\%.

hl (c)
AMBRIDGE TEXTS
IN APPLIED
MATHEMATICS

Nonlinear
Systems

P.G.DRAZIN
 9 1 %S7

1/99 ^ ^
STRAND PRICE
$ 18503

Nonlinear systems

5676 54-
 Cambridge texts in applied mathematics
Maximum and minimum principles: a unified approach with applications
M. J. SEWELL

Introduction to numerical linear algebra and optimization

P. G. CIARLET

Solitons: an introduction
P. G. DRAZIN AND R. S. JOHNSON

The kinematics of mixing: stretching, chaos and transport

J. M. OTTINO

Integral equations: a practical treatment,

from spectral theory to applications
D. PORTER AND D. S. G. STIRLING

Perturbation methods
E. J. HINCH

Nonlinear systems
P. G. DRAZIN
Nonlinear systems

P. G. DRAZIN
Professor of Applied Mathematics
University of Bristol

Cambridge
UNIVERSITY PRESS
Published by the Press Syndicate of the University of Cambridge
The Pitt Building, Trumpington Street, Cambridge CB2 1RP
40 West 20th Street, New York NY 10011-4211, USA
10 Stamford Road, Oakleigh, Melbourne 3166, Australia

© Cambridge University Press 1992

First published 1992

Reprinted with corrections 1993, 1994

Printed in the United Kingdom by Bell and Bain Ltd., Glasgow

British Library cataloguing in publication data available

Library of Congress cataloguing in publication data

Drazin, P. G.
Nonlinear systems / by P. G. Drazin.
p. cm. - (Cambridge texts in applied mathematics; 10)
Includes bibliographical references and index.
ISBN 0 521 40489 4-ISBN 0 521 40668 4 (pbk)
1. Nonlinear theories. 2. Differential equations, Nonlinear.
3. Chaotic behavior in systems. I. Title. II. Series.
QA427.D7 1992
515'.355-dc20 91-27971 CIP

ISBN 0 521 40489 4 hardback

ISBN 0 521 40668 4 paperback

AO
To Susannah and Adam
• ’ .
 Contents

Preface page xj

1 Introduction 1
1 Nonlinear systems, bifurcations and symmetry breaking 1
2 The origin of bifurcation theory 3
3 A turning point 6
4 A transcritical bifurcation 10
5 A pitchfork bifurcation 13
6 A Hopf bifurcation 19
7 Nonlinear oscillations of a conservative system 22
8 Difference equations 28
9 An experiment on statics 35
Further reading 37
Problems 38

2 Classification of bifurcations of equilibrium points 48

1 Introduction 48
2 Classification of bifurcations in one dimension 49
3 Imperfections 55
4 Classification of bifurcations in higher dimensions 59
Further reading 63
Problems 63

3 Difference equations 68
1 The stability of fixed points 68
2 Periodic solutions and their stability 72
3 Attractors and volume 74
3.1 Attractors 74
3.2 Volume 80
4 The logistic equation 81
5 Numerical and computational methods 94
6 Some two-dimensional difference equations 96
7 Iterated maps of the complex plane 103
Further reading 109
Problems 109
 viii Contents

*4 Some special topics 125

1 Cantor sets 125
2 Dimension and fractals 127
3 Renormalization group theory 132
3.1 Introduction 132
3.2 Feigenbaum’s theory of scaling 135
4 Liapounov exponents 140
Further reading 143
Problems 144

5 Ordinary differential equations 149

1 Introduction 149
2 Hamiltonian systems 153
3 The geometry of orbits 155
*4 The stability of a periodic solution 157
F urther reading 161
Problems 161

6 Second-order autonomous differential systems 170

1 Introduction 170
2 Linear systems 172
3 The direct method of Liapounov 178
4 The Lindstedt-Poincare method 181
5 Limit cycles 186
6 Van der Pol’s equation 190
Further reading 197
Problems 199

7 Forced oscillations 214

1 Introduction 214
2 Weakly nonlinear oscillations not near resonance:
regular perturbation theory 215
3 Weakly nonlinear oscillations near resonance 219
4 Subharmonics 225
Further reading 229
Problems 230

8 Chaos 233
1 The Lorenz system 233
2 Duffing’s equation with negative stiffness 246
*3 The chaotic break-up of a homoclinic orbit: Mel’nikov’s method 251
4 Routes to chaos 261
5 Analy sis of time series 264
Further reading 277
Problems 277
 Contents IX

* Appendix: Some partial-differential problems 283

Answers and hints to selected problems 290

Bibliography and author index 301
Motion picture and video index 309
Subject index 310
 ■

I
 Preface

This book is an introduction to the theories of bifurcation and chaos. It

treats the solution of nonlinear equations, especially difference and ordi¬
nary differential equations, as a parameter varies. This is a fascinating
subject of great power and depth, which reveals many surprises. It requires
the use of diverse parts of mathematics - analytic, geometrical, numerical
and probabilistic ideas - as well as computation. It covers fashionable
topics such as symmetry breaking, singularity theory (which used to be
commonly called catastrophe theory), pattern selection, chaos, predictabil¬
ity, fractals and Mandelbrot sets. But it is more than a fashionable subject,
because it is a fundamental part of the theory of difference and differential
equations and so destined to endure. Also the theory of nonlinear systems
is applied to diverse and countless problems in all the natural and social
sciences, and touches on some problems of philosophy.
The writing of the book evolved with lecture courses I have given to
final-year undergraduates at the University of Bristol and to graduates at
the University of Washington and Florida State University in the USA
over the last decade. I hope that others wili enjoy this book as our students
have enjoyed the courses.
Most of the equations treated in traditional mathematics courses at
university are linear. These linear algebraic, ordinary differential, partial
differential and integral equations are solved by various powerful methods,
which essentially depend upon the principle of superposition. However,
this book is about nonlinear equations. Using little more than linear alge¬
bra and advanced calculus, we shall introduce the theories of bifurcation,
imperfections, free oscillations, forced oscillations, and chaos. (A knowledge
of the elements of the phase plane is a desirable pre-requisite of Chapter 1,
but the phase plane is treated from the beginning in Chapter 6.) The treat¬
ment is suitable for final-year undergraduates or first-year postgraduates,
whether they major in mathematics, physics, chemistry, engineering, mete¬
orology, oceanography or economics, provided that they have already
mastered linear algebra and advanced calculus and they are eager to learn.

xi
XU Preface

My primary aim is to introduce simply the mathematical properties of

nonlinear systems as an integrated theory, rather than to present isolated
fashionable topics. A secondary aim is to give an impression of the diverse
applications of the theory, without detracting from the primary aim. The
approach is to discuss topics in as concrete a way as possible, using worked
examples and problems to motivate and illustrate general principles. Few
general results are proved, and many results are merely made plausible. I
have tried to tell the truth and nothing but the truth, not to tell the whole
truth, by quoting results and simplifying where it seems desirable; I hope
that I have not oversimplifed the material. For a single volume or lecture
course to cover so much requires a superficial treatment of many points. I
am conscious particularly that the analytic rigour is deficient and that the
treatment of nonlinear ordinary differential equations is less thorough and
systematic than in several good books. This is a price to pay for including
so much other material.
Chapter 1 is an elementary introduction to, and summary of much of,
Chapters 2, 3, 5 and 6. So an instructor might base a short course on
Chapter 1 with occasional excursions into other chapters, might go straight
through the book from beginning to end, or might start with Chapter 2
and then go straight on to the end. Chapter 4 contains miscellaneous topics
which, although of relevance and interest, might be omitted from a course.
There are, no doubt, other strategies on which a successful course may be
based.
More-advanced parts of the text, which might well be omitted at a first
reading, are denoted by asterisks; these parts are paragraphs, examples,
problems, subsections, sections, Chapter 4 and the Appendix. The end of a
worked example is denoted by a small square. The system of equation
numbers is such that (1.2.3) denotes the third equation of §2 of Chapter 1,
and so forth; the first number, denoting the chapter, is omitted when the
equation cited is in the same chapter as the citation, and the second num¬
ber, denoting the section, is omitted when the equation is in the same sec¬
tion as the citation. Thus equation (1.2.3) would be described as (3) in §1.2,
as (2.3) in §1.3, and as (1.2.3) in Chapter 2. The second problem of Chapter 3
is denoted by Q3.2, and its answer by A3.2, and so forth.
The problems are of very variable difficulty, but are ordered according
to the subject matter of the main text rather than according to their diffi¬
culty. Asterisks, hints, references and brief answers are provided to help
readers recognize and overcome the difficulties. Many applications of non¬
linear systems are introduced in the problems without any attempt to
study the modelling carefully. The aim is to emphasize the mathematical
 Preface xiii

aspects, but to motivate the mathematics by use of models and to lead

readers towards the modelling.
I thank Dr C. J. Budd (for some suggestions), Mr W. M. Challacombe
(for help in preparing Figs. 8.2, 8.3), Dr K. J. Falconer (for analytical
advice and criticism of a draft of the book), Miss Alice James and Mr M.
Woodgate (for help in preparing Fig. 3.11), Mr B. Joseph (for help in pre¬
paring Fig. 8.8), Mr. T. Milac (for help in preparing Fig. 3.6), Ms C. R.
Pharoah (for drafting and re-drafting so patiently almost all the figures),
Dr P. L. Read (for help in preparing Figs. 8.18, 8.19), Dr D. S. Riley (for
help in proof-reading), Dr Susan C. Ryrie (for much criticism of drafts of
this book and for help in preparing Figs. 8.14-17), Mr P. Shiarly (for help
in word processing and preparing Figs. 3.12, 3.13), Mr. K. Slater (for help
in preparing Figs. 4.5, 6.7), Dr M. Slater (for copious and careful criticism
of a draft of the book), and to the students of the University of Bristol, the
University of Washington and Florida State University with whom I have
learnt so much. Please do not blame them for the defects of the book
because, of course, the responsibility is mine alone.

Bristol P. G. Drazin

Additional note for 1994 reprinting

Some minor improvements and corrections have been made in this impres¬
sion. In particular, I thank Dr A. A. Wheeler for help in improving Fig. 7.3.
Also some problems have been added. I take the opportunity to express my
thanks to the staff of Cambridge University Press, especially the copy-editor
Mrs Bowring, for the encouragement they have given to me and for the high
quality of their work on the book. To see the quality, look around.

Bristol PGD
 • - !

'

. * '
 1
Introduction

Begin: to have commenced is half the deed. Half yet remains: begin again on
this and you will finish all.

Ausonius (Epigrams no. xv)

1 Nonlinear systems, bifurcations and symmetry breaking

A nonlinear system is a set of nonlinear equations, which may be algebraic,
functional, ordinary differential, partial differential, integral or a combina¬
tion of these. The system may depend on given parameters. Dynamical
system is now used as a synonym of nonlinear system when the nonlinear
equations represent evolution of a solution with time or some variable like
time; the name dynamical system arose, by extension, after the name of the
equations governing the motion of a system of particles, even though the
nonlinear system may have no application to mechanics. We may also
regard a nonlinear system as representing a feedback loop in which the
output of an element is not proportional to its input. Nonlinear systems
are used to describe a great variety of phenomena, in the social and life
sciences as well as the physical sciences, earth sciences and engineering.
The theory of nonlinear systems has applications to problems of econom¬
ics, population growth, the propagation of genes, the physiology of nerves,
the regulation of heart-beats, chemical reactions, phase transitions, elastic
buckling, the onset of turbulence, celestial mechanics, electronic circuits
and many other phenomena. This introduction to nonlinear systems, then,
is an introduction to a great variety of mathematics and to diverse and
numerous applications. We shall emphasize the mathematical aspects of
nonlinearity which arise so often in applications, many of which give rise to
surprising results. We shall see, in particular, that as a parameter changes
slowly a solution may change either slowly and continuously or abruptly
and discontinuously. A metaphor for such an abrupt change is the proverb
‘It is the last straw that breaks the camel’s back.’
In applications of the theory of nonlinear systems we are usually inter¬
ested in enduring rather than transient phenomena, and so in steady states.
Thus steady solutions of the governing equations are of special importance.

1
 2 Introduction

Of these steady solutions only the stable ones, i.e. those which, when dis¬
turbed slightly at some instant, are little changed for ever afterwards, cor¬
respond to states which persist in practice, and so are usually the only ones
observable. It follows that a state may change abruptly not only if it ceases
to exist but also if it becomes unstable as a parameter changes slowly.
We shall also see that as a result of instability a small cause may have a
large effect in the sense that a small disturbance at a given instant may
grow and become significant such that after a long time the behaviour of
the system depends substantially on the nature of the disturbance, however
small the disturbance was. For example, a spherical pendulum with the
bob finely balanced directly above its point of suspension may be dest¬
abilized by a jog, even by a gentle breath on it; further, the direction and
timing of the ensuing motion of the pendulum depend strongly on the
very small disturbance of the unstable position of equilibrium. Lorenz de¬
scribed this in a metaphor in which the unstable prairie atmosphere might
be triggered by the flutter of the wings of a butterfly in a distant jungle, and
thereby a devastating tornado might arise; this is called the butterfly effect.
A bifurcation occurs where the solutions of a nonlinear system change
their qualitative character as a parameter changes. In particular, bifurca¬
tion theory is about how the number of steady solutions of a system
depends on parameters. The theory of bifurcation, therefore, concerns all
nonlinear systems and thence has a great variety of applications. A bifurca¬
tion, contradicting Linnaeus’s assertion that ‘Nature does not proceed by
jumps’, may confound intuition, so many applications of the theory are
important. We shall see how a bifurcation of a nonlinear system and the
onset of instability of a solution usually occur at the same critical value of
a parameter governing the system. Bifurcation may also occur at other
values of the parameter.
Bifurcation is often associated with what is called symmetry breaking.
For this, the system has some sort of symmetry, i.e. the nonlinear problem
is invariant under some group of transformations. The symmetry implies
that the set of all the solutions is invariant under the transformations, but
not that each solution is invariant. Symmetry is broken at bifurcation if all
solutions are symmetric when a parameter is greater (or less) than a critical
value but some are asymmetric when the parameter is less (or greater)
than that value. For example, when a circular pond dries out, the mud
at the bottom shrinks, and the pattern of the cracks of the mud which
develops does not have circular symmetry. The causes of pattern selection
are closely related to the bifurcations and the stability of the solutions of
the governing equations.
 1.2 The origin of bifurcation theory 3

Also oscillations occur in many applications, so periodic solutions and

their stability are important too. Further, aperiodic unsteady solutions
may occur as seemingly random solutions with stationary statistical pro¬
perties, these are chaotic solutions or, simply, chaos. A chaotic solution also
may be stable in the sense that it persists even when the solution is per¬
turbed slightly at some time; this is a dynamic equilibrium. Much of the
book concerns chaos, although no formal definition of chaos is given for all
systems. It seems that, for the present at least, chaos is best treated like a
beard in the following sense. For most chins, there is no doubt whether a
chin is clean-shaven or sports a beard; but an occasional chin provokes
controversy as to whether it is bearded or its owner has temporarily omit¬
ted to shave. In short, there is general, but not complete, agreement on
what is and what is not a beard, and the word ‘beard’ is found useful in
describing chins. In a similar way, we shall learn how to recognize chaos
when we see it, but may not agree precisely what to call it.
The theory of stability and bifurcations is introduced in this chapter by
giving first some informal definitions and a little historical background
and then a few illustrative examples. Next, nonlinear oscillations and
difference equations are introduced. Thus this chapter provides an over¬
view of the whole book by detailed consideration of some, albeit elemen¬
tary, examples. These examples may at first be studied as separate prob¬
lems, but their importance lies in their being canonical cases typical of
wide classes of bifurcations. Most of the ideas and methods used later will
be raised in this chapter, although their later treatment will be more sys¬
tematic, more general and more thorough. However, chaos will not be
introduced until Chapter 3; it will be seen not as an isolated phenomenon
but as a common property of an important subclass of solutions of non¬
linear systems. Indeed, ideas arising in all the chapters will be seen to be
linked in many different ways. They will be built up bit by bit and drawn
together finally in Chapter 8. An overall theme is how the stable solutions
of a nonlinear system vary as a parameter is increased.

2 The origin of bifurcation theory

The ideas of bifurcation theory arose slowly and imperceptibly at first,
being almost as old as algebra itself. At the simplest, we may view the
quadratic equation,

x2 — a — 0,

as an example. The roots ±^/a are real for a > 0 and are a complex conju-
4 Introduction

Fig. 1.1 A sketch of the bifurcation diagram for the equation x2 — a = 0.

gate pair for a < 0. We may say that there is a change in the character of
the solutions at a = 0, where there is a repeated root x = 0. There is also
symmetry breaking in the sense that the quadratic equation is invariant
under complex conjugation, and that each solution is invariant under com¬
plex conjugation (i.e. is real) for a > 0 but no solution is invariant under
complex conjugation for a < 0. If we confine our attention to real solu¬
tions, then there are two for a > 0, one for a = 0 and none for a < 0. We
have sketched the parabola in the {a, x)-plane to illustrate the real solutions
in Fig. 1.1; this is an example of what is called a bifurcation diagram. The
horizontal axis gives the value of the parameter and the vertical axis gives
the variable to measure the solutions. We shall seen in §3 that this diagram
may be useful in describing the solutions of a closely related ordinary
differential equation.
Less-trivial examples arose naturally in the solution of problems of par¬
ticle dynamics. Thus bifurcations came to be considered in the seventeenth
century, although our point of view is rather different now. For one exam¬
ple (Andronow & Chaikin 1949, p. 75), the number and character of the
positions of equilibrium of a bead, constrained to move on a smooth circu¬
lar wire in a uniformly rotating vertical plane under the action of gravity,
depend on the rate of rotation, as is shown in Fig. 1.2 and in Q1.8.
It may, however, be said that the significance of bifurcations came to be
recognized first in the eighteenth century. Euler’s (1744) work on the equi¬
librium and buckling of an elastic column under a load and D’Alembert’s
work on the figures of equilibrium of a rotating mass of self-gravitating
fluid are the foundations of bifurcation theory.
 1.2 The origin of bifurcation theory 5

mau>2s\nO

Fig. 1.2 The bead on the smooth wire has one position (at 9 = 0) of stable
equilibrium if a>2 < g/a and two (with equal and opposite angles 6 =
± arccos(g/aoj2)) if co2 > g/a.

The problem which D’Alembert worked on has been studied from the
seventeenth to the twentieth century. Consider a mass of a uniform incom¬
pressible fluid which has constant uniform rotation and is subject to its
own gravitational attraction. If the fluid is in dynamic equilibrium as a
balance between pressure, centrifugal force and self-gravitation, then what
is the figure of the body formed by the fluid? Answers to this question have
been used to model a star and a planetary system in formation. Newton,
Maclaurin, Jacobi, Liouville, Dirichlet, Dedekind, Riemann, Poincare,
Liapounov, Elie Cartan and many other great mathematicians and scien¬
tists have worked on the problem. It has stimulated a lot of enduring
mathematics, although the physical applications of the problem have
proved disappointing because important physical processes (such as the
relation between the pressure and density of the star, and the generation of
energy in the interior) were neglected in the model. When the rotation is
weak, the figure is an oblate spheroid, first calculated by Maclaurin. As the
rate of rotation increases beyond a critical value (which, by dimensional
analysis, may be seen to be proportional to the square-root of the product
of the gravitational constant and the density of the fluid) ellipsoidal figures
of equilibrium, first calculated by Jacobi, occur with three different semi¬
axes. Symmetry about the axis of rotation is thereby broken. Poincare
(1885a, p. 270) first used the French word bifurcation in its present sense.
He found that a sequence of pear-shaped figures of equilibrium branches off
6 Introduction

Fig. 1.3 Sketches of some figures of equilibrium of a rotating self-gravitating

mass of fluid, (a) Maclaurin spheroid, (b) Jacobi ellipsoid, (c) Poincare pear-
shaped body.

the sequence of Jacobi ellipsoids, just as the Jacobi ellipsoids branch off the
Maclaurin spheroids (see Fig. 1.3). Poincare also did a lot of work on
the general mathematical theory of bifurcation. Later Liapounov (1906)
worked on the problem, and was thereby led to develop a general method
of perturbation in bifurcation theory, now known as the Liapounov-
Schmidt method.

3 A turning point
In this section and the next four sections we treat the most important types
of bifurcation by describing the steady solutions of some simple ordinary
differential equations, their dependence on a parameter, and their stability.
Stability is defined formally in Chapter 5; for the present an intuitive
notion of stability will suffice. The notion and importance of stability were
recognized by Lagrange in the eighteenth century, but it is difficult to
summarize them more clearly than Clerk Maxwell did in a lecture in 1873
(cf. Campbell & Garnett 1882, p. 440):

When ... an infinitely small variation of the present state will alter only by a
small quantity the state at some future time, the condition of the system,
whether at rest or in motion, is said to be stable; but when an infinitely small
variation in the present state may bring about a finite difference in the state
of the system in a finite time, the condition of the system is said to be unsta¬
ble.
It is manifest that the existence of unstable conditions renders impossible
the prediction of future events, if our knowledge of the present state is only
approximate, and not accurate.

For our first example to consider in detail, we take the ordinary differ-
 1.3 A turning point 7

ential equation,

We shall treat this only as a simple mathematical problem here, although

the importance and many applications of essentially similar problems may
be appreciated later. We shall often regard t as the time, although it is more
properly regarded as the independent variable of the equation.
An equilibrium point, or a critical point, of equation (1) is a steady solu¬
tion, i.e. a value, X say, such that there exists a solution x for which x{t) =
X for all t. Therefore dx/dt = 0 and it follows at once from equation (1)
that

X = ±Ja. (2)

Note that we seek only real solutions, so that there are two steady solu¬
tions for a > 0, one for a = 0, and none for a < 0. To examine the stability
of the solution x = X for a > 0, we rewrite equation (1) without approxi¬
mation as

dx'
— = ~(2X + x')x', (3)

where we define the perturbation of the equilibrium point as x' = x — X. It

is plausible that we may find the stability of the solution x = X for a > 0
by supposing that x' is small, linearizing the equation, and solving it. (In
fact, this method can be rigorously justified for quite general systems, ex¬
cept at the margin of stability, i.e. except for those values of a separating
linearly stable and unstable solutions. Here we can see that a small non¬
linear term cannot change the sign of the right-hand side of equation (3),
and so cannot change the stability, unless X = 0.) Therefore we consider
the linearized system,

dx'
— 2Xx’. (4)
df

This equation has the solution x'(f) = x’0est, where Xq is the given initial
value x'(0) of x' and the exponent is seen at once to be

s=-2X. (5)

It follows that if X = yja then x'{t) -*■ 0 as t -> oo for all x'0. Therefore all
small initial perturbations of the equilibrium point X = y]a remain small
for all time and the point is stable. We can similarly see that the point
 8 Introduction

X = —yja is unstable because a small initial disturbance grows until it

becomes no longer small.
The solution X = 0 for a = 0 is at the margin of stability of the lin¬
earized system, in the sense that there are arbitrarily close values of a for
which the solution is stable, and there are arbitrarily close values of a for
which it is unstable. However, the linearized system of a given nonlinear
system is not in general sufficient to determine the stability of the null
solution at the margin of stability itself, because the terms nonlinear in x'
might make unstable the solution which is just linearly stable. For this
simple equation (1), indeed, we may contradict the stability given by the
linear theory, because we may solve the nonlinear problem explicitly at the
margin a = 0 of stability. In that case equation (1) becomes

and we deduce readily that if x(0) = x0 then

(6)

0 as t -*• oo for all x0 ^ 0

— oo asr-> l/( — x0)forallx0 < 0

Therefore the equilibrium solution x = 0 for a = 0 is unstable, because a

small perturbation may grow so much that it eventually ceases to be small.
Indeed, this is an example of blow-up, for which a solution attains an
infinite limit in a finite time.
It is convenient to illustrate the steady solutions in a bifurcation dia¬
gram, i.e. to plot the equilibrium points versus the value of the parameter in
the (a,x)-plane. The diagram for equation (1) is shown in Fig. 1.4. It is
conventional to draw stable steady solutions as continuous curves and
unstable ones as dashed curves. We have also added broad arrows to indi¬
cate the evolution of a solution x(t) with time for a fixed value of a.
The point (0,0) is an example of what is called a simple turning point, a
fold, or a saddle-node bifurcation (cf. Q6.3). The number of steady solutions
changes as a increases through zero, and the stability of the solution
changes as X increases through zero in this example. The function a(X),
defined by 0 = a- X2, is well-behaved for all X, even though X{a) is a 2-1
function for a > 0 and does not exist for a < 0. We shall see that turning
points with the same qualitative character arise frequently in many kinds
of nonlinear systems.
This example is so simple that even if a * 0 we may again confirm our
results by explicit integration of equation (1). By separation of variables.
 1.3 A turning point 9

Fig. 1.4 A turning point: the bifurcation diagram for dx/dt — a — x2. The stable
solutions are denoted by a continuous curve and the unstable by a dashed
curve. The arrows indicate the directions of change of solutions with time.

we see that

It follows that

i/2 (x o + fll/2 tanh(a1/2f)\

if a > 0
\u1/2 + x0tanh(a1/2f)/
x(t) = - (7)
_ -u/2 (xo ~ ( —a)1/2 tan{( —u)1/2f}\
if a < 0
V(-fl)1/2 + x0tan{(-a)1/2f}/

This gives the solution x(f) for all t ^ 0, or, if the solution becomes
singular, until the solution becomes singular, in terms of the given initial
value x0 of x. It follows that

a112 as t oo if a > 0, x0 > —a112

— a112 as t oo if a > 0, x0 = — a1/2

x(t) > .
—oo as t -*■ (a)~1/2 arctanh( — all2/x0) if a > 0, x0 < —a112

—oo asf-»( —a)_1/2arctan{—( —a)1/2/x0} ifa<0

(8)
Note that the limits are independent of the initial value x0 of x, although
which limit is attained does depend on x0. Note also that the solution
10 Introduction

may become infinite after a finite time t, the value of t at the singularity
depending on x0 as well as the parameter a which occurs in the differential
equation; such a movable singularity is characteristic of nonlinear equa¬
tions, in contrast to the fixed singularities of linear equations. Fig. 1.4
shows the limits more easily, but the formulae provide extra information
about the time that is required for x(t) to progress to its infinite limit.

4 A transcritical bifurcation
The logistic equation,

dn , 7
— = an - bn1, (1)
dt w

arose as a simple model in the theory of population growth (Verhulst 1838,

Pearl & Reed 1920). It may represent the growth of a population of a given
species, the number of individuals of the species being approximated by the
real variable n. When n is sufficiently small the population grows or dies
out exponentially according to whether a is positive or negative respective¬
ly. If the population grows then, after a while, it will become so numerous
that its food supply becomes inadequate or its predators thrive and there¬
fore its growth rate will be reduced. It follows that the exponential growth
rate will be moderated in a way represented, at least approximately, by the
nonlinear term for some positive value of b. However, for a mathematical
exposition, we shall here not restrict the signs of n and b.
The equilibrium points are

n - 0 for all b, and n = a/b for all b # 0, for all a. (2)

To examine the stability of the null solution, we linearize equation (1) and
find

dn
= an. (3)
dr

Therefore we take n(t) oc es' and deduce that

s = a. (4)
This at once gives the general solution

n(t) = n0eat, (5)

where n0 is the value of n at t — 0. This solution decays exponentially to
 1.4 A transcritical bifurcation 11

n
/\ /

'> ''
/ ' 0 a

\' 'f

Fig. 1.5 The bifurcation diagram of the linearized system dn/dt = an in the
(a, n)-plane. Note that the n-axis represents all the stable solutions n = n0 for
a = 0. Arrows indicate the directions of change of some solutions with time.

zero if a < 0, grows exponentially to infinity if a > 0, and remains constant

if a = 0. Therefore the null solution of the linearized system (3) is stable if
a ^ 0 and unstable if a > 0, as is illustrated in the bifurcation diagram of
Fig. 1.5.
We may similarly show that the solution n = a/b is unstable if a < 0 and
stable if a > 0.
In fact we may solve the nonlinear equation (1) explicitly in simple terms.
It may be solved by separation of variables. However, it is a Bernoulli
equation which may also be solved by rewriting it without approximation
as

1 dn a
= b.
n2 dt n

i.e.

d(l/n) a
+ - b
dt n

(It can easily be verified that our conclusion (6) is true even if n = 0 and \/n
does not exist.) This is a linear equation in \ jn with elementary solution,

1 if a # 0

n(t)
rig1 + bt if a = 0

where n0 = n(0) # 0. Therefore

12 Introduction

ann
if a # 0
bn0 + (a — bn0)e~
n{t) = < (6)
«o
if a = 0
1 + bn0t

\a/b if a > 0 and bn0 > 0]

as t -* oo. (7)
[ 0 if a ^ 0 and bn0 > a\

Also
In {1 + a/{-bn0)} if a # 0, bn0 < min(0, a))
n(t) -*■ +oo as t
1 K~bn0) if a = 0, bn0 <0 J ’
(8)

the limit being -oo if b > 0 and +oo if b < 0. Note that, as in §3, the limits
are independent of the initial value n0 of n, although which limit is attained
does depend on n0, and that the solution may become infinite after a finite
time, the value of t at the singularity depending on n0 as well as the coeffi¬
cients a and b which occur in the differential equation.
In summary, we draw the bifurcation diagrams for the two cases b > 0
and b < 0 in Fig. 1.6. Again, the bifurcation diagrams give most of the
essential results easily and clearly. This type of bifurcation, characterized
by the intersection of two bifurcation curves, is called a transcritical bifur¬
cation. Examples of transcritical bifurcation occur for many types of non¬
linear system, not only for ordinary differential equations. Of course, the
case b = 0 is given in Fig. 1.5. Also, if the model were of the growth of
a population of n individuals, then we would impose the condition that
n ^ 0, and so only the upper halves of the bifurcation diagrams would be
relevant.

(a) (b)
Fig. 1.6 Transcritical bifurcations: the bifurcation diagrams of the system
dn/dt = an- bn2 in the (a, r+plane for (a)b> 0 and (b) b < 0.
 1.5 A pitchfork bifurcation 13

You may have noticed that equation (1) has the same form as equation
(3.3), which is an exact transformation of equation (3.1). The parameters a
and b in equation (1) are the analogues of — 2yja and 1 respectively in
equation (3.3). However, we use a as the abscissa in each bifurcation dia¬
gram, and insist on the reality of the abscissa. Thus our taking the positive
square-root yja in §3 gives a turning point, whereas there is a transcritical
bifurcation in this section.

5 A pitchfork bifurcation
Landau (1944), investigating nonlinear stability of steady flow of a New¬
tonian fluid, postulated a model equation which is essentially

^ = ax - bx\ (1)
dr

where a, b are real parameters. Landau’s model was substantiated in the

1960s, but we shall not dwell on the modelling here. We shall merely take
equation (1) as an instructive mathematical example and solve it. Note that
it differs from equation (4.1) in having a cubic rather than a quadratic
nonlinearity.
The equilibrium points are easily seen to be

x = 0 for all a, (2)

and

x — ±A for all a, b such that a/b > 0, (3)

where we define A = (a/b)112.

To examine the stability of the null solution, take the linearized system,

dx
= ax, (4)
dr

and deduce the general solution,

x(r) = x0efll, (5)

as before, where x0 is the initial value of x. Therefore the null solution of

the linearized system is stable if and only if a ^ 0. This result is illustrated
in Fig. 1.5. However, we shall see very soon that when a = 0 the null
solution of the nonlinear equation (1) is stable if b > 0 and unstable if
b <0.
In fact we can again solve the nonlinear equation explicitly, replacing
14 Introduction

the dependent variable x of equation (1) by 1/x2, finding a linear equation

in 1/x2, and deducing the elementary solution. Thus we rewrite equation
(l)as
1 dx a
x3 dr x2

i.e.
dx-2
+ 2ax~2 = 2b,
dr

and deduce that

fb ■ 2 at
+ I *o2-e if a # 0
x 2(r) = < a ",
x0 2 + 2 bt if a = 0
J

where x0 is the value of x at t = 0. Therefore

ax"
if a ^ 0
bxo + (a — bxo)e 2at
x2(r) = - (6)
xl
if a = 0
1 + Ibx^t

This gives x(r) = {x2(r)}1/2 sgnx0, where the sign function is defined by
sgnx = 1 if x > 0, sgnx = 0 if x = 0, and sgnx = -1 if x < 0. Now vari¬
ous cases arise for separate treatment.

(i) b > 0 and a > 0. In this case it follows that

x(r) -* A sgnx0 as r -> oo.

The limit of x has the same sign as x0 but is otherwise independent of the
initial value x0 of x. It can be seen that the nonlinear term -bx3 of equa¬
tion (1) reduces the exponential growth of small unstable disturbances of
the null solution eventually, and leads to attainment by x(r) of the constant
value A or — A. This is called equilibration. The temporal development of x
is illustrated in Fig. 1.7. This case is an example of what is called bistability,
i.e. of the existence of two stable equilibrium points of the same system.

(ii) b > 0 and a < 0. In this case the nonlinear term in equation (1) rein¬
forces the exponential damping of the linearized system, albeit very weak¬
ly, and

x(r)->0 asr-»co for all x0.

 1.5 A pitchfork bifurcation 15

Fig. 1.7 A sketch of the development and equilibration of some solutions of

equation (1) when a, b > 0.

Fig. 1.8 A supercritical pitchfork bifurcation: a sketch of the bifurcation dia¬

gram in the (a, x)-plane for b > 0. Arrows indicate the directions of change of
some solutions with time.

In summary of cases (i) and (ii) for b > 0, we draw the bifurcation dia¬
gram of Fig. 1.8. There is a bifurcation at the origin, with a unique real
point x = 0 of equilibrium for a ^ 0 but three points of equilibrium for
a > 0. The null solution is stable if and only if a ^ 0, and the solutions
x = ± A are stable whenever they exist, i.e. for all a > 0. There is said to be
supercritical bifurcation in this case, because the bifurcated solutions arise
16 Introduction

as a increases above its critical value, zero, at which the null solution is
marginally stable.

(iii) b < 0 and a > 0. In this case the nonlinear term on the right-hand side
of equation (1) has the same sign as the linear one, and reinforces the
exponential growth of small disturbances of the null point. It can be seen
from equation (6) that the solution ‘blows up’ after a finite time, i.e.

x(t) -*• oo as t -* (2a)-1 ln(l — a/bxq).

(iv) b < 0 and a < 0. In this case the two terms on the right-hand side of
equation (1) have different signs and so may have zero sum. Therefore the
equilibrium points x — ±A exist, and they act as ‘thresholds’ so that if
— A < x0 < A then x(f) -> 0 as t -*• oo but if ±x0 > A then x(r) -*• +oo
respectively as t -*■ ( —2a)_1 ln{xo/(xo — A2)}. This is illustrated in Fig. 1.9.

In summary of cases (iii) and (iv) with b < 0, we draw the bifurcation dia¬
gram Fig. 1.10. There is a unique steady solution x = 0 if a ^ 0 but three
steady solutions x = 0, ± A if a < 0. The null solution is stable if and only
if a < 0, and the two solutions x — ±A are unstable whenever they exist,
i.e. for all a < 0. There is said to be subcritical bifurcation. There is also said

Fig. 1.9 A sketch of the development and equilibration of some solutions when
a, b < 0.
 1.5 A pitchfork bifurcation 17

Fig. 1.10 A subcritical pitchfork bifurcation: a sketch of the bifurcation dia¬

gram in the (a, x)-plane for b < 0. Arrows indicate the directions of change of
some solutions with time.

to be subcritical instability of the null solution to disturbances of finite

amplitude, because a solution grows if xl > A2, although the null solution
is stable to disturbances for which *o < A2, when the parameter a is less
than its critical value, zero, for instability.
These are examples of pitchfork bifurcations, named after the appearance
of the bifurcation diagrams in Figs. 1.8 and 1.10. Pitchfork bifurcations
often occur for systems with some symmetry, as a manifestation of symme¬
try breaking. Equation (1) is of invariant form if x is replaced by — x, so is
the pair of solutions x = A, — A, but a single solution, either x = A or
x = — A, is not invariant. We have in fact met an example of a pitchfork
bifurcation already in §2: the bead on the circular wire in a vertical plane
which rotates has one or three positions 6 of equilibrium according to
whether the dimensionless parameter a(o2/g is respectively less or greater
than one; the bifurcation is described locally by a pitchfork although the
differential equation is of second order (Q1.8).
Equation (1) can readily be seen to be equivalent to the equation,

= 2ax2 — 2 fix4, (7)

dt
and therefore to

^ = 2 ay- 2 by2,
 18 Introduction

where y = x2 ^ 0. Thus there is the upper half of a transcritical bifurcation

in the (a, y)-plane for fixed b, as is shown in §4.
In summary of §§3-5 we note that each example is a treatment of a
first-order differential equation of the form

^ = F(a,x), (8)

which admits a steady solution such that x(r) = X for all t if

F(a,X) = 0. (9)

This determines the equilibrium points X for each value of a; there may be
no zero X of F, one zero or many zeros according to both F and the value
of a. It is informative to plot all the curves x = X(a) in the (a,x)-plane
as the bifurcation diagram. In considering the stability of an equilibrium
point X, the value of a is fixed, and so it is convenient to suppress a in the
analysis below. Suppose then that

% = FM (10)

for a well-behaved function F, and F(X) = 0. To find the evolution of all

small perturbations x' = x - X of X, expand as a Taylor series

F(x) = F(X + x')

= F(X) + x'F'(X) + 0(x'2) as x' - 0

- x'F'(X) + 0(x'2),

and neglect the very small terms of order x'2. Thus the linearized form of
equation (10) is

dx'
= (id
This has general solution

x(t) = Xoes(, (12)

where s = F (2Q. It follows that if F'(X) < 0 then all small perturbations of
the equilibrium point decay monotonically ast->oo, and X is stable, and
also that if F’(X) > 0 then all small perturbations grow monotonically and
so X is unstable. The evolution of all solutions of equation (10) can be
traced globally by considering the signs of F(x).
The illustrations of a turning point, a transcritical bifurcation and a
pitchfork bifurcation in §§3-5 by use of first-order equations give the es-
 1.6 A Hopf bifurcation 19

sence of these bifurcations of steady solutions for higher-order equations.

However, the time dependence of the solutions we have found is rather
special; for example, we shall see in §7 that the equation

d2x
■jt = ax — bxJ
dr
has unsteady solutions which are rather different from those of equation
(1). But first we shall see in the next section that some bifurcations which
may occur for second- and higher-order systems do not occur for first-
order systems.

6 A Hopf bifurcation
For the next example we shall consider the system of ordinary differential
equations,

dx dv
_ = -y + (fl - x2 - y2)x, — = x + (a - x2 - y2)y, (1)

for all real a. The solution will introduce an important aspect of bifurca¬
tion: the branching of a time-periodic solution from a steady one.
First seek the equilibrium points and test their stability, as before. On
putting dx/dt = dy/dr = 0, it follows readily that the only steady solution
is the null solution,

x = 0, y = 0. (2)

Investigating the stability of this solution, we linearize equations (1) for

small x and y to find

dx
— = -y + ax, —
dr
,
dy = x -(- ay.
(3)

The general solution of this system is a linear combination of the normal

modes with x, y oc est, say x(t) = estu, y(t) = es,v. These satisfy the equation

Ju = 5U (4)

to determine the eigenvalues s and eigenvectors u = [u, r]T of the 2x2

matrix

J=

Therefore
20 Introduction

0 = det(J - si)

a —s —1
1 a — s

= (a — s)2 + 1.

Therefore
s = a ± i. (5)

Therefore there is linear stability if Re(s) < 0 for both eigenvalues, i.e. if
a < 0, and instability if a > 0.
This example has been chosen so that again it may be solved explicitly
in simple terms. If we use polar coordinates so that x = r cos 9, y = r sin 6
for r ^ 0, then x + iy = re'6 and so

d(rel9) dx . dy
dt dt + 1 dt

= — y + ix + (a — x2 — y2)(x -I- iy), (6)

i.e.

+ ^r^je'e = *re'e + (<* ~ r2)rew.

Therefore, on dividing by e‘° and equating real and imaginary parts,

^ l 2 \ de , (7)
dr = r(a-r)’ dr1-

It follows from the method of §5 that

arl
2\a-2at
if a # 0
ro+(<*~ We
r2(t) = ^ (8)
r
ro2
if a = 0
1 + 2rlt

d — t -f- 90, (9)

where r — r0, 0 = 60 at t = 0. Note the similarity of equation (5.1) and
equation (6), which can be written as

dz
= iz + (a - \z\2)z, (10)
dt

where z = x + iy (seealsoQl.il).
 1.6 A Hopf bifurcation 21

Fig. 1.11 A Hopf bifurcation: sketches of phase portraits of the system (1) in the
(x,y)-plane. Arrows indicate the directions of some orbits as time increases,
(a) a 0. (b) a > 0.

A solution can be represented as a point x = (x, y) by use of Cartesian

coordinates in a plane. Then, for a ^ 0, all solutions x(t) -► 0 as t -> oo.
Each trajectory or orbit, i.e. each locus of the point \(t), spirals into the
origin of the (x,y)-plane counter-clockwise as oo. See Fig. 1.11 for a
phase portrait of the orbits in the (x,y)-plane. For a > 0 the origin becomes
an unstable focus. A new stable periodic solution, namely

x = Jacosit + 90), y = Jasinit + 0o), (11)

arises as a increases through zero. Such a solution is called a limit cycle,

because it is a periodic solution approached by other solutions in the
limit as f -> oo. It is represented by a closed curve, in this case the circle
x2 + y2 — a, in the phase plane of the orbits. It can be seen in Fig. 1.12 that
there is supercritical bifurcation in this example, the periodic solutions
bifurcating from the null solution as a increases through zero. (For details
of the theory of the phase plane see Chapter 6, where a more thorough
account of two-dimensional differential systems is given ab initio.)
The essence of a transcritical bifurcation and a pitchfork bifurcation is
that the real eigenvalue s of the unique least stable mode increases through
zero as a parameter a increases (or decreases) through a critical value, say
ac, and one or two new steady solutions arise. In contrast, for a Hopf
bifurcation, exemplified in this section, the real part of a pair of eigenvalues
of the least stable complex conjugate modes increases through zero as a
increases or decreases through ac, and a time-periodic solution arises; these
points are illustrated schematically in Fig. 1.13. Hopf (1942) showed that
22 Introduction

Fig. 1.12 Sketch of the bifurcation diagram in the (a, x)-plane for the system (1).
The dotted curves x = ±Ja give the maximum and minimum of x(t) over a
period of the stable oscillation.

(«) (b)

Fig. 1.13 Behaviour of eigenvalues s of the linearized systems at bifurcation as

the parameter a moves through its critical value ac. Dots denote the location of
s when a = ac, and arrows the directions of change of s as a increases (or
decreases), (a) Saddle-node, transcritical or pitchfork bifurcation, (b) Hopf
bifurcation.

this type of bifurcation occurs quite generally for systems of nonlinear

differential equations.

7 Nonlinear oscillations of a conservative system

We have noted that first-order differential equations may contain the es¬
sence of saddle-node, transcritical and pitchfork bifurcations but are too
 1.7 Nonlinear oscillations of a conservative system 23

simple to exhibit other bifurcations, or to describe oscillations which often

occur with higher-order systems. Indeed, we have already seen that a Hopf
bifurcation occurs only for a system of at least two first-order equations. (It
is shown in Chapter 5 that a second-order equation is equivalent to a
system of two first-order equations.) To introduce nonlinear oscillations
now, we shall take another example, namely equations of the form

We shall assume that / is as smooth as the arguments below require. This

equation may be regarded as the equation of motion of a particle moving
in a line under the force field / per unit mass with potential V, where

(2)

on taking some lower bound x0 for definiteness.

An equilibrium point X is a zero of /, i.e. a local extremum of V To
consider the stability of the point X, define x' = x — X, and linearize equa¬
tion (1) for small perturbations x' to get

(3)

This gives simple harmonic motion, and therefore stability, if f'(X) < 0,
but exponentially growing solutions in general, and therefore instability, if
f'(X) > 0. These results may be compared with those for the first-order
equations of the form dx/dt = /(x) in §§3-5. In each case the equilibrium
solutions X are the zeros of /. However, the solution x(t) approaches or
leaves X monotonically for a first-order equation according as f'{X) is
negative or positive respectively, whereas x(t) oscillates about or (in gener¬
al) leaves X according as f'(X) is negative or positive respectively. For the
second-order system (1) the point (X,0) is a centre in the phase plane of
(x, dx/dt) if f\X) < 0, and a saddle point if f'{X) > 0.
Equation (1) has the elementary integral

(4)

where £ is a constant of integration such that E = ^Xq if x = x0 and

dx/dt = x0 at t = t0. This is the energy equation of a particle in the force
field /. It shows that F(x) ^ E. Also equation (4) can be written as
2

(5)
 24 Introduction

where F(x) = E — F(x). This specifies the magnitude of dx/dt in terms of x

for each solution but not the sign of dx/dt. Therefore x increases or de¬
creases monotonically until it approaches a zero of F; this shows that the
behaviour of x(r) near a zero of F is important.
Suppose then that x approaches a simple zero X! of F, so F(xx) = 0 and
/(x^ = F'ixx) # 0. Therefore

dx
= ±{2f(xl)(x - xj}1/2 + 0(|x-X!|3/2) as x -+ x1;

so x — Xj has the same sign as /(xj). We can integrate this iteratively to

deduce that

x(t) = x1 +if(xi)(t-t1)2 + 0{(t-t1)*} • as t-*tu (6)

where is the time when x = xl9 i.e. x(tj) = x1. Therefore x has a simple
minimum or maximum xx at tt according as /(x2) is positive or negative
respectively.
Next suppose that x approaches a double zero xx of F, so F(xx) =
F'(xj) = 0and/'(xi) = /’"(xj) / 0. Now equation (5) gives/'(x^ > Oand

-fa = ±{/'(xi)}1/2(^ ~ ^i) + O {(x-x^2} as x -* Xj.

On integration this gives

x(t) - Xj ~ constant x exp[± {/'(xj)}^^ asf->+oo. (7)

respectively. Thus a solution may approach the limit Xj, the double zero of
F, in an infinite time. Note that in this case Xj is an unstable point of
equilibrium which may be approached, by a special solution, as time tends
to infinity (see in Example 1.2 how a simple pendulum may come to rest at
its unstable point of equilibrium directly above its point of suspension if its
initial velocity has a special relationship to its initial angle).
These ideas give the qualitative character of many solutions. Suppose,
for example, that F(x) > 0 for Xj < x < x2 and F has simple zeros Xj and
x2. Taking X! < x0 < x2 and x0 > 0, without loss of generality, we see that
x increases monotonically until it reaches x2. There it changes direction in
accord with equation (6), and then decreases monotonically until it reaches
Xj. Thereafter it changes direction at xx, increases monotonically, and
passes through x0. When it passes through x0 it will have the same value x0
of dx/dt as it did initially, so that the same motion will repeat itself, giving a
periodic solution. The period is
 1.7 Nonlinear oscillations of a conservative system 25

T= dr
o

where the circle on the integral sign denotes that the integration is to be
taken ‘around’ the orbit for one period.

(8)

because dx/dt is positive (at the same point x) over one half of the period
and negative but with the same magnitude over the other half.
We can picture the solutions of the system (1) by regarding it as the
equation of motion of a particle of unit mass on a smooth plane wire with
equation z = V(x)/g, where z is the height, x is the distance along the wire,
and g is the acceleration due to gravity. This gives the energy equation (4),
shows that the equilibrium points where dF/dx = 0 are stable if F has a
minimum and unstable otherwise, and makes apparent the occurrence of
nonlinear oscillations about stable points. This is illustrated in Fig. 1.14 of
the graph of V(x) — E = — F(x); note that the motion is confined to values
of x where F ^ 0.
Note that the equations dx/dr = /(x) and d2x/dr2 = /(x) have the same

Fig. 1.14 Sketch of a graph of a ‘typical’ function V(x) — E for a given motion.
Motion may occur only where K(x) ^ E. Other motions for the same field may
occur for different values of E.
26 Introduction

equilibrium points, namely the zeros of /. Moreover a zero X is stable for

each equation if f'(X) < 0 and unstable for each equation if f'(X) > 0. It is
for these reasons that the equations dx/dt = F{a,x) and d2x/dt2 = F(a,x)
have the same bifurcation diagram. So we meet turning points, trans-
critical points and pitchforks for second- as well as for first-order equa¬
tions. However, the temporal development of solutions of second-order
equations is different from that of solutions of first-order equations. In
particular, stability of an equilibrium point is monotonic for a first-order
equation but simple harmonic motion for the corresponding second-order
equation.

Example 1.1: a turning point. Suppose that

Then there are equilibrium points at x = ± yja for all a ^ 0, as in §3,

but the solutions do not all vary monotonically with time. The ‘energy’
integral,

jv2 + jx3 — ax = E,

gives the orbits in the phase plane of x and v = dx/dt, different orbits being
in general given by different values of E. If a > 0 then E = + §a^Ja at
the equilibrium points (± yja, 0) respectively, and (yja, 0) is a centre and
(- yja, 0) is a saddle point. The closed orbit through (— Ja, 0) is said to be a
separatrix, because it separates the closed orbits, which represent periodic
solutions, from unbounded orbits. Note also that the saddle point is con¬
nected with itself by the separatrix; for this reason the separatrix is an
example of what is called a homoclinic orbit. The separatrix has equation

W = ^(x + %/a)2(2v/a - x).

A periodic solution has period

r=2 P"__
J*. {2(£ + ax - jx3)}1/2’

where xt and x2 are the zeros of E + ax - ^x3 for -fa^/a < E < f a^Ja.
Note that T -* 00 as E ± %aja, the separatrix being a limit of periodic
orbits as the period tends to infinity. If a 0 then all the orbits are un¬
bounded and x(t) -*• -00 as t 00, except for the unstable point of equilib¬
rium at the origin when a = 0. The phase portraits for a > 0, a = 0 and
a < 0 are sketched in Fig. 1.15. □
 1.7 Nonlinear oscillations of a conservative system 27

(c)
Fig. 1.15 Sketch of the phase portrait of d2x/dt2 = a x2 for (a) a > 0,
(b) a = 0, and (c) a < 0.

Example 1.2: the simple pendulum. The equation of motion of a simple

pendulum of length / is

,d20
1-^2 =-g sine, (9)

where 9 is the angle the pendulum makes with the downward vertical, and
has energy equation

1/d 9
= g(cos 9 — cos a)/l, (10)
2\dt

where the amplitude a of the solution is defined as the positive value of 9

where d0/'dt = 0. See the phase portrait in Fig. 1.16. Note the saddle points
at ((2n + l)7r,0) for /i = 0, ±1, ±2,..., and the closed orbits in the region
between each pair of orbits, called heteroclinic orbits, connecting neigh-

plane.
 28 Introduction

bouring saddle points. Now each closed orbit represents a periodic solu¬
tion. This gives a nonlinear oscillation of period

de
T= 2
a {2g(cos9 — cos a)//}1/2

d0
=4 mw
!0 {4(sin2^a — sin210)}1/2

f*72 2sinjucos^d^
= 4(1/0):1,2
|0 cos^0{4(sin2ja — sin2 ja sin2 f)}1/2 ’

on substituting sin = sin \a sin <f>,

= 4(//sr)1/2 'Kl2 d(t>

0 cos

= 4(//0)1/2K(sin2ia), (11)
where the complete elliptic integral of the first kind is defined as
'n/2
d<f>
K(m) = for 0 ^ m < 1. (12)
(1 — msin2 <f>)1/2

Therefore
rn/2
T = 4{l/g)1/2 I (1 -I- |sin2 jasin2(/> + 0(sin4^a)} d<f> asa->0

= 4(//0)1/2{i7i + ^a27r + 0(a4)}

= 27t(//0)1/2{l -I- i^a2 + 0(a4)}. (13)

This gives the leading term in the correction to the period 2n(l/g)112 of
oscillations of small amplitude. Note that the period increases with ampli¬
tude, as we might have anticipated, because the pendulum moves more
slowly as it rises. □

8 Difference equations
To emphasize that there are nonlinear systems other than ordinary
differential equations, we shall next introduce nonlinear difference equa¬
tions. We shall see that they have bifurcations which are very similar to
those of ordinary differential equations, and that many concepts and
methods are useful in understanding both kinds and, indeed, other kinds
of nonlinear systems. Difference equations are mathematically interesting
 1.8 Difference equations 29

and have many important applications. They also are, in many ways, more
elementary and fundamental than differential equations, involving a dis¬
crete independent variable rather than a continuous one; therefore they
perhaps deserve more attention than they usually get from instructors and
students, although the applications of differential equations are even more
important.
A difference equation, recurrence equation or iterated map is in general
of the form

xn+i = F(x„,n) for n = 0,1,2,..., (1)

where x„ e IRm and F: IRm x Z —► IRm. Thus difference equations are func¬
tional, sometimes algebraic, systems which correspond to differential equa¬
tions. We regard the integral variable n as corresponding to the indepen¬
dent real variable t of a differential equation, take x0 as a given initial value
of the dependent variable x„, and consider the behaviour of x„, especially
as n —> oo. We shall further study functions F which also depend on param¬
eters, and how the solutions {x„}, i.e. the sequences {x0, xl5...}, change
qualitatively with those parameters.
In fact a difference equation may arise from taking a finite-difference
approximation to a differential equation, or a differential equation may
arise as a continuum approximation to a discrete process. For example, if
x(t) satisfies the logistic differential equation,

— = ux - fix2, (2)

then we may choose h as a suitable small positive number, define x„ =

x{nh) for n = 0, 1,..., and approximate the derivative dx/dt at t = nh by
the Euler forward difference (xn+1 — xn)/h. It follows that

x„+1 - x„
= ax. bxl

approximately, i.e. that

xn+1 = F(a,b,h,xn), (3)

where F(a, b, h, x) = x + h(ax — fix2). This little example illustrates the im¬
portant point that most computers are digital rather than analogue, deal¬
ing with discrete rather than continuous variables, and so it is the rule for a
differential equation to be ‘solved’ by first taking a difference equation
which approximates it and then solving the difference equation. In this way
we can compute the stable equilibrium points of a differential equation of
30 Introduction

the form

T,=m <4>
where xeIR and /: IR -*• 1R, by iterating the solutions of the difference
equation

*„+i = xn + hf(x„). (5)

Numerical analysts have studied the efficiency of this method and its con¬
vergence as the ‘time’ step h-* 0. The method may be easily generalized for
systems of ordinary differential equations with x e !Rm and f: [Rm -> IRm, and
for functions f which depend on t as well as x. Also other finite-difference
approximations to the differential equation may be used successfully.
Yet if we seek only equilibrium solutions of equation (4) then we need
seek only the zeros of/ which may be computed by use of any one of many
other difference equations. A well-known equation follows by the Newton-
Raphson method. For this we assume that/is continuously differentiable,
estimate xn as a zero X of / and construct an improved estimate xn+1
by approximating the curve with equation y = f(x) by its tangent y =
f(x„) + (* — xn)f (x„) at the point (x„,/(xn)). This gives the next approxi¬
mation to X as

*n+i =xH- f{xn)/f'{xn) (6)

if /'(*„) # 0. We guess x0 from what knowledge of / we have at the start,

and compute successive approximations xl5 x2, ..., stopping when we
have reason to believe that x„ approximates the exact zero X with the
desired accuracy. The choice of x0 and the convergence of the process as
n -* oo are considered in many textbooks of numerical analysis.
Differential equations also determine difference equations in a very
different way. Suppose that we consider, for example, the solutions x(t) e
IR3 of a given differential system,

dx
* - G(x)’ <7>

where G: IR3 -*• IR3. Note that this system is autonomous, i.e. G does not
depend on the independent variable t explicitly. Then we may take any
point x„ = (x„,yn,0) in the plane with equation z = 0, and define another
point xn+1 in the plane as follows. Set x(0) = xn, integrate the differential
system to find the orbit of x in the phase space, i.e. find the locus of x(t) in
(R for t > 0. Let xn+1 be the point where the orbit next crosses the plane
 1.8 Difference equations 31

Fig. 1.17 A sketch of the return map.

z = 0 in the same sense. Of course, the orbit may never cross the plane
again, so xn+1 may not exist. In this way we define a map F: !R2 R2 such
that xn+1 = F(x„). It is called the Poincare map or the first return map of the
plane z = 0 for the differential system (7), as illustrated in Fig. 1.17. We can
similarly define a return map for any non-planar smooth surface in R3. We
can also define return maps for a differential system of order higher than
three, for which the phase space has dimension higher than three. Return
maps are very useful as distillations of the complicated orbits in a higher¬
dimensional space, even though they are found by discarding much infor¬
mation about those orbits (Poincare 1882, p. 253).

Example 1.3: a ‘stroboscopic’ return map. A Poincare section may also be

taken in a line. Thus consider the system

= -y + (a-x2 - y2)x, — = x + (a - x2 - y2)y,

and define the Poincare map P. R -> R as the map of successive intersec¬
tions of the orbit x(t) and the x-axis in the same sense. We solved the
system in §6, and so can see that if x(0) = (x„,0) then 9 — t if x„ > 0 and
9 = t + n if x„ < 0. Therefore the orbit cuts the x-axis in the same sense, i.e.
from below if x„ > 0, next when t = 2n. This gives

xn+1 = r(27i)sgn(x„)

= + (« - x2)e-4*a}]1/2
32 Introduction

if a # 0, i.e. xn+1 = P(xn) where

P(x) = x[a/{x2 + (a — x2)e-4"a}]1/2. □

Again, difference equations arise directly from modelling natural phe¬

nomena. A famous example is the logistic difference equation, or quadratic
difference equation,

Nn+1 = aN„ - bN2. (8)

This equation is often used with a > 1, b > 0 to model the growth of a
population whose nth generation has N„ individuals. The ideas of the mod¬
elling are similar to those for the logistic differential equation treated in §4.
It follows that if b = 0 then

K = aNn_j = a2N„_2 = ■■• = anN0, (9)

i.e. then the population grows exponentially with n. The nonlinear term,
— bN2, however, represents the moderation of growth due, perhaps, to the
limitation of the food supply or the concomitant increase in the number of
predators. We can simplify the nonlinear equation (8) a little by trans¬
forming the dependent variable to x„ = bNJa, so that

*„+i = bNn+l/a

= b(aNn - bN2)/a

= ax„(l - x„). (10)

The theory of linear difference equations is quite general and systematic.

It is analogous to the theory of linear ordinary differential equations in
many ways. On the other hand the theory of nonlinear difference equations
is largely a collection of special methods. We shall show, however, that
bifurcation theory provides a framework with which some aspects of non¬
linear difference equations may be examined systematically.
The analogue of a steady solution of a differential equation is a fixed
point of a difference equation. We say that X is a fixed point of the
difference equation xn+1 = F(x„) or of the map F: IRm -► IRm if

F(X) = X. (11)

It follows that if x0 = X then

x„ = X forn = 1,2,_ (12)

 1.8 Difference equations 33

By extension, we allow points at infinity to be fixed points; for example,

-oo is a fixed point if m = 1 and F(x) = - x2.

Example 1.4: the fixed points of the logistic map. Consider the logistic
equation (10). It is associated with the logistic map F:RxR-»R such that
F(a,x) = ax(l — x). If x e [0,1] and a e [0,4] then 0 ^ F(a,x) ^ 1, so we
usually shall take 0 ^ x0 ^ 1 and 0 ^ a ^ 4, and confine our attention to
0 ^ x„ ^ 1. Now equation (10) has fixed points X such that

X = aX( 1 - X)

and therefore

X = 0, or (a - l)/a if a ^ 0. □ (13)

Next let us examine, by analogy with the stability of an equilibrium

point of a differential equation, the stability of a fixed point X of a one¬
dimensional difference equation of the form

xn+i = f(*n) for n = 0,1,..., (14)

where F: U -* 1R is continuously twice-differentiable. We have F(X) = X

by definition of the fixed point. Define the perturbation of the fixed point
as

x'„ = xn- X.

Therefore equation (14) becomes

X + x'n+1 = F(X + x')

= F(X) + x'nF'(X) + 0(x;2) as x'„ -> 0.

Therefore the linearized system is

x'n+l=F(X)x'n. (15)

Its solution is

x'n = F'{X)x'n.1 = ... = {F(X)}nx'0.

Therefore, as n -*• oo, we find x'->0 if |F(AT)| <1, x' is bounded if
\F(X)\ = 1, and x' -* oo if |/r'(A')| > 1. Therefore the null solution of the
linearized system is stable if and only if \F'(X)\ ^ 1. We deduce that the
fixed point of the nonlinear system (14) is stable if |F'P0| < 1 and unstable
if |F'(AT)| > 1. The linearized system gives x' = x(, if F'(X) = 1 and x' =
 34 Introduction

(— 1)"xq if F'(X) = — 1 and hence stability if |F'(Ar)| = 1; but in fact the

linearized system is not sufficient to determine the stability of the fixed
point of the nonlinear system in this case, because nonlinear terms may
cause growth or decay of solutions when |.F'(Ar)| = 1 (as they do for solu¬
tions of an ordinary differential equation in §§3-5 when F\X) = 0).

Example 1.5: stability of the fixed points of the logistic map. For the
logistic map F(x) = ax( 1 — x), we find F'{X) = a( 1 — 2X). For the fixed
point X = 0 we find F'(X) = a and deduce that it is stable for — 1 < a < 1;
in fact (see Q1.32) it is also stable for a = 1. For X = (a — 1 )/a, we find
F'(A') = a{l — 2(a — l)/a} = 2 — a and deduce that it is stable for 1 <
a < 3; in fact it is also stable for a = 3 (see §3.4). □

We have seen in §§6, 7 that differential equations may have periodic

solutions as well as equilibrium points. Similarly, difference equations may
have periodic solutions as well as fixed points. To see that, suppose that
there exist a function F:R-*R and distinct points Xu X2 such that X2 =
F(Xi), Xi = F(X2). Then, choosing x0 = X2, we find that xx = F(x0) =
F(X2) = Xux2 = F(Xl) = F(XJ = X2,..., x2r_! =X1,x2r = X2,... for
r = 2, 3,.... We say that the set {X1,X2} of two points is a solution of
period two or a two-cycle of F. There may similarly occur solutions of
periods three, four etc.

Example 1.6: a two-cycle of the logistic difference equation. If there exists a

solution of the logistic equation (10) with period two then

X1 = aX2{ 1 - X2)
and
X2 = aXfl - Xfj

= a{aX2( 1- X2)} [1 - {aX2( 1 - X2)}],

i.e.

aX2(X2 - 1 + 1 /a){a2Xi - a(a + 1)X2 + a + 1} = 0.

This gives the two fixed points (with X2 = 2ft), 0 and (a — 1 )/a, which we
have found in Example 1.4, and also two other real roots,

XUX2 = [a + 1 + {(a + l)(a - 3)}1/2]/2a, (16)

for a < -1 or a > 3. This solution of period two bifurcates from the fixed
point X = § at a = 3; this is an example of what is called a flip bifur-
 1.9 An experiment on statics 35

Fig. 1.18 Some of the bifurcation diagram of the logistic difference equation
x„+i = ax„(l — x„). Stable fixed points are denoted by a continuous curve, un¬
stable fixed points by a dashed curve, stable two-cycles by a dotted curve, and
unstable two-cycles by a dot-dashed curve.

cation. (There is also a flip bifurcation at X = 0, a = — 1.) The periodic

solution is in fact stable for 3 < a ^ 1 + J6 = 3.4495, and bifurcates fur¬
ther when it becomes unstable at a = 3.4495. Therein lies a long story
which we shall take up in Chapter 3. The story so far is summarized in
the bifurcation diagram in the (a,x)-plane shown in Fig. 1.18. Note the
transcritical bifurcation at (1,0) and the flip bifurcation at (3, f). □

9 An experiment on statics
We shall next describe a simple experiment, which you may easily perform
yourself, and interpret it qualitatively in terms of bifurcations in order to
illustrate a general way of thinking about nonlinear systems. The experi¬
ment concerns elasticity, although we shall use only physical intuition for
our interpretation. Consider the experiment which is shown in Fig. 1.19. It
was proposed for the present purpose by Benjamin (cf. Iooss & Joseph
1990, §11.11). As the length / of the arch of wire increases, the strength of
gravity relative to the elastic forces will increase. Therefore, when / is small,
the upright symmetric position 6 = 0 of the arch of wire is stable. However,
as / exceeds some critical value, /c say, the arch will become unstable and
lean on one side or the other. Looking at the experiment, you may see that
the arch falls quite quickly to its new asymmetric position of equilibrium,
and that the new position cannot be made very close to the upright by
making / — lc sufficiently small. If the board and the wire have a con¬
figuration symmetric in ±6 then the side on which the arch leans will be
determined by the initial conditions. As / increases further, the arch will
36 Introduction

(«) V>)
Fig. 1.19 A sketch of the experiment, (a) Side elevation. (i>) Front elevation.

lean more steeply. For these lengths / the only stable positions of equilibri¬
um are the two in which the arch leans on one side or the other, with
equal but opposite angles ± 9. Further experiments will show that there is
hysteresis, in the sense that if l is decreased slowly from a value greater than
/c then the arch will remain in an asymmetric position when / is decreased
below /c. Finally, when / is decreased below another critical value, l0 say,
the arch will spring back to its original upright position 9 = 0. This verbal
description corresponds to the bifurcation diagram of Fig. 1.20(a). We may
plausibly regard it as a manifestation of an equation of motion of the form,

Fig. 1.20 (a) A sketch of the bifurcation diagram in the (/, 0)-plane for the ideal¬
ized symmetric experiment. The arrows denote a hysteresis loop. (b) A sketch of
the bifurcation diagram for a slightly asymmetric experiment.
 Further reading 37

with steady solutions / = L(9) such that F(L{9),9) = 0 for -n <9 <n,
say (although this differential equation is a crude model).
The assumed symmetry of the experiment implies that F(l, -9) = F(l, 9)
for all 9, and therefore that the bifurcation curves l = L(9) in Fig. 1.20(a)
are symmetric about the /-axis. The symmetry breaking which occurs
where / = lc is associated with the subcritical pitchfork bifurcation there.
The two turning points at / = /0 are symmetrically placed. You may ob¬
serve, however, that in your experiment the wire is a little bent so that the
arch always leans on the same side, and so deduce that the configuration of
your apparatus is not exactly symmetric. We say that a system which is
only slightly asymmetric is imperfect, whereas the idealized symmetric
system is perfect. In practice, the experiment will be imperfect however
carefully the apparatus is built. So it is instructive and useful to consider
bifurcation diagrams for imperfect systems. A bifurcation diagram for the
imperfect system is suggested in Fig. 1.20(h); test it with your own experi¬
ment. You could improvise an experiment with a shoe-lace instead of the
wire and your hand instead of the board.

The simple examples of this chapter have been used to introduce several
important concepts and methods of bifurcation theory. In the following
chapters we shall develop the theory more systematically and in greater
detail, and also introduce some advanced concepts and methods. However,
the types of bifurcation we have already met will be seen to occur often in
a wide variety of mathematical problems which may represent a wider
variety of applications. They even occur for nonlinear partial differential
systems, in which the dependent variables belong to a function space of
infinite dimension. We shall see that the essential properties of many bifur¬
cations may be represented in a phase space of dimension as low as one
or two. It is for this reason that the simple bifurcations we have already
exemplified are so common and so useful. There are, however, important
bifurcations which must be represented in phase spaces of dimensions
higher than two, and these bifurcations require deeper understanding and
describe more complicated phenomena.

Further reading
§1.1 This book is an introduction to the theory of nonlinear systems from
the point of view of an applied mathematician. Pippard (1985) introduces
38 Introduction

the subject from the point of view of a physicist, and Devaney (1989) from
the point of view of a pure mathematician. In addition to these clear and
illustrative general texts we shall recommend papers and books as they are
needed, point by point and section by section. Devaney’s (VI989) video is a
vivid introduction to the mathematical theory of chaos.
The ‘butterfly effect’ was described by Lorenz in a lecture at Washing¬
ton DC in 1970 (but note also the remarkable story by Bradbury (1953)).
§1.2 For an extensive account of the figures of equilibrium of a rotating
mass of gravitating fluid and their stability, read Chandrasekhar (1969).
§1.3 Stability is formally defined in §5.1. We shall show more carefully in
Chapter 6 why the linearized system gives the stability of the solution of
the nonlinear second-order differential system except when the linearized
system is at the margin of stability. However, a rigorous proof for ordinary
differential systems of arbitrary order is given by, e.g., Coddington &
Levinson (1955, Chap. 13).
§1.6 We assume that the reader is familiar with the elements of the
theory of differential equations of second order, of the phase plane and of
the types of singularity at a point of equilibrium; these will be discussed
further in Chapter 6.
§1.8 For an introduction to the theory of linear difference equations, the
book by Bender & Orszag (1978) is recommended.

Problems
Q1.1 A transcritical bifurcation. Sketch the bifurcation diagram of the equation,

dx
— = x(a - c - abx).

in the (a,x)-plane for given positive constants b and c, indicating which

steady solutions are stable.
Q1.2 A logistic equation with a cubic term. Show that if
dx
— = ax — bx2 + cx3
df

for x(0) > 0, a, b > 0, b2 > 4ac, then x f X as t —► oo provided that x(0) is
sufficiently small, where X = a/b + 0(a2) as a 10 for fixed b, c.
[Drazin & Reid (1981, p. 458).]
Q1.3 Another transcritical bifurcation. Sketch the bifurcation diagram of the
equation,
dx
— = -x(x2 - 2bx - a),
 Problems 39

in the (a, x)-plane for a given positive constant b, indicating which steady
solutions are stable.
Discuss the states of the system which could be observed in practice
when a oscillates very slowly between -2b2 and 2b2 if solutions with x < 0
are prohibited. [Assume that the system has ‘noise’, i.e. that small perturba¬
tions occur frequently.]
Q1.4 Slow passage through a transcritical bifurcation. Show that if

a = ef, e > 0, and x = x0 at t = t0 then

x(r) = x0expect2) lexp&tl) + *o exp(jes2)ds

*0

Deduce that if x0 > 0 and t0 = T0/Je < 0 then x(f) -♦ 0 as e j 0 for fixed t,
T0 < 0, and x(f) -> a as e [ 0 for fixed a > 0.
Discuss the behaviour of such a solution for which e is small, t0 is large
and negative, and x0 > 0, interpreting it as giving the stable quasi-static
solution x(t) = 0 for t <0 and x(f) = a for t > 0 with a time of order e_1/2 for
transition from the former to the latter.
[Note that when t < 0 the solution x(f) approaches so close to the quasi-
statically stable null solution that when t > 0 it takes a long time to leave
the quasi-statically unstable null solution and approach the quasi-statically
stable solution X = a. This would not occur in practice if there were noise in
the system. Lebovitz & Schaar (1975), Haberman (1979).]
Q1.5 The lemniscate. Sketch the bifurcation diagram of the equation,

dx
= 2(a2 — x2) — (a2 + x2)2,
di

in the (a, x)-plane, indicating which steady solutions are stable, and identi¬
fying two turning points and a transcritical bifurcation.
Q1.6 The folium of Descartes. Sketch the bifurcation diagram of the equation,

dx , ,
— = x + a — 3 ax,
at

in the (a, x)-plane, indicating which steady solutions are stable, and identi¬
fying a turning point and a pitchfork bifurcation.
Q1.7 The Landau equation: a cautionary example. Show that if

dx
— = a (a/66)1/2 sin{(6h/a)1/2x},
dt

a/b > 0 and x(0) > 0 is sufficiently small, then x(f) -> n(a/6b)1/2 as t -» oo.
Sketch the bifurcation diagram in the (a, x)-plane for fixed b > 0.
 40 Introduction

[Note that dx/dt = ax — bx3 + 3b2x5/40a + ... but that x(t)-f* A =

(a/b)112 as t -*■ oo and a J, 0, because then the coefficients of the terms in the
fifth and higher powers of x become large. Drazin & Reid (1981, p. 458).]
Q1.8 A particle on a rotating circular wire. A particle moves around a smooth
circular wire of radius a which is fixed relative to a vertical plane. Gravity
acts on the particle, and the plane rotates with constant angular velocity co
about a vertical diameter of the circle. It is given that

d2e . n o
a—T = — gsmd + aar cos0sin0,
at

where the radius through the particle makes an angle 9 with the downward
vertical.
Find the number of the positions of equilibrium (i.e. steady solutions)
and their stability as the dimensionless parameter aa>2/g varies. Sketch the
bifurcation diagram in the plane of aa>2/g and 9.
[Andronow & Chaikin (1949, p. 75).]
Q1.9 Action of a spring on a particle on a circular wire. A particle of mass m is
constrained to move around a smooth circular wire of radius a fixed in a
vertical plane. The particle is acted on by gravity and is attached to one end
of a spring of natural length l(<2a) and stiffness constant k, the other being
fixed to the highest point of the circle. The equation of motion is given to be

d 29
= k(2acos^0 — /)sin j9 — mg sin 9,

where 9 is the angle between the downward vertical and the radius to the
particle.
(a) Find the number of the equilibrium points and their stability as k
varies. Sketch the bifurcation diagram in the (k, 0)-plane
(b) Re-examine the solution to the problem, regarding l/2a and ak/mg as
independent dimensionless parameters. Sketch the bifurcation dia¬
gram in the (//2a, 0)-plane (for fixed ka/mg > 0). Describe the bifurca¬
tion surface in the three-dimensional space of ak/mg, l/2a and 9.
Q1.10 The rotation of a weighted pulley. A light wheel of radius a has a uniform
semicircular rim of mass M, and may rotate freely in a vertical plane about a
horizontal axis through its centre. A light string passes around the wheel
and suspends a mass m. It is given that this system is governed by the
equation,

d29
(M - m)a2-^ = ag(m - 27r“lMsin0),

where 9 is the angle between the downward vertical and the diameter
through the centre of mass of the heavy rim.
Find the equilibrium points, and conditions for their existence and sta-
 Problems 41

bility. Sketch the bifurcation diagram in the (k, 0)-plane, where k = m/M.
What happens to the wheel when k is large?
Ql.l 1 The complex Landau equation. Show that if

b\z\2z

for complex z, a, b, then

d|z|2 d(phz)
= 2ar|z|2 - 26r|z|4, = a -, ~ ^I
dr dr

where z — |z|e'ph2, a — ar -f ia„ b — bT + ib{ in terms of modulus, phase, and

real and imaginary parts.
Q1.12 A stable limit cycle. Showthatif

dx
= — y + x(l — x2 - y2)/(x2 + y2)1'2,
dr

— = x + y( 1 -x2 - y2)/(x2 + y2)1/2

then x(r) = cos(r + 0O), y(t) = sin(r + 90) represents a stable limit cycle.
[Nemytskii & Stepanov (1960, p. 24).]
Q1.13 A semistable limit cycle. Show that if

^ = -y + x(l - x2 - y2)2, ^ = x + y(l - x2 - y2)2,

then x(r) = cos(r + 0O), y(r) = sin(r + 90) represents a semistable limit cycle,
i.e. it is a periodic solution which is stable on one side and unstable on the
other (and so is unstable).
[Nemytskii & Stepanov (1960, p. 25).]
Q1.14 An infinity of limit cycles. Show that if

~ = -y + */((x2 + y2)1/2), ^ = * + y/((x2 + y2)1/2),

where /is defined by /(r) = sin {l/(r2 - 1)} for r # ± 1 and /(± 1) = 0, then
limit cycles are given by x(r) = r;cos(r + 60), y(r) = r, sin(r + 90) for j = 0, 1,
..., where r0 = 1 and r7 = (1 + \/jn)112 for j = 1,2,.... Which are stable?
[Nemytskii & Stepanov (1960, p. 26).]
Q1.15 A Hopf bifurcation. Given that x — r cos 9, y = r sin 9 for r ^ 0, and

dx , dv
— = -y + x{l — f(x,y)/a], ~ = x + y{ 1 -f(x,y)/a}

for a > 0 and a continuous function /, show that

— = {1 — f(r cos 9, r sin 9)/a}r,

 42 Introduction

Deduce that the null solution is stable if a < /(0,0) and unstable if
0 < /(0,0) < a. Show that if 0 < a < /(x, y) for all x, y then the solution x(r),
y(t) -» 0 as t -* oo for all initial conditions.
Verify that if f(x,y) = g(r) and a = g(r0) for some positive differentiable
function g and some positive constant r0 then there exist solutions of the
form
x(f) = r0 cos(f + 60), y(t) = r0 sin(t + 60).

Show that the orbit of these periodic solutions is stable if g'(r0) = 0 and
unstable if g'(r0) < 0.
Further, given that g(r) = 1 + 1/{1 + (r - l)2}, sketch the bifurcation
diagram in the first quadrant of the (a,r0)-plane, and sketch the phase
portraits in the (x, y)-plane for each of the qualitatively different cases which
arise.
Q1.16 A conservative system with periodic solutions. Show that if

d2x
dF + sgn*-°

then £(dx/df)2 + |x| = E for some constant E ^ 0. Sketch the phase portrait
in the (x, dx/df)-plane. Show that each orbit has period 4(2£)1/2 for E > 0.
[Note that, by definition, sgn x = 1 if x > 0, = 0 if x = 0, and = -1 if
x < 0.]
Q1.17 The oscillation of a nonlinear spring. Show that if

then there are oscillations of amplitude a with period

4K{ea2/(2-€a2)}
(1 - jea2)l/2 ’

where K is the complete elliptic integral of the first kind defined by

K(m) = | 77-1-2 1/2 for m < 1.

Jo (1 - msm2</>)112
Deduce that
T = 2n{l + |ea2 + 0(e2a4)} as€a2^0.

[Cf. Example 6.6.]

Q1.18 The relativistic harmonic oscillator. It is given that the relativistic one¬
dimensional motion of a particle of rest mass m0 and velocity v = dx/df
is governed by the equation
 Problems 43

where c is the speed of light and k a positive constant. Supposing that a is the
amplitude of an oscillation, so that x = + a where v = 0, deduce the mass-
energy integral in the form,

Show further that there is an oscillation of period

where / — 1 + e(a2 — x2) and e = k/2m0c2. Hence show that

T = 2n(m0/k)1/2{l + feu2 + 0(e2a4)} as ea2 -► 0,

i.e. in the Newtonian limit as c -> oo.

Q1.19 Solitary waves. Seek a wave solution of permanent form for the Korteweg-
de Vries equation.

by assuming that u(x,t) — f(x — ct) for some constant c and function /.
Deduce that

if'2=P + W2 + Af+B

for some constants A, B of integration.

Seeking a solitary wave, such that f(X), f'(X), f"(X) -> 0 as X -*• oo,
show that

f(X) = -$csech2{$y/c(X - Aq)} for -oo <X < oo,

where X0 is some real constant and c ^ 0.

[Korteweg & de Vries (1895); cf. Drazin & Johnson (1989, Chap. 2).]
QUO Smooth step-like waves of permanent form. It is given that the substitution of
u(x, t) — f(x — ct) into some nonlinear partial differential equation for u
gives rise, on integration, to an ordinary differential equation of the form
if'2 = F(/); where c is taken to be a constant, F" is continous, F has double
zeros at ft, f2, F(/) > 0 for fy < f < f2, and /x and f2 are some constants.
Show that there exist two monotonic solutions / such that f(X) -» ft as
X -* +oo and f(X) -+ f2 as X -* ±oo respectively.
[Such solutions are called topological solitons.]
Q1.21 The elastica. (a) It is given that the angle \p of deflection of a static uniform
elastic beam at a distance s from one end is governed by the equation,
 44 Introduction

where ansa positive parameter. Show that

jip'2 — to2 cos 1p = constant

for each solution, where ip' = dip/ds. Sketch the phase portrait of the solu¬
tions in the (ip, ip' )-plane.
(b) Suppose, moreover, that the beam has length /, is clamped at one end,
so that ip = 0 at s = 0, and free at the other end, so that ip' = 0 at s = /.
Solving the linearized equation for small tp together with the boundary
conditions, show that to = (n + f)n/l is an eigenvalue for n = 0,1,_
Show that if ip increases monotonically from 0 to a (where a is a given
positive angle, not necessarily small) as s increases from 0 to /, then

,_ r d4
" Jo (1 — /c2sin2^)1/2’-

where k = sin \a. [Hint: substitute sin jip = k sin ^.] Deduce that

col = jn{l + £sin2 ja + 0(a4)} asa^O.

Considering next solutions which are not monotonic, show that

'*/2
tol = (2 n + 1) for n— 1,2,...
(1 — k2 sin2 <p)1/2

for orbits which go around the origin of the phase plane and for which
ip = a at s = l. Sketch the bifurcation diagram in the (to2, a)-plane, where
ip = a at s = / but ip need not be a monotonic function of s.
Q1.22 An equation transformable to a conservative system. Show that if

d2x
+ k(x) = /(*)
dt2

then

1 /dx\2

2m(x) ( dt ) + = £,

where V(x) = ~^m(u)f(u)du,m(x) = exp {2 J5k(u)du}. Hence show that

d2y
d? "
where y = ft (m(u))1/2 du, g(y) = (m(x))1/2/(x).
Q1.23 An elementary proof of a fixed-point theorem. Prove that if F: IR -»• IR is
continuous and there exists a closed interval I such that I <= F(I) then there
exists X e I such that F(X) = X.
[Hint: either use the intermediate-value theorem or sketch the graphs of
y — F(x) a°d y = x in the (x,y)-plane for x e I.]
 Problems 45

Q1.24 Iterated approximation to a square-root. Using the Newton-Raphson meth¬

od to approximate a zero of f(a, x) = x2 — a for a > 0, show that successive
approximations to Ja may be given by

*-+i = 2(xn + a/xn) for n = 0,1,....

Hence evaluate the square-root of 55 to three significant figures.

Q1.25 Convergence of the Newton—Raphson method. Given that f is a twice contin¬
uously differentiable function on [a, b] such that f(a)f(b) < 0 and /'(x) # 0
for all x e (a, b), show that / has a unique zero X in (a, b).
Show that if xn+1 = F(x„) for n = 0, 1,..., where F(x) = x - f(x)/f'(x),
and xn, xn+1 e(a,b), then

M
\xn+x-X\^—\xn-X\2,

where m = mina<JC<l, |/'(x)| and M = maxa<x<fc |/"(x)|.

Q1.26 Fibonacci numbers, the golden mean and a continued fraction. Defining the
Fibonacci numbers x„ by

*n+2 = *„+!+*„ for n = 0,1,...,

x0 = 0 and xx = 1, and defining yn = xjxn+l, deduce that

yn+1 = V(i + yn)-

Defining the golden mean as a = 5(^5 — 1), = 0.618, prove that

y„~* a as n -> 00.

Deduce that the continued fraction

= a.
1 +
1 +
1 +

Q1.27 Another continued fraction. Given a > 0, seek to compute X = yja as fol¬
lows. Estimate r > 0 as a first approximation to yja and define 5 = a — r2 so
that X = r + 5/(r + X). Then iterate xn+1 = r 4- d/(r + x„) with x0 = r.
Find the fixed points of the difference equation and whether they are
stable.
Hence express y/a as a continued fraction.
Use the difference equation also to evaluate the square-root of 101 to six
significant figures.
Q1.28 Enhancement of stability of a fixed point. Define G: IR x R -♦ IR by

G(a,x) = aF(x) + (1 — a)x,

 46 Introduction

given the map F: IR -♦ IR. Show that, for a ^ 0, X is a fixed point of F if and
only if it is a fixed point of G. Find those values of a for which AT is a stable
fixed point of G when F is continuously differentiable.
In particular, take F(x) = 50 — cosh x. Show graphically that there is
a unique positive fixed point X of F, and numerically that 4 < X < 5. De¬
duce that X is an unstable fixed point of F, but a stable fixed point of G if
a = 0.02. [You are given that cosh 4 = 27.31, sinh4 = 27.29, cosh 5 = 74.21
and sinh 5 = 74.20.]
Define H over an interval where F' ^ 1 by

H(x) = x - {x - F(x)}/{1 - F'(x)}.

Deduce that, in this interval, A" is a fixed point of H if and only if it is a

fixed point of F. Show that AT is a stable fixed point of H for all twice
differentiable F.
Q1.29 Stability of the null solution of a finite-difference approximation to a differ¬
ential equation. Suppose that

x„+i = xn + hf(x„) for n = 0,1,...,

where /: IR -> IR is continuously differentiable, /(0) = 0 and h > 0. Show

that the fixed point AT = 0 is stable if /'(0) < 0 and unstable if /'(0) > 0
when h is sufficiently small.
[Note that this agrees with the condition for stability of the null solution
of the differential equation dx/dt = /(x).]
Q1.30 The logistic map. Define F:U x U->Uby F(a,x) = ax{l - x). Show that if
0 < a < 4 then F maps the interval [0,1] into itself and if a = 4 then F maps
the interval onto itself. Show that if a > 1 and either x0 < 0 or x0 > 1 then
x„ -*• -oo as n —► oo, where xn+1 - F(a, xj.
Q1.31 Elementary transformation of the logistic equation. Show that if yB+1 =
1 - by2„ and x„ = &a - \)yn + \ for n = 0, 1,..., where b = ±a2 ~ \a, then
xn+1 = ax„(l - xB). Show that if zn = by„ then zn+1 = b - zjj.
Q1.32 The asymptotic solution of the logistic equation in a special case. Show that if

*»+i = *n(l - Xj for n = 0,1,...,

and 0 < x0 < 1 then x„ ~ l/n as n -> oo.
[Putnam (1967).]
Q1.33 The explicit solution of the logistic equation in a special case. Show that if

*B+i = 2x„(l - x„) forn = 0,1,...,

and y„ = 1 — 2x„, then yB+1 = yHence or otherwise show that

xB = i{l — (1 — 2x0)2"}.
Deduce that limB^ xB = \ if and only if 0 < x0 < 1.
[Bender & Orszag (1978), p. 53).]
 Problems 47

Q1.34 The limit of the solution of a difference equation. Show that if xn # 2 and

1
xn + l for n = 0,1,...,
2-x„
then

1)jc0
n + l — nx0
and therefore x„ = 1.
[Putnam (1947).]
Q1.35 A linear difference equation. Suppose that

xn+2 + 2bxn+1 + cx„ = 0 for n = 0,1,...,

where real b, c ^ b2, x0 and x1 are given. To solve this problem, show that

*n+l ^Xn,

where the column vector x„ = X"

L*n + 1
and the 2 x 2 matrix A = r ‘i.
L-c 2b j
Deduce that x„ = Anx0.
Show that there exist £t, £2 such that x0 = + £2u2 for all x0, where
Sj is the eigenvalue of A belonging to the eigenvector for j = 1, 2, and
thence that x„ = fj j"u, + £2S2U2- Hence or otherwise prove that

xn = ^{-b + (b2- c)112}" + Z2{-b-(b2- c)1/2}",

where ^,^2 = 2{x0 ± (bx0 + xx)/(b2 — c)1/2} respectively.

Deduce that if xn+2 = x„ + xn+x, x0 = 0 and xx = 1 then the Fibonacci
numbers x„ = [{^5 + 1)}” - {—i(V5 - l)}"]/>/5 ~ {i(V5 + 1)}"/V5 as
n —► co.
Q1.36 Equality of slopes of a second-generation map at its fixed points. Define the
second-generation map G of F by G(x) = F(F(x)), where F is a continuously
differentiable function, i.e. F e Cl(—oo, oo), and suppose that F has a two-
cycle {A, T} such that X = F(Y) and Y = F(X). Deduce that X and Y are
fixed points of G, and that G'(X) = F'(X)F'(Y) = G'(7).
Q1.37 Some cycles of a map. Defining F by F(x) = (1 — x)/(l + ax) for a > 0, x ^
— 1/a, show that F maps the interval [0,1] onto itself. Find all the fixed
points of F and ascertain whether each is stable, according to the values of a.
Find all the two-cycles of F. What other cycles of F are there?
Q1.38 Brouwer’s fixed-point theorem in one dimension. Prove that if F: IR —> IR is
continuous, I is a closed interval, and F(I) C I, then F has at least one fixed
point in I.
[Hint: use the intermediate-value theorem.]
 2
Classification of bifurcations of equilibrium points

Not chaos-like together crush’d and bruis’d.

But, as the world, harmoniously confused:
Where order in variety we see,
And where, tho’ all things differ, all agree.
Alexander Pope (Windsor Forest)

1 Introduction
We have met autonomous differential equations, i.e. equations which do
not involve the independent variable explicitly, of the form

% -FW- <»

where F: [Rm -> [Rm. Then the equilibrium points, sometimes called the criti¬
cal points, are the zeros X of F, i.e. the points X such that F(X) = 0. Like¬
wise we have met difference equations of the form

x«+i = G(x„) forn = 0,1,..., (2)

where G: !Rm -> !Rm, and sought fixed points X such that X = G(X). There¬
fore finding the fixed points of a difference equation is equivalent to finding
the equilibrium points of an ordinary differential system, on identifying
x ~ G(x) = F(x) for all x e lRm. Further, the bifurcations of the fixed points
of a difference equation, as a parameter changes, are equivalent to the
bifurcations of an ordinary differential system. Indeed, we have met a
transcritical bifurcation for a difference equation, and will soon meet turn¬
ing points and pitchfork bifurcations in examples of difference equations
with a parameter.
So we shall in this chapter seek the solutions x of nonlinear functional
equations of the form

F(a,x) = 0 (3)

as / parameters a vary, where F: R' x (Rm -* (Rm. We define a bifurcation

point, or a branch point, as a solution (a0, X0) of equation (3) such that the

48
 2.2 Bifurcations in one dimension 49

number of solutions x of (3) in a small neighbourhood of X0 changes when

a varies within a small neighbourhood of a0. You may verify that this
definition gives turning points, transcritical points and pitchforks as bifur¬
cation points. (The definition needs extension for a nonlinear system with
solutions which evolve with time, to allow for the change of the qualitative
character of the unsteady solutions as a parameter changes.)
The zeros of F cannot, of course, in general be given explicitly as ele¬
mentary functions of a by solving equation (3). We may be able to find the
zeros analytically in a few special cases, but usually must resort to numeri¬
cal methods of solution. However, we can examine the local properties of
the solutions systematically and thereby classify the types of bifurcation.
The classification becomes complicated as the number of types grows with
/ and m, but nonetheless there are important particular and general results
known.

2 Classification of bifurcations in one dimension

First treat the case m = 1, beginning with the subcase 1 = 1. Thus we seek
solutions x = X(a) of the equation

F(a, x) = 0,

for the map F: U x IR -► IR. These solutions may be represented by curves

in the bifurcation diagram of the (a, x)-plane.
To apply local analysis, we suppose that one point (a0, Y0) of the solu¬
tion is known somehow, i.e. that

F(cio’X0) = 0. (1)
Then we may expand F as a Taylor series about this point in order to find
neighbouring solutions, provided that F is infinitely differentiable with
respect to a and x at (a0, Y0):

0 = F(a,Y) (2)

= F(a0,X0) + (X- X0)Fx(a0,X0) + (a - a0)Fa(a0,X0)

+ \(X-X0)2Fxx(a0,X0) + --

= (X - X0)Fx0 + (a - a0)Fa0 + {{X - X0)2Fxx0 + • • •, (3)

when we use equation (1) and denote evaluation of the partial derivatives
of F at (a0, X0) by the subscript zero. Therefore

X(a) = X0 - (a - a0)Fa0/Fx0 + o(a - a0) as a -*• a0, (4)

50 Classification of bifurcations

provided that Fx0 / 0. This heuristic argument gives the equilibrium solu¬
tions near (a0,X0) in the (a, x)-plane, in fact the tangent to the curve at
{a0,X0) in the bifurcation diagram. It can be proved, by use of the implicit
function theorem, that the curve x = X(a) exists in a neighbourhood of
(a0, X0) if both F is continuously differentiable with respect to a and x in a
neighbourhood of (a0, AT0) and Fx0 * 0.
If Fx0 = 0 then equation (3) gives

X(a) = X0± {—2(a - a0)Fa0/Fxx0y>2 + o((a - a0)1'2) (5)

as (a — a0)sgn(Fa0/Fxx0) f 0, provided that Fxx0 ^ 0. Note that we require

a — a0 to have the appropriate sign to ensure that the two solutions (5)
(representing a turning point at (a0, X0)) are real.
An infinity of further special cases can be seen to arise according to
which of the leading partial derivatives of F vanish at (a0, X0).

Example 2.1: an exponential map. Consider the difference equation xn+1 =

G(a,x„), where G(a,x) = aex; then the fixed points X of G are the zeros of
F(a, x) = aex — x. We see, by inspection, that one zero is given by a = 0,
X = 0. So, expanding about this point, we find

0 = F(a,X)

= a( 1 + X + \X2 + ±X3 + •••)- X.

Therefore
X = a -f- aX + jaX * + • • •

= a + a2 + 0(a3) asa-+0.

Note also that F(a, x) = 0 is equivalent to x = In x — lna, and that

x/ln x -> oo as x -> oo. So there is another fixed point for which

Ar(a)~-lna asajO.

We also see by inspection that a zero of F is given by X = 1 when

a = 1/e. To find other solutions nearby, we expand

0 = F(a,*)

= *We. !) + (*- 1)^ 1/e, 1) + (o - l/e)Fa(l/e, 1)

+&X - l)2^x(l/e, 1) + {X - 1 )(a - l/e)Fax(l/e, 1) + --

= 0 + 0 + (a - l/e)e + {(X - l)2 + e(X - l)(n - 1/e)

+ i(2f-l)3 + ---.
 2.2 Bifurcations in one dimension 51

Fig. 2.1 Sketch of the bifurcation diagram for F(a, x) = aex - x, i.e. of the curve
with equation aex = x in the (a, x)-plane.

Therefore

(X - l)2 = 2(1 - ae) + 2(X - 1)(1 - ae) - ±(X - l)3 + • • •.

Therefore

X = 1 ± {2(1 — ae)}1/2[l ± {|(1 — ae)}1/2 + 0(1 — ae)] as a f 1/e.

Observe that Fx( 1/e, 1) = 0 and there is a turning point at A' — 1, a = 1/e in
the bifurcation diagram in the (a,x)-plane, as sketched in Fig. 2.1.
*Note that an expansion of the form X{a) = 1 + (a — l/e)X1 +
(a — l/e)2A"2 + • • • is invalid as a-* 1/e. However, we may expand

a = 1/e + eax + e2a2 + ••• and X{a) = 1 + eXx + e2X2 + •••

as e -*• 0, (6)

where the new small parameter e is introduced instead of a — 1/e. This

gives

0 = aex - X

= (1/e + eax + •••)exp(l + eXx +•••)- (1 + eXx + •••)

= (1 + eeax + •••){ 1 (eXx + • • •) 4- 2(eAi + ■ *■ )2 + '■ ■}

-\-eXx- •••

= eeaj + e2(ea2 + eax Xx + jXf) + 0(e2) as e -► 0.

c
52 Classification of bifurcations

Equating coefficients of e and e2, we deduce that

flj =0, Xy — —2 ea2

respectively. There is some freedom of choice of the coefficients a2,a3,

because the same solution X may be expressed differently by the pair (6) of
expansions. Requiring X to be real, we may take a2 as any negative num¬
ber. It is convenient to choose a2 = — 1 and a3 = a4 = ... = 0 as a means
of defining e in terms of a - 1/e. Then Xx = ±(2e)1/2, and

X = 1 ± (2e)1/2e -I- 0(e2) as e -* 0,

where a = 1/e — e2, in agreement with the solution found otherwise above.
□

When the first several terms of the Taylor series (3) vanish, the local
behaviour of the equilibrium curve near (a0,X0) is determined by the first
terms which do not vanish. The main special cases may be classified as
follows.

(i) A regular point (a0, X0) of the equilibrium curve is one for which Fx0 ^
0 or Fa0 / 0. Then the implicit function theorem shows that X(a) or a(A')
respectively exists and is continuously differentiable in a neighbourhood of
(a0, X0). For an example (see Fig. 2.2(a)), take

Fig. 2.2 Sketches of examples of typical kinds of bifurcation points, (a) A re¬
gular point: * - X0 + (a - a0) - (a - a0)3 = 0. (b) A regular turning point:
(x - X0) -(a- a0) = 0. (c) A transcritical bifurcation: (x - X0)2 - (a - a0)2
= 0. (d) A singular turning point: {(x - X0)2 - (a - a0)}(x - *„) = 0. (e) A
cusp: (x - X0)2 - (a - n0)3 = 0.
 2.2 Bifurcations in one dimension 53

F(a,x) = x - X0 + a - a0 - (a - a0)3.

(ii) A regular turning point or simple turning point (a0, X0) of an equilibrium
curve is a regular point at which da/dX changes sign and Fa0 # 0. For
example (see Fig. 2.2(h)), take

F{a, x) = (x - X0)2 - (a - a0).

Note that this is a special case of a regular point, that X(a) is a 2-1 function
for a > a0 but does not exist for a < a0, and that Fx0 = 0.

(iii) A singular point is a point on an equilibrium curve which is not a

regular point, i.e. it is a point (a0, X0) where Fx = Fa = 0.

(iv) A double point is a singular point (a0,X0) through which pass two and
only two branches of the equilibrium curve with distinct tangents. This
requires F2x0 > Faa0Fxx0. Typically there is a transcritical bifurcation at a
double point. For example (see Fig. 2.2(c)), take

F(a,x) = (x - X0)2 - (a- a0)2.

(v) A singular turning point is a double point at which da/dX changes sign
on one branch of the equilibrium curve with distinct tangents. We need
cubic terms in the Taylor expansion of F to distinguish this special case of a
double point. This gives a pitchfork bifurcation at (a0,X0). For example
(see Fig. 2.2(d)), take

F{a,x) = {(x - X0)2 - (a- a0)}(x - X0).

(vi) A cusp is a point of second-order contact between the two branches of

the curve. For example (see Fig. 2.2(e)), take

F{a,x) = (x- X0)2 - (a- a0)3.

(vii) A conjugate point is an isolated singular point. For example, take

F(a,x) = (x - X0)2 + {a- a0)2.

This exhausts the essentially distinct cases in which at least one of the
second derivatives of F is non-zero at the equilibrium point (a0, A^). There
are, however, other cases of higher-order singularity.

Next consider cases / > 1 and m = 1. It can be seen, as for the case / = 1,
that, for well-behaved functions F:U‘x IR —»• IR, a bifurcation may arise
only at those values of a for which F has a multiple zero x, i.e. where
54 Classification of bifurcations

F( a, x) = 0 and Fx( a, x) = 0.

We can in principle eliminate x from these two equations to get a set, called
the bifurcation set, of values of a. So the bifurcation set is a subset of IR'. (It
is possible, with this informal definition, that there is a point of the bifurca¬
tion set at which there is no bifurcation; for example, if F(a, x) = (x — a)2
then a = 0, x = 0 is not a bifurcation point.)

Example 2.2: the cusp bifurcation set. Suppose that

F(a, b, x) = 4x3 — lax + b.

Then we may picture F — 0 as a surface of height x above the (a, b)-plane,

shown in Fig. 2.3. There are one, three, or, exceptionally, two, real zeros x
according to the values of a and b. The number changes on the bifurcation

Fig. 2.3 The cusp catastrophe, (a) The folded surface F{a,b,x) = 0 in the
(a, h,.v)-space and (b) the projection in the (a, h)-plane of its points with ‘vertical’
tangents, and with a cusp at (0,0).
 2.3 Imperfections 55

set, where there are folds, and the tangent plane to the surface is vertical, i.e.
perpendicular to the (a, 6)-plane. Thus to find the bifurcation set, we solve
F = F' = 0, i.e.

4x3 - lax + b = 0 and 12x2 - 2a = 0

for a, b. Therefore

x = ±(£a)1/2 for a ^ 0, and ±4(£a)3/2 + 2a(£a)1/2 + 6 = 0,

i.e.

27 62 = 8a3.

This gives a cusp at the origin in the (a, 6)-plane. □

3 Imperfections
In §1.9 we introduced an example of an imperfection. In general, we use
mathematical models with various idealizations, such as symmetry, steadi¬
ness, plane boundaries and unbounded domains, which are not found pre¬
cisely in the ‘real’ world they are designed to represent. So it is important
to consider the effect of small ‘irregularities’ on idealized models. An irreg¬
ularity may be important if it changes not only the quantitative character
of the solutions a little but also the qualitative character of the solutions.
If the qualitative character of the set of all the solutions of a system is
changed by an infinitesimal perturbation of the system (or, rather, at least
one infinitesimal perturbation of a specified class) then the system is said to
be structurally unstable (cf. Andronow & Pontryagin 1937, Andronow &
Chaikin 1949, p. 337). (This is in contrast to the usual concept of instability
which concerns the evolution in time of solutions of the same system but
infinitesimally different initial values.) The theory of imperfections treats
the change in the character of the set of solutions when an extra small
parameter is introduced into the system, and thereby assesses the limita¬
tions of the simplifications in a model of a natural phenomenon, as we did
in §1.9.
To substantiate these ideas, we may consider the one-dimensional evo¬
lution equation of the form

= F(a,x,S), (1)

where <5 is the extra small parameter in addition to the usual parameter a.
We call the system with <5 = 0 perfect and with small S # 0 imperfect. Then
 56 Classification of bifurcations

the equilibrium points and their bifurcations for S = 0 may be perturbed

by putting

F(a,x,S) = 0 (2)

and expanding F as a Taylor series in S as well as a — a0 and x — X0.

Example 2.3: the structural stability of a turning point. Consider the perfect
system with

F(a, x, 0) = x2 — a.

Then the imperfect system may be expanded about the turning point (0,0)
of the perfect system by use of a Taylor series:

0 = F(a,x,S)

= x2 - a + Fd0S + Fxdox3 + Fad0aS + jFxxd0x23 + • ■ •

= (1 + 2^FxxS0)(x + \3Fxd0)2 - (1 - 3FaS0)(a - SFd0)

+ 0(S2,Sa2,Sx3) as a, x, <5-*0,

on collecting the terms in a way to show that the nose of the parabola in
the (a,x)-plane is preserved for small 3 and neglecting very small terms.
Therefore the bifurcation curve F(a,x,3) = 0 has the same qualitative be¬
haviour near (0,0) in the (a, x)-plane for small <5, although the turning point
is translated, rotated and magnified a little. Thus a turning point is struc¬
turally stable. The imperfection changes the position of almost all points of
the curve a little but does not change the topological character of the curve
locally. The perfect and imperfect systems both have a nose-shaped bifur¬
cation curve near (0,0) in the (a, x)-plane: ‘Nose is a nose is a nose is a nose’
as Gentnude Stein (1922) put it. □

Example 2.4: the structural instability of a transcritical bifurcation. Similarly

take

F(a,x,S) = x2 - a2 + S.

Then the perfect system, i.e. F(a, x, 0) = 0, has a transcritical bifurcation in

the (a, x)-plane at (0,0). However, if 3 ^ 0 there are two separate solutions;
both have a turning point if 3 > 0. We can see the bifurcation diagrams in
Fig. 2.4 in the (a, x)-plane for different values of 3 as sections of the bifurca¬
tion surface F(a, x, <5) = 0 in (a, x, 3)-space. □
 2.3 Imperfections 57

(a) (b)

Fig. 2.4 An imperfect transcritical bifurcation: sketches of the equilibrium

curves x = ± (a2 - S)i/2 for (a) d>0,(b)d = 0, and (c) S < 0.

Example 2.5: the structural instability of a pitchfork bifurcation. Next take

F{a,x,S) = x(x2 — a) + 3,

and consider solutions F(a,x,S) = 0. (Note that this is essentially the same
function F as in Example 2.2, but now we look at it from a different point of
view.) If 3 = 0 then there is a supercritical pitchfork bifurcation at (0,0) in
the (a, x)-plane. If, however, 3^0 then there are two separate branches of
the equilibrium curve in the (a, x)-plane, one with a turning point. Rather
than solve the cubic for x, it is easier to calculate a = x2 + 3/x as x varies
for fixed 3. See the curves in Fig. 2.5, noting that the symmetry in ±x for
3 = 0 is broken for 3 / 0. □
58 Classification of bifurcations

Example 2.6: an isola bifurcation. The topological nature of structural sta¬

bility implies at once that a conjugate point is structurally unstable, be¬
cause an isolated point can be smoothly changed to a small closed curve
or to nothing. The canonical example of this is given by F(a, x, <5) =
a2 + x2 — S. Then for S > 0 the bifurcation curves are circles with centre
(0,0) and radii y/3 in the (a,x)-plane. For <5 = 0 the curve is the conjugate
point (0,0), and for 3 < 0 there is no solution of F(a,x,S) — 0. An isolated
closed curve, such as one of the circles, in a bifurcation plane is called an
isola.
Another example of an isola bifurcation is given by F(a,x,S) — \x2 +
jo3 — a — f + S, shown in Fig. 2.6. (Note the geometrical analogy with
Example 1.1, but different interpretation of the variables.) If S — 0 then
there is a transcritical bifurcation at (-1,0) and a turning point at (2,0)
because F(a,x, 0) = \x2 — j(a + 1)2(2 — a). If 5 = § then there is an iso¬
lated singular point at (1,0) and an unbounded branch of the bifurcation
curve with a turning point at (-2,0), because F{a, x, §) = \x2 +
— 1 )2(a + 2). If 0 < <5 < § then there are two separate branches of the
bifurcation curve, one an isola and the other unbounded; if <5 < 0 or S > §

Fig. 2.6 An isola: sketches of the bifurcation diagrams for F(a,x,6) = 0 in the
(a,x)-plane, where F(a,x,6) = jx2 + ±a3 - a - § + 6. (a) 6 < 0. (b) 6 = 0
(c)0<6<Ud)6 = Ue)6>i
 2.4 Bifurcations in higher dimensions 59

then there is a single unbounded bifurcation curve with a turning point on

the a-axis. o

4 Classification of bifurcations in higher dimensions

For the general case of l control variables, or parameters, a and m state
variables, or behaviour variables, x we have

F(a, x) = 0, (1)

for F: R' x lRm -+ |Rm. If we know one solution (a0, X0) such that

F(a0, X0) = 0 (2)

and F is infinitely differentiable at (a0,X0) then we may expand equation

(1) as the Taylor series

0 = Fi0 + X (“j ~ a}0) dJi

j=i
F X C^k — -^fco) dA +
8aj. k=l l8xu

for i — 1,2 where the suffix zero denotes evaluation at (a0, X0), i.e.

o = 7=1
X Maj ~ «/o) + k=l
X uxk - xk0) +
— A(a — a0) + J(X — X0) + • • •, (3)

say, where AtJ = [dFJdaf]0 is the element of an m x / matrix A in its ith

row andyth column, and the m x m Jacobian matrix J has elements Jik =
[SFJdx^Q. It follows heuristically that

X(a) = X0 — J_1A(a - a0) + o(|a - a0|) asa^a0 (4)

if J is non-singular. Again, the implicit function theorem may be used to

prove that if (a) equations (1), (2) hold and F is continuously differentiable
at (a0, X0) and (b) J is non-singular, then there exists a unique solution X(a)
which is continuously differentiable near (a0, X0).

* Example 2.7: a pitchfork bifurcation. Suppose that F(a, x) = [siny — tanx,

ax - y - x2y]T for all a e IR and x = [x,y]T e IR2, and seek X(a) such that
F(a, X(a)) = 0. By inspection, we see that X = 0 is a zero of F for all a. Also
note the symmetry of the system whereby it is invariant under reflection in
the origin in the (x,y)-plane, i.e. under the transformation (x,y)->( — x, — y).
To seek solutions near a given solution a = a0, X0(a0) = 0 such that
F(a0, X0(u0)) = 0, take the Taylor series
60 Classification of bifurcations

0 = F(a, X(n)) (5)

~dF~ ~BF~
= F0 + (a - a0) + X + •••
_dci_ 0 _d\_

= JX + nonlinear terms in (a — a0) and X,

because F0, [<5F/dd]0 = 0, where the Jacobian matrix is

and we use the subscript zero to denote evaluation at a = a0, x = X(a0),

= 0. Therefore det J = 1 — a0; so J is invertible if and only if a0 / 1.
Therefore the implicit function theorem gives a unique solution (in fact
X(u) = 0) in the neighbourhood of (a0,0) except when a0 = 1. So we antici¬
pate a bifurcation where a = 1.
To find all the neighbouring solutions X(a) when a — 1 is small, we first
note that

1
J =
-1

when a0 = 1, and that then J has eigenvalues 0,-2 belonging to the re¬
spective eigenvectors [1,1]T, [ — 1,1]T.
Next rewrite equation (5), without approximation, in the form

(T-sin Y) — (X — tanX)
JX (6)
-(a - 1)X + X2Y

This form is convenient because the small terms on the left-hand side are
linear in a — l, X, Y, but when a Taylor expansion is taken the very small
terms of quadratic or higher order in a - 1, X, Y are all on the right-hand
side.
We see that, to linear order, X may be any member of the null space of J,
i.e. X is an arbitrary multiple of the eigenvector [1,1]T with zero eigen¬
value. We anticipate that the nonlinear terms will determine what the small
multiple of the eigenvector X is as a -»• 1. Accordingly define

e = [l,l]X = X+ Y, (7)
and assume that we may expand

X(a) - eX1 + €2X2 + •••, a = 1 + eal + e2a2 -|- ase-^0. (8)

 2.4 Bifurcations in higher dimensions 61

(This is a generalization of the second method used in Example 2.1.) Now

substitute the expansions (8) into equations (6) and (7), and equate co¬
efficients of successive powers of e.
The coefficients of e give

*!-*! = 0, x1 - y, = 0,
and
Xi + Yt = 1.

These equations have the unique solution Xx = j[l, 1]T, thereby confirm¬
ing the form of solution which we anticipated.
The coefficients of e2 give

^2 X2 = 0, X2 — Y2 = — UjXj = — 2ol,

and

x2 + y2 = 0.
Therefore a2 = 0 for the compatibility of the first two of these equations,
and we deduce that X2 = [0,0]T.
The coefficients of e3 give

— Y3 = — a2X2 — a2X2 + Xf Y2 = —\a2 + |

and

X3 + Y3 = 0.

The solvability condition, i.e. compatibility condition, of the first two of

these equations gives a2 = 5, and then we deduce that X3 = ^[— 1,1]T.
We may proceed in this way to find as many terms in the expansions of a
and X as we wish. So far we have found that

X(«) = *e[l, 1]T + 1,1]T + 0(€4), a = 1+W + 0(e3)

as e -»■ 0. This is enough to give the most important result, namely that in
addition to the null solution there are two solutions such that

X(a)~ ±|{3(u- 1)}1/2[1,1]T as a 11.

This describes a supercritical pitchfork bifurcation at a = 1, X = 0. Note

that the definition (7) of e normalizes the solution and renders its expres¬
sion unique. □
62 Classification of bifurcations

When the Jacobian matrix J is singular there are various forms of bifur¬
cation according to its co-rank as well as the higher terms in the Taylor
series. This is the subject of singularity theory, at one time called catastro¬
phe theory. We shall merely summarize it briefly here. The co-rank of J,
rather than its rank, describes the topological behaviour of the solutions as
a varies near a0, because many components of x may be ‘passive’, i.e. the
bifurcations may occur in a subspace of [Rm with dimension much less than
m. (Also the nullity of J is more easily generalizable in the case m = oo
when [Rm is replaced by an infinite-dimensional function space.) We seek to
retain only the ‘active’ components of x in identifying the essential topolog¬
ical character of the bifurcations, and these lie in the null space of J. Simi¬
larly we define the co-dimension to be the dimension of the subspace of [R( in
which a varies to give the essential character of a bifurcation. For example,
the turning point or fold of §1.3 has co-dimension one even though it may
be embedded in a space with more than one parameter (cf. Example 2.2).
We also seek to make smooth invertible transformations of the coordi¬
nates to reduce the bifurcations to simple canonical forms. Thus we are
interested in equivalence classes of bifurcations with the same topological
character.
Thom (1975) considered F = —grad V, i.e. F = — VV, for a potential
function F(a, x) and showed that a structurally stable bifurcation is topo¬
logically equivalent to one of a few canonical forms. On taking = 0 and
X0 = 0 without loss of generality, the forms are as listed in Table 2.1, it
being more compact to give the potential than all the components of F.
Some familiar bifurcations may be seen in this table. Transcritical and
pitchfork bifurcations do not appear in their own right because they are
not structurally stable. However, they do appear. A pitchfork bifurcation,

Table 2.1. Canonical forms of structurally stable potential singularities of

co-dimension ^ 4: the seven elementary catastrophes.

co-rank co-dimension V name

1 1 3X3 +ax fold or turning point
1 2 |x4 +\ax2 + bx cusp
1 3 5X5 + 3 ax3 + \bx2 + cx swallow’s tail
2 3 x3 + y3 + cxy — ax — by hyperbolic umbilic
2 3 x3 — 3xy2 + c(x2 + y2) — ax — by elliptic umbilic
1 4 gx6 + 5 ax4 + 3 bx3 + \cx2 + dx butterfly
2 4 x2y + £y4 + cx2 + dy2 — ax — by parabolic umbilic
 Problems 63

for example, arises as a special case (b = 0) of the cusp; as we saw in detail

in Examples 2.2 and 2.5, a small change in b destroys the pitchfork; but
each cusp catastrophe has one special case which gives a pitchfork.

In this chapter we have examined the behaviour of fixed points, i.e. steady
solutions, as parameters vary. In the remainder of the book we shall exam¬
ine the behaviour of unsteady solutions. So, having studied statics, we shall
now go on to study dynamics.

Further reading
§2.2 Courant (1936, pp. 127-9) gives the elementary theory of singular
points on curves.
§2.3 Iooss & Joseph (1990) describe systematically the analytic theory,
including the theory of imperfections, with many examples and exercises.
§2.4 Thom (1975) describes the elementary bifurcations, emphasizing
geometrical aspects of the theory and their applications. Golubitsky &
Schaeffer (1985) is a valuable treatise on bifurcations.

Problems
Q2.1 A turning point. Consider the difference equation

x„+i =G(a,x„) for n = 0,1,...,

where G: IR x [R -> K is a well-behaved function of both its arguments. Us¬

ing subscripts x and a to denote partial derivatives, show that if G(0, A^) =
*o, G*(0, A'q) = 1, Ga(0, Xq) ^ 0, G„(0, Xq) ^ 0 then there is a turning point
at a = 0, x = Jfo, and that if, moreover, Ga(0, A'q) < 0, Gxx(0,X0) > 0 then
there are two fixed points for 0 < a « 1 but none for 0 < —a « 1.
Q2.2 A flip bifurcation. Consider the difference equation

x„+i = G(a,x„) forn = 0, 1,...,

where G:IRx[R-»IRisa well-behaved function of both its arguments.

Show that if G(0,0) = 0, Gx(0,0) ^ 1, Ga(0,0) # 0 then there is a unique fixed
point near to x = 0 when a is near to zero.
Defining H by H(a, x) = G(a, G(a, x)) — x, show that Hx(a, x) =
Gx(a,G(a,x))Gx(a,x) - 1, Ha(a,x) = Ga(a,G{a,x)) + Gx(a, G(a,x))G„(a,x).
Deduce that if, moreover, Gx0 = — 1 then H0 = Ha0 = Hx0 = Hxx0 = 0, on
using the subscript zero to denote evaluation at a = 0, x = 0. *Hence show
that in general there is a flip bifurcation of solutions of the difference equa¬
tion with map G at a = 0, x = 0.
64 Classification of bifurcations

Q2.3 An integro-differential equation. Given that

^ + ^a-2j {y(x)}2dxjy = 0 and y(0) = y(l) = 0,

show that all solutions are of the form y(x) = A sin nnx, where A is a con¬
stant and n a positive integer. Hence find all the solutions, and sketch the
bifurcation diagram in the (a, A)-plane.
[Griffel (1981,§11.6).]
Q2.4 Imperfect bifurcations. Sketch the curves F(a, x, (5) = 0 in the (a, x)-plane for
all ‘typical’ values of (5 in the cases
(i) F(a, x, S) = x(x - a) + S,
(ii) F(a,x,8) = x(a — x — x2) + 8,
(iii) F(a,x,8) = x2 + ax + a2 — 8.
Q2.5 The structural instability of a pitchfork bifurcation. Consider the roots of
F(a,x,8) = 0, where F is defined by F(a,x,8) = x3 4- <5x2 — ax. Show that
the pitchfork bifurcation at (0,0) in the (a, x)-plane for 8 = 0 becomes a
transcritical bifurcation for small <5, and that there is a turning point at
( —4<52, — 2<5)- Sketch the bifurcation curves in the (a,x)-plane for 8 > 0.
Q2.6 A model of bifurcation of some steady flows. Consider the bifurcation di¬
agrams of the equation F(l,b,c,x) = 0 in the (b, x)-plane for various values
of c, where F is defined by

F(a,b,c,x) = x3 — 2ax2 — (b — 3)x + c

for all real a, b, c, x. Show that if c — 0 then there is a transcritical bifurca¬

tion but if c ^ 0 then there is a ‘primary’ and a ‘secondary’ branch of
the equilibrium curve such that the primary branch has a fold only when
— ii < c < 0. Sketch the curves for c = 0.4, 0, —0.1 and Show that
the bifurcation set, i.e. the set of values of b and c such that F(l,b,c,x) =
Fx( 1, b, c, x) = 0, is a curve in the (b, c)-plane with equation

(27c- 18h + 38)2 = 4(36 — 5)3.

Show that this curve has a cusp at b = f, c = Sketch the curve.

Show that, for all values of a, the cusp is at b = 3 — fa2, c = — ^aVand
hence that it lies on the curve with equation (3 — b)3 = 27c2.
[Benjamin (1978, p. 17) suggested this as a model of the bifurcation of
toroidal vortices in a liquid between coaxial rotating cylinders.]
Q2.7 A problem of particle dynamics which exhibits a structurally unstable bifurca¬
tion. A particle moves along a smooth wire which is fixed in a vertical plane.
Gravity acts on the particle, and the plane rotates with constant angular
velocity io about a vertical axis in the plane. Take Cartesian coordinates
(x, z), where Oz is the upward vertical, so that the wire has an equation of the
form z = /(x) and the axis of rotation has equation x = 8a > 0. You are
given that the equation of motion of the particle then is
 Problems 65

(1 + f 2)W + ^ (jfi) ~ ®2(* ~ Sa) + Of' = 0.

Show that there may be equilibrium at the point x = X, where X is any
root of gf'(X) = a>2(X — 8a), and that this point is stable if a>2 < gf"(X) and
unstable if a>2 > gf"(X).
Show that if/(x) = \a(x/a)2{\ + (x/a)2} and b is defined by b = au>2/g,
then there is supercritical pitchfork bifurcation at b = 1 and X = 0 when
<5 = 0, but that the bifurcation has a different character when 0 < 5 « 1.
Sketch the bifurcation curves in the (b, x)-plane for both cases. Further,
show that for all real 6 there is marginal stability and bifurcation where
2182b2 = 2(b — 1 )3, the equation of a curve with a cusp at (0,1) in the
((5, h)-plane.
[Drazin & Reid (1981, p. 463).]
Q2.8 The van der Waals equation and the cusp catastrophe. The pressure P, vol¬
ume v and temperature T of a sample of a gas are given to satisfy the
equation of state
(P + a/v2)(v — b) = RT

for positive constants a, b, R. Regarding this as a cubic which determines v

as a function of P for a fixed value of T, show that the three roots of the cubic
coincide with value vc, when P = PC,T = Tc, where the critical temperature
is Tc = %a/21bR, the critical pressure is Pc = a/21b2 and the critical volume is
vc = 3b. Sketch the graphs of the function P(v) in the first quadrant of the
(v, P)-plane for T > Tc, T = Tc, T < Tc, showing that the function is single¬
valued if T ^ Tc but not if T < Tc.
Defining n = P/Pc — 1, <j> = vjv — 1 and 9 = T/Tc — 1, rewrite the van
der Waals equation as F(0, n, <f>) = 0, where F is defined by

F(0, n, <t>) = </>* + i(80 + n)<t> + §(40 - n).

Hence show that there is a double zero v of F if

81(40 -n)2 = -(80 + 7t)3,

which gives a cusp at (0,0) in the (0,7t)-plane.

[In the derivation of the van der Waals (1873) equation it is assumed that
the distribution of the molecules of the gas in space is uniform. This as¬
sumption is invalid for a real gas near its critical point, so the equation
is a poor physical model where its mathematical properties are most
interesting.]
Q2.9 Landau's theory of second-order phase transitions. It is given that the
thermodynamic potential $ of a uniform sample of a pure substance at
constant pressure is modelled by an analytic function of temperature T and
the square of a measure q of the order of the sample, so that

<J>(T,>?2) = <D 0(T) + A(T)q2 + B(T)q4 + - ,

66 Classification of bifurcations

where A(T) ~ a(T — Tc) as T -» Tc, B > 0, a > 0, and Tc is the transition
temperature of the substance. Equilibrium may occur at a stationary point
of <l> with respect to variations of rj, and is stable at a minimum of <t>.
Show that there is stable equilibrium with rj = 0 as T j 7"c and t] ~
{a(Tc - T)/2B{TC)}112 as Tc. Deduce plausibly that near transition the
entropy S = —(8<&/dT)p and the specific heat Cp = T(dS/dT)p are given ap¬
proximately by the forms

S0(T) C0(T)
S(T) — CAT) =
S0(T) - a2(Tc - T)/2B' C0(T) + a2T/2B

T>Tq
for
T< T

[This theory, with a discontinuous specific heat, is useful, although the

assumption that the thermodynamic potential is an analytic function of
tj is invalid physically, and leads to results which are not very accurate.
Landau & Lifshitz (1980, Chap. XIV).]
Q2.10 Weiss 's theory of ferromagnetism. The magnetic moment M per unit volume
of a uniform sample of material in equilibrium is given to be

M = N^tanh(A^M//cBT),

where N is the number of unpaired electrons per unit volume, g is the

magnetic moment of an electron, A is a measure of the self-interaction of the
magnetic moments of the electrons, kB is Boltzmann’s constant, and T is the
temperature of the material. Show that

m = tanh(m/f),

where m = M/Ng, t = T/Tc and the Curie temperature Tc = N/.g2/kB. Show

that m = 0 is the unique solution for t ^ 1 but that there are three solutions
tn for t < 1. Sketch the bifurcation diagram in the (f, m)-plane. Deduce that
M = 0 is the unique solution for T > Tc but that there are two extra solu¬
tions M for T < Tc such that

M{T) ~ ± {(3/cbN/A)(Tc - T)}112 as T j Te.

[Weiss (1907) explained the ferromagnetism of a material below its Curie

temperature, although his original work, in the absence of a knowledge of
quantum mechanics, was oversimplified. Physical arguments show that the
solution M = 0 is stable for T > TQ> and that the other two solutions are
stable for T < Tc. See Kittel (1986, Chap. 15).]
Q2.11 Phase instability. Consider the equation

— = (a - |z|2)z + e.
 Problems 67

where z is a complex function of the real variable t but a and e are real
constants. Expressing z = re'fl for non-negative modulus r and real phase 0,
show that

dr d0 esinfl
= (a — r2)r 4- ecosfl.
d; dt r

Investigate the steady solutions and their stability. Sketch the bifurcation
curves in the (a, r)-plane for fixed e > 0, e = 0 and e < 0.
[Note that at a turning point an eigenvalue in general has a real part
which changes sign, but that both branches of the bifurcation curve may
represent unstable solutions if another eigenvalue has positive real part at
the turning point.]
Q2.12 Generalized Newton-Raphson method. Given that f: IT -» [Rm is continu¬
ously differentiable, that its Jacobian matrix J(x) at x is invertible for all x,
and that X is a zero of f; show that a linear approximation to f near X gives
rise to the difference equation

xn+i = F(x„) forn = 0,1,...,

in order to find successive approximations to X, where F(x) =

x — J 1 (x)f(x); and deduce that x„ -» X as n -> oo if x0 is sufficiently close
to X.
Q2.13 The swallow's tail. Take V = 5.x5 -F jax3 + jbx2 + ex and find F =
— dK/d.v. Consider the bifurcation set in (a, b, c)-space defined by F =
dF/dx = 0. Sketch the intersections of the set and three planes a = constant,
for a < 0, a = 0, and a > 0. Hence sketch a perspective view of the bifurca¬
tion set in (a, b, c)-space.
[Thom (1975, §5.3).]
 3
Difference equations

Where do we come from? What are we? Where are we going?

Paul Gaugin ('D'ou venons-nous? Que sommes-nous? Ou allons-nous?', paint¬
ing of 1897, now in Museum of Fine Arts, Boston, MA)

1 The stability of fixed points

To continue our detailed examination of the ideas introduced in Chapter 1,
we shall next consider the dynamics of difference equations, i.e. of iterated
maps, because they are in most respects the simplest of nonlinear systems.
We noted that the methods of solution of difference equations are similar
to the methods of solution of differential equations, and we related how a
fixed point is the analogue of an equilibrium point of a system of ordinary
differential equations. In this section we shall elaborate the analogy and
then examine the stability of fixed points, much as we examined the
stability of equilibrium points. Periodic solutions and their stability are
examined in §2, and some fundamental analysis of difference equations
is given in §3. The logistic difference equation is treated at length in §4,
not only because it is a fascinating equation whose solutions have a rich
structure but also because it is used as a prototype of many nonlinear one¬
dimensional difference equations. Some relevant numerical methods are
introduced in §5. Iterated two-dimensional maps are described in §6 and
iterated maps of the complex plane in §7.
First note that it is sufficient to consider a system of first-order differ¬
ence equations rather than an equation or equations of higher order. For
an mth-order difference equation,

xn+m ~ F(xn, xn+l,..., xn+m_j, n) for n = 0,1,...,

where F: IRm x Z -► (R, can be expressed as m simultaneous first-order

equations yn+1 = G(y„, n) for y„ e R" and G: Um x Z -* Rm by defining y„ =
[xn>xn+i>---iXn+m-i]T a°d G{y,n) = [y2,y3,..., yn+m_1, F(yx,...,ym, n)]T,
where y = [yi,...,ym]T. The converse is false, i.e. m first-order equations
are, in general, not equivalent to one mth-order equation.

68
 3.1 The stability of fixed points 69

Example 3.1: reduction of a second-order equation to two first-order equa¬

tions. It is easier to understand this idea by looking at a simple example.
So consider the second-order linear difference equation,

x„+2 + 2 bxn+l + cx„ = 0,

and define

" ~yn i"

y„
_y«2_

Then

yn2
_*„+2_ _~2byn2 ~cynl_
say, where we define G by G([y1,y2]T) = [y2,-lby2 - cy^.a

Next note that we may consider autonomous systems, i.e. those that do
not depend explicitly on the independent variable n, without loss of gener¬
ality. To show this consider the system

x„+i = F(x„,n)

for F: IRm x Z -* IRm. It may be expressed in the autonomous form,

y„+i = G(y„),
where y„ = [xn,y„.m+1]T e Um+1 and G(y) = [F(x,ym+1), 1 + ym+1]T and
yo.m+i = 0- This gives G: [Rm+1 -* IRm+1. The essential idea here is to extend
the definition of x„ e IRm to y„ s IRm+1 by the addition of an extra compo¬
nent which we arrange to be n itself.

Example 3.2: reduction to an autonomous system. Again a simple example

shows the idea more clearly. To put the equation

xn+1 = x2„- n

into autonomous form, we define y„ = [y„i,>'n2]T and take yn+1 =

l>ni - ynn 1 + yniV for n = 0, 1,..., where y0 = [x0,0]T. Then it follows
that y„ = [x„, n]T and that xn+1 = yn+1 t = x2 — n, as required. □

Now, to consider the stability of a fixed point, it is sufficient to confine

our attention to an autonomous system of first-order difference equations.
First it will help to recall the theory for the case m = 1 in §1.8. There we
tacitly assumed the intuitive notion of stability of a fixed point of a differ-
70 Difference equations

ence equation, namely that if the initial solution is sufficiently close to the
fixed point then the solution remains close to the fixed point thereafter.
This notion can be formalized mathematically in many ways which are
similar but not the same. Here we shall adopt a definition which is essen¬
tially due to Liapounov (1892): a fixed point X of F: IRm -► IRm is stable if for
all e > 0 there exists <5(e) such that

|F"(x0) — X| < e for m = 1,2,... (1)

for all x0 such that |x0 — X| < S, where x„ = F"(x0) is defined iteratively
by xn+1 = F(xn). The definition means that the set of solutions {x„} is
uniformly continuous at X for all n > 0 with respect to variations of x0.
Thus, for example, when m = 3, ‘the fixed point X is stable’ means that
‘given any small sphere with centre X and radius e, there exists a second
sphere with centre X and radius S such that if x0 lies in the second sphere
then Xj, x2,... all lie in the first sphere’ (in exceptional cases the second
sphere is the same as the first, but in general it is smaller). The fixed point X
is asymptotically stable if it is stable and x„ -» X as n -+ oo. It is globally
stable if it is stable and if for all e > 0 and for all x0 there exists N(e) such
that |F"(x0) — X| < efor n = N, N + 1,_It is metastable if it is stable but
not globally stable, i.e. if it is stable to all small initial perturbations x0 — X
but not to some which are not small.
Generalizing the theory of §1.8, take the fixed point X of the difference
equation xn+1 = F(xn) and define its perturbation as

K = x„ - X. (2)
Therefore the difference equation becomes

X + <+1 = F(X + x'„)

= F(X) + Jx'n + o(|xJ) as |xj,| -+ 0,

if F is continuously twice differentiable, where J = [grad F]x is the m x m

Jacobian matrix of F evaluated at X, i.e. the matrix with element [dF-Jdx^x
in the ith row and j th column. Now X = F(X). Therefore

Xn + 1 = (3)

the linearized system, arises in the limit as |xj -> 0. It is plausible that the
linearized system will govern the behaviour of the perturbation x^ when it
is small, and so determine the stability of the fixed point of the nonlinear
equation. In any event, equation (3) has solutions of the form

< = <?"u. (4)

 3.1 The stability of fixed points 71

where
Ju = qu, (5)

i.e. where u is an eigenvector and the multiplier q is the corresponding

eigenvalue of the real m x m Jacobian matrix J. More generally, we find m
eigenvalues qj belonging to eigenvectors uj, where the m eigenvectors form
a complete set in lRm, although they may occur in complex conjugate pairs.
It follows that, given x'0 e Rm, there exist numbers f2,..., (each is
either real or one of a complex conjugate pair, cf. Q3.50) such that

xo = X fa. (6)
j=i
(For the special case where J does not have m linearly independent eigen¬
vectors, this expansion is not possible for all Xq, and special treatment is
needed, cf. Q3.4.) Therefore
m

K = X ^juj for n = 0,1,.... (7)

7= 1

We deduce that the fixed point X is linearly stable if and only if x^ is

bounded as n -* oo for all Xq, and therefore is linearly stable if \qj\ < 1 for
j = 1, 2, ..., m and unstable if \qj\ > 1 for at least one value of j. It can
be shown under quite general conditions that the solution is nonlinearly
stable when it is linearly stable except possibly at the margin of linear
stability.

Example 3.3: the logistic map. Cf. Example 1.5. For the map F(x) =
ax(l — x) we have F: 1R -*• IR and the Jacobian matrix J is a 1 x 1 matrix
which is simply the derivative [dF/dx]x = a( 1 — 2X'). The eigenvalue is
q = a( 1 — 2X). Therefore q = a for the fixed point X = 0. Therefore the
point is stable for — 1 < a < 1. Similarly, the fixed point X = (a — l)/a has
q = 2 — a and is stable for 1 < a < 3. Note (cf. Example 1.5) that q = 1
and there is a transcritical bifurcation when a = 1, but q = — 1 and there is
a flip bifurcation when a = 3. □

Example 3.4: the two-dimensional case. If m = 2 then

<= £i4?Ui + £2q>2-

If qx and q2 are real and distinct then the eigenvectors and u2 are real
and independent, so we may use Cartesian components x' = £,xq\ and
y'n = £2q2 to express the linear perturbation as

< = *>! +y>2. □

72 Difference equations

In the special case of marginal stability we may order the multipliers so

that 1^1 = \q2\ = ■■■ = \qk\ = 1 and 1 > |<?k+1| ^ ••• ^ \qm\ for some inte¬
ger k such that 1 ^ k ^ m. Then, in the limit as n -*■ oo, \'n = \n — X e E,
where E is the fc-dimensional subspace of !Rm spanned by the eigenvectors
u1? u2,..., uk. Note that E is independent of x'0, although xj, does depend
on Xq.
Bifurcation often occurs at the onset of instability as a parameter varies.
There is a close relationship between the bifurcation of a fixed point as a
parameter changes (§2.4) and the stability of a fixed point at a given value
of a parameter. The same Jacobian matrix plays a role in each of the local
analyses, because each involves a Taylor expansion about the same fixed
point at the same value of the parameter.
Suppose that a nonlinear system depends on a parameter a such that
the fixed point X(a) is stable for a < ac and unstable for a > ac. Then there
is marginal stability when a = ac, as described for the linearized system
above. There are centre manifold theorems which, under a variety of quite
general hypotheses, show that for sufficiently small values of a — ac solu¬
tions of the nonlinear system are such that, in the limit as n -*■ oo, x„ — X
belongs to a /c-dimensional manifold. For real systems, an eigenvalue is
either real or one of a complex conjugate pair. So, for k = 1 and q = 1, we
expect, in general, a turning point or a transcritical bifurcation at a = ac,
with xn — X -*• as n -*• oo and a -*> ac. For k = 1 and q = —1 we ex¬
pect a flip bifurcation at a = ac, with x„ - X ~ (— l)"f-Uj as n -► oo and
a -► ac. For k = 2, qx = e'8 for real 6 and q2 = qlt we except a Hopf-like
bifurcation, with x„ - X ~ {1eto#u1 + ?1e_i',eu1 as n -> oo. A Hopf bifurca¬
tion, supercritical or subcritical, of the nonlinear system leads to an orbit
along a closed curve in a two-dimensional submanifold of IRm.

2 Periodic solutions and their stability

In §1.8 we met a solution of period two as well as fixed points. More
generally, we say that if there exist a map F: lRm -> |Rm and distinct points
Xr for r = 1, 2,..., p such that Xr+1 =F(Xr)forr= l,...,p- landXi =
F(Xp) then the set S = {X1,X2,...,Xp} of p points is a solution of the
difference equation xn+1 = F(x„) with period p or a p-cycle. Thus, for a
p-cycle, p is the least positive integer such that xB+p = x„ for all n > 0. Note
that F(S) = {F(Xt),F(X2),...,F(Xp)} = {X2,X3,...,X1} = S, so we say
that S is an invariant set under the operation of the map F, just as a fixed
point is.
 3.2 Periodic solutions and their stability 73

The stability of a p-cycle S may be defined by replacing X by each point

of S in the above definition of stability of a fixed point, so that S is stable if
F"(x0) is in one of the p hyperspheres with radius e and centre X; for; =
1,..., p if x0 is in one of the p hyperspheres with radius S and centre X,.
To consider the stability of a p-cycle it is helpful to consider the pth-
generation map, i.e. the map F composed with itself p times. For the
simplest example, consider a two-cycle {X, Y} such that

Y = F(X), X = F(Y), (1)

where F: lRm —► [Rm, and define the second-generation map G of F by G(x) =

F2(x) = F o F(x) = F(F(x)) for all x e ffT. Note that a fixed point of G
must be either a fixed point of F (when Y = X) or a point of a two-cycle of
F (when Y / X).
A two-cycle {X, Y} of a continuous map F is stable if and only if the
fixed point X (or Y) of the second-generation map G is stable. To under¬
stand this note that if x0 lies within a distance S of X implies that G"(x0) lies
within a distance e of X for all n > 0 then it follows by continuity that there
exists S' such that F"(x0) lies within a distance e of either X or Y for all
n > 0 if x0 lies within a distance S' of either X or Y.
We shall suppose that F is continuously twice differentiable. Then the
linearized system for G at X is

x;+1 = K(X)X;,

where K(X) is the Jacobian matrix of G, i.e. the matrix with element
[dGJdXj^x in the ith row and;'th column. Now

dG^Xj) _ dF^jxj))
dxj dxj

~8_F[ ~sf;
_dxk_ F(x) 8 Xj

on differentiating a function of functions, so

K(X) = J(F(X))J(X)

= J(Y)J(X),

where J is the Jacobian matrix of F. This gives stability of the two fixed
points, X and Y, of G if all the eigenvalues of the m x m product matrix
K(X) have moduli less than one, and instability if the modulus of at least
one eigenvalue is greater than one. It follows that the two-cycle of the
first-generation map F is stable or unstable respectively.
74 Difference equations

Example 3.5: the logistic map. In §1.8 we found that the logistic map
F(x) = ax{ 1 — x) has the two-cycle X, Y = [a + 1 + {(a + l)(a — 3)}1/2]/2a
for a > 3. This gives F: M-+U with dF/dx as the 1 x 1 Jacobian matrix.
The multiplier of the second-generation map is the product

dF dF
a =
dx X dx

= a(l - 2X) x a(l - 2y)

= a2{l - 2(X + T) + 4*T}

= a2{l — 2 (a + 1 )/a + 4 (a + 1 )/a2}

= 4 + 2 a — a2.

Therefore the two-cycle is stable for 1 > \q\ = |4 + 2a — a2|, i.e. for
3 < a < 1 + ^6 = 3.4495, and unstable for a > 1 + yj6. Note further that
the second-generation map has q = - 1 at a = 1 + ^6, where it has a flip
bifurcation (and so there is a bifurcation of the first-generation map from a
stable two-cycle to a stable four-cycle). □

In summary, we note that, for a given map F and initial point x0, either
(a) x0 is a fixed point, (b) F"(x0) is a fixed point for some positive integer n,
i.e. x0 is an eventually fixed point of F, (c) F"(x0) tends to a fixed point as
n^> co, i.e. x0 is an asymptotically fixed point, (d) x0 is a point of period p
for some integer p ^ 2, (e) F"(x0) is a point of period p for some positive
integer n, i.e. x0 is an eventually periodic point, (f) F"(x0) tends to a solution
of period p, so F"(x0) has a subsequence which converges to a point of
period p, i.e. x0 is an asymptotically periodic point, or (g) F"(x0) is aperiodic
and does not approach a p-cycle, i.e. none of the previous cases is applica¬
ble. In the last case we sometimes say that the orbit is chaotic.

3 Attractors and volume

3.1 Attractors
In studying difference equations of the general form xn+1 = F(xn) for
n ~ 0’ •••> where F: IRm —* IRm, we are often interested in sequences
(x05Xj,...} for given F for all x0, i.e. in the orbits {x„} = {F"(x0)}. In
particular, we are interested in the behaviour of xn as n —► oo, just as we
are interested in solutions x(t) of differential equations as the independent
 3.3 Attractors and volume 75

variable t -* co. If there exists x^, and an infinite subsequence {x„ } for
integers nr such that 0 < nx < n2 < n3... and x„r - Xqo as r - oo then we
say that x^ is a cluster point, a point of accumulation or a limit point, of the
sequence {x„}. It follows, for example, that if x„ -► X as n -* oo then X is the
cluster point of {x„}, and if {x„} = {± -± f, -f,|, -§, f, -f,...} then 1
and — 1 are the cluster points.
Let C be the set of cluster points of the orbit {F"(x0)} for given F and x0.
The set C is called the co-limit set of the orbit, because it is essentially the
limit of the orbit for large positive n, and co is the last letter of the Greek
alphabet. Then F(C) = C, because {F"+1(x0)} has the same set of cluster
points as {F"(x0)}, i.e. C is an invariant set of the map F. Thus the concept
of an co-limit set embraces both a fixed point and a p-cycle. The co-limit set
C is closed, because the limit of a sequence of points of the set is a limit of
some subsequence of points of the orbit and so belongs to C.
We define the attractor A of F and x0 as the set C when all orbits near C
have the same set C of cluster points. Thus all attractors are co-limit sets
but not all co-limit sets are attractors. The definition of attractor may be
made precise in various slightly different ways, but our informal definition
will suffice here. It follows that A is a closed invariant set of F. We may
define an oc-limit set and a repeller of the orbit (F"(x0)} similarly by taking
cluster points in the limit as n-* — oo.
*A less informal definition is that the set A c= (Rm is an attractor of F if 3
an open set N c Rm such that V x e N the distance between F"(x) and A
tends to zero as n tends to infinity and there is no proper subset of A
satisfying these conditions. Thus F"(x) approaches all points of A and no
proper subset of A is an attractor.
Different authorities define an attractor in slightly different ways, but an
attractor is essentially what we would find after a long time, by starting at
any point near it and then computing the iterations of the map, whereas an
co-limit set is an ideal property of a single orbit. Note that an attractor may
be a single point, i.e. an asymptotically stable fixed point, or p points, i.e. a
cycle of period p. We have met examples of fixed points and of two-cycles,
i.e. cycles of period two, above. In §4 we shall show that some limits of
orbits are aperiodic and so an attractor may be an infinity of points.

Example 3.6: some attractors of the logistic map. The previous examples
give the following. The fixed point X = 0 of the logistic map F{a, x) =
ax(l — x) is an attractor for — 1 < a < 1; it is in fact a cluster point of {x„}
for all x0 e [0,1]. For 1 < a < 3, the point X = 0 is a cluster point of {x„},
but only for x0 = 0 or 1; the fixed point X = (a — l)/a is the attractor and
76 Difference equations

X = 0 the repeller. For 3 < a < 1 + yj6, the two-cycle is an attractor and 0
and (a — 1 )/a are repellers. □

For a given map F and attractor A, we define the domain of attraction,

or basin of attraction, D(A) as the set such that if x0 e D then A is the set of
cluster points of the sequence {x„}, i.e.

D(A) = {x: the set of cluster points of (F"(x)} is contained in A}.

It follows that D(A) is an invariant set of F, i.e. F(D(A)) = D(A), because

if x is in D(A) then F"(x) is in D(A). Also if F is continuous then D(A) is
an open set; because an orbit tending to A lies in D(A) and an orbit in
D(A) tends to A, and so if F"(x) -> A as n -*■ oo then F"(x + y) -*■ A for all
sufficiently small |y|, by continuity.
Suppose that xn+1 = F(a,xn) for n = 0, 1, ..., where F: [R1 x [Rm -► (Rm.
Then a0 e R' is said to be a bifurcation value if the topological character of
the set of attractors or repellers of F changes when a varies in a small
neighbourhood of a. This concept of bifurcation embraces a change in the
number or the stability of the fixed points of F as a varies through a0.

Example 3.7: some domains of attraction of the logistic map. Given

F{a,x) = ax(l — x), consider what points are mapped where; in particular
it is helpful to find what are the maps and inverse maps (i.e. the images and
pre-images) of the fixed points and to remember that the map is continu¬
ous. Then careful algebra shows that D(0) is the open interval ((a — l)/a,
1/a) when 0 < a < 1, and D((a — l)/a) is the open interval (0,1) when 1 <
a < 3. When 3 < a < 1 + yj6 the domain of attraction of the two-cycle is
the union of the two open intervals (0,(a — 1 )/a) and ((a — l)/a, 1).
These results may be more easily understood geometrically, by sketch¬
ing the curve y = F(a, x) and the line y = x in the (x, y)-plane for fixed a,
than algebraically. In the construction of Fig. 3.1 (see also Fig. 3.3(h)) the
intersections of the curve y =/(x) and the line y = x give the points (A, X),
where f{X) = X, and thence the fixed points. The construction can be seen
to determine geometrically the point (/(x), 0) when given the point (x,0).
This enables us to see qualitatively the fixed points, the pre-images of the
fixed points, the two-cycles etc. and thence the domains of attraction. □

Example 3.8: the Bernoulli shift (cf. von Neumann 1951). Let 6n+l be the
fractional part of 26n for n = 0, 1,..., i.e. consider the difference equation
9n+l — o(6n), where the function o is defined by

ct(x) = 2x modulo 1 and 0 ^ <r(x) < 1. (1)

 3.3 Attractors and volume 77

Fig. 3.1 The geometrical construction of the fixed points of a map/of the real
line, and of the map f(x) of a given value x.

Therefore <r(x) = 2x if 0 ^ x < \ and a(x) — 2x — 1 if ^ < x < 1. The rela¬

tionship of 9n+1 to 9n by this ‘sawtooth’ map is illustrated in Fig. 3.2.
We see at once that 0 is the unique fixed point. Rather than setting out
methodically to find the p-cycles, and then their stability, in turn, we can
find all the dynamics of the difference equation as follows. We shall meet
our first example of chaos.
Note that 0 < 91 < 1 so that we may take 0 ^ 90 < 1 with little loss of
generality. Then we can express 90 = dj2 + d2/22 + d3/23 + as a bi¬
nary number with digits dj = 0 or 1 for j = 1,2,.... Now it can be seen that

Fig. 3.2 The sawtooth map, <f> = a(0\ where a(6) is the fractional part of 26.
78 Difference equations

the binary expression of 0n+1 is the binary expression of 6n after the removal
of the first digit (zero or one) to the right of the binary point. (This is called
the Bernoulli shift after Jakob Bernoulli’s (1713) posthumous study of inde¬
pendent random events with only two possible outcomes, e.g. coin tossing.)
So if 60 has a binary expression of finite length, i.e. if 60 = q/2s for some
positive integers s and odd q, then 6n = 0 for all n > s. Also if 90 is rational
and has a binary expression of infinite length then 90 — q/2sr for some
integers q, s 5s 0, r ^ 3 such that q and 2sr are coprime and r is odd; there¬
fore 9„ = q/2s~nr modulo 1; therefore 9n+p = 9„ for all n ^ s, where the
period p is such that 1 ^ p ^ r, by the pigeon-hole principle. If 90 is
irrational then {9„} is an aperiodic sequence.
For an example, first take 90 — 3; then {9n} = {5,5,3, a two-cycle.
Again, if 90 = \ then {0„} = {j, 0,0,...}, with an eventually fixed point. If
0O = 5 then {9„} = {£,f, f, f,f,...}, a four-cycle. If 0O = 6 then 9l = £ and
the eventually periodic sequence proceeds as above. If 90 = j or f then the
sequence has period three, and so forth. All the periodic sequences are
unstable, because each rational value of 90 is arbitrarily close to an irratio¬
nal value of 90 which generates an aperiodic sequence {0„}. Each aperiodic
sequence in fact goes either through or arbitrarily close to each point of the
interval [0,1],
It follows for this map o that there is an infinity of different co-limit sets
C according to the values of 90; C is finite if 90 is rational and infinite if 90 is
irrational. With the definition of an attractor which we have chosen, no set
C is an attractor, because an initial point 90 in each neighbourhood of C
may give a sequence {0n} with a different set of cluster points.
Let us consider a wider issue next. Two infinite sequences {0^} and {0"}
can be generated by iterating the sawtooth map (1) with irrational initial
values 9'0 and 0q respectively. You can see that, however small 9q — 9'0 is,
0« — 0« may have almost any value between zero and one for sufficiently
large values of n, because the value of 9" - 9'n depends only on the digits
in the binary expressions of 0„ and 9’0 which lie more than n places to the
right of the binary points. When the difference 9” - 9'n is small it in gen¬
eral doubles each time n increases by one, because 0; - 9'„ = 2"(0q - 9'0)
modulo 1. Thus the value of <rn(0o) for a large value of n depends very
sensitively on the initial value 0O. This property, with exponentially grow¬
ing separation of neighbouring orbits, is called sensitive dependence on ini¬
tial conditions (after Ruelle 1979). Indeed, there is a periodic solution for a
rational value of 0O and an aperiodic one for an irrational value of 0O, and,
of course, the sets of rationals and irrationals are each dense in the interval
[0,1]. We may say that an aperiodic solution rapidly ‘forgets’ the initial
condition as n increases. Sensitive dependence on initial conditions is
 3.3 Attractors and volume 79

sometimes used as a characteristic to define chaos, although a satisfactory

precise definition is elusive. This issue is elaborated in §4.4.
There is an important geometrical application of the sawtooth map.
Each point on the circumference S1 of a circle may.be represented by its
polar angle (/> and so uniquely by 9 = </>/2n modulo 1, where 0 ^ 6 < 1.
Then the map represents a doubling of the polar angle <f> of a point on the
circle, and in this representation is a continuous map. □

Example 3.9: the rotation map. Another simple, but instructive, map is the
piecewise linear map, F: [0,1) x [0,1) -> [0,1) defined by

F(a, x) = x + a modulo 1 and 0 ^ F < 1.

We take 0 ^ a < 1 without loss of generality. We may again regard this as

a continuous map of the circle S1 in which a point of the circumference is
rotated by a constant angle 2na, because a point with angle 27rx„ is mapped
to the point with angle 27txn+1 = 2nxn + 2na.
Note that x„ = Fn(a, x0) = x0 + na modulo 1.
If a = 0 then F is the identity map. Therefore all x e [0,1) are fixed
points and stable. However, none is an attractor because they are not
asymptotically stable.
If a > 0 is rational then a = q/p for coprime integers 0 < q < p. There¬
fore x„ - x0 = nq/p modulo 1. Therefore x„ / x0 for 0 <n<p and
xp = x0. Therefore {x„} has period p, and the p-cycle is {x0,x0 + 1/p,
x0 + 2/p,...,x0 + (p - 1 )/p} in some order. This cycle is stable, but not
asymptotically stable, because a small change of x0 merely translates the
cycle a little.
Conversely, if {x„} has period p, then 0 = xp- x0 = pa modulo 1, i.e. pa
is an integer. Therefore a is rational.
It may be intuitively sensed that as an irrational a is approximated by a
rational q/p with large p the sequence {x„} is approximated by a large
p-cycle which covers the interval [0,1) uniformly. This can be stated more
carefully as follows.
^Suppose next that a is irrational and 0 < £ < 1. Then x„ goes arbitrari¬
ly close to £. We shall show that either there exist non-negative integers m,
n such that x0 + na — m = in which case x„ = £; or x0 + na — m # £ for
all m,n^t 0. In the former case, one and only one member of {x„} equals £.
In the latter case, it may be proved that there exists an infinite subsequence
{nr} of the integers such that 0 ^nl <n2 <••■ and x„r -> £ as r -» oo, i.e.
that £ is a cluster point of {x„}. To prove this, first choose any integer r > 0
and note that at least two of the r 4- 1 points x0, xt,..., xr must lie in one of
80 Difference equations

the r equal subintervals [0,1/r),[l/r,2/r),...,[(r — l)/r, 1), by the pigeon¬

hole principle. Denote these two points by xh xm for / < m, and define
e = xm — x,, so |e| < 1/r. Therefore xm_, = x0 + (m — l)a = x0 + xm — x, =
Xq + € modulo 1. Similarly, xk(m_D = x0 + ke. Therefore there exists k such
that the distance of xk(m_,) from £ is less than |e|. Therefore |x„p — £1 < 1/r,
where nr = k(m — /). So we can construct a subsequence {xni, x„2,...} such
that |x„r — £1 < 1/r for r = 1,2,.... The result follows. Thus in this case the
co-limit set is the interval [0,1], but is not an attractor. □

3.2 Volume
Another important, but rather different, concept is that of how the volume
of a set of points is mapped. Consider a set S„ cr Rm and a given continu¬
ously differentiable map F: IRm -*• [Rm. Then we may define Sn+1 as the set
F(S„), i.e the points xn+1 = F(x„) for all x„ e S„. Also define pn as the hyper¬
volume, i.e. the measure, of the set S„ of points; so pn is the total length of
the points of the set on the real line if m = 1, the total area of the points of
the set in the plane if m = 2, or the total volume of the points of the set if
m = 3. Now the local ratio of the hypervolume pn+1 to p„ is the modulus of
the Jacobian det J(x„), where the m x m Jacobian matrix J has element
dFJdxj in its ith row and jth column. Therefore if the modulus is every¬
where less than one then p„ decreases monotonically as n increases; in this
case the hypervolume pn shrinks as n increases, and the dimension of the
attractor must be less than m.
If | det J | = 1 everywhere then = /i0 for all n, i.e. the hypervolume n„ is
independent of n, and we say that F is a volume-preserving map or an
area-preserving map as is appropriate, or simply a measure-preserving map.
More generally, F is a measure-preserving map if the measure of all sets S is
the measure of the set of all points mapped to S, i.e. p{S) = ^(F-1(S)) for
all S c [Rm. The definition of attractor we have taken, with the condition
that all neighbouring orbits are ‘drawn’ into the attractor, implies that a
measure-preserving map does not have an attractor.
If F preserves volume and orientation, e.g. if F is a rotation about an
axis, then det J = 1; if F preserves volume but reverses orientation, e.g. it is
a reflection, then det J = — 1.

Example 3.10: an area-preserving map. Take x„ = (x„,yn)e U2, F: 1R2 -+

R2, where the rotation R(x) = (y, -x), the shear S(x) = {x,y + /(x)),

F(x) = S(R(S(x») = (y +/(x), -x+f(y +f(x)))

and/is differentiable. Therefore the Jacobian matrix is

 3.4 The logistic equation 81

J(x) = /'W 1
+f'(x)f'(y +/(x)) f'(y +/M)
and hence the Jacobian is det J = 1 for all x and y. Therefore F is an
area-preserving map. □

Example 3.11: the product of the eigenvalues. If the eigenvalues of the

Jacobian matrix of a map F:[Rm->[ir at x are qlt q2, ..., qm then
• • • Qm = det (dFi/dxj). It follows that if the map is area-preserving then
the product of the eigenvalues is ±1. □

4 The logistic equation

Many of the important properties of bifurcation of nonlinear systems are
exhibited by the logistic difference equation, so we shall examine it in de¬
tail, finding immensely greater richness and depth of results than might be
anticipated from so simple an equation. Some of the concepts which arise
in this way will be elaborated in later sections. We shall see that the logistic
equation not only is a prototype of a large class of one-dimensional non¬
linear difference equations but also shares many properties of higher¬
dimensional difference equations and other nonlinear systems. It may be
said that the logistic difference equation is used in the theory of iterated
maps as the fruit fly, Drosophila, is used in genetics.
We again take

x„+i =F(a,xn), for n = 0,1,..., (1)

where
F(a, x) = ax( 1 — x) for all a ^ 0. (2)

If a ^ 0 and x > 1 or x < 0 then F(a, x) < 0, so let us here take 0 ^ x0

^1 and 1 max0s;jt<1 F(a,x) = \a, i.e. a ^ 4, in order that F maps
the interval [0,1] into itself. We seek to examine the sequences {x„},
mostly for given values of a e [0,4] and x0 e [0,1]. We have found in
§1.8 that the fixed points of F are X = 0 for all a and X = (a — l)/a for
a # 0. The fixed point X = 0 is stable if 0 ^ a ^ 1, and X = (a — 1 )/a
is stable if and only if 1 < a ^ 3. Also there is a two-cycle with Xx, X2 =
[a + 1 ± {(a + l)(a — 3)}1/2]/2u for 3 < a and it is stable if and only if
3 <a^ 1 ^6.
It is helpful to examine graphically how the fixed points and the period¬
ic solutions arise, and also see why they become unstable. The graphical
method will show that the qualitative behaviour of the logistic map is
shared by all maps of the form F(a,x) = af(x) for a smooth and convex
82 Difference equations

Fig. 3.3 The line y = .x and some curves y = af(x) for various values of a, where
f(x) = x(l — x). (a) a = \, l, 2, 3, 4. (b) The graphical construction of x„+1 from
x„, for the case a = j.

function / such that /(O) = 0, f{b) = 0 for some b > 0, and f(x) > 0 for
0 < x < b.
First examine the graphs of y = F(a, x) for various values of a illustrated
in Fig. 3.3(a). Each cuts the line y — x in the fixed points. It can be seen at
once that there is one fixed point X = 0 for all a and also another for a > 1,
because the convex curve lies below its tangent y = ax at the origin if and
only if a ^/'(0) = 1.
Secondly consider the graphical construction of xn+I from x„ illustrated
in Fig. 3.3(h). When the tangent y — ax to the curve y = F(a, x) at the fixed
point X — 0 lies below the line y = x, it can be seen that x„+1 > x„ and x„
approaches the fixed point as n —► oo. Similarly, if the tangent lies above
the line y = x, i.e. if a > 1, then the fixed point is unstable. Also it can be
seen that the slope of the tangent to the curve at the other fixed point X =
(a — l)/a determines the stability likewise.
The geometrical effect of the map can be seen to be equivalent to non-
uniform stretching (when a > 2) of the interval [0,1] followed by folding
back onto the subinterval [0, \d] to give the curve y = F(a, x). This stretch¬
ing and folding is characteristic of maps with infinite attractor sets (cf. §6).
Next look at the graphs of some of the curves with equations y = G(a, x)
for various values of a, where G is the second-generation map defined by
G(a,x) = F(a,F(a,x)) = F2(a,x) for all x, as illustrated in Fig. 3.4(a). Note
that G is a quartic. Each curve cuts the line y = x in the fixed points of the
second-generation map, and thereby gives the two-cycle of F as well as
the fixed points of F. When a =s 1 the curve cuts the line only at the fixed
 3.4 The logistic equation 83

Fig. 3.4 (a) The line y = x and curves y = G(a, x) for the second-generation map
G = F2 of the logistic map F, with a = 1, 2, 3.2. (b) The line y = x and some
curves y = H(a, x) for the fourth-generation map H = F4 of the logistic map F,
with a = 3, 3.7.

point X = 0. When a > 1 it also cuts the line in the other fixed point, X =
{a — 1 )/a. The curve y = G(3,x) touches the line at the fixed point X = §.
When a > 3 the curve y = G(a, x) cuts the line at the origin and three other
points, the middle of which is a fixed point of F and the rest of which are
the two-cycle of F, so the two-cycle exists if and only if a > 3. This illus¬
trates the geometrical essence of a flip bifurcation and period doubling. If
the slope of the curve y = G(a, x) is less than one in modulus where it
intersects y = x, then the intersection represents a stable fixed point of G,
and therefore a stable fixed point or two-cycle of F.
It is also instructive to look at the graphs of some higher-generation
maps F2'. Note that Fq is a polynomial of degree 2q because F is a polyno¬
mial of degree two. Two graphs of the fourth-generation map H are
illustrated in Fig. 3.4(h), where H is defined by H(a,x) = G(a,G(a,x)) =
F4(a,x). This suggests that four-cycles arise as a increases. Indeed, the
graph of the curve with equation y = Fq(a, x) oscillates more rapidly as q
increases, so we anticipate repeated period doubling as a increases.
These and other properties are illustrated by the special cases in which
we can solve the logistic difference equation more-or-less explicitly by
analysis: a = -2 (Q3.19), a = 0 (trivial), a = 1 (Q1.32), a = 2 (Q1.33), and
a = 4 (later in this section).

*Example 3.12: local analysis of a flip bifurcation. We have just seen the
flip bifurcation from a general geometrical point of view, having seen it in
explicit analytic terms for the logistic map in Example 1.6. Another point

D
84 Difference equations

of view is that of perturbation theory. We shall describe here quite a gener¬

al theory of the flip bifurcation, and apply it to the logistic map.
The linear theory of stability of the fixed point X = (a — 1 )/a of the
logistic map F(a,x) = ax{ 1 — x) shows (Example 3.3) that X becomes un¬
stable, the multiplier q decreasing through —1, as a increases through
ac = 3. It gives x„ = X + ( — l)"xo + 0(x'02) as x„ -> 0, where a = ac; so
there is a solution of period two, with x2n = x0 = X + x'0 and x2„+i =
X — x'0iorn = 0,1,_
These results of the linear theory guide our choice of the ansatz (or
assumed form) of the perturbation theory of the flip bifurcation as a -*■ ac,
for the ansatz must allow for the appropriate behaviour. The start of the
problem is that there is a fixed point X(a) with a multiplier q(a) such that
q{ac) = — 1, i.e.

F(a, X) = X for a near ac and Fx(ac, Xc) = -1,

where Xc = X(ac) and a subscript x denotes a partial derivative with re¬

spect to x. We assume that there is also a two-cycle {X1,X2} as a -*• ac
(either from above or from below, but not both) such that

F(a, Xk) = Xj, where j, k = 1,2 or 2,1, (3)

j — *c + e*/l + £2Xj2 + ’" for; =1,2 (4)

a = ac + eax + e2a2 H- as e -> 0. (5)

It only remains to substitute the series (4) and (5) into equation (3),
equate coefficients of successive powers of e, and find the coefficients ax,
*n, ^2i> a2> *i2> ••• in turn. The form of the series (4) and (5), with the
introduction of the new small parameter e, rather than series in powers of
a — ac, is motivated in Example 2.1: the more flexible form (5) allows the
expansion (4) to be effectively either in powers of (a — ac)1/2 or in powers of
(ac - a)1/2 rather than in powers of a - ac, and so to match the properties
of the solution. The expansion (5) will be seen not to be unique, although
the choice of a2, a3,... does not affect the solution itself, of course.
Now substitution of series (4), (5) into equation (3) gives

Fc + e(alFac 4- XklFxc) + e (a2Fac + Xk2Fxc + 2a2Faac + a1XklFaxc

+ 2XkiFxxc) + ••• = Xc + eXn + €2Xj2 + •••,

where the subscript c denotes evaluation of a partial derivative at a = ac

and X = Xc. But Fc = Xc and Fxc = — 1. Therefore
 3.4 The logistic equation 85

€(fliF-C ~ *«) + ^Foc ~ Xk2 + \a\FoaC + axXkXFaxc + hX2xFxxe) + -

= eATjj + e2^ + ' ‘' •

Equating coefficients of e, we find

-*n + X2x = axFac (6)

twice.
Equating coefficients of e2, we find

a2FaC - Xk2 + \a\Faac + axXkiFaxc + $X?xFxxc = Xj2.

Therefore

X12 + X22 = a2Fac + WiFaac + + WXFXXC

and

X12 + X22 = a2Fac + ia2Faac 4- axX21Faxc + \X22lFxxc.

Therefore

a i X 21 Faxc + jX21Fxxc = ax XlxFaxc + jXxxFxxc.

In general X2x ^ Xxx because X2 ^ Xx. Therefore

Xu + X2x = —2axFaxJFxxc (7)

unless Fxxc = 0. It follows from equations (6) and (7) that either ax = 0 or
— 2Faxc/Fxxc = Fac. Assuming the general case, in which the latter equality
is not satisfied, we deduce that

ax = 0.
Therefore

X2x — —Xxx and Xx2 + X22 — a2Fac + iX2lFxxc.

Equating coefficients of e3 in equation (3) (and assuming that Faa = 0 in

order to simplify the calculation a little), we deduce that

^3Fac Xk2 + <2 2 Xk i Faxc -4- XkxXk2 Fxxc + (>Xkx Fxxxc Xj2.

Therefore

X13 + X23 = a3Fac + a2XkxFaxc + XklXk2Fxxc + eXkXFxxxc for k = 1,2.

Therefore

a2^21^a*c + X21X22Fxxc + ^X2XFXXxc = a2XXXFaxc + XlxX12FXXC

+ hXfxFxxxc.
86 Difference equations

Therefore

^■a2^axc + (X12 + X22)Fxxc + 1^1 l^jcxxc = 0

becauseX21 = — Xtl, ^0. Therefore

a2(2Faxc + FacFxxe) + \X\x FX2XC + Fxxxc = 0, (8)

a formula sufficient to give the leading approximation to the behaviour of

the two-cycle as a -*■ ac.
To verify this, we apply the results to the logistic map, for which
F{a,x) = ax( 1 - x), ac = 3 and Xc = f, and so Fac = f, Faxc = Fxxc =
— 6 and Fxxxc = 0. Therefore equation (8) gives X\ 1 = %a2. Therefore

xj~i~ i(-1 y(a - 3)1/2 as a J. 3 for; = 1,2,

in agreement with Example 1.6. □

Another important approach to the problem is the numerical one. We

can calculate the periodic solutions of a difference equation in various
ways. A simple program to calculate the periodic solutions and many other
properties of the logistic equation with a computer is described in the next
section. You can easily confirm for yourself the results in Table 3.1 of
attractors appropriate to various intervals of the parameter a. You may
infer the occurrence of an infinite sequence of period doubling with values
ar at bifurcations such that

as r -*■ oo, (9)

ar ar+1

and therefore

Table 3.1. Some attractors of the

logistic map for a ^ 0

Interval Attractor

0 a ^ 1 X =0
1 < a ^ a, = 3 X — (a — 1 )/a
at < a ^ a2 = 3.449 21- cycle
a2 < a ^ a3 = 3.544 22- cycle
a3 < a ^ a4 = 3.564 23- cycle

ar < a < af+1 2r-cycle

 3.4 The logistic equation 87

ar = ao0 — AS r + o((Tr),

where ax = 3.5700... and S = 4.6692.... A sequence with an asymptotic

property of the form (9) for the same value of d is called a Feigenbaum
sequence, after Feigenbaum (1978), who showed that the qualitative char¬
acter of the period doubling and the exact value of the constant S in rela¬
tion (9) are independent of the particular map F chosen, although the
constants a^ and A do depend on F. So a Feigenbaum sequence is charac¬
teristic of a wide class of maps with the same form as F and not just of
the logistic map. In this sense <5 is a universal constant. Indeed, period
doubling and Feigenbaum sequences will be seen to occur for difference
equations which are not measure-preserving, whether they are one- or
higher-dimensional, and also for many nonlinear ordinary and partial
differential equations. We shall show some details of the period doubling in
the limit as r -kx> and demonstrate the universality of the limit of the
period doubling in §4.3.
The properties are even more interesting when a is increased above a^.
All the solutions of periods 2r which originated for a < a^ persist for a >
fl»o- but they are unstable and in addition there is an attractor with an
infinite number of points x for most values of a> ax. Such an attractor
gives an aperiodic solution, and is called a strange attractor (after Ruelle &
Takens 1971). It is often called a chaotic solution, or simply chaos (after Li
& Yorke 1975), because, as we shall see, the solution may be regarded as a
sample of a random variable. It may help to imagine the chaos as being due
to the infinity of repellers, each of which repels the solution in its orbit in
the manner in which the ball is repelled by the pins in a pin-ball ma¬
chine; this model is crude but has some essence of the truth. The attractor is
called a Cantor set because it is topologically equivalent to a famous set
defined by Cantor (1883). In the next chapter we shall elaborate the aspects
of set theory which arise. For the present we add that a Cantor set of points
is uncountable and its points are disconnected, in the sense of there being
no interval contained in the set, although to each point of the set there is
another point arbitrarily close, and each convergent subsequence of the set
converges to a member of the set, i.e. a Cantor set is closed. You will find
the nature of the attractor easier to understand if you use the computer
program described in the next section and you read the set theory in the
next chapter.
Although chaos is the rule for a > a^, there are ‘windows’ i.e. small
intervals of a, for which a stable periodic attractor exists. In particular you
may like to use the computer program of the next section to demonstrate
88 Difference equations

period

Fig. 3.5 A sketch of part of the bifurcation diagram in the (a, x)-plane where
there are orbits of periods 3 x 2r for r = 0, 1, .... Dotted curves denote the
stable three- and six-cycles, and dot-dashed curves the unstable three-cycles.

that there is a stable three-cycle or (3 x 2r)-cycle for 3.8284 < a < 3.8495.
Look at Fig. 3.5, and note the origin of the stable and unstable three-cycles
due to a turning point of the third-generation map F3. In fact, the solution
undergoes period doubling so that solutions of periods 3 x 2r arise as a
increases, and the critical values of a at which there is bifurcation form
another Feigenbaum sequence! These three-cycles originate because F3 is
a polynomial of degree eight, and so F3 has eight fixed points, which are
real or occur in complex conjugate pairs; two are the fixed points of F for
all values of a; and three complex conjugate pairs become real as a in¬
creases through 1 + JS = 3.8284 (Q3.17). Further, there is an infinity of
windows in which p-cycles are stable for 3 < a < 4, such that for each
integer p there is some interval of a for which an attractor is a p-cycle. For
example, a stable five-cycle arises when a increases through 3.7382. Of
course, some of the windows in which there exists a stable periodic solution
are very narrow indeed.
The bifurcation diagram for 0 ^ a ^ 4 is summarized in Fig. 3.6. Note
the many windows for a > 3.6.
The case a = 4 is rather special (von Neumann 1951), because then
maxo<*< i F(4,x) = 1 and F maps the interval [0,1] onto itself. This prop¬
erty helps us to ‘solve’ difference equation (1). We find that according to the
 3.4 The logistic equation 89

Fig. 3.6 Computer-drawn diagrams of the attractors of F(a,x) = ax(l - x) in

the (a, x)-plane for (a) 0 < a ^ 3.9 and (b) 3.5 < a ^ 4.

initial value x0 there may be an unstable orbit of period p for all positive
integers p. Thus there is a countable infinity of periodic orbits. These unsta¬
ble periodic orbits are the continuation of all the periodic orbits which
arose for 3 < a < 4. However, almost all initial values x0 give orbits of
points x„ with a set of cluster points which is the interval [0,1] itself. These
results may be shown as follows.
Suppose that
= P(0n), (10)
where
P(9) = sin2 nO (11)
90 Difference equations

(b)
Fig. 3.6 (cont.)

Note that P has period one and is a 2 — 1 mapping of the interval [0,1]
onto itself, as shown in Fig. 3.7, so in general there are two possible values
of 0„ for a given value of x„ e [0,1], Next suppose that 0n+1 is the fractional
part of20„,i.e. that

0„+i = 20„ modulo 1 and 0 ^ 0n+1 < 1, (12)

and 0 ^ 0O < 1. Now relations (10) and (11) give

x„+1 = sin2 7i0n+1

= sin2 2n0„,

because 0n+1 is given by equation (12), and P has period one,

 3.4 The logistic equation 91

Fig. 3.7 The 2 - 1 mapping P of the interval [0,1] onto itself, for P(Q) = sin2 nO.

= 4 sin2 n9n cos2 nd„

= 4x„(l - xj

= F(4,xn). (13)

We can now regard the original nonlinear difference equation (13) as

equivalent to the simpler piecewise-linear equation (12) by use of the trans¬
formation (10). So we may in principle solve equation (13) for a given value
of x0 by first inverting equation (10) to find a value of 0O, then solving
equation (12) as in §3 to find 0„, and finally finding x„ from equation (10), so
that x„ = sin2(2";r0o).
We have met the sawtooth map (12) in Example 3.8. Here we note that a
fixed point X of the logistic equation (13) is the map (11) either of the fixed
point 9 = 0 of equation (10) or of a two-cycle of equation (12), because P is
a 2 — 1 map and /*(1 — 9) = P(9) for all 9. In fact X = 0 corresponds to
9-0 and X = | to the two-cycle, 02m_1 = 3 and 02m = f for m = 1,2,....
The two-cycle of equation (13) must similarly correspond to a two-cycle
or a four-cycle of equation (12). In fact the two-cycle Xx, X2 = £(5 ± y/5)
= sin2{(3 ± l)7r/10] = sin2[(7 + l)7i/10] corresponds to the four-cycle
04m-3 = 5> 04m-2 = 5> ^4m-l = I &nd 04m = J.
92 Difference equations

Let us consider two wider issues next. Although they are of general
importance, they can be simply illustrated by the logistic equation with
a = 4. First recall from Example 3.8 that if there are two infinite sequences
{0'} and {0"} generated by relation (12), with irrational initial values 9'0
and 0q respectively, then 0" — 0' = 2"(0q — 9'0) modulo 1. There is a corre¬
sponding property of the sequences {x„} that the value of x„ for large
values of n depends sensitively on the initial value x0, with exponentially
growing separation of neighbouring orbits. Indeed, there is a periodic solu¬
tion if x0 corresponds to a rational value of 0O and an aperiodic one if x0
corresponds to an irrational value of 0O, and, of course, the sets of rationals
and irrationals are each dense in the interval [0,1]. We have demonstrated
the sensitivity to initial conditions for the special value a = 4 because the
sensitivity can be made explicit, but a similar sensitivity occurs for other
values of a for which there is an attractor which is infinite. The attractor is
the same set for all initial values within the domain of attraction, but the
position of x„ within the attractor for a given large value of n depends
sensitively upon x0. Also many other one-dimensional and higher-dimen¬
sional nonlinear ordinary differential equations and other nonlinear sys¬
tems show sensitive dependence on initial conditions.
Secondly, we see how ideas of probability are useful. Each difference
equation determines x„ precisely in terms of x0, yet in some respects a
sequence {x„} which has an infinite set of cluster points may be regarded as
a sequence of random numbers. Indeed, piecewise-linear difference equa¬
tions of the form <f>n+l = p(j>„ modulo q for integers p and q have been used
in computers to generate so-called random numbers. We can illustrate this
with equation (12). It can be proved that if 0O is irrational then the infinite
sequence {0„} is uniformly distributed, i.e. there exists

lim {(number of times 0„ e [0, c] for n - 1,2,..., N)/N} = c — b

N~* oo

for all b, c such that 0 ^ b ^ c ^ 1. In the language of probability theory

we may say that the probability that 0„ lies in the interval [h, c] is c — b, i.e.
0n is a sample of a random variable 0 which satisfies

P{b ^ 0 ^ c) = c - b. (14)

Now P(1 — 0) = P(9) and P is a 2 — 1 function, as shown in Fig. 3.7, so

relation (10) gives x„ as a sample of a random variable X such that
 3.4 The logistic equation 93

when b and c lie on the same side of i.e.

P(B ^ X ^ C) = 27t-1(arcsinx/C - arcsin^/fl), (15)

on putting nb = arcsin^/B and nc = arcsin^/C. In particular, this gives the

probability distribution function of the random variable X as

F(x) = P(X ^ x)

= 2n~' arcsin^/x. (16)

Also the probability density function of X is given by

fix) = F\x)

1
7r{x(l — x)}1/2’

This function gives the probability f(x)dx that X lies in the elemental
‘interval’ [x,x dx], and so shows that x„ is found much more often near
the ends than it is at the middle of the interval [0,1]. (Note that the vari¬
able X and these functions F and / are different from others which have
been denoted above by the same letters.) We see that, although a particular
value x„ may depend sensitively on x0, the statistical properties of the
sequence {x„} are the same for all values of x0 which correspond to irratio¬
nal values of 90. The Cantor nature of the attractor for other values of a
complicates the probability distribution function, because it may be dis¬
continuous, and a fortiori not differentiable, at an infinity of points. None¬
theless probabilistic ideas are still useful.
The philosophical implications of these ideas deserve consideration.
For a given value of x0 the sequence {x„} is in principle determined com¬
pletely and precisely by an algorithm, namely iteration of the map F. Yet in
practice the sequence may be indistinguishable from a sequence of random
numbers. In laboratory experiments, however careful, observations are not
absolutely precise, and there is no possibility of distinguishing an irrational
from some rational value of a datum. Also numerical errors, such as round¬
off error, are inevitable in the practice of processing the data to find x„+1
from x„. So we could not predict all the future of a chaotic system, even if
we had an exact model on which to base our predictions. The errors in
prediction grow exponentially in time (or, rather, in n) because of our lack
of exact knowledge of the present, so that doubling the accuracy of our
measurements and data processing will avail only a little: we may predict
with confidence only the near future of chaotic solutions and their statisti-
94 Difference equations

cal properties over a long time. It is because of this, perhaps, that such
chaos is sometimes called deterministic chaos.

In summary of this section, we state that each co-limit set for the logistic
difference equation is one of four types, (a) It may be a finite attractor as a
stable periodic solution or fixed point, for example when a = 3.2. (b) It may
be an unstable fixed point or p-cycle, for example the two-cycle when a =
3.5. (c) The co-limit set may be an infinite attractor as a Cantor set with
chaos, for example when a — 3.6. (d) It may be an interval with no
attractor, for example when a = 4. For a given value of a, the co-limit sets
may be of more than one type according to the value of x0 (see Q3.18 for an
example). The qualitative behaviour of the solutions of the logistic equa¬
tion can be discerned to be independent of the symmetry and of the simple
parabolic form of the map, because most of the arguments we have used do
not depend on these special properties of the map, although the quantita¬
tive behaviour does. Therefore the ideas and qualitative results of this sec¬
tion are widely applicable to all one-dimensional maps of similar form.

5 Numerical and computational methods

Computers are invaluable tools with which to investigate nonlinear sys¬
tems. Numerical results, especially when they are presented graphically,
are useful aids to learning. Also they may suggest conjectures and stimu¬
late new lines of mathematical enquiry, just as observations of natural
phenomena have done for millennia. The conjectures may be proved math¬
ematically afterwards. The aim of this section is to encourage readers to
make their own numerical experiments and use them.
First, however, a warning should be given. Sometimes numerical experi¬
ments may suggest false results, because they usually model a mathema¬
tical ideal only approximately, and small errors may be important when
a solution is unstable or chaotic. We have seen how a differential equa¬
tion may be modelled by a difference equation in numerical calculations.
This leads to truncation errors. Most computers are digital rather than
analogue, using discrete rather than continuous variables. So no fixed- or
floating-point variable, however accurate the computer is, can represent a
real variable exactly. This leads to round-off errors. We have already met
sensitive dependence on initial conditions, for which it may be significant
whether a number is rational or irrational, and so can see an inherent
limitation of computers. Indeed, when solving the logistic equation on any
digital computer, there is only a finite number of discrete values to repre-
 3.5 Numerical and computational methods 95

Table 3.2. List of the BASIC program ‘LOGIS’ (You may find it more
convenient to replace lower by upper case letters throughout.)

10 INPUT “aO, astep, ia’\ aO, astep, ia%

11 REM Type in initial value of a, step for values of a, number of steps
20 MODEO
30 FOR i% = 0 TO ia%
40 a = aO + astep*i%
50 PRINT TAB(60,5), a
60 x = .501
61 REM Take xO = 0.501
70 FOR n% = 1 TO 192
80 x = a*x*(l. —x)
81 REM Specify the logistic map
90 IF n% < 65 THEN SOUND 1, -10,255*x, 2 ELSE PLOT 69, i%*16, x*1000
91 REM If n < 65 play a note of frequency representing x; if n ^ 65 plot
the pixel for current values of a (horizontal coordinate), x (vertical coordinate)
100 NEXT n%: NEXT i%
110 STOP

sent a real variable, so every solution of the equation calculated must be

eventually periodic! Nonetheless, a solution with a very large period usual¬
ly represents well many aspects of an aperiodic solution.
Next we present a computer program to give the bifurcation diagram of
the logistic equation. It will give Figs. 1.18, 3.5, 3.6 and many others which
you may choose for yourself. The program ‘LOGIS’, listed in Table 3.2, is
in BBC BASIC, but you will find it easy to translate into another dialect of
BASIC or into FORTRAN if you know these computer languages.
This program calculates x„, where xn+1 = ax„( 1 - x„) and x0 a \ (line
60) for various values of a. First for a = aO, the values of xlt x2,..., x64 are
recorded (line 90) as a succession of short notes, a higher note for a higher
value. This serves both to let the value of n become ‘large’ before results are
plotted on the screen and to demonstrate some properties audibly. You
can ‘hear’ convergence, whether it is monotonic or oscillatory, fast or slow;
you can hear the difference between a two-cycle and a four-cycle; you can
hear the strange ‘melody’ of a Cantor set. Next the 128 values of x65, x66,
..., x192 are plotted (line 90) in a vertical line on the screen. Of course, two
values may be plotted as the same point on the screen if the values are too
close together to be resolved. The scale has been chosen so that x = 0 at
the bottom of the screen and x = 1 near the top. The whole procedure is
next repeated ia times, starting with the value aO + astep of a, and increas¬
ing by astep each time so that the final value is aO + astep x ia. (Note that
96 Difference equations

ia is represented by the integer variable ia% in the program.) The vertical

lines, each of 128 values of x„, are plotted according to the values of a,
starting with aO at the left-hand side of the screen and moving right as a
changes. Thus a is the abscissa and x is the ordinate for the points plotted.
The current value of a is shown (line 50) near the top right-hand corner of
the screen.
Run the program and type the values of aO, astep and ia in response to
the cue (line 10). You are recommended first to try aO = 0, astep = 0.1 and
ia = 40 to plot the whole of the bifurcation diagram of Fig. 1.18. Look at
the crude attempt to plot the Cantor sets. Next you might try aO = 3.4,
astep = 0.02 and ia = 30 in order to examine the period doubling and
strange attractors a little more closely. You could re-plot Fig. 3.5 by taking
aO = 3.82, astep = 0.001 and ia = 30. You might also like to change line 80
of the program to examine other difference equations.

6 Some two-dimensional difference equations

There are many interesting one-dimensional difference equations for func¬
tions F(a, x) other than the logistic map. A few are given in the problems at
the end of this chapter. However, in this and the next section we shall
discuss some two-dimensional difference equations of the form,

*«+i =f(xn,yn), y„+i = g(xn,yn) forn = 0,1,....

Sometimes we write them as

x«+i — F(x„) forn = 0,1,...,

where F: IR2 -> IR2. Higher-dimensional spaces permit a greater variety of

properties than one-dimensional ones do. In particular, the occurrence of
strange attractors in IR2 can be seen geometrically by use of a horseshoe
map. Smale (1967) showed how a class of maps F: S -»• IR2 give attractors
which are the Cartesian product of an interval and a Cantor set, where S is
the unit square [0,1] x [0,1], The essential idea is the qualitative one of
taking F as a map of S which is the result of first contracting S in the
vertical, then stretching it in the horizontal direction, and finally folding it
over back into S. Then F(S) resembles a horseshoe, as shown in Fig. 3.8.
Consideration of f)“=0 F"(S) shows that it converges to the attractor of F
for almost all initial points of S, and that the attractor is like a Cantor set in
a vertical section but like an interval in a transverse section.

Example 3.13: the baker's transformation. This is a simple example of a

map similar to a horseshoe, although it is a discontinuous map. Consider
 3.6 Some two-dimensional difference equations 97

Fig. 3.8 Sketches of (a) the square S with vertices A, B, C and D; (b) contracting,
stretching and folding; (c) the horseshoe F(S), where A' = F(A) etc • (d) F2(SV
(e) F3(S) in S.

the map F of the half-open square [0,1) x [0,1) onto itself where F(x) =
(cr(x),g(a, x,y)).

<t(x) = 2x modulo 1 and 0 ^ <x(x) < 1,

2 ay for 0 ^ x <
g(a, x, y) modulo 1 and 0 ^ g < 1.
May + 1) for 5 ^ x < lj

As in Example 3.8 of the Bernoulli shift, the sawtooth map a maps the
interval [0,1) onto itself, and leads to chaotic orbits <r"(x) independently of
y when x is irrational.
Note that (0,0) is the only fixed point of F, and it is unstable, as for the
Bernoulli shift. There is, in fact, an infinity of unstable p-cycles.

The Jacobian matrix of F is J

o for almost all x, so det J = a.
-e 2«J
Therefore F is area preserving if and only if a = ± 1.
Consider first the case a — 1. Then the geometrical nature of the map is
illustrated in Fig. 3.9(a)-(c), where it is seen that the map is equivalent to a
horizontal stretching and vertical contraction, followed by a vertical cut¬
ting and stacking. This resembles the preparation of dough, so F is often
called the baker’s transformation.
When 0 < a < 1, areas are contracted by F. We show F(S) and F2(S) in
Figs. 3.9(d), (e), where S is the unit square. Iterating, we see plausibly that
F"(S) has 2" slices of thickness (ja)n each, like mille feuilles (or filo) pastry.
98 Difference equations

Fig. 3.9 (a) The unit square S; (b) the baker’s transformation by contracting,
stretching, cutting and stacking; and (c) F(S) for a = 1 .(d) F(S); and (e) F2(S) for
0 < a < 1.

In the limit as n -* oo, this gives an attractor with the topological character
of the interval [0,1) in the x-direction and of a Cantor set in the ^-direction.
□

Example 3.14: the Henon map. Next we shall follow Henon (1976) and
• Henon & Pomeau (1976), who considered the pair of difference equations
with
f(x,y) = 1 + y — ax2, g(x,y) = bx (1)

for given parameters a and b. This is not a horseshoe map, but it is some¬
what similar geometrically, and it allows us to follow the details closely. It
is easy to show that there are two fixed points X = (X, 7) if a > a0 =
—1(1 — b)2 and a # 0, namely

X = lb- 1 ± {(1 -b)2 + 4a}1/2]/2a, Y = bX. (2)

We see that there is a turning point at (a0, -(1 - b)/2a0) in the bifurcation
diagram in the (a, x)-plane for fixed b ^ 1.
A fixed point is stable if the 2x2 Jacobian matrix,

-2 aX 1
J(X) =
b 0

has eigenvalues, qx and q2, such that \ql \ < land|g2| < l.Now

ql,q2= -aX ±(a2X2 + b)112.

 3.6 Some two-dimensional difference equations 99

It follows that the fixed point with X = —(1 — b)/2a + (a - a0)1/2/a is sta¬
ble in the limit as a j a0 when — 1 < b < 1, but the other fixed point X —
-(1 - b)/2a - (a - a0)ll2/a is unstable as a [ a0. It can be further shown
, that the former fixed point is stable for a0 < a < ax and -1 < b < 1,
where a, =4(1 — b)2, there being a flip bifurcation as a increases through
. The other fixed point is always unstable. As a increases beyond a, there
are two-cycles (Q3.34), period doubling with a Feigenbaum sequence,
strange attractors interspersed with windows of stable periodic solutions;
and finally infinity becomes the attractor. This is illustrated in Fig. 3.10. In
fact the structure of the windows of stable p-cycles amid chaos is similar to
that seen in Fig. 3.6(6) for the logistic map, but is complicated by the
coexistence of other attractors and their ‘collision’ as a increases for fixed b.
The Jacobian is det J(x) = — b for all x. Therefore the map is a uniform
contraction for — 1 < b < 1, area-preserving for b = +1, and a uniform
expansion for |6| > 1.
Henon & Pomeau (1976) interpreted the map geometrically as the prod¬
uct of successively a folding, a contraction (when -1 < b < 1), and a re¬
flection in the line y = x, and showed that the map is a canonical form to
which any quadratic map with a constant Jacobian may be reduced by a
linear transformation. If a < a0 or if a is sufficiently large then x„ -♦ 00 as
n-^ co. Otherwise x„ -> 00 or A according to the value of x0, where A is an

ORUL COLLEGE
100 Difference equations

attractor. (Remember that there may be more than one attractor for the
same pair of values of a, b.) Henon & Pomeau showed that, for some values
of a and b, a certain domain of the (a, x)-plane is mapped into itself and
therefore that the points of the attractor are bounded within the domain,
although they may be uncountable. Computations indicate that it is a
strange attractor. A characteristic of strange attractors in dimensions
higher than one is that the average, over an attractor, of the modulus of the
Jacobian is less than one and the average of the modulus of at least one
eigenvalue is greater than one.
The numerical results of Henon & Pomeau are very revealing. They
worked mostly with a = 1.4 and b = 0.3, for which there is a strange
attractor. They took x0 near the unstable fixed point (0.63135448...,
0.18940634...), though this initial point is little better than many others,
and plotted several points x„, as illustrated in Fig. 3.11(a). The other fixed
point (the one which would be unstable for all b) is (—1.13135...,
— 0.339406...), a short distance from the attractor. The point x„ wanders
over the attractor in an apparently random fashion. They next enlarged a
region near the first fixed point and plotted those of a large number of
calculated points x„ which happen to lie in the region, as shown in Fig.
3.11(6). They then repeated this process, enlarging a region within the first
region and plotting those of a very large number of calculated points which
lie in the new region, as shown in Fig. 3.11(c). They repeated this process
once more, calculating even more points, as shown in Fig. 3.11(d). Note
the self-similar pattern which they discovered: the successive enlargements
look like one another, and it appears that the process of enlargement may
be continued ad infinitum (as for Cantor’s middle-thirds set, which is dis¬
cussed in §4.1). The set appears to have a similar structure on all small
length scales. Also the set is dense like a curve along one of the ‘braids’, but
is like a Cantor set in a transverse direction, so that the attractor is locally
the Cartesian product of a line and a Cantor set.
You can prepare these and other plots for yourself by use of the simple
program ‘HENON’ listed in Table 3.3. It is in BBC BASIC, but may be
translated easily. First the parameters are read (line 10). The value of
SCALE is to be chosen to determine the magnification of the plot. A fixed
point is calculated in line 40. The initial values, x0 = 0 and y0 = 0, are
assigned in line 50. The first 300 points x„ are computed but not plotted in
order that the initial values may be ‘forgotten’. Thereafter the points (x„, y„)
are plotted on the video screen, with the x-axis horizontal, the y-axis verti¬
cal, and the fixed point at the middle of the screen. As SCALE is increased
both the number of points computed and the magnification are increased.
 3.6 Some two-dimensional difference equations 101

0.4

~1 0 -0.5 0.0 0.5 1.0 1 5

(«)

0.21

0.20

0.19

y 0.18

0.17

0.16

0.52 0.54 0.56 0.58 0.60 0.62 0.64 0.66 0.68

(b) x

Fig. 3.11 (a) A plot of 4 x 103 successive points x„ of the Henon map for a =
1.4, b = 0.3, with x0 = 0.63135448, y0 = 0.18940634. (b) Enlargement of the
rectangle shown in (a) when 7.5 x 104 points are computed. The cross denotes
a fixed point, (c) Enlargement of the rectangle shown in (b) when 1.5 x 106
points are computed. (d) Enlargement of the rectangle shown in (c) when 107
points are computed.
 102 Difference equations

0.192

0.191

0.190

0.189

0.188

0.187

0.186

0.6225 0.6250 0.6275 0.6300 0.6325 0.6350 0.6375 0.6400

(c) x

0.1900

0.1898

0.1896

0.1894

y 0.1892

0.1890

0.1888

0.1886

0.6295 0.6300 0.6305 0.6310 0.6315 0.6320 0.6325 0.6330 0.6335

id) x

Fig. 3.11 (cont.)

You might start with a = 1.4, b = 0.3 and SCALE = 200. Then, in fol¬
lowing runs, fix these values of a and b but increase SCALE bit by bit to
1000000, say. Again, you might fix b = 0.3 and SCALE = 200, but de¬
crease a bit by bit from 1.4. You might like to experiment for yourself
afterwards. In particular, you might like to change lines 80 and 90 in order
to study a different two-dimensional difference equation. □
 3.7 Iterated maps of the complex plane 103

Table 3.3. List of the BASIC program HEN ON’{You may find it more
convenient to use capital letters throughout.)

10 INPUT “a, b, SCALE”, a, b, SCALE

11 REM Type in values of a, b and the scale of the phase plane to be plotted on
line 120
20 MODEO
30 C = b - 1.
40 X = :5*(C + SQR(C*C + 4.*a))/a: Y = b*X
41 REM (X, Y) is an unstable fixed point to centre plot
50 xn = 0: yn = 0
60 NN% = 300 + SCALE*2.
70 FOR n% = 1 TO NN%
80 xnl = xn
90 xn = yn + 1. -a*xn*xn: yn = b*xnl
91 REM The Henon map
100 IF n% < 301 GOTO 130
110 U = (xn - XfSCALE + 512: V = (yn - Y)*SCALE + 512
111 REM The horizontal and vertical coordinates of the pixel to be plotted to
represent (xn, yn) in the (x, y)-plane
120 PLOT 69, U, V
130 NEXT n%
140 PRINT a, b, xn, yn, NN%
150 STOP

7 Iterated maps of the complex plane

Iterated analytic maps of the complex plane into itself are a special and
important class of two-dimensional difference equations. They have been
studied for over a century. The difference equation

z„+i = F{zn) forn = 0,1,...,

where F: C -► C is analytic, represents the iteration of the map F of the

complex plane into itself. We may write z„ = x„ + iyn and take the real and
imaginary parts of the complex difference equation to get a real pair of
difference equations for x„ and yn. In considering fixed points and their
stability, it is, however, more convenient to use the fact that F is an analytic
function of the complex variable z. Thus Z is defined as a fixed point of F if
it is a zero of F{Z) - Z, and it can be shown, by the argument of §1.8, to be
stable if |F'(Z)| < 1 and to be unstable if |F(Z)| > 1.

Example 3.15: the logistic map. Iterate the map

F(a,z) = az{ 1 - z).

104 Difference equations

where F: C x <C -*• <C. The fixed points 0 and (a - 1 )/a have the same forms
as for the real case of §1.8, and they are stable if \a\ < 1 and 12 - a\ < 1

respectively. The periodic solutions likewise have the same forms as for the
real logistic equation, but here they exist for all complex a # 0 because
we have removed the restriction that all the solutions are real. For some
purposes it is more convenient to transform z and a so that we replace the
logistic difference equation by

Ztl + \ = G(b, Zlt)’

where G(b,z) = z2 — b and b = \a2 — \a (Q1.31). It is found that G may

map certain sets onto themselves for a given value of b, i.e. that G has
invariant sets. These sets may be finite, when composed of fixed points and
periodic solutions, or infinite. The infinite sets may be curves or have a
more complicated structure. □

Example 3.16: the square map. In the special case b = 0, we find zn+l =
G(z„), where G(z) = z2, i.e.

z„+1=z2 for n = 0 1 , ,....

It can be seen that the fixed points are Z = 0,1, oo, and that
_ _ _2 _
_4 _8 _
zn ~ zn-l ~ zn—2 ~ zn-3 ~
_ ... ~
_ zO
,2" •
Therefore z„ -> 0 as n -> oo if |z0| < 1, z„ -> oo if |z0| > 1, and z„ e S1 if
z0 e S1, where S1 is the unit circle with equation |z| = 1. Thus the origin
and infinity are stable fixed points with domains of attraction D(0) =
{z: |z| < 1} and D(oo) = {z: |z| > 1}, and 5D(0) = S1 = 5D(oo). Also 1 is
an unstable fixed point. Further, G(SX) = S1 = G_1(SX), i.e. all the images
and pre-images of all the points of S1 lie in S1. If z0 = exp(27ri0o) then z„ =
exp(27ri0n) = exp(2n7ri0o), because 6n is mapped by the Bernoulli shift (of
Example 3.8). Therefore there are unstable p-cycles of all periods p in S1. In
addition there are aperiodic solutions dense in S1. o

Let R: C -> C be a quadratic map with two attractors, Ax and A2, say,
each a finite set; then dD(A2) = dD(Aj), and we call dD(Aj) the Julia set J
of R. Thus, if z e J then /?"(z) does not tend to a limit point unless z is an
unstable fixed point of R or a pre-image of the fixed point. The notion of a
Julia set can be generalized for all rational maps R. It can be shown that
(a) R(J) = J = K~X(J), i.e. the set of images and pre-images of all points of J
lie in J; (b) J # 0, i.e. J is not empty, and J is closed; (c) unstable p-cycles
are dense in J; and (d) J contains no point which belongs to an attractor.
 3.7 Iterated maps of the complex plane 105

Example 3.17: some Julia sets. The following two examples have been
chosen for their simplicity to illustrate the fundamental ideas, but they are
atypical; Julia sets usually have a complicated structure, (i) Suppose that
R(z) = \(z + \/z). Then R(Z) = Z gives the fixed points Z = ± 1 (or oo).
Now R'(z) = |(1 - 1 /z2). Therefore R'(± 1) = 0. Therefore Ax = {1} and
^2 — {— 1} are attractors. It can be shown that there are no others, be¬

cause D(l) = {z: Re(z) > 0} and D(-1) = {z: Re(z) < 0}. Then J is the
imaginary axis. The four properties of J listed above may be verified for
this example, (ii) Take R(z).== z2. Then the fixed points are Z = 0, 1. Now
R (z) = 2z, so that Z = 0 is stable and Z = 1 unstable. Also infinity is a
stable fixed point. So we may take Aj = {0} and A2 = {oo}. Now D(0) =
I2- Izl < 1} ar,d D(oo) = {z: | z | > 1}, so J = S1, where S1 is the set of points
(z: \ A - 1}, i-e. the unit circle with centre 0. If R(z) = z2 + a for a # 0, then
J has a complicated structure, but is close to the circle S1 when a is small. In
fact part (i) is equivalent to (ii) in the sense that if we define G by w =
(z - l)/(z + 1) and G(w) = (F(z) - 1}/{F(z) + 1}, where F(z) = \(z + 1/z),
then G(w) = w2. □

Consider next the map F: CxC-»C defined by F(a, z) = z2 - a. Then

Fz(a, 0) = 0 and there is an attractor of F(a, z) at z = oo for all a e C. Ac¬
cordingly, the Mandelbrot set of F is defined as

M = {a: a e Qlim^^ Fn{a,0) > oo}.

Thus M is the set of values of the complex parameter a for which the given
initial point z = 0 does not lie in the domain of attraction of infinity. (The
initial point z = 0 is chosen because Fz(a,0) = 0.) Mandelbrot (1982, pp.
188-9) shows computer-drawn diagrams of M for F(a,z) = z2 - a and
initial point z = 0, and for F{a,z) = az( l- z) and z = \. The boundaries
<3M in these two cases appear to be complicated sets with a self-similar
character (see §4.1).
We shall end by mentioning an important theorem. Julia (1918) and
Fatou (1919) showed that the Julia set J of F(a, z) is connected if and only if
Fn(a, 0) > go, i.e. if and only if a belongs to the Mandelbrot set of
F{a,z) for z = 0. So if a belongs to the Mandelbrot set then J, although a
complicated set with many apparently isolated parts, is in fact joined to¬
gether. Look for this in Fig. 3.12. Also see the Mandelbrot set in Fig. 3.13,
and the beautiful coloured pictures widely available in the literature. Note
the finer scales appearing as the picture is enlarged; there is, in fact, an
infinite regression of finer and finer scales. The set itself is shown in black;
an explanation of the significance of the colouring is indicated below.
106 Difference equations

Fig. 3.12 Tableau of the Julia set of F(a,z) = z2 + a, in the rectangle —0.3194417
sS Rez ^ —0.3193553, —0.4452514 ^ Imz < —0.4451650 for a =
— 0.2232 — 0.7296i. The alternation of black and white denotes different rates
of divergence to infinity, but many different rates of divergence are denoted by
black and many other rates by white; so the set itself is not any black or white
region.

Perhaps the best way to enjoy the pictures aesthetically as well as to

understand them is to make your own.
Tables 3.4 and 3.5 list some programs in BBC BASIC to compute Julia
and Mandelbrot sets. They have been written to be simple and effective for
many purposes rather than fast or good for all purposes. The program
‘JULIA’ gives a plot in colour of the rates of divergence rather than the
Julia sets themselves for the iterated map F(a,z) = z2 + a. To run the pro-
Fig. 3.13 Parts of the Mandelbrot set for F(a, z) = z2 + a. The set itself is
coloured black; the meaning of the other colours is indicated in the program
'MANDEL' of Table 3.5. (a) In the rectangle, —2.25 Rea ^ 2.25, —1.25
Ima '£ 1.25. (b) In the rectangle, -1.23 =£ Rea =£ -1.1, 0.25 Ima =£
0.358.

facing p. 106
 3.7 Iterated maps of the complex plane 107

Table 3.4. List of the BASIC program ‘JU LI A ’

10 XS% = 80: YS% = 128: NIT% = 200: M = 4

11 REM These specify the numbers of steps for x, y; the maximum number of
iterations; and the effective size of infinity
20 INPUT “AR,Ar\AR,AI
21 REM These are the real and imaginary parts of a
30 INPUT “XMIN, XMAX, YMIN, YMAX”, XN, XX, YN, YX
31 REM These specify the rectangle of the complex plane for z = x + iv
40 MODEO y
50 GAPX = (XX - XN)/XS%: GAPY = (YX - YN)/YS%
51 REM These evaluate the steps for x and y
60 FOR NY% = 0 TO YS% - 1
70 FOR NX% = 0 TO XS% - 1
80 X = XN + GAPX*NX%: Y = YN + GAPY*NY%: COUN% = 0
81 REM This specifies the coordinates of the pixel (in the complex z-plane) whose
colour must be found next
90 COUN% = COUN% + 1
100 X2 = X*X: Y2 = Y*Y
110 Y = 2.*X*Y + AI: X = X2 - Y2 + AR
111 REM Lines 100, 110 replace z = x + iy by z2 + a
120 IF X2 + Y2 < M AND COUN% < NIT% THEN GOTO 90
121 REMX2 + Y2 = |z|2
130 C% = 7*COUN%/NIT%
140 GCOL 0, 7 — C%
141 REM Lines 120-140 determine crudely the colour of the pixel according to
how many iterations COUN% it has taken for |z|2 to exceed M, a rough estimate
of infinity.
150 PLOT 69, NX%*8, NY%*4
160 NEXT NX%: NEXT NY%
170 STOP

gram you need to answer the cue (line 20) by typing in the real and the
imaginary parts of a to specify the set. The next four numbers specify the
Cartesian coordinates of the four vertices of the rectangle in the complex
z-plane in which the program will represent the Julia set. You might try, for
your first run, to type in 0.32, 0.043 and thereafter -2., 2., -1.5, 1.5. For
your second run try typing in -0.12375, 0.56508 and thereafter -2., 2.,
— 1.5,1.5. Then you might care to experiment for yourself.
Beware of the slowness of the program. The program may take a long
time to run completely, although it has been written to fill only a quarter of
the screen in order to reduce the running time. You may alter the size of the
picture or change the resolution according to whether you prefer a good
big picture or a rapid program. You may be able to improve the colour
108 Difference equations

Table 3.5. List of the BASIC program MANDEL’

10 RS% = 80: IS% = 128: NIT% = 100: M = 4

11 REM These specify the numbers of steps for ar, at; the maximum number of
iterations; and the effective size of infinity
20 INPUT “ARMIN, ARMAX, AIMIN, AIMAX”, ARN, ARX, AIN, AIX
21 REM These specify the rectangle of the complex plane for a = at + ia{
30 MODEO
40 GAPR = (ARX - ARN)/RS%: GAPI = (AIX - AIN)/IS%
41 REM These evaluate the steps for ar and a{
50 FOR NI% = 0 TO IS% - 1
60 AI = AIN + GAPI*NI%
70 FOR NR% = 0 TO RS% - 1
80 AR = ARN + GAPR*NR%
81 REM This specifies the coordinates of the pixel (in the complex a-plane) whose
colour must be found next
90 X = 0: Y = 0: COUN% = 0
100 COUN% = COUN% + 1
110 X2 = X*X: Y2 = Y*Y
120 Y = 2.*X*Y + AI: X = X2 - Y2 + AR
121 REM Lines 110,120 replace z = x + iybyz2 + a
130 IF X2 + Y2 < M AND COUN% < NIT% THEN GOTO 100
131 REMX2 + Y2 = |z|2
140 C% = COUN%/10
150 IF C% = 5 THEN C% = 4
160 IF C% = 6 OR C% = 7 THEN C% = 5
170 IF C% = 8 OR C% = 9 THEN C% = 6
180 IF C% > 9 THEN C% = 7
190 GCOL 0, 7 — C%
191 REM Lines 140-190 determine the pixel colour from the value of COUN%
200 PLOT 69, NR%*8, NI%*4
210 NEXT NR%: NEXT NI%
220 STOP

scheme, which has been chosen more for simplicity of the program than for
aesthetics. You may use the program for different maps by changing lines
100-120.
The program ‘MANDEL’ computes and plots the Mandelbrot set for
the map F{a,z) = z2 + a. To run the program you need to answer the cue
(line 20) by typing in four numbers to specify the Cartesian coordinates of
the four vertices of the rectangle in the complex a-plane in which the pro¬
gram will plot the Mandelbrot set. You might try the input -2.25, 2.25,
-1.25, 1.25 for a start, and for your next run -1.23, —1.1, 0.25, 0.358 to
emulate Fig. 3.13.
 Problems 109

Again, you may wish to improve the program. In particular you may
plot sets for other maps by changing lines 110,120.

Further reading
§3.1 Devaney (1989) gives an excellent account of difference equations.
§3.3 The properties of the sawtooth map may be found from the prop¬
erties, largely due to Borel, of the binary representation of numbers. These
are described by Hardy & Wright (1979, Chap. 9), who, in particular, prove
the statistical properties used in §4.
§3.4 A fuller account of the iteration of one-dimensional maps is the
rigorous treatment by Collet & Eckmann (1980).
§3.7 The literature of iterated maps of the complex plane includes clas¬
sic papers by Cayley (1879), Fatou (1906, 1919) and Julia (1918) as well
as recent analytical and numerical work. Accounts by Mandelbrot (1982)
and Falconer (1990) are recommended in order to learn more than is in
the short account here. Some beautiful coloured pictures of Julia and
Mandelbrot sets are shown by Peitgen & Richter (1986) and Peitgen et al.
(VI990), and some more-advanced computer programs to make your own
pictures of Julia and Mandelbrot sets are given by Peitgen & Saupe (1988).

Problems

Q3.1 Instability of a fixed point. Suppose that F(X) = X, where F: [Rm-» IR™
Then show that X is unstable if and only if 3 e > 0 such that V(5>03n>0
and x0 e !Rm such that |x0 - X| < 3 and |F"(x0) - X| ^ e.
Q3.2 An unstable fixed point which might seem to be asymptotically stable. With
plane polar coordinates (r,0) such that x = rcos 6, y = rsin 9, r ^ 0, and
— n<9^n,a. map F: IR2 ->• IR2 is defined by

F(r,0) = (rsin|0/sin0,|0) if0/O, n,

F(r,0) = (ir,0)

Show that if 9o*0,n then x„ = F"(ro,0o) lies on the circle with centre
x = 0, y = rj2 sin 0O and radius r0/2 sin 0O for all n. Deduce that x„ -► 0 as
n-> oo for all 0O. Show that nonetheless there are points x0 near 0 such that
|x„| may be arbitrarily large (when 0O is close enough to tt and n = 1). De¬
duce that 0 is an unstable fixed point of F in the sense of Liapounov.
110 Difference equations

Q3.3 Frechet derivatives. If V is a normed linear vector space, T: V -> V is a

nonlinear operator, X e V, and there exists a linear operator J:V-»V such
that

T(X + ex) = T(X) + eJ(x) + o(e) as e 0

for all x e V then J is called the Frechet derivative of T at X.

(i) Show that if F: [R -»IR is a differentiable function then the Frechet

derivative of F at X e IR is the multiphcative function J: R -> IR de¬
fined by J(x) = F'(X)x V x e IR. (So the Frechet derivative is a gener¬
alization of the ordinary derivative.)
(ii) Show that if F: IRm -* [Rm is differentiable at X e then J is the
Jacobian matrix of F at X.
(iii) Show that if T(x) = sin{x'(0)} for all xeC1 [0,1] then J at X is given
by J(x) = cos{X'(0)}x'(0).
(iv) Show that if T(x) = {t3x3(f) — x2(t)x'(t)} dt for all xeC'fO, 1]
then J at leC'fOjl] is given by J(x) = 3 Jo t3-X’2(r)oc(t)dr —
lX2(t)x(t)V0.
*(v) Use Taylor’s theorem to show that if T:C‘( — oo, oo) ->C'(— oo, oo) is
defined by

T(x(f)) = — ax(x( — t/a.))

for all x e C‘( — oo, oo), where a is a given positive constant, then J at A"
is defined by

J(y(0) = -aX'(X(-t/x))y(-t/<x) - ay(X{-t/ot))

for all y e C( — oo, oo). '

Q3.4 A linear difference equation whose matrix does not have a complete set of

eigenvectors. Show that the real 2x2 matrix A = ^ I has a double

—a 2a J
eigenvalue q = a. Show further that if a ^ 0 then the eigenvalue belongs to a
unique (except for a multiplicative constant) eigenvector [l,a]T, and there
exist real £l5 £2 such that

*o = Si + s2
for all x0 e (R2.
Deduce that if xn+1 = Ax„ for n = 0,1,..., and x0 is as above, then

x„ = (Si - nS2)a"
■a + S2«n
 Problems 111

Q3.5 Stability of a periodic solution. Show that the p-cycle {X,, X2,... ,X„} of the
continuously differentiable map F:U-*U is stable if \F'(X,)F'(X,)
F'(Xp) |<1. 2'"'
Q3.6 A difference equation with a supercritical pitchfork bifurcation and flip
bifurcations. Suppose that xn+1 = F(a,xn) for n = 0, 1,where F(a,x) =
ax - x3. Find the fixed points and for what values of a each point exists.
Find for what values of a each is linearly stable.
Show that if X, = F(a,X2), X2 = F^X,) then X„ X2 are roots of the
equation

X(X2 -a + 1)(X2 - a - 1)(X4 - aX2 + 1) = 0.

Hence or otherwise find all the two-cycles of F. For what values of a is each
cycle linearly stable?
Sketch the bifurcation diagram in the (a, x)-plane, indicating clearly the
fixed points and the two-cycles.
Q3.7 A difference equation with a subcritical pitchfork bifurcation and flip bifurca¬
tions. Suppose that xn+1 = F(a, xn) for n = 0,1,..., where F(a, x) = ax + x3.
Find the fixed points and for what values of a each point exists. Find for
what values of a each is linearly stable.
Show that if X, = F(a, X2), X2 = F^XJ then X„ X2 are roots of the
equation
X(X2 + a + 1)(X2 + a - 1)(X4 + aX2 + 1) = 0.

Hence or otherwise find all the two-cycles of F. For what values of a is each
cycle linearly stable?
Sketch the bifurcation diagram in the (a, x)-plane, indicating clearly the
fixed points and the two-cycles.
Q3.8 Monotonic maps and their p-cycles. Consider differentiable F: IR -> IR. (a)
Show that if F'(x) > 0 for all x then F has no p-cycle for p ^ 2. (b) Show that
if F'(x) < 0 for all x then F has a unique fixed point, and F has no p-cycle for
P S* 3.
[Hint for the last part of (b): consider the sign of the derivative of the
gth-generation map. Devaney (1989, p. 59).]
Q3.9 Cluster points of the sawtooth map. Show that if o(Q) is defined as the frac¬
tional part of 2d and 90 = 2~JU+l\ then 0O is irrational and the cluster
points of the sequence {<r"(0o)} are 2~r for r = 1,2,....
Q3.10 A pitchfork bifurcation. Given that F(a,x) = x[a/{x2 + (a — x2)e~4na}]1/2,
as in Example 1.3, show that if a < 0 then F has a stable fixed point X = 0
and if a > 0 then F has an unstable fixed point X = 0 and two stable fixed
points X = ±y/a. Find F"(a,x). What are the domains of attraction of the
stable points?
Q3.11 Shuffling cards. A pack, i.e. a deck, of 2s playing cards is given a riffle shuffle
by cutting the pack into two halves and then interleaving the halves, so that
112 Difference equations

the bottom card returns to the bottom. Denoting the cards of the pack by
the numbers 0, 1,..., 2s - 1, where 0 represents the bottom card, 1 the next
card above,..., and 2s — 1 the top card before the shuffle; and denoting by
F(r) the position, above the bottom after the shuffle, of the rth card before
the shuffle; show, by a plausible interpretation of the shuffle, that

F(r) = 2r modulo 2s — 1, 0 ^ F(r) < 2s — 1 for 0 < r < 2s — 1,

and F(2s — 1) = 2s — 1.
Deduce that if a pack of 52 cards is riffle shuffled eight times then the
pack ends up in the same order as it started.
Q3.12 Reflection of a light ray by the interior of a cylinder. Consider the path of a
light ray which is reflected repeatedly by a circular cylinder whose interior
surface is silvered. Suppose that the path lies in a plane perpendicular to the
axis of the cylinder, and take polar coordinates such that the ray’s nth
reflection is at a point with polar angle 9 = 6n, for n = 0,1,....
Show that if the angle of reflection is a at first then

0„+i = 9„ + n - 2a.

Deduce that the path of the ray is closed if and only if a/n is rational.
Q3.13 Twist maps. The map F: IR2 -* R2 is defined by F(x,y) = (f(x,y),g(x,y)),

f(x, y) = x cos ip — y sin ip, g(x, y) = x sin \p + y cos ip,

where ip is a differentiable function of (x2 + y2)1/2.

Show that the map is area preserving.
Express the map as a transformation of the polar coordinates r and 9,
where x = rcos0, y = rsin0, and show that a circle with centre at O is
invariant under the map.
Describe geometrically the iterations of the map when ip is a constant,
according to the various values of that constant.
[Moser (1973) showed the importance of perturbations of such maps.]
Q3.14 A logarithmic map. Consider the difference equation xn+1 = af{xn) for n =
0, 1,..., where/(x) = lnx and a > 0. Show, graphically or otherwise, that
there are 2, 1 or 0 fixed points according as a ^ e respectively. Suppose that
a > e and let the fixed points be Xt and X2 > Xx; then find which is stable
and find the domain of attraction of each stable fixed point.
Q3.15 Stability of fixed points of difference equations which represent equilibrium
points of an ordinary differential equation. First just read this paragraph.
There are several methods to find the equilibrium points of an ordinary
differential equation. The Newton-Raphson method to find an equilibrium
point of the equation
 Problems 113

gives the difference equation

x„+i = N(x„) for n = 0,1,...,

where N(x) = x -f(x)/f'(x). Alternatively, although the equilibrium points

are just the zeros of/, we may take forward finite differences with a chosen
small positive time step h and approximate dx/df in the differential equation
^ (*n+i — x„)/h to give

xn+I = F(x„),

where F(x) = x + hf(x). This may give a stable equilibrium point by taking
lim«-oo *n-or an iterative approximation thereto.
Now find N and F for the equation

b> these two methods. Show that the steady solutions given by X = + ^Ja
for a ^ 0 are fixed points of the appropriate two difference equationsrand
find whether they are stable fixed points. What are the domains of attrac¬
tion of the stable fixed points for each of the two difference equations?
Q3.16 A qualitative difference between the set of solutions of an ordinary differential
equation and of its finite-difference approximation. Show that if

then x(t) — x(0)/{l + 2x2(0)r}1/2 for t ^ 0. Deduce that X = 0 is the only

equilibrium point and that it is globally asymptotically stable.
Show that if

xn+i=x„-hx* forn = 0,1,...,

where h > 0, then X = 0 is the only fixed point and it is stable. What is its
domain of attraction? Show further that {-(2/h)l/2,(2/h)112} is a two-cycle
and it is unstable.
[Stuart (1990).]
Q3.17 Some three-cycles. Show that if {X, Y,X} is a three-cycle of a differentiable
map F: R - R, i.e. Y = F(X), Z = F(Y), X = F(Z), and if G = F3 is the
third-generation map of F, then

G'(X) = G'(Y) = G'(Z).

Defining F by F(a,x) = ax(l - x) for x e K, show that if A' belongs to a

three-cycle of F then X is a root of the sextic equation,

a6*6 - (3a + 1 )a5X5 + (3a + 1 )(a + 1 )a*X4 - (a3 + 5a2 + 3a + l)a3A:3

+ (2a + l)(a2 + a + 1 )a2X2 - (a + l)(a2 + a + 1 )aX + (a2 + a + 1) = 0.

114 Difference equations

*Hence or otherwise show that two real three-cycles of the logistic map exist
if a > + 1 or a < —(-JS — 1).
[Hint: it may help you to let the three-cycles be {X1,Yl,Z1} and
{X2, Y2,Z2}, to define tj = XjYjZj, u} = £ YjZj and Vj = £ Ar} for j = 1, 2,
and then to find the quadratic equation whose roots are vlt v2. Myrberg
(1958, p. 13).]
Q3.18 The mapping of a subinterval onto itself by the logistic map. Consider the
logistic map F(a,x) = ax( 1 — x) and its second-generation map G(a,x) =
F(a, F(a, x)) over the interval 0 < x ^ 1.
Show that G(a,j) = fea2(4 ~ a) and that if a > 2 then G(a,x) ^ G{a,\)
for all x sufficiently close to i.e. that G has a local minimum at x = j.
Deduce that if 4 is the root of ^a2(4 — a) = 1/a such that 3 < a < 4
then G(y4,i^/42(4 — A)) = (A — \)IA. Hence or otherwise show that G(A,x)
maps the subinterval [1/4, (A — 1 )/4] onto itself. Deduce that F(A,x) maps
the subinterval [1/4,54] onto itself, [0,1/4] onto [0,(4 — l)/4] and
[i4,1] onto [0,1/4].
Discuss briefly the attractors of F for a = 4.
[In fact 4 * 3.6790. It may be helpful to sketch graphs of G and F, and
some important points and their maps, for a = 4. Also numerical experi¬
ments may suggest the nature of the attractors.]
Q3.19 The explicit solution of the logistic difference equation in a special case. Show
that the logistic map

F(x) = — 2x(l — x)

maps the closed interval [ — 2,1] onto itself.

Show that if x„ = \ + cos 27t0n, dn+1 = 2dn modulo 1 and 0 ^ d„ < 1 then
x„+1 = F(x„). Deduce that if xn+1 = F(x„), — \ < x0 < § and 0O is a root of
cos2tt0o = x0 — ^ then x„ = \ + cos(2"+1rt0o). Discuss briefly the periodic
and aperiodic solutions {x„}.
Q3.20 Some statistical properties of the logistic difference equation. The mean x of a
sequence {x„} is defined by x = lim^*, {(xx + x2 + ••• + xn)/N} and the
variance Var(x) by Var(x) = x2 - (x)2. Deduce that if xn+1 = 4x„(l - x„)
for n = 0,1,... and {x„} is an aperiodic sequence then x = { and Var(x) =
Q3.21 Bifurcations of the logistic map for a < 0. Take F(a,x) = ax(l — x) and use a
computer to investigate the bifurcations for a < 0 by iterating F"(a, j). Find,
in particular, the period doubling and chaos as a decreases from — 1.
Q3.22 The tent map. Define the map F by

F(a,x) = a(^ - |x - £|)

for a > 0. Sketch the curve y = F(a, x) in the (x, y)-plane for a ‘typical’ posi¬
tive value of a. Find the fixed points of F and their stability for all a > 0.
Sketch a bifurcation diagram in the (a, x)-plane. Find a two-cycle for a = 2,
and show that {f, f, f} is a three-cycle.
 Problems 115

Show that

F"(a, c + e) — Fn(a, c) = ± a"e as e -* 0

for fixed c ^j,a and n. Discuss briefly the relevance of this result to sensi¬
tive dependence on initial conditions.
Q3.23 Second-generation tent map. Show that if G(a,x) = F2la x) F(a x) =
a(2 ~ la ~ *1), and 1 < a < 2 then

forO x < l/2a

a(l — ax) for l/2a < x ^ {

G(a, x) =
a(l — a + ax) for 5 < x < 1 - l/2a

a2(l — x) for 1 - l/2a x ^ 1

Sketch the graph y — G(a, x) in the (x, y)-plane. Hence or otherwise find how
many solutions of period two the difference equation xB+i = F(a, x„) has for
each value of a e (1,2], and state which are stable.
Q3.24 Topological conjugacy of the tent and logistic maps. Two maps F and G are
said to be topologically conjugate if there exists a smooth invertible map H
such that F = H~lGH.
Show that the logistic map, defined by F(x) = 4x(l - x) for 0 < x < 1,
and the tent map, by G(x) = 2(j — jj — x|), are topologically conjugate, on
taking H(x) = 2n~l arcsin^x. Deduce that if x„+1 = G(xJ for n = 0, 1,...
and y„ = sin2(^7rxB) then yB+1 = f(yB). What is the solution x„ in this case?
Q3.25 A class of one-dimensional difference equations, (a) Suppose that 0B+1 = 28„
modulo 1 and 0 8n < 1 for n = 0,1,..., and define

xB = sn2(2/C(m)0„|m) forO^mcl,

where sn is the Jacobian elliptic function with parameter m, and K is the

complete elliptic integral of the first kind. Then show that

X-.+1 = 4/(m,x„), where/(m,x) = x(l - x)(l - mx)/(l - mx2)2.

[You need have no knowledge of elliptic functions to solve this problem

/
correctly; the hints will suffice. It may help to think of sn as a generalization
of sin and K(m) as a generalization of with the extra variable m; indeed,
sn(z|0) = sinz and K(0) = \n. Hint: you are given that sn(z|m) is an odd
function of z with period 4K(m) and

sn2(2z|m) = 4sn2(z|m){l - sn2(z|m)}{l - msn2(z|m)}/{l - wisn4(z|w)}2.]

Hence show that f(m, x) has only one maximum for 0 < x < 1, where its
value is and that f(m, 0) = 0,/(m, 1) = 0, fx(m, 0) = 1.
[Hint: you are given further that sn(z|m) is a monotonically increasing
continuous function of z for 0 < z < K(m), that sn(0|m) = 0 and

E
116 Difference equations

sn{K(m)\m) = 1. In fact, f(m,x) is a convex function of x over 0 < x < 1 if

0 < m < 1.]
(b) Consider the difference equations of the form xB+1 = af(m,xn),
where 0 < x0 < 1, 0 < a ^ 4 and 0 < m < 1. Ascertain the number and
stability of the fixed points by graphical means. Describe qualitatively the
bifurcations of the system as a increases for fixed m.
Use a computer to verify your description, finding some of the attractors
for negative as well as positive values of m.
[Bristol (1987).]
Q3.26 The explicit solution of a one-dimensional difference equation. Consider the
map F defined by

F(x) = 2x/(l + x2)

for —oo < x < oo. Show that F maps (—oo, oo)onto [—1,1]. Find the fixed
points of F and whether they are stable.
Show that if xn+1 = F(x„) for n = 0,1,... and -oo < x0 < oo then there
exists a real number zx such that Xj = tanhzt provided that x0 # ±1. De¬
duce that x„ = tanh(2"“1z1). Find the attractors of F and their domains of
attraction.
Q3.27 Another one-dimensional difference equation. Define F by

F(a, x) = an~l sin nx

for a ^ 0. Show that there are fixed points X = 0 for all a and that X =
± U(a)/n, where U is the zero of u — a sin u such that < U < n. Sketch a
graph showing clearly the values of a for which the second and third fixed
points exist. What other fixed points are there?
For what values of a is the origin a stable fixed point? Show that
the second and third fixed points are unstable when a > ax, where ax =
Ux /sin Ux and Ux is the zero of tan u + u such that \n < Ux < n.
Describe qualitatively the bifurcations as a increases from zero to n. You
should use graphs of F and some higher-generation maps, but may also use
heuristic arguments and quote any relevant results.
Describe briefly the various sets of cluster points when a = n.
Q3.28 Period doubling of an exponential map. Define F by F(a, x) = aex for — oo <
x < oo, a < 0. Show that F has a unique fixed point X(a), which is stable if
— e ^ a ^ 0 and unstable if a < —e. Show that X(a) ~ a as a 10, Y(a) =
— 1 + (a + e)/2e 4- 0{(a + e)2} as a-*— e, and X(a)-ln( — a) as
a -> —oo.
Defining the second-generation map G of F by G(a, x) = F(a, F(a, x)),
show that Gx = FG, Gxx = F(1 + F)G, Gxxx = F(1 + 3F + F2)G, where the
subscript denotes partial differentiation with respect to x. Deduce that
Gx(a,X) = X2, Gxx(a, X) = X2(l + X), Gxxx(a,X) = *2(1 + 3Y + X2). By
graphical consideration of the intersections of the line y = x and the curve
 Problems 117

y — G(a,x) in the (x,y)-plane or otherwise, show that a stable two-cycle

bifurcates from X as a decreases through — e.
Show that if x = G(a, x) and x # X then

(X + 1)(X - 1) + \X2(X + 1 )(x - X)

+ £*2(1 -1- 3X + X2)(x — X)2 -|-= 0

as x -♦ X. Hence or otherwise deduce that the two-cycle is given by

XlfX2 = — 1 ± {— 6(a + e)/e}1/2 + 0(a + e) as a f — e.

Sketch the bifurcation diagram in the (a,x)-plane, indicating the fixed

points and the two-cycle.
Q3.29 Arnold tongues for the sine map. Define the fractional part as the function
/: R -» [0,1) such that

/(x) = x modulo 1 and 0 < 1.

(a) The rotation map F: [0,1) -> [0,1) is defined by

F(x) =f(x + a)

for a e [0,1). Find the fixed points X, if any,' of F, showing for what
values of a each exists.
Prove that F has two-cycles {X, T} if and only if a = the two-
cycles then being given by Y = X -I- \ for all X e [0,^).
(b) The sine map G : [0,1) -+ [0,1) is defined by

G(x) = /(x -I- a -I- (27t)_1 £>sin27rx)

for fixed parameters a e [0,1) and b e [0, oo). Sketch the curve y =
a + (27t)-1 6 sin 2nx in the (x,y)-plane for 0 < x ^ 1, 0 < a < j and 2na <
b <2n(\ — a). Hence or otherwise show that then G has two fixed points,
and find them in explicit terms of elementary functions. Show further that
G has no fixed point if b < 2na and one if b = 2na.
Consider next the two-cycles {Z, T} of G, taking 0 X < Y < 1 with¬
out loss of generality. Show that

X + a + (2rt)_1 6 sin 2nX = Y, (1)

Y + a + (2n)~l bsin2nY = X + 1 (2)

for sufficiently small b. It is given that the series

a = \ + bax + b2a2 + ...,

X(a) = + bXx + b2X2 + ..., Y(a) = Y0 + bYt + b2Y2 + ...

converge for sufficiently small b. Equating coefficients of b in equations (1)

and (2), deduce that
118 Difference equations

a, =0, y, = A, + (2;t) 1 sin 2nX0.

Equating coefficients of b2, deduce that if

\ - b2/$n + 0{b3) < a < { 4- b2/Sn + 0(b3) as b J, 0.

then there exists a two-cycle for all X0 e [0, \) for sufficiently small b > 0.
Summarize your results in a sketch of the regions of the (a,h)-plane
where there exist fixed points of G and where there exist two-cycles.
[There are similar regions of the (a, h)-plane where p-cycles of G exist for
p = 2, 3, ..., and small b. These regions join up and chaos ensues as b
increases. The regions are called Arnol'd tongues. Cf. Arnol’d (1961, §12).]
Q3.30 An exponential map. Find the fixed points of F(a,x) = xeatl~x) and their
stability for all a > 0. Discover the stable periodic orbits for a = 2.3, 2.6 and
2.9 by numerical experiments.
Q3.31 A cubic one-dimensional map. Given that F(a,x) = ax(l — x2), show that F
maps the interval [0,1] into itself if 0 < a < 3^/3/2 = 2.598. Show that the
fixed points are X = 0 for all a and X = ±{(a — l)/zz}1/2 for all a < 0 and
a > 1. Show that X = 0 is stable for — 1 < a < 1 and X = {(a — 1 )/a}1/2
for 1 < a < 2. Sketch the bifurcation diagram in the (a, x)-plane.
Use a computer to find period doubling at values ar of a for r = 1,2,...,
and verify that

(al~l-ar)/(ar-ar+l)^> 8 asr—>°°,

where 8 is Feigenbaum’s universal constant, and that 0* ~ 2.3202.

Q3.32 Superstable periodic solutions of a one-dimensional map. A fixed point X of
the smooth map /: R -»IR is called superstable if it is, according to linearized
theory, the most stable a fixed point can be, i.e. if it is quadratically stable.
Deduce that f'(X) = 0.
Similarly, a p-cycle {A-,,...,Xp} is superstable if it corresponds to a
superstable fixed point of fp. Show then that f'(Xj) = 0 for j equal to one
value of 1, 2,..., p. Deduce that there is a superstable p-cycle of / if p is the
least positive integer such that fp(xm) = xm, where xm is a simple maximum
of/.
Use a computer to calculate the first few superstable 2r-cycles of
F(a,x) = ax(l — x), basing your method on a numerical scheme (the method
of false position, for example) to solve F2r(a, 2) = 2 for a iteratively. Defining
Ar as the value of a at which the the 2r-cycle is superstable, calculate the first
few values of (Ar - /4r_,)/(Ar+1 - Ar).
Similarly calculate the value of a at which the three-cycle is superstable
by solving F3(a, 5) =
[It may help, and is instructive, to plot the graph of y = Fp(a,%) - ^ in
the (a, y)-plane before solving Fp(a, 2) = 2 iteratively, in order to find the
 Problems 119

value of a at which the p-cycle is superstable. Note that it is easier to calcu¬

late Ar accurately than the value ar at which there is period doubling, be¬
cause when a is close to a, the 2'-cycle is weakly stable and so solutions
converge to it slowly.]
Q3.33 Another form of the Henon map. Show that if xB+1 = l + yn - ax* an(j
yn+i = bxn for n = 0, 1,.... then un+1 = a + bvn - u2 and vn+l = u„, where
Un = and v„ = ayjb for b ^ 0.
Discuss the solutions of the difference equation in the (u,t;)-plane for
various values of a when 6 = 0.
Q3.34 Two-cycles of the Henon map. Consider the map F: !R2 -»R2 defined by
F(x) = (f(x,y),g(x,y)),

f(x,y)= 1 +y-ax2, g(x,y) = bx

for all x = (x,y), where - 1 < b < 1. Show that if X = (X, Y) is a fixed point
of the second-generation map F2 then

{aX2 + (1 -b)X - 1} {a2X2 - a(l - b)X + (1 - b)2 - a} = 0.

Deduce that the two-cycle {X,,X2} ofF is given by

Xi.Xa = ({1 - b ± 2(a - aj)1/2}/2a,6{1 - b + 2(a - at)l,2}/2a)

respectively for all a > a, = |(1 - b)2. Show further that the two-cycle is
stable if at < a < a2 = (1 — b)2 -f j(l + b)2 and unstable ifa > a2.
Q3.35 The location of an invariant set of the Henon map. Show that if a = 1.4,
6 = 0.3 then the quadrilateral with vertices (-1.33,0.42), (1.32,0.133),
(1.245, -0.14),(-1.06, -0.5) is mapped into itself by equations (6.1).
[Henon & Pomeau (1976).]
Q3.36 The Lozi map. A piecewise linear map F: R2 -*■ R2 is defined by F(x) =
(/(x), g(x)), where x = (x, y),

f(x,y) = 1 + y — a|x|, 0(x,y) = 6x

for given real parameters a, 6.

Show that the fixed points of the map F are X+ for all a > -(1 - 6) and
X- for all a > 1 - 6, where X+ = (1/(1 + a - 6), 6/(1 + a - 6)) and X. =
(- l/(a + 6 - 1), — 6/(a + 6 - 1)).
Henceforth assume that -1 < 6 < 1. Deduce that X_ is always un¬
stable. Given that X+ is stable for -(1 - 6) < a s$ a,(6), find the function
a, of 6.
Show that if {X,, X2} is a two-cycle of F, a2 * (1 — 6)2, X, = (X,, V]),
and X2 = (X2, Y2), then X, X2 <0. Hence or otherwise find each two-cycle
of F and the condition satisfied by a and 6 for it to exist.
Sketch the fixed points and two-cycles of F in the (a, x)-plane for fixed 6,
indicating clearly what is a fixed point and what a two-cycle.
120 Difference equations

[It may help to regard the map as an analogue of the Henon map (Lozi
1978).]
Q3.37 The Lozi map again. A piecewise linear map F: R2 —»R2 is defined by
F(x) = (/(x),gf(x)), where x = (x,y),

f(x,y) = 1 + y - fl|x|, g(x,y) = bx

for given parameters a, b > 0.

Show that F maps the right half-plane to the upper half-plane.
Defining the points U = (0, b(q — 1)/<?(1 + a — b)), Z =
((1 - <?)/( 1 + a - b),0), and q = ~^{a + (a2 + 4b)1/2}, where a > |1 - b\,
show that F(U) = Z and that the fixed point X+ lies in the line UZ. Find the
coordinates of V = F(Z)and W = F(V).
♦Deduce that F(A) c A, where A is the triangle with vertices Z, U and V,
ifO < b < 1 and b + 1 < a < 2 — $b.
[Misiurewicz (1980) proved that there is a strange attractor in A. Com¬
pute it and show it graphically for, say, a = 1.7 and b = 0.5 (Lozi 1978).]
Q3.38 A quadratic area-preserving map. Define the rotation map R: R2 -» R2, the
shear map S: R2 -» R2 and T: R2 -*• R2 by

R(x) = (x cos a — y sin a, x sin a -I- y cos a), S(x) = (x, y — x2), T = RS,

where x = (x, y) and a is the angle of the rotation.

Show that

T(x) = (xcosa — (y — x2)sina,xsina -I- (y — x2)cosa)

and T is area preserving.

Show that if a -> 2n — a then T( — x,y) T(x,y). Deduce that it is suffi¬
cient to consider only 0 < a < n.
Show that if a = 0 then T"(x) = (x,y — nx2) and if a = n then T2 is the
identity map.
By expressing T_1 = S^R-1 or otherwise, show that the inverse map
of T is given by T“‘(x) = (xcosa + ysina, —xsina -I- ycosa -I- (xcosa +
ysin a)2).
Show that if 0 < a < 0 then the fixed points of T are 0 and (2 tan(^a),
2tan2(^a)). Find the stability characteristics of each point, and state for
what values of a each is stable.
Show that there is no two-cycle of T.
Using a computer, experiment with iterations of the map T.
[Henon (1969).]
Q3.39 Area-preserving maps of the plane and period doubling. Show that the map
F: R2 -» R2, where F(x) = (F(x,y),G(x,y)),
 Problems 121

F(*.F) = -y + /(x), G(x,y) = X -/(F(x,y)),

is area preserving for all differentiable functions /.

Consider iterated maps of this form for /(x) = ax - (1 - fl)x2, finding all
the fixed points and conditions for their linear stability. Show, in particular,
that the origin is stable for — 1 < a ^ 1. Suggest a plausible reason for
anticipating a flip bifurcation at a = - 1. Sketch a bifurcation diagram in
the (a, x)-plane.
Verify by use of a computer that there is period doubling at a = a,
for r = 1, 2, where and (ar_, - ar)/(ar - ar+l) -* 8 as r-> oo,
a * « — 1.266, and <5 % 8.271.
[In fact this new 5 is a universal constant for period doubling of measure¬
preserving maps (Bountis 1981), analogous to Feigenbaum’s constant <5.]
Q3.40 The standard map. Consider the difference equations pn+l = F(k,pn,xn),
xn+i = G(fc,p„,xJ for n = 0, 1, ..., corresponding to the map of the unit
square of the (p, x)-plane into itself, where

k
F(k,p,x) = p — —sin2nxmodulo 1 and 0 < F < 1,
2n

G(k, p,x) = x + F(k, p, x) modulo 1 and 0 < G < 1.

Show that the map is area preserving for all k.

Show that when k = 0 each orbit lies in the line p = p0 and the solution
is periodic if p0 is rational and aperiodic if p0 is irrational.
Use a computer to examine how these orbits change as k increases from
zero for various values of x0 and p0.
Q3.41 The cat map. Consider the map F: T2 -» T2 of the unit square modulo 1, i.e.
of the torus in which 2nx and 2ny may represent the angles around two
circles which thread the torus in perpendicular directions, defined by
F(x) = (f(x,y),g(x,y)), where

f(x,y) = x + y, 0(x, y) = x + 2y modulo 1

andO ^/(x^^x.y) < 1.

Show that the eigenvalues of the Jacobian matrix are quq2 = i(3 ± y/5)
for all x, y, and that the map of T2 is area preserving.
Prove that (0,0) is the fixed point of F, and that the two-cycles of F are
{(5> 5)>(f» l)}> {(§> s)’ (s> f)}-
Show that if X, Y are rational then X = (X, F) is a fixed point of Fp for
some positive integer p. Show that if X is a fixed point of Fp then X, Y are
rational.
Deduce that each fixed point of Fp is unstable for p = 1,2.
[This map is called the cat map because Arnol’d & Avez (1968, p. 6) first
posed it with a sketch of how a silhouette of a cat’s head is distorted, cut up,
and transposed by F and F2 (see Fig. 3.14).]
122 Difference equations

Fig. 3.14 Sketches of the cat map. (a) The cat’s head in the unit square. (b) The
map of the head before moduli one are taken, (c) The final map of the head.

Q3.42 A map with an invariant curve. Defining a map F: IR2 -* IR2 by

F(x, y) = (ax,ab(y -f(x)),

where /: IR -»IR is differentiable, and a, b > 0, show that F maps the line
x = x0 into the line x = ax0. What are the fixed points of F, and for what
values of a, b does each exist?
Show that if g: IR -> IR is a function of Weierstrass type defined by

g(x) = X a~bkf(akx) fora > 1

k-0

then the curve y = g(x) is mapped into itself by F. Find the Jacobian of F at
a typical point of the curve, and deduce that the curve is a repeller.
Taking f(x) = sin27tx and ae {2,3,...}, modify F so that it maps a
cylinder into itself and find an invariant closed curve of the modified map.
[Falconer (1990, p. 196).]
Q3.43 Canonical form of a quadratic map of the complex plane. Show that if
zB+1 = az2 + 2bz„ + c, w„ = az„ + b, d = b2 — ac — b and a ^ 0 then wn+,
= w„2 - d.
Q3.44 Stability of a fixed point of an analytic map of the complex plane. Suppose
that the map F: € -»C is analytic at Z, where Z is a fixed point of F,
i.e. Z = F(Z). Then show that Z is stable if |F'(Z)| < 1 and unstable if
\F'(Z)\ > 1.
 Problems 123

Find the fixed points of F(z) — z2 + a for all complex a, and for what
values of a each is stable.
Show that if Z is a fixed point of F2 then

(Z2 - Z + a)(Z2 + Z + a+ 1) = 0.

Hence find the two-cycle of F. For what values of a is it stable?

Sketch the region of the complex a-plane in which either a fixed point or
the two-cycle is stable. What does this remind you of?
Q3.45 Analytic map of the complex plane. Suppose that F: C -» C is analytic and
generates a map G: IR2 -♦ IR2 by taking real and imaginary parts in the usual
way. Deduce that the Jacobian determinant of G is |dF/dz|2. What are the
eigenvalues of the Jacobian matrix?
Q3.46 The cubic map. Defining F: C - C by F(z) = z3 for all z, find each fixed
point of F, and whether it is stable. Find the domain of attraction of each
stable point.
Find each two-cycle of F and whether it is stable.
Given the difference equation z„+i = F(zJ for n = 0,1,..., express z„ =
rnexP(>0J for rn ^ 0 and real 9„, and thence determine the difference equa¬
tions satisfied by r„, 9n. Describe all the other cycles briefly.
Find at least one closed curve mapped by F onto itself.
Q3.47 The real part of a Mandelbrot set. Consider what real values of a lie in the
Mandelbrot set M of F(a, z) = z2 + a. Show first that (|, oo) c M, then that
[ — 4-4) c M, and [ — f, —|) c: M. What happens ifa < — f?
Q3.48 Some simple maps of the plane. Consider the following four maps F: IR2 -* IR2
corresponding to maps H: C -+ C, where F = (F, G), z = x + iy, H(z) =
F(x, y) + iG(x, y). Prove in each case that F and H are equivalent, and that
F is area preserving.
Deduce that all compositions of the four maps F are area preserving.
(a) Translations: F(x) = (x + a,y + b) for real a, b\ H(z) = z + a + ib.
(b) Reflection in the x-axis: F(x) = (x, — y); H(z) = z.
(c) Rotation about the origin: F(x) = (xcosa — y sin a, x sin a + ycosu)
for real a; //(z) = e“z.
(d) Shear in the x-direction: F(x) = (x -(- /(y),y) for real differentiable
function /; H(z) = z + /(Im(z)).
Q3.49 Fixed points of a map and their stability. Suppose that

xB+, = ka -t- mxr„ ifxB>0, forn = 0,1,...,

where m, r > 0. Defining p > 0, u„ > 0, b by m = p'~\ un = px„, b = akp,

deduce that

un+l = F(b,r,u„),

where F(b, r, u) = b + ur for all u > 0. Taking r = 2 in turn, find the fixed
124 Difference equations

points of F for all b and determining for what values of b each exists and
each is stable.
Show, by graphical means or otherwise, that the qualitative results are
typical for r < l,r > 1 respectively.
Q3.50 The initial-value problem. Considering the eigenvalue r, belonging to the
eigenvector vy of the adjoint of the Jacobian matrix J such that

JTvy = ryv;

for) = 1,2,..., m, show that in general the eigenvalues may be ordered so that
r, = qt and then vju* = 0 if k*j.
Deduce that, in equation (1.6), = v)x^/v)u;.

Q3.51 A closure problem. Considering aperiodic orbits of the logistic equation

xn+] = axn(l~x„),

and assuming that there exists the rth momenta limw__ forr= 1,
2, ..., show that x2= (a - 1 )x/a. In particular, deduce that if a- 4 and x. = y
then Var(jc) = j.
Show also that
= o2j^( 1 - 2x„ + xt), x„+2 = a2x„ {1 - (a+1) ar„ + 2cvFn -axsn},
and deduce that each of these two equations gives
x = a\2jF - x*)Ha - 1 )2{a + 1)
if a * ± 1. What prospect is there of evaluating J as a function of a by using
higher-moment equations as well as those above?
Q3.52 Conjugate maps. Suppose that x„ = P{6„), xn+l = F{x„) and 0„+l = 26„ for n = 0,
1,..where 0 < P(6) < 1 and P has period 1. Given P, verify F in each of the
following cases.
(a) P(0)= 1-11 -2d\,F(x)= 1-11-2x1.
(b) P{ 6) = 4 1 - 6), F(x) = 4( 1 - jc)'1/2 {1 - (1 - x)'1/2}.
Q3.53 Area-preserving maps. You are given a map of the plane defined by lx, y) —»
{fix, y), g(x, y)) for smooth /, g: IR2 —> IR. Show that the map is area¬
preserving if and only if there exist smooth F, G: R —> IR, y/: IR2 —» IR, such
that

f(x, y) = F(y/{x, y)), g(x, y) = G(y/(x,y)).

Q3.54 Some two-cycles. Find all the two-cycles of the baker’s transformation of
Example 3.13 for 0< a < 1.
Q3.55 A bound on a Julia set. Show that if/fz) = z2 + a and Izl > max(2, lal) then l/(z)l
> Izl. Deduce that/”(z) ->~asn->»>, and thence that the Julia set of a is a
subset of the disc Izl < max(2, lal).
 Some special topics

One gets a similar impression when making a drawing of a rising cumulus

from a fixed point; the details change before the sketch can be completed. We
realize thus that: big whirls have little whirls that feed on their velocity, and
little whirls have lesser whirls and so on to viscosity....
L.F. Richardson (Weather Prediction by Numerical Process, p. 66)

1 Cantor sets
We have met some strange attractors and asserted that they were ‘Cantor
sets’. We shall now define Cantor sets more carefully. They are called Can¬
tor sets because they share some important properties of Cantor's middle-
thirds set, K say (Cantor 1883, p. 590). To define K we first define a se¬
quence of subsets K„ of the unit interval, as shown in Fig. 4.1. Thus we
define K0 = [0,1], K, = [0,±] u [f, 1], K2 = [0,£| u [|,i] u [f,|] u [§, 1]
and so forth. Note that K„ is the union of 2" closed subintervals of [0,1]
of the form [r/3",(r +1)/3"], for appropriate integers r, and so the sub¬
intervals have length 3_". The middle third of each subinterval of K„ is
removed to give Kn+1, so that Kn+1 <= K„ c [0,1]. Then K is the set of all
points common to K0,K1,K2,...,i.e.K = n“=0Kn.
Even this set, in spite of its being so simply constructed, appears as a
repeller of a one-dimensional map (namely the tent map shown in Q4.4).
Another definition of K is the set of all numbers expressible in the terna¬
ry form O.Xjx2x3... = xJ3 + x2/32 + x3/33 + ..., where the digit x„ = 0
or 2 for n = 1, 2, .... (Note that the same real number may have two
ternary forms, e.g., 5 may be expressed as either 0.1 or .
0 0222... and so
belongs to K.) It can be seen that each point of Kx has the ternary form
. x2x3... if it lies in [0,3] = [0,0.0222...] or
0 0 . x2x3... if it lies in
02

[f, 1] = [0.2,0.222...], where now we place no restriction on x2, x3,_

Similarly each point of K2 has the ternary form 0.00x3x4..., 0.02x3x4...,
0.20x3x4 ..., or 0.22x3x4 ... according to which of the four subintervals of
K2 it lies in. Extending this idea by induction, we see that the two defini¬
tions of K are equivalent.
Cantor’s middle-thirds set K has an uncountable infinity of points. To

125
126 *Some special topics

0 Ko 1

0 3
1

•» V •» Yl2 y * I

Fig. 4.1 Construction of Cantor’s middle-thirds set K = P)“=0 K„.

prove this suppose that K is countable and use reductio ad absurdum.

Thus suppose that all the members of K can be ordered as x1 = 0.xn
xi2*i3---> x2 — 0-x2ix22x23---> • ■ •, say, in ternary form. Now define y =
O-y1.V2.V3 • • • >where ym = 0 if xmm = 2 and ym = 2 if xmm = 0. Then 0 < y <
1, y e K and x„ # y V n. This proves the falsity of the supposition that K is
countable.
Nonetheless, K clearly contains no subinterval, however small, of [0,1],
and [0,1] contains an infinity of subintervals which do not intersect K.
Indeed the length of K is zero. To prove this, note that the length of K0 is 1,
of Kx is f, and the length of Kn+1 is two thirds of the length of K„. There¬
fore the length of K„ is (§)". Also KcK„ implies that K is not longer than
K„. Therefore

0 ^ length of K

^ length of K„ for n = 1,2,...

= (!)"
-* 0 as n -* 00.

Therefore the length of K is zero. Therefore K contains no interval, i.e. K is

totally disconnected.
The set K is closed, i.e. all the cluster points of K belong to K. To prove
this note that each of the K„ is closed and that the intersection of a set of
closed sets is closed.
Each point xeKisa cluster point of K. To prove this suppose that
 4.2 Dimension and fractals 127

x — O.Xjx2... in ternary form, and define yn = O.X!x2 ...x„. Then yn e K

and yn -* x as n -* oo. If x = 0.x tx2 ...xm has a ternary expression which
terminates then define yn = x + 2/3m+" if x # 1 and yH = x - 2/3" if x = 1.
Cantor defined a perfect set as a set which is both closed and such that
each point of the set is a cluster point.
Note that two of the thirds of KB+1, when magnified by a factor of three,
are identical to the whole of K„. This is an example of a self-similar or
scaling property. Many interesting sets are self-similar like K.
A set which is an uncountable subset of the points of IRm, is totally
disconnected, and is perfect is called a Cantor set; it may or may not be
self-similar.
The problems at the end of this chapter show that there is nothing very
special in Cantor’s choosing to remove the middle third of each interval
rather than the middle rth for some r such that 0 < r < 1. Also the ideas of
self-similarity may be used to construct Cantor sets in IRm for m > 1. Again,
self-similarity is not an essential property of a Cantor set, although it is a
useful property in constructing a Cantor set.

2 Dimension and fractals

To each set E c: Rm there is assigned a topological dimension, d say, which is
an integer such that 0 ^ d ^ m. We are familiar with the facts, for example,
that for a finite set of points d = 0, for a parabola d = 1, for the interior of a
triangle d = 2, for the interior of a sphere d = 3, and for [Rm itself d = m.
This notion of dimension goes back to Euclid’s definitions of point, line,
surface and solid, and to Descartes’ coordinate geometry because it is the
minimum number of coordinates needed to describe all the points of E.
This notion of topological dimension has been formalized in the twentieth
century on the basis of the idea of Poincare (1905, Chap. Ill, §3) to let a
point have zero dimension, to relate the dimension of a space to the way it
can be divided by boundaries, and to use mathematical induction. Thus
curves can be divided by cuts which are of zero dimension (i.e. by points)
and so have one dimension, surfaces can be divided by cuts of one dimen¬
sion (i.e. by curves) and so have two dimensions, etc. Cantor’s middle-
thirds set had no place in Euclid’s Elements, of course, but it may be said to
have topological dimension d = 0 because it is a subset of IR with zero
length (and so has neither breadth nor length).
However, there are some useful generalizations of this definition of di¬
mension. We shall give here an informal definition of the dimension, D say,
of a set E c: IRm. We shall see that D = d for ‘simple’ sets and d ^ D < m,
128 *Some special topics

Fig. 4.2 Covering of (a) a line segment (in R) of length / by N lines of length e,
(b) a circle of radius / and centre 0 by n squares of side e, and (c) a line segment
(in R2) of length / by N squares of side e.

but D is not an integer or D / d for some less simple sets. The ideas come
from the definition of Hausdorff dimension (Hausdorff 1919).
We shall in fact use box-counting dimension, or box dimension, which
is somewhat similar to Hausdorff dimension; the dimensions have equal
values for many of the sets we shall meet. To introduce the definition of box
dimension, first take an example in which E is a straight line segment of
length / in IR (see Fig. 4.2). Then the number N of equal lines of length e
necessary to cover all the points of E is given by N(e) ~ /e-1 as e -+ 0. Of
course, we could use as many intervals of length e as we wished to cover the
line segment, but if overlap is avoided and wastage at the ends is minimized
then the number N satisfies the limit. Next take E as the interior of a circle
of radius l in IR2. Then the number N of equal squares of side € necessary to
cover E is such that N(e) ~ nl2e~2 as e -» 0. These and other examples with
simple sets E give a relation of the form N(e) ~ Ve~d as e -» 0 for fixed V,
where d is the topological dimension of E in IRm and E is covered by hyper¬
cubes of side e. This property suggests the generalization that if a set E is
covered in this way such that

N(e) ~ Ve~D as e -* 0 (1)

for some numbers D and V then D is the dimension of E, whether D is or is

 4.2 Dimension and fractals 129

not an integer. Thus we define

lnN{e))
D = lim (2)
«4-o

if the limit exists.

Example 4.1: the box dimension of Cantor's middle-thirds set. We consider

first the set K„ of 2" closed subintervals of length 3_". So to cover K„ with
lines of length e = 3~" we need N(e) = 2" of them. Of course K may be
covered in the same way, because Kn+X c K„ and K = f)®=0 K„. Now we
seek D such that equation (1) is satisfied. We find e —»0 as n —* oo and

N(e) ~ 2" as n -» oo
_ gnln2

_ ^gnln3yn2/ln:

where
D = In 2/ln 3 = 0.63093...,

a result due to Hausdorff (1919, p. 172).

By the definition we require e to be a continuous variable which tends to
zero, although here we have used e = 3_" as the discrete variable n -> oo. If
e # 3~" for any integer n, then the covering of K may be ‘wasteful’, but the
limit is essentially the same. □

Example 4.2: the von Koch curve. An interesting curve S shaped like a
snowflake was proposed by von Koch (1904). It may be defined as follows.
Let S0 be the three sides of an equilateral triangle, each side having length
one. Then the length of the curve S0 is three. Let Sx be the curve for which
the middle third of each straight part of Sx is replaced by two equal straight
lines directed outwards, so that the replacement is two sides of an equilat¬
eral triangle of which the middle third is the base. It is easiest to see the
definition of Sx by looking at Fig. 4.3. Similarly we replace the middle
thirds of each straight part of Sx to define S2, and define S„ iteratively.
Then S may be defined as the limit of S„ as n-> oo, i.e. S = (X: 3 a point
X„ e S„ for n = 0,1,..., & X„ -*• X as n -» oo}. (Note that SB+X <4= S„. So the
limiting set S is not H?=o S„.) Now here S„ consists of 3 x 4" line segments,
each of length 3-". Therefore S„ can be covered by 3 x 4" equal squares of
side € = 3“". This gives
130 *Some special topics

Fig. 4.3 The curves S0, S, and S2 leading to the definition of the von Koch
curve S.

N(e) ~ 3 x 4" as n -*• oo

= 3e~D

if D = In4/ln3 = 1.26186. This gives the box dimension D of S plausibly.

Also the length of S„ is given by

L(€) = eJV(e) (3)

= 361-0.

Of course L(e) -*• oo as n -*• oo, i.e. as e -*• 0. Lastly note that the von Koch
curve is self-similar in the sense that if a twelfth of S„+1 is magnified by a
factor of three then it becomes identical to a third of S„, for n = 0,1,_
The von Koch curve resembles the shape of many coastlines and con¬
tours in some ways. Indeed, Richardson (1961) examined the western
coastline of Britain and discovered empirically that it was approximately
self-similar over many scales, the self-similarity being in a statistical sense
that the coastline looks similar on any scale of magnification unless a
specific feature can be recognized. He found that the length L of the coast¬
line satisfies the relation

L(e) * Ke1_0 (4)

for 10 km < e< 1000 km, where D « 1.25, by counting on maps, for vari¬
ous values of e, the number N = L/e of steps of length e from point to point
along the coast. He did this by ‘walking’ dividers whose points were a
distance e apart.
In the theory of box dimension we take the limit as e -*• 0, but, of course,
Richardson did not use measurements on scales smaller than those re¬
solved by the maps available to him, let alone on the scales of metres,
 4.2 Dimension and fractals 131

microns or infinitesimals. Nonetheless, the theoretical model of a Cantor

set of non-integral dimension is a useful description of the observations
made on a wide range of scales, and suggests an investigation of the geolog¬
ical reasons for different values of the dimension of different coastlines and
contours. □

Mandelbrot (1975) defined a fractal as a set whose dimension is strictly

greater than its topological dimension, i.e. D > d. However, the spirit of his
approach is that, like life, a fractal is an easier thing to sense than to define
rigorously. The essential property of a fractal is that its structure varies on
all small length scales. This is seen, for example, in Cantor’s middle-thirds
set and in the attractor of the Henon map. Similarly, definitions of strange
attractor vary, but for most practical purposes we may regard attractors
which are Cantor sets and fractal attractors as equivalent to one another
and to a strange attractor. It is certainly helpful to describe as fractal a set
whose box dimension is an integer but whose topological dimension is
different: for example, the plane set of Q4.ll is fractal although its box
dimension is one.
*We have simplified the concept of Hausdorff dimension in defining
box dimension D. The essential difference is that for box dimension the set
E is covered by hypercubes whereas for Hausdorff dimension the covering
sets are quite general in shape and size. For a more formal approach to
Hausdorff dimension, we define the Hausdorjf-s-measure of E c [Rm as

where E is covered by sets U, of diameter less than e and the infimum is

taken over all such €-coverings. If there exists D such that Jfs(E) = 0 for all
s > D, and Jfs(E) = oo, for s < D, then D is the Hausdorff dimension of E;
and we may often identify JfD(E) as being proportional to the quantity V
introduced above. There are also other definitions of dimension which are
not always integers. For the sets we shall meet, the values of the dimensions
are mostly equal.
It is usually impossible to evaluate rigorously the dimension of a fractal,
except for self-similar sets such as those in our examples. However, we can
compute the dimension of a given set more easily. Many methods to com¬
pute D have been devised, because the calculation of D by use of the cover¬
ing method is very time consuming. In these methods sometimes different
definitions of dimension are used (although different definitions can occa¬
sionally give different values of the dimensions).
132 ♦Some special topics

3 Renormalization group theory

3.1 Introduction
This section is an elaboration of some ideas raised in §3.4, so a review of
that section would make a good foundation to the reading of this one.
Note, in particular, Table 3.1 and Fig. 3.6. For new results first note that in
fact not only is the nature of period doubling universal but so, in a sense to
be seen soon, is the order of the p-cycles which arise at the bifurcations as
the parameter increases. Thus Fig. 3.6 has the same qualitative appearance
for a wide class of maps of which the logistic map is just one. This follows
from a remarkable theorem.

SarkovskiVs theorem. If F: (R -*■ IR is continuous, F has a k-cycle and /o k

in the following ordering of all the positive integers, then F also has an
/-cycle:

1<i 2 <i 22 <i 23 24 c • • •

23-9<i 23-7<i 23-5<i 23-3

• • • o 22 • 9 <] 22 • 7 <i 22 • 5 o 22 • 3

•••<i2-9<i2-7<]2-5<i2-3

■••<9o 7<i 5<a 3.

This powerful theorem with so few hypotheses is due to Sarkovskii

(1964); a simpler proof related by Devaney (1989) is more accessible. The
theorem is valid only for one-dimensional maps. The converse of the theo¬
rem is in fact also true, i.e. if / <i k then there exists a continuous function
F: IR -*■ IR such that F has a cycle of period / but not one of period k.
Note that first the powers of 2 are listed in ascending order, then the
products of the powers of 2 (in descending order) and the odd numbers (in
descending order). The theorem means, for example, that if F has a 10-cycle
then it also has a 176-cycle, because 176 = 24 • 11 2 • 5 = 10. In particu¬
lar, it implies that if F has a k-cycle where k is not a power of 2 then F has
an infinity of cycles, and if F has a finite number of cycles then all their
periods are powers of 2. It also has the following corollary.
 4.3 Renormalization group theory 133

Corollary. If F has a three-cycle then F has an /-cycle for all positive

integers /.

This astonishing corollary has been epitomized as ‘period three implies

chaos (Li & Yorke 1975). To understand the background to the epitome,
recall that, although the theorem tells nothing about the stability of the
/-cycles, experience of the logistic map in §3.4 suggests that almost all if
not all the /-cycles will be unstable. So the cycles will play the role of
the repellers in the metaphor of the pin-ball machine. Also recall that the
logistic map F(a, x) = ax(l — x) has stable cycles in the ‘windows’ of its
parameter a. For example, it can be seen in Fig. 3.5(b) that F has two stable
six-cycles, first a six-cycle on its own account and secondly a six-cycle from
the period doubling of the three-cycle. The six-cycle is visited by Fn(a,x) in
different orders in each of the two cases. Sarkovskii’s theorem does not
cover the multiplicity of a cycle of a given period, so it does not imply a
universal order of the appearance of cycles at the bifurcations of a differ¬
ence equation as a parameter increases. The theorem suggests period
doubling of a k-cycle to 2" • k-cycles for/c — 3, 5,... as well as 2; in fact each
of these sequences of period doublings leads to chaos with a Feigenbaum
relation of the form (3.4.9) for each value of k. Again, it is possible that only
a finite sequence of flip bifurcations occurs as a parameter increases, in
which case there is no route to chaos by period doubling.
Now we move on to examine the detailed structure of period doubling.
It is a good example of self-similarity. Period doubling is found to be
characterized by a universal scale a for the state variable x as well as the
scale <5 for the parameter a. The structure of the period doubling is there¬
fore revealed by renormalization, the name being used for 40 years by
theoretical physicists to describe groups of scaling transformations in the
theories of particle physics and of phase transitions. To explain renormal¬
ization group theory, we shall first introduce the concept of superstability
and then the scales themselves.
Numerical calculations of the value ar of a at which a 2r-cycle arises
from a flip bifurcation are especially difficult, because the cycle is very
weakly stable when a is near to ar and so computations over a long time are
needed to calculate the eigenvalue accurately. However, calculations of the
value Ar of a at which the 2r-cycle is most stable are much easier. Accord¬
ingly we say that a cycle is superstable if it is as linearly stable as it
can be, e.g. if the eigenvalue of F2r is q = 0 at each point of the 2r-cycle
134 *Some special topics

{X2, X2,..., X2r}. For an example of superstability with a cycle of period

one, the Newton-Raphson method to calculate a fixed point is superstable
and so converges very rapidly once a close approximation to a fixed point
has been found.

Example 4.3: the logistic map. It is shown in §3.4 that if F(a, x) = ax(l — x)
then there is a stable fixed point, i.e. a 2°-cycle, X = (a — 1 )/a for 1 < a ^
ax = 3 with eigenvalue q = 2 — a. Therefore A0 = 2 because q = 0 when
a = 2; then X = \. Also there is a stable 2x-cycle {XUX2} when
ax < a ^ a2 = 1 + V6 with = [dF2(a,x)/dx]Xi = Fx(a>X1)Fx(a,X2) =
4 + 2a — a2. Therefore A2 is the zero of q such that a2 < Al < a2, i.e.
Ax = 1 + V5 = 3.236; then X, = j,X2 = i(l + J5). □

Recall that, by use of the chain rule, the multiplier q determining the
stability of the 2r-cycle can be shown to have the same value f] j=i Fx(a> Xj)
at each point of the 2r-cycle, so that q = 0 if and only if the derivative of F
vanishes at one point of the cycle. Therefore, if F is a smooth convex
function with a simple maximum at Xm then q = 0 if and only if Xj = Xm
for one value of j, i.e. if and only if Xm belongs to the 2r-cycle, i.e.

F2r(A„XJ = Xm.

Thus Ar is a value such that

F2r(Ar,Xm) = 0.

In fact if F is a smooth convex function then Ar is the unique value; indeed,

as a increases from ar to Ar to ar+l, F2'{a, A^) decreases monotonically
from 1 to 0 to —1.
We have established that if a = Ar then Xm belongs to the 2p-cycle. So
the other points are Fj(Ar, Xm) for j = 1, 2,..., 2r — 1. Of these points the
closest to Xm is F2r~'(Ar,Xm). To see why this is true, first note that each
member of a 2r_1-cycle of F is a fixed point of Fr and the 2r-cycle of F
contains two 2r_1-cycles of F. The 2r-cycle of F bifurcates from the 2r_1-
cycle of F as a increases through aT and the line y = x cuts the curve
y = F2’ (a,x) at the two-cycle as well as the fixed point of F2' 1 (see, e.g.,
Fig. 3.4). So, as a increases through ar, the two points of the two-cycle of
F2' ' separate from the fixed point and one another. However, this leaves
Xu the point that becomes Xm when a = Ar, closest to F2'~1(a,Xl) and
these two points of the 2r-cycle stay closest as a increases to Ar. To prepare
to investigate the scaling of the separations of the points {Xt, X2,..., X2r}
of a 2r-cycle for large r, define
 4.3 Renormalization group theory 135

x,l
1-

Fig. 4.4 Sketch (not to scale) in the (a, x)-plane of the bifurcation diagram of a
one-dimensional map, showing the flip bifurcations and superstable 2r-cycles.

dT = F2'-\Ar, XJ -Xm for r = 1,2,..., (1)

the distance from Xm to the nearest other member of the superstable 2r-
cycle. Then the location of the flip bifurcations and superstable cycles is
summarized in Fig. 4.4.

3.2 Feigenbaum’s theory of scaling

We are now ready to describe Feigenbaum’s theory of period doubling,
although in addition a knowledge of the elements of applied functional
analysis will.help. It is interesting that Feigenbaum’s (1978) paper was
rejected by the first journal to which it was submitted (Cvitanovic 1984,
p. 244). Feigenbaum calculated Ar and dr numerically for several values
of r and for a few functions F and concluded that

A, = am - Bd~r + o(S~r), dr ~ D/( — a)r as r -*• oo, (2)

where B, D are constants which depend upon the map F, but S = 4.6692...
and a = 2.5029... are ‘universal’ constants which do not. This shows that a
is the x-scale of the route to chaos by period doubling much as <5 is the
a-scale. The scaling of dr can be expressed as

lim (— otY{F2r(Ar+l,Xm) - Xm] = -D/a. (3)

r-* oo

This leads to the further hypothesis that the limit

136 ♦Some special topics

gi(x - Xm)= \im {glr(x - XJ} (4)

r-»x

exists, where glr(x — Xm) = ( — ct)r{F2r(Ar+i,Xm + (x — Xm)/( — aY) — Xm},

for we see that (3) implies that ^(O) = -D/a, and calculations of
glr(x — Xm) for quite low values of r (rather than infinity) seem to confirm
the existence of a limit gx independent of F. Then the scaling of x — Xm
shows that only the behaviour of F near to its maximum determines gx and
so it is this behaviour which is responsible for the universality of gx.
To make the notation a little less cumbersome it is convenient to trans¬
late the origin of x to the maximum Xm of F. So henceforth we shall simply
put Xm = 0 without loss of generality.

Example 4.4: the logistic map. If F(a,x) = ax( 1 — x) then we may replace
x - Xm = x - \ by x to get the new function F(a, x) = a(i — x2), ensuring
that the maximum of F is now at x = 0. Then we find that

0lo(x) = F(Ax,x) = - x2),

0i i(*) = (-a)F2(,42,x/(-a))

= (-a)A2{i - Alii - x2/a2)2}

= <xA2{iiiA\ - 1) - jA\x2/a2 -I- A22x*/a4}.

We can similarly find glr(x) for r = 2, 3, etc. and plot the curves y = £lr(x)
in the (x,y)-plane to see that a limiting function g{ seems to emerge as r
increases. □

The essence of this scaling of x is captured by the operator T defined by

Ti/dx) = — a^(i//( —x/a)) (5)

for all continuous functions ij/. Then

T0i(x) - -CLgx(gx{-xlz))

= -« lim {(-a)rF2r(4r+1,(-a)rF2r(/lr+1,x/(-a)r+1))}.
r-*ao

Now define ^ by (j>iy) = (-a)rf2r(/lr+1,y/(-a)r), so </>2(y) = </>(<j>(y)) =

(-a)rF2r+1(/4r+i,y/( —a)r). Then taking y = x/( — a), we deduce that

Tgx(x)= lim {(-a)r+1F2'+'(/lr+1,x/(-a)r+1)}

r-*ao

= lim {(— <x)qF2q(Aq, x/( — a)4)},

= 0o(*)> say.
 4.3 Renormalization group theory 137

Similarly, it can be shown that

T0*(x) = gk-i (x) for k = 2,3,..., (6)

where gk is defined by

9k(x) = lim {(— a)rF2r(/4r+Jk,x/( — a)r)}. (7)

r-+ oo

Taking the limit as k -► oo in equation (6), we conclude plausibly that there

exists a function
g{x) = lim gk(x) (8)
k-»ao
such that
T 0 = 0, (9)

i.e. that there exists a ‘fixed point’ g of the nonlinear functional operator
T. The famous equation (9) was discovered in a discussion between
Cvitanovic and Feigenbaum (1978, p. 46). We shall incidentally show later
that the fixed point g is unstable.
Although we in fact know a from numerical solutions of difference equa¬
tions, its proper status at this stage of the theory is a constant to be deter¬
mined from equation (9). To find it first note that if g(x) is a solution of
equation (9) then so is pg{x/p) for all p / 0. So we may, by convention,
choose a particular value of p such that

0(0) = 1. (10)

Then, on putting x = 0 into equation (9) and using (5), it follows that

«=-1/0(1). (11)

Feigenbaum (1979) verified numerically the above scaling structure, and

sought to find g as an even function by expanding 0(x) as a series in powers
of x2, truncating the series, and equating coefficients of successive powers
of x2 in equation (9). In this way he found that

g(x) = 1 - 1.52763x2 + 0.10482x4 - 0.02671x6 + —, (12)

and a = —1/0(1) = 2.5029.... Thus a appears as a sort of nonlinear eigen¬

value of the functional equation (9).

Example 4.5: quadratic approximation to g. To solve equation (9) for all

x, where g is an even function and 0(0) = 1, and then find a = —1/0(1)
approximately, assume that 0(x) = 1 + bx2 for some constant b and
neglect all higher powers of x. Then substitution into equation (9) gives
138 *Some special topics

1 + bx2 = — a[l + b{ 1 + b(-x/a)2}2]

= — a(l + b + 2 b2x2/a2 + b3x*/a*).

Equating coefficients of x° and x2, and neglecting the term in x4, we find
that

1 = -a(l + b), b = — 2b2/ot.

Therefore a = -1/(1 + b) = -1/(1 - £a). This gives a quadratic equa¬

tion for a with solution a = 1 ± ^3. But we require a > 1. Therefore a =
1 + ^3 = 2.73..., and b = — |a = — 1.37.... It is a crude approximation,
but the example shows how to calculate g and a to higher approximations.
□

Next we move on to find the scaling of the parameter a in the route to

chaos by period doubling, evaluating S. We shall show that

gk{x) — g(x) ~ constant x <5~ku1(x) as k -* oo, (13)

where d is the eigenvalue belonging to the first eigenfunction u1 of the

linear operator Jg defined as the Frechet derivative of the nonlinear opera¬
tor T evaluated at the ‘point’ g in the space of continuous functions.

Example 4.6: calculation of the Frechet derivative of T. The Frechet deriva¬

tive of the operator T ‘at’ iJ/ is defined by linearization of T about i/f, i.e.
by the equation

T(\j/ -(- e^) = Ti/f + eJ^ + o(e) ase->0 (14)

for all (well-behaved) functions <j>. To find we expand

T(iA + e<t>) = — a(i^ + cMM-x/a) + e^(-x/a))

= — ou/di/K —x/a) + €</>( —x/a.)) - ae^(ip(-x/ot) -I- €</>(-x/a))

= -atf#(-*/a)) ~ #'(^(-x/a))-e^(-x/a)) + 0(c2)

— — x/a)) + 0(e2) ase-»0,

on assuming that i//, <t> are well-behaved and taking a Taylor series,

= T^ + eJ^ + 0(e2),

where J ^ is defined by

J+<J> = — ai//(i/f(—x/a))^(-x/a) - a^(-x/a)). □ (15)

 4.3 Renormalization group theory 139

To proceed to find the n-scaling as a -* a^, we expand

F{a,x) = F(ao0,x) + (a-aJf(x) + O{{a-aJ2} as a^a^, (16)

where /(x) = Fa(a0O,x). (Of course, if F{a,x) = af{x) then equation (16) is
exact for all a on omission of the remainder term 0{(a - a*,)2}.) Therefore
the Taylor expansion of the operator T acting on equation (16) gives

TF(a, x) = — txF(a, F(a, -x/a))

= —<xF(a, Fia^, -x/a) + (a - ajf(-x/a) + ■••)

= -aF(aQ0,F(a00, -x/a)) - a (a - a00)/(F(a00, -x/a)) + •••

- a {a - aCD)f( — x/ix)Fx(aO0,F(aa0, -x/a)) + •••

= TFla^x) + (a - + 0{(a - O2} asa-^a^.

On iteration, this process gives

TkF(a,x) = T*F{a00,x) + (a - OV,F(<Uii)/(x) + 0{(a - aj2}

= 0(x) + (a - aJJkgf{x) + 0{(a - aj2} + o(l) (17)

as a-* ax, k -* oo.

To simplify equation (17) consider the eigenvalue problem

J gu = ku

and suppose that it has eigenvalue kj belonging to the eigenfunction u7 for

j = 1, 2,...» where {u7} is a complete set of continuous functions over the
interval on which / is positive, for example a complete set for C[ — j, £] if
F(a,x) = — x2). Then

f(x) = X
i=l
</«/(*)

for some constants Therefore

J‘/W = X ^jUj(x)
7=1

~ £i^iui(x) as/c-+oo,

if we assume that ^ / 0 and we may take |A, | > |Ay| for j = 2, 3,.... Then
let 5 = A, and/i(x) = Therefore equation (17) gives

T*F(n,x) = g(x) + (a - aJSkh(x) + o(l) + 0{{a - aj2}

as k -* oo, a -+
140 *Some special topics

Therefore

TF(Ar,0) - 0(0) ~ (Ar - aJSrh{ 0) as r -> oo. (18)

Now

TF(Ar,0)= —<xF(Ar,F(Ar,0))

= —<xF2(Ar,0).

On iteration, this gives

TrF(,4r,0) = (— a)rF2r(/4r, 0).

But Xm belongs to each superstable cycle, so F2\Ar, Xm) = Xm, and,

after translation of the maximum to the origin, this gives F2r(Ar, 0) = 0.
Therefore

TF{Ar,0) = 0.

Therefore relation (18) gives

Ar - ax-g(0)/5rh{0) as r ^ oo

= S~r/h( 0),

which was anticipated in the first of relations (2).

Feigenbaum (1980) also examined the Fourier spectrum of (Fn(x0)} for
2r-cycles as r -> oo.
All this, then, is Feigenbaum’s heuristic theory of scaling of x and a in
the route to chaos by period doubling. The astonishing ubiquity of
Feigenbaum’s sequence in period doubling of maps of IRm for m > 1, of
solutions of differential equations, and of phenomena in laboratory experi¬
ments, stems from this theory for one-dimensional maps.

4 Liapounov exponents
In studying chaotic solutions (§§3.3, 3.4) we have met sensitive dependence
on initial conditions and met simple examples of neighbouring orbits
which separate exponentially. To be more formal we may define an in¬
finite invariant set S of a map F: IR -» IR to have sensitive dependence on
initial conditions if there exists S > 0 such that for all x e S and all neigh¬
bourhoods N (however small) of x there exists y e N and n > 0 such that
|F"(x) - F"(y)| > 3. So neighbouring orbits, however close initially, sepa¬
rate from one another, although each keeps close to the invariant set.
It is, moreover, a characteristic of neighbouring chaotic orbits that their
 4.4 Liapounov exponents 141

separation is an exponential function on average, though not necessarily an

exact exponential function. It is this rapid separation which makes it im¬
possible in practice to predict the behaviour of a chaotic solution far into
the future. This is in contrast to the behaviour of an orbit near an attractor
which is a fixed point or a periodic solution. These ideas can be quantified
by use of what are called Liapounov exponents.
Consider then a continuously differentiable map F: R ->■ R and suppose
that there exists k such that

|F"(x0 + e) - F"(x0)| ~ eenX as e -+ 0, n -> oo

provided that ce"'1 -»0 also, i.e.

dF"(x0)
e ~ee"A as n -* oo,
dx0

to express the average exponential separation of the orbit starting at

x0 + e from the orbit starting at x0. Therefore

dF*(x0) |
k = lim
N-'ao dx0 j

= Hm {N-1ln|F'(xJV_1)F'(x^_2)...F'(x0)|},

where x„ = Fn(x0), on differentiating a function of a function and using

induction (cf. Q3.5),

= lim X In|F'(x„)|l. (1)

N-ao „=o J

This shows that A is a measure of the exponential separation of the neigh¬

bouring orbits averaged over all points of an orbit around an attractor.
We now may formally define the Liapounov exponent k of an invariant
set of F by the limit (1), if it exists. Sometimes ex is called a Liapounov
multiplier or Liapounov number. In general k depends on the initial point x0
of the orbit, but it is the same for almost all x0 in the domain of attraction
of a given attractor. We see that for a stable cycle k < 0 and neighbouring
orbits converge (Q3.5), but that for a chaotic attractor k > 0. The
Liapounov exponent may be interpreted in terms of information theory as
giving the rate of loss of information about the location of the initial point
x0 (Shannon & Weaver 1949) or in terms of Kolmogorov entropy as
measuring the disorder of the system (Kolmogorov 1959).
In general k can only be found by computation, but it can be evaluated
analytically in some simple cases.
142 *Some special topics

Fig. 4.5 Computed values of the Liapounov exponent for the logistic map
F{a,x) = ax(l — x) for 2.9 < a < 4.

Example 4.7: some simple one-dimensional maps, (i) First take the linear
mapF(a,x) = ax. Therefore Fn(a,x) = anxand |dF"(a,x)/dx| = \an\ = enln|01.
Therefore A = ln|a|. If the fixed point X = 0 is stable then \a\ ^ 1 and
A ^ 0. Of course if a > 1 then orbits separate exponentially as they are
repelled by the fixed point, but there is no attractor and no chaos.
(ii) For the Bernoulli shift in §3.3 we found that <xn(0o + e) — <t"(0o) =
2"e as € -► 0 for fixed n, where 0O 6 (0,1) is fixed and a is the sawtooth map.
Therefore A = In 2. This is of the essence of sensitive dependence, although
the co-limit set of an aperiodic sequence is an invariant set but not an
attractor because it does not attract all neighbouring points. None the less,
it seems natural to define chaos so that the aperiodic solutions of the
Bernoulli shift are chaotic.
(iii) For the logistic map F(n, x) = ax(l — x) computations are in gener¬
al necessary to find A. Some results are shown in Fig. 4.5. Note how the
windows where a corresponds to a stable cycle have A < 0 and the chaotic
attractors have A > 0. Also note the stable three-cycle and its period doubl¬
ing as a increases above 3.8284. The imperfect resolution of the diagram
smooths out the fine structure of the curve. □
 Further reading 143

There is no widely accepted definition of chaos for all nonlinear systems,

but sensitive dependence is often used as a defining property. We may
define F to be chaotic on an infinite invariant set S if F has sensitive depen¬
dence on initial conditions, if periodic points of F are dense in S, and if F is
topologically transitive, i.e. if for all pairs of open sets A, B c= S there exists
n > 0 such that F"( A) nB#0. The condition of topological transitivity is
added to ensure that the invariant set S is not decomposable into the union
of two or more invariant sets.
Liapounov exponents may also be defined for higher-dimensional
maps. Suppose that F: Rm -»• lRm is a continuously differentiable map with
m x m Jacobian matrix J(x), the element in the ith row and 7th column
being dFJdxy Let q1(FN{x)),...,gm(Fw(x)) be the m eigenvalues of J(FAr(x)),
ordered according to decreasing magnitudes of their moduli. Then the m
Liapounov exponents of an orbit (FN(x)} as it approaches an attractor are
defined by

Aj = Oflnl4;(F*(x))l) for; = 1,2,..., m, (2)

= lim (i
v-oo
X
\Jy „=o
ln|<Z/(F"(x))|Y
J
(3)

on using the chain rule with the Jacobian matrix of F*. Of course it is the
largest of the Liapounov exponents which will in general give the exponen¬
tial separation of neighbouring orbits.
Kaplan & Yorke (1979) conjectured that the Hausdorff dimension of an
attractor in [Rm is given by

D = k + (t/)l IVil, (4)

where > k2 > • ■ • > Am and k is the largest integer such that Yj=i A, > 0.
It is now known that equation (4) is correct for most of the attractors likely
to be met in practical problems but that the right-hand side of (4) is only
an upper bound for the Hausdorff dimension. The advantage of computing
D by means of formula (4) is that it is quicker to compute Liapounov
exponents than to compute D by covering methods.

Further reading
§4.2 For a development of the theory of dimension and fractals, the
book by Falconer (1990) is recommended. It covers many aspects with
rigour and without excessive technical difficulties.
144 *Some special topics

§4.3 Collect & Eckmann (1980) prove the properties of period doubling
of one-dimensional maps.
§4.4 Falconer (1990, Chap. 13) treats the theory of Liapounov expo¬
nents with greater rigour.

Problems

Q4.1 Self-similarity of Cantor's middle-thirds set. Let K denote Cantor’s middle-

thirds set, and z: [0,1] -» [0,1) the map defined by t(x) = 3x modulo 1,
0 < t < 1. Then show that z maps K n [0,j] onto K n [0,1), K n [0,£]
onto Kn [0,j], etc.
Q4.2 Cantor’s middle-thirds set. (i) Prove that 4 belongs to Cantor’s middle-
thirds set K. (ii) Defining x = 2£®=1 show that x e K and that x is
irrational.
Q4.3 Self-similarity of a middle-halves set. Define the map z: R -*• [0,1) by t(x) =
4x modulo 1 and 0 < z < 1. Then show that z maps S and Sn [0,5] onto S
(less the single point 1), where S is the middle-halves set of Q4.7.
Q4.4 Invariance of Cantor’s middle-thirds set under a tent map. Consider the map
F: (1,00) x R -*• R defined by

F(a,x) = a{\ - |x- ^|).

Find the fixed points of F for a > 1 and their linear stability. Show that if
x < 0 then F(a, x) = ax and that if x > 1 then F(a, x) < 0. Deduce that
F"(x) -> —00 as n -* 00 if x < 0 or x > 1.
Taking a = 3, show that if 3 < x < § then F(3,x) > 1, that if either 0 <
x<3or§sgxi%l then 0 ^ F(3,x) < 1, and that if either 5 < x < § or
■§ < x < f then 3 < F(3, x) < f. Hence or otherwise show plausibly that
Cantor’s middle-thirds set is invariant under the map F.
[Mandelbrot (1982, p. 181).]
Q4.5 Invariance of a Cantor set under the logistic map. Consider the map
F: (4, 00) x IR —► IR defined by F(a,x) = ax(l — x). Show that if x < 0 then
F(a,x) < ax and that if x > 1 then F(a,x) < 0. Deduce that F"(x) -*■ —00 as
n -* 00 if x < 0 or x > 1.
Show that if \ - (1 - 4/a)1/2 < x < \ + (1 - 4/a)1'2 then F(a,x) > 1.
Speculate on the nature of the invariant set of F in the interval [0,1].
[Fatou (1906), Mandelbrot (1982, pp. 182, 192).]
Q4.6 A devil’s staircase. You are given that the continuous piecewise linear func¬
tion F„ is defined over the interval [0,1] by F„(0) = 0, F„'(x) = l//„ if x e K„
and F^(x) = 0 if x $ K„, where K„ is the union of the 2" closed intervals of
length 3“" used to define Cantor’s middle-thirds set K and /„ is the length of
Kn. Also F is the function defined by F(x) = lim^^ F„(x) pointwise for
0 < x < 1. Further, C„ is defined as the curve in the (x,y)-plane with equa-
 Problems 145

tion y = F„(x) for 0 ^ x < 1, i.e. as the graph of F„, and C is defined as the
graph of F.
(a) Show that F„(l) = 1, Fn($) = \ for n = 0,1,.... Sketch C0 and C2.
(b) Find the length of C„ and deduce that the length of C is 2.
(c) Show that F'(x) = 0 if x $ K and that F is continuous over [0,1].
Deduce that F is not differentiable atxeK.
(d) Interpret F as the probability distribution function of a random vari¬
able which is uniformly distributed over K.
Q4.7 A middle-halves set. Define S0 = [0,1], S, = [0,*] u [|, 1], S2 = [0,^] u
[l6. i] u [4, f§] [ft. 1], • • •, so that SB+1 is S„ after the removal of the mid¬
dle half of each subinterval, and thence define S = Pj“=0 SB.
Show that each number x e S has the quaternary form x = O-x^Xj...,
where the digit xB is either 0 or 3 for n = 1,2,....
Show that the box dimension of S is j.
[Cf. Falconer (1985, p. 15).]
Q4.8 The box dimension of the Sierpihski gasket. Let S0 be the interior of an
equilateral triangle with sides of length /. Regard S0 as the union of the
interiors of four equilateral triangles with sides length \l, and define Sj as S0
less the interior of the middle of the four triangles. Similarly, remove the
middles of the three triangles of S, to construct S2, and so forth (see Fig.
4.6). Define S = p|*=oSn.
Show that the box dimension of S is D = In 3/ln 2 = 1.58496.
[Sierpinski (1915), Eggleston (1953).]
Q4.9 The box dimension of self-similar sets. Define the set S as S = P)*=0 S„,
where S„ c: [R, S0 is the unit interval, and Sn+1 is similar to S„ such that SB+1
contains R times as many equal subintervals as S„ does, the length of each
being of r times the length of those of SB. Deduce that the box dimension is

D = lnR/ln(l/r).

State the values of R and r for Cantor’s middle-thirds set and for the von
Koch curve. Define the set S c [0,1] such that if x e S has decimal expres¬
sion x = 0.x j x2... then each digit x„ is even, and show that D = log10 5.

Fig. 4.6 Sketch of the sets for Q4.8.

146 *Some special topics

Fig. 4.7 Sketch of the sets for Q4.10.

Justify the application of the formula displayed above to the box dimen¬
sion of self-similar nesting sets in IRm, generalizing the definition of r.
[Mandelbrot (1982, p. 37).]
Q4.10 The Sierpihski carpet. Define self-similar sets S„ c= IR2 as follows, for n = 0,
1,_First define S0 = {(x,y): 0 ^ x,y ^ 1}, the unit square; Sj = {(x,y):
(x, y) e S0 & x or y $ (^, §)}, i.e. S t is the union of eight of the nine subsquares
of S0 with sides of length 3, the central subsquare being removed; S2 as S t
with the central subsubsquare of side 1/32 removed from each subsquare of
Sj, as shown in Fig. 4.7; and so forth. Hence define S = H*=o S„.
Show that the area of S is zero and that the box dimension of S is
D = In 8/ln 3 = 1.8928.
[Sierpinski (1916).]
Q4.11 A self-similar plane set with dimension of a line segment. Let S0 be the square
[0,1] x [0,1]. Divide S0 into 16 subsquares with sides of length i, and
define Sj as the union of the four corner subsquares. Similarly, remove the
12 non-corner subsquares of sides of length 1/42 from each subsquare of SL
to get S2, as shown in Fig. 4.8, and define S3, S4,... similarly. Then define
s=nr-«s..
Show that S = {(x,y): x = 0.xxx2... & y = 0.y1y2... in quaternary
form, where x„, y„ = 0 or 3 for all n}.
Show that the box dimension of S is D = 1.
Show that the orthogonal projection of S„ onto the line y = — ^x is the
segment AB, where A = ( —f,^) and B = (f, -§). Divide AB into 4" equal

□ S 0 E|
Q El E3 El

□ □ □ □
□ m □ □
S, s2
Fig. 4.8 Sketch of the sets for Q4.11
 Problems 147

subintervals. Deduce that to each subinterval there corresponds a sub¬

square of S„ and to each subsquare of S„ there corresponds a subinterval
ofAB.
[Cf. Falconer (1985, p. 16).]
Q4.12 The Cartesian product of Cantors middle-thirds set and an interval. Show
that if the plane set E is the Cartesian product of Cantor’s middle-thirds set
K and the unit interval, i.e. if E = {(x,y): x e K,0 < y ^ 1}, then the box
dimension of E is D = 1 + In 2/ln 3.
What is the box dimension of the Cartesian product of K and an m-
dimensional hypercube with sides of unit length?
Q4.13 The baker's transformation. The map F: S S is defined by F(x) = (<r(x),
g(a,x,y)), where S = [0,1) x [0,1) is the half-open unit square, x = (x,y),

<x(x) = 2x modulo 1 and 0 < o < 1,

and

. , f iay for 0 < x < , ,

g(a,x,y) = < Vmodulol and 0 < g < 1.
(ilay+l) forfs$x<lJ

Show plausibly that if 0 < a < 1 then the attractor of the orbit {F"(x)} is
topologically equivalent to the Cartesian product of the interval 0 < x < 1
and a Cantor set in the y-direction, and the box dimension of the attractor is
D = 1 + In 2/(ln 2 — In a).
[See Example 3.13.]
Q4.14 A non-uniform Cantor set. Let S0 = [0,1], = [0,^] u [§, 1], S2 =
[0. Te] v [e» i] u [§, 4] u [§, 1], • • •, whereby the right-hand half of the left-
hand half and the left-hand third of the right-hand half of each subinterval
of S„ is removed to form SB+1. The set S is defined by S = P)*=oS„. Show
that the box dimension of S is D = In 2/ln 3.
Q4.15 Proof of a very special case of Sarkovskii’s theorem. Prove that if F: R -► IR
is continuous and F has a two-cycle then F has a fixed point.
[Hint: it may help to consider the sign of F(x) - x.]
Q4.16 Proof of a special case of the converse of Sarkovskii’s theorem. Define the
continuous function F: [1,5] -» [1,5] such that F(l) = 3, F2(l) = 4,
F3(l) = 2, F4(l) = 5, F5(l) = 1, and F is piecewise linear; and note that F
has the five-cycle {1,3,4,2,5}. Sketch the graph of y = F(x) in the (x,y)-
plane for 1 < x < 5. Show that F has a unique fixed point and evaluate it.
Show that F3[l,2] = [2,5], F3[2,3] = [3,5], F3[4,5] = [1,4], and F3 is
monotonically decreasing on [3,4]; hence or otherwise prove that F3 has no
three-cycle.
♦Q4.17 An approximate model of period doubling. Consider period doubling for the
logistic difference equation in the form xB+1 = F(a, x„), where

F(a, x) = 1 — ax2.

F
148 *Some special topics

Show that the fixed points are X± = {— 1 ± (1 4- 4a)112}/2a respectively,

and that X+ becomes unstable at a flip bifurcation as a increases through
Show further that the second-generation map is given by F(a, F(a, x)) =
1 — a + 2a2x2 — a3x4.
We next assume that, in the Feigenbaum sequence, a stable 2r-cycle be¬
comes unstable as a increases through ar, that ar-*ax, that the limit is
determined by the behaviour of F(a, x) near its maximum where x = 0, and
that there are scaling properties as r -» oo. Accordingly we may assume that
x„ and a — are small, and proceed to develop an approximate model as
follows.
Renormalize x„ by the factor a0, where a0 = — l/(a — 1) and neglect x4
to deduce that xn+2/oc0 as 1 — bx(xj<x0)2, where bx = </>(a) and </>(a) =
2a2(a — 1). Write this as xn+2 = 1 — bxx2, by change of notation. Repeat
this scaling argument to show similarly that xn+2, = 1 — brx2 approximately,
where br = <j>(br_x). Deduce that if the 2r-cycle becomes unstable when a —
ar, then br = f and \ = ^r~1(ar) approximately. This gives ar -> ax, where
= </>(ax) approximately, i.e. ax w j(l + J3) = 1.37, and ar -+ a as
r -> go, where a = — 1 /(a^ — 1) * —2.8.
From the approximate relation ar = </>(ar+l), deduce that

(ax - ar+l)/(ax - ar)^S asr-oo,

where 3 * « 4 + ^3 = 5.73.
[Landau & Lifshitz (1987, §32).]
Q4.18 The tent map. Show that the Liapounov exponent of the tent map defined
in Q3.22 is ln|a|.
Q4.19 The baker’s transformation. Show that the Liapounov exponents are Xx =
In 2 > 0 and X2 = — ln(2/u) < 0 for the baker’s transformation defined in
Example 3.13 and Q4.13. Verify that formula (4.4) gives the box dimension
of the attractor correctly.
Q4.20 Arnold's cat map. Define F: T2 -> T2, where T2 = IR2/Z2 is the torus, by

F(x,y) = (x + y modulo 1, x + 2y modulo 1).

Show that F is area-preserving.

Find the Liapounov exponents of F.
[Falconer (1990, p. 196).]
 5
Ordinary differential equations

Time present and time past

Are both present in time future,
And time future contained in time past.
If all time is eternally present
All time is unredeemable.
What might have been is an abstraction
Remaining a perpetual possibility
Only in a world of speculation.
What might have been and what has been
Point to one end, which is always present.
T.S. Eliot (Burnt Norton, I)

1 Introduction
We have introduced ordinary differential equations in Chapter 1, and
considered some bifurcations of their equilibrium points. We shall next
consider the points, their stability, attractors and bifurcations more sys¬
tematically and in greater depth. In this chapter we make a few general
remarks and definitions, and consider a few general properties, before ex¬
amining special classes of equations in detail in later chapters. We shall
observe many similarities but some dissimilarities between the properties
of difference and differential equations.
We can represent an mth-order differential equation as a system of m
first-order equations, much as we did for difference equations. (Again, the
converse is in general false.) So we need consider only

^ = F(x,t) (1)
df

for F: IRm x U -> IRm. We shall assume that F is sufficiently well behaved for
there to exist a unique solution x(t) of equation (1) such that x(0) = x0 for
given x0 e IRm; moreover we assume that x(f) is nonsingular, at least for t
less than some finite value depending on x0 as well as F. The system (1) is
said to be autonomous if F does not depend on the independent variable t
explicitly, and to be non-autonomous otherwise. We can transform a non-

149
150 Ordinary differential equations

autonomous system (1) in Rm to an autonomous system in IRm+1, much as

we did for difference equations. To do this, identify y = [x,ym+1]T e IRm+1,
choose ym+i(0) = 0 and write equation (1) as the suspended system.

dy
(2)
dt

The (m + l)th component ensures that ym+1(t) = t for all t > 0. Therefore

dy
= G(y),
dt

say, where G: (Rm+1 -♦ IRm+1 is defined by G(y) = [F(y), 1]T and x =

[yx,...,ym]T. So we have been able to replace the explicit dependence
of F on t by an equivalent implicit dependence of G on t. As a result, we
may consider only autonomous systems, without loss of generality.

Example 5.1: Dujfing's equation. Suppose that

d2x ,
-7-5- + x — = T cos cot
dr

for real parameters T and 00. Then define Xj = x, x2 = dx/dt, x3 by

dx3/dt = 1 and x3(0) = 0, and x = [x1,x2,x3]T e [R3. Therefore
x3(t) = t on integration, dxjdt = dx/dt = x2, and dx2/dr = d2x/dr2 =
T cos tut — x + 5X3 = T cos cox3 — Xj + yx3. Therefore dx/dt = F(x),
where F(x) = [x2, Tcos cux3 - xx + |x3,1]T. a

Example 5.2: equivalence of an mth-order equation to m first-order equa¬

tions. Given that x(m) =/(x,x',x",...,x(m_1),t), where x' = dx/dt etc., de¬
fine y = [x,x',...,x(n,~1)]Te !Rm. It follows that dy/dt = [x',x",...,x(m)]T
= F(y, t), say, where we define F(y, t) = [y2, y3,...,ym,f(y, t)]T e Rm.
It is conventional to write computer routines to integrate a system of m
ordinary differential equations rather than an mth-order equation, so the
method of this example is important in computing solutions of a differ¬
ential equation. □

Consider orbits (x(t)} by solving the equation

dx
dF = F(x) (3)

for t > 0, with initial condition x(0) = x0 for given x0 e (FT. The orbit may
be regarded as the path of a particle which starts at a given point x0 and
moves with velocity F in the phase space for t > 0. To describe the
 5.1 Introduction 151

paths it is helpful to define the map <j>s: IRm x [R -► IFT, called the flow,
which maps each point x0 in phase space for each ‘time’s to x(s), where x(f)
is the solution of the above system such that x(0) = x0. Thus the flow <J», is
the analogue of the iterated map F" for a difference equation xn+1 = F(x„).
Note that <M<Mxo)) = <MX(0) = x(s + t) = <k+r(xo)> and <M<Mxo)) =
<{>s+r(x0), similarly, so that ^(J), = <|>s+t = <j>,<|>s, just as FmF" = Fm+" = F"Fm
for m, n = 0,1,2,.... Therefore all the flows <|>( for t > 0 are the elements of
an Abelian semigroup (not a group because the inverse = <J»_r may not
exist).
An equilibrium point, or critical point, X is a zero of F, for if x0 = X and
F(X) = 0 it follows that x(f) = X for all t > 0. Note that an equilibrium
point is a fixed point of the flow <J>, for all t.
An equilibrium point X is said to be stable if V e > 0 3 <5(e) such that

|x(f) — X| < e V r > 0

V x0 such that |x0 — X| < S. This definition of stability, analogous to the

definition of stability of a fixed point of a difference equation in Chapter 3,
is that the solutions x are uniformly continuous for t > 0 at x = X with
respect to variations of initial points x0. The point is asymptotically stable
if it is stable as above and moreover

x(t) -> X as t -> oo

V x0 such that |x0 — X| < 5. It is globally asymptotically stable if it is stable

and x(f) -*■ X for all x0 e IRm.
To consider the stability of a given point X of equilibrium we may,
without loss of generality, suppose that X = 0. To show this, define the
perturbation
x = x - X.

Therefore
dx' dx
dt dt

= F(x)
= F(X + x')

= H(x'),

say. Then H(0) = F(X) = 0, as required.

So we shall examine the stability of the null solution of equation (3),
assuming that F(0) = 0. Stability concerns small perturbations, so we shall
linearize the system by approximating F for small x. Now if F is continu-
 152 Ordinary differential equations

ously twice differentiable, then

F(x) = Jx + 0(x2) asx->0,

where the m x m constant matrix J is defined as the Jacobian matrix with

element [5Ff/dx^]0 in its ith row and jth column. Therefore the linearized
system of (3) at X = 0 is

To solve this linearized equation, we use the method of normal modes,

assuming that x(f) = es,u. Therefore

Ju = su; (5)

so the exponent s is the eigenvalue of J belonging to eigenvector u. Let us

suppose that the real matrix J has eigenvalue Sj belonging to eigenvector u,
for;' = 1, 2,..., m and that the set {ux,...,um} spans IRm. (The case when
J does not have m independent eigenvectors requires special treatment.)
Then for all x0 e IRm there exist £l5 <^2,..., such that

x0 = I j. (6)
j~i
It follows that the solution of equation (4) is
m
x(t)=L^exp{sjt)Uj for t ^ 0. (7)
j=i

(Note that Sj and u7- are either a real triplet or one of a complex conjugate
pair of triplets because J and x0 are real.) We infer asymptotic stability,
with exponential decay of all modes if Re(s,) < 0 for j = 1, 2, ..., m and
instability of X if Re(s,) > 0 for at least one value of j. It is said that the
equilibrium point 0 is linearly stable or infinitesimally stable if it is a stable
point of equilibrium of the linearized system (4), whether it is a stable or an
unstable point of equilibrium of the nonlinear system (3).
A cluster point or limit point x*, of an orbit (x(f)} as t -* oo is defined as a
point such that V e, t > 0 3 f! such that

lx(*i) - xool < e and tx > T.

We may qualify e by ‘however small’ and x by ‘however large’. The defini¬

tion implies the existence of a sequence tx < t2 < • • • such that

X(*J -* xao and tn -* oo as n -* oo.

In particular, if x(f) —* X as t —► oo then X is an equilibrium point which is a

cluster point of the orbit. The set of cluster points is called the co-limit set of
 5.2 Hamiltonian systems 153

the orbit. An attractor is an co-limit set to which all neighbouring orbits

tend as t -> oo. Thus an attractor may be an asymptotically stable point of
equilibrium or a solution of period T, i.e. a solution for which T is the least
number such that x(f + T) = x(r) V t, to which neighbouring orbits tend.
The domain of attraction, or basin of attraction, of an attractor A <= is
defined as

D(A) = {x0: dx/df = F(x), x(0) = x0 & the co-limit set of {x(t)}

is contained in A}.

An oc-limit set and a repeller may be defined as analogues of co-limit set and
attractor respectively as t -*■ — oo.
To consider bifurcations next, suppose that

l=F(a'x)
for F: 1R' x IRm -»• !Rm. If the topology of the phase portrait changes with a
in all neighbourhoods of a0 then we say that there is a bifurcation at a0,
and a0 is a bifurcation value of a. This definition, at the cost of vagueness,
deals with the bifurcations of unsteady as well as steady solutions of the
differential system.

2 Hamiltonian systems
Next we examine the evolution of the volume of a set of points governed by
equation (1.3), and, in particular the case in which the volume is constant
for all initial sets.
First trace the evolution of a single orbit for a little time. Now

x(t + h) = x(f) -I- hF(x(t)) + o{h) as h -*■ 0

if F is continuously differentiable. Therefore the Jacobian of the transfor¬

mation <J)A: x(t) i-» x(r + h) is

= 1 + h div F 4- o(h) as h -> 0,

where div F, i.e. V • F, = [dFJdx;]x(f) = trace J(x), because the determinant

consists of ones in the diagonal with a term of order h added to each
element (whether the element is on or off the diagonal). It follows (§3.3.2)
that the system is measure-preserving, i.e. that the volume of a set of the
‘particles’ is independent of t, if
154 Ordinary differential equations

div F = 0. (1)

This may be recognized as the equation of continuity of an incompressible

fluid composed of the particles.
An especially important subset of measure-preserving systems is the set
of Hamiltonian systems, important because of their omnipresence in the
theories of classical and quantum mechanics.
A Hamiltonian system is of the form

dq _ dH dp _ _ dH
dt dp ’ d£ dq

for a Hamiltonian function, or simply Hamiltonian, H(p,q,t). It arises

(Q5.12) from a generalized form of Newton’s laws of motion, where p e IR"
is the generalized momentum and q e IR" the generalized coordinate of a
mechanical system of degree of freedom n; then H is usually the sum of the
kinetic and potential energies. We may rewrite the system in the form

^ = F(x,0,

where F: IR2" x IR -*■ OR2", x = (q, p) and F = (dH/dp, -dH/dq). We see that

divF =i±(f)+i± = 0.
i=i dqt \dpij k\ dpi

It follows that a Hamiltonian system preserves volume in the 2n-dimen-

sional phase space. This is Liouville's theorem of analytical dynamics.

Example 5.3: the simple harmonic oscillator. When H(p,q) = \{p2/m +

mto2q2) we find

dq dH dp dH
mco2q.
d t = TP=plm’ d7

d2q
so ^ + “4 = 0.0

Note that if a Hamiltonian system is stationary, i.e. if H = H(p, q), then

dH dH dp dH dq
 5.3 The geometry of orbits 155

Therefore H is constant. If we identify H as the total energy of the mechani¬

cal system then this represents the conservation of energy. Accordingly
we sometimes call systems with div F = 0 measure-preserving or non-
dissipative, and those with div F < 0 dissipative. In this way measure pre¬
servation is associated with frictionless systems. We see that each orbit is
confined to the hypersurface, H{p, q) = constant, in IR2" if H is independent
of t; this hypersurface has topological dimension 2n — 1 in general.

3 The geometry of orbits

In the phase plane of a two-dimensional autonomous system, the attract¬
ors are no more complicated than a point or a closed curve, giving a peri¬
odic solution, because of the special topology of a plane: if an orbit starts
inside or outside a closed orbit then it remains inside or outside respective¬
ly for all time, because two orbits can cross only at an equilibrium point,
i.e. at a zero of F. We shall discuss the phase plane in detail in the next
chapter. First, we shall take a glimpse at attractors in IRm for m 3, where a
closed curve does not have an inside and an outside, and see that chaotic
attractors are possible. Note, however, that a non-autonomous second-
order system is equivalent to a third-order system and so may have chaotic
attractors; also F(x, r) may have different values at the same point x of the
phase plane at different times on an orbit, so the orbit may cross itself.
When m 3 attractors may have rich structures. A closed orbit still
represents a periodic solution. Also an orbit may, for example, wind
around a torus T2. To illustrate this possibility we can consider a geometri¬
cal model without reference to a particular differential system. Suppose
then that a point P moves on T2 <= IR3 with angles shown in Fig. 5.1

4>\

0 2n 0
(a) (b)
Fig. 5.1 (a) The torus T2 and its polar angles 6 and 4>. (b) Map of the torus onto
the half-open square [0,2n) x [0,2n).
156 Ordinary differential equations

Fig. 5.2 Orbits on a torus, {a) Periodic solution x(t) and fixed point of the
Poincare map. (b) Periodic solution x(t) and two-cycle of the Poincare map.
(c) Quasi-periodic solution x(f) cutting circle densely in the section of the
Poincare map.

such that
9 = 2nkt and </> = 2nlt for t ^ 0,

where k, l are real. To picture the orbits on the torus it is helpful to use the
Poincare map on the plane section 9 = 0. Now P returns to the plane
whenever 2nkt = 2wr for n = 1,2,..., i.e. when t = n/k, at which times (f> =
2nln/k. Therefore the Poincare map is the rotation map of points on the
circle cut by the torus on the plane 9 = 0, </>n+x = </>„ + 2nl/k, or </>„+l/2n =
</>„/2n -(- l/k (see Example 3.9). It follows that the orbit of the Poincare map
and thence the orbit of P on the torus is periodic if and only if l/k is
rational; in that case l/k = q/p for coprime integers p and q, and P winds p
times around the torus in the 9-way and q times in the ^-way before it
returns to its starting point. A few cases, relating the Poincare map to
orbits on the torus, are sketched in Fig. 5.2. See also Poincare (1885b,
p. 226).
The case with irrational l/k gives an example x(t) of what is called a
quasi-periodic function with two fundamental frequencies, namely 2nk and
2nl. Quasi-periodicity is a natural generalization of periodicity. A quasi-
periodic function x: R -► IRm is defined as one for which there exists a func¬
tion P: IR" -*• IRm and fundamental frequencies fx, /2,..., f„ such that

(a) \{t) = P(f,t,f2t,...,//) for all t, where P has period 2n in each of its n
arguments, and
(b) n ^ 2 and no frequency is a rational multiple of another.

(Note that we need condition (b) to ensure that quasi-periodic functions

differ from periodic functions; because if n = 1 or, for example, /, = 2f2 =
••• = nf„, then x becomes a periodic function.)
We have not met analogues of quasi-periodic solutions in Chapter 3
because difference equations have solutions of only integral periods, and
all integers are rational multiples of one another.
 5.4 The stability of a periodic solution 157

Example 5.4: some quasi-periodic functions, (i) x(t) = cos t + 3 cos y/21
gives a quasi-periodic function x: IR -*• IR with fundamental frequencies
1 and yj2. (ii) x(f) = [5 cos 2t sin nt, 1 + sin2f + sin37tf]T gives a quasi-
periodic function x: IR -► IR2 with fundamental frequencies 2 and n. □

Reverting to our discussion of the orbit of P on the torus, we may

envisage a topologically equivalent attractor, so that neighbouring orbits
approach the given attractor as t -* oo. Then the Poincare map on a plane
section would give a closed curve as an attractor, and the map of the curve
would be topologically equivalent to the rotation map.
More complicated solutions (x(t)} c [Rm for m ^ 3 may approach
attractors of non-integral dimension. They are sometimes called strange
attractors and found to be associated with chaotic solutions which have
sensitive dependence on initial conditions. The attractors with m = 3, let
alone m > 3, have not been classified yet.

In summary of this chapter so far, the concepts of steady solution, periodic

solution, quasi-periodic solution, stability, a flow, an orbit, its co-limit set,
an attractor, and a domain of attraction have been described. These are the
fundamental concepts of the qualitative theory of ordinary differential
equations. A change in the topological character of an attractor as a pa¬
rameter varies is a bifurcation, which may lead to symmetry breaking or to
chaos. Detailed applications of these concepts will be developed in the
remainder of the book, starting with a description of the stability of a
periodic solution in the next section.

*4 The stability of a periodic solution

As a parameter increases, a periodic solution of a nonlinear differential
system may become unstable and bifurcate, thereby leading to a quasi-
periodic solution with two fundamental frequencies, just as a steady so¬
lution may become unstable and lead to a periodic solution at a Hopf
bifurcation. The bifurcation sequence of an equilibrium point, followed by
a periodic solution and then a quasi-periodic solution is represented geo¬
metrically by attractors in phase space as the sequence of fixed point, S1
and T2 when the parameter increases. So the stability of a periodic solution
is fundamental to this sequence of bifurcations.
We shall in this section first define the stability of a periodic solution
and then sketch a theory of finding when a periodic solution is stable. The
theory is a well-known method to treat the linear stability of a periodic
158 Ordinary differential equations

solution, analogous to the method of normal modes to treat the linear

stability of a steady solution. It is called Floquet theory; physicists some¬
times call it Bloch theory.
First we shall define the stability of a periodic solution in two different
ways. Suppose that the differential system

£ = F(x), (1)

for F: [Rm -*■ IRm, has a known solution X of period T, so that X(f 4- T) =
X(f) for all t. Then X is represented by a closed curve in phase space. The
solution is said to be stable in the sense of Liapounov, or simply stable, if we
replace X by X(r) in the definition of stability of an equilibrium point in §1,
i.e. if x(r) — X(t) is small for all t ^ 0 whenever x(0) — X(0) is sufficiently
small. It follows then that the limit cycle of §1.6 is stable.
However, this definition of stability is often stricter than is desirable,
because we are often more concerned with the orbit in phase space rather
than where on the orbit the solution is at a given time. To understand this,
consider the example of the system

^ = — y{x2 + y2)112, ^ = x(x2 + y2)1'2.

On taking plane polar coordinates such that x = r cos 6, y = r sin 9, r ^ 0,

it follows easily (§6.1) that

d6
r.
dt

Therefore
r(t) = r0, d{t) = r0t + 0O,

i.e.

x(t) = r0 cos (r0t + d0), y{t) = r0 sin(r0f + 0O),

and the solution has period 2n/r0. This solution, X say, is unstable, because
if there is a neighbouring solution, Xt say, such that r = r0 + 5, 6 = d0 at
t = 0 for S > 0 then

|Xi(f) - X(t)| = {(r0 + d)2 - 2r0(r0 + d)cos(dt) + r2}1'2

does not remain small for all t > 0, however small 1X^0) - X(0)| = 6 may
be. Nonetheless, the orbit of in the phase plane is the circle r = r0 + 5,
which does remain close to the orbit r = r0 of X for all t > 0 when S is
 5.4 The stability of a periodic solution 159

sufficiently small. It is only the phases 6 of \x and X which do not remain

close. So we might wish to deem this solution X stable by defining stability
differently.
Accordingly, a solution X is said to be orbitally stable, or stable in the
sense of Poincare, if the orbits (x(f)} for t ^ 0 of all neighbouring solutions
remain close to the orbit of X in phase space. Then stability in the sense of
Liapounov implies stability in the sense of Poincare. However, stability in
the sense of Poincare does not imply stability in the sense of Liapounov,
the system of the previous paragraph being a counterexample: the periodic
solution described there is unstable but orbitally stable.
The orbit {X(f)} for t ^ 0 of a given solution, periodic or otherwise, may
be pictured as a curve surrounded by a tube of radius e in phase space Rm.
If all orbits starting at t = 0 within distance S of X(0) remain within the
tube for all r > 0 then X is orbitally stable. If, moreover, there is only small
‘shear’ between the orbit of X and neighbouring orbits within the tube
then X is stable in the sense of Liapounov.
The essence of Floquet theory is to treat the stability of the periodic
solution X of system (1) as follows and as illustrated in Fig. 5.3. Consider an
orbit of x(f) which starts at x(0) near X(0), so that the perturbation x' =
x — X of the orbit is governed by the linearized system of (1). Then
L: x'(0) i—* \'{T) generates a linear map of all x(0) near X(0) and the evolu¬
tion of the solution x(r) for t ^ T may be found by iterating the map L.
Suppose that L has eigenvalues qx, q2, ■ ■ ■, qm■ Therefore L"{x'(0)} -* 0 as

Fig. 5.3 Sketch of the orbits of {X(r)}, {x(t)} and the linear map
L: x(0) - X(0) i- x(T) - X(0).
160 Ordinary differential equations

n -*■ oo, i.e. x(nT) approaches X(nT) = X(0), if\qj\ < 1 for j = 1,2,..., m; in
this case (x(f)} is attracted by the closed orbit {X(f)} and so the periodic
solution X is asymptotically stable (and orbitally stable). Similarly, if \qj\ >
1 for at least one eigenvalue qs then X is unstable (but could be orbitally
stable if the eigenvector u, were tangential to the orbit of X at t = 0).
The method may be expressed in more detail analytically. The lin¬
earized system of (1) for small perturbations x' = x — X is

^ = J(X(!))x', (2)

where the Jacobian matrix J has elements \_dFi/dxj']x as usual. Now the
linearized system (2) has coefficients which are functions of t with period T.
There is a complete analytic theory of the solution of such linear systems
due to Floquet (1883). In particular (cf. Q5.17), it can be proved that for
each fundamental matrix <I> of system (2), i.e. for each m x m nonsingular
matrix such that

— = J(X(t))0>,

there exists an m x m nonsingular matrix P with period T and a constant

m x m matrix R such that

<D(r) = P(r)e,R. (3)

We can in general diagonalize R by use of similar matrices. Then there exist

m linearly independent solutions of the Floquet system (2) such that

= Pj(t) exp(s/) for; = 1,2,..., m, (4)

where pj(t + T) = pj(t) for all t, and Sj is an eigenvalue of R. The Floquet

exponent Sj is related to the Floquet multiplier qj by = exp(SjT). So the
condition for linear stability is that \qj\ 1, i.e. Re(sj) ^ 0, for all;. To
calculate R, and hence the Floquet exponents, it is necessary to integrate m
independent solutions of the system (2) over only one period, from t = 0 to
T, say.
If a solution X of period T depends on a parameter such that a Floquet
multiplier qx decreases through — I as the parameter increases, then the
perturbation develops a period double that of the basic solution X. This is
called a subharmonic instability. It may lead, as the parameter increases, to
a bifurcation with the origin of a solution with double the period of the old
one. Again, if qx, q2 = qx have a modulus which increases through one as
the parameter increases then there may be a bifurcation with the origin of
 Problems 161

a quasi-periodic solution whose fundamental frequencies are 2n/T and

Im(lng1)/7’ initially; such a bifurcation is similar to a Hopf bifurcation in
the sense that the solution acquires an extra frequency (and phase) at the
bifurcation.

Further reading
§5.1 We have assumed that the right-hand side of the differential system
(1) is well enough behaved for the solutions of the initial-value problem to
exist and to be unique. Many good books on the analytic theory of ordi¬
nary differential equations, e.g. Coddington & Levinson (1955, Chap. 1)
give the relevant conditions and theorems. Also some conditions for which
stability or instability for the linearized system determine the stability of an
equilibrium point are stated and proved by, e.g., Coddington & Levinson
(1955, Chap. 13).
Poincare’s contributions to the theory of ordinary differential equations
which we describe in Chapters 5-8 are prodigious. Much of his work was
first published in a series of papers in the 1880s, much in a book (Poincare
1892,1893,1899). In addition to finding the results which bear his name, he
framed the ideas and coined the names of node, focus, saddle and limit
cycle, discovered the recurrence theorem (Q5.14), anticipated chaos (§8.3),
and conceived index theory, rotation number and the reduced three-body
problem. The development of his ideas has been traced by Gray (1992).
§5.4 For further reading on Floquet theory, Coddington & Levinson
(1955, pp. 78-81, 218-20) and Jordan & Smith (1987, pp. 245-60) are
recommended.

Problems
Q5.1 A Poincare map of an equivalent autonomous system. Show that the non-
autonomous first-order equation

dz
— = —Z + € COS t
dt

is equivalent to the autonomous second-order system,

dz d0
— = —z + €cos0, — = 1, with 0(to) = t0.
dt dt

Represent the orbit (x(f)} c S1 x R with cylindrical polar coordinates

(r, 0, z) as points (1,0(t), z(t)) on the surface of a cylinder with unit radius, so
that (1,0,z) and (1,0 + 2«7t,z) coincide for n— +1, ±2,.... Then construct
162 Ordinary differential equations

the Poincare map P®°: E -*E, where E = {(r, 0, z):r — 1,9 = 0O} is the line
of intersection of the orbits with the half-plane 9 = 90, deducing that

P®°(l,0o,z) = (l,0o,ze-2* + 2^(1 - e_2*)(cos0o + sin0o)).

Find the fixed point Z of P®° and show that Z -* 0 as e -» 0 for all 0O. For
what values of 0O, e is Z stable?
Q5.2 Gradient systems. Show that if V: |Rm —► IR is continuously differentiable,

dx
— = -grad F(x),

and x*, belongs to the co-limit set of an orbit, then x^ is an equilibrium

point.
Q5.3 An unstable point of equilibrium which is linearly stable. Show that if

— = 2x2y — = — 2xy2
dt dr y ’

then the origin is an equilibrium point. Show also that xy is constant on

each orbit. Deduce that the origin is unstable.
Show that the origin is a stable point of equilibrium for the linearized
system, namely

Q5.4 Euler’s equations of motion of a rigid body about a fixed point. It is given that
a rigid body freely rotating with angular velocity a) = (a)l,a>2,a>3) about
a fixed point (i.e. a point fixed to an inertial frame) is governed by the
equations
da). dca,
A-—= (B - C)(D2a>3, B-— = (C - A)(o3(ox,
at at

da) 3
C— = (A - B)a)1co2,

where A, B, C are the principal moments of inertia of the body about its
fixed point. Show that

T= UAcoj + Ba)\ + Co)2), h2 = A2co2 + B2co2 + C2a>2

are constants of the motion.

Show that the steady solution tOj = n, a)2 = a)3 = 0 is stable, with small
oscillations of period 2nn~l {BC/(A - B)(A - C)}1/2, if either A > B > C or
A < B < C, and that it is unstable if B > A > C. Discuss the relevance of
this result to the motion of a tennis racket which is spun and thrown into the
air.
 Problems 163

Q5.5 The Rdssler system. Consider the system

dx dy dz
Tt‘~y-Z’ dt = x + ay' +

Find the equilibrium points, and conditions for their existence. Sketch the
bifurcation diagram in the (c, x)-plane for fixed a,b> 0, naming the bifurca¬
tion points.
[Rdssler (1976).]
Q5.6 The pitchfork bifurcation of the Lorenz system. Show that X = 0 is an equi¬
librium point of the system
dx
— = [tr(y - x),rx - y - zx,-bz + xyf

for x = [x,y, z]T, and positive parameters r, b, a. Linearize the system about
the null solution to find the eigenvalues s and eigenvectors of the normal
modes proportional to e5'.
Show that the eigenvalue st of one mode satisfies ~ o(r — 1 )/(o + 1)
and that a corresponding eigenvector u j -» [ 1,1,0]T as r -» 1.
(a) Show that the system for steady solutions may be written without
approximation as
x - y = 0,

x — y = zx — (r — l)x,

bz = xy.

Also define S = j[l, l,0]x, so that

x + y = 2S.

Now, assuming that there exist steady solutions of the form

x = c5x, + S2x2 + ..., r = l + drx + 52r2 + •■■ as d -* 0

for fixed b and <x, substitute these expansions into the four equations, and
show that x, = [1,1,0]T and r, = 0. Find r2 and x2.
(b) Next define e = (r — 1)1/2, A = [1, l,0]x/2€andu = c2f forr > l,and
assume that there exist unsteady solutions of the form

x(f) = ex^u) 4- €2x2(u) + ..., A'= F0(A) + eFfA) -1- asejO,

where a prime denotes differentiation with respect to u. Show first that the
system may be expressed as

x - y = -€2x’/o,

x — y = zx — e2x 4- e2y',

bz = xy — €2z\
164 Ordinary differential equations

without approximation. Then show that

x, = All, 1,0]T, F0(A) = -4r>t(l - A2/b).

o + l
Also findx2.
[See §8.1.]
Q5.7 Nonlinear oscillations of a conservative system. Show that the equation

d2x
+ V'(x) = 0
dr2

of §1.7 can be represented by the Hamiltonian H(p,q) = \p2 + V(q).

Q5.8 Mathieu's equation. Show that the equation,

d2x
—^ + (a — 2q cos 2 t)x = 0,
dr

where a, q are constants, is given by the Hamiltonian H{y,x) =

{(ax2 + y2) — qx2 cos 21.
Q5.9 The Henon-Heiles system. Find dp,/df, dp2/dt, dqjdt, dq2/dt for the
system with Hamiltonian H(p1,p2,^1,^2) = {(p\ + p\ + q\ + q\) +
- id-
[Henon & Heiles (1964).]
Q5.10 The Toda lattice. Find the equations of the system with Hamiltonian

H(Pi,qi) =

[This is a model of a monatomic crystal as a lattice of particles con¬

nected by nonlinear springs (Toda 1967).]
Q5.ll Finite-difference schemes of approximation of a differential equation. Sup¬
pose that

where F: IR2 —► IR2. This differential equation may be approximated by the

difference equation
x»+i = x„ + hf(h,x„)

for small h > 0 and a function f: IR x IR2 -♦ R2 such that f(/j,x) =

F(x) + o(h) as h->0. Various functions f are chosen by numerical ana¬
lysts, for example, f(/i,xn) = F(x„) (explicit Euler scheme), f(h,xn) = F(x„+1)
(implicit Euler scheme), f(h,x„) = i{F(x„) + F(xn+1)} (trapezoidal scheme),
and f = F(x„ + {hf) (implicit mid-point scheme). It can be shown by, say,
expanding f in powers of h, that area preservation of the differential equa¬
tion does not imply area preservation of the difference equation unless f is
chosen carefully.
Taking F(x) = [y, -x]T, show that if
 Problems 165

dx dy
dd7=~*

then the system is area preserving and x2 + y2 is constant on each orbit.

(a) Show that by the explicit Euler scheme

■^o+i ■*« "b y>i+i ~ y>■ ^-*n>

and x2+1 + y2+I = (1 + l»2)(x2 + y2).

(b) Show that by the implicit Euler scheme

^n + l “i“ hyn +1* .Vn +1 = yn ^*n + l>

and x2+1 + y2+t = (x2 + y2)/(l + h2).
(c) Show that by the trapezoidal scheme

+ jh(y„ + yB+1), yB+1 = y„ - $h{xn + xB+1),

the system is area preserving, and xB+I + yB+1 = x2 + y2.

(d) Show that by the implicit mid-point scheme

x„+i = *„ + Hyn ~ {hxn)/(\ + i/i2), yB+1 = y„ - h(x„ + \hyn)/(\ + ih2),

the system is area preserving, and xB+1 + yB+1 = x2 + y2.

Q5.12 Deduction of Hamilton's equations from Lagrange’s equations. You are given
that a system of particles with one degree of freedom is governed by the
variational principle,

<5 L(q, q, t) dt = 0,

where q is the generalized coordinate, q = dq/dt and L is the Lagrangian.

Deduce that

d dL dL _
df dq dq

[The solution q(t) from t = t, to t2 gives an extremum of the integral for

given function L and values <7(^1) and q{t2). Also L-T—V, where Tis the
kinetic energy and V is the potential energy of the system.]
Define the Hamiltonian H by H(p, q, t) = pq — L(q, q, t) and the new vari¬
able p, called the generalized momentum, by p = dL/dq, where q is considered
to be some function of p and q. Thence show that the differential

cH dH dH
— dp + — dq + dt = dH
cp cq ~di

dL\.. .. dLA 8La

P~si) I** p~*
for all dp, dq, dt. Deduce that
 166 Ordinary differential equations

dq dH dp dH dH dL
dt dp' dt V ~dt ~dt’

Q5.13 ABC flows. Consider the flow of a fluid with velocity distribution u: IR3 -*
R3 such that u(x) = (Bcosy + C sin z,Ccosz 4- /tsinx,v4cosx + Bsiny)
and x = (x, y, z) for given real constants A, B, and C. Show that curl u = u.
[*It follows that the flow is a Beltrami flow, a Beltrami flow being defined as
one for which curl u is parallel to u.] Show also that

tan X tan Y tan Z = — 1, A2 = B2 sin2 Y + C2 cos2 Z

at a stagnation point X = (X, Y, Z), i.e. a point where u(X) = 0. Taking A, B,

C > 0, deduce that there is no stagnation point in the flow unless there
exists an acute-angled triangle with sides of lengths A, B and C. Show that if
{X, Y, Z) is a stagnation point then so is (X + n,Y + n,Z + n).
Denote the half-open cube [0, 2jt) x [0,2rt) x [0,27r)byT3 = R3/(27tZ)3.
Then consider the flow in T3 given by the orbits for t > 0, where

dx i \
it - UW'

x(0) = (x0,y0,z0), and the components of x are reduced periodically

modulo 2n so that 0 < x0, y0, z0 < 2n. Show that the flow is volume
preserving (i.e. incompressible) and that a stagnation point is an equilibrium
point.
Show that if A = B = C # 0 then there are eight equilibrium points,
namely j7t), their cyclic permutations, (|7r,^7t,|7t) and
n). By examination of the linearized system of differential equa¬
tions, show that (57c, f 7r, is unstable.
Deduce that if B = C = 0 and A ^ 0 then

x(r) = (x0,y0 + /4tsinx0,z0 + /lrcosx0) modulo 2n.

Use this solution to determine the Poincare map of the half-open face of the
cube with z = 0 and 0 < x, y < 2n, i.e. show that if x(0) = (xB,yn,0) then the
orbit next cuts the face in the point (x„+,, yn+,, 0), where

*n + l = y„+1 = y„ + 2n tan x„ modulo 2ti.

and 0 ^ yB+, < 2n, provided that /4cosx„ > 0. Hence describe briefly the
iterates of the Poincare map according to the values of the initial point
(*o> y0. zo)-
[Henon (1966), Dombre et al. (1986).]
*Q5.14 Invariant set of a measure-preserving map and recurrence. Consider the
dx
system — = F(x), where F: -*■ R" is continuously differentiable and
dt
div F = 0 for all x. Suppose further that there exists a bounded invariant set
 Problems 167

I of the flows <j>„ and V0cl is a set of points with positive measure, i.e.
Mvo) > 0, and define V(t) = «|>,V0 for all t > 0.
Show first, by reductio ad absurdum or otherwise, that at least two of the
sets V(0), V(l), V(2),... have a point in common. Deduce that there exists
t, > 0 such that V(0) and V(f,) have a common point. Hence or otherwise
show that there exists an infinite sequence t, < t2 < ... such that t„ -* oo as
n —* <x> and V(0) and V(rn) have a common point for n = 1, 2,_Deduce
that if x(0) c I then x(f) lies arbitrarily close to x(0) for a sufficiently large
value of t. [Poincar6(1890, §8).]
. 15 Stability in the senses of Liapounov and Poincare, (a) Show that if

then the periodic solution A(f) = a cost is stable in the senses of both
Poincare and Liapounov.
(b) Show that if
d2x
—y + sin x = 0,
dr

then a periodic solution of amplitude a is stable in the sense of Poincare but

unstable in the sense of Liapounov. [Cf. Example 1.2.]
16 Stability of a periodic solution. Show that the equation

d2x dx dx
+ b + x — a + x = 0
df2 dt df

can be expressed as the system

dr d6
= b(a — r2)rcos2 0, — = — 1 + \b{a — r2) sin 26,
dt df

where x = r cos 6, dx/df = rsin0, r ^ 0. Deduce that for a > 0 there is a

solution of period 2n given by r = Ja, 9 = f0 — f.
Linearizing the system about the solution X(f) = Ja cost, and defining
r' = r — yja, show that

r'(t) = r'(0)exp{ — ab(t + ^sin2f)}.

Deduce that the limit cycle X is stable in the sense of Liapounov if, more¬
over, b > 0.
Defining the Poincare map Pa: Z -*• £ of the system by Pa{r(0)} = r(2n)
for all r(0) ^ 0, where I is the half-line 9 = 0 in the phase plane, show that
yja is an asymptotically stable fixed point of Pa for all a,b> 0.
17 Floquet theory. Consider the system

= A(f)x(f),
dt

where x is an m x 1 vector and A is a real m x m matrix of period T.

168 Ordinary differential equations

Defining O as the fundamental matrix of the system such that

d<D(t)
—^ = A(f)<D(t), 0(0) = I,

I as the m x m unit matrix, and *P(f) = <l)(t + T), show that

and hence that <D(f + T) = 0(f)C, for some nonsingular matrix C. Deduce
that C = 0(7). Assuming that the nonsingularity of C implies that there
exists R such that C = eTR, define P(f) = <D(r)e_,R and prove that P has
period T. Deduce that <D(f) = P(t)e'R.
[Note that integration over a period T determines P, and then P deter¬
mines the fundamental matrix and hence all solutions for — oo < t < oo.]
The Floquet multipliers are defined as the eigenvalues q2, q2,..., qm of C
and the Floquet exponents as the eigenvalues s,, s2,..., sm of R. Show that
qj = exp(SjT) for) = 1,2,m if the exponents are ordered appropriately.
Prove the Jacobi-Liouville formula that

detO(T) = qlq2...qm = exp<^ trace A(t)dt>,

and deduce that no Floquet multiplier is zero.

[Cf. Coddington & Levinson (1955, Chap. 3, §5).]
Q5.18 Meissner's equation. Find the Floquet multipliers of the equation,

where /(f) = 1 if 0 < t < \ and f(t) = — 1 if 5 < t < 1, and / has period 1,
for a > 0.
Q5.19 Hill’s and Mathieu's equations, (a) Suppose that x satisfies Hill’s equation.
namely

where P is a real function of period n. Use the results of Q5.17 to show that
qt,q2 are either both real or complex conjugates and that qxq2 = 1. Deduce
that there exist independent solutions xltx2 such that x/t + n) = qjx/t) for
j — 1, 2, and that if the null solution of the equation is on the margin of
stability then Xj has period n (synchronous) or 2 n (subharmonic) for j = 1 or
2.
[Cf. Coddington & Levinson (1955, Chap. 8, §4).]
(b) Seek marginally stable subharmonic solutions of Mathieu’s equation.
 Problems 169

d*x
+ (a — 2e cos 2t)x = 0,
d7

for small e by expanding

a(€) = 1 + ea{ + e2a2 + x(t,e) = x0(t) + €x,(f) + ...

as € -»0. First take x0(t) = cost and show that a, = 1, x,(t) = -|cos31.
Then take x0(r) = sin t and fuida^x,.
*Q5.20 Mathieu's equation. Show that each solution x of the equation,

d2x
—T + (a — 2e cos 2 t)x = 0,
dt

is either even or odd itself, or may be expressed as the sum of an even

solution and an odd solution.
Seeking a solution x of period n to determine the margin of stability of
the null solution, assume that x is even and
co

x(t) = A2ncos2nt
n=0

for some constants A0, A2,.... Deduce that

oAq eA2 ~ 0,

(a - 4)A2 - €(2A0 + A4) = 0,

(a - 4n2)A2n - <=(/l2(1_2 + >l2n+2) = 0 for n ^ 2.

Find those points on the margin of stability a = a2„(e) in the (e,a)-plane

on which there is an even solution of period n by assuming that a2n =
4n2 + ea„i + €2an2 + ...,x„(t) - cos2nt + ex„,(f) + €2x„2(t) + ... as e -♦ 0,
for n = 0,1,_Deduce that

a0(e) = -W + °(f3)-

a2(e) = 4 + fe2 + 0(€3),

a2n(e) = 4n2 + e2/2(4n2 - 1) + 0(e3) for n ^ 2 as e -► 0.

Q5.21 A flow. Deduce from §1.3 that if

dx
dr
then the flow map is given by 0, (x) = x/( 1 + xt) for appropriate values of x, t.
Verify that <j)s (0, (x)) = 0J+, (x). Find 071 (x), stating for what values of x, t it
exists.
 6
Second-order autonomous differential systems

Some instances of typical relaxation oscillations are: the aeolian harp, a

pneumatic hammer, the scratching noise of a knife on a plate, the waving of a
flag in the wind, the humming noise sometimes made by a water-tap, the
squeaking of a door, ... the tetrode multivibrator, the periodic sparks ob¬
tained from a Wimshurst machine, ... the intermittent discharge of a con¬
denser through a neon tube, the periodic re-occurrence of epidemics and of
economical crises, the periodic density of an even number of species of ani¬
mals living together, and the one species serving as food for the other, the
sleeping of flowers, the periodic re-occurrence of showers behind a depres¬
sion, the shivering from cold, menstruation, and, finally, the beating of the
heart.

B. van der Pol & J. van der Mark (Phil. Mag. 1928)

1 Introduction
An important class of systems of ordinary differential equations is that of
second-order autonomous systems, i.e. systems of the form

dx dv
d7 = F(x,y)’ dt = (!)

They are important because they are fundamental to much of the be¬
haviour of higher-order systems and because they may be used to illustrate
much of the behaviour simply. They are important also because they have
many applications; indeed, many mechanical, electronic, chemical and bio¬
logical phenomena have been successfully modelled by systems of the form
(1). Poincare thoroughly investigated their properties, by both analytical
and geometrical methods. In particular, he devised the use of the phase
plane, examining orbits in the (x, y)-plane, i.e. the curves with equation

dy _ G(x,y)
dx F(x,y)‘

He also looked at the local behaviour of orbits near equilibrium points,

classifying all types of these points (Poincare 1881,1882).

170
 6.1 Introduction 171

An important subclass of systems of the form (1) is the class of those for
which F(x, y) = y and therefore

(3)

Conversely, any second-order differential equation (3) can be put in the

form (1) with dx/dt = y and dy/dt = G(x,y). It is customary for computer
software to give numerical algorithms to integrate differential equations in
the canonical form (1), not (3). However, many well-known nonlinear oscil¬
lations are modelled by equations of the form (3), for example

d2x dx
^y+/W-37 + 0(x) = O (Lienard’s equation), (4)
dt

d2x 1 / dx\ dx
+ €<-[ — ) — 1 >— + x = 0 (Rayleigh’s equation), (5)
d7 [3\dtJ dt

d2x . dx
(van der Pol’s equation). (6)
d?+e(x -,)* + x=0

We shall meet these and other examples later.

It will be occasionally easier to use plane polar coordinates, such that
x = r cos 0, y = r sin 0 for r 0, rather than Cartesian coordinates. Then it
can be shown by elementary calculus that

dx dv
x— + y- (7)
df dt

so that equations (1) become

dr
= cos 9{F(r cos 0,r sin 0)} + sin 0{G(rcos 0, rsin0)}, (8)
dt

dd
r 1 cos 0 (G(r cos 0,r sin 0)} r 1 sin 9{F(r cos 0,r sin 0)}. (9)
dt

We shall sometimes use the equations in polar form.

The qualitative character of all the solutions of system (1) can be epito¬
mized by a sketch of the solution curves in the (x,y)-plane, i.e. by a phase
portrait. It is desirable that the sketch includes one each of all the topologi¬
cally different kinds of orbit. In principle this requires the integration of
equation (2). In practice sketching phase portraits is partly an art based on
experience. The practical rules are to seek any simple symmetry the system
has, to find all the equilibrium points, to find the local behaviour of the
172 Second-order differential systems

orbits near each equilibrium point (by use of the solutions of the linearized
system), to identify any simple solution curve (e.g., a coordinate axis), and
to make such other deductions as are apparent (e.g., about orbits at infinity).
Also equation (2) gives the slopes of the curves everywhere. There should
come a stage of this process (for textbook and class examples, if not for a
complicated example from real life, which requires numerical integration)
when these details can be synthesized in your mind so that a global picture
of the solution curves is understood, perhaps in a flash of realization. It
must be added that the advent of accessible and easy numerical programs
with simple graphics on cheap computers has already made computation a
valuable complement to these traditional methods.
The rest of the chapter covers aspects of this problem, substantiating
some of the details which make up a phase portrait; the examples provide a
little experience of sketching phase portraits. The local behaviour of the
orbits near an equilibrium point is found in §2 by explicit solution of the
linearized equations. The Liapounov direct method of proving stability or
instability of an equilibrium point is given in §3. In §5 limit cycles, which
correspond to closed curves in the phase plane, are explained and the
Poincare-Bendixson theorem is given to prove their existence. Some asym¬
ptotic methods to give quantitative results are introduced in §4 and §6. It
will become increasingly apparent, in sketching phase portraits, that the
equilibrium points and the limit cycles are very important in determining
the global pattern of the orbits.

2 Linear systems
This section is a review of the linear theory, some of which has already been
used in Chapter 1.
Suppose that the system (1.1) has a point X of equilibrium, i.e. that
F(X) = 0, where X = [X, T]T and F = [F, G]T. Without loss of generality
we may translate the equilibrium point to the origin so that X = 0. Then
F(0,0) = G(0,0) = 0. To consider orbits of (1.1) near 0, assume that F is
continuously twice differentiable, so

F{x,y) = ax + by + 0{x2 + y2),

G(x, y) = cx + dy + 0(x2 + y2) as x, y -> 0,

where a = \_dF/dx]0, b = [dF/dy]0, c = [5G/dx]0, and d = [(dG/dy]0. Then

the behaviour of the system (1.1) may in general be approximated locally
by that of the linearized system
 6.2 Linear systems 173

(1)

where the column vector x = [x, y]T and the 2 x 2 matrix J = We

shall examine this approximation more critically later. For the present we
shall classify the types of equilibrium points of the system (1) according to
the properties of J.
The linear system (1) with constant coefficients can be solved in general
by using the method of normal modes, i.e. by seeking solutions \(t) = es'u
for a constant vector u. Then

Ju = su,

i.e. s is the eigenvalue of J belonging to eigenvector u. Therefore

0= “ — /
c a — s

= s2 - ps + q, (2)

where p = trace J = a + d and q = det J = ad — be. Therefore s = sl or

s2, where

si,s2 = HP ± (P2 ~ 4<?)1/2}- (3)

Therefore, if p2 # 4q, the general solution of (1) is

x(t) = Cx exp(s1t)^1 J + C2exp(s2t)^2J (4)

for arbitrary constants C, and C2, where the eigenvalue Sj belongs to eigen¬
vector u7 = [Uj, VjY for j = 1,2, i.e. where

auj + bvj = SjUj, cuj + dvj = SjVj.

This explicit solution, and the solution for the case p2 = 4q of a double
eigenvalue, enable us to classify all the equilibrium points according to the
topology of the phase portraits of neighbouring orbits. The results are as
follows.

(i) Node: p2 > 4q and q > 0. Here sx, s2 are real, distinct and of the same
sign. We take sx> s2 without loss of generality. Therefore

x(f) ~ Q exp(s1t)[u1,i;1]T as t -* oo
174 Second-order differential systems

Fig. 6.1 (a) Sketch of the phase portrait of a stable node. On reversal of the
arrows of time, we get the phase portrait of an unstable node. (b) Sketch of the
phase portrait of a saddle point.

if Cx # 0, and so y/x -*■ v1/u1 as t -*• oo. This describes what is called a node,
stable if < 0, and so if p < 0, and unstable if p > 0. See Fig. 6.1.

(ii) Saddle point: p2 > 4q and q < 0. Here sl5 s2 are real and of opposite
signs, so we may take sx > 0 > s2 without loss of generality. The orbits
near 0 resemble hyperbolae.

(iii) Focus or spiral point: p2 < 4q and p ^ 0. Here st and s2 are a com¬
plex conjugate pair with non-zero real part, so we may take st = s2 =
i{p + i(4q - p2)1/2}. Then

\{t) = Cep,/2[cos(/?f + y),K cos{fit + y -f <5)]T,

where C, y are arbitrary real constants, p = {q - ip2)1'2, and K, S are real

constants determined by a, b, c and d. Therefore the origin is a stable point
of equilibrium if p < 0 and unstable if p > 0. See Fig. 6.2.

These are the generic, i.e. typical, cases, for which no two eigenvalues are
equal and the real part of no eigenvalue is zero. If the real parts of all
eigenvalues are non-zero then the equilibrium point is said to be hyperbol¬
ic. A hyperbolic point of equilibrium can be shown to be structurally stable
in the sense that a small well-behaved perturbation such as weak non¬
linearity will not change the topology of the orbits near the point, so that a
stable hyperbolic point will not thereby become unstable nor an unstable
hyperbolic point become stable. There are a few special cases that remain
to be described.
 6.2 Linear systems 175

Fig. 6.2 (a) Sketch of the phase portrait of a stable focus. For an unstable focus
the arrow is reversed. (b) Sketch of the phase portrait of a stable improper
node. For an unstable improper node, the arrows are reversed.

(iv) Improper node, inflected node, or degenerate node: p2 = 4q and q > 0.

Here s2 = sx is real. The solution can in general be shown to give a limiting
form of a node because the lines y = v1x/u1 and y = v2x/u2 coincide. The
node is stable if p < 0 and unstable if p > 0. See Fig. 6.2.

(v) Centre or vortex: p2 < 4q and p = 0, i.e. p = 0 and q > 0. Here sl5
s2 = + i<?are purely imaginary. Then

x(f) = C[cos{qt + y),Kcos(qt + y + <5)]T’

where C, y are arbitrary real constants and K, S are real constants deter¬
mined by a, b, c and d. The orbits are ellipses with centre at the origin. See
Fig. 6.3.

(«) (b) (c)

Fig. 6.3 (a) Sketch of the phase portrait of a centre. (f>) Sketch of the phase
portrait of a degenerate node for p < 0 in case (vi). The arrows are reversed if
p > 0. (c) Sketch of the phase portrait of a stable sink with a < 0. The arrows
are reversed for a source with a < 0.
176 Second-order differential systems

(vi) Degenerate node or improper node: p2 > 4q and q = 0, i.e. p2 > 0 and
q = 0. This is a limiting case of (i) above, for which we may put = 0 and
s2 = p. Therefore

x(r) = C1[h1,u1]t + C2ep,[u2,v2y,

and the orbits are straight lines parallel to y = v2x/u2, approaching the
line y = ulx/v1 as t -*■ oo if p < 0 and leaving it if p > 0. See Fig. 6.3.

(vii) Source, sink, proper node or star point: p2 = 4q, q>0,b = c = 0, and a =
d* 0. Here dx/dt = ax, dy/dt = ay. Therefore x(t) = x0ea' and y(t) = y0eat. This
gives radial orbits, the origin being stable (a sink) if a < 0 and unstable (a
source) if a > 0. See Fig. 6.3.

All the results are summarized in Fig. 6.4. The generic cases of an equilib¬
rium point are those represented by points not on the parabola p2 = 4q,
the p-axis or the q-axis. Different authorities choose slightly different sets
of names for the types of equilibrium points.
A node, a saddle point and a focus are defined for a nonlinear system

ASYMPTOTICALLY STABLE q UNSTABLE

stable nodes stable foci unstable foci / unstable nodes

saddle points saddle points

Fig. 6.4 Summary in the (p, <?)-plane of the classification of points of equilibrium
at x = 0, y = 0 for the linear system dx/dt = ax + by, dy/dt = cx + dy, where
p — a + d and q = ad — be. A disc • denotes that the origin is stable, and a
circle o that it is unstable.
 6.2 Linear systems 177

similarly, according to the topological character of the orbits near the

equilibrium point; thus a node is an equilibrium point which all the nearby
orbits in the plane approach (or leave) and all except possibly two ap¬
proach (or leave) in the same direction, a saddle point is an equilibrium
point which is approached by two orbits as t -*■ oo and two as r -*■ —oo,
a focus is an equilibrium point which each nearby orbit in the plane
approaches (or leaves) along a spiral, and a centre is an equilibrium point
surrounded by closed orbits. A stable node and a stable focus are at¬
tractors, and an unstable node and an unstable focus are repellers. A
saddle point is neither an attractor nor a repeller because, although almost
all orbits in its neighbourhood leave the neighbourhood, two orbits ap¬
proach the equilibrium point. A centre, although stable, is not asymptoti¬
cally stable and therefore is not an attractor because orbits in its neigh¬
bourhood do not approach the centre.
The nonlinear system

^ = Jx + m

and the linearized system have in general the same topological character
of orbits near 0 if \ is well-behaved and % = o(|x|) as x -* 0 (cf. §3 and
Coddington & Levinson (1955, Chap. 13)). So the linear theory gives the
orbits near an equilibrium point. It is for this reason that the linear theory
is so important in the nonlinear theory. The exceptions to the general rule
arise when the equilibrium point is not hyperbolic, i.e. when p2 = 4q, p = 0
or q = 0, because weak nonlinearity may change one kind of equilibrium
point into another when the linear theory is on the margin between two
kinds of point. For example, weak nonlinear dissipation may change a
centre into a stable focus. We have met and will meet similar ideas in other
examples and problems.

Example 6.1: a focus. If

dx dv
— = ax — cy, — = cx + ay
dr dr

then p = a + d = 2a and q - ad - be = a2 + c2. Therefore p2 - 4q =

— 4c2 < 0. Therefore 0 is a focus, stable if a < 0 and unstable if a > 0. In
fact, each orbit is an equiangular spiral with r(r) = r0ea' and 9{t) = 90 + cr,
where r, 6 are plane polar coordinates such that x = r cos 9 and y = r sin 9.
(When c = 1, this becomes the linearized system of §1.6.) □
178 Second-order differential systems

3 The direct method of Liapounov

We shall next describe the direct method of Liapounov (1892), a powerful,
but not omnipotent, method of proving the stability or instability of an
equilibrium point of a nonlinear system of differential equations. It is
sometimes called the second method of Liapounov (we shall not describe the
first method, which is almost forgotten). First an example will motivate the
method.

Example 6.2: the energy method. Consider the linear system,

dx
— = cx + ay for a < 0, (1)
dt

of Example 6.1 from a different point of view, defining H(x) = ^(x2 + y2).
Therefore

= (ax2 — cxy) + (cxy + ay2)

= 2aH

^0, (2)

with equality if and only if x = y = 0. Therefore

H(x(t)) = H(x( 0))e2a'

-*• 0 as t -*• oo

for all x(0). Therefore

x(t) -*• 0 as t -*• oo,

i.e. the origin is a globally asymptotically stable point of equilibrium of

system (1).
Next consider a nonlinear perturbation of the system (1), i.e.

dx dy
— = ax-cy + f(x,y), — = cx + ay + rj(x,y) (3)

for a < 0, where f, q are continuously differentiable functions such that

{£(x,>>)}2 + M*,y)}2 = 0{(x2 + y2)2} as x -► 0.

 6.3 The direct method of Liapounov 179

Now we find that

dH „
— = 2aH + x£ + yrj

< aH

for sufficiently small x / 0 in order that aH + x£ + ytj < 0. (Remember

that a < 0.) Therefore H < H(0)ea\ and so H(x(t)) -*■ 0 as t -*■ oo for all
small x(0), and the origin is asymptotically stable. □

Now let us come to the direct method itself. We shall apply it to the
system

~ = F(x), (4)

where F: IRm -*■ IRm and F(0) = 0. The method is essentially the same for all
m, so we shall give it for general m, but emphasize the case m = 2. We shall
show that if a scalar function H with special properties may be found then
the origin is a stable point of equilibrium. Finding the function is a matter
of experience, and of trial and error, but the function is a generalization of
the energy of a mechanical system, and the argument we use is similar to
the one that if energy is dissipated and there is no available potential ener¬
gy then an equilibrium point is stable.
First we need a definition. A function /: IRm -+ IR is positive definite (or
semidefinite) if /(x) > 0 (or ^ 0 respectively) for all x # 0 and /(0) = 0.
Now we may state the theorem (or, rather, three theorems).

Liapounov’s theorem. If 0 is an equilibrium point of system (4) and there

exists a function H: lRm -> (R such that

(i) H and its partial derivatives are continuous,

(ii) H is positive definite, and
(iii) (a) — F • grad H is positive semidefinite,
(b) — F • grad H is positive definite, or
(c) F • grad H is positive definite,

then the equilibrium point is (a) stable, (b) asymptotically stable, or (c)
unstable respectively.

We shall not prove the theorem, which requires some care in case (a);
but the essential idea is that dH/dt = grad H ■ dx/dt = F • grad H and so
that H increases or decreases monotonically with t if F • grad H is positive or
180 Second-order differential systems

negative. Because H is positive definite, x decreases or increases according

to whether H increases or decreases everywhere near 0. The function H is
called a Liapounov function. The only difficulty of the method lies in finding
H. Even if a Liapounov function exists, it may be very hard to find.

Example 6.3: a Liapounov function. Consider the system

dx dy 3
-x- 2y2.
dt d7 = xy~y

By inspection, the origin is an equilibrium point. For the Liapounov func¬

tion try H(x, y) = y(x2 + ay2) for a value of a to be determined. Then a > 0
in order that H is positive definite. Therefore

dH
— -
17 • grad
= F AU
H
dt

= (—x — 2y2)x + (xy - y3)ay

= -(x2 + 2y4),

on choosing a = 2,
<0,

with equality if and only if x = y = 0. Therefore the origin is an asym¬

ptotically stable point of equilibrium, o

Example 6.4: a Hamiltonian system. Consider the system

dp _ dH dq dH
dt dq ’ dt dp ’

where p, q e (R", the Hamiltonian function is of the form

H(p,q) = T(p,q) + F(q),

K(0) = 0, and Tis a positive definite quadratic form in p. (This represents a

mechanical system with kinetic energy T and potential energy V) Suppose
further that q = 0 is a simple minimum of V. Then p = q = 0 gives an
equilibrium point. Therefore H is positive definite near 0 and dff/dt = 0,
i.e. dH/dt is positive semidefinite. Therefore the Hamiltonian H serves as a
Liapounov function and the origin is a stable point of equilibrium.
Lagrange (1788, Part II, Section 5.15) is famous for (among other things)
first proving that if V has a minimum then equilibrium is stable; Liapounov
(1892, §25) proved that otherwise the equilibrium is unstable. □
 6.4 The Lindstedt-Poincare method 181

4 The Lindstedt-Poincar6 method

There are many asymptotic methods to find approximately the behaviour
of weakly nonlinear oscillations. We shall introduce one in this section,
only describing it by an example. A canonical equation for oscillations
of a conservative system is

(1)

where / is a continuously differentiable nonlinear function such that

/(0) = /'(0) = 0, and e is a parameter. (Note that any equation of the
form d2y/dt2 = g(y), with g(Y) = 0 and g'(Y) < 0, may be reduced to the
form (1). To see this, please remember from §1.7 that a general second-
order equation representing a conservative system may be expressed in
the form d2y/dr2 = g(y) for a well-behaved function g, and that there may
be oscillations about a stable point Y of equilibrium if g(Y) = 0 and
g'(Y) < 0. On our defining x = y — Y, t = {—g'{Y)}112t and e/(x) =
x + g(x + Y)/{— g'{Y)}, this leads to equation (1) with a function / such
that f{x) = 0{x2) as x-* 0.) When e is small there is weak nonlinearity.
Recall that when e = 0 there is simple harmonic motion with period 2n
independently of the amplitude of the oscillation. We anticipate, from
experience of §1.7, that when e is small there are periodic solutions whose
periods depend on e and the amplitude.
On first thoughts, one might seek to solve this problem by expanding
the solution
x(t, e) = x0(f) + exx (t) -I- e2x2{t) H- as e -» 0, (2)

substituting this regular expansion into equation (1), equating coefficients

of e°, e1, e2,..., and solving the resulting equations for x0, xl5 x2,..., in
turn. This method, however, may give a nonuniformly valid approxima¬
tion, as € -»0, over an infinite interval of time. To understand this non¬
uniformity, a simple linear example will suffice.

Example 6.5: nonuniformity of a regular perturbation of simple harmonic

motion. Suppose that

(3)

and

x = a, — =0 at t = 0. (4)
dt
182 Second-order differential systems

Then substituting expansion (2) into equation (3) and initial conditions (4),
and equating coefficients of e° gives

d2x0
+ x0 = 0
df2

and

x0 = a, at f = 0.

Therefore

x0(f) = a cost.

Next, equating coefficients of e, we find that

d2xx
+ xx = — 2x0 = —2a cost
df2

and

xi = 0, =0 at t = 0.
df

Therefore

xt(f) = —at sin t.

This gives

x(t) = ajcosf — ef sinf + 0(e2)} ase->0. (5)

We now see that, however small e is, exj(f) is as large as x0(t) after long
enough a time (when ef 1). So the solution (5) may converge uniformly to
the exact solution x(f, e) as e -*• 0 if f belongs to a given finite interval, but
not if t belongs to an infinite interval.
In this example of simple harmonic motion we can see in detail what
is happening, because we may at once write down explicitly the exact
solution.

x(f,e) = acos{(l + e)f}

= a cos ef cos f — a sin ef sin f.

The standard power series of cosef and sinef give all terms in the expan¬
sion (2), the first two of which can be recognized in approximation (5). □
 6.4 The Lindstedt-Poincare method 183

Poincare (1893, §123) recognized that this nonuniformity could be re¬

solved by expanding the frequency of the periodic solution as well as the
solution itself. Although he acknowledged Lindstedt’s (1883) use of this
idea, the method is often named after Poincare alone. The idea serves to
reduce the problem to one over a finite interval of time, namely the period
of the oscillation, and so to avoid the nonuniformity of the limits as e -> 0
and t -* oo. We shall describe the method by an example.

Example 6.6: a nonlinear spring. The equation

(6)

governs a soft spring if e > 0, Hooke’s law if e = 0, and a hard spring if

e < 0. It has energy integral

The equilibrium points are X = 0 for all e, and X = ± 1/^/e for all € > 0.
So there is a pitchfork bifurcation at x = 0, l/e — 0 in the (l/e,x)-plane.
The point x = dx/dt = 0 can easily be seen to be a centre, and x = ± 1 /yjc,
dx/dr = 0 a saddle point in the phase plane. The phase portraits are
sketched in Fig. 6.5. The separatrices have equation dx/d t =

(«)
Fig. 6.5 Sketch of the phase portrait of the nonlinear spring (6) in the (x,dx/dt)-
plane. (a) The soft spring e > 0. (b) The hard spring e < 0.
184 Second-order differential systems

±(je)l,2(x2 — 1/e). Motions of small amplitude correspond to closed

curves for all e, and so are periodic. The period, given by equation (1.7.8),
depends upon both e and the amplitude (which determines the energy E) in
general. However, if e = 0 there is simple harmonic motion of period 2n for
all amplitudes.
This sets the scene. Now we shall describe the Lindstedt-Poincare
method to find the periodic solutions for small e. Define co as the unknown
frequency of a solution, so the period is In/w, and define t = cot. Then we
may express equation (6) as

co2 + x — ex3 = 0 (7)

without approximation. The chosen periodic solution will also satisfy the
condition of periodicity,

x(t + 2n, e) = x(t, e) for all t. (8)

We apply the method by expanding the frequency

co = (o0 -1- ecoj + €2co2 H- (9)

as well as the solution

x(r,e) = x0(t) + ex^r) + €2x2(t) + ••• ase->0. (10)

Further, we may define

u = x(0,e) (11)

where, by translation of time, we take

dx
^=0 atr=0 (12)

without loss of generality. Then a is usually the amplitude of an oscillation.

Now it is straightforward to substitute both expansions (9) and (10) into
equation (7), periodicity condition (8), and initial conditions (11) and (12).
Equating coefficients of c°, we find

7 d2x0
“° d7r + x» = 0’

X0(t + 2a) - Jt0(r) for all T, x0 = a, ^ = 0 at t = 0.

dr
The differential equation has general solution
 6.4 The Lindstedt-Poincare method 185

*o(*) = a0cos(T/co0) + b0s in(t/a>0)

for some constants a0, b0. Then the periodicity and initial conditions give

CO0 = 1, X0(t) = flCOST.

Equating coefficients of e, we find that

d2*i , ~ d2x0
+ Xj = — 2 to, + Xq
dr2 dr

= (2o>, a + fa3)cos t + fa3 cos 3t, (13)

dxj
Xj(t + 27r) = Xj(t) for all r, X! = —— = 0 atr = 0.
dt

If the term in cos t were not to vanish, then there would be a secular term
j(2cujfl + |a3)Tsinr in the particular integral of xls and so exl would be
unbounded and xY could not satisfy the periodicity condition. It follows
that the coefficient of cost on the right-hand side of the equation for xt
does vanish, and therefore

tt>i = - fa2.

Therefore
Xj(t) = al cost + bx sinT — ^u3cos3t

for some constants ax and b1. The initial conditions for X! determine them,
finally giving

Xi(t) = j2U3(cost — cos3t).

This gives the uniformly valid approximation,

x(f,€) = a cos cot + j2€a3(coscot — cos 3cot) -I- 0{e2a5)

co = 1 — |ca2 + 0(c2aA) as e -*■ 0.

We may find higher approximations by proceeding to find co2, x2, co3,... in

a similar way, but the leading corrections to simple harmonic motion are
already found. (See also Q1.17.)
That the coefficient of cos t on the right-hand side of equation (13) van¬
ishes can be deduced alternatively by use of Lagrange’s identity, namely

f2* T dv dul2*
J. <^-»Md< = [u--C-Jo 04)

for all continuously twice differentiable functions u, v, where the self-

186 Second-order differential systems

adjoint linear operator L is defined by L = d2/dr2 + 1. The identity can be

deduced easily by integrating by parts. It becomes, on putting u = cos t,
v = x\ and using equation (13),

f2* dx.
cost^c^a + fa3)cost + fa3cos3t}dr = cost——I- Xj smt
dr

i.e.
j(2wl a + fa3)= 0,

on using the periodicity condition for xl. This gives the equation for w1 as
above. □

It should be noted that the Lindstedt-Poincare method is applicable to

periodic solutions of many more differential systems than those of the form
(1). Indeed, it is applicable to partial as well as ordinary differential systems
of many kinds and orders, when a linear oscillation is well enough known
to perturb. In particular the method is suitable for weakly nonlinear solu¬
tions of equations of the form

d2x
+ x = e/(x, dx/df)

in the limit as e -*■ 0.

5 Limit cycles
We have seen that the system of equations

dx dv
dt=F(X,y)’ dt = G(x’y) (1)

may have not only equilibrium points but also periodic solutions which
model nonlinear oscillations; and that a periodic solution is represented by
a closed curve in the phase plane. Further, damped or negatively damped
oscillations may tend to a periodic solution; this limiting periodic oscilla¬
tion is called a limit cycle. It is an attractor. An example was given in §1.6.
Note that an orbit of an autonomous second-order system (1) for well-
behaved F, G cannot cross itself except at an equilibrium point, where F and
G both vanish. Also, because the system is well-behaved, all the orbits near
each point except an equilibrium point are all in the same direction F =
(F, G), and the equilibrium points are only of the types we have described; so no
orbit can have a very complicated pattern near a point. By geometrical
intuition it follows that each orbit in the phase plane goes to either (a)
 6.5 Limit cycles 187

infinity, (b) an equilibrium point (which, like a saddle point, need not be
stable), (c) a point of itself (when it is a periodic solution) or (d) a limit cycle.
This is the essence of the proof of the Poincare—Bendixson theorem, which we
shall only state. It is a theorem effectively showing that two-dimensional
autonomous systems do not have chaotic solutions.

The Poincare-Bendixson theorem. If D is a closed bounded region of the

(x, y)-plane, and a solution of system (1) for well-behaved F is such that
x(t) e D for all t ^ 0, then the orbit either is a closed path, approaches a
closed path as t -*■ oo, or approaches an equilibrium point.

The theorem shows that the classification of attractors of well-behaved

autonomous two-dimensional systems is complete, there being no other
attractors than equilibrium points and limit cycles. The remainder of the
theory concerns the number, nature and location of the equilibrium points
and limit cycles, and thence the phase portrait. Indeed, Hilbert’s sixteenth
problem concerns the number and properties of limit cycles for system (1)
when F, G are polynomials. Some deep problems remain open to this day,
even when F, G are quadratic. However, many useful properties of limit
cycles are elementary, and we shall introduce some in this and the next
section, mostly by use of examples.

Example 6.7: proof of the existence of a limit cycle. Suppose that

£=x~y~x+2>'2)> Tt=x+y~ +y
It is helpful to use plane polar coordinates r, 6. Then equations (1.7) give

7dd
r~r = x{x + y - y(x2 + y2)} - y{x - y - x(x2 + 2y2)}
dt

x2 + y2 + xy3

r2 + \rA sin2 6 sin 29,

dr
x{x - y - x(x2 + 2y2)} -l- y{x + y - y(x2 + y2)}
rdf ~

x2 + y2 - (x2 + y2)2 - x2y2

r2 -I- r4(l + £ sin 26).

Therefore dr/dt > 0 for all 6 if r < r, and dr/dt < 0 for all 6 if r > r2, where
rt = 2/y/5 and r2 = 1. Therefore if x(0) e D, where D is the annulus defined
188 Second-order differential systems

by D = {x: rx ^ r r2}, then x(r) e D for all t > 0. Further d0/df ^ 0 in D,

so there is no equilibrium point in D. Then the Poincare-Bendixson theo¬
rem gives (at least) one limit cycle in D. □

* Example 6.8: weakly nonlinear analysis of a Hopf bifurcation. The onset

of a limit cycle at a Hopf bifurcation can be analysed asymptotically by use
of the Lindstedt-Poincare method. To illustrate this we take the system

dx 2 dy 2
— = —y + ax 4- xyz, — = x + ay - x . (2)
dr dr

It can be seen by inspection that the origin is an equilibrium point, and its
linearized equations are the same as those of §1.6. There we found the onset
of instability as a increased through zero, with eigenvalues s = ± i at a = 0.
So here also we anticipate the existence of a limit cycle of period 2n/co such
that o) -*• 1 and the amplitude of the cycle vanishes as a -* 0 from either
above (supercritical bifurcation) or below (subcritical bifurcation).
To find the limit cycle asymptotically as a -*■ 0, define r = cot, although
co is not known yet, and deduce without approximation that

cox' = —y + ax + xy2, coy' = x + ay — x2, (3)

where a prime denotes differentiation with respect to r. The periodicity

condition is that

x(t + 27i, a) = x(t, a) for all t, (4)

where x = (x, y). It is convenient to translate the origin of time, if necessary,

so that x(t) attains a maximum at z = 0; this gives, without loss of generali¬
ty, a ‘maximum’ condition that

x'(0,a) = 0. (5)

Next assume that

x(t, a) = a1/2x1/2( t) + ax^x) + a3/2x3/2(r) + ■■■,

co(a) = co0 + al/2col/2 + acol H- as a -+ 0.

The appropriateness of this ansatz to represent the true behaviour of the

limit cycle for small values of a is essential for its success; experience of
the Hopf bifurcation of §1.6 might lead to the choice of the ansatz, but
the self-consistency evident after the calculation is more convincing. Now
equate coefficients of a112, a, a3'2,... by turn in equations (3), (4), (5).
The coefficients of a1/2 give
 6.5 Limit cycles 189

0ioxui + y 1/2 — ~ xm — 0,

x i/2 (t + In) = x1/2(t) for all r, x'1/2(0) = 0.

On eliminating y1/2, it follows that

Lx1/2 = 0,

where the linear differential operator L = cood2/dt2 + 1. Therefore

*l/2(*) = ^1/2 COS(t/a)0 + <5l/2)

for some constants A1/2, Sl/2 of integration. The periodicity condition now
gives

co0 = 1,

and the maximum condition gives <51/2 = 0. Therefore L = d2/dt2 + 1,

x1/2(t) = All2cosx,

and

yinW = ~xi/2(*) = A.\a sint,

where A1/2 has to be determined later.
Next, coefficients of a give

x'l + = — wl/2xl/2» y'l ~~ X1 ~ ^1/2^ 1/2 — xl/2>

x,(t + 2n) = Xj(t) forallt, xi(0) = 0.

Therefore

Lxj — tu1/2(y1/2 x1/2) + x2/2

= 2co1/2/l1/2cost + i/42/2(l + cos2t).

Therefore

Xj(t) = Ai cos(t + <5t) + ctf1/2/41/2Tsinr + A\a(^ — £cos2t)

for some constants Alt of integration. The periodicity condition implies

that the secular term vanishes and therefore

t^l/2 — 0>

and then the maximum condition implies that = 0. Therefore

Xj(t) = /IjCOST + A2m(\ - £cos2t),

y,(r) = —xi(i) = /tjsinr - \A\l2s\nlx.

 190 Second-order differential systems

Next, coefficients of a312 give

x3/2 + ^3/2 = ~0Jlxl/2 + Xl/2 + Xl/2-Vl/2>

y 32
' / ,~ x3/2 = — ^iyi/2 + .V 1/2 + 2x1/2X1,

x3/2(t -I- 2tt) = x3/2(t) for all r, x;/2(0) = 0.

Therefore

Lx3/2 = tOi(yi/2 + x1/2) + x1/2 — ^ 1/2 + xi/2yf/2 + 2x1/2y1/2y1/2 + 2x1/2xl

= 2Al/2((o1 cost — sint) + /l1/2/li(l + cos2t)

+ A31/2( — | sint + |sin3t + § cost — £cos3t).

In order that x3/2 has period 2rc there must be no secular term in its particu¬
lar integral, and therefore the coefficients of sin t and cos t on the right-
hand side of the above equation must vanish, i.e.

— 2>41/2 — £A\j2 = 0, 2Al/2col -I- ^A\/2 = 0.

Therefore

A f/2 = — 8, (ol =

The fact that A\j2 < 0 implies that there is a real periodic solution only
when a < 0, so that

cost
x(t,a) ~ ( — 8a)1/2
sin t

a>(a) = 1 + a + o(a) as a f0,

i.e.

x(f, a) ~ (— 8a)1/2 cos art, y(t,a) ~ (~8a)1/2 sin art asa|0.

These results give the leading asymptotic properties of the limit cycle near
the subcritical Hopf bifurcation at a = 0. Further properties can be found
by solving for x3/2, co3/2, x2,..., in turn. □

6 Van der Pol’s equation

This section is devoted to a famous equation of nonlinear oscillations,
which will serve as an example of a limit cycle and as an example for
explaining some asymptotic methods. In modelling an electrical circuit
with a thermionic valve, i.e. a vacuum tube, van der Pol (1926) derived an
 6.6 Van der Pol’s equation 191

equation of the form

d2x /2 dx
^j- + e(x2- 1)—+ x = 0. (1)

It may be rewritten as the system

dx dv
= v, = —x — e(x2 — l)y; (2)
dr dr

or, on using equations (1.8), (1.9), as

^ = — e(r2cos20 — l)rsin20, ^ = — 1 — e(r2 cos2 d — I)cos0sin0.

dr dr
(3)
Note first that equation (1) is of invariant form under the transforma¬
tions (a) r -» - r and e-* -e, and (b) x -» -x. It follows that the set of all
orbits is similarly invariant, although an individual orbit need not be. Also
the reverse of sign of e will reverse the direction of time, so we may consider
6^0 without loss of generality (remembering that damping for e < 0 cor¬
responds to negative damping for € > 0).
Rayleigh (1883, 1894, Vol. I, §68a) modelled some nonlinear vibrations
by an equation of the form

d2u
+ € + u = 0. (4)
dr2

On differentiation with respect to t, this becomes

d2« du
dt1 + dt “ °'

We see, on identifying x = du/df, that equations (4) and (1) are equivalent.
If e = 0, van der Pol’s equation (1) gives simple harmonic motion of
period 2n, with general solution

x(t) = acos(f - t0) (5)

for arbitrary real amplitude a and phase t0.

The only equilibrium point of equation (1), i.e. system (2), is at the origin
of the (x, y)-plane. It is an unstable focus for e > 0 (but a stable focus for
e < 0 and a centre for e = 0).
System (3) gives

— = — c(x2 - l)r sin2 6.

dt
192 Second-order differential systems

Therefore, when e > 0, the damping is positive where x2 > 1 but negative
where x2 < 1. So large amplitudes are damped ‘on the whole’, although
small ones are amplified. We anticipate that there is a stable limit cycle.
Careful use of the Poincare-Bendixson theorem leads to a proof that there
exists a limit cycle, and it is unique. Indeed, the result is a special case of
a more general theorem for Lienard’s equation, which we state without
proof.

Theorem. The equation

d2x „ , dx
dF + /WdF + 9W = 0

has a periodic solution which is unique (up to translations of r), and this
solution is asymptotically orbitally stable, if

(a) /, g are continous,

(b) / is an even function,
(c) F(x) < 0 for 0 < x < a, F(x) > 0 and is increasing for x > a (so
F(x) = 0 only at x = 0, ±a), where F is defined by F(x) = j£/(u)du,
(d) g is an odd function and x^(x) > 0 for all x # 0.

Next we shall proceed to find this limit cycle for van der Pol’s equation
by asymptotic methods, first for small e and then for large e.
We shall find the limit cycle as e -► 0 by the method of averaging. This
powerful asymptotic method in the theory of weakly nonlinear oscillations
is due to Kryloff & Bogoliuboff (1943), who modified a method of van der
Pol (1922). We shall first introduce the method of averaging briefly by a
digression on a linear equation which we can solve explicitly for all e.

Example 6.9: damped simple harmonic motion. Consider the equation

d2x . dx
+ 2e — + x = 0.
dF dr

Let v = dx/dr and take polar coordinates in the phase plane, i.e. the (x, u)-
plane. Then, without approximation, we deduce from equations (1.8), (1.9)
that

dr
— 2 er sin2 6, -j- = — 1 — c sin 29.
dr dr

It is now evident that r is nearly constant and d0/dr nearly -1 when e is

small, so the motion is nearly simple harmonic with a circular orbit with
 6.6 Van der Pol’s equation 193

centre O in the phase plane. So if the angle is 9 at time t, then the time is
approximately t + 2n at angle 9 — 2n. However, dr/dt oscillates with 9 in
each circuit around O and these oscillations have a non-zero mean, so that
their cumulative effect may change r by an amount which is not small after
a long enough time (of order e_1). We can approximate these cumulative
effects for small e by averaging the right-hand side of the equation for dr/dt
around a typical circuit. Accordingly define

o(t) =
* r(t + s)ds ds,

where the dummy variable s of time increment is such that the angle goes
from 9 to 9 — 2n in the interval of integration,

f° , , d9 IC° d9
J2/(f + S)d0/dS/L d0/ds

= 1 + °(€) P r(t + s)d9 as e -+ 0,

2n Jo

because df?/ds + 1 = 0(e). Therefore

da _ 1 + 0(e)
df 271 j: dr(t -I- s)
ds
d9

1 + 0(e)
2n

=-ea + 0(e2)
r<- 2er sin2 9)d9

ase->0,

because r = a + 0(e). Similarly, we define

a)(t) = Q
d9(t + s)
ds J
ds I (J) ds

=±r2" d
2" Jo
9(t + s)
ds
d 9 + 0(e2) as£->0.

Therefore, on averaging,
i P*
ID + 1 (—ersin20)d0 + 0(e2)
2n Jo
= 0(e2) ase-+0;

it so happens that the first-order correction to a> is zero. Thus da/dt =

— ea, + 1 = 0 to leading order; and so a{t) ~ a0e “, 9(t) -»t0 - t as
194 Second-order differential systems

e -* 0. Therefore

x(t, e) ~ a0e el cos(t — t0) as e ->• 0.

Compare this with the exact solution. For — 1 < e < 1, it is simply the
solution of a focus in §2:

x(t,e) = a0e-tfcos{(l - e2)1/2(t - t0)},

v(t,e) = -a0e_€'[(l - €2)1/2sin{(l - e2)1/2(f - t0)}

+ ecos{(l - e2)1/2(t - t0)}],

for which

0(f,e) = — arctan[(l — e2)1/2tan{(l - e2)1/2(t — t0)} -I- e]. □

The method of averaging is suitable to find asymptotically the weakly

nonlinear oscillations not only for equations of the form

d2x
+ w2x = ef(x,dx/dt,e),
dr

but also for ordinary differential equations of higher order, and to find
asymptotically a wide variety of weakly nonlinear waves for partial differ¬
ential equations.
Let us return to van der Pol’s equation. We again anticipate that when
0 < e « 1 the limit cycle is close to some solution (5) of the simple harmon¬
ic motion found for e = 0; then x = r cos 0 and v = r sin 0, where r and
d0/dr are approximately constant. Equations (3) suggest that dr/dt = 0(e)
and dO/dt = — 1 + 0(e) as e -♦ 0, in confirmation of this. However, if we
wait a long time, of the order e-1, the small effects of the terms of order e
may accumulate and change r and 6 by quantities of order one. To find
these quantities, we average the contributions to dr/dt and dd/dt over the
approximate period 2n of the simple harmonic motion. This gives the lead¬
ing approximation to the slow changes of dr/dt and dO/dt and thence to the
changes of r and 6 over long times of order e-1.
Accordingly define the averages of r and d0/df as in the example above
and deduce from equations (3) that

da e C2n
dt = ~2i j„ -i(l -cos20)}d0 + O(e2)

= 2ta(l - |a2) + 0(e2),

(6)
because r is within order e of a for all t; and similarly
 6.6 Van der Pol’s equation 195

€ P2"
CO + 1 (r2cos20 - I)cos0sin0d0 + 0(e2)
2n Jo
= 0(e2) as€-+0. (7)
Equations (6) and (7) give

~t = iea(l - ^a2), co = - 1 (8)

approximately for small e. Equation (8) for da/dr is familiar from §1.5; its
solution at once gives

a(t) -*2 as t -*■ oo

for all a(0) > 0. This gives the asymptotic form of the limit cycle,

x(r) = 2 cos(f — t0) + 0(e) as e j 0. (9)

The orbits in the phase plane are approximately circular as they spiral
clockwise about the origin but spiral inwards or outwards slowly towards
the circle (9).
To find the limit cycle when e » 1 it is slightly easier to consider
Rayleigh’s equation (4) rather than van der Pol’s equation. We write the
system as

du
dr
x. - e(ix3 - x).

Further, define w = u/e and F(x) = jx3 — x. Therefore

dx dx/dt edx/d t
dw dw/dt du/dr

, w + F(x)
= —e -.
x

This gives the orbits in the (w, x)-plane for all e.

We now see that dx/dw -» oo as e -+ oo except near the curve w =
- F(x). Of course, dx/dw = 0 on the curve for all e. Therefore the direction
of each orbit in the (w, x)-plane is nearly parallel to the x-axis except near
the curve, the slope being large and negative where (w + F(x)}/x > 0 but
large and positive where {w + F(x)}/x < 0. So the slope of an orbit
changes sign at the w-axis and near the curve w = — F(x).
It follows that if an orbit starts at t = 0 off the curve then it goes ‘verti¬
cally’ towards the curve, moves near the curve clockwise around the origin
196 Second-order differential systems

Fig. 6.6 Sketch of the limit cycle for van der Pol’s or Rayleigh’s equation in the
(w, x)-plane for very large e.

(clockwise because w increases where x = edw/dt > 0 when e > 0) until the
tangent to the curve is vertical. There it must leave the neighbourhood of
the curve, moving vertically until it nears the curve again, and so forth, as
shown in Fig. 6.6. This gives a closed orbit Q1Q2Q3Q4Q1, which is the
limit cycle in the (w, x)-plane as e -* 00.
A more refined approximation is needed to find the details of the solu¬
tion near the curve w = — F{x). However, we have already found the
leading-order approximation to the limit cycle, and may use it to estimate
the period. On integrating around the cycle, the period may be expressed
as

x
 Further reading 197

because x = du/dt = edw/dr.

as € -+ oo,

because ‘dw’ is negligible along QtQ2 and Q3Q4 and symmetry implies
that there are equal contributions to the integral along Q2Q3 and Q4Ql5

= 2e[In x - \x2~\\

= (3 — 2 In 2)e

« 1.614c.

because x = 0 if and only if dw/dt = 0, and in the exact theory w — w0 oc x2

as x -*■ 0 on an orbit (see §1.7).)
The solution changes rapidly along the vertical tangents and
Q3Q4 and slowly along the segments Q4Qt and Q2Q3 of the curve; this
has led to the name relaxation oscillation for this kind of stable limit cycle
for large e, in which the solution slowly, as if languidly, builds up and then
rapidly changes.
Numerical calculations show that the limit cycle of van der Pol’s equa¬
tion in the (x, u)-plane and the (f,x)-plane for e = 10, a fairly large value,
are as in Fig. 6.7. Note that v = dx/dt = d2u/dt2, w = u/e and the point Qj
is passed at the same time tj as the point Q7, for j = 1, 2, 3, 4; and observe
the correspondence between Figs. 6.6 and 6.7.

Further reading
§6.3 A proof of Liapounov’s direct method is given in many good books
on ordinary differential equations, e.g., Jordan & Smith (1987, Chap. 10).
§6.4 For further reading on weakly nonlinear oscillations and asym¬
ptotic methods, the book by Kevorkian & Cole (1981, Chaps. 2, 3) is
recommended.
198 Second-order differential systems

(A)
Fig. 6.7 The limit cycle of van der Pol’s equation for e = 10, T = 19.08. (a) The
(x, u)-plane. (b) The (r, x)-plane.
 Problems 199

§6.5 Coddington & Levinson (1955, Chap. 16) prove the Poincare-
Bendixson theorem (cf. Poincar6 1882, Chap. VI; Bendixson 1901).
§6.6 Jordan & Smith (1987) give more details of van der Pol’s equation.

Problems
Q6.1 Lienard's construction. Show that the equation

d2x dx
¥+% + «w=°

is equivalent to the system

dx dy
= -g(x).
at'-™ dt

where F(x) = jj/(u)du. Then the (x,y)-plane is called the Lienard plane.
Suppose next that ^(x) = x for all x. Then show that the direction of the
orbit at a point P in the Lienard plane may be constructed as follows. Draw
the line through P parallel to the y-axis, and let Q be the point where the line
cuts the curve with equation y = F(x). Draw the line through Q parallel to
the x-axis, and let R be the point where this line cuts the y-axis. Then the
orbit through P is in the direction perpendicular to RP.
[Lienard (1928).]
Q6.2 Coulomb damping, i.e. motion of a particle under action of a spring with dry
friction. Suppose that

d2x
m = — Ax — Fsgn(dx/dr),
dF
where m, A, F > 0 and the function sgn is defined by sgn(u) = 1 if v > 0,
sgn(0) = 0, and sgn(i;) = — 1 if v < 0. Sketch the phase portrait in the
(x, dx/dr)-plane.
Q6.3 The quintessence of a saddle-node bifurcation. Consider the system

dx dy
a — x = -y-
dr dr

Show that there is a stable node and a saddle point if a > 0, but no equilibri¬
um point if a < 0. Sketch the bifurcation diagram in the (a, x)-plane, and the
phase portraits in the (x, y)-plane for a > 0, a = 0 and a < 0.
Q6.4 Epidemics. A simple model of an epidemic of a disease is the system,

dx dy
— = -cxy, — = cxy - by,

where x(r) represents the number of individuals in a population who are

200 Second-order differential systems

liable to infection, y(t) is the number who are infectious at time t, b > 0 is the
rate of recovery (or death) from the disease, and c > 0 is a rate of infection.
Show that if x(0) < b/c then the number y of infectious individuals will
decrease monotonically to zero, but if x(0) > b/c then the number will
increase monotonically until the number x of susceptible individuals de¬
creases to b/c.
[Kermack & McKendrick (1927).]
Q6.5 The Lotka-Volterra equations. Given that the growth of a population of x
individuals of a species of prey and y individuals of a species of predator is
governed by the equations

dx dy
— = x(a-cy), — = -y(b-cx),

for constants a, b, c > 0, show that X = (0,0) is a saddle point and X =

(b/c, a/c) is a centre. Show further that small oscillations about the centre
have period 2n/(ab)1'2. Prove that dy/dx = -y(b - cx)/x(a - cy), and inte¬
grate this equation. Hence or otherwise sketch the phase portrait in the first
quadrant of the (x, y)-plane.
[Lotka (1920) used these equations to model the chemical reactions
D + X -+ 2X, E + Y^E + F, X + Y-^2Yina well-stirred reaction vessel,
where x denotes the concentration of molecule X and y of Y, where D,
E are abundant molecules, and where a, b, c are the reaction coefficients.
Volterra (1926) used these prey-predator equations to model the population
of fish in the Adriatic Sea.]
Q6.6 Fishing. Suppose that a population of superpredators, e.g. fishermen, preys
with equal intensity / on both species such that a is replaced by a- f and b
by b + / in the above model (Q6.5). Then deduce that in stable equilibrium
the predator species y is decreased but the prey species x is increased by the
superpredators.
Q6.7 A damped simple pendulum. Find the equilibrium points of the equation

,d20
/ -j~2 + 2k — + g sin 6 = 0
dr2 df

and classify them, i.e. name their type, for all g, k, l > 0. Sketch the phase
portrait in the (6, d0/df)-plane for each topologically different case.
Q6.8 A nonlinear discontinuous restoring force. Find the equilibrium points of the
equation

d2x dx
diI + 2kdi+ {s8n(dx/d')}x = 0
for all k 0, where the function sgn is defined by sgn x = 1 if x > 0, sgn 0 =
0, sgnx = -1 if x < 0. Sketch the phase portrait in the (x,dx/df)-plane for
k = 0,0 < k < l,k = l,k > 1.
 Problems 201

Q6.9 The Brusselator. Biochemical oscillations are modelled by the system

dx
= a — x — bx + x2y,
dt dt

Find the equilibrium points and their stability for all a, b > 0.
Q6.10 Phase portraits and bifurcation. Treat each of the following real nonlinear
systems in turn for all real a, (a) noting the symmetries, if any, of the system,
(b) finding all the equilibrium points and for what values of a they exist,
(c) ascertaining the stability or instability of each point for each value of a,
(d) classifying the local behaviour of the orbits near each equilibrium point
for each value of a, (e) noting any special properties of the phase portrait,
(f) sketching the qualitative features in a phase portrait for each ‘typical’
value of a, and (g) sketching the bifurcation diagram in a suitable plane, e.g.
the (a, x)-plane.

dx
(i) “-*’+** {, -~y~
dt
dx dy
(ii) (a + x)y, — = —ax + x2 + y2.
dt at
dx
(iii) (a + x)y, -j- = a(y - x) + x2 + y2.
dt at
dx dy
(iv) a(x - y) - x2 + y2, — = (a + x)y.
dt
dx
(v) 3a(x -y)-x2 + y2, ~^ = x(a - y).
dt
dx dy
(vi) x(a-x- y), — = 2a(y - x) + y(x - y).
dt dt

dx dv
(vii) a(3x - 5y) - x2 + y2, — = x(2a - y).
dt dt

[Bristol (1983,1990), Iooss & Joseph (1990, Chap. V).]

Q6.11 Weakly nonlinear theory of a transcritical bifurcation. Consider the system

dx
dt
= ay-y‘
sH -2 y+i>

and define x = [x, y]T.

(a) Show that x = 0 is a solution for all a. Show that it is unstable for
a > 0, an unstable normal mode behaving like [2, l]Texp(^at) as a ->0.
Sketch the phase portrait of the solutions near x = 0 for the cases a < — 1,
— l<a<0, a>0, naming the type of singularity at the origin of the phase
plane in each of the three cases.
(b) Find each of the other steady solutions, stating the values of a for
which it exists. Find the values of a for which it is unstable.
202 Second-order differential systems

(c) Sketch the bifurcation diagram in the (a, x)-plane for the steady solu¬
tions of the system, and name the type of each of the bifurcations of the
solutions.
(d) To investigate nonlinear solutions near one bifurcation, show that the
system becomes
dx
0= d—
aA
u °y+ y’
-
x-2 y = a—-\x2-,
d>' 1 2
du

without approximation, where u = at. Defining A such that [2, l]x = aA,
i.e.

2x + y = aA,

and assuming that a solution may be expanded as

x(a,t) = axfu) + a2x2(u) + ••• asa->0,

where

dA
= F0(A) + aFl(A) +
du

show that x, = %A[2,1]T and F0(A) = \A — j^A2. Find x2 in terms of A.

Q6.12 A sphere of combustible material in equilibrium. It is given that the equilibri¬
um of a sphere of a uniform combustible solid is governed by the dimension¬
less system.
d2<f> 2 d<t>
dr2 + r dr + = °’ (1)

d^
—=0 atr = 0, <j> = Q at r = 1. (2)

This is a bifurcation problem to determine all dimensionless temperature

distributions (f> for all positive values of the Frank-Kamenetskii parameter S.
Defining x = Sr2e*, y = rd<t>/dr + 2, t = lnr, deduce that an equivalent
system is

dy 7
ii = 2 ~ y’ (3)

where

y) (0,2) as t -» — oo, x = S at t = 0. (4)

Show that equation (3) has a stable focus at (2,0) in the (x,y)-plane,
which solution for all t corresponds to a solution of (1) for all 5 such that
(j> = 0 at r = 1 only if 8 = 2.
 Problems 203

Sketch the phase portrait of the solutions of equation (3) in the right-
hand half of the (x, y)-plane, given that there exists an orbit H connecting
(0,2) to (2,0).
Show that if a solution (x,,(t), yh(t)) of equation (3) traverses the orbit H
and corresponds to a solution of boundary conditions (4) then xfc(0) = S.
*Hence or otherwise sketch the bifurcation diagram in the (<5, ^(0))-plane.
[Budd (1989).]
Q6.13 A condition for stability. Show that if

where x e (Rm, F: IRm -» Rm, F(0) = 0, and

uTF(u) ^ 0

for all u e Rm, then 0 is a stable point of equilibrium.

Deduce that the solution 0 of the linear system

is asymptotically stable if all the eigenvalues of A + AT are negative.

*Q6.14 Liapounov function of a difference equation. Show that if F: IRm -> IRm is con¬
tinuous, F(0) = 0, and there exists a continuous positive definite function
H: Rm -> 0? such that H(F(x)) < H(x) for all x in a neighbourhood of 0, then
0 is a stable fixed point of F.
Q6.15 Asymptotic stability of an equilibrium point of Lienard's equation. Suppose
that
d2x „ dx
d? + /MS + 9W“°-

where /, g are continuous, #(0) = 0, and f(x) > 0, xg(x) > 0 in a neighbour¬
hood of the origin excluding x = 0 itself. Show that the system is equivalent
to
dx dy
= -g(x).
a-’-** dt

where F(x) = Jo/(u)du, and that there is an equilibrium point at the origin
of the (x, y)-plane.
Taking H(x,y) = jy2 + G(x) as a Liapounov function, where G(x) =
g(u) du, show that the origin is asymptotically stable.
Q6.16 A linear system, (a) Find explicitly the general solution of the system.

dx dy , . dz
= — 3x, = — 2y — z.
dr dt “ ~y + 2z' dt

Deduce that the origin is asymptotically stable.

204 Second-order differential systems

(b) Prove that the origin is asymptotically stable by Liapounov’s direct

method.
Q6.17 A fundamental matrix of a linear system. Suppose that

(1)

where x is a 2 x 1 column vector and A(t) is a 2 x 2 real matrix. Define 4>i(t)

as the solution of the system (1) such that 4>i(0) = [1,0]T and <J>2(0 as the
solution such that <J»2(0) = [0,1]T. Then a fundamental matrix <D of (1) is
defined as the 2 x 2 matrix with columns <j>,, 4>2, i.e. <D(t) = [<M0> 4*2(0]-
Show that dO/df = AO and 0(0) = I, the unit matrix, and hence that
x(0 = <l>(0x(0).
Deduce that if A is constant then O(f) = 1 + rA/1! + t2A2/2\ + •••,
— etA, say.
Q6.18 A Liapounov function for an asymptotically stable linear system. In Q6.17,
suppose further that A is constant and has eigenvalue sk < 0 belonging to
the eigenvector u* for k = 1, 2, and that the two eigenvectors are indepen¬
dent. Deduce that there exists a constant real invertible 2x2 matrix P such

that P_1AP

Defining X = [Xl,X2Y such that x = PX, deduce that the system (1) of
Q6.17 is equivalent to

d^_s* (2)
dt ~SiXu dt _S2*2‘

Defining

H(x) = X2 + X2, (3)

show that H is positive definite and that - dH/dt is positive definite if x

satisfies system (1). Hence prove the asymptotic stability of the null solution
of system (1) (although this result follows directly in the present case because
the solution of the linear system (1) can be found explicitly).
Q6.19 The asymptotic stability of the null solution for both a nonlinear system and
its linearized system. Suppose yet further that

dx
— = Ax + $(x) (4)

for the same real constant matrix of Q6.18, where |£(x)| = o(|x|) as x -► 0.
Using the same Liapounov function H of equation (3) of Q6.18, prove that
the null solution of system (4) is asymptotically stable if the null solution of
its linearized system is stable.
 Problems 205

Q6.20 Instability of an equilibrium point of a nonlinear system. Suppose that

— = ax + by + £(x,y), — = cx + dy + q(x, y). (5)

and p,q> 0, where £(x,y), tj(x,y) = o(|x|) ,

as x -»0 p = a + d, q = ad — be.
Show that the origin is an unstable point of equilibrium of the linearized
system.
Define K by -K(x,y) = (ax + by)2 + (cx + dy)2, and show that

K(r cos 9, r sin 6) ^ mr2 for all 9,

where

m = j[a2 + b2 + c2 + d2 — {(a2 — b2 + c2 — d2)2 + 4(ab + cd)2}1/2] > 0.

Defining H by

H(x,y) = (ax + by)2 + (cx + dy)2 + g(x2 + y2),

show that

H(x, y) >(m + q)(x2 + y2) for all x, y,

grad H ■ dx/df > mp(x2 + y2) for sufficiently small x # 0.

Hence prove that the origin is an unstable point of equilibrium of the non¬
linear system (5).
Q6.21 Liapounov function for a damped hard spring. Consider the equation

for k, l > 0. Defining the Liapounov function H(x, dx/dr) = \(dx/dt)2 +

^x2 + ilx4, show that x(t) -* 0 as t -* oo for all initial conditions.
Q6.22 Asymptotic stability of an equilibrium point. Show that the null solution of
the system

^ * 2 3 ^ 2 2 3
— - 2xy* - Xs, rri*y-y

is asymptotically stable.
[Hint: try a Liapounov function of the form H(x,y) = |(x2 + ay2).]
Q6.23 Instability of an equilibrium point. Show that the null solution is the only
equilibrium point of the system

dx 2 , 3
— = xy2 + x2y + x3,
dr

and that it is unstable.

206 Second-order differential systems

Q6.24 An example of stability with an unstable linearized system. Show that the
origin is a stable point of equilibrium of the system

dy
— x
dt

but that it is an unstable point of equilibrium of the linearized system.

[Hint: it may help to use a Liapounov function of the form H(x, y) =
xm + by".']
Q6.25 An application of Liapounov’s method to a class of ordinary differential sys¬
tems. Show that the null solution of the system

dx dy
-j7 = y - xf(x,y), ~n = ~x- xf{x,y).

where /(0,0) = 0 and / is continuous near the origin O of the (x, y)-plane, is

(a) stable if / is positive semidefinite at O,

(b) asymptotically stable if / is positive definite at O, and
(c) unstable if —/ is positive definite at O.

Hence show that the null solution of the system

dx . . dy
— = y - x(x4 + y4), — = -x - y(x4 + y4)

is asymptotically stable.
Q6.26 A Liapounov functional. Consider the nonlinear diffusion equation

for k > 0, together with the boundary conditions that u(x) = 0 at x = 0,

7i. Defining the functional H: C[0,n] -» R by H(v) = Jgic2dx, show that
dH(u)/dt = — {k(du/dx)2 + u4} dx < 0 and hence that the null solution is
stable (in a mean sense).
Show further, by the calculus of variations or otherwise, that 0
jo 5(dw/<9x)2dx - H(w) for all functions w e C2[0,7t] such that w(x) = 0 at
x = 0, 7T.

Deduce that dH{u)/dt -H(u) and thence that the null solution of
equation (1) is in fact asymptotically stable (in the mean).
[Liapounov’s method has been generalized so as to apply to partial
differential equations (Zubov 1957).]
Q6.27 A nonlinear oscillation. Consider the equation

d2x
-nr + x - ex|x| = 0.

Find its equilibrium points, and for what values of € each exists. Find for
 Problems 207

what values of e each is stable, and classify the type of the point. Sketch the
bifurcation diagram in the (e, x)-plane.
Show that

1 1
+ -x2 --€|x|3 = E

for some constant E of integration.

Sketch the phase portraits for ‘typical’ values of e, i.e. for some values
which give rise to all topologically different portraits.
Show that there exist periodic oscillations for all e, and that the period of
the oscillation of amplitude a is

dx
fea3- x2 + fex3)1/2'

Deduce that

T = 2n 4- fea + 0(e2a2) as ea -> 0.

Q6.28 Another nonlinear oscillation. Consider the equation

for an asymmetric spring. Find its equilibrium points, and for what values of
€ each exists. Find for what values of e each is stable, and classify the type of
the point. Sketch the bifurcation diagram in the (e, x)-plane.
Use the Lindstedt-Poincare method to find the leading terms in the
expansion of the periodic solutions. Explain the signs of the corrections to
the period and to the time average of x for small e.
Q6.29 The precession of the perihelion of a planet in the general theory of relativity.
It is given that the equation of the orbit of a planet with plane polar coordi¬
nates (r, 6) is

d2u 1
^ + u = 7 + e/«,

where u = 1/r, the sun is fixed at the origin r = 0, l = h2/GM and e =

3GM/c2l. Here G is the gravitational constant, c is the velocity of light, M is
the mass of the sun, and h is the angular momentum per unit mass of the
planet about the sun.
Show that there is a centre at (uN,0) and a saddle point at (ur,0)
in the (u,du/d0)-plane, where uN(e) = \/l + e/l + 0(e2/l) and ur(e) =
1/e/ — l/l + 0(e/l) as ej.0, i.e. in the Newtonian limit. Sketch the phase
portrait and identify the region of orbits representing solutions which are
periodic functions of 6. (N.B. Periodicity of an orbit in time is a very different
property.)
208 Second-order differential systems

Define <f> — cod, where 2n/a) is the period of an orbit, and assume that
a>(e) = 0)0 + ea», + e2a)2 + •••. Hence show that a>0 = 1 and to1 = — 1.
Deduce that the planet is at perihelion, i.e. that its distance from the sun is a
local minimum, at successive angles 6 differing by 2n/a> = 2n + 2ne + 0(e2)
as e 10.
[This gives the precession of the perihelion of the planet by the angle
2ne = 6nGM/c2l approximately each revolution, where / is close to the
mean radius of the orbit. This result is used in one of the classic tests of
Einstein’s general theory of relativity.]
Q6.30 A periodic solution of Rayleigh’s equation. It is given that the equation

d2x fdx 1 (dx

dH+x = e|d7-3U
has a unique periodic solution for all e ^ 0, such that dx/dt = 0, d2x/dt2 < 0
at t = 0. Then use the Lindstedt-Poincare method to show that the solution
is given by

x(e, t) ~ 2 cos a>t ase->0,

where a> = 1 — -I- 0(e2).

Q6.31 Bendixson's criterion for the non-existence of a periodic solution. Show that if

dx dy
F(x,y), G(x,y),
dt dr

for continuously differentiable functions F, G, and dF/dx + dG/dy ^ 0 in a

domain D c [R2, then there is no closed orbit in D.
[Bendixson (1901).]
Show that if

dx dy ,
— = 2x - xyz + cosy, — = -y — x2y + sinx
at dr

then there is no closed orbit inside the circle x2 + y2 = 1.

Q6.32 Liapounov and Poincare-Bendixson. Consider the nonlinear system

J = y + x3 - x{x2 + y2)2, ^ = -x + y3 - y(x2 + y2)2.

(a) Show that the origin is an equilibrium point. Linearize the system
at the origin, sketch the phase portrait of the linearized system, and state
whether the origin is a stable or unstable point of it.
Use Liapounov’s direct method to show that the origin is an unstable
point of equilibrium of the nonlinear system.
(b) Show that the given system has a limit cycle.
[Bristol (1989).]
 Problems 209

Q6.33 Liapounov, Poincare-Bendixson and Poincare. Consider the nonlinear

system

^L = ax + y- xf(x2 + y2), ^ = -x + ay - yf(x2 + y2),

where a is a real parameter, / is continuous, /(0) = 0 and f(x2 + y2) >

(x2 + y2)1'2.
(a) Show that the origin is the only equilibrium point. Linearize the
system at the origin and find what type of point it is according to the values
of a.
Use the Liapounov function H(x, y) = ^(x2 + y2) to show that the origin
is a stable point of equilibrium of the nonlinear system if a < 0 and unstable
if a > 0. For what values of a is it certainly asymptotically stable?
(b) Show that there exists a stable limit cycle of the system if a > 0.
(c) Take the special case with /(r2) = r for all r ^ 0 and with a > 0.
Find the limit cycle explicitly by solving the system.
A Poincare map P: IR2 -»IR2 is defined for the system in this case as
follows: take x(0) = x0, integrate the system from t = 0 to t = 2n to deter¬
mine x, = x(27t), and let xt = P(x0) for all x0, where x(t) = (x(f),y(t)). Show
that

ax
P(x) = -2na *
Ix| -I- (a - |x|)e

[Bristol (1988).]
Q6.34 Demonstration of the existence of a limit cycle. By considering the directions
in which orbits cross suitable closed curves, show that the system

” = —x — y + (x2 + 2y2)x, ^ = x - y + (x2 + 2y2)y,

df at

has at least one periodic solution.

Q6.35 Demonstration of the existence of another limit cycle, (a) Given the equation,

d2x
dF + S8”x = °’
find explicitly the solution such that

x = 0, dx/df = 4 T at t = 0,

where T > 0. State the amplitude and period of the solution in terms of T.
(b) Sketch the phase portrait of the equation

^ + e(x2 - 1)^ + sgnx = 0

dr dt

forO < e « 1.
210 Second-order differential systems

(c) Show that the equation

+ e(x2 — 1)^ + tanh ax = 0,

dr dt

for a, e > 0, has one and only one periodic solution (except for a translation
of time). Is it stable?
Note that tanh ax -» sgnx as ax -* +oo, and comment on the relation¬
ship between the phase portraits for parts (b), (c) of this question.
Q6.36 Demonstration of the existence of yet another limit cycle. Show that the
equation
dx
+ x3 = 0
dt

has at least one periodic solution.

Q6.37 Change in the topological character of the orbits near a singular point owing
to nonlinearity. Show that if

dx _ 2y dy _ 2x
dt X + ln(x2 -I- y2)’ dt ^ ln(x2 + y2)’

then
dr _ _ dO _ 1
dt ’ dt lnr’

where x = rcos0, y = rsin0, r > 0. Hence find 9 as a function of r and

describe the character of the orbits near the origin.
Describe the character of the orbits of the linearized system, namely

dx dy

[Nemytskii & Stepanov (1960, p. 85).]

Q6.38 A Tokens-Bogdanov bifurcation. Show that the system

dx dy
= y. = a + by + x2 + xy
dt dt

has equilibrium points at X± = ( + (-a)1'2,0) if a < 0, and at (0,0) if a = 0,

there being no equilibrium point if a > 0.
Show that X+ is a saddle point, and that X_ is unstable if b > (~a)l/2
and stable if b < (-a)1'2. Moreover, classify the types of the equilibrium
point X_ according to the values of a < 0, b.
*Using the Lindstedt-Poincare method, define e = b — ( — a)1'2, and
assume that x = (x, y) has period 2n/a>, and

x = xo + fl/2xi/2 + exi + e3/2x3/2 + • • •, co = co0 + €1/2<u1/2 + ecoj +•••

ase ->0, where x0 = X_. Hence verify that co0 = {2(—a)l/2}1/2 anda)l/2 = 0,
 Problems 211

find x1/2, and show that to, = (10 + o>q)/6o)0. Deduce that there is, for fixed
a < 0, a suhcritical Hopf bifurcation, in the sense that these periodic solu¬
tions occur when 0 < (— a)1'2 — b « 1.
Deduce that there is a turning point (with one eigenvalue s = 0) at a = 0,
b t*0, and a Hopf bifurcation (with eigenvalues s= ±i(2b)i/2) at b =
( — a)112, a < 0.
[The system exhibits a coincidence of turning points and Hopf bifurca¬
tions, called a Tokens-Bogdanov bifurcation, at a = b = 0. Cf. Gucken-
heimer & Holmes (1986, §7.3).]
Q6.39 The method of averaging. Show that if

d2x
+ x = ef(x, dx/df)
df2
then

da
sin0/(acos0,asin0)d0 -I- 0(e2),
df
♦2*
€
to + 1 a-1 cos Of (a cos 6, a sin 9) 69 + 0(e2) as e -»0,
27C o

where x = r cos 9, dx/dt = r sin 6, a is the average of r, and a> is the average
of d0/dr around one circuit of the nearly circular orbit in the (x,dx/dt)-
plane.
Q6.40 A self-excited oscillation. Show that the origin of the (x, u)-plane is an equi¬
librium point of the equation

u ^
d2x dx
-jt ~€(1 — lxl) + * = 0,
df

where v = dx/dt. Name the type of the equilibrium point, and state when it
is stable according to the value of e. What other equilibrium points are
there?
You may assume that there is one and only one limit cycle when € > 0,
and that it is stable. Show that the limit cycle is given by

x(c, t) -* 47t cos(f — f0) asejO.

Show that the equation is equivalent to the system

dw x
_ = -€{w + F(x)},
df ~ e’

where F(x) = jx|x| — x. Sketch the curve w = - F(x) in the (w, x)-plane and
indicate the locus of the limit cycle as e -* oo. Hence or otherwise show that
the period T of the limit cycle is given by

F(e) ~ 2{V2 - ln(l + Jl)}e as€->oo

H
212 Second-order differential systems

Q6.41 Another self-excited oscillation. Consider the equation

Show that a limit cycle x(e, t) is given by

x(e, t) -»• 23/4 cos(r — t0) as e 10.

Show that the equation is equivalent to the system

dw x dx ,
dT ' ? d7 “ ~'{w + F(x))’

where F(x) = ^x5 — x. Sketch the curve w = — F(x) in the (w,x)-plane and
indicate the locus of the limit cycle as e -* oo. Hence or otherwise show that
the period T of the limit cycle is given by

T(e) ~ 2{J(a4 — 1) — lna}e ase->ao,

where a is the positive root of the quintic equation a5 — 5a — 4 = 0.

[In fact a « 1.65. You may assume that Jo” sin2 9cos4 Odd = |7r.]
Q6.42 Yet another self-excited oscillation. Consider the equation

d2x dx
3?+% + 1 = t
where /(x) = — 1 if — 1 < x ^ 1 and /(x) = 1 if |x| > 1. Write down the
explicit general solutions for the four cases, |x| > 1 and — 1 ^ x ^ 1, for
0 < e < 2 and e > 2. Sketch the phase portraits in the (x, u)-plane for 0 <
e < 2 for e > 2, where v = dx/dt, deducing plausibly that there is a limit
cycle.
Show by the method of averaging that the limit cycle x(e, t) is given by

x(e, t) -> aecos(t — t0) asejO,

where ae > 1 is the root of the equation 2 sin 2a + n = 4a for a, and 0 < a =
arccos(l/a) < \n.
Show that the differential equation is equivalent to the system

dw x dx
-£{w + F(x)},
df e’ dr

where F(x) = —x for |x| < 1 and F{x) = x — 2sgnx for |x| > 1. Sketch the
curve w = — F(x) in the (w, x)-plane and indicate the locus of the limit cycle
as e -* oo. Hence or otherwise show that the period T of the limit cycle is
given by

T(e)~2€ln3 ase-»oo.
 Problems 213

Q6.43 Rayleigh’s equation. Apply the method of averaging directly to Rayleigh’s

equation.

d2x
, , +€ + x = 0,
dt2

to find the amplitude and frequency of the limit cycle as € -»0. You need
only find the terms up to those of order e.
Q6.44 Simple pendulum. Writing the equation

— =-gr'sin0

in the form

^+ 0= em,

where t = (l/g)v2t and ef(0) = 0 — sin 0 (by introducing artificially a small

positive parameter c), use the method of averaging to find the frequency of
the periodic solutions of small amplitude approximately, giving the correc¬
tion term in the square of the amplitude of the oscillation.
Q6.45 An approximation to a limit cycle. Use the method of averaging to approxi¬
mate the solutions of the equation

as e -»0, finding the amplitude and frequency of the oscillations for small e.
Q6.46 An amplitude equation. Show that if

d2x iJdx\
— + jc =0

and a is the average of {x2 + (dx/dt)2}1/2 over the cycle of an oscillation,

then

da 1 ,, ,
— ~ -—ea3(a2 — 6) as€-+0.
dt 16

Hence find the amplitude and stability of the limit cycle for small e.

Q6.47 Equilibrium points of a Hamiltonian system. Show that if a second-order sys¬

tem has a continuously twice-differentiable Hamiltonian Hip, q) and dHIdp =
dH/dq = 0 at a given point, then it is a point of equilibrium. Find the Jacobian
matrix J, and show that trace J = 0, det J = {d2HI dp2) (i^H/dq2) - (d2Hldpdq)2.
Deduce that the point is a saddle if det J < 0 and a centre if det J > 0.
 7
Forced oscillations

The thing that hath been, it is that which is.

Ecclesiastes i, 9.

1 Introduction
Chapter 6 is an account of many and varied properties of free oscillations
which are solutions of two-dimensional autonomous systems of ordinary
differential equations. We saw there that the Poincare-Bendixson theorem
prohibits chaos, but saw in Chapter 5 that a two-dimensional non-autono-
mous system is equivalent to a three-dimensional autonomous system, and
so may have chaotic solutions: the proof of the Poincare-Bendixson theo¬
rem becomes invalid because for a non-autonomous system orbits may
have different directions at the same point of the phase plane at different
times. So in this chapter we shall examine forced oscillations, i.e. solutions
of non-autonomous second-order equations of the form

(1)

We call F the forcing term, thinking of the equation as a model of an

oscillation of a particle of unit mass with a restoring and damping force /
driven by an external force F which is a periodic function of time. We shall
see that forced oscillations have even more and more-varied properties
than do free oscillations; in particular some forced oscillations are chaotic.
To illustrate the properties of forced oscillations we shall examine in
detail Duffing's equation, namely

-^2 +& + otx + Sx3 = r cos cot, (2)

for real parameters k, a, <5, T, co. (Duffing (1918) used the equation to model
nonlinear mechanical vibrations.) With so many parameters it will help to
consider various special cases, beginning in this section with a review of
some properties of the linear form of the equation. First consider
214
 7.2 Weakly nonlinear oscillations not near resonance 215

d2x
-r~., + ax = T cos cot (3)
dr

for a > 0, an equation which represents forced simple harmonic motion

without damping. The general solution is

.. , _ . , r cos cot
x(f) = /I cos wat + Bsin Jat +-ifco ^ a,
,
(4)
a — co

for arbitrary constants /l, B. This is a quasi-periodic function in general,

but a periodic function if co is a rational multiple of Jix. The response, or
forced oscillation, is the particular integral T cos cot Ha- co2) proportional to
the forcing term and grows without bound as co2 —> a. It is in phase with the
forcing if co2< a and K out of phase if co2 > a.
.. . „ . Tt sin cot
x(f) — A cos cot + B sin cot H-—— (5)
2co

and there is said to be resonance; it can be seen that the response is a

secular term which grows without bound as t increases, but which is finite
at any given value of t.
Damping diminishes this response. If

d2x dx _
—T + k — + ax = T cos cot, (6)
dr dt

and k > 0,0 < k2 < 4a, then

x(f) = e *'/2{/lcos(a — {k2)l/2t + Bsin(a — jk2)1/2t}

T{(a — co2)coscot + kco sin cot}

(7)
+ (a — co2)2 + k2co2

Now the response is of order T/kco when co2 = a, so there is a resonance

‘peak’ with finite height T/kco. The complementary function, namely
e~kt/2{/4cos(a - \k2Yl2t + Bsin(a — ik2)l/2t), is called the transient, be¬
cause after a long time it decays, leaving only the response to the forcing.
We shall next examine how nonlinearity, like damping, modifies resonance
and renders the response finite.

2 Weakly nonlinear oscillations not near resonance:

regular perturbation theory
To introduce the effects of nonlinearity on resonance, first consider an
illustrative example of Duffing’s equation with no damping and weak
 216 Forced oscillations

nonlinearity,

d2x , _
—r + ft^x — ex* = T cos f, (1)
dr

on taking a> = 1 without loss of generality (because this is merely a rescal¬

ing of the unit of time), and rescaling a, <5. Let us ignore the transients, and
confine our attention to synchronous oscillations, i.e. to those solutions
x(e, t) of equation (1) with the same period 2n as the forcing. Also, to ap¬
proximate weak nonlinearity, we shall expand

x(e, t) = x0(t) + exj(f) + e2x2(t) A- as e ->■ 0 (2)

for fixed ft, T. Substituting expansion (2) into equation (1), we find that

d2xn d2Xi i ,
-jpp + t-faT +-1- ^2(*o + exx A-) — e(x0 4- exx A-)3 = Tcost.

Equating coefficients of successive powers of e, we find

d2x0
-I- ft2x0 = T cos t, (30)
dt2

d2xx
+ ft2xx = Xo, (31)
dr2

d2x2
A- ft2x2 = 3xoXt, (3a)

and so on.
Also the periodicity condition

x(e, t + 2n) = x(e, t) for all t, e, (4)

gives similarly the general condition

x„(t + 2n) = x„(t) for all t, n. (5J

Therefore equation (30) gives

x0(t) = A cosilt + Bsinftt + -2C°S 1

if ft + 1. Also periodicity condition (50) now gives

T cost
*o(0 = (6)
ft2-l

if ft is not an integer.
 7.2 Weakly nonlinear oscillations not near resonance 217

Next equation (3 x) becomes

d2xt 2 _ r3(3 cos t + cos 3r)

+ 1 = 4(Q2 - l)3 •

This and periodicity condition (5X) give

3T3 cos t T3 cos 31 ^

“ 4(Q2 - l)4 + 4(ft2 - l)3(ft2 -9)' ( ’

We can proceed to find x2, x3,... in turn to determine the unique solution
consistent with the hypotheses we have made. In fact the method is suc¬
cessful provided that ft is not an odd integer, because if ft is an odd integer
then some equation (3J has a forcing term in cosftf on its right-hand side
to drive a secular term in x„. Also the expansion (2) in powers of e is not
uniformly valid when ft is close to an integer, i.e. is not valid in the case of
resonance which we set out to examine. At any rate, let us review the results
of this theory by drawing a response diagram as in Fig. 7.1. It is informa¬
tive to plot the amplitude a0 = r/(ft2 — 1) of the response as a function of
the natural frequency ft of the free oscillations; however, a change of sign of
a0 merely corresponds to a phase shift of n in t, so the magnitude of a0 is a
better measure of the response to the forcing T than is a0 itself.

Example 7.1: weakly forced oscillations of van der Pol’s equation. There
are many asymptotic methods to solve weakly nonlinear problems (cf.
Kevorkian & Cole 1981), but not sufficient space to illustrate them all here.
However, to illustrate another asymptotic method as well as important
phenomena of forced oscillations, consider the equation

Fig. 7.1 Sketch of the bifurcation diagram in the(Q,|a0|)-plane for fixed T # 0,

i.e. of the curve | a01 = |r/(fi2 — 1)|.
 218 Forced oscillations

d2x dx ]
y + x = e<y cos cot + (1 — x2) — (8)
dr

for y, €, co > 0, co # 1. To solve the equation for small e, suppose that

x = x0 + €Xj + e2x2 H- as e J 0, (9)

where x0 = r cos(t — <f>) for slowly varying functions r, <f> of time, so that r, <j)
are functions of et. Note that the limit cycle of van der Pol’s equation
without forcing (i.e. with y = 0) has period 2n for small e, and that the
forcing has period 2n/co.
Substitution of expansion (9) into equation (8) gives

dr d<f>
cos(r — </>) + + 2- sin(r — <f>)
dr dr

+ 6 -I- 0(e2) = ey cos cur + e(l + 0(e2)

= ey cos cur — jer(4 — r2)sin(r — <j>)

+ jer3 sin 3(r - <f>) + 0(e2). (10)

Therefore, on equating terms of order e, it follows that

d2xx
+ xt y cos cur - |r(4 - r2) sin(r -</>) + |r3 sin 3(r - </>)
~dP~

2r d 6 2 dr
-7drcos('~^) + idrsin(' (11)

Remember that r, ^ vary slowly, so that each term above is of the same
order. To ensure that Xj is bounded, i.e. that there is no secular term like
rsin(t — <f>) or rcos(r — (j>) in x1? the coefficients of cos(r — <f>), sin(r — <f>) on
the right-hand side of the above equation vanish. Therefore

d<j> , dr
— = 0(e2), — = ier(4 - r2) + 0(e2) asejO. (12)

This is just what we found (equations (6.6.8)) for the limit cycle of van der
Pol’s equation without forcing, so that r -* 2 as t -*• oo unless r = 0 at t = 0.
It follows that

d2xt
+ Xj = y cos cot + £r3 sin 3(r — <f>).
dr2
 7.3 Weakly nonlinear oscillations near resonance 219

Therefore

x(e,£) = r cos(f -</>)- ~er3sin3(t -</>)- ^ C°S + o(e). (13)

Note that the solution approaches the sum of the stable limit cycle of
van der Pol’s equation without forcing and a forced oscillation. However,
as cd —► 1 the forced oscillation of period 2n/co dominates the free oscilla¬
tion of period In - this is called entrainment. This leads to the invalidity of
the method of approximation, an issue which will be taken up in Example
7.2. □

3 Weakly nonlinear oscillations near resonance

Next we shall resolve the nonuniformity of the validity of the weakly non¬
linear approximation for fixed forcing T, showing that the resonant
response is large but finite for weakly nonlinear oscillations. This will be
illustrated with the example,

d2x
+ x = e(y cos t — /?x + x3) (1)

x(e, t + 2n) = x(c, t) for all t. (2)

This equation may be used to represent the forcing of a soft spring, or of a

simple pendulum when the term in sinx (i.e. sin0) is approximated by
x — gx3. The terms have been chosen not only so that there is both weak
nonlinearity and closeness to resonance but also so that these two effects
balance as e -> 0 for fixed constants /?, y, i.e. that we take the distinguished
limit as € -> 0.
Here substitution of expansion (2.2) in powers of e into equation (1)
yields first

+ x0 = 0.

Therefore

x0(£) = a0 cos t + b0 sin t

for some constants a0,b0. This solution satisfies the periodicity condition
for all a0,b0.
Next, coefficients of e in equation (1) give
220 Forced oscillations

d2xt 3
—7-5- + Xj = 7 cos t - px0 + Xo
dr

= y cos t — /J(a0 cos t + b0 sin 0 + (ao cos t + b 0 sin t)3

= {y- Pa0 + |a0(ao + &o)}cos£ + b0{-p + |(a£ + b£)}sint

+ kao(ao ~ 3bo)cos 31 + kb0(3a% — b%)sin 31.

The periodicity condition, x^f + 2tt) = xx(£), requires that the secular
terms, involving t sin t and t cos t, in xx must vanish and therefore that the
coefficients of cos t and sin t respectively on the right-hand side of the equa¬
tion of Xj must vanish, i.e. that

a0{P ~ l(«o + &o)} = 7, b0{p - l(al + b%)} = 0.

Therefore

b0 = 0, a0(p - |al) = y. (3)

This gives

x0(£) = a0 cos t, (4)

xt(£) = ax cos t + bx sint — ^aocos3f, (5)

where a1,bl may in fact be determined uniquely in terms of a0 by requiring

that the secular terms at the next approximation be zero.
We have now found the leading terms in the approximation to the
weakly nonlinear forced oscillations near resonance. Let us examine their
behaviour. The leading approximation x0 to the response is in phase (or n
out of phase if a0 has the opposite sign to ey) with the forcing, because
b0 = 0; this follows from the fact that the nonlinear term is an odd function
of x. The cubic (3) for a0 has been discussed in Examples 2.2 and 2.5.
It has three real roots if p > (16y2/9)1/3, two if p = (16y2/9)1/3 and one if
P < (16y2/9)1/3. It gives a pitchfork bifurcation in the (/?,a0)-plane when
y = 0 and an imperfect one for y ^ 0, as shown in Fig. 7.2. Note that
ao = — (4y/3)1/3 if P = 0; a0 = 0 or ±(4/1/3)1/3 if y = 0; and a0 ~ y/p as
y -»0 for P # 0.
In terms of the dimensional form.

d2x
+ ax + <5x3 = T cos cat, (6)
dtJ

of Duffing’s equation without damping we may choose a time scale so that

t = cot and deduce that
 7.3 Weakly nonlinear oscillations near resonance 221

r= o

Fig. 7.2 Sketch of the bifurcation curves in the (/?, |a0|)-plane for fixed y.

d2x r
H—*x H—;r COS T,
dt* co to* W

i.e.

= e(ycost — fix 4- x3).

on identifying T/to2 = ey, a/to2 = 1 + e/J and e = —d/co2. This shows ex¬
plicitly that there is weak nonlinearity and closeness to resonance when
t is small for fixed /}, y. Now the second of equations (3) gives to2 =
a + \6al - Y/a0 if a0 # 0. If T = 0 then a0 = Oor ± {4(to2 - a)/3S}112.
We may plot the above results in the (to, |a0|)-plane as shown in Fig. 7.3,
as is conventional to give the response according to the tuning, showing,
for fixed S < 0, the bifurcation curves for constant values of T. We see that,

Fig. 7.3 Curves T = constant in the (to,|a0|)-plane near to the point (Va, 0):
to2 = a + iSao - r/a0 for 8 < 0. It can be shown that there is onset of instability
at the turning point as indicated by the broken curve.
222 Forced oscillations

for fixed T, the resonance peak at co = Ja is made finite. It can be seen

that near resonance there may be one periodic response Trout of phase with the
forcing or three periodic responses, of which one is n out of phase and two
are in phase with the forcing. We have only considered solutions with
the same period as the forcing here, but in fact they are stable as indicated
by continuous curves and unstable as indicated by the broken curve in
Fig. 7.3. It can be seen that if co were to increase slowly then the solution
would abruptly increase its amplitude when it became unstable, and that
hysteresis might occur, just as we would anticipate for a cusp catastrophe.
If we were to include weak damping in equation (1), i.e. if we were to
start instead with

then the above asymptotic method would give the amplitude equations

*b0 + a0{P -l(a% + bl)} = y, Ka0 - b0{p - l(al + b%)} = 0 (8)

instead of (2). The damping leads to a phase difference between the forcing
and the response, with b0 # 0. However, it can be seen by adding the
squares of equations (8) that

ro{K2 +{P~lrl)2} =y2 (9)

instead of equation (3), where the amplitude of the response is now r0 =

(flo + ho)1/2. Note the relation of this result for fixed k to the cusp catas¬
trophe of Example 2.2. We again plot the resulting response curves T =
constant in the (co, r0)-plane as shown in Fig. 7.4. In fact there is hysteresis
owing to the instability of the solution denoted by the broken curve.

Fig. 7.4 Curves T = constant, for fixed k > 0, in the (co,r0)-plane: co2
a + \8al - f/a0 for 5 < 0.
 7.3 Weakly nonlinear oscillations near resonance 223

Example 7.2: weakly forced oscillation of van der Pol’s equation near reso¬
nance. Suppose that

f dx 1
•jp- -I- x = ejycost + (1 - x2) — - px> (10)

as ej.0 for fixed p, y. First note the similarity to Example 7.1, the only
difference being that the limit cycle of the unforced van der Pol equation
with period 27r/(l + e/1)1/2 and the forcing with period 2n are nearly in
resonance here.
The equation may be solved asymptotically by using the expansion,

x = x0 + ext + e2x2 + ••• asejO, (11)

of Example 7.1. However, we shall take the form

x0 = a cost + b sin t (12)

here for slowly varying functions a, b instead of the polar variables r =

(a2 + b2)112, <f>, assuming that da/dr, db/dt = 0(e) as e 10, after van der Pol
(1922). This will serve to display a slightly different method and also solve
the problem in hand a little more conveniently. The (a, b)-plane is called the
van der Pol plane.
Substitution of expansion (11) into equation (10) now gives

dh\ dtA •
2— cost + — 2 — ) sin t + e + 0(e2)
dtj dtj

= e<ycost — fix0 + (1 + 0(e2).

Equating terms of order e, we find

d2xx 2 db 2 da dxo
+ Xj = ycosf — Px0 —- — cost + - —sint + (1 — *o)^rr
(

dt2 e dt e dt dt

n, , . ,
2 db 2d a .
= ycost — Biacost + osmf)-—cost H-— sint
' r e dt e dt

+ (\ — ^r2)(hcost — a sint) — ia{3b2 — a2)sin3t

— i(3a2 — h2)cos3t. (13)

In order to ensure that xt is bounded, the coefficients of sin t, cos t on the

right-hand side of equation (13) vanish. Therefore

~ = gC(4 - r2)a + \ePb, ^ = ^e(4 - r2)b - \epa + \ey. (14)

at at
224 Forced oscillations

The equilibrium points (a0, b0) of system (14) give periodic solutions (12)
approximating solutions of equation (10):

i(4 - rl)a0 + pb0 = 0, Pa0 - £(4 - r%)b0 = y. (15)

Squaring and adding, we find that

ro{A(4 — r,)2 + p2} = y2, (16)

a familiar form of equation for the amplitude of a resonant oscillation

(cf. equation (9)).
Now system (14) enables us to investigate the stability of its equilibrium
points corresponding to the orbital stability of the solutions of period 2n of
equation (10). As in §6.2, we calculate the Jacobian matrix at (a0, b0),

j _ Ti(4 — ro ~ 2ao) jP — 4a0b0 ~1

L ~2p~ Uob0 £(4 -rl- 2&o)J'

Therefore

p = trace J = ^e(2 - r£),

q = det J = ie2{-^(4 - r^)(4 - 3rl) + p2}.

Now, by using the results encapsulated in Fig. 6.4, the stability charac¬
teristics of the periodic solution (a0,b0) may be found according to the
values of /?, rl sketched in Fig. 7.5. Note that p > 0 below the line with
equation rl = 2, that q > 0 outside the ellipse with equation
 7.4 Subharmonics 225

^(ro — §)2 + 3/12 = 1, and that p2 = 4q on the lines with equation /•£ =
16/?2.
When there is a stable point (a0, b0) of equilibrium of system (14) there is
a stable solution of equation (10) with the period of the forcing, not the
period of the natural oscillation of the unforced system. This is another
example of entrainment. It is related to the synchronization of two clocks
when fixed on the same board, an apparently mysterious phenomenon first
reported by Huygens in the seventeenth century. □

4 Subharmonics
In solving problems of forced oscillations, we have hitherto confined our
attention to one simple equation and the weakly nonlinear theory of its
synchronous oscillations, i.e. its solutions with the same period as the
sinusoidal forcing term. This theory may similarly be applied to different
equations and to different solutions: for example, by taking a more general
periodic forcing term as a Fourier series rather than a single term, by
looking at the transients, and by ascertaining the stability or instability of
the periodic forced oscillations. Also there may be a periodic response
whose period is not the period itself, but an integral multiple of the period
of the forcing term. For a forcing term T cos cot there often, but not always,
exists a nonlinear oscillation of period 2nn/co, for an integer n greater than
one, called a subharmonic of order 1 /n. A brief example will serve to illus¬
trate this point.
Consider a forced hard spring governed by the equation,

d2x
—T + ax + <5x3 = r cos cot, (1)
at

for a, 6 > 0, i.e. by

+ ax + Sx3 = T cos t, (2)

where x = cot. There is in fact (because the nonlinearity involves only the
odd function x3) no subharmonic of order so we shall at once seek a
subharmonic of order 3, for which

x(t + 6n) = x(t) for all t. (3)

We may find x by weakly nonlinear theory for small <5, because we know
that such periodic solutions exist when 5 = 0, a = §a>2 and there is simple
harmonic motion. Accordingly, we expand
 226 Forced oscillations

x(/?,t) = x0(t) + <5xx (t) + <52x2(t) + •••, (4)

co(S) = co0 4- Saji + S2co2 + • • • as <5 -»0 for fixed a. (5)

Therefore coefficients of S° in equation (2) give

: d2x0
(On + ax0 = T cos r. (60)

Equation (60) has general solution

p COS T
*o(T) = a cos(y/<XT/co0) + bsin(y/(XT/(o0) + -—
a — (On

if (Oq # a. Then the periodicity condition (3) gives

(o0 = 3y/ix,

as anticipated from the linear theory. Therefore

r cost
x0(t) = acosjt + hsin^t —
8a ’
where a and b have yet to be found.
Next, coefficients of <5X in equation (2) give

,2d2*i d2x0
(°o~^r + axi = —2(o0(o1
dt2
*o (6,)

T cos tN
= — 2a>0(o1 ^ — 5a cos jt — %b sin ^t +
8a j

-l{°(“2 + b2 + 3&. ~^°2 - ^)jcosit

+ terms in cos t, sin t and other harmonics,

because cos3 jt = 4cost + f cos 31, sin3^t = f sinjt — jsint, etc. In order
that Xj has period 671, we require that its secular terms vanish. Therefore
the coefficients of cos jt and sin jt in the above equation vanish, i.e.

a{“2 + 1,2+s? - ~ w=«• (7.)

b{a2+b2 + m~ Tia,°m') + £ah = °- (7.)

 7.4 Subharmonics 227

The latter equation has one root b = 0, for which the former equation gives
either the null solution a = 0 or

8
2-27WoWi=0. (8,)
8a ° + 32a

If ft # 0 then the difference of a times (7S) and b times (7C) implies that

b = + Jla,

so that equation (7S) now gives

4a2 + — a + (82)
4a 32?-27“o“I=0-

Equations (8j), (82) have pairs of roots a differing by a factor of —2, be¬
cause we can rewrite (82) as

2 - =0.
(-2a)2-8S(-2‘')+32?
So each equation has two, one or no real roots according to the values of
T/a and oj0 col « (co2 — 9a)/<5.
We may invert the problem and regard co as a given forcing frequency
and find whether subharmonics of order 5 exist according to the values of
a, <5, and T. Also the stability of these subharmonics may be examined by
use of Floquet theory and the weakly nonlinear approximation; but it will
suffice here to state that some are stable.

Example 7.3: a weakly nonlinear oscillation near subharmonic parametric

resonance. Consider the equation

d2x
-jp- + {1 + €(/? -(- cos 2t)}x + x2 = 0. (9)

There is a rich variety of solutions of this equation, but here we confine our
attention to subharmonic oscillations of period 2n as e -*• 0 for fixed /?.
Thus impose the condition
x(e, t + 2n) = x(e, t) for all t. (10)

Without loss of generality, it is convenient to choose the phase, i.e. translate

the origin of time, so that

^= 0 at t = 0. (11)

Note that equation (9) is not of the form (1.1), because the coefficient of the
228 Forced oscillations

linear term x is a function of time with period n. The subharmonic reso¬

nance of order \ is of a slightly different kind, called parametric resonance,
because the period 27t/(l + e//?)1/2 of linear free oscillations is close to twice
the period n of the coefficient.
The problem may be solved asymptotically by using the expansion,

x(e, t) = e1/2x1/2(t) + €Xi(t) + e3/2x3/2(£) + • • • as e -*• 0. (12)

The reason for choosing powers of e1/2 rather than e will emerge later When
it becomes evident that weak nonlinearity and linear resonance affect the
solution at the same order, i.e. it becomes evident that the distinguished
limit has been achieved. How does one choose the ansatz without fore¬
knowledge? In practice, experience with trial and error is usually success¬
ful, at least in the long run, in finding a self-consistent asymptotic solution.
Now substitute expansion (12) into equations (9), (10) and equate co¬
efficients of e1/2. Therefore

d2x1/2 dx 1/2
-dp-+*>«
= 0, = 0 at t = 0.
d£

Therefore
*1/2 (0 = r1/2cost,
for some constant r1/2 to be found later.
Next, coefficients of e in equation (9) give

d2xx 2

= — ir2/2(l + cos2£).

The solution of this equation and condition (11) is

x^t) = —ir2/2(l — jcos2f) + ^cost.

Next, coefficients of e3/2 give

d2x.
Y2- + *3/2 = —2x1/2x1 - (0 + cos2£)x1/2
dt

— r3/2( 1 — 5 cos 2£)cos t — 2 r1/2rx cos21

— rl/2(P + cos2r)cosf

= rf/2(|cost - 5 cos 3t) - r1/2rx( 1 + cos2£)

— ri/2(P cos t -|- 5 cos 31 + j cos t).

 Further reading 229

To annihilate the secular terms in r3/2, as usual we require the coefficient of

cos t to be zero, i.e.

6^"i/2 - (P + i)rll2 = 0.

Therefore r1/2 = ±{f(/J + j)}1/2. It follows, in order that the solution be

real, that if P 4- j > 0 then e > 0, and the solution of period 2n is

*(M) ~ {s(P + i)e}1/2cost asej.0,

or the same solution n out of phase. If, however, P + ^ < 0 then e ] 0, and if
P = — 2 ^en a different method of approximation must be used. □

Let us review the results of this chapter. Duffing’s equation, and other
equations of forced oscillations similarly, have solutions with a very rich
structure. We have sought and found, by use of weakly nonlinear theory,
solutions with the same period as the forcing and solutions whose period is
an integral multiple of the period of the forcing. These solutions may also
exist for strong nonlinearity, and may be stable or unstable. In short there
may be many periodic solutions which attract and which repel neighbour¬
ing orbits in phase space. There may be bifurcations as the periodic solu¬
tions come and go, become stable and unstable, when the parameters vary.
In particular, if a synchronous oscillation becomes unstable and is suc¬
ceeded by a stable subharmonic oscillation of order | as a parameter in¬
creases then there is period doubling. All this is a background which might
lead us to anticipate the occurrence of chaos for strongly nonlinear solu¬
tions, and we shall see in §8.2 that chaotic solutions do indeed occur; but
first in the next chapter we shall introduce chaos with an even more famous
equation than Duffing’s.

Further reading

§7.2 Many good books cover the theory of this chapter. In particular,
Jordan & Smith (1987, Chap. 5) describe the theory of weakly nonlinear
forced oscillations not near resonance.
§73 Jordan & Smith (1987, Chap. 5) also describe the theory of weakly
nonlinear oscillations near resonance.
§7.4 Jordan & Smith (1987, Chap. 7) describe subharmonics and the
stability of forced oscillations.
230 Forced oscillations

Problems
Q7.1 A forced periodic oscillation. Consider the solutions of the equation

d2x 3
—+ x3 = r cos t
dr

which have period 2n. Taking x = a and dx/dt = 0 at t = 0 (without loss

of generality), so that x(t) = a cos t + higher harmonics, and neglecting
the higher harmonics, show that an approximate solution is given by
3a3 -4a = 4T.
Q7.2 Another forced periodic oscillation. Consider the solutions of the equation

d2x
—T + sgn x = r cos cot
dr
which have period 2n/co. Assuming that x(t) = a cos cot + b sin cot, and neg¬
lecting higher harmonics, show that an approximation to periodic solutions
is given by a = — (nT — 4)/nco2 if T > 4/n and by a = — (nT + 4)/nco2 if
T < -4/n.
Q7.3 A forced periodic oscillation for a nonlinear spring. Show that if

d2x ' -
-vs- + fix + ex-2 = T cos t
dr
and x has period 2n then

T cosr eT2 / 1 cos 21 \

x(e, t) + 0(e2)
n2 -1 2(n2 - l)2\i? + n

as € -* 0 for fixed non-integral Q.

Q7.4 Forced oscillations for a damped nonlinear oscillator with negative linear
stiffness. Show that if
d2x dx
"jr + ^ -fa + <*x3 = 0
dr2 dt

and a, /?, 5 > 0, then there is a saddle point at the origin and, if b1 < 80,
there are stable foci at (± (0/<x)1/2,0) in the (x, dx/dr)-plane.
Show that if, however,

d2x dx
-j~2 + o --px + ax3 = T cos cot
dr dr
then there are three oscillations of period 2n/co such that

x(r) = r {<5(0 sin cor - (co2 + P) cos cot}/{{co2 + fl)2 + (52<w2} + 0(r2),
x(t) = ±(/J/a)1/2 + T{bcosin cot - (co2 - 20)cos cot}/{(co2 - 2/S)2 + b2co2}

+ o(r2)
as f -»0 for fixed a, (}, S > 0.
 Problems 231

Q7.5 Resonance of a nonlinear oscillation. Find the leading approximations to the

solutions of the equation

d2x
+ fi2x — ex2 = T cos t
dr

with period 2n in the limit as e -> 0 (a) when T and non-integral fi are fixed
and (b) when ft2 = 1 + e/i and T = ey for fixed /?, y.
[Jordan & Smith (1987, p. 132).]
Q7.6 A subharmonic oscillation. Supposing that

d2x
—-5- + fl2X + €X2 = r COS t
dr

and x has period 4n, and expanding Q = Q0 + eQ, + e2Q2 + ••• and
x(e,t) = x0(r) -f ex^t) -I- as e ->0, show that Q0 = \ and x0(t) =
a cos jt + b sin jt — f T cos t. Thence find two equations satisfied by a, b, fij,
and solve them.
[Hint: Lagrange’s identity gives [udv/dt — vdu/dt]o" = Jo" {w(d2i;/dt2 +
— u(d2u/dr2 ^u)} dt for all continuously twice differentiable functions
u, v. Tryv = X! withw = cos\t and sinjtin turn.]
Q7.7 An exact subharmonic solution and its stability. Show that the equation.

d2x
+ x + <5x3 = T cos 3cot,
dr2

has a periodic solution x = X, where X(t) = (4r/<5)1/3coscor, co2 =

1 + 3(r2<5/4),/3.
Take x = X + x' for small x', and deduce that x' satisfies a linearized
equation of the form

d2x'
+ (a — 2<5fCOs2i)x' = 0,
dr2

where t = cot. Evaluate a, q in terms of S, /?, co.

*Use well-known properties of Mathieu’s equation (cf. Abramowitz &
Stegun 1964, Chap. 20) to deduce that the solution X is stable.
[McLachlan (1956, p.241).]
Q7.8 Another exact subharmonic solution and its stability. Show that the equation,

d2x
4- ax + 5x2 = r cos 4r,
dr2

for a, S, T > 0, has a solution x = X, where X(t) = C + A cosh, a2 =

16 + 4T<5, A = ± (2r/<5)‘/2, C = (4 - a)/2<5 < 0.
Show that the linearized equation for small perturbations x' of the
periodic solution X has the form
232 Forced oscillations

dV
+ (a — 2q cos 2 t)x' = 0.
df2

Evaluate a, q in terms of T, S.
[McLachlan (1956, p. 242).]
Q7.9 Yet another exact periodic solution. Given that the equation

d2x
+ (fi2 — 2 q cos 2t)jc + ex3 = 0
df2

has a solution X such that X(t) = a cos t for real constants a, q, e, Q, find q, fi
in terms of a, e.
Expressing x = X + x' and linearizing, show that

d2x' ,
~^2~ + (1 + ea2 + ea2cos2t)x' = 0.

*Deduce that X is stable if e > 0.

 8
Chaos

For the want of a nail the shoe was lost.

For the want of a shoe the horse was lost.
For the want of a horse the rider was lost.
For the want of a rider the battle was lost,
For the want of a battle the kingdom was lost—
And all for the want of a horseshoe-nail.
Benjamin Franklin (Expanded excerpt from Poor Richard's Almanack,
February 1752)

1 The Lorenz system

We shall in this section introduce chaotic solutions of ordinary differential
systems by detailed study of one system, which is often used as a prototype
for the study of chaos. In §2 Duffing’s equation with negative stiffness will
be discussed briefly: its periodic solutions, their instabilities, and the onset
of chaos. This leads to a quite general treatment of a common cause of the
onset of chaos in §3. The last two sections bring together many parts of this
book. The various sequences of bifurcations leading to chaos as a parame¬
ter of a differential system is increased are summarized in §4. Finally, the
diagnosis of bifurcations and chaos on the basis of experimental measure¬
ments of an unknown dynamical system is discussed. This is a task of great
importance and wide applicability; it is the inverse problem to all those
considered earlier in the book, requiring induction rather than deduction.

Lorenz (1963) studied a model of two-dimensional convection in a hori¬

zontal layer of fluid heated from below:
dx
-=-<rx + ay, (U

dy
— = rx-y-zx, (1„)

dz
— = -bz + xy. (U

233
 234 Chaos

in which x represents the velocity and y, z the temperature of the fluid at

each instant, and r, a, b are positive parameters determined by the heating
of the layer of fluid, the physical properties of the fluid, and the height of
the layer. These equations, the Lorenz system, with three dependent vari¬
ables and three parameters, have a great diversity of solutions with a com¬
plicated structure, more complicated than can be described in the few
pages available here. However, we shall describe the chief properties, em¬
phasizing the origin and nature of chaos.
*To specify the model a little more, we add that x, y, z are coefficients
of Fourier components of the velocity and temperature fields which are,
by use of no better than a fair approximation to solutions of the govern¬
ing equations of motion and heat, assumed to vary with height and the
horizontal coordinate in prescribed ways. In detail, the stream function
is il/{x*,z+,t) = x(t)sin(ax+/d)sin(nz+/d) and the perturbation of the
temperature field, i.e. the difference of the temperature from that of a state
of rest with a uniform vertical temperature gradient, is d{x^,z^,t) =
y(t)cos(axlt./d)sm(nz*/d) + z(t)s\n(2nz^/d), where x* is the horizontal co¬
ordinate, z* is the vertical coordinate, d is the depth of the layer of the fluid,
and a is a dimensionless wavenumber. Then r is the Rayleigh number of the
layer of fluid divided by its critical value for the onset of instability of the
basic state of rest (i.e. of the state with x = y = z = 0), a is the Prandtl
number of the fluid, and b = 4/(1 + a2). The Rayleigh number is a dimen¬
sionless measure of the imposed temperature difference across the layer,
representing the ratio of the destabilizing buoyancy forces to the stabilizing
forces due to molecular diffusion of momentum and heat. The Prandtl
number is the ratio of the coefficients of kinematic viscosity and thermal
diffusion of the fluid.
First note the invariance of the system under the transformation (x, y, z)
~+( — x,—y, z), i.e. under reflection in the z-axis. This symmetry expresses
the left-right symmetry of the thermal convection of the fluid. Also it can
be seen at once that the z-axis is an orbit, i.e. ifx = y = 0atr = 0 then
x = y = 0 for all t > 0; moreover, the orbits on the z-axis tend to the origin
as r —► oo.
We continue in the usual way by finding the equilibrium points and
their linear stability, and sketching a bifurcation diagram. So, putting
dx/dr = 0, we deduce that y = x, rx - y - zx = 0, xy — bz = 0. A little
elementary algebra now shows that an equilibrium point is either (a) the
origin 0:

x = y = z =0 for all r, (2)

or (b) one of the points C, C':
 8.1 The Lorenz system 235

x — y — +{b{r — 1)}1/2, z = r— 1 respectively for r > 1. (3)

Note the pitchfork bifurcation of the null solution at r = 1 (cf. Q5.6), if we

regard a, b as fixed and vary r. The other equilibrium points C, C' are
symmetrically placed with respect to the z-axis.
Linearization of the system for the null solution gives
dx'
— = -ax' + ay', (4J

(4,)

dz'
— bz'. (4J
d7
Therefore

z'(r) = z0e bt (5)

and x'(t), y'(r) oc esf, where

0 = S + <7
-r s + 1

= s2 + (a + l)s + cr(l — r).

Therefore

s = + 1) ± %{{o + l)2 + 4<r(r - 1)}1/2. (6)

Therefore the null solution is unstable (i.e. Re(s) > 0 for at least one mode)
if r > 1, and stable (i.e. Re(s) < 0) if r < 1; in fact, the three eigenvalues Sj
are such that sx > 0 > s2 and s3 = —b when r > 1, so that the origin is a
saddle point in three-dimensions.
Next linearize about the other equilibrium points C, C', defining
perturbations

x' = x + {b{r- 1)}1/2, / = y + {b(r - 1)}1/2, z'= z - (r - 1)

for r > 1. Therefore

dx'
— = ay -ax. (7.)

dL = x' -y' + {b(r — l)}1/2z', (7y)

= ± {b(r - l)}1/2(x' + y') - bz'. (7.)

236 Chaos

Therefore x',y',z' oc e*', where

s+a —a 0

0= -1 s+ 1 ±{b(r- 1)}1/2

+ {b(r — 1)}1/2 + {b(r-1)}1/2 s + b

= s3 + (p 4- b + l)s2 + b(o + r)s + 2ba(r — 1) (8)

= /(s), say, for all s.

To find the behaviour of the zeros sx, s2, s3 of the cubic / as r increases
from 1, and hence the stability of the equilibrium points, we must do a little
intricate, but elementary, algebra. Note that the coefficients of the cubic are
all positive, because r > 1 in order that the equilibrium points C, C' exist.
Therefore /(s) > 0 for all s ^ 0. Therefore there is instability (Re(s) > 0)
only if there are two complex conjugate zeros of /. Now it is easy to see
that when r = 1 the three zeros are s = 0, —b, — (a + 1), and therefore
there is linear stability or a margin of stability. The first zero gives s ~
— 2<r(r — l)/(cr + 1) as r j 1, so stability is lost in the limit as r approaches 1
from above. As r increases from 1, instability can set in only where Re(s) = 0,
and so where two zeros are = ia>, s2 = — ico for some real co. But the sum
of the three zeros of the cubic / is

si + s2 + s3 = —(coefficient ofs2 in /)

= —(a + b + 1).

Therefore

s3 — —(a + b + 1)

on the margin of stability, where sx = ico, s2 = -io. Therefore, on the

margin,

0 =/(-(* + &+ 1))

= {— (o + b + l)}3 + (<r + b + 1){ — (a + b + l)}2

+ b(a + r){ —(o- + b + 1)} + 2b<r(r - 1)
= r{-b{a + b + 1) + 2bo) - ba(it + b + 1) - 2ba,
i.e.

<t(<t + b 4- 3)
say. (9)
 8.1 The Lorenz system 237

So instability can arise only if <r, b are such that rc > 1. Thus the points C,
C' are stable if and only if either

(a) a <b -l-l and 1 < r or

(b) a > b + 1 and 1 < r < rc.

In fact, if there is instability then as r increases from 1 the following hap¬

pens: sx decreases from zero until it coalesces with s2 (when = s2 < 0),
they become a complex conjugate pair, and eventually their real part in¬
creases through zero, while s3 remains negative for all r > 1. We see that
each of the points C, C', when unstable, has one negative eigenvalue and
two complex conjugate eigenvalues, so that neighbouring orbits spiral out¬
wards from the point in the plane spanned by the eigenvectors ut, u2 at the
same time as they approach the point in a direction parallel to u3; this equi¬
librium point is an example of what is called a saddle-focus.
Weakly nonlinear theory shows that in fact there is a subcritical Hopf
bifurcation at C, C' when r = rc if rc > 1, for a range of values of a, b. So for
these values, which we shall assume in the account below, there is no at¬
tractor apparent as r increases through rc.
Let us summarize our results in a bifurcation diagram, Fig. 8.1, for this
case. Then, to prepare to understand what happens when r increases above
rc, we shall establish a few more elementary results.
We can use Liapounov’s theorem to show that the null solution is glob¬
ally asymptotically stable when r < 1 and stable if r = 1; for take the

Fig. 8.1 Sketch of the bifurcation diagram in the (r, x)-plane for the Lorenz
system with fixed b, a when rc = a(a + b + 3)/(a — b — 1) > 1.
238 Chaos

Liapounov function H(x,y,z) = \(x2 + ay2 + oz2). Then

dH dx dy dz
--— = x— + ay— + oz —
dt dr dt dt

= ax( — x + y) + <ry(rx — y — zx) + az( — bz + xy)

= —ox2 + <t( 1 + r)xy — oy2 — obz2

= -jo{ 1 + r)(x - y)2 - jo(l - r)(x2 + y2) - obz2

^0

for all x, y, z when r ^ 1, with equality if and only ifx = y = z = 0 when

r < 1 and x = y, z = 0 when r = 1.
We can deduce that the orbits at infinity are directed towards the neigh¬
bourhood of the origin as t increases, by a similar argument. We form

jtj{x2 + y2 + (z - r - o)2}

dy , dz
^_+(z_r_<7)_

= — erx(x — y) + y(rx — y — zx) — (z — r — o)bz + (z — r — o)xy

= — ox2 — y2 — bz2 + b(r + <r)z

<0 as x -> oo.

So %{x2 + y2 + (z — r — o)2} is a positive definite function which de¬

creases as t increases when x is large. Therefore the orbits there move
towards (0,0, r 4- o) in phase space.
Next, taking dx/dt = F(x) for the Lorenz system (1), we form

divF = TxG{y ~ x^ + dy(rx - y - zx) + fo(~bz + xy)

= -(o + b+l) (10)

<0.

So if n(t) is the volume of a set of points in the phase space of (x, y, z) at time
t and each point of the volume evolves according to the Lorenz system (1)
then n(t) -* 0 as t —► oo. It follows that the volume of any attractor must be
zero. Also there can be no source of volume such as a repelling equilibrium
point or limit cycle, because the sum of the real parts of the three eigen¬
values is negative. Therefore there is no quasi-periodic orbit; because a
 8.1 The Lorenz system 239

quasi-periodic orbit traverses the surface of a torus, and the torus is an

invariant surface, which cannot contain a diminishing volume without an
interior source.
The next task should be to sketch the phase portraits for various values
of r by synthesizing all the bits of information we have gleaned, just as we
did for two-dimensional autonomous systems in Chapter 6. However, the
task is much more difficult now because the topology of attractors is intrin¬
sically more complicated and more variable in three-dimensional space
than in two-dimensional space, and because visualization in our minds,
or sketching on paper, is more difficult for three-dimensional than two-
dimensional objects.
To approach the issue slowly, we next ask what the attractors are when
r > rc. We have found that no equilibrium point is stable; all orbits come in
from infinity; volumes n shrink as t increases; no attractor has finite vol¬
ume; and there is no quasi-periodic attractor. As r increases to rc there is a
subcritical Hopf bifurcation. Then where do the orbits go to when r > rc?
Numerical experiments reveal that there is a strange attractor with fractal
dimension D such that 2 < D < 3; it seems to be topologically equivalent
to the Cartesian product of a plane surface and a Cantor set. You can get
an impression of it by looking at the time series in Fig. 8.2 and the orbits in

Fig. 8.2 The graph of y(t) for r = 28, a = 10, b = f, x(0) = (0,1,0) and 0 ^ t <
30. Note that the ^-coordinates of C, C' in this case are ± 8.48.
240 Chaos

Fig. 8.3 Projections of part of an orbit (the same as in Fig. 8.2 but for
30 < t 70) for r = 28, a = 10, b = f in (a) the (x,z)-plane, (b) the (x, y)-plane,
and (c) the (y, z)-plane. Note that the coordinates of C, C' in this case are
(±8.48, ±8.48,27).
 8.1 The Lorenz system 241

40-

20-

(c)
Fig. 8.3 (cont.)

Fig. 8.3. The results come from numerical integration of the system (1),
with x(0) = (0,1,0) and r = 28, a = 10, b = f. (You can easily verify that
rc = 24.74 and C, C' = (±8.48, ±8.48,27) with these values of a, b.) First
the orbit goes once around C and then spirals out several times around C'.
It gets close to the attractor at about t = 20. The attractor can be seen as a
number of spirals around C followed by a number of spirals around C' and
so forth; the sequence of numbers of spirals seems as if it were a random
sequence. To visualize the attractor it may help to think of the groove on
two inter-connected gramophone discs which are rather warped - or to
think of a butterfly’s wings.
Calculations for r = 28, a = 10, b = f also show that neighbouring
orbits near the attractor separate exponentially on average, so that two
orbits which start very close together soon lose all correlation with one
another. Thus there is sensitive dependence on initial conditions. So if the
results of Figs. 8.2, 8.3 were recalculated with a different computer or a
slightly different program then the new results would probably diverge
from the ones shown until the two calculated orbits x(f) would soon be¬
come uncorrelated, although each would traverse the same attractor.
In fact the strange attractor can be found not only for some r > rc but
also for rt < r ^ rc, where rt > 1. So when rt <r < rc the strange attractor
 242 Chaos

period halving period halving

EP 2EPs 2EPs + SA SA SAorLC/LC SAorLC LC SAorLC/LC

6 1 24!o6 24!74 30.1 99.52 100.79 145 166 214.36 %

Fig. 8.4 Symbolic diagram specifying the attractors of the Lorenz system for
a = 10, b = f and 0 < r < oo. A stable equilibrium point is denoted by ‘EP’, a
stable limit cycle by ‘LC’, a strange attractor by ‘SA’, and the alternation of
attractors (usually with a complicated sequence of bifurcations) as r increases
by ‘or’.

coexists with the attractors C, C'. If r were to oscillate quasi-statically in an

interval including and rc then there would be hysteresis, with attractor C
or C' as r increases to rc, strange attractor as r increases further, strange
attractor as r decreases through rc to rt, and then attractor C or C' as r
decreases further. Also note that the strange attractor originates at r = rx
without being associated with the onset of instability of another attractor.
As r increases from zero to infinity there is a very complicated sequence
of bifurcations and attractors. In particular, there is a period halving of
stable limit cycles as r increases, whereby chaos ceases; this is the inverse
process of period doubling and it is governed by a Feigenbaum sequence
associated with the universal constant <5. Finally, as r becomes large a limit
cycle becomes the unique attractor (see Q8.3 for the limit cycle as r -*• oo).
An impression of the sequence of bifurcations is given in Fig. 8.4, based on
numerical calculations for a = 10, b = f. It is schematic and somewhat
simplified.
There is an even richer variety of nonlinear phenomena as r varies for
other pairs of values of a, b.
The ‘cause’ of chaos can be seen by examining the orbits geometrically.
To understand the changing topology of the orbits as r varies, try to see
computer animations of the orbits (e.g. Stewart FI987). There is space
here only to give an inkling of the cause. First we need to recognize that
the equilibrium points and the limit cycles are fundamental to the to¬
pology, just as they are in the phase plane. However, more topologically
different points of equilibrium occur in three-dimensions than occur in
two-dimensions according to whether the three eigenvalues are real or
complex and have positive or negative real parts.
*Let us see where the orbits near 0, C and C’ go to and come from when
1 < r < rc. (Remember that rc = 24.74 when a = 10, b = f.) Then 0 is un¬
stable, with > 0 > s2 > s3: it is a saddle point in three-dimensions. So
the orbits leaving 0 are parallel and anti-parallel to an eigenvector Uj
corresponding to sl5 and the orbits entering 0 locally lie in the space
 8.1 The Lorenz system 243

spanned by u2 and u3, i.e. a plane. Let us follow the orbits leaving 0 and
trace back the orbits entering 0. It helps to make two definitions first. If a
point lies on an orbit which tends to an equilibrium point X of an autono¬
mous differential system as t -*■ oo then the point lies in what is called the
stable manifold of X, denoted by WS(X); similarly, if a point lies on an orbit
which tends to X as t-* — oo then the point lies in what is called the unsta¬
ble manifold of X, denoted by WU(X). Thus if x(0) e WS(X) then the orbit
(x(t)} c WS(X); again, if an orbit starts at x(0) in WU(X) then the orbit
(x(r)} remains in WU(X) for —oo < t < oo. It may help to note that if X is
stable then WU(X) does not exist and WS(X) is the three-dimensional do¬
main of attraction of X (or, more generally, the m-dimensional domain in a
phase space IRm). Thus, the unstable and stable manifolds are in a sense
generalizations, suitable for an unstable point of equilibrium, of the do¬
main of attraction of a stable point of equilibrium. It follows that, for the
Lorenz system when r > 1, the manifold Wu(0) is a curve with tangent
parallel to ut at 0, and Ws(0) is a two-dimensional surface with tangent
plane at 0 spanned by u2 and u3. (When r = 1, ut = (1,1,0), u2 = (1, — cr, 0)
and u3 = (0,0,1).) Note that no orbit can cross a stable or an unstable
manifold except at an equilibrium point, so the two-dimensional surface
Ws(0) divides the orbits of [R3 into two sets. When r is not too large the
structure of Ws(0) is fairly simple. However, we shall see that, as r increases,
the structure soon develops twists and forms ‘sheets’, and becomes hard to
visualize.
*At first for r > 1 the curve Wu(0) leaves 0 in each of the octants of C, C'
and ends at C, C respectively. However, when r = r0 the curve Wu(0)
leaves 0 in each direction and returns to 0 in the surface Ws(0), as shown
in Fig. 8.5(a): two homoclinic orbits between 0 and itself are established.
(There are two because of the symmetry of the Lorenz system.) An orbit
connecting an equilibrium point with itself, approaching the point as t -*■
± oo, is called a homoclinic orbit, or homoclinic connection. It can be seen
that a homoclinic orbit corresponds to a solution of infinite ‘period’, be¬
cause although the connection is a closed curve its solution never reaches
the saddle point 0 in a finite time (cf. Example 1.1). It is this special proper¬
ty of the Lorenz phase portrait which defines r0 and is important in deter¬
mining the changing topology of the orbits. As r increases above r0 two
unstable limit cycles (of finite period) arise from the homoclinic orbits (it is
these which become the limit cycles which coalesce with the equilibrium
points C, C' as r increases to rc). By solving the Lorenz system numerically,
it can be shown that if a = 10, b = f then r0 = 13.93 approximately. When
r increases through r0 the unstable limit cycles serve to repel Wu(0) so that

j
244 Chaos

Fig. 8.5 A perspective sketch of the unstable manifold W“(0) in the phase space
of the Lorenz system, (a) The two symmetrically located homoclinic orbits
when r = r0. (b) The two symmetrically located heteroclinic orbits when r = r,.

the branch of Wu(0) leaving 0 in the octant of C ends not at C but at C', and
similarly the branch leaving in the octant of C' ends at C. As r increases
further, orbits cross over from C to C' and back an increasing number of
times before finally spiralling in to either C or C'. This is sometimes called
pre-chaos because it is difficult to distinguish an orbit that spirals around C
and C' repeatedly for a long time before eventually tending to one of those
points from an orbit that spirals around them for ever.
* Almost all orbits end at either C or C' for values of r such that these are
the only attractors (the exceptions are those orbits on Ws(0) and those on
the stable manifolds of the two unstable limit cycles). When r = rx, the
curve Wu(0) leaves 0 in one octant, eventually enters the unstable limit
cycle about C, and leaves 0 in the opposite octant and eventually enters the
unstable limit cycle about C, as shown in Fig. 8.5 (b): two heteroclinic
orbits between 0 and the limit cycles are established, a heteroclinic orbit
being an orbit connecting two invariant sets as t -* ± oo. It is this property
which defines rt and marks the onset of chaos. (The two heteroclinic orbits
are established at the same value of r because of the symmetry of the
system.) For a = 10, b = f it is found numerically that = 24.06. The
chaotic attractor originates as r increases above rt, so that for rx < r < rc
the chaotic attractor coexists with the attractors C, C' but for 1 < r ^ rx
the only attractors are C, C'.
* These facts are the bones of the skeleton which supports the flesh of the
‘body’ of the orbits. Try to picture the chaotic orbits in the three-dimen¬
sional phase space, and see the origin of chaos. It is far from easy.
 8.1 The Lorenz system 245

*Example 8.1: perturbation of a homoclinic orbit. Another inkling of how

the break-up of a homoclinic orbit, as a parameter varies, leads to chaos is
given in this example of an autonomous system in the phase plane. Howev¬
er, we find asymptotically conditions for the formation, not of chaos, but of
a stable limit cycle as the parameter varies a little.
The problem is this. Suppose that

(11)

where F: IR x IR2 -► [R2 is well behaved and there is a saddle point, at 0 say.

where 2,, X2 >0. Further suppose that there is a homoclinic orbit through
0, in the first quadrant, say, when a = 0.
Then the linearized equations at 0 are

and their solutions are

x(t) = x0 exp^j t), y{t) = y0 exp( - X21).

Thus orbits near 0 resemble rectangular hyperbolae, having equations

Consider next the orbit of system (11) which enters the small square
whose sides have equations x = ±e, y = ±e at (x„,€) in the first quadrant.
If 0 < e « 1 then the orbit leaves the square at (e, y„) approximately, where,
on defining r = A2/2l5 the above equation of the orbits gives

Now let us follow the orbit globally for small a. We suppose that when
a = 0 the unstable manifold leaving 0 tangentially to the positive x-axis
returns along the positive y-axis to form the homoclinic orbit. For small a
we may linearize perturbations of this orbit, and deduce in principle that
the orbit leaving the square at (e, y„) returns to the square at (xn+1, e), where

x„+1 = ka + lyn
246 Chaos

for some constants k, l. The linearization is valid because the orbit is every¬
where close to the homoclinic orbit. Therefore

xn+1 = ka + mx' forx„>0 (12)

where m = /e1_r. This difference equation gives us the dynamics, and, in

particular, shows that there is a stable limit cycle if the difference equation
has a stable fixed point. The methods of Chapter 3 show that, according to
the sign of r — 1 and value of a/cm1/(r_1), there may exist a stable fixed point
(see Q3.49).
Of course, a chaotic solution cannot arise in this example of a plane
autonomous system. However, now it is not difficult to conceive of a simi¬
lar problem in IR3, such that there is a cube with sides of length 2e centred
at a saddle point on a homoclinic orbit, and the perturbation of that orbit
for small a generates a map on a face of the cube which has a chaotic
attractor (cf. Glendinning & Sparrow 1984). □

2 Duffing’s equation with negative stiffness

Duffing’s equation offers another good illustration of an ordinary differ¬
ential system with chaos. In modelling the lateral vibrations of a beam
under a temporally periodic load, Holmes (1979) was led to consider the
equation,
d2x dx
7t + 3 —-fix + ax3 = / cos cot
d t2 dt (1)

for P > 0, i.e. for negative stiffness. The sign of the linear stiffness (- ft) is
important, because equation (1) admits no oscillations in the absence of
nonlinearity and forcing, whereas the form of Duffing’s equation with posi¬
tive stiffness in Chapter 7 does admit them.
However, equation (1) admits nonlinear oscillations in the absence of
damping and forcing, i.e. the equation

d2x
— px + ax =0 (2)
dr

admits them. The solutions of equation (2) can easily be found by the
methods of §1.7 and Chapter 6. The phase portrait in the (x, u)-plane is
sketched in Fig. 8.6(a), where v = dx/dt. With the addition of positive
damping, we consider

d2x dx
dF + idr/fot + “ 0 (3)
 8.2 Duffing’s equation with negative stiffness 247

Fig. 8.6 Sketches of the phase portraits in the (x, u)-plane for / = 0, a > 0,
P> 0. (a) Equation (2), with 6 = 0. (b) Equation (3), with <5 > 0.

for 3 > 0. Then there are no oscillations, the centres at (±(£/a)1/2,0) in the
phase plane becoming stable foci, as shown in Fig. 8.6(b).
*Note that the saddle point at the origin of the phase plane of Fig. 8.6(a)
has one orbit leaving it in the first quadrant, and that this orbit coincides
with the orbit approaching it in the fourth quadrant, i.e. its unstable mani¬
fold Wu(0) coincides with its stable manifold Ws(0) to form a separatrix
enclosing closed orbits. By the invariance of equation (2) under the trans¬
formations x > —x, t -> — t, the phase portrait is symmetric about each
axis. Therefore there is a similar separatrix in the third and second quad¬
rants. In other words, there are two symmetrically placed homoclinic con¬
nections of the saddle point with itself. (Note that the orbits form a figure of
eight, whereas the two homoclinic orbits of the Lorenz system, though also
symmetric, are like the outline of a butterfly’s wings.) The presence of
damping breaks up the homoclinic orbits, the saddle point at the origin
being structurally stable but the centres at (±(/?/a)1/2,0) becoming stable
foci. If 3 > 0 then the unstable manifold Wu(0) leaves the origin in the first
quadrant but approaches the focus at ((/?/a)1/2,0), and the stable manifold
Ws(0) approaching 0 in the fourth quadrant comes from infinity. Equation
(3) is symmetric under the transformation x-► — x, but not under time
reversal because of the damping with 3 / 0.
Equations (2), (3) have no chaotic solutions because they are autono¬
mous second-order differential equations, but, as we shall see, equation (1)
is equivalent to an autonomous third-order system and does have some
chaotic solutions as well as noh-chaotic ones.
The weakly nonlinear theory of forced oscillations gives insight into
some of the solutions. There is space enough here to consider in detail only
the synchronous oscillations, i.e. those with the same period 2n/a) as the
forcing. Take 3 = 0 at first, so that
248 Chaos

— /fa 4- ax3 = /coscot, (4)

and approximate the forced oscillations for small / by van der Pol’s meth¬
od as

x(t) = c(t) + a(t) cos cot -I- b(t) sin cot, (5)

where a, b, c are slowly varying functions. Then on substituting approxi¬

mation (5) into equation (4), neglecting the very small terms d2a/dt2,
d2b/dt2 and the higher harmonics, and equating coefficients of sin cut,
cos cot, 1 in turn, we find
*\

2<x>-^= —(P + co2)b + 3ccbc2 + fa br2,

dt

2co~ = (^ + co2)a — 3a c2a — \a.ar2 + /, (6)

dt

d2c
-r-K — pc — ixcJ — 4acr ,
dr2

where r = (a2 + b2)112.

This four-dimensional system has equilibrium points where

c(ac2 — /? + far2) = 0,

a(-p - co2 + 3ac2 + far2) = /, «- (7)

b{ — ji — co2 -f 3ac2 + far2) = 0.

Therefore equation (4) has small synchronous oscillations (a) about the
saddle point of equation (2) at the origin, given by

a(faa2 - fi - co2) =f, b = c = 0; (8)

and (b) about the other two equilibrium points (the centres) of equation (2),
given by

a(2p — co2 — ^a a2) =/, b = 0, ac2 = ft — fa a2 for a2 < 2/?/3a.

(9)

As seen in §7.3, the presence of damping can modify the cubic governing
the amplitude of an oscillation. It can be shown similarly that if S ± 0 then
equation (1) has small synchronous oscillations (a) about the saddle point
of equation (3) at the origin, given by

r2{(far2 — /? — co2)2 + S2co2} = f2. c = 0; (10)

 8.2 Duffing’s equation with negative stiffness 249

Fig. 8.7 The curves (a), (b) in the (/,r)-plane representing weakly nonlinear
solution of equation (1) of period 2n/co for a = 100, /? = 10, <5 = 1, w = 3.76.

and (b) about the foci of equation (3), given by

r2{(2/J - co2 - ^ar2)2 + S2co2} = /2, ac2 = /? - far2 for r2 < 2/?/3a.
(11)
Fig. 8.7 displays the curves (10), (11) in the (/, r)-plane for ‘typical’ values of
a, /J, 3, (o. Note that the curve (b) representing oscillations about the two
foci ends at r = (2/?/3a)1/2, where it meets the curve (a) representing oscilla¬
tions about the saddle point.
Of course, a periodic solution will be found after numerical integration
of its differential equation for a long time only if it is stable. So to find the
stability of the above weakly nonlinear solutions of equation (4) with 3 =
0, first linearize the fourth-order system about its equilibrium point of type
{a) with a = a0,b = c = 0. Then the linearized system is

0 — ($ — co2 4- faflo

Ifi + co2 — \ <xal 0

d2c'
w = (ft - MW-
Therefore the point is stable if ctal > §/? and 21<x2a% — 48(/J + co2)<xal +
I6((i + co2)2 > 0, i.e. if f/1 < aa„ < f(/? 4- to2) or ^(/? + to2) < atJo- These
results can be simply generalized for the case of damping (5 > 0). It follows
that in Fig. 8.7 the curve (a) represents unstable oscillations of period 2n/co
for r < (2/?/3a)1/2, =0.26, as well as between its two turning points (at
r = 0.34,0.56). Also the stability of solutions of equation (4) of type (b), i.e.
of the oscillations about the two centres of equation (2) without forcing,
•250 Chaos

can be found similarly. This provides a framework on which an interpreta¬

tion of the numerical results for the forced oscillations can be based. How¬
ever, there are weakly nonlinear oscillations which are not synchronous,
and the strongly nonlinear phenomena are more complicated.
When / 7^ 0 it is often convenient to use an autonomous third-order
system equivalent to (1) rather than to use (1) itself. So consider the sus¬
pended system,

dx dv , dd
—- = v, -—= (ix — Sv — txx* + f coscod, — = 1, (12)
dt dt dt

where OeS1, i.e. where a point on the circumference of the circle S1 of

length 2n/w is represented by its polar angle <x>9, and 6 -I- 2nn/co is
identified with 6 for n = +1, ±2, ...; this periodicity of 6 is helpful to
represent the periodicity of the forcing. The structural stability of equation
(3) ensures that the orbits of system (12) are topologically equivalent of
those of equation (3) in the phase space 1R2 x S1 of (x, v,9) for sufficiently
small /. Thus there are (cf. Q7.4) two stable attracting oscillations close to
(±(/?/a)1/2,0) and an oscillating saddle-type orbit close to (0,0).
In order to help to describe the orbits, next define a Poincare map P^°:
X -> Z, where £ = {(x, v,9) e IR2 x S1: 9 = t0 e [0,27i/cu)}, such that P^°
maps a point on an orbit of equation (1) at time t0 to the point at time
t0 + 27t/o> on the same orbit. Thus Pj-0 is effectively a stroboscopic map of
the (x,v)-plane. The value of t0 is significant because the forcing term /
cos cot varies with its phase. However, Pq° is defined and can be seen to be
independent of f0, being merely the flow ^2„/t0 for equation (3); it has three
fixed points, at the equilibrium points (0,0), (±(/?/a)1/2,0) of equation (3). A
solution of equation (1) having period 2pn/co for a positive integer p corre¬
sponds to a p-cycle of the plane map P/°, with different p-cycles for different
values of t0 in general.
*We can define the stable manifold MS(X) and unstable manifold MU(X)
of a fixed point X of a map F by analogy with the stable and unstable
manifolds of an equilibrium point of an autonomous differential system.
Thus MS(X) is the invariant set of F such that if x e MS(X) then F"(x) -♦ X
as n -+ oo, and MU(X) is the invariant set of F such that if x e MU(X) then
Fn(x) -> X as -co. Now we have noted that a fixed point X of P£>
coincides with an equilibrium point of equation (3). Also the stable mani¬
fold MS0(X) of the fixed point X of P'° coincides with the stable manifold
WS(X) of the equilibrium point X of equation (3); similarly the unstable
manifold Mg(X) of P£> coincides with WU(X) of equation (3). When, how-
 8.3 The chaotic break-up of a homoclinic orbit 251

ever, / is small but not zero, P}° has three fixed points which are no more
than close to the three equilibrium points of equation (3).
On the basis of the weakly nonlinear theory above, it can readily be
appreciated that, with development of subharmonic solutions and chaos,
the structure of the bifurcations of the solutions of equation (1) as /, a, /J, <5
and co vary is complicated. Indeed, it is astonishingly complicated. The
origin of some of the chaos is explained in Example 8.2 (other chaos arises
from period doubling). The large number of parameters makes it impracti¬
cal to describe here more than a few of the bifurcations. We shall describe
the bifurcations only as / varies for fixed a = 100, ^ = 10, S = l,co = 3.6,
after Holmes (1979). He sought the attractors by integrating the equation
numerically for long times and for various values of /. He found by the
method of averaging that the weakly nonlinear theory of the oscillations of
period 2n/co is a good approximation when / < 0.5 or / > 2.5, and that the
oscillations are stable then. Thus he computed the oscillations about the
origin for values of r above that at the upper turning point (i.e. for r >
0.56), and the oscillations about the foci for 0 < / < 0.95. At f x 0.95 the
oscillations of type (b) start period doubling. Strange attractors appear for
1.1 ;$/ ^ 2.5, but with windows of stable periodic oscillations, including
an oscillation of period 107t/ct», i.e. period 5, for 1.15 </ < 1.2.
A few solutions are shown in Fig. 8.8 of the phase plane. In Fig. 8.8(a)
the orbit of a large synchronous oscillation about all three fixed points is
seen coexisting with a strange attractor shown up by the stroboscopic map.
In Fig. 8.8(6) the orbit of a period-three solution circles first about the
right-hand focus, then about the left-hand focus, and finally about all three
equilibrium points of equation (3).
Many more details about Duffing’s equation and its chaotic solutions
have been given by Holmes (1979) and others (cf. Guckenheimer & Holmes
(1986, §2.2)).

*3 The chaotic break-up of a homoclinic orbit: Mel’nikov’s method

A common cause of chaos in ordinary differential systems, namely the
formation and break-up of a homoclinic connection as a parameter is
varied, is explained in this section. The essential topological aspects of the
break-up will be described and some general methods will be used, but at
the same time we shall take an illustrative case and show how to ascertain
the presence of chaos by an asymptotic method due to Mel’nikov (1963).
Consider then the case of a Hamiltonian system of one degree of free-
252 Chaos

(«)

Fig. 8.8 Sketches of orbits and Poincart: maps in the (jc, o)-plane of equation (2.1)
for cl = fi = at = 1, / = 0.3 (after Guckenheimer & Holmes (1986, Fig. 2.2.5)).
(a) A large stable orbit of period In jut and a stroboscopic Poincare map of
a strange attractor for 6 = 0.15. (b) A large stable orbit of period 6n/w for
<5 = 0.22, with three fixed points shown by circles.
 8.3 The chaotic break-up of a homoclinic orbit 253

dom with a periodic perturbation, namely

^ = F(x) + ef(x, f), (1)

where x = [x,y]T e 1R2, F: IR2 - IR2 with F = [dH/dy, -dH/dx]T for some
function H(y, x), f: IR2 x IR -*• U2 with f(x, t + T) = f(x, t) for all x, t, and F,
f are well-behaved. Thus F is Hamiltonian, and f has period T and may or
may not be Hamiltonian. The methods can be applied to perturbations of
systems with non-Hamiltonian F and to differential systems of order
higher than two, but equation (1) will serve well for illustration.
For e = 0 we find the basic system, i.e.

(2)

it is an autonomous Hamiltonian system, so its orbits lie on curves with

equation H(y,x) = E for different constants E. We shall in addition as¬
sume that the basic system has a saddle point, X0 say, with a homoclinic
connection. Let {q0(t)} be the homoclinic orbit, so that q0 satisfies equa¬
tion (2) and

q0(t) ->• X0 asr-^+oo. (3)

Note that qo(0) may be chosen to be any given point of the orbit, so that q0
is defined uniquely only up to a translation in time. Of course, q0(r) lies in
both the stable WS(X0) and unstable manifolds WU(X0) of X0 for equation
(2) for all t, and WU(X0) = WS(X0) along the homoclinic orbit. Equation (2)
having a two-dimensional phase space, these manifolds are merely plane
curves.
For e # 0 it is helpful to consider the suspended system, namely

(4)

This autonomous system, with the three-dimensional phase space of (x, 9),
is (cf. §5.1) equivalent to (1). We define a Poincare map P'°: Z'° -> Z'°for the
suspended system (4), where 2> = {(x, 6): 9 = t0e [0, T)}, as follows. Take
a point (x0,r0) as initial point of an orbit of the system, and integrate the
system from t = f0 to r = tQ + T to get the solution (P'°(x0), t0 + T); thus
P'° is again defined as a stroboscopic map. (We shall regard t0 as a variable
later, because / cos cot and hence P'° depend on their phases.) When e is
small, P'° has a fixed point, at X*'0, say, where X^° = X0 + 0(e) as e 0.
This fixed point corresponds to a periodic orbit of system (4), as illustrated
254 Chaos

Fig. 8.9 (a) The fixed points X'° of the map P'° and X0 of the map P£> in the
(x, y)-plane. (b) The periodic orbit of system (1) corresponding to the fixed point
X'° and the periodic orbit of system (2) corresponding to the fixed point X0.
Note that, by periodicity, the plane 6 = T should coincide with the plane 0 = 0.

in Fig. 8.9. For small e, the fixed point is a saddle point of the Poincare
map, and the distance from the periodic orbit to the unperturbed orbit, i.e.
the line x = X0, is of order e.
Recall that the stable and unstable manifolds of the equilibrium point
X0 of the basic system (2) for e = 0 coincide (see also Fig. 8.10(a)). This
leads us to enquire about the analogous behaviour for small e # 0. Before
using perturbation theory to find this behaviour quantitatively, let us
examine the possible types of behaviour geometrically. First denote the
stable manifold of the fixed point X^° of the Poincare map P'° by M^X*'0),
and the unstable manifold by M“(X'f°). In the present problem with a
two-dimensional phase space for equation (1), these manifolds are plane
curves. They depend on t0, moving and ‘waving’ with period T as t0 varies.
In particular, MS0(X0) = WS(X0) and Mg(X0) = WU(X0), so MS(X0) =
MS0(X0) along the homoclinic orbit. Also note that if x e M*(X'f°) then the
images (P'°)"(x) -► X'f° exponentially as n -► oo; similarly, if x e M“(X'°)
then the pre-images (P'°)n(x) -*• X‘° as n ->■ -oo.
The stable manifold Mse(X',°) and the unstable manifold M“(X'e°) either
do or do not intersect. If they do not intersect, then they are as shown in
Fig. 8.10(h), or (c). If they do intersect, then in general they intersect trans¬
versely, as shown in Fig. 8.10(d), and as first envisaged by Poincare (1899,
§397) with the prescience of a genius: ‘Let us seek to visualize the pattern
formed by the two curves and their infinite number of intersections... The
 8.3 The chaotic break-up of a homoclinic orbit 255

Fig. 8.10 Sketches of some possible configurations of the stable and unstable
manifolds of a fixed point of a plane map. (a) The homoclinic connection of X0.
(b), (c) Non-intersecting manifolds of X'°. (d) Transversely intersecting mani¬
folds of Xj°: a homoclinic tangle.

intersections form a kind of grid... with an infinitely tight mesh; each curve
never intersects itself, but must fold upon itself in a very complicated way
in order to intersect infinitely often the vertices of the grid’. This pattern is
now called a homoclinic tangle.
Let us see why a homoclinic tangle appears as it does. If the stable and
unstable manifolds of the Poincare map intersect at one point, x, say, then
all the images and pre-images of x must lie in each of the manifolds, be¬
cause each manifold is an invariant set of the map. Therefore the manifolds
intersect an infinity of times. Moreover, because an orbit on one of the
manifolds approaches or leaves the saddle point X'° of the map exponen¬
tially, successive images or pre-images get closer and closer together as
they approach the saddle point. Also continuity implies that a point on one
side of the stable manifold Mse(X'°) is mapped by P'° to a point on the same
side; similarly, a point is mapped to a point on the same side of the unstable
manifold. It follows that one ‘lobe’ between the two manifolds is mapped
by P'° into another lobe an even number of lobes ahead. So for systems (1)
which are area-preserving, or approximately so, the areas of neighbouring
lobes are equal, or nearly equal. Therefore the height of the lobes increases
256 Chaos

as they approach the saddle point, because their base decreases exponen¬
tially. The large height may lead to more complicated intersections of the
two manifolds than those depicted in Fig. 8.10(d), although a manifold
cannot intersect itself except at a fixed point, because an orbit of the differ¬
ential system (4) is unique for given initial conditions. This is an heuristic
explanation of the configuration of Fig. 8.10(d), which has been sketched
for illustration rather than quantitative detail. It should also be borne in
mind that the intersection of the two manifolds may be even more intricate
if they intersect on the ‘other’ side of the saddle point.
Consider next the successive images of points in the interior of a small
square near the saddle point, under iteration of the Poincare map. The
centre of the square will move slowly near the saddle point, while the
square itself is repeatedly contracted in the direction of the stable eigen¬
vector (i.e. in the direction parallel to the stable manifold) and stretched in
the direction of the unstable eigenvector (i.e. in the direction parallel to the
unstable manifold). Eventually the square will ‘escape’ from the neighbour¬
hood of the saddle point, move more rapidly around (near the homoclinic
orbit if e is small), and then return to the neighbourhood of the saddle
point if the manifolds intersect. As the square moves around, it will in
general be folded. So it may return near to its original position as a horse¬
shoe map (cf. §3.6) of its original shape, as indicated in Fig. 8.11; this prop¬
erty is, in fact, implied in general by the existence of the homoclinic tangle.
This leads to chaos, the square returning again and again as the iterated
horseshoe map of its original self.
After this qualitative description of the breaking of a homoclinic orbit
 8.3 The chaotic break-up of a homoclinic orbit 257

as e varies from zero, let us find the details quantitatively for small e. The
success of the asymptotic method, due to Mel’nikov (1963), depends on
knowledge of the global properties of the basic system (2); of course, in the
present case this system is integrable. It is already apparent that a crucial
issue is whether the stable and unstable manifolds intersect transversely,
leading to a homoclinic tangle and chaos. The method gives the distance
between the manifolds approximately, and hence a condition for their in¬
tersection.
Before launching into the technical details of the Mel’nikov method
with an elaborate notation, let us review the essence of the method. We
have already chosen arbitrarily a point on the homoclinic orbit of the basic
system (2), and taken the phase of the solution q0 so that qo(0) is the chosen
point. We shall proceed to find, when e is small, the distance between the
stable manifold M*(X‘e°) and the unstable manifold M“(X'(°) near qo(0) for
all t0. Although the distance depends on both qo(0) and the phase t0, we
shall, for convenience, regard it as a function of t0 for fixed qo(0). Thus we
shall find whether there exists a value of t0 such that the manifolds M*(X'€°),
M“(X'°) intersect transversely near our chosen point qo(0) in the section
Z‘°. This is equivalent to finding whether there exists a qo(0) such that the
manifolds intersect transversely in the fixed section Z'°, because t0 is mere¬
ly a phase of q0 and a translation in time is equivalent to a translation
around the homoclinic orbit. So, if there exists a value of t0 such that the
manifolds M*(X‘f°), M“(X'e°) intersect transversely an infinity of times then
they do so not only for this value of t0 but for all values of t0.
First find the equations of the orbits, say (qs({t, t0), f}, {q“(t, t0), t}, which
satisfy equation (4) and lie in the extensions of the stable and unstable
manifolds respectively in the three-dimensional phase space. There exist
uniformly valid expansions of these solutions of system (1) of the form

qS(f, t0) = q0(t - to) + eqs,■(*, t0) + 0(e2) for ^ as c _ 0 (5)

q“(t, t0) = q„(* - t0) + «qi(t. to) + 0(e2) for t ^ t0j
We can in principle find qse, and q“ likewise, by regular perturbation theory.
Noting that q* is a solution of system (1) and q0 of (2), linearize (1) about q0
for small e to get

_ J(,0(, - ,„)),■,«.<„) + f(q0(r - »„),«> (6’)

dr

for t ^ t0, where J is the Jacobian matrix of F. In addition, we require that

qsi(t, t0) -*■ lim^o {(x'e° - xo)/*} as t -* oo in order that qs€(t, f0) -*• K°-Also
q“ satisfies a similar equation (6U) for t ^ t0 and similar boundary condi-
258 Chaos

Fig. 8.12 The distance between the orbits on the stable and unstable manifolds
near q0(0) at time t0.

tion as t -► — oo. It will be seen that the form of equations (6s), (6U) but,
fortunately, not their explicit solutions qsl5 q“, is needed to find the condi¬
tion for the onset of chaos.
We can define the distance between Ms(X'e°) and M“(X'€°) by measuring
it, at least for small e, in the direction normal to the homoclinic orbit of the
basic system (1) at the point qo(0). Thus we define the displacement

d(to) = ^o) — ^o) (7)

= e{fli(^o> ^o) — <li(^o! to)} + ^(e2) as e —► 0.

For small e, the two points q"(f0,f0), q^o^o) are in general slightly
displaced from qo(0), and the manifolds are nearly tangential to the
homoclinic orbit at qo(0), as shown in Fig. 8.12. So we may resolve the
displacement d in the direction of the normal to the homoclinic orbit
at qo(0) in order to measure the distance between the manifolds. Now,
if F = [Ft, F2]t then the unit outward normal vector is n =
[ — F2(q0(0)), F1(qo(0))]T/|F(qo(0))|. So the Mel’nikov distance between the
two manifolds at q0(0) is defined as

D(t0) = dn (8)

= €F(qo(0)) a (q?(t0,t0) - q\(t0,t0)}

+ 0(e2) as e -♦ 0, (9)
|F(qo(0))|

where the wedge product of any given pair of vectors a = [>i,a2]T, b =

[&!, h2]T is defined as a a b = axh2 - a^.
To evaluate D, it helps to eliminate q“, qsj from expression (9) by using
equations (6U), (6s), because q?, qs are not known explicitly. This may be
done by first defining

As(t,f0) = F(q0(r-f0)) Aq\(t,t0). (10s)

 8.3 The chaotic break-up of a homoclinic orbit 259

It follows, on differentiating by parts, that

dAs(r, t0)
—37— = J(q0(f - t0))F(q0(f - to)) A qs,(t,t0)

dqsi(t,t0)
+ F(q0(f - t0)) A
dt

= J(q0(t - t0))F(q0(f - t0)) a qsi(t, t0)

+ F(q0(t - t0)) a {J(q0(t - t0))q\(t,t0) + f(q0(t - t0),t)}

= trace{J(q0(t - t0))}F(q0(t - t0)) a qi(t, t0)

+ F(q0(t - t0)) a f(q0(t - t0), t)

= F(q0(t - t0)) a f(q0(t - t0), t), (11)

because (Ja) a b+a a (Jb) = (trace J) (a a b) identically, and trace J =

8Fl/8x -I- dF2/dy = d2H/dxdy — d2H/8ydx = 0. Therefore

As(t0, t0) = As(oo, t0) - I

Jt0
F(q0(t - f0)) a f(q0(t - t0), t) dt. (12s)

Now As(oo,t0) = lim,^ {F(q0(t - t0)) a ql(t,t0)} = 0, because qs(t,t0) is

bounded and F(q0(t — t0)) -»• F(X0) = 0 as t -*• oo. Therefore

As(t0,t0) = - I 0
F(q0(t - t0)) a f(q0(t - f0), t) dt. (13s)

Similarly,

Au(t0,t0) F(q0(t - t0)) a f(q0(t - t0),t)dt. (13u)

Therefore equation (9) gives

eM(t0)
D(t0) - + 0(e2) ase-+0, (14)
|F(qo(0))|

where the Mel'nikov function is defined as

Af(t0)= I F(q0(t - t0)) a f(q0(t - t0),t)dt. (15)

J -OO

It can be seen that in general if M has a simple zero, x say, then D has a
simple zero near r for small e, and therefore the stable and unstable mani¬
folds intersect transversely at the point corresponding to t0 = t. This
implies that there is an infinity of intersections and a homoclinic tangle as
260 Chaos

in Fig. 8.10(d), and therefore that there is chaos. Conversely, if M has no

zero then D has no zero for small e, the manifolds do not intersect, as
shown in Fig. 8.10(6) or (c), and there is no chaos. Note also that D and M
in fact depend upon qo(0) as well as t0, but that they have zeros for all qo(0)
or for none.
It can easily be shown from equation (15) that M has period T, because f
has. This also follows from the identity P'0+r = P'° of the Poincare maps,
and verifies the result that one intersection of the stable and unstable mani¬
fold implies an infinity of intersections.
Finally, note that if f were Hamiltonian or were a constant then expres¬
sion (15) could be simplified a little; and if F were not Hamiltonian then
equation (11) would not follow, and the expression of D would be more
complicated, but the method would still be applicable. Again, if the order
of the differential system (1) were higher than two then the essence of the
ideas above would be applicable but their expression more complicated.

Example 8.2: Dujfing's equation. We next apply Mel’nikov’s method to

Duffing’s equation (2.1) with negative stiffness, after Holmes (1979), and
find conditions for the onset of some of the chaos described in §2. Equation
(2.1) with weak damping and forcing may be written in the form

dx dy
dt
= y, — = x — x + e(y cos cot — <5y),
dt
(16)

where y, 6, co > 0 and 0 < e « 1. On putting F = [y,x — x3]T, f =

[0, y cos cot — <5y]T, it can be seen that equations (16) have the form (1).
For e = 0, there is a saddle point at 0 in the phase plane, with two
symmetric homoclinic orbits (see Fig. 8.6(a)). The orbits of the basic system
(2) have equation H(y,x) = E, i.e. \y2 - \x2 + £x4 = E. It follows that the
homoclinic orbits have equation y2 - x2(l — |x2), and solutions

<lo(t) = ±(V2secht, — ^2 sech r tan t), (17)

on taking qo(0) = (± ^2,0). Therefore the Mel’nikov function is

m0) = F(q0(* - fo)) A f(q0(f - t0),t)dt

yo(t - t0) {y cos cot - 3y0{t t0)} d t

cos co(s + t0) - <5y0(s)}ds,

 8.4 Routes to chaos 261

on substituting s — t — t0 (which is an obvious, but frequently useful, sub¬

stitution to simplify the Mel’nikov integral),

= yj2y sin cot0 j* sech s tanh ssincosds — 2S \ sech2stanh2sds,

J -00

for the right-hand homoclinic orbit with y0(s) = - J2 sech s tanh s,

= V27rywsech(^7r«)sin(a>r0) - §<5. (18)

Therefore if 2yj2S cosh(^7rcy) < 3nyco then M has simple zeros and there is
chaos for small e # 0, whereas if 2yj2S cosh{^tcco) > 3nya> then M(t0) < 0
for all t0 and there is no chaotic break-up of the homoclinic orbit. □

4 Routes to chaos
We have met a few sequences of bifurcations leading to chaos as a parame¬
ter, say a, of a nonlinear system varies. There are many more sequences.
However, at the risk of oversimplification of research that is continuing, it
may be said that this transition to chaos occurs in one of four different
ways, i.e. the sequences may be divided in four classes.

(i) Subcritical instability. On this route a ‘familiar’ attractor, i.e. a point,

periodic or quasi-periodic attractor, becomes unstable as a slowly in¬
creases or decreases through a critical value, ac say, and the system then
‘jumps’ rapidly to a strange attractor which is not a continuous extension
of the familiar attractor. An example of this occurs for the Lorenz system
as r increases through rc; there is hysteresis because the fixed points are
stable if 1 < r < rc and the chaotic attractor is stable if r0 < r, where 1 <
r0 < rc. Indeed, the local theory shows that, as a increases through a value
ac where there is a subcritical turning point, pitchfork bifurcation or Hopf
bifurcation, the attractor ceases to exist so that a solution must go some¬
where else; it may therefore go abruptly either to another ‘familiar’
attractor or to a strange attractor.

(ii) A sequence of bifurcations. A route of transition to turbulent motion

of a fluid was first charted by Ruelle & Takens (1971), and later revised
by Newhouse, Ruelle & Takens (1978). Before the work of Ruelle &
Takens, the prevailing belief was that transition to turbulence occurs after
an infinity of bifurcations of increasingly complicated quasi-periodic flows
as a parameter, e.g. the Reynolds number, increases. In contrast, Ruelle &
Takens conjectured that a quasi-periodic solution with more than three
262 Chaos

fundamental frequencies is in general unstable, so that turbulence would

ensue after only a few bifurcations.
Turbulence is another of those things which are usually easier to iden¬
tify in practice than to define, but it may be said that it is a flow of a real
fluid which is chaotic in space as well as time. Our concern here is with the
onset of chaos rather than turbulence, and the two phenomena have been
confused, so let us study this route only in terms of differential equations.
We have seen how a stable steady solution (a point attractor in phase
space) becomes a periodic one (a closed curve S1) at a Hopf bifurcation,
and how the periodic solution may become unstable at another bifurcation
and be succeeded by a stable quasi-periodic solution with two fundamental
frequencies (on the surface of a torus T2) as the parameter a varies. Such
bifurcations may recur, with the appearance of a quasi-periodic attractor
with three, or even four, fundamental frequencies. It now seems that quasi-
periodic solutions with five or more frequencies are in general unstable.
Also, as a varies, the frequencies in general vary and thereby become ratio¬
nally related, so that for a quasi-periodic attractor with two fundamental
frequencies /l5 f2 the relation f2 = pfi/q becomes true for some integers
p and q; then the solution has become periodic and is no longer quasi-
periodic. Thereafter the ratio of the frequencies may remain constant as a
varies further, in which case we say the frequencies are locked (cf. Q3.29,
and the somewhat similar phenomenon of entrainment in Chapter 7). Then
after further variation of a, the ratio f2lfx may start again to vary until
chaos ensues. So the graph of versus a may be like a devil’s staircase
(cf. Q4.6) - the ratio f2/ft may vary continuously and monotonically, from
its value at the bifurcation in which the quasi-periodic attractor first re¬
places the periodic attractor, taking fixed rational values as a increases in a
sequence of intervals.
This whole route, somewhat simplified, may be summarized by the
scenario:

steady ——■* periodic -> quasi-periodic -»■ quasi-periodic -*■ chaos

P f .cl r r . t2 r r r . Tt

as a varies monotonically. The main properties of this route to chaos are

that chaos ensues after a short sequence of bifurcations with continuous
changes of the attractors.
But we have not yet specified the crucial bifurcation, namely the one
which brings chaos. The onset of chaos itself may follow the break-up of a
homoclinic orbit between an unstable point of equilibrium and itself (i.e. a
limiting orbit leaving the point and returning to it) or of a heteroclinic orbit
 8.4 Routes to chaos 263

between two equilibrium points. The Lorenz system at r — rx gave an example

of this in §1, although the chaotic attractor originates then without
continuous evolution from a quasi-periodic solution.

(iii) Period doubling. Feigenbaum analysed the infinite succession of peri¬

od doubling en route to chaos for one-dimensional difference equations.
This route is found in many differential systems, with the same universal
scaling (^although the phenomenon is essentially two-dimensional for
differential equations, because a Floquet multiplier cannot vanish (Q5.17)
and a complex multiplier must be one of a complex conjugate pair, and so
a multiplier has to go from 1 to — 1 in the unit circle of the complex plane
without passing through the origin as a varies monotonically from ar to
ar+j). Period doubling at a — ac is illustrated in Fig. 8.13, where the orbits
in the phase space of the variable x and the graphs of x(t) are sketched, x
being a typical component of x.

CO|QC
The Lorenz system as r decreases through 99.52 or 214.36 for a = 10, b —
is a differential system with period doubling. The logistic and the Henon
maps give classic examples of period doubling for difference equations.

(iv) Intermittent transition. On this route, first charted by Pomeau &

Manneville (1980), a limit cycle becomes unstable as the parameter a in¬
creases through a critical value, ac say. As a increases to ac the limit cycle

Fig. 8.13 Sketch of orbits in x-space and in the (t, x)-plane (a), (b) before (a ^ ac)
and (c), (d) after (a > ac) period doubling.
264 Chaos

becomes unstable to smaller and smaller disturbances. At a = ac it

coalesces with a similar unstable limit cycle, as at a turning point of a
Poincare map. For small positive a — ac the same cycle persists most of the
time, but is occasionally and rarely interrupted by ‘bursts’ which are not
small. As a — ac increases the bursts occur more frequently but do not
change much in magnitude. The average time between the ‘random’ bursts
tends to infinity like (a — ac)_1/2 as a j ac.
An example of this occurs for the Lorenz system as r increases through
166, when a = 10, b = f. Another example occurs for the logistic map
F(a,x) = ax(l — x) as a decreases through ac = 1 4- y/S. For a > ac there
is a stable and an unstable three-cycle; the cycles coalesce at a = ac, where
F3 has a turning point. For 0 < ac — a « 1 there is chaos of this intermit¬
tent character; although there is no three-cycle, there is a ‘ghost’ of the
three-cycle near which x„ may linger for many iterations before eventually
moving away, then returning to linger again, and so forth.
The difference equation xn+1 = x„ — x2 — a has a turning point at
a = 0, x = 0, and illustrates the non-chaotic phase of intermittency. There
are a stable fixed point (X = ( —a)1/2) and an unstable fixed point =
-( — a)1'2) if a < 0, one weakly stable fixed point {X = 0) if a = 0, and no
fixed point if a > 0. However, if 0 < a « 1 the origin acts as a ‘ghost’ fixed
point in the following sense. A point x„ stays very close to the origin for a
long time of the order of a-1/2 as a j 0, before eventually moving away from
the origin (Q8.10). It is this long proximity to the ghost of a fixed point of
a one-dimensional map near a turning point which represents the non-
chaotic phase of intermittency when the solution of a differential system is
nearly periodic, the one-dimensional map being a Poincare map of the
differential system. However, this particular difference equation is deficient
as a model of the chaotic phase of intermittency because it cannot reinject
x„ back near the origin once x„ has left the neighbourhood of the origin.

5 Analysis of time series

An important practical problem, with wide ramifications, is the interpreta¬
tion of signals from a nonlinear system. A ‘system’ may be a natural phe¬
nomenon, a laboratory experiment, or a numerical experiment. Often we
know little about the structure of a system, but can control some of the
parameters which specify it and can measure its output, or some of its
output. The ‘output’ is effectively the solution as a function of time. The
measurements are of various quantities, either over an interval of time or at
 8.5 Analysis of time series 265

a sequence of instants. If we are ignorant of the nature of the system, i.e. if

we do not know the equations which govern it, then we may seek to use the
measurements to learn about the nature, the physical mechanisms, etc. In
short the system may be like a ‘black box’, so we need to be able to use
measurements of its output to find the states of the system, their stability
and bifurcations as the control parameters vary. This may tell us some¬
thing about what is ‘inside’ the black box. These ideas can be applied to
phenomena ranging from sunspot cycles or weather patterns to growth of
a population of foxes or stock exchange prices. Here we shall discuss only
such nonlinear systems as we have met before in this book.
It is usually easy to recognize a stable steady state of a system, although
sometimes the presence of ‘noise’ may make an equilibrium point difficult
to identify unequivocally, especially if it is only weakly stable.
A periodic solution may be detectable by ‘eye’, i.e. merely by looking at
the graph, or a list, of measurements of one state variable as a function of
time. Confirmation of the periodicity may come from detecting the same
period, albeit different phases, in measurements of other state variables.
Again, noise may confuse matters. So may errors of measurement and
inaccuracy of the numerical processing of the data.
Quasi-periodic solutions are difficult to recognize as such merely by
looking at a graph of the output and chaos is even more difficult. Indeed,
the distinction between the time signal of a system in a chaotic state and
noise, i.e. the random errors of measurement and data processing as well as
the extraneous perturbations of the system itself, is not easy in principle or
practice.
We need more-objective, or, rather, less-subjective, methods to identify
the attractor by examination of the output. Some common methods of
diagnosis of the signal from a nonlinear system are summarized as follows.
Let us suppose that the signal measured is the time series d(t).

(i) Examine the time series by eye.

(ii) Examine correlations, e.g. d(s + t)d{s) = T-1 jjd(s -I- t)d(s)ds, where
T is a ‘long’ time. The graph of d(s -I- t)d(s) as a function of t may be
revealing. If d is periodic or quasi-periodic then so is its autocorrelation
d(s + t)d(s). But if d is chaotic and has zero mean then its autocorrelation
decays rapidly with t because of sensitive dependence on initial conditions
and the consequent effective independence of two parts of a solution unless
they occur at nearly the same time.
266 Chaos

(iii) Plot a phase plane, with a state variable d(t) as abscissa and d(t + t) as
ordinate, where t is the independent variable acting as the parameter of
the orbit and t is some suitable positive number. This is called a scatter
diagram in statistics, and the method of delays in the theory of nonlinear
systems. Packard et al. (1980) and Takens (1981) have proposed and ana¬
lysed this method of constructing a vector of whatever dimension is desired
from the time series of a single scalar. We need t to be neither so small that
d(t + t) is closely correlated with d(t), nor so large that d(t + r) is indepen¬
dent of d(t). We should be wary of aliasing in case, for example, there were a
periodic solution and t happened to be the period. It is often a good way to
choose r so that it is the least value for which the average d(s + c)d(s) = 0,
another is to use trial and error. An alternative method is to plot a phase
plane with two state variables, dft) and d2(t), say, as abscissa and ordinate;
provided, of course, that two variables can be measured. Then we have a
phase plane which gives some projection of an orbit, even though we do
not know the dimension of the attractor traversed by the orbit or of the
phase space in which the orbit belongs. However, experience at looking at
such projections helps the diagnosis of an attractor as a limit cycle (with a
closed curve in the plane), a quasi-periodic or a chaotic solution.

(iv) Make a Poincare section of the phase plane above or of other phase
spaces.

(v) Take a Fourier transform of the output. Given measurements of a state

variable, d say, over an interval of length T, we may compute the Fourier
series of d. Fast Fourier transforms offer an efficient computational meth¬
od to do this, so that a signal from an experiment may be processed ‘on
line’. Suppose then that a state variable d(t) has been measured for 0 ^ t ^
T. In practice there is a finite number of measurements at discrete instants;
say, then, that we measure d„ = d(nh) for n = 0,1,.... 2*, where h = T/2N is
the sampling time and N is an integer. Therefore the Fourier series may be
found as an approximation to the Fourier transform by taking h quite
small and T quite large. The accuracy of the spectrum of high frequencies is
limited by the smallness of the sampling time h, and of low frequencies by
the length T of the interval of measurement; we do not expect to resolve
frequencies/ > h~l or f < T~l.
It is common practice to plot the power spectrum Px(f) of a time series
x(t), where Px(f) = \x(f)\2 and x is the complex Fourier transform of x.
For a sinusoidal function there is a single peak of the spectrum at its
frequency fx. This is characteristic of a periodic solution just after its origin
at a Hopf bifurcation. But in general, for a well-behaved periodic function
 8.5 Analysis of time series 267

there are peaks of exponentially diminishing amplitude at fx (fundamental

frequency), 2fx (first harmonic), 3/j (second harmonic), etc. In an ideal mod¬
el with a continuum of precise measurements over an infinite interval, these
peaks have the zero width and infinite height of a delta function, but noise
in the system, in the measurements, and in processing of the signal gives
narrow peaks of finite height in practice.
To get a glimpse at the nature of the spectrum of a quasi-periodic func¬
tion, consider the example with d(t) = ax cos mfx t + a2 cos nf2t for integers
m, n and incommensurate fundamental frequencies ft, f2. Then take a
typical nonlinear term, say,

d2{t) = af cos2 njfi t + 2a xa2 cos mfxt cos nf2t + a\ cos2 nf2t

= |(a2 + aj) + jaj cos 2mfx t + jaj cos2 lnf2t

+ aia2{cos(w/1 + nf2)t + cos{mfx - nf2)t).

This, and other nonlinear interactions similarly, may generate not only
peaks at 2m/j, 2nf2 but also at ±mfx ± nf2. There are several examples in
§6.4 and Chapter 7 of this weakly nonlinear generation. Thus nonlinearity
in the equations of a system may generate not only all harmonics but also
all sum and difference frequencies. These can be identified in the Fourier
spectrum of output from a system, and hence lead to a diagnosis of a
quasi-periodic attractor.
Chaotic solutions have spectra with broad bands rather than isolated
peaks, and with a high noise level, but nevertheless may have prominent
peaks and a lot of structure. Chaos is not formless, and is usually far from
white noise.

*(vi) The dimension D of an attractor can be found in principle by measur¬

ing only a single state variable, say d, because the measurements of d(t)
imply the measurements of all derivatives of d in principle and thus of all
the state variables x e IRm of the solution. This may be better done by
constructing a Takens vector (d(t), d(t + x), ..., d(t + rt)), because numer¬
ical differentiation is subject to error. However, an infinite time series with
infinite precision is needed to specify the attractor with certainty, and in
practice experiments yield a finite time series with a lot of noise, so calcula¬
tions of D may be difficult unless D is 0, 1 or 2. It is usually easier, though,
to calculate the fractal dimension of the attractor from the output of a
numerical system because the noise is low. We see then that D = 0 for a
steady solution, D = 1 for a periodic solution on S1, D = 2 for a quasi-
periodic solution on T2 with two fundamental frequencies, and so forth,
but that D may have a non-integral value for a chaotic solution. Thus D
268 Chaos

gives the effective degree of freedom of the solution or the dimension of the
submanifold in which the attractor lies.

*(vii) Find the Liapounov exponents of the system. These are often a
means of calculating D. They also may serve to confirm the presence of
chaos.

Example 8.3: looking at time series and their analyses. To appreciate these
methods it is helpful to see their application to some specific cases. First
examine analyses of time series found by numerical integration of the
Rossler system. This is the system of Q5.5, and we have taken parameters
a = b = 0.2 (although these details are not important here). Now look at
the results of both time series and power spectra in Fig. 8.14. They suggest
that there is a periodic attractor when c = 2.6, that it has undergone period
doubling before c increases to 3.5 and quadrupling before c reaches 4.1.
The peaks of the spectra are not of infinite height and zero width because of
round-off and truncation errors in both the production and analysis of the
time series (of finite duration). The resultant noise dominates the signal at
the level of the troughs between the peaks. However, the peaks of the
fundamental and its harmonics are clearly identifiable in Fig. 8.14(a). Also
the appearance of the subharmonic of order \ is clearly discernible in Fig.
8.14(6), and of the subharmonic of order \ in Fig. 8.14(c). This analysis is
confirmed by the phase plots of Fig. 8.15.
Next look at the time series and its power spectrum in Fig. 8.16. The
data come from Ryrie’s (1992) numerical integration of a high-order sys¬
tem of ordinary differential equations devised by Curry et al. (1984) to
model thermal convection of a layer of fluid heated from below. The solu¬
tion appears to be quasi-periodic. It is not always possible to identify the
fundamental frequencies fr,f2 unambiguously in practice. However, at the
onset of a quasi-periodic attractor following the instability of a periodic
solution as a parameter increases, we may identify as the fundamental
frequency of the periodic solution and f2 as the frequency of its linear
instability (*given by the Floquet exponent), so that fx is the frequency of
the highest peak in the spectrum of the newly created quasi-periodic solu¬
tion. As the parameter increases further, it may become more difficult to
identify /, and f2 from the spectrum. In any event, a plausible interpreta¬
tion (Ryrie 1992) of Fig. 8.16 is that/! « 6.73, corresponding to the highest
peak, and f2 % 2.51.
The next case is of time series and a Fourier spectrum found by numeri¬
cal integration of the Lorenz system (1.1) for r = 28, o = 10, b = f, shown
 8.5 Analysis of time series 269

(a)

(b)

10

x 0

-10

(c)
Fig. 8.14 Solutions of the Rossler system dx/dt = -y — z, dy/dt = x +
dz/dt = | + z(x - c) as a time series x(t) and the logarithm of its power spec¬
trum versus frequency, In Px(f), where Px is the square of the modulus of the
complex Fourier transform of x, for (a) c = 2.6, a periodic, (b) c = 3.5, a period-
doubled, and (c) c = 4.1, a period-quadrupled solution.
270 Chaos
 8.5 Analysis of time series 271

-10
-10 10
x

(c)

Fig. 8.15 Solutions of the Rossler system dx/df = —y — z, dy/dr = x + ^y,

dz/dt = \ + z(x — c) in the (x, y)-plane for (a) c = 2.6, a periodic, (b) c = 3.5,
a period-doubled, and (c) c = 4.1, a period-quadrupled solution. Each orbit
moves clockwise as t increases.

Fig. 8.16 A quasi-periodic solution found by integrating a high-order differ¬

ential system which models themal convection: a time series ^(t) and the loga¬
rithm of its power spectrum versus frequency. In P^f) (after Ryrie (1992)).
272 Chaos

Fig. 8.17 The Lorenz system (1.1) for r = 28, a = 10, b = f: the time series x(t)
and the logarithm of its power spectrum versus frequency, In Px(f).

in Fig. 8.17. The erratic motion about the two unstable points of equilibri¬
um is indicated by the time series, and leads to the broad-band power
spectrum with a large number of closely packed peaks which fall off rapidly
as their frequency increases. This is characteristic of chaos. (The calcula¬
tions are not accurate enough nor sampled often enough to resolve the
high-frequency oscillations.)
These cases are of data computed by numerical solution of systems of
ordinary differential equations. As is typical of such cases, there is a very
low level of noise and the systems happen to be known. This is in contrast
to cases with data from laboratory experiments, even careful ones. Obser¬
vations of natural phenomena, such as meteorological and biological data,
usually have even more noise than laboratory observations do.
So the last case is of laboratory observations. They have been made by
Read et al. (1992) in experiments on the motion of a liquid in a differentially
heated rotating annulus, a model of the large-scale motion of the Earth’s
atmosphere. Fig. 8.18 represents one run of the experiment. Fig. 8.18(a)
shows a short excerpt of the time series for the heat transfer H (in watts)
across the liquid as well as the temperature T(in degrees Celsius) at a point
fixed in the middle of the annular container. Fig. 8.18(h) shows the power
spectrum PT of the temperature as a function of frequency / (in radians per
second). Fig. 8.18(c) shows a phase portrait of the orbit constructed in a
special way from the time series for T, and Fig. 8.18(d) shows a Poincare
section of the orbit. Read et al. explain the details. They interpret the series
as indicating that the system is in a quasi-periodic state with two funda¬
mental frequencies, because of the positions of the peaks in the spectrum,
because Fig. 8.18(c) looks like the projection of an orbit winding around a
torus, and because Fig. 8.18(d) looks like the section of a torus. Fig. 8.19
similarly represents another run. Fig. 8.19(a) shows some of the time series
 8.5 Analysis of time series 273

pT kr*

io^6

o 0.5

(b) /
Fig. 8.18 Analysis of experimental data of Read et al. (1992, run (e)) for a
quasi-periodic flow, (a) Time series for T°C and H watts versus f seconds.
(b) Power spectrum Pr(f). (c) Phase portrait for T. (d) Poincare section.
274 Chaos

(0

L.s

(4)
Fig. 8.18 (cont.)
 8.5 Analysis of time series 275

Fig. 8.19 Analysis of experimental data of Read et al. (1992, run (b)) for a
chaotic flow, (a) Time series for T°C and H watts versus t seconds. (f>) Power
spectrum PT(f). (c) Phase portrait for T. (d) Poincare section.

K
276 Chaos

■i i-1-r

(c)

t-r n

(<0
Fig. 8.19(co«f.)
 Problems 277

for T and H, Fig. 8.19(b) shows PT, Fig. 8.19(c) shows the phase portrait
constructed from the series for T, and Fig. 8.19(d) shows its Poincare sec¬
tion. Read et al. interpret the series as indicating that the system is in a
chaotic state, because of the broad-band spectrum, the convoluted phase
plot of the orbit, and the diffuse Poincare section. □

It is important to remember that, even if the phenomenon observed is

well-modelled by a system dx/dt = F(a, x) for F: IR' x IRm -*• IRm, in general
/, m, a, x and F are unknown, and we may be unable to measure x directly,
but instead measure p data of the form d(a, t) = g(a, x(f)) over some inter¬
vals of time for some unknown function g: IR' x [Rm ->• IRP. This poses a
formidable inverse problem: to find F, or at least something about F, or
even about one orbit x(f), from knowledge of d(a, t). It is as if we spend our
lives in a cave and seek to understand the world outside merely by looking
at some shadows cast upon the wall of the cave.

Further reading
§8.1 Sparrow (1982) describes the geometrical theory of the Lorenz sys¬
tem at length, although the existence of chaos for the system has yet to be
proved rigorously.
§8.2 Guckenheimer & Holmes (1986, Chap. 2) describe solutions of
Duffing’s equation with negative stiffness.
§8.3 Lichtenberg & Lieberman (1983, Chap. 7) and Guckenheimer &
Holmes (1986, Chap. 4) describe Mel’nikov’s method with greater genera¬
lity, rigour and detail.
§8.5 There is a well-developed statistical theory of the output of sys¬
tems, though these are traditionally assumed to be linear. The books by
Priestley (1981,1988) are recommended to learn this theory. There is also a
large literature on signal processing. King & Drazin (1992) have recently
reviewed the analysis of time series.

Problems
Q8.1 Oscillations about neutral equilibrium points of the Lorenz system. Show that
if r = rc, =ct(ct + b + 3)/(a — b — 1) > 1 then the eigenvalues governing
the stability of the equilibrium points C, C' of the Lorenz system are s =
±i{2bo(o + 1)/(ct — b - 1)}1/2, — (<r + b + 1). Find the eigenvectors belong¬
ing to these eigenvalues.
*Q8.2 Weakly nonlinear theory of periodic solutions of the Lorenz system. Defining
x = [x,y,z]T, rc = o(o + b + 3)/(a - b - 1) > 1, Xe = Yc = {b(rc - 1)}1/2,
Z„ = - L xc = [xc,Yc,zcy, x = x Xc, to, =
278 Chaos

2bo(o + 1)/(<t — b — 1)}1/2, e = |r — rc|, t = cot for real a>, J=

— CT <X 0

1 —1 —Xc, and X' = dX/dr, show that the Lorenz system may
y, xc -b _
se rewritten without approximation as

JX - wX' = [0, ZX + e(Xc + X), -xry,

according to whether r — rc = ± e.
Let u = [u, v, w]T denote an eigenvector (unique up to an arbitrary
multiplicative constant) of J and ^ of JT belonging to the same eigenvalue
icoc.
Assuming that there exists a solution X of period 2n in t such that

a> = wc + ecu, + ..., X = €1/2X1/2 + eX, + e3/2X3/2 + • • • as e -*• 0,

show first that a solution for X1/2 is given by

xi/2 = all2ueix + al/2 ue“h

for some complex constant a1/2 to be determined later.

Show secondly that

JXj o)cX1 — [0,Z1/2X1/2 + Xc, —X1/2yi/2]T.

Using the above solution for X1/2, deduce that

X, = Ujue" + axUQ~'n + 1 + |a1/2|2u10 + a?/2u12e2it + a?/2u12e"2it

for some complex constant a1, where

Ju10 = [0,wu + vvu, — (uv + ut;)]T,

Ju12 - 2icocu12 = [0,wu, -uu]T,

J1 = [0,-Xc,0]T.

Thirdly show that

^X3/2 — WcX3/2 = WlX'1/2 + [0, +X1/2 + Zl/2Xl + Z1Xl/2,

~(xmYi + xl y1/2)]T.
Now seek only the component, say qe‘\ of X3/2 proportional to eiT and
evaluate the expression Jq — io>cq. Taking the product of this expression
with ^T, deduce the solvability condition that

ia^u1^ = [0, ±u — (w/ + un) — |a1/2|2(vvu12 + uw12 + wul0 + uw10),

(um + vl) + |«i/2|2(ut;12 + vul2 + uvi0 + uu10)]£,

on using a (fairly) natural notation. Use this complex scalar condition to

 Problems 279

find both of the real quantities |a1/2|2 and a»,. Deduce a condition for the
periodic solutions to be subcritical, and verify numerically that they are
indeed subcritical when a = 10, b = §.
[Beware of the length of the algebraic manipulation needed to answer
this question.]
Q8.3 The Lorenz system for large r. Transform the system

dx dy dz
^= -.rx-y-zx, jt~xy-bz

into the system

d£ _ dn dC
^-{ii-«»(: + «>

without approximation, where e = r~1/2, £ = ex, rj = e2oy, £ = o(e2z — 1),

r — t/e.
Putting e = 0 for fixed <!;, t], £, show that

£2 - 2C = 2A, r\2 + ? = B2

for constants A, B of integration, and thence that

= F(a
where F is defined by F(£) = C + A£2 - and C is another constant.
Deduce that there exists a solution £(t) of period 2j|2 {F(^)}_1/2d^, where
fi, £2 are simple zeros ofF such that F(£) > Ofor^ < £ < <*2.
[In fact £(t) can be found in elementary terms of elliptic functions.]
Q8.4 The Rikitake two-disc dynamo. It is given that the self-generation of the
Earth’s magnetic field is modelled by the equations,

dx, dx2 dv
—- = -vx, + yx2, — = -vx2 + (y-a)x1, — = l-x,x2,

where X!, x2 represent the currents in two large eddies in the Earth’s core, y
the angular velocity of one eddy, a the positive constant difference of the
angular velocities of the two eddies, and v the ratio of the mechanical time-
scale to the magnetic diffusion time-scale of the dynamo formed by the two
rotating eddies and their electromagnetic interaction.
Show that the two equilibrium points are

x, = ±k, x2 = ±k~i, y = vk2,

where v(k2 — k~2) = a, and analyse their stability.

Show that ifx = (xj,x2,y) and the system is expressed as dx/dt = F then
div F = — 2v. Deduce that the volume of an attractor is zero.
Show that if v = 0, a = 0, then there exists a solution with xx = A cosh u,
x2 = A sinh u, ^y2 = u — \A 2 cosh 2u + B for constants A, B of integration.
 280 Chaos

Use a computer to explore some chaotic solutions of the sytem for a,

v > 0.
[Rikitake (1958), Allan (1962), Cook & Roberts (1970). The equilibrium
points represent the coincidence of the geomagnetic north pole with the
geographic north and south poles, and chaotic solutions the wandering of
the geomagnetic pole and its reversals.]
*Q8.5 The non-chaotic break-up of a homoclinic orbit. Show that the equilibrium
point of the equation

d2x dx
+ k-- x + x2 = 0
dt2 dt

at the origin of the (x, dx/df)-plane has a homoclinic orbit when k = 0.

Sketch the phase portraits for — 1 <k<0,k = 0,0<k< 1, labelling clearly
the homoclinic orbit, the stable manifolds and the unstable manifolds of the
origin.
*Q8.6 The stable manifold of a map of the unit square. The stable manifold MS(X) of
a fixed point X of a map F is defined as the set of points x such that
limn-co F"(x) = X.
Show that if the piecewise linear map F: T2 -* T2 is defined by

F(x) = [y,x + y]T modulo 1,

where x = [x,y]T, then 0 is a fixed point with eigenvalues j(l + yj5). Find
the corresponding eigenvectors. Deduce that Ms(0) is dense in T2.
*Q8.7 A weakly forced simple pendulum. It is given that the equation of motion of a
forced simple pendulum is

d20
+ sin 9 = e(a + y cos cot).
dt2

for a, y > 0.
Show that if e = 0 then there is a pair of heteroclinic orbits connecting
saddle points at (±7i,0) in the (0,^)-plane, where <f> = dQ/dt. Deduce that
one of these orbits is given by

0o(t) = 2 arctan(sinh t), = 2 sech t.

Show that the Mel’nikov function for the perturbation problem of the
intersection near (0,2) of the unstable manifold from (— 7t, 0) and the stable
manifold to (rc, 0) can be expressed as

Af(t0) <t>o(t — t0)(a + y cos cot) dt.

and deduce that

M(r0) = 2n{oi + y sech (\txjo) cos (cur0)}.

Hence show that there is chaos for small e if y > a cosh(^7i<y).

 Problems 281

*Q8.8 A Mel'nikov function of a Hamiltonian system which may be identically

zero. Show that if there is a perturbed Hamiltonian of the form H(y, x, t) =
Ho(y>x) + €/i(y, x, t) then the system’s Mel’nikov function may be expressed
as

M(t0) = {H0(q0(t - t0)), /i(q0(r - f0), t)} dr,

J -oo

on defining the Poisson brackets such that

m h) = d^--d-^-
°’ dx dy dy dx

Show that if //0(y,x) = ^(x2 + y2) — 3X3, h{y,x,t) = \x2 cos cut then the
resultant system is

dx dy
-j- = y, -~r = —x + x—ex cos cot.
df df

Show that if e = 0 then there is a centre at (0,0) and a saddle point at

(1,0) in the (x, y)-plane, with a homoclinic orbit given by

xo(0 = 2{3tanh2(^r) - 1}, y0(t) = jtanh(jt)sech2(jt).

Deduce that the Mel’nikov function is

Mfro) = -|sin(tor0) | sin(ajr)tanh(^t)sech2(^t){3tanh2(^t) - 1} dr.

'f

Define
f co r 00
11 = sin 2 cos tanh s sech2 s ds, /2 = sin 2 cos tanh3 s sech2 s ds,
J ~ao J -ao

and integrate f twice by parts to deduce that 3/2 = (2 — co2)/, and thence
that

M(t0) = f(l - aj2)sin(<ut0)/1.

Now deduce that if co = 1 then M(t0) = 0 for all t0, and comment on
whether chaos can be shown to occur for small e in this case.
*Q8.9 A Mel'nikov function of distance along a basic heteroclinic orbit. Suppose
that

dx . dy . .
—- = smxcoshy, — = — cos x smh y-f € sin cut.

Show that if e = 0 then there are saddle points at (mt, 0) in the (x, y)-plane
for ft = 0, ±1, ±2, .... and that (0,0) and (n,0) are connected by the
heteroclinic orbit (x0(t), 0), where

sin{x0(t)} = sech[t + ln{tan(|xm)}] for all t.

282 Chaos

Consider, when e is small, the separation near the point (xm,0) of the
unstable manifold from (0,0) and the stable manifold to (n, 0) of the
Poincare map P®, and show that the Mel’nikov function may be expressed
as

= —7tsin[a)ln{tan(^xm)}] sech(^Trcu).

[Cox et al. (1990).]

Q8.10 The non-chaotic phase of intermittency. Suppose that xn+, = F(a, xj for
n — 0,1,, where F(a, x) = x — x2 — a.
(a) Show that if a < 0 then there are two fixed points ±( —a)1/2, that
—( — a)112 is always unstable, but that ( — a)1/2 is stable if — 1 < a and its
domain of attraction is D(( —a)1/2) = (—( — a)112,1 + ( —a)1/2).
(b) Show that if a = 0 then there is a unique fixed point X = 0, and that
it is (weakly) unstable.
(cj Show that if a > 0 then x„ -» — oo as n -* oo for all x0. Show that if
x0 = 0 then

x„ = —na — £(n — l)n(2n — l)a2 — p3(n)nsa3 — p4(n)n7a4 —

as a 10 for fixed n, where p} are certain functions such that p,(n) = 0(1) as
n -» oo for j = 3, 4, — Deduce that x„ = 0(na) as a j. 0 for fixed n2a, and
thence that x„ = o(l) until n is of order of magnitude of a~l/2.
Q8.11 Autocorrelations. Show that if d(t) = c + a cos ft + bsin/f then d{s + t)d(s)
= c2 + ^(a2 + h^cosA^and that if d(t) = c + ax cos/,f + a2cosf2t for
fi ^ ±/i thend(s + t)d(s) = c2 + \ajcosf{t + ^a|cos/2t.
Q8.12 An exact integral of the Lorenz system. Show that the system (1.1) gives

dt
Deduce that if b = 2a, x0 = x(0), zq = z(0) then

x2-2az = (xq - 2oz0)e-2a’

for t > 0; discuss whether chaos may occur in this case.
[Segur (1982, p. 275).]
Q8.13 Moments of the Lorenz system and the closure problem. Given that the
moments xi = limr_^ | V (t) dr J etc. exist for j =1,2,..., integrate the
Lorenz system (1.1) to deduce that

y =x, zx = (r- l)x, xy = bz =x^ = (yT + b^/r forr*0.

Show further that x = 0 if the solution has the same symmetries as the sys¬
tem. What prospect is there of evaluating x2 etc. as functions of r, a, b by tak¬
ing these and other moments of the system and solving the resultant moment
equations algebraically?
 * Appendix: Some partial-differential problems

We have seen that bifurcations and chaos for a system of difference or

ordinary differential equations often occur in lower dimensions than the
dimension of the system. Similarly, although a partial differential system
has an infinite dimension, its bifurcations and chaos often occur in a mani¬
fold of low finite dimension. Indeed, turning points, transcritical bifurca¬
tions, pitchfork bifurcations, Hopf bifurcations, limit cycles etc. arise for
partial differential systems. This can be demonstrated in many cases by use
of one of a few perturbation techniques, for example the Liapounov-
Schmidt reduction or centre manifold theory. The essence of these tech¬
niques is to consider perturbations of marginal stability in which the values
of both the parameters and the state variables are close to those corre¬
sponding to marginal stability, and in which the effects of these two kinds
of perturbations are balanced asymptotically. At the margin of stability,
the number of eigenvalues whose real parts are zero is usually small, so
that their eigenfunctions span a low-dimensional space; all components of
an initial disturbance not in this space being strongly damped. The centre
manifold of a weakly nonlinear system is tangential to this space as the
margin of stability is approached.
The fact that phenomena of interest occur in a low-dimensional mani¬
fold makes the dynamics much easier to understand, but it seems that some
phenomena, for example turbulent motion of a fluid, cannot be represented
in a low-dimensional manifold.
In this brief introduction to nonlinear systems, we do not treat partial
differential systems in detail. However, a few problems which illustrate
instability and bifurcations are given below.

QA.l Derivation of an infinite ordinary-differential system from a partial-

differential problem. Consider the real equation

du d2u
dt dx2

283
284 * Appendix

with periodic boundary conditions,

«(x + 2n, t) = u(x, t) for all x, t.

Show that the problem is invariant if x -* x + a for a real constant a. Tak¬

ing the complex Fourier series

u(x,t)= £ un(t)einx,
n= —oo

show that = un and

^=-„2m+ £ m(n — rn)umun_m.

Deduce that the last equation is invariant if u„ -»einau„.

QA.2 Another nonlinear diffusion equation and a pitchfork bifurcation, (a) Consider
the system

du
ft \ - k~2
-m d2u . u = 0 at z = 0,7t,
dt

for k > 0 and some well-behaved function /. Suppose that it has a steady
solution u = U(z). Expressing u = U 4- u', and linearizing for a small per¬
turbation deduce that

du' d2u'
—-/'(t/)u'= u' = Oatz = 0,7i.

Taking normal modes of the form u'(z, t) = d(z)eSI, derive the eigenvalue
problem

d2d
^2 + K 1 {f'(U) ~ s}tf = 0, u = 0 at z = 0, ?t.

(b) Supposing further that /(0) = 0, /'(0) > 0, and U(z) = 0 for all z,
find all the eigenvalues s, and deduce that the null solution is stable if
K> K =/'(0) and the marginally stable mode is A = sin z.
(c) Show that if f(u) = sin u then kq = 1.
To perturb the marginally stable solution for this f first define e2 =
kc —k for e > 0 and x = e2r, so that

2du j.d2u
6 dt ~slnu = (Kc~e )fa2> u = Oatz = 0,7t.

Then assume that

“(z, K) = cA(t) sin z + e2u2(z, t) + e3u3(z, t) + • ■ ■ as € -»0,

and equate coefficients of e, e2, e3 in the above system to deduce that

 Some partial-differential problems 285

[Matkowsky (1970).]
QA.3 The Swift-Hohenberg model. Consider the real equation

for —oo < x < oo and t ^ 0, where R is a positive parameter. Find the
stability of the null solution by the method of normal modes and Fourier
components, i.e. by linearizing the equation, assuming that u oc es'+i*x for
real wavenumber k, and finding s as a function of R and k. Plot the values of
k versus R when the condition for marginal stability is satisfied. What is the
critical value Rc of R above which there is instability for at least one value
of kl
Find other solutions which are independent of x and t, and identify a
pitchfork bifurcation of the null solution.
Assuming that u has period 2n in x, and expressing the solution of the
equation as the complex Fourier series

u(x,t)= £ un(t)einx,
n= —oo

show that u_„ = un. Find an infinite system of ordinary differential equa¬
tions for {u„}. Verify the stability of the null solution and find when the
other solutions are stable.
[Swift & Hohenberg (1977) proposed this partial differential equation as
a model of nonlinear Rayleigh-Benard convection, i.e. convection of a hor¬
izontal layer of fluid heated from below, the same flow which led Lorenz
(1963) to his very different system. The null solution u = 0 represents the
state of rest of the fluid, and the other equilibrium points represent steady
convection cells.]
QA.4 A nonlinear Schrodinger equation. Given that

.du d2u ,
'aF + a? + '“I "“
show that there is a solution u - U where

U(t) = aexp{ia2(t - f0)}

for real amplitude a and phase t0.

Show that the linearized equation for a small perturbation u' is
286 *Appendix

where u = U + u' and «' is the complex conjugate of u'. Taking normal
modes of the form

u'(x,t) = U(r)/(t) cos{k(x - x0)}

for wavenumber k > 0 and real phase x0, show that

f' = i{(a2-k2)f + a2f}.

By expressing this complex equation as a real pair of equations for the

real and imaginary parts of /, or otherwise, deduce that / oc eSI where
s2 = k2{2a2 — k2), and hence that the periodic solution U is unstable if
k < y/2a.
[This equation governs the behaviour of a weakly nonlinear wavepacket,
and has many applications, e.g., in the theories of water waves and of plas¬
mas. Cf. Drazin & Johnson (1989).]
QA.5 The Ginzburg-Landau equation. Given that

du d2u
— = u - (1 + iR)|u|2u + (1 4- ib)j-z

and

—=0 at x = 0, /
ox

for real parameters R, b, and /, show that a solution is given by u = U,

where

l/(r) = exp{-iR(t - r0)}

for an arbitrary phase t0.

Show that the linearized equation governing the stability of the periodic
solution U is

d2u'
— = u' - (1 + iR)(2u' + U2W) + (1 + ib)~.

Hence show that there are normal modes of the form

u'(x, t) = U(t)f(t)cos(nnx/l)

for n = 0, 1,2,.... Find / and deduce that if b < 0 then the solution U is
linearly stable provided that R < Rc, where

Rc= ~{l +(l +b2)n2/2l2}/b.

[Kuramoto & Koga (1982). The steady form of the equation arose origi¬
nally in the theory of superconductivity (Ginzburg & Landau 1950), but
the unsteady form was found by Gor’kov & Eliashberg (1968); others have
applied it to weakly nonlinear modulation of unstable waves of many kinds.]
 Some partial-differential problems 287

QA.6 The chemical basis of morphogenesis. It is given that the concentrations u, v

of two interacting species of molecules satisfy nonlinear diffusion equations
and boundary equations of the dimensionless form

du dv
— =f{u,v) 4- V2u, r^=g(u, v) + dV2v in D,

8u dv
on 5D,
dn dn

where D is the domain occupied by the molecules, d is the (positive) ratio of

their diffusion coefficients, /, g are given well-behaved functions represent¬
ing the interaction, and the Laplacian operator is V2 = d2/dx2 -I- d2/dy2 +
d2/dz2.
Suppose that U, V are constants such that f(U, V) = g(U, V) = 0, and
let u', v' be a small perturbation of this spatially uniform equilibrium solu¬
tion. Then find the linearized problem satisfied by u', v'.
Assuming that the spatial eigenvalue problem

V2p + k2p = 0 in D, dp/dn = 0 on dD,

has known solution k, p(x), show that the normal modes of the linearized
problem have the form u'(x, t) = i2es‘p(x), v'(x, t) = ves'p(x), where

s + k2 -fu(U, V) -fv(U, V)
-gu(U,V) s + dk2 - gv(U, V)

Find U, V when f(u, v) = y(a — u + u2v), g(u, v) = y(b — u2v) for con¬
stants y, a, b > 0. Find the eigenvalues k and eigenfunctions p for the one¬
dimensional problem with D = {x: 0 < x < nl}. Discuss the stability of the
solutions in this case.
[Turing (1952) initiated such modelling, showing that diffusion can
cause instability; Murray (1989, Chap. 14) gives a modern review.]
QA.7 Spontaneous combustion of a slab. It is given that the temperature <j> of a
one-dimensional slab of a uniform solid is governed by a dimensionless
problem of the form

d<f> d2<t>
+ 8e*, <f> = Oat* = ± 1,
~dt dx2

where the Frank-Kamenetskii parameter 8 > 0.

Seeking steady solutions of the form <f> = F(x), verify that

F(x) = 2 ln(cosh c sech cx),

where c is such that <5 = 2c2 sech2 c, coshc = exp(^m), <j>m = F(0). Sketch
the bifurcation diagram in the (<5,^m)-plane, identifying a turning point at
(2(cq — l),21n(coshc0)), where c0 « 1.2 is defined as the positive root of
ctanhc = 1.
288 ♦Appendix

Show that the linearized problem for small perturbations of this steady
solution is

= ^4 + 6cF<t>’, <*' = 0atx=±l.

dt dx

Taking normal modes with <t>'(x, f) = <p(x)est, obtain the Sturm-Liouville

problem,
d (b - A A
—^ + (2c2 sech2 cx — s)<f> = 0, ^ = 0 at x = +1.
dx2

Deduce that s is real, and so s increases through zero at any margin of

stability.
Defining y(x) = tanh cx, verify that if s = 0 then the eigenfunction is

Hx) = iyln{(l + y)/(l - y)} - 1

and coth c tanh(coth c) = 1, i.e. c = c0.

Show that <t> is an associated Legendre function of y for general values
of s.
[Buckmaster & Ludford (1982, Chap. 12) describe the model of the slab
and its steady solutions.]
QA.8 The Proudman- Johnson equation. Consider the system,

d3f _ 1 d*f , ,d3f dfd2f 5/_ f __u1

dtdy2 Rdy*+^dy3 8y dy2’ * +1’5y 3t y ±1,

for positive parameter R.

Supposing that f = F is a steady solution of the system for a given
function F(y, R), taking small unsteady perturbations g so that / = F + g,
and linearizing, show that

d3g 1 84g 83g 82g dg

—— =-- + f _ F' z _ p" — 4- F"'a = 0 at y = ± 1,
dtdy2 R dy4 dy3 dy2 dy

where a prime denotes differentiation with respect to y. Then, taking nor¬

mal modes with g(y, t, R) = es'G(y, R), derive the eigenvalue problem,

Giv + R(FG"' - F'G" - F"G' + F"'G) = RsG", G = G' = 0 at y = ± 1.

Deduce that the steady solution F is unstable if there exists an eigenvalue

such that Re s > 0.
Show that a solution is given by F(y) = y for all R. ^Deduce that

G = bty + b2 + exp(iRy2){b3U(s - i^/Ry) + b4V(s - IjRy)}

for some constants bt, b2, b3, b4, where U, V are the standard parabolic
cylinder functions. Hence or otherwise show that there is stability for R <
Rt, where = 0,andsoR, a 4.51.
 Some partial-differential problems 289

[The system governs the unsteady non-parallel flow, at Reynolds num¬

ber R, of a viscous incompressible fluid in a channel with suction at moving
porous parallel walls (cf. Watson et al. 1990).]
.9 Blow-up. Consider the nonlinear diffusion equation

du d2u ,

where du/dx = 0 at x = 0, 1 and u(x,0) = u0(x) for 0 < x ^ 1 for a given

smooth real function u0.
Taking Fourier expansions

00 00

u = 2a0 + X ancos nnx = \ £ a»*'n*x for real a-n = an,

ft—l n=
-ao

00 00

du/dx = £ bn sin nnx, d2u/dx2 = $c0 + t c„ cos nnx

»=1 n=l

at time t > 0, show that bn = —nna„, cn = —n2n2a„. Deduce that

da 00
~Tt~ = + 2 I aman_m for t > 0.
m=- oo

Now show that

Deduce that if ao(0) > 0 then there exists a positive constant k such that

a0(t) ^ 2/(k — f) for all r > 0

for which the solution exists. Hence show that if JoU0(x)dx > 0 then the
solution u(x, t) becomes singular in a finite time.
[This kind of singularity is called blow-up, and is described more gener¬
ally for nonlinear partial differential equations by, for example. Palais
(1988).]
 Answers and hints to selected problems

By day and night he measured and calculated; covered enormous quantities

of paper with figures, letters, computations, algebraic symbols; his face,
which was the face of an apparently sound and vigorous man, wore the
morose and visionary stare of a monomaniac; while His conversation, with
consistent and fearful monotony, dealt with the proportional number n....
Thomas Mann (The Magic Mountain, Chap. VII, The great god dumps’)

The answer, if one is given, to a problem is denoted by the prefix A; for

example, the answer to Q 1.1 is Al.l. In some cases a hint to the solution is
given. Of course, when a reference is given in the text it is more useful than
the brief answer here.

Chapter 1

Al.l X = 0 V a is stable for a < c. X = (a — c)/ab Va ^ 0 is stable if a > c; trans-

critical bifurcation at (c, 0).
A1.2 x(0)| X as f -mx), where X = h[l — {1 — 4ac/b2}1/2]/2c is least positive
zero of right-hand side of equation.
A1.3 X = 0 V a is stable if a < 0. X = b + (b2 + a)112 V a > — b2 is stable. X =
b — (b2 + a)1/2 Va Js — b2 is stable V a > 0. Hysteresis in oscillations.
A 1.4 J[oexp(3es2)ds ~ a_1N/eexp(ja2/e)ase|.0forfixeda > 0, T0 < 0.
A 1.5 F(a,x) — 2(a2 — x2) — (a2 -(- x2)2 is symmetric in ±a, & in ±x. F(a,x) —
2r2 cos 20 — r4 with a = r cos 0, x = r sin 0. F(a, x) = 0 has turning points
at (± sj2,0), transcritical bifurcation at (0,0).
A 1.6 F(a,x) = x3 + a3 — 3ax is antisymmetric in a, x. F(a,x)-0 has
asymptote x + a = — 1, turning point at (22/3,21/3), transcritical bifurcation
at (0,0).
A 1.8 d20/df2 = a>2f(k,0), where k = ato2/g > 0, 0 ^ 0 < 2n, f(k,0) =
sin0(cos0 — l/k). f(k,&) = 0 implies 0 = 0, 7tV/cor© = ±arccos(l/fc)

290
 Answers and hints to selected problems 291

V k > 1. Test sign of fe(k,&) for stability. © = 0 stable V k < 1; 0 = n

always unstable; © = ±arccos(l/fe) stable V/c > 1.
A1.9 d20/dr2 =f(k,6), where f(k,0) = (am)~l {k(2a cos — /)sin — mgfsinfl},
0 ^ 9 < 2n, a, g, k, l, m> 0. f(k, 0) — 0 implies 0 = OVfcor© = + a V
k>mg/(a — \l), where a = 2arccos{lc//2(afc — mg)}. 0 = 0 stable V
k < mg/(a — £/), © = + a stable V k > mg/(a — 5/).
A1.10 0 = aor7r — aV/c< 2/n, where 0 < a = arcsin(^Trfc) < 571; former equilib¬
rium is always stable, latter unstable, turning point at (2/tt, ^7t).
A1.12 dr/dt = 1 — r2, d0/dt = 1.
A1.13 dr/dr = r(l - r2)2,d9/dt = 1.
A1.14 dr/dt = rf(r),d9/dt = 1,/(»)) = 0, r; stable only for even;'.
A1.15 dr/dt = r(l —f/a), dO/dt = 1. dr/dt = r{1 — g(r)/a}, g(r0) = a. (0,0) is
stable for 0 < a < §, unstable for a > f, with Hopf bifurcation at (§,0). For
j < a < 2 3a stable (r0 < 1) & an unstable (r0 > 1) limit cycle; for 1 < a < §
3 an unstable (r0 > 1) limit cycle.
A1.16 T = 4jo l/{2(£ — x)}1/2 dx.
A 1.17 ^(dx/dr)2 + jx2 — |ex4 = \a2 — £ea4. Substitute x = acos</> in T=
jodx/[(a2 - x2){l - je(a2 + x2)}]1/2.
A1.18 T = 4c_115/2(1 + ea2 cos2 0)a cos 0/(ea cos 0 + je2a4cos40)1/2 d0 if x =
a cos 9.
A 1.19 A = B = 0 for a solitary wave.
A 1.21 (a) Cf. Example 1.2. (b) Linearized equation is \p" + a>2\p = 0, so ip =
sin cus. (os = I# dt///{4(cos 1p — cos a)}1'2.
A 1.23 Express [a,h] = I, [m, M] = F(I). Then F(I) 3 I implies that m < a < b <
M. Therefore 3 c, d e I such that F(c) = a, F(d) = b. Therefore F(c) — c <
0 ^ F(d) — d, & intermediate-value theorem gives at least one zero of F(x)
— x in [c,d],
A1.25 By 2nd mean-value theorem, 3 £e(a,b) such that 0 =f(X) = /(x„) +
(X - xn)f'(xn) + \{X - x„)2/"(£). But f(x„) = (xn - xn+1)f'(xn). Therefore
\xn+1-X\ = ±f"m*n-X\2/f'(xn).
A 1.27 F(x) = r + 6/(r + x). X = ±Ja. F'(±s/a) = (r + yja)/(r ± Ja). Ja is
stable, — yja is unstable.
A1.28 X is stable fixed point of G if - 2 < a {F'(X) - 1} < 0.
A 1.29 Define F(x) = x + hf(x). |F'(0)| < 1 iff -2 < hf'(0) < 0.
A1.32 To estimate x„ as n -* 00 heuristically, assume that x„ ~ bnp as n -*• 00, test
this ansatz for consistency, & deduce that p,b = 1.
A1.35 s,, s2 = —b ± (b2 — c)1/2, uy- = [l,s,]T. b — —j,c= — 1, x0 = 0, Xj = 1,
s„s2 = i(l ± V5),^i,^= ±1/V5.
A 1.36 On differentiating a function of a function, G'(X) = F'(X)F'(F(X)) =
F'(X)F'(Y), =G'(Y) similarly.
A1.37 X = {— 1 ± (1 4- a)ll2}/a; former point is (just) stable, the latter unstable.
F2(x) = x V x ^ — 1/a, so only cycles are the 2-cycles {x,(l — a)/(l + ax)} V
x # - 1/a, X.

L
292 Answers and hints to selected problems

Chapter 2

A2.2 Use implicit-value theorem. A zero of H is either a fixed point of G or a

member of a 2-cycle of G.
A2.3 ,4 = 0 or a = A + n2n2. Pitchfork bifurcations at (n2n2,0).
A2.4 (i) Imperfect transcritical bifurcation, (ii) Imperfect transcritical bifurcation
& turning point, with hysteresis for <5 < 0. (iii) Isola for 3 > 0, isolated point
for 5 — 0, no solution for 3 < 0.
A2.5 F = 0 gives x = 0 or parabola (x + ^)2 = a + j<52 with nose at a = —id2,
x = — j8; i.e. turning point at ( — id2, — jd), transcritical bifurcation at (0,0)
if <5 # 0.
A2.6 F(l,b, 0,x) = 0 gives x = 0 or parabola (x — l)2 = b — 2, with transcritical
bifurcation at b = 3, x = 0 & turning point at b = 2, x = 1. F = Fx = 0 gives
3{2 + 3c — fb — fa(3 — — b)}2 = 4(b + fa2 — 3)3, a curve in (b,c)-
plane with cusp at b = 3 — fa2, c = — yja3. On eliminating a, (b — 3)3 =
-21c2.
A2.7 x = X, dx/df = d2x/dt2 = 0 gives gf'(X) — a>2(X — da) — 0. Linearization
gives [1 + {/'(A')}2]d2x'/dr2 -I- {gf"(X) — a>2)x' = 0, & so stability
according to sign of gf"(X) — a>2. For the given /, using dimensional
analysis, define u = X/a so F(u,b,8) = 0, where F(u,b,S) = u3 - — l)u
+ jdb. Imperfect pitchfork bifurcation at (1,0) in the (b,u)-plane. Eliminate
x from F = Fu = 0 to find cusp. Cf. Examples 2.2 & 2.5.
A2.8 The triple root occurs when the cubic has the form (v — i>c)3 = 0. F = F^ =
0 gives the equation of curve in (0,7t)-plane with cusp at origin.
A2.11 If e > 0 then (a — r2)r + e = O& 0 = O, n respectively, & so there is half
(r ^ 0) of an imperfect pitchfork in the (a, r)-plane for 9 = 0, & half for 9 = n;
however 9 = 0 is stable & 9 = n unstable V a, so the turning point for 9 = n
separates 2 unstable solutions.
A2.12 f(x) = f(X) + J(X)(x — X) + nonlinear terms. The Jacobian matrix of F at
X is 0.
A2.13 Equation of intersection (with x as a parameter) is b = — 4x3 — 2ax,
c = 3x4 + ax2.

Chapter 3

A3.4 Use mathematical induction.

A3.5 Xx is a stable fixed point of G = Fp if 1 > |G'(A',)|.
A3.6 X = 0 V a, stable if -1 < a < 1; or X = ±(a - 1)1/2 V a > 1, stable if 1 <
a <2. The three 2-cycles are {(a + 1)1/2, -(a + l)l/2} V a > -1, always un¬
stable; {± [f{a + (a2 - 4)1/2}]1/2, ± [f {a - (a2 - 4)1'2}]1/2} V a > 2, stable
if 2 < a < J5. Pitchfork bifurcation at (1,0), flip bifurcations at (-1,0),
(2, + 1) in (a, x)-plane.
 Answers and hints to selected problems 293

A3.7 X — 0 V a, stable if — 1 < a < 1; or X = + (1 — a)i/2 V a < 1, always unsta¬

ble. The three 2-cycles are {(- 1 - a)m, -(-1 - a)112} V a < -1, stable if
-2 < a < -1; {+ + (a2 - 4)1/2}]1'2, ± [*{-a - (a2 - 4)1'2}]1'2}
V a < — 2, stable if —y/5 < a < —2.
A3.8 (a) Assume 3 p-cycle with F(X}) = XJ+U F(Xp) = Xx & use reductio ad
absurdum. Take Xt < X2 without loss of generality. Then monotonic prop¬
erty implies that X2 < X3 < • • • < Xp < Xu a contradiction, (b) F(x) - x
decreases more rapidly than — x & so has unique zero. Note that if G(x) =
Fq(x) then G'(x) > 0 V x if q is even & G'(x) < 0 if q is odd, & use previous
parts of question.
A3.10 F"(a,x) = x/[a/{x2 + (a - x2)e_4*B‘,}]1/2. D(0) = (-oo, oo) if a < 0;
= (0, oo), D( —^/a) = (-oo,0) if a > 0.
A3.11 F8(r) = 28r = (5 x 51 + l)r = rmodulo 51.
A3.12 Use rotation map.
A3.13 r -*• r, 0 -* 6 + i//(r).
A3.14 Xj is unstable, X2 stable; D(-Y2) = (^,,00). 3 turning point at (e,e) in the
bifurcation diagram in the (a, x)-plane.
A3.15 N(x) = j(x + a/x), X = ±yja, N'(±yja) = 0 so (super-)stability,
D( — \Ja) — (—00,0), D^a) = (0, 00). F(x) = x + h(a - x2), X = ±yja,
F'(±yja) = 1 + 2y/ah, so Ja is stable (if 0 < h < 1 /^ja) & —yja is unstable,
D(Vfl) = (~Va>\/a + !/A), D(-co) =<-00, -Ja)u(Ja + l//i,00).
A3.17 Eliminate Y, Z from Y - aX( 1 — X),Z = aT(l — T), X = aZ( 1 — Z) to get
octic, and then discard factors X(aX — a + 1). 1112 = {a2 + a + 1 )/a6,
vt + v2 = (3 a + 1 )/a, ut + u2 + vt + v2 = (3 a + l)(a + l)/a2, t, + t2 +
utv2 + u2v 1 = (a3 + 5a2 + 3a -I- 1 )/a3, tj — Uj + v} = (a3 — l)/a3, v2 =
2 (a2 4- a + l)/a2.
A3.18 F(a,j) = 3a, Gx(a,x) = a2( 1 — 2x)(l — 2ax -I- 2ax2), so if a > 2 then G has
two equal maxima at x = j{1 ± (1 — 2/a)1/2} & a minimum at x = 3.
G(A,±A2(4 - A)) = G(A, 1/A) = {A - 1 )/A, G(A,(A - 1 )/A) = (A - 1 )/A.
G(A,{)=1/A. F(A,l/A) = (A-l)/A. F(A,iA)=l/A. F(A,(A - 1)/A) =
(A - 1)/A. F(A, x) ^ F(A, i) = \A. 1/A < $ < (A - 1 )/A.
A3.19 Let P(6) = 3 + cos2n0 = x. Therefore P(26) = F(x) by elementary trigo¬
nometry.
A3.20 x = Jox/(x)dx = Josin2(7t0)d0 = 3, x2 = |oX2/(x)dx = Josin4(7t0)d0 = §.
A3.22 X = 0 V a, stable V a < 1; X = a/(a + 1) V a > 1, unstable; or 0 < X < 3 for
a = 1, always stable. {§,f}.
A3.23 The 2-cycle, {a/(l + a2),a2/(l + a2)}, is unstable Va > 1.
A3.24 H~l(x) = sin2(^7tx), FH~l(x) = sin2(7ix).
A3.25 (a) xn+, = sn2(2/C(m)0B+1|m) = sn2(4/C(m)0„|w) = F(m,xn). (b) F(m,x) =
±sn2(4K(m)0\m), where x = sn2(2K(m)6„\m).
A3.26 X = 0 is unstable, X = ±1 are (super-)stable. D(1) = (0, 00), D( — 1) =
(-oo,0).
A3.27 X = 0 V a, stable for — 1 < a < 1; X = U/n V a > 1, marginally stable if
a cos U = — 1. Note analogy with logistic map, & incompleteness of anal-
294 Answers and hints to selected problems

ogy because, e.g., F is odd function of x. Pitchfork bifurcation at (1,0) in

(a, x)-plane, period-doubling etc. as a increases to n. If a = n then F maps
[0,1] onto itself, & unstable periodic orbits are dense in [0,1]. Also succes¬
sion of turning points as a increases.
A3.28 Graphs show uniqueness of A < 0. X = ln( — AT) — ln( — a) ~ — ln( — a) as
a -* —oo. By chain rule, Gx(a, x) = Fx(a, F(a, x))Fx(a, x) = F(a, F(a, x))F(a, x)
= G(a,x)F(a,x). Therefore Gx(a, X) = X2, etc. Therefore G = — 1, Gx = 1,
Gxx = 0, Gxxx = — 1 at a = — e, X = — 1, so line y = x has triple intersec¬
tion with curve y = G(a,x) at x = — 1 if a = — e, with onset of 3 intersec¬
tions for a < — e. Xlt X2 are roots of {x — G(a,x)}/(x — X) = 0, where
the Taylor series G(a,x) = A'-t-(x — Ar)A'2-l-2(x — Ar)2A'2(l-|-Ar) +
g(x — A')3A'2(1 + 3X + X2) + ..., so 3 subcritical flip bifurcation at
(— e, — 1) in (a, x)-plane.
A3.29 (a) See Example 3.9. (b) G(X) = X iff a + (In)-1 sin2jtX = 0, i.e. X =
1 — (27t)-1 sin(27ia/6) or \ + (2n)~1 sin(2na/b). Y0 = X0 + j as in part (a).
a2 = (87t)_1 sin(47rA0), so — 1/87T < a2 < 1/871.
A3.30 X = 0 V a, unstable for a > 0; X = 1 V a, stable for 0 < a ^ 2. Stable 2-cycle
for a = 2.3,4-cycle for a = 2.6, chaos for a — 2.9.
A3.31 Cf. Q3.6. Note analogy with logistic map, although F is odd function of x so
bifurcation diagram is symmetric in ±x, with pitchfork bifurcation at (1,0),
flip bifurcation at (— 1,0) in (a, x)-plane.
A3.34 K = J(Xj)J(X2). = detK = b2, q1 + q2 = trace K — 4(1 — b)2 —
4a + 2b. <7, = 1, q2 = b2 when a - a,; q1 = —\,q2= —b2 when
a = a2. Note also that q2 = q.i, |<7il = \b\ when (1 - b)2 + ^b( 1 - b) a ^
(1 - b)2 + 6(1 + b).
A3.36 a,(b) = 1 - b. X+ = (a + 1 - b)/{a2 + (1 - b)2},
X_ = — (a + b — 1 )/{a2 + (1 — b)2} V a > 1 — b.
A3.37 V = ((1 ~b + aq)/q( 1 + a - b), 6(1 - q)/( 1 + a - 6)).
A3.38 0 is always stable, because q = e±ia. Other point is always unstable, because
q has form 6 ± (62 — c)1/2 with 6 > 1, real c. Fixed points of T2 are these
two, & two complex points.
A3.41 If X, Y are rational, express X = a/r, Y = 6/r for integers a, 6, r, not neces¬
sarily coprime. Then F"(X) = (c/r,d/r) for some integers c, 4; but 3 only r2
points of the form (c/r,d/r)\ use the pigeon-hole principle.
A3.42 (0J(0)/(ab - 1)) V a # 1, (.X, Y) V Y, where A is a zero of /, for a = 1.
a 0
F(x, g(x)) = (ax, g(ax)). J = q 1 = a, qi = a", so toil.
-abf'(x) ab
|<j| > 1, & there is instability. F(x) = (axmod l,al’(y —f(x)) gives a contin¬
uous map of the cylinder because/(l) =/(0).
A3.44 Z = Z±, =^{1 ± (1 — 4a)1/2} = ^(1 + r1/2eifl/2), say, where a = ^ + rei9 for
r Si 0, -n < 0 < n. Stability if 1 > |F'(Z)|2, where F'(Z±) = 1 ± (1 - 4a)1/2.
Therefore Z+ is unstable V a, Z_ is stable if r < cos2 $9, i.e. a lies in a
cardioid in the complex plane. 2-cycle {± (-a -1)1/2} is stable if
\a + 11 < a circle. Cf. Mandelbrot set.
 Answers and hints to selected problems 295

A3.45 Use Cauchy-Riemann relations. qx,q2 = F'(z),F'(z).

A3.47 Cf. Q3.44 & Fig. 3.13(a). a e M iff Fn(a, 0) -f* oo iff 0 £ D(oo) for real map
F(a,x) = x2 + a (which is equivalent to logistic map). F has fixed points
x± = f{l ± (1 — 4a)1/2} V a < X+ is always unstable; A'_ is stable &
0 e D(X_)for — | < a < f. {Xx,X2} is a 2-cycle of F for a < — f, where AT,,
x2 = 2O ± (-3 - 4a)1/2}. {A',,X2} is stable & 0 e D({.Y,, A"2}) for -f
a < — |. As a decreases there is period doubling.

Chapter 4

A4.2 (i) ± = 0.020202.... (ii) x = 0.200020....

A4.3 If x e S n [0, f) then 4x e S because the point of the quaternary expression of
x is moved one place to the right. Similarly, if y e S then f y e S n [0, J], etc.
A4.4 X = 0, a/(a + 1) both unstable. F maps (if) to (1,§], & [0,f], [f, 1] each
onto [0,1] repeatedly.
A4.6 (a) Piecewise linear function Fn increases by 2~n over each of 2" subintervals
of length 3-", /„ = (§)". (b) Length of C„ = total horizontal + total sloping
length = 1 - /„ + (1 + Z2)1/2. (c) |F„(y) - F„(x)| < 2|y - x|(ln3)/(,n2» V x,
y e [0,1], V n.
A4.7 ZV(e) = 2" when e = 4~".
A4.9 N(e)ccRn, eocr", D = lim^^ {In jV/ln(e-1)}. For K, R = 2, r = f; for
von Koch curve, R = 4, r = f. For decimal set, R = 5, r = 10-1, D =
In 5/ln 10 = logi0 5. Same method works in IRm if e = length of sides of
hypercubes.
A4.ll S„ has N = 4" subsquares with sides of length e = 4~". Subsquares of S,
project onto 4 equal subintervals AC^ CjC2, C2C3, C3B, where
Ci = ( —1\)> 2o)> c2 — (5, —1\)), c3 — (5, —4), & so forth.
A4.12 D = m — 1 + In2/ln3.
A4.13 F"(S) has 2" strips of unit length (in x-direction) & height (fa)" in y-direction.
Therefore N ~ {l/(2a)"} x 2" as n -> 00 if e = (2a)n.
A4.15 F(X2) — X2 = — {FfA1!) — Xx}, so continuous function F(x) — x has at
least one zero between Xt, X2, by intermediate-value theorem.
A4.16 Only fixed point of F is X = F3 has no fixed point in [1,3] u [4,5].
F3[3,4] = F2[2,4] = F[2,5] = [1,5], being monotonically decreasing map
of interval in each case. Therefore F3 has a unique fixed point in [3,4],
which can only be X.
A4.19 D = 1 + 2i/|22| in this case with m = 2, k = 1, > 0 > A2.
A4.20 At,A2 = In{f(3 ± 75)}.

Chapter 5

A5.1 z(r) = {z0 - fe(cos 90 + sin 0O)} exp(0o - t) -I- fe(cos t + sin t) if z = z0 at
t = 90. Z — fe(cos 90 + sin 90) is always stable.
296 Answers and hints to selected problems

A5.2 V e > 0 3 {f„} such that |x(rj - xj < € & r„ f oo as n -► oo. Therefore, V
e' > 0, | K(x(r„)) - K(x,J| < e' as n -* oo. Therefore grad F(x(t„))->0 as
n-* oo.
A5.4 Linearized system is da/j/dr = 0, dco'2/dr = (C — A)na>'3/B, d(o'3/dt =
(A — B)a>'2/C. A rigid body is stable if it spins steadily about its greatest or
least principal axis of inertia, but unstable if it spins about its intermediate
principal axis of inertia.
A5.5 A" = j{c ± (c2 — 4ab)l/2}, Y = —Z = —j{c ± (c2 — 4ab)1,2}/a if c2 > 4ab.
Turning points at (+ 2(ab)l/2, ±(ab)112) in (c, x)-plane.
A5.6 See §8.1. (a) r2 = l/b, x2 = [0,0, l/h]T. N.B. Exact solution is x =
[(5,<5,<52/h]T, r = 1 + S2/b.(b) x2 = [0,0,A2/bY.
A5.12 Use Euler-Lagrange equation. dL/dq = (d/dr) (dL/dq) = dp/dt.
A5.13 For linearized problem of instability, s = — A/y/2 or J2A twice. Poincare
map is identity map for x & rotates y by 2n tan x (see Example 3.9).
A5.15 (a) General solution x(r) = (A + a)cos(t + t) is close to X at t = 0 for small
constants a, t. |x(t) — A'(r)! = 0(a, t) V r as a, t -> 0. (b) Use Example 1.2.
A5.16 dr'/dt = — (2ab cos2 t)r', d9'/dt = (yjab sin 2t)r'. Pa(-Ja + r') —
yja + r'exp( — 2nab) + 0(r'2) as r' -+0.
A5.17 P(r + T) = <F(r + T’)e~(t+r)R = 0(t)erRe-(,+r)R = P(r). If Ru = su & / is a
polynomial then /(R)u = f(s)u. d(det 0)/dr = (trace A) det O.
A5.18 Let xt, x2 be solutions of Meissner’s equation such that X! = 1, xt = 0,

x2 = 0, x2 = 1 at t = 0. Then O = Xl Xl . Solving piecewise for 0 < t <

LA1
i i <
2> 2 ^r ^ 1, find Xj(l) = cosjucosh^fl — sin ^asinh^u and also
x,(l) = a(cosjasinh^a — sin^acosh^a). Similarly find x2(l) =
(sin^acoshja + cos^asinh^a)/a, x2(l) = sin|asinhja + cos^acosh^a.
Therefore 0 = det(0(l) — ql) = q2 — 2bq + 1, where b = 2{xi(l) + x2(l)} =
cos2 a cosh2a. Therefore qlt q2 = b ±{b2 - 1)1/2. N.B. q{ = q2= ± 1 for
values of a such that cos cosh \a = ±1.

A5.19 (a) If x =
x
dx/dr ] then A -
~P(t) 0
0

a
0(1) is a real 2 x 2 matrix so its

eigenvalues qt, q2 are real or a complex conjugate pair, traceA(r) = 0 V t.

Therefore qxq2 = detOfa) = constant = det0(0) = 1. Therefore 0 =
det(<D(7t) - ql) = q2 - {x^rc) + x2(n)}q + 1, & (x^) + x2(7t)}2 < 4 im¬
plies stability. At margin of stability qt = q2 = +1 (for period n, 2n
respectively).(b) Ifx0(r) = sinrthenu, = -l,x1(t)= -|sin3r.

Chapter 6

A6.1 If P = (x, y) then Q = (x, F(x)), R = (0, F(x)), RP = (x,y - F(x)) dx/dr =
(y — F(x), — x), RP-dx/dr = 0.
A6.2 Orbits are alternately semiellipses in upper & . lower halves of (x,dx/dr)-
plane, with centres at (- F/A, 0), (F/A, 0) respectively, but end where dx/dr =
0, — F/A ^ x ^ F/A.
 Answers and hints to selected problems 297

A6.3 (— -y/a, 0) is a saddle point, (Ja, 0) is a stable node.

A6.4 y = — x 4- bc~l lnx 4- constant.
A6.5 Linearized equations at (0,0) are dx'/dt = ax', dy'/dt = - by'; & at (b/c, a/c)
are dx /dr = —by', dy'/dr = ax', so d2x'/dt2 = —abx'. Separate variables,
integrate, & take exponents to find xby“ = kec^x+y) for constant k of integra¬
tion on each orbit.
A6.7 Equilibrium point (nn, 0) is a saddle point if n is even; it is a stable focus for
gl > k2, a stable node for gl < k2 if n is odd.
A6.8 Equilibrium points are at (X, 0) V X, such that each orbit ends where it meets
x-axis. Equation is piecewise linear in upper & lower half planes.
A6.9 (a,b/a) is stable if b < a2 + 1 or unstable if b > a2 + 1, & node if
(a2 + 1 — b)2 > 4a2 or focus if (a2 + 1 — b)2 < 4a2.
A6.10 (i) Equilibrium points are (0,0) V a, saddle point V a > 0, stable node V
a < 0; and (± -Ja, 0) V a > 0, stable nodes. Supercritical pitchfork bifurca¬
tion at x = y = a = 0.
(ii) (0,0) is a centre, (a, 0) a saddle point V a ^ 0. Transcritical bifurcation
at (0,0) in (a,x)-plane. System is invariant if (x,a)-»( —x, — a) or (y,f)->
( — y, — r). x = —a is an invariant line. If a = 0, exact integral is y2 =
x2(lnx + C).
(iv) Equilibrium points are (0,0) V a, (a,0), (-a, - a), (-a,2a) V a # 0.
y = 0 is invariant line. System is independent of a if (x, y, t) -*• (x/a, y/a, at),
so take a = 0, 1 without loss of generality. If a = 0, exact integral is
y2(y2 — 2x2) = C. If a = 1, (0,0) is unstable node & (1,0), (— 1, — 1), ( — 1,2)
are saddle points.
(vii) Equilibrium points are (0,0) V a & (0,5a) V a# 0. System is
invariant if (x, y, t, a) -»(— x, — y, — t, — a), so take a ^ 0 without loss of gen¬
erality. If a = 0, exact integral is x2 = 2y2(C - In y). (0,0), (0,5a) are unsta¬
ble foci if a > 0.
A6.11 (a) 0 is stable focus if a < — 1, stable node if — 1 < a < 0, saddle point if
a > 0. (b) ( — 2,0) is unstable if a < 0. (— 1 ± (1 + 4a)1/2,a) exist for a >
— i, former is unstable for — ^ < a < 0, latter for a > 0. (c) Transcritical
bifurcations at (0,0), (0,-2), turning point at (-1,-1) in (a,x)-plane.
(d) Equate coefficients of a in the three equations, solve for x,. Equate
coefficients of a2 to find F0, x2 = A(l — A)[\, — 2]T/50.
A6.12- <(> = ln(2/<5r2). Saddle point at (0,2).
A6.13 0 3s xTF(x) = jd(xTx)/df. Therefore |x(t)| < |x(0)|. Taking F(x) = Ax,
deduce xTF(x) = |xT(A + AT)x ^ 0 V x.
A6.14 Try reductio ad absurdum. It may help to find elsewhere a proof of
Liapounov’s theorem (a) on p. 179 & adapt it to a difference equation.
A6.16 (a) x(r) = x0e"', y(t) = roe"'cos(0o — 2t), z(t) = — roe~'sin(0o — 21).
(b) H(x) — x2 + y2 + z2 will do.
A6.17 Verify results & use uniqueness of solutions of differential equations.
A6.20 K(x,y) = jr2{a2 + b2 + c2 + d2 + (a2 — b2 + c2 — d2) cos 20 +
 298 Answers and hints to selected problems

2(ab + cd) sin 20}. dK/dO = 0 if tan 20 = 2(ab + cd)/(a2 — b2 + c2 — d2).

dH/dt = 2pK + 0(|x| • |£|) > pK as x - 0.
A6.21 dH/dt = -2k(dx/dt)2.
A6.22 Take a — 2, say, so that dH/dt = — 2(x2 — ^)2 — 2y2/25.
A6.23 Try H(x) = x2 + axy + by2 & find suitable constants a, b.
A6.24 m = 4, n = 2, b — 2 will do.
A6.25 Try H(x) = x2 + y2.
A6.26 If (5|L(w, wjdx = 0 & L(w, wj = j(w2 — w2), then Euler-Lagrange equa¬
tion is wxx + w = 0. Minimizing function is w(x) = sinx. Deduce inequality
& use it to show dH(u)/dt < —kH(u) unless u(x, t) = 0 almost everywhere.
A6.27 Centre at (0,0) V e is stable, saddle points at (± 1/e, 0) V e ^ 0 are unstable.
A6.28 Centre at (0,0) V e is stable, saddle point at (l/e,0) V e ^ 0 is un¬
stable. x(e,t) = acoscot + ea2(£cos2cot + jcoscot — 2) + 0(e2a3), a> =
1 — -h^a1 -l- 0(e3a3) as ea -* 0. 2n/a> > 2n, (co/2n) jo*/® x(e, t) dt < 0 be¬
cause periodic orbit for e ^ 0 goes nearer saddle point, where motion is
slow, than it would for e = 0.
A6.29 uN, ur = {1 +(1 — 4e)1/2}/2e/. Linearization gives a centre, saddle point
respectively if e < j. Problem is essentially same as Q6.28 with x =
u — uN: C32d2x/d<l>2 = elx2 — (1 — 4e)1/2x.
A6.30 Cf. Kevorkian & Cole (1981, §3.1.2).
A6.31 Suppose a closed orbit C encloses a domain S c D & use reductio ad
absurdum. A = dF/dx + dG/dy is of the same sign in D because it is non-zero
& continuous. Therefore 0 # js Adxdy = Jc(Edy — Gdx), by divergence
theorem, =0 because C is an orbit.
A6.32 (a) Origin is centre, so stable. Take H = r2. Therefore dH/dt =
2r4(l - 2sin2 20 - r2) ^ 0 if r2 < (b) Take D = {x: \ < r < 1}. There¬
fore dr/dt > 0 on outside circle of annulus D & dr/dt < 0 on inside circle
ofD. d0/df = -(1 + \r2 sin2 40) ^ OinD.
A6.33 (a) 0 is only equilibrium point because d0/df ^ 0. dH/dt = 2H{a — /(r2)}.
(b) Define p as least zero of /(r) = a > 0, & take D = {x: e < r < p - e} for
e so small that dr/dt < 0 on r = e & dr/dt > 0 on r = p - e. (c) dr/dt =
r(a - r), dd/dt = -1 implies that x(t) = a cos(0O - t), y(t) = a sin(0o - t) is
stable limit cycle.
A6.34 dr/dt = -r{l - r2(l + sin20)}, d0/dr = 1. Take D = {x: 1/^/2 < r < 1}.
In fact, explicit solution is 0 = t - t0, r(f) = r0/[e28 + r^{^(e2fl + sin 20 -
cos0) — §(e28 — 1)}]1/2 ~ l/{j(sin20 — cos0) + f}1/2 as 0-»-oo, giving
the unstable limit cycle.
A6.35 (a) Cf. Q1.16. x(f) = $Tt + ±t2 for —jT ^ t < 0, x(r) = \Tt - \t2 for 0 ^
1 < 2T, where period T = 4(2u)l/2. (b) Unstable focus at (0,0). Orbits spiral
anticlockwise out from (0,0) & in from infinity to stable limit cycle, (c) Use
theorem of §6.6.
A6.36 Define H(x) = }x4 + \y2, where y = dx/dt, & show that dH/dt > 0 for small
|x| & dH/dt > 0 for large |x|. Bound 0 by small and large level curves of H.

0
 Answers and hints to selected problems 299

A6.37 0 = ln|lnr| gives focus at origin with tight spiral, whereas linearized system
gives sink at origin with radial straight orbits.
A6.38 xl/2 — a cos cot,yl/2 — — aco0 cos cot, where a3 = 4a>o.
A6.40 da/dt ~ ^ca(l — 4a/3n), co = — 1 + 0(e2).
A6.41 da/dt ~ |ea(l — £a4),co = — 1 + 0(e2).
A6.42 da/dt = 0(e2) if a < 1, da/dt = ea(j + n~x sin 2a — 2oL/n) if a > 1, where
a - arccos(l/a), co = — 1 + 0(e2).
A6.44 da/dt = 0(a2), to = — 1 + yga2 + 0(a3) as a -*0. In agreement with equa¬
tion (1.7.13), this gives T = 2n(l/g)ll2/\co\ = 2tt(//^)1/2{ 1 + ±a2 + 0(a3)}.
A6.45 da/dt ~ —2ea( 1 — 8a/37t), co = — 1 + 0(e2). Limit cycle is x(e, t) -*
|7rcos(r - f0).
A6.46 a -»^6, |<w| -♦ 1 as e -► 0.

Chapter 7

A7.1 d2x/dr2 + x3 = ( — a + fa3) cost + h.h. ifx(t) = a cost -I- h.h.

A7.2 Express x(t) % a cos cat + b sin cot =ur cos(cot — t0). Therefore sgnx(t) =
Iao + E(ancosncot + b„sinncot), where a„ as 4(wr)_1 cos(n-r0), bn as
4(«7t)_1 sin(nt0). Therefore sgn x(f) as 4(a cos cot + b sin cot)/nr.
A7.4 Expand x = x0 + Tx1-(-..., assume x„ has period In/co. Therefore x0 =
0, ±(f}/a)112.
A7.5 Expand x = x0 4- ext -I-..., assume x„ has period 27t. (a) x0 =
Tcosr/(fi2 - 1),x, = ±T2(Q2 - 1)-2{Q-2 + (Q2 - 4)-1cos2t}.(b) x0 =
a0cost + b0sint, where annihilation of secular terms in x, gives b0 = 0,
a0 = y/P-
A7.6 a(iQ, - IT) = 0, &(*«, + f T) = 0.
A7.7 a = 2 - \/oj2, q = - 1 + 1/co2.
A7.8 a = 4,q = + (2T/<5)1/2.
A7.9 q = f ea2, 0 ^ Q2 = 1 — ea2. Therefore 1 + ea2 < 2, q < f, & Abramowitz
& Stegun (1964, Fig. 20.1) give stability for c > 0.

Chapter 8

A8.3 See §1.7.

A8.5 Saddle point at (0,0) & centre at (1,0) for k = 0; saddle point at (0,0) & focus
at (1,0) for 0 < k2 < 1, the focus being stable for 0 < k < 1 & unstable for
— 1 < k < 0.
A8.8 If M(t0) is identically zero then we can deduce only that D(f0) = 0(e2) as
€ -»■ 0, & cannot ascertain whether D has a simple zero without proceeding
to a higher approximation.
A8.9 Note that ln{tan(2Xm)} -» +oo as xm -»0, n respectively, & so the separa¬
tion of the zeros of M tends to zero near the saddle points.
300 Answers and hints to selected problems

Appendix

AA.3 Rc = min(l - *2)2 = 0. _ dujdt = {R - (n2 - 1 )>. - Xp.,= -oo upuqun_p_q.

AA.5 d//df = —(1 4- /?)(/ +/) — (1 + ib)k2f, where k = tin/l. Let / = g + ih, &
get real pair of equations for g, h. Take g, h oc e5'; thence s = — 1 — k2 ±
(1 - 2Rbk2 - h2fc4)1/2.
AA.6 u', = fu(U, K)u' + /„(£/, F)d + W, v[ = gu(U, V)u' + gv(U, V)v! +
dW2v', du'/dn = dv'/dn = 0. U = a + b, V = b/(a + b)2. p(x) = coskx
for/c = 0, l/l,2/l,— s2 + {(1 + d)k2 + y(a + b)2 — y(b — a)/(a + b)}s +
d/c4 + yk2{(a + b)2 - (b - a)d/(a + b)2} + y2(a + b)2 = 0.
AA.7 F"(x) = — 2c2sech2cx = — SeF. Lagrange’s identity gives (s — s)ji, <j><j>dx =
[<£d^/dx — ^d^/dx]L, = 0 & thence real s. (1 — y2)d2^/dy2 — 2yd^/dy +
{2 — sc 2/U — y2)}4> = 0 is the associated Legendre equation of degree 1 &
order
 Bibliography and author index

The numbers in square brackets following each entry give the pages of this book on
which the entry is cited. I have been unable to inspect copies of a few of the works
cited, so their citations are second- (or even third-) hand.

Abramowitz, M. & Stegun, I. A. 1964 Handbook of Mathematical Functions. Wash¬

ington, DC: Natl Bureau Standards. [231,299]
Allan, D. W. 1962 On the behaviour of systems of coupled dynamos. Proc. Camb.
Phil. Soc. 58,671-93. [280]
Andronow, A. A. & Chaikin, C. E. 1949 Theory of Oscillations. Princeton Univer¬
sity Press. English version of Russian book of 1937. [4,40, 55]
Andronow, A. A. & Pontryagin, L. 1937 Syst&mes grossiers. Dokl. Akad. Nauk
SSSR 14, 247-50. [55]
Arnol’d, V. I. 1961 Small denominators. I. Mappings of the circumference onto
itself. Izvest. Akad. Nauk Ser. Mat. 25,21-86. English transl. in Amer. Math. Soc.
Transl. (2) 46 (1965), 213-84. [ 118]
Arnol’d, V. I. & Avez, A. 1968 Ergodic Problems of Classical Mechanics. New York:
Benjamin. [121]
Bender, C. M. & Orszag, S. A. 1978 Advanced Mathematical Methods for Scientists
and Engineers. New York: McGraw-Hill. [38,46]
Bendixson, I. 1901 Sur les courbes definies par des equations differentielles. Acta
Math. 24, 1-88. [199, 208]
Benjamin, T. B. 1978 Bifurcation phenomena in steady flows of a viscous fluid I.
Theory. Proc. Roy. Soc. Lond. A359,1 -26. [64]
Bernoulli, J. 1713 Ars Conjectandi. Basel: Impensis Thurnisorum, fratrum. [78]
Bountis, T. C. 1981 Period doubling bifurcations and universality in conservative
systems. PhysicaDi, 577-89. [121]
Bradbury, R. 1953 The Sound of Thunder. On pp. 100-13 of The Golden Apples of
the Sun. Garden City, NY: Doubleday. [38]
Bristol 1983, 1987-90 Examination questions on final-year mathematics course
Nonlinear Systems at University of Bristol. [116,201,208,209]
Buckmaster, J. D. & Ludford, G. S. S. 1982 Theory of Laminar Flames. Cambridge
University Press. [288]
Budd, C. J. 1989 Applications of Shilnikov’s theory to semilinear elliptic equations.
SIAM J. Math. Anal. 20, 1069-80. [203]

301
 302 Bibliography and author index

Campbell, L. & Garnett, W. 1882 Life of James Clerk Maxwell. London: Macmillan.
[6]
Cantor, G. 1883 Ueber unendliche, lineare Punktmannichfaltigkeiten. Math. Ann.
21, 545-91. Also Gesammelte Abhandlungen (1932), ed. E. Zermelo, pp. 165-209,
Berlin: Teubner. [87,125]
Cayley, A. 1879 The Newton-Fourier imaginary problem. Amer. J. Math. 2, 97.
Also Collected Math. Papers 10 (1896), pp. 405-06, Cambridge University
Press. [109]
Chandrasekhar, S. 1969 Ellipsoidal Figures of Equilibrium. New Haven, CT: Yale
University Press. [38]
Coddington, E. A. & Levinson, N. 1955 Theory of Ordinary Differential Equations.
New York: McGraw-Hill. [38, 161, 168, 177, 199]
Collet, P. & Eckmann, J.-P. 1980 Iterated Maps on the Interval as Dynamical
Systems. Boston, MA: Birkhauser. [109, 144]
Cook, A. E. & Roberts, P. H. 1970 The Rikitake two-disc dynamo system. Proc.
Camb. Phil. Soc. 68, 547-69. [280]
Courant, R. 1936 Differential and Integral Calculus, vol. 2. London: Blackie. [63]
Cox, S. M., Drazin, P. G., Ryrie, S. C. & Slater, K. 1990 Chaotic advection of
irrotational flows and of waves in fluids. J. Fluid Mech. 214, 517-34. [282]
Curry, J. H., Herring, J. R., Loncaric, J. & Orszag, S. A. 1984 Order and disorder in
two- and three-dimensional Benard convection. J. Fluid Mech. 147, 1-38. [268]
Cvitanovic, P. (ed.) 1984 Universality in Chaos. Bristol, England: Adam Hilger.
[135]
Devaney, R. L. 1989 An Introduction to Chaotic Dynamical Systems. 2nd edn. Red¬
wood City, CA: Addison-Wesley. [38,109,111,132]
Dombre, T., Frisch, U., Greene, J. M., Henon, M., Mehr, A. & Soward, A. M. 1986
Chaotic streamlines in the ABC flows. J. Fluid Mech. 167,353-91. [166]
Drazin, P. G. & Johnson, R. S. 1989 Solitons: an Introduction. Cambridge Univer¬
sity Press. [43,286]
Drazin, P. G. & Reid, W. H. 1981 Hydrodynamic Stability. Cambridge University
Press. [38,40,65]
Duffing, G. 1918 Erzwungene Schwingungen bei Veranderlicher Eigenfrequenz.
Braunschweig: Vieweg. [214]
Eggleston, H. G. 1953 On closest packing by equilateral triangles. Proc. Camb.
Phil. Soc. 49,26-30. [145]
Euler, L. 1744 De Curvis Elasticis. Appendix on pp. 245-310 of Methodus Inveni-
endi Lineas Curvas, Lausanne: Bousquet. [4]
Falconer, K. J. 1985 Geometry of Fractal Sets. Cambridge University Press [145
147]
Falconer, K. J. 1990 Fractal Geometry. Chichester, England: Wiley. [109 122 143
144, 148] ’ ’ ’
Fatou, P. 1906 Sur les solutions uniformes de certaines equations fonctionnelles.
Comptes Rendus(Paris) 143, 546-8. [109, 144]
Fatou, P. 1919 Sur les equations fonctionnelles. Bull. Soc. Math. (France) 47 161-
271. [105, 109]
 Bibliography and author index 303

Feigenbaum, M. J. 1978 Quantitative universality for a class of nonlinear transfor¬

mations. J. Statist. Phys. 19,25-52. [87,135,137]
Feigenbaum, M. J. 1979 The universal metric properties of nonlinear transforma¬
tions. J. Statist. Phys. 21,669-706. [137]
Feigenbaum, M. J. 1980 The transition to aperiodic behavior in turbulent systems.
Commun. Math. Phys. 77,65-86. [140]
Floquet, G. 1883 Sur les Equations diff6rentielles lin6aires il coefficients p£ri-
odiques. Ann. Sci. Ecole Norm. Sup. 12,47-89. [160]
Ginzburg, V. L. & Landau, L. D. 1950 On the theory of superconductivity.
Zhur. Eksper. Teor. Fiz. 20, 1064-82 (in Russian). English transl. in Collected
Papers of L. D. Landau (1965), ed. D. ter Haar, pp. 546-68, Oxford: Pergamon.
[286]
Glendinning, P. & Sparrow, C. 1984 Local and global behavior near homoclinic
orbits. J. Statist. Phys. 35,645-96. [246]
Golubitsky, M. & Schaeffer, D. G. 1985 Singularities and Groups in Bifurcation
Theory I. New York: Springer-Verlag. Appl. Math. Sci. 51. [63]
Griffel, D. H. 1981 Applied Functional Analysis. Chichester, England: Ellis Hor-
wood. [64]
Guckenheimer, J. & Holmes, P. J. 1986 Nonlinear Oscillations, Dynamical Systems,
and Bifurcations of Vector Fields. 2nd edn. New York: Springer-Verlag. Appl.
Math. Sci. 42. [211,251,252,277]
Haberman, R. 1979 Slowly varying jump and transition phenomena associated
with algebraic bifurcation problems. SIAM J. Appl. Math. 37,69-105. [39]
Hardy, G. H. & Wright, E. M. 1979 An Introduction to the Theory of Numbers. 5th
edn. Cambridge University Press. [109]
Hausdorff, F. 1919 Dimension und ausseres Mass. Math. Ann. 79, 157-79. [128,
129]
Henon, M. 1966 Sur la topologie des lignes de courant dans un cas particulier.
Comptes Rendus(Paris) 262, 312-14. [166]
Henon, M. 1969 Numerical study of quadratic area-preserving mappings. Quart.
Appl. Math. 27, 291-312. [120]
Henon, M. 1976 A two-dimensional mapping with a strange attractor. Commun.
Math. Phys. 50,69-77. [98]
Henon, M. & Heiles, C. 1964 The applicability of the third integral of motion: some
numerical experiments. Astron. J. 69,73-79. [164]
Henon, M. & Pomeau, Y. 1976 Two strange attractors with a simple structure. On
pp. 29-68 of Turbulence and Navier-Stokes Equations, ed. R. Temam. Berlin:
Springer-Verlag. Lecture Notes in Math. 505. [98,99,100,119]
Holmes, P. J. 1979 A nonlinear oscillator with a strange attractor. Phil. Trans. Roy.
Soc. Lond. A292,419-48. [246,251,260]
Hopf, E. 1942 Abzweigung einer periodischen Losung von einer stationaren Lo-
sung eines Differentialsystems. Ber. Math.-Phys. Klasse Sachs. Akad. WTss.
Leipzig 94, 1-22. English transl. on pp. 163-93 of The Hopf Bifurcation and its
Applications, ed. Marsden, J. E. & McCracken, M. (1976), New York: Springer-
Verlag. [21]
 304 Bibliography and author index

Iooss, G. & Joseph, D. D. 1990 Elementary Stability and Bifurcation Theory. 2nd
edn. New York: Springer-Verlag. [35,63,201]
Jordan, D. W. & Smith, P. 1987 Nonlinear Ordinary Differential Equations. 2nd edn.
Oxford: Clarendon Press. [161,197,199,229,231]
Julia, G. 1918 Memoire sur l’iteration des fonctions rationelles. J. Math. Pures
Appl. (7) 4,47-245. Also Oeuvres (1968), 121-319, Paris: Gauthier-Villars. [105,
109]
Kaplan, J. L. & Yorke, J. A. 1979 Chaotic behavior of multidimensional difference
equations. On pp. 204-27 of Functional Differential Equations and Approxima¬
tion of Fixed Points, ed. Peitgen, H.-O. & Walther, H.-O., Berlin: Springer-
Verlag. Lecture Notes in Math. 730. [143]
Kermack, W. O. & McKendrick, A. G. 1927 A contribution to the mathematical
theory of epidemics. Proc. Roy. Soc. Lond. A1 IS, 700-21. [200]
Kevorkian, J. & Cole, J. D. 1981 Perturbation Methods in Applied Mathematics.
New York: Springer-Verlag. Appl. Math. Sci. 34. [197,217,298]
King, G. P. & Drazin, P. G. (eds.) 1992 Interpretation of time series from nonlinear
systems. Physica D 58. [277]
Kittel, C. 1986 Introduction to Solid State Physics. 6th edn. New York: Wiley. [66]
Koch, H. von See von Koch, H.
Kolmogorov, A. N. 1959 Entropy per unit time as a metric invariant of automor¬
phisms. Dokl. Akad. Nauk SSSR 124, 754-55. [141]
Korteweg, D. J. & De Vries, G. 1895 On the change of form of long waves advancing
in a rectangular canal, and on a new type of long stationary waves. Phil. Mag.
(5) 39,422-43. [43]
Kryloff, N. M. & Bogoliuboff, N. N. 1943 Introduction to Non-linear Mechanics.
Princeton University Press. Annals Math. Studies 11. English version of papers,
especially Introduction a la mecanique non-lineaire: les methodes approchees
et asymptotiques’. Ukrainska Akad. Nauk Inst. Mec., Chaire de Phys. Math.
Annales 1-2(1937). [192]
Kuramoto, Y. & Koga, S. 1982 Anomalous period-doubling bifurcations leading to
chemical turbulence. Phys. Lett. A92,1-4. [286]
Lagrange, J. L. 1788 Mecanique Analytique. Paris: Courcier. [180]
Landau, L. D. 1944 On the problem of turbulence. Comptes Rendus Acad. Sci.
U.S.S.R. (Doklady) 44, 311-14. Also Collected Papers (1965), ed. D. ter Haar,
pp. 445-60, Oxford: Pergamon. [13]
Landau, L. D. & Lifshitz, E. M. 1980 Statistical Physics, Part I. 3rd edn. London:
Pergamon. [66]
Landau, L. D. & Lifshitz, E. M. 1987 Fluid Dynamics. 2nd edn. London: Pergamon
[148]
Lebovitz, N. R. & Schaar, R. J. 1975 Exchange of stabilities in autonomous sys¬
tems. Stud. Appl. Math. 54,229-60. [39]
Li, T.-Y. & Yorke, J. A. 1975 Period three implies chaos. Amer. Math. Monthly 82
985-92. [87,133] ’
Liapounov, A. M. 1892 The general problem of the stability of motion. Commun.
Soc. Math. Kharkov. Transl. from Russian to French in ‘Probleme general de
 Bibliography and author index 305

la stabilite du mouvement’, Ann. Fac. Sci. Toulouse (2) 9, (1907) 203-475; re¬
printed in Annals Math. Studies 17 (1947), Princeton University Press. [70, 178,
180]
Liapounov, A. M. 1906 Sur les figures d’equilibre peu differentes des ellipsoides
d’une masse liquide homogene donnee d’un mouvement de rotation. 1: Etude
generate du probleme. Zap. Akad. Nauk (St Petersburg) 1,1-225. [6]
Lichtenberg, A. J. & Lieberman, M. A. 1983 Regular and Stochastic Motion. New
York: Springer-Verlag. Appl. Math. Sci. 38. [277]
Lienard, A. 1928 Etude des oscillations entretenues. Rev. Gen d'Elect. 23, 901-12,
946-54. [199]
Lindstedt, A. 1883 Beitrag zur Integration der Differentialgleichungen der Sto-
rungstheorie. Mem. Acad. Imp. Sci. St Petersbourg, (7), 33, no. 4,1-20. [183]
Lorenz, E. N. 1963 Deterministic nonperiodic flow. J. Atmos. Sci. 20, 130-41. [233,
285]
Lotka, A. J. 1920 Undamped oscillations derived from the law of mass action. J.
Amer. Chem. Soc. 42,1595-9. [200]
Lozi, R. 1978 Un attracteur etrange du type attracteur de Henon. J. Phys. (Paris) 39
(C5), 9-10. [119,120]
Mandelbrot, B. B. 1975 Les Objets Fractals: Forme. Hasard et Dimension. Paris:
Flammarion. [131]
Mandelbrot, B. B. 1982 The Fractal Geometry of Nature. San Francisco, CA: Free¬
man. [105, 109, 144, 146]
Matkowsky, B. 1970 A simple nonlinear dynamical stability problem. Bull. Amer.
Math. Soc. 76,620-5. [285]
McLachlan, N. W. 1956 Ordinary Non-linear Differential Equations in Engineering
and Physical Sciences. 2nd edn. Oxford: Clarendon Press. [231,232]
Mel’nikov, V. K. 1963 On the stability of the centre for time-periodic perturbations.
Trans. Moscow Math. Soc. 12, 1-57. [251, 257]
Misiurewicz, M. 1980 Strange attractors for the Lozi mappings. Ann. N.Y. Acad.
Sci. 357, 348-58. [120]
Moser, J. 1973 Stable and Random Motions in Dynamical Systems with Special
Emphasis on Celestial Mechanics. Princeton University Press. Annals Math.
Studies 77. [112]
Murray, J. D. 1989 Mathematical Biology. Berlin: Springer-Verlag. [287]
Myrberg, P. J. 1958 Iteration von Quadratwurzeloperationen. Annales Acad. Sci.
Fennicae A I Math. 259, 1-16. [114]
Nemytskii, V. V. & Stepanov, V. V. 1960 Qualitative Theory of Differential Equa¬
tions. Princeton University Press. English version of Russian book of 1949. [41,
210]
Neumann, J. von See von Neumann, J.
Newhouse, S., Ruelle, D. & Takens, F. 1978 Occurrence of strange axiom A at¬
tractors near quasi-periodic flows on Tm, m ^ 3. Commun. Math. Phys. 64, 35-
40. [261]
Packard, N. H., Crutchfield, J. D., Farmer, J. D. & Shaw, R. S. 1980 Geometry from
a time series. Phys. Rev. Lett. 45,712-16. [266]
306 Bibliography and author index

Palais, B. 1988 Blowup for nonlinear equations using a comparison principle in

Fourier space. Comm. Pure Appl. Math. 41,165-96. [289]
Pearl, R. & Reed, L. J. 1920 On the rate of growth of the population of the United
States since 1790 and its mathematical representation. Proc. Nat. Acad. Sci. 6,
275-88. [10]
Peitgen, H.-O. & Richter, P. H. 1986 The Beauty of Fractals. New York: Springer-
Verlag. [109]
Peitgen, H.-O. & Saupe, D. (ed.) 1988 The Science of Fractal Images. New York:
Springer-Verlag. [109]
Pippard, A. B. 1985 Response and Stability. Cambridge University Press. [37]
Poincare, J. H. 1881 Memoire sur les courbes definies par une equation differen-
tielle. J. Math. Pures Appl. (3) 7, 375-422. Also Oeuvres 1 (1928), ed. P. Appell,
pp. 3-53, Paris: Gauthier-Villars. [170]
Poincare, J. H. 1882 Memoire sur les courbes definies par'une equation differen-
tielle. J. Math. Pures Appl. (3) 8, 251-96. Also Oeuvres 1 (1928), ed. P. Appell,
pp. 53-84, Paris: Gauthier-Villars. [31,170, 199]
Poincar6, J. H. 1885a Sur l’6quilibre d’une masse fluide animee d’un mouvement de
rotation. Acta Math. 7, 259-380. Also Oeuvres 7 (1952), ed. J. Levy, pp. 40-140,
Paris: Gauthier-Villars. [5]
Poincare, J. H. 1892 Les Methodes Nouvelles de la Mecanique Celeste, vol. I. Paris:
Gauthier-Villars. English transl. of all 3 vols. (1967) as NASA TT F-450, TT
F-451, TT F-452, Washington, DC: NASA. [161]
Poincare, J. H. 1893 Les Methodes Nouvelles de la Mecanique Celeste, vol. II. Paris:
Gauthier-Villars. [161,183]
Poincare, J. H. 1899 Les Methodes Nouvelles de la Mecanique Celeste, vol. III.
Paris: Gauthier-Villars. [161,254]
Poincare, J. H. 1905 La Valeur de la Science. Paris: Flammarion. [127]
Pol, B. van der See van der Pol, B.
Pomeau, Y. & Manneville, P. 1980 Intermittent transition to turbulence in dissipa¬
tive dynamical systems. Commun. Math. Phys. 74,189-97. [263]
Priestley, M. B. 1981 Spectral Analysis of Time Series. London: Academic. [277]
Priestley, M. B. 1988 Non-linear and Non-stationary Time Series Analysis. London:
Academic. [277]
Putnam 1947 Problem 1. Amer. Math. Monthly 54,401. [47]
Putnam 1967 Problem A-3. Amer. Math. Monthly 74,772. [46]
Rayleigh, J. W. S. 1883. On maintained vibrations. Phil. Mag. 15, 229-35. Also Sci.
Papers2(1900), 188-93, Cambridge University Press. [191]
Rayleigh, J. W. S. 1894 The Theory of Sound. 2nd edn. London: Macmillan. [191]
Read, P. L., Bell, M. J., Johnson, D. W. & Small. R. M. 1992 Quasi-periodic and
chaotic flow regimes in a thermally driven, rotating annulus, J. Fluid Mech.
238, 599-632. [272,273,275, 277]
Richardson, L. F. 1961 The problem of contiguity: an appendix to Statistics of
Deadly Quarrels. General Systems Yearbook 6, 139-87. Also Collected Papers 2
(1993), ed. O. M. Ashford, H. Chamock, P. G. Drazin, J. C. R. Hunt, P. Smoker &
I. Sutherland, Cambridge University Press. [130]
 Bibliography and author index 307

Rikitake, T. 1958 Oscillations of a system of disk dynamos. Proc. Camb. Phil Soc
54, 89-105. [280]
Rossler, O. E. 1976 An equation for continuous chaos. Phys. Lett. A57 397-98
[163]
Ruelle, D. 1979 Sensitive dependence on initial condition and turbulent behavior
of dynamical systems. Ann. NY. Acad. Sci. 316,408-16. [78]
Ruelle, D. & Takens, F. 1971 On the nature of turbulence. Commun. Math. Phys. 20,
167-92. [87,261]
Ryrie, S. C. 1992 Unsteady three-dimensional Benard convection: light-scattering,
statistics and chaos. Fluid Dyn. Res. 9,19-57. [268,271]
Sarkovskii, A. N. 1964 Coexistence of the cycles of a continuous mapping of the
real line into itself. Ukrain. Math.Zh. 16,61-71. [132]
Shannon, C. E. & Weaver, W. 1949 The Mathematical Theory of Communication.
Urbana, IL: University of Illinois Press. [141]
Sierpinski, W. 1915 Sur une courbe dont tout point est un point de ramification.
Comptes Rendus (Paris) 160,302-5. [145]
Sierpinski, W. 1916 Sur une courbe cantorienne qui contient une image biunivoque
et continue de toute courbe donnee. Comptes Rendus (Paris) 162,629-32. [146]
Smale, S. 1967 Differentiable dynamical systems. Bull. Amer. Math. Soc. 73, 747-
817. [96]
Sparrow, C. 1982 The Lorenz Equations: Bifurcations. Chaos, and Strange At¬
tractors. New York: Springer-Verlag. Appl. Math. Sci. 41. [277]
Stein, G. 1922 Sacred Emily. A poem from the book Geography and Plays. Boston,
MA: Four Seas. [56]
Stuart, A. M. 1990 The global attractor under discretization. On pp. 211-26 of
Continuation and Bifurcations: Numerical Techniques and Applications, ed. Roose,
D., De Dier, B. & Spence, A., Dordrecht, Holland: Kluwer Academic. [113]
Swift, J. & Hohenberg, P. C. 1977 Hydrodynamic fluctuations at the convective
instability. Phys. Rev. A15,319-28. [285]
Takens, F. 1981 Detecting strange attractors in turbulence. On pp. 366-81 of Dy¬
namical Systems and Turbulence, ed. Rand, D. A. & Young, L.-S., New York:
Springer-Verlag. Lecture Notes in Math. 898. [266]
Thom, R. 1975 Structural Stability and Morphogenesis. Reading, MA: Benjamin
English version of French book of 1972. [62,63,67]
Toda, M. 1967 Vibration of a chain with nonlinear interaction. J. Phys. Soc. Japan
22,431-36. [164]
Turing, A. M. 1952 The chemical basis of morphogenesis. Phil. Trans. Roy. Soc.
Lond. B237,37-72. [287]
van der Pol, B. 1922 On oscillation hysteresis in a simple triode generator. Phil.
Mag.(6)43,700-19. [192, 223]
van der Pol, B. 1926 On 'relaxation-oscillations’. Phil. Mag. (7)2,978-92. [190]
van der Waals, J. D. 1873 On the Continuity of Gaseous and Liquid States of Matter.
Doctoral thesis, University of Leiden (in Dutch). English transl. on pp. 121-239
of On the Continuity of Gaseous and Liquid States of Matter, ed. Rowlinson, J. S.
(1988), Amsterdam: North-Holland. [65]
308 Bibliography and author index

Verhulst, P. F. 1838 Notice sur la loi que la population suit dans son accroissement.
Corr. Math. Phys. 10, 113-21. [10]
Volterra, V. 1926 Variazioni e fluttuazioni del numero d’individui in specie animali
conviventi. Mem. Accad. Naz. Lincei 2, 31-113. English transl. on pp. 409-48 of
Animal Ecology by Chapman, R. N. (1931), New York: McGraw-Hill. [200]
von Koch, H. 1904 Sur une courbe continue sans tangente, obtenue par une con¬
struction geometrique elementaire. Arkiv for Mat. Astron. Fys. 1, 681-704.
[129]
von Neumann, J. 1951 Various techniques used in connection with random digits.
J. Res. Nat. Bur. Stand. 12, 36-8. Also Collected Works 5 (1961), ed. A. H. Taub,
pp. 768-70, Oxford: Pergamon. [76. 88]
Waals, J. D. van der See van der Waals, J. D.
Watson, E. B. B., Banks, W. H. H., Zaturska, M. B. & Drazin, P. G. 1990 On
transition to chaos in two-dimensional channel flow symmetrically driven by
accelerating walls. J. Fluid Mech. 212,451-85. [289]
Weiss, P. 1907 L’hypothese du champ moleculaire et la propriete ferromagnetique.
J. Phys. (Paris) 6, 661-90. [66]
Zubov, V. I. 1964 Methods of A. M. Lyapunov and Their Application. Groningen,
Holland: Noordhofl". English transl. of Russian book of 1957. [206]

Additional bibliography
Gor’kov, L. P. & Eliashberg, G. M. 1968 Generalization of the Ginzburg-Landau
equations for non-stationary problems in the case of alloys with paramagnetic
impurities. Sov. Phys. JETP 27, 328-34. [286]
Gray, J. 1992 Poincare, topological dynamics, and the stability of the solar system.
On pp. 503-24 of An Investigation of Difficult Things, ed. P. M. Harman & A. E.
Shapiro, Cambridge University Press. [161]
Poincare, J. H. 1885b Sur les courbes definies par les equations differentielles. J.
Math. Pures Appl. (4) 1, 167-244. Also Oeuvres 1 (1928), ed. P. Appel, pp.
90-161, Paris: Gauthier-Villars. [156]
Poincare, J. H. 1890 Sur le probleme des trois corps et les equations de la dynamique.
Acta Math. 13, 1-270. Also Oeuvres 7 (1952), ed. J. Levy, pp. 262-A19. Paris:
Gauthier-Villars. [167]
Segur, H. 1982 Solitons and the inverse scattering transform. On pp. 200-77 of
Topics in Ocean Physics, ed. A. R. Osborne & P. Malanotte Rizzoli, Amsterdam:
North-Holland. [282]
 Motion picture and video index

Various films and videos about nonlinear systems have been made. Some, which
may be rented or bought, are listed below. It it very instructive to see, in particular,
computer-made animations of solutions of nonlinear systems in order to visualize
their geometrical properties.

The following film may be obtained on application to Aerial Press, Inc., PO Box
1360, Santa Cruz, CA 95601, USA. Aerial Press also sells various relevant pro¬
grams on discs suitable for PCs and compatible computers.
Stewart, H. B. F1987 The Lorenz system. 25 mins., 16 mm, colour. [242]

The following video may be obtained on application to the American Mathe¬

matical Society, PO Box 6248, Providence, RI02940, USA.
Devaney, R. L. V1989 Chaos, fractals and dynamics Computer experiments in
mathematics. 60 mins., colour. [38]

The following video may be obtained on application to W. H. Freeman, 20 Beau¬

mont Street, Oxford OX1 2NQ, England.
Peitgen, H. O., Jurgens, H., Saupe, D. & Zahlten, C. V1990 Fractals: an animated
discussion. 63 mins., colour. [109]

Also much computer software, both in the public domain and for sale, is a source
of fun and an important aid to learning.

309
 Subject index

ABC flows, 166 saddle-node, 8,22,199

accumulation, point of, 75 subcritical, 16, 111, 188,190,211,237,239
aliasing, 266 supercritical, 15, 21,61, 111, 188
a-limit set, 75, 153 Takens-Bogdanov, 210
area-preserving map, 80, 120-1, 154, 165 transcritical, 10-13,18,21-2, 38-9, 52-3,
Arnol'd tongues, 117-18 56-7,72,201,283
Arnol'd'scat map, 121-2, 148 turning-point, 8-10. 21 -2, 26-7. 39, 50-3,
associated Legendre function, 288, 300 56,63,72,283
asymptotic method, bifurcation diagram, 4,9, 11-12, 15, 17, 22,
averaging, 192-5. 211-13 35-6, 51 -2,57-8,66,88-90,99, 135,
Licnard's, 195 221,237
Lindstedt-Poincare, 181-6, 188-90, 207- bifurcation point, 48
8, 277-9 bifurcation sequence, 63, 83-6, 133, 242,
Mel'nikov’s, 257-61,280-2 261-3
van der Pol’s, 192, 248 bifurcation set, 54,64,67
weakly nonlinear, 50-2, 59-61,83-6, 163, bifurcation value, 76, 153
188-90, 194-5,201-2.210-13,215-31, binary numbers, 78
247, 277-9 bistability, 14
asymptotic stability, 70, 151-2, 204-5 Bloch theory, 158
asymptotically fixed point, 74 blow-up, 8-10, 12, 16, 289
asymptotically periodic point, 74 box-counting dimension, 128-31, 145-8
attraction, branch point, 48
basin of. 76, 153 Brusselator, 201
domain of, 76, 104-5,111-13,123,153, 282 butterfly catastrophe, 62
attractor, 75,82, 86,99, 104-5, 153, 155,177, butterfly effect, 2. 38
186-7.238-9
strange, 87.96, 125,131,241-2,251 Cantor sets, 87,94-6.98, 100.125,127,239
autocorrelation, 265, 282 Cantor’s middle-thirds set, 100,125,144-5,
autonomous system, 30,48.69, 149-50, 161. 147
170, 186-7 cards, shuffling, 111-12
averaging, method of, 192-5, 211-13 cardioid, 294
cat map, 121-2, 148
baker’s transformation. 96-8, 147, 148 catastrophe,
basin of attraction, 76, 153 butterfly, 62
behaviour variables, 59 cusp, 54-5,62-5,222
Beltrami flow, 166 elliptic umbilic, 62
Bendixson’s criterion, 208 fold, 8,55.62
Bernoulli shift, 76-9,90,97, 104, 111-12, hyperbolic umbilic, 62
114-15, 142. 147 parabolic umbilic, 62
Bernoulli's equation. 11 swallow's tail, 62,67
bifurcation. 2, 5, 23. 58, 76.283 catastrophe theory, 62
flip. 34-5.63,83-6. 133. 135 Cauchy-Riemann relations, 295
fold, 8, 55,62, 64 cave, Plato's metaphor of, 277
Hopf, 19-22.41.72,188-90,237,262. centre, 23,26,175
277-9 centre manifold, 72,283
pitchfork, 13-17,21 -2, 52-3,57,59-65, chaos, 3,74,79,87,133,143,187,233-77
111, 163,283 deterministic, 93-4

310
 Subject index 311

routes to, 86-7,133,135,242,245,254-7, equation

261-4 quadratic, 32,104
chemical basis of morphogenesis, 287 two-dimensional, 31,80-1,96-110,
chemical reactions, 200,287 119-24,155-6
circle maps, 79 differential equation, 7,23,149 etc.
closed set, 126-7 differentially heated rotating annulus, 272
cluster point, 75, 111, 126,152 diffusion equations, 206,283-9
coastline’s fractal shape, 130-1 dimension, 127
co-dimension, 62 box-counting, 128-31,145-8
collision of attractors, 99 fractal, 131,239,267
combustion, 202,287 Hausdorff, 128,131,143
compatibility condition, 61,278 topological, 127
complete elliptic integral, 28,42,44,115 direct method of Liapounov, 178-80,203-6,
computational methods, 29,94-6 208-9,237-8
computer programs, 95,103,106-9 disconnected set, 87,126
condition, dissipative system, 155
compatibility, 61,278 distinguished limit, 219,228
solvability, 61,278 domain of attraction, 76,104-5,111-13,
conjugacy, topological, 115 123,153,282
conjugate maps, 115 double point, 53
conjugate point, 53,58 doubling, period, 34-5,63,83-6,133,160,
connection, 229,242,263
heteroclinic, 27,244,262,280-1 dry friction, 199
homoclinic, 26,243-7,253-7,260,262, Duffing’s equation, 150,214-17,219-22,
280-1 225-7,230,246-51,260-1
conservation of energy, 23,154-5 dynamical system, 1
conservative system, 23-8,42,164-5 dynamo, Earth’s 279
continued fraction, 45
contraction, 96-9 eigenvalue, 20-1,47,71,98,110,138-9,152,
control variables, 59 159-60,173,235,278,284-8
convection, thermal 233-4,268,271-2, eigenvector, 47,71,110,138-9,152,159—
285 60,173,278,284-8
coordinate, generalized, 154,165 elastica, 4,43-4
co-rank, 62 elliptic functions, 115,279
correlation, 265-6,282 elliptic integral, 28,42,44,115
Coulomb friction, 199 elliptic umbilic, 62
countable set, 125-7 energy, 155
critical point, 7,48,151 kinetic, 154,165,180
critical pressure, 65 potential, 23-5,154,165,180
critical temperature, 65 energy equation, 23-4,43,155,183
critical volume, 65 energy integral, 23-4,43,155,183
cubic map, 123 energy method, 178
Curie temperature, 66 entrainment, 219,262
cusp, 52-5,62-5,222 entropy, Kolmogorov, 141
cusp catastrophe, 54-5,62-5,222 epidemics, 199
cycle, 34,73 equations,
limit, 21,41,167,186-90,192,208- Bernoulli's, 11
13.242 difference, 29,32,46,48,68 etc.
differential, 7,23,27-30,149 etc.
damped oscillation, 186,192-4,200,215, diffusion, 283-9
219-20.222,230,246-50 Duffing’s, 150,214-17,219-22,225-7,
degenerate node, 175-6 230,246-51.260-1
delays, method of, 266-7 Euler-Lagrange, 298
Descartes, folium of, 39 Euler's, 162
deterministic chaos, 93-4 Ginzburg-Landau, 286
devil's staircase, 144,262 Hamilton’s, 154,165,180,253
difference equations, 29,32,46,48,68 Hill’s, 168
logistic, 32-5; see also logistic difference integro-differential, 64
312 Subject index

equations (cont.) flow,

Korteweg-de Vries, 43 Beltrami, 166
Lagrange’s, 16S incompressible, 154,166
Landau’s, 13,39,41 focus, 21,174-7
Lienard’s, 171,192,199,203 fold, 8,55,62
logistic difference, 32; see also logistic folding, 82,96-9,256
difference equation folium of Descartes, 39
logistic differential, 10-13,29 forced oscillation, 214-32,246-50
Lorenz’s, 163,233,272,277-9 forcing term, 215
Lotka-Volterra, 200 Fourier transform, 266,268-9,271-3,
Mathieu’s, 164,168-9,231-2 275
Meissner’s, 168 fractal, 131
nonlinear diffusion, 206,284-9 fractal dimension, 131,239,267
nonlinear Schrodinger, 285 fraction, continued, 45
partial differential, 206,283-300 fractional part, 76-7,90
Proudman-Johnson, 288 Frank-Kamenetskii parameter, 202,287
Rayleigh’s, 171,191,195,208,213 Frechet derivative, 110,138
recurrence, 29 free oscillations, 214,217,228
van der Pol’s, 171,217-19,223-5 frequencies, fundamental, 156-7,215,262,
van der Waals, 65 267-8,272
equilibration, 14-15 frequency locking, 117,262
equilibrium point, 7,48,151 friction, 199
Euler finite-difference schemes, 29 frictionless system, 155
explicit, 29,46,113,164-5 function,
implicit, 164-5 associated Legendre, 288,300
Euler strut, 4,43-4 elliptic, 115,279
Euler-Lagrange equation, 165,298 Hamiltonian, 154,180,253
Euler’s equations of motion, 162 implicit, 48-50,59
eventually fixed point, 74 Liapounov, 178-80,203-6,208-9,238
eventually periodic point, 74 Mel’nikov, 259-61,280-2
explicit finite-difference schemes, 29,46,113, parabolic cylinder, 288
164-5 periodic, 21,34,72,153,156,158 etc.
exponents, positive definite, 179
Floquet, 160,168,268 positive semidefinite, 179
Liapounov, 141-3,148,268 probability density, 93
exponents of normal modes, 7,10,18-19, probability distribution, 93,145
152,173,235-6,284-9 quasi-periodic, 156-7,215,262,265,
267-8,271-2
fast Fourier transform, 266 functional, Liapounov, 206
Feigenbaum sequence, 86-8,99,118,133, fundamental frequencies, 156-7,215,262,
135,148 267-8,272
ferromagnetism, 66 fundamental matrix, 160,168,204
Fibonacci numbers, 45,47
films, 242,309 generalized coordinate, 154,165
finite-difference schemes, 29-30,46,113, generalized momentum, 154,165
164-5 geomagnetism, 279
first return map, 31 Ginzburg-Landau equation, 286
fixed point, 32,48 globally asymptotic stability, 70,151
asymptotically, 74 golden mean, 45
eventually, 74 gradient systems, 62,162
stability of, 33-4,70-1 growth of population, 10,32,200
fixed singularity, 10
fixed-point theorem, 44,111 Hamiltonian, 154,165,180,253
flip bifurcation, 34-5,63,83-6,133,135 Hamiltonian function, 154,180,253
Floquet exponent, 160,168,268 Hamiltonian system, 154,180,253,281
Floquet multiplier, 160,168 Hamilton's equations, 165
Floquet theory, 158-61,167-9,227,268 hard spring, 183,205,225
flow, 151,167 harmonic motion, simple, 23,154,162,165,
 Subject index 313

181-2,191-4,215 Koch snowflake curve, 129-30

harmonics, 267-8 Kolmogorov entropy, 141
Hausdorff dimension, 128,143 Korteweg-de Vries equation, 43
Henon map, 98-103,119
Henon-Heiles system, 164 Lagrange’s equations, 165
heteroclinic connection, 27,244,262,280-1 Lagrange’s identity, 185,231,300
heteroclinic orbit, 27,244,262,280-1 Lagrange’s theorem, 180
Hilbert’s sixteenth problem, 187 Lagrangian, 165
Hill’s equation, 168 Landau’s equation, 13,39,41
homoclinic connection, 243-6,253 Landau’s theory of phase transition, 65-6
homoclinic orbit, 26,243-7,253-7,260,262, Legendre function, associated, 288,300
280-1 lemniscate, 39
homoclinic tangle, 254-6 Liapounov exponents, 141-3,148,268
Hooke’ law, 183 Liapounov function, 178-80,203-6,208-9,
Hopf bifurcation, 19-22,41,72,188-90,211, 238
237,262,277-9 Liapounov functional, 206
horseshoe map, 96-7,256 Liapounov multipliers, 141
hyperbolic umbilic, 62 Liapounov number, 141
hyperbolic point of equilibrium, 174,177 Liapounov-Schmidt method, 6,283
hysteresis, 36,222,242,261,290 Liapounov’s direct method
for a difference equation, 203
image, 76,104,254-6 for an ordinary differential system, 178-
imperfect system, 37,55-9,64 80,203-6,208-9,237-8
imperfection, 55-9 for a partial differential equation, 206
implicit finite-difference schemes, 164-5 Liapounov’s second method, 178-80,203-6,
implicit function theorem, 50,59 208-9,237-8
implicit mid-point scheme, 164-5 Liapounov’s sense of stability, 70,73,109,
improper node, 175-6 151,158,167
incompressible fluid, 154,166 Liapounov’s theorem, 179,237
infinitesimally stable, 152 Lienard’s construction, 199
inflected node, 175-6 Lienard’s equation, 171,192,199,203
information theory, 141 Lienard’s method, 195
initial conditions, sensitive dependence on, Lienard’s plane, 199
78-9,92-3,115,140-3,241,265 limit,
instability, distinguished, 219,228
infinitesimal, 152 Newtonian, 43,207
linear, 38,152 limit cycle, 21,41, 167, 186-90, 192,
phase, 66 208-13, 242
structural, 55-8,64-5,250 limit point, 75,152
subcritical, 17,211,237,261,279 Lindstedt-Poincare method, 181-6,188—
integral, elliptic, 28,42,44,115 90,207-8,277-9
integro-differential equation, 64 linear stability, 7, 38, 152
intermittent transition, 263-4,282 linearized system, 7,10,13,18-19,23,33,70,
invariant set, 72,75-6,104,119,121-2,140, 73,152,173,235
144,166-7 Liouville's theorem, 154
isola, 58 locking, frequency, 117,262
iterated map, 29,103,112,121 logistic difference equation, 32,104
attractors, 75-6,86,114
Jacobi ellipsoid, 6 bifurcation diagram, 35,88-90, 135
Jacobian, 80-1,152 etc. computer program, 95
Jacobian determinant, 80-1,97,99,153,173 cycles, 34,82-3,113-14
Jacobian elliptic functions, 115,279 domains of attraction, 76
Jacobian matrix, 19,59,70,72,110,123,152, fixed points, 33
160,173 Liapounov exponents, 142
Jacobi-Liouville formula, 168 period doubling, 34,82-3,136,147-8
Julia set, 104,106,107 stability of cycles, 74
stability of fixed points, 34
superstability, 118,133-4
kinetic energy, 154,165,180 logistic differential equation, 10-13,29
314 Subject index

logistic map, 33,46,71; see also logistic Mel’nikov’s method, 257-61,280-2

difference equation metastability, 70
Lorenz system, 163,234,272,277-9 method,
Lotka-Volterra equations, 200 asymptotic, see asymptotic method
Lozi map, 119-20 averaging, 192—5,211 — 13
energy, 178
Maclaurin spheroid, 6 Liapounov’s direct, 178-80,203-6,208-
magnetic pole, 280 9,237-8
Mandelbrot set, 105,108-9,123 Liapounov-Schmidt, 6,283
manifold, Lindstedt-Poincare, 181-6,188-90,
centre, 72,283 207-8, 277-9
stable, 243-4,247,250,253-8, 280 Mel’nikov’s 257-61,280-2
unstable, 243-4,247,250,253-8 Newton-Raphson, 30,45,67,112,134
maps, normal modes, 7,10,19,152,173,235-6,
area-preserving, 80,120-1,154,165 284-8
baker’s, 96-8,147-8 numerical, 28-30,46,94,113,164-5
cat, 121-2,148 Poincare-Lindstedt, 181-6,188-90,
circle, 79 207-8, 277-9
conjugate, 115 van der Pol’s, 192,248
cubic, 123 method of averaging, 192-5,211-13
first return, 31 method of delays, 266-7
Henon, 98-103,119 method of Liapounov, direct, 178-80,203-
horseshoe, 96-7,256 6,208-9,237-8
iterated, 29, 103,112,121 modes, normal, 7,10,19,152,173,235-6,
logistic, 33,46, 71; see also logistic 284-8
difference equation momentum, generalized, 154,165
Lozi, 119-20 morphogenesis, chemical basis of, 287
measure-preserving, 80, 120-1,154, 165 motion pictures, 242, 309
Poincare, 31,161-2,166-7,250-6,260, movable singularity, 10
266 multipliers, Floquet, 160,168
reflection, 123
return, 31 Newton-Raphson method, 30,45,67,112,
rotation, 79-80,112,117,120,123,156 134
sawtooth, 76-7, 79,91,111-2 Newtonian limit, 43,207
second-generation, 47,73,82-3,111,115- node, 173-7,199
17 degenerate, 175-6
shear, 80,120,123 improper, 175-6
sine, 117 inflected, 175
square, 104 proper, 176
standard, 121 stable, 173-7
stroboscopic, 31, 162,250-6,260 unstable, 173-7
tent, 114-5,125,144,148 noise, 39,265-7
topologically conjugate, 115 non-autonomous system, 149,161
translation, 123 non-dissipative system, 155
twist, 112 nonlinear Schrodinger equation, 285
margin of stability, 7,38,169,283 nonlinear system, 1 et seq.
Mathieu’s equation, 164, 168-9,231-2 normal modes, 19,152, 173,235-6,284-8
matrix, numerical methods, 29-30,46,94 113
fundamental, 160,168,204 164-5
Jacobian, 19, 59, 70,72, 110,123,152,160
mean, 114 w-limit set, 75,80,152-3
mean, golden, 45 open set, 76
measure-preserving map, 80,120-1,154, orbit, 21,30,150-3,155-6
165 heteroclinic, 27,244,262,280-1
measure-preserving system, 153-5 homoclinic, 26,243-7,253-7,260,262,
Meissner's equation, 168 280-1
Mel’nikov distance, 258 planetary, 207
Mel’nikov function, 259-61,280-2 relativistic, 207
 Subject index 315

orbital stability, 159-60,167,224 260,266

oscillations, Poincare pear-shaped body, 6
damped, 186,192-4,200,215,219-20, Poincare-Bendixson theorem, 186-8,192,
222,230,246-50 209,214
forced, 214-32,246-50 Poincare-Lindstedt method, 181-6,188—
free, 214,217,228 90,207-8,277-9
linear, 23,154,162,165,181-2.191-4,215 Poincare's sense of stability, 159-60,167
etc. points,
nonlinear, 22-8,41-3,186,191,206-7 etc. accumulation, 75
relativistic, 42-3 cluster, 75, 111, 126, 152
relaxation, 197 conjugate, 53, 58
self-excited, 190-2,194-7,209-13 critical, 7,48,151
simple harmonic, 23,154,162,165,181-2, double, 53
191-4,215 equilibrium, 7,48, 151
synchronous, 168,216,225,229,247 fixed, 32,48
weakly nonlinear, 181,188-90,194-5, hyperbolic, 174,177
207-8,211-13,227 limit, 75,152
periodic, 34,72,74,156
parabolic umbilic, 62 regular, 52
parabolic cylinder function, 288 saddle, 26,174,235,242,247,250,253,
parameter, 2,4,48,55,59,72 280-1
parametric resonance, 228 singular, 53
part, fractional, 76-7,90 spiral, 174
partial-differential problems, 206,283-9 star, 176
pattern selection, 2 turning, 8-10,18,21 -2,26-7,38-9,50-3,
pendulum, simple, 27-8,213,219 56,63,72
perfect set, 127 point of equilibrium, 7,48,151
perfect system, 37,55 Poisson brackets, 281
perihelion, 208 population growth, 10,32,200
period doubling, 34-5,63,83-6,133,160, positive definite function, 179
229,242,263 positive semidefinite function, 179
see also subharmonic potential, 23,62
period halving, 242 potential energy, 23-5,154,165,180
‘period three implies chaos’, 133 power spectrum, 266,268-9,271-3,275
periodic function, 21,34,72,156,158 etc. precession of perihelion, 208
periodic point, 34,72,74,156 etc. pre-chaos, 244
asymptotically, 74 pre-image, 76,104,254-5
eventually, 74 prey-predator equations, 200
periodic solution, 21,24,34,72,153,155-61, probability density function, 93
186 etc. probability distribution function, 93,145
see also limit cycle Proudman-Johnson equation, 288
perturbation, 7, 18, 151,284
regular, 181 quadratic difference equation, 32,104
phase instability, 66 q uasi-periodic function, 156-7,215,238,
phase plane, 21,155 262,265,267-8,271
phase portrait, 21,171,201 quaternary number, 145
phase space, 30
phase transitions, 65-6,133 random number, 92-3,145
pigeon-hole principle, 78,80,294 random variable, 87,92-3,145
pin-ball machine, 87,133 Rayleigh-Benard convection, 234,268,285
pitchfork bifurcation, 13-17,21 -2,52-3,57, Rayleigh’s equation, 171,191,195,208,213
59-65,111,163,283 reactions, chemical, 200,287
plane, recurrence, 166-7
Lienard, 199 recurrence equation, 29
phase, 21,155 reflection map, 123
van der Pol, 223 regular perturbation, 181
planetary orbits, 207 regular point, 52
Poincare maps, 31,161-2,166-7,250-6, relativistic harmonic oscillator, 42-3
 316 Subject index

relativistic planetary orbits, 207 Mandelbrot, 105,108-9,123

relaxation oscillation, 197 co-limit, 75,80,152-3
renormalization group theory, 133-40 perfect, 127
repeller, 75-6,87,125,133,153,177 totally disconnected, 87, 126
resonance, 215 uncountable, 125-7
parametric, 228 shadows on the wall, Plato’s metaphor of,
response, 215 277
return map, 31 shear map, 80,120, 123
stroboscopic, 31,162,250-6, 260 shuffling cards, 111-12
riffle shuffle of cards, 111-12 Sierpinski carpet, 146
Rikitake dynamo, 279 Sierpinski gasket, 145
Rossler system, 163,268-9 similarity, self-. 100.105. 127, 130-1.133,
rotating annulus, differentially heated, 272 144-6
rotation map, 79-80, 112,117, 120, 123,156 simple harmonic motion, 23. 154, 162, 165.
routes to chaos, 86-7, 133,135,242,254-7, 181-2, 191-4,215
261-4 simple pendulum, 27-8,213,219
Ruelle-Takens-Newhouse route to chaos, simple turning point, 8-10, 18, 21-2, 38-9,
261 50-3,56-7 '
sine map, 117
saddle point, 23,26,174, 235, 242, 247, 250, singular point, 53
253,280-1 singularity,
saddle-focus, 237 fixed, 10
saddle-node bifurcation, 8,22,199 movable. 10
Sarkovskii’s theorem, 132-3,147 singularity theory, 62
sawtooth map, 76-7,79,91, 111-2 sink, 175-6
scaling, 127,130-1,133,135,148,263 snowflake curve, 129-30
scatter diagram, 266 soft spring, 183, 219
Schrodinger equation, nonlinear, 285 solitary wave, 43
second method of Liapounov, 178-80,203- soliton, topological, 43
6, 208-9,237-8 solvability condition, 61,278
second-generation map, 47, 73,82-3, 111, source, 175-6
115-17 spectrum,
secular terms, annihilation of, 185, 189-90, Fourier, 266, 268-9, 271-3, 275
218,220,226,229 power, 266, 268-9, 271-3. 275
self-excited oscillations, 190-2, 194-7, 209- spiral point, 174
13 spontaneous combustion of a slab. 287-8
self-similarity, 100, 105,127,130-1, 133, spring,
144-6 hard, 183, 205,225
semistability, 41 linear, 183
sensitive dependence on initial conditions, nonlinear. 42, 183, 205,219, 225
78-9,92-3,115,140-3,241,265 soft, 183,219
separatrix, 26, 183,247 square map, 104
sequence, stability,
bifurcation, 63, 83-6, 261-3 asymptotic, 70, 151-2,204-5
Feigenbaum, 86-8,99,118, 133, 135, 148 globally asymptotic, 70, 151
uniformly distributed, 92 infinitesimal, 152
set, Liapounov, 70,73, 109, 151, 158, 167
a-limit, 75,153 linear, 7, 38, 152
bifurcation, 54,64,67 margin of, 7, 38, 169,283
Cantor, 87,94-6,98,100,125,127.239 orbital, 159-60,167,224
Cantor’s middle-thirds, 100,125,144-5, Poincare, 159-60, 167
147 structural, 55-8,64-5,250
closed, 126-7 supercritical, 15
countable, 125-7 stable manifold, 243-4,247,250,253-8,280
disconnected, 87,126 stable node, 173-7
invariant, 72, 75-6,104, 119,121-2,140, stagnation point, 166
144,166-7 standard map, 121
Julia, 104,106, 107 star point, 176
 Subject index 317

state variables, 59 transcritical bifurcation, 8-10,21-2, 26-7,

stiffness, 40,230,246 38-9, 50,52-3,56,63,72
strange attractor, 87,96,125,131,241-2,251 transient, 215, 225
stretching, 82,96-8,256 transition to chaos, 261
stroboscopic return map, 31,162, 209, transitive, topologically, 143
250-6, 260 translation, 123.184,188.210,227,253,257
structural stability, 55-8,64-5,250 trapezoidal scheme, 164-5
Sturm-Liouville problem, 288 turbulence, 261-2,283
subcritical bifurcation, 16,111,188,190,211 turning point, 6-10. 18, 21 -2,26-7, 38-9,
subcritical instability, 17,211,237,261,279 50, 52-3, 56,63, 72
subharmonic, 160, 168,225-9, 231, 268 regular, 53
subharmonic instability, 160,263,268 simple, 8, 53
see also period doubling singular, 53
superconductivity, 286 twist map, 112
supercritical bifurcation, 15,21,61, 111,188 two-dimensional difference equations, 31,
supercritical stability, 15 80-1,96-110,119-24, 155-6
superstability, 118,133
suspended system, 150, 161,250,253
umbilic,
swallow's tail, 62,67
elliptic, 62
Swift-Hohenberg model, 285
hyperbolic, 62
symmetry breaking, 2,5,17,35,57
parabolic, 62
synchronous oscillation, 168, 216, 225, 229, uncountable set, 125-7
247
uniformly distributed sequence, 92
system, universal constants,
autonomous, 30,48,69, 149-50,161,170, a, 133, 135-40
186-7 <5,86, 118, 121, 133, 135, 139-40
conservative, 23-8,42,164-5 unstable manifold, 243-4, 247,250,253-8
dissipative, 155 unstable node, 173-7
dynamical, 1
gradient, 62, 162
Hamiltonian, 154,180,253,281 van der Pol plane, 223
imperfect 37, 55-9,64 van der Pol’s equation, 171,217-19, 223-5
linearized, 7,10,13,18-19,23,33,70,73, van der Pol’s method, 192, 248
van der Waals equation, 65
152,173,235
Lorenz, 163,234,268,272,277-9 variance, 114
variables,
non-autonomous, 149,161
behaviour, 59
nonlinear, 1 el seq.
control, 59
perfect, 37,55
random, 87,92-3, 145
Rossler, 163, 268-9
state, 59
suspended, 150,161,250,253
videos, 38, 109, 309
volume-preserving map, 80
Takens vector, 266-7
von Koch snowflake curve, 129-30
Takens-Bogdanov bifurcation, 210
Volterra equations, 200
tent map, 114-15, 125, 144, 148
vortex, 175
ternary number, 125,127
thermal convection, 233-4, 268, 271-2,285
time series, 198,239,264-77 wave, solitary, 43
Toda lattice, 164 weakly nonlinear approximations, 50-2,
tongues, Arnol’d, 117-18 59-61,83-6, 163,181,188-90, 194-5,
topological conjugacy, 115 201 -2,210-13,215-31,247,277-9
topological dimension, 127 weakly nonlinear oscillations, 181,188-90,
topological soliton, 43 194-5,207-8,211-13,227
topologically transitive, 143 wedge product, 258
torus, 121,148,155-7,166,239,262,280 Weierstrass function, 122
totally disconnected set, 87,126 Weiss’s theory of ferromagnetism, 66
trajectory, 21 windows, 87,99,133,142,251
 '

.
The theories of bifurcation, chaos and fractals as well as equilibrium,
stability and nonlinear oscillations, are-parts of the theory of the evolution
of solutions of nonlinear equations. A wide range of mathematical tools
and ideas are drawn together in the study of these solutions, and the
results applied to diverse and countless problems in the natural and social
sciences, even philosophy.
The text evolves from courses given by the author in England and the
United States. It introduces the mathematical properties of nonlinear
systems, mostly difference and differential equations, as an integrated
theory, rather than presenting isolated fashionable topics. Topics are
discussed in as concrete a way as possible and worked examples and
problems are used to explain, motivate and illustrate the general prin¬
ciples. The essence of these principles, rather than proof or rigour, is
emphasized. More advanced parts of the text are denoted by asterisks,
and the mathematical prerequisites are limited to knowledge of linear
algebra and advanced calculus, thus making it ideally suited to both senior
undergraduates and postgraduates from physics, engineering, chemistry,
meteorology etc. as well as mathematics.

Cambridge Texts in Applied Mathematics

EDITORS
Professor H. Aref, Institute of Geophysics and Planetary Physics, University of
California at San Diego, USA.
Professor D.G. Crighton, Department of Applied Mathematics and Theoretical
Physics, Cambridge, UK.
The aim of this series is to provide a focus for publishing textbooks in applied
mathematics at the advanced undergraduate and beginning graduate level. It is
planned that the books will be devoted to covering certain mathematical techniques
and theories and exploring their applications.
The main audience for the series will be in departments of applied mathematics,
engineering science, engineering or physics. It is important to recognise that for
students, clear and careful exposition and a sound pedagogic presentation are
frequently more relevant than surveying and synthesising recent developments.
Authors will be encouraged to provide plenty of worked examples and exercises, of
varying degrees of difficulty, for the reader. Books in the series should provide a solid
understanding of how a given method can usefully be applied to help solve problems
in physics and engineering.

Cover design by The Pinpoint Design Company

Cambridge
UNIVERSITY PRESS