Mechanics, SpringerVerlag. (Russian original, Moscow, 1974.)
Arnold, V. 1. (1983) Geometrical Methods in the Theory of Ordinary Differential Equations, SpringerVerlag.
Arnold, V. 1. & Avez, A. (1968) Ergodic Problems of Classical Mechanics, W. A. Benjamin. Aronson, D. G., Chory, M. A., Hall, G. R. & McGehee, R. P. (1983) Bifurcations from an invariant circle for twoparameter families of maps of the plane; a computer assisted study, Commun. Math. Phys., 83, 30354.
Arrowsmith, D. K. & Place, C. M. (1982) Ordinary Differential Equations, Chapman & Hall. Arrowsmith, D. K. & Place, C. M. (1984) Bifurcations at a cusp singularity with applications, Acta Applicandae Mathematicae, 2, 10138.
Aubry, S. & Le Daeron, P. Y. (1983) The discrete FrenkelKontorova model and its extensions, Physica, 8D, 381422.
Barnett, S. (1975) Introduction to Mathematical Control Theory, Oxford University Press. Birkhoff, G. D. (1968a) An extension of Poincare's last geometrical theorem, G. D.
Birkhoff Collected Mathematical Papers, Vol. 2, pp. 25266, Dover.
BirkholT, G. D. (1968b) Sur quelque courbes Iermee remarquables, G. D. Birkhoff Collected Mathematical Papers, Vol. 2, pp. 41843, Dover.
Bogdanov, R. I. (1981a) Bifurcation of the limit cycle of a family of plane vector fields, Sel. Math. Sov., 1(4), 37388.
Bogdanov, R. I. (1981b) Versal deformation of a singularity of a vector field on the plane in the case of zero eigenvalues, Sel. Math. Sov., 1(4),389421.
Bowen, R. (1970) Markov partitions for Axiom A dilTeomorphisms, Amer. J. Math., 92, 72547.
Bowen, R. (1978) On Axiom A difleornorphisms, CBMS Regional Conference Series in Mathematics, 35, AMS Publications, Providence.
Carr, I. (1981) Applications of Centre Manifold Theory, SpringerVerlag.
Casdagli, M. (1987a) Periodic orbits for dissipative twist maps, Ergod. Th. & Dynam. Sys., 7, 16573.
Casdagli, M. (l987b) Rotational chaos in dissipative systems, Research Report No. 103, Dept of Mathematics, University of Arizona.
6.8.4
REFERENCES
6.8.5
6.8.6
6.8.7
6.8.8
414
References
References
415
Chenciner, A. (1982) Points homoclines au voisinage d'une bifurcation de Hopf degeneree de difTeomorphismes de R2, C. R. Acad. Sci., Paris, 294,26972.
Chenciner, A. (1983) Bifurcations de difTeomorphismes de R2 au voisinage d'un point fixe elliptique, Les Houches, Session XXXVI, 1981, Chaotic Behaviour of Deterministic Systems, ed. G. looss, R. H. G. Heileman & R. Stora, pp. 275348, NorthHolland.
Chenciner, A. (1985a) Bifurcations de pointes fixes elliptiques: I  Courbes invariantes, Publ. Math. I.H.E.S., 61,67127.
Chenciner, A. (1985b) Bifurcations de points fixes elliptiques: II  Orbites periodiques et ensembles de Cantor invariants, Inventiones Math., 80, 81106.
Chenciner, A. (1987) Bifurcations de points fixes elliptiques: III  Orbites periodiques de 'petites' periodes et elimination resonannante des couples de courbes invariantes, Publ. Math. I.H.E.S., 66,591.
Chillingworth, D. R. J. (1976) Differential Topology with a View to Applications, Pitman. Chirikov, B. V. (1979) A universal instability of manydimensional oscillator systems, Phys. Rep., 52(5), 263379.
Chow, S. N. & Hale, J. K. (1982) Methods of Bifurcation Theory, SpringerVerlag. Collet, P. & Eckmann, J. P. (1980) Iterated Maps on the Interval as Dynamical Systems,
Birkhauser, Boston.
Cvitanovic, P. (1984) Unil'ersalily in Chaos, Adam Hilger.
Devaney, R. L. (1986) An Introduction to Chaotic Dynamics, Benjamin.
Dumortier, F. (1977) Singularities of vector fields on the plane, J. Diff. Eqns., 23, 53106. Dumortier, F. (1978) Singularities of Vector Fields, Monograjias de Matematica, No 32,
Instituto de Matematica Pura e Aplicada, Rio de Janeiro.
Eckmann, J. P. (1983) Les Houches, Session XXXVI, 1981, Chaotic Behaviour of Deterministic Systems, ed. G. looss, R. H. G. Heileman & R. Stora, pp. 455510, NorthHolland.
Gardner, M. (1961) Second Scientific American Book of Mathematical Puzzles and Diversions, Simon & Shuster.
Golubitsky, M. & Guillemin, V. (1973) Stable Mappings and Their Singularities, SpringerVerlag.
Greene, J. M. (1979) A method for determining a stochastic transition, J. Math. Phys., 20(6), 1183201.
Guckenheimer, J. (1979) Sensitive dependence on initial conditions for onedimensional maps, Commun. Math. Phys., 70,13360.
Guckenheimer, J. &. Holmes, P. (1983) Nonlinear Oscillations, Dynamical Systems and Bifurcations of Vector Fields, SpringerVerlag.
Guckenheimer, J., Oster, G. & Ipatchki, A. (1977) The dynamics of density dependent
population models, J. Math. Bioi., 4, 10147.
Gumowski, J. & Mira, C. (1980) Dynamique Chaotique, Cepadues Editions. Hale, J. K. (1969) Ordinary Differential Equations, Wiley.
Hardy, G. H. (1979) An Introduction to the Theory of Numbers, p. 125, Oxford University Press.
Hassard, B. D., Kazarinoff, N. D. & Wan, WH. (1981) Theory and Applications of the Ropf Bifurcation, Cambridge University Press.
Heileman, R. H. G. (1980) Self generated chaotic behaviour in nonlinear mechanics, Fundamental Problems in Statistical Mechanics, Vol. 5, ed. E. G. D. Cohen,
pp. 165233, NorthHolland. '
Henon, M. (1969) Numerical study of quadratic areapreserving mappings, Quart. Appl.
Math., 27,291311.
Henon, M. (1976) A twodimensional mapping with a strange attractor, Commun. Math.
PllYs., 50, 6977.
Henon, M. (1983) Les Houches, Session XXXVI, 1981, Chaotic Behaviour of Deterministic Systems, ed. G. Iooss, R. H. G. Helleman & R. Stora, pp. 57170, NorthHolland.
Hirsch, M. W. & Smale, S. (1974) Differential Equations, Dynamical Systems and Linear
Algebra, Academic Press.
Hirsch, M. W., Pugh, C. C. & Shub, M. (1977) Invariant Manifolds, SpringerVerlag. Hockett, K. & Holmes, P. (1986) Josephson's junction, annulus maps, Birkhoff attractors, horseshoes and rotation sets, Ergod. Th. and Dynam. Sys., 6, 20539. looss, G. (1979) Bifurcations of Maps and Applications, North Holland Math. Stud., No. 36. Jordan, D. W. & Smith, P. (1977) Nonlinear Ordinary Differential Equations, Oxford University Press.
Katok, A. (1982) Some remarks on BirkhofT and Mather twist map theorems, Ergod. Th.
& Dynam. Sys .. 2, 18594.
KoImogorov, A. N. (1957) General theory of dynamical systems and classical mechanics,
Proc. Int. Conf. Math., I, 31533, NorthHolland.
Lang, S. (1966) Introduction to Diophantine Approximations, Addison Wesley.
Le Calvez, P. (1986) Existence d'orbites quasiperiodiques dans les attracteurs de Birkhoff, Commun. Math. Phys., 106, 38394.
Li, TY. & Yorke,J. A. (1975) Period three implies chaos, Amer. Math. Monthly, 82, 98592. Lichtenberg, A. J. & Lieberman, M. A. (1982) Regular and Stochastic Motion,
Springer Verlag.
Lorenz, E. N. (1963) Deterministic nonperiodic flow, J. Atmo~. Sci., 20, 13041. MacKay, R. S., Meiss, J. D. & Percival, I. C. (1984) Transport in Hamiltonian Systems, Physica, 13D, 5581.
MacKay, R. S. & Percival, I. C. (1985) Converse KAM: theory and practice, Commun.
Math. Phys., 98, 469512.
Manning, A. (1974) There are no new Anosov diffeomorphisms on tori, Amer. J. Math.,
96,4229
Marsden, J. E. & McCracken, M. (1976) The Hop! Bifurcation OIld its Applications,
SpringerVerlag.
Mather, J. N. (1967) Anosov diffeomorphisms, Bull. Amer. Math. Soc., 73(6),7925. Mather, J. N. (1982) Existence of quasiperiodic orbits for twist homeomorphisms of the annulus, Topology, 21(4), 45767.
Mather,J. N. (1984) Nonexistence of invariant circles, Ergod. Th. & Dynam. Sys .. 2, 397. May, R. M. (1976) Simple mathematical models with very complicated dynamics, Nature, 261, 45967.
May, R. M. (1983) Nonlinear problems in ecology and resource management, Les Houches, Session XXXVI, 1981, Chaotic Behaviour of Deterministic Systems, ed. G. looss, R. H. G. Heileman & R. Stora, pp. 51563, NorthHolland.
Misiurewicz, M. (1983) Maps of an interval, Les Houches, Session XXXVI, 1981, Chaotic Behaviour of Deterministic Systems, ed. G. looss, R. H. G. Heileman & R. Stora, pp. 56590, NorthHolland.
Moser, J. K. (1962) On invariant curves of areapreserving mappings of an annulus,
Nachr. Akad. Wiss., Gouinqen, Math. Phys. Kl., 2, 120.
Moser, J. K. (1968) Lectures on Hamiltonian systems, Mem. A. M. S., 81, 160. Moser, J. K. (1969) On the construction of almost periodic solutions for ordinary differential equations, Proc. Int. Conf. on Functional Analysis alld Related Topics, pp. 607, Tokyo.
Moser, J. K. (1973) Stable and random motions in dynamical systems, Alln. Math.
Studies, 77, Princeton University Press.
Newhouse, S. E. (1979) The abundance of wild hyperbolic sets and nonsmooth stable
416
References
sets for diffeomorphisms, Publ. Math. I.H.E.S., SO, 10151.
Newhouse, S. E. (1980) Lectures on Dynamical Systems, Prog. in Math., 8, pp. 1114,
Birkhauser, Boston.
Nitecki, Z. (1971) Differentiable Dynamics, MIT Press.
Niven, I. (1956) Irrational Numbers, Math. Assoc. Amer., Menasha, Wisconsin.
Pal is, 1. & Takens, F. (1977) Topological equivalence of normally hyperbolic systems, Topology, 16, 33545.
Peixoto, M. M. (1962) Structural stability on twodimensional manifolds, Topology, 2,
10121.
Percival, I. C. (1982) Chaotic boundary of a Hamiltonian map, Physica, 6D, 6777. Percival, I. C. & Richards, D. (1982) Introduction to Dynamics, Cambridge University Press. Rand, D. A. Institute of Mathematics, University of Warwick, private communication. Robbin, J. (1972) Topological conjugacy and structural stability in discrete dynamical
systems, Bull. Amer. Math. Soc., 78, 92352.
Rossler, O. E. (1979) Continuous chaosfour prototype equations, Bifurcation Theory and Applications in Scientific Disciplines, ed. O. Gurel & O. E. Rossler, Ann. N. Y. Acad. Sci., 316, pp. 37692.
Ruelle, D. (1989) Elements of Differentiable Dynamics and Bifurcation Theory, Academic Press.
Riissrnann, H. (1959) Uber die Existencz einer Normalform inhaltstreuer elliptischer Transformationen, Math. Ann., 137,6477.
Riissmann, H. (1970) Uber invariante Kurven differenzierbarer Abbildungen eines Kreisringes, Nachr. Akad. Wiss., Gottinqen II, Math. Phys. Kl., 67105.
Sarkovskii, A. N. (1964) Coexistence of cycles of a continuous map of a line into itself, Ukrailliall Math. J., 16,6171.
Siegel, C. L. & Moser, J. K. (1971) Lectures 011 Celestial Mechanics, SpringerVerlag. Smale, S. (1965) Diffeomorphisms with many periodic points, DijJerential and
Combinatorial Topology, ed. S. Cairns, pp. 6380, Princeton University Press.
Smale, S. (1966) Structurally stable systems are not dense, Amer. J. Math., 88, 4916. Smale, S. (1967) Differentiable dynamical sytems, Bull. Amer. Math. Soc., 73,747817. Sotomayor,1. (1974) Generic oneparameter families of vector fields on two dimensional
manifolds, Publ. Math. I.H.E.S., 43,546.
Sparrow, C. (1982) The Lorenz Equations, SpringerVerlag. Sternberg, S. (1969) Celestial Mechanics Part II, W. A. Benjamin.
Takens, F. (1971) A CIcounterexample to Moser's Twist Theorem, lndag. Math., 33, 37986.
Takens, F. (1973a) Introduction to global analysis, Commltn. Math.lnst., Rijksulliversiteit Utrecht, 2, IIII.
Takens, F. (1973b) Unfoldings of certain singularities: Generalized Hopf bifurcations, J.
DijJ. Eqns., 14,47693.
Takens, F. (1974a) Singularities of vector fields, Publ. Math. I.H.E.S., 43, 47100. Takens, F. (1974b) Forced oscillations and bifurcations, Commun. Math. Inst.,
Rijksuniversiteit Utrecht, 3, 159.
Taylor, J. B. (1968) unpublished.
Thompson.J. M. T. & Stewart, H. B. (1986) Nonlinear Dynamics and Chaos, Wiley. Van Strien, S. 1. (1979) Centre manifolds are not Coo, Math. Z., 166, 1435.
Whitley, D. C. (1983) Discretedynamical systems in dimensions one and two, Bull.
Lond. Math. Soc., 15, 177217.
Wiggins, S. (1988) Global Bifurcations and Chaos, SpringerVerlag.
actionangle variables, 48 additional resonant terms areapreserving planar map, 308 rational rotation, 86, 258 timedependent vector field, 91
Anosov automorphism
and Anosov diffeomorphisms, 134 chaotic basic set, 159
definition, 133
and dynamical systems, 120 homoclinic points, 1368. 182 periodic points. 134
Anosov diffeomorphism
conjugacy to automorphism. 134 and dynamical systems, 133 structural stability, 135
areapreserving planar map Birkhoff normal form, 305 complex form, 306
and Henon areapreserving map, 169 normal form, 308
Arnold's circle map, 248
and dissipative standard map, 349 and unfoldings of rotations, 257
Arnold tongues definition, 252 symmetry of, 293
and unfoldings of rotations, 257 atlas. 2
attracting set, 346
chaotic, 161
attractor, 346
strange, 161
AubryMather Theorem, 338 axiomA diffeomorphism, 158
Baker's transformation, 184 basic sets
Anosov automorphism, 159 Decomposition Theorem, 1589 horseshoe diffeomorphism, 158 spinning diffeomorphism, 160
INDEX
bifurcation, local. 191 bifurcation curve, 207 bifurcation point, 190 biharmonic oscillator, 48 Birkhoff attractor, 348 Birkhoff normal form
for areapreserving planar map, 305 for Hamiltonian, 303
Birkhoff periodic orbit, 336
Birkhoff periodic points, 309
of type (p,q), 336
Birkhoff set, 347 rotation interval, 347 Birkhoff Theorem. 338 blowingup
for cusp singularity, 107 directional, 105
polar, 102
in xdirection, 106 in ydirection, 106
Bogdanov map, 359 Bogdanov points, 378
Calabi invariant, 345
C'norm for vector field, 123, 125 canonical polar coordinate system, 305 canonical transformation, 45
Cantor set
definition, 333, 386
in double invariant circle bifurcation, 375 invariant
for areapreserving twist homeomorphism, 338
for circle homeomorphism, 3325
for horseshoe diffeomorphism, 139~7; dynamics on. 14954, 184
in SmaleBirkhoff Theorem, 165 for spinning diffeomorphism, 160 centre eigenspace, 94
centre manifold
differentiability of, 97
This action might not be possible to undo. Are you sure you want to continue?