You are on page 1of 265

O rdinary Differential

Equations
Qualitative Theory

Luis B a rre ira


C la u d ia V ails

G raduate Studies
in M athem atics
V o lu m e 137

A m e rican M ath e m atical So cie ty


Ordinary Differential
Equations
Qualitative Theory
Ordinary Differential
Equations
Qualitative Theory

Luis Barreira
Claudia Vails

Translated by the authors

Graduate Studies
in Mathematics
Volume 137

^/^PHIOIM
HVS>
American Mathematical Society
y Providence, Rhode Island
EDITORIAL COMMITTEE
David Cox (Chair)
Daniel S. Freed
Rafe Mazzeo
Gigliola Staffilani
This work was originally published in Portuguese by 1ST Press under the title Equagoes
Diferenciais: Teoria Qualitativa by Luis Barreira and Claudia Vails, 1ST Press 2010,
Institute Superior Tecnico. All Rights Reserved.
The present translation was created under license for the American Mathematical
Society and published by permission.

Translated by the authors.

2010 Mathematics Subject Classification. Primary 34-01, 34Cxx, 34Dxx, 37Gxx, 37Jxx.

For additional information and updates on this book, visit


w w w .a m s .o r g /b o o k p a g e s / g sm -1 3 7

L ib ra ry o f C on g ress C a ta lo g in g -in -P u b lica tio n D a ta


Barreira, Luis, 1968-
[Equagoes diferenciais. English]
Ordinary differential equations : qualitative theory / Luis Barreira, Claudia Vails ; translated
by the authors.
p. cm. - (Graduate studies in mathematics ; v. 137)
Includes bibliographical references and index.
ISBN 978-0-8218-8749-3 (alk. paper)
1. Differential equations-Qualitative theory. I. Vails, Claudia, 1973- II. Title.

QA372.B31513 2010
515'.352-dc23 2012010848

C o p y in g and reprin tin g. Individual readers of this publication, and nonprofit libraries
acting for them, are permitted to make fair use of the material, such as to copy a chapter for use
in teaching or research. Permission is granted to quote brief passages from this publication in
reviews, provided the customary acknowledgment of the source is given.
Republication, systematic copying, or multiple reproduction of any material in this publication
is permitted only under license from the American Mathematical Society. Requests for such
permission should be addressed to the Acquisitions Department, American Mathematical Society,
201 Charles Street, Providence, Rhode Island 02904-2294 USA. Requests can also be made by
e-mail to reprint-perm ission@am s.org.
2012 by the American Mathematical Society. All rights reserved.
The American Mathematical Society retains all rights
except those granted to the United States Government.
Printed in the United States of America.
0 The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at h ttp : //www. ams. o r g /
10 9 8 7 6 5 4 3 2 1 17 16 15 14 13 12
Contents

Preface IX

Part 1. Basic Concepts and Linear Equations

Chapter 1. Ordinary Differential Equations 3


1. 1. Basic notions 3
1.2. Existence and uniqueness of solutions 9
1.3. Additional properties 21
1.4. Existence of solutions for continuous fields 32
1.5. Phase portraits 35
1. 6. Equations on manifolds 48
1.7. Exercises 53

Chapter 2. Linear Equations and Conjugacies 57


2. 1. Nonautonomous linear equations 57
2.2. Equations with constant coefficients 63
2.3. Variation of parameters formula 75
2.4. Equations with periodic coefficients 78
2.5. Conjugacies between linear equations 85
2 .6 . Exercises 97

Part 2. Stability and Hyperbolicity

Chapter 3. Stability and Lyapunov Functions 105


3.1. Notions of stability 105
VI Contents

3.2. Stability of linear equations 108


3.3. Stability under nonlinear perturbations 113
3.4. Lyapunov functions 116
3.5. Exercises 123

Chapter 4. Hyperbolicity and Topological Conjugacies 127


4.1. Hyperbolic critical points 127
4.2. The Grobman-Hartman theorem 129
4.3. Holder conjugacies 139
4.4. Structural stability 141
4.5. Exercises 143

Chapter 5. Existence of Invariant Manifolds 147


5.1. Basic notions 147
5.2. The Hadamard-Perron theorem 149
5.3. Existence of Lipschitz invariant manifolds 150
5.4. Regularity of the invariant manifolds 157
5.5. Exercises 167

Part 3. Equations in the Plane

Chapter 6. Index Theory 171


6.1. Index for vector fields in the plane 171
6.2. Applications of the notion of index 176
6.3. Index of an isolated critical point 179
6.4. Exercises 181

Chapter 7. Poincare-Bendixson Theory 185


7.1. Limit sets 185
7.2. The Poincare-Bendixson theorem 190
7.3. Exercises 196

Part 4. Further Topics

Chapter 8. Bifurcations and Center Manifolds 201


8.1. Introduction to bifurcation theory 201
8.2. Center manifolds and applications 206
8.3. Theory of normal forms 215
8.4. Exercises 222
Contents Vll

Chapter 9. Hamiltonian Systems 225


9.1. Basic notions 225
9.2. Linear Hamiltonian systems 229
9.3. Stability of equilibria 231
9.4. Integrability and action-angle coordinates 235
9.5. The KAM theorem 239
9.6. Exercises 240

Bibliography 243

Index 245
Preface

The main objective of this book is to give a comprehensive introduction


to the qualitative theory of ordinary differential equations. In particular,
among other topics, we study the existence and uniqueness of solutions,
phase portraits, linear equations and their perturbations, stability and Lya
punov functions, hyperbolicity, and equations in the plane.
The book is also intended to serve as a bridge to important topics that
are often left out of a second course of ordinary differential equations. Exam
ples include the smooth dependence of solutions on the initial conditions,
the existence of topological and differentiable conjugacies between linear
systems, and the Holder continuity of the conjugacies in the Grobman-
Hartman theorem. We also give a brief introduction to bifurcation theory,
center manifolds, normal forms, and Hamiltonian systems.
We describe mainly notions, results and methods that allow one to dis
cuss the qualitative properties of the solutions of an equation without solving
it explicitly. This can be considered the main aim of the qualitative theory
of ordinary differential equations.
The book can be used as a basis for a second course of ordinary differen
tial equations. Nevertheless, it has more material than the standard courses,
and so, in fact, it can be used in several different ways and at various levels.
Among other possibilities, we suggest the following courses:

a) advanced undergraduate/beginning graduate second course: Chap


ters 1-5 and 7-8 (without Sections 1.4, 2.5 and 8.3, and without the
proofs of the Grobman-Hartman and Hadamard-Perron theorems);
b) advanced undergraduate/beginning graduate course on equations in
the plane: Chapters 1-3 and 6-7;

IX
Preface

c) advanced graduate course on stability: Chapters 1-3 and 8-9;


d) advanced graduate course on hyperbolicity: Chapters 1-5.
Other selections are also possible, depending on the audience and on the
time available for the course. In addition, some sections can be used for
short expositions, such as Sections 1.3.2, 1.4, 2.5, 3.3, 6.2 and 8.3.
Other than some basic pre-requisites of linear algebra and differential
and integral calculus, all concepts and results used in the book are recalled
along the way. Moreover, (almost) everything is proven, with the excep
tion of some results in Chapters 8 and 9 concerning more advanced topics
of bifurcation theory, center manifolds, normal forms and Hamiltonian sys
tems. Being self-contained, the book can also serve as a reference or for
independent study.
Now we give a more detailed description of the contents of the book.
Part 1 is dedicated to basic concepts and linear equations.
In Chapter 1 we introduce the basic notions and results of the the
ory of ordinary differential equations, in particular, concerning the
existence and uniqueness of solutions (Picard-Lindelof theorem) and
the dependence of solutions on the initial conditions. We also estab
lish the existence of solutions of equations with a continuous vector
field (Peanos theorem). Finally, we give an introduction to the de
scription of the qualitative behavior of the solutions in the phase
space.
In Chapter 2 we consider the particular case of (nonautonomous)
linear equations and we study their fundamental solutions. It is of
ten useful to see an equation as a perturbation of a linear equation,
and to obtain the solutions (even if implicitly) using the variation
of parameters formula. This point of view is often used in the book.
We then consider the particular cases of equations with constant co
efficients and equations with periodic coefficients. More advanced
topics include the dependence of solutions on the initial con
ditions and the existence of topological conjugacies between linear
equations with hyperbolic matrices of coefficients.
Part 2 is dedicated to the study of stability and hyperbolicity.
In Chapter 3, after introducing the notions of stability and asymp
totic stability, we consider the particular case of linear equations, for
which it is possible to give a complete characterization of these no
tions in terms of fundamental solutions. We also consider the partic
ular cases of equations with constant coefficients and equations with
periodic coefficients. We then discuss the persistence of asymptotic
stability under sufficiently small perturbations of an asymptotically
Preface XI

stable linear equation. We also give an introduction to the theory of


Lyapunov functions, which sometimes yields the stability of a given
solution in a more or less automatic manner.
In Chapters 4-5 we introduce the notion of hyperbolicity and we
study some of its consequences. Namely, we establish two key re
sults on the behavior of the solutions in a neighborhood of a hyper
bolic critical point: the Grobman-Haxtman and Hadamard-Perron
theorems. The first shows that the solutions of a sufficiently small
perturbation of a linear equation with a hyperbolic critical point are
topologically conjugate to the solutions of the linear equation. The
second shows that there are invariant manifolds tangent to the sta
ble and unstable spaces of a hyperbolic critical point. As a more ad
vanced topic, we show that all conjugacies in the Grobman-Hartman
theorem are Holder continuous. We note that Chapter 5 is more tech
nical: the exposition is dedicated almost entirely to the proof of the
Hadamard-Perron theorem. In contrast to what happens in other
texts, our proof does not require a discretization of the problem or
additional techniques that would only be used here. We note that
the material in Sections 5.3 and 5.4 is used nowhere else in the book.

In Part 3 we describe results and methods that are particularly useful in the
study of equations in the plane.
In Chapter 6 we give an introduction to index theory and its ap
plications to differential equations in the plane. In particular, we
describe how the index of a closed path with respect to a vector field
varies with the path and with the vector field. We then present sev
eral applications, including a proof of the existence of a critical point
inside any periodic orbit, in the sense of Jordans curve theorem.
In Chapter 7 we give an introduction to the Poincare-Bendixson
theory. After introducing the notions of a-limit and w-limit sets,
we show that bounded semiorbits have nonempty, compact and con
nected a-limit and w-limit sets. Then we establish one of the impor
tant results of the qualitative theory of ordinary differential equa
tions in the plane, the Poincare-Bendixson theorem. In particular,
it yields a criterion for the existence of periodic orbits.

Part 4 is of a somewhat different nature and it is only here that not every
thing is proved. Our main aim is to make the bridge to important topics
that are often left out of a second course of ordinary differential equations.

In Chapter 8 we give an introduction to bifurcation theory, with


emphasis on examples. We then give an introduction to the theory
of center manifolds, which often allows us to reduce the order of an
XU Preface

equation in the study of stability or the existence of bifurcations. We


also give an introduction to the theory of normal forms that aims
to eliminate through a change of variables all possible terms in the
original equation.
Finally, in Chapter 9 we give an introduction to the theory of Hamil
tonian systems. After introducing some basic notions, we describe
several results concerning the stability of linear and nonlinear Hamil
tonian systems. We also consider the notion of integrability and the
Liouville-Arnold theorem on the structure of the level sets of inde
pendent integrals in involution. In addition, we describe the basic
ideas of the KAM theory.
The book also includes numerous examples that illustrate in detail the new
concepts and results as well as exercises at the end of each chapter.

Luis Barreira and Claudia Vails


Lisbon, February 2012
Part 1

Basic Concepts and


Linear Equations
Chapter 1

Ordinary Differential
Equations

In this chapter we introduce the basic notions of the theory of ordinary


differential equations, including the concepts of solution and of initial value
problem. We also discuss the existence and uniqueness of solutions, their
dependence on the initial conditions, and their behavior at the endpoints
of the maximal interval of existence. To that effect, we recall the relevant
material concerning contractions and fiber contractions. Moreover, we show
how the solutions of an autonomous equation may give rise to a flow. Finally,
we give an introduction to the qualitative theory of differential equations,
with the discussion of many equations and of their phase portraits. We
include in this study the particular case of the conservative equations, that
is, the equations with an integral. For additional topics we refer the reader
to [2, 9, 13, 15].

1.1. Basic notions

In this section we introduce the notions of a solution of an ordinary differ


ential equation and of initial value problem, and we illustrate them with
various examples. We also show how the solutions of an autonomous equa
tion (that is, an equation not depending explicitly on time) may give rise to
a flow.

1.1.1. Solutions and initial value problem s. We first introduce the


notion of a solution of an ordinary differential equation. Given a continuous
function / : ) ^ R in an open set > C R x R , consider the (ordinary)
1. Ordinary DiSerential Equations

differential equation
x' = f{t,x). (1.1)
The unknown of this equation is the function x = x{t).

D efinition 1.1. A function x: (a,b) )M of class (with a > oo and


b < +oo) is said to be a solution of equation (1.1) if (see Figure 1.1):
a) {t,x{t)) 6 D for every t 6 (a, 6);
b) x'{t) = f{t,x{t)) for every t (a, 6).

F ig u re 1.1. A solution x = x{t) o f the equation x' = f { t , x).

Exam ple 1.2. Consider the equation


x' X + 1. (1.2)
If X = x{t) is a solution, then
(e*x)' = e*x + e^x' = e*{x + x') = e*t.
Since a primitive of is e*(t 1), we obtain
e*x(t) = e*(f 1) + c
for some constant c R. Therefore, the solutions of equation (1.2) are given
by
x{t) = t 1 + ce *, t G R, (1.3)
with c G R.

Now we consider an equation in R^.

Exam ple 1.3. Consider the equation


{x,y)' = (y, - x ) , (1.4)
which can be written in the form
\x' = y,
(1.5)
1 y' = -X .
1.1. Basic notions

If {x{t),y{t)) is a solution, then


(x^ + y^Y = 2xx' + 2yy' = 2xy + 2 y{ -x ) = 0. (1.6)
Thus, there exists a constant r > 0 such that

+ y(i)^ =
(for every t in the domain of the solution). Writing
x{t) = rcosO{t) and y(t) = rsin0(t),
where 0 is a differentiable function, it follows from the identity x' = y that
x'{t) = r9'{t)smd{t) = rsin^(t).
Hence, 0'{t) = 1, and there exists a constant c M such that 9{t) = t+c.
Therefore,
(a;(t), y(t)) = (r cos(t + c ),r sin(t + c)), t 6 M, (1.7)
with c R. These are all the solutions of equation (1.4).

E xam ple 1.4. It follows from (1.5) that


X = [X ) = y = X.
On the other hand, starting with the equation x" = x and writing y x',
we obtain y' = x" = x. In other words, we recover the two identities
in (1.5). Hence, the equations (1.5) and x" = x are equivalent.
More generally, writing
X = ( X i,...,X f c ) = (x ,a :',...,a ;(''-i)),
the equation
xW = /(t ,x ,a ;',...,x ( ^ - i))
can be written in the form X' = F(t, X), where
F {t ,X ) = { X 2 , X 3 , . . . , X k - i J { t , X ) ) .

One can also consider differential equations written in other coordinates,


such as, for example, in polar coordinates.

E xam ple 1.5. Consider the equation

(r' = r,
( 1.8)
]9' = 1
written in polar coordinates (r,9). Since x = rcos9 and y = rsin^, we
obtain
x' = r' cos 9 r9' sin ^ = r cos 9 rsin.9 = x y (1.9)
and
y' = r'sin^ + r9' cos 9 = rsin9 + rcos9 = y + x. (1-10)
1. Ordinary Differential Equations

Thus, equation (1.8) can be written in the form

f x' = x - y ,
( 1.11)
(y' = a: + y.
Incidentally, one can solve separately each equation in (1.8) to obtain the
solutions
f r(t) = ce\
( 1.12)
^0(f) = t d,
where c, d G R are arbitrary constants.

Now we introduce the notion of initial value problem.

D efinition 1.6. Given {tQ,xo) D, the initial value problem

(x' = f{t,x),
(1.13)
|a;(to) = *0
consists of finding an interval (a, 6) containing to and a solution x : (a, b)
R" of equation (1.1) such that x{to) = xq. The condition x{to) = xq is called
the initial condition of problem (1.13).

Exam ple 1.7 (continuation of Example 1.2). Consider the initial value
problem
\x' = X + 1,
(1.14)
{:
[a;(0) = 0.
Taking t = 0 in (1.3), we obtain a;(0) = 1 + c, and the initial condition
yields c = 1. Hence, a solution of the initial value problem (1.14) is
x(t) = t 1 + e tG R.

Exam ple 1.8 (continuation of Example 1.3). Consider equation (1.4) and
the corresponding initial value problem

f(a;,y)' = (l/.
(1.15)
\{x{tQ),y(to)) = {xo,yo),

where to R and (a;o,yo) G Taking t = to in (1.7), we obtain


(o, yo) = {r cos(-to + c), r sin(-to + c )) .
Writing the initial condition in polar coordinates, that is,
(a:o,2/o) = ('cos0o,rsin6'o), (1.16)
we obtain c = t o+ 9o (up to an integer multiple of 27t). Hence, a solution of
the initial value problem (1.15) is
{x{t),y{t)) = (r cos{t + to + 6o),rsin{t + to + Oq)), t G R. (1-17)
1.1. Basic notions

An equivalent and more geometric description of the solution (1.17) of


problem (1.15) is the following. By (1.16) and (1.17), we have
x{t) = r cos $0 cos{t + to) r sin Oqsin(t + to)
= xo cos(t - to) + yo sin(t - to),
and analogously (or simply observing that y x'),
y{t) = -Xo sin(t - to) + yo cos(t - to).
Thus, one can write the solution (1.17) in the form
^Xo''
(1.18)

where
cost sint'
R{t) =
sint costy
is a rotation matrix. Indeed, as shown in Example 1.3, each solution of
equation (1.4) remains at a fixed distance from the origin.

The following is a characterization of the solutions of an ordinary differ


ential equation.

P rop osition 1.9. Let / : H > be a continuous function in an open set


D c R X ]R^. Given (to,xo) E D, a continuous function x: (a, 6) R
in an open interval {a,b) containing to is a solution of the initial value
problem (1.13) if and only if

x(t) = Xo -H / f{s,x{s)) ds (1.19)


Jto
for every t (a, b).

P roof. We first note that the function 1 / ( t , x(t)) is continuous, because


it is a composition of continuous functions. In particular, it is also integrable
in each bounded interval. Now let us assume that x = x(t) is a solution of
the initial value problem (1.13). For each t (a, 6), we have

x{t) Xo = x{t) x(to) = I x'{s) ds = I f{s, x{s)) ds,


JtQ JtQ
which establishes (1.19). On the other hand, if identity (1.19) holds for
every t (a, 6), then clearly x{to) xo, and taking derivatives with respect
to t, we obtain
x'{t) = f{t,x{t)) (1.20)
for every t G (a, 6). Since the function 1 f { t , x{t)) is continuous, it follows
from (1.20) that x is of class C^, and hence it is a solution of the initial
value problem (1.13).
1. Ordinary Differential Equations

1.1.2. N otion o f a flow. In this section we consider the particular case


of ordinary differential equations not depending explicitly on time, and we
show that they naturally give rise to the notion of a flow.

D efinition 1.10. Equation (1.1) is said to be autonomous if / does not


depend on t.

In other words, an autonomous equation takes the form


x' - f(x),
where / : D > K is a continuous function in an open set Z> C M".
Now we introduce the notion of a flow.

D efinition 1.11. A fanaily of transformations (pt'.MP' K for t 6 K such


that (fio = Id and
(pt+s ^(ptOips for t,s e R (1-21)
is called a flow.

Exam ple 1.12. Given y G M , the family of transformations R ^ E


defined by
<Pt{x) = x + ty, t G M, a; K"
is a flow.

The following statement shows that many autonomous equations give


rise to a flow.

P rop osition 1.13. If f : R^ is a continuous function such that each


initial value problem
x' = / ( ) ,
( 1.22)
{:a;(0) = XQ
has a unique solution x{t, a;o) defined for t G R, then the family of transfor-
motions defined by
<Pt{xo) = x(t,xo) (1-23)
is a flow.

P ro o f. Given s G R, consider the function y : R ^ R defined by


y{t) = x{t + s,xo).
Clearly, y{0) = x { s, xq) and
y'{t) = x'{t + s, xo) = f(x{t + s, a;o)) = f{y{t))
for t G R. In particular, y is also a solution of the equation x' f{x).
By hypothesis each initial value problem (1.22) has a unique solution, and
hence,
y{t) = x{t,y{Q)) = x{t,x{s,xo)),
1.2. Existence and uniqueness o f solutions

or equivalently,
x(t + s, xq) = x(t, x(s, xo)), (1-24)
for every t, s G M and xq M . Using the transformations <pt in (1.23), one
can rewrite identity (1.24) in the form

<Pt+s(xo) = (<pto<Ps)(xo)-
On the other hand, <po(xo) = x (0, xq) xq, by the definition of x(t,xo).
This shows that the family of transformations <pt is a flow.

Exam ple 1.14 (continuation of Examples 1.3 and 1.8). We note that equa
tion (1.4) is autonomous. Moreover, the initial value problem (1.15) has a
unique solution that is deflned in M, given by (1.17) or (1.18). Thus, it fol
lows from Proposition 1.13 that the family of transformations (pt deflned by

(Pt(xo,yo) = (r cos(-t + 0o), r sin (-t -1- ffo))


is a flow. Using (1.18), one can also write
<Pt(xo,Po) = (x(t),p(t)) = (xo,po)Ji(t)*,
where E(t)* is the transpose of the rotation matrix R(t). Hence, identity
(1.21) is equivalent to
R(t + s)* = R(s)*E(t)*,
or taking transposes,
R(t + s) = R(t)R(s),
for t, s K. This is a well-known identity between rotation matrices.

Exam ple 1.15 (continuation of Example 1.5). Taking t = 0 in (1.12), we


obtain r(0) = c and 0(0) = d. Thus, the solutions of equation (1.8) can be
written (in polar coordinates) in the form
(r(t),0(t)) = (roe^t + eo),

where (ro,0o) = ('(0)> ^(0))- K f t is the flow given by Proposition 1.13


for the autonomous equation (1.11), then in coordinates (x,p) the point
{r{t),6(t)) can be written in the form
(^t (^'o cos 00)^0 sin 00) = (r(t) cos 0(t),r(f) sin 0(f))
= (roe* cos(f -|- 0o), roe* sin(f -|- 0o))

1.2. Existence and uniqueness of solutions

In this section we discuss the existence and uniqueness of solutions of an


ordinary differential equation. After formulating the results, we make a
brief digression on auxiliary material that is used in the proofs of these and
of other results in the book.
10 1. Ordinary Differential Equations

1.2.1. Form ulation o f the P ica rd -L in d e lo f theorem . We first recall


the notion of a compact set.
D efinition 1.16. A set K C M' is said to be compact if given an open cover
of K, that is, a family of open sets {Ua)aei C K* such that Uae/ 3 K,
there exists a finite subcover, that is, there exist N e N and a i , . . . , Ojv I
such that U jli Uai D K.

We recall that a set AT C is compact if and only if it is closed and


bounded.
Now we introduce the notion of a locally Lipschitz function.
D efinition 1.17. A function / : D >E in a set D C M x K is said to be
locally Lipschitz in x if for each compact set K C D there exists L > 0 such
that
- f{t,y)\\ <L||a;-y|| (1.25)
for every (t,x), (t,y) K.

Using this notion, one can formulate the following result on the existence
and uniqueness of solutions of an ordinary differential equation.
T h eorem 1.18 (Picard-Lindelof theorem). If the function f : D >W^ is
continuous and locally Lipschitz in x in an open set D c M. x E , then for
each {to,xo) G D there exists a unique solution of the initial value prob
lem (1.13) in some open interval containing to.

The proof of Theorem 1.18 is given in Section 1.2.3, using some of the
auxiliary results described in Section 1.2.2.
In particular, all functions are locally Lipschitz.
P rop osition 1.19. If f is of class C^, then f is locally Lipschitz in x.

P roof. We first show that it is sufficient to consider the family % of compact


sets of the form I x J c D, where / C E is a compact interval and
J = {a :G E : ||a;-p|| < r }
for some p E " and r > 0. To that effect, we show that each compact set
K <z D is contained in a finite union of elements of X. Then, let AT C D
be a compact set. For each p E K, take Kp E X such that p E intATp.
Clearly, K C (Jpe^^ it Kp, and thus, it follows from the compactness of AT
that there exist AT G N and pi,...,PN K such that K C ( J i l i l^^tKp^.
Hence, K c [J ili Kp^ c D, which yields the desired result.
Now let / X J be a compact set in X and take points (t, x), {t, y) E I x j .
The function p : [0,1] D defined by
9{s) = f { t ,y s{x - y))
1.2. Existence and uniqueness o f solutions 11

is of class C^, because it is a composition of functions. Moreover,

/(t , a;) - f{t, y) = p (l) - ?(0) = f g'{s) ds


Jo

^ Jo
where d f jdx is the Jacobian matrix (thinking of the vector x y os a
column). Thus,

\\f{t>x)-f{t,y)\\< sup ^ { t , y + s{x - y)) \x-y\\


se(o,i]
(1.26)
< sup \ x -y \l
qeJ

because the line segment between x and y is contained in J. Now we observe


that since / is of class C^, the function

(t, q) i->

is continuous, and hence.

L \= max < +00.

It follows from (1.26) that


\\f{t,x) - f{t,y)\\ < L||a;-y||
for every (t,a:), (t, y) & I x J . This shows that / is locally Lipschitz in x.

The following is a consequence of Theorem 1.18.

T h eorem 1.20. If f : D > R is a function of class in an open set


D c R x R"', then for each {to,xo) D there exists a unique solution of the
initial value problem (1.13) in some open interval containing tq.

P roof. The result follows immediately from Theorem 1.18 together with
Proposition 1.19.

However, locally Lipschitz functions need not be differentiable.

E xam ple 1.21. For each a:, y 6 R, we have


||a;| - |y|| < \x-y\.
Hence, the function / : R^ > R defined by f{ t,x ) = |a;| is locally Lipschitz
in X, taking L = 1 in (1.25) for every compact set K C R^.

We also give an example of a function that is not locally Lipschitz.


12 1. Ordinary Differential Equations

E xam ple 1.22. For the function / : H defined by f( t ,x ) = we


have
1
| /( t .x ) - /( t ,0 ) | = |x"/ -0^/=| = |x-0|.

Since l/|x|^/^ >+oo when x 0, we conclude that / is not locally Lipschitz


in X in any open set > C intersecting the line R x {0 }.

The following example illustrates that if / is continuous but not locally


Lipschitz, then the initial value problem (1.13) may not have a unique solu
tion.
E xam ple 1.23 (continuation of Example 1.22). If x = x(t) is a solution of
the equation x' = x^/^, then in any open interval where the solution does
not take the value zero, we have x'/x^/^ = 1. Integrating on both sides, we
obtain x(t) = (t -f c)^/27 for some constant c G R. One can easily verify
by direct substitution in the equation that each of these solutions is defined
in R (even if vanishing at t = c). In particular, x(t) = ^ jTl is a solution
of the initial value problem
fx ' = x2/3,
\x{0) = 0,
which also has the constant solution x{t) = 0.

On the other hand, we show in Section 1.4 that for a continuous func
tion / the initial value problem (1.13) always has a solution, although it
need not be unique.

1.2.2. C on tractions in m etric spaces. This section contains material of


a more transverse nature, and is primarily included for completeness of the
exposition. In particular, the theory is developed in a pragmatic manner
having in mind the theory of differential equations.
We first recall some basic notions.
D efinition 1.24. Given a set X , a function d: X x X RJ is said to be
a distance in X if:
a) d{x, y) = 0 if and only if x = y;
b) d{x,y) = d{y,x) for every x,y X;
c) d{x, y) < d{x, z) + d{z, y) for every x ,y , z E X.
We then say that the pair {X, d) is a metric space.

For example, one can define a distance in R*^ by

d((xi,...,Xn),(yi,...,yn)) = ^5^(t-2/i)^^ , (1-27)


1.2. Existence and uniqueness o f solutions 13

or

d{{xi,...,Xn),{yi,...,yn)) = m a x {| x i-y i| : i = 1 ,... ,n }, (1.28)

among many other possibilities.

D efinition 1.25. Let (X,d) be a metric space. Given x e X and r > 0, we


define the open ball of radius r centered at x by

B{x,r) = { y X : d{y,x) < r},

and the closed ball of radius r centered at x by

B{x,r) ^ {y e X : d(y,x) < r}.

A particular class of distances is obtained from the notion of norm.

D efinition 1.26. Let A be a linear space (over M). A function X ^


is said to be a norm in X if:

a) ||x|| = 0 if and only if x = 0;


b) ||Ax|| = |A| ||x|| for every A 6 K and x X ;
c) ||x + y|| < ||x|| + ||j/|l for every x , y e X .
We then say that the pair (X, H-H) is a normed space.

For example, one can define a norm in R by


1/2
||(xi,...,x,^ (1.29)
^ i=l /
or
||(xi,...,Xn)|| = max{|xi| :i = l , . . . , n } , (1.30)
among many other possibilities.

P rop osition 1.27. If the pair (X, H'll) is a normed space, then the function
d: X X X RJ defined by d{x,y) = ||x y|| is a distance in X.

P roof. For the first property in the notion of distance, it is sufficient to


observe that

d{x, y) = 0 ||x - y|| = 0 x - y = 0,

using the first property in the notion of norm. Hence, d{x,y) = 0 if and
only if X = y. For the second property, we note that
d{y,x) = ||y-x|| = ||-(x-y)||
= |-lM |x-y|| = ||x-y||,
14 1. Ordinary Differential Equations

using the second property in the notion of norm. Hence, d{y,x) = d{x,y).
Finally, for the last property, we observe that
d{x,y) = ||a;-yl| = \\{x - z) + {z - y)\\
< ||a;-2:|| + ||^-y||
= d{x,z) + d{z,y),
using the third property in the notion of norm.

For example, the distances in (1.27) and (1.28) are obtained as described
in Proposition 1.27, respectively, from the norms in (1.29) and (1.30).
Now we introduce the notions of a convergent sequence and of a Cauchy
sequence.

D efinition 1.28. Let (x)n = (aJn)nN be a sequence in a metric space


(X,d). We say that:
a) {xn)n is a convergent sequence if there exists x X such that
d{xn, x) > 0 when n )oo;
b) {xn)n is a Cauchy sequence if given e > 0, there exists p N such
that
d{xn, Xm) < for every n , m > p.

Clearly, any convergent sequence is a Cauchy sequence.

D efinition 1.29. A metric space {X, d) is said to be complete if any Cauchy


sequence in X is convergent.

For example, the space R with the distance in (1.27) or the distance
in (1.28) is a complete metric space. The following is another example.

P rop osition 1.30. The set X = C{I) of all bounded continuous functions
x: I MT in a set I is a complete metric space with the distance
d(x,y) = sup{||x(t) -y(t)|| : t G l } . (1.31)

P roof. One can easily verify that d is a distance. In order to show that X
is a complete metric space, let (xp)p be a Cauchy sequence in X. For each
t 7, we have
\\Xp{t) - Xg(t)|| < d(Xp, Xg), (1.32)
and thus (xp(t))p is a Cauchy sequence in R . Hence, it is convergent (be
cause all Cauchy sequences in R are convergent) and there exists the limit
x{t) = lim Xp{t). (1.33)
p ^oo

This yields a function x: I R". For each t,s I, we have


||x(t) - x(s)|| < ||x(t) - xp(t)l| + ||xp(t) - Xp(s)|| -I- ||xp(s) - x(s)||. (1.34)
1.2. Existence and uniqueness o f solutions 15

On the other hand, it follows from (1.32) (and the fact that (xp)p is a Cauchy
sequence) that given e > 0, there exists r 6 N such that

||a;p(t) - a;g(t)|| < (1.35)

for every t E I and p , q > r . Letting g ^ oo in (1.35), we obtain

||xp(t) - a;(t)|| < (1.36)

for every t E I and p > r . Hence, it follows from (1.34) (taking p = r) that

||a;(t) x(s)|| < 2 + | | r(0 r(s)|| (1-37)

for every t,s E I. Since the function Xr is continuous, given t E I, there


exists (5 > 0 such that

||a;r(t) Xr(s)|| < whenever ||t s|| < 5.

Hence, it follows from (1.37) that

||o;(t) x(s)|| < 3 whenever ||f s|| < 6,

and the function x is continuous. Moreover, it follows from (1.36) that given
> 0, there exists r G N such that
\\x{t)\\ < \\xp{t)-x{t)\\ + ||a;p(t)||
< + sup { ||xp(t) II : t E 1} < + 0 0

for every p > r , and hence x E X. Furthermore, also by (1.36),

d{xp,x) = sup {||a;p(t) a:(t)|| : t E l } < e

for every p > r, and thus d(xp, a;) >0 when p oo. This shows that X is
a complete metric space.

Now we recall the notion of Lipschitz function.

D efinition 1.31. A function s : / ^ M in a set / C is said to be


Lipschitz if there exists L > 0 such that

||a;(t)-x(s)|| < L||<-5|| (1.38)

for every t,s E I.

Clearly, all Lipschitz functions are continuous.

P rop osition 1.32. The set Y C C{I) of all bounded Lipschitz functions
with the same constant L in (1.38) is a complete metric space with the
distance d in (1.31).
16 1. Ordinary Differential Equations

Proof. In view of Proposition 1.30, it is sufficient to show that if (xp)p is a


Cauchy sequence in Y (and thus also in C{I)), then its limit x satisfies (1.38).
Then, let {xp)p be a Cauchy sequence in Y. We have

|xp(t) - a;p(s)|| < L p - s||, t,s l (1.39)

for every p N. On the other hand, by (1.33), for each t / we have


Xp{t) x{t) when p >oo, and thus it follows from (1.39) that x e Y.

Now we consider a particular class of transformations occurring several


times in the proofs of this book.

D efinition 1.33. A transformation T: X X in a metric space (X,d) is


said to be a contraction if there exists A (0,1) such that

diT{x) ,T{y )) <\d {x ,y)


for every x,y E X.

Exam ple 1.34. Let A be an n x n matrix and let T : R " ^ K be the linear
transformation defined by T{x) Ax. Consider the distance d in (1.27),
obtained from the norm
/ ___
n K 1/2
1/2
L,---,a;)|| = (
11(^1,- J .
^ i=l /
We have
d{T{x),T{y)) = \\Ax-Ay\\
= P ( ^ - y ) l l < \\Md{x,y),
where
II .III ll-^ll (1.40)
Ip II
In particular, if ||A|| < 1, then the transformation T is a contraction.
For example, if A is a diagonal matrix with entries A i , . .. ,A in the
diagonal, then

||A|| = sup
(xi,...,Xn)#0 IK^li **) ^n) II
< max {|Ai| : i = 1,... ,n }.

Thus, if |Ai| < 1 for i = 1 ,... ,n, then ||A|| < 1 and T is a contraction.

We say that a;o G X is a fixed point of a transformation T: X X if


T{ xq) = xq. Contractions satisfy the following fixed point theorem, where
T is defined recursively by = T o T for n e N.
1.2. Existence and uniqueness o f solutions 17

T h eorem 1.35. I f T : X ^ X is a contraction in a complete metric space


{X,d), then T has a unique fixed point. Moreover, for each x X the
sequence {T^{x))n converges to the unique fixed point ofT.

P ro o f. Given x E X, consider the sequence Xn = T'^{x) for n 6 N. For


each m, n G N with m > n, we have
diXm i n) ^ d ( Xr r i i ^ m l ) "h d { X j f i \ , X m 2 ) "I" ' "I" d( ^Xf i +l , Xn)

< + + \'^)d{T{x), x)
. (1.41)
= -d{T{x),x)
1 -A

< ^^d {T{x),x).

Thus, {xn)n is a Cauchy sequence in X, and since the space is complete the
sequence has a limit, say x q G X . Now we note that

d{T{xn),T{xo)) < Xd{xn,xo) -> 0


when n 00, and hence, the sequence {T{xn))n converges to T{ xq). But
since T{xn) = Xn+i, the sequence {T{xn))n also converges to xq- It follows
from the uniqueness of the limit that T{ xq) = xq, that is, xq is a fixed point
of the transformation T.
In order to show that the fixed point is unique, let us assume that yo E X
is also a fixed point. Then

d{xo,yo) = d{T{xo),T{yo)) < \d{xo,yo). (1.42)

Since A < 1, it follows from (1.42) that xq = yo- The last property in the
theorem follows from the uniqueness of the fixed point, since we have already
shown that for each x E X the sequence (T^{x))n converges to a fixed point
ofT .

It follows from the proof of Theorem 1.35 (see (1.41)) that if xq G is


the unique fixed point of a contraction T: X >X, then
A*^
d(T^(x),xo) < d(T(x),x)

for every x X and n G N. In particular, each sequence (T (x))n converges


exponentially to xq.
Motivated by the last property in Theorem 1.35, we introduce the fol
lowing notion.

D efinition 1.36. A fixed point xo G A" of a transformation T: X X is


said to be attracting if for each x G A we have T (x) > xq when n -> oo.
18 1. Ordinary Differential Equations

Now we consider a more general situation in which it is still possible


to establish a fixed point theorem. Given metric spaces X and Y (not
necessarily complete), consider a transformation S: X x Y > X x Y of the
form
S{x, y) = (T{x), A{x, y)), (1.43)
where T: X - X and A: X x Y > Y. Each set {a;} x Y" is called a
fiber of X X Y. We note that the image of the fiber { s } x Y under the
transformation S is contained in the fiber { T ( x ) } x Y.
D efinition 1.37. The transformation S in (1.43) is said to be a fiber con
traction if there exists A (0,1) such that
dY{A{x,y),A(x,y)) < Xdyiy^y) (1.44)
for every x E X and y^y e Y.

The following is a fixed point theorem for fiber contractions.


T h eorem 1.38 (Fiber contraction theorem). For a continuous fiber con
traction S, if xq ^ X is an attracting fixed point of T and yo is a fixed
point ofy\-^A(xo^y)j then (xo,yo) is an attracting fixed point of S.

P roof. Let dx and dy be, respectively, the distances in X and Y, We define


a distance d in X x F by
d{{x,y),{x,y)) = dx{x,x) + dY(y,y).
It follows from the triangle inequality that
d{S'^{x,y),{xo,yo)) < d{S'^{x,y),S'^{x,yo))+d{S^{x,yo),{xo,yo)). (1.45)
We want to show that both terms in the right-hand side of (1.45) tend to
zero when n >oo. Given x e X, define a transformation Ax'. Y -> Y by
Ax(y) = A(x,y). Clearly,
S{x,y) = {T{x),Ax{y)).
Moreover, for each n N we have

S^{x,y) = {T^{x),Ax,niy)), (1-46)


where
Ax^n A'jpn-l^j,^ O O o Ax
For the first term in (1.45), it follows from (1.44) and (1.46) that
d(5"(x,j/),S'"(x,yo)) = dY{Ax,n{y),Ax,n{yQ)) < y^dY{y,yo) 0 (1.47)
when n > oo. For the second term, we note that
d{S^{x,yo),{xo,yo)) < dx{T^ix),xo)dY{Ax,n{yo),yo)- (1-48)
Since xq is an attracting fixed point of T, we have dx{T^{x), x q ) 0 when
n oo. In order to show that (xq, yo) is an attracting fixed point of S, it
1.2. Existence and uniqueness o f solutions 19

remains to verify that the second term in the right-hand side of (1.48) tends
to zero when n oo. To that effect, note that

iyo))yo)
n 1

- O o ^ T ^ { x ) ) ( y o ) y (-^ T "-l(i) O OA x i + l ( x ) ) { y o ) )
i=0
n1
* ^<^y(^T*(x)(yo),yo))
1=0

where o o = Id for i = n 1. Since yo is a fixed point


of A xq, we have

Ci := dy(^T^x)(yo),yo) = dy(^(T *(),yo),^(a;o,yo))- (1.49)

On the other hand, since the transformation A is continuous (because S is


continuous) and xq is an attracting fixed point of T, it follows from (1.49)
that the sequence Cj converges to zero when i -4 oo. In particular, there
exists c > 0 such that 0 < Cj < c for every i N. Thus, given /c G N and
n > k, we have

i=0 i=0 i=k


k1 n
< c A * + sup Cj ^ 2
i=o i=k
^nk+1 ^

Therefore,
^ ^ 1
lim su p^ A^*Ct < supcj ----- r 0
n-oo j> k 1 A
when k oo. This shows that

d Y { ^ x , n { y o ) , y o ) 0 when n ^ oo,

and it follows from (1.45), (1.47) and (1.48) that (a;o,yo) is an attracting
fixed point of S.

1.2.3. P r o o f o f the theorem . Now we use the theory developed in the


previous section to prove the Picard-Lindelof theorem.

P r o o f o f T h eorem 1.18. By Proposition 1.9, the initial value problem


(1.13) consists of finding a function x G C{a,b) in an open interval (a, 6)
20 1. Ordinary Differential Equations

containing to such that

a ;(t)= a :o + / f{s,x{s)) ds (1.50)


Jto
for every t {a,b). Here C{a,b) is the set of all bounded continuous func
tions y: {a,b) M . We look for x as a fixed point of a contraction.
Take constants a < t o < b and /3 > 0 such that

K := [o, b] X B{xo,l3) C D, (1.51)

where
B{xo,P) = {yeR ^-. \\ y- xo \\< p}.
Also, let X C C{a, b) be the set of all continuous functions x : (o, 6) K
such that
||x(t) xoll ^ t ^ () b)-
We first show that X is a complete metric space with the distance in (1.31).
Given a Cauchy sequence {xp)p in X, it follows from Proposition 1.30 that
it converges to a function x C(a,b). In order to show that x X , we note
that
||x(t) - xoll = lim ||xp(t) - xoll < /5
p >oo

for t G (a, b), since ||xp(t) xqH< for t G (a, b) and p G N. Moreover, there
is no loss of generality in looking for fixed points in X and not in C(a,b).
Indeed, if x : (a, b) is a continuous function satisfying (1.50), then

||a;(t)-xo|| < I / f(s,x(s))ds


II Jto
< |t to|M < (6 o)M ,
where
M = max{||/(t,x)|| : (t,x) G X } < -f-oo, (1.52)
because / is continuous and K is compact (recall that any continuous func
tion with values in R, such as (t, x) i->- ||/(i,x)||, has a maximum in each
compact set). This shows that if x G C{a,b) satisfies (1.50), then it belongs
to X for some 0.
Now we consider the transformation T defined by

T(x){t) = x o + [ f{s,x{s))ds
Jto
for each x G X . We note that t T (x)(t) is continuous and

||r(x)(t) Xoll < I f f{s,x{s))ds < { b - a)M.


II Jto
1.3. Additional properties 21

For b a sufficiently small, we have (b a)M < 0 and thus T{X) C X.


Moreover, given x,y X,

||r(x)(0 -T(y)(t)|| < I [ [ f { s , x { s ) ) - f { s , y { s ) ) ] d s


II Jto

< 1 / i ' l k ( s ) - 2/(s)||ds


I Jto
< ( b - a ) L d {x , y ) ,
where L is the constant in (1.25) for the compact set K in (1.51) and d is
the distance in (1.31) for I = (a,b). Hence,
d{T{x),T{y)) < {b - a)Ld{x,y)
for every x,y X. For b a sufficiently small, we have (6 a)L < 1, in
addition to {b a) M < P, and T is a contraction in the complete metric
space X . By Theorem 1.35, we conclude that T has a unique fixed point
X E X. This is the unique continuous function in (o, b) satisfying (1.50).

1.3. Additional properties

In this section we study some additional properties of the solutions of an


ordinary differential equation, including the dependence of the solutions on
the initial conditions and what happens when a given solution cannot be
extended to a larger interval.

1.3.1. Lipschitz d ependence on the initial conditions. We first de


scribe how the solutions depend on the initial conditions for a continuous
function / that is locally Lipschitz.
We start with an auxiliary result.

P rop osition 1.39 (Gronwalls lemma). Letu,v: [a, 6] he continuous


functions with v > 0 and let c EM.. If
rt
u { t ) < c + / u{s)v{s)ds (1.53)
Ja
for every t E [o, 6], then

u{t) < cexp / v{s) ds


Ja
for every t E [a, 6].

P ro o f. Consider the functions

-^(0 = [ u{s)v{s) ds
Ja
22 1. Ordinary Differential Equations

and

^ ( 0 = [ v{s)ds.
Ja

Clearly, R{a) = 0 and by hypothesis,

R'{t) = u{t)v(t) < {c + R{t))v{t).

Therefore,
R'(t) - v(t)R{t) < cv{t)
and

- v{t)R(t))

< cv{t)e~^^^\

Since R{a) = 0, we obtain

e~^^*^R{t) < j dr
Ja

= -ce-^ (r) "= = c ( l - e - ^ W ) ,


T=a

and hence,
R{t) < - c.
Thus, it follows from (1.53) that

u{t) < c + R{t) <

for every t 6 [a, 6]. This yields the desired result.

Now we establish the Lipschitz dependence of the solutions on the initial


conditions.

T h eorem 1.40. Let f : D ^W^ be continuous and locally Lipschitz in x


in an open set D c E x R . Given D, there exist constants
/3,C > 0 and an open interval I containing to such that for each X2 R
with 11! X2W< 0 the solutions xi{t) and X2{t), respectively, of the initial
value problems

x' = f{t,x), x' = f{t,x).


and (1.54)
x{to) = Xl x{to) = X2

satisfy
lla:i(t) - X2 {t)\\ < C\\xi - X2 W for t G I. (1.55)
1.3. Additional properties 23

P roof. By the Picard-Lindelof theorem (Theorem 1.18), the initial value


problems in (1.54) have unique solutions xi(t) and X2{t) in some open in
terval I containing to (which we can assume to be the same). Moreover, by
Proposition 1.9, we have

Xi{t) =^xi+ f f{s, x{s)) ds (1.56)


Jto
f o r t l and i = 1,2. For simplicity of the exposition, we consider only
times t > to. The case when t < to can be treated in an analogous manner.
We define a function y : 7 >R by

y{t) = x i { t ) - X 2 {t).

For t > tO) it follows from (1.56) that

||y(t)|| < Ik i-a : 2 ||-I- [ \\f{s,xi{s)) - f{s,X2 {s))\\ds


Jto

<\\x i - X 2\\+ L [ ||y(s)||(is,


Jto
for some constant L (because / is locally Lipschitz in x). Letting

^i(t) = l|y(t)ll. ^(t) = and c = ||xi - X2II,


it follows from Gronwalls lemma (Proposition 1.39) that

u(t) < cexp / u(s) ds = ||xi X2||<^L{t-to)


Jto
for t 7n[to) + 00), which establishes the inequality in (1.55) for these values
of t.

1.3.2. S m ooth d ependence on the initial conditions. In this section


we show that for a function / of class the solutions are also of class
on the initial conditions. The proof uses the Fiber contraction theorem
(Theorem 1.38).
We first establish an auxiliary result on the limit of a sequence of differ
entiable functions.

P rop osition 1.41. If a sequence of differentiable functions fn- U R "


in an open set 17 C R is (pointwise) convergent and the sequence dfn of
their derivatives is uniformly convergent, then the limit of the sequence fn
is differentiable in U and its derivative is the limit of the sequence dfn in U.

P roof. Let f : U ^W^ and g \ U - ^ Mn be, respectively, the (pointwise)


limits of the sequences / and dfn, where Mn is the set of n x n matrices
24 1. Ordinary Differential Equations

with real entries. Given x, R such that the line segment between x and
x + h is contained in C/, we have

fn(x + h) - fn(x) = j ^ f n { x + th)dt = dx+thfnhdt (1.57)

for every n N. On the other hand, since the sequence dfn converges
uniformly, given 5 > 0, there exists p G N such that

{ll*^2//n <^3//mil V ^ <5


for m , n > p. Hence, it follows from (1.57) that
I [fn{x + h ) - fn{x)) - {fm{x + h) - fm{x)) |
~ / (d x + th fn d x + th fm ^ h d t
II J o

< [ \\dx+thfn - dx+thfm\\ dt\\h\\ < <5||/l||

for m , n > p. Letting m ^ oo, we obtain


II{fn{x + h ) ~ fn{x)) - { f { x + h ) ~ f{x)) || < 5||/i|| (1.58)
for every n > p.
Now we show that / is differentiable. Given x, h G R such that the line
segment between x and x + h is contained in U, for each e G (0,1) we have
/ ( x + eh) - /( x ) f{x + eh) - /(x ) fn{x + eh) - fn{x)
- g{x)h <

fn{x + eh) - fn{x)


+ dxfnh

+ \\dxfnh - g{x)h\\.
(1.59)
Since g is the limit of the sequence dfn, given 5 > 0, there exists g G N such
that
sup {\\dyfn - g{y)\\:y U } < 6 (1.60)
for n > q. Now, take n > max{p, g}. Since fn is differentiable, for any
sufficiently small e we have
fn{x + eh) - fn{x)
dxfnh <5|H|. (1.61)

It follows from (1.59) together with (1.58), (1.60) and (1.61) that
/ ( x + eh) - / (x)
-g{x)h <35||h||.
e
Therefore, the function / is differentiable at x, and dxfh = g{x)h for each
h G R , that is, df = g.
1.3. Additional properties 25

Now we establish the dependence of the solutions on the initial con


ditions.
T h eorem 1.42. If the function /:>-> K is o / class in an open set
D cM. X R , then for each {tQ,xo) D the function (t,x) <p{t,x), where
(/?(, xo) is the solution of the initial value problem (1.13), is of class in
a neighborhood of {to,xo).

P roof. We divide the proof into several steps.

Step 1. Construction of a metric space. Given (to)o) -D, take constants


a,P > 0 such that
S := [^0 o, to + cc] X B{ xq, 2S) C D. (1.62)
Since the set S is compact and / is continuous, we have
K ;= max {||/(t, x)|| : (t,x) 5 } < -l-oo
(recall that a continuous function with values in R has a maximum in each
compact set). If necessary, we take a smaller a > 0 such that a K < 13, and
we consider the set X of all continuous functions (p: U B{ xq,20) in the
open set
U = {tQ-a,tQ + Oi)x B{ xq,^), (1.63)
equipped with the distance
d{(p,ip) = sup{||(^(t,x) ~'tp{t,x)\\ : (t,x) e U } . (1.64)
Repeating arguments in the proof of the Picard-Lindeldf theorem (The
orem 1.18), it follows easily from Proposition 1.30 that X is a complete
metric space.

Step 2. Construction of a contraction in X . We define a transformation T


in X by

T((fi){t,x) = x + [ f{s,ip{s,x))ds
Jto
for each {t,x) U. We first show that T{(p) is a continuous function.
Clearly, t i->- T{<p)(t,x) is continuous for each x 6 B{ xq,S). Given t
(to cc, to + Q:), we show that the function x T{ip){t, x) is also continuous.
For each x ,y g B{ xo,S)> we have

\\T{<p){t,x) -T{(p){t,y)\\ < \\x-y\\ + I f ||/(s,y5(s,x)) - f{s,(p{s,y))\\ds


\Jto

<\\x-y\\ + L\ [ |l<^(s,x)-(/3(s ,2/)||(ts


I Jto
(1.65)
for some constant L. Indeed, by Proposition 1.19, the function / is locally
Lipschitz in x, and hence, one can take L to be the constant in (1.25) for
26 1. Ordinary Differential Equations

the compact set S in (1.62). Moreover, each function (/? X is uniformly


continuous on the compaxit set

U = [ t o - a, to + a] X B{xo,P),
because (p is continuous. Thus, given 5 > 0, there exists > 0 such that

for every s [to a, to + a] and ||x y|| < e. Without loss of generality,
one can always take e < 5. Hence, it follows from (1.65) that

\\T{p>){t,x)-T{^p){t,y)\\ < e + L|t-to|(5 <5{l-\-La)


whenever ||o; 3/|| < e. This shows that the function x T{(p){t,x) is
continuous. On the other hand,

||T(v?)(t,a;)-o|| < ||x-xo|| + | / ||/(s, v?(s,aj))|| ds


\Jto
< ||a: - xoll + a K < P + P = 2P,
and hence T{X) C X.
Now we verify that T is a contraction. We have

\\T{<p){t,x) - T{i>){t,x)\\ < f \\f{s,<p{s,x)) - f{s,i>{s,x))\\ds


Jto ( 1. 66)
< Lad{(p,rl}),

with the same constant L as in (1.65), and thus,

d{T{(p),T{ip)) < Lad{ip,ip).

For any sufficiently small a, we have La < 1, in addition to a K < /?, and
thus T is a contraction. By Theorem 1.35, we conclude that T has a unique
fixed point (po E: X, which thus satisfies the identity

(po{t,x) = x + [ f{s,<po{s,x))ds (1-67)


Jto
for every t {to a, to + a) and x B{xo, /?). By Proposition 1.9, it follows
from (1.67) that the function 1 1-)- cpo{t,xo) is a solution of the initial value
problem (1.13). In particular it is of class C^. To complete the proof, it
remains to show that ipo is of class in x.

Step 3. Construction of a fiber contraction. Again, let be the set of


n X n matrices with real entries. Also, let Y be the set of all continuous
functions M , with {7 as in (1.63), such that

sup {|l$(t,a;)|| : (t,x) U} < +oo


1.3. Additional properties 27

(the norm o;)|| is defined by (1.40)). It follows from Proposition 1.30


that F is a complete metric space with the distance

d ( $ ,^ ') : = sup{||$(t,a:) - S'(t,x)|l : (t,x) Uj .


Now we define a transformation ^4 in X x F by

A((p, $)(t, x) = Id + |^(s, x ))$ (s, x) ds,

where Id Mn is the identity matrix. We also consider the transformation S


in (1.43), that is,

We first show that S is continuous. By (1.66), the transformation T is


continuous. Moreover,
\\A{if, $ )(t,x ) -^(i/,^')(t,x)||

=
II
y
df
-^{s,ip{s,x))^s,x)ds-
/* Of
J
(s,V(s,a:))^'(s,x)ds

< 1 ^ |^ (s,< ^ (s,x ))[$ (s,x ) - ^ (s,x )]d s


( 1. 68)
+ 1 ^ ^ |^ (s,v ?(s,x )) - ^ ( s , i p { s , x ) ) ^ ^{ s ,x ) ds

< aMd{^, 3')


df
+ f qJ s,(p ( s, x )) - -^{s,'lp{s,x)) ||J'(s,x)||ds
'to
where
M := max : (t, x) G [7 ^ < + 00.
{
Taking (p = rp in (1.68), we obtain

d(yl(v?,$),^(V3,$)) < a M d { ^ , ^ ) . (1.69)

In particular, the function $ ^) is continuous for each tp E X. On


the other hand, since df/dx is uniformly continuous on the compact set U,
given 5 > 0, there exists e > 0 such that
d f, , d f. ,
-(s ,x )--(s ,y ) <5

for s [to ~ oCfto + Q:] and ||x y|| < e. Hence, if d{(p,rp) < e (see (1.64)),
then
^{ s, ( p { s, x ) ) - |^(s,V'(s,a;)) <5
28 1. Ordinary Differential Equations

for s G (to <^)to + cc) and x B(xo,0). It follows from (1.68) with $ = ^
that
\\A{(p,^){t,x) - A{'tp,^){t,x)\\ < a5sup{||#(t,a:)|| : (t,x) e U } ,
and thus also
d{A{(p,^),A{7p,^)) < a(5sup{||$(t,a:)|| : (t,x) G 17},
whenever d{(f, < e. This shows that the function tp i-)- A{ip, $) is con
tinuous for each $ G F, and hence A{ X x F ) c F. Moreover, for any
sufficiently small a we have aM < 1, and it follows from (1.69) that 5 is a
fiber contraction.
Step 4- Uniform convergence and regularity. By the Fiber contraction
theorem (Theorem 1.38), given (<^, $ ) G X x F, the sequence S^{(p,^)
converges to (po, $o) when n >oo, where $o is the unique fixed point of the
contraction $ A{ipQ, $ ) in the complete metric space F. In other words,
the sequence of functions 5'"(^, $ ) converges uniformly to the continuous
function (^o> ^o) in the open set U when n oo.
Since $o is continuous, if one can show that
dpQ
{t,x) = ^o{t,x), (1.70)
dx
then ipo is of class in x, and thus it is of class C^. To that effect, consider
the functions ipi and given by (pi{t,x) = x and $i(t,a:) = Id. For each
n G N, we define a pair of functions by

These can be obtained recursively from the identities

(Pn+l{t,x) = X + [ f{s,(pn{s,x))ds (1.71)


Jto
and
U df
^n+l {t,x) = ld + J {s, cpn{s , x)) ^n{s , x) ds . (1.72)

One can use induction to show that the functions </? are of class in x for
each n G N, with

5>n+l d
^ f^ r
dx
Jt f(^^Pn{s,x))ds
(1.73)
= ld + ^{s,(pn{s,x))^^{s,x)ds.

For this, it is sufficient to observe that if g>n is of class in a;, then the func
tion {s,x) i-> f{s,ipn{s,x)) is also of class and thus one can use (1.71)
1.3. Additional properties 29

together with Leibnizs rule to obtain the differentiability of (pn+i and iden
tity (1.73). This implies that (pn+i is of class C^.
Now we show that
^^"'(t, x) = ^n(t, x) (1.74)
dx
for each n N. Clearly,
d(pi
(t,x) = Id = $ i(t, x).
dx
In order to proceed by induction, we assume that (1.74) holds for a given n.
It then follows from (1.73) and (1.72) that
d(Pn+l , f df
(t,a;) = Id-f- J ^{s,(pn{s,x))^n{s,x)ds = ^n+l{t>x).
dx
This yields the desired identity.
Finally, given t G {to a,to + oc), we consider the sequence of functions
fn{x) = fnitt x). By (1.74), we have dxfn = ^n(i> x). On the other hand, as
we already observed, by the Fiber contraction theorem (Theorem 1.38) the
sequences / and dfn converge uniformly respectively to v^o(^) ) and $o(^> )
Hence, it follows from Proposition 1.41 that the derivative {d(po/dx){t^x)
exists and satisfies (1.70). This completes the proof of the theorem.

1.3.3. M axim al interval o f existence. In this section we show that each


solution of the initial value problem (1.13) given by the Picard-Lindelof
theorem (Theorem 1.18) can be extended to a maximal interval in a unique
manner.

T h eorem 1.43. If the function / ; >M is continuous and locally Lips-


chitz in x in an open set D c M .x R , then for each {to, xq) G D there exists
a unique solution tp: {a,b) > R of the initial value problem (1.13) such
that for any solution x : Ix ^ of the same problem we have Ix C (o, b)
and x{t) = <p{t) for t G Ix.

P roof. We note that J = \JxIx is an open interval, because the union of any
family of open intervals containing to is still an open interval (containing to)-
Now we define a function J -> R as follows. For each t G Ix, let
(p{t) = x{t). We show that the function ip is well defined, that is, p{t) does
not depend on the function x. To that effect, let a;: /x )R and y : ly >R'^
be solutions of the initial value problem (1.13). Also, let I be the largest
open interval containing to where x = y. We want to show that I I x D l y
Otherwise, there would exist an endpoint s of 7 different from the endpoints
of IxH ly. By the continuity of x and y in the interval Ixtlly, we would
have
p := lima;(t) = limy(t).
t-^s ' ' t^s
30 1. Ordinary Differential Equations

On the other hand, by the Picard-Lindelof theorem (Theorem 1.18) with


(to^xo) replaced by (5,p), there would exist an open interval (s a ,s + a) C
Ix n ly where x = y. Since (s a, s + a) \ / ^ 0 , this contradicts the fact
that I is the largest open interval containing where x = y. Therefore,
I = IxO ly and X = y \in Ix O l y Clearly, the function (^: J is a
solution of the initial value problem (1.13). This completes the proof of the
theorem.

In view of Theorem 1.43, one can introduce the following notion of max
imal interval (of existence) of a solution.
D efinition 1.44. Under the assumptions of Theorem 1.43, the maximal
interval of a solution x: I oi the equation x = f{t^x) is the largest
open interval (a, 6) where there exists a solution coinciding with x in I.

We note that the maximal interval of a solution is not always R.


E xam ple 1.45. Consider the equation x^ = x^^. One can easily verify (for
example writing x^/x"^ = 1 and integrating on both sides) that the solutions
are x{t) = 0, with maximal interval R, and

m =

with maximal interval (oo, c) or (c,+ oo), for each c M. More precisely,
for xq = 0 the solution of the initial value problem
ix' = x^,
[a;(to) = a;o
is x{t) 0, with maximal interval M, for xq > 0 the solution is

x{t) =

with maximal interval (00, to + 1/a^o) and, finally, for xq < 0 the solution
is again given by (1.75), but now with maximal interval (to + l/a^O) + 00).

The following result describes what happens to the solutions of an ordi


nary differential equation when at least one endpoint of the maximal interval
is a finite number.
T h eorem 1.46. Let f : D - ^ W ^ h e continuous and locally Lipschitz in x in
an open set D c M . x R . If a solution x{t) of the equation x' = f{ t ,x ) has
maximal interval (a,b), then for each compact set K C D there exists s > 0
such that {t,x{t)) E D \ K for every
t (o, o -|- ) U (6 (I'f)
(when a = 00 the first interval is empty and when b = -l-oo the second
interval is empty).
1.3. Additional properties 31

P roof. We consider only the endpoint b (the argument for a is entirely


analogous). We proceed by contradiction. Let us assume that for some
compact set K G D there exists a sequence {tp)p with tp b when p t oo
such that
{tp,x{tp}) K for every p N.
Since K is compact, there exists a subsequence (tfcp)p such that {tkp,x{tkp))
converges to a point in K when p -> oo (we recall that a set ii' C is
compact if and only if any sequence (pp)p c K has a convergent subsequence
with limit in K). Let

(6,xo) = lim {tkp,x{tkj,)).


p ^oo ^

Now we consider the initial condition x{b) = xq and a compact set


Kap := [fca, 6 + a] X B{ xq, P) C D
as in the proof of the Picard-Lindelof theorem (Theorem 1.18), for some
constants a , P > 0 such that 2Ma < P, where
M = sup (||/(t,a;)|| : (t,x) G K^p}.
Moreover, for each p G N, we consider the compact set

Lp = hp - /2,ifcp + ot/2] X B{x{tkp),P/2) C D.


For any sufficiently large p, we have Lp c and hence,
2sup {||/(t,x)|| : X G L p}o'/2 < 2Ma/2 < P/2.
Thus, proceeding as in the proof of the Picard-Lindelof theorem, we find
that there exists a unique solution

y- {tkp ~ CK/2,tfcp + oc/2) )R "


of the equation x' = f(t,x) with initial condition y{tkp) = x{tkp). Since
tkp -t- Oi/2 > b for any sufficiently large p, this means that we obtain an
extension of the solution x to the interval (o, tkp+a/2). But this contradicts
the fact that b is the right endpoint of the maximal interval. Therefore, for
each compact set K G D there exists e > 0 such that (t, x{t)) E D \ K tor
every t E {b e, b).

The following is an application of Theorem 1.46.

Exam ple 1.47. Consider the equation in polar coordinates

(1.77)

Proceeding as in (1.9) and (1.10), we obtain


x' = r cos 6 2r sin 0 = x 2y
32 1. Ordinary Differential Equations

and
y' = r sin 0 + 2r cos 0 = y + 2x.
We observe that the assumptions of Theorem 1.46 are satisfied for the open
set ) = RxR^. Hence, for each compact set K = [c,d]xB C D, where H is a
closed ball centered at the origin, there exists e > 0 such that (t, x(t),y{t))
D \ K , that is,
t^[c,d] or {x{t),y{t)) ^ B, (1.78)
for t as in (1.76). On the other hand, by the first equation in (1.77), the
distance of any solution to the origin decreases as time increases. This
implies that each solution (x{t),y{t)) with initial condition (x(to), y{to)) ^ B
remains in B for every t > to- Hence, it follows from (1.78) that t ^ [c,d]
for any sufficiently large t in the maximal interval of the solution. But since
d is arbitrary, we conclude that for any solution with initial condition in B
the right endpoint of its maximal interval is +oo.

1.4. E x is te n c e o f s o lu tio n s fo r co n tin u o u s field s

In this section we show that for a continuous function / the initial value
problem (1.13) always has solutions. However, these need not be unique (as
illustrated by Example 1.23).
We first recall a result from functional analysis.

P rop osition 1.48 (Arzela-Ascoli). Let<pk- (>i>) ^ R be continuous func


tions, for k e N, and assume that:
a) there exists c > 0 such that

sup {||(^fc(t)|| : t E (a,b)} < c for every k e N\

b) given e > 0, there exists 6 > 0 such that ||v?fc(t) 7fc(s)l| < e for
every k e N and t,s E (a, b) with |t s| < 5.
Then there exists a subsequence of {^Pk)k converging uniformly to a contin
uous function in (a, b).

P roof. Consider a sequence {tm)m C [a, 6] that is dense in this interval.


Since {ipk{h))k is bounded, there exists a convergent subsequence {^pi{ti))k-
Similarly, since {<Ppi{t2))k is bounded, there exists a convergent subsequence
{iPp2{t2))k- Proceeding always in this manner, after m steps we obtain a
convergent subsequence (9?p|"(tm))fc of the bounded sequence Up^m-i{tm))k-
Now we define functions Vfe: (a, b) E for A: 6 N by

M i ) =Tp^(i)-
1.4. Existence o f solutions for continuous fields 33

We note that the sequence {i>k{,tm))k is convergent for each m N, because


with the exception of the first m terms it is a subsequence of the convergent
sequence (<^p(tm))fe-
On the other hand, by hypothesis, for each e > 0 there exists (5 > 0 such
that
WAit) - V-feWII = Ibpfc(i) - Ppk{s)\\ < e (1.79)
for every k N and t, s G (o, b) with |t s] < 5. Given t, tm (a, b) with
\t tm\ < S and p,q e N, it follows from (1.79) that
ll'^p(t) 'tpq{t)\\ < ll'0p(t) ~ '*/p(im)|l
h llVp(^m) ~ '0 g(^m)|| + HVg(^m) ~ '0 g(OII (1.80)
< 2e + llVp(^m) '^/g(^m)||-

Since the sequence {ipk{tm))k is convergent, there exists Pm N such that

||'0 p(^m) ~ '0 g(^m)|| ^ (i-'^i-)


for every p,q > Pm- Moreover, since the interval (a, 6) can be covered by
a finite number N of intervals of length 26, and without loss of generality
centered at the points ti,t2, ... ,tN, it follows from (1.80) and (1.81) that

\\ipp{t) - ipg{t)\\ < 3e

for every t G (a, b) and p,q > m a x{p i,. .. , p n }- This shows that (V>p)p is a
Cauchy sequence in the complete metric space C{a, b) with the distance d
in (1.31) (see Proposition 1.30). Thus, the sequence converges uniformly to
a continuous function in (a, 6). This completes the proof of the proposition.

Now we use Proposition 1.48 to establish the existence of solutions for
any ordinary differential equation with a continuous right-hand side.

T h eorem 1.49 (Peano). If f : D is a continuous function in an open


set I? C M X R , then for each (toi^^o) G D there exists at least one solution
of the initial value problem (1.13) in some open interval containing to.

P ro o f. We consider again the equivalent problem in (1.50), and we take


a < to < b and P > 0 such that the conditions (1.51) and (6 o)M < P
hold, where M is the constant in (1.52). Given a > 0, we define recursively
a continuous function Xa' {a, b) R by

xo - f[s,Xa{s + a))ds, te{a,to~ a),


Xoft) = < xo, t G [to ~ O', to o], (1.82)
,X0 + Jta+af{s,Xa{s-a)) ds, t G [to + cx,b).
34 1. Ordinary DifFerential Equations

For example,

Xc {t) = x o + f f{s,X a{s-a))ds = xo+ f f{s,xo)d.


Jto+a Jto-ha
for t E [to+ ce, ^0 + 2a], and
rto-a rt
Xa { t ) = x o - f{s,Xa{s + a ) ) d s ^ X Q + f{s,xo) ds
Jt Jtoa
for t [to - 2a, to a]. We have

||a(i) - a^oll < [ I


I Jto^a
||/(s,a;a(sTQ:))||<is < { b - a)M < p,

and thus,
IkaWII < Ikoll + ^ -
Moreover,

lIrCaW -a(s)|| < I f \\f{u,Xoci:u ^ a )) lld u (1.83)


Js

This shows that the two conditions in Proposition 1.48 are satisfied for any
sequence of functions {xam)mi where {otm)m C M"*" is an arbitrary sequence
such that am \ 0 when m oo. Therefore, there exists a subsequence
{Pm)m of (am)m such that {xp^)m Converges uniformly to a continuous
function x: (a,b) >R'.
It remains to show that is a solution of the initial value problem (1.13).
By (1.83), we have
-a;(s)|| < \\xpAs - M - X0^(s)\\ + \\xff^{s) - x{s)\\
<M\pm\ + \\xp,is)-x{s)\\,
and thus, the sequence of functions s xp^{s Pm) converges uniformly
to X. One can show in a similar manner that the sequence of functions
^ + ^rn) also converges uniformly to x. On the other hajid, it
follows from (1.82) that

Xpmit) = X 0 + [ f {s , Xfi^{s - Pm)) ds


*^to-\-/3m
for t e [to>^)) and letting m oo, we obtain

x { t ) - x o + [ f{s,x{s))ds (1.84)
Jto
for t [to, b). Analogously, it follows from (1.82) that
to-Pm
/ f[s,Xp^{s + Pm))ds
1.5. Phase portraits 35

for t G (a, to]) and letting m >oo, we obtain identity (1.84) for t (o, to]-
Finally, it follows from Proposition 1.9 that x is a solution of the initial value
problem (1.13).

1.5. P h a se p o r tr a its

Although it is often impossible (or very difficult) to determine explicitly


the solutions of an ordinary differential equation, it is still important to
obtain information about these solutions, at least of qualitative nature. To
a considerable extent, thiscan be done describing thephase portrait of
the differential equation. We note that in thissection weconsider only
autonomous equations.

1.5.1. O rbits. Let / : D > M be a continuous function in an open set


D and consider the autonomous equation
x' = f{x). (1.85)
The set D is called the phase space of the equation.

D efinition 1.50. If x = x{t) is a solution of equation (1.85) with maximal


interval I, then the set {x(t) : t G 1} C D is called an orbit of the equation.

We first consider a very simple class of orbits.

D efinition 1.51. A point xq G D with / ( xq) = 0 is called a critical point


of equation (1.85).

P rop osition 1.52. If xq is a critical point of equation (1.85), then { xq} is


an orbit of this equation.

P ro o f. It is sufficient to observe that the constant function x: M ^ E


defined by x{t) = xq is a solution of equation (1.85).

Exam ple 1.53. By Example 1.45, the equation x' = x^ has the solutions
x{t) = 0, with maximal interval E, and x{t) = l / ( c t), with maximal
interval (oo, c) or (c, +oo), for each c E. Thus, we obtain the orbits {0 }
(coming from the critical point 0),

{ l / ( c t ) : t G (o o ,c )} = E'*

and
{ l / ( c t) : t G { c , + o o )} = E .

Now we show that in a sufficiently small neighborhood of a noncritical


point one can always introduce coordinates with respect to which the orbits
are parallel line segments traversed with constant speed.
36 1. Ordinary Differential Equations

T h eorem 1.54 (Flow box theorem). Let f : D - ^ W ^ b e a function of


class in an open set D Given a point p E D with f(jp) ^ 0, there
exists a coordinate change y g{x) in a neighborhood ofp transforming the
equation x' = f {x) into the equation y' = v for some v e W^\ { 0}.

P roof. Let be the solution of the equation x' f{x) with initial con
dition a;(0) = xo, in its maximal interval. Since f{p) ^ 0, some component
of /(p ) is different from zero. Without loss of generality, we assume that the
first component is nonzero, and we define new coordinates y = (y i, . . . , yn)
by
x = F (y) = (pj,i(p-|-y),
where p = (p i, ...,Pn) and y = (0,y2,y s ,,Pn)- By Theorem 1.42, the
function F is of class in some neighborhood of zero. Moreover, F (0) = p.
Now we compute the Jacobian matrix doF. The first column of doF is
given by
d
(p + y)|j,=o = (P + y)) U o = /(^(O)) = fip)-

Since (po = Id, the remaining columns are

^<Pj/i(P + y) = ^<P?/i(Pl.P2 + y2,P 3+y3,---,P n + yn)|^^0 = ei

for i = 2, . . . , n, where ( e i,. . . , Cn) is the canonical basis of R . Hence,


doF = (/(p )e 2 Cn)
is a lower triangular matrix with nonzero entries in the diagonal (because by
hypothesis the first component of /(p ) is nonzero). Thus, doF is nonsingular
and F is indeed a coordinate change in a neighborhood of zero. Now let
(xi (t) , . . . , Xfi(t)) = ^t(p by)
be a solution of the equation x' = /( x ) , where q = {0,q2,qz, ... ,yn), and
write it in the new coordinates in the form y(t) = (yi(t), ,yn(t)). Since

M P + Q ) = P'ivit)) = Vyi{t){Pl,P2 + y2(i), ,Pn + Vn{t)),


we conclude that
y{t) = (i,q2,q3,-.-,qn)-
Taking derivatives, we finally obtain y' = ( 1 ,0 ,..., 0).

Now we consider other types of orbits (see Figure 1.2).

D efinition 1.55. An orbit {x{t) : t E (a, 6)} that is not a critical point is
said to be:
a) periodic if there exists T > 0 such that x{t + T) = x{t) whenever
t,t -\-T E {a, b)]
1.5. Phase portraits 37

b) homoclinic if there exists a critical point p such that x{t) p when


t >o+ and when t ^b~\
c) heteroclinic if there exist critical points p and q with p q such that
x{t) p when t -> a"*" and x{t) q when t -^b~.

F ig u re 1.2. 1) periodic orbit; 2) homoclinic orbit; 3) heteroclinic orbit.

We show that all orbits in Definition 1.55 and more generally all orbits
contained in a compact set have maximal interval R.

D efinition 1.56. A solution x = x{t) is said to be global if its maximal


interval is M.

P rop osition 1.57. Any solution of equation (1.85) whose orbit is contained
in a compact subset of D is global.

P ro o f. Let F : R x D he the function F{t,x) = f{x). Equation (1.85)


can be written in the form
x' = F{t,x).
Given an orbit {x(t) : t ( o , 6) } contained in some compact set K C D,
let us consider the compact set [m,m] x K C R x D. It follows from
Theorem 1.46 that there exists e > 0 such that
(t,x{t)) 0 [m,m] X K
for every t G (a, a + e) U (6 gr, 6). Hence, there exists e > 0 such that
t ^ [m,m] for the same values of t (since the orbit is contained in K).
Letting m oo, we conclude that a = oo and b +oo.

The following is a consequence of Proposition 1.57.


38 1. Ordinary Differential Equations

P rop osition 1.58. Critical points, periodic orbits, homoclinic orbits and
heteroclinic orbits are always obtained from global solutions.

1.5.2. Phase portraits. The phase portrait of an autonomous ordinary


differential equation is obtained representing the orbits in the set D, also in
dicating the direction of motion. It is common not to indicate the directions
of the axes, since these could be confused with the direction of motion.

Exam ple 1.59 (continuation of Example 1.53). We already know that the
orbits of the equation x' = x^ are {0 }, M"'" and R . These give rise to the
phase portrait in Figure 1.3, where we also indicated the direction of motion.
We note that in order to determine the phase portrait it is not necessary to
solve the equation explicitly.

F ig u re 1.3. Phase portrait o f the equation x ' = x^.

Exam ple 1.60. Consider the equation x' = |a;|. The only critical point is
the origin. Moreover, for a: 7^ 0 we have x' = || > 0, and hence, all solutions
avoiding the origin are increasing. Again, this yields the phase portrait in
Figure 1.3, that is, the phase portraits of the equations x' = x^ and x' = |a:|
are the same. However, the speed along the orbits is not the same in the
two equations.

Examples 1.59 and 1.60 illustrate that different equations can have the
same phase portrait or, in other words, the same qualitative behavior. We
emphasize that phase portraits give no quantitative information.

E xam ple 1.61. Consider the equation


{x,y)'= {y,-x) (1.86)
(see Example 1.3). The only critical point is (0,0). Moreover, by (1.6), if
(x, y) is a solution, then (x^ + y^)' = 0, and thus, other than the origin
all orbits are circles. In order to determine the direction of motion it is
sufficient to consider any point of each orbit. For example, at (x, y) = (, 0)
we have {x,y)' = (0, x), and thus, the phase portrait is the one shown in
Figure 1.4. We note that it was not necessary to use the explicit form of the
solutions in (1.7).

It is sometimes useful to write the equation in other coordinates, as


illustrated by the following example.
1.5. Phase portraits 39

F ig u re 1.4. Phase portrait o f the equation { x ,y ) = (y, x ).

E xam ple 1.62 (continuation of Example 1.61). Let us write equation (1.86)
in polar coordinates. Since x r cos 9 and y = r sin 6, we have

r '= ( V i ^ y -
2 -^x^ + y2 r
and
{y/x)' y'x - x'y
6' = farctan') = -
\ X/ 1 + {y/xY
Thus, equation (1.86) takes the form

r' = 0,

{ 9' = -\
in polar coordinates. Since r' = 0, all solutions remain at a constant dis
tance from the origin, traversing circles (centered at the origin) with angular
velocity 9' = \. Thus, the phase portrait is the one shown in Figure 1.4.

Exam ple 1.63. Consider the equation

{ x , y ) ' = {y,x). (1.87)


The only critical point is the origin. Moreover, if (x, y) is a solution, then
(x^ y^)' 2xx' 2yy' = 2xy 2yx = 0.
This shows that each orbit is contained in one of the hyperbolas defined
by x^ y^ = c, for some constant c ^ 0, or in one of the straight lines
bisecting the quadrants, namely x = y and x = y. In order to determine
the direction of motion in each orbit, it is sufficient to observe, for example.
40 I. Ordinary Differential Equations

that a;' > 0 for y > 0, and that aj' < 0 for y < 0. The phase portrait is the
one shown in Figure 1.5.

F ig u re 1.5. Pheise portrait o f the equation {x ,y )' = {y ,x ).

Exam ple 1.64. Consider the equation

(x' = y{x^ + 1),


( 1.88)
\y' = x{x^ + 1).

As in Example 1.63, the origin is the only critical point, and along the
solutions we have
(x'^ - y^Y = 2a;a;' - 2yy'
= 2xy{x^ + 1) ~ 2yx{x^ + 1) = 0.

Since the signs of y{x^ + 1 ) and x{x^ + 1) coincide respectively with the signs
of y and x, the phase portrait of equation ( 1.88) is the same as the phase
portrait of equation (1.87) (see Figure 1.5).

E xam ple 1.65. Consider the equation

n.89i
|y' = x{x^ - 1).

The critical points are obtained solving the system

y{x^ 1) = x{x^ 1) = 0,
1.5. Phase portraits 41

whose solutions are (0,0), (l,y ) and (l,y ), with y G K. Moreover, along
the solutions we have
{x^ - y^)' = 2xx' - 2yy'
= 2xy{x^ 1) ~ 2yx{x^ 1) = 0,
and thus, each orbit is again contained in one of the hyperbolas defined
by x^ y^ c, for some constant c ^ 0, or in one of the straight lines
X = y and x = y. The phase portrait of equation (1.89) is the one shown
in Figure 1.6, where the direction of motion can be obtained, for example,
from the sign of y' x{x^ 1); namely,
y '> 0 for a; G (1,0) U (1 ,+oo)

and
y' <Q for a; G (oo, 1) U (0,1).
In particular, it follows from the phase portrait that any orbit intersecting
the vertical strip
L = {{x,y) G 1 < a: < 1}
is contained in L and is bounded, while any other orbit is unbounded.

F ig u re 1.6. Phase portrait o f equation (1.89).

The following example uses complex variables.

Exam ple 1 .6 6 . Consider the equation

fa;' = a;^ - y^,


(1.90)
[ y' = 2xy.
42 1. Ordinary Differential Equations

Letting z = x + iy, one can easily verify that equation (1.90) is equivalent
to z' = where z' = x' + iy'. Dividing by ^ yields the equation

i.
r2
whose solutions are
1
z{t) = - (1.91)
t+ c
with c 6 C. In order to obtain x{t) and y(t), it remains to consider the real
and imaginary parts of z(t) in (1.91). Namely, if c = a + z6 with o, 6 M,
then

One can use these formulas to verify that the phase portrait is the one shown
in Figure 1.7.

1.5.3. Conservative equations. In this section we consider the particu


lar case of the differential equations conserving a given function. We first
introduce the notion of conservative equation.

D efinition 1.67. A function D -> R of class that is constant in no


open set and that is constant along the solutions of the equation x' = f(x)
is said to be an integral (or a first integral) of this equation. When there
exists an integral the equation is said to be conservative.

We note that a function E of class is constant along the solutions of


the equation x' = f{ x) if and only if

j^E{x{t)) = 0

for any solution x = x{t). More generally, one can consider integrals in
subsets of D and integrals that are not of class C^.
1.5. Phase portraits 43

Exam ple 1.68. It follows from identity (1.6) that the function
defined by E{x,y) = is an integral of equation (1.4), and thus, the
equation is conservative.

Exam ple 1.69. Let E: D R he a function of class that is constant


in no open set. Then the equation in H x D C given by

x' = dE/dy,
(1.92)
y' = - d E j d x
is conservative and the function E is an integral. Indeed, if (x, y) is a solution
of equation (1.92), then
d . dE , dE ,
(1.93)
= -y 'x ' + x'y' = 0,

which shows that E is an integral. Incidentally, equation (1.4) is obtained


from (1.92) taking
x^ + y2
E{x,y) =

E xam ple 1.70. Now we consider equation (1.92) with

E{x,y) = xy(x + y - 1), (1.94)

which can be written in the form


x' = x^ + 2xy X,

{y' = - 2xy - y^ + y.
In view of (1.93), in order to obtain the phase portrait of equation (1.95)
(1.95)

we have to determine the level sets of the function E in (1.94). These are
sketched in Figure 1.8, where we have also indicated the direction of motion.
One can easily verify that the critical points of equation (1.95) are (0,0),
(1,0), (0,1) and (1/3,1/3). Now we show that the orbits in the interior of
the triangle determined by (0, 0), (1 , 0) and (0, 1), in a neighborhood of the
critical point (1/3 ,1/3), are in fact periodic orbits. Writing X = x 1/3
and Y = y 1/3, we obtain

^^fev) = ( x + i ) ( r + ^ ) ( : v + K - i )

= - 4 + + Y^) + (X + Y)XY.
2i{ o
Since the quadratic form

+ XT + = 1 1/2
(1.96)
1/2 1
44 1. Ordinary Differential Equations

F ig u re 1.8. Phase portrMt o f equation (1.95).

is positive definite (the 2 x 2 matrix in (1.96) has eigenvalues 1/2 and 3/2),
it follows from Morses lemma (Proposition 9.16) that there is a coordinate
change ( X , ? ) = g{X,Y) in neighborhoods of (0,0) such that

E(x,y) = - ^ + ^{X^ + X Y + Y^).

Since the level sets of E are closed curves in the variables ( X , y ) , in a


neighborhood of (0, 0), the same happens in the variables {x,y), now in a
neighborhood of (1/3 ,1/3).

Exam ple 1.71. Consider the equation

ix' = y,
(1.97)
\y' = f{x),
where
f{x)=^x{x-l){x-3). (1.98)
The critical points are (0, 0), ( 1, 0) and (3 ,0). Now we consider the function

E{x,y) = ^y^ - f{s)ds

I 2 I 4 .4 0 3o
= + 3^=' -
If (x,y) = {x{t),y(t)) is a solution of equation (1.97), then

j^E{x,y) = yy' - f{x)x' = yf{x) - f {x )y = 0, (1.99)

which shows that the function E is an integral. In order to obtain the phase
portrait, we have to determine the level sets of the function E. One can
1.5. Phase portraits 45

show that these are the curves sketched in Figure 1.9, which determines the
phase portrait. In particular, there exists a homoclinic orbit connecting the

F ig u re 1.9. Phase portrait o f equation (1.97) with f given by (1.98).

In order to determine the level sets of the function E, we proceed as


follows. We first write equation (1.97) in the form
x" = y' = fix).
This corresponds to apply a force f{ x) to a particle of mass 1 moving without
friction. For example, f{x) can be the gravitational force resulting from the
particle being at a height f{x). Actually, in order to obtain the phase
portrait it is not important what f {x ) really is, and so there is no loss
of generality in assuming that f{ x) is indeed the gravitational force. In
this context, identity (1.99) corresponds to the conservation of the total
energy E{x, y), which is the sum of the kinetic energy y^/2 with the potential
energy
= f{s)ds = ~ x ^ + ^x^

The function V has the graph shown in Figure 1.10, which can be obtained
noting that
V'{x) = x(x l)(a; 3).
The phase portrait can now be obtained from the energy conservation.
For example, it follows from Figure 1.10 that a ball dropped with velocity
2/ = 0 at a point x in the interval (0, (8 \ /l0)/3) starts descending and after
some time passes through the point x = 1, where it attains maximal speed.
Then it starts moving upward, in the same direction, up to a point x' G
(0, (8 \/l0)/3) with V{x') = V (x), where again it attains the velocity y = 0.
This type of behavior repeats itself indefinitely, with the ball oscillating
46 1. Ordinary Differential Equations

F ig u re 1.10. Graph o f the function V in Example 1.71. The points o f


intersection with the horizontal aocis are 0, (8 \ /l0 )/3 and (8 + \ /l0 )/3 .

back and forth between the points x and x'. This corresponds to one of
the periodic orbits in Figure 1.9, in the interior of the homoclinic orbit.
On the other hand, a ball dropped with velocity y > 0 at a point x (0,1)
with /2 + V {x) = 0 (by Figure 1.10 the function V is negative in the
interval (0, 1)) travels along the same path up to a point x' with V{x') =
V{x), although now without zero velocity, because of the conservation of
energy. Instead, it continues traveling up to a point x" G (l,+ o o ) with
-\-V{x) = V{x"), where it finally attains zero velocity (the existence of
this point follows from Figure 1.10 and the conservation of energy). Then
the ball starts traveling backwards, in particular, passing through the initial
point , and approaching the origin indefinitely, without ever getting there,
since
y'^ + V{x) = Q = E{0,0).
Analogously, a ball dropped with velocity y < 0 at a point x 6 (0 , 1) with
y^/2 + F(o;) = 0 exhibits the same type of behavior. Namely, it approaches
the origin indefinitely, without ever getting there. This corresponds to the
homoclinic orbit in Figure 1.10. The remaining orbits can be described in
an analogous manner.

E xam ple 1.72. Now we consider equation (1.97) with

f {x ) = x{x - l)(a: - 2). (1.100)


The critical points are (0,0), (1,0) and (2,0). Moreover, the function

E{x,y) = ^y^ - f{s)ds


1 101)
( .

= ^2/^ - + x^-x^
1.5. Phase portraits 47

is an integral of equation (1.97). The phase portrait is the one shown in


Figure 1.11. In particular, there exist two heteroclinic orbits connecting the
critical points (0, 0) and (2, 0).

F ig u re 1.11. Phase portrait o f equation (1.97) with / given by (1.100).

In order to determine the level sets of the function E in (1.101), we


proceed in a similar manner to that in Example 1.71. In this case the
potential energy is given by

V(x) = f f{s) ds 7 ^ + x^,


Jo 4
and has the graph shown in Figure 1.12.

F ig u re 1.12. Graph o f the function V in Example 1.72. The points o f


intersection with the horizontal axis are 0 and 2.
48 1. Ordinary Differential Equations

Now we use the conservation of energy to describe the phase portrait.


In particular, for a ball dropped with velocity y = 0 at a point x G ( 0 ,2 ) ,
we obtain an oscillatory movement between x and 2 x. Indeed,

V(x) + x^ x^ = ^x^(x 2)^,

and hence V(2 x) = V(x). This corresponds to one of the periodic orbits
in Figure 1.11. On the other hand, a ball dropped with velocity y > 0 at
a point X G ( 0 , 2) with energy y^/2 + F (x) = 0 approaches indefinitely the
point 2, without ever getting there, since E(2, 0) = 0. Also, considering the
movement for negative time, we obtain the heteroclinic orbit in the upper
half-plane in Figure 1.11. The heteroclinic orbit in the lower half-plane
corresponds to the trajectory of a ball dropped with velocity y < 0 at a
point X G ( 0 , 2) with energy y^/2 -\-V{x) = 0. The remaining orbits in
Figure 1.11 can be described in an analogous manner.

1.6. Equations on manifolds

This section is a brief introduction to the study of ordinary differential equa


tions on (smooth) manifolds. After recalling the relevant notions concerning
manifolds and vector fields, we consider ordinary differential equations de
fined by vector fields on manifolds and we give several examples.

1.6.1. B asic notions. In this section we recall some basic notions concern
ing manifolds. We first introduce the notion of a differentiable structure.

D efinition 1.73. A set M is said to admit a differentiable structure of


class and dimension n if there exist injective maps ipi: Ui M defined
in open sets Ui C E for i G / such that:

a)
b) for any i,j G I w ith V = Pi{Ui)r\(pj{Uj) 0 the preimages >~^(V)
and are open sets in ' \ and the map ^o (^j is of class

It is sometimes required that the family (Ui,(pi)i^i is maximal with re


spect to the conditions in Definition 1.73, although one can always complete
it to a maximal one. Each map ipi: Ui M is called a chart or a coordinate
system. Any differentiable structure induces a topology on M ; namely, a set
A c M is said to be open if c R is an open set for every i G /.
Now we introduce the notion of a manifold.

D efinition 1.74. A set M is said to be a manifold of class and dimen


sion n if:
a) M admits a differentiable structure of class and dimension n;
1.6. Equations on manifolds 49

b) when equipped with the corresponding induced topology, the set M


is a Hausdorff topological space with countable basis.

We recall that a topological space is said to be Hausdorff if distinct


points have disjoint open neighborhoods, and that it is said to have a count
able basis if any open set can be written as a union of elements of a given
countable family of open sets.

Example 1.75. The real line R is a manifold of dimension 1 when equipped


with the differentiable structure composed of the single chart </?: R > R
given by (p(x) = x.

Example 1.76. The circle


= {{x,y) G R^ : = 1}
is a manifold of dimension 1. A differentiable structure is given by the maps
(-l,l)^ 5 ^ z = l,2,3,4
defined by
(pi{x) = (x, V l -a ;2),

(P2{x) = ( a ; , - \ /l -a ;2),
( 1. 102)
ipsix) = { y / l - x ^ , x ) ,

(P4{x) = { - V i - x^,x).

One can show in a similar manner that the n-sphere


5^^ = (x G R +^ : ||x|| = 1},
where
/n + l \ 1/2

11(^1....... ^+i)ii = (

is a manifold of dimension n.
Example 1.77. The n-torus T " = 5^ x x is a manifold of dimension n.
A differentiable structure is given by the maps '0: (1 ,1) T defined by
1p{xi,...,Xn) = {lllixi),...,i>n{xn)),
where each il^i is any of the functions </?i, ^3 and in (1 .102).

Example 1.78. Let (p: U R^ be a function of class in an open


set [/ C R . Then the graph
M = {{x,ip{x)) : X G 17} C R X R"^
is a manifold of class and dimension n. A differentiable structure is given
by the map 0 : 17 0 (x) = (x,y?(x)).
R x R^ defined by
50 1. Ordinary Differential Equations

Exam ple 1.79. Let E: U E"* be a function of class in an open set


{7 C E . A point c E* is said to be a critical value of E if c = E{x) for
some X e U such that dxF: E " )E"^ is not onto. One can use the Inverse
function theorem to show that if c E^ is not a critical value of E, then
the level set
= { x e U : E{x) = c }
is a manifold of dimension n m.
For example, ii E: U > E is an integral of a conservative equation
x' = f{x), then for each c G E that is not a critical value of E, the level set
E~^c is a manifold of dimension n 1.

1.6.2. V ector fields. In order to introduce the notion of a tangent vector,


we first recall the notion of a differentiable map between manifolds.

D efinition 1.80. A map / : M > AT between manifolds M and N is said


to be differentiable at a point a: G M if there exist charts (p: U M and
Tp: V N such that:
a) a: G p{U) and C i>{Vy,
b) 'ip~^ o / o is differentiable at ip~^{x).
We also say that / is of class in an open set W C M if all maps tf}~^ofo<p
are of class in
Now we introduce the notion of a tangent vector to a curve. Let M
be a manifold of dimension n. A differentiable function a : (e, e) M is
called a curve. Also, let D be the set of all functions / : M ^ E that are
differentiable at a given point x G M.

D efinition 1.81. The tangent vector to a curve a: (s,e) -> M with


a(0) = a: at t = 0 is the function F: D ->-R defined by
d (f o a)
H f) = dt t=o
We also say that F is a tangent vector at x.

One can easily verify that the set TxM of all tangent vectors at x is a
vector space of dimension n. It is called the tangent space of M at x. The
tangent bundle of M is the set
T M = {{x,v) : X G M ,u G TxM}.
One can show that T M is a manifold of dimension 2n. A differentiable
structure can be obtained as follows. Let ip\ U )M be a chart for M and let
( x i , ... ,x) be the coordinates in U. Consider the curves aii (e,e) M
for i = 1, . . . , n defined by
ai{t) = <p{tei),
1.6. Equations on manifolds 51

where (e i,...,e n ) is the standard basis of R . The tangent vector to


the curve o-j at t = 0 is denoted by djdxi. One can easily verify that
{d/dx\,.. ., dfdxn) is a basis of the tangent space N o w we con
sider the map : 17 x R ^ TM defined by

tl}{xi,...,Xn,y\,...,yn) =
i=l
for some chart (p: U M. One can show that the maps of this form define
a differentiable structure of dimension 2n in TM.
Now we are ready to introduce the notion of a vector field.
D efinition 1.82. A vector field on a manifold M is a map X \M TM
such that X{x) T^M for each x G M.

Using the notion of differentiability between manifolds, one can also


speak of a differentiable vector field.

1.6.3. D ifferential equations. Let M be a manifold. Given a continuous


vector field X : M -^T M, consider the ordinary differential equation
x' = X{x). (1.103)
More generally, one can consider time-dependent vector fields.

D efinition 1.83. A curve x: (a,b) M of class is said to be a solution


of equation (1.103) if x'{t) = X(x{t)) for every t (o, 6), where x'{t) is the
tangent vector to the curve s x{t -I- s) at s = 0.

Using charts one can establish the following manifold version of the
Picard-Lindelof theorem (Theorem 1.18).

T h eorem 1.84. If X is a differentiable vector field on M , then for each


(to,xo) R X M there exists a unique solution of the initial value problem
ix' = X{x),
| (to ) = Xq
in some open interval containing to.

More generally, all local aspects of the theory of ordinary differential


equations in R can be extended to arbitrary manifolds. In addition to
the Picard-Lindelof theorem, this includes all other results concerning the
existence of solutions (such as Peanos theorem) and the dependence of
solutions on the initial conditions (see Theorems 1.40 and 1.42).
On the other hand, some nonlocal aspects differ substantially in the case
of differential equations on manifolds. In particular, the following result is
a consequence of Theorem 1.46.
52 1. Ordinary Differential Equations

T h eorem 1.85. I fX is a differentiable vector field on a compact manifold,


then all solutions of the equation x' = X (x) are global, that is, all solutions
are defined for t G M.

In a similar manner to that in Proposition 1.13, this implies that for a


differentiable vector field on a compact manifold the solutions of the equation
x' X (x) give rise to a fiow. More precisely, the following statement holds.

T h eorem 1 .8 6 . If X is a differentiable vector field on a compact mani


fold M and x{t,xo) is the solution of the initial value problem

W = X{x),
|a;(0) = XQ,

then the family of transformations <pt: M M defined by <pt(xo) = x(t,xo)


is a flow.

The following are examples of differential equations on manifolds.

Exam ple 1.87. Let M = S^. Given a vector field X : M > TM, each
vector X(x) is either zero or is tangent to the circle. If X does not take the
value zero, then the whole 5^ is an orbit of the equation x' = X(x). If X
has exactly p zeros, then there are 2p orbits.

Exam ple 1 .8 8 . Let M = 5^ c and consider a vector field X : M - TM


such that for x E other than the North and South poles the vector X (x)
is nonzero and horizontal. Then the equation x' = X (x) has critical points
at the North and South poles and all other orbits are horizontal circles
(obtained from intersecting the sphere with horizontal planes).

Exam ple 1.89. Let M T^ = S^xS^. Since 5^ can be seen as the interval
[0, 1] with the endpoints identified, the torus can be seen as the square
[0,1] X [0,1] with two points x ,y E Q identified whenever x y E l? . Thus,
a differential equation in the torus can be seen as the equation

(1.104)
{: y' = 9{x,y)

in for some functions / , ^ : )R such that

f{x-{-k,y-\-l) = f { x ,y ) and g { x k , y 1) = g{x,y)


for every a;, y G R and k,l e Z.
For example, if / = a and y = ,0 for some constants a, /3 G R, then the
solutions of equation (1.104) give rise to the flow

^tix, y) = {x-\- at, y + fit) mod 1 .


1.7. Exercises 53

If a = 0, then the image of each solution is a vertical periodic orbit. If


a ^ 0 and P/a Q , then the image of each solution is a periodic orbit with
slope P/a. Finally, a ^ 0 and P/a ^ Q, then the image of each solution
is dense in the torus.

A detailed study of the theory of differential equations on manifolds


clearly falls outside the scope of the book. In particular, some topological
obstructions influence the behavior of the solutions of a differential equation.
For example, any vector field X on S'^ has zeros and thus, the equation
x' = X{x) has critical points. We refer the reader to [3, 1 1 ] for detailed
treatments.

1.7. Exercises

E xercise 1.1. For the equation x' = f{x), show that if cost is a solution,
then sin t is also a solution.

E xercise 1.2. For the equation x" f{x):


a) show that if 1 /(1 + 1) is a solution, then 1 /(1 t) is also a solution;
b) And / such that 1 /(1 + 1) is a solution.

E xercise 1.3. Find an autonomous equation having t ^ /(l+ t ) as a solution.

E xercise 1.4. For the equation {x,y)' (y,x) in E^, writing


x{t) = rcosh^(t) and y(t) = r sinh. 6(t),
with r 6 E, show that 6{t) = t + c for some constant c E.
E xercise 1.5. Let / : E E be a Lipschitz function and let p: be
a continuous function. Show that for the equation

(x' = f{x),
\y' = 9{x)y
the initial value problem (1.13) has a unique solution.

Exercise 1 .6 . Let / : (o, 6) ^ E \ {0 } be a continuous function,


a) Show that the function g: (a, 6) ^ E defined by

is invertible for each xq M.


b) Show that the initial value problem (1.13) has a unique solution.
Hint: For a solution x(t), compute the derivative of the function t ^
g{x{t)).
54 1. Ordinary Diiferential Equations

Exercise 1.7. Let / : be a function of class and let (, ) be an


inner product in R .

a) Show that if (f(x),x) < 0 for every x R , then each solution of


the initial value problem (1.13) is defined for all t > to-
b) Show that if 5 : R R is a differentiable function such that
9{x) > ||x|p and {f{x),Vg{x)) < 0
for every x R " \ {0 } (where ||xp = {x,x) and Vg is the gradient
of g), then each solution of the initial value problem (1.13) is defined
for all t > to.

E xercise 1.8. Write the equation

I x' = -a y ,
[y ' = ax
in polar coordinates.

Exercise 1.9. Write the equation

{ x' = ex y x(x^ + y^),


y' = x + e y - y{x^ + y^)
in polar coordinates.

Exercise 1.10. Let u,v,w: [a,5] >R be continuous functions with to > 0
such that ^
u{t) < v{t) + / w{s)u{s) ds
Ja
for every t G [0 , 6]. Show that

u{t) < v{t) + J w{s)v{s) exp ^ J to(tt) du^ ds

for every t [a, 6].

Exercise 1.11. Let T: X X he a transformation of a complete metric


space X such that is a contraction for some m G N. Show that:
a) T has a unique fixed point xq GX ;
b) for each x E X the sequence T (x) converges to xq when n > 00.

Exercise 1 .1 2 . Show that the function A 1-^ ||.A|| defined by (1.40) is a norm
in the space M of n x n matrices with real entries (see Definition 1.26).

Exercise 1.13. Show that the norm H-H defined by (1.40) satisfies
11^511 < P ll \\B\\
for every A, B E Mn-
1.7. Exercises 55

E xercise 1.14. Consider the set Z c C{I) (see Proposition 1.30) of all
Lipschitz functions x: I > R in a given bounded set / C R^, with the
same constant L in (1.38), such that a:(0) = 0 (when 0 I). Show that Z
is a complete metric space with the distance
r
d{x, y) = sup - :t l\ {0 }
}
Exercise 1.15. Let (pt be a flow defined by an equation x' = f{x). Say how
/ can be obtained from the flow.

Exercise 1.16. Show that if / : R ^ R is a bounded function of class C^,


then the equation x' = f {x ) defines a flow such that each function ipt is a
homeomorphism (that is, a bijective continuous function with continuous
inverse).

Exercise 1.17. Given T > 0, let / : R x R ^ R be a Lipschitz function


such that
f{t,x) = f{t + T,x) for (t ,a :)e R x R .
Show that for each (to,xo) G R x R " the solution x{t,to,xo) of the initial
value problem (1.13) has maximal interval R and satisfies the identity

x{t, to, xq) = x{t + T,to + T, xq) for t G R.

Exercise 1.18. Sketch the phase portrait of the equation x" = x^ x, that
is, of the equation (x, y)' = (y, x^ x) in R^.

Exercise 1.19. Sketch the phase portrait of the equation:

\x' = y{y^ - x"^),


a)
[y' = -a:(y2 - x^);

lx' = - y + x{l - x ^ - 2/2),


b)
[ 2 / ' = a; + 2/(1 - a ;2 - 2/2);

^x' a;(10 a;2 y^),


c)
y = y (l - a ;2 - y 2).

Exercise 1 .2 0 . Find all periodic orbits of each equation in Exercise 1.19.

Exercise 1.21. For each A G R, sketch the phase portrait of the equation:
a) x" = \x{x 1);
b) x" = x{x A);
(x' = - y + x{X-x'^ - 2/2),
1 y '= a; + 2/ ( A - x 2 - 2/2).
56 1. Ordinary Differential Equations

E xercise 1 .2 2 . Find an integral for the equation


fx' = x + 3y^,
= -2x-y.

Exercise 1.23. Consider the equation in polar coordinates


fr' = 0,
yO' = (r^ l)(r^ cos^ 0 + r sin0 + 1).
a) Sketch the phase portrait.
b) Find all global solutions.
c) Find whether the equation is conservative.
d) Show that in a neighborhood of the origin the periods T (r) of the
periodic orbits satisfy
T{r) = 27T + ar^ + o(r^),
for some constant o ^ 0 (given a function R ^ R such that
g{x)/x^ 0 when a; ^ 0, we write g{x) = o{x'^)).

Solutions.
1.2 b) f {x ) = 2x^.
1.3 [(a:'+ a;)/(2 - a:')]' = 1, that is, {x,yy = {y,2{y - l){y - 2)/{2 + x)).
1.8 {r,0y (0,a).
1.9 {r,0y = (er r, 1).
1.15 f {x ) = {d/dt)(pt{x)\t=o.
1.20 a) There are no periodic orbits.
b) {(a;, y) G R^ : a;2 + = 1}.
c) There are no periodic orbits.
1 .2 2 E{x, y) = x'^ + xy + y^.
1.23 b) All solutions are global,
c) It is conservative.
Chapter 2

Linear Equations and


Conjugacies

This chapter is dedicated to the study of the qualitative properties of linear


ordinary differential equations and their perturbations. We first introduce
the notion of a fundamental solution, which contains complete information
about the solutions of a linear equation. In particular, we describe the fun
damental solutions of the equations with constant coefficients and periodic
coefficients. We then give a complete description of the phase portraits of
the equations with constant coefficients in dimension 1 and dimension 2.
We also present the variation of parameters formula for the solutions of
the perturbations of a linear equation. Finally, we introduce the notion of
a conjugacy between the solutions of two linear differential equations with
constant coefficients, and we discuss the characterization of the differen
tiable, topological and linear conjugacies in terms of the coefficients of the
equations. For additional topics we refer the reader to [4, 15, 17].

2.1. N o n a u to n o m o u s lin ea r e q u a tio n s

In this section we consider equations in R of the form

x' = A{t)x, (2.1)

where the matrices A{t) Mn vary continuously with t G R (we recall that
Mn is the set of n x n matrices with real entries). Equation (2.1) is called
a linear equation.

2 .1 .1 . Space o f solutions. One can easily verify that f{ t,x ) = A{t)x is


continuous and locally Lipschitz in x. It follows from the Picard-Lindeldf
58 2. Linear Equations and Conjugacies

theorem (Theorem 1.18) that the initial value problem (1.13), that is,

| x ' = A W x,
[x{to) = xo,
has a unique solution in some open interval containing to-

P rop osition 2 .1 . All solutions of equation (2.1) have maximal interval R.

P roof. We consider only times t > to- The argument for t < to is entirely
analogous. If x = x{t) is a solution of equation (2.1), then

x(
:{t) = x{to) + [ A{s)x{s) ds
Jto
for t in some open interval containing to? thus,

'WII < IN(fo)ll + / ||.4(5)||.|| i (5)||<Js


a;
Jto
for t > to in that interval. It follows from Gronwalls lemma (Proposi
tion 1.39) that

l|x(()ll<IW(o)l|exp f \\A(,s)\\ds,
Jto
for the same values of t. Since the function t A{t) is continuous, the
integral 11-4(5)11 ds is well defined and is finite for every to and t. Thus,
it follows from Theorem 1.46 that the maximal interval of the solution x
is M.

Using the notion of global solution in Definition 1.56, one can rephrase
Proposition 2.1 as follows: all solutions of equation (2.1) are global.
The following result shows the set of all solutions of equation (2.1) is a
linear space.

P rop osition 2 .2 . If X \ , X 2 : are solutions of equation (2.1), then


cixi -I- C2X2 is a solution of equation (2.1) for each ci,C2 R.

P roof. We have
( c i X i + C 2 X 2 )' = C \x 'i C2X2

= C l A { t ) x i + C2A{t)x2
= A { t ) { c \ X i + C2X2),
and thus, c \ X \ -I- 020:2 is a solution of equation (2.1).

Exam ple 2.3. Consider the equation


0 1 X
(2.3)
-1 0
2.1, Nonautonomous linear equations 59

whose solutions are given by (1.7), with r > 0 and c G M. We have


_ / r cos c cost + r sine sin t
(^) / cos c sin t + r*sin c cos
f cost \ . /sint^
= r cos cl . ^ + r sin c
\smtJ \cost
This shows that the set of all solutions of equation (2.3) is a linear space of
dimension 2, generated by the (linearly independent) vectors
(co st,sin t) and (sin t, cost).

More generally, we have the following result.

P rop osition 2.4. The set of all solutions of equation (2.1) is a linear space
of dimension n.

P roof. Let e i , . . . , be a basis of R^. For i = 1 ,..., n, let = Xi{t) be


the solution of the initial value problem (2.2) with xq = Ci. For an arbitrary
xo G R^, the solution of problem (2.2) can be obtained as follows. Writing
xo = CiCi, one can easily verify that the function
n
x{t) =
i=l
is the solution of equation (2.1) with x{to) = xq. In particular, the space of
solutions of equation (2.1) is generated by the functions Xi{t),... ,Xn{t). In
order to show that these are linearly independent, let us assume that
n
'Y^CiXi{t) = 0
i=l
for some constants c i , . . . , c^ 6 R and every t R. Taking t = to, we obtain
S r= i hence ci = C2 = = Cn = 0 (because e i , . . . , Cn is a
basis). Thus, the functions a;i(t),. . . , Xn{t) are linearly independent.

2.1.2. Fundamental solutions. In order to describe all solutions of equa


tion (2.1) it is useful to introduce the following notion.

Definition 2.5. A function X{t) with values in M whose columns form


a basis of the space of solutions of equation (2.1) is called a fundamental
solution of the equation.

We note that any fundamental solution is a function X : R Mn satis


fying
X \t) = A{t)X{t) (2.4)
for every f G R, with the derivative of X (t) computed entry by entry.
60 2. Linear Equations and Conjugacies

Now let X{t) be a fundamental solution of equation (2.1). For each


vector c G R , the product X{t)c is a linear combination of the columns
of X{t). More precisely, if x i {t ),... ,Xn(t) are the columns of X{t) and
c = ( c i ,.. .,Cn), then
n
X { t ) c ^ '^CiXijt).
i= l

This shows that the solutions of equation (2.1) are exactly the functions of
the form X{t)c with c G R , where X{t) is any fundamental solution.

E xam ple 2 .6 . Let / : R ) R be a function of class such that the


equation x' = f {x ) has only global solutions. Given a solution (pt{x) with
= X, consider the matrices

^{t) =
The equation
y' =
is called the linear variational equation of x' = f{x) along the solution (ftix).
It follows from the identity

^V?t(a:) = f{ipt{x))

that if the map {t^x) is sufficiently regular, then


d
-^dx<pt d(p^(jc)f dx^t A.{t)dx^t

This shows that the function X (t) = dx<pt satisfies the identity

X'{t) = A{t)X{t),

and thus also


{X{t)c)' = A{t){X{t)c)
for each c G R . Since

X (0 ) = dx<po = dxld - Id,

the columns of X{t) are linearly independent, and hence X{t) is a funda
mental solution of the linear variational equation of x' = f {x ) along the
solution ipt{x).

Now we establish an auxiliary result.

P rop osition 2.7. If X{t) is a fundamental solution of equation (2.1), then


its columns are linearly independent for every t G R.
2.1. Nonautonomous linear equations 61

P r o o f. Let , Xn(t) be the columns of a fundamental solution X(t).


Given to G M, we show that the vectors xi(to), , n(to) are linearly inde
pendent. Assume that
n
'^CiXiito) = 0
i=l
for some constants c i , . . . , c K. Then the solution
n
(2.5)
i=l
of equation (2.1) satisfies x{to) = 0, and it follows from the uniqueness of
solutions that x{t) 0 for every t R. Since the functions a:i(t),. . . , Xn{t)
form a basis of the space of solutions (because X{t) is a fundamental so
lution), it follows from (2.5) that ci = = Cn = 0. Thus, the vectors
xi (to)) ) Xn{to) are linearly independent.

Proposition 2.7 has the following consequence.

P rop osition 2.8. The solution of the initial value problem (2.2) is given by

x{t) = X{t)X(to)~^xo, t M, (2.6)


where X{t) is any fundamental solution of equation (2.1).

P ro o f. It follows from Proposition 2.7 that the matrix X{to) is invertible,


and thus, the function x{t) in (2.6) is well defined. Moreover,

x{to) = X (to)A (to) ^a;o = a^O)


and it follows from identity (2.4) that

x'{t) = X\t)X{to)-'^xo
= A{t)X{t)X{to)-'^xo = A{t)x{t).
This shows that x{t) is the solution of the initial value problem (2.2).

Any two fundamental solutions are related as follows.

P rop osition 2.9. If X{t) andY{t) are fundamental solutions of equation


(2.1), then there exists an invertible matrix C Mn such thatY(t) = X{t)C
for every t M.

P ro o f. Since X (t) and Y(t) are fundamental solutions, each column of Y(t)
is a linear combination of the columns of X (t) and vice-versa. Hence, there
exist matrices C,D Mn such that

Y{t) = X { t ) C and X{ t ) = Y{ t ) D
62 2. Linear Equations and Conjugacies

for every < M. Thus,

Y(t) = X {t)C = Y{t)DC. (2.7)

On the other hand, by Proposition 2.7, the columns of the matrix Y{t) are
linearly independent for each t G K. Therefore, Y{t) is invertible and it
follows from (2.7) that DC = Id. In particular, the matrix C is invertible.

It also follows from Proposition 2.7 that \iX{t) is a fundamental solution
of a linear equation, then detX{t) ^ 0 for every t e R . In fact, we have a
formula for the determinant.

T h eorem 2 .1 0 (Liouvilles formula). If X{t) is a fundamental solution of


equation (2.1), then

det X(t) = det X(to) exp f tr A(s) ds


Jto
for every t, to G M.

P roof. One can easily verify that

^ d e tX (t) = ^ d e t A ( i ) , (2.8)
i=l
where Di{t) is the n x n matrix obtained from X{t) taking the derivative of
the ith line, and leaving all other lines unchanged. Writing X(t) = {xij{t))
and A(t) = (oij(t)), it follows from identity (2.4) that
n

k=l
for every t M and i ,j = l , . . . , n . In particular, the ith line of Di{t) is
/ n n \

(^il (0) > I ^ ^ik(t)^kl (Q? ? ^ik(i)^kn(l') j


.A;=l k=l

^ ^^ik(l')(^kl{l')^ j ^kn(l'))
k=l
Since the last sum is a linear combination of the lines of X{t), it follows
from (2.8) and the properties of the determinant that

d ^
det X(t) = ^ au(t) det X(t)
CuU
i=l
= tr A(t) d etX (f).
2.2. Equations with constant coefficients 63

In other words, the function u{t) = detX{t) is a solution of the equation


u' = trA (t). Therefore,

u{t) = u{to) exp tr A{s)ds,


Jto

which yields the desired result.

2 .2 . Equations w ith constant coefficients

In this section we consider the particular case of the linear equations in


of the form
x' = Ax, (2.9)

where .A is an n x n matrix with real entries. We note that the function


f{ t,x ) = Ax is of class C^, and thus, it is continuous and locally Lipschitz
m X.

2.2.1. E xponential o f a m atrix. In order to solve equation (2.9), we first


recall the notion of the exponential of a matrix.

D efinition 2 .1 1 . We define the exponential of a matrix A Mn by


oo
( 2. 10)
A;=0
with the convention that e = Id.

P rop osition 2 .1 2 . The power series in (2.10) is convergent for every ma


trix A.

P ro o f. One can easily verify that ||j4*'|| < ||>1||*^ for each k and thus,

E inl-^11s E
OO

k=0
OO

k=0
..

<+~- (211)
In other words, the series in (2.10) is absolutely convergent, and thus it is
also convergent.

Exam ple 2.13. If A = Id, then A'^ = Id for each k e N, and thus.

1
e = l i d = eld.
^ A:!
fc=o
64 2. Linear Equations and Conjugacies

Exam ple 2.14. If ^4 is an n x n diagonal matrix with entries A i,. . . , An in


the diagonal, then

OO 1 /A j 0
V - V
^ k\ ^ k\
k=0 k=0 U A^
1 \A: 0 \
' Z^fc=0

0 Y^k=0 E^n)
/e^i 0 \

0 e-^/

Exam ple 2.15. Consider the matrix

0 Id
^ \ -ld 0

where Id is the identity. We have

A^ = ld, A^ = A, ^2 = -I d , A ^ ^ -A , A'^^ld,

and thus = A'^ for each k e N . Hence,

At

gtt ^ g < , e** e .


IdH------- r:----- A
2 2i
= costid + sintj4.

In order to describe a somewhat expedited method to compute the ex


ponential of a matrix, we first recall a basic result from linear algebra.

T h eorem 2.16 (Jordan canonical form). For each A Mn, there exists an
invertible n x n matrix S with entries in C such that

(R i 0
S~^AS = 2 12
( . )
U Rk
2.2. Equations with constant coefEcients 65

where each block R j is an Uj x nj matrix, for some Uj < n, of the form


fXj 1 0 \

Rj (2.13)
1
VO /
where Xj is an eigenvalue of A.

Using the Jordan canonical form in Theorem 2.16, the exponential of a


matrix can be computed as follows.

P rop osition 2.17. If A is a square matrix with the Jordan canonical form
in (2.12), then
0 \
= Se^ = S (2.14)
U
P ro o f. We have {S ^^45)*^ = S ^A'^S for each fc 6 N. Therefore,
OO .

k=0

= S - ' { ^ l ^ y = S-^e^S.

On the other hand, one can easily verify that the exponential of a matrix in
block form is obtained computing separately the exponential of each block,
that is.
fe^^ 0 \
S~^AS

lo g-Rfcy
This yields the desired result.
When A consists of a single block R j as in (2.13), the exponential
can be computed explicitly as follows.

P rop osition 2.18. If a matrix A G Mn has the form A = Aid + N, where


/O 1 0 \

N = (2.15)
1
\o 0 /
66 2. Linear Equations and Conjugacies

then
I Id + + - N ^ + +
1 N'n1 (2.16)
( n - 1)!

P roof. Since the matrices Aid and N commute, one can show that

(see Exercise 2.4). By Example 2.14, we have = e^Id. On the other


hand, since N'^ = 0, we obtain
oo .4 nn 1
X

^k\ ^k\
k=0 k=0
and thus,
n 1
e'' = e^Id V iw * =

This yields identity (2.16).


2 .2 .2 . Solving the equations. In this section we describe the relation
between a linear equation x' = Ax and the exponential of a matrix. We
start with an auxiliary result.

P rop osition 2.19. We have (e^*)' = Ae-^* for each t M, with the deriv
ative of e^* computed entry by entry.

P roof. We have

k=0
Since

E ^E n11*^11"=<+~>
CX) .. OO

k=o k=o
each entry of the matrix is a power series in t with radius of conver-
gence +oo. Hence, one can differentiate the series term by term to obtain
oo .
(e^*)' = E I'k t^-^A'^
k=l **
OO .

= Ae^
for each t R.

Now we find all solutions of equation (2.9).


2.2. Equations with constant coefRcients 67

P rop osition 2 .2 0 . Given {to, xq) M x M , the solution of the equation


x' Ax with initial condition x{to) = xq is given by
x{t) = e^^^-^'>xo, t R. (2.17)

P ro o f. By the Picard-Lindelof theorem (Theorem 1.18), the desired solu


tion is unique, and by Proposition 2.1 its maximal interval is M. To complete
the proof, it is sufficient to observe that for the function x{t) in (2.17) we
have
x{to) = e^^xo = e^XQ = Idcco = xq,
and by Proposition 2.19,
x'{t) = = Ax{t).
This shows that x{t) is the desired solution.

2.2.3. Phase portraits. In this section we describe the phase portraits of


all linear equations with constant coefficients in M and in R^. The following
example gives a complete description for scalar equations.

E xam ple 2.21. Consider the scalar equation x' = ax, where a G R. Its
phase portrait, shown in Figure 2.1 , depends only on the sign of the con
stant a. Namely, when a = 0 there are only critical points. On the other
hand, when o ^ 0 the origin is the only critical point, and the remaining
orbits R"*" and R have the direction indicated in Figure 2.1 .

a= 0

a> 0

o < 0

F ig u re 2.1. Phase portrait of the equation x' = ax.

Now we consider linear equations in R^ and we describe their phase


portraits based on the Jordan canonical form.

E xam ple 2.22. We first consider a matrix A G M 2 with Jordan canonical


form
(Xr 0\
VO X2J
for some Ai,A2 R (we emphasize that we are not assuming that A is of
this form, but only that its Jordan canonical form is of this form). Let
68 2. Linear Equations and Conjugacies

vi,V2 G be eigenvectors of A associated respectively to the eigenvalues


Ai and A2- The solutions of the equation x' = Ax can be written in the form
x{t) = cie^^*ui + C2e^'^*V2, t G M, (2.18)
with ci,C2 G M. Now we consider several cases.

Case Ai < A2 < 0. For C2 ^ 0, we have


x{t)
V2 (2.19)
C26\2t
when t + 00, and hence, the solution x(t) has asymptotically the direction
of V2 (when t + 00). This yields the phase portrait shown in Figure 2.2,
and the origin is called a stable node.

F ig u re 2.2. Case Ai < A2 < 0.

Case Ai > A2 > 0. Analogously, for C2 ^ 0 property (2.19) holds when


t 00, and hence, the solution x{t) has asymptotically the direction of V2
(when t > 00). This yields the phase portrait shown in Figure 2.3, and
the origin is called an unstable node.

Case Ai < 0 < A2. Now there is expansion along the direction of V2 and
contraction along the direction of vi. The phase portrait is the one shown
in Figure 2.4, and the origin is called a saddle point.

Case Ai = 0 and A2 > 0. The solution in (2.18) can now be written in the
form
x{t) = civi + C2e^^*V2, t G M, (2.20)
with ci,C2 G R. Thus, for C2 7^ 0 the solution traverses a ray with the
direction of V2. On the other hand, the straight line passing through the
2.2. Equations with constant coefEcients 69

F ig u re 2.3. Case Ai > A2 > 0.

F ig u re 2 .4 . Case Ai < 0 < A2.

origin with the direction of vi consists entirely of critical points. This yields
the phase portrait shown in Figure 2.5.

Case Ai = 0 and A2 < 0. The solution is again given by (2.20). Analo


gously, the phase portrait is the one shown in Figure 2.6.

Case Ai = A2 = A < 0. The solution in (2.18) can now be written in the


form
x{t) = e^*xo, t M,
with xo M^. We note that there is no privileged direction, unlike what
happens in Figure 2.2, where there is an asymptote. Thus, the phase portrait
is the one shown in Figure 2.7, and the origin is called a stable node.
2.2. Equations with constant coefficients 71

Case Ai = A2 > 0. Analogously, the phase portrait is the one shown in


Figure 2.8, and the origin is called an unstable node.

F ig u re 2 .8 . Case Ai = A2 > 0.

Case Ai = A2 = 0. In this case there are only critical points.

Now we consider the case of nondiagonal Jordan canonical forms with


real eigenvalues.

E xam ple 2.23. Consider a matrix A G M 2 with Jordan canonical form

(ol)
for some A G K. One can easily verify that the solutions of the equation
x' = Ax can be written in the form
x{t) = [(c i + C2t)vi + C2U2]e^ , t G E, (2.21)
with ci,C2 M, where ^;l,U2 G \ {0 } are vectors such that
Av\ = Avi and Av2 = Avi + V2- 2 22
( . )
Whenever ci ^ 0 or C2 0, we have
x{t)
/ \T7 ^ "^1
(ci + C2t)e^*
when t >+ 00 and when t >00. Now we consider three cases.

Case A < 0. The phase portrait is the one shown in Figure 2.9, and the
origin is called a stable node.

Case A > 0. The phase portrait is the one shown in Figure 2.10, and the
origin is called an unstable node.
72 2. Linear Equations and Conjugacies

Vl

F ig u re 2.9. Case A < 0.

Vl

F ig u re 2.10. Case A > 0.

Case A = 0. The solution in (2.21) can now be written in the form

x{t) = {civi + C2V2) + C2tvi, t E M,

with ci,C2 G M, where vi,V2 G \ {0 } are vectors satisfying (2.22). In


particular, the straight line passing through the origin with the direction of
2.2. Equations with constant coefficients 73

v\ consists entirely of critical points. The phase portrait is the one shown
in Figure 2.11.

F ig u re 2.11. Case A = 0.

Finally, we consider the case of nonreal eigenvalues. We recall that for


matrices with real entries the nonreal eigenvalues always occur in pairs of
conjugate complex numbers. Moreover, the eigenvectors associated to these
eigenvalues are necessarily in \ and thus, it is natural to use complex
variables.

Exam ple 2.24. Let A M 2 be a 2 x 2 matrix with eigenvalues a + ih


and a ib, for some 6 7^ 0. The solutions of the equation x' = Ax in are
given by
x{t) = + C2C t R,
with ci,C2 G C, where vi,U2 \ { 0} are eigenvectors associated respec
tively to a + ib and a ib. Since the matrix A has real entries, one can take
V2 = vT- Indeed, it follows from Av\ = (a 4- ib)v\ that

Av\ = Av\ = (a -1- ib)vi = (o ib)v\y


and thus, v\ is an eigenvector associated to the eigenvalue o ib. Hence,
taking C2 = cT, we obtain
x{t) = e*[ci cos{bt) -I- ciisin(6t)]ui -I- e*[cTcos(6t) cTi sin(6t)]wr
(2.23)
= 2e*cos(6t)Re(cit;i) 2e sin(6t) Im (civi).

We note that the vectors

ui -- 2Re(ciVi) and U2 = 21m(civi)


are in R^. Now we consider three cases.
74 2. Linear Equations and Conjugacies

Case a = 0. In this case the solution in (2.23) can be written in the form

x{t) = cos{bt)ui + sm{bt)u2-

Thus, the phase portrait is the one shown in Figure 2.12 or the one obtained
from it by reversing the direction of motion (this corresponds to change the
sign of b). The origin is called a center.

F ig u re 2.12. Case a = 0.

Case a > 0. The phase portrait is the one shown in Figure 2.13, and the
origin is called a stable focus.

F ig u re 2.13. Case a > 0.

Case a < 0. The phase portrait is the one shown in Figure 2.14, and the
origin is called an unstable focus.
2.3. Variation o f parameters formula 75

2.3. V a ria tio n o f p a ra m e te rs fo rm u la

In this section we present a result giving the solutions of a class of perturba


tions of the linear equation (2.1). We continue to assume that the matrices
A(t) Mn vary continuously with f G R.

T h eorem 2.25 (Variation of parameters formula). Let 6: R > R be a


continuous function. Given (to>a^o) R x R , the solution of the initial
value problem
ix' = A{t)x + b{t),
(2.24)
|a:(to) = xo

has maximal interval R and is given by

x(t) = X{t)X{to)-'^xo + f X { t ) X { s ) - % s ) ds, (2.25)


Jto

where X{t) is any fundamental solution of equation (2.1), with the integral
computed component by component.

P ro o f. We first observe that the function

f{ t ,x ) = A{t)x + b{t)

is continuous and locally Lipschitz in x. Thus, by the Picard-Lindeldf theo


rem (Theorem 1.18), the initial value problem (2.24) has a unique solution.
Moreover, since the function s i-> X{t)X{s)~^b{s) is continuous (because the
entries of the inverse of a matrix B are continuous functions of the entries
of B), the integral in (2.25) is well defined and is finite for every to and t.
Thus, it follows from Theorem 1.46 that the function x{t) is defined for all
76 2. Linear Equations and Conjugacies

t g R. Moreover, x{to) = xq, and using (2.4) we obtain

x'{t) = X'{t)X{to)-'^xo + f* X ' { t ) X { s ) - \ s ) ds + X{t)X{t)-^b{t)


Jto

= A{t)X{t)X{to)-'^xo + A{t) f X(t)X{s)-^bis) ds + b{t)


Jto
= A{t)x{t) + b{t).

This yields the desired result.

The equations

x' = A{i)x and x' = A{t)x + b{t)

are often referred to, respectively, as homogeneous equation and nonhomo-


geneous equation.
We also obtain a Variation of parameters formula in the particular case
of perturbations of the autonomous linear equation (2.9).

T h eorem 2.26 (Variation of parameters formula). Let 6: R R be a


continuous function. Given (to, xo) R x R ", the solution of the initial
value problem
x' = Ax + b{t),
(2.26)
{: x{to) = Xo
has maximal interval R and is given by

a;(t) = ds. (2.27)


Jto

P ro o f. By Theorem 2.25, the initial value problem (2.26) has a unique


solution, with maximal interval R. Hence, it is sufficient to observe that for
the function x{t) in (2.27) we have

x(to) = e^a;o = Idxo = xo,


as well as

x'{t) = t Ae^(*-)6(s) ds + e^^*-%(t)


Jto

= + e ^ ( * - ^ ^ b ( s ) ds^ + b (t)

= Ax(t) + b{t).

This completes the proof of the theorem.


2.3. Variation o f parameters formula 77

We note that Theorem 2.26 is not an immediate consequence of Theo


rem 2.25. This is due to the fact that if X{t) is a fundamental solution of
equation (2.1), then in general may not be equal to X {t s),
although this property holds for equations with constant coefficients (see
Exercise 2.3).
Now we establish an auxiliary result.

P rop osition 2.27. If a matrix A has only eigenvalues with negative


real part, then there exist constants c , d > 0 such that
< ce-<^\ t > 0, (2.28)
with the norm in (1.40) obtained from any given norm ll-H in R .

P ro o f. Since the matrix A has only eigenvalues with negative real part, it
follows from (2.14) and (2.16) that for each entry aij{t) of there exist
constants Cij, dij > 0 and a polynomial pij such that
\aij (t) I < Cije~^^A (i) I, t > 0.

When Pij is not identically zero, we have

limsupylog(e"^'^ |py(t)|) = - d i j ,
t-^+OO ^
and hence, given e > 0, there exists Cij > 0 such that
e~^^i%j{t)\ < CijC-^'hj-e)t^ j. > 0.

This yields the inequality


|aii(i)| < CijCije~^hi~e)t^ t>0.
Taking e sufficiently small so that
d := min {dij e : i ,j = 1,... ,n] > 0,
we obtain
|ay(t)| < ce t > 0, (2.29)
where
c = max {cijCij : i ,j = 1 ,..., n }.
Now we recall that all norms in R* are equivalent, that is, given two norms
II-II and II'll', there exists C > 1 such that
c - lM < iH '< c | | x | |
for every x R . Since Mn can be identified with R*^^, taking m = n^
it follows from (2.29) that inequality (2.28) holds for some constant c.

The following is an application of Theorem 2.26 to a particular class of


equations of the form x' = Ax + b{t).
78 2. Linear Equations and Conjugacies

Exam ple 2.28. Under the hypotheses of Theorem 2.26, let us assume that
the matrix A has only eigenvalues with negative real part and that the
function b is bounded. Using (2.28), it follows from Theorem 2.26 that the
solution X = x{t) of the initial value problem (2.26) with to = 0 satisfies

lk(t)|| < c e - ^ ||a;o||+ t ds


Jq
for t > 0, where K = sup^>o ll^(^)ll- Thus,

|Wt)||<oe- ||xoll + ^ ( l - e - * ) A :

for t > 0. In particular, the solution is bounded in R"*".

2.4 . Equations w ith periodic coefficients

In this section we consider the particular case of the linear equations with
periodic coefficients. More precisely, we consider equations in R'^ of the form
x' = A{t)x, (2.30)
where the matrices A{t) G vary continuously with t G R, and we assume
that there exists T > 0 such that
A{t + r ) = A{t) (2.31)
for every t G R.

D efinition 2.29. Given T > 0, a function F : R is said to be T-


periodic if
F{t + T) = F{t) for every t G R.
The function F is said to be periodic if it is T-periodic for some T > 0.

According to Definition 2.29, the constant functions are T-periodic for


every T > 0.

Exam ple 2.30. Consider the equation


x' - a{t)x,
where a: R R is a periodic continuous function. The solutions are given
by ^
x{t) = exp o(s) d^x{to), t G R.

For example, when a{t) = 1 we have x{t) = e* *a;(to), and the only periodic
solution is the zero function. On the other hand, when a(t) = cos t we have
x{t) =
and for a;(to) ^ 0 the solution is a nonconstant 27r-periodic function.
2.4. Equations with periodic coefRcients 79

Now we describe the fundamental solutions of the linear equations with


periodic coefficients. We continue to denote by M the set of n x n matrices.

T h eorem 2.31 (Floquet). If A: E ^ is a T-periodic continuous func


tion, then any fundamental solution of equation (2.30) is of the form
X{t) = P{t)e^\ (2.32)
where B and P{t) are n x n matrices for each t E, with
P (t + T) = P (t), teR . (2.33)

P ro o f. It follows from (2.4) that if X(t) is a fundamental solution of equa


tion (2.30), then
X'{t + T) = A{t -h T)X{t + T) = A{t)X{t -h T) (2.34)
for t G E. Hence, Y{t) = X{t-\- T) is also a fundamental solution of equa
tion (2.30), and by Proposition 2.9 there exists an invertible n x n matrix C
such that
X {t + T ) = X {t )C (2.35)
for t E. On the other hand, since C is invertible, there exists an n x n
matrix B such that = C. It can be obtained as follows. Let S be an
invertible n x n matrix such that S~^CS has the Jordan canonical form
in (2.12). For each matrix Rj = A^-Id -|- Nj, with ^ 0 (because C is
invertible) and Nj an Uj x nj matrix as in (2.15), we define

log Rj = lo g A,- Id-b


aJ J
( _ l ) m + l ^ :m
3
^{logXj)ld-\-J2 m\f
m=l
where logAj is obtained from branch of the complex logarithm. One can
verify that = R j for each j (see for example [16]). This implies that
the matrix
''log R\ 0
B = ^S (2.36)
0 log Rkj
satisfies = C. Now we define n x n matrices
P{t) = X{t)e~^^
for each t e R. It follows from (2.35) (see also Exercise 2.3) that
P{t -I- T) = X(t +

= X(t)e-^* = P{t).
80 2. Linear Equations and Conjugacies

This completes the proof of the theorem.


We note that the matrix B in (2.32) is never unique. Indeed, if = C,
then
exp [(5 + m{2m/T)ld)T] = C
for every m g Z. We also observe that the matrix P(t) is invertible for every
t 6 M, because by Proposition 2.7 all matrices X{t) are invertible.
We continue to consider a T-periodic continuous function A{t) and we
introduce the following notions.

D efinition 2.32. Given a fundamental solution X{t) of equation (2.30):


a) an invertible matrix C M such that X{t + T) = X{t)C for every
t R is called a monodromy matrix of equation (2.30);
b) the eigenvalues of a monodromy matrix C are called characteristic
multipliers of equation (2.30);
c) a number A G C such that is a characteristic multiplier is called
a characteristic exponent of equation (2.30).

In fact, both the characteristic multipliers and the characteristic expo


nents are independent of the monodromy matrix that was used to define
them. This is an immediate consequence of the following result.

P rop osition 2.33. Let .A: R ) Mn he a T-periodic continuous function.


If X (t) and Y (t) are fundamental solutions of equation (2.30) with mon
odromy matrices, respectively, C and D, then there exists an invertible n x n
matrix S such that
S-'^CS = D. (2.37)

P roof. The monodromy matrices C and D satisfy


X{t-\-T) = X {t)C and Y{t-i-T) = Y { t ) D (2.38)
for every t G R. On the other hand, it follows from (2.34) that X{t-\-T) and
Y{t -\-T) are also fundamental solutions. Thus, by Proposition 2.9, there
exists an invertible n x n matrix S such that
Y{t -HT) = X {t -1- T)S
for every t G R. Therefore,
Y(t + T) = X{t)CS = Y{t)S~'^CS. (2.39)
Comparing (2.39) with the second identity in (2.38) yields (2.37).

Now we give some examples of linear equations with periodic coefficients,


starting with the particular case of constant coefficients.
2.4. Equations with periodic coefficients 81

E xam ple 2.34. When there exists a matrix A G Mn such that A{t) = A
for all t E, one can take any constant T > 0 in (2.31). It follows from
Floquets theorem (Theorem 2.31) that there exist matrices B and P{t) such
that
= P{t)e^^ (2.40)
for every t G R. Since At and AT commute, we have
g4(t+T) ^

(see Exercise 2.3), and thus C = is a monodromy matrix of equa


tion (2.30). Now let S be an invertible n x n matrix such that has
the Jordan canonical form in (2.12). Proceeding as in (2.36), one can take

B = i5 1 o g (S -lg 4 T .S)S~^
T
1
= ^<51oge(^ -^AS)Tg-\

/log 0

log 6^*=^
0 \

RkTj
0 \
S~^ = A,
Rkj
and it follows from (2.40) that P{t) = Id for every t G R.
For example, taking T = 1, a monodromy matrix of the equation x' = Ax
is C = e^. Hence, the characteristic multipliers are the numbers G C,
where A i,. . . , A are the eigenvalues of A, and the characteristic exponents
are A i,. . . , A, up to an integer multiple of 27tz.
Exam ple 2.35. Consider the linear equation with 27r-periodic coefficients
/x Y _ / l 2 -|- sin t ^
\yj ~ VO - c o s t /( 2 + sint)y '
One can easily verify that the solutions are given by

<2/(0/ \c/(2-|-sint)y V l/(2 + sint) 0^


with c, d G R. Hence, a fundamental solution is
-1
^ (0 =
1/(2 -t- sinf)
82 2. Linear Equations and Conjugacies

which yields the monodromy matrix


C = X(0)"^X(27 t)
^0 2\ f - 1 o27T

.1 2 ^ 1 / 2 0
Therefore, the characteristic multipliers are 1 and

The following result shows that the characteristic exponents determine


the asymptotic behavior of the solutions of a linear equation with periodic
coefficients.

P rop osition 2.36. Let A : R > Mn be a T-periodic continuous function.


Then X is a characteristic exponent of equation (2.30) if and only if e^*p{t)
is a solution of equation (2.30) for some nonvanishing T-periodic function p.

P roof. We first assume that e^^p{t) is a solution of equation (2.30) for


some nonvanishing T-periodic function p. It follows from Floquets theorem
(Theorem 2.31) that there exists xq R such that
e^*'p{t) P{t)e^^XQ
for t G R. Moreover, it follows from (2.33) that

that is.
.Bt BtBT
e^^P(t)e'^*xo = P(t)e-*e XQ.
Therefore,
P(t)e'*(e^^-e^^Id)xo = 0.
Since P{t)e^*' is invertible for each t R, the number e^^ is an eigenvalue
of the monodromy matrix and A is a characteristic exponent.
Now we assume that A is a characteristic exponent. Then e^^ is an
eigenvalue of It follows immediately from Propositions 2.17 and 2.18
that if jx is an eigenvalue of B, then is an eigenvalue of with the
same multiplicity. This implies that A is an eigenvalue of P, up to an integer
multiple of 27ri. Hence, there exists xq E R^\{0} such that B xq = Ascq, and
e^^XQ = e^^XQ
for t R. Multiplying by P(t), we obtain the solution
P{t)e^^XQ = e^^P{t)xQ = e^*p{t),
where p{t) = P{t)xo- Clearly, the function p does not vanish, and it follows
from (2.33) that
p{t + r ) = P{t + T)xo = P{t)xQ = p{t).
This completes the proof of the proposition.
2.4. Equations with periodic coefficients 83

The following example shows that in general the eigenvalues of the ma


trices A{t) do not determine the asymptotic behavior of the solutions.
E xam ple 2.37 (Markus-Yamabe). Consider the equation x' A{t)x in
with
. / . _ 1 / 2-1-3 cos^t 2 3 sin t cos t''
2 y2 3 sintcost 2-1-3sin^t
One can easily verify that the eigenvalues of the matrix A{t) are {li^/7)/4
for every t 6 M. In particular, they have negative real part. On the other
hand, a fundamental solution is
cos t e * sin t
X {t) = (2.41)
e*/^sint e~*cost
and so there exist solutions growing exponentially. Incidentally, by Propo
sition 2.36, it follows from (2.41) that the characteristic exponents are 1/2
and 1.

Proposition 2.36 also allows one to establish a criterion for the existence
of T-periodic solutions.

P rop osition 2.38. Let A: M ) Mn be a T-periodic continuous function.


If 1 is a characteristic multiplier of equation (2.30), then there exists a
nonvanishing T-periodic solution.

P ro o f. Let A G C be a characteristic exponent such that = 1. By


Proposition 2.36, there exists a nonvanishing T-periodic function p such
that x{t) = e^*p{t) is a solution of equation (2.30). We have
x{t -f T) = -h T)
= e^^e^p(t) = x{t)
for every t G R, and the function x is T-periodic.

Now we describe how one can obtain some information about the char
acteristic multipliers and the characteristic exponents without solving ex
plicitly the equation.

P rop osition 2.39. Let >1: R M be a T-periodic continuous function.


If the characteristic multipliers of equation (2.30) are pj e^i'^, for j =
1 , . . . ,n, then
pi
rT
J J p j = exp / tr.4(s)ds (2.42)
j=l ''0
and
1 /-T
= / tryl(s)ds (mod 27ri/T). (2.43)
T Jo
84 2. Linear Equations and Conjugacies

P roof. Let X {t) be a fundamental solution of equation (2.30) with mon-


odromy matrix C. It follows from (2.35) and Liouvilles formula (Theo
rem 2.10) that
det X {t + T)
d e tC =
detX (t)
(2.44)

/
t+T
rt+T
tr>l(s) ds = exp JP i

tr.4(s)ds,

where the last identity is a consequence of the periodicity of A. Since the


characteristic multipliers are the eigenvalues of C, identity (2.42) follows
immediately from (2.44). For identity (2.43), we first note that

n p ,= n e ^ ^ ^ = e x p (r x :A A
j= l j= l \ j= l /

Thus, it follows from (2.42) that

exp [ T 1 = exp / trA{s)ds.


\ j=i J
This yields the desired result.

Exam ple 2.40. Let A{t) be a T-periodic continuous function. We discuss


the existence of unbounded solutions in terms of the sign of
n
s= Rey^ Aj,
3=1
where the numbers Xj are the characteristic exponents. If s > 0, then there
exists j such that Re Xj > 0, and it follows from Proposition 2.36 that there
is an unbounded solution in R'*'. On the other hand, if s < 0, then there
exists j such that Re Xj < 0, and it follows from Proposition 2.36 that there
is an unbounded solution in R .

Exam ple 2.41. Consider the linear equation

\x' = x + y,
(2.45)
= X + (cos^ t)y.
We note that the matrix
1 1
A{t) =
1 cos^ t
is TT-periodic. Since
/*7T /7T
/ tr A(s) d s = (1 -|- cos^ s) ds > 0,
Jo Jo
2.5. Conjugacies between linear equations 85

it follows from (2.43) that


1
Re(Ai -|- A2) = / (1 "I" cos^ s) ds > 0,
7T Jq
where Ai and A2 are the characteristic exponents. Hence, by Example 2.40,
there exists a solution of equation (2.45) that is unbounded in R"*".

Exam ple 2.42. Consider the equation


x" + p{t)x = 0,
where p; M R is a T-periodic continuous function. Letting x' = y, the
equation can be written in the form
/
0 lA fx^
(2.46)
-p {t) 0) I y ,
We note that the matrix

A(t) =
-p(t) 0
is T-periodic. Since tr>l(t) = 0 for every t G R, it follows from Proposi
tion 2.39 that
rT
P1P2 = exp / tr j4( s) ds 1,
Jo
where pi and p2 are the characteristic multipliers. Taking A G C such that
= Pi and = p2, it follows from Proposition 2.36 that there exist
solutions e^*pi(t) and e~^*p2{t) of equation (2.46), where pi and p2 are
nonvanishing T-periodic functions.

2.5. Conjugacies between linear equations

In this section we describe how to compare the solutions of two linear differ
ential equations with constant coefficients. To that effect, we first introduce
the notion of conjugacy.

2.5.1. N otion o f con ju gacy. We begin with an example that illustrates


the problems in which we are interested.

E xam ple 2.43. Consider the scalar equations


x' = X and y' = 2y. (2.47)
The solutions are given respectively by
x{t) e*x{0) and y' = e^*y{0),
and have maximal interval R. The phase portrait of both equations in (2.47)
is the one shown in Figure 2.15.
86 2. Linear Equations and Conjugacies

F ig u re 2 .15. Pheise portrait o f the equations in (2.47).

This means that from the qualitative point of view the two equations
in (2.47) cannot be distinguished. However, the speed along the solutions
is not the same in both equations. Thus, it is also of interest to compare
the solutions from the quantitative point of view. More precisely, we want
to find a bijective transformation /i : R >M such that

h{e^x) = e^^h{x) (2.48)

for t, X R. Identity (2.48) can be described by a commutative diagram.


Namely, for each t G R we define transformations i^t, : R R by

ipt{x) = e*x and tptix)

One can easily verify that the families of transformations (ft and ipt are flows
(see Definition 1.11). Identity (2.48) can now be written in the form

h o(ft = tptoh,

which corresponds to the commutative diagram


<pt

I'*
>f>t

When identity (2.48) holds, the solution x{t) = e*x{0) of the first equa
tion in (2.47) is transformed bijectively onto the solution y{t) = e^ h(x(0))
of the second equation. In particular, we must have h{0) = 0, since the
only critical point of the first equation must be transformed onto the only
critical point of the second equation (or h would not transform, bijectively,
solutions onto solutions).
Assuming that h is differentiable, one can take derivatives with respect
to t in (2.48) to obtain

h'{e*x)e^x = 2e^*h(x).

Taking t = 0 yields the identity


h'{x) 2
= for X G R \ {0 }
h{x)
2.5. Conjugacies between linear equations 87

(we note that h(x) 7^ 0 for a; 7^ 0, because /i(0) = 0 and h is bijective).


Integrating on both sides, we obtain
log|/r(a;)| = log(a;^) + c
for some constant c This is the same as
h(x) = ax^
for some constant a E \ {0 }, with a; > 0 or a; < 0. In order that h is
bijective, we take
ax if a; > 0,
h{x) = < 0 if a; = 0, (2.49)
bx^ if a: < 0,
where o, 6 R are constants such that ab < 0. One can easily verify that
h is differentiable in R, with h'{0) = 0. Incidentally, the inverse of h is not
differentiable at the origin.

Now we consider arbitrary linear equations. More precisely, in an anal


ogous manner to that in Example 2.43, we consider the linear equations
x' = Ax and y' = By (2.50)
in R , where A and B are n x n matrices. The solutions are, respectively,
x{t) = e^ a:(0) and y{t) = e^*'y{0),
and have maximal interval R.

D efinition 2.44. The solutions of the equations in (2.50) are said to be:
a) topologically conjugate if there is a homeomorphism /i: R ^ R
(that is, a bijective continuous transformation with continuous in
verse) such that
h{e^^x) = e^^h{x) (2.51)
for every t R and a; 6 R ;
b) differentially conjugate if there is a diflFeomorphism h (that is, a bi
jective differentiable transformation with differentiable inverse) such
that identity (2.51) holds for every t G R and x E R ;
c) linearly conjugate if there is an invertible linear transformation h
such that identity (2.51) holds for every t e R and x E R .
We then say, respectively, that h is a topological, differentiable and linear
conjugacy between the solutions of the equations in (2.50).

In an analogous manner to that in Example 2.43, for each t E we


define transformations R ^ R by
(pt(x) = and ipt{x) = e^^x.
88 2 . Linear Equations and Conjugacies

Then identity (2.51) can be written in the form

ho(pt = iptoh,
which corresponds to the commutative diagram

ipt

Clearly, if the solutions of the equations in (2.50) are linearly conjugate,


then they are differentially conjugate (since any invertible linear transforma
tion is a diffeomorphism), and if the solutions are differentially conjugate,
then they are topologically conjugate (since any diffeomorphism is a home-
omorphism).

2 .5 .2 . Linear conjugacies. In this section we describe some relations be


tween the various notions of conjugacy. We first show that the notions of
linear conjugacy and differentiable conjugacy are equivalent for the linear
equations in (2.50).

P rop osition 2.45. The solutions of the equations in (2.50) are differentially
conjugate if and only if they are linearly conjugate.

P roof. As we observed earlier, if the solutions are linearly conjugate, then


they are differentially conjugate.
Now let h: K R " be a differentiable conjugacy. Taking derivatives
in identity (2.51) with respect to x, we obtain

d^At^he'^* = e^^dxh,
where dxh is the Jacobian matrix of h at the point x. Taking x = 0 yields
the identity
= e- C, (2.52)
where C = doh. We note that C is an invertible n x n matrix. Indeed,
taking derivatives in the identity h~^{h{x)) = x, we obtain

dh(o)h-^C = ld,
which shows that the matrix C is invertible. For the invertible linear trans
formation R'^ R given by g{x) = Cx, it follows from (2.52) that

5 (e^*x) = e^^g{x)

for every t R and x G R . In other words, 5 is a linear conjugacy.


2.5. Conjugacies between linear equations 89

By Proposition 2.45, in what concerns the study of conjugacies for lin


ear equations it is sufficient to consider linear conjugacies and topological
conjugacies.

Exam ple 2.46 (continuation of Example 2.43). We already showed that


the bijective differentiable transformations h: R -> R that are topologi
cal conjugacies between the solutions of the equations in (2.47) take the
form (2.49) with ab < 0. Since none of these functions is linear, there are no
linear conjugacies and it follows from Proposition 2.45 that there are also no
differentiable conjugacies. Indeed, as we already observed in Example 2.43,
the inverse of the function h in (2.49) is not differentiable.

Now we establish a criterion for the existence of a linear conjugacy in


terms of the matrices A and B in (2.50).

P rop osition 2.47. The solutions of the equations in (2.50) are linearly
conjugate if and only if there exists an invertible n x n matrix C such that

A = C~^BC. (2.53)

P ro o f. We first assume that the solutions are linearly conjugate, that is,
identity (2.51) holds for some invertible linear transformation h{x) = Cx,
where C is thus an invertible matrix. Then identity (2.52) holds and taking
derivatives with respect to t, we obtain

CAe^^ = Be^^C.

Finally, taking t = 0 yields CA = BC, and identity (2.53) holds.


Now we assume that there exists an invertible n x n matrix C as in (2.53).
Then
x' = Ax x' = C~^BCx (Cx)' = B{Cx).
Solving the first and third equations, we obtain

x{t) = e"^*a;(0) and Cx{t) = e^*Ca:(0)

for every t K and x(0) 6 M . Therefore,

Ce^*x(0) = e-*C'x(0),

and since the vector x(0) is arbitrary we obtain (2.52). Hence, identity (2.51)
holds for the invertible linear transformation h{x) = Cx.

It follows from Proposition 2.47 that the solutions of the equations in


(2.50) are linearly conjugate if and only if A and B have the same Jordan
canonical form.
90 2. Linear Equations and Conjugacies

2.5.3. T op olog ica l conjugacies. In this section we consider the notion


of topological conjugacy. Since it is not as stringent as the notions of dif
ferentiable conjugacy and linear conjugacy, it classifies less equations. In
other words, the equivalence classes obtained from the notion of topologi
cal conjugacy have more elements. However, for the same reason, it allows
us to compare linear equations that otherwise could not be compared. For
example, one can obtain topological conjugacies between the solutions of
equations whose matrices A and B have different Jordan canonical forms,
in contrast to what happens in Proposition 2.47 with the notion of linear
conjugacy.
A detailed study of topological conjugacies falls outside the scope of the
book, and we consider only a particular class of matrices.

D efinition 2.48. A square matrix A is said to be hyperbolic if all of its


eigenvalues have nonzero real part.

We recall that the eigenvalues of a matrix vary continuously with its


entries (for a proof based on Rouches theorem see, for example, [5]). Thus,
a matrix obtained from a sufficiently small perturbation (of the entries of)
a hyperbolic matrix is still hyperbolic.

D efinition 2.49. For a square matrix A, we denote by m{A) the number


of eigenvalues of A with positive real part, counted with their multiplicities.

Now we establish a criterion for the existence of topological conjugacies


between equations with hyperbolic matrices.

T h eorem 2.50. Let A and B be hyperbolic n x n matrices. If

m{A) = m(B), (2.54)

then the solutions of the equations in (2.50) are topologically conjugate.

P roof. We first assume that m{A) 0. We then explain how one can obtain
the result from this particular case. We construct explicitly topological
conjugacies as follows.
Step 1. Construction of an auxiliary function. Define a function 5 : R R
by
POO

q (x )= /
Jo
where H-H is the norm in R'^ given by
n \ 1/2
||(xi,...,a:n)|| = ( '^ x :
2 \
i=l
2.5. Conjugacies between linear equations 91

Since A has only eigenvalues with negative real part, by Proposition 2.27
there exist constants c,/n>0 such that ||e"^*|| < for every t > 0. This
implies that
fO O POO

/ dt < / < + 00,


Jo Jo
and thus, the function q is well defined. Denoting by B* the transpose of a
matrix we observe that
P^oo
OO

q{x) = / {e^*'x)* e^^xdt


Jo (2.55)
x*{e^^y e^^xdt = x*Cx,
= /
where POO

Jo
C= / *e^^dt

is an n X n matrix. This shows that g is a polynomial of degree 2 in


without terms of degree 0 or 1. In particular,
/ a; \ ^
(2.56)
llN M IMP
for a; 0. Thus, taking
a = min{g(a;) : ||a:|| = 1 } and P = max{g(a;) : ||x|| = 1},
it follows from (2.56) that
a||a;f < q{x) < P\\x\^. (2.57)

Step 2. Construction of the conjugacy. Let x y o . We note that


||e"^*a;|| >0 when s +oo
and
||e"^x|| >+ 0 0 when s -> oo. (2.58)
The first property follows immediately from Proposition 2.27. For the second
property, we note that the matrix is nonsingular (see Exercise 2.3), and
hence 0 for x 7^ 0. Now we consider the root spaces of A, that is,
= {x C : (A AId)^x = 0 for some A: G N}
for each eigenvalue A of A. We recall that = 0_j^ F\, where the direct
sum is taken over all eigenvalues of A, counted without their multiplicities.
Hence,
E^ = 0 (R nF A ).
A
In particular, given x \ {0 }, there exists at least one eigenvalue A of A
such that the component of x in R n F\ is nonzero. Thus, by (2.16), in
92 2. Linear Equations and Conjugacies

order to establish property (2.58) it is sufficient to show that all nonzero


entries of the matrix
Bt 1 _^7711jym
= e^Hld + tN + --- +
(m 1)!
where B = Aid + iV is an m x m Jordan block, have the same property. To
that effect, we note that if A has negative real part and p{t) is a nonzero
polynomial, then
\e^^p{t)\ > +0 0 when t oo.

We proceed with the proof of the theorem. It follows from (2.57) that
the image of the function s i-> q{e^^x) is On the other hand, we have
poo pc
q(e^^x)= / \\e^^*+^^xfdt= / dt, (2.59)
Jo Js

and thus, the function s q{e^^x) is strictly decreasing. In particular,


there exists a unique tj, G R such that q{e^^^x) = 1. Now we define a
transformation h : R >R " by

e~-*e"^*^a;/g(e^ 'x )i/2 if x 7^ 0,


h{x) = (2.60)
0 if X = 0,
where
poo
q {x )= ||e^*x|pdt. (2.61)
Jo
Since
g(g>lOx-<)e^*x) = q{e^^"^x) 1,
we have = tx t. Therefore,
^ - B { t x - t ) p A { U - t ) At
X
h{e^*'x) = q,(gy4(txf) QAtg.'^1/2

= e^*h(x).
g(gAtxx)l/2

which establishes identity (2.51). In order to show that h is a topological


conjugacy, it remains to verify that h is a homeomorphism.
Step S. Continuity of h. It follows from (2.59) and the Implicit function
theorem that the function x ^- tx is differentiable (outside the origin). On
the other hand, proceeding as in (2.55), one can show that 5 is a polynomial
of degree 2 in R , without terms of degree 0 or 1. Thus, it follows from (2.60)
that the transformation h is differentiable outside the origin. Now we show
that h is continuous at x = 0. Proceeding as in (2.56) and taking
a = min{g(x) : ||x|| = 1 } and ^ max{g(x) : ||x|| = 1 },
2.5. Conjugacies between linear equations 93

we obtain
a||a;f < q{x) < B\\x\\^. (2.62)
By (2.62), in order to show that h is continuous at the origin it is sufficient
to prove that q{h{x)) 0 when a; > 0. By (2.60), for a; 0 we have

q{e
q{h{x)) =
q{e^*^^x)
On the other hand, it follows from (2.61) that
poo
q{e~^^^e^^^x) = / < ||e -**||^g(e^ *a;),
^0
and hence,
q{h{x))<\\e-^^^f. (2.63)
Now we observe that, by (2.59),
poo
q{e^*^x) = / ||e'^*a;|pdt = 1.
Jtx
Thus, tx 00 when x > 0. Since B has only eigenvalues with negative
real part, it follows from (2.63) that q{h{x)) > 0 when x 0. This shows
that h is continuous in E .

Step 4- Construction of the inverse of h. We define a transformation


g \E " -> E by

{ q-A sxqBsx'x/g(e'*x)^/^
^ if X ^ 0,
g{x) =
0 if X = 0,

where Sx is the unique real number such that


poo
q{e^^^x) = / ||e^ x|pdt = 1
^ Sx

(it can be shown in a similar manner to that for tx that the number Sx exists
and is unique). We have

r
g{e^ -h{x)) = / dt = 1 ,
/o g(e^*x)
and the uniqueness of implies that 5/1(3.) = ix- Thus, for x ^ 0 we have
g g Btx ^-Atx^
g(h{x)) =
g,(gBia;/j^(x))l/2g(g/lta:x)l/2
X
q,(gBtj;/j(3.))l/2g(gAt*x)l/2
94 2. Linear Equations and Conjugacies

Since

q{e^*^h{x)) = q q(e,Atxy.y.l2

q{e^^^x) 1
q{e^^^x)
we obtain g{h{x)) = x iox x ^ Q. We also have ^(h(0)) = 5 (0) = 0, which
shows that g is the inverse of h. In order to show that g is continuous one
can proceed as for h, simply interchanging the roles of A and B. Summing
up, h is a homeomorphism.

Step 5. Reduction to the case m{A) = 0. Finally, we describe how to reduce


the general case to the case of matrices with m{A) = 0. Making appropriate
changes of coordinates, one can always assume that the matrices A and B
have, respectively, the forms

('V 1 ) (It b_) (2.64)

where the indices + correspond to the eigenvalues with positive real part
and the indices correspond to the eigenvalues with negative real part. It
follows from (2.54) that the matrices A^ and have the same dimension,
say n+, and that the matrices A - and B - also have the same dimension,
say n_. We write x = (x+,x-), with x+ and x_ By (2.64),
the equations in (2.50) can be written, respectively, in the forms

x'+ = A+X+, = B+a;+,


and (2.65)
x'_ = A -X - x'_ = B -X -.

By the result for m{A) = 0 and the corresponding result for eigenvalues
with positive real part, there exist topological conjugacies and h_, re
spectively, between the solutions of the equations

x'^ = A^x+ and x'^ = B+a;+,

and between the solutions of the equations

x'_ = A - X - and x'_ = B -x~ .

One can easily verify that the transformation h : E > M" defined by

h {x + ,x -) = (h + {x + ),h -{x -)) (2.66)

is a topological conjugacy between the solutions of the equations in (2.65).



The following is an application of Theorem 2.50.
2.5. Conjugacies between linear equations 95

E xam ple 2.51. Consider the matrices


-1 1 -1
^ l= (~ n ^ .) > A 2 = ( ~ n = . -^4 =
-1 -1 0

The corresponding linear equations x' = AiX in have the phase portraits
shown in Figure 2.16. One can easily verify that the four matrices Ai are
hyperbolic, with m(Ai), respectively, equal to 0, 0,0 and 1 for i = 1 , 2,3,4.
Hence, it follows from Theorem 2.50 that the solutions of the first three
equations are topologically conjugate and that they are not topologically
conjugate to the solutions of the last equation. On the other hand, since the
four matrices have different Jordan canonical forms, it follows from Propo
sitions 2.45 and 2.47 that the solutions of these equations are neither differ
entially conjugate nor linearly conjugate.

A4

Figure 2.16. Phase portraits of the equations in Example 2.51.


96 2 . Linear Equations and Conjugacies

The proof of Theorem 2.50 can be used to construct topological conju


gacies explicitly. We describe briefly the construction. For each x ^ 0, let
tx K be the unique real number such that

dt = 1 (2.67)
/
Jtx

(it is shown in the proof of the theorem that tx is well defined). Then a
topological conjugacy h: R " M between the solutions of the equations
x' = Ax and y' = By is given by (2.60), that is,

_ j if x ^ 0,
h(x) = ( 2.68)
0 if a: = 0.

The following is a specific example.

Exam ple 2.52. Consider the matrices

"-1 0 -1 1
and B=
0 -1 0 -1

We have

At _ ( e * 0 _ /e -
e = and eBt
0 e - VO e - 'J

Writing x = (y, z), we obtain

We^^xf = e~^\y'^ +

and it follows from (2.67) that

\e^^x\\^dt = \e-'^^-{y^ + z^) = l.


fJtx A

Thus,

^ 1,

and
2 . ^2
e =
^ [ y + I log 2^
0 \ } \z
2.6. Exercises 97

In order to determine h, we note that


roo poo II
+ tz
/ \\e^^e^^-xfdt= / e - dt
Jo Jo W z
poc
= / e 2J2 + z^)dt
+ 2 tyz+ i^z^
Jo
= e-2tx + 2 y ^ + i^

+ yz + - z

Finally, by (2.68), we take h(0,0) = 0 and


y2 + ^;2 / z, r + ^
h{y,z) =
y2 + +

for {y,z) ^ (0, 0).

The problem of the existence of topological conjugacies is very different


in the case of nonhyperbolic matrices, as the following example illustrates.

Exam ple 2.53. Consider the equations

= and
\y = a x [y = by
for some constants a, 6 > 0. One can easily verify that they have the same
phase portrait; namely, the origin is a critical point and the remaining orbits
are circular periodic orbits centered at the origin and traversed in the neg
ative direction. However, when a ^ b the periods of the periodic orbits are
different in the two equations. Namely, in polar coordinates the equations
take respectively the forms

fr ' = 0, f r ' = 0,
and
\e' = a [0' = b,
and thus, the periods of the periodic orbits are, respectively, 2-n/a and 27c/b.
When a ^ b , this prevents the existence of a topological conjugacy between
the solutions of the two equations, because it would have to transform peri
odic orbits onto periodic orbits of the same period.

2.6. Exercises

E xercise 2 .1 . Compute for the matrix


0 -2
A=
0
98 2. Linear Equations and Conjugacies

E xercise 2 .2 . Find necessary and sufficient conditions in terms of a matrix


A M such that for the equation x' = Ax:
a) all solutions are bounded;
b) all solutions are bounded for t > 0;
c) all solutions converge to the origin.

E xercise 2.3. For a matrix A 6 Mn, consider the equation x' = Ax.
a) Use (2.10) to show that = Id for each t M. Hint: Compute
the derivative of the function t !-)
b) Show that
^A(t-s) ^ ^At^-As for every t,s e (2.69)

Hint: Take derivatives with respect to t.


c) Use Theorem 2.25 and identity (2.69) to give an alternative proof of
Theorem 2.26.
d) Show that dete"^ =

E xercise 2.4. Given matrices A ,B . Mn, show that if


[A,[A,B]] = [B ,[A ,B ]]^ i),
where [A, B] = BA AB, then
gAtgBf ^ ^{A+B)t^[A,B\tV^^ t e M.

Hint: Show that


x{t) =
is a solution of the equation x' = t[A, B]x for each xq

Exercise 2.5. Given a matrix A M , let


I g ifA glA _ Q %A
COSA = ------ T------ and sin A =
2i
Compute these functions for the matrices
^0 0 0^
B= and (7=11 0 0
.0 1 0>

Exercise 2 .6 . Consider the equation x' = a(t)x for a continuous function


a:
a) Find all solutions of the equation.
b) Identify the following statement as true or false: There exists a non-
rT
vanishing T-periodic solution if and only if Jq o(s) ds = 0.
2.6. Exercises 99

c) Identify the following statement as true or false: There exists an


unbounded solution in R"*" if and only if / q a(s) ds ^ 0 for some t > 0.

Exercise 2.7. For equation (2 .1), show that if trA(t) = 0 for every t R,
then given n linearly independent solutions x i,. ..,Xn, the volume deter
mined by the vectors x i ( x ) ,. . . , Xn{t) is independent of t.

Exercise 2.8. Given continuous functions f ,g : R )R, solve the equation

,yj ~ [g{t) m j U ;
Exercise 2.9. Let o , 6 : R > R be T-periodic continuous functions. Show
that if the equation x' = a{t)x has no T-periodic solutions other than the
zero function, then the equation x' = a{t)x -|- b{t) has a unique T-periodic
solution.

E xercise 2 .1 0 . Show that the equation


' x' = X cos^ t z sin(2t),
< y' = X sin(4t) -t- y sin t 4z,
2' = xsin(5t) 2 cost
has at least one unbounded solution.

E xercise 2 .1 1 . Show that there exist functions f ,g : such that the


equation
(x' = f(x ,y ),
(2.70)
\y' = g{x,y)
has the phase portrait in Figure 2.17.

E xercise 2 .1 2 . Construct topological conjugacies between the solutions of


the equations:
x' = 2x, [ x' = 3x,
and
y = -y,
x' = 2x y,
and y' = - 2y,
z' = 2z.

E xercise 2.13. Let A: R j >Mn be a continuous function and let x{t) be


a solution of the equation x' = A{t)x. Show that:
a) for each t > 0,

|x(t)l|<||x(0 )||exp f\\A{s)\\ds-,


Jo
100 2. Linear Equations and Conjugacies

F ig u re 2 .17. Phase portrait o f equation (2.70).

b) it r iMMii ds < oo, then ||x(t)|| converges when t +oo.

E xercise 2.14. Show that if x{t) and y{t) are, respectively, solutions of the
equations
x' = A{t)x and y' = A{t)*y
in M , then {x{t),y{t)) is independent of t, where (,) is the usual inner
product in R"'.

E xercise 2.15. Show that if X {t) is a fundamental solution of the equation


x' = A{t)x, then Y(t) = is a fundamental solution of the equation
y' = -A (t)*y.

E xercise 2.16. Show that if A{t) = A{t)* for every t M, then ||a;p is
an integral of the equation x' = A{t)x.

Exercise 2.17. Verify that the function h in (2.66) is a topological conju-


gacy.

Exercise 2.18. Let / : R x R"' x R^ R be a function of class such


that
/ ( t + T ,x,A ) = f{t,x ,X )
for every {t, x,X) G R x R x R^. Assuming that x(t) is a T-periodic solution
of the equation x' = f{t, x, 0) and that the only T-periodic solution of the
linear variational equation
2.6. Exercises 101

is the zero function, show that given e > 0, there exists 5 > 0 such that if"
llAll < S, then there exists a unique T-periodic solution x\{t) of the equation
x' = f{t, X, A) satisfying
-a;(t)|| < for t e

Solutions.
cos(\/2t) \/2 sin(-\/2<)A
2 .1 ---
sin(V^f) / \ / 2 cos(V^t) J '
2.2 a) A has only eigenvalues with zero real part and diagonal Jordan
block.
b) A has no eigenvalues with positive real part and each eigenvalue
with zero real part has diagonal Jordan block.
c) A has only eigenvalues with negative real part.

2.5 cosB = sinB =

sin (7

2.8 a:(t) -aexp/Q|(/(s)+5(s))ds + 6exp/jJ(/(s)


y{t) aexpjQ(f(^s)+g(s))dsbexpjQ[f(^s'^_g^g'j'jds, with a, 6
2 .1 2 a) h{x,y) = {sgnx- |xp/^,sgri2/- \y\^^^).
b) h{x,y,z) = if{x ,y ),g {z )), where g{z) = ggn^ \z\^/^ and
0
f{x ,y ) = if (x,y) = (0, 0),
I log 2- ^ if (s, y) 7^ (0, 0)
Part 2

Stability and
Hyperbolicity
Chapter 3

Stability and Lyapunov


Functions

In this chapter we introduce the notions of stability and asymptotic sta


bility for a solution of an ordinary differential equation. In paiticular, for
nonautonomous linear equations we characterize the notions of stability and
asymptotic stability in terms of the fundamental solutions. We also show
that the solutions of a sufficiently small perturbation of an asymptotically
stable linear equation remain asymptotically stable. Finally, we give an in
troduction to the theory of Lyapunov functions, which sometimes allows one
to establish in a more or less automatic manner the stability or instability
of a given solution. For additional topics we refer the reader to [20, 24, 25].

3.1. Notions of stability

Given a continuous function f : D ^ in an open set C M x E , consider


the equation
x' = f{t,x ). (3.1)
We assume that for each (to, xq) D there exists a unique solution x{t, to, xq)
of the initial value problem

ix' = f{t,x ),
|a;(to) = XQ.

3.1.1. Stability. In this section we introduce the notion of stability for a


solution of equation (3.1). Essentially, a solution x{t) is stable if all solutions
with sufficiently close initial condition remain close to x{t) for all time.

7^
106 3. Stability and Lyapunov Functions

D efinition 3.1. A solution x{t, to, xq) of equation (3.1) defined for all t > to
is said to be stable if given e > 0, there exists 5 > 0 such that if l|xo^oll <
then:

a) the solution x{t,to,xo) is defined for all t > to',


b) ||a:(t,to,a:o) - a;(t,to,So)H < e for t > to-

Otherwise, the solution x{t, to, xo) is said to be unstable.

The following are examples of stability and instability of solutions.

E xam ple 3.2. Consider the equation

X' = y ,
(3.2)
y' = - x - y.

If (x, y) is a solution, then

{x^ + y y = 2xy + 2 y {-x - y ) = -2y^ < 0,

and thus, the conditions in Definition 3.1 are satisfied for the zero solution
(with to arbitrary and xo = 0). This shows that the critical point (0, 0) is
a stable solution. Alternatively, note that the matrix of coefficients of the
linear equation (3.2) has eigenvalues (1 it iy/3)/2, both with negative real
part.

E xam ple 3.3. Consider the equation in polar coordinates

r' = 0,
(3.3)
{ e' = /( r ) ,

where / is a positive function of class with f'{ro) 7^ 0 for some ro > 0.


The phase portrait is the one shown in Figure 3.1: the origin is a critical
point and the remaining orbits are circular periodic orbits centered at the
origin.
It follows from (3.3) that each periodic orbit has period 2n/f{r). Since
/'(r o ) 7^ 0, for r 7^ To sufficiently close to ro the corresponding periodic orbit
is traversed with angular velocity /( r ) 7^ f{ro)- This yields the following
property. Let x{t) and xo(t) be solutions of equation (3.3) such that x{to)
and xo(to) are, respectively, on the circles of radius r and tq. Given x{to)
arbitrarily close to xo(to)) there exists t > to such that x{t) and xo{t) are on
the same diameter of the periodic orbits, but on opposite sides of the origin.
This shows that the second condition in Definition 3.1 is not satisfied, and
thus the solution xo(t) is unstable.
3.1. Notions o f stability 107

F ig u re 3.1. Phase portrait o f equation (3.3).

3.1.2. A sy m p totic stability. Now we introduce the notion of asymptotic


stability for a solution.

D efinition 3.4. A solution x(t, to, xq) of equation (3.1) defined for all t > to
is said to be asymptotically stable if:
a) x(t,to,xo) is stable;
b) there exists o: > 0 such that if ||o;o ~ ^o|| < <^, then
||a;(t, to, xo) x(t, to, xo) ||>0 when t -> +oo.

The following example shows that for a solution to be asymptotically


stable it is not sufficient that the second condition in Definition 3.4 is satis
fied.

Exam ple 3.5. Consider the equation in polar coordinates

r' = r ( l r),
{ (3.4)
e' = sin2(0/ 2).

Its phase portrait is shown in Figure 3.2. We note that the critical point
(1,0) is a solution satisfying the second condition in Definition 3.4 but not
the first one (it is sufficient to consider, for example, the solution on the
circle of radius 1 centered at the origin).

Exam ple 3.6. Consider the equation

\r' = r ( l r),
(3.5)
I e' = s in 2 e.
108 3. Stability and Lyapunov Functions

F ig u re 3 .2 . Phase portrait o f equation (3.4).

Its phase portrait is shown in Figure 3.3. We note that since the angular
velocity O' does not depend on r, for any ray L starting at the origin the set

Lt = {x{t,tQ,xo) : xo L }
is still a ray (starting at the origin) for each t > to. This implies that each
solution outside the straight line y = 0 is asymptotically stable. On the
other hand, all solutions on the straight line y = 0 are unstable.

F ig u re 3.3. Phase portrait o f equation (3.5).

3.2. Stability o f linear equations

In this section we consider the particular case of the linear equations

x' = A{t)x, (3.6)


3.2. Stability o f linear equations 109

where is an n x n matrix varying continuously with t K. After


studying the general case, we consider the particular cases of the equations
with constant coefficients and periodic coefficients.

3 .2 .1 . N onau ton om ou s linear equations: general case. We first show


that in what concerns the study of the stability of linear equations it is
sufficient to consider the zero solution.

P rop osition 3.7. Let A: R -> M be a continuous function. For equa


tion (3.6), the zero solution (with arbitrary initial time to) is stable (respec
tively asymptotically stable) if and only if all solutions are stable (respectively
asymptotically stable).

P ro o f. We divide the proof into steps.

Step 1. Reduction to the zero solution. Let X {t) be a fundamental solu


tion of equation (3.6). By Theorem 2.8, the solution of the initial value
problem (2.2) has maximal interval R and is given by

x{t, to, xo) - X {t)X (to) ^xo (3.7)

It follows from (3.7) that the zero solution (with initial time to) is stable if
and only if given e > 0, there exists <5 > 0 such that

||A(t)A(to) ^xol| < e when ||xo|| < ^ (3.8)

for t > to. Since X(t)X(to)~^ is a linear transformation, this is the same as

||X(t)X(to)"Ho - o)|| < e when ||o - o|| < S,


or equivalently,

||a:(t,to,a:o)-a;(t,to,o)|| < e when ||o; - xqII <

for t > to and xo R . Therefore, the zero solution (with initial time <o) is
stable if and only if all solutions (with initial time to) are stable.
For the asymptotic stability, we note that since X (t)X (to ) ^ is linear,
we have
lim ||X(t)X(to) ^a;o| | = 0 when ||a:o|| < a (3.9)

if and only if

lim ||A(t)X(to) ^(xo - a;o)|| = 0 when ||xo-xo||<Q:


t>H-oo

for every xo R". This shows that the zero solution (with initial time to)
is asymptotically stable if and only if all solutions (with initial time to) are
asymptotically stable.
no 3. Stability and Lyapunov Functions

Step 2. Independence from the initial time. It remains to verify that the
stability and asymptotic stability of the zero solution are independent from
the initial time to- To that effect, take ti ^ to and note that
X{t)X{ti)-^ = X {t )Xit o)-^ X{t o)Xiti)-\
It follows from (3.8) that given e > 0, there exists <5 > 0 such that
||X(t)X(ti) ^a;o|| < e when ||X(to)^(ti) ^a;o|| < <5,
and thus,

||X(t)X(ti)"^a;o|( < e when ll^oll < ||x(to)X(ti)-i||

for t > to- If ti > to, then property (3.10) holds for t > ti. On the other
hand, for ti < to the function t i-)- X {t)X{ti)~^ is continuous in [ti,fo]-
Taking
\\Xito)X{h)-
6<
inaxt6[ti,eo) ll^ (0 ^ (^ i) Mr
we obtain
||X(to)X(ti) ^xo|| < max ||X(t)X(ti) Ml llo|| < S < e
te[ti,to]
for t 6 [ti, to] and xo as in (3.10). Therefore, property (3.10) holds for t > t\.
This shows that if the zero solution is stable with initial time to, then it is
stable with any initial time.
Now we assume that the zero solution is asymptotically stable with initial
time to. It follows easily from (3.9) that
lim X (t) = 0,
t >^+oo
and thus.
lim ||Z(t)X(ti) - ^ | | = 0
t>+00
for every ti M and xo M . This shows that the zero solution is asymp
totically stable with any initial time.

It follows from Proposition 3.7 that for equation (3.6) the zero solution
(with arbitrary initial time to) is unstable if and only if all solutions are
unstable.
In view of Proposition 3.7 it is natural to introduce the following notion.

D efinition 3.8. Equation (3.6) is said to be stable (respectively, asymptot


ically stable or unstable) if all its solutions are stable (respectively, asymp
totically stable or unstable).

Now we give a characterization of the notions of stability and asymptotic


stability for the nonautonomous linear equation x' = A{t)x.
3.2. Stability o f linear equations 111

T h eorem 3.9. Let ^4: R ^ be a continuous function and let X {t) be a


fundamental solution of equation (3.6). Then the equation is:
a) stable if and only if sup {|lX(t)|| : t > 0 ] < + oo;
b) asymptotically stable if and only if
||X(t) | | >-0 when t + o o .

P ro o f. It follows from Proposition 3.7 and (3.8) that the zero solution is
stable (with arbitrary initial time to) if only if given > 0, there exists
(5 > 0 such that
||X(t)X(0) ^a;o|| < when ||a;o|| < (5,
for t > 0. Thus, if the zero solution is stable, then
||X(t)X(0)-ia:o|
||X(t)X(0)"^|| = sup
xo^O l o||
||X(t)X(0)-H.5xo/||xo||)||
S ||<5o/lko||||
_ ||X(t)X(0)-So||
7=s l yoll 4
for t > 0, and hence,
sup {||X(t)|| : t > 0 } < sup {||X(t)X(0) ^|| : t > 0 }||X(0)||
(3.11)
< ffl< + o o .

On the other hand, if the supremum in (3.11) is finite, then there exists
C > 0 such that
||X(t)||<C for t > 0,
and thus.

||X(t)X(0) ^xoll < when ||xo|| <


C||X(0)-i| r
for t > 0. This establishes the first property.
For the second property, we first note that if ||.X(t)|| > 0 when t +oo,
then
sup {||X(t)|| : t > O} < + 00 ,
and hence, by the first property, the solution is stable. Moreover,
||X(t)X(0) ^ x o | | 0 when t ->+oo, (3.12)
for every xq R , and thus, equation (3.6) is asymptotically stable. On the
other hand, if the equation is asymptotically stable, then property (3 .12)
holds for ||xo|| sufficiently small, and thus ||^(t)|| 0 when t -> +oo.
112 3. Stability and Lyapunov Functions

3.2.2. C onstant coefficients and p eriod ic coefficients. In this sec


tion we consider the particular cases of the linear equations with constant
coefficients and periodic coefficients. We start with the case of constant
coefficients.

T h eorem 3.10. For a square matrix Aj the equation x' Ax is:


a) stable if and only if A has no eigenvalues with positive real part and
each eigenvalue with zero real part has a diagonal Jordan block;
b) asymptotically stable if and only if A has only eigenvalues with neg
ative real part;
c) unstable if and only if A has at least one eigenvalue with positive real
part or at least one eigenvalue with zero real part and a nondiagonal
Jordan block.

P roof. The claims follow easily from the Jordan canonical form in The
orem 2.16 combined with the description of the exponential of a Jordan
block in Proposition 2.18. Indeed, if S is an invertible square matrix satis
fying (2.12), then

oRit n
= S
oRk^

where Rj is the nj x nj matrix in (2.13). On the other hand, it follows from


Proposition 2.18 that

gRjt _ gA,t ^ld + tNj + ^ t ^Nf + 1


{rij - 1)!
where Nj is the rij x nj matrix in (2.15). This shows that each entry of the
matrix is of the form where p is a polynomial of degree at most
rij 1. The desired result now follows immediately from Theorem 3.9.

Now we consider linear equations with periodic coefficients.

T h eorem 3.11. Let >1: R ^ M be a T-periodic continuous function and


let B he the matrix in (2.32). Then the equation x' A(t)x is:
a) stable if and only if there are no characteristic exponents with positive
real part and for each characteristic exponent with zero real part the
corresponding Jordan block of the matrix B is diagonal;
b) asymptotically stable if and only if there are only characteristic ex
ponents with negative real part;
3.3. Stability under nonlinear perturbations 113

c) unstable if and only if there is at least one characteristic exponent


with positive real part or at least one characteristic exponent with
zero real part such that the corresponding Jordan block of the ma
trix B is not diagonal.

P ro o f. By Floquets theorem (Theorem 2.31), any fundamental solution of


the equation x' = A{t)x is of the form (2.32). By (2.33), the matrix function
P{t) is T-periodic, and thus, each property in Theorem 3.9 depends only on
the term e-*. That is.
sup{||X(t)|| : t > 0} < + 0 0 o sup {||e^*|| : t > O} < +oo.
and
lim ||e ||= 0.
t^-+oo" "
This shows that the stability of the equation x' = A{t)x coincides with the
stability of the equation x' = Bx. Since the eigenvalues of the matrix B
are the characteristic exponents (up to an integer multiple of 2m/T)^ the
desired result follows now immediately from Theorem 3.10.

3.3. Stability under nonlinear perturbations

In this section we consider a class of nonlinear perturbations of an asymp


totically stable linear equation. We start with the case of perturbations of
linear equations with constant coefficients.
T h eorem 3.12. Let A be an n x n matrix having only eigenvalues with
negative real part and let g : MxW^ be continuous and locally Lipschitz
in X. If g{t, 0) = 0 for every t G R, and

(3.13)
teK Ikll
then the zero solution of the equation
x' = Ax 4- g{t, x) (3-14)
is asymptotically stable. Moreover, there exist constants C, A, 5 > 0 such that
for each to M and each solution x{t) of equation (3.14) with ||a;(to)|| < S,
wc have
lk(t)||<C'e-^(*->)||a;(to)|| for t > t o . (3.15)

P roof. Since A has only eigenvalues with negative real part, by Proposi
tion 2.27 there exist constants c,ij.> 0 such that
||e^*|| < ce-^ (3.16)
for t > 0. On the other hand, by (3.13), given e > 0, there exists <5 > 0 such
that
||9 ((,*)||<e||x|| (3.17)
114 3. Stability and Lyapunov Functions

for every f 6 M and a: G M " with |la ;|l < 6. Now let x{t) be the solution of
equation (3.14) with x{to) = xq. By the Variation of parameters formula in
Theorem 2.26, we have

x{t) = xq +
f e'^^*~^^g{s,x{s)) ds (3.18)
Jto
for t in the maximal interval of the solution. Moreover, given xq G M with
Ikoll < 5, the solution x{t) satisfies ||x(t)|| < S for any t > to sufficiently
close to to> say for t G [to>^i]- Hence, it follows from (3.16) and (3.17) that

||a;(t)|| < ce-^(*-**)||a:o||+ f ce->^^^-^h\\x{s)\\ds, (3.19)


Jtn
or equivalently,

e^ ||a;(t)||<ce'^ '>||xo||+ T cee^||a;(s)|| ds,


Jto
for t G [tO )ii]- By Gronwalls lemma (Proposition 1.39), we obtain
e^ ||x(t)|| <
that is,
||a:(t)|| < (3.20)
for t G [t0)ii]- Assuming, without loss of generality, that c > 1 and taking
e > 0 so small that p + ce < 0, it follows from (3.20) that if ||o|| < ^/c,
then ||x(t)|| < 3 for t G [toi^i]- Thus, there exists <2 > i\ such that the
solution x{t) is defined in the interval [to,t2] and satisfies ||a;(t)|| < 5 for
t G [t0)^2]- One can repeat this procedure indefinitely (without changing e
and 5) to conclude that there exists an increasing sequence tn such that x{t)
is defined in the interval [t0)^n] and satisfies ||a:(t)|| < <5 for t G [to>^n]> for
each n G N.
Now let b be the supremum of all sequences tn- If 6 < +oo, then we
would have ||(6 )|| < <5, which contradicts to Theorem 1.46. Thus, b = +oo
and the solution x{t) is defined in [to,+oo). Moreover, it satisfies (3.20) for
allt > to, which yields inequality (3.15). This shows that the zero solution
is asymptotically stable.

The following is an application of Theorem 3.12.

T h eorem 3.13. Let /: R be a function of class and let xq G R


be a point with f(xo) = 0 such that dxof has only eigenvalues with negative
real part. Then there exist constants C,X,d > 0 such that for each to ^ R
the solution x{t) of the initial value problem
(x' = fix),
[ x(to) = xo
3.3. Stability under nonlinear perturbations 115

satisfies
lk(i) - xoll < - a;o|| (3.21)
for every t > to and xq M with ||xo a:o|| < d.

Proof. We have

= f(x) = dxofix - xo) + f{ x) - dxof{x - xo). (3.22)


Letting y = x x q , equation (3.22) can be written in the form

y' = Ay + g{t,y), (3.23)

where A = d^of and

9{i, y) = / ( o + y ) ~ dxofy-
By hypothesis, the matrix A has only eigenvalues with negative real part.
Moreover, g{t, 0) = / ( xq) = 0 for every t M, and since / is of class C^, we
have
g{t> y) _ f{xo + y ) ~ /( o ) - d^ofy
-^ 0
?R llvll toll
when y 0. In other words, the hypotheses of Theorem 3.12 are satisfied.
Hence, it follows from (3.15) that there exist constants C, A,<J > 0 such that
for each to e R and each solution y{t) of equation (3.23) with ||y(to)|| < d,
we have
||y(i)||<C'e-"( - ')||2/(to)|| for t>to.
This establishes inequality (3.21).

One can obtain corresponding results for nonlinear perturbations of


nonautonomous linear equations. The following is a version of Theorem 3.12
in this general context.

T h eorem 3.14. Let A: E > be a continuous function and let X(t) be


a fundamental solution of the equation x' = A{t)x such that

||Z(t)X(s)-^||

for some constants c, ^ > 0 and every t > s. If the function 5 : E x E >E
is continuous and locally Lipschitz in x, satisfies g{t,0) = 0 for every t E,
and property (3.13) holds, then the zero solution of the equation
x' = A{t)x + g{t,x) (3.24)

is asymptotically stable.

P roof. We follow closely the proof of Theorem 3.12, replacing identity (3.18)
by an appropriate identity. Namely, if x{t) is the solution of equation (3.24)
116 3. Stability and Lyapunov Functions

with a;(to) = xq, then it follows from the Variation of parameters formula in
Theorem 2.25 that

x{t) = X{t)X{to)-'^xo+ f X{t)X{s)-'^g{s,x{s))ds (3.25)


Jto
for t in the maximal interval of the solution. On the other hand, by (3.13),
given e > 0, there exists 5 > 0 such that inequality (3.17) holds for every
t G K and a; G M with ||a;|( < 6. Now take ti > to and xq G M with ||a;o|| < 5
such that x{t) is defined in the interval [to,ti] and satisfies ||a;(t)|| < S for
tG It follows from (3.25) and (3.17) that inequality (3.19) holds for
t [tO) ti]- Now one can repeat the arguments in the proof of Theorem 3.12
to conclude that if ||xo|| < S/c, then the solution x{t) is defined in [to, +oo)
and satisfies
b(t)l| <
for t > t o (assuming that c > 1 and that e > 0 is so small that p + ce < 0).
In particular, the zero solution is asymptotically stable.

3.4. Lyapunov functions

This section is an introduction to the theory of Lyapunov functions, which


sometimes allows one to establish the stability or instability of a given so
lution in a more or less automatic manner.

3.4.1. Basic notions. We first recall the notion of locally Lipschitz func
tion.

D efinition 3.15. A function / : ) >^E in an open set D C R is said to


be locally Lipschitz if for each compact set K c D there exists L > 0 such
that
\\f{x)-f{y)\\<L\\x-y\\ (3.26)
for every x ,y E K.

Let / : D > E be a locally Lipschitz function. One can easily verify


that / is locally Lipschitz if and only if the function g : M x D >MT defined
by g{t,x) = f i x ) is locally Lipschitz in x. Moreover, by (3.26), any locally
Lipschitz function is continuous. Now let ^t{xo) = x{t,xo) be the solution
of the initial value problem

x' = fix ),
(3.27)
{:a;(0) = XQ,
which in view of the Picard-Lindelof theorem (Theorem 1.18) is well defined.
Given a differentiable function V: D define a new function V: D
by
Vix) = W i x ) fix).
3.4. Lyapunov functions 117

We note that

''( I ) = (3.28)

Now we introduce the notion of a Lyapunov function for a critical point


of the equation x' = f{x).

Definition 3.16. Given xq E D with f{xo) = 0, a differentiable function


V: D - M. is called a Lyapunov function for xq if there exists an open set
U C D containing xq such that:

a) V{ xq) = 0 and F ( ) > 0 for a; 17 \ {a;o};


b) F(a;) < 0 for x e U.

A Lyapunov function is called a strict Lyapunov function if the second con


dition can be replaced by F(a;) < 0 for x 17 \ { xq}.

Example 3.17. Consider the equation

f x' = -X -I- y,
\y' = -X -

The origin is the only critical point. We show that the function F : M
given by
V{x,y) = x^ -I-
is a strict Lyapunov function for (0,0). We have F (0,0) = 0 and V { x , y ) > 0
for (x,y) 7^ (0,0). Moreover,

V(x, y) = (2x, 2y) ( - X + y , - x - y^)


= - 2(x2 -Fy^) < 0

for (x,y) ^ (0, 0).

3.4.2. Stability criterion. The existence of a Lyapunov function (respec


tively, a strict Lyapunov function) for a critical point of a differential equa
tion x' = / ( x ) allows one to establish the stability (respectively, the asymp
totic stability) of that point.

Theorem 3.18. Let / : D R be a locally Lipschitz function in an open


set jD C R and let xq E D be a critical point of the equation x' = f{x).
a) If there exists a Lyapunov function for xq, then xq is stable.
b) If there exists a strict Lyapunov function for xq, then xq is asymp
totically stable.
118 3. Stability and Lyapunov Functions

P roof. We first assume that there exists a Lyapunov function for xq in some
open s e t U c D containing xq. Take e > 0 such that B{xo,s) C U, and

m = m in {y (x ) : x dB{xo,e)}.

Since V is continuous (because it is locally Lipschitz) and V > 0 in the set


B{xo,e) \ {a;o}, there exists 5 e (0,e) such that

0 < m a x {y (x ) : x G S (x o ,5 )} < m. (3.29)

On the other hand, it follows from (3.28) that the function 1 V { ( p t { x ) ) is


nonincreasing (in the maximal interval I of the solution). Indeed, proceeding
as in the proof of Proposition 1.13, we obtain tpt = (pt-s^s for t sufficiently
close to s, and thus.

at (3.30)
= viM ^)) < 0
for s E I. Hence, it follows from (3.29) that any solution (pt{x) of the initial
value problem (3.27) with x B{ xq,S) is contained in B{xo,e) for every
t > 0 in its maximal interval. This implies that each solution ipt{x) with
X E B{xo,S) is defined for all t > 0, and thus the critical point xq is stable.
Now we assume that there exists a strict Lyapunov function for xq. It
remains to show that (pt{x) xq when t +oo, for any point x E B{ xq, a)
with a sufficiently small. Proceeding as in (3.30), we conclude that for each
X E U\{ xq} the function 1V{ ( pt { x) ) is decreasing (in the maximal interval
of the solution). Now let (tn)n be a sequence of real numbers with tn +oo
such that {<Ptni^))n converges, and let y be the limit of this sequence. Then

V(.>tix)) \ V{y) when n -> oo,

because t !-> V{(pt{x)) is decreasing. Moreover,

V{cpt{x))>V{y) for t > 0. (3.31)

Now we assume that y ^ xq. Then V{(ps{y)) < V{y) for every s > 0. Taking
n sufficiently large, one can ensure that </?t+s(a;) = ((a;)) is as close
as desired to (ps{y), and thus also that V{ipt+s(,^)) is as close as desired
to V{ips{y)) < V { y ) . Hence,

V{<Ptn+s{x)) < V { y ) ,
but this contradicts to (3.31). Therefore, y = xq, and we conclude that
(pt{x) Xq when t +oo. Cl

The following are applications of Theorem 3.18.


3.4. Lyapunov functions 119

Exam ple 3.19. Consider the equation

x' = y -
(3.32)
{ y' = -x^.
This can be seen as a perturbation of the linear equation {x,yY = (y, 0),
which has the phase portrait in Figure 2.1 1 . We consider the critical point
(0, 0) of equation (3.32), and the function
V{x,y) = x^ + 2y^.
We have V (0 ,0) = 0 and V{x,y) > 0 for (x,y) 7^ (0,0). Moreover,
V{x, y) = (4x^, 4y) (y - xy^, -x^)
= 4x^y 4x^y^ 4x^y = 4x^y^ < 0.
Hence, F is a Lyapunov function for (0,0), and it follows from Theorem 3.18
that the origin is stable.

E xam ple 3.20. Consider the equation

x' = x^ X y,
(3.33)
[y = X.
We discuss the stability of the critical point at the origin. Equation (3.33)
can be written in the form

+ I{ ] (3.34)

Since the 2 x 2 matrix in (3.34) has eigenvalues (li-\/3)/2, both with neg
ative real part, it follows from Theorem 3.13 that the origin is asymptotically
stable.
One can also use Lyapunov functions to study the stability of the origin.
However, it is not always easy to find a strict Lyapunov function (when it
exists). For example, consider the function V(x,y) = x"^ + y^. We have
V (x, y) = (2x, 2y) (x^ - x - y, x) = 2x^(x - 1),
and thus, V{x,y) < 0 in a sufficiently small neighborhood of (0,0). Hence,
V is a Lyapunov function, and it follows from Theorem 3.18 that the origin
is stable. However, this does not show that the origin is asymptotically
stable. To that effect, consider the function
W{x,y) = x^ -t-x y + y^.
We have W (0,0) = 0. Moreover,
1 1/2^
W{x,y) =
1/2 (3.35)
120 3. Stability and Lyapunov Functions

Since the 2 x 2 matrix in (3.35) has eigenvalues 1/2 and 3/2, the matrix is
positive definite. In particular, W{x,y) > 0 for (x,y) ^ 0. We also have
W{x,y) = (2x + y,x + 2y) {x^ - x - y , x )
= 2x^ - x ^ + x^y - x y - y " ^

=- (l/2 f )
Since the 2 x 2 matrix in (3.36) is positive definite, there exist constants
a, 6 > 0 such that

-a | | (x ,y )f < -
1 )2 'f) 0 ^
for every (a;,y) M^. Moreover,
2x^ + x^y
0 when (x, y) 0.
II ( ^ . 9) IP ^

Therefore, given e > 0, we have


W{x,y)
a e < < -b + e
(.y)IP
for any sufficiently small (x, y) ^ 0. Taking e so small that 6 + e < 0, we
obtain W(x,y) < 0 for any sufficiently small (x,y) ^ 0. Hence, it follows
from Theorem 3.18 that the origin is asymptotically stable.

Exam ple 3 .2 1 . Consider the equation


x" + f ix ) = 0, (3.37)
where / : K -> R is a function of class with /(O) = 0 such that
xf{x) > 0 for x^O (3.38)
(that is, f { x ) and x always have the same sign). This corresponds to ap
ply a force f {x ) to a particle of mass 1 that is moving without friction.
Condition (3.38) corresponds to assume that the force always points to the
origin.
Equation (3.37) can be written in the form

x' = y,
(3.39)
y' = - f { x ) ,
and (0,0) is a critical point. We use a Lyapunov function to show that the
origin is stable. Namely, consider the function

V(^,y) = ^y'^ + f{s)ds, (3.40)


3.4. Lyapunov functions 121

which corresponds to the sum of the kinetic energy y^/2 (recall that the
particle has mass 1) with the potential energy Jq f{s) ds. We have V (0 ,0) =
0 and V{x,y) > 0 for (x,y) 7^ (0,0), due to condition (3.38). Moreover,

V{x, y) = (/(a;), y) (y, - f i x ) ) = 0,


and thus, V is a Lyapunov function for (0,0). Hence, it follows from Theo
rem 3.18 that the origin is stable. Incidentally, along the solutions we have

^ V { x , y ) = yy' + f(x)x'

= - y f { x ) + f{ x) y = 0,
and thus, equation (3.39) is conservative. This corresponds to the conserva
tion of energy.

E xam ple 3.22. Given e > 0, consider the equation


x" -f ex' + fix) = 0, (3.41)
with / : as in Example 3.21. Equation (3.41) can be written in the
form
\x' = y,
= -fix ) - ey.
We consider again the function V in (3.40), which satisfies F (0, 0) = 0 and
Vix,y) > 0 for ix,y) ^ (0,0). Moreover,
Vix, y) = ifix), y) (y, - f i x ) - ey) = -ey^ < 0,
and F is a Lyapunov function for (0,0). Hence, it follows from Theorem 3.18
that the origin is stable.

3.4.3. Instability criterion. We conclude this chapter with the descrip


tion of an instability criterion for the critical points of an equation x' = fix).
The criterion is analogous to the stability criterion in Theorem 3.18.

T h eorem 3.23. Let f : D - ^ W ^ be a function of class in an open set


D C M" and let xq E D be a critical point of the equation x' = fix). Also,
let V: U ^ M. be a function of class in a neighborhood U C D of xq
such that:
a) V(xo) = 0 and F(a;) > 0 for x U \ {a:o};
b) V takes p o s itiv e v a lu e s in a n y n e ig h b o rh o o d o f xq.

Then the critical point xq is unstable.

P ro o f. Let A C U he a neighborhood of xq and let <ptix) be the solution


of the initial value problem (3.27). If in each neighborhood A there exists a
solution iptix) that is not defined for all t > 0, then there is nothing to show.
122 3. Stability and Lyapunov Functions

Hence, one can assume that all solutions ipt{x) with x e A ave defined for
all t > 0.
Now take y E A with V(y) > 0 (which by hypothesis always exists).
Since V{x) > 0 for x U \ {a;o}, proceeding as in (3.30) we conclude that
the function t V{(ft{y)) is increasing whenever (pt{y) U. Thus, since
y(a;o) = 0, the solution (pt{y) does not come near xq, that is, there exists a
neighborhood B of xq such that (pt{y) ^ B for t > 0. Now we assume that
the solution does not leave A and we define
m = m f{V(ipt{y))-t>0}.
Since V = V V f is continuous and A \ B is compact, we have
m > inf (l^() : a; \H} > 0
(because continuous functions with values in R have a minimum in each
compact set). We also have
> V { y ) + mt for t>0.
Thus, there exists T > 0 such that
V{(pr{y)) > m ax{F(a;) : x G A},
and hence ipriy) ^ A. This contradiction shows that there exist points x
arbitrarily close to xq (because by hypothesis V takes positive values in any
neighborhood of xq) such that the solution <pt(x) leaves the neighborhood A.
Therefore, the critical point xq is unstable.

The following is an application of Theorem 3.23.

E xam ple 3.24. Consider the equation


ix' = Zx + y^,
\y' = - 2 y + x^,
for which (0, 0) is a critical point, and the function
V{x,y) - x^ - y ^ .
Clearly, V (0,0) = 0 and V takes positive values in any neighborhood of (0,0).
Moreover,
V{x, y) = (2a;, -2y ) (3a; + y^, - 2 y + x^)
= 6a;^ + + 2xy^ 2x^y.
Since
V{x,y)
1 when ( x , y ) (0, 0),
6x^ +
we have V{x,y) > 0 for any sufficiently small (x,y) ^ (0,0). Hence, it
follows from Theorem 3.23 that the origin is unstable.
3.5. Exercises 123

3.5. Exercises

E xercise 3.1. Find all stable, asymptotically stable and unstable solutions
of the equation:
a) x' = x{x 2);
b) x" + 4x = 0.

E xercise 3.2. For a function R " ^ E of class (7^, consider the equation
x' = V 5 (x).
a) Show that if is a nonconstant solution, then p o tt is strictly in
creasing.
b) Show that there are no periodic orbits.
c) Determine the stability of the origin when g{x,y) = x^ + y^.
d) Determine the stability of the origin when g{x, y) x^y^.

E xercise 3.3. Consider the equation in polar coordinates

fr ' = /( r ) ,

M ' = l,
where
= /^ s in (l/r2 ), r -^ 0 ,
^ \0, r = 0.
Show that the origin is stable but is not asymptotically stable.

E xercise 3.4. Determine the stability of the zero solution of the equation:

x' = X + xy^
a)
[y' = X - 2/ - x^ - y^]

x' = X + x'^ + y^,


b)
y = 2x - 3y-|-y^;
x' = X -H2x(x + y)^,
c)
y' = -y ^ -l-2 y 3 (x -| -y )^ ;

x' = x^ 3xy^,
d)
y = 3x^y - y^.

Exercise 3.5. Let o: R ^ R be a T-periodic continuous function.


a) Find the characteristic exponent of the equation x' = a{t)x.
b) Find a necessary and sufficient condition in terms of the function a
so that the zero solution is asymptotically stable.
124 3. Stability and Lyapunov Functiom

Exercise 3.6. Compute

A(x) = limsup Y log||x(i)||


t-^+oo t
(with the convention that log 0 = - o o ) for each solution x{t) of the equation:
a) x" + x = 0;
b) x' = [a + 6(sin log t + cos log t)\x with o, 6 G M.
Exercise 3.7. Given an n x n matrix A, define x- [~o, +oo) by

x{v) = limsup - logll^'^ull.


n>-+oo ^
Show that:
a) x{oiv) = x{v) for a ^ 0;
b) x(v + w) < m ax{x(w ),x(t)};
c) if x{v) + x{w), then x(v + w) = m a x{x (v ),x(i)};
d) X takes only finitely many values.
Exercise 3.8. Show that if {xi{t),X2{t)) is a nonzero solution of the equa
tion
{ x'l = [1.01 - (sinlogt-I-coslogt)]a;i,
x'2 = [1.01 4- (sinlogt -I- coslogt)]aj2,
then
limsup-log||(a;i(<),X2(i))II < 0.

Exercise 3.9. Consider the equation


y{ = [ - 1.01 - (sinlogt + coslogt)]yi,
{
j /2 = [-1-01 -I- (sin log t -I- coslogt)]2/2 + J/i^-
li
a) Verify that each solution iyi{t),y2{t)), with t > 0, can be written in
the form
iyi{t) =
\y2{t) = c2e - i i*+W + ci2e - i oi*+ (0
for some constants ci, C2 and s, where a{t) = t sin log t.
b) Taking e G (0, 7t/ 4) and defining tk = for fc G N, verify that
3a(r) > 3rcose for r G [tke~^,tk],
and conclude that
A ' ' g-3o(r)-1.0lT g -3 a (r)-1 .0 lT ^ ^ > gg(3cose-1.01)tfc
Js Jtke~^
for any sufficiently large fc G N, where c > 0 is a constant.
3.5. Exercises 125

c) Show that there exists a solution {yi{t),y2{t)) such that

limsupylog||(yi(t),y2(i))|| > 0.

Hint: Consider the times t = tkc".

Exercise 3.10. Let / : M -> M be a function of class C such that the


equation x' = f {x ) defines a fiow (fit in M .
a) Show that

(pt{x) = x + f{x) t + ^{dxf)f{x)t^ + o{t^).

b) Verify that
det dxift = 1 + div f {x )t + o{t).
c) Given an open set ^ C R and t G R, show that

^y{ (p t{ A) )= f d iv /,
di Jm a )
where y, denotes the volume in R .
d) Show that if d iv / = 0, then the equation x' = f {x ) has neither
asymptotically stable critical points nor asymptotically stable peri
odic solutions.

Solutions.
3.1 a) The solutions in (o o ,2) are asymptotically stable and the solu
tions in [2, -l-oo) are unstable.
b) All solutions are stable but none are asymptotically stable.
3.2 c) Unstable,
d) Unstable.
3.4 a) Asymptotically stable.
b) Asymptotically stable.
c) Asymptotically stable.
d) Unstable.
3.5 a) (1/r) Jq a{s) ds.
b) Jq a{s) ds < 0.
3.6 a) \{x) = 0.
b) A(x) = o -H |6|.
Chapter J,.

Hyperbolicity and
Topological
Conjugacies

This chapter is dedicated to the study of hyperbolicity and its consequences,


particularly at the level of stability. After a brief introduction to the notion
of hyperbolicity, we establish a fundamental result on the behavior of the
solutions in a neighborhood of a hyperbolic critical point the Grobman-
Hartman theorem. It shows that the solutions of a sufficiently small per
turbation of an equation with a hyperbolic critical point are topologically
conjugate to the solutions of its linear variational equation. We also show
that the topological conjugacy is Holder continuous. For additional topics
we refer the reader to [7, 15, 19, 23].

4.1 . Hyperbolic critical points

In this section we introduce the notion of hyperbolic critical point.


We recall that a square matrix is said to be hyperbolic if all its eigen
values have nonzero real part (see Definition 2.48). Now let / : R > R be
a function of class C^.

D efinition 4.1. A point xq R with / ( xq) = 0 such that the matrix dxof
is hyperbolic is called a hyperbolic critical point of the equation x' = f{x).

Given a hyperbolic critical point xq R of x' = /( x ) , consider the


linear equation
x' = Ax, where A = dx^f-
128 4. Hyperbolicity and Topological Conjugacies

We recall that its solutions are given by


x(t) = t R.
D efinition 4.2. Given a hyperbolic critical point xq E M of the equation
x' = f(x), we define the stable and unstable spaces of xq, respectively, by
= {a; R : -> 0 when t +oo}
and
= {a; R : e'^^x 0 when t oo}.

P rop osition 4.3. If xq R is a hyperbolic critical point of the equation


x' = f{x), then:
a) and are linear subspaces o /R with E^ E'^ = R ";
b) for every x E E^, y E E'^ and t EM., we have
e^^x E E^ and e^^y E E^.

P ro o f. Since the matrix A = dxof has no eigenvalues with zero real part,
its Jordan canonical form can be written in the form
0
0 Au^
with respect to some decomposition R = F 0 where ^4^ and Au cor
respond respectively to the Jordan blocks of eigenvalues with negative real
part and the Jordan blocks of eigenvalues with positive real part. It follows
from Proposition 2.27 that
0 when t y hex),
for X E F^, and that
-> 0 when t y (X),
for X E F'^. Hence,
ps _ ps and pu ^ pu
which establishes the first property.
For the second property, we first recall that

for every t, r R (see Exercise 2.3). In particular, if a; F* and t R, then


gAr _ gi4t (gAr^)

Since e^'^x )0 when r -> -l-oo, we obtain


gAt(gAr^) Q J.

and it follows from (4.1) that e^^x E F. One can show in an analogous
manner that if j/ F and t R, then e^^y E FF.
4.2. The Grobman-Hartman theorem 129

By Proposition 4.3, the spaces and associated to a hyperbolic


critical point form a direct sum. Hence, given a; M , there exist unique
points y E E^ and z E E'^ such that x = y + z. We define Pg,Py,: R > R
by
PgX = y and PuX = .2. (4.2)
One can easily verify that Pg and Pu are linear transformations with

Pg{R^) = E^ and Pu{M.^) = E^.

Moreover, Pg = Pg and P^ = Pu, that is, Pg and are projections, respec


tively, over the spaces E^ and One can also show that

P = ReG* + ImG* and = R eG -b Im G ,

where G and G are the subspaces of generated by the root spaces,


respectively, of eigenvalues with negative real part and eigenvalues with
positive real part. More precisely,

G = {a; C : (dxof AId)*x = 0 for some A: G N, A G C with Re A < O}

and

G = {a: G C : {d x o f AId)^a; = 0 for some A: G N, A G C with ReA > O}.

4.2. The G robm an -H artm an theorem

Let / : R R*^ be a function of class G^ such that the equation x' f{x )
has a hyperbolic critical point xq. In this section we show that the solutions
of the equations
x' = f {x ) and y' = dx^fy
are topologically conjugate, respectively, in neighborhoods of xq and 0. More
precisely, and in an analogous manner to that in Definition 2.44, this means
that if and (pt{z) are, respectively, the solutions of the initial value
problems
ix' = f{x), iy' = dxofy,
(4.3)
[x(0) = z \y{0) = z,
then there exists a homeomorphism h: U - V , where U and V are, respec
tively, neighborhoods of xq and 0, such that h{xo) = 0 and

HM^)) = M K ^ )) (4.4)

whenever z,>pt{z) E U. This guarantees that the phase portraits of the


equations in (4.3) are homeomorphic, respectively, in neighborhoods of xq
and 0 (see Figure 4.1).
130 4. Hyperbolicity and Topological Conjugaciet

F ig u re 4.1. Topological conjugacy between the solutions o f the equa


tions in (4.3).

4 ,2 .1 . P e r t u r b a t io n s o f h y p e r b o lic m a t r ic e s . W e f ir s t e s t a b lis h a re

s u lt o n th e e x is t e n c e o f t o p o lo g ic a l c o n ju g a c ie s fo r t h e p e r t u r b a t io n s of a

lin e a r e q u a t io n = Ax w it h a h y p e r b o lic m a t r ix A.

T h e o re m 4 .4 . Let A be an n x n hyperbolic matrix and let g:MP' be


a bounded function such that g{0) = 0 and

lli^ ( )- i? (y )ll < < ^ l | a : - y | | (4 .5 )

fo r every x , y M ". For any sufficiently small S, there exists a unique


bounded continuous function 77: E > R such that 77( 0 ) = 0 and

ho = iptoh, i R , (4 .6 )

where = Id + 77 and il)t is the flow determined by the equation

x' = A x + g{x). (4 .7 )

Moreover, h is a homeomorphism.
4.2. The Grobman-Hartman theorem 131

P roof. We divide the proof into several steps.


Step 1. Existence of global solutions. We first show that equation (4.7)
defines a flow. Each solution x{t) satisfies

x{t) + [ e'^^*~^^g(x(s)) ds (4.8)


Jto
for t in the corresponding maximal interval Ix (we note that the function
(t, x) Ax + g{x) is continuous and locally Lipschitz in x, and thus one
can apply Theorem 1.43). It follows from (4.5) with y = 0 and (4.8) that

||x(t)|| <ell^ll(*-*o)||x(to)||+
Jto
for t > to in Ix- By GronwalPs lemma (Proposition 1.39), we obtain

e-ll"ll*||x(t)|| < e-ll^llo||x(to)||e^( -*\


or equivalently,
||x(t)||<e(ll^ll+^)( - o)||a;(to)||, (4.9)
for t >t o in Ix- This implies that each solution x{t) is defined in [to,+oo).
Otherwise, the right endpoint b of the interval Ix would be finite, and by (4.9)
we would have
||x(r)|| < e(ll^+^ll)(-*<)||x(to)||,
which contradicts to Theorem 1.46 on the behavior of solutions at the end
points of the maximal interval. One can show in a similar manner that all
solutions are defined in the interval (oo,to]-
Step 2. Construction of an auxiliary function. Let X q be the set of bounded
continuous functions 77: E > E with 77(0) = 0. It follows from Proposi
tion 1.30 that Xo is a complete metric space with the distance
d(77,0 = sup{||77(x)-^(x)|| : x E " } . (4.10)
We define a transformation G in X q by
f +00 f-\
r+ oo
G{rj){x) = / Pse^^T){e~^'^x)dT - PuS rj{e^'^x) dr
Jo Jo
(*+ 00
/
t r-\-oo

-00 Jt
(4.11)
for each 77 6 X q and x E"', where Pg and Pu are, respectively, the pro
jections over the stable and unstable spaces (see (4.2)). We first show that
the transformation G is well defined. Since the matrix A is hyperbolic, by
Proposition 2.27 there exist constants c,fj,>0 such that
llP.e^^ll < ce~i^^ and (4.12)
132 4. Hyperbolicity and Topological Conjugacies

for T> 0. Thus,


\\ P .e^'ri(e-''l\<ce-n \riU (4.13)
and
||F.e- ',(e''^|| < ce-'-'Ihlloo (4.14)
for r > 0, where
H loo := sup { I|i)(a!) II ; I K }.

Since ?? G X o, we have Halloo < +00- Hence, it follows from (4.13) and (4.14)
that
P+OO /*+oo
/ dr + / \\Pue~^'^r}{e'^'^x)\\dT
Jo Jo (4.15)
P+OO P+OO
ce '^||7/||oodr +
ce '*'^||7||oodr =
- Jo
/ io' ' M
and the transformation G is well defined. Moreover, by (4.15), we have

l|GW(j^)ll < ^ h llo o

for each x R ", and the function G{ri) is bounded for each t} Xo- Now we
show that G{ t]) is continuous. Given a; e R and e > 0, it follows fi'om (4.15)
that there exists T > 0 such that
f+OO />+oo
/ ||PsC'^^^(e '^^a;)||dr+ / \\Pue~^'^r]{e^'^x)\\dT < e.
Jt Jt
For each a;, y G R", we have

||G(t7)( x ) - G{T}){y)\\ < 2 e + f \\Pse^'^[r}{e~^^x) - r}{e~^'^y)]\\ dr


Jo

+ r ||Pe-^"[y(e^"x) - vie^-^y)] ||dr


Jo (4.16)
fT
<2e+ I ce ^'^||77(e "^'^a:) r]{e~^'^y)\\ dr
Jo
+ f ce~i^^Me'^^x)-rj{e'^^y)\\dr.
Jo
For each 5 > 0, the function (t, x) i-> ri{e~^*'x) is uniformly continuous on the
set [0, T] X B{x,S) (because continuous functions are uniformly continuous
on each compact set). Hence, there exists 5 > 0 such that
||T?(e~^^a;) -7?(e~^^y)|| < e
for r G [0, T] and y G B{x,S). Analogously, there exists <5 > 0 (which one
can assume to be the same as before) such that
l|?7(e^a:)-77(e^^y)|| < e
4.2. The Grobman-Hartman theorem 133

for r G [0,T] and y G B(x,S). It follows from (4.16) that


\\G{t]){x) - G{r])(y)\\ < 2e + 2cTe
for y G B{x, 5). This shows that the function G(r}) is continuous. Moreover,
by (4.11), we have G{r]){0) = 0 and thus G(Xo) c X q.
Step 3. Equivalent form of identity (4.6). Given rj e X q, let
9v(x) = 9 {h(x)),
where h = I d + v- We note that the function is bounded, because g is
bounded. It follows from (4.5) that

lli^?(a:) - 9v(y)\\ ^ <^11- y\\ + <J||r/(a;) - r]{y)\\.


Moreover,
Pt?(0) = 9{h{0)) = 5(0) = 0,
and thus 5^ G X q.

Lem m a 4.5. Identity (4.6) is equivalent to G(p^) = rj.

P r o o f o f the lem m a. If <^(5,,) = rj, then it follows from Leibnizs rule that

d
e

= e

= e- A t s { Pse'^"9n{^^^x)dr- Put-^^gr,{e^'^x)dr^l^^

= e~^*{Psgr){x) + Pu9v(^))
= e-^^g{h{e^*x)).
(4.17)
On the other hand,
d
^e-^%{e^^x) = -Ae-^%{e^^x) + e~^^^h{e^^x),
C/U C/C
and it follows from (4.17) that

-Ahie^^x) + A ) = g{h{e^^x)),

that is.
^h(e^*x) = Ah{e^^x) + g{h{e^*x)).
at
134 4. Hyperbolicity and Topological Conjugacies

This shows that the function y{i) = h{e^*x) satisfies equation (4.7). Since
y(0) = h{x), we obtain
h{e^^x) = 'ijjt{h{x))
for every x M , which yields identity (4.6).
Now we assume that (4.6) holds. Since h Id + rj, one can rewrite this
identity in the form
ld +T] ipt o ho (4.18)
On the other hand, by the Variation of parameters formula in Theorem 2.26,
we have
ipt{x) = e"^^x + [ g{ipT{x)) dr. (4.19)

Using (4.6) and (4.19), it follows from (4.18) that


rj{x) = 0t(/i(e ^ x)) - X

= e^ fi(e-" x) + r e^( -^)^(V>r(^(e " x))) dr - X


Jo
(4.20)
= e^*''q{e~^*'x) + f dr
Jo
= e ^ ^ g { e - ^ ^ + [ e^^g{h{e-^-^x)) dr.
Jo
Moreover, it follows from (4.13) that

||P,e^ 77(e-^*x)|| < ce-^%\\oo ^ 0


when t > + 00, and thus, by (4.20),
^+oo
Psil{x) = / PsC^'^g{h{e~^'^x)) dr. (4.21)
Jo
On the other hand, it follows from the third equality in (4.20) that

e ^7?(x) = 7?(e~^ x) + f e~^'^g{h(e^^'^~^^x))dT.


Jo
Substituting x by e^*x, we obtain

e ^*7(e^*x) = T]{x) + [ e~^'^g{h{e^'^x)) dr. (4.22)


Jo
Analogously, it follows from (4.14) that

||Pe-^*77(A)||<ce-^*||r?||oo^O
when t >00, and thus, by (4.22),
f+OO
PuVix) = - Pue~^'^g{h{e^'^x)) dr. (4.23)
Jo
4.2. The Grobman-Hartman theorem 135

Adding (4.21) and (4.23), we finally obtain


r]{x) = PsVix) + PuV{x)
r+oo f+oo
= / PsS^'^g{h{e~^'^x)) dr Py,e~^'^g{h{e^'^x)) dr

= G{goh)(x) = G{gri){x).
This yields the desired result.

Step 4- Construction of the conjugacy. Now we show that the equation


G{grj) = ri has a unique solution rj.
Lem m a 4.6. For any sufficiently small S, there exists a unique function
g e X o satisfying G{grf) = g.

P r o o f o f the lem m a. We define a transformation F in X q by

F{ri) = G(g^). (4.24)

Since grj X q for rj X q, we have F{Xo) C A q (because G{Xo) c X q).


Now we show that F is a contraction. For each X q, we have

Joo
+00

It follows from (4.5) that


/ [^^(e^("-*)x) - p(e^("- )a;)] dr.

||s,(e'"'- )x) - ss(e' <'-x)|| < 5|k(e''<'- >x) - {(e^W-Ox)!!


< Sd{i],^),

and hence, using (4.12), we obtain

||F(t/)( x ) - F{0{x)\\ < f ce-f^^*-'^'>6d{v,0 dr


JOO
+00

/
2c(5 .
ce~^^^~'^^5d{'q,
.
= d{ri,^).
dr

Thus,
2c5
d{F{rj),F{0) ^
and F is a contraction for any sufficiently small 5. Since X q is a complete
metric space and F (X q) C X q, it follows from Theorem 1.35 that there
exists a unique g E X q such that F(??) V-
136 4. Hyperbolicity and Topological Conjugacies

By Lemma 4.5, the function r) given by Lemma 4.6 satisfies identity (4.6).
It remains to show that /i = Id + ?7 is a homeomorphism.
Step 5. Existence of the inverse. We first show that h is surjective. Given
y M", the equation h{x) = y has a solution of the form x = y z ii and
only if
z = -^{y-l- z). (4.25)
Now we recall Brouwers fixed point theorem: aay continuous function
H: B B in a closed ball B C M has at least one fixed point (for a
proof see, for example, [10]). Applying this theorem to the function

z !-> H{z) = -T]{y + z)

in the closed ball B = B{0, ||7?||oo)) we conclude that there exists at least
one point z B satisfying (4.25). This shows that h is surjective.
Now we show that h is injective. Given x,y such that h{x) = h{y),
it follows from (4.6) that

h{e^^x) = h{e^*y)
for t G R, and thus,

e^*(x - y) = r){e^*'x) + r]{e^^y). (4.26)

If Ps{x y) 7^ 0 or Py,{x y) ^ 0, then the left-hand side of (4.26) is


unbounded while the right-hand side is bounded. This contradiction shows
that
- y ) = Pu{x - y) = 0
and hence x = y.

Step 6. Continuity of the inverse. We first recall the Invariance of domain


theorem: ii f : U -> R is an injective continuous function in an open set
C/ C R , then V = f(JJ) is open and /|!7: 17 > F is a homeomorphism (for
a proof see, for example, [10]). Since the function h: R ^ R is continuous
and bijective, it follows from the theorem that /i is a homeomorphism.

4.2.2. H y p erb olic critical points. Now we establish the result described
at the beginning of Section 4.2 as a consequence of Theorem 4.4.

T h eorem 4.7 (Grobman-Hartman theorem). Let / : R -> R " be a func


tion of class and let xq R 6e a hyperbolic critical point of the equa
tion x' = f{x). If if}t{z) and <pt{z) are, respectively, the solutions of the ini
tial value problems in (4.3), then there exists a homeomorphism h: U ^V,
where U and V = h{U) are, respectively, neighborhoods of xq and 0, such
that h{xo) = 0 and h{'ijjt(z)) = <pt[h{z)) whenever z,ipt{z) G U.
4.2. The Grobman-Hartman theorem 137

P ro o f. Without loss of generality, we assume that xq = 0. Indeed, making


the change of variables y = x xq, the equation x' = /( ) becomes y' = F(y),
where F(y) = f(xo + y). Moreover, F(0) = f(xo) = 0.
Taking a;o = 0, we write
f (x ) = Ax + g(x),
where
A = dof and g{x) = f { x ) - Ax.
We note that the matrix A is hyperbolic and that the function g is of class .
Moreover, 5 (0) = 0. In order to apply Theorem 4.4 it remains to verify
property (4.5). However, in general, this property may fail, which leads us
to modify the function g outside a neighborhood of 0. More precisely, given
5 > 0 as in Theorem 4.4, take r > 0 so small that
sup{||da;^|| : X B{0,r)} < 5/Z (4.27)
(this is always possible, because the function x 1-^ dxg is continuous and
= 0). Since ^(0) = 0, it follows from the Mean value theorem that
sup{|fy(a:)|| : x G B (0 ,r )} < 5r/Z. (4.28)
Now we consider a function a : R > [0,1] of class such that:
a) a{x) = 1 for a; G B{0,r/Z);
b) a{x) = 0 for a; G R \ H (0,r);
c) sup {||da;a|| : x G R } < 2/r.
For the function p : R R of class defined by
g{x) = a{x)g(x), (4.29)
we have 5 (0) = 0. Moreover, by (4.27) and (4.28), we obtain
= \\dxag{x) + a{x)dxg\\
< sup ||da:a;|| sup |fy(a;)||+ sup ||d*5 ||
leE" leB(o.r) leK" (4.30)
2 5r S .

Hence, it follows from the Mean value theorem that property (4.5) holds for
the function g. Thus, one can apply Theorem 4.4 to obtain a homeomor-
phism h ld + g satisfying
ho = t G R,
where 4>t{z) is the solution of the initial value problem

{ x' = Ax + g(x),
g(0) -- z.
138 4. Hyperbolicity and Topological Conjugacies

But since g coincides with g on the ball B (0 ,r/3 ), we have ipt{z) =


whenever ||2;|| < r/3 and ||Vi('2:)|| < r/3 . This establishes identity (4.4) for
any t M and z G M such that ||z|| < r/3 and ||^t()|| < r/3 .

4.2.3. Stability under p ertu rbations. It should be noted that Theo


rems 4.4 and 4.7 do not require the stable and unstable spaces to have
positive dimension, that is, the theorems also include the case when one of
them is the whole space M . This observation has the following consequence.

T h eorem 4.8. Let A be a n n x n hyperbolic matrix and let 5 : R -> M" be


a bounded function with g{0) = 0 such that property (4.5) holds for every
a;,y g R .
a) If A has only eigenvalues with negative real part and 6 is sufficiently
small, then the zero solution of equation (4.7) is asymptotically sta
ble.
h) If A has at least one eigenvalue with positive real part and 6 is suf
ficiently small, then the zero solution of equation (4.7) is unstable.

P roof. It follows from (4.6) that

ipt = ho e^*' oh.-1 (4.31)

where h is the homeomorphism given by Theorem 4.4. Since A has only


eigenvalues with negative real part, we have ||e'^*|| > 0 when t ) -l-oo.
Thus, since h{0) = 0, it follows from (4.31) that the origin is asymptotically
stable. Analogously, when A has at least one eigenvalue with positive real
part, we have ||e"^ || + 0 0 when t -t-00, and it follows from (4.31) that
the origin is unstable.

The following result is a consequence of Theorem 4.7.

T h eorem 4.9. Let /: R >R be a function of class and let xq G R


be a hyperbolic critical point of the equation x' = f{x).
a) If the matrix dx^f has only eigenvalues with negative real part, then
the critical point xq is asymptotically stable.
b) If the matrix dxgf has at least one eigenvalue with positive real part,
then the critical point xq is unstable.

P roof. It is sufficient to consider identity (4.4) and proceed in a similar


manner to that in the proof of Theorem 4.8.

We note that the first property in Theorem 4.9 is also a consequence of


Theorem 3.13.
4.3. Holder conjugacies 139

4 .3 . H o ld e r c o n ju g a c ie s

It is natural to ask whether the homeomorphisms in Theorems 4.4 and 4.7


have more regularity. Unfortunately, in general, it is not possible to obtain
conjugacies of class C^, although a detailed discussion falls outside the scope
of the book. It is, however, possible to show that those homeomorphisms
are always Holder continuous. More precisely, we have the following result.

T h eorem 4.10. Let A be a hyperbolic n x n matrix and let M >M


be a bounded function with g{0) = 0 such that property (4.5) holds. Given
a sufficiently small a G (0,1) and K > 0, there exists 6 > 0 such that the
homeomorphism h in Theorem 4-4 satisfies

\\h{x)-h{y)\\ < - y||

for every , y 6 E xvith ||a; y|| < 1.

P roof. Let X q be the set of bounded continuous functions r;: R )R with


r^(0) = 0. Given a (0,1) and K > 0, we consider the subset Xa C X q
composed of the functions rj G X q such that

\\v{x)-v(y)\\<K\\x-yr

for every x, y G R with ||x y|| < 1. Using Proposition 1.30, one can easily
verify that Xa is a complete metric space with the distance in (4.10). Thus,
in order to prove the theorem it suffices to show that the transformation F
in (4.24) satisfies F{Xa) C Xa- Indeed, we already know from the proof of
Theorem 4.4 that F is a contraction in X q.
Take y G Xa. We have
||h(x) - h{y)\\ = ||x - y + ffix) - ri(y)\\
< l k - y | | + N (a ;)-^ (y )ll-

Since rj G Xa, whenever ||x y|| < 1 we obtain

||h(x) - h{y)\\ < \\x - y|| + K\\x - y||


(4.32)
< ( l + F )| | x -y | r

Now we consider the difference


W = F( t7)(x) - F(7?)(y)

- L 00
+00
jT ^h)]dr,
140 4. H yperbolicity and Topological Conjugacies

where = g oh. Prom the proof of Theorem 4.4, we know that g^ ^ X q.


Thus, it follows from (4.5) and (4.12) that

IIW ^ II < c r e - ^ ( - ^ ) 5 | l7 7 ( e ^ ( ^ - )a :) - r7 (e ^ (^ " V ) I I dr


J OO
-oo
+00
+ c
/ e-M0-^)5||;,(e^(^-*)a;) - 77(e" (-*)y)|| dr.

On the other hand, by (2.11), we have ||e"^*|| < for t M. Hence,


by (4.32), we obtain

||W|| < c5(l + K) f e-^(*-^)||e^(^-*)(a: - y)|| dT


J OO
+0O

/ e-^( -^)||e^(^- )(x -y )| | dr

< c5{\ + K)\\x - y|r f dr


J OO
+00

= c5{l + K)\\x - y|| r


/ dr
J OO
+00

For any sufficiently small a, we have p o:||.A|| > 0, and hence,


/
2c5(l + K ) ,
IIW^II < lk -y ir (4.33)
/^ - P I I
Thus, for any sufficiently small 5, it follows from (4.33) that

||F(77)(x)-F(r?)(y)||<iir||x-yr
for every x ,y . M with ||x y|| < 1. This completes the proof of the
theorem. D

Under the assumptions of Theorem 4.10 one can also show that

| | h -H x )-h -n y )ll< i^ lk -y r
for every x^y E MP with ||x y|l < 1, eventually making a and 6 sufficiently
smaller.
Now we consider the particular case of a hyperbolic critical point. The
following result is a simple consequence of Theorems 4.7 and 4.10.

T h eorem 4.11. Let / : be a function of class and let G


be a hyperbolic critical point of the equation x' = fix). If and
4.4, Structural stability 141

are, respectively, the solutions of the initial value problems in (4.3), then for
the homeomorphism h in Theorem J^.l, given a sufficiently small a G (0,1)
and K > 0, there exists d > 0 such that
||/i(x) - % ) | | < K\\x-y\\
for every x ,y g U with \\x y\\ < d.

4.4. Structural stability

We also establish a result on the existence of topological conjugacies for


arbitrary linear and nonlinear perturbations.

T h eorem 4.12 (Structural stability). Let be a function of


class and let xq be a hyperbolic critical point of the equation x' = f{x).
// 5 : is a function of class and
a := sup ||/(a;) - i^(x)|| + sup \\dxf - dxg\\
xeMV' xew^
is sufficiently small, then the solutions of the equations
x^ = f {x ) and x' g{x)
are topologically conjugate, respectively, in neighborhoods ofxo andxg, where
Xg is the unique critical point of = g{x) in a sufficiently small neighbor
hood of Xq.

P ro o f. We first show that the equation = g{x) has a unique critical


point Xg in a sufficiently small neighborhood of xq, provided that a is suffi
ciently small. To that effect, consider the function H: x E^ E^
defined by

where (M ) is the set of all functions ^ : R >R of class such that

||f?||oi := sup ||f?(a;)|| + sup ||4i?|| < +oo.

One can show that the Implicit function theorem holds in the space C^(R )
when equipped with the norm IHIci. Now we show that H is of class C^,
taking the norm
ll(5,y)ll = IMIci + IMI
in C^(R ) X R . For each {g,y), (h,z) G C'^(R ) x R and e G R, we have
H{{g, y) + e{h, z)) - H{g, y) _ {g + eh){y + ez) - g{y)

^ gjy + g -g ) - 9{y)
+ h{y + ez)
e
dygz + h{y)
142 4. Hyperbolicity and Topological Conjugacies

when e 0. Take
A := H {{g , y) 4-{ h , z ) ) - H {g , y) - h{y) - dygz

= {9 + h)(y + z ) - g{y) - h{y) - dygz


= [g{y + ^) - g{y) - dygA + [h{y + z ) ~ h{y)].
Since
h{y + A - h{y) = / d y + tz h z d t,
Jo
we have
1 1 % + 2 ) - % ) II < ||/i||cip||,
and thus,
INI . \\gjy + A - gj y) - dygzW ||h||ci|NI .
Il(/i,^)l|- Nl INIci + INI
when ||(h, ^)|| 0. This shows that the function H has a derivative, given
by
dig,y)H(h, z) = h{y) + dygz. (4.34)
In order to verify that H is of class (7^, we note that
\\(d(g,y)H d(^ig)H){h, z)||
< ||h(y) - /i(y)|| + \\(dyg - dyg)z\\

< ll^llci lb - y|l + 1 1 % - % l l Ibll + 1 1 % - % l l Ibll


< Ibllcilb - i/ll + lb - 5llci|bll + 1 1 % - % l l Ibll.
and thus,

\\dlg,y)H - d^g,y)H\\ < lb - y|| + lb - p IIc-i + 11% - % l l - (4-35)


Since dg is continuous, it follows readily from (4.35) that the function dH
is also continuous. Moreover, by (4.34), we have
dH
(5 , y ) z = d(p,3;)i3'(0, z ) = d y g z ,

and thus.
^ (/,a ;o ) = % / .

Since the matrix dx(,f is invertible (because it is hyperbolic), it follows from


the Implicit function theorem that for g 6 (^^(M ) with ||/p||c7i sufficiently
small, there exists a unique Xg in a sufficiently small neighborhood of xq
such that g{xg) = 0. Moreover, the function ^ is continuous.
Now we consider the pairs of equations
x' = f {x ) and y' = dxofy (4.36)
and
x' = g{x) and y' = dxggy- (4.37)
4.5. Exercises 143

By the Grobman-Hartman theorem (Theorem 4.7) there exist topological


conjugacies between the solutions of the equations in (4.36) respectively in
neighborhoods of the points xq and 0, as well as between the solutions of
the equations in (4.37) respectively in neighborhoods of the points Xg and 0.
To complete the proof, it is sufficient to observe that since the eigenvalues
of a matrix vary continuously with its entries and the function g Xg is
continuous, for ||/5||c7i sufficiently small the matrix dx^g is also hyperbolic
and
midxgg) = m{dxof),
where m{A) is the number of eigenvalues of the square matrix A with pos
itive real part, counted with their multiplicities. Hence, it follows from
Theorem 2.50 that the solutions of the equations y' = dx^fy and y' = dxggy
are topologically conjugate. Since the composition of topological conjugacies
is still a topological conjugacy, we obtain the desired result.

4 .5 . E x ercises

E xercise 4.1. Find all hyperbolic critical points of the equation defined by:
a) f{x) = x { x - 1);
b) fix) = x^;
c) f{ x ,y ) = (x + y , y - x ^ ) .
E xercise 4.2. Determine the spaces E and of the origin for the func
tion:
/I -3 0 0 \ x\
3 1 0 0 y
a) f{ x, y ,z ,w ) =
0 0 -2 1 z
lo 0 -1 -2 / \wj
b) f{ x, y ) = { x - y ^ , x ^ - y ) .
E xercise 4.3. Under the assumptions of Theorem 4.4, show that if a trans
formation h satisfying (4.6) is differentiable, then:
a) dgAtxhe = dfi^x)'^tdxh]
b) dgAtj.hAe^^x = Ah{e^^x) + g{h{e^*'x))\
c) = doiptB, where B = doh;
d) dxhAx = Ah(x) + g{h{x)).
E xercise 4.4. Verify that h{x, y) = (x, y-bx^/5) is a differentiable conjugacy
between the solutions of the equations

x' = X,
and
\y' = - y
144 4. Hyperboiicity and Topological Conjugacies

Exercise 4.5. Discuss the stability of the zero solution of the equation

a;' = a: 2sina; + y,

{;y ^ y + xy.
Exercise 4.6. Consider the equation

ix' = yf{x,y),
(4.38)
= -xf{x,y),

where f { x ,y ) = {x^ + - l)(y - - 2).


a) Sketch the phase portrait.
b) Show that all solutions are global.
c) Find all homoclinic orbits.
d) Find all periodic orbits.
e) Show that in a neighborhood of the origin the periods T (r) of the
periodic orbits, where r = V x ^ " + ^ , satisfy T (r) = tt + ar^ + o(r^)
for some a ^ 0.
f) Find whether the solutions of equation (4.38) are topologically con
jugate to the solutions of the linear equation

= 3//(0,0),
\y' = - x / ( 0, 0)
in a neighborhood of the origin.

Exercise 4.7. Consider the equation

f x ' = yy(x,y),
\y' = -x y (x ,y ),
where
y(x,y) = (x^ + y^ - l)[y^ - (x^ -I- 2f][x^ - (y^ + 2f].
a) Sketch the phase portrait.
b) Find all global solutions.
c) Find all homoclinic orbits.
d) Find all periodic orbits.
e) Show that in a neighborhood of the origin the periods T{r) of the
periodic orbits satisfy T'{r) 7^ 0 for some r 7^ 0.

Exercise 4.8. Consider the equation x" -I- sinx = 0.


a) Find an integral for the equation.
b) Sketch the phase portrait.
4.5. Exercises 145

c) Determine the stability of all critical points.


d) Find whether there exist periodic orbits of arbitrarily large period.
e) Find whether there exist periodic orbits of arbitrarily small period.
f) Find whether there exist periodic orbits of period tt^.

Solutions.
4.1 a) 0 and 1.
b) There are no critical points.
c) (0,0) and (1,1).
4.2 a) E^ = {(0 ,0 )} X and = R^ x {(0 ,0 )}.
b) = {0 } X R and = R x {0 }.
4.5 It is unstable.
4.6 c) { ( x , y ) 6 R 2 : x 2 + j/2 = 4 } \ { ( 0 , 2 ) } .
d) {(x ,y ) R^ : + y^ r^} with r G (0,1) U (1,2).
f) They are not.
4.7 b) All solutions are global.
c) There are no homoclinic orbits.
d) {() y) R^ : x^ + y^ = r^} with r (0,1) U (1,2).
4.8 a) y^/2 cosx.
c) ((2n + 1)7T, 0), n G Z is unstable and {2mr, 0), n G Z is stable but
not asymptotically stable.
d) Yes.
e) No.
f) Yes.
Chapter 5

Existence of Invari2int
Manifolds

In this chapter we continue the study of the consequences of hyperbolicity.


In particular, we establish the Hadamard-Perron theorem on the existence of
invariant manifolds tangent to the stable and unstable spaces of a hyperbolic
critical point. The proofs use the fixed point theorem for a contraction in
a complete metric space (Theorem 1.35) and the Fiber contraction theorem
(Theorem 1.38). For additional topics we refer the reader to [7, 18, 19, 23].

5.1. Basic notions

Let / : R > R be a function of class and let xq be a hyperbolic


critical point of the equation x' = f{x ). Also, let and <pt{z) be,
respectively, the solutions of the initial value problems in (4.3). By the
Grobman-Hartman theorem (Theorem 4.7), there exists a homeomorphism
h: U F , where U and V are, respectively, neighborhoods of xq and 0,
such that h{xo) = 0 and the identity

H M ^ )) = rt{h{z)) (5.1)

holds whenever U. In particular, h~^ transforms each orbit


(pt{h{z)) of the linear equation y' = Ay, where A dxof, into the orbit
tpt{z) of the equation x' = f{x ). This yields the qualitative behavior in
Figure 4.1. Moreover, the stable and unstable spaces and (see Defi
nition 4.2) are transformed by the homeomorphism h~^ onto the sets

V^ = h~\E ^nV ) and F = -h n F). (5.2)

147
148 5. Existence o f Invariant Manifolds

Without loss of generality, one can always assume that the open set V
satisfies

n V )cE ^n V and <p-t{E^ f)V) C ETnV (5.3)

for every t > 0. Indeed, using the construction in the proof of Theorem 2.50,
one can show that there exist open sets 5 c E^ and 5 C containing
the origin such that

and c
for every t > 0. Namely, given r > 0, one can take
P+OO
x e E >: / ||e"a ;f d t < r
Jo }
and
5
= {x e E ^ : J \\e^^xfdt<r'
}
Then the open set V = x 5 satisfies (5.3) for every t > 0. Moreover,
taking r sufficiently sinall, one can assume that U = h~^{B^ x 5 ) is the
open set in the statement of the Grobman-Hartman theorem (Theorem 4.7).
In what follows, we always make these choices.

P rop osition 5.1. Let / : R )R be a function of class and let xq be


a hyperbolic critical point of the equation x' = f{x). Then
V^ = { x e U : M ^ ) e U fo r t > 0 } (5.4)

and
V^ = { x U : ^t{x) e U fo r t < O}. (5.5)
Moreover, for each t > 0 we have

M V ") C and V>-t(^ ) C V^. (5.6)

P roof. We first observe that

E ^ d { x e V : <pt{x) y for t > 0}


and
E T d { x V : ipt(x) e V i o r t < O}.
This follows from the Jordan canonical form of the matrix dxof- Hence,
by (5.3), we obtain

E^nV = { x V : F for f > 0}

and
E ^ n v = { x e v : <pt{x) e V i o r t < o }.
5.2. The Haxiamard-Perron theorem 149

Using (5.1) and (5.2) one can write


= h~\E^ n v )
= |/r ^(x) h~^{V) : {(pt o h){h~^{x)) U for t > 0}
= {/r ^(o;) U : {ho'if}i){h~^{x)) G V for t > 0 }
= {/i"Ha;) e U : Mh~Hx)) e U fort > 0 } ,
which establishes (5.4). Identity (5.5) can be obtained in an analogous
manner. Finally, the inclusions in (5.6) follow easily from (5.3). Indeed,
applying h~^ to the first inclusion we obtain
{h~^ o cpt){E^ n F ) C h~\E^ n F ) = F
for t > 0, but since h~^ o = 0^ o h~^, we conclude that
^tiV^) = {^toh-^ ){E^n V)
= {h-^oipt){E^nV) C F
for t > 0. The second inclusion in (5.6) can be obtained in an analogous
manner.

5.2. The Hadamard-Perron theorem

In fact, if / is of class C'^, then the sets F* and F in (5.2) are manifolds of
class C*, respectively, tangent to the spaces E^ and E^ at the point xq (see
Figure 5.1). This means that in a neighborhood of xq the sets F and F are
graphs of functions of class C'^. This is the content of the following result.

T h eorem 5.2 (Hadamard-Perron). Let / : R ^ R 6e a function of


class C'^, for some /s G N, and let x q be a hyperbolic critical point of the
equation x' = f{x). Then there exists a neighborhood B of x q such that
the sets F n H and F D H are manifolds of class containing x q and
satisfying
(F n H) = F; and T^o iV " n B ) = E^.

We shall prove the theorem only for A: = 1 (for the general case see,
for example, [2]). The proof (for A; = 1) consists of showing that in some
neighborhood B of x q the sets F fl H and OB are graphs of functions
<p:E ^nB ^E ^ and ip -.E '^ n B ^ E ^ (5.7)
of class C^, that is,
V^nB = { { x , ^{ x ) ) : x E^nB} ,
V ^ n B ^ {{i}{x),x)-.xeE rn B ].
The argument consists essentially of two parts: first, we show that there
exist Lipschitz functions <p and ip (see Definition 1.31) satisfying (5.8), and
150 5. Existence o f Invariant Manifolds

F ig u re 5.1. Manifolds V and V .

then we show that these functions are of class C^. The techniques used
in both parts of the proof are also of different nature. For clarity of the
presentation, in Section 5.3 we prove that there exist Lipschitz functions tp
and Ip, and we establish their regularity in Section 5.4.
Motivated by Theorem 5.2 we introduce the following notions.
D efinition 5.3. For a function / : R >R of class and a hyperbolic
critical point xo of the equation a;' = f(x), the sets D B and DB
in Theorem 5.2 are called, respectively, (local) stable manifold and (local)
unstable manifold of xq.

Due to the inclusions in (5.6), it is also common to say that the stable
and unstable manifolds are invariant manifolds.

5.3. Existence of Lipschitz invariant manifolds

In this section we establish the existence of Lipschitz functions ip and ip as


in (5.7) that satisfy the identities in (5.8).
T h eorem 5.4. Let / : R > R be a function of class and let xo be
a hyperbolic critical point of the equation x' = f{x). Then there exist a
neighborhood B of xq and Lipschitz functions
(p :E ^ n B ^ E ^ and i p - . E T n B ^ E ^ (5.9)
satisfying the identities in (5.8).
5.3. Existence o f Lipschitz invariant manifolds 151

P ro o f. We only consider the set n B, since the result for D B can


then be obtained in a simple manner. Namely, let p: R M be the
function g{x) = f{x). We note that if ipt{xo) is the solution of the equation
x' = f{x) with x(0) = xq, for t in the maximal interval Ix^ = (a, b), then

^ ip - t {x o ) = - /( V -t(o))
for every t M. and xq G R such that t e Ixo- This shows that
'^t{xo) = 'i/j-tixo) is the solution of the equation x' = g{x) with x(0) = xq,
for t 6 (6, a). It follows from the identity dxo5 = ~dxof that xq is also a
hyperbolic critical point of the equation x' = g{x), with stable and unstable
spaces, respectively, E'^ and E^. Now let Vg and Vg be the correspond
ing sets obtained as in (5.2). By Proposition 5.1, there exists an open set
17 C R such that
= {x G 17 : ^t(x) G 17 for t > 0 }
= { x G 17 : ipt{x) G 17 for t < O},
and hence DB = for any sufficiently small neighborhood B of xq.
Thus, the result for n B follows from the result for P n B with respect
to the function g.
Moreover, without loss of generality, from now on we assume that xq = 0.
This can always be obtained through a change of variables (see the proof of
Theorem 4.7).
Step 1. Equivalent formulation of the problem. Let us consider coordinates
(x,y) E^ X E^ (recall that B 0 = R "'). For the equation x' = /( x ),
it follows from (5.6) that if (z,(p(z)) G P is the initial condition at time
to = 0, then the corresponding solution remains in P for all t > 0, and thus
it is of the form (x(t), ip{x{t))), for some function x(t).
As in the proof of Theorem 4.7, we write
/( x ) = Ax + g{x),
where
A = dof and g{x) = / ( x ) Ax.
We also consider the function g in (4.29), which satisfies (4.30). For any
sufficiently small z, it follows from the Variation of parameters formula in
Theorem 2.26 that
rt
x{t) = Pse'^^z + [ '^'>g{x{T),(p{x{T)))dT (5.10)
Jo
and

(p{x{t)) = Pue^^cp{z) + f Be^( '^)p(x(r),v?(x(r)))dr (5.11)


7o
152 5. Existence o f Invariant Manifolds

for t > 0 (because then the solution belongs to the neighborhood of zero
where g and g coincide). The idea of the proof is to use these identities and
a fixed point problem to establish the existence of a Lipschitz function ip.
Step 2, Behavior of the stable component Consider the set Y of all contin
uous functions ip: E'^ such that (/?(0) = 0 and
\W{x)-ip{y)\\<\\x-y\\, x,y E \ (5.12)
It is a complete metric space with the distance
\\(f{x) - 1p{x)
= sup|J : a; \ {0 }
uc I }
(see Exercise 1.14). Given (/? G F, let a; = x^p be the unique function of
class satisfying (5.10). In order to verify that it exists and is unique, we
note that since the right-hand side of (5.10) is differentiable in t, one can
take derivatives to obtain

x ' { t ) = A P sC ^ ^ z + a [ Pge'^^*~'^^g{x{T), (p ( x { t ) ) ) d r + P s g {x {t), (p{x{t)))


Jo
= A x {t) + P sg{x{t),(p{x{t))).

Now we define
hp{x) = Psg{x,(p{x)). (5.13)
Taking r = 0 in (4.12) we obtain ||Ps|| < c. Hence, it follows from (4.30)
and (5.12) that
\\h^{x) - h^{y)\\ < c5\\{x,g>{x)) - {y,<p{y))\\
= c5\\{x - y, <fi{x) - (^(y))||
<c5\\x-y^-\-c5\\ip{x)-(p{y)W
< 2c5\\x - y\\.
The Picard-Lindelof theorem (Theorem 1.18) ensures that there exists a
unique solution of the equation
x' = Ax + h^{x) (5.15)
for each initial condition with values in E^. By (5.10), the function x = x,p
takes only values in E. Moreover, it follows from (4.12), (4.30) and (5.10)
that
||l(()|| < ce-'||z|| + / ce-'<-')||(l(T).v(l(T)))||<ir
Jo
in the maximal interval of the solution x{t). Since
||v?(a;(r))|| = ||v?(a;(r)) - 9p(0)|| < ||a;(T)||,
we obtain ^
||x(t)|| < ce-^ ||^|| f ce-^(*-^)2(5||a;(T)|| dr.
Jo
5.3. Existence of Lipschitz invariant manifolds 153

Hence, it follows from Gronwalls lemma (Proposition 1.39) that

e-llxWII < c\\z\\e^\


or equivalently,
l|x(t)|| < (6.16)
for t > 0 in the maximal interval of the solution. Eventually taking 6 smaller
so that p + 2c6 < 0, one can proceed as in the proof of Theorem 3.12 to
show that the function x{t) is well defined and satisfies (5.16) for every t > 0 .
Step 3. Some properties of the function x^p. Given z^z e let Xip and x^p be
the solutions of equation (5.15), respectively, with Xy,(0) = 2: and x<^(0) = z.
Since
\\g{x^{T),ip{x^{r))) - g{x^{T),(f{x^{T)))\\
< <5(||a;^(r) - x^(r)|| + \\^{x ^{ t )) - (^(x^(r))||) (5.17)
< 25\\x^{t ) - x^{t )\\,

writing
p{t) = \\x^{t) - x^{t)\\
it follows from (5.10) and (4.12) that

p{t) < ce~'^*'\\z z\\ + f c e ^^* '^ ^ 2 ( 5 p ( r ) dr


Jo
for t > 0. By Gronwalls lemma, we obtain
p{t) < c e ( - ^ + 2 c 5 ) t ||_ ^ _ ^11 (5 18)

for t > 0.
Given ip, ip E Y, let Xy, and x^j, be, respectively, the solutions of the
equations (5.15) and
x' = Ax + h , p { x )
with X(^(0) x.0 (O) = We have
||p(x^(r),(^(x^(r))) - g{x^{T),
0 (x^(r)))||
< \\9 M t ) , p { x ,^{ t ) ) ) - g { x ^ { T ) , i p { x ^ { T ) ) ) \ \
+ \\9Mr),ip{x^{T))) - g{x^{T),ip{x,p{T)))\\
< 5\\p {x,p{t )) - 0(x^(r))|l + (51|x^(r) - x^(r)|| (5.19)
< S y M r ) ) - iP{x^{t ))\\ + 5\\iP{x^{t )) - ip{x,p{T))\\
+ (5||x^(r)-x^(r)||
< (5d((^,V)||a:<^(r)|| + 2(5||x^(r) -x^(r)||.
Writing
p{t) = llx^(f) -x^(t)||.
154 5. Existence o f Invariant Manifolds

it follows from (5.10) and (5.16) that

p(t) < ce-'*(*-^)(5d(<^, V)ce(-^+2ci)rii^ll ^ f* ce~f^{t-r)2Sp(r) dr


Jo Jo
p2c5r\r=t pt
<c2,5d((^,V')e-^ V r l k l l + / ce(-^+2c5)(t-r)25p(.^)^^
Jq

^ Jo
that is,

g(M-2c5)t^(^) < ld{<p,'iP)\\z\\ + f 2c5e^>^-^<^^^p{r) dr.


^ Jo
Hence, it follows from Gronwalls lemma that

or equivalently.

P(t) < (5.20)


for every t > 0.

Step 4- Behavior of the unstable component Now we show that

p{x^{t)) = Pu^*p { z )+ f Py,e^^^ '^^g{x^{T),(p{x^{T)))dT (5.21)


Jo
for every t > 0 if and only if
f+O O
<P{^) = - Pue~^'^g{x^{T),p{x^{T)))dT (5.22)
Jo
for every t > 0. We first show that the integral in (5.22) is well defined.
Indeed, it follows from (4.12), (5.12) and (5.16) that

\\Pue~'^^g{x^{t),p{x^{t)))\\ < ce~>^^25\\x^{t)\\


< ce-^ 2<5ce(-'^+2^) ||z|| (5.23)
= 2c25e(-2^+2c<5)t||^||_

Since 2p + 2c5 < 0, we obtain


f+ o o r+oo
/ \\Pue~'^'^g{x^{T),ip{x^{T)))\\dT < 2c^6 dr < +oo,
Jo Jo
which shows that the integral in (5.22) is well defined.
Now we assume that identity (5.21) holds for every t > 0. Then

(p{z) = Pue~^^(p{x^(t)) - [ Pue~^'^g{x^{T),ip{x^{r)))dT. (5.24)


Jo
5.3. Existence o f Lipschitz invariant manifolds 155

Proceeding as in (5.23), we obtain


\\Pue-^^ip{x^{t))\\ < ce-^ ce(-^+2ci)i||^||
< c2e(-2^+2c5)t||_^|| 0

when t + 00. Thus, letting t > +oo in (5.24), we conclude that (5.22)
holds. On the other hand, if identity (5.22) holds for every t > 0, then

P e ^ V (^ )+ [ Pue"^^*~'^^9{x^{T),<p{x^{T)))dT
Jo
+00

/ Pue^^^~'"^Hx<p{r),(p{x^{r))) dr.

We show that the right-hand side of this identity is equal to ip{x^{t)). De


noting by Ft{z) the solution of equation (5.10), which is also the solution of
the autonomous equation (5.15) with initial condition F q { z ) z , it follows
from Proposition 1.13 that
X^{ t ) = F r {z ) = F r -t{,F t{z)) = F r - t { x ^ { t ) ) .

We note that identity (5.22) can be written in the form


f+O O
cp{z) = - / Pue-^^g{Fr(z), <p(Fr(z))) dr. (5.25)
Jo
Hence, it follows from (5.25) that
r+ o o
Pue'^^*~'"'>gix^(T),(p{x^{T))) dr
- /
+00

/ +00
Pue^^^-'^'>g(Fr-t{x^{t)), cp {F r-t(x ^ m ) dr

PuC ^'~g{Fr{x^{t)), (p{Fr{x^{t)))) dr


- f

Step 5. Existence of the function g>. Now we use identity (5.22) to show
that there exists a unique function <p g Y satisfying (5.21). To that effect,
we define a transformation T in y by
r+oo
T{(p){z) = - Pue~^'^g{x^{T),(p{x^{T)))dT. (5.26)
Jo
Since p(0) = 0 and x^{ t ) = 0 when 2 = a:y,(0) = 0, we have T((/?)(0) = 0.
Given z,z G E^, let x^ and x^ be again the solutions of equation (5.15),
respectively, with x<^(0) = 2: and x^{Qi) = z. It follows from (5.17) and (5.18)
that
\\g{x^{r),<pMr))) - 9{x^{r),9>{x^{r)))\\ < 2c5e( ^+2='^)^||z - f ||,
156 5. Existence o f Invariant Manifolds

and thus,
/+00
\\T{^){z)-T{<p){z)\\< / - z|| dr
Jo
c^6
<
H cS
for any sufficiently small 6. This guarantees that T{Y) c Y.
Now we show that T is a contraction. Given E Y and z E E^,
let and be the solutions, respectively, of the equations (5.15) and
x' = Ax + h^{x) with x^(0) = x.^(0) = 2:. It follows from (5.19), (5.20)
and (5.16) that

||p(x^(r), p {x ^{ t ))) - 9 {x^{ t ), tp{x^{T)))\\


< 6d{p,'ip)\\x^{T)\\ + c5d{(p,tp)\\z\\e^~^'^^^'>'^

and thus,
^+00
\\T{p){z) -T{ip){z)\\ < / ce~>^'^2c6d{<p,ip)\\z\\e^~f^+^^'>'^dT
Jo
/+00
= / 2c^6d{p,'ip)\\z\\e^~^^'^^^^'^dr
Jo
c^6
< d{p,-ip)\\z\\
p 2c5
This implies that

(?5
d {T {p),T m < d{p,tp). (5.27)
p, 2c6
For any sufficiently small 5, the transformation T is a contraction. Thus,
by Theorem 1.35 there exists a unique function p e Y such that T{p) = v?,
that is, such that identity (5.22) holds. This shows that there exists a
unique function p e Y satisfying (5.10) and (5.11). Since ^ ^ in the ball
B = 5 (0 , r/3 ) (see the proof of Theorem 4.7), we obtain

V^r\B = {{z, p{z)) .zEE^r\B).

This completes the proof of the theorem.

It also follows from the proof of Theorem 5.4 that along F D 5 the
solutions decrease exponentially. More precisely, given .2 5* D 5 , let

t ^ {x^{t),p{x^{t)))
5.4. Regularity o f the invariant manifolds 157

be the solution of the equation x' = f {x ) with x^{0) = z. It follows from


(5.12) and (5.16) that
ll(x^((),>(i(()))ll < 2||xy(()|| < 2 c e '-'+=*) ||2||, t > 0.

5.4. Regularity of the invariant manifolds

In this section we conclude the proof of the Hadamard-Perron theorem


(Theorem 5.2). Namely, we show that the functions (p and V* in (5.9) are of
class in a neighborhood of xo-

T h eorem 5.5. Let / : R R be a function of class and let x q be a


hyperbolic critical point of the equation x' = f{x). Then the functions p
and rp in Theorem 5.4 a,re of class in a neighborhood of x q .

P ro o f. As in the proof of Theorem 5.4, we only consider P n B, since


the argument for P n B is entirely analogous. Moreover, without loss of
generality, we assume that x q = 0.
Step 1. Construction of an auxiliary transformation. Let L{E^,E'^) be the
set of all linear transformations from E^ to Also, let Z be the set of all
continuous functions ^ L{E^,E^) such that $(0) = 0 and
sup {||$(z)|| : 2; G < 1.
Using Proposition 1.30, one can easily verify that equipped with the distance
d ($ ,^ ) = sup{||$(2) - *'(2)11 : z e E ^ }
the set Z is a complete metric space. Given p E Y and z E E^ (see the
proof of Theorem 5.4 for the definition of the space Y), let x = be the
solution of equation (5.15) with x^(0) = 2 . For simplicity of the exposition,
we introduce the notations
y^{T,z) = {x^{t ),(p{x:p{t )))
and
G(r, 2, 5, $ ) = ^)) + (5.28)

where (x,y) E E^ x Ef^. Now we define a linear transformation A((^, $ ) ( 2)


for each (y?, $ ) T x Z and 2 G ?* by
/ +00
A{<p, $ ) ( 2) = - / Pue-^^G{r, 2, p, $ )IP ( t ) dr, (5.29)
Jo
where
W = ^L{E^,E^)
is the matrix function of class determined by the identity

W{t) = Pse^* + f Pse^^^-^^Gir, 2, cp, $)TP(r) dr. (5.30)


Jo
158 5. Existence o f Invariant Manifolds

Writing
C{t,z) = A + P sG{ t, z ,(p ,^), (5.31)
for each x E the function u{t) = W{t)x is the solution of the linear
equation u' = C{t, z)u with initial condition u(0) = x. Since the function
in (5.13) is Lipschitz (see (5.14)), it follows from Theorem 1.38 that the maps
{t,z) Xipit) and are continuous. Thus,

(t, z) G{t, z, (f, $ ) and (t, z) i-> C{t, z)

are also continuous, since they are compositions of continuous functions


(see (5.28)). Hence, the solution u has maximal interval E and the function
W is well defined.
Now we show that the transformation A is well defined. It follows from
(4.12) and (4.30) that
f*+oo
f+O O
e ^^ G(r,2:, $)W '( t )||dr
Jo (5.32)
f+O O
<2cS e ^^||W(r)||dr.
Jo
On the other hand, by (5.30), we have

||W(t)|| < ce-^ + 2c5 dr.


Jo
Multiplying by it follows from Gronwalls lemma (Proposition 1.39) that

||W(t)|| < (5.33)

for t > 0, and by (5.32) we conclude that


/'+00 ^2;:
B < 2c^5 / < (5.34)
Jo jj, c6
This shows that the transformation A is well defined and
c2(5
P ( ,^ )(^ )ll< fi cS < 1

for any sufficiently small 6. Since dog = dog = 0, ?(0) = 0 and = 0 for
z = 0, it follows from (5.29) that A{ifp, $ ) ( 0 ) = 0.
We also show that the function z A{tp,^){z) is continuous for each
(y?, ^) e Y X Z. Given z,z E E^, let x^ and x,p be the solutions of equa
tion (5.15), respectively, with x^(0) = z and x^{0) = z. We also consider
the matrix functions

Wz = Wz,^,^ and Wz = Wz,<p,i>.


5.4. Regularity o f the invariant manifolds 159

Given e > 0, it follows from (5.32) and (5.34) that there exists T > 0 such
that
f*+oo
p-j-oo
/ \\Pue-^-^G{r,z,ip,^)W,{T)\\dr<e
Jt
and
p+oo
/ ||Pe-^^G (r, z, ip, ^ ) W , i r ) \ \ dr < e.
Jt
Thus,
\\ A {ip ,^){z) - A{ip, $ )(z )
(5.35)
< c r e-^"||G(r, z, ip, ^ ) W ,{r) - G{r, z, ip, $)W^,(r)l| d r + 2e.
Jo
Now we show that the function

(r, z) !-> G { t , z , ip, ^)W z(t ) (5.36)

is continuous. In view of the former discussion, it remains to show that the


function z i-> Wz{t ) is continuous. To that effect, we consider the linear
equation u' = C{t, z)u, with the matrix C{t, z) as in (5.31). Since

W z (t) = Id E ^ + r C(s, z ) W z ( s ) ds
Jo
and

W z{t) = Ids* + r (7(5, z ) W z { s ) ds,


Jo
we have

W z{t) - W z{t) = f \ c { s , z) - C ( s , z)]W z{s) d s


Jo (5.37)
+ / C{s,z){Wz{s)-Wz{s))ds.
Jo
Take A, r > 0. Since {t, z) i-> C{t, z) is uniformly continuous on the compact
set [0, A] X 5 (0 , r), given e > 0, there exists S (0, r) such that

\\C{t,z) - C'(t,z)|| < e

for t G [0, A] and z, z G 5(0 , r) with (|z z|| < 5. Moreover,

||<7(t,z)|| < A, where A = ||A|| + 2(5||5s||.

Hence, it follows from (5.31), (5.33) and (5.37) that

\\W z{t) - W-z{ t)\\ < + \ f l|IT,(s) - W^,(s)|| d s


p - 2cd Jo
160 5. Existence o f Invariant Manifolds

for t [0, A] and z,z e B(0, r) with H^: z|| < S. By Gronwalls lemma, we
finally obtain
ce ce o A A
H - 2c6 n - 2c5
for t [0, A] and z,z B{0,r) with ||z z\\ < 6. This implies that the
function (t,z) i-)- Wz(t) is continuous in each set [0, A] x 5(0 , r) and thus
also in R j x 5 (we already know that the function is of class in t, since it
is a solution of the equation u' = C{t, z)u). Therefore, the function in (5.36)
is also continuous, since it is a composition of continuous functions. Hence,
it is uniformly continuous on the compact set [0, T] x 5(0 , r) and there exists
S (0, r) such that
||G(r, z, ^)W,{r) - G{ t, z , $)W 5(r)|| < e

for t [0, T] and z,z & 5(0 , r) with \\z z|| < 5. It follows from (5.35) that
ce
$)(^) - A{if, $)(2;)|| < + 2e

for z,z E 5 (0 , r) with H^; z\\ < 6, and the function z i->- A{(p,^){z) is
continuous (because it is continuous in each ball 5 ( 0 ,r)). This shows that
A{Y x Z ) c Z .
Step 2. Construction of a fiber contraction. Now we show that the trans
formation S: Y X Z ^Y X Z defined by

is a fiber contraction. Given (p Y, e Z and z E let


and
be the matrix functions determined by identity (5.30). We have
\\A{p, $ )(z ) -yl(^,')(z)||
r+oo
<c e~'^^G{T,z,cp,^)W^{T) - G{T,z,<p,^)W^{T)\\dr
p-\-oo
<c e-^"5(||W$(r)-W^(r)||
Jo
+ ||$(x^(r))W$(r) - '^(x^(r))W^(r)||) dr
r-\-oo (5.38)
<c / e-^"<5(||W*(r)-W^(T)||
Jo
+ ||$(x^(r))||-||W$(r)-W^(r)||
+ ||$(x^(r))-^(x^(r))||.|lW^(r)||)dr
r+ oo
< c e-^^<5(2||W$(r) - W^(r)l| -k d($, ^)||Wq/(r)||) dr.
5.4. Regularity o f the invariant manifolds

Now we estimate the norm of

w(t) = W^(t) -
(5 33) obtain
In an analogous manner to that in (5.38), using

ft
|lto(t)||<c / e-f^^*-'^hS\\w{T)\\dT + cS d i^ ,'^ ) ^
rt , /* _;.(t-r)g-(A-2<='^)^dr
<2cS / e-^^^^-^'>\\w{T)\\dr + c^Sdi^,'^) ^
Jo i>t
< 2 c ( J e - ^ r e '^ ^ | | t ( r ) | | d r + c^5d(^,
Jo
< 2 c (5 e -^ * J e ' ^ ^ | | u ;( r ) | | dr + -Cd ( $ , -r\ f-/X+2c<^)^

which yields the inequality

{t)\\<2c5 j e'" ||ry(r)||dr+


Jo
For the function
2c6t
a{t) = ^d{^,'H)e

we have

a'{t) + 2c(5e^*||w(t)|| ^ <a^(t) _j_ 2cSef^^\\w{t)\\


a{t) + 2c(5/o e/*^||u)(r)|| dr " ^(t) a(t) + 2c5/g eM^||u;(T)|| dr
a {t)
< + 2c(5,
a(t)

and integrating on both sides yields the inequality

log (o:(t) + 2cS J e^'^||r(;(T)|| d r ) - logo;(0) < logo:(t) - loga(O) + 2c6t.

Taking exponentials we obtain

< ct(t) + 2c6 /*e^n|u;(r)|| dr < a{t)e^^^\


Jo

or equivalently,

\\W^{t) - W^{t)\\ < |d($, (5.39)


162 5. Existence o f Invariant Manifolds

By (5.33) and (5.39), it follows from (5.38) that


r+oo
\\A{ip, $ )(z ) - A{ip, 'J ') ( z ) || < c^5d{^, ^ ) / dr
Jo
r+oo
g-(2M-2ci)r
+ c^-5d{^, / (
Jo
r+oo
= 2cHd{^,'if) / e-2(M-2c5)r^,
Jo
c^5
H - 2c5
This shows that for any sufficiently small 5 the transformation 5 is a fiber
contraction.

Step 3. Continuity of the transformation S. Given ip,'ip E Y, ^ E Z and


2 E^, let
and

be the matrix functions determined by identity (5.30). We have

\\A{y>, $)(2) ->1(V',$)(^)||


r +00
<c e
Jo
r +00 -^T
<c e \\W^ir)\\dr
Jo
r+oo
+ c \\W^ir)-W^{T)\\dT
i
r+oo
+ c ll^(^('r))|| W^ij)\\dT
i
r+oo
+ c ||$(%(r)) - $(x^(r))|| \\W^{t )\\dr
(r+oo
+ c^ ||$(a;^(r))|| \\W^{t ) - W^{t )\\dr,
1
and thus,

\\A{^p, ^){z) - A{tp,^){z)\\


r+oo
<c2 / ,^{2/jb2cS)r dr
Jo
r+oo
+ C2 / ,^-(2fjL-2c6)r dr
Jo
5.4. Regularity o f the invariant manifolds 163

+ 2cS / e-^^||W^(r) - W^r)]] dr


Jo
/*+oo
+ c^S / - ^(a;v-(T))|| dr,
Jo
using also (5.33). Take e > 0. It follows from (4.30) that there exists T > 0
such that
+00
,2 / ^{2fjb2c6)T dr

,2 / +00 ^ {2 i i 2 c 6 )t
dr
i :
r-\-oo
r+oo
+ 2c5 e~>^^\\W^{r) - W^{t )\\dr
/*+oo
+ (^6 e-(2/^-2-^)"||$(x^(r)) - $(x^(r))|| dr
r+ o o
<10c^6J e-(2#^-2=^)^dr<e

for every z E E^. Hence,

\\A{>p, ^){z) - A{'ip,^){z)


/^ ^ - { 2 / jL - 2 cS )t
< c^ / | (v .(n .))-g M ^ .)) dr
Jo
+ c2 / e-(2M-2c5)r ^ ( y v ( r , z ) ) - ^ { y t { r , z ) ) dr
I

+ c + \\W^{T)-W^(r)\\dr
r - ( If(w (n ^ ))
r
+ c V {2fjL2c5)r ll^(v W ) - W )ll dr + e.
Jo
(5.40)

Now we recall that the function g vanishes outside a closed ball 5(0 , r)
(see (4.29)). Hence, in (5.40) it is sufficient to take r G [0,T] and z G 5
such that

On the other hand, again because 5 = 0 outside 5 (0 , r), for any sufficiently
large 5 > 0 the solutions of the equation x' = Ax + h^{x) (see (5.13)) with
initial condition a;(0) G 5 such that ||x(0)|| > 5 , satisfy

{x{t),(p{x{t)))\\ > r
164 5. Existence o f Invariant Manifolds

for every (p G Y and t 6 [0, T] (because then they are also solutions of the
equation x' = Ax). Hence,

\\y<pij,z)\\, \\yi>(T,z)\\ > r

for every r [0, T] and z E with ||^:|| > R. This implies that for r
as in the proof of Theorem 4.7 all integrals in (5.40) vanish for ||z|| > R.
Therefore,

cT
< c sup ^-{2fi2c5)r dr
zB( , ) J/o
o r Jo

+ c sup ^-{2ii-2 c5)t dr


f
zeB{0,R) Jo (5.41)

+ 2c(5 sup f e ^'^\\W^{t ) - W.^{T)\\dT


zeB(o,R) J o

+ c^6 sup f - $(x.^(r))|| dr + e.


zE
z .B(0^R
B( o,r )
Jo

Since the functions dg and $ are uniformly continuous on the closed ball
5 ( 0 , 2cR), there exists 77 > 0 such that

dg, X 9g
+ < s (5.42)

and

||$(to) $(ty)|| < (5.43)

for w,w e B{0,2cR) with ||rw - U7|| < 77. Moreover, it follows from (5.12)
and (5.20) that

\MT,z)-y^{T,z)\\ < 2\\x ^{ t ) - x ^{ t )\\

< cd{(p,i})\\z\\e^~^'^^'^'>'^ < cRd((p,ip),

for every r > 0 and z G 5 , and any sufficiently small S such that /j,-4c5 > 0.
Now we estimate Wi^(t) W)/,(t). It follows from (5.30) that

< c [ e"'"(* '^)||(7(T,z,<73,$)W<^(r)-G(r,z,7/),$)W^(r)||dr


Jo
5.4. Regularity o f the invariant manifolds 165

< I e -(2 M -2 c 5 )r
^ {y A r,z))-^ M r.z)) dr
7
+ c2 / e-(2M-2ci)r dr
/
+ 2c5 f e->^^\\W^{T) - W ^ { t )\\ dT
Jo
+ cH r dr.
Jo
By (5.16), we have
k(p('r)|| < c||^:|| < cR,
and thus,
||yv ('^>^)ll ^ 2 c||2 || < 2cR
for (^ y , r > 0 and G B{0,R). Hence, it follows from (5.42), (5.43)
and (5.44) that

\\W^{t) - W^{t)\\ < C^e f e-(2M-2ci)r g-(2,i-2ci)r


Jo Jo
+ 2c5 [ e~^'^\\W^p{r) - W^{r)\\ dr
Jo

for t > 0, whenever d{(p^'ip) < r]/{cR). By Gronwalls lemma, we obtain

for t > 0, and it follows from (5.41), (5.42) and (5.43) that

(?5e
2/i 2c5

2ix-2c5\ 2iJi-4c6j
whenever d{<p,ip) < rj/{cR). This shows that the transformation A is con
tinuous, and the fiber contraction S is also continuous (we already know
that T is a contraction; see (5.27)).

Step 4- Smoothness of the function ip. We first show that if 0 G F is of


class and T is the transformation in (5.26), then T (^ ) is of class and

d,Tii^) = A{'iP,di^){z). (5.45)


166 5. Existence o f Invariant Manifolds

To that effect, we note that if ip is of class , then the function h^: E^


defined by
h^{x) = Psg{x,ip{x))
(see (5.13)) is also of class C^. Thus, it follows from Theorem 1.42 applied
to the equation
x' = Ax + h^{x)
that the function u defined by u{t,z) = x-^{t) is of class C^. Moreover, for
$(z) = dzip it follows from (5.30) that W{t) = du/dz. Indeed, substituting
this function W in the right-hand side of (5.30) (with tp replaced by ip) we
obtain

F.e-*+^ (If z))<i.,^g) ir


= Ps6^* + Pse^(*~'^)^5 (y^(r, z)) dr

^ z ^ z)) dr^
dx^(r) du
dz
using also the Variation of parameters formula in Theorem 2.26. It follows
from the uniqueness of solutions of a differential equation that W(t) =
dujdz. Moreover, applying Leibnizs rule in (5.30), we obtain
P+OO
A{ip, dip){z) = - / dz [Pe"^'5(y^(r, z))] dr = dzT{ip).
Jo
Since t->- A{ip,dip){z) is continuous, we conclude that T{ip) is of class C^.
Now we consider the pair (i^Ot^o) = ( 0 ) 0 ) ^ ^ ^ Clearly, $o('2^) =
dz(fo- We define recursively a sequence ((^n> ^*n) by

(Vn-l-l) ^ n -l-l) <S'(<^n) ^ n ) = A {ip n ,^ n ))-

Assuming that pn is of class and ^n{z) dzPni it follows from (5.45)


that the function Pn+i is of class and

dzPn+l dzT(^Pfi) A[P ji, $7i) (5.46)

Moreover, if p and $ are, respectively, the fixed points of T and i->-


A{p,^)., then, by the Fiber contraction theorem (Theorem 1.38), the se
quences pn and converge uniformly respectively to p and $ in each
compact subset of E^. It follows from (5.46) and Proposition 1.41 that p is
differentiable and dzP = $(.z). Since $ is continuous, this shows that p is
of class .
5.5. Exercises 167

5.5. Exercises

E xercise 5.1. Sketch the proof of Theorem 5.2 for k = 2, that is, show that
if the function / is of class C^, then there exists a neighborhood B of xq
such that y n 5 is the graph of a function of class

E xercise 5.2. Consider a diflFeomorphism / : of class such that


/(O) = 0 and dof = ( o &) i with 0 < o < 1 < 6. Given 5 > 0, let
V = { x e B(0, S) : r { x ) e B(0,5) for n N}.
Show that there exist 5 and A with a < A < 1 such that if (x, y) V, then
||/"(a;,y)|| < A'"||(x,y)||, n e N.
Hint: Write / in the form
f{x, y) = (ox + g{x, y), 6x + h{x, y))
and look for V as a graph
V = {(x,v?(x)) : X G (-<5, (5)}
of some Lipschitz function cp: (-S, 6) with (/?(0) = 0.
Part 3

Equations in the Plane


Chapter 6

Index Theory

In this chapter we introduce the notion of the index of a closed path with
respect to a vector field in the plane. On purpose, we do not always present
the most general results, since otherwise we would need techniques that
fall outside the scope of the book. In particular, we discuss how the index
varies with perturbations of the path and of the vector field the index does
not change provided that the perturbations avoid the zeros of the vector
field. As an application of the notion of index, we establish the existence
of a critical point in the interior of any periodic orbit in the plane (in the
sense of Jordans curve theorem). For additional topics we refer the reader
to [4, 9, 19].

6.1. Index for vector fields in the plane

In this section we introduce the notion of the index of a closed path with
respect to a vector field in the plane, and we discus how it varies with
perturbations of the path and of the vector field.

6.1.1. Basic notions. We first introduce the notion of regular path.

D efinition 6 .1 . A continuous function 7 : [0,1] is called a regular


path if there exists a function a : (a, b) R^ of class in some interval
(a,b) D [0,1] such that a{t) = 7 (t) and a'{t) 7^ 0 for t [0,1]. The image
7 ([0, 1]) of a regular path is called a curve.

For simplicity of the exposition, we also introduce the following notion.

D efinition 6 .2 . A regular path 7 : [0,1] > R^ is called a closed path if


7 (0) = 7 ( 1) and the restriction 7 |(0 , 1) is injective (see Figure 6.1).

171
172 6. Index Theory

We note that the image 7 ([0, 1]) of a closed path has no intersections
other than the initial and final points 7 (0) = 7 (1).
Consider a vector field / : of class and write / = ( / i , / 2))
where / i , / 2 : > E are the components of / . We recall that the line
integral of / along a regular path 7 is defined by

I l o dt,

where j{t) = (7iW ,72(0)-

D efinition 6.3. Given a closed path 7 such that 7 ([0, 1]) contains no zeros
of / , the number
1 / ' / 1V / 2 - / 2V /1
/? + / !
is called the index of 7 (with respect to / ) .

Now we give a geometric description of the notion of index. To that


effect, we note that in a neighborhood of each point x 6 7 ([0, 1]), the angle
between / (a;) and the positive part of the horizontal axis is given by

arctan(/2(a:)//i(a;)), fi{x) > 0,


7t/2, fi{x) = 0 and f 2(x) > 0,
0{x) = <
arctan(/2( x ) // i (x)) + tt, / i (x) < 0,
61
( . )

-"/2, /i(x ) = 0 a n d /2(x) < 0,

where arctan denotes the inverse of the tangent with values in (-7r /2, 7r /2).
Now we define

( 6 . 2)
6.1. Index for vector Gelds in the plane 173

We note that although the function 6 is only defined locally and up to an


integer multiple of 2tt, the gradient VO is defined globally.
P rop osition 6.4. I/j: [0,1] is a closed path such that 7 ([0, 1]) con
tains no zeros of f , then Ind/ 7 = N { f , j ) .

P ro o f. For the function 0 in (6.1), we have


/lV/2 - /2V/1
V9 =
/f + M
and thus,

ve

= ^ / v % ( ) ) '/(()< (6.3)

0{j{t)) dt.
27t Jo dt
This yields the desired result.
It follows from the definition of N{f, 7 ) in (6.2) and Proposition 6.4 that
the index is always an integer.
Exam ple 6.5. Consider the phase portraits in Figure 6.2. One can easily
verify that
Ind/ 7 i - 0, Ind/ 72 = 1
in the first phase portrait, and that
Ind/ 7 3 = 0, Ind/ 7 4 = 1
in the second phase portrait.

We note that if a closed path is traversed in the opposite direction, then


the index changes sign. More precisely, given a closed path 7 ; [0, 1] ^
we consider the closed path 7 : [0, 1] > defined by

( - 7 ) ( 0 = 7 (1 - t)-
Since 7 and 7 have the same image, if the curve 7 ([0, 1]) contains no zeros
of / , then In d /(7 ) is also well defined and
In d /(7 ) = In d /7 .

Now we verify that in a sufficiently small neighborhood of a noncritical


point the index is always zero.
P rop osition 6 .6 . Let be a function of class C^. If xq e R"^ is
such that f{xo) ^ 0, then Ind/ 7 = 0 for any closed path 7 whose image is
contained in a sufficiently small neighborhood ofxQ.
174 6. Index Theory

F ig u re 6 .2 . Phase portraits for Example 6.5.

P roof. We first note that by continuity the function / does not take the
value zero in a sufficiently small neighborhood of xq. Thus, for any closed
path 7 in this neighborhood, the image 7 ([0, 1]) contains no zeros of / and
the index Ind/ 7 is well defined. Moreover, also by the continuity of / , in
a sufficiently small neighborhood of xq the function 9 in (6.1) only takes
values in some interval [a, 6] of length less than 2ir, and thus, it follows from
the first integral in (6.2) that iV (/,7 ) = 0. By Proposition 6.4, we conclude
that Ind/ 7 = 0.

6 .1 .2 . P erturbations o f the path and o f the vector field. In this


section we study how the index varies with perturbations of the path 7 and
of the vector field / .
We first recall the notion of homotopy.

D efinition 6.7. Two closed paths 70, 7 1 : [0,1] > are said to be homo
topic if there exists a continuous function H : [0, 1] x [0, 1] such that
(see Figure 6.3):
a) H{t, 0) = 7 o(t) and H{t, 1) = 7 i(t) for every t G [0,1];
b) i?(0, s) = H{l,s) for every s G [0,1].
We then say that is a homotopy between the paths 70 and 7 1 .

The following result describes how the index varies with homotopies of
the path.

P rop osition 6 .8 . The index of a closed path with respect to a vector field
of class does not change under homotopies between closed
paths whose images contain no zeros of f .
6.1. Index for vector fields in the plane 175

P ro o f. Let H : [0,1] x [0,1] ^ be a homotopy between two closed paths


70 and 71 such that H{[0, 1] x [0, 1]) contains no zeros of / . Since the function
{t, s) !-> f(7s(t)) is uniformly continuous on the compact set [0, 1] x [0, 1]
(because continuous functions are uniformly continuous on compact sets),
given e > 0 and s [0, 1], there exists a neighborhood Is of s in [0 , 1] such
that
llf(7r(t))-f(7sm i<e
for every t [0,1] and r e Is- This implies that iV (/,7 r) = In d /7 s for
r /s (because the index is an integer). Hence,

In d /7 r = In d /7 s for r /.

Since the interval [0,1] is compact, there exists a finite subcover ,...,
of [0,1], with Si < S2 < < Sn. Therefore,

Ind/ 70 = Indf js! = Ind/ 7^2 = = Ind/ 7 s = Ind/ 7 1 ,

which yields the desired result.


Now we consider perturbations of the vector field / .

P rop osition 6.9. Let F : E^ x [0,1] ^ E^ 6e a continuous function such


that X !-> F{x, s) is of class for each s [0, Ij. If-y is a closed path such
that F{x, s) for every x E 7 ([0, Ij) and s E [0,1], then Ind^o 7 = Indi?i 7 ,
where Fs{x) = F{x, s).

P ro o f. We proceed in a similar manner to that in the proof of Proposi


tion 6.8. Since the function (t,s) 1-^ F('y(t),s) is uniformly continuous on
[0,1] X [0,1], given > 0 and s E [0,1], there exists a neighborhood Jg of s
in [0, 1] such that
||F ( 7 ( t ) , r ) - F ( 7 ( t ) , ) ||< e
176 6. Index Theory

for every t G [0,1] and r G Js- This implies that Indj?^ 7 = Indp^ 7 for
r E Js- Since the interval [0,1] is compact, there exists a finite subcover
Jsi,---, Jsm. of [0) 1]) with < S2 < < Sm- Therefore,
Indfo 7 = IndF.i '7 = Ind^^^ 7 = = IndF, 7 = M fi 7,
which yields the desired result.

6.2. Applications o f the notion o f index

We give several applications of the notion of index.

6.2.1. P eriod ic orbits and critical points. In this section we show that
any periodic orbit has a critical point in its interior (in the sense of Jordans
curve theorem), as an application of Proposition 6.8. We first recall the
notion of a connected set.

D efinition 6.10. A set t/ C is said to be disconnected if it can be


written in the form U = A\J B ioi some nonempty sets A and B such that
A n B = A n B = 0.
A set 17 C is said to be connected if it is not disconnected.

We also introduce the notion of a connected component.

D efinition 6.11. Given U C M^, a set A C 17 is called a connected compo


nent of U if any connected set B C U containing A is equal to A.

Now we recall Jordans curve theorem (for a proof see, for example, [10]).

P rop osition 6 .1 2 (Jordans curve theorem). I f j : [0,1] zs o continu


ous function mth'y{0) = 7 (1) such that-yKO, 1) is injective, thenR^\j{[0, Ij)
has two connected components, one bounded and one unbounded.

The bounded connected component in Proposition 6.12 is called the


interior of the curve 7 ([0, Ij). One can now formulate the following result.
P rop osition 6.13. Let f : be a function of class and let
7 : [0, 1] > 6e a closed path. jljfInd/ 7 7^ 0, then the interior of the
curve 7 ([0, 1]) contains at least one critical point of the equation x' = f(x).

P roof. Let us assume that the interior U of the curve 7 ([0, 1]) contains no
critical points. Since 6 is of class and V0 = {dd/dXydOjdy), it follows
from Greens theorem that

0 ^ In d ,. = V . = [ | ( D - 1 ( g ) ] = 0.

This contradiction shows that U contains at least one critical point.


6.2. Applications o f the notion o f index 177

Proposition 6.13 has the following consequence.

P rop osition 6.14. If f : is a function of class C^, then in the


interior of each periodic orbit of the equation x' = f {x ) there exists at least
one critical point.

P ro o f. Let 7 be a closed path whose image 7 ([0, 1]) is a periodic orbit.


We first show that Ind/ 7 = 1 . We note that Ind/ 7 = In d j7 , where g
is defined in a neighborhood of 7 ([0, 1]) by g{x) = /(a;)/||/(a;)|| (clearly,
/ does not take the value zero on the periodic orbit). This follows readily
from Proposition 6.4, because the function 0 in (6.1) takes the same values
for / and g. Moreover, without loss of generality, we assume that the curve
7 ([0, 1]) is contained in the upper half-plane and that the horizontal axis
is tangent to 7 ([0, 1]) at the point 7 (0), with 7 ^(0) pointing in the positive
direction of the axis (if the curve 7 is traversed in the opposite direction, then
one can consider the curve 7 and use the identity Ind/ 7 In d /(7 )).
Now we define a function u : A in the triangle

A = {(t ,s ) G [0,1] X [0,1] : f < s}

by

[ - 5 (7 (0)), (f,s) = (0,1),


v{t,s) = ^ gintit)), t = s,
[(7 (s )-7 (t))/| l7 (s )-7 (t)| | , t < s and (t, s) ^ (0, 1).

One can easily verify that the function v is continuous and that it does not
take the value zero. Let a{t, s) be the angle between v{t, s) and the positive
part of the horizontal axis. Clearly, o:(0,0) = 0, because 7 ^(0) is horizontal
and points in the positive direction. Moreover, since 7 ([0, 1]) is contained
in the upper half-plane, the function [0, 1] 9 s a (0, s) varies from 0 to it.
Analogously, the function [0 ,1] 9 1 1-> a{t, 1) varies from tt to 2tt.
On the other hand, since v does not take the value zero in A, it follows
from Proposition 6.13 that Indi,5A = 0. This shows that the function
[0,1] 9 1 1-9 a{t, t) varies from 0 to 2n. Now we observe that aft, t) coincides
with the angle 5(7 (f)) between gi'yft)) and the positive part of the horizontal
axis, by the definition of v. Therefore,

and it follows from Proposition 6.4 that Ind/ 7 = 1-


Applying Proposition 6.13, we conclude that there exists at least one
critical point in the interior of the periodic orbit.
178 6. Index Theory

The following is a generalization of Proposition 6.13 to vector fields of


class for closed paths whose image is a circle.

P rop osition 6.15. Let f : E ? ^ be a function of class and let


7 : [0,1] M? be a closed path whose image is a circle. If Ind/ 7 ^ 0,
then the interior of the curve 7 ([0, 1]) contains at least one critical point of
the equation x' = f{x).

P ro o f. We assume that the interior of the curve 7 ([0, 1]) contains no critical
points. Now let a be a closed path in the interior of 7 ([0, 1]) traversed in
the same direction as 7 . Then the function H : [0,1] x [0,1] -> R^ defined by
H{s,t) sy^t) + (1 - s)a{t)
is a homotopy between a and 7 . Moreover, i? ([0 ,1] x [0,1]) contains no
critical points. It follows from Proposition 6.8 that
Indy a = Indy 7 7^ 0.
On the other hand, if the diameter of the set a ([0 ,1]) is sufficiently small,
then it follows from Proposition 6.6 that Indy a = 0. This contradiction
shows that there exists at least one critical point in the interior of the
curve 7 ([0, 1]).

6.2.2. B rou w ers fixed point theorem . Here and in the following sec
tion we provide two applications of Proposition 6.9. The first is a proof of a
particular case of Brouwers fixed point theorem. Namely, we only consider
functions of class in the plane.

P rop osition 6.16. If f


. is a function of class and B is
a closed ball such that f { B ) c B, then f has at least one fixed point in B.

P roof. Eventually making a change of variables, one can always assume


that
B = {{x,y) :x^ +y^ < l } .
Now we consider the transformation 5 : R^ R^ defined by g{x) = x f{x).
We want to show that g has zeros in B. If there are zeros on the boundary
of B, then there is nothing to prove. Thus, we assume that there are no zeros
on the boundary of B, and we consider the function F : R^ x [0,1] > R^
defined by
F{x,t) = tf{x) X.
We note that F{x, 1) = g{x) 7^ 0 for x E dB, by hypothesis. Moreover, for
t [0, 1) we have ||t/(x)|| < 1 , and thus, one cannot have tf{x) = x when
||x|| = 1. This shows that F{x,t) 7^ 0 for every x dB and t E [0,1]. Now
we consider the closed path 7 : [0, 1] R^ defined by
7(t) = (cos(27rt),sin(27rt)).
6.3. Index o f an isolated critical point 179

that traverses the boundary of B in the positive direction. Then the condi
tions in Proposition 6.9 are satisfied, and we conclude that
Indp 7 = Ind^i 7 = Indigo 7 = India 7 = 1 .

Hence, it follows from Proposition 6.15 that there exists at least one zero
of g in the interior of the ball B.

6.2.3. Fundam ental th eorem o f algebra. Now we give a proof of the


Fundamental theorem of algebra, again as an application of Proposition 6.9.

P rop osition 6.17 (Fundamental theorem of algebra). Given o i , ... ,On


C, the equation 2 -t- -|- + a = 0 has at least one root in C.

P ro o f. Identifying C with E^, we define a function F : C x [0, 1] >C by


F{z, t) = z'^ + t{aiz^~^ -I------- h o).
Moreover, given r > 0, we consider the closed path 7 : [0,1] -> C defined by
7 (t) = Now we assume that
r > 1 -|- |oi| |On|.
For z 7 ([0 ,1]) and t [0,1], we have
|z"| ^n1
> |aiK-^ + |a2K-i-F + |an|r ^
n2
> \ai\r^-^ + \a2K-'^ + + |n|

> | a iz " ^ -I- + 0,n\


> t|aiz ^ + + On|)
since r > 1. This shows that F (z, t) ^ 0 for z 7 ([0, 1]) and t [0,1]. Thus,
letting Ft{z) = F{z, t) for each t G [0,1], it follows from Proposition 6.9 that
Indi?j 7 = IndFo 7- (6-4)
On the other hand, one can easily verify that IndFo7 = n 7^ 0. It follows
from (6.4) that Ind^i 7 7^ 0. Hence, by Proposition 6.15, there exists at least
one zero of Fi in the interior of 7 ([0, 1]), that is, the polynomial
Fi(z) = z -b a iz " ^ -I------- h a
has at least one root with |z| < r.

6 .3 . I n d e x o f an is o la te d c r itic a l p o in t

In this section we introduce the notion of the index of an isolated critical


point of the equation x' = f{x), for a vector field / ; ^ E^ of class C^.

D efinition 6.18. A critical point xq of the equation x' = f{x) is said to be


isolated if it is the only critical point in some neighborhood of xq.
180 6. Index Theory

Given an isolated critical point xq of the equation x' = f{x), take e


sufficiently small such that the ball B(xo,e) contains no critical points be
sides Xq, and consider the closed path 7 ^: [0, 1] > defined by

7e(t) = X q + (cos(27rt),sin(27Tt)).

By Proposition 6.8, the integer number In d /7 e is the same for any suffi
ciently small e > 0, and one can introduce the following notion.

D efinition 6.19. Given an isolated critical point x q of x' = /( x ) , the index


of Xq (with respect to / ) is the integer number In d /7 e, for any sufficiently
small > 0. We denote it by Ind/ x q .

Exam ple 6.20. Consider the phase portraits in Figure 2.16. The origin is
an isolated critical point in all of them. One can easily verify that the index
is 1 in the case of the saddle point and 1 in the remaining phase portraits.

We show that in order to compute the index of any closed path it is


sufficient to know the index of the isolated critical points. For simplicity of
the proof, we consider only vector fields of class C^.

T h eorem 6.21. Let / : > be a function of class and let j be a


closed path with positive orientation. If the equation x' = f (x ) has finitely
many critical points x i , . . . , X n in the interior of the curve 7 ([0, 1]), then
n
Ind/ 7 = ^ I n d / Xj.
i= l

F ig u re 6.4. Paths 7 and 7 i for i = 1 , . . . , n.

P roof. Let U be the interior of the curve 7 ([0, 1]) and take > 0 sufficiently
small such that B{xi,e) C U for t = 1 , .. ., n. We also consider the closed
paths 7 i : [0, 1] >R^ given by

7 i(t) = Xi -I-(cos(27rt),sin(27rt))
6.4. Exercises 181

for i = 1 , .. ., n (see Figure 6.4). It follows from Greens theorem that

= [ ^ ( 1 ) - l(S)]
where D = U \ I J ^ i B{xi,e). Thus, by (6.3), we obtain

I n d , 7 = = E l 0 / T i -

*^ 7 i= l i= l

Since the interior of each curve 7 i([0, 1]) contains no critical points besides Xi
we have Ind/ 7 ^ = IndfXi for i = 1 , .. ., n. Thus,

Ind/ 7 = ^ Ind/ 7 i = ^ Ind/ Xi,


i=l i=l
which yields the desired result.
6.4. Exercises

Exercise 6 .1 . Consider the equation in polar coordinates

r' = r cos 0,
{
O' = sin0.
a) Sketch the phase portrait.
b) Determine the stability of all critical points.
c) Find the index of all isolated critical points.

Exercise 6.2. For each a,PE'R, consider the equation

ix' = y,
|y' = - x /4 + ay - P{x^ + 4y^)y - {x^ + 4y^)^y.
a) Find the index of the origin when a = 1 .
b) Find whether the equation is conservative when a ^ O .
c) Show that if a = = 0, then the origin is stable.
d) Show that if a > 0, then there exists at least one periodic orbit.

Exercise 6.3. Consider the equation in polar coordinates

ir' = r(l + COS0),


- r(l cos6).

a) Sketch the phase portrait.


b) Determine the stability of all critical points.
c) Find the index of all isolated critical points.
182 6. Index Theory

d) Find whether there are global solutions that are not critical points.

E xercise 6.4. Consider the equation

u = u uv,
{ /
V = u v V.

a) Find the index of the origin.


b) Show that
H{u, v) = u + v log(uu)
is an integral in the quadrant {( , u) : n, u > 0}.
c) Show that there exist infinitely many periodic orbits. Hint; Verify
that H has a minimum at (1,1).
d) Sketch the phase portrait.

E xercise 6.5. Show that if ^ ^ M are bounded continuous functions,


then the equation
ix' = y + f{x,y),
\y' = - x + g{x,y)
has at least one critical point.

E xercise 6 .6 . Let F : M > R be a bounded continuous function.


a) Show that there exists a; R such that F { F{x)) x. Hint: Use
Exercise 6.5.
b) Show that there exists a; R such that sin(l sin^(l x^)) = x.

E xercise 6.7. Find the index of the origin for the vector field
f{x, y) = {2x + y + x^ + xy^, x + y - y ^ + a ;V )-

E xercise 6 .8 . Let / ; R^ -4 R^ be a function of class C^. Show that if the


equation x' = f{x) has a periodic orbit 7 , then the following alternative
holds: either div / = 0 in the interior [/ of 7 (in the sense of Jordans curve
theorem) or d iv / takes different signs in U. Hint: Write / = ( / i , / 2) and
note that by Greens theorem,

/ div / = / ( - / 2, / i ) .
JU J'l

E xercise 6.9. Show that if /,^ : R )R are functions of class then the
equation
\x' = f{y),
\y' = g{x) + y^
has no periodic orbits.
6.4. Exercises 183

E xercise 6 .1 0 . Consider the equation


x" = p{x)x' + q{x).
Use Exercise 6.8 to show that if p < 0, then there are no periodic orbits.

E xercise 6.11. Consider the equation


fx' = y ( l - h x - y ^ ) ,
= x(l + y - x^).
Show that there are no periodic orbits contained in the first quadrant.

Exercise 6 .1 2 . Find whether the equation has periodic solutions:


a) x^^ x^ H1 = Oj
''a ; ' + _ 1 = Q ,

b)
y' + a; + 1 = 0.

Solutions.
6.1 b) The origin is the only critical point and is unstable,
c) 2.
6.2 a) 1 .
b) It is not.
6.3 b) (0,0) is unstable.
c) 1.
d) There are.
6.4 a) -1 .
6.7 - 1 .
6 .1 2 a) It has not.
b) It has not.
Chapter 1

PoincareBendixson
Theory

This chapter is an introduction to the Poincar^Bendixson theory. After


introducing the notion of invariant set, we consider the a-limit and w-limit
sets and we establish some of their basic properties. In particular, we show
that bounded semiorbits give rise to connected compact a-limit and w-limit
sets. We then establish one of the important results of the qualitative the
ory of differential equations in the plane, the Poincare-Bendixson theorem,
which characterizes the a-limit and w-limit sets of bounded semiorbits. In
particular, it allows one to establish a criterion for the existence of periodic
orbits. For additional topics we refer the reader to [9, 13, 15, 17].

7.1. Limit sets

Let the function / : R ^ M be continuous and locally Lipschitz (see Defi


nition 3.15). Then the equation

x' = /(a:) (7.1)

has unique solutions. We denote by v?t(xo) the solution with a;(0) = xo, for
t Ixo> where Ixq is the corresponding maximal interval.

7.1.1. Basic notions. We first introduce the notion of invariant set.

D efinition 7.1. A set A C R " is said to be invariant (with respect to


equation (7.1)) if (^t(x) A for every x G A and t Ix-

185
186 7. Poincare-Bendixson Theory

Exam ple 7.2. Consider the equation

X' = y ,
(7.2)
1y' = -X.
Its phase portrait is the one shown in Figure 1.4. The origin and each circle
centered at the origin are invariant sets. More generally, any union of circles
and any union of circles together with the origin are invariant sets.

We denote the orbit of a point a; M (see Definition 1.50) by

7 (a;) = 7 /(a;) = {(pt{x) : < / * } .

It is also convenient to introduce the following notions.

D efinition 7.3. Given x E M , the set

7 +(aj) = 7 +(x) = {<pt{x) : t G 4 n R+}

is called the positive sem iorbit of x, and the set

7 (x) = ' y j{ x ) = {(pt{x) : t G 4 n R }

is called the negative sem iorbit of x.

One can easily verify that a set is invariant if and only if it is a union of
orbits. In other words, a set >1 C R is invariant if and only if

^ = U l{x)-
xeA

Now we introduce the notions of a-limit and w-limit sets.

D efinition 7.4. Given x E R , the a -lim it and oj-lim it sets of x (with


respect to equation (7.1)) are defined respectively by

a{x) = af{x) = Pi 7-(y)


ye-y^x)
and
w (x ) = U)f{x) = P i 7 + (y ).
y-y(x)

E xam ple 7.5 (continuation of Example 7.2). For equation (7.2), we have

a(x) = co(x) = 7'^(a;) = J~(x) = j(x)

for any x G R^.


7.1. Limit sets 187

E xam ple 7.6. Consider the equation in polar coordinates

{ r' - r (l r), (7.3)


e' = 1 .

Its phase portrait is the one shown in Figure 7.1. Let

5 = {a; G : ||a;|| = l } .

We have
a{x) (jj{x) = 'j'^{x) = 'y~{x) = 7 (a;)
for X G {(0 ,0 )} U S,

a{x) = { ( 0, 0)} and uj{x) = S

for X G with 0 < ||x|| < 1, and finally,

a{x) = 0 and u>{x) S

for X G with Hxl| > 1.

E xam ple 7.7. For the phase portrait in Figure 7.2 we have:

a(xi) = w(xi) = 0 ;

o:(x2) = 0 , w(x2) = {g };

a (x 3) = w(x3) = {p };

Q;(x 4) -- W(X4) = lix A } -


188 7. Poincare-Bendixson Theory

F ig u re 7.2. Phase portrait for Example 7.7.

7.1.2. A d d ition al properties. In this section we establish some proper


ties of the a-limit and w-limit sets for equation (7.1).

P rop osition 7.8. If the positive semiorbit 7 '^(a;) of a point x E MP is


bounded, then:
a) u){x) is compact, connected and nonempty;
h) y E <jj{x) if and only if there exists a sequence tk Z' +oo such that
^tk () y ^hen k oo;
c) ft{y) Lo{x) for every y e u{x) and t > 0;
d) inf{||(^((a;) ?/|| : y w (x)} 0 when t )-foo.

P ro o f. Let K = 7 +(a;). It follows readily from the definition of w-limit set


that u)(x) is closed. On the other hand, w(x) C K and thus u{x) is also
bounded. Hence, the tu-limit set is compact.
Moreover, since the semiorbit y'^{x) is bounded, we have E+ C Ix (by
Theorem 1.46), and thus,

^ip^ )= n
t>o
where
M = Ws{x) : s > t ) .
Identity (7.4) yields the second property in the proposition. Indeed, if y
uj{x), then there exists a sequence tk Z' +oo such that y for N.
Thus, there is also a sequence Sk +oo with sk > tfe for A: N such that
Pski^) y when A; > oo. On the other hand, if there exists a sequence
tk + 00 as in the second property in the proposition, then y 6 for
7.1. Limit sets 189

k N, and hence,
OO

y n = n
k=l OO
because At C At> for t > t'.
Now we consider a sequence {tpk{x))k contained in the compact set K.
By compactness, there exists a subsequence with tk Z' +oo, con
verging to a point of K. This shows that u{x) is nonempty.
Now we show that u>{x) is connected. Otherwise, by Definition 6.10,
we would have uj{x ) = A\J B for some nonempty sets A and B such that
AC\B Ar\B = 0. Since w(x) is closed, we have
A = A n u{x) = Ar\{AU B)
= ( A n A) li ( A n B) = A
and analogously B = B. This shows that the sets A and B are closed, and
hence, they are at a positive distance, that is,
5 := inf {||a 6|| : a A,b E: 5 } > 0.
Now we consider the set

inf I
= { yeco{x)
One can easily verify that (70 i f is compact and nonempty. Hence, it follows
from the second property in the proposition that C n K nuj(x) ^ 0. But
by the definition of the set C we know that C7 O i f does not intersect w(x).
This contradiction shows that w(x) is connected.
In order to verify the third property in the proposition, we recall that, by
the second property, if y G o;(x), then there exists a sequence tk -l-oo such
that y when k oo. By Theorem 1.40, the function y i-> ipt{y) is
continuous for each fixed t. Thus, given f > 0, we have

^tk+t{y) = M>tk(y)) My)


when k OO. Since tk + t -l-oo when A: >oo, it follows from the second
property that (pt{y) w(x).
Finally, we establish the last property in the proposition. Otherwise,
there would exist a sequence tk +oo and a constant ^ > 0 such that
inf Jktfc (a;)-yII > 5 (7.5)

for A: G N. Since the set K is compact, there exists a convergent subsequence


(pt'^{x))k of (PtkiMik C K, which by the second property in the proposition
has a limit p G w(x). On the other hand, it follows from (7.5) that
\\^t'^{x)-y\\ > S
190 7. Poincare-Bendixson Theory

for every y u>{x) and A; N. Thus, ||py|| > <5for y G oj{x), which implies
that p ^ (jj{x). This contradiction yields the desired result.

We have an analogous result for the a-limit set.

P rop osition 7.9. If the negative semiorbit 7 (x) of a point x ts


bounded, then:
a) a{x) is compact, connected and nonempty;
b) y a{x) if and only if there exists a sequence tk \ 00 such that
V'tk(3^) y when k 00;
c) <Pt(y) a{x) for every y G a{x) and t < 0;
d) inf{||^t(a;) y|| : y G a(a;)} 0 when t 00.

P roof. As in the proof of Theorem 5.4, let g: M M be the function


g{x) = f{x). We recall that if (pt{xo) is the solution of the equation
x' = f {x ) with a;(0) = x q , for t in the maximal interval Ixq = (o,,b), then
fpt(xo) <P-t(a:o) is the solution of the equation x' = g{x) with a;(0) = x q ,
for t G (6, a). This implies that

7s(a^) = 'Ifix) and 7 + {x) = 7 ^ (x) (7.6)

fo r e v e r y x G and thus.

n 7 /(y )= n 7/(a:) =Ws()- (7.7)


y e jf(x ) yeygix)

Now we assume that the negative semiorbit 7 ^ (a:) is bounded. Since


7^() = 7 / ()i it follows from Proposition 7.8 that:

a) u)g{x) is compact, connected and nonempty;


b) y G iVg{x) if and only if there exists a sequence tk Z' + 0 0 such that
itki^) y when A; - 7 00;
0) Vt(y) ^ ^gix) for every y G ojg{x) and t > 0;
d) inf {W'tftix) y|| : y L0g(x)} ^ 0 when t >+ 00.

Since t/'t = <P-ti the proposition now follows readily from (7.7) and these
four properties.

7.2. The Poincare-Bendixson theorem

Now we turn to and we establish one of the important results of the qual
itative theory of differential equations: the Poincare-Bendixson theorem.
7.2. The Poincare-Bendixson theorem 191

7 .2 .1 . Intersections w ith transversals. We first establish an auxiliary


result. Let / : )> be a function of class C^. Also, let L be a line
segment transverse to f. This means that for each x e L the directions of L
and f{x) generate R^ (see Figure 7.3). We then say that L is a transversal
to / .

F ig u re 7.3. Orbits in the neighborhood o f a transversal.

P rop osition 7.10. Given x the intersection oj{x) n L contains at


most one point.

P roof. Assume that cj{x ) fl L is nonempty and take q u>{x) n L. By


Proposition 7.8, there exists a sequence tk +oo such that <pt^(x) q when
A: ^ oo. On the other hand, since L is a transversal to / , it follows from the
Flow box theorem (Theorem 1.54) that for each y sufficiently close to L
there exists a unique time s such that <Ps{y) G L and tpt{y) 0 L for t G (0, s)
when s > 0, or for t (s, 0) when s < 0 (see Figure 7.3); in particular, for
each A: G N there exists s Sk ss above such that Xk = ^tk+sk () ^
Now we consider two cases: either (xk)k is a constant sequence, in which
case the orbit of x is periodic, or {xk)k is not a constant sequence. In
the first case, since the orbit of x is periodic, the w-limit set cj{x ) = ^{x)
only intersects L at the constant value of the sequence {xk)ky and thus
u{x) r\ L = {g}. In the second case, let us consider two successive points
of intersection Xk and Xk+i, that can be disposed in L in the two forms in
Figure 7.4. We note that along L the vector field / always points to the
same side (in other words, the projection of / on the perpendicular to L
always has the same direction). Otherwise, since / is continuous, there
would exist at least one point z L with f{z) equal to zero or with the
direction of L, but then L would not be a transversal. We also note that
the segment of orbit between Xk and Xk+i together with the line segment
between these two points form a continuous curve C whose complement
192 7. Poincare-Bendixson Theory

F ig u re 7.4. Intersections with the transversal L.

M?\C has two connected components, as a consequence of Jordans curve


theorem (Proposition 6.12). The bounded connected component is marked
in gray in Figure 7.4. Due to the direction of / on the segment between Xk
and x/c+i (see Figure 7.4), the positive semiorbit is contained in the
unbounded connected component. This implies that the next intersection
x/c+2 does not belong to the line segment between Xk and Xk-^-i- Therefore,
the points x/., Xk+i and XkJ\-2 are ordered on the transversal L as shown in
Figure 7.5. Due to the monotonicity of the sequence {xk)k along L, it has
at most one accumulation point in L and hence uo{x)r\L = {g}.

7.2.2. T h e P oin ca re-B en d ix son theorem . The following is the main


result of this chapter.

T h eorem 7.11 (Poincare-Bendixson). Let / : -> be a function of


class C^. For equation (7.1), if the positive semiorbit 'y~^{x) of a point x is
bounded and lo {x ) contains no critical points, then ui(x) is a periodic orbit.

P ro o f. Since the semiorbit j'^{x) is bounded, it follows from Proposition 7.8


that uj{x) is nonempty. Take a point p e lj{x ). It follows from the first and
third properties in Proposition 7.8, together with the definition of cj-limit
set, that u){p) is nonempty and oj{p) C u{x). Now take a point q G oj{p). By
hypothesis, q is not a critical point, and thus there exists a line segment L
containing q that is transverse to / . Since g G a;(p), by the second property
in Proposition 7.8, there exists a sequence tk +oo such that ^tk{p) Q
when k ^ oo. Proceeding as in the proof of Proposition 7.10, one can always
7.2. The Poincare-Bendixson theorem 193

F ig u re 7.5. Intersections Xk, Xk+i and Xk+2 with the transversal L.

assume that <Ptk(p) L for k N. On the other hand, since p to(x), it


follows from the third property in Proposition 7.8 that <ptk(p) a; (a;) for
A: G N. Since <ptk(p) S u>(x) fl L, by Proposition 7.10 we obtain

Ttfc(p) 9 every k N.

This implies that 7 (p) is a periodic orbit. In particular, j(p) c oj(x).


It remains to show that co(x) = -yip). If uj(x)\j(p) ^ 0 , then since u>(x)
is connected, in each neighborhood of -/(p) there exist points of cu(x) that
are not in -y(p). We note that any neighborhood of 'y(p) that is sufficiently
small contains no critical points. Thus, there exists a transversal L' to f
containing one of these points, which is in u>{x), and a point of 7 (p). That
is, u){x) n L' contains at least two points, because 7 (p) C w(x); but this
contradicts Proposition 7.10. Therefore, w(x) = j{p) and the w-limit set
of X is a periodic orbit.

One can obtain an analogous result to Theorem 7.11 for bounded nega
tive semiorbits.
194 7. Poincare-Bendixson Theory

T h eorem 7.12. Let f : > be a function of class C^. For equa


tion (7.1), if the negative semiorbit 'y~{x) of a point x is bounded and a{x)
contains no critical points, then a{x) is a periodic orbit.

P roof. As in the proof of Proposition 7.9, consider the function 5 :


defined by g{x) = f {x ) and the equation x' = g{x). By (7.6) and (7.7),
we have
7 7 (^) = i t ^
The result is now an immediate consequence of Theorem 7.11.

Exam ple 7.13 (continuation of Example 7.6). We already know that equa
tion (7.3) has a periodic orbit (see Figure 7.1), namely the circle of radius 1
centered at the origin. Now we deduce the existence of a periodic orbit as
an application of the Poincare-Bendixson theorem. Consider the ring

Z? = < X G
1 Ixll < 2
2 < }
For r = 1/2 we have r' - 1/4 > 0, and for r = 2 we have r' = 2 < 0.
This implies that any orbit entering D no longer leaves D (for positive
times). This corresponds to the qualitative behavior shown in Figure 7.6.
In particular, any positive semiorbit 7 "''(x) of a point x G > is contained in D

\ X

/ \

F ig u re 7.6. Behavior in the boundary o f D.

and hence it is bounded. Moreover, it follows from (7.3) that the origin is
the only critical point. By the Poincare-Bendixson theorem (Theorem 7.11),
we conclude that w(x) is a periodic orbit for each x G >.

Exam ple 7.14. Consider the equation

{ x' = x(x^ Sx 1) y,
(7.8)
y' = y{x^ + y"^ - Sx - 1) + X,
7.2. The Poincare-Bendixson theorem 195

which in polar coordinates takes the form


J r ' = r(r^ 3r cos ^ 1),
[O = 1.
For any sufficiently small r, we have
3r cos 6 1 < 0 ,
and thus r' < 0. Moreover, for any sufficiently large r, we have
3r cos ^ 1 > 0,
and thus r' > 0. On the other hand, the origin is the only critical point.
Now we use an analogous argument to that in Example 7.13. Namely, for
ri > 0 sufficiently small and V2 > 0 sufficiently large, there are no critical
points in the ring
>' = {a; : n < ||a;|| < V2}.
Moreover, any negative semiorbit 'y~{x) of a point x D' is contained in D',
and hence it is bounded. It follows from Theorem 7.12 that a{x) C D' is a
periodic orbit for each x e D'. In particular, equation (7.8) has at least one
periodic orbit in D'.

Now we formulate a result generalizing the Poincare-Bendixson theorem


to the case when u>{x) contains critical points.

T h eorem 7.15. Let f : E ? ^ E ? b e a function of class C^. For equa


tion (7.1), if the positive semiorbit 7 *'(a;) of a point x is contained in a
compact set where there are at most finitely many critical points, then one
of the following alternatives holds:
a) (jj{x) is a critical point;
b) co(x) is a periodic orbit;
c) u{x) is a union of a finite number of critical points and homoclinic
or heteroclinic orbits.

P ro o f. Since uj{x) C 7 "' (a;), the set u>{x) contains at most finitely many
critical points. If it only contains critical points, then it is necessarily a
single critical point, since by Proposition 7.8 the set uj{x) is connected.
Now we assume that w(x) contains noncritical points and that it contains
at least one periodic orbit -yip). We show that oj{x) is the periodic orbit.
Otherwise, since w{x) is connected, there would exist a sequence {xk)k C
u){x) \ j{p) and a point xq 7 (p) such that Xk >xq when k oo. Now
we consider a transversal L to the vector field / such that xq G L. It
follows from Proposition 7.10 that u>(x) D L = { xq}. On the other hand,
proceeding as in the proof of Proposition 7.10, we conclude that 'y'^{xk) C
196 7. Poincare-Bendixson Theory

ui(x) intersects L for any sufficiently large k. Since cv(x) n L = {a:o}, this
shows that Xk E j(xo) = jip) for any sufficiently large k, which contradicts
the choice of the sequence (xk)k- Therefore, u(x) is a periodic orbit.
Finally, we assume that u j ( x ) contains noncritical points but no periodic
orbits. We show that for any noncritical point p E ui(x) the sets ui(p) and
a(p) are critical points. We only consider u(p), because the argument for
a(p) is analogous. Let p u>(x) be a noncritical point. We note that
uj(p) C uj(x). If g u){p) is not a critical point and L is transversal to /
containing q, then, by Proposition 7.10,
u>{x) r\L uj(p) r\L { 9};
in particular, the orbit 7 '^(p) intersects L at a point x q - Since y'^{p) C u>{x),
we have xq = q, and thus 7 *'(p) and w(p) have the point q in common.
Proceeding again as in the proof of Proposition 7.10, we conclude that
u{p) = 7 (p) is a periodic orbit. This contradiction shows that u{p) con
tains only critical points and since it is connected it contains a single critical
point.

We recall that by Proposition 7.8 the set oj{x) is connected. Under


the assumptions of Theorem 7.15, this forbids, for example, that ui{x) is a
(finite) union of critical points.
One can also formulate a corresponding result for negative semiorbits.

7.3. E x e rcise s

E xercise 7.1. Consider the matrices


/4 1 0 0 o\ /-I 1 0 0 0 \
0 4 1 0 0 0 -1 0 0 0
A= 0 0 4 0 0 and B = 0 0 2 0 0
0 0 0 1 1 0 0 0 0 -3
^0 0 0 -1 V lo 0 0 3 0 )
a) For the equation x' = Ax show that a{x) = {0 } for every a; K.
b) For the equation x' = Bx show that a solution x is bounded if and
only if x(0) G {0}^ x

Exercise 7.2. By sketching the phase portrait, verify that there exist equa
tions in the plane with at least one disconnected w-limit set.

E xercise 7.3. Consider the equation

x' = x"^ xy,


{ y ' = y2 _ a;2 _ ^

a) Show that the straight line a; = 0 is a union of orbits.


7.3. Exercises 197

b) Find whether there exist other straight lines passing through the
origin and having the same property.

E xercise 7.4. For each e G M, consider the equation in polar coordinates


r ' = r(l r),

{ 9' = sin^ 9 + e.

a) Sketch the phase portrait for each G R.


b) Find all values of e for which the equation is conservative.
c) Find the period of each periodic orbit when e = 1 .
d) Find whether the smallest invariant set containing the open ball of
radius 1/2 centered at ( 1, 0) is an open set when e = 0.

E xercise 7.5. Consider the equation


fx' = x^

a) Show that there is an invariant straight line containing (0, 0).


b) Show that there are no periodic orbits,
c) Sketch the phase portrait.

E xercise 7.6. For the function B{x, y) = x y { l x y), consider the equation

and =
ay ox
a) Find all critical points and verify that the straight lines x = 0 and
y = 0 are invariant.
b) Show that the straight line x + y = 1 is invariant,
c) Find an invariant compact set with infinitely many points,
d) Sketch the phase portrait.

E xercise 7.7. Given a function / : R^ R of class (7^, consider the equa


tion x' = V /(x ). Show that any nonempty w-limit set is a critical point.

Exercise 7.8. Verify that there exists an autonomous equation in R^ with


a periodic orbit but without critical points.

Solutions.
7.3 b) There are none.
7.4 b) The equation is conservative for no values of e.
c) The only periodic orbit is the circle of radius 1 centered at the
origin and its period is l/(sin^ 9 + I)d9 = v^ tt.
d) It is open.
198 7. Poincare-Bendixson Theory

7.6 a) (0,0), (0,1), (1,0) and (1/3,1/3).


c) Triangle determined by (0,0), (0,1) and (1,0).
Part 4

Further Topics
Chapter 8

Bifurcations and
Center Manifolds

This chapter gives a brief introduction to bifurcation theory. We begin with


the description of several examples of bifurcations. In particular, among
others we consider the Hopf bifurcation, which corresponds to the appear
ance (or disappearance) of a periodic orbit. We then give an introduction to
the theory of center manifolds, which often allows one to reduce the order of
an equation in the study of the existence of bifurcations. Center manifolds
are also useful in the study of the stability of a critical point, by reducing
the problem to the study of the stability on the center manifold. Finally, we
give an introduction to the theory of normal forms, which aims to eliminate,
through an appropriate change of variables, all possible terms in the original
equation. For additional topics we refer the reader to [8 , 12, 14].

8.1. Introduction to bifurcation theory

We start with an example that illustrates the type of problems considered


in bifurcation theory.

Exam ple 8.1. Consider the equation

x' = X,

{ y' = {1 + e)y,

with e R. The phase portrait for each value of e is shown in Figure 8.1.
( 8. 1)

We are interested in knowing for which values of eo R there exists e


in an arbitrarily small neighborhood of eo such that the solutions of equa
tion (8.1) for o and e are not differentially conjugate (see Definition 2.44).
202 8. Bifurcations and Center Manifolds

6= 0

-1 < < 0 = 1 < -1

F ig u re 8 .1 . Phase portrait o f equation (8.1) for each e e R.

Prom the phase portraits in Figure 8.1 it is clear that this happens for = 0
and = 1 . Moreover, it follows from Propositions 2.45 and 2.47 that the
same happens for all remaining values of , because the matrix

0
1+

of the linear equation (8.1) has different eigenvalues for different values of e.
Hence, the solutions of equation (8.1) for any two different values of are
not differentially conjugate. Again, this shows that differentiable conjugacies
are somewhat rigid, since they distinguish phase portraits that clearly have
the same qualitative behavior (such as the first three phase portraits in
Figure 8.1).
Now we consider the analogous problem for the notion of topological con-
jugacy. In other words, we want to know for which values of q K there
exists in an arbitrary small neighborhood of q such that the solutions
of equation (8.1) for q and are not topologically conjugate (see Defini
tion 2.44). In this case, it follows readily from Theorem 2.50 that this only
8.1. Introduction to bifurcation theory 203

occurs for e = 1. We then say that a bifurcation occurs in equation (8.1)


at 6 = - 1.

Now we formalize the concept of bifurcation, also allowing topologi


cal conjugacies up to a time change along each orbit. Consider a function
/: X )MP of class and the equation
x' = f{x ,e) , (8.2)
for each value of the parameter e G
D efinition 8 .2 . We say that a bifurcation does not occur in equation (8.2)
at e = o if for each arbitrarily close ei G R^ there exist a homeomorphism
h: R -4- R and a continuous function r : R x R" > R with t r{t,x)
increasing for each a; G R'^ such that
HM x)) ^ (8.3)
for every t G R and x G R , where tpt{z) and tpt{z) are, respectively, the
solutions of the initial value problems
fx ' = /(x ,e o ), b ' = /(x , i),
(8.4)
|x(0) = z t(0) 2;.

In other words, a bifurcation occurs at = q if in any arbitrarily small


neighborhood of sq there exists 1 such that the solutions of the equations
x' = /(x , o ) and x' = /(x , i)
are not transformed into each other by a homeomorphism preserving orien
tation.

Exam ple 8.3. Consider the equation


fx' = (1 -I- ^)y,
(8.5)
|j/' = - ( 1 -l-^)x.
The origin is a critical point and the remaining orbits are circular periodic
orbits centered at the origin and of period 2ir/(1+e^). Indeed, equation (8.3)
can be written in polar coordinates in the form
fr' = 0.
0' = - ( l + e 2)

Since
Pt = V't(l+e)/(l+2),
with (ft and ipt defined by (8.4), taking
h{x) = X and r(t) = t ( l - f q) / ( 1 + e?)
yields identity (8.3). This shows that no bifurcations occur in equation (8.5).
204 8. Bifurcations and Center Manifolds

E xam ple 8.4. Consider the equation


x' = sx x^. (8.6)
Clearly, x 0 and x e are critical points. The phase portrait is the one
shown in Figure 8.2. One can easily verify that the only bifurcation occurs
at s = 0. It is called a transcritical bifurcation and corresponds to the
collision of two critical points, one stable and one unstable, that exchange
their stability after the collision.

< 0

e= 0

> 0

F ig u re 8 .2 . Phase portrait o f equation (8.6).

Exam ple 8.5. Consider the equation


x' = e x^. (8.7)
The number of critical points depends on the sign of e. For < 0 there
are no critical points, for = 0 the origin is the only critical point, and
finally, for > 0 there are two critical points, namely \/ and ^/e. The
phase portrait is the one shown in Figure 8.3. Clearly, the only bifurcation
in equation (8.7) occurs at = 0. It is called a saddle-node bifurcation (see
also Example 8.6) and corresponds to the collision of two critical points, one
stable and one unstable, that are annihilated after the collision.

< 0

= 0

> 0
-x /i

Figure 8.3. Phase portrait of equation (8.7).


8.1. Introduction to bifurcation theory 205

Exam ple 8 .6 . Consider the equation

lx' = x^,
( 8.8)
\y' = y-
We note that the first component coincides with equation (8.7), and thus,
the phase portrait of equation (8.8) is the one shown in Figure 8.4. Again,
the only bifurcation occurs at e = 0. It can be described as the collision of a
saddle point and a node for e > 0, which disappear for e < 0. This justifies
the name of the bifurcation in Example 8.5.

e< 0

F ig u re 8.4. Phase portrait o f equation (8.8).

E xam ple 8.7. Consider the equation

x' = ex x^. (8.9)

For e < 0 the origin is the only critical point, while for e > 0 there are
two critical points, namely and \/e. The phase portrait is the one
shown in Figure 8.5. One can easily verify that the only bifurcation occurs
at = 0. It is called a pitchfork bifurcation and corresponds to the creation
(or annihilation) of two critical points, one stable and one unstable.

< 0

> 0
-V e 0

Figure 8.5. Phase portrait of equation (8.9).


206 8. Bifurcations and Center Manifolds

Exam ple 8 .8 . Consider the equation

{ x' ex y x{x^ + y^), ( 8 . 10)


y' = X + ey - y{x^ + y^),

which in polar coordinates takes the form

(8.11)

We note that the first component in (8.11) was already considered in Ex


ample 8.7 (although now we are only interested in the nonnegative values
of the variable, because r > 0). The phase portrait of equation (8.10), or of
equation (8.11), is the one shown in Figure 8.6. One can easily verify that
the only bifurcation in equation (8.10) occurs at e = 0. It is called a Hopf
bifurcation and corresponds to the creation (or annihilation) of a periodic
orbit.

F ig u re 8 .6 . Phase portrait o f equation (8.10).

There are many other bifurcations, but a systematic study of bifurcation


theory clearly falls outside the scope of the book (for detailed treatments we
refer the reader to [8 , 1 2 ]).

8.2. Center manifolds and applications

In this section we give a brief introduction to the theory of center manifolds,


and we illustrate how it can be of help in bifurcation theory.

8 .2 .1 . B asic notions. Let / : R R be a function of class and


let xq be a critical point of the equation a;' = f(x). Unlike what happens
in Section 4.1, here we do not assume that xq is a hyperbolic critical point.
We continue to write A = dxof-
8.2. Center manifolds and applications 207

D efinition 8.9. We define the stable, unstable, and center spaces of xq,
respectively, by

E- = /aj M \ {0 } : limsup \ log ||e'^ x|| < o l U {0 },


t f-^ + o o t )

E' = |a; e M \ {0 } : limsup log ||e^ a;|| < o l U {0 },


1. t-^ -o o \t\ J

= la; R \ {0 } : limsup log ||e^*x|| = o l U {0}.


t t^oo 1^1 J

In other words, E^ and E'^ contain the initial conditions (other than the
origin) whose solutions have some exponential behavior, while E'^ contains
the initial conditions whose solutions have no exponential behavior. One can
easily verify that in the case of a hyperbolic critical point the sets E^ and
in Definition 8.9 coincide with the stable and unstable spaces introduced in
Definition 4.2.

P rop osition 8 .1 0 . If xq is a critical point of the equation x' = f{x), then:


a) E^, E'^ and E'^ are subspaces o /R and E^ E^ E'^ R ;
b) for every t G R we have
e^\E^) C E\ e^^E^) c E^ and e^\E^) c E^.

P ro o f. One can proceed in a similar manner to that in the proof of Propo


sition 4.3. Namely, the Jordan canonical form of the matrix A dx^f can
be written in the block form
0 0'
0 Ay, 0
0 0 Ac
with respect to the decomposition E^E'^E^. The matrices >ls, Au
and Ac correspond, respectively, to the Jordan blocks of eigenvalues with
negative, positive and zero real part.

When E^^ = {0 }, that is, when the critical point xq is hyperbolic, the
Grobman-Hartman theorem and the Hadamard-Perron theorem (Theo
rems 4.7 and 5.2) describe with sufficient detail the phase portrait of the
equation x' = f{x) in a neighborhood of xq. In particular, the solutions
of the equations x! = f{x) and y' = Ay are topologically conjugate, re
spectively, in neighborhoods of xq and 0. Moreover, there exist invariant
manifolds (in the sense that we have the inclusions in (5.6)) that contain xq
and are tangent, respectively, to the stable and unstable spaces E^ and E^.
In addition. Theorem 4.12 shows that any sufficiently small perturba^
tion of a vector field with a hyperbolic critical point has a homeomorphic
208 8. Bifurcations and Center Manifolds

phase portrait. Therefore, there are no bifurcations in the neighborhood of


a hyperbolic critical point under sufficiently small perturbations. Hence,
bifurcations may only occur when ^ { 0}.

8.2.2. C enter m anifolds. We start the discussion of the nonhyperbolic


case (when ^ {0 }) with a result that is analogous to the Hadamard-
Perron theorem (Theorem 5.2).

T h eorem 8 .1 1 (Center manifold theorem). If xq g KP' is a critical point of


the equation x' = f{ x ) for a function / : R of class C^, with A; N,
then there exist manifolds W, and of class containing xq such
that:
a) = E^, = E'^ and = E<^;
b) the solutions of the equation x' = f {x ) with initial condition in,
and remain in these manifolds for any sufficiently small
time.

Moreover, the manifolds and are uniquely determined by these prop


erties in any sufficiently small neighborhood of xq.

The proof of Theorem 8.11 falls outside the scope of the book (see [6]
for details). However, it corresponds to an elaboration of the proof of the
Hadamard-Perron theorem (Theorem 5.2). We note that Theorem 8.11
contains Theorem 5.2 as a particular case (when xq is a hyperbolic critical
point).

D efinition 8 .1 2 . The manifolds VP, VP and VP are called, respectively,


stable, unstable, and center manifolds.

Unlike what happens with the stable and unstable manifolds, the center
manifold VP in Theorem 8.11 may not be unique.

E xam ple 8.13. Consider the equation

ix' = x^,
( 8. 12)
1 y ' = -y -

The origin is a critical point and for the function f(x, y) = {x^, y) we have

^(0.0) / = ( o -l)

Thus,
E = { 0} x R , E = {0 } and E = R x {0}.
8.2. Center manifolds and applications 209

One can easily verify that for the initial condition (x(0),j/(0)) = (xo,yo)
equation (8.12) has the solution

(x(t),y(t))=

Eliminating t we obtain y{x) = (yoe which yields the phase por


trait in Figure 8.7.

F ig u re 8.7. Phase portrait of equation (8.12).

Since each orbit in the half-plane x R is tangent to the horizontal


axis, there are infinitely many center manifolds, given by
= {{x,r]c{x)) : a; G R }
where
' i f a : < 0,
ric{x) =
^0 if X > 0
for an arbitrary constant c R. One can easily verify that each manifold
is of class (7 (but is not analytic).

8.2.3. A pplications o f center m anifolds. We describe briefly in this


section how the center manifolds given by Theorem 8.11 can be used in the
study of the stability of a (nonhyperbolic) critical point.
Consider variables x, y, z parameterizing W x x in a sufficiently
small neighborhood of x q . One can show, with an appropriate generalization
of the Grobman-Hartman theorem together with Theorem 2.50 that the
solutions of the equation x' = / ( x ) are topologically conjugate to those of
x' = X,
y' = y, (8.13)
[z' = F{z)
210 8. Bifurcations and Center Manifolds

for some function F (see [6] for details). When ^ {0 } it follows from
(8.13) that the critical point xo is unstable. Now we assume that E'^ = {0 }.
In this case, if E*^ = { 0}, then a;o is asymptotically stable. On the other
hand, if E = { 0} but E^^ { 0}, then the stability of xq coincides with the
stability of the origin in the equation
z' = F{z) (8.14)
(assuming that a;o is represented by z = 0). In summary, we have three
cases:
a) if FF' ^ { 0}, then xq is unstable;
b) if = { 0} and E*^ = { 0}, then xq is asymptotically stable;
c) if E^ = { 0} and E*^ ^ { 0}, then the stability of xo coincides with
the stability of the origin in equation (8.14).
In the third case, it is sufficient to study the behavior on the center manifold.

Exam ple 8.14. Consider the equation

\x' = - x + y'^,
(8.15)
\y ' = y ^ ~ x ^ .

The origin is a critical point and for the function

/(a:, 2/) = (-a: + y^y^ - x ^ )


we have
-1 0
<^(0,0) / 0 0
Thus
E*=Rx{0}, E = {0 } and E^= { 0 } x R ,
and we are in the third case. Since / is of class C, it follows from the Center
manifold theorem (Theorem 8.11) that there exist manifolds W* and of
class for any 6 N. Moreover,

= {{<p{y),y) - y e ( -( 5 ,5 ) }

for some function (^: (5,5) >E of class with ^ (0) = (p'{0) = 0, because
06 and ToW*^ = E"^. Thus, one can write the Taylor series

ip{y) = ay^ + by^ + --- (S.ig)


(up to order k). Substituting x = (p{y) in (8.15), we obtain

x' = <p'{y)y' = <p'{y){y^ - (p{y)^),


or equivalently,
~ ^ {y) + = <^'(y)(y^ - <p(y)^).
8.2. Center manifolds and applications 211

Using (8.16) yields the identity


-ay^ - b y ^ -------- \-y^ ^ {2ay + 3by'^ + ... ){y'^ - a^j/^------ ).
Equating terms of the same order in this identity, we obtain
a + 1 = 0, b = 2a , . . . ,
that is,
a = 1, b = 2, ___
Hence,
<f{y) = - 2y3 + ... , (8.17)
and it follows again from (8.15) that the equation z' = F{z) takes the form

y' = y'^- y>{yf = y^ - y ^ ------ . (8.18)


In a sufficiently small neighborhood of y = 0 the sign of y' coincides with
the sign of y^, and thus the origin is unstable in (8.18). According to the
previous discussion, this shows that the origin is unstable in equation (8.15).
Moreover, it follows from (8.17) and (8.18) that in a sufficiently small neigh
borhood of the origin the center manifold and the motion along are
those shown in Figure 8.8.

F ig u re 8.8. Center manifold o f equation (8.15).

Although it is not necessary for the study of the stability of the origin,
one can also use the former procedure to approximate the stable manifold.
Namely, we have
TU* = { (x ,V '( a :)) :x ( -( 5 ,5 ) }
for some function ip: {6, (5) > E of class with ip{0) = ip'{0) = 0, because
0 G VF and TqVF = E. Then one can write
ip{x) = ax^ + H----- . (8.19)
212 8. Bifurcations and Center Manifolds

Substituting y = 'tp{x) in (8.15), we obtain


y' = tp'{x)x' = tp'{x){-x + rp{x)"^),
and thus,
'ip{x)^ x^ = ij}'{x){x +
Using (8.19) yields the identity
o?x^H----- = {2ax + 3^x^ H----- ){x + ci^x^ H------).
Equating terms in this identity, we obtain
- 1 = -2 a , 0 = - 3 /3 ,...,
and thus,
ip{x) - + Ox^ H----- .
This shows that in a neighborhood of the origin the phase portrait is the
one shown in Figure 8.9.

F ig u re 8 .9. Phase portrait o f equation (8.15) in a neighborhood o f the origin.

Exam ple 8.15. Consider the equation

{ x' = x^ xy, (8.20)


y' = - y + x^.
The origin is a critical point and for the function
/( x , y) = (x^ - xy, - y + x^)
we have

d ( o , o ) /= ( o _ i)-
Hence,
= {0 } X R, E^ = {0 } and E^ = R x {0 }.
8.2. Center manifolds and applications 213

Since f(0 ,y ) = (0, - y ) , the function / is vertical on the vertical axis. This
implies that the vertical axis is invariant. By the uniqueness of the stable
manifold in Theorem 8.11, we conclude that = E^.
Now we consider the center manifold
= {(x,i/;(x)) : X (-(5,<5)},
where
^(a;) = ax^ + ^----- . (8.21)
Substituting y = ijj{x) in equation (8.20), we obtain
y' = ip'{x)x' = i!}'{x){x^ ~ xi){x)),
and thus,
-ip{x) + x'^ = ip'{x){x'^ - X'^{x)).
Using (8.21) yields the identity
-ax^ - b x ^ -------- h = (2ao; + Sbx^ -\----- ){x^ - ax^ - b x ^ ------- ),
and we conclude that
0 = 1, 2o = 6, ___
Thus,
'ip{x) = x^ 2x^ + ,
and
x' x^ xip{x) = x^ x^ + 2x^ H----- .
Therefore, the origin is unstable in equation (8.20). In a neighborhood of
the origin the phase portrait is the one shown in Figure 8.10.

E = VF

F ig u re 8.10. Phase portrait o f equation (8.20) in a neighborhood o f the origin.

Now we illustrate with an example how center manifolds can be used to


simplify the study of bifurcation theory.
214 8. Bifurcations and Center Manifolds

Exam ple 8.16. Consider the equation

{ x' = ex + y^,
( 8.22)
y' = - y +

Adding a third component e' = 0, we obtain

(8.23)

Since the 3 x 3 matrix in (8.23) has eigenvalues 0, 0 and 1, it follows from


the Center manifold theorem (Theorem 8.11) that there exist manifolds W
and that are tangent respectively to the stable and center spaces

E = {0 } X R X {0 } and = R x {0 } x R.

It follows from the discussion in the beginning of this section that in order to
determine the stability of the origin in (8.23) (and thus in (8.22)) it suffices
to study the behavior on the center manifold (or more precisely, on any
center manifold). It can be written in the form

= {{x,ip {x,e),e) : x , e e ( - 5 ,5 ) }

for some function (p: (5,5) x (5,5) > R with ^(0,0) = 0 and d(o,o)V^ = 0-
For y = ip{x, e) we have

and by (8.22), we obtain


dip
{ex -x'^ + y^) = ip(x,e) + x . (8.24)
dx
Writing
<p{x, e) = ax^ + hxe + ce^ H----- ,
it follows from (8.24) that

(2ax + be-\----- ){ex + (p{x, e)^) = ax^ bxe ce^ ----- 1- x^,

which yields a = 6 = c = 0. Hence, by the first equation in (8.22) we obtain

x' = ex x^ + (p{x, e)^ = ex x^ H-----.

The behavior of the solutions of equation (8.22) in a neighborhood of the


origin can now be deduced from the equation x' = e x x^, which was already
studied in Example 8.7.
8.3. Theory o f normal forms 215

8.3. Theory of normal forms

As noted earlier, there are many types of bifurcations and it is desirable to


have some kind of classification. While a corresponding systematic study
falls outside the scope of the book, it is at least convenient to have a proce
dure, as automatic as possible, that can simplify a large class of equations
before we study the existence of bifurcations. The theory of normal forms,
which we introduce in this section, provides such a procedure.
We begin with an example.

Exam ple 8.17. For each e M, consider the scalar equation


x ' = e x + /(x),
where / : R ^ K is a function of class C with /(O) = f'{0) = 0. Writing
the Taylor series of f , we obtain
x ' = e x + a x ^ + b x ^ -\----- . (8.25)
Now we consider a change of variables of the form
x = y + ay^ + Py^ ----- . (8.26)
If the series in (8.26) has a positive radius of convergence (in particular, if it
is a polynomial), then indeed it defines a change of variables in a suflSciently
small neighborhood of the origin. Substituting (8.26) in (8.25), we obtain
y'{l -I- 2ay + SPy"^ -\----- ) = e{y + ay^ + Py^) + a{y^ -1- 2ay^) + by^ -\------ ,
and thus,
ey + {ea -I- a ) y ^ -----
+ eP + {2aoi + b)y^ ^
y' = 1 + 2ay -b _|_
Using the identity
oo

which holds for |r| < 1, after some computations we obtain


y' = {ey + {ea + a)y'^ -!-)
X [l - 2 a y - SPy"^-------- f- {2a y + SPy'^ -|-----)^ -------] (8.27)
= ey + { a - ea)y^ + {b + 2ea^ - 2eP)y^ H----- .
In particular, choosing the coefficients a and P so that the coefficients of y^
and y^ in (8.27) vanish, that is, taking
a j o 6
a - and P = - 5- + r -
e e^ 2e
for e 7^ 0, we obtain the equation
y' = ey + o{y^).
216 8. Bifurcations and Center Manifolds

Now we consider the general case. Given a matrix A G M and a function


/ : R ^ R of class C with /(O) = 0 and dof = 0, consider the equation
x' Ax + f{x ).
In a similar manner to that in Example 8.17, we consider a change of vari
ables X = y + h{y), where h: R > R is a differentiable function with
h(0) = 0 and doh = 0. Then
x' = y' + dyhy' = (Id -I- dyh)y'.
Now we recall that

(-B )* ^ = (Id -b 5 )-i (8.28)


A:=0
for any square matrix B with ||5|| < 1. Since doh = 0, it follows from (8.28)
that for any sufficiently small y the matrix Id -1- dyh is invertible, and one
can write
y' = (Id -I- dyh)~^x'
= (Id A dyh)~'^ [A{y + h{y)) A f{y A h(y))].
It also follows from (8.28) that
y ' ^ { l d - d y h A - - - ) [Ay -h Ah{y) A f { y A h (y))]. (8.29)
Now let Hm,n be the set of all functions p: R " R such that each compo
nent is a homogeneous polynomial of degree m, that is, a linear combination
of polynomials of the form x^^ with
A: N, i i , ... e { 1 ,. . . , n} and mi-I------- f-mfe = m.
We note that Hm,n is a linear space. Write
/ = /2 + /3 + --- (8.30)
and
/i = Id + /^2 H ^3 I" *** )
where fm^gm ^ for each m > 2. It follows from (8.29) that
y' = Ay A [Ah2{y) A h iv ) ~ Dyh2Ay]
(8.31)
A ^ [A h k {y ) A 9k{y) - DyhkAy],
fc=3
where each function : R R depends only on /i2, . . . , hk-\- In order to
eliminate all possible terms in the right-hand side of (8.31), we try to solve,
recursively, the equations
Dyh2Ay - Ah2{y) = h {y ) (8.32)
and
DyhkAy - Ahkiy) = 9k(y) (8-33)
8.3. Theory o f normal forms 217

for fc > 3, finding successively functions h2,hs,. .. Equations (8.32) and


(8.33) are called homological equations.
For eeich m > 2 and n N, we define a map
. TT H,m^n
A m,n

by
{L ^ h ){y) = D yh A y-A h {y).

One can easily verify that is a linear transformation. The homological


equations can be written in the form
kyTii
L\^h2 = f 2 and L'^^hk = 9k, k > 3. (8.34)

E xam ple 8.18. Let m n = 2 and A = ( - i o)- Each function h H2,2


can be written in the form

h{x, y) = [ax^ + bxy + cy'^, dx^ + exy + fy'^),

for some constants a, 6, c, d, e and / . In particular, d im if2,2 = 6. We obtain

/^2ax + 6y 6 a ; + 2cy\
( y \ _ ( dx"^ + exy +
VA )\ iV) \2dx-\-ey ex + 2 fy ) \x ) \ax^ bxy cy^
(2xy
= a
V x2
-x^ 2
+ d
2xy -2 x y
22
and thus, the linear transformation L2 is represented by the matrix

/O -1 0 -1 0 0 \
2 0 -2 0 -1 0
0 -1 0 0 0 -1
1 0 0 0 -1 0
0 1 0 2 0 -2
^0 0 1 0 1 0 /

with respect to the basis of i?2,2 given by

xy 0
0 xy Of
We note that in order to solve the homological equations in (8.34) it
is sufficient to find the inverse of each transformation L ^ ", when it exists.
But since is linear, in order to decide whether the transformation is
invertible it is sufficient to compute its eigenvalues.
218 8. Bifurcations and Center Manifolds

P rop osition 8.19. If X\,... ,Xn are the eigenvalues of A (counted with
their multiplicities), then the eigenvalues of L'2'^ are the numbers

^k + (8.35)
i= l
with
/c 6 { 1 , . . . ,n}, m i , ... ,mn NU {0} and mi + ----- |-mn = m.
Moreover, the eigenvectors of associated to the eigenvalue in (8.35) are
the polynomials
y .m i
(8.36)
where Vk is an eigenvector of A associated to the eigenvalue A*.

The following is a particular case.

E xam ple 8.20. For the diagonal matrix

Ai 0
A=
\0 ^n/
if e i , . . . , Cn is the standard basis of R , then
J^m,n(^mi . . . x^^ek)(xi, . . . , Xn)
( 0

0 Ai^i'
(mia:^ ^ ^x^--Xn) (m ^ x f x(f;^ ^)
0 \An3^n)

\ 0
-x^^-'-x^XkCk
= ( m iA i + + n i n X n j x ^ ^ x ^ ^ e k ^ x^^ XkCk

~f^ miAj - Afc^ x^^ Xn"ek.


\ i=i '
This shows that x'ff^ek Hm,n is an eigenvector of associated
to the eigenvalue in (8.35).

By Proposition 8.19, one may not be able to solve a homological equation


in (8.34) if some of the numbers in (8.35) are zero. This motivates the
following notion.
8.3. Theory o f normal forms 219

D efinition 8 .2 1 . The vector (Ai,. . . , An) is said to be resonant if there exist


integers k E {1,... ,n} and m i , , m 6 M U {0}, with mi + -----1- m > 2,
such that

Afc ^ ymiXj.
i=l

When there are no resonant vectors one can solve all homological equa
tions and thus eliminate all terms fm in (8.30). We recall that for a function
p : R > we write g{x) = o(||x||^) if g{x)/\\x^^ > 0 when x ^ 0.

T h eorem 8 .22 (Poincare). Let A be ann x n matrix and let / : R )M


be a function of class C with /(O) = 0 and dof = 0. If the vector formed
by the eigenvalues of A is not resonant, then for each A: G N the equation
x' = Ax + fix) can be transformed into
v' = Av + o(|M )

by a change of variables x = y + h{y).

P roof. Since the vector (Ai,. . . , A) formed by the eigenvalues of A is not


resonant, all the eigenvalues of in (8.35) are different from zero. Thus,
each linear transformation L ^ ' is invertible and one can obtain successively
solutions of all homological equations in (8.34) up to order k.

Now we describe briefly what happens when there are resonant vectors.
In this case one can only invert the restriction of L to the subspace E of
Hm,n generated by the root spaces corresponding to the nonzero eigenvalues
in (8.35). More precisely, consider the decomposition Hm,n = E F , where
F is the subspace of Hm,n generated by the root spaces corresponding to
the zero eigenvalues in (8.35). Since

L^'^E = E and = F,

one can write the linear transformation in the form

Le 0
0 Lp
with respect to the decomposition Hm,n = E F. We also write

- (ff and h2= [^F

The first homological equation in (8.34) takes the form

iL E hi = f i , i h i = L-^^fi,
\LFhP = f i [LFh^ = fi'.
220 8. Bifurcations and Center Manifolds

Substituting h f in (8.31), we finally obtain


y' = Ay + M y ) - { L : ^ ^h2){y) + - -
= A y + fiiy ) - + f i ( y ) - { L p h ^ X y )
= Ay + (y) - {LFh^){y) +
The terms of degree 2 with values in F cannot be eliminated, unless we
have already / ; f = 0 from the beginning. Continuing this procedure for the
remaining homological equations in (8.34), in general only the terms of the
form (8.36) cannot be eliminated.

E xam ple 8.23. Consider the equation


V _ / 0 / a:
+ (8.37)
^y) - 1,-2 Oj \y,
The eigenvalues of the 2 x 2 matrix in (8.37) are 2i, and thus, in order to
determine whether there are resonant vectors we have to solve the equation
mi2i + m2(2i) = 2i,
that is,
mi m2 = 1,
with m i, m2 N U {0 } and mi 4- m2 > 2. The solutions are
(mi, m2) = (1,2),(2,1),(2,3),(3,2),....
By the former discussion, this implies that by a change of variables only the
terms of the form
xy^, x^y, x^y^, x^y'^, . ..
cannot be eliminated. In other words, denoting by u and v the components
in the new coordinates, equation (8.37) takes the form

u' = 2v + aiuv^ + a2 U^v + asu^v^ + Q4 U%^ H------,

{ v' = -2 u + biuv"^ + b2U^v + bzu^v^ + b^u^v^ ^-----

for some constants Oi, 6i M with N.

Exam ple 8.24. Consider the equation


2 0
0 1 + (8.38)

The eigenvalues of the 2 x 2 matrix in (8.38) are 2 and 1, and thus, in


order to determine whether there are resonant vectors we have to solve the
equations
2mi + m2 = 2 and 2mi + m2 = 1 ,
with m i , m2 G N U {0} and mi + m2 > 2. One can easily verify that
(mi, m2) = (0,2) is the only solution. This implies that denoting by u and v
8.3. Theory o f normal forms 221

the components after an appropriate change of variables, equation (8.38)


takes the form

{ u' = 2u +
u' = u +

for some constants a, /3 M. Moreover, one can take /3 0, since by Propo


sition 8.19 the solution {m i,m2) = (0,2) corresponds to the eigenvector
(y^,0) (see (8.36)), which has the second component equal to zero.

We conclude this section with a general result concerning the perturba


tions of linear equations whose matrices have only eigenvalues with positive
real part.

T h eorem 8.25. Let A be an n x n matrix having only eigenvalues with


positive real part and let / : R " )E 6e 0 function of class C with /(O) = 0
and dof = 0. Then for each k N the equation x' = Ax + f {x ) can he
transformed by a change of variables x = y + h{y) into

y' = Ay + p{y) + o{\\yf),

where p is a polynomial.

P ro o f. In order to verify whether there are resonant vectors, consider the


equation

Xj y ^'rriiXi (8.39)
i=l
We note that
n
Re Xj = rrii Re A^. (8.40)
i=l
Taking a vector ( m i , . . . , m) such that

maxi Re
E
i=l
rrii >
mini Re Xi

(we recall that by hypothesis the eigenvalues have positive real part), we
obtain

rrii Re Xi > rui min Re Xi


i=l i=l
> maxReAi > ReA^.
i
This shows that identity (8.40), and thus also identity (8.39), do not hold
for the vector {m\,... ,mn). Therefore, there exist at most finitely many
222 8. Bifurcations and Center Manifolds

resonant vectors ( m i , . . . , rrin), taken among those such that


m axi Re Aj
<
mim Re Xi
i=l
This yields the desired result.

8.4. E x e rcise s

Exercise 8 .1 . Consider the equation

{ x' = xy + ax^ + bxy^,


y' = - y + cx^ + dx^y.

a) Find a center manifold of the origin up to third order,


b) Determine the stability of the origin when a + c < 0.

Exercise 8 .2 . Consider the equation


\x' = 2 x + p { x , y ) ,
\y' = y + q{x,y),
where p and q are polynomials without terms of degrees 0 or 1.
a) Find the stable, unstable and center manifolds of the origin.
b) Show that by an appropriate change of variables the equation can
be transformed into z' 2z-\- cw^, w' = w.

Exercise 8.3. For each e R, consider the equation


(x' = y - x ^ ,
[y ' = e y - x^.
a) Use a Lyapunov function to show that the origin is stable when
e = 0.
b) Find a center manifold of the origin up to fifth order,
c) Determine the stability of the origin for each e 0.
d) Find a center manifold of the origin when e = 1.
e) Find whether there exist heteroclinic orbits when = 1.

E xercise 8.4. Consider the equation


ix' = y + x^,
= ex - y2.
a) Find whether there exist periodic orbits contained in the first quad
rant when e = 1.
b) Determine the stability of all critical points when e = 1,
8.4. Exercises 223

c) Determine the stability of the origin when 6: = 0.


d) Find whether any bifurcation occurs for some e > 0,

E xercise 8.5. For each (e, 5) G consider the equation


x' = ex y + Sx(x^ +

{;
y = x - e y + Sy{x^ + y^).
a) Determine the stability of the origin for each pair (e, 5) with |e| < 1.
Hint: For the function
V{x, y) = x^ + y^ - 2exy
we have
V {x, y) = 25{x^ + y'^)V{x, y).
b) For each pair (s,S) with |e| > 1, show that no bifurcations occur in
a neighborhood of the origin.
c) For each pair (e, S) with 5 < 0, show that each positive semiorbit is
bounded.

Solutions.
8 .1 a) = {(x, cx^ + o(x^)) : x 6 (5,5)}.
b) The origin is unstable.
8 .2 a) = { ( 0, 0)}, = {-5,6) x {-5,5) and = { ( 0, 0)}.
8.3 a) {x^ + 2]p)' 4{x^x' + yy') = 4a: < 0.
{-5,6) X {-6,6) if e = 0,
b) =
({x,x^/e + 3/e + ) : X e (5,5)} if e 7^ 0.
c) Unstable.
d) tU= = { ( x , a ; 3 ) : a ; e ( - 5 , 5 ) } .
e) There exist.
8.4 a) There are none.
b) (0, 0) are ( 1, - 1) are unstable.
c) Unstable.
d) There are no bifurcations.
8.5 a) Unstable for 5 > 0, stable but not asymptotically stable for 5 = 0,
and asymptotically stable for 5 < 0 .
Chapter 9

Hamiltonian Systems

In this chapter we give a brief introduction to the theory of Hamiltonian


systems. These are particularly important in view of their ubiquity in phys
ical systems. After introducing the basic notions of the theory, we establish
some results concerning the stability of linear and nonlinear Hamiltonian
systems. We also consider the notion of integrability and, in particular, the
Liouville-Arnold theorem on the structure of the level sets of independent
integrals in involution. In addition, we briefly describe the basic ideas of
the Kolmogorov-Arnold-Moser theory. For additional topics we refer the
reader to [1, 3, 2 2 , 25].

9.1. Basic notions

Let H : -> R be a function of class C^, called a Hamiltonian. More


generally, one can consider Hamiltonians in an open subset of R^ . We
write the coordinates of R^"' in the form (q,p), where q = {qi,... ,qn) and
p = (pi,...,Pn).

D efinition 9.1. The system of equations

/ , . , dH / s . 1 ,V
= (9.1)

is called a Hamiltonian system with n degrees of freedom.

The following are examples of Hamiltonian systems.

Exam ple 9.2. Let [/: R > R be a function of class C^. Letting q' = p,
the equation q" = U'{q) is equivalent to the Hamiltonian system with one

225
226 9. Hamiltonian Systems

degree of freedom determined by the Hamiltonian

H{q,p) = T
(9.2)

It is easy to verify that H is constant along the solutions of the equation

Incidentally, the terms p^/2 and U{q) in (9.2) correspond respectively to the
kinetic and potential energies.
Exam ple 9.3. The Kepler problem is a special case of the two-body prob
lem and can be described by a Hamiltonian system with two degrees of
freedom; namely, consider the Hamiltonian
\\pf (9.4)
H{q,p) = ^ - U(q)

in for the potential U{q) = M/lkll> with p E The corresponding


Hamiltonian system is
pq
q' = P , P' = 13- (9.5)

Now we write system (9.1) in the form


x> = X h {x ) = JWH(x), (9.6)
where x = (q,p) E R^ ,

VH =
'dH ^ ^ dH
d q i " ' dqn d p i ' " ' ' dpn

and
0 Id^
J=
^-Id 0
Here Id E Mn is the identity matrix. In particular, we would like to know
which transformations preserve the Hamiltonian structure, in the following
sense. Making the change of variables x = q>{y), the Hamiltonian sys
tem (9.6) is transformed into
y' = i<P*X){y),
where
{^*XH){y) = {dyq>r^XHi<p{y)).
We would have <
p *X h = XHotp, and hence also

y' = Xnoipiy),
which would correspond to the preservation of the Hamiltonian structure, if
and only if
{dy^)-^JVH{q>{y)) = J ^ { H o ^){y)
9.1. Basic notions 227

for every y E Since

V{ Ho (p ){ y ) = {dy(p)*VH{(p{y)),

the identity (p*Xn = X ho<


p holds whenever

{dyy>)-^J = J{dyy^r. (9.7)

Since J~^ = J, identity (9.7) can be written in the form

^dy^'^ J d y ^ J. (9.8)

D efinition 9.4. A differentiable transformation : V R^ in an open set


V C R^" is said to be canonical if identity (9.8) holds for every y e V.

Exam ple 9.5. We show that for n = 1 a differentiable transformation


V?: F > R^ is canonical if and only if det dyip = 1 for every y E V . Writing

a b
dyif =
c d
identity (9.8) is equivalent to

o h\ ( Q a b 0 ad be
c d V- 1 c d -ad + 6c 0 )(- ;)
The last equality holds if and only if

det dy(p ad be = 1.

This shows that for n = 1 the canonical transformations are the differen
tiable transformations preserving area and orientation.

Now we show that the solutions of a Hamiltonian system induce canoni


cal transformations preserving volume and orientation. Let H : R^ -> R be
a Hamiltonian of class Given x E R^", we denote by <pt{x) the solution
of the Hamiltonian system (9.6) with initial condition ipo{x) = x.

P rop osition 9.6. Given a Hamiltonian H of class C^, for each sufficiently
small t the map ipt is a canonical transformation preserving volume and
orientation.

P roof. We have
= X h M x )). (9.9)

Since H is of class it follows from (9.9) and an appropriate generalization


of Theorem 1.42 for vector fields that
d ,
228 9. Hamiltonian Systems

This shows that dx^pt is a solution of the linear variational equation of the
Hamiltonian system along the solution (pt{x). Moreover, since (po{x) = x,
we have dx^o = Id- Now let
ot{i^ = (^dx^t) Jdx^t'
We want to show that a{t) = J. Taking derivatives with respect to t, we
obtain
^ (0 i^(pt(x)^Hdx^t) Jdx^t + {dx^t) Jd(p^(^x)^Hdx^t
{dxVt) [{^(pt{x)^H) d -\- Jdfp^^x^Xfj^dx^ti

and it follows from the identity dyXn = JdyH that


a'it) = +
Since J*J = Id and = Id, we obtain
a'(t) = = 0.
On the other hand, since dx(po = Id, we have a(0) = J and thus a{t) =
q;(0) = J. This shows that (pt is a canonical transformation. Since

det(da;^i)* = deldxpu
it follows from identity (9.8) that |detda;(^tP = 1- But since detd^yo =
detid = 1, it follows from the continuity of 1 <ptix) that detda;<^t = 1 for
any sufficiently small t. This establishes the desired result.
Exam ple 9.7. Consider the Hamiltonian system with one degree of freedom

\Q = P ,
1 p' = -Q,
determined by the Hamiltonian

H(q,p) 2^ 2'
By Example 1.8, we have
>t(qo,Po) = (po sin t + qocost, po cos t - qo sin t)
cost sin A / qo
sint cos t j YPo>
and clearly,
J .J , / cost
detdto,.,=det l
sint'
cost = 1.
We also note that for any differentiable function F : -i given a
solution x(t) of equation (9.6) we have

^ F (a ;(t)) = VF(x(t)) Xff(x(t)) = (VF(x(t)))*JVH(x(t)).


9.2. Linear Hamiltonian systems 229

D efinition 9.8. The Poisson bracket {F, G} of two differentiable functions


F, G : R is defined by

{F .G }(x ) = ( V F ( x JVG(x) = g

If a;(f) is a solution of the Hamiltonian system (9.6) and F is a differen


tiable function, then

Therefore, in order that F is an integral of equation (9.6) (see Defini


tion 1.67) it is necessary that {F, JI) = 0. Incidentally, we have {If, If} = 0,
and thus any Hamiltonian is constant along the solutions of its Hamiltonian
system.

9.2. Linear Hamiltonian system s

In this section we consider a particular class of Hamiltonians. Namely, given


a 2n X 2n symmetric matrix S, consider the Hamiltonian

If(x) = L * S x .
Clearly, is a polynomial of degree 2 without terms of degree 0 or 1. The
corresponding Hamiltonian system, called a linear Hamiltonian system, is
x' = X h {x ) JSx.
D efinition 9.9. A matrix B = JS with S symmetric is said to be Hamil
tonian. A matrix A is said to be canonical if
A*JA = J. (9.10)

We give several examples.


Exam ple 9.10. It follows from (9.8) that if F is a canonical
transformation, then dy(p is a canonical matrix for each y V .

Exam ple 9.11. We show that a matrix B is Hamiltonian if and only if


B * J - IJ B = 0. (9.11)
Indeed, if B is Hamiltonian, then
B*J -IJB = S*J*J -IJ^S = S* - S = 0.
On the other hand, if identity (9.11) holds, then
i-JB)* = B * { - j y = B*J = - J B ,
and the matrix S JB is symmetric. Moreover, JS = J^B = B and
the matrix B is Hamiltonian.
230 9. Hamiltonian Systems

Exam ple 9 .1 2 . We show that any exponential of a Hamiltonian ma


trix H is a canonical matrix. Let B = JS, with S symmetric, be a Hamil
tonian matrix, and consider the function
P{t) =
We have P{0) = J and

= {e^^)*S*J*Je^^ + {e^yj^Se^^
= {e^ySe^^ - {e^ySe^^ = 0,
because the matrix S is symmetric. This shows that fi{t) J for t M, and
each matrix is canonical.

Now we describe some properties of the eigenvalues of Hamiltonian ma


trices and canonical matrices.

P rop osition 9.13. Let B be a Hamiltonian matrix. If A is an eigenvalue


of B, then A, A and A are also eigenvalues of B, with the same multiplic
ity as A. Moreover, ifO is an eigenvalue of B, then it has even multiplicity.

P roof. It follows from (9.11) that B = J~^{B*)J. Hence, the matrices


B and B* are similar and they have the same eigenvalues (with the same
multiplicity). Thus, if A is an eigenvalue of B, then A is an eigenvalue of
B* and hence also of B (with the same multiplicity). Since B is real, we
conclude that A and A are also eigenvalues (with the same multiplicity).
In particular, the nonzero eigenvalues of B form groups of two or four, and
thus if 0 is an eigenvalue, then it has even multiplicity.

P rop osition 9.14. Let A be a canonical matrix. If X is an eigenvalue of A,


then 1/A, A and 1/A are also eigenvalues of A, with the same multiplicity
as A. Moreover, if 1 and 1 are eigenvalues of A, then each of them has
even multiplicity.

P roof. It follows from (9.10) that A is invertible and A* = JA~^J~^. In


particular, the matrices A~^ and A* are similar and they have the same
eigenvalues (with the same multiplicity). Thus, if A is an eigenvalue of A,
then 1/A is an eigenvalue of A~^ and hence also of A*. This shows that 1/A
is an eigenvalue of A (with the same multiplicity as A). Since A is real, we
conclude that A and 1/A are also eigenvalues (with the same multiplicity).
The last property follows immediately from the previous ones.

By Theorem 3.10, a linear equation x' = Bx is asymptotically stable if


and only if B has only eigenvalues with negative real part, and is stable if
and only if B has no eigenvalues with positive real part and each eigenvalue
9.3. Stability o f equilibria 231

with zero real part has a diagonal Jordan block. It follows easily from
Proposition 9.13 that for Hamiltonian matrix B the linear equation x' = Bx
is not asymptotically stable. Moreover, if the equation is stable, then all
eigenvalues of B are purely imaginary and have a diagonal Jordan block.

9.3. Stability of equilibria

In this section we study the stability of the critical points of a nonlinear


Hamiltonian system. Let H: R be a Hamiltonian of class and
consider the corresponding Hamiltonian system x' = X h {x ).
D efinition 9.15. A point xq R^ is said to be an equilibrium of the
Hamiltonian H if VH { xq) = 0. An equilibrium xq of H is said to be
nondegenerate if the Hessian matrix is nonsingular.

We note that the equilibria of H are critical points of x' = X h {x ).


We recall that the index of a bilinear transformation F : R^ x R^" > R,
such as, for example, (u, u) i-> u*d"^^Hv, is the maximal dimension of the
subspaces E C R^ such that F is negative definite in E x E. The following
statement shows that the form of a Hamiltonian i f in a neighborhood of a
nondegenerate equilibrium xq is completely determined by the index of the
Hessian matrix d^^H (for a proof see, for example, [21]).

P rop osition 9.16 (Morses lemma). If H : R^ > R is a function of


class and xq R^"' is a nondegenerate equilibrium of H, then there
exists a change of variables g: B{0,r) > V, where V is a neighborhood
ofxo, such that g{0) = xq and
A 2n
{ H o g ) { y i , . . . , y 2n) = H { x o ) - ' ^ y i + ^ yf
z=AH-1

for every ( yi ,... ,y2n) H(0,r), where A is the index of

Now we present a first result on the stability of the equilibria of a non


linear Hamiltonian system. The proof uses Morses lemma.
T h eorem 9.17. Let H be a Hamiltonian of class C^. If xq is an equilibrium
of H with (positive or negative) definite Hessian matrix then xq is a
stable but not asymptotically stable critical point of the equation x' = X h {x ).

P ro o f. By hypothesis, the index of the Hessian matrix ^ or 2n.


Hence, it follows from Morses lemma that there exists a change of variables
g: 5(0 , r) > V, where F is a neighborhood of xq, such that y(0) = xq and
2n
{H o g){yi,.. . ,p2n) = H { x o ) Y ^ y f
2=1
232 9. Hamiltonian Systems

for every {yi,... ,y2n) G B{0,r). This shows that for c sufficiently close
to H { x q ), with c > H { x q ) for positive definite, and c < H { x q ) for
dl^H negative definite, the level sets
V c ^ { x e V : H(x) = c}
are diffeomorphic to 2n-dimensional spheres that approach the point xq
when c > H{ xq). On the other hand, since H is constant along the solutions
of its Hamiltonian system, any solution >pt{x) with initial condition x K
remains in Vc for every t G M (we note that since Vc is compact, the maximal
interval of any of these solutions is K). This yields the desired result.
Exam ple 9.18 (continuation of Example 9.2). Consider the Hamiltonian H
in (9.2). It follows from Theorem 9.17 that if qo is a strict local minimum
of U with U"{qo) > 0, then (^OtO) is a stable but not asymptotically stable
critical point of equation (9.3). Indeed, since qo is a local minimum of U,
we have U'{qo) = 0 and hence (9o>0) is a critical point of equation (9.3).
Moreover, since U"{qo) > 0, the Hessian matrix

(Tj> 0^
= d,(go,0)(C ,P) - ( Q

is positive definite.

Now we consider the case when is not definite. In order to study


the stability of the equilibrium xq it is convenient to consider the linear
variational equation
y' = dxoXny-
D efinition 9.19. The eigenvalues of the Hamiltonian matrix dx^Xa =
Jd'^^H are called the characteristic exponents of at xq.

It follows from Theorem 4.9 that if some characteristic exponent of X jj


at Xq has positive real part, then xq is an unstable critical point of the equa
tion x' = X h {x). Since the matrix dx^Xn is Hamiltonian, it follows from
Proposition 9.13 that a necessary condition for the stability of an equilib
rium Xq is that all characteristic exponents of X h at xq are purely imaginary,
say equal to i A i , . . . , iA . Now we give a sufficient condition for stability.

P rop osition 9 .2 0 . Let xq be an equilibrium of H with characteristic expo


nents
i A i , . . . , fAn.
(f |Ai|,.. . , |An| are nonzero and pairwise distinct, then xq is a stable but not
asymptotically stable critical point of the equation x' = X h {x).

P roof. Without loss of generality, we assume that


Ai > > An > 0. (9.12)
9.3. Stability o f equilibria 233

Now let w i , . . . , Wn,wi,... ,Wn G \ {0 } be the eigenvectors associated


respectively to the eigenvalues iAi, . . . , iA, iAi, . . . , iXn of the matrix
B = Jd^^H. We write
Uj = Rewj and Vj =
One can easily verify that u i ,... ,Un,vi,... ,Vn is a basis of and it
follows from the identity Bwj = iXjWj that
Buj = XjVj and Bvj = XjUj (9.13)
for j = 1 , . . . , n (because B is real). We also assume, without loss of gener
ality, that
UjJvj = 1 for j = 1,... ,n. (9-14)

Now we consider the bilinear transformation (u, v) = u*Jv in x


We have
{u,v) (u,v)* v*J*u = v*Ju = {v,u) (9.15)
for u,v E In particular, taking u = u, we obtain
(uj,Uj) {vj, Vj) = 0 for y = 1 , .. ., n. (9.16)
Moreover, it follows from (9.14) and (9.15) that
{uj,Vj) = ~{vj,Uj) = 1 for j = 1,... ,n. (9.17)
We also have
{Bwj ,Wk)= iXj {wj ,Wk), (9.18)
and using (9.11), we obtain
{Bwj,wk) = WjB*Jwk = -WjJBwk = -iXk{wj,Wk). (9.19)
Thus,
(Aj + Xk){wj,Wk) = 0,
and since Aj -I- A^ > 0, we conclude that
{wj,Wk) = 0 for j,k = l , . . . , n .
Substituting wj = uj + ivj and Wk =Uk + ivk yields the identity
(u j, Uk) + i{uj, Vk) + i{vj,Uk) - {vj,Vk) = 0. (9.20)
Proceeding as in (9.18) and (9.19) with wj replaced by wJ, we obtain
i{-Xj + Xk){wj,wk) = 0 for j, k = l , . . . , n .
Thus, it follows from (9.12) that {wj,Wk) 0 for j k, and hence,
(uj,uk) + i{uj,Vk) - i{vj,Uk) + {vj,Vk) = 0 (9.21)
for j 7^ k. Adding and subtracting (9.20) and (9.21) yields the identities

{Uj,Uk) "I" i{Uj,Vk) 0


234 9. Hamiltonian Systems

and
i{vj,Uk) - {vj,Vk} = 0
for j ^ k. Taking the real and imaginary parts, we finally obtain

{uj,Uk) = {uj,Vk) = {vj,Uk) = (vj,Vk) = 0 (9.22)

for j ^ k. It follows from identities (9.16), (9.17) and (9.22) that the ma
trix C with columns u i , ... ,Un,vi,... ,Vn satisfies

C*JC = J.

In other words, C is a canonical matrix. Moreover, by (9.13), we have

/ Ai \

Xrt
C~^BC =
Al

and hence.

A l
An
C*JBC = JC~^BC = -
Al

XnJ

Therefore, the change of variables x xq Cy transforms the Hamiltonian H


into the function

y ^ H{xo) + ^{Cyfdl^HCy + o(||yf)

= H { x o ) - \ y * C * J B C y + o{\\y\\^)

= H { x^ ) - \ y* J C - ^ B C y + o{\\yf)

= H{ xq) + ^ ^ Aj(g^ -|-p|) + odlylA,


j= i

where y = ( ? i , . , 9n)Pi) >Pn)- The desired result can now be obtained


as in the proof of Theorem 9.17.
9.4. Integrability and action-angle coordinates 235

9.4. Integrability and action-angle coordinates

The solutions of a Hamiltonian system rarely can be obtained explicitly.


In this section we consider a class of Hamiltonians for which this is possible.
Let H : -> R be a Hamiltonian of class C^.

D efinition 9 .2 1 . If there exists a canonical transformation into variables


(0, / ) G T X R such that the Hamiltonian H can be written in the form
H{q,p) = h{I), then the variables (0,1) are called action-angle coordinates.

In action-angle coordinates, the Hamiltonian system can be written in


the form
e' = oj{I), J' = 0,
where u{I) = V/i(/). Hence, the solutions are given explicitly by
6{t) = Oo + tu{Io), I{t) = Jo,
where 6q = 0(0) and I q I (0).

E xam ple 9.22. Consider the Hamiltonian

H{q,p) = 1 ^ 2 ^ \ -fO}- (9-23)


i=i
One can pass to action-angle coordinates by applying the transformation
Qj cosOjy pj = ^/^sin0j, j = 1,... ,n.
By Example 9.5, this transformation is canonical, because the derivative of
(qjyPj) has determinant 1 for j = 1 , . . . , n. In the new coordinates
(0,7) the Hamiltonian H{6,I) = H{q,p) is given by
n

m i ) =
j= i
The corresponding Hamiltonian system is
J ' = 0, 0'=A,', j = n.

In general, a Hamiltonian cannot be written in action-angle coordinates.


However, there is a class of Hamiltonians, called completely integrable (or
integrable in the sense of Liouville), for which locally there exist action-angle
coordinates. More precisely, for a Hamiltonian H : R^" > R we describe
how the existence of n integrals F i, . . . , satisfying certain properties in a
set U C R^" has important consequences concerning the existence of local
action-angle coordinates and the structure of the solutions in each level set
Mc = [x U : Fj{x) = Cj for j = 1 , .. ., n}, (9.24)
where c = ( c i , . . . , c) R . We first introduce some notions.
236 9. Hamiltonian Systems

D efinition 9.23. Given n differentiable functions F \ , . .. , Fn: -> M, we


say that they are:
a) independent in 17 C if VFi(a:),. . . , S/Fn{x) are linearly indepen
dent for each x G 17;
b) in involution in 17 C R^ if {Fi, F j } = 0 in U fov i, j = 1 , ... ,n.

We also introduce the notion of complete integrability.

D efinition 9.24. A Hamiltonian H is said to be completely integrable (or in-


tegrable in the sense of Liouville) in a set 17 C R^" if there exist n integrals
F \ , . . . , F n of equation (9.6) that are independent and in involution in U.

Exam ple 9.25 (continuation of Example 9.22). Consider the Hamilton


ian H in (9.23). One can easily verify that the functions

H j { q , p ) = X j i q ] + p j ) / 2 ,

for j = 1 , . . . , n, are integrals, and that they are independent and in involu
tion in R^ . Hence, the Hamiltonian H is completely integrable.

Exam ple 9.26 (continuation of Example 9.3). Consider the Hamiltonian H


in (9.4) and the angular momentum

L {q ,p ) =qm - q2Pi- (9.25)

One can easily verify that L is an integral. Indeed, it follows from (9.5) that
if x{t) = (q{t),p(t)) is a solution, then

L{x{t)) = q[p2 + qiP2 - ?2Pi - q2P'i


pq2 , pqi ^
= P1P2 - 9 l | j ^ - P2P1 + q 2 j ^ = 0.

Moreover, the integrals H and L are independent outside the origin and
{H, L} = 0. For the first property, we observe that

pq \
V H = { p , and V L = {p2, - p i , - 92, 9i)-
M V
Hence VH WL = 0, and V i7 and V L are linearly independent outside the
origin. For the second property, we note that

SH L \ ~ aTT dL dH dL dH dL
dq\ dp\ dq2 dp2 dpi dq\ dp2 dq2
= ]|^(-92) + - P2(-Pi) = 0.

Hence, the Hamiltonian in (9.4) is completely integrable outside the origin.


9.4. Integrability and action-angle coordinates 237

Now we formulate the following important result without proof (for de


tails see, for example, [19]). We denote by T = E"'/Z"' the n-dimensional
torus.

T h eorem 9.27 (Liouville-Arnold). For a Hamiltonian H that is completely


integrable in an open set U C let F i , . .. ,Fn'. U R be integrals of
equation (9.6) that are independent and in involution in U. Then:
a) each set Me in (9.24) is an n-dimensional manifold that is invariant
under the solutions of the equations x' = Xpi{x) for i . ,n;
b) if all solutions of these equations in Me have maximal interval R,
then each connected component of Me is diffeomorphic to x R *^
for some 0 < k < n, and there exist action-angle coordinates in a
neighborhood of Me; moreover, the solutions of the equation xf =
X h {x ) in Me induce the trajectories

t (Vo + toj mod 1 , 2/0 + tu)


in X where u> = lo{c) E R*^ and v = v{c) 6 R ~*^.

We first consider an example in R^.

Exam ple 9.28. The Hamiltonian H in (9.2) is completely integrable in a


neighborhood of any point (qo, 0) R^ where qo is a strict local minimum
of U. Moreover, by Morses lemma (Proposition 9.16), up to a change of
variables in a neighborhood of (go.O) the level sets Me in (9.24) are circles.
This corresponds to taking n = k = 1 in Theorem 9.27, in which case the
sets Me are diffeomorphic to the torus = 5^.

Now we present an example illustrating the complexity that can occur in


the decomposition into level sets in Theorem 9.27. The details fall outside
the scope of the book, and so we partially omit them.

E xam ple 9.29 (continuation of Example 9.26). Consider the Hamiltonian


H in (9.4) and the angular momentum L in (9.25). Introducing the canonical
transformation

^ = \/<li+Qh ^ = arctan , pr = Pe 9iP2 - 92Pi,


^ 91 V g T + 92
^
one can write

+ and L = ps.

The Hamiltonian system can now be written in the form


, dH OH , dH , dH
''= W r' We' =
238 9. Hamiltonian Systems

that is,
r' =Pr,
e' = pe/r^,
Pr = -Pe/f'^ +
/ e = 0-
For the integrals F\ = H and F2 = L, we have the level sets
Me = {{r,e,pr,pe) 6 R ''' X 5 ^ X : H{r,pr,pe) = a,pe = b}

= | ( t, e,pr, ft) e IR'* X -5^ X :2 ~ r ^

where c = (o,ft). Since pi + h^/r^ > 0, the region in the variables {r,6)
where there exist solutions is given by

Ra = [{r, 6) eR -^

Using the symbol A ^ B to indicate the existence of a diffeomorphism


between A and B, one can show that:
a) if /i < 0, then Ra = 0 ior a < 0, and Ra { - p / a ,+ o o ) x 5^ for
o > 0;
b) if p 0, then i?a = 0 for a < 0, and Ra R '*' x for o > 0;
c) if p > 0 , th e n Ra ~ R'*' x fo r a > 0, and Ra (0 , p/a) x fo r
a < 0.

Now we define
S a= (J U r , e ) ,
(Ty0)GRa
where
Ia{r, 9) = | (r , 9,Pr,P0) 6 R+ x x R^ : +Pg/r^ = 2 | .

The set Ia(r,9) is an ellipse, a point or the empty set, respectively, if (r,9)
is in the interior, in the boundary or in the exterior of Ra- One can show
that:
a) if p > 0, then S a ^ S ^ \ 5^ for a < 0, and S'a 5^ \ {S^ U 5^) for
a > 0;

b) if p = 0, then = 0 for a < 0, and 5o ~ 5^ \ {S^ U S^) for a > 0;


c) if p < 0, then 5a = 0 for a < 0, and 5a ~ 5 \ 5^ for a > 0.
Taking into account that
Me = 5a n(R + X 5^ X R X {b}),
one can then describe each level set up to a diffeomorphism.
9.5. The KAM theorem 239

An interesting particular case of the Liouville-Arnold theorem (Theo


rem 9.27) occurs when the level sets Me are compact. In this case each
connected component of Me is diffeomorphic to the torus T (because R* is
not compact). Moreover, there is no need to assume as in Theorem 9.27 that
the solutions of the equations x' = Xpiix) have maximal interval R, since
then this property is automatically satisfied. In this context we introduce
the following notion.
D efinition 9.30. The trajectories in a torus T diffeomorphic to a level
set Me, given by
t (ai -b tu>i,... ,Oin + tUn) mod 1, (9.26)
are called quasi-periodic trajectories with frequency vector w = (wi,... ,w).
The vector oj is said to be nonresonant if {k,(J) ^ 0 for every A: Z \ {0}.

One can easily obtain the following result.


P rop osition 9.31. Let H be a completely integrable Hamiltonian. For each
level set Me diffeomorphic to a torus T , if u> is nonresonant, then each
trajectory in (9.26) is dense in the torus.

9.5. The K A M theorem

Although the completely integrable Hamiltonian systems are relatively rare,


there are still many examples of Hamiltonian systems that are obtained
from small perturbations of Hamiltonians of the form h(I), using action-
angle coordinates. A concrete example is given by the motion of the planets
in the solar system, which is close to the problem of motion without taking
into account the interactions between planets (but only between the Sun
and each planet). This last system is completely integrable, since it consists
of several two-body problems (see Examples 9.3 and 9.26).
In this context it is natural to introduce the following notion. We recall
that for a function 5 : R R we write g{e) = 0{e) if there exists a constant
C > 0 such that |^(e)| < C\e\ for every e R.
D efinition 9.32. A parameterized family of Hamiltonians He = He{9,I),
with {0, / ) T X G for some G C R , is said to be almost-integrable if it is
of the form
He{e,i) = h{i) + M e , i ) , with M e , i ) = o{e), (9.27)
for every e 6 R in some neighborhood of zero.

Let He be an almost-integrable family of Hamiltonians. The Hamilton


ian system defined by Hg is

9' = c ( / ) + ^ ( , / ) ,
240 9. Hamiltonian Systems

where u>{I) Vh{I). For e = 0 all n-dimensional tori T / = T' x { / } with


I lG are invariant. Now we assume that
det dtjh ^ 0 for I G. (9.28)
This condition guarantees that the transformation 7 w(I) is a local dif-
feomorphism, and hence all tori T / can be (locally) parameterized by the
components of the vector w(7). We have the following fundamental result
(for details see, for example, [22]).

T h eorem 9.33 (KAM theorem). Let Hg be an analytic Hamiltonian in


T"' X G satisfying (9.27) and (9.28). Given r > n 1 and 7 > 0, there exists
eo = 0 (7 ^) such that if jej < Sq, then for each torus T / = x { / } with
7
|(A:,a;(7))|> k e r ^ \ { 0}
iiA;ir
there exists an invariant torus of the Hamiltonian Hg, with the same fre
quency vector u}{I), such that the maximal distance from Tj to T / is 0{e/'y).
Moreover, the volume of the complement m T x G of the set covered by the
tori Tj is 0 (7 ).

Theorem 9.33 says that under the nondegeneracy condition in (9.28),


if e is sufficiently small, then many of the n-dimensional invariant tori are
preserved in the perturbed system, up to a small deformation. When n = 2
one can deduce from Theorem 9.33 that the perturbed system has bounded
trajectories, because the tori (which in this case are closed curves) separate
the phase space (by Jordans curve theorem). For n > 3 , in general this
property may not hold.

9.6. Exercises

Exercise 9.1. Show that a differentiable transformation ip is canonical if


and only if
{ F o G o ^ } = { F , G] o p
for any differentiable functions F and G.

E xercise 9 .2 . For the flow pt defined by a Hamiltonian system, use Liou-


villes formula (Theorem 2.10) to show that deld^pt = 1

E xercise 9.3. Show that:


a) if a matrix B is Hamiltonian, then tr F = 0;
b) if a matrix A is canonical, then det A = 1;
c) for n = 1 a matrix B is Hamiltonian if and only if tr F = 0;
d) fo rn = 1 a matrix A is canonical if and only if det A = 1 .
9.6. Exercises 241

E xercise 9.4. Verify that the equation x" + sin a; = 0 is a Hamiltonian


system with one degree of freedom.

E xercise 9.5. Let H : R be a Hamiltonian of class such that the


corresponding Hamiltonian system defines a flow cpt in R^"^. Show that:
a) i f is constant along the solutions;
b) (ft preserves volume in R^ , that is, ^i{ipt{A)) = fi{A) for every open
set A C R^ and t G R, where /x denotes the volume in R^ ;
c) there exist neither asymptotically stable critical points nor asymp
totically stable periodic solutions.

Exercise 9.6. For a function / : R R of class C^, assume that the


equation x' = f (x) defines a flow (ft preserving volume in R and that there
exists a bounded set .4 C R*^ such that <Pt{A) = A for every t G R.
a) Given a nonempty open set U <Z A, show that not all sets
P2(U), (P3{U),.. . are pairwise disjoint.
b) Show that for each nonempty open setU <Z A there exist x U and
n G N such that ip2n(x) G U.
Bibliography

[1] R. Abraham and J. Marsden, Foundations o f Mechanics^ Benjamin/Cummings, 1978.


[2] H. Amann, Ordinary Differential Equations, Walter de Gruyter, 1990.
[3] V. Arnold, Mathematical Methods o f Classical Mechanics, Springer, 1989.
[4] V. Arnold, Ordinary Differential Equations, Springer, 1992.
[5] R. Bhatia, Matrix Analysis, Springer, 1997.
[6] J. Carr, Applications o f Centre Manifold Theory, Springer, 1981.
[7] C. Chicone, Ordinary Differential Equations with Applications, Springer, 1999.
[8] S.-N. Chow and J. Hale, Methods o f Bifurcation Theory, Springer, 1982.
[9] E. Coddington and N. Levinson, Theory o f Ordinary Differential Equations, McGraw-
Hill, 1955.
[10] T. Dieck, Algebraic Topology, European Mathematical Society, 2008.
[11] B. Dubrovin, A. Fomenko and S. Novikov, M odem Geometry - Methods and Appli
cations: The Geometry and Topology o f Manifolds, Springer, 1985.
[12] J. Guckenheimer and P. Holmes, Nonlinear Oscillations, Dynamical Systems, and
Bifurcations of Vector Fields, Springer, 1983.
[13] J. Hale, Ordinary Differential Equations, Robert E. Krieger, 1980.
[14] J. Hale and H. Kogak, Dynamics and Bifurcations, Springer, 1991.
[15] P. Hartman, Ordinary Differential Equations, SIAM, 2002.
[16] N. Higham, Functions o f Matrices: Theory and Computation, SIAM, 2008.
[17] M. Hirsch and S. Smale, Differential Equations, Dynamical Systems, and Linear Al
gebra, Academic Press, 1974.
[18] M. Irwin, Smooth Dynamical Systems, Academic Press, 1980.
[19] A. Katok and B. Hasselblatt, Introduction to the M odem Theory o f Dynamical Sys
tems, Cambridge University Press, 1995.
[20] J. La Salle and S. Lefschetz, Stability by Liapunovs Direct Method, With Applications,
Academic Press, 1961.
[21] J. Milnor, Morse Theory, Princeton University Press, 1963.

243
244 Bibliography

[22] J. Moser, Stable and Random Motions in Dynamical Systems^ Princeton University
Press, 1973.
[23] J. Palis and W . de Melo, Geometric Theory of Dynamical Systemsj Springer, 1982.
[24] D. Sanchez, Ordinary Differential Equations and Stability Theory^ Dover, 1968.
[25] F. Verhulst, Nonlinear Differential Equations and Dynamical Systems, Springer, 1996.
Index

a-limit set, 186 multiplier, 80


u;-limit set, 186 chart, 48
closed
action-angle coordinates, 235 ball, 13
almost-integrable Hamiltonian, 239 path, 171
Arzela-Ascoli theorem, 32 compact set, 10
asymptotic stability, 107 complete metric space, 14
asymptotically stable completely integrable Hamiltonian, 236
equation, 110 conjugacy
solution, 107 differentiable - , 87
attracting point, 17 linear - , 87, 88
autonomous equation, 8 topological - , 87, 90, 127
connected
ball component, 176
closed - , 13 set, 176
open - , 13
conservative equation, 42
bifurcation, 203
contraction, 12, 16
H o p f - 206
fiber - , 18
pitchfork - , 205
convergent sequence, 14
saddle-node - , 204
coordinate system, 48
theory, 201
critical point, 35
transcritical - , 204
hyperbolic - , 127, 136
Brouwers fixed point theorem, 178
isolated -,1 7 9
critical value, 50
canonical
curve, 50, 171
matrix, 229
transformation, 227
Cauchy sequence, 14 degree o f freedom, 225
center, 74 differentiable
manifold, 201, 206, 208 conjugacy, 87
manifold theorem, 208 map, 50
space, 207 structure, 48
characteristic differential equation, 3
exponent, 80, 232 on manifold, 48

245
246 Index

disconnected set, 176 Hamiltonian, 225


distance, 12 almost-integrable - , 239
completely integrable - , 236
equation matrix, 229
asymptotically stable - , 110 system, 225
autonomous - , 8 heteroclinic orbit, 37
conservative - , 42 homoclinic orbit, 37
homogeneous - , 76 homogeneous equation, 76
homological - ,2 1 7 homological equation, 217
linear - , 57 homotopic paths, 174
linear variational - , 60 homotopy, 174
nonautonomous - , 57 Hopf bifurcation, 206
nonhomogeneous 76 hyperbolic
stable - ,1 1 0 critical point, 127, 136
unstable - ,1 1 0 matrix, 90
equilibrium, 231 hyperbolicity, 127
nondegenerate - , 231
existence of solutions, 9, 33 independent functions, 236
exponential, 63 index, 171, 172, 180
theory, 171
fiber, 18
initial
contraction, 18
condition, 6
contraction theorem, 18
value problem, 3, 6
first integral, 42
instability criterion, 121
fixed point, 16
integrability, 235
Floquets theorem, 79
integral, 42
flow, 8
invariance o f domain theorem, 136
box theorem, 36
invariant
focus
manifold, 147, 150
stable - , 74
set, 185
unstable - , 74
isolated critical point, 179
formula
Liouvilles - , 62 Jordan canonical form, 64
variation of parameters - , 75, 76 Jordans curve theorem, 176
frequency vector, 239
function K A M theorem, 239, 240
Lipschitz - , 15
locally Lipschitz - , 10, 116 lemma
Lyapunov - , 105, 116, 117 Gronwalls - , 21
periodic - , 78 Morses - , 231
strict Lyapunov -,1 1 7 line integral, 172
functions linear
in involution, 236 conjugacy, 87, 88
independent - , 236 equation, 57
fundamental Hamiltonian system, 229
solution, 59 variational equation, 60
theorem of algebra, 179 Liouvilles formula, 62
Liouville-Arnold theorem, 237
global solution, 37 Lipschitz function, 15
Grobman-Hartman theorem, 129, 136 locally Lipschitz function, 10, 116
Gronwalls lemma, 21 Lyapunov function, 105, 116, 117

Hadamard-Perron theorem, 149 manifold, 48


Index 247

center 208 quasi-periodic trajectory, 239


invariant 147, 150
stable 150, 208 regular path, 171
unstable - , 150, 208 resonant vector, 219
matrix
canonical - , 229 saddle point, 68
Hamiltonian - , 229 saddle-node bifurcation, 204
hyperbolic - , 90 semiorbit
monodromy - , 80 negative - , 186
maximal interval, 29, 30 positive - , 186
metric space, 12 sequence
monodromy matrix, 80 Cauchy - , 14
Morses lemma, 231 convergent - , 14
set
negative semiorbit, 186 a-limit - , 186
node a;-limit - , 186
stable - , 68, 69, 71 compact - , 10
unstable - , 68, 71 connected - ,1 7 6
nonautonomous equation, 57 disconnected -,1 7 6
nondegenerate equilibrium, 231 invariant - , 185
nonhomogeneous equation, 76 solution, 3, 4, 51
nonresonant vector, 239 asymptotically stable - , 107
norm, 13 fundamental - , 59
normal form, 215 global - , 37
stable - , 106
open ball, 13 unstable - , 106
orbit, 35, 186 solutions
heteroclinic - , 37 differentially conjugate - , 87
homoclinic - , 37 linearly conjugate - , 87
periodic - , 36 topologically conjugate - , 87
space
path center - , 207
closed - , 171 complete metric - , 14
regular - , 171 metric - , 12
Peanos theorem, 33 o f solutions, 57
periodic stable - , 128, 207
function, 78 unstable - , 128, 207
orbit, 36 stability, 105, 113
phase asymptotic - , 107
portrait, 35, 38, 67 criterion, 117
space, 35 theory, 105
Picard-Lindelof theorem, 10 stable
pitchfork bifurcation, 205 equation, 110
Poincar^Bendixson focus, 74
theorem, 190, 192 manifold, 150, 208
theory, 185 node, 68, 69, 71
point solution, 106
attracting - , 17 space, 128, 207
critical - , 35 strict Lyapunov function, 117
fixed 16
Poisson bracket, 229 tangent
positive semiorbit, 186 bundle, 50
248 Index

space, 50
vector, 50
theorem
Arzela-Ascoli - , 32
Brouwers fixed point - , 178
center manifold - , 208
fiber contraction - , 18
Floquets - , 79
fiow box - , 36
Grobman-Hartman - , 129, 136
Hadamard-Perron - , 149
invariance of domain - , 136
Jordans curve - , 176
KAM - , 239, 240
Liouville-Arnold - , 237
Peanos - , 33
Picard-L indelof- , 10
Poincare-Bendixson - , 190, 192
theory
of normal forms, 215
Poincare-Bendixson - , 185
topological conjugacy, 87, 90, 127
transcritical bifurcation, 204
transversal, 191

uniqueness of solutions, 9
unstable
equation, 110
focus, 74
manifold, 150, 208
node, 68, 71
solution, 106
space, 128, 207

variation of parameters formula, 75, 76


vector
field, 51
frequency - , 239
nonresonant - , 239
resonant -,2 1 9
Selected T itles in This Series
137 Luis B a rreira and C lau dia Vails, Ordinary Differential Equations, 2012
136 A rsh a k P etrosy an , H en rik S h ahgholian , an d N in a U raltseva, Regularity of Free
Boundaries in Obstacle-Type Problems, 2012
135 P ascal C h errier and A lb e r t M ilan i, Linear and Quasi-linear Evolution Equations in
Hilbert Spaces, 2012
134 J ea n -M a rie D e K on in ck an d F loria n L u ca, Analytic Number Theory, 2012
133 J effrey R a u ch , Hyperbolic Partial Differential Equations and Geometric Optics, 2012
132 T eren ce T ao, Topics in Random Matrix Theory, 2012
131 Ian M . M u sson , Lie Superalgebras and Enveloping Algebreis, 2012
130 V iv ia n a E ne and J u rgen H erzog , Grobner Bsises in Commutative Algebra, 2011
129 S tuart P. H astings and J. B r y c e M c L e o d , Classical Methods in Ordinary Differential
Equations, 2012
128 J. M . L an d sberg, Tensors: Geometry and Applications, 2012
127 J effrey S trom , Modern Classical Homotopy Theory, 2011
126 T eren ce T ao, An Introduction to Measure Theory, 2011
125 D ro r V a rolin , Riemann Surfaces by Way of Complex Analytic Geometry, 2011
124 D a v id A . C ox , J oh n B . L ittle, an d H en ry K . Sch enck, Toric Varieties, 2011
123 G re g o ry Eskin, Lectures on Linear Partial Differential Equations, 2011
122 T eresa C re sp o and Z b ig n iew H a jto , Algebraic Groups and Differential Galois Theory,
2011
121 T obias H oick C o ld in g and W illia m P. M in ico z z i, II, A Course in Minimal Surfaces,
2011
120 Q in g H an , A Basic Course in Partial Differential Equations, 2011
119 A le x a n d e r K o ro s te le v and O lga K orostelev a , Mathematical Statistics, 2011
118 H al L. Sm ith and H orst R . T h ie m e , Dynamical Systems and Population Persistence,
2011
117 T eren ce T ao, An Epsilon of Room, I: Real Analysis, 2010
116 Joa n C erd^ , Linear Functional Analysis, 2010
115 J u lio G on za lez-D fa z, Ig n a cio G a rcia -J u ra d o, and M . G loria F iestras-Jan eiro,
An Introductory Course on Mathematical Game Theory, 2010
114 J osep h J. R o tm a n , Advanced Modern Algebra, 2010
113 T h om a s M . L ig gett, Continuous Time Markov Processes, 2010
112 Fredi T roltzsch , Optimal Control of Partial Differential Equations, 2010
111 S im on B ren d le, Ricci Flow and the Sphere Theorem, 2010
110 M atth ias K reck , Differential Algebraic Topology, 2010
109 J oh n C . N eu , Training Manual on Transport and Fluids, 2010
108 E n riqu e O u terelo and Jesus M . R u iz, Mapping Degree Theory, 2009
107 J effrey M . L ee, Manifolds and Differential Geometry, 2009
106 R o b e r t J. D averm an and G era rd A . V en em a , Embeddings in Manifolds, 2009
105 G iovan ni L eon i, A First Course in Sobolev Spaces, 2009
104 P a o lo A lu fii, Algebra: Chapter 0, 2009
103 B ran k o G riin b au m , Configurations of Points and Lines, 2009
102 M a rk A . P in sk y , Introduction to Fourier Analysis and Wavelets, 2002
101 W a rd C h en ey and W ill L ight, A Course in Approximation Theory, 2000
100 I. M a rtin Isaacs, Algebra, 1994
99 G era ld T eschl, Mathematical Methods in Quantum Mechanics, 2009
98 A le x a n d e r I. B o b e n k o and Y u ri B . Suris, Discrete Differential Geometry, 2008
97 D av id C. U llrich, Complex Made Simple, 2008
96 N . V . K ry lo v , Lectures on Elliptic and Parabolic Equations in Sobolev Spaces, 2008
95 L eon A . Takh tajan, Quantum Mechanics for Mathematicians, 2008
SELECTED TITLES IN THIS SERIES

94 Jam es E. H u m p h rey s, Representations of Semisimple Lie Algebras in the BGG


Category 0, 2008
93 P eter W . M ich o r, Topics in Differential Geometry, 2008
92 I. M a rtin Isaacs, Finite Group Theory, 2008
91 L ou is H alle R ow en , Graduate Algebra: Noncommutative View, 2008
90 L arry J. G erstein , Basic Quadratic Forms, 2008
89 A n th o n y B o n a to , A Course on the Web Graph, 2008
88 N ath an ial P. B row n an d N aru taka O zaw a, C*-Algebras and Finite-Dimensional
Approximations, 2008
87 Srikanth B . Iyengar, G ra h a m J. L euschke, A n to n L eykin, C lau d ia M iller, E zra
M iller, A n u ra g K . Singh, an d U li W a lth er, Twenty-Four Hours of Local
Cohomology, 2007
86 Y ulij Ilyashenko and Sergei Y akovenko, Lectures on Analytic Differential Equations,
2008
85 J oh n M . A lo n g ! and G ail S. N elson , Recurrence and Topology, 2007
84 C h aralam b os D . A lip ra n tis and R a b e e T ou rk y , Cones and Duality, 2007
83 W olfg a n g E belin g, Functions of Several Complex Variables and Their Singularities, 2007
82 Serge A lin h a c and P a trick G era rd , Pseudo-differential Operators and the
Nash-Moser Theorem, 2007
81 V . V . P ra solov , Elements of Homology Theory, 2007
80 D avar K h osh n ev isa n , Probability, 2007
79 W illia m Stein, Modular Forms, a Computational Approach, 2007
78 H arry D y m , Linear Algebra in Action, 2007
77 B en n ett C h ow , P en g Lu, and Lei N i, Hamiltons Ricci Flow, 2006
76 M ich ael E. T aylor, Measure Theory and Integration, 2006
75 P eter D . M iller, Applied Asymptotic Analysis, 2006
74 V . V . P ra solov , Elements of Combinatorial and Differential Topology, 2006
73 Louis H alle R ow en , Graduate Algebra: Commutative View, 2006
72 R . J. W illia m s, Introduction to the Mathematics of Finance, 2006
71 S. P. N ov ik ov and I. A . T aim a n ov, Modern Geometric Structures and Fields, 2006
70 Sean D in een , Probability Theory in Finance, 2005
69 Sebastian M on tiel an d A n to n io R o s , Curves and Surfaces, 2005
68 Luis C affarelli and S a n d ro Salsa, A Geometric Approach to Free Boundary Problems,
2005
67 T . Y . Lam , Introduction to Quadratic Forms over Fields, 2005
66 Y uli E idelm an, V ita l! M ilm a n , and A n ton is T so lo m itis, Functional Analysis, 2004
65 S. R a m a n a n , Global Calculus, 2005
64 A . A . K irillov , Lectures on the Orbit Method, 2004
63 S teven D ale C u tk osk y, Resolution of Singularities, 2004
62 T . W . K orn er, A Companion to Analysis, 2004
61 T h om a s A . Iv ey and J. M . L an d sberg, Cartan for Beginners, 2003
60 A lb e r to C a n d el and L aw ren ce C on lon , Foliations II, 2003
59 Steven H . W ein tra u b , Representation Theory of Finite Groups: Algebra and
Arithmetic, 2003
58 C e d ric V illan i, Topics in Optimal Transportation, 2003
57 R o b e r t P la to, Concise Numerical Mathematics, 2003

For a complete list of titles in this series, visit the


AMS Bookstore at w w w .a m s .o r g /b o o k s t o r e /.
' A 'V

This textbook provides a comprehensive introduction to the qualitative theory


of ordinary differential equations. It includes a discussion of the existence and
uniqueness of solutions, phase portraits, linear equations, stability theory, hyperbo-
licity and equations in the plane.The emphasis is primarily on results and methods
that allow one to analyze qualitative properties of the solutions without solving
the equations explicitly. The text includes numerous examples that illustrate in
detail the new concepts and results as well as exercises at the end of each chapter.
The book is also intended to serve as a bridge to important topics that are often
left out of a course on ordinary differential equations. In particular, it provides
brief introductions to bifurcation theory, center manifolds, normal forms and
Hamiltonian systems.

IS B N '9 7 8 - 0 - 8 2 1 8 - 8 7 4 9 - 3