(Applied Mathematical Sciences 35) Jack Carr (Auth.) - Applications of Centre Manifold Theory-Springer-Verlag New York (1981)

Applied Mathematical Sciences I Volume 35
Jack Carr
Applications of
Centre Manifold Theory
Springer-Verlag
New York Heidelberg Berlin
Jack Carr
Department of Mathematics
Heriot-Watt University
Riccarton, Currie
Edinburgh EH14 4AS
Scotland
Library of Congress Cataloging in Publication Data
Carr, Jack.
Applications of centre manifold theory.
(Applied mathematical sciences; v. 35)

"Based on a series of lectures given in the
Lefschetz Center for Dynamical Systems in the
Division of Applied Mathematics at Brown University
during the academic year 1978-79"-Pref.
1. Manifolds (Mathematics) 2. Bifurcation
theory. I. Title. II. Series: Applied mathematical
sciences (Springer-Verlag New York Inc.); v. 35.
QA1.A647 vol. 35 [QA613] 510s [516'.07] 81-4431
AACR2
All rights reserved.
No part of this book may be translated or reproduced in any

form without written permission from Springer-Verlag.
The use of general descriptive names, trade names, trademarks,

etc. in this publication, even if the former are not especially
identified, is not to be taken as a sign that such names, as
understood by the Trade Marks and Merchandise Marks Act, may
accordingly be used freely by anyone.
© 1981 by Springer-Verlag New York Inc.
Printed in the United States of America
9 8 7 6 5 4 3 2 1
To my parents
PREFACE
These notes are based on a series of lectures given in
the Lefschetz Center for Dynamical Systems in the Division of

Applied Mathematics at Brown University during the academic
year 1978-79.
The purpose of the lectures was to give an introduction
to the applications of centre manifold theory to differential

equations. Most of the material is presented in an informal
fashion, by means of worked examples in the hope that this
clarifies the use of centre manifold theory.
The main application of centre manifold theory given

in these notes is to dynamic bifurcation theory. Dynamic
bifurcation theory is concerned with topological changes in

the nature of the solutions of differential equations as para-
meters are varied. Such an example is the creation of periodic
orbits from an equilibrium point as a parameter crosses a
critical value. In certain circumstances, the application of
centre manifold theory reduces the dimension of the system

under investigation. In this respect the centre manifold
theory plays the same role for dynamic problems as the
Liapunov-Schmitt procedure plays for the analysis of static

solutions. Our use of centre manifold theory in bifurcation
problems follows that of Ruelle and Takens [57] and of Marsden
and McCracken [51].
In order to make these notes more widely accessible,

we give a full account of centre manifold theory for finite
dimensional systems. Indeed, the first five chapters are de-
voted to this. Once the finite dimensional case is under-

stood, the step up to infinite dimensional problems is
essentially technical. Throughout these notes we give the
simplest such theory, for example our equations are autono-
mous. Once the core of an idea has been understood in a
simple setting, generalizations to more complicated situations
are much more readily understood.
In Chapter 1, we state the main results of centre mani-

fold theory for finite dimensional systems and we illustrate
their use by a few simple examples. In Chapter 2, we prove
the theorems which were stated in Chapter 1, and Chapter 3

contains further examples. In Section 2 of Chapter 3 we out-
line Hopf bifurcation theory for 2-dimensional systems. In
Section 3 of Chapter 3 we apply this theory to a singular per-

turbation problem which arises in biology. In Example 3 of
Chapter 6 we apply the same theory to a system of partial dif-
ferential equations. In Chapter 4 we study a dynamic bifurca-
tion problem in the plane with two parameters. Some of the
results in this chapter are new and, in particular, they con-
firm a conjecture of Takens [64]. Chapter 4 can be read in-
dependently of the rest of the notes. In Chapter 5, we apply
the theory of Chapter 4 to a 4-dimensional system. In Chap-
ter 6, we extend the centre manifold theory given in Chapter
2 to a simple class of infinite dimensional problems. Fin-
ally, we illustrate their use in partial differential equa-
tions by means of some simple examples.
I first became interested in centre manifold theory
through reading Dan Henry's Lecture Notes [34]. My debt to
these notes is enormous. I would like to thank Jack K. Hale,
Dan Henry and John Mallet-Paret for many valuable discussions
during the gestation period of these notes.

This work was done with the financial support of the
United States Army, Durham, under AROD DAAG 29-76-G0294.
Jack Carr
December 1980
TABLE OF CONTENTS
Page
CHAPTER 1. INTRODUCTION TO CENTRE MANIFOLD THEORY . . . 1
1.1. Introduction . . . . . . . . . . . . . . . . 1
1.2. Motivation . . . . . . . . . . . . . . 1
1.3. Centre Manifolds . . . . . . . . . . . . . . 3
1.4. Examples . . . . . . . . . . . . . . . 5
1.5. Bifurcation Theory . . . . . . . . . . . . . 11
1.6. Comments on the Literature . . . . . . . . . 13
CHAPTER 2. PROOFS OF THEOREMS . . . . . . . . . . . . . 14
2.1. Introduction . . . . . . . . . . . . . . 14
2.2. A Simple Example . . . . . . . . 14
2.3. Existence of Centre Manifolds . . . . . . . 16
2.4. Reduction Principle. . . . . . . . 19
2.5. Approximation of the Centre Manifold . . . . 25
2.6. Properties of Centre Manifolds . . . . 28
2.7. Global Invariant Manifolds for Singular
Perturbation Problems. . . . . . . . . 30
2.8. Centre Manifold Theorems for Maps. . . . . . 33
CHAPTER 3. EXAMPLES . . . . . . . . . . . . . . . . . . 37
3.1. Rate of Decay Estimates in Critical Cases. . 37

3.2. Hopf Bifurcation . . . . 39
3.3. Hopf Bifurcation in a Singular Perturbation
Problem. . . . . . . . . . . . . . . . 44
3.4. Bifurcation of Maps . . . . . . . . . . . . . 50
CHAPTER 4. BIFURCATIONS WITH TWO PARAMETERS IN TWO

SPACE DIMENSIONS . . . . . . . . . . . . . . 54
4.1. Introduction . . . . . . . . . . . . . . . . 54
4.2. Preliminaries . . . . . . . . . . . . . . . . 57
4.3. Scaling . . . . . . . . . . . . . . . . . . . 64
4.4. The Case e >0 . . . . . . . . . . . . . . 64
4.5. The Case c < 0 . . . . . . . . . . . . . . 77
4.6. More Scalin . . . . . . 78
4.7. Completion of the Phase Portraits. . . . . . 80
4.8. Remarks and Exercises. . . . . . . . . . . . 81
4.9. Quadratic Nonlinearities . . . . . . . . . . 83
CHAPTER S. APPLICATION TO A PANEL FLUTTER PROBLEM . . . 88
5.1. Introduction 88
5.2. Reduction to a Second Order Equation . . . 89
5.3. Calculation of Linear Terms. . . . . . . . . 93
5.4. Calculation of the Nonlinear Terms . . . . . 95
Page
CHAPTER 6. INFINITE DIMENSIONAL PROBLEMS. . . . . . . . 97
6. 1 . Introduction . . . . . . . . . . . . . . . . 97
6.2. Semigroup Theory . . . . . . . . . . . . . . 97
6.3. Centre Manifolds . . . . . . . . . . . . . . 117
6.4. Examples . . . . . . . . . . . . . . . . . . 120
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . 136
INDEX . . . . . . . . . . . . . . . . . . . . . . . . . . 141
CHAPTER 1
INTRODUCTION TO CENTRE MANIFOLD THEORY
1.1. Introduction
In this chapter we state the main results of centre
manifold theory for finite dimensional systems and give some

simple examples to illustrate their application.
1.2. Motivation
To motivate the study of centre manifolds we first

look at a simple example. Consider the system
x = ax3, y = -y + y2, (1.2.1)
where a is a constant. Since the equations are uncoupled

we can easily show that the zero solution of (1.2.1) is
asymptotically stable if and only if a < 0. Suppose now that
x = ax 3 + x2y
(1.2.2)
y = -y + y + xy - x3.
Since the equations are coupled we cannot immediately decide
if the zero solution of (1.2.2) is asymptotically stable, but
we might suspect that it is if a < 0. The key to understand-

ing the relation of equation (1.2.2) to equation (1.2.1) is
t
2 1. INTRODUCTION TO CENTRE MANIFOLD THEORY
an abstraction of the idea of uncoupled equations.
A curve, say y = h(x), defined for lx1 small, is
said to be an invariant manifold for the system of differen-
tial equations
x = f(x,Y), Y = g(x,Y), (1.2.3)
if the solution (x(t),y(t)) of (1.2.3) through (x0,h(x0))

lies on the curve y = h(x) for small t, i.e., y(t) =
h(x(t)). Thus, for equation (1.2.1), y - 0 is an invariant

manifold. Note that in deciding upon the stability of the
zero solution of (1.2.1), the only important equation is
x = ax3, that is, we only need study a first order equation

on a particular invariant manifold.
The theory that we develop tells us that equation

(1.2.1') has an invariant manifold y = h(x), jxj small, with
h(x) = 0(x2) as x + 0. Furthermore, the asymptotic stability
of the zero solution of (1.2.2) can be proved by studying a
first order equation. This equation is given by
u = au + u2h(u) = au3 + 0(u4), (1.2.4)
and we see that the zero solution of (1.2.4) is asymptotically
stable if a < 0 and unstable if a > 0. This tells us that

the zero solution of (1.2.2) is asymptotically stable if
a < 0 and unstable if a > 0 as we expected.
We are also able to use this method to obtain estimates

for the rate of decay of solutions of (1.2.2) in the case
a < 0. For example, if (x(t),y(t)) is a solution of (1.2.2)
with (x(0),y(0)) small, we prove that there is a solution
u(t) of (1.2.4) such that x(t) u(t)(1+o(1)),
y(t) - h(u(t))(1+o(1)) as t -
1.3. Centre Manifolds 3
1.3. Centre Manifolds
We first recall the definition of an invariant mani-

fold for the equation
x - N(x) (1.3.1)
where x E1Rn. A set S d1Rn is said to be a local invari-
ant manifold for (1.3.1) if for x0 E S, the solution x(t)
of (1.3.1) with x(O) = x0 is in S for Iti < T where
T > 0. If we can always choose T = then we say that S
is an invariant manifold.
Consider the system
Ax + f(x,y)
(1.3.2)
By + g(x,y)
where x E1Rn, y E1Rm and A and B are constant matrices
such that all the eigenvalues of A have zero real parts
while all the eigenvalues of B have negative real parts.

The functions f and g are C2 with f(0,0) - 0,
f'(0,0) = 0, g(0,0) = 0, g'(0,0) = 0. (Here, f' is the
Jacobian matrix of f.)
If f and g are identically zero then (1.3.2) has

two obvious invariant manifolds, namely x - 0 and y = 0.
The invariant manifold x = 0 is called the stable manifold,

since if we restrict initial data to x = 0, all solutions
tend to zero. The invariant manifold y = 0 is called the
centre manifold.
In general, if y - h(x) is an invariant manifold for
(1.3.2) and h is smooth, then it is called a centre manifold

if h(0) - 0, h'(0) - 0. We use the term centre manifold in
place of local centre manifold if the meaning is clear.
If f and g are identically zero, then all solutions
of (1.3.2) tend exponentially fast, as t to solutions
of
x - Ax. (1.3.3)
That is, the equation on the centre manifold determines the
asymptotic behavior of solutions of the full equation modulo
exponentially decaying terms. We now give the analogue of
these results when f and g are non-zero. These results
are proved in Chapter 2.
Theorem 1. There exists a centre manifold for (1.3.2),
y - h(x), lxj < 6, where h is C2.
The flow on the centre manifold is governed by the

n-dimensional system
u = Au + f(u,h(u)) (1.3.4)
which generalizes the corresponding problem (1.3.3) for the
linear case. The next theorem tells us that (1.3.4) contains

all the necessary information needed to determine the asymp-
totic behavior of small solutions of (1.3.2).
Theorem 2. (a) Suppose that the zero solution of (1.3.4) is

stable (asymptotically stable) (unstable). Then the zero solu-
tion of (1.3.2) is stable (asymptotically stable) (unstable).
(b) Suppose that the zero solution of (1.3.4) is
stable. Let (x(t),y(t)) be a solution of (1.3.2) with
(x(O),y(O)) sufficiently small. Then there exists a solution
u(t) of (1.3.4) such that as t i m,
x(t) - u(t) + O(e- Yt)

11.3.5)
yt)
y(t) - h(u(t)) + O(e
1.4. Examples 5
where y > 0 is a constant.
If we substitute y(t) = h(x(t)) into the second equa-

tion in (1.3.2) we obtain
h'(x)[Ax + f(x,h(x))] = Bh(x) + g(x,h(x)). (1.3.6)
Equation (1.3.6) together with the conditions h(0) = 0,
h'(0) = 0 is the system to be solved for the centre manifold.

This is impossible, in general, since it is equivalent to
solving (1.3.2). The next result, however, shows that, in prin-

ciple, the centre manifold can be approximated to any degree
of accuracy.
For functions $:]Rn +]Rm which are CI in a neigh-

borhood of the origin define
(MO) (X) _ O'(x)[Ax + f(x,h(x))] - B$(x) - g(x,h(x))

Note that by (1.3.6), (Mh)(x) = 0.
C1
Theorem 3. Let $ be a mapping of a neighborhood of the
origin in ]Rn into ]Rm with $(0) = 0 and $'(0) = 0.
Suppose that as x _ 0, (M$)(x) - O(jxjq) where q > 1. Then

as x - 0, jh(x) - $(x)l = O(Ixl q)
1.4. Examples
We now consider a few simple examples to illustrate the
use of the above results.
Example 1. Consider the system
x - xy + ax3 + by2x
Y - -Y + cx2 + dx2Y.
By Theorem 1, equation (1.4.1) has a centre manifold y = h(x).
To approximate h we set
(M0)(x) = m'(x)[xm(x) + ax3 + bxm2(x)l + m(x) - cx2 - dx2m(x)
If '(x) = 0(x2) then (MO) (x) = 4(x) - cx2 + 0(x4). Hence,
if m(x) = cx2, No) (x) = 0(x4), so by Theorem 3, h(x) =
cx2 + 0(x4). By Theorem 2, the equation which determines the
stability of the zero solution of (1.4.1) is
u = uh(u) + au3 + buh2(u) = (a+c)u3 + 0(u5).
Thus the zero solution of (1.4.1) is asymptotically stable if
a + c < 0 and unstable if a + c > 0. If a + c = 0 then
we have to obtain a better approximation to h.
Suppose that a + c = 0. Let 4(x) = cx2 + c(x) where
fi(x) = 0(x4). Then (M4)(x) = c(x) - cdx4 + 0(x6). Thus, if
m(x) = cx2 + cdx4 then (M4)(x) = 0(x6) so by Theorem 3,

2 4 6
h(x) = cx + cdx + 0(x ). The equation that governs the sta-
bility of the zero solution of (1.4.1) is
u = uh(u) + au3 + buh2(u) = (cd+bc2)u5 + 0(u7).
Hence, if a + c = 0, then the zero solution of (1.4.1) is
asymptotically stable if cd + bc2 < 0 and unstable if
cd + bc2 > 0. If cd + bc2 = 0 then we have to obtain a
better approximation to h (see Exercise 1).
Exercise 1. Suppose that a + c = cd + bc2 = 0 in Example 1.
Show that the equation which governs the stability of the
zero solution of (1.4.1) is u = -cd2u7 + 0(u9).
Exercise 2. Show that the zero solution of (1.2.2) 1% asymp-
totically stable if a < 0 and unstable if a n.

1.4. Examples 7
Exercise 3. Suppose that in equation (1.3.2), n = 1 so that

IyIq)
A = 0. Suppose also that f(x,y) = axp + O(IxIp+l +
where 2q > p + 1 and a is non-zero. Show that the zero
solution of (1.3.2) is asymptotically stable if a < 0 and
p is odd, and unstable otherwise.
Example 2. Consider the system
x = ex - x3 + xy
(1.4.2)
Y = -Y + Y2 - x2
where c is a real parameter. The object is to study small
solutions of (1.4.2) for small lei.
The linearized problem corresponding to (1.4.2) has
eigenvalues -1 and c. This means that the results given
in Section 3 do not apply directly. However, we can write
(1.4.2) in the equivalent form
3
x = ex - x + xy
y= -y+y 2
- x
2
(1.4.3)
When considered as an equation on ]R3 the ex term in
(1.4.3) is nonlinear. Thus the linearized problem correspond-
ing to (1.4.3) has eigenvalues -1,0,0. The theory given in

Section 3 now applies so that by Theorem 1, (1.4.3) has a two
dimensional centre manifold y = h(x,c), Ixl < 61, Iel < a2'
To find an approximation to h set
(M4)(x,c) - mx(x,e)[ex-x3+xm(x,e)] + m(x,e) + x2 - 02(x,e).
'then, if 4(x,c) - -x2, (M4)(x,e) = O(C(x,c)) where C is a
homogeneous cubic in x and c. By Theorem 3,

h(x,c) - -x2 + O(C(x,c)) . Note also that h(O,c) = 0 (see
Section 2.6). By Theorem 2 the equation which governs small
solutions of (1.4.3) is
u = cu - 2u3 + O(IuIC(u,e))
(1.4.4)
The zero solution (u,c) = (0,0) of (1.4.4) is stable so
the representation of solutions given by Theorem 2 applies

here. For -62 < e < 0 the solution u = 0 of the first
equation in (1.4.4) is asymptotically stable and so by Theorem
2 the zero solution of (1.4.2) is asymptotically stable.
For 0 < c < 62, solutions of the first equation in
(1.4.4) consist of two orbits connecting the origin to two
small fixed points. Hence, for 0 < c < 62 the stable mani-
fold of the origin for (1.4.2) forms a separatrix, the unstable
manifold consisting of two stable orbits connecting the origin

to the fixed points.
Exercise 4. Study the behavior of all small solutions of

w + + ew + w3 = 0 for small C.
Example 3. Consider the equations
y = -y + (y+c)z
(1.4.5)
ei = y- (Y+1) z
where c > 0 is small and 0 < c < 1. The above equations
arise from a model of the kinetics of enzyme reactions [33].

If e = 0, then (1.4.5) degenerates into one algebraic equa-
tion and one differential equation. Solving the algebraic
equation we obtain
1.4. Examples 9
z= (1.4.6)
Y Y
and substituting this into the first equation in (1.4.5) leads
to the equation
_
(1.4.7)
y 1
where A = 1 - c.
Using singular perturbation techniques, it was shown
in [331 that for c sufficiently small, under certain condi-
tions, solutions of (1.4.5) are close to solutions of the de-

generate system (1.4.6), (1.4.7). We shall show how centre
manifolds can be used to obtain a similar result.
Let t = CT. We denote differentiation with respect
to t by ' and differentiation with respect to T by '.
Equation (1.4.5) can be rewritten in the equivalent form
Y' Ef(Y,w)
w' _ -w + y2 - yw + ef(Y,w) (1.4.8)
where f(y,w) = -y + (y+c)(y-w) and w = y - z. By Theorem
1, (1.4.8) has a centre manifold w = h(y,c). To find an ap-
proximation to h set
(MO)(Y,E) = Emy(Y,E)f(Y,m) + 0(Y,E) - y2 + y (Y, e) - ef(Y, m) -
If 0 (Y,e) = y2 - Acy then (Mm)(Y,c) O(1Y13 + IcI3) so
that by Theorem 3,
h(Y,e) - y2 - aey + O(IY13 + 1el3).
By Theorem 2, the equation which determines the asymptotic

behavior of small solutions of (1.4.8) is
u' = ef(u,h(u,e))
or in terms of the original time scale
u = f(u,h(u,c)) _ -A(u-u2) + O(Ieul + lul3). (1.4.9)
Again, by Theorem 2, if a is sufficiently small and y(0),
z(0) are sufficiently small, then there is a solution u(t)
of (1.4.9) such that for some y > 0,
y(t) = u(t) + 0(e Yt/E)

(1.4.10)
z(t) = y(t) - h(y(t),c) + O(e yt/e).
Note that equation (1.4.7) is an approximation to the equation

on the centre manifold. Also, from (1.4.10), z(t) = y(t) -
y2(t), which shows that (1.4.6) is approximately true.
The above results are not satisfactory since we have to
assume that the initial data is small. In Chapter 2, we show
how we can deal with more general initial data. Here we
briefly indicate the procedure involved there. If y0 # -1,
then
(y,w,e) _ (y0,y0(l+y0) 1,0)
is a curve of equilibrium points for (1.4.8). Thus, we ex-
pect that there is an invariant manifold w = h(y,c) for
(1.4.8) defined for a small and 0 < y < m (m = 0(1)),
and with h(y,c) close to the curve
w = y2 (1+y)- 1. (1.4.11)
For initial data close to the curve given by (1.4.11), the
stability properties of (1.4.8) are the same as the stability
properties of the reduced equation
6 - f(u.h(u,e)).
1.5. Bifurcation Theory 11
1.5. Bifurcation Theory
Consider the system of ordinary differential equations
w = F (w, c)
F(O,c) s 0
where w EIRn+m and c is a p-dimensional parameter. We
say that c = 0 is a bifurcation point for (1.5.1) if the
qualitative nature of the flow changes at c = 0; that is, if
in any neighborhood of c = 0 there exist points and

c1
e2 such that the local phase portraits of (1.5.1) for
E = C1 and c = c2 are not topologically equivalent.
Suppose that the linearization of (1.5.1) about w = 0
is
w = C(e)w. (1.S.2)
If the eigenvalues of C(0) all have non-zero real parts

then, for small 10, small solutions of (1.5.1) behave like
solutions of (1.5.2) so that c = 0 is not a bifurcation
point. Thus, from the point of view of local bifurcation
theory the only interesting situation is when C(0) has eigen-
values with zero real parts.
Suppose that C(0) has n eigenvalues with zero real
parts and m eigenvalues whose real parts are negative. We
are assuming that C(0) does not have any positive eigen-
values since we are interested in the bifurcation of stable

phenomena.
Because of our hypothesis about the eigenvalues of
(:(0) we can rewrite (1.5.1) as

X = Ax + f(x,y,c)
By + g(x,y,c) (1.5.3)
E= 0
where x EIRn, y EIRm, A is an n x n matrix whose eigen-
values all have zero real parts, B is an m x m matrix

whose eigenvalues all have negative real parts, and f and
g vanish together with each of their derivatives at
(x,y,E) _ (0,0,0).
By Theorem 1, (1.5.3) has a centre manifold y = h(x.E),
1xI < 61, IEI < d2. By Theorem 2 the behavior of small solu-
tions of (1.5.3) is governed by the equation
u - Au + f(u,h(u,c),c)
(1.5.4)
E = 0.
In applications n is frequently 1 or 2 so this is a
very useful reduction. The reduction to a lower dimensional
problem is analogous to the use of the Liapunov-Schmidt pro-

cedure in the analysis of static problems.
Our use of centre manifold theory in bifurcation prob-

lems follows that of Ruelle and Takens [57] and of Marsden
and McCracken [51]. For the relationship between centre
manifold theory and other perturbation techniques such as
amplitude expansions see [14].
We emphasize that the above analysis is local. In
general, given a parameter dependent differential equation it
is difficult to classify all the possible phase portraits.

For an example of how complicated such an analysis can be, see
[661 where a model of the dynamic behavior of a continuous
stirred tank reactor is studied. The model con%i%tt of is

1.6. Comments on the Literature 13
parameter dependent second order system of ordinary differ-
ential equations. The authors show that there are 3S pos-
sible phase portraits!
1.6. Comments on the Literature
Theorems 1-3 are the simplest such results in centre
manifold theory and we briefly mention some of the possible
generalizations.
(1) The assumption that the eigenvalues of the lin-

earized problem all have non-positive real parts is not
necessary.
(2) The equations need not be autonomous.

(3) In certain circumstances we can replace 'equilibrium
point' by 'invariant set'.
(4) Similar results can be obtained for certain classes
of infinite-dimensional evolution equations, such as partial
differential equations.
There is a vast literature on invariant manifold theory
[1,8,22,23,27,28,30,32,34,35,42,44,45,48,51]. For applications
of invariant manifold theory to bifurcation theory see [1,14,
17,18,19,24,31,34,36,37,38,47,48,49,51,57,65). For a simple
discussion of stable and unstable manifolds see (22, Chapter

13) or (27, Chapter 3].
In Chapter 2 we prove Theorems 1-3. Our proofs of

Theorems 1 and 2 are modeled on Kelly (44,45). Theorem 3 is
a special case of a result of Henry [34] and our proof follows

his. The method of approximating the centre manifold in
Theorem 3 was essentially used by Hausrath (32) in his work on
stability in critical cases for neutral functional differen-

tial equations. Throughout Chapter 2 we use methods that
generalize to infinite dimensional problems in an obvious way.
CHAPTER 2
PROOFS OF THEOREMS
2.1. Introduction
In this chapter we give proofs of the three main

theorems stated in Chapter 1. The proofs are essentially
applications of the contraction mapping principle. The pro-
cedure used for defining the mappings is rather involved, so
we first give a simple example to help clarify the technique.

The proofs that we give can easily be extended to the corres-
ponding infinite dimensional case; indeed essentially all we
have to do is to replace the norm 1.1 in finite dimensional
space by the norm 11.11 in a Banach space.
2.2. A Simple Example
We consider a simple example to illustrate the method
that we use to prove the existence of centre manifolds.
Consider the system
x1 = x2, x2 = 0, k = -Y + g(xl,x2), (2.2.1)
where g is smooth and g(xl,x2) = 0(x2+x2) as (xl,x2) +
(0,0). We prove that (2.2.1) has a local centre manifold.
1 e
2.2. A Simple Example is
Let *:IR2 IR be a C function with compact support

such that p(xl,x2) = 1 for (xl,x2) in a neighborhood of
the origin. Define G by G(xl,x2) = (xl,x2)g(xl,x2). We
prove that the system of equations
xl = x2, x2 = 0, Y = -y + G(xl,x2), (2.2.2)
has a centre manifold y = h(xl,x2), (xl,x2) E IR2. Since
G(xl,x2) = g(xl,x2) in a neighborhood of the origin, this
proves that y = h(xl,x2), x2 + x2 < 6 for some 6, is a
local centre manifold for (2.2.1).
The solution of the first two equations in (2.2.2) is
xl(t) = z1 + z2t, x2(t) = z2, where xi(0) = zi. If
y(t) = h(xl(t),x2(t)) is a solution of the third equation in
(2.2.2) then
ft-h(z = -h(zI+z2t,z2) + G(zl+z2t,z2). (2.2.3)
To determine a centre manifold for (2.2.2) we must single out

a special solution of (2.2.3). Since G(xl,x2) is small for
,ill xl and x2, solutions of the third equation in (2.2.2)
behave like solutions of the linearized equation y = -y.
l'he general solution of (2.2.2) therefore contains a term

t.
like e- As t + -, this component approaches the origin
perpendicular to the zl,z2 plane. Since the centre manifold
I,; tangent to the zl,z2 plane at the origin we must elim-
mate the a-t component, that is we must eliminate the com-

l-onent that approaches the origin along the stable manifold
.is t - To do this we solve (2.2.3) together with the

kondition
lim h(zl+z2t,z2)et - 0. (2.2.4)

t y -.
16 2. PROOFS OF THEOREMS
Integrating (2.2.3) between and 0 and using
(2.2.4) we obtain
0
h(zl,z2) = f- esG(zl+z2s,z2)ds.
1 m
By construction, y = h(zl,z2) is an invariant manifold for

(2.2.2). Using the fact that G has compact support and
that G(xl,x2) has a second order zero at the origin it
follows that h(zl,z2) has a second order zero at the origin;

that is, h is a centre manifold.
2.3. Existence of Centre Manifolds
In this section we prove that the system
x = Ax + f (x, y)
(2.3.1)
y = By + g(x,y)
has a centre manifold. As before x EIRn, y EIRm, the eigen-

values of A have zero real parts, the eigenvalues of B
have negative real parts and f and g are C2 functions

which vanish together with their derivatives at the origin.
Theorem 1. Equation (2.3.1) has a local centre manifold
y = h(x), lxi < 6, where h is C2.
Proof: As in the example given in the previous section, we
prove the existence of a centre manifold for a modified equa-
tion. Let p:IRn + [0,11 be a Cm function with ,y(x) = 1
when lxi < 1 and y,(x) = 0 when lxi > 2. For c > 0
define F and G by
F(x,y) = f(x*(E),y), G(x,y) - g(xP(E),y)
The reason that the cut-off function p is only a function of

2.3. Existence of Centre Manifolds 17
x is that the proof of the existence of a centre manifold

generalizes in an obvious way to infinite dimensional prob-
lems.
We prove that the system
x = Ax + F(x,y)
(2.3.2)
By + G(x,y)
has a centre manifold y = h(x) , x EIRn, for small enough c.
Since F and G agree with f and g in a neighborhood of
the origin, this proves the existence of a local centre mani-
fold for (2.3.1).
For p > 0 and pl > 0 let X be the set of Lipschitz

functions h:IRn +]Rm with Lipschitz constant pl, Ih(x)I < p
for xEIRn and h(0) = 0. With the supremum norm 11-11, X

is a complete space.
For h E X and x0 EIRn, let x(t,x0,h) be the solu-
tion of
x = Ax + F(x,h(x)), x(O,xO,h) = x0. (2.3.3)
The bounds on F and h ensure that the solution of (2.3.3)
exists for all t. We now define a new function Th by
r 0
(Th)(x0) = 1- e- Bs G(x(s,x0,h),h(x(s,x0,h)))ds. (2.3.4)
If h is a fixed point of (2.3.4) then h is a centre mani-
fold for (2.3.2). We prove that for p,pl, and c small
enough, T is a contraction on X.
Using the definitions of F and G, there is a con-
tinuous function k(c) with k(O) = 0 such that

IF(x,y)I + IG(x,y)I ck(c),
IF(x,y) - F(x',y')I c k(c)[Ix-x'I + Iy-y'II, (2.3.5)
IG(x,y) - G(x',y')I c k(c)[Ix-x'I + Iy-y'I],
for all x, x' E]Rn and all y, y' E]Rm with IyI, Iy'I < c.
Since the eigenvalues of B all have negative real
parts, there exist positive constants S,C such that for
s < 0 and y E IRm,
le- Bsyl
< CesslyI. (2.3.6)
Since the eigenvalues of A all have zero real parts, for
each r > 0 there is a constant M(r) such that for x E IRn
and s E IR,
IeAsxl
< M(r)erlsllxl. (2.3.7)
Note that in general, M(r) + m as r + 0.
If p < c, then we can use (2.3.5) to estimate terms
involving G(x(s,x0,h),h(x(s,x0,h))) and similar terms. We
shall suppose that p < c from now on.
If x0 E]Rn, then using (2.3.6) and the estimates on
G and h, we have from (2.3.4) that
ITh(x0)I < CS lck(c). (2.3.8)
Now let x0, x1 E]Rn. Using (2.3.7) and the estimates

on F and h, we have from (2.3.3) that for r > 0 and
t < 0,
M(r)e-rtlx0-x1I
Ix(t,x0,h) - x(t,xl,h)I <
0 r(s-t)Ix(s,x0,h)
+ (l+pl)M(r)k(c) I e - x(s,xl,h)Ids.
it
By Gronwall's inequality, for t < 0,

2.4. Reduction Principle 19
M(r)Ixl-x0Ie-Yt,
Ix(t,x0,h) - x(t,x1,h)I < (2.3.9)
where y = r + (l+pl)M(r)k(c). Using (2.3.9) and the bounds
on G and h, we obtain from (2.3.4)
1Jx0-x1I
ITh(x0) - Th(xl)I < C(M(r) +pl)k(c)(s-y) (2.3.10)
if c and r are small enough so that 0 > y.
Similarly, if hl, h2 E X and x0 E]Rn, we obtain
+(1+pi)M(r)k(c)r-1(S-y) 1]
ITh1(x0)-Th2(x0)I < Ck(c)(S
(2.3.11)
IIh1-h2II
By a suitable choice of p,p1,c and r, we see from
(2.3.8), (2.3.10) and (2.3.11) that T is a contraction on
X. This proves the existence of a Lipschitz centre manifold
for (2.3.2). To prove that h is C1 we show that T is a
contraction on a subset of X consisting of Lipschitz dif-
ferentiable functions. The details are similar to the proof

given above so we omit the details. To prove that h is C2
we imitate the proof of Theorem 4.2 on page 333 of (22].
2.4. Reduction Principle
The flow on the centre manifold is governed by the
n-dimensional system
u = Au + f(u,h(u)). (2.4.1)
In this section we prove a theorem which enables us to relate
the asymptotic behavior of small solutions of (2.3.1) to solu-

tions of (2.4.1).
We first prove a lemma which describes the stability

properties of the centre manifold.
Lemma 1. Let (x(t),y(t)) be a solution of (2.3.2) with
I(x(0),y(0))I sufficiently small. Then there exist positive

C1 and u such that
Iy(t) - h(x(t))I < C1e-utly(0) - h(x(0))I
for all t > 0.
Proof: Let (x(t),y(t)) be a solution of (2.3.2) with
(x(0),y(0)) sufficiently small. Let z(t) - y(t) - h(x(t)),
then by an easy computation
i = Bz + N(x,z) (2.4.2)
where
N(x,z) = h'(x)[F(x,h(x)) - F(x,z+h(x))] + G(x,z + h(x))
- G(x,h(x)).
Using the definitions of F and G and the bounds on h,
there is a continuous function 6(c) with 6(0) = 0 such
that IN(x,z)I < 6(c)Izi if IzI < c. Using (2.3.6) we ob-
tain, from (2.4.2),

t
e-'(t-s)Iz(s)Ids
Iz(t)I < Ce-stlz(0)I + C6(c)I0
and the result follows from Gronwall's inequality.
Before giving the main result in this section we make
some remarks about the matrix A. Since the eigenvalues of
A all have zero real parts, by a change of basis we can put
A in the form A = A + A2 where A2 is nilpotent and

1
A t
le 1 xI - IxI. (2.4.3)
Since A2 is nilpotent, we can choose the basis such that

IA2xI < (S/4)IxI, (2.4.4)
where S is defined by (2.3.6).

We assume for the rest of this section that a basis
has been chosen so that (2.4.3) and (2.4.4) hold.

stable (asymptotically stable) (unstable). Then the zero
solution of (2.3.1) is stable (asymptotically stable) (un-

stable).

(x(O),y(O)) sufficiently small. Then there exists a solu-
tion u(t) of (2.4.1) such that as t m,
x(t) - u(t) + 0(e-Yt)

0(e-Yt)
(2.4.5)
y(t) = h(u(t)) +
where y > 0 is a constant depending only on B.
Proof: If the zero solution of (2.4.1) is unstable then by

invariance, the zero solution of (2.3.1) is unstable. From
now on we assume that the zero solution of (2.3.1) is stable.

We prove that (2.4.5) holds where (x(t),y(t)) is a solution
of (2.3.2) with I(x(0),y(0))I sufficiently small. Since F
and G are equal to f and g in a neighborhood of the
origin this proves Theorem 2. We divide the proof into two

steps.
I. Let u0 E]Rn and z0 E]Rm with I(u0,z0)I suf-
ficiently small. Let u(t) be the solution of (2.4.1) with

u(0) - u0. We prove that there exists a solution (x(t),y(t))
of (2.3.2) with y(0) - h(x(0)) - z0 and x(t) - u(t),

y(t) - h(u(t)) exponentially small as t + m
II. By Step I we can define a mapping S from a
neighborhood of the origin in IRn+m into IRn+m by
S(u0,z0) _ (x0)z0) where x0 = x(0). For I(x0,z0)l suf-
ficiently small, we prove that (x0)20) is in the range of S.
I. Let (x(t),y(t)) be a solution of (2.3.2) and
u(t) a solution of (2.4.1). Note that if u(O) is suffici-
ently small,
u - Au + F(u,h(u)) (2.4.6)
since solutions of (2.4.1) are stable. Let z(t) - y(t) -
h(x(t)), $(t) - x(t) - u(t). Then by an easy computation
t - Bz + N($+u,z) (2.4.7)
AO + R($,z) (2.4.8)
where N is defined in the proof of Lemma 1 and
R($,z) - F(u+$,z+h(u+$)) - F(u,h(u)).
We now formulate (2.4.7), (2.4.8) as a fixed point

problem. For a > 0, K > 0, let X be the set of continuous
functions $: [0,m) +IRn with I,(t)eatl < K for all t > 0.
If we define II$II - sup{14(t)eatl: t > 0}, then X is a com-

plete space. Let (u0,z0) be sufficiently small and let u(t)
be the solution of (2.4.6) with u(0) - u0. Given $ E X
let z(t) be the solution of (2.4.7) with z(0) - z0.
Define T$ by
A1(t-s)
(T$)(t) - -
e [A20(s) + R(0(s),z(s))1ds. (2.4.9)
t
We solve (2.4.9) by means of the contraction mapping prin-
ciple. If $ is a fixed point of T, then retracing our

steps we find that x(t) = u(t) + 4(t), y(t) - z(t) + h(x(t))
is a solution of (2.3.2). We can take a to be as close to
S as we please at the cost of increasing K and shrinking
the neighborhood on which the result is valid. For simplicity
however, we take K - 1 and 2a - S where 0 is defined by
(2.3.6).
Using the bounds on F,G,h and the fact that

N($,0) - 0, there is a continuous function k(c) with
]Rn
k(0) - 0 such that if 01,02 E and zl,z2 EIRm with
Izil < c, then
IN($1,z1) - N($2,z2)I k(c)[Iz11101-021 + Izl-z21]

(2.4.10)
IR($1)z1) - R($2,z2)I k(c)[1z1-z2l + 141-$21]
From (2.4.7),
t
<
CIzOIe-Ot
+ Ck(e)j e-O(t-s)Iz(s)Ids
Iz(t)I
0
where we have used (2.3.6) and (2.4.10). By Gronwall's in-
equality
sl t
Iz(t)I CIz0Ie (2.4.11)
where S1 = 0 - Ck(e). From (2.4.9), if c is sufficiently
small,
e-at + CIzOIe- sls)ds

< k(e)1W(e-as + at
IT,(t)I < e-
it
where we have used (2.4.3), (2.4.4), (2.4.10) and (2.4.11).
Hence T maps X into X.
Now let 01,02 E X and let zl,z2 be the correspond-
ing solutions of (2.4.7) with zi(0) - z0. We first estimate

w(t) - zl(t) - z2(t). From (2.4.7) and (2.4.10),
t t s) [Izl(s)II01(s)-02(s)I
e
I w ( t )I < C k ( c ) I + Iw(s)lids.
0
Using (2.4.11),
t
e-S(t-s)Iw(s)Ids
Iw(t) I < Clk(c)IIO1-02IIe-St + Ck(c)IO
where C1 is a constant, so that by Gronwall's inequality
Iw(t) I = Iz1(t)-z2(t)I < C1k(c)II01-0211 e Slt. (2.4.12)
Using (2.4.4) and (2.4.12), for c sufficiently small,
ITO1(t)-T02(t)I < 21101-0211 + k(c)fm(I01(s) - 02(s)I
+ Izl(s) - z2(s)I)ds < a101-0211
where a < 1.
The above analysis proves that for each (u0,z0) suf-
ficiently small, T has a unique fixed point. If U is a
neighborhood of the origin in ]Rn+m then it is easy to re-
peat the above analysis to show that T: X x U i X is a con-
tinuous uniform contraction. This proves that the fixed point
depends continuously on u0 and z0.
II. Define S by S(u0,z0) (x0,20) where
x0 = u0 + 0(0). Since 0 depends continuously on u0 and
z0, S is continuous. We prove that S is one-to-one, so
that by the Invariance of Domain Theorem (see [11] or [601)
S is an open mapping. Since S(0,0) = 0, this proves that
the range of S is a full neighborhood of the origin in ]Rn+m
Proving that S is one-to-one is clearly equivalent
to proving that if u0 + 00(0) = ul + 01(0) then u0 = u1
and 00(0) = 01(0). If u0 + 00(0) - ul + 41(0) then the
initial values for x and y are the same, so that by

2.5. Approximation of the Centre Manifold 2S
uniqueness of solution of (2.3.2), u0(t) + 00(t) = ul(t) +
01(t) for all t > 0, where ui(t) is the solution of

(2.4.6) with ui(0) = ui. Hence, for t > 0,
u0(t) - ul(t) = O1 (t) - 00(t). (2.4.13)
Since the real parts of the eigenvalues of A are all zero,

limlul(t) - u0(t)lect = m for any c > 0 unless u1(0) _
tam
at
u0(0). Also, 10i(t)l < e- for all t > 0. It now follows
from (2.4.13) that S is one-to-one and this completes the
proof of the theorem.
2.5. Approximation of the Centre Manifold
For functions 0:IRn +IRm which are C1 in a neigh-
borhood of the origin define
(MO) (x) = O'(x)[Ax + f(x,0(x))] - BO (x) - g(x,0(x))
Theorem 3. Suppose that 0(0) = 0, 0'(0) = 0 and that
(Mo)(x) = 0(1x1q) as x + 0 where q > 1. Then as x + 0,
lh(x) - O(x)l = 0(Ixlq)
Proof: Let 6: Rn Rm be a continuously differentiable

function with compact support such that 0(x) = O(x) for
lxj small. Set
N(x) = 6' (x) [Ax + F(x,6(x))] - BO(x) - G(x,6(x)), (2.5.1)
where F and G are defined in Theorem 1. Note that

N(x) = O(1xjq) as x + 0.
In Theorem 1, we proved that h was the fixed point of

a contraction mapping T: X + X. Define a mapping S by
Sz - T(z+e) - 0; the domain of S being a closed subset

Y c X. Since T is a contraction mapping on X, S is a
contraction mapping on Y. For K > 0 let
Y - { z E X : l z ( x ) I < K j x j q for all x E I R n}.
If we can find a K such that S maps Y into Y then we
will have proved the theorem.
We first find an alternative formulation of the map S.
For z E Y let x(t,x0) be the solution of
ii = Ax + F(x,z(x) + 6(x)), x(O,xO) - x(0). (2.5.2)
From (2.3.4)
0
(T(z+6))(x0) - - e-BsG(x(s,x0),z(x(s,x0)) + O(x(s,x0)))ds.
f
r 0
d[e-BsO(x(s,x0))]ds
-6(x0) -I-
1 1
u-s
0
_ (e-Bs[BO(x(s,x0)) O(x(s,x0))]ds.
- as
Writing x for x(s,x0) etc., from (2.5.1) and (2.5.2)
B6 (x) - ds (x) = B6 (x) - 6' (x) [Ax + F (x, z (x) + 6 (x)) ]
_ -N(x) - G(x,6(x)) + 6' (x)[F(x,6)
- F(x,z(x) + 6(x))].
Using Sz - T(z+6) - 6 and the above calculations
0
(Sz)(x0) = r- e-BsQ(x(s,x0),
z(x(s,x0)))ds (2.5.3)
where x(s,x0) is the solution of (2.5.2) and
Q(x,z) - G(x,6+z) - G(x,6) - N(x) + 6'(x)[F(x,6)

(2.5.4)
F(x,6+z)].
2.5. Approximation of the Centre Manifold 27
We now show that S maps Y into Y for some K > 0.
By choosing 6 suitably, we may assume that I0(x)I < E
for all x E]Rn. Since N(x) = O(Ixlq) as x + 0,
IN(x)I < Cllxlq, x E]Rn (2. S. S)
where C1 is a constant. Now
IQ(x,z)I_
IQ(x,o)I + IQ(x,z) - Q(x,o)I

(2.5.6)
= IN(x)I + IQ(x,z) - Q(x,0)I.
We can estimate IQ(x,z) - Q(x,0)I in terms of the Lipschitz
constants of F and G. Using (2.3.5), there is a continuous
function k(c) with k(0) - 0, such that
IQ(x,z) - Q(x,0)I < k(c)IzI (2.5.7)
for IzI < c. Using (2.S.S), (2.5.6), (2.5.7), for z E Y
and x E fn, we have that
IQ(x,z)I < C11xIq + k(c)Iz(x)I

(2. S.8)
< (C1 + Kk(c)) IxIq.
Using the same calculations as in the proof of Theorem 1, if
x(t,x0) is the solution of (2.5.2), then for each r > 0,
there is a constant M(r) such that
Ix(t,x0)I < M(r)Ix0Ie Yt, t < 0 (2.5.9)
where y - r + 2M(r)k(c).
Using (2.3.6), (2.5.8) and (2.5.9), if z E Y,
C(C1+Kk(c))(M(r))q(S-gy)-llxOIq
-
C2Ix0Iq
I($z)(x0)I <
provided c and r are small enough so that 0 - qy > 0.

By choosing K large enough and c small enough, we have
that C2 < K and this completes the proof of the theorem.
2.6. Properties of Centre Manifolds
(1) In general (2.3.1) does not have a unique centre
manifold. For example, the system x - -x3, y = -y, has the

two parameter family of centre manifolds y - h(x,cl,c2)
where
c
1
exp(
1
x
-2
), x > 0
h(x,cl,c2) = 0 x - 0
1 - 2
c2 exp(- x ), x < 0.
However, if h and h1 are two centre manifolds for
(2.3.1), then by Theorem 3, h(x) - h1(x) - O(jxjq) as x + 0

for all q > 1.
(2) If f and g are Ck, (k > 2), then h is Ck
(441. If f and g are analytic, then in general (2.3.1)
does not have an analytic centre manifold, for example, it
is easy to show that the system
x - -x3, y = -y + x2 (2.6.1)
does not have an analytic centre manifold (see exercise (1)).
(3) Centre manifolds need not be unique but there are
some points which must always be on any centre manifold. For
example, suppose that (x0,y0) is a small equilibrium point
of (2.3.1) and let y - h(x) be any centre manifold for
(2.3.1). Then by Lemma 1 we must have y0 - h(x0). Similarly,
if r is a small periodic orbit of (2.3.1), then r must lie

on all centre manifolds.
2.6. Properties of Centre Manifolds 29
(4) Suppose that (x(t),y(t)) is a solution of (2.3.1)

which remains in a neighborhood of the origin for all t > 0.
An examination of the proof of Theorem 2, shows that there is
a solution u(t) of (2.4.1) such that the representation
(2.4.5) holds.
(5) In many problems the initial data is not arbit-

rary, for example, some of the components might always be
nonnegative. Suppose S c7Rn+m with 0 E S and that (2.3.1)
defines a local dynamical system on S. It is easy to check,
that with the obvious modifications, Theorem 2 is valid when

(2.3.1) is studied on S.
Exercise 1. Consider
X -x 3 y -y + x2. (2.6.1)
Suppose that (2.6.1) has a centre manifold y - h(x), where
h is analytic at x = 0. Then
h(x) - E anxn
n-2
for small x. Show that a2n+1 ' 0 for all n and that
an+2 = nan for n - 2,4,..., with a2 = 1. Deduce that
(2.6.1) does not have an analytic centre manifold.
Exercise 2 (Modification of an example due to S. J. van
Strien [63]). If f and g are Cm functions, then for
each r, (2.3.1) has a Cr centre manifold. However, the
size of the neighborhood on which the centre manifold is de-
fined depends on r. The following example shows that in
general (2.3.1) does not have a Cm centre manifold, even if

f and g are analytic.
Consider
x -ex - x3, Y = -Y + x2, 0. (2.6.2)
Suppose that (2.6.2) has a C' centre manifold y = h(x,c)
for lxi < 6, jej < 6. Choose n > 6-l. Then since
h(x,(2n)-1) is Cw in x, there exist constants
al,a2.... ,a2n such that
2n
1) -
h(x,(2n) alxi + O(x2n+1)
i=1
for lxi small enough. Show that ai - 0 for odd i and
that if n > 1,
(2i)(2n)-1)a2i
(1 - = (2i-2)a2i-2' i - 2,...,n
(2.6.3)
a2 0.
Obtain a contradiction from (2.6.3) and deduce that (2.6.2)

does not have a C centre manifold.
Exercise 3. Suppose that the nonlinearities in (2.3.1) are
odd, that is f(x,y) = -f(-x,-y), g(x,y) - -g(-x,-y). Prove
that (2.3.1) has a centre manifold y = h(x) with h(x) -
-h(-x). [The example -x3, y = -y, shows that if h is
any centre manifold for (2.3.1) then h(x) -h(-x) in
general.]
2.7. Global Invariant Manifolds for Singular Perturbation

Problems
To motivate the results in this section we reconsider
Example 3 in Chapter 1. In that example we applied centre
manifold theory to a system of the form

31
2.7. Global Invariant Manifolds
Y' - Ef(Y,w)
w' = -w + y2 - yw + Ef(Y,w) (2.7.1)
where f(0,0) = 0. Because of the local nature of our results

on centre manifolds, we only obtained a result concerning
small initial data. Let v - -w(l+y) + y2, s S(T), where

s'(T) = 1 + y(T); then we obtain a system of the form
Y' = Egl(Y,v)
v' = -v + Eg2(Y,v) (2.7.2)
E, -0
where gi(0,0) - 0, i = 1,2. Note that if y j -1, then

(y,0,0) is always an equilibrium point for (2.7.2) so we ex-
pect that (2.7.2) has an invariant manifold v - h(y,c) de-
fined for -1 < y < m, say, and c sufficiently small.
Theorem 4. Consider the system
x' - Ax + Ef (x,y, E)
Y' - By + Eg(x,Y,E) (2.7.3)
E' = 0
where x E]Rn, y E]Rm and A,B are as in Theorem 1. Suppose

also that f,g are C2 with f(0,0,0) - 0, g(0,0,0) - 0.
Let m > 0. Then there is a 6 > 0 such that (2.7.3) has an
invariant manifold y - h(x,E), jxj < m, Ic < 6, with
Ih(x,E)l < Cici, where C is a constant which depends on
m,A,B,f and g.
Proof: Let p:]Rn + [0,1) be a C_ function with p(x) - 1
if lxi < m and y(x) - 0 if lxi > m + 1. Define F and
C by
F(x,y,c) - ef(xip(x),y,e), G(x,y,c) = eg(xi(x),y,e).
We can then prove that the system
x' = Ax + F(x,y,c)
(2.7.4)
y' = By + G(x,y,c)
has an invariant manifold y = h(x,e), x E]Rn, for IcI suf-
ficiently small. The proof is essentially the same as that
given in the proof of Theorem 1 so we omit the details.
Remark. If x = (x1,x2,...9xn) then we can similarly prove
the existence of h(x,c) for mi < xi < mi.
The flow on the invariant manifold is given by the
equation
u' - Au + cf(u,h(u,c)). (2.7.5)
With the obvious modifications it is easy to show that the

stability of solutions of (2.7.3) is determined by equation
(2.7.5) and that the representation of solutions given in

(2.4.5) holds.
Finally, we state an approximation result.
Theorem 5. Let $:]Rn+l i]Rm satisfy $(0,0) = 0 and
J(M$)(x,e)1 < Cep for jxl < m where p is a positive
integer, C is a constant and
(M$)(x,e) - Dx$(x,e)[Ax + ef(x,$(x,e))] - B$(x,e)
- eg(x,$(x,e)).
Then, for 1xI < m,
jh(x,e) - $(x,e)l < C1ep
for some constant C1.

2.8. Centre Manifold Theorems for Maps 33
Theorem 5 is proved in exactly the same way as Theorem
3 so we omit the proof.
For further information on the application of centre

manifold theory to singular perturbation problems see Fenichel
[24] and Henry [34].
2.8. Centre Manifold Theorems for Maps
In this section we briefly indicate some results on

centre manifolds for maps. We first indicate how the study
of maps arises naturally in studying periodic solutions of
differential equations.
Consider the following equation in ]Rn
A = f(x,A) (2.8.1)
where f is smooth and A is a real parameter. Suppose that

for A = A0, (2.8.1) has a periodic solution y with period
1'. One way to study solutions of (2.8.1) near y for
A - A01 small is to consider the Poincare map P(A). To
define P(A) let y be some point on y, let U be a local
cross section of y through y and let U1 be an open

neighborhood of y in U. Then P(A): U1 i U is defined by
I'(a)(z) - x(s), where x(t) is the solution of (2.8.1) with
x(0) = z and s > 0 is the first time x(t) hits U.
(See [51] for the details).
If P(A) has a fixed point then (2.8.1) has a periodic

,,rhit with period close to T. If P(A) has a periodic
point of order n,(P(A)kz / z for 1 < k < n - 1 and
I'r'(a)n - z) then (2.8.1) has a periodic solution with period
,lose to nT. If P(a) preserves orientation and there is a

,losed curve which is invariant under P(a) then there exists
an invariant torus for (2.8.1).
If none of the eigenvalues of the linearized map
P'(A0) lie on the unit circle then it can be shown that
P(a) has essentially the same behavior as P(A0) for
Ix - small. Hence in this case, for IA - a0I small,

X01
solutions of (2.8.1) near y have the same behavior as when
A - a0. If some of the eigenvalues of P'(A0) lie on the
unit circle then there is the possibility of bifurcations
taking place. In this case centre manifold theory reduces
the dimension of the problem. As in ordinary differential
equations we only discuss the stable case, that is, none of
the eigenvalues of the linearized problem lie outside the

unit circle.
n+m + n+m
Let T: IR IR have the following form:
T(x.Y) - (Ax + f(x,y), By + g(x,y)) (2.8.2)
where x E]Rn, y E]Rm, A and B are square matrices such
that each eigenvalue of A has modulus 1 and each eigen-
value of B has modulus less than 1, f and g are C2
and f,g and their first order derivatives are zero at the
origin.
Theorem 6. There exists a centre manifold h:]Rn + ]Rm for
T. More precisely, for some c > 0 there exists a C2 func-
tion h: ]Rn + ]Rm with h(0) = 0, h'(0) - 0 such that
jxl < c and (xl,yl) - T(x,h(x)) implies yl - h(xl).
In order to determine h we have to solve the equation
(x1,h(xl)) - T(x,h(x))
By (2.8.2) this is equivalent to

2.8. Centre Manifold Theorems for Maps 35
h(Ax + f(x,h(x))) = Bh(x) + g(x,h(x)).
n
For functions $: IR i Dm define M$ by
(M$)(x) = $(Ax + f(x,h(x))) - B$(x) - g(x,h(x))
so that Mh - 0.
Theorem 7. Let $: IR
n
i
Dm be a C1 map with $(0) = 0,
01(0) = 0 and (M$)(x) = O(Ixlq) as x i 0 for some q > 1.
Then h(x) = $(x) + O(Ixlq) as x i 0.
We now study the difference equation
xr+l = Axr + f(xr''r)

(2.8.3)
yr+l = Byr +
As in the ordinary differential equation case, the asymptotic

behavior of small solutions of (2.8.3) is determined by the
flow on the centre manifold which is given by
ur+l = Aur + f(ur,h(ur)). (2.8.4)

stable (asymptotically stable) (unstable). Then the zero solu-
tion of (2.8.3) is stable (asymptotically stable) (unstable).

stable. Let (xr,yr) be a solution of (2.8.3) with
(x1,yl)
sufficiently small. Then there is a solution ur of (2.8.4)
such that Ixr - url < KBr and Iyr - h(ur)I < KBr for all
r where K and B are positive constants with B < 1.
The proof of Theorem 6 and the stability claim of

Theorem 8 can be found in [30,40,51]. The rest of the asser-
tions are proved in the same way as the ordinary differential
equations case.
Exercise 4. Show that the zero solution of
xr+l = -xr + taxi + xryr
_ 1 2
yr+1 2 yr - axr + xryr
is asymptotically stable if a > 0.

CHAPTER 3
EXAMPLES
3.1. Rate of Decay Estimates in Critical Cases
In this section we study the decay to zero of solutions

of the equation
i + i + f(r) = 0
where f is a smooth function with
f(r) = r3 + ar5 + 0(r7) as r + 0, (3.1.2)
where a is a constant. By using a suitable Liapunov func-

tion it is easy to show that the zero solution of (3.1.1) is
asymptotically stable. However, because f'(0) = 0, the rate
of decay cannot be determined by linearization.
In [10] the rate of decay of solutions was given using

techniques which were special to second order equations. We
show how centre manifolds can be used to obtain similar re-

sults.
We first put (3.1.1) into canonical form. Let
x = r + i, y = i, then
x -f(x-y)
(3.1.3)
y - -y - f(x-y).
37
38 3. EXAMPLES
By Theorem 1 of Chapter 2, (3.1.3) has a centre manifold
y = h(x). By Theorem 2 of Chapter 2, the equation which
determines the asymptotic behavior of small solutions of
(3.1.3) is
u - -f(u - h(u)). (3.1.4)
Using (3.1.2) and h(u) - 0(u2),
u = -u3 + 0(u4). (3.1.5)
Without loss of generality we can suppose that the solution
u(t) of (3.1.5) is positive for all t > 0. Using L'Hopital's

rule,
ju(t)s-3ds.
-1 - lim - lim t-1
t- u J 1
Hence, if w(t) is the solution of
w - -w3, w(0) - 1, (3.1.6)
then u(t) - w(t+o(t)). Since
t-1/2
w(t) - 1 + Ct-3/2 + 0(t-5/2) (3.1.7)
7
where C is a constant, we have that
u(t) - t-1/2 + o(t-1/2). (3.1.8)
To obtain further terms in the asymptotic expansion of
u(t), we need an approximation to h(u). To do this, set
(M$)(x) - -$'(x)f(x-$(x)) + $(x) + f(x-$(x)).
If $(x) = -x3 then (M$)(x) - 0(x5) so that by
Theorem 3 of Chapter 2, h(x) - -x3 + 0(x5). Substituting this
into (3.1.4) we obtain

3.2. Hopf Bifurcation 39
u = -u3 - (a+3)u5 + 0(u7) (3.1.9)
Choose T so that u(T) a 1. Dividing (3.1.9) by u3, inte-

grating over [T,t] and using (3.1.8), we obtain
rt
w-1(u(t)) - t + constant + (3+a) ft u (3.1.10)
T
where w is the solution of (3.1.6). Using (3.1.8) and

(3.1.9),
ft
ftu2(s)ds -
u s ds + ft 0(u4(s))ds
T
u(s) T
T
(3.1.11)
- -ln t-1/2 + constant + 0(1).
Substituting (3.1.11) into (3.1.10) and using (3.1.7),
u(t) t-1/2 t-3/2[(a+3)ln t + C]

+ o(t-3/2) (3.1.12)
4/7
where C is a constant.
If (x(t),y(t)) is a solution of (3.1.3), it follows
from Theorem 2 of Chapter 2 that either (x(t),y(t)) tends to
zero exponentially fast or
x(t) = tu(t), y(t) = Tu3(t)

where u(t) is given by (3.1.12).
3.2. Hopf Bifurcation
There is an extensive literature on Hopf Bifurcation
[1,17,19,20,31,34,39,47,48,51,57,59] so we give only an out-

line of the theory. Our treatment is based on [19].
Consider the one-parameter family of ordinary dif-
ferential equations on ]R 2
40 3. EXAMPLES
z = f(x,a), xE IR2
such that f(0,a) = 0 for all sufficiently small a. Assume

that the linearized equation about x = 0 has eigenvalues
y(a) ± iw(a) where y(O) = 0, w(0) = w0 + 0. We also assume
that the eigenvalues cross the imaginary axis with nonzero
speed so that y'(0) + 0. Since y'(0) + 0, by the implicit

function theorem we can assume without loss of generality
that y(a) = a. By means of a change of basis the differen-
tial equation takes the form
i = A(a)x + F(x,a), (3.2.1)
where
r a -w(a)
A(a) =
1 w (a) a
F(x,a) = O(lx12).
Under the above conditions, there are periodic solu-
tions of (3.2.1) bifurcating from the zero solution. More
precisely, for a small there exists a unique one parameter
family of small amplitude periodic solutions of (3.2.1) in
exactly one of the cases (i) a < 0, (ii) a = 0, (iii) a > 0.
However, further conditions on the nonlinear terms are re-

quired to determine the specific type of bifurcation.
Exercise 1. Use polar co-ordinates for
xl = axl - wx2 + Kxl(x2 + x2)
x2 = wxl + ax2 + Kx2(x2 + x2)
to show that case (i) applies if K > 0 and case (iii) ap-
plies if K < 0.
To find periodic solutions of (3.2.1) we make the sub-
stitution
x1 = Cr cos 0, x2 = er sin 0, a + ca, (3.2.2)
where c is a function of a. After substituting (3.2.2)
into (3.2.1) we obtain a system of the form
r = e[ar + r2C3(O,ac)I + e2r3C4(O,ac) + O(e3)

(3.2.3)
6 = m0 + O(e) .
We now look for periodic solutions of (3.2.3) with c - 0
and r near a constant r0. If C3 and C4 are independent
of 0 and the higher order terms are zero then the first
equation in (3.2.3) takes the form
c[ar + $r2] + e2Kr3. (3.2.4)
Periodic solutions are then the circles r = r0, where r0
is a zero of the right hand side of (3.2.4). We reduce the
first equation in (3.2.3) to the form (3.2.4) modulo higher
order terms by means of a certain transformation. It turns
out that the constant B is zero. Under the hypothesis K
is non-zero, it is straightforward to prove the existence of

periodic solutions by means of the implicit function theorem.
The specific type of bifurcation depends on the sign of K
so it is necessary to obtain a formula for K.
Let F(x,a) = [F1(xl,x2,a), F2(xl,x2,a)]T and let
F(xiIx2)a) = BZ(xl,x2,a) + B3(xl,x2,a) + 0(x4+x4) (3.2.5)
where Bi is a homogeneous polynomial of degree i in
(xl,x2). Substituting (3.2.2) into (3.2.1) and using (3.2.5)

we obtain (3.2.3) where for i - 3,4,
42 3. EXAMPLES
Ci(6,a) _ (cos 0)B1i _1(cos 6, sin 6,a)

(3.2.6)
+ (sin 0)Bi_1(cos 6, sin 6,a).
Lemma 1. There exists a coordinate change
r - r + cul(r,6,a,c) + c2u2(r,6,a,c)
which transforms (3.2.3) into the system
r = car + c2 r3K + O(c3)

(3.2.7)
6 - m0 + 0(c)
where the constant K is given by

2n
K - (1/2n)f [C4(6.0) - w01C3(6.0)D3(6.0)]dO (3.2.8)
0
where C3 and C4 are given by (3.2.6) and
D3(6,0) - (cos 6)BZ(cos 6, sin 6,0) (sin 6)B2(cos 6, sin 6,0).
The coordinate change is constructed via averaging. We
refer to [19] for a proof of the lemma.
If K = 0 then we must make further coordinate changes.

We assume that K + 0 from now on.
Recall that we are looking for periodic solutions of

(3.2.7) with c + 0 and r near a constant r0. This sug-
gests that we set a = -sgn(K)c and r0 = IKJ -1/2. The next

result gives the existence of periodic solutions of
r - c2[-sgn(K)r + r3K] + 0(c3)

(3.2.9)
e'W0+0(c)
with r - r0 small.
Lemma 2. Equation (3.2.9) has a unique periodic solution for

c small and r in a compact region either for c > 0 (when
K< 0) or c < 0 (when K > 0) . Also
r = r0 + 0(c), 6(t,c) - mot + 0(c)
and the period of the solution i(c) is given by
t(c) - (21r/wo) + 0(c).
The periodic solution is stable if K < 0 and unstable if
K > 0.
Lemma 2 is proved by a simple application of the impli-
cit function theorem. We again refer to [19] for a proof.
Lemma 2 also proves the existence of a one parameter
family of periodic solutions of (3.2.1). We cannot immedi-

ately assert that this family is unique however, since we may
have lost some periodic solutions by the choice of scaling,

i.e. by scaling a + ac and by choosing c - -sgn(K)a.
In order to justify the scaling suppose that

x1 = R cos 6, x2 = R sin 6 is a periodic solution of (3.2.1)
bifurcating from x1 - x2 - 0. Then R satisfies
R = aR + 0(R2).
When R attains its maximum, = 0 so that a - 0(R). This
justifies the scaling a + ac. A similar argument applied to
periodic solutions of (3.2.7) justifies the choice of c.
Theorem. Suppose that the constant K defined by (3.2.8) is
nonzero. Then (3.2.1) has a unique periodic solution bifur-
cating from the origin, either for a > 0 (when K < 0) or
a < 0 (when K > 0). If x - R cos 6, y = R sin 6, then the
periodic solution has the form

44 3. EXAMPLES
R(t,a) _
JaK-lll/2 + O(JaI)
9(t,a) - w0t + 0(Jajl/2)
with period t(a) _ (2n/w0) + O(Iaj 1/2). The periodic solu-
tion is stable if K < 0 and unstable if K > 0.
Finally, we note that since the value of K depends
only on the nonlinear terms evaluated at a - 0, when apply-
ing the above theorem to (3.2.1) we only need assume that the
eigenvalues of A(a) cross the imaginary axis with non-zero
speed and that
A(0) =
with m0 non-zero.
3.3. Hopf Bifurcation in a Singular Perturbation Problem
In this section we study a singular perturbation prob-
lem which arises from a mathematical model of the immune re-
sponse to antigen [521. The equations are
ez [x3 + (a - )x + b -]
a m 6(1-x) - a - ylab (3.3.1)
ylab + y2b
where e,6,yl,y2 are positive parameters. In the above model

a and b represent certain concentrations so they must be
non-negative. Also, x measures the stimulation of the system
and it is scaled so that jxj < 1. The stimulation is assumed
to take place on a much faster time scale than the response
so that c is very small.
3.3. Hopf Bifurcation in a Singular Perturbation Problem 45
The above problem was studied in [52J and we briefly

outline the method used by Merrill to prove the existence of
periodic solutions of (3.3.1). Putting c = 0 in the first
equation in (3.3.1) we obtain
x3 + (a - 4)x + b - m 0. (3.3.2)
Solving (3.3.2) for x as a function of a and b we ob-
tain x = F(a,b) and substituting this into the second equa-

tion in (3.3.1) we obtain
6(1-F(a,b)) - a - ylab
2 (3.3.3)
-Ylab + b.
Y2
Using 6 as the parameter, it was shown that relative to a

certain equilibrium point (a0,b0) a Hopf bifurcation takes
place in (3.3.3). By appealing to a result in singular per-
turbation theory, it was concluded that for c sufficiently
small, (3.3.1) also has a periodic solution.
We use the theory given in Chapter 2 to obtain a simi-

lar result.
Let (x0,a0,b0) be a fixed point of (3.3.1). If
b0 + 0 then and x0,b0 satisfy

a0 = Y2/Y1
3
Y
1
x0 + (Y1 - 2)x0 + b0 - = 0,
Y
6(1-x 0) - Y2 - Y2b0 = 0.
2 1
Recall also that for the biological problem, we must have

b0 > 0 and jx0I < 1. We assume for the moment that x0 and
b0 satisfy (3.3.4) and these restrictions. The reality of
these solutions are considered later. We let a0 = Y2/Yl
for the rest of this section.

46 3. EXAMPLES
Let y = a - a0, z - b - b0, w - -i(x-x0) - x0y - z
where ip = 3x2 + a0 - . Then assuming ip is non-zero,
Ew = g(w,Y,z,E)
Y - f2(w,Y,z,E) (3.3.5)
i - f3(w,Y,z,E)
where
g(w,Y,z,E) - fl(w,Y,z,E) - Ex0f2(w,y,z,E) - Ef3(w,Y,z,E)
fl(w,Y,z,E) - -ipw + N(w+x0Y+z,Y)
i-1
f2(w,Y,z,E) - (2 -1x0 - 1 - ylb0)y + (. - y2)z
+
2
lw - ylyz
f3(w,Y,z,E) e -ylb0y - ylyz
-p-263
N(e,Y) - + 3i-1x062 - ye.
In order to apply centre manifold theory we change the
time scale by setting t = Es. We denote differentiation with

respect to s by ' and differentiation with respect to t
by Equation (3.3.5) can now be written in the form
w' - g(w,Y,z,E)
y' = Ef2(w,y,z,E)
(3.3.6)
z' = Ef3(w,y,z,E)
E' = 0.
Suppose that * > 0. Then the linearized system corresponding
to (3.3.6) has one negative eigenvalue and three zero eigen-

values. By Theorem 1 of Chapter 2, (3.3.6) has a centre mani-
fold w = h(y,z,c). By Theorem 2 of Chapter 2, the local be-
havior of solutions of (3.3.6) is determined by the equation

y' = Ef2(h(y,z,E),y,z,E)
(3.3.7)
z' = Ef3(h(y,z,E),y,z,E)
or in terms of the original time scale
y - f2(h(y,z,E),y,z,E)
(3.3.8)
i = f3(h(y,z,E),y,z,E).
We now apply the theory given in the previous section to show

that (3.3.8) has a periodic solution bifurcating from the
origin for certain values of the parameters.
The linearization of the vector field in (3.3.8) about
y - z = 0 irs given by
*-1
>y 1x0 - 1 - Ylb0
-
Y2
2 2
J(E) -
0
-Y1b0
If (3.3.8) is to have a Hopf bifurcation then we must have
trace(J(E)) - 0 and 2 *-I - Y2 > 0. From the previous analy-

sis, we must also have that x0,b0 are solutions of (3.3.4)
with 1x01 < 1, b0 > 0 and ip > 0. We do not attempt to ob-
tain the general conditions under which the above conditions
are satisfied, we only work out a special case.
Lemma. Let Y1 < 2Y2. Then for each E > 0, there exists
6(E), x0(E), b0(E) such that 0 < 2x0(E) < 1, b0(E) > 0,
> 0, 6(E)iy-1 - 2Y2 > 0, trace (J(E)) = 0 and (3.3.4) is
satisfied.
Proof: Fix Y1 and Y2 with Y1 < 2Y2. If x0,b0,6 sat-
isfy the second equation in (3.3.4) then
6
Y
trace(J(0)) - T[x0>V 1 - 1 (1-x0)J.
2Y
48 3. EXAMPLES
It is easy to show that there is a unique x0(0) E (0,-) that
satisfies trace(J(O)) = 0. Clearly i > 0 for this choice
of x0(0). We now obtain b0(0) and 6(0) as the unique
solution of (3.3.4) and an easy computation shows that
b0(0) > 0, 6(0) > 0 and 6(0)i-1 - 2Y2 > 0. By the implicit
function theorem, for c, x0 - x0(0), b0 - b0(0) sufficiently
small, there exists 6(e,x0,b0) = 6(0) + O(e) such that
trace(J(c)) = 2 i-1x0 - 1 - y1b0 + O(e) = 0.
After substituting 6 = 6(e,x0,b0) into (3.3.4), another

application of the implicit function theorem gives the re-
sult. This completes the proof of the lemma.
From now on we fix Y1 and Y2 with yl < 2Y2. Using

the same calculations as in the lemma, for each c and 6
with a and 6 - 6(c) sufficiently small there is a solution

x0(e,6), b0(e,6) of (3.3.4). Writing x0 = x0(e,6(e)) and
trace(J) as a function of 6, we have that
8
Y16(e)1-1 3
a6(trace(J(6)))
6=6(e) (f x0) [6y1x0
+ (2Y2-y1)x0 - 12y1x0] + O(e) < 0,
if c is sufficiently small. Hence, the eigenvalues of J(6)
cross the imaginary axis with non-zero speed at 6 - 6(c).
Let y1 = R(e)z, z1 - m(c)y where
R(e) - (y1b0(e))-1/2 + O(e), m(c) = [(6/2)x-1-Y2]1/2 + O(e).
For 6 = 6(c), (3.3.8) in these new coordinates becomes
Y1m- 1
y1 ` -w0z1 - (e)y1z1
(3.3.9)
1(e)y1,e)-Y1L-1(c)y1zI
z1 = w0y1 + m(e)(6/2)*-1h(m-1(e)z1,
where
w2(c) = Y1b0(c)[(6/2),y-1 + O(c).

Y2]
To apply the results of the last section, we need to
calculate the K(c) associated with (3.3.9). We shall show

how K(0) can be calculated; if K(0) is non-zero then K(c)
will be non-zero also. To calculate K(0) we need to know

the quadratic and cubic terms in (3.3.9) when c = 0. Thus,
we have to find h(y,z,0) modulo fourth order terms.
Let
(MO)(Y,z) ` -g(,(Y,z),Y,z,O) (3.3.10)
Then by Theorem 2 of Chapter 2, if we can find 0 such that
(MO)(Y,z) = O(y4+z4) then h(Y,z,0) = 0(Y,z) + O(y4+z4).

Suppose that
(3.3.11)
0 _ 02 + 03
where 0j is a homogeneous polynomial of degree j. Substi-
tuting (3.3.11) into (3.3.10) we obtain
(Mo)(Y,z) = *02(Y,z) - 3iy-1x0(x0y+z)2 + Y(XOY+z)

+ O(IYI3 + IzJ3).
Hence, if
1y(x0y+z)
02(Y,z) = 3iy-2x0(Xoy+z)2 - (3.3.12)
then (Mo)(y,z) = O(IyJ3 + jzJ3). Substituting (3.3.11) into
(3.3.10) with g2(y,z) given by (3.3.12), we obtain
(Mo)(Y,z) _ ip03(Y,z) + i-2(XOY+z)3 6iy-1x0(XOY+z)02(Y,z)
+ Y02(Y,z) + O(Y4+z4).
Hence,
50 3. EXAMPLES
h(y,z) 02(y,z) - i-3(x0y+z)3 + 6iy-2x0(xOY+z)02(Y,z)
*-1y02(Y,z)
+ O(Y4+z4)
where 02(y,z) is defined by (3.3.12). K(0) can now be cal-
culated as in the previous section. The sign of K(0) will
depend on yl and y2. If K(0) is non-zero then we can
apply Theorem 1 of Section 2 to prove the existence of
periodic solutions of (3.3.1). The stability of the periodic
solutions is determined by the sign of K(0). If K(0) is
zero then we have to calculate K(e).
3.4. Bifurcation of Maps
In this section we give a brief indication of some re-
sults on the bifurcation of maps. For more details and
references the reader is referred to the book by Iooss [40].
Let V be a neighborhood of the origin in ]Rm and
let F V -,]Rm be a smooth map depending on a parameter
u E]Rq with F(0) = 0 for all V. If bifurcation is to
take place then the linearized problem must be critical. We
assume that F6 has simple complex eigenvalues A(0), a(0)
with IA(0)I - 1, A(0) 0 ±1, while all other eigenvalues of
FO' are inside the unit circle. By centre manifold theory
all bifurcation phenomena take place on a two-dimensional

manifold so we can assume m = 2 without loss of generality.
We first consider the case when u is a one-dimen-
sional parameter and we assume that a(u) crosses the unit
circle, that is
A(u) 0 when u = 0.
U 1-1
If An 0 1 for n - 3 or 4 then in general Fu has an

3.4. Bifurcation of Maps 51
invariant closed curve bifurcating from 0 [57,58]. More-
over, if an = 1, n > 5, then in general Fu does not have
any period points of order n bifurcating from 0 [40].
However, if u is a two-dimensional parameter, then in gen-
eral Fu has periodic points of order n for u in a small
region [4,64]. The method of proof used in [4] and [64] re-
lies on approximating Fu by the time one map of a certain

differential equation; we outline a more direct proof.
Suppose that
Fu (w) - A(u)w + O(IwI2), w E1R2, u E IR2
where A(0) has eigenvalues a(u), a(u) with an(0) - 1,
n > 5. For definiteness we assume that n - 2p + 3 for some
positive integer p; the case n = 2p + 4 is treated in the
same way. Then by changing coordinates [40], Fu has the
normal form
2k
Fn(z) - An(u)z + + b(u)z2p+2 + R(z,z,u)
U
z
k-1 a k(u)l11
where z = x + iy, R(z,z,u) - O(jzj 2p+3) and ak(u), b(u)

are complex numbers. We assume that after a change of para-
meter, An(u) - 1 - u where u - ul + iu2. We also assume

that al(0) and b(0) are nonzero. If z = reie is a fixed
point of Fu then
h(r,O,u) = 0
where
ul + iu2 -
J1k+ b()r2eie + O(r22).
By redefining u + u(a1(u))-1, we can assume without loss of

generality that aI(u) - 1. Given 0 E [0,27T), ul > 0 and
52 3. EXAMPLES
u2 small enough, the equation
Re(h(r,e,u)) - 0
has a unique (small) solution r - r(ul'u2,O) with
r(ul,u2,0) - ul/2(1 + O(u)) (3.4.1)
Substituting this into the equation
Im(h(r,e,u)) = 0 (3.4.2)
we obtain
Im(ak(u))r2k + Im(b(u)e-in9)r2p+1 + O(r2P+2)

u2 =
k=2
Let v2(ul,u2,0 J2kIm(a(u))r2k. We now make the change

o f parameter c = el + ic = ul + i(u2 - u2). By (3.4.1),
2
the map u -+ c is one-to-one for u small enough. Equation
(3.4.2) now takes the form
e2 - Im(b(u)e-in9)e12p+l)/2(1 + O(el/2 + Ie2I)).
Since b(0) 0 0, by the implicit function theorem the above
equation has n distinct solutions 0r(e), r - 1,...,n, for

< Cei2p+1)/2 n
Ie2l where C is a constant. Thus Fu has
(n-2)/2
fixed points for u in a region of width ON
If A3(0) = 1, then for one parameter families of maps,
in general, there are bifurcating fixed points of order 3
and no invariant closed curves [40,41). The case A4(0) = 1
is more complicated. In this case the map can be put into
the normal form
Fu(z) - a(u)z + b(u)z2z + c(u)z3 + O(IzIS)

3.4. Bifurcation of Maps 53
where a(u) - i + O(u). If u E IR then in general either

fixed points of order 4 or an invariant closed curve bifur-
cate from 0; the actual details depend on a(u), b(u) and
c(u) [40,41,681. If u is a two-dimensional parameter then

Fu is studied by investigating bifurcations of the differen-
tial equation
A(u)z + B(u)z2z + C(u)z3 (3.4.3)
where A(u), B(u) and C(u) are chosen such that the time
one map T(u) of (3.4.3) satisfies
Fu (z) - iT(u) (z) - O(Iz15)
The bifurcation diagrams given in [41 for equation (3.4.3)
seem only to be conjectures. Bifurcations of (3.4.3) are

studied in [541 in the case A(u) - u E with bifurcation
parameters Re(u) and Re(B(u)).
We note that the above bifurcation results do not ap-

ply if the map has special properties. For example, F
U
could represent the phase flow of a periodic Hamiltonian sys-
tem so that Fu is a symplectic map. The following result
is due to Moser [531: if F is a smooth symplectic map with
F(0) - 0, then in general F has infinitely many periodic
points in every neighborhood of the origin if the linearized
map F'(0) has an eigenvalue on the unit circle. For an ap-
plication of this result to the existence of closed geodesics

on a manifold see [461.
CHAPTER 4
BIFURCATIONS WITH TWO PARAMETERS IN TWO DIMENSIONS
4.1. Introduction
In this chapter we consider an autonomous ordinary dif-

ferential equation in the plane depending on a two-dimensional
parameter c. We suppose that the origin x = 0 is a fixed
point for all c. More precisely, we consider
z f(x,E), x E ]R2, a El,E2) E ]R2,

(4.1.1)
f(O,E) E 0.
The linearized equation about x = 0 is
z = A(c)x,
and we suppose that A(0) has two zero eigenvalues. The ob-
ject is to study small solutions of (4.1.1) for (E1,E2) in
a full neighborhood of the origin. More specifically, we wish

to divide a neighborhood of c = 0 into distinct components,
such that if e,- are in the same component, then the phase
portraits of (4.1.1)E and (4.1.1)- are topologically equival-

ent. We also want to describe the behavior of solutions for
each component. The boundaries of the components correspond
S4
4.1. Introduction 55
to bifurcation points.
Since the eigenvalues of A(O) are both zero we have
that either (i) A(O) is the zero matrix, or (ii) A(O) has
a Jordan block,
There is a distinction between (i) and (ii) even for the study
of fixed points. Under generic assumptions, in case (ii),
equation (4.1.1) has exactly 2 fixed points in a neighborhood
of the origin. For case (i) the situation is much more com-
plicated [21,29].
Another distinction arises when we consider the eigen-

values of A(c). We would expect the nature of the eigen-
values of A(c) to determine (in part) the possible type of
bifurcation. If
then the eigenvalues of A(c) are always real so we do not
expect to obtain periodic orbits surrounding the origin. On
the other hand if
(4.1.2)
then the range of the eigenvalues of A(e) is a neighborhood

of the origin in , that is, if z is a small complex num-
ber then A(e) has an eigenvalue z for some C.
We shall assume from now on that A(e) is given by

(4.1.2).
56 4. BIFURCATIONS WITH TWO PARAMETERS
Takens [64,65] and Bodganov [see 3] have studied nor-
mal forms for local singularities. Takens shows, for example,
that any perturbation of the equation
2 2
xl = x2 + x1,
is topologically equivalent to
2 2
1 = x2 + xl, x2 = E1 + E x - xl
2 1
for some E1,E2. There are certain difficulties in applying
these results since we must transform our equation into nor-
mal form, modulo higher order terms.
In [47, p. 333-348], Kopell and Howard study (4.1.1)
under the assumptions that A(E) is given by (4.1.2) and that

2
a f
-2 2(0,0) + 0
ax1
where f2 is the second component of f. Their approach con-
sists of a systematic use of scaling and applications of the
implicit function theorem.
In this chapter, we use the same techniques as Kopell
and Howard to study (4.1.1) when the nonlinearities are cubic.
Our results confirm the conjecture made by Takens [64] on the
bifurcation set of (4.1.1).
Similar results are given in [4] together with a brief
outline of their derivation.
The results on quadratic nonlinearities are given in
Section 9 in the form of exercises. Most of these results can

be found in Kopell and Howard [47].
4.2. Preliminaries 57
4.2. Preliminaries
Consider equation (4.1.1) where A(e) is given by
(4.1.2). We also suppose that the linearization of f(x,e)
is A(e)x, and that
f(O,e) = 0, f(x,e) = -f(-x,e). (4.2.1)
The object is to study the behavior of all small solutions of

(4.1.1) for a in a full neighborhood of the origin. Equa-
tion (4.1.1) is still too general, however, so we shall make

some additional hypotheses on the nonlinear terms. Set
f = (fl,f2)T
3 3
- f2(0,0)' s = f2(0,0
axl axlx2
(Hl) a+ 0
(H2) $+ 0
3
f1(0,0) = 0.
(H3) ax3
1
(Hl) implies that for small c, (4.1.1) has either 1
or 3 fixed points. Under (Hl), it is easy to show that by

a change of co-ordinates in (4.1.1), we can assume (H3) (see
Remark 1). We assume (H3) in order to simplify the computa-
tions.
Under (Hl) we can prove the existence of families of

periodic orbits and homoclinic orbits. Under (Hl)-(H3) we can
say how many periodic orbits of (4.1.1) exist for fixed E.
The sign of $ will determine the direction of bifurcation
and the stability of the periodic orbits, among other things.

From now on we assume (Hl)-(H3).
Level Curves of H(yl,y2) a (y2/2)

- (y2/2) + (y4/4).
Figure 1
E1
Bifurcation Set for the Case a < 0, 8 < 0.
Figure 2
The main results are given in Figures 2-5. Sections

3-8 of this chapter show how we obtain these pictures. The
pictures for $ > 0 are obtained by using the change of
variables x2 + -x2, E2 + -E2 and t + -t.
E2
E1
Bifurcation Set for the Case a > 0, 8 < 0.
Figure 3
The cases a > 0 and a < 0 are geometrically dif-

ferent. The techniques involved in each case are the same
and we only do the more difficult case a < 0. The case
a > 0 is left for the reader as an exercise.
From now on we assume a < 0 in addition to (Hl)-(H3).
Note that this implies that locally, (4.1.1) has 1 fixed
point for E1 < 0 and 3 fixed points for E1 > 0.
REGION 1 REGION 2
REGION 3
REGION 4
Phase Portraits for the Case a < 0, a < 0.
Figure 4
ON L1
REGION 5
ON L2
Figure 4 (cont.)
REGION 6
Figure 4 (cont.)
REGION 1
REGION 2
REGION 3
REGION 4
Phase Portraits for the Case a > 0, $ < 0.
Figure S
4.3. Scaling
We scale the variables in equation (4.1.1) so that the
first components of the non-zero fixed points are given by
±1 + 0(E). To do this we introduce parameters u,6, scaled
variables yl,y2 and a new time t by the relations
62l0111/2
6 = lEla-1 1/2, E2 = lall/26u, xl = 6yl, x2 = Y29
t = lal-1/26-1T.
For (u,6) in a neighborhood of the origin, (E1,E2) belongs

to a region of the form {(El,E2): 1Ell
< E0, El I
2
(constant)E 2}. The yi are assumed to lie in a bounded set,
say lyil < M. A further discussion of the scaling is given
in Section 6.
After scaling (4.1.1) becomes
Yl = Y2 + 62g1(u,6,Y)
(4.3.1)
y2 = sgn(E1)Y1 + - y1 + 6yyly2 + 62g2(u,6,Y)
uy2
where the dot means differentiation with respect to t,
y = slat-1/2 and gi(u,6,y) = 0(1). The size of the bounds

on the gi depends on M,u,6. We write t for T from now
on.
The cases E1 > 0 and E1 < 0 are treated separately.
4.4. The Case El > 0
With E1 > 0, equation (4.3.1) becomes
Yl = Y2 + 62g1(u,6,Y)
(4.4.1)
2
Y2 = y1 + zY2 - y1 + 6yyly2 + 6 92(u,6,Y)
Let H(yl,y2) (y2/2) Then along

- (y2/2) +
(y4/4).
4.4. The Case El > 0 65
solutions of (4.4.1),
H(Y1,Y2) = VY2 + 6yy2Y2 + 0(62). (4.4.2)
Note that for u = 6 = 0, H is a first integral of (4.4.1).

The level curves of H(yl,y2) = b consist of a figure
of eight if b = 0, and a single closed curve if b > 0 (see
Figure 1). For b > 0, the curve H(yl,y2) = b passes
through the point yl = 0, y2 = (2b)1'2. For b > 0 and 6
sufficiently small, we prove the existence of a function
u = u1(b,6) = -yP(b)6 + 0(62) such that for b > 0, (4.4.1)
with u = ul(b,6) has a periodic solution passing through the

point yl = 0, y2 = (2b)1'2, and with ul(0,6), (4.4.1)
has a figure of eight solutions.

For fixed u,6, the number of periodic solutions of
(4.4.1), surrounding all three fixed points, depends upon the
number of solutions of
u = ul(b,6) = -yP(b)6 + 0(62). (4.4.3)
We prove that P(b) - m as b - m and that there exists

b1 > 0 such that P'(b) < 0 for b < b1 and P'(b) > 0
for b > b1. These properties of P(b) determine the number

of solutions of (4.4.3).
Suppose, for simplicity of exposition, that ul(b,6)
-yP(b)6 and that y < 0. If 0 < b2 < bl, then there exists
b3 > b1 such that ul(b2,6) = ul(b3,6). Hence, if

ul(b216), then (4.4.1) has two periodic solutions, one
passing through yl = 0, y2 = (2b2)1'2, the other passing
through yl - 0, y2 = u > ul(0,6), then (4.4.1)

(2b3)1"2.
If
has one periodic solution surrounding all three fixed points.

Finally, if u = u(b1,6), then the periodic solutions coincide.
In Figure 4, the periodic solutions surrounding all
three fixed points in regions 3-5 correspond to the periodic
solutions of (4.4.1) which are parametrized by u = ul(b,6),
b > b1. Similarly the "inner" periodic solutions in region 5
are parametrized by u = ul(b,6), 0 < b < bl. The curve L1
in (el,e2) space corresponds to the curve u = ul(0,6).
Similarly the curve L2 corresponds to the curve u = ul(bl,6)
(see Figure 2).
In general ul(b,6) is not identically equal to
-yP(b)6, but the results are qualitatively the same. For
example, we prove the existence of a function bl(6) = bl +
0(6), such that if u,6 satisfy u = ul(bl(6),6), then equa-
tion (4.4.3) has exactly one solution. The curve
u = ul(bl(6),6) which is mapped into the curve L2 in
(el,e2) space corresponds to the points where the two
periodic solutions coincide.
If b < 0, then the set of points for which
H(yl,y2) = b consists of two closed curves surrounding the
points (-1,0) and (1,0). For 0 < c < 1, we prove the
existence of a function u = u2(c,6) = -yQ(c)6 + 0(62) such
that for u = u2(c,6), (4.4.1) has a periodic solution surrounding
the point (1,0) and passing through yl = c, y2 = 0. Using

f(x) = -f(-x), this proves the existence of a periodic solu-
tion surrounding the point (-1,0) and passing through
yl = -c, y2 = 0.
We also prove that Q'(c) > 0 for 0 < c < 1. Let
6 > 0 and suppose u satisfies
u2(0,6) < -sgn(y)u < u2(1,6) (4.4.4)

4.4. The Case el > 0 67
Then the equation u = P2 (c,6) has exactly one solution.
Hence, for fixed u,6 satisfying (4.4.4), equation (4.4.1)
has exactly one periodic solution surrounding (1,0). The
region in (u,6) space corresponding to (4.4.4) is mapped
into region 4 in (el,e2) space (see Figure 2).
Lemma 1. For 6 sufficiently small, there exists a function
u = u(6) = -(4/5)6y + 0(62) such that when u(6), (4.4.1)
has a homoclinic orbit.
Proof: Let S(u,6), U(u,6) be the stable and unstable mani-

folds of the fixed point (0,0) in (4.4.1). These manifolds
exist, since for u,6 sufficiently small, (0,0) is a saddle
[271. Let H(u,6,+) be the value of H(yl,y2) when U(u,6)
hits y2 = 0, yl > 0. Similarly, H(u,6,-) is the value of

H(yl,y2) when S(u,6) hits y2 = 0, yl > 0. H(u,6,±) are
well defined since stable and unstable manifolds depend con-

tinuously on parameters.
Let I(u,6,+) denote the integral of H(yl,y2) over
the portion of U(u,6) with yl > 0, y2 > 0 from yl =
y2 = 0 to y2 = 0, yl > 0. Then
H(u,6,+) = I(u,6,+). (4.4.5)
Similarly, I(u,6,-) denotes the integral of H(yl,y2) over
the portion of S(u,6) with yl > 0, y2 < 0 from yl = y2 = 0
to y2 = 0, yl > 0, so that H(u,6,-) = I(u,6,-).
Equation (4.4.1) has a homoclinic orbit (with yl > 0)
if and only if
H(u,6,+) - H(u,6,-) = 0. (4.4.6)
We solve (4.4.6) by the implicit function theorem.

Using (4.4.2) and (4.4.5),
H(u,6,+) f (iy2 + y6y2y2)dt + O(u2 + 62 (4.4.7)
where the above integral is taken over the portion of U(0,0)
from yl - y2 = 0 to y2 = 0, yl > 0. Similarly,
(1YZ + y6y2y2)dt + O(u2 + 62), (4.4.8)
where the integral is taken over the portion of S(0,0) from
0 to Y2 = 0, yl > 0. Using (4.4.7), (4.4.8) and

yl = Y2 =
U(0,0) = -S(0,0), we obtain
H(u,6,+) = -H(u,6,-) + O(u2 + 62) (4.4.9)
Using (4.4.7) and (4.4.9),
. (H1(0,0,+) - H1(0,0,-)) - 2f y2dt > 0,
so that by the implicit function theorem, we can solve (4.4.6)
for u as a function of 6, say u - u(6). We now show how
to get an approximate formula for u(6). We can write equa-
tion (4.4.6) in the form
of+y2dt + y6f+ y2Y2dt + O(u2 + 62) = 0.
4
Hence, using (4.4.1) and y2 = +[yl2 (yl/2)]1/2, we obtain
yly2dyl
y6f
U= + 0(6 2 ) = - 4 5 + 0(6 2 )
+ y2dy1
This completes the proof of Lemma 1.
Lemma 1 proves the existence of a homoclinic orbit of
(4.1.1) when (E1,E2) lies on the curve L1 given by
1
E2 - -(4/5)lal + O(Ei/2).
8E1
4.4. The Case el > 0 69
Using f(x,e) = -f(-x,e), when (el,e2) lies on L1, equa-
tion (4.1.1) has a figure of eight solutions.

We now prove the existence of periodic solutions of
(4.4.1) surrounding all three fixed points. In the introduc-
tion to this section, we stated that (4.4.1) has a periodic

solution passing through yl = 0, y2 = (2b) 1/2 for any b > 0.
In Lemma 2, we only prove this for "moderate" values of b.
The reason for this is that in (4.3.1) the gi are bounded
only for y in a bounded set. In Section 6, we show that by

a simple modification of the scaling, we can extend these re-
sults to all b > 0.
Lemma 2. Fix b > 0. Then for 0 < b < b and 6 suffici-
ently small, there exists a function u ul(b,6) -yP(b)6 +
0(62) such that if u = ul(b,6) in (4.4.1), then (4.4.1) has
a periodic solution passing through yl = 0, y2 = (2b)1/2.
As b - 0 the periodic solution tends to the figure of
eight solutions obtained in Lemma 1.
Proof: Let H(u,6,b,+) be the value of H(yl,y2) when the
orbit of (4.4.1) which starts at yl - 0, y2 = (2b) 1/2 inter-
sects y2 = 0. Similarly, H(u,6,b,-) is the value of
H(yl,y2) when the orbit of (4.4.1) which starts at yl = 0,

1/2
y2 = -(2b) is integrated backwards in time until it inter-
sects y2 = 0. Then (4.4.1) has a periodic solution passing
through yl = 0, y2 - (2b) 1/2 if and only if
H(u,6,b,+) - H(u,6,b,-) = 0. (4.4.10)
Let I(u,6,b,+) denote the integral of H(yl,y2) over
the portion of the orbit of (4.4.1) with y2 > 0, starting at
yl - 0, y, - (2b)1/2 and finishing at y2 - 0, yl > 0.

Similarly, I(u,6,b,-) is defined by integrating backwards
in time. Thus,
H(u,6,b,±) = b + I(11,6,b,±). (4.4.11)
Using (4.4.2) and (4.4.11),
H(u,6,b,+) - b + J(uy2 + y6y2y2)dt + 0(u2+62) (4.4.12)
where the above integral is taken over the portion of the or-
1/2
bit of (4.4.1) with u = 6 = 0 from yl = 0, y2 = (2b)
to yl = c, y2 = 0 where
4b = c4 - 2c2. (4.4.13)
Similarly, I(u,6,b,-) = -I(u,6,b,+) + O(u2 + 62), so
that equation (4.4.10) may be written in the form
J(uy2 y6y2y2)dt + O(u2 + 62) = 0. (4.4.14)

+
Hence by the implicit function theorem, we can solve (4.4.14)
to obtain u = -yP(b)6 + 0(62) where
2
d
P(b) = yly2 yl (4.4.15)
y2dy1
In order to prove that the periodic solution tends to

the figure of eight solutions as b + 0, we prove that
H(u,6,b,±) - H(u,6,±) as b - 0. (4.4.16)
This does not follow from continuous dependence of solutions

on initial conditions, since as b + 0 the period of the
periodic solution tends to infinity. The same problem occurs
in Kopell and Howard [47, p. 339] and we outline their method.
For yl and y2 small, solutions of (4.4.1) behave like

4.4. The Case Cl > 0 71
solutions of the linearized equations. The proof of (4.4.16)
follows from the fact that the periodic solution stays close
to the solution of the linearized equation for the part of the
solution with (yl,y2) small and continuous dependence on

initial data for the rest of the solution.
Lemma 3. P(b) + - as b + m.
Proof: The integrals in (4.4.15) are taken over the curve
y2 In (y1/2) + 2b]1/2, from yl = 0 to yl = c where
c is defined by (4.4.13). Thus, JO(b)P(b) = Jl(b) where
Ji(b) = Jc w2i(w2 - (w4/2) + 2b)1/2dw. (4.4.17)

0
Substituting w = cz in (4.4.17) we obtain
r1
Ji(b) = c2 (cz)2ig(z)dz
J
0
where g(z) _ [(z2-1) + (c2/2)(l-z4)]1'2. Since g(z) <
g(c-1) for 0 < z < 1, we have that JO(b) < Dlc3 for some
positive constant D1. Similarly, there exists a positive

constant D2 such that Jl(b) > D2c5. The result now follows.
Lemma 4. There exists b1 > 0 such that P'(b) < 0 for
b < b1 and PI(b) > 0 for b > b1.
Proof: It is easy to show that P'(b) + as b + 0. Hence,
by Lemma 3 it is sufficient to show that if P'(bl) = 0 then
P"(bl) > 0.
Let r(w) _ [w2 - (w4/2) + 2b]1/2. Differentiating

(4.4.17) with respect to b we obtain
c 2i
dw. (4.4.18)
Ji J r w
0
Integrating by harts in .J0 we obtain

c 4 2
JO = wr w dw. (4.4.19)
1
0
Also
1c r2 w w4 2 + 2b]
J = w = (c [w2 dw. (4.4.20)
0 0 r(w) r w
O
Similarly,
1c rc 2
rww
3J1 =
0
iw,6ww4 w = 3
0 rwj [w2 - (w4/2)+2b]dw. (4.4.21)
Using (4.4.18) - (4.4.21) we can express J0 and J1 in
terms of JI and J. A straightforward calculation yields
3J0 = 4bJ0 + J1
(4.4.22)
15J1 = 4bJ6 + (4+12b)Ji.
Suppose that P'(b1) = 0. Then JO(b1)P"(b1) _
J111(b1) - P(b1)J0"(b1). Using (4.4.22) we obtain
4b1(4b1+l)[J"(b1) - P(b1)J3(b1)]
(4.4. 23)
J6(b1) [P2(b1) + 8b1P(b1) - 4b1] .
Hence P"(b1) has the same sign as
P2(b1) + 8b1P(b1) - 4b1. (4.4.24)
Since P'(b1) = 0 we have that Ji(b1) = P(b1)J;(b1). Using

(4.4.22) we obtain
5P2(b1) + 8b1P(b1) - 4P(b1) - 4b1 = 0. (4.4.25)
Using b1 > 0, it is easy to show that (4.4.25) implies that
P(b1) < 1. Using (4.4.24) and (4.4.25), P"(b1) has the same
sign as P(b1) - P2(b1). This proves that P"(b1) > 0 as
required.
4.4. The Case e1 > 0 73
Lemma S. For 6 sufficiently small there exist b1(6) =
b1 + 0(6), b2(6) = b2 + 0(6), where P(0) = P(b2), with the

following properties:
Let 6 > 0.
(i) if u = u1(b1(6),6), then the equation
u = 'l(b,6) (4.4.26)
has exactly one solution.
(ii) If y < 0 and u1(b1(6),6) < u < ul(b2(6)P6)p

then equation (4.4.26) has exactly two solutions b3(6), b4(6)
with b3(6) = b3 + 0(6), b4(6) = b4 + 0(6), where b3 and
b4 are solutions of
ly 1.
P(b) = -ud (4.4.27)
A similar result holds if y > 0.
(iii) If y < 0 and u > ul(b2(6),6), then (4.4.26)

has exactly one solution b5(6) = b5 + 0(6), where b5 is
the unique solution of (4.4.27).
(iv) If y < 0 and u < u1(b1(6),6), then (4.4.27)
has no solutions. A similar result holds for y > 0.
Proof: By Lemma 4, there exists b2 > 0 such that P(b2) _
P(0) and P'(b2) > 0. Set g(z,6) = 6-1[111(b2+z,6) -
ul(0,6)], for 6 + 0 and g(z,O) = 0. Then g(z,6) _

P'(b2)z + 0(161 + z2). By the implicit function theorem there
exists z(6) = 0(6) such that g(z(6),6) = 0. Hence, if
b2(6) = b2 + z(6) then u1(b2(6),6) = u1(0,6). The existence
of b1(6) is proved in a similar way. The rest of the lemma
follows from the properties of P(b).
We now prove the existence of periodic solutions of

(4.4.1) currnunding a jingle fixed point.
Lemma 6. For 0 < c < 1 and 6 sufficiently small, there
exists u = u2(c,6) _ -yQ(c)6 + 0(62) such that if
u - u2(c,6), then (4.4.1) has a periodic solution passing
through yl - c, y2 - 0. As c - 0 the periodic orbits tend
to the homoclinic orbit obtained in Lemma 1.
Proof: Let H(u,6,c,+) be the value of H(yl,y2) when the

orbit of (4.4.1) starting at yl - c, y2 - 0 intersects
y2 - 0, y2 > 0. Similarly, H(u,6,c,-) is the value of

H(yl,y2) when the orbit of (4.4.1) starting at yl = c,
y2 = 0 is integrated backwards in time until it intersects
y2 - 0. Equation (4.4.1) will have a periodic orbit passing

through yl - c, y2 - 0 if and only if
H(u,6,c,+) - H(u,6,c,-) - 0. (4.4.28)
Using the same method as in Lemma 2, we can rewrite
equation (4.4.28) as
G(u,6,c) - j(uy2 + y6y2y2)dt + O(u2 + 62) - 0 (4.4.29)
where the above integral is taken over the curve
y2 = r(yl) = [y1 - (y4l /2) + (c4/2)-c2]1/2, (4.4.30)
from yl = c, y2 - 0 to y2 = 0 again. By (4.4.29)
au G(0,0,c) - f y2dt > 0. (4.4.31)
Thus, for fixed c we can solve (4.4.29). We cannot solve

(4.4.29) uniformly in c however, since as c - 1 the right
hand side of (4.4.31) tends to 0. We use a method similar to
that used by Kopell and Howard in [47, p. 337-338] to obtain
u2(c,6) for 0 < c < 1.
4.4. The Case c1 > 0 75
Equation (4.4.1) has a fixed point at (yl,y2) -
(1,0) + O(IiiI,I61). For u,6 sufficiently small we make a

change of variables yj = hj(yl)y2,u,6) so that the fixed
point is transformed into (yl,y2) = (1,0). An easy calcula-
tion shows that if we make this transformation, then the only
change in (4.4.1) is in the functions gl and g2. We sup-
pose that the above change of variables has been made and we
write y, for y,.
The curves H(yl,y2) = H(c,0) can be written in the
form
4H(Y1,Y2)+1 - 2y2 + (y1-l) 2(yl+l)2 - (c-1)2(c+l)2. (4.4.32)
Thus, for c close to 1 the closed curves are approximately
y2 + 2(yl-l) 2 - 2(c-1) 2. (4.4.33)
So, instead of equation (4.4.29) we consider the equation
G(1,6,c) = (c-1)-2G(c,u,6) - 0. (4.4.34)
If we prove that G(u,6,c) is bounded for 0 < c < 1 and
that (3/3u)_(0,0,c) is bounded away from zero for 0 < c < 1,
then we can solve (4.4.34) uniformly in c.
Now
H(Y1,Y2) = VY2 + 6(yy12

y + 6Y2g2 + 6(y3l-yl)g1].
Since the fixed point is at (1,0), H(yl,y2) is bounded above

and below by quadratic forms in y2 and yl - 1. By (4.4.32),
4H(yl,y2) + 1 is bounded above and below by quadratic forms

in y2 and yl - 1. Hence there exist functions gi(u,6)
O(juj + 161) such that

gl(u,6) < at ln(4H(yl,y2) + 1) g2(u,6)
Integrating the above inequality over the curve given by
(4.4.30) we obtain
exp(gl(11,6)T) < [4H(u,6,c,+)+1]/[4H(c,0)+1]

(4.4.35)
< exp(g2(u,6)T),
where T is a bound for the time taken to trace the orbit.
Using (4.4.29), (4.4.32) and (4.3.35) we see that 6(11,6,c)
is bounded. The fact that (3/3v)-(c,0,0) is bounded away

from zero follows easily from (4.4.32). This proves that for
6 sufficiently small and 0 < c < 1, we can solve (4.4.32)
to obtain u - u2(c,6) - -yQ(c)6 + 0(62) where
J0(c)Q(c) = J1(c),
d
Ji(c) - 1 w2ir(w)dw
c
and r(w) is defined by (4.4.30) and r(d) - 0, d > c.
The fact that the periodic solution tends to the homo-

clinic orbit is proved in the same way as the corresponding
result in Lemma 2.
Using f(x) = -f(-x), Lemma 6 proves the existence of

periodic solutions of (4.4.1) surrounding (-1,0).
Lemma 7. Q'(c) > 0 for 0 < c < 1.
Proof: We write Q,J0 and J1 as functions of b where
4b - c4 - 2c2. Since (db/dc) < 0 for 0 < c < 1, we must

prove that Q1(b) < 0 for -1 < 4b < 0. Following the same
procedure as in Lemma 4, we find that J0,J0,, J1,Ji satisfy
equation (4.4.22). Thus, if Q'(bl) - 0 then
4.5. The Case c1 < 0 77
5Q2(bl) + 8b1Q(bl) - 4Q(bl) - 4b1 = 0. (4.4.36)
Since -1 < 4b1 < 0, the roots of (4.4.36) are less than 1.
Hence, if Q'(bl) - 0 then Q(bl) < 1. Also, from (4.4.23),

if Q'(bl) - 0 then Q"(bl) has the same sign as 4b1 -
8b1Q(bl) - Q2(bl). Using (4.4.36) and Q(bl) < 1, this im-
plies that Q"(bl) < 0. Since Q(-1/4) - 1, Q(O) - 4/5, this

shows that Q1(b) < 0 for -1 < 4b < 0. This completes the
proof of the lemma.
4.5. The Case El < 0
With < 0, equation (4.3.1) becomes

El
yl = Y2 + 6g1(u,6,Y)
(4.5.1)
Y2 - -Yl + uY2 - yl + 6Yyly2 + 62g2(v,6,Y)
Let H1(yl'y2) = (y2/2) + (y1/2) + (y4l/4). Then along solu-
tions of (4.5.1), H1 - yZ + 6yy2y2 + 0(62). Using the same

methods as in the previous lemmas, we prove that (4.5.1) has
a periodic solution passing through y1 - c > 0, y2 = 0, if
and only if u = u3(c,6) - -yR(c)6 + 0(62), where
JO(c)R(c) = J1(c),
c
J1(c) - r w2ir(w)dw,
0
r(w) = [2b - (w4/2) - w211/2, 4b - c4 + 2c2.
In order to prove that for fixed u,6, equation (4.5.1)
has at most one periodic solution we prove that R is strictly
monotonic.
Lemma 8. R'(c) > 0 for c > 0.

Proof: We write R,JO,J1 as functions of b. It is suffici-
ent to prove that R'(b) > 0 for b > 0.
Using the same methods as before, we show that
3J0 - 4bJ6 - J1
(4.5.2)
15J1 - (12b+4)Ji - 4bJ6
3wwdw
- 4bJ6 - 4Ji. (4.5.3)
to
Now R' has the same sign as S where S - 15[JiJO - J0,J1].
By (4.5.2),
S - (8b-4)JO'Ji - 5(Ji)2 + 4b(Jo,) 2.
By (4.5.3), bJ0 > J. Hence, for b > 2,
S > (J,)2b-1(8b-4-5b+4)
- 3(Ji)2 > 0.
Similarly, if 0 < b < 2, then S > 3b2(Jo)2 > 0. This com-

pletes the proof of Lemma 8.
4.6. More Scaling
In Section 4 we proved the existence of periodic solu-
tions of (4.4.1) which pass through the point (0,(2b)1/2).
Lemma 3 indicates that we may take b to be as large as we

please. However our analysis relied on the fact that the gi
are 0(1) and this is true only for y in a bounded set.
Also, our analysis in Sections 4 and S restricts E1,E2 to a

region of the form {(E1,E2): 1c11 < EO, E1 < (constant) c } .
To remedy this we modify the scaling by setting

6 = 1Ela-111/2h-1, where h is a new parameter with, say,
0 < 4h < 1. The other changes of variables remain the
same. Note that if u,6 lie in a full neighborhood of the
4.6. More Scaling 79
origin then (c1,e2) and (xl,x2) lie in a full neighbor-
hood of the origin.
After scaling, (4.1.1) becomes
Yl - Y2 + 62g1(11,6,h,y)
(4.6.1)
y2 - h2sgn(el)Y1 + 11Y2 - yl + 6yyiy2 + 62g2(i,6,h,Y)
For 0 < 4h < 1 and 11,6 sufficiently small, the gi are
0(1) for y in a bounded set.
Let c1 > 0. Let H2(Y1OY2) _ (Y2/2) - (h2y2/2) +
(y4/4). Then along solutions of (4.6.1), H2 - 11y2 +

y6y2y2 + 0(112 + 62). Following the same procedure as in Sec-
1 2
tion 4, we find that (4.6.1) has a periodic solution passing

through yl = 0, y2 = 1, if and only if
11 = h2111(b,6) = -yh2P(b)6 + 0(62),
where 2b - h-4. An easy computation shows that as h 0,
h2P(b) = K + 0(h2), where J0K = 2 J1,

r1
Ji = fl w
0
Thus, for small h, (4.6.1) has a periodic solution passing
through yl = 0, y2 - 1 if and only if 11 - -yK6 +

0(h2161 + 62). In particular, when el - 0, E2 > 0, B < 0,
(4.1.1) has a periodic solution passing through xl - 0,
x2 =
IaI(-BK)-1c2
+ 0(c3/2)
For < 0, a similar analysis shows that (4.6.1) has

c1
a periodic solution passing through yl = 1, y2 - 0, if and
only if
h2u2(h-1,6) -yh2Q(h-1)6
u = = + 0(62),
and that as h 0, h2Q(h-1) - (K//f) + 0(h2).

4.7. Completion of the Phase Portraits
Our analysis in the previous sections proves the exist-
ence of periodic and homoclinic orbits of (4.1.1) for certain
regions in (el,e2) space. We now show how to complete the
pictures.
Obtaining the complete phase portrait in the different
regions involves many calculations. However, the method is
the same in each case so we only give one representative
example. We prove that if y < 0, then the "outer" periodic
orbit in region 5 is stable. We use the scaling given in
Section 3.
Since we are in region 5, ul(0,d) > u > ul(bl(d),d).
Fix u and d and let b2 be the solution of u = ul(b,d)
with b2 > bl(d). Then we have to prove that the periodic
solution r of (4.4.1) passing through yl = 0, y2 = (2b2)1/2
is stable.
Let b3 be the solution of u = ul(b,d) with
b3 < bl(d). Then there is a periodic solution passing through

yl = 0, y2 = (2b3)1/2.
In fact, we prove that any solution
of (4.4.1) starting "outside" this periodic solution (and in-
side some bounded set) tends to r as t - m.
Let b > b3. Let c(b,+) be the value of yl when
the orbit of (4.4.1) starting at yl = 0, y2 = (2b) 1/2 first
hits y2 = 0. Similarly, c(b,-) is the value of yl when
the orbit of (4.1.1) starting at yl = 0, y2 = (2b) 1/2 is
integrated backwards until it hits y2 = 0. Using f(x,c) _
-f(-x,c), the solution passing through (0,(2b)1/2), spirals
inwards or outwards according to the sign of c(b,+) + c(b,-).
Using the calculation in Lemma 2, c(b,+) + c(b,-) has the

4.8. Remarks and Exercises 81
same sign as H(u,d,b,+) - H(u,d,b,-) which in turn has the

same sign as
S = ul(b2,d) - ul(b,6).
Using the properties of ul(b,b) given in Lemma 5, S is
positive if b < b2 and negative if b > b2. The result now

follows.
4.8. Remarks and Exercises
Remark 1. Consider equation (4.1.1) under the hypothesis

that the linearization is given by x = A(c)x where A(c)
is defined by (4.1.2). Suppose also that (4.2.1) and (Hl)

hold. Make the change of variables x = B(c)x, where
B(c) = I - ra-1A(e) and
a3f1(0,0) a3f2(0,0)
r =
3
ax 3X3
The map x - x will be one-to-one for c sufficiently small.

Using the fact that A(e) and B(c) commute, it is easy to
show that the transformed equation satisfies all the above
hypotheses and that in addition it enjoys (H3). In particu-
lar, if the transformed equation is
x = F(x,e)
then
a3F1(0,0)
ax 0
a3F2(0,0) a3f2(0,0)
3
-3
ax ax
1 1
ax
-
a3F2(0,0) a3f2(0,0)
ax +
3a3f1(0,0)
a3
Remark 2. We have assumed that f(x,e) = -f(-x,e). If we
assume that this is true only for the low order terms then we
would obtain similar results. For example, on L1 we would
get a homoclinic orbit with x1 > 0. Similarly on another
curve Li we would get a homoclinic orbit with x1 < 0. In
general L1 and Li would be different although they would

have the same linear approximation u = -(4/5)6y.
Remark 3. Suppose that we only assume (Hl) and (H3). Then
we can still obtain partial results about the local behavior

of solutions. For example, in Lemma 2 we did not use the
hypothesis B nonzero. Hence, for each b > 0, (4.1.1) has
a periodic solution through x1 = 0, x2 =

IaI-1/2(2b)l/2c1
for some e1 and e2 with e1 > 0. However, we cannot say
anything about the stability of the periodic orbits and we
cannot say how many periodic orbits (4.1.1) has for fixed e1
and 62.
Exercises
(1) Suppose that a and B are negative. Prove
that for (61,62) in regions 5 and 6, (4.1.1) has a connect-

ing orbit. (Hint: Use the calculations in Lemma 1 and
Lemma 6.)
(2) Suppose that a and g are negative and that

(61,62) is in region 3. Let U be the unstable manifold of
the point (0,0) in (4.1.1). Prove that U is in the re-
gion of attraction of the periodic orbit.

4.9. Quadratic Nonlinearities 83
(3) Suppose that a is positive and 8 is negative.
Show that the bifurcation set and the corresponding phase por-
traits are as given in Figures 3 and S.
4.9. Quadratic Nonlinearities
In this section we discuss the local behavior of solu-
tions of (4.1.1) when the nonlinearities are quadratic. Most
of the material in this section can be found in [47].

Suppose that the linearization of (4.1.1) is x = A(e)x
where A(e) is given by (4.1.2) and that
2
f(O,e) = 0, fl(0,0) 0,
2 2
ax2
a - f2(O,O) 0, 8
ax1ax2 f2(O,O) , 0.
1
We also assume that 8 > 0, a < 0; the results for the other
cases are obtained by making use of the change of variables,
t - -t, e2 -1. -e2, x1 -. ±x1, x2 -. +x2,

Introduce parameters u,6, scaled variables y1,y2
and a new time t by the relations
a-1 1/2 2 a 1/263

d = l
e1 1
x1 = 6 Y1, x2 = l l Y2,
tlal-1/26-1
e2 = 'all/26u, t =
We assume that e1 > 0 in what follows, see Exercise 8 for
the case e1 < 0.
After scaling, (4.1.1) becomes
y1 = y2 + 0(62)
(4.9.1)
2
Y2 ' Y1 + "Y2 - y1 + 6YYly2 + 0(6)
2
where the dot means differentiation with respect to t, the

yi lie in a bounded set and y - a a,-1/2. Note that by
making a change of variable we can assume that (1,0) is a
fixed point of (4.9.1) for p,6 sufficiently small.
The object of the following exercises is to show that
the bifurcation set is given by Figure 6 and that the associa-
ted phase portraits are given by Figure 7.
Exercises.
(4) Let H(y1,Y2) (Y2/2) - (y2/2) + (y /3). Show

=
that along solutions of (4.9.1),
(5) Prove that there is a function u = ul(c,6),
0 < c < 1, such that if u = u1(c,6) in (4.9.1) then there

is a periodic solution through (c,0) if 0 < c < 1 and a
homoclinic orbit if c = 0.
(6) Prove that u1(c,6) - -yP(c)6 + 0(62) where
J0(c)P(c) = J1(c),
(c1
Ji(c) = Jc w1R(w)dw,
R(w) _ [w2 - (2/3)w3 + 2b] 1/2,
6b - c2(2c-3), R(c1) - 0, c1 > c.
(7) Prove that P(0) = (6/7), P(l) = 1 and P'(c) > 0
for 0 < c < 1. Deduce that for fixed u,6, (4.9.1) has at
most one periodic solution. (To prove that P'(c) > 0 we can
use the same techniques as in the proof of Lemma 4. An al-
ternative method of proving P'(c) > 0 is given in [47].)

(8) If el < 0, then after scaling (4.1.1) becomes
Yl ° Y2 + 0(62)
2 2
y2 = -yl + PY2 - yl + y6Y1Y2 + 0(6 )
Put yl s z - 1, u = p - y6. Then
a y2 + 0(62)
y2 = z + PY2 - z2 + 6yzy2 + 0(62)
which has the same form as equation (4.9.1) and hence trans-
forms the results for eI > 0 into results for the case
E1 < 0.
(9) Show that the bifurcation set and the correspond-
ing phase portraits are as given in Figures 6-7.
E1
Bifurcation Set for the Case a < 0, 8 > 0.
Figure 6
REGION 1
REGION 2
Phase Portraits for the Case a < 0, 8 > 0. (For the phase
portraits in regions 4-6 use the transformations in Exer-
cise 8.)
Figure 7
ON L
REGION 3
Figure 7 (cont.)
CHAPTER 5
APPLICATION TO A PANEL FLUTTER PROBLEM
5.1. Introduction
In this chapter we apply the results of Chapter 4 to a

particular two parameter problem. The equations are
z = Ax + f(x) (5.1.1)
where
x - [xl,x2,x3,x4]T, f(x) = [f 1(x),f2(x),f3(x),f4(x)]T,
r 0 1 0 0 1
al b1 c 0
A --
0 0 0 1
-c 0 a2 b2
f1(x) = f3(x) = 0,
f2(x) = xlg(x),
f4(x) = 4x3g(x),
2g(x) = -n4(kx2 + axlx2 + 4kx2 + 4ax3x4),
c = 8p,
3 a j
-n2j2[n2j2 + r],
bj - -[an4j4 + T 6]; a - 0.005, 6 = 0.1,
as
5.2. Reduction to a Second Order Equation 89
k > 0, a > 0 are fixed and p,r are parameters. The above
system results from a two mode approximation to a certain

partial differential equation which describes the motion of a
thin panel.
Holmes and Marsden [36,38] have studied the above equa-

tion and first we briefly describe their work. By numerical
calculations, they find that for p = p0 = 108, r = r0 =
-2.237r2, the matrix A has two zero eigenvalues and two
eigenvalues with negative real parts. Then for lp-p01 and
Ir-r01 small, by centre manifold theory, the local behavior

of solutions of (5.1.1) is determined by a second order equa-
tion depending on two parameters. They then use some results
of Takens [64] on generic models to conjecture that the local

behavior of solutions of (5.1.1) for lp-p01 and Ir-r0l
small can be modelled by the equation
u + au + bu + u2u + u3 = 0
for a and b small.
Recently, this conjecture has been proved by Holmes
[37], in the case a = 0, by reducing the equation on the

centre manifold to Takens' normal form. We use centre mani-
fold theory and the results of Chapter 4 to obtain a similar
result.
5.2. Reduction to a Second Order Equation
The eigenvalues of A are the roots of the equation
a4 + d1A3 + d2a2 + d3A + d4 = 0 (5.2.1)
where the di are functions of r and p. If A has two
zero eigenvalue% then d3 - d4 - 0. A calculation shows that

90 S. APPLICATION TO A PANEL FLUTTER PROBLEM
if d3 - d4 - 0, then
a1a2 + c2 = 0
(5.2.2)
a1b2 + b1a2 - 0
or in terms of r and p,
64
4n4(n2+r)(4n2+r) + p2 = 0 (5.2.3)
(16an4+6p1/2)(n2+r) + 4(an2+6p1/2)(4n2+r) - 0. (5.2.4)
We prove that (5.2.3), (5.2.4) has a solution r - r0,

p p0'
From (5.2.3) we can express p in terms of r. Sub-
stituting this relation into (5.2.4) we obtain an equation

H(r) = 0. Calculations show that H(r1) < 0, H(r2) > 0 where
rl - -(2.225)r2 and r2 = -(2.23),r2, so that H(r0) = 0 for
some r0 E (r2,r1). Further calculations show that (5.2.3),
(5.2.4) has a solution r0,p0 with 107.7 < p0 < 107.8.
In the subsequent analysis, we have to determine the

sign of various functions of r0 and p0. Since we do not
know r0 and p0 exactly we have to determine the sign of

these functions for r0 and p0 in the above numerical
ranges.
When r = r0, p = p0, the remaining eigenvalues of A

are given by
(b1+b2) t [(b1-b2)2 + 4(a1+a2)]1/2

X3,4
and a calculation shows that they have negative real parts
and non-zero imaginary parts.
We now find a basis for the appropriate eigenspaces
when r = r0, p - p0. Solving Av1 - 0 we find that

5.2. Reduction to a Second Order Equation 91
vl - [1,0,-al/c,0]T. (5.2.5)
The null space of A is in fact one-dimensional so the can-
onical form of A must contain a Jordan block. Solving
A2v2 - 0 we obtain
v2 = [0,1,-b1/c,-a1/c]T, Av2 = v1. (5.2.6)
The vectors vi and v2 form a basis for the generalized
eigenspace of A corresponding to the zero eigenvalues.
Similarly, we find a (real) basis for the space V spanned
by the eigenvectors corresponding to A3 and A4. Solving
Az - A3z, we find that V is spanned by v3 and v4 where
2v3 - z + z, 2v4 = i(z-z),
z - [1,A3,w,A3w]T (5.2.7)
we - b A - b1b2 + a2.
2 3
Let A0 denote the matrix A when I' = r0 and
p = p0. Let S = [vl,v2,v3,v4] where the vi are defined

by (5.2.5), (5.2.6) and (5.2.7) and set y = S-lx. Then
(5.1.1) can be written in the form
jy = By + S-1(A-A0)SY + F(Y,r,P) (5.2.8)
where F(y,I',p) = S-lf(Sy),
0 1 0 0 1
0 0 0 0
0 0
pl p2
0 0
-p2 pl JI
and where aS - Pi + ip2, pl < 0, p2 t 0.

Then for Ir-r0I and Ip-p0I sufficiently small
(5.2.8) has a centre manifold y3 = hl(yl'y2'r'p)'
h2(yl,y2,r,p). The flow on the centre manifold is gov-

y4 =
erned by an equation of the form
yl 0 1 Y1 Y1
] = [ ][ E(r,P) [
Y2 0 0 y2 y2
(5.2.9)
+ N(Y1,Y2,r,P)
where E(r,p) is a 2 x 2 matrix with E(r0,p0) - 0 and
N(yl,y2,r,p) contains no linear terms in yl or y2. We
show that there is a nonsingular change of variables (r,p)
(c1,c2) for (r,p) close to (r0,p0) and a r,p dependent
change of variables such that the lin-

(yl,y2) i (yl'y2)
earized equation corresponding to (5.2.9) is
yl 0 1
yl
(5.2.10)
Y2 E1 c2 y2
The transformation is of the form

(yl,y2) i (yl'y2)
Identity + O(Ir-r0I +Ip-POI) and r - rot p = p0 is mapped
into el = 0, c2 = 0. After these transformations, (5.2.9)
takes the form
yl a 0 1 'l + N(y1,Y2,r0,P0)
Y2 c2 Y2
E1
(S. 2.11)
+ N(Y1,Y2,r,P)
where we have dropped the bars on the yi. The function N
will contain no linear terms in yl or y2 and
N(yl'y2'r0,p0) = 0. Since the nonlinearities in (5.1.1) are
cubic, the same will be true of N and A.

5.3. Calculation of Linear Terms 93
Let
N(Y1,Y2,ro,PO) _ [N1(Y1,Y2,ro,PO), N2(Y1,Y2,r0,P0)]T (5.2.12)
and let
3
al = -s N2(o,o,ro,PO)
ayl
3
ay3 N1(o,o,ro,PO)
(5.2.13)
1
3
S = a N2(o,o,r0,PO)
aylay2
8 = 3r + S.
Using the results in Chapter 4 (see, in particular, Remark 1

in Section 8), if a1 and 0 are non-zero, we can deter-
mine the local behavior of solutions of (5.2.11). By Theorem
2 of Chapter 2, this determines the local behavior of solu-
tions of (5.2.8).
5.3. Calculation of the Linear Terms
From (5.2.1), trace(A) = d3(r,p), det(A) = d4(r,p)
where
d3(r,P) = 4n2(n2+r) (ant+r) + 69 p2
d4(r,P) - n2(16an4+6p1/2)(n2+r) + 4n2(air2+6P1/2)(4n2+r)
Calculations show that the mapping
(r,p) (d3(r,P),d4(r,P)) (5.3.1)
has non-zero Jacobian at (r,p) = (ro,po).
Define the matrix C(r,p) by

C(r,p) = + E(r,p)
and let J be the value of the Jacobian of the mapping
(r,p) -. (trace(C(r,p)), det(C(r,p))), (5.3.2)
evaluated at (r0,p0). By considering the 4 x 4 matrix
B + S-1(A-AO)S, it is easily seen that J is a non-zero
multiple of the Jacobian of the mapping given by (5.3.1).
Hence J is non-zero, so by the implicit function theorem we
can use E1 = -det(C(r,p)), E2 ' trace(C(r,p)) as our bifur-
cation parameters. Approximate formulae for E1 and E2
can easily be found if so desired.
Let C(r,p) - [cij],
M =
[:11 :12]
yl -M Y1
Y2 Y2
J
Then the linearized equation corresponding to (5.2.9)
is
yl 0 1
yl
y2 E1 E2
Y2
Note that M is equal to the identity matrix when
e1 = E2 = 0.
5.4. Calculation of the Nonlinear Terms 95
5.4. Calculation of the Nonlinear Terms
We now calculate the nonlinear term in (5.2.9) when

r - rot p - p0. Since the nonlinearities in (5.1.1) are
cubic, the centre manifold has a "cubic zero" at the origin.
Using x - Sy, on the centre manifold
xl - Yl + 03, x2 y2 + 03
-a bl (5.4.1)
x3 - yl - E Y2 + 03, x4 al
c Y2 + 03
where 03 - 0(IY1I3 + IY2I3).
Let S-1 - [t..] and let
Fj(Y1,Y2) - fj(Y1,Y2,hl(Y1,Y2,r0,PO)).
Then using the notation introduced in (5.2.12),
N1(Y1,Y2,r0,PO) ' t12F2(Y1,Y2) + t14F4(Y1,Y2)

(5.4.2)
N2(Y1,Y2,r0,PO) - t22F2(Y1,Y2) + t24F4(Y1,Y2)
Using (5.4.1),
al 8ka1b1 2
n4 3 2 3 2
F2(Y1,Y2) ' - [kyl + 4k(c) yl + ayly2 + yly2
c
+ 4aal 2
Y1Y2]
2
+ terms in
yly2 + 05
4
-4c 1
F2(Yl,y2) + 2bcn rkyiy2
F4(Yl,y2) = + 4k( c1)2 YlY21
+ terms in yly2 and y3 + 05
where 05 - 0(Iyl1S + IY21 ). Note also that since r - r0

and p - pot (5.2.2) holds.
From (5.2.13) and (5.4.2),

96 5. APPLICATION TO A PANEL FLUTTER PROBLEM
a,
-T- k(l + 4 (c) 2(t22
- 74
al - 4 ac t24)
4 a a
r k(1 + 4(c)2)(t12 4
- c t14)
2
_ _n4 8kalbl 4aa1 al
2ir4
(a + + )(t22 - 4 c t24) + -3b1kt24(c 2+4a1).
2
Routine calculations show that
t12 = m2b2(2a1-b1b2)
t22 = mat
t14 = m2c(bl+b2)
t24 = -mc
m = (al+a2-blb2)-1 < 0.
Using (5.2.2) and numerical calculations, we find that

a
t22 - 4 c t24 = m(4a1+a2) > 0
t14 a b1m2(2a2-bZ-4a1) > 0

t12 - 4 ac
so that al and r are negative. Similarly, the coefficient
of k in is
7 a1b1m(12a1+a2) < 0.
c
Hence a1 and a are negative and the local behavior of solu-
tions of (5.1.1) can be determined using the results of
Chapter 4.
CHAPTER 6
INFINITE DIMENSIONAL PROBLEMS
6.1. Introduction
In this chapter we extend centre manifold theory to a

class of infinite dimensional problems. For simplicity we
only consider equations of the form
w = Cw + N(w), w(O) E Z
where Z is a Banach space, C is the generator of a strongly

continuous semigroup on Z and N: Z - Z is smooth. In the
next section we give a brief account of semigroup theory.

For additional material on semigroup theory see [6,43,50,55,
561. For generalizations to other evolution systems see
[34,51).
6.2. Semigroup Theory
In earlier chapters we studied centre manifold theory
for the finite dimensional system
w - Cw + N(w). (6.2.1)
The most important tools for carrying out this program were:
97
98 6. INFINITE DIMENSIONAL PROBLEMS
(1) The solution of the linear problem = Cw and estimates

of exp(PCt), exp((I-P)Ct), where P is the projection
onto the space associated with eigenvalues of C with
zero real parts.
(2) The variation of constants formula
rt
w(t) = exp(Ct)w(O) + exp(C(t-s))N(w(s))ds
J
0
for the solution of (6.2.1) and Gronwall's inequality.
To carry out this program for partial differential
equations we have to study ordinary differential equations in
infinite dimensional spaces. We first study linear problems.
Let Z be a Banach space and C a linear operator
from some domain D(C) of Z into Z. We wish to solve the

problem
w = Cw, t > 0, (6.2.2)
w(0) = w0 E Z. (6.2.3)
Suppose that for each w0 E Z the above equation has a
unique solution w(t). (We define later the meaning of the
term solution). The solution w(t) is a function of t and
w0 and we write w(t) = T(t)w0. Then T(t) is a linear
mapping from Z into Z with T(0) = I. If w(t) is a
solution of (6.2.2) - (6.2.3), then for any s > 0, w(t)
w(t+s) = T(t+s)w0 is a solution of (6.2.2) with w(0) - w(s).
Hence T(t+s)w0 = T(t)T(s)w0. If solutions depend continu-
ously on initial data then T(t)wn T(t)w whenever wn - w

in Z, so that T(t) must be bounded. Finally, we require
that T(t)w0 + w0 as t - 0+.
6.2. Semigroup Theory 99
Definition. A one parameter family T(t), 0 < t < of
bounded linear operators from Z into Z is a (strongly

continuous) semigroup if
(i) T(O) = I,
(ii) T(t+s) = T(t)T(s), t,s > 0 (semigroup property),
(iii) IIT(t)w - wII- 0 as t - 0 for every w E Z.
Example 1. Let L(Z) denote the set of bounded linear opera-

tors from Z into Z and let C E L(Z). Define T(t) by
n
T(t) = eCt L Cn.
n=0
The right hand side converges in norm for each t > 0 and it
is easy to verify conditions (i), (ii), (iii). Thus eCt
defines a semigroup.
Example 2. Let Z be a Banach space of uniformly continuous
bounded functions on [0,m) with the supremum norm. Define
(T(t)f)(8) = f(O+t), f E Z, 8 > 0, t > 0.
Conditions (i) and (ii) are obviously satisfied and since
IIT(t)f - f11 = sup{If(0+t) - f(8)I: 8 > 0} - 0 as t - 0+,
(iii) is satisfied. Hence T(t) forms a semigroup.
Definition. The infinitesimal generator C of the semigroup

T(t) is defined by
T(t)z - z
Cz = lim+
t+0 t
whenever the limit exists. The domain of C, D(C), is the set

of all elements z E Z for which the above limit exists. We
also say that C generates T(t).

In Example 1 the infinitesimal generator is C E L(Z)
while in Example 2 the infinitesimal generator is
(Cf) (e) = f' (e) , D(C) - {f: f' E Z}.

Exercise 1. If C E L(Z) prove that IleCt - 111- 0 as
t - 0.
Exercise 2. Let Z - R2 and define T(t): Z - Z by
(e-ntan),
T(t)(an) = (an) E Z, t > 0.
Prove that T(t) is a semigroup on Z with infinitesimal
generator given by
C(an) = (-nan), D(A) - {(an): (nan) E Z).
Does II T (t) - III - 0 as t - 0+?

Exercise 3. Prove that if T(t) is a semigroup then
IIT(t)II < Mewt, t > 0, (6.2.4)
for some constants w > 0, M > 1. (Hint: Use the uniform

boundedness theorem and property (iii) to show that IIT(t)11
is bounded on some interval [O,e]. Then use the semigroup
property.)
Exercise 4. Find a 2 x 2 matrix C such that if
Il eCt Il < M for all t > 0 then M > 1.
Exercise S. Let C be the generator of a semigroup T(t)
which satisfies IIT(t)II < Mewt, t > 0. Prove that C - mI
is the generator of the semigroup S(t) - e-WtT(t) and that
11 S(t) 11 < M for all t > 0.

Exercise 6. If T(t) is a semigroup, prove that for each
w E Z, T(t)w is a continuous function from [0,m) into Z.
Let T(t) be a semigroup with generator C. Now
h-1[T(t+h)w - T(t)w] = h-1[T(h) - I]T(t)w

(6.2.5)
- T(t)h-1[T(h)w - w].
If w E D(C) then the right side of (6.2.5) converges to
T(t)Cw as h - 0+. Thus the middle term in (6.2.5) con-
verges as h - 0+ and so T(t) maps D(C) into D(C). Also
the right derivative satisfies
dT t w = CT(t)w - T(t)Cw, t > 0, w E D(C). (6.2.6)

t
Similarly, the identity
h-1[T(t)w - T(t=h)]w - T(t-h)h-1[T(h)w - w]
shows that (6.2.6) holds for the two-sided derivative.
Exercise 7. For w E Z, prove that w(t) E D(C) where
t
w(t) = T(s)w ds,
J0
and that t-1w(t) - w as t + 0+. Deduce that D(C) is
dense in Z.
Exercise 8. If wn E D(C), then from (6.2.6)

t
T(t)wn - wn = T(s)Cwnds.
J0
Use this identity to prove that C is closed.
Equation (6.2.6) shows that if w0 E D(C) and C is
the generator of a semigroup T(t), then w(t) = T(t)w0 is a
solution of the Cauchy problem (6.2.2) - (6.2.3). In applica-

102 6. INFIMITE DIMENSIONAL PROBLEMS
tions to partial differential equations, it is important to
know if a given operator is the generator of a semigroup. To
see the problems involved, suppose that C generates the con-
traction semigroup T(t), that is IIT(t)II < 1 for all t > 0.
The Laplace transform
R(a)w = J e-AtT(t)w dt, (6.2.7)

0
exists for Re(A) > 0 and IIR(A)II< A 1 for A > 0. Since
T(t) is in some sense exp(Ct), we expect R(A) to be the
resolvent of C, that is, (XI - C)-1. This is indeed the case
and is easy to prove. The main problem is in finding the in-
verse Laplace transform, that is, given C such that
(XI - C)-1 exists for A > 0, say, is C the generator of a
semigroup? The basic result is the Hille-Yosida Theorem.
Theorem 1 (Hille-Yosida Theorem). A necessary and sufficient
condition for a closed linear operator C with dense domain
to generate a semigroup of contractions is that each A > 0
is in the resolvent set of C and that
I IR(A) I I = I I ( A I - C ) - 1 11 <A1 for A > 0. (6.2.8)
The reader is led through the proof of the above
theorem in the following exercises.
Exercise 9. Let T(t) be a contraction semigroup and for
A > 0 define R(A) as in (6.2.7) for w E Z. Use the identity
Ah
h-1(T(h)-I)R(A)w = h-'[(e - 1)j e-AtT(t)wdt
0
h
- eah
1
e-AtT(t)w dt]
0
to prove that R(A) maps Z into D(C) and that

(XI - C)R(A) = I.
Prove also that for w E D(C) , R(X) (XI - C)w = w and deduce
that R(X) is the resolvent of C.
Exercise 10. Suppose that C is a closed linear operator on
Z with dense domain such that (0,m) is in the resolvent
set of C with (6.2.8) satisfied. For A > 0 define
C(X) E L(Z) by
C)-1 C)-1
C(X) = XC(XI - = X2(XI - - XI.
Prove that
(i) C(a)w - Cw as A for w E D(C),
(ii) Il exp (tC(X)) II < 1 and

Ilexp(tC(X))w - exp(tC(u))wII < tlIC(X)w - C(u)wII
for t > 0, X,u > 0 and w E Z.
Use the above results to define T(t) by
T(t)w = lim exp(C(A)t)w

aim
and verify that T(t) is a semigroup of contractions on Z
with generator C.
Exercise 11. Let Z = R2 and let {an} be a sequence of

complex numbers such that a = sup{Re(an): n = 1,2,...} <
Define C by
C(an) = (an an ), D(C) = {(an): (an an) E Z}.
Use the Hille-Yosida Theorem to prove that C - aI generates
a contraction semigroup and deduce that C generates a semi-
group T(t) with IIT(t)II < eat for t > 0.

Remark. The above version of the Hille-Yosida Theorem char-
acterizes the generators of semigroups which satisfy

eWt
IIT(t)II < (see Exercise 5). A similar argument gives the
characterization of the generators of all semigroups.
Remark. Let C be a closed operator with dense domain such

that all real nonzero A are in the resolvent set of C and
II(AI - C)_111.' IAI 1 for such A. Then C is the generator
of a group T(t) , t EIR, that is, T(0) = I, T(t+s) = T(t)T(s)
all t,s E IR and T(t)z - z as t + 0 for each z E Z.

Example 3. Let C be a selfadjoint operator on a Hilbert
space H with C < 0 . Then 11 (XI - C) 1 1 1 < A 1 for A > 0,
so by the Hille-Yosida Theorem, C generates a contraction

semigroup.
Example 4. Let A = iC where C is a selfadjoint operator
on a Hilbert space H. Then for all nonzero real A,
I I (AI -A)_111 < IAI so that A generates a unitary group.

Let f E C([0,T];Z) and let C be the generator of a
semigroup T(t) on Z. Then formally the solution of
w = Cw + f(t), 0 < t < T, w(0) = w0, (6.2.9)
is given by the variation of constants formula
t
w(t) = T(t)w0 + I
T(t-s)f(s)ds. (6.2.10)
0
If f is continuously differentiable and w0 E D(C) then it
is easy to check that w(t) defined by (6.2.10) satisfies
(i) w(t) is continuous for 0 < t < T,
(ii) w(t) is continuously differentiable for
0 < t < T,
(iii) w(t) E D(C) for 0 < t < T and (6.2.9) is sat-
isfied.
A function w(t) which satisfies (i), (ii), (iii) is

said to be a strong solution of (6.2.9). It is easy to check
that such a solution is unique in this class. If w0 t D(C)
or f is only continuous, then in general w(t) defined by
(6.2.10) is not in D(C) so (6.2.9) does not make sense. We
now define what we mean by a solution in this case.
Definition. A function w E C([0,T];Z) is a weak solution
of (6.2.9) if w(0) = w0 and if for every v E D(C*), the
function <w(t),v> is absolutely continuous on [0,T] and
dt<w(t),v> = <w(t),C*v> + <f(t),v> a.e.
where C* is the adjoint of C and < , > denotes the pair-
ing between Z and its dual space.
The proof of the following theorem can be found in 191.
Theorem 2. The unique weak solution of (6.2.9) is given by

(6.2.10).
A semigroup T(t) is said to be an analytic semigroup

if in addition the map t - T(t)z is real analytic on (0,W)
for each z E Z. Analytic semigroups have many additional

properties. In particular, if C is the generator of an analy
tic semigroup T(t), then T(t) maps Z into D(C) for each
t > 0. Thus if f is continuously differentiable, then w(t)
defined by (6.2.10) is a strong solution of (6.2.9) for each

w0 E Z if T(t) is an analytic semigroup. An example of a
generator of an analytic semigroup is a nonpositive selfad-

joint operator and in general, solutions of parabolic equations
are usually associated with analytic semigroups. Hyperbolic
problems, however, do not in general give rise to analytic
semigroups. To see this, suppose that C is the generator
of a group T(t) and that T(t) maps Z into D(C) for
t > 0. By the closed graph theorem, CT(l) E L(Z) and hence
C = CT(1)T(-l) is a bounded linear operator.
In applications, it is sometimes difficult to prove

directly that C + A is a generator, although C is a gen-
erator and A is in some sense small relative to C. The
simplest such result is the following.
Theorem 3. Let C be the generator of a semigroup (group)
T(t) and let A E L(Z). Then C + A generates a semigroup

(group) U(t). If IIT(t)II < McWt for t > 0 then
IIU(t) II < Me(-+MIIAII)t for t > 0.

Outline of the Proof of Theorem 3: If A + C is the genera-
tor of U(t), then by the variation of constants formula, U(t)
must be the solution of the functional equation

t
U(t) - T(t) + T(s)AU(t-s)ds. (6.2.11)
J
0
Equation (6.2.11) is solved by the following scheme:
U(t) - F Un(t) (6.2.12)

n-0
where U0(t) - T(t) and
rt
Un+1(t) - T(s)AUn(t-s)ds.
0
A straightforward argument shows that U(t) defined by
(6.2.12) is a semigroup with generator A + C.
Exercise 12. Prove Theorem 3 in the case M - 1, w - 0 in

the following way: let B = A + C so that B is a closed

operator with dense domain D(B) - D(C). Show that for
A > IIAII, IIA(AI - C)_111 <1 so that (I - A(AI - C)-1)-1
exists for A > IIAII Hence show that II (AI - B)_111 <
(A - IIAII)-1 for A > IIAII and use the Hille-Yosida Theorem.
Example 5 (Abstract Wave Equation). Let A be a positive

selfadjoint operator on a Hilbert space H and let B E L(H).
Consider the equation
v + B4 + Av - 0. (6.2.13)
Define a new Hilbert space Z = D(A1/2) x H with inner pro-

duct < , > defined by
<w,w> - (A1/2w1,A1"2w1) + (w2,w2)
where ( , ) is the inner product in H and w = [wl,w21,
w = [wl,w21. Equation (6.2.13) can now be written as a first

order system on Z,
w = Cw (6.2.14)
where
C -
D(C) = D(A) x D(A1/2). If B - 0 then C is skew-selfad-
joint so by Example 4 it generates a group. For B # 0, C is
is the sum of a generator of a group and a bounded operator

so by Theorem 3 it generates a group.
As a concrete example of the above, consider the wave

equation
vtt + vt - Ov = 0, x E SI, v(x,t) = 0, x E BSI,
where D is a bounded open subset of IRn with boundary BSI.
This equation has the form (6.2.13) if we set H L2(SI),
D(A) _ {w E H: -Ow E H, w = 0 on 801, A = -0, B I. In

this case Z = H01 (0) x L2(SI) (see [2] for a definition of
1
the Sobolev space H0 (S2)).
Let T(t) be a contraction semigroup on Hilbert space
H with generator C. Then for s > 0,
IjT(s+t)zII _ IIT(t)T(s)zjl < IjT(t)zjj
so that IIT(t)zII is nonincreasing. For z E D(C), v(t)
11T(t) zII 2 is differentiable in t and
v1(t) _ <T(t)Cz, T(t)z> + <T(t)z, T(t)Cz>.
Since v'(0) < 0, we have that
<Cz,z> + <z,Cz> < 0, z E D(C). (6.2.15)
A linear operator C with dense domain is said to be dissi-
pative if (6.2.15) holds.
Exercise 13. If C is dissipative and is the generator of a
semigroup, prove that the semigroup is a contraction.
Theorem 4 (Lumer-Phillips Theorem). Let C be a dissipative
operator on the Hilbert space H. Suppose that for some

a > 0 the range of BI - C is H. Then C is the generator
of a contraction semigroup.
Exercise 14. Prove the above theorem. (Hint: deduce from
(6.2.15) that II (XI-C) z II > Ali z it for A> 0 and z E D(C).

Use this and the fact that the range of BI - C is II to
deduce that C is closed. Show that the non-empty set

(X > 0: range of XI - C is H) is open and closed and use
the Hille-Yosida Theorem.)
Remark. There is also a Banach space version of the above
result which uses the duality map in place of the inner pro-
duct.
Exercise 15. Let C be the generator of a contraction semi-
group on a Hilbert space and let B be dissipative with

D(B) D(C) and
IIBwll < alUCwll + bllwll , w E D(C)
for some a with 0 < 2a < 1. Use the inequality

IIB(AI - C) lwll < (2a + bX 1)Ilwll to show that AI - B - C is
invertible for large enough A. Deduce that B + C generates
a contraction semigroup.
Exercise 16. Let C be a closed linear operator on a Hil-

bert space H and suppose that C and C* are dissipative.
Prove (i) the range of I - C is closed in H,
(ii) <h - Ch,z> - 0 for all h E D(C) implies z = 0. De-
duce that the range of I - C is H so that by Theorem 4,
C generates a contraction semigroup on H.
If all the eigenvalues of a matrix C have negative

real parts then Ilexp(Ct)II < Met, t > 0, for some M > 1
and w > 0. The corresponding question in infinite dimen-

sions is: if C is the generator of a semigroup T(t), does
Re(a(C)) < 0 imply that IIT(t)II < Me-wt, t > 0, for some
w > 0 and M > 1? (a(C) denotes the spectrum of Q. In
general, the answer is no. Indeed, it can be shown that if

a < b then there is a semigroup T(t) on a Hilbert space
such that sup Re(a(C)) - a and IIT(t)II = ebt [69]. Thus to
obtain results on the asymptotic behavior of the semigroup in
terms of the spectrum of the generator, we need additional
hypotheses.
W t
If e 0 is the spectral radius of T(t), that is
eWOt - lim IIT(t)kIIl/k,
then a standard argument shows that for m > m0 there exists
M(w) such that IIT(t)II < M(w)eWt. Hence we could determine

the asymptotic behavior of T(t) from the spectral radius
of T(t). However the spectrum of T(t) is not faithful to
the spectrum of C, that is, in general the mapping relation
a(T(t)) - exp(ta(C)) (6.2.16)
is false. While (6.2.16) is true for the point and residual

spectrum (55] (with the possibility that the point 0 must
be added to the right hand side of (6.2.16)) in general we

only have
continuous spectrum of T(t) exp(t(continuous spectrum of C)).
R2
Example 6. Let Z - and C(xn) - (inxn), D(C) =
{(xn) E Z: (nxn) E Z}. Then C generates the semigroup
T(t) given by T(t)(xn) _ (eintxn). The spectrum of C is
{in: n - 1,2,...,} while the spectrum of T(l) is the unit
circle so (6.2.16) is false.
For special semigroups, e.g., analytic semigroups,
(6.2.16) is true (the point zero must be added to the right
hand side of (6.2.16) when the generator is unbounded), but

this is not applicable to hyperbolic problems. For the prob-

lem studied in this chapter, the spectrum of the generator
consists of eigenvalues and the associated eigenfunctions
form a complete orthonormal set. Thus the asymptotic be-
havior of the semigroup could be determined by direct compu-

tation. It seems worthwhile however to give a more abstract
approach which will apply to more general problems.
Let C be the generator of a semigroup T(t) with
IIT(t)II< McWt. Hence the spectrum of T(t) lies in the disc
IaI < exp(ort). Let A E L(Z). Then A + C generates a semi-
group U(t). Suppose that there are only isolated eigen-

values of A + C in Q - {A: Re(A) > w}. Then if A E Q,
eXt is an eigenvalue of U(t). We prove that if A is com-

emt}
pact then the spectrum,of U(t) in the set {A: IXI >
consists only of points e?t, A E Q.
Theorem 5 (Vidav [67], Shizuta [62]). Let C be the genera-
tor of a semigroup T(t) with IIT(t)II < McWt and let
A E L(Z) be compact. Then A + C generates a semigroup
U(t) such that U(t) - T(t) is compact.
Proof: From Theorem 3, A + C generates a semigroup U(t)
which is a solution of the functional equation
rt
U(t) - T(t) + T(s)AU(t-s)ds. (6.2.17)
I
0
We prove that the map s - f(s) = T(s)AU(t-s) is continuous
in norm on [O,t]. Thus the integral in (6.2.17) converges
in norm. Since f(s) is a compact linear operator and the

set of compact operators in L(Z) is closed in the uniform
operator topology, U(t) - T(t) is compact.
To prove the claim on the continuity of f we first
prove that s -r AU(t-s) is continuous in norm. Let
t > s > 0 and h1 > 0 with h1 sufficiently small. Then
the set
{A(U(t-s-h) - U(t-s))w: IIwII = 1, IhI < hl}
has compact closure. For e > 0 it can be covered with a
finite number of balls with radius a and centres at
wl,w2,...,wn. Let g(s) = AU(t-s). Then for IIwII = 1,
Ilg(s+h)w - g(s)wII < IIA(U(t-s-h) - U(t-s)) (w-wk) II

+ IIA(U(t-s-h) - U(t-s))wkll
Hence if 6 is chosen such that
IIA(U(t-s-h) - U(t-s))wkll < E
for IhI < 6 and k = 1,2,...,n then IIg(s+h) - g(s)II <
(constant) E. The continuity of f follows from similar
reasoning applied to
f(s+h) - f(s) - T(s+h)[g(s+h)-g(s)] + (T(s+h)-T(s))AU(t-s).
This completes the proof of the theorem.
Since U(t) - T(t) is compact, the essential spectrum
of U(t) and T(t) coincide [43]. (It is important to note
that we are using Kato's definition of essential spectrum in
this claim.) Hence we have proved:
Corollary to Theorem S. The spectrum of U(t) lying outside

emt
the circle IxI = consists of isolated eigenvalues with
finite algebraic multiplicity.

The above result is also useful for decomposing the

space Z. Suppose that the assumptions of Theorem S hold.
Since there are only isolated points in the spectrum of U(t)
outside the circle IXI = eWt, there is only a finite number

of points in the spectrum of A + C in the halfplane
Re(X) > m + e for each e > 0. Let A1,...''n denote these
eigenvalues for some fixed e > 0 and let P be the corres-
ponding projection. Then we can decompose U(t) into the
sum PU(t) + (I-P)U(t) and for any 6 > e there exists
M(6) such that ii(I-P)U(t)ii < M(6)eN+6)t Finally, PZ is
finite dimensional and so it is trivial to make further de-

compositions of PU(t).
We now show how the above theory can be applied to

hyperbolic problems. Let H be a Hilbert space with inner
product ( , ). Let A be a positive selfadjoint operator
on H such that A-1 is defined on all of H and is com-
pact. Consider the equation
v + 2a' + (A+B)v = 0
where B E L(H) and a > 0. Let Z be the Hilbert space
D(A1/2) x H with inner product < , > as defined in Example
S. We can recast the above equation in the form
w = Cw
r0 I
C =
A-B -2aI ].
As in Example 5, C is a bounded perturbation of a skew-
selfadjoint operator and so C generates a group S(t). Let
1
I 01 I 0
U(t) S(t)
IJ
I
aI aI I
Then U(t) is a group with generator
L
-A
aI I
-aI
r 0
a2I-B
0
Cl + C2
and the compactness of the injection D(A1/2) -r H implies
that C2: Z + Z is compact. Thus the above theory applies.
Since C1 generates a group U1(t) with HU1(t)II < e-at,
the asymptotic behavior of U(t) (and hence S(t)) depends
on a finite number of eigenvalues.
The above theory applies, for example, to the case

H = L2(0), where fl is a bounded domain in IRn, A . -D,
D(A) _ {v E H: Av E H and v - 0 on 9c} since by standard
elliptic theory, A-1 is compact.
Example 7. Consider the coupled set of wave equations

Tr
utt + taut - uxx + of g(x,s)u(s,t)ds - By = 0

0 (6.2.18)
vtt + 2avt - vxx . 0
for 0 < x < Tr with u = v = 0 at x = 0,Tr, where
g(x,s) = E n 2 sin nx sin ns.

n=1
As in Example 5 we can write (6.2.18) as a first order system
w = Cw
on Z = (H0(0,7r) x is a bounded pertur-

L2(O,Tr))2. Since C
bation of a skew-selfadjoint operator it generates a

group T(t).
An easy computation shows that the eigenvalues of C

- a n) 1/2,
are an = -a ± (a2 n = 1,2,..., where an = n2 or
an = n2 + (Bn)/(2n2). For a > 0 and B small enough the

real parts of all the eigenvalues are negative so we expect
that IIT(t)II < Me--t, t > 0, for some w > 0. This formal
argument can be rigorized by applying the above theory.
We now suppose that a = 0. In this case, for B
sufficiently small all the eigenvalues of C are purely ima-
ginary and they are simple. By analogy with the finite di-
mensional situation we expect that IIT(t)II < M, t > 0, for
some constant M. We show that this is false.
Consider the following solution of (6.2.18):
u(x,t) = 2n-1 m(cos mt - cos amt)sin mx

(6.2.19)
v(x,t) = m-l sin mx cos mt
where m is an integer and am = m2 + (Bn)/(2m2). Note that
the initial data corresponding to (6.2.19) is bounded inde-
pendently of in. Let tm 2m3n. Then for large m,
2
cos mtm - cos amtm = 1 - cos(n26) + 0(m 4) > constant,
so that Iux(x,tm)I > (constant)m2. Thus IIT(tm)II
(constant)m2.
The above instability mechanism is associated with the
fact that for a = 0, B # 0, the eigenfunctions of C do not
form a Riesz Basis. For further examples of the relationship
between the asymptotic behavior of solutions of w = Cw and
the spectrum of C see [15,16].
We now consider the nonlinear problem

w = Cw + N(w), w(0) = w0 E Z, (6.2.20)
where C is the generator of a semigroup T(t) on Z and
N: Z - Z. Formally, w satisfies the variation of constants
formula
rt
w(t) = T(t)w0 + T(t-s)N(w(s))ds. (6.2.21)
J
to
Definition. A function w E C([O,T];Z) is a weak solution
of (6.2.20) on [0,T] if w(0) = w0; E L1([O,T];Z)
and if for each v E D(C*) the function <w(t),v> is ab-
solutely continuous on [0,T] and satisfies
Ut <w(t),v> = <w(t),C*v> + <N(w(t)),v> a.e.
where C* is the adjoint of C and < , > denotes the pair-
ing between Z and its dual space.
As in the linear case, weak solutions of (6.2.20) are
given by (6.2.21). More precisely:
Theorem 6 [9]. A function w: [0,T] -. Z is a weak solution
of (6.2.20) on [O,T] if and only if E L1([O,T];Z)
and w is given by (6.2.21).
As in the finite dimensional case, it is easy to solve
(6.2.20) using Picard iteration techniques.
Theorem 7 [61]. Let N: Z -. Z be locally Lipschitz. Then
there exists a unique maximally defined weak solution
w E C([O,T);Z) of (6.2.20). Furthermore, if T < W then
I Iw(t) I I -- as t-T . (6.2.22)
As in the finite dimensional case, (6.2.22) is used as
a continuation technique. Thus if for some w0 E Z, the

solution w(t) of (6.2.20) remains in a bounded set then
w(t) exists for all t > 0.
6.3. Centre Manifolds
Let Z be a Banach space with norm We consider

ordinary differential equations of the form
w = Cw + N(w), w(0) E Z, (6.3.1)
where C is the generator of a strongly continuous linear
semigroup S(t) and N: Z - Z has a uniformly continuous
second derivative with N(0) = 0, N'(0) = 0 [N' is the
Frechet derivative of N].
We recall from the previous section that there is a
unique weak solution of (6.3.1) defined on some maximal inter-
val [0,T) and that if T < W then (6.2.22) holds.
As in the finite dimensional case we make some spec-
tral assumptions about C. We assume from now on that:
(i) Z = X ® Y where X is finite dimensional and Y
is closed.
(ii) X is C-invariant and that if A is the restric-
tion of C to X, then the real parts of the
eigenvalues of A are all zero.
(iii) If U(t) is the restriction of S(t) to Y, then
Y is U(t)-invariant and for some positive con-
stants a,b,
ae-bt,
IIU(t)1i< t > 0. (6.3.2)
Let P be the projection on X along Y. Let
B - (I-P)C and for x E X, y E Y, let

f(x,y) = PN(x+y), g(x,Y) ' (I-P)N(x+Y). (6.3.3)
Equation (6.3.1) can be written
z Ax + f(x,y)
(6.3.4)
Y By + g(x,y).
An invariant manifold for (6.3.4) which is tangent to

X space at the origin is called a centre manifold.
Theorem 8. There exists a centre manifold for (6.3.4),
y = h(x), lxi < 6, where h is C2.
The proof of Theorem 8 is exactly the same as the
proof given in Chapter 2 for the corresponding finite dimen-

sional problem.
The equation on the centre manifold is given by
u - Au + f(u,h(u)). (6.3.5)
In general if y(O) is not in the domain of B then y(t)
will not be differentiable. However, on the centre manifold
y(t) - h(x(t)), and since X is finite dimensional x(t),
and consequently y(t), are differentiable.
stable (asymptotically stable) (unstable). Then the zero
solution of (6.3.4) is stable (asymptotically stable)
(unstable).

II(x(0),y(0))II sufficiently small. Then there exists a solu-
tion u(t) of (6.3.5) such that as t -

x(t) - u(t) + O(e-Yt)

(6.3.6)
Y(t) - h(u(t)) + O(e_Yt)
where Y > 0.
The proof of the above theorem is exactly the same as

the proof given for the corresponding finite dimensional
result.
Using the invariance of h and proceeding formally
we have that
h'(x)[Ax + f(x,h(x))] - Bh(x) + g(x,h(x)). (6.3.7)
To prove that equation (6.3.7) holds we must show that h(x)
is in the domain of B.
Let x0 E X be small. To prove that h(x0) is in
the domain of B it is sufficient to prove that

U(t)h(x0) - h(x0)
lim
ti0+ t
exists. Let x(t), y(t) = h(x(t)) be the solution of

(6.3.4) with x(0) = x0. As we remarked earlier, y(t) is
differentiable. From (6.3.4)
r0 t
y(t) = U(t)h(x0) + U(t-T)g(x(T),Y(T))dT,

1
so it is sufficient to prove that
t
lim+ i U(t-T)g(x(T),y(T))dT
0
exists. This easily follows from the fact that U(t) is a
strongly continuous semigroup and g is smooth. Hence
h(x0) is in the domain of B.

Theorem 10. Let 0 be a C1 map from a neighborhood of the
origin in X into Y such that 0(0) - 0, 0'(0) = 0 and
O(x) E D(B). Suppose that as x _ 0, (Mo)(x) = O(jxjq),
q > 1, where
(MO) (x) _ 0' (x) [Ax + f (x, 0) 1 - BO (x) - 9 (x, 0 (x) )
Then as x - 0, IIh(x) - O(x) 11

= O(Ixlq).
The proof of Theorem 10 is the same as that given for
the finite dimensional case except that the extension
6: X -r Y of 0 must be defined so that 8(x) is in the
domain of B.
6.4. Examples
Example 8. Consider the semilinear wave equation
vtt + vt - vxx - v + f(v) ` 0, (x,t) E (0,Tr) x (0,°°)

(6.4.1)
v= 0 at x - O,Tr
where f is a C3 function satisfying f(v) - v3 + 0(v4)
as v -r 0. We first formulate (6.4.1) as an equation on a
Hilbert space.
Let Q - (d/dx) 2 + I, D(Q) - H2(0,Tr) f1 H1(O,Tr). Then

Q is a self-adjoint operator. Let Z - H1(0,Tr) x L2(0,Tr),
then (6.4.1) can be rewritten as
w = Cw + N(w) (6.4.2)
where
w2
Cw N(w)
Qwl-w2 - f(wl)
Since C is the sum of a skew-selfadjoint operator and a

6.4. Examples 121
bounded operator, C generates a strongly continuous group.
Clearly N is a C3 map from Z into Z.
The eigenvalues of are an = [-1 ±

(5-4n2)1/2]/2.
ai = 0 and all the other eigenvalues have real part less
than 0. The eigenspace corresponding to the zero eigen-
value is spanned by ql where
1
ql(x) I Js in x.
0
To apply the theory of Section 3, we must put (6.4.2)

into canonical form. We first note that Cq2 - -q2 where
1
q2(x) ]sin x
and that all the other eigenspaces are spanned by elements of
the form an sin nx, n > 2, an EIR2. In particular, all

other eigenvectors are orthogonal to ql and q2. Let
X - span(gl), V - span(gl,g2), Y span(q 2) ® V1, then
Z - X ® Y. The projection P: Z X is given by
wl
P = + w2)gl (6.4.3)
1
w2
L
where
n
2
w. = w. (6) sin e do.
1
0
Let w = sql + y, s EIR, y E Y and B = (I-P)C. Then we can
write (6.4.2) in the form
sql = PN(sgl+y)
(6.4.4)
y - By + (I-P)N(sgl+y)
By Theorem 8, (6.4.4) has a centre manifold y - h(s),
h(O) = 0, h'(0) - 0, h: (-6,6) - Y. By Theorem 9, the equa-
tion which determines the asymptotic behavior of solutions of

(6.4.4) is the one-dimensional equation
sql - PN(sgl + h(s)). (6.4.5)
Since the nonlinearities in (6.3.4) are cubic, h(s) - 0(s3),
so that
rn
s = n f(s+0(s3))sin 46 d9
1
0
or
s = -33 s3 + 0(s4). (6.4.6)
Hence, by Theorem 9, the zero solution of (6.4.4) is asymptoti-
cally stable. Using the same calculations as in Section 1 of
Chapter 3, if s(O) > 0 then as t - -,
s(t) = ( t) 1/2 + o(t-1/2). (6.4.7)
Hence, if v(x,t) is a solution of (6.4.1) with v(x,O),
vt(x,O) small, then either v(x,t) tends to zero exponen-
tially fast or
v(x,t) - ±s(t)sin x + 0(s3) (6.4.8)
where s(t) is given by (6.4.7).
Further terms in the above asymptotic expansion can be
calculated if we have more information about f. Suppose that
f(v) = v3 + av5 + 0(v7) as v - 0. In order to calculate an
approximation to h(s) set
(MO)(s) ` O'(s)PN(sgl+0(s)) - BO(s) - (I-P)N(sgl+0(s))

(6.4.9)
6.4. Examples 123
where 0:IR + Y. To apply Theorem 10 we choose O(s) so that
(Mo)(s) = 0(s5). If O(s) = 0(s 3) then
MO(s) = -BO(s) - (I-P)N(sgl) + 0(s5)

(6.4.10)
-BO(s) - s3g2 - 0 ]q + o(s5)

1 1
where q(x) = sin 3x. If
s
(s) - ag2s3 + 1
]qs3 (6.4.11)
s2
then substituting (6.4.11) into (6.4.10) we obtain
(MO) (s) = ag2s3 +

- S3 0
801+02
2
qs3 - s3g2 -
1
q + 0(s5)
Hence, if a - 3/4, 01 - 1/32, 02 0, then MO(s) - 0(s5),
so by Theorem 10
1
3 ]qs3 + 0(s5).
h(s) = g2s3 + 3 (6.4.12)
0
Substituting (6.4.12) into (6.4.5) we obtain
-3s3 _ 213 + 5a s5 + 0 s7
s = 2T y a ( )
4 1
The asymptotic behavior of solutions can now be found using

the calculations given in Section 1 of Chapter 3.
Example 9. In this example we apply our theory to the equa-
tion
1
vtt + vt + vxxxx - [B + (2/r4)J (vs(s,t))2dslvxx = 0,

0
(6.4.13)
with v - vxx - 0 at x - 0,1 and given initial conditions

v(x,0), vt(x,0), 0 < x < 1. Equation (6.4.13) is a model for
the transverse motion of an elastic rod with hinged ends, v
being the transverse deflection and B a constant. The
above equation has been studied by Ball [7,8] and in particu-
lar he showed that when a = -Tr2 the zero solution of
(6.4.13) is asymptotically stable. However, in this case the
linearized equations have a zero eigenvalue and so the rate
of decay of solutions depends on the non-linear terms. In
[13] the rate of decay of solutions of (6.4.13) was found
using centre manifold theory. Here we discuss the behavior
of small solutions of (6.4.13) when B + Tr2 is small.
As in the previous example we formulate (6.4.13) as an
ordinary differential equation. We write ' for space deri-
vatives and for time derivatives. Let
Z = H2(0,1) f1 H1(0,1) x L2(0,1), Qv = -V,,,, + By",
wl w2
w= C w= N (w) _
w2 Qww2 [f:1] )
2w 1
Tr 0
Then we can write (6.4.13) as
w = Cw + N(w). (6.4.14)
It is easy to check that C generates a strongly continuous
group on Z and that N is a C" map from Z into Z.
If A is an eigenvalue of C then we must have a non-
trivial solution u(x) of
u,,,, - Bu" + (),+A2)u = 0
u(0) = u"(0) = u(1) = u"(1) = 0.

6.4. Examples 125
An easy computation then shows that A is an eigenvalue of
C if and only if
2A = -1 ± [1 - 4 (n4Tr4+Bn2Tr2) ] 1/2.
Let E = r20 + Tr4. Then the eigenvalues of C are an(E),

where 2A1(E) = -1 + [1-4E]1/2, A2(E) = -1 - A1(E), and all
the rest of the an(E) have real parts less than zero for
E sufficiently small. The eigenspaces corresponding to
al(E) and A2(E) are spanned by ql and q2 where
1 1
ql(x) sin Trx, q2(x) _ sin Trx
1
(E) 2
(E)
while the eigenspaces corresponding to An(E) for n > 2
are spanned by elements orthogonal to ql and q2.
Let X = span(gl), V = span(gl,g2), Y = span(g2) 0 V

then Z = X 0 Y and the projection P: Z - X is given by
wl
PI = (XI(E) - )2(E)) 1(w2-a2(E)WI)gl
W2
I
where
1
wj = 2 j0 wj (9)sin re d9.
Let w = sql + y, where s EIR and y E Y. Then we
can write (6.4.14) in the form
Sgl = A1(E)sg1 + PN(sgl+y)

= By + (I-P)N(sgl+y) (6.4.15)
= 0.
By Theorem 8, (6.4.15) has a centre manifold y = h(s,E),
IsI < 6, IEI < CO. Using h(s,E) = O(s2+JESI), if N2(w)
is the second component of N(w) then

1
N2(sgl+h(s,e))(x) _ 2 [JO s2Tr4cos2Trede]Tr2s sin Trx
+ O(s4 + ics31)
= -s3sin Trx + O(s4 + Ies3I).
Hence PN(sgl+h(s,e)) _ (-s3+O(s4+1Cs31))gl, so that by
Theorem 9, the asymptotic behavior of small solutions of
(6.4.15) is determined by the equation
s - al(e)s - s3 + O(les3I + Is41). (6.4.16)
We can now determine the asymptotic behavior of small solu-
tions of (6.4.14). For 0 < e < 6, solutions of (6.4.14) are
asymptotically stable. For -6 < c<0, the unstable manifold
of the origin consists of two stable orbits connecting two
fixed points to the origin. (See Figure 1.)
Exercise 17. For c = 0, show that equation (6.4.16) can be
written as
s = -s3 - 3s5 + O(s7)
Phase Portrait for Small Negative c
Figure 1
6.4. Examples 127
Example 10. Consider the equation
utt + tut - uxx + a2v + f(u,v) = 0

(6.4.17)
vtt + 2vt - vxx - u + g(u,v) - 0
for (x,t) E (0,Tr) x (O,Tr) with u - v - 0 at x - 0,7r,

where f(u,v), g(u,v) have a second order zero at u = v - 0.
For a = 2, we show that the linearized problem has two purely

imaginary eigenvalues while all the rest have negative real
parts. We then use centre manifold theory to reduce the prob-
lem of bifurcation of periodic solutions to a two-dimensional

problem.
Let w - (u,v,u,S)T, then we can write (6.4.17) as
w = Cw + N(w) (6.4.18)
on Z - (H0(O,Tr))2 x (L2(0,n))2. Let
uxx + av
2
-v - u
xx
If u is an eigenvalue of A then A is an eigenvalue of
C where a2 + 2A + u - 0 and all the eigenvalues of C
arise in this way. An easy calculation shows that the eigen-

values of C are given by
A -1 ± (1-n2tia)1"2, n - 1,2,.... .
For a - 2, the eigenvalues of C are al - i, al -i, while
all the rest have negative real parts. Also
(Re A1(2)) > 0
so that al(a) and a1(a) cross the imaginary axis with non
zero speed. It is now trivial to apply centre manifold theory
to conclude that for a - 2 small, the behavior of small
solutions is determined by an equation of the form
r a 1l
s Is + J(s,a) (6.4.19)
L 1 a
where s E ]R2, a is a real parameter and J(s,a) - O(s2).
To apply the theory in Section 2, Chapter 3, we need to calcu-
late the quadratic and cubic terms in J(s,O). To do this we
need to put (6.4.18) into canonical form and to calculate the
centre manifold when a - 2. From now on we let a - 2.
On the subspace {r sin nx: r E]R4} the operator C
can be represented by the matrix Cn where
0 0 1 0
0 0 0 1
C
n
-n2 -4 -2 0
L 1 -n2 0 -2 J
Note that the eigenvalues of C are given by the eigenvalues
of Cn for n = 1,2,... . To put (6.4.18) into canonical
form we first find a basis which puts C1 into canonical
form. Calculations show that if

-2 0 0
ql 0 q2 1 q3 -1
0 -2 -2
-1 0 2
-2
0
q4
4
L 1J
then Clgl - -q2, Clg2 ql, Clg3 = -2q3 + q4, Clg4 - -q3-2q4.
Let w - Qz where Q - [g1,g2,q3,q4}, then we can rewrite

6.4. Examples 129
(6.4.18) as
i = Q-1CQz + Q-1N(Qz)
(6.4.20)
Let X= s1,s2 E IR}, V =

rl,r2 E]R}, Y = V 0 [X ® V]1 where ,y(x) = sin x. Then
Z = X 0 Y and the projection P on X along Y is given by
wl
w2
Pw =
0
0 J
2
fn
w. = wj(e)sin a de.
1
0
Let z - [s1,s200,0]T + y, si EIR, y E Y, then we can
write (6.4.20) in the form
0 1
Is + <PQ-1N(Qz),4>
-1 0
(6.4.21)
y = By + (I-P)Q 1N(Q z)
where B - (I-P)Q 1CQ, 11,1,0,0]T,y and s = [s1,s21T.
By Theorem 8, (6.4.21) has a centre manifold y h(s).
From w = Qz we have that w1 = -2z1 - 2z4 and
w2 - z2 - z3. Let Q-1 = (tij),
F(z) f(-2z1 - 2z4, z2 - z3)

t13 t14
g(-2z1 - 2z4, z2 z3)
G(z) J L t23 t24 -
Suppose that F(z) a F3(z) + O(IzI 4) and G(z) = G3(z) +

O(1zI4) where F3 and G3 are homogeneous cubics. Then if
J1(s), J2(s) denote the first two components of

on the centre manifold,
r n
J1(s) - n J F3(slsin 6, s2sin 6,0,0)sin 6 d6 + O(Isl 4)
0
with a similar expression for J2(s). Hence, on the centre
manifold
0 1 s1 + [Jl(s)
(6.4.22)
1 0 s2 J2(s)
and we can apply the theory given in Section 2 of Chapter 3.
If the constant K associated with (6.4.22) is zero (see
Section 2 of Chapter 3 for the definition of K) then the
above procedure gives no information and we have to calculate
higher order terms.
If F(z) = F2(z) + F3(z) + O(1z14) and G(z) - G2(z) +
G3(z) + O(Izl 4) where F2 and G2 are homogeneous quad-

ratics, then the calculation of the nonlinear terms is much
more complicated. On the centre manifold, z1 - s1iy + O(s2),
z2 = s2 + O(s2), z3 - O(s2), and z4 = O(s2). The terms of
order s2 make a contribution to the cubic terms in
PQ-1N(Qz). Hence, we need to find a quadratic approximation
to h(s). This is straightforward but rather complicated so
we omit the details.
Example 11. Consider the equations
BA-1u2
ut - Duxx + (B-l)u + A2v + 2Auv + u2v +
BA-1u2
vt = 6Dvxx - Bu - A2v - 2Auv - u2v - (6.4.23)
u = v - 0 at x - 0,1,
where A,B,6,D are positive. The above equations come from
a simplified model of a chemical reaction with u + A and
6.4. Examples 131
v + BA-1 as the chemical concentrations [5,12].
We study (6.4.23) on Z = (H0(011))2. Set
2
d
D + (B-1) A2
dx
W = , C-
C] B 9D
d
2
- A2
dx
N(w) - 1 (2Auv + u2v + BA lug)

C
then we can write (6.4.23) as
w - Cw + N(w). (6.4.24)
We analyze the situation in which for some value of the
parameters A,B,9,D, C has two zero eigenvalues such that
the restriction of C to the zero eigenspace has a Jordan
block. The bifurcation of static solutions where C has two
zero eigenvalues and the restriction of C to the zero eigen-
space is zero, has been studied in [25,26].
On the subspace {r sin nirx, r E R2} the operator C
can be represented by the matrix
rn2n2D + B - 1 A2
C =
n
-B -9Dn2n2 - A2
The eigenvalues of C are given by the eigenvalues of Cn
for n = 1,2,... . We suppose that two of the eigenvalues of
C are zero while all the rest have negative real parts.
For simplicity we assume that the eigenvalues of C1 are zero.
If C1 is to have two zero eigenvalues then

trace(C1) - B - 1 - rr2D - A2 - Oir2D 0

(6.4.25)
det(C1) - A2B - (B-1-ir2D) (A2+Oir2D) = 0.
We make the following hypotheses:
(Hl) There exists AO,BO,9O,D0 such that (6.4.25) is satis-
fied and the real parts of the rest of the eigenvalues
of C are negative.
(H2) For A,B,0,D in a neighborhood of AO,BO,9O,DO, we can
parametrize trace(C1) and det(C1) by
trace(C1) = E2, det(C1) _ -E1. (6.4.26)
The first hypothesis is satisfied, for example, if

D0 > 0 and
00 = 1, A0 ,r2 DO, B0 - (1 + rr2D0)2. (6.4.27)
If we vary B and 0 and keep A - A0, D = DO, then the map-
ping (B,9) + (det(C1), trace(C1)) has a non-zero Jacobian
at B = BOB 0 = 00 if AO,BO,00 are given by (6.4.27), so
by the implicit function theorem, (H2) is satisfied. In
order to simplify calculations we assume (6.4.27) from now on.
Let X = {siy: s i 2 , Y = X1 where iy(x) = sin rrx,

then Z = X ® Y. By Theorem 8, the system
if Cw + N (w)
e= 0
has a centre manifold h: (neighborhood of X x]R2) Y,
where we have written c _ (E1,E2). On the centre manifold,
the equation reduces to
s = C1(E)s + l(siy + h(s,E)) (6.4.28)

6.4. Examples 133
where s = (sl,s21T and
r0 1
Ni(z) = 2 Ni(z(6))sin n9 d9.

1
We treat the linear and nonlinear parts of (6.4.28) separately.

If ql = [1,P]T, q2 = [-p,l]T where p =
then C1(0)g1 = 0 and C2(0)g2 = (A2+B)ql. Let Q = (g1,g2]'

then
0 A2+B
Q 1C1(0)Q =
0 0
Let Q-1C1(e)Q = T = (tip) and let
1 0 1 0
M(e) = = 2 + 0(e)
0
A +B
t11 t12
Then for c sufficiently small, M(c) is nonsingular and
0 1
M(e)Q-1C1(e)QM-1(E)
_
[-detT) trace(T)
0 1 0 1
det(C1) trace(C1) el e2
by (6.4.26). Let s = QM-1(e)r, then (6.4.28) becomes
r 0
I
1 _
r + M(e)Q 1N(QM 1(e)r,
r + h(s,c)). (6.4.29)
L el e2
To check the hypotheses of Section 9 of Chapter 4 we

only need calculate the nonlinear terms when c = 0. Using
the fact that h(s,0) = 0(s2), routine calculations show that
M(0)Q 1N(QM-1(0)r
+ h(s,O)) - [P1,P2ITR(rl,r2)
+ 0(jr1I + I r2I3)
L34 6. INFINITE DIMENSIONAL PROBLEMS
where
j(l+p2)-1(l+p)(A2+B)
p1 - 3n(l+p2)-1(1-P), p2
R(r1,r2) a a1r1 + a2r1r2 + a3r2
BA-1 a2 - 2(A2+B)-1(B3/2A-2+A-BA-1).
al -
2B1/2,
Using (6.4.27), al = (r2D0) 1(1-n4D2) so that a1 is non-
zero if n2D0 + 1. We assume that n2D0 } 1 from now on.

1,
Note also that since 1 + p =-(n2D0) we have that p2
is non-zero.
To reduce (6.4.29) to the form given in Section 9,

Chapter 4,we make the substitution
p - (I-p1p21A(e))r. (6.4.30)
Substituting (6.4.30) into (6.4.29) and using the above cal-
culations we obtain
0 1
p + F(p,c) (6.4.31)
Lel e2
a2F2(0,0) a2F2(0,0)
a = ate- a B -
ap- p - = 2alp1 + a2p2
p
1
-=0
a2F1(0,0)
apl
We have already checked that a + 0. For most values
of DO,B is nonzero [B is only zero when D0 is a solu-
tion of a certain algebraic equation]. If a + 0 then we
can apply the theory given in Section 9, Chapter 4 to obtain
the bifurcation set for (6.4.23). If B - 0 then the theory
given in Section 9, Chapter 4 still gives us part of the

6.4. Examples 135
bifurcation set; the full bifurcation set would depend on
higher order terms.
Remark. If we vary 9 in (6.4.23) the theory given in Sec-
tion 3 does not apply since the map (9,v) - is not

evxx
even defined on the whole space. However, it is easy to
modify the results of Section 3 to accommodate the above sit-
uation. (See, for example, Exercises 1-2 in Section 3.4 of
[34].)
REFERENCES
1] R. Abraham and J. Marsden, Foundations of Mechanics,

2nd ed., Benjamin, New York (1978).
:2] R. A. Adams, Sobolev Spaces, Academic Press, New York

(1975).
[3] V. I. Arnold, Bifurcations in versal families, Russian

Math. Surveys 27, 54-123 (1972).
[4] V. I. Arnold, Loss of stability of self-oscillations

close to resonance and versal deformations of equivari-
ant vector fields, Functional Analysis and its Appl. 11,
No. 2, 1-10 (1977).
[5] J. F. G. Auchmuty and G. Nicolis, Bifurcation analysis

of nonlinear reaction-diffusion equations, Bull. Math.
Biology 37, 323-365 (1975).
[6] A. V. Balakrishnan, A lied Functional Analysis, Applica-

tions of Mathematics, Vol. pringer er ag, Berlin
and New York (1976).
[7] J. M. Ball, Stability theory for an extensible beam,

J. Diff. Egns. 14, 399-418 (1973).
[8] J. M. Ball, Saddle point analysis for an ordinary dif-

ferential equation in a Banach space and an application
to dynamic buckling of a beam, Nonlinear Elasticity,
J. Dickey (ed.), Academic Press, New York
[9] J. M. Ball, Strongly continuous semigroups, weak solu-
tions, and the variation of constants formula, Proc.
Amer. Math. Soc. 63, 370-373 (1977).
[10] J. M. Ball and J. Carr, Decay to zero in critical cases

of second order ordinary differential equations of
Duffing type, Arch. Rat. Mech. Anal. 63, 47-57 (1976).
136
References 137
[11] M. Berger, Nonlinearity and Functional Analysis,

Academic Press, New or
[12] J. Boa and D. S. Cohen, Bifurcation of localized dis-
turbances in a model biochemical reaction, SIAM J. Appl.
Math. 30, 123-135 (1976).
[13] J. Carr and Naji-Al-Amood, Rate of decay in critical

cases, J. Math. Anal. and Appl. 75, 242-250; II, Ibid.,
293-300
[14] J. Carr and R. G. Muncaster, The application of centre

manifolds to amplitude expansions, submitted to SIAM J.
Appl. Math.
[15] J. Carr and M. Z. M. Malhardeen, Beck's Problem, SIAM

J. Appl. Math. 37, 261-262 (1979).
[16] J. Carr and M. Z. M. Malhardeen, Stability of noncon-

servative linear systems, S rin er Lecture Notes in
Math., Vol. 799, Springer er ag, Berlin and New York
MO).
[17] N. Chafee, The bifurcation of one or more closed orbits
from an equilibrium point of an autonomous differential
system, J. Diff. Egns. 4, 661-679 (1968).
[18] N. Chafee, A bifurcation problem for a functional dif-
ferential equation of finitely retarded type, J. Math.
Anal. Appl. 35, 312-348 (1971).
[19] S. N. Chow and J. Mallet-Paret, Integral averaging and
bifurcation, J. Diff. Egns. 26, 112-159 (1977).
[20] S. N. Chow and J. Mallet-Paret, Fuller index and Hopf
bifurcation, J. Diff. Egns. 29, 66-85 (1978).
[21] S. N. Chow, J. Mallet-Paret, and J. K. Hale, Applications
of generic bifurcation, I, Arch. Rat. Mech. Anal. 59,
159-188 (1975); II, Ibid. 62, (1976).
[22] E. A. Coddington and N. Levinson, Theory of Ordinar
Differential Equations, McGraw-Hill, New York (1955).
[23] N. Fenichel, Persistence and smoothness of invariant
manifolds for flows, IndianaUniv. Math. J. 21, 193-226
(1971).
[24] N. Fenichel, Geometric singular perturbation theory for
ordinary differential equations, J. Diff. Egns. 31,
53-98 (1979).
[25] M. Golubitsky and D. Schaeffer, A theory for imperfect

bifurcation via singularity theory, Commun. Pure Appl.
Math. 32, 21-98 (1979).
[26] M. Golubitsky and D. Schaeffer, Bifurcation analysis

near a double eigenvalue of a model chemical reaction,
MRC Rep. 1859, Univ. of Wisconsin (1978).
58 REFERENCES
27] J. K. Hale, Ordinary Differential Equations, Wiley-

Interscience, New York 1
28] J. K. Hale, Critical cases for neutral functional dif-

ferential equations, J. Diff. Egns. 10, 59-82 (1970).
29] J. K. Hale, Generic bifurcation with applications,

Nonlinear Analysis and Mechanics: Heriot-Watt Symposium
Ced. R. J. Knops) Vol. 1, Pitman, London (1977).
30] P. Hartman, Ordinary Differential Equations, Second

Edition, P. artman
31] P. Hassard and Y. H. Wan, Bifurcation formulae derived

from center manifold theory, J. Math. Anal. Appl. 63(1),
297-312 (1978).
32] A. R. Hausrath, Stability in the critical case of pure

imaginary roots for neutral functional differential
equations, J. Diff. Egns. 13, 329-357 (1973).
33] F. G. Heinenken, H. M. Tsuchiya and R. Aris, On the

mathematical status of the pseudo-steady state hypothe-
sis of biochemical kinetics, Math. Biosci. 1, 95-113
(1967).
:34] D. Henry, Geometric Theor of Parabolic Equations, to

be published by pringer er ag.
:35] M. Hirsch, C. Pugh and M. Schub, Invariant Manifolds,

Springer Lecture Notes in Math., o. 5B3, Springer-
Verlag, Berlin and New York (1977).
[36] P. J. Holmes, Bifurcations to divergence and flutter in

flow induced oscillations - a finite dimensional analy-
sis, J. Sound and Vib. 53, 471-503 (1977).
[37] P. J. Holmes, Coupled flutter and divergence in auto-

nomous nonlinear systems, to appear in A.S.M.E.
Symposium on Nonlinear Dynamics.
[38] P. J. Holmes and J. E. Marsden, Bifurcations to diver-

gence and flutter in flow induced oscillations - an
infinite dimensional analysis, Automatica 14(4), 367-
384 (1978).
[39] E. Hopf, Abzweigung einer periodischen Losung von einer

stationaren Losung eines Differential systems, Ber.
h. Phys., K1. Sachs Akad. Wiss. Leipzig 94, T2
Mat_
[40] G. Iooss, Bifurcation of Maps and A lications, North-

Holland Mathematics Studies, Vol. North-Holland
Publishing Company, Amsterdam, New York and Oxford
(1979).
References 139
[41) G. Iooss and D. D. Joseph, Bifurcation and stability

of nT-periodic solutions branching from T-periodic
solutions at points of resonance, Arch. Rat. Mech. Anal.
66, 2, 135-172 (1977).
[42) M. C. Irwin, On the stable manifold theorem, Bull.

London Math. Soc. 2, 196-198 (1970).
[43) T. Kato, Perturbation Theory for Linear 0 erators,
Springer er ag, Berlin and New or
[44) A. Kelley, The stable, center-stable, center, center-

unstable, and unstable manifolds, J. Diff. Egns. 3,
546-570 (1967).
[45) A. Kelley, Stability of the center-stable manifold,

J. Math. Anal. Appl. 18, 336-344 (1967).
[46) W. Klingenberg, Lectures on Closed Geodesics, Springer-

Verlag, Berlin and New York
[47) N. Kopell and L. N. Howard, Bifurcations and trajector-
ies joining critical points, Adv. Math. 18, 306-358
(1976).
[48) 0. E. Langford, Bifurcation of periodic solutions into

invariant tori: the work of Ruelle and Takens,
S rin er Lectures Notes in Mathematics, Vol. 322,
pp. 159-19Z, pringer- er ag, Berlin and New York (1972).
[49) J. E. Marsden, Qualitative methods in bifurcation
theory, Bull. Amer. Math. Soc. 84, 1125-1147 (1978).
[50) J. E. Marsden and T. J. R. Hughes, Topics in the mathe-
matical foundations of elasticity, Nonlinear Anal sis
and Mechanics: Heriot-Watt Symposium (ed. Knops)
Vol. itman, London (1978).
[51) J. E. Marsden and M. F. McCracken, The Hopf Bifurcation
and Its A lications, Applied Math. sciences, Vol. 19,
pringer er ag, Berlin and New York (1976).
[52) S. J. Merrill, A model of the stimulation of B-cells by

replicating antigen, I., Math. Biosci. 41, 125-141
(1978) ; II. Ibid., 41, 143-155 (1978). -
[53) J. Moser, Proof of a generalized form of a fixed point
theorem, S rin er Lecture Notes in Math., Vol. 597,
Springer- er ag, Berlin and New York 977).
[54) A. I. Neishtadt, Bifurcation of the phase pattern of an

equation system arising in the problem of stability
loss of selfoscillations close to 1 : 4 resonance.
P.M.M. 42, 896-907 (1978).
[55] A. Pazy, Semigroups of linear operators and applica-

tions to partial differential equations, University of
Maryland Lecture Notes, Vol. 10, (1974).
40 REFERENCES
56] M. Reed and B. Simon, Methods of Modern Mathematical

Physics, Vol. II: Fourier Analysis, a -a ointness,
Academic Press, New or (1975).
57] D. Ruelle and F. Takens, On the nature of turbulence,

Comm. Math. Phys. 20, 167-192 (1971).
:58] R. J. Sacker, On invariant surfaces and bifurcation of

periodic solutions of ordinary differential equations,
New York Univ. IMM-NYU, 333 (1964).
[59] D. H. Sattinger, Topics in Stability and Bifurcation

The pry, Springer Lecture Notes in Math., Vol. ,
pier-Verlag, Berlin and New York (1973).

[60] J. T. Schwartz, Nonlinear Functional Analysis, Gordon
and Breach Science Publishers, New York (1969).
[61] I. Segal, Nonlinear semigroups, Ann. Math. 78, 339-364

(1963).
[62] Y. Shitzuta, On the classical solutions of the Boltz-
mann equation, to appear in Commun. Pure Appl. Math.
[63] S. J. van Strien, Center manifolds are not C", Math. Z.

166, 143-145 (1979).
[64] F. Takens, Forced Oscillations and Bifurcations,

Communication Math. Institute, ij suniversiteit,
Utrecht (1974).
[65] F. Takens, Singularities of vector fields, Publ. Inst.

Hautes Etudes Sci. 43, 47-100 (1974).
[66] D. A. Vaganov, N. G. Samoilenko and V. G. Abramov,

Periodic regimes of continuous stirred tank reactors,
Chem. Engrg. Sci. 33, 1133-1140 (1978).
[67] I. Vidav, Spectra of perturbed semigroups with applica-

tions to transport theory, J. Math. Anal. Appl., 30,
264-279 (1970).
[68] Y. H. Wan, Bifurcation into invariant tori at points of

resonance, Arch. Rat. Mech. Anal. 68, 343-357 (1978).
[69] J. Zabczyk, A note on Co semigroups, Bulletin de

L'Academie Polonaise des Sciences, Serie e-3 ssciences
mat aastr. et p ys., Z3, 895-898 (197S).
INDEX
Analytic semigroup, 105, 106, Contraction mapping, 14, 17, 19,

110 22, 24-26, 102-104, 108-109
Asymptotic behavior, 4, 8-12 Coupled equations, 1

19, 29, 32, 37-39, 46, 89,
114-115, 118, 122-123, 126 Dissipative operator, 108-109
128
Eigenvalues, 3, 7, 11-13, 16,
Asymptotic stability, 1, 2, 18, 20, 25, 34, 40, 44, 46-47,
4, 6-8, 21, 35-36, 118, 50, 53-55, 89-91, 98, 109,
122, 124, 126 111-112, 114-115, 117, 120,
124-125, 127-128, 131-132
Averaging, 42
Equilibrium point, 8, 13, 28,
Autonomous equations, 13 55, 57, 59, 65-67, 69, 73-75,
82, 84, 126, 131
Bifurcation (see also
Hopf bifurcation), 11-12, Evolution equation, 13, 97
40-41, 43, 47, 50, 52-53,
55-59, 82, 84-85, 94, 127, First integral, 65
131, 134-135
Generator, 97, 99-134
Cauchy problem, 101
Geodesics, 53
Center manifold
Gronwall's inequality, 18, 20,
approximation of, 2, 5-7, 23-24, 98
9, 13, 25, 32, 35, 38,
49, 120, 122-123 Group, 106-107, 113-114, 124
definition, 3 Hamiltonian system, 53

existence of, 1-4, 6-7, Hille-Yosida theorem, 102-104,
9, 12-17, 29-34, 38, 46, 107, 109
89, 93, 97, 117-118, 122,
124-125, 127-129, 132 Homoclinic orbit, 57, 67-68, 74-
75, 80, 82, 84
flow on, 2-6, 8, 10, 12,
19, 21-22, 29, 32, 35, Hopf bifurcation, 39-50
38, 46-47, 89, 92, 118-
119, 126, 128-129, 132 Implicit function theorem, 40-
41, 43, 48, 52, 56, 67-68,
properties of, 19-20, 28-30, 70, 73, 94, 132
95, 118
Infinitesimal generator (see
Chemical reactions, 8, 12, generator)
130-131
Instability, 2, 4, 6-7, 21, 35,
Closed graph theorem, 106 43-44, 115, 118
Compact operator, 111-112, Invariance of domain theorem, 24

114
Invariant manifold, 2-3, 13, 16,
Continuation, 116 30-34
141
.42 INDEX
Invariant set, 13, 33-34, Separatrix, 8

51-53, 117
Skew-selfadjoint operator, 107,
Jacobian, 3, 93-94, 132 113, 120
Laplace transform, 102 Spectral radius, 110
Liapunov function, 37 Spectrum, 109-113, 115, 117
Liapunov-Schmidt procedure, Stability, 2, 4, 6, 8, 19, 21-

12 22, 35, 43-44, 50, 57, 118
Linearization, 7, 11-13, 15, Stable manifold, 3, 8, 13, 15,

34, 37, 40, 44, 46-47, 50, 67
53-54, 57, 71, 81-82, 88,
92-94, 127 Strong solution, 104-105
Lumer-Phillips theorem, 108 Symplectic mapping, 53
Neutral functional differen- Uncoupled equations, 1

tial equations, 13
Uniform boundedness theorem,
Nilpotent, 20 100
Normal form, S2, 56, 89 Unstable manifold, 8, 13, 67,

82, 126
Parabolic equations, 105
Weak solution, 105, 116
Partial differential equa-
tions, 13, 98, 102
Periodic solutions, 28, 33,

40-45, 47, 50-51, 55, 57,
65-67, 69-71, 73-80, 82,
84, 127
Perturbation, 9, 12, 30, 33,

44-45, 113-114
Phase portraits, 11-13, 54,

60-63, 80, 82, 84-87, 126
Poincare map, 33
Rate of decay, 2, 4, 10, 37,

124
Reduction of dimension, 12,

19, 89
Resolvent, 102-103
Saddle point, 67
Selfadjoint operator, 104-

105, 107, 113, 120
Semigroup, 97-134

(Applied Mathematical Sciences 35) Jack Carr (Auth.) - Applications of Centre Manifold Theory-Springer-Verlag New York (1981)

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Applied Mathematical Sciences 35) Jack Carr (Auth.) - Applications of Centre Manifold Theory-Springer-Verlag New York (1981)

Uploaded by

Copyright:

Available Formats

Applied Mathematical Sciences I Volume 35

Library of Congress Cataloging in Publication Data

(Applied mathematical sciences; v. 35)

All rights reserved.

No part of this book may be translated or reproduced in any

The use of general descriptive names, trade names, trademarks,

© 1981 by Springer-Verlag New York Inc.

Printed in the United States of America

These notes are based on a series of lectures given in

the Lefschetz Center for Dynamical Systems in the Division of

The purpose of the lectures was to give an introduction

to the applications of centre manifold theory to differential

The main application of centre manifold theory given

bifurcation theory is concerned with topological changes in

critical value. In certain circumstances, the application of

centre manifold theory reduces the dimension of the system

theory plays the same role for dynamic problems as the

Liapunov-Schmitt procedure plays for the analysis of static

In order to make these notes more widely accessible,

voted to this. Once the finite dimensional case is under-

simplest such theory, for example our equations are autono-

mous. Once the core of an idea has been understood in a

simple setting, generalizations to more complicated situations

are much more readily understood.

In Chapter 1, we state the main results of centre mani-

their use by a few simple examples. In Chapter 2, we prove

the theorems which were stated in Chapter 1, and Chapter 3

line Hopf bifurcation theory for 2-dimensional systems. In

Section 3 of Chapter 3 we apply this theory to a singular per-

Chapter 6 we apply the same theory to a system of partial dif-

ferential equations. In Chapter 4 we study a dynamic bifurca-

tion problem in the plane with two parameters. Some of the

results in this chapter are new and, in particular, they con-

firm a conjecture of Takens [64]. Chapter 4 can be read in-

dependently of the rest of the notes. In Chapter 5, we apply

the theory of Chapter 4 to a 4-dimensional system. In Chap-

ter 6, we extend the centre manifold theory given in Chapter

2 to a simple class of infinite dimensional problems. Fin-

ally, we illustrate their use in partial differential equa-

tions by means of some simple examples.

I first became interested in centre manifold theory

through reading Dan Henry's Lecture Notes [34]. My debt to

these notes is enormous. I would like to thank Jack K. Hale,

Dan Henry and John Mallet-Paret for many valuable discussions

during the gestation period of these notes.

CHAPTER 1. INTRODUCTION TO CENTRE MANIFOLD THEORY . . . 1

CHAPTER 2. PROOFS OF THEOREMS . . . . . . . . . . . . . 14

3.1. Rate of Decay Estimates in Critical Cases. . 37

CHAPTER 4. BIFURCATIONS WITH TWO PARAMETERS IN TWO

CHAPTER S. APPLICATION TO A PANEL FLUTTER PROBLEM . . . 88

CHAPTER 6. INFINITE DIMENSIONAL PROBLEMS. . . . . . . . 97

In this chapter we state the main results of centre

manifold theory for finite dimensional systems and give some

To motivate the study of centre manifolds we first

x = ax3, y = -y + y2, (1.2.1)

where a is a constant. Since the equations are uncoupled

Since the equations are coupled we cannot immediately decide

if the zero solution of (1.2.2) is asymptotically stable, but

we might suspect that it is if a < 0. The key to understand-

an abstraction of the idea of uncoupled equations.

A curve, say y = h(x), defined for lx1 small, is

said to be an invariant manifold for the system of differen-